More work to disentangle bset.c from the rest of the code:
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

More disentangling bset.c from the rest of the bcache code - soon, the
sorting routines won't have any dependencies on any outside structs.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Only use extent comparison for comparing extents, so we're not using
START_KEY() on other key types (i.e. btree pointers)
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

More refactoring:

node() -> bset_bkey_idx()
end() -> bset_bkey_last()
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Getting away from KEY_PTRS and moving toward KEY_U64s - and getting rid of magic
2s

Also - split out the part that checks against journal entry size so as to avoid
a dependancy on struct cache_set in bset.c
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

In the process of disentagling/libraryizing bset.c from the rest of the
bcache code.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

It was a single element mempool before, it's slightly cleaner to just use a real
mempool.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Used this fixed code to find and fix the bug fixed by
a4d885097b0ac0cd1337f171f2d4b83e946094d4.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

That was a terrible name for a macro, add some better helpers to replace it.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Now that we've got code for raid5/6 stripe awareness, bcache just needs
to know about the stripes and when writing partial stripes is expensive
- we probably don't want to enable this optimization for raid1 or 10,
even though they have stripes. So add a flag to queue_limits.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

This error path shouldn't have been hit in practice.. and we've got reworked
reserve code coming soon so that it shouldn't _ever_ be bit... but if we've got
code for this error path it should be correct.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

We need a reserve for allocating buckets for new btree nodes - and now that
we've got multiple btrees, it really needs to be per btree.

This reworks the reserves so we've got separate freelists for each reserve
instead of watermarks, which seems to make things a bit cleaner, and it adds
some code so that btree_split() can make sure the reserve is available before it
starts.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Also flesh out the documentation a bit
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Another minor performance optimization
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Unnecessary since a bucket that has dirty pointers pointing to it can
never be invalidated - and skipping it is a measurable performance
boost, since the bucket gen will usually be a cache miss.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Signed-off-by: NKent Overstreet <kmo@daterainc.com>

We were unnecessarily waiting on a journal write to complete when we just needed
to start a journal write and start setting up the next one.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

The real fix is where we check the bytes we need against how much is
remaining - we also need to check for a journal entry bigger than our
buffer, we'll never write those and it would be bad if we tried to read
one.

Also improve the diagnostic messages.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

The code that handles overlapping extents that we've just read back in from disk
was depending on the behaviour of the code that handles overlapping extents as
we're inserting into a btree node in the case of an insert that forced an
existing extent to be split: on insert, if we had to split we'd also insert a
new extent to represent the top part of the old extent - and then that new
extent would get written out.

The code that read the extents back in thus not bother with splitting extents -
if it saw an extent that ovelapped in the middle of an older extent, it would
trim the old extent to only represent the bottom part, assuming that the
original insert would've inserted a new extent to represent the top part.

I still haven't figured out _how_ it can happen, but I'm now pretty convinced
(and testing has confirmed) that there's some kind of an obscure corner case
(probably involving extent merging, and multiple overwrites in different sets)
that breaks this. The fix is to change the mergesort fixup code to split extents
itself when required.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>
Cc: linux-stable <stable@vger.kernel.org> # >= v3.10

The old writeback PD controller could get into states where it had throttled all
the way down and take way too long to recover - it was too complicated to really
understand what it was doing.

This rewrites a good chunk of it to hopefully be simpler and make more sense,
and it also pays more attention to units which should make the behaviour a bit
easier to understand.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

There is a possibility for a bucket to be invalidated by the allocator
while moving_gc was copying it's contents to another bucket, if the
bucket only held cached data. To prevent this moving checks for
a stale ptr (to an invalidated bucket), before and after reads.
It it finds one, it simply ignores moving that data. This only
affects bcache if the moving_gc was turned on, note that it's
off by default.
Signed-off-by: NNicholas Swenson <nks@daterainc.com>
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Garbage collector needs to check keys in the writeback keybuf to
make sure it's not invalidating buckets to which the writeback
keys point to.
Signed-off-by: NNicholas Swenson <nks@daterainc.com>
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Removed gc_move_threshold because picking buckets only by
threshold could lead moving extra buckets (ei. if there are
buckets at the threshold that aren't supposed to be moved
do to space considerations).

This is replaced by a GC_MOVE bit in the gc_mark bitmask.
Now only marked buckets get moved.
Signed-off-by: NNicholas Swenson <nks@daterainc.com>
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Signed-off-by: NNicholas Swenson <nks@daterainc.com>
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Signed-off-by: NNicholas Swenson <nks@daterainc.com>
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Signed-off-by: NNicholas Swenson <nks@daterainc.com>
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Dirty data accounting wasn't quite right - firstly, we were adding the key we're
inserting after it could have merged with another dirty key already in the
btree, and secondly we could sometimes pass the wrong offset to
bcache_dev_sectors_dirty_add() for dirty data we were overwriting - which is
important when tracking dirty data by stripe.

NOTE FOR BACKPORTERS: For 3.10 (and 3.11?) there's other accounting fixes
necessary that got squashed in with other patches; the full patch against 3.10
is 408cc2f47eeac93a, available at:
  git://evilpiepirate.org/~kent/linux-bcache.git bcache-3.10-writeback-fixes
Signed-off-by: NKent Overstreet <kmo@daterainc.com>
Cc: linux-stable <stable@vger.kernel.org> # >= v3.10

diff --git a/drivers/md/bcache/btree.c b/drivers/md/bcache/btree.c
index 2a46036..4a12b2f 100644
--- a/drivers/md/bcache/btree.c
+++ b/drivers/md/bcache/btree.c
@@ -1817,7 +1817,8 @@ static bool fix_overlapping_extents(struct btree *b, struct bkey *insert,
 			if (KEY_START(k) > KEY_START(insert) + sectors_found)
 				goto check_failed;

-			if (KEY_PTRS(replace_key) != KEY_PTRS(k))
+			if (KEY_PTRS(k) != KEY_PTRS(replace_key) ||
+			    KEY_DIRTY(k) != KEY_DIRTY(replace_key))
 				goto check_failed;

 			/* skip past gen */

We're just waiting on kthread_should_stop(), nothing else, so
interruptible sleep was wrong here.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

at the beginning (schedule_timout_interuptible) and others
do his on their own

This prevents wrong load average calculation (load of 1 per thread)
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

An old array block could have its reference count decremented below
zero when it is being replaced in the btree by a new array block.

The fix is to increment the old ablock's reference count just before
inserting a new ablock into the btree.
Signed-off-by: NJoe Thornber <ejt@redhat.com>
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Cc: stable@vger.kernel.org # 3.9+

The old behaviour, returning -EINVAL if a ref_count of 0 would be
decremented, was removed in commit f722063e ("dm space map: optimise
sm_ll_dec and sm_ll_inc").  To fix this regression we return an error
code from the mutator function pointer passed to sm_ll_mutate() and have
dec_ref_count() return -EINVAL if the old ref_count is 0.

Add a DMERR to reflect the potential seriousness of this error.

Also, add missing dm_tm_unlock() to sm_ll_mutate()'s error path.

With this fix the following dmts regression test now passes:
 dmtest run --suite cache -n /metadata_use_kernel/

The next patch fixes the higher-level dm-array code that exposed this
regression.
Signed-off-by: NJoe Thornber <ejt@redhat.com>
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Cc: stable@vger.kernel.org # 3.12+

The module parameter stats_current_allocated_bytes in dm-mod is
read-only.  This parameter informs the user about memory
consumption.  It is not supposed to be changed by the user.

However, despite being read-only, this parameter can be set on
modprobe or insmod command line:
modprobe dm-mod stats_current_allocated_bytes=12345

The kernel doesn't expect that this variable can be non-zero at module
initialization and if the user sets it, it results in warning.

This patch initializes the variable in the module init routine, so
that user-supplied value is ignored.
Signed-off-by: NMikulas Patocka <mpatocka@redhat.com>
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Cc: stable@vger.kernel.org # 3.12+

Some module parameters in dm-bufio are read-only. These parameters
inform the user about memory consumption. They are not supposed to be
changed by the user.

However, despite being read-only, these parameters can be set on
modprobe or insmod command line, for example:
modprobe dm-bufio current_allocated_bytes=12345

The kernel doesn't expect that these variables can be non-zero at module
initialization and if the user sets them, it results in BUG.

This patch initializes the variables in the module init routine, so that
user-supplied values are ignored.
Signed-off-by: NMikulas Patocka <mpatocka@redhat.com>
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Cc: stable@vger.kernel.org # 3.2+

Commit f494a9c6 ("dm cache: cache
shrinking support") broke cache resizing support.

dm_cache_resize() is called with cache->cache_size before it gets
updated to new_size, so it is a no-op.  But the dm-cache superblock is
updated with the new_size even though the backing dm-array is not
resized.  Fix this by passing the new_size to dm_cache_resize().
Signed-off-by: NVincent Pelletier <plr.vincent@gmail.com>
Acked-by: NJoe Thornber <ejt@redhat.com>
Signed-off-by: NMike Snitzer <snitzer@redhat.com>

Micro benchmarks that repeatedly issued IO to a single block were
failing to cause a promotion from the origin device to the cache.  Fix
this by not updating the stats during map() if -EWOULDBLOCK will be
returned.

The mq policy will only update stats, consider migration, etc, once per
tick period (a unit of time established between dm-cache core and the
policies).

When the IO thread calls the policy's map method, if it would like to
migrate the associated block it returns -EWOULDBLOCK, the IO then gets
handed over to a worker thread which handles the migration.  The worker
thread calls map again, to check the migration is still needed (avoids a
race among other things).  *BUT*, before this fix, if we were still in
the same tick period the stats were already updated by the previous map
call -- so the migration would no longer be requested.
Signed-off-by: NJoe Thornber <ejt@redhat.com>
Signed-off-by: NMike Snitzer <snitzer@redhat.com>

A thin-pool may be in read-only mode because the pool's data or metadata
space was exhausted.  To allow for recovery, by adding more space to the
pool, we must allow a pool to transition from PM_READ_ONLY to PM_WRITE
mode.  Otherwise, running out of space will render the pool permanently
read-only.
Signed-off-by: NJoe Thornber <ejt@redhat.com>
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Cc: stable@vger.kernel.org

If the thin-pool transitioned to fail mode and the thin-pool's table
were reloaded for some reason: the new table's default pool mode would
be read-write, though it will transition to fail mode during resume.

When the pool mode transitions directly from PM_WRITE to PM_FAIL we need
to re-establish the intermediate read-only state in both the metadata
and persistent-data block manager (as is usually done with the normal
pool mode transition sequence: PM_WRITE -> PM_READ_ONLY -> PM_FAIL).
Signed-off-by: NJoe Thornber <ejt@redhat.com>
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Cc: stable@vger.kernel.org

Rename commit_or_fallback() to commit().  Now all previous calls to
commit() will trigger the pool mode to fallback if the commit fails.

Also, check the error returned from commit() in alloc_data_block().
Signed-off-by: NJoe Thornber <ejt@redhat.com>
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Cc: stable@vger.kernel.org