Remove dm_get_mapinfo() because no target uses it. Targets can allocate
per-bio data using ti->per_bio_data_size, this is much more flexible
than union map_info.
Leave union map_info only for the request-based multipath target's use.
Also delete the unused "unsigned long long ll" field of union map_info.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Make the function dm_table_alloc_md_mempools static because it is not
called from another file.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
In theory copying the space map root can fail, but in practice it never
does because we're careful to check what size buffer is needed.
But make certain we're able to copy the space map roots before
locking the superblock.
Signed-off-by: Joe Thornber <ejt@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Cc: stable@vger.kernel.org # drop dm-era and dm-cache changes as needed
The persistent-data library used by dm-thin, dm-cache, etc is
transactional. If anything goes wrong, such as an io error when writing
new metadata or a power failure, then we roll back to the last
transaction.
Atomicity when committing a transaction is achieved by:
a) Never overwriting data from the previous transaction.
b) Writing the superblock last, after all other metadata has hit the
disk.
This commit and the following commit ("dm: take care to copy the space
map roots before locking the superblock") fix a bug associated with (b).
When committing it was possible for the superblock to still be written
in spite of an io error occurring during the preceeding metadata flush.
With these commits we're careful not to take the write lock out on the
superblock until after the metadata flush has completed.
Change the transaction manager's semantics for dm_tm_commit() to assume
all data has been flushed _before_ the single superblock that is passed
in.
As a prerequisite, split the block manager's block unlocking and
flushing by simplifying dm_bm_flush_and_unlock() to dm_bm_flush(). Now
the unlocking must be done separately.
This issue was discovered by forcing io errors at the crucial time
using dm-flakey.
Signed-off-by: Joe Thornber <ejt@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Cc: stable@vger.kernel.org
Discard block size not being equal to cache block size causes data
corruption by erroneously avoiding migrations in issue_copy() because
the discard state is being cleared for a group of cache blocks when it
should not.
Completely remove all code that enabled a distinction between the
cache block size and discard block size.
Signed-off-by: Heinz Mauelshagen <heinzm@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
If the discard block size is larger than the cache block size we will
not properly quiesce IO to a region that is about to be discarded. This
results in a race between a cache migration where no copy is needed, and
a write to an adjacent cache block that's within the same large discard
block.
Workaround this by limiting the discard_block_size to cache_block_size.
Also limit the max_discard_sectors to cache_block_size.
A more comprehensive fix that introduces range locking support in the
bio_prison and proper quiescing of a discard range that spans multiple
cache blocks is already in development.
Reported-by: Morgan Mears <Morgan.Mears@netapp.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Acked-by: Joe Thornber <ejt@redhat.com>
Acked-by: Heinz Mauelshagen <heinzm@redhat.com>
Cc: stable@vger.kernel.org
dm-era is a target that behaves similar to the linear target. In
addition it keeps track of which blocks were written within a user
defined period of time called an 'era'. Each era target instance
maintains the current era as a monotonically increasing 32-bit
counter.
Use cases include tracking changed blocks for backup software, and
partially invalidating the contents of a cache to restore cache
coherency after rolling back a vendor snapshot.
dm-era is primarily expected to be paired with the dm-cache target.
Signed-off-by: Joe Thornber <ejt@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Uninlined nested functions can cause crashes when using ftrace, as they don't
follow the normal calling convention and confuse the ftrace function graph
tracer as it examines the stack.
Also, nested functions are supported as a gcc extension, but may fail on other
compilers (e.g. llvm).
Signed-off-by: John Sheu <john.sheu@gmail.com>
gc_gen was a temporary used to recalculate last_gc, but since we only need
bucket->last_gc when gc isn't running (gc_mark_valid = 1), we can just update
last_gc directly.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
This was originally added as at optimization that for various reasons isn't
needed anymore, but it does add a lot of nasty corner cases (and it was
responsible for some recently fixed bugs). Just get rid of it now.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
This changes the bucket allocation reserves to use _real_ reserves - separate
freelists - instead of watermarks, which if nothing else makes the current code
saner to reason about and is going to be important in the future when we add
support for multiple btrees.
It also adds btree_check_reserve(), which checks (and locks) the reserves for
both bucket allocation and memory allocation for btree nodes; the old code just
kinda sorta assumed that since (e.g. for btree node splits) it had the root
locked and that meant no other threads could try to make use of the same
reserve; this technically should have been ok for memory allocation (we should
always have a reserve for memory allocation (the btree node cache is used as a
reserve and we preallocate it)), but multiple btrees will mean that locking the
root won't be sufficient anymore, and for the bucket allocation reserve it was
technically possible for the old code to deadlock.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
With the locking rework in the last patch, this shouldn't be needed anymore -
btree_node_write_work() only takes b->write_lock which is never held for very
long.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
Add a new lock, b->write_lock, which is required to actually modify - or write -
a btree node; this lock is only held for short durations.
This means we can write out a btree node without taking b->lock, which _is_ held
for long durations - solving a deadlock when btree_flush_write() (from the
journalling code) is called with a btree node locked.
Right now just occurs in bch_btree_set_root(), but with an upcoming journalling
rework is going to happen a lot more.
This also turns b->lock is now more of a read/intent lock instead of a
read/write lock - but not completely, since it still blocks readers. May turn it
into a real intent lock at some point in the future.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
This isn't a bulletproof fix; btree_node_free() -> bch_bucket_free() puts the
bucket on the unused freelist, where it can be reused right away without any
ordering requirements. It would be better to wait on at least a journal write to
go down before reusing the bucket. bch_btree_set_root() does this, and inserting
into non leaf nodes is completely synchronous so we should be ok, but future
patches are just going to get rid of the unused freelist - it was needed in the
past for various reasons but shouldn't be anymore.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
This means the garbage collection code can better check for data and metadata
pointers to the same buckets.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
This will potentially save us an allocation when we've got inode/dirent bkeys
that don't fit in the keylist's inline keys.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
Change the invalidate tracepoint to indicate how much data we're invalidating,
and change the alloc tracepoints to indicate what offset they're for.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
Deadlock happened because a foreground write slept, waiting for a bucket
to be allocated. Normally the gc would mark buckets available for invalidation.
But the moving_gc was stuck waiting for outstanding writes to complete.
These writes used the bcache_wq, the same queue foreground writes used.
This fix gives moving_gc its own work queue, so it was still finish moving
even if foreground writes are stuck waiting for allocation. It also makes
work queue a parameter to the data_insert path, so moving_gc can use its
workqueue for writes.
Signed-off-by: Nicholas Swenson <nks@daterainc.com>
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
The on disk bucket gens are allowed to be out of date, when we reuse buckets
that didn't have any live data in them. To deal with this, the initial gc has to
update the bucket gen when we find a pointer gen newer than the bucket's gen.
Unfortunately we weren't doing this for pointers in the journal that we're about
to replay.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
On recovery we weren't correctly keeping track of what journal buckets had open
journal entries, thus it was possible for them to be overwritten until we'd
written all new journal entries.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
In order to avoid wasting cache space a partial block at the end of the
origin device is not cached. Unfortunately, the check for such a
partial block at the end of the origin device was flawed.
Fix accesses beyond the end of the origin device that occured due to
attempted promotion of an undetected partial block by:
- initializing the per bio data struct to allow cache_end_io to work properly
- recognizing access to the partial block at the end of the origin device
- avoiding out of bounds access to the discard bitset
Otherwise, users can experience errors like the following:
attempt to access beyond end of device
dm-5: rw=0, want=20971520, limit=20971456
...
device-mapper: cache: promotion failed; couldn't copy block
Signed-off-by: Heinz Mauelshagen <heinzm@redhat.com>
Acked-by: Joe Thornber <ejt@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Cc: stable@vger.kernel.org
During demotion or promotion to a cache's >2TB fast device we must not
truncate the cache block's associated sector to 32bits. The 32bit
temporary result of from_cblock() caused a 32bit multiplication when
calculating the sector of the fast device in issue_copy_real().
Use an intermediate 64bit type to store the 32bit from_cblock() to allow
for proper 64bit multiplication.
Here is an example of how this bug manifests on an ext4 filesystem:
EXT4-fs error (device dm-0): ext4_mb_generate_buddy:756: group 17136, 32768 clusters in bitmap, 30688 in gd; block bitmap corrupt.
JBD2: Spotted dirty metadata buffer (dev = dm-0, blocknr = 0). There's a risk of filesystem corruption in case of system crash.
Signed-off-by: Heinz Mauelshagen <heinzm@redhat.com>
Acked-by: Joe Thornber <ejt@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Cc: stable@vger.kernel.org
This has been a relatively long-standing issue that wasn't nailed down
until Teng-Feng Yang's meticulous bug report to dm-devel on 3/7/2014,
see: http://www.redhat.com/archives/dm-devel/2014-March/msg00021.html
From that report:
"When decreasing the reference count of a metadata block with its
reference count equals 3, we will call dm_btree_remove() to remove
this enrty from the B+tree which keeps the reference count info in
metadata device.
The B+tree will try to rebalance the entry of the child nodes in each
node it traversed, and the rebalance process contains the following
steps.
(1) Finding the corresponding children in current node (shadow_current(s))
(2) Shadow the children block (issue BOP_INC)
(3) redistribute keys among children, and free children if necessary (issue BOP_DEC)
Since the update of a metadata block's reference count could be
recursive, we will stash these reference count update operations in
smm->uncommitted and then process them in a FILO fashion.
The problem is that step(3) could free the children which is created
in step(2), so the BOP_DEC issued in step(3) will be carried out
before the BOP_INC issued in step(2) since these BOPs will be
processed in FILO fashion. Once the BOP_DEC from step(3) tries to
decrease the reference count of newly shadow block, it will report
failure for its reference equals 0 before decreasing. It looks like we
can solve this issue by processing these BOPs in a FIFO fashion
instead of FILO."
Commit 5b564d80 ("dm space map: disallow decrementing a reference count
below zero") changed the code to report an error for this temporary
refcount decrement below zero. So what was previously a harmless
invalid refcount became a hard failure due to the new error path:
device-mapper: space map common: unable to decrement a reference count below 0
device-mapper: thin: 253:6: dm_thin_insert_block() failed: error = -22
device-mapper: thin: 253:6: switching pool to read-only mode
This bug is in dm persistent-data code that is common to the DM thin and
cache targets. So any users of those targets should apply this fix.
Fix this by applying recursive space map operations in FIFO order rather
than FILO.
Resolves: https://bugzilla.kernel.org/show_bug.cgi?id=68801
Reported-by: Apollon Oikonomopoulos <apoikos@debian.org>
Reported-by: edwillam1007@gmail.com
Reported-by: Teng-Feng Yang <shinrairis@gmail.com>
Signed-off-by: Joe Thornber <ejt@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Cc: stable@vger.kernel.org # 3.13+
i) by the time DM core calls the postsuspend hook the dm_noflush flag
has been cleared. So the old thin_postsuspend did nothing. We need to
use the presuspend hook instead.
ii) There was a race between bios leaving DM core and arriving in the
deferred queue.
thin_presuspend now sets a 'requeue' flag causing all bios destined for
that thin to be requeued back to DM core. Then it requeues all held IO,
and all IO on the deferred queue (destined for that thin). Finally
postsuspend clears the 'requeue' flag.
Signed-off-by: Joe Thornber <ejt@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
The spin lock in requeue_io() was held for too long, allowing deadlock.
Don't worry, due to other issues addressed in the following "dm thin:
fix noflush suspend IO queueing" commit, this code was never called.
Fix this by taking the spin lock for a much shorter period of time.
Signed-off-by: Joe Thornber <ejt@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Ideally a thin pool would never run out of data space; the low water
mark would trigger userland to extend the pool before we completely run
out of space. However, many small random IOs to unprovisioned space can
consume data space at an alarming rate. Adjust your low water mark if
you're frequently seeing "out-of-data-space" mode.
Before this fix, if data space ran out the pool would be put in
PM_READ_ONLY mode which also aborted the pool's current metadata
transaction (data loss for any changes in the transaction). This had a
side-effect of needlessly compromising data consistency. And retry of
queued unserviceable bios, once the data pool was resized, could
initiate changes to potentially inconsistent pool metadata.
Now when the pool's data space is exhausted transition to a new pool
mode (PM_OUT_OF_DATA_SPACE) that allows metadata to be changed but data
may not be allocated. This allows users to remove thin volumes or
discard data to recover data space.
The pool is no longer put in PM_READ_ONLY mode in response to the pool
running out of data space. And PM_READ_ONLY mode no longer aborts the
pool's current metadata transaction. Also, set_pool_mode() will now
notify userspace when the pool mode is changed.
Signed-off-by: Joe Thornber <ejt@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
If a thin metadata operation fails the current transaction will abort,
whereby causing potential for IO layers up the stack (e.g. filesystems)
to have data loss. As such, set THIN_METADATA_NEEDS_CHECK_FLAG in the
thin metadata's superblock which:
1) requires the user verify the thin metadata is consistent (e.g. use
thin_check, etc)
2) suggests the user verify the thin data is consistent (e.g. use fsck)
The only way to clear the superblock's THIN_METADATA_NEEDS_CHECK_FLAG is
to run thin_repair.
On metadata operation failure: abort current metadata transaction, set
pool in read-only mode, and now set the needs_check flag.
As part of this change, constraints are introduced or relaxed:
* don't allow a pool to transition to write mode if needs_check is set
* don't allow data or metadata space to be resized if needs_check is set
* if a thin pool's metadata space is exhausted: the kernel will now
force the user to take the pool offline for repair before the kernel
will allow the metadata space to be extended.
Also, update Documentation to include information about when the thin
provisioning target commits metadata, how it handles metadata failures
and running out of space.
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Joe Thornber <ejt@redhat.com>
Commit b5330655 ("dm thin: handle metadata failures more consistently")
increased potential for the pool's mode to be changed in response to
metadata operation failures.
When the pool mode is changed it isn't synchronized with the mode in
pool_features stored in the target's context (ti->private) that is used
as the basis for (re)establishing the pool mode during resume via
bind_control_target.
It is important that we synchronize the pool mode when it is changed
otherwise the pool may experience and unexpected mode transition on the
next resume (especially if there was no new table load).
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Acked-by: Joe Thornber <ejt@redhat.com>
Commit 55494bf2947dccdf2 ("dm snapshot: use dm-bufio") broke snapshots.
Before that 3.14-rc1 commit, loading a snapshot's list of exceptions
involved reading exception areas one by one into ps->area and inserting
those exceptions into the hash table. Commit 55494bf2947dccdf2 changed
it so that dm-bufio with prefetch is used to load exceptions in batchs.
Exceptions are loaded correctly, but ps->area is left uninitialized.
When a new exception is allocated, it is stored in this uninitialized
ps->area which will be written to the disk. This causes metadata
corruption.
Fix this corruption by copying the last area that was read via dm-bufio
into ps->area.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Since DM_DEBUG_BLOCK_STACK_TRACING is a DM_PERSISTENT_DATA config option
move it from drivers/md/Kconfig to drivers/md/persistent-data/Kconfig.
Doing so fixes indentation for other DM config options.
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
When remapping a block to the cache's fast device that is larger than
2TB we must not truncate the destination sector to 32bits. The 32bit
temporary result of from_cblock() was being overflowed in
remap_to_cache() due to the logical left shift.
Use an intermediate 64bit type to store the 32bit from_cblock() result
to fix the overflow.
Signed-off-by: Heinz Mauelshagen <heinzm@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Cc: stable@vger.kernel.org
It was always intended that a user could provide a thin metadata device
that is larger than the max supported by the on-disk format. The extra
space would just go unused.
Unfortunately that never worked. If the user attempted to use a larger
metadata device on creation they would get an error like the following:
device-mapper: space map common: space map too large
device-mapper: transaction manager: couldn't create metadata space map
device-mapper: thin metadata: tm_create_with_sm failed
device-mapper: table: 252:17: thin-pool: Error creating metadata object
device-mapper: ioctl: error adding target to table
Fix this by allowing the initial metadata space map creation to cap its
size at the max number of blocks supported (DM_SM_METADATA_MAX_BLOCKS).
get_metadata_dev_size() must also impose DM_SM_METADATA_MAX_BLOCKS (via
THIN_METADATA_MAX_SECTORS), otherwise extending metadata would cap at
THIN_METADATA_MAX_SECTORS_WARNING (which is larger than supported).
Also, the calculation for THIN_METADATA_MAX_SECTORS didn't account for
the sizeof the disk_bitmap_header. So the supported maximum metadata
size is a bit smaller (reduced from 33423360 to 33292800 sectors).
Lastly, remove the "excess space will not be used" warning message from
get_metadata_dev_size(); it resulted in printing the warning multiple
times. Factor out warn_if_metadata_device_too_big(), call it from
pool_ctr() and maybe_resize_metadata_dev().
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Acked-by: Joe Thornber <ejt@redhat.com>
An invalid ioctl will never be valid, irrespective of whether multipath
has active paths or not. So for invalid ioctls we do not have to wait
for multipath to activate any paths, but can rather return an error
code immediately. This fix resolves numerous instances of:
udevd[]: worker [] unexpectedly returned with status 0x0100
that have been seen during testing.
Signed-off-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Cc: stable@vger.kernel.org
The code was using sectors to count the number of sectors it was zeroing... but
then it passed it to bio_advance()... after it had been set to 0. Amusing...
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
dm_pool_close_thin_device() must be called if dm_set_target_max_io_len()
fails in thin_ctr(). Otherwise __pool_destroy() will fail because the
pool will still have an open thin device:
device-mapper: thin metadata: attempt to close pmd when 1 device(s) are still open
device-mapper: thin: __pool_destroy: dm_pool_metadata_close() failed.
Also, must establish error code if failing thin_ctr() because the pool
is in fail_io mode.
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Acked-by: Joe Thornber <ejt@redhat.com>
Cc: stable@vger.kernel.org
When restoring bi_end_io, increase bi_remaining before retrying the bio
to avoid BUG_ON(atomic_read(&bio->bi_remaining) <= 0) in bio_endio().
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
