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

The BUG_ON at the end of __bch_btree_mark_key can be triggered due to
an integer overflow error:

BITMASK(GC_SECTORS_USED, struct bucket, gc_mark, 2, 13);
...
SET_GC_SECTORS_USED(g, min_t(unsigned,
	     GC_SECTORS_USED(g) + KEY_SIZE(k),
	     (1 << 14) - 1));
BUG_ON(!GC_SECTORS_USED(g));

In bcache.h, the SECTORS_USED bitfield is defined to be 13 bits wide.
While the SET_ code tries to ensure that the field doesn't overflow by
clamping it to (1<<14)-1 == 16383, this is incorrect because 16383
requires 14 bits.  Therefore, if GC_SECTORS_USED() + KEY_SIZE() =
8192, the SET_ statement tries to store 8192 into a 13-bit field.  In
a 13-bit field, 8192 becomes zero, thus triggering the BUG_ON.

Therefore, create a field width constant and a max value constant, and
use those to create the bitfield and check the inputs to
SET_GC_SECTORS_USED.  Arguably the BITMASK() template ought to have
BUG_ON checks for too-large values, but that's a separate patch.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>

When deciding what order to reuse buckets we take into account both the bucket's
priority (which indicates lru order) and also the amount of live data in that
bucket. The way they were scaled together wasn't as correct as it could be...
this patch improves and documents it.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Soon, bset.c won't need to depend on struct btree.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

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>

More refactoring:

node() -> bset_bkey_idx()
end() -> bset_bkey_last()
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>

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>

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

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>

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>

The new bio_split() can split arbitrary bios - it's not restricted to
single page bios, like the old bio_split() (previously renamed to
bio_pair_split()). It also has different semantics - it doesn't allocate
a struct bio_pair, leaving it up to the caller to handle completions.

Then convert the existing bio_pair_split() users to the new bio_split()
- and also nvme, which was open coding bio splitting.

(We have to take that BUG_ON() out of bio_integrity_trim() because this
bio_split() needs to use it, and there's no reason it has to be used on
bios marked as cloned; BIO_CLONED doesn't seem to have clearly
documented semantics anyways.)
Signed-off-by: NKent Overstreet <kmo@daterainc.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Martin K. Petersen <martin.petersen@oracle.com>
Cc: Matthew Wilcox <matthew.r.wilcox@intel.com>
Cc: Keith Busch <keith.busch@intel.com>
Cc: Vishal Verma <vishal.l.verma@intel.com>
Cc: Jiri Kosina <jkosina@suse.cz>
Cc: Neil Brown <neilb@suse.de>

Bcache has a hack to avoid cloning the biovec if it's all full pages -
but with immutable biovecs coming this won't be necessary anymore.

For now, we remove the special case and always clone the bvec array so
that the immutable biovec patches are simpler.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

More testing ftw! Also, now verify mode doesn't break if you read dirty
data.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

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

The old scanning-by-stripe code burned too much CPU, this should be
better.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

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

It never really made sense to expose this, so just kill it.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Big garbage collection rewrite; now, garbage collection uses the same
mechanisms as used elsewhere for inserting/updating btree node pointers,
instead of rewriting interior btree nodes in place.

This makes the code significantly cleaner and less fragile, and means we
can now make garbage collection incremental - it doesn't have to hold a
write lock on the root of the btree for the entire duration of garbage
collection.

This means that there's less of a latency hit for doing garbage
collection, which means we can gc more frequently (and do a better job
of reclaiming from the cache), and we can coalesce across more btree
nodes (improving our space efficiency).
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Couple changes:
 * Consolidate bch_check_keys() and bch_check_key_order(), and move the
   checks that only check_key_order() could do to bch_btree_iter_next().

 * Get rid of CONFIG_BCACHE_EDEBUG - now, all that code is compiled in
   when CONFIG_BCACHE_DEBUG is enabled, and there's now a sysfs file to
   flip on the EDEBUG checks at runtime.

 * Dropped an old not terribly useful check in rw_unlock(), and
   refactored/improved a some of the other debug code.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Now, the on disk data structures are in a header that can be exported to
userspace - and having them all centralized is nice too.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Just reorganizing things a bit.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Lots of stuff has been open coding its own btree traversal - which is
generally pretty simple code, but there are a few subtleties.

This adds new new functions, bch_btree_map_nodes() and
bch_btree_map_keys(), which do the traversal for you. Everything that's
open coding btree traversal now (with the exception of garbage
collection) is slowly going to be converted to these two functions;
being able to write other code at a higher level of abstraction  is a
big improvement w.r.t. overall code quality.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

This simplifies the writeback flow control quite a bit - previously, it
was conceptually two coroutines, refill_dirty() and read_dirty(). This
makes the code quite a bit more straightforward.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

We needed a dedicated rescuer workqueue for gc anyways... and gc was
conceptually a dedicated thread, just one that wasn't running all the
time. Switch it to a dedicated thread to make the code a bit more
straightforward.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

At one point we did do fancy asynchronous waiting stuff with
bucket_wait, but that's all gone (and bucket_wait is used a lot less
than it used to be). So use the standard primitives.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

We never waited on c->try_wait asynchronously, so just use the standard
primitives.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

This was the main point of all this refactoring - now,
btree_insert_check_key() won't fail just because the leaf node happened
to be full.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Originally I got this right... except that the divides didn't use
do_div(), which broke 32 bit kernels. When I went to fix that, I forgot
that the raid stripe size usually isn't a power of two... doh
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Works kind of like the ext4 setting, to panic or remount read only on
errors.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

The old asynchronous discard code was really a relic from when all the
allocation code was asynchronous - now that allocation runs out of a
dedicated thread there's no point in keeping around all that complicated
machinery.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

Background writeback works by scanning the btree for dirty data and
adding those keys into a fixed size buffer, then for each dirty key in
the keybuf writing it to the backing device.

When read_dirty() finishes and it's time to scan for more dirty data, we
need to wait for the outstanding writeback IO to finish - they still
take up slots in the keybuf (so that foreground writes can check for
them to avoid races) - without that wait, we'll continually rescan when
we'll be able to add at most a key or two to the keybuf, and that takes
locks that starves foreground IO.  Doh.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>
Cc: linux-stable <stable@vger.kernel.org> # >= v3.10
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

The alloc kthread should've been using try_to_freeze() - and also there
was the potential for the alloc kthread to get woken up after it had
shut down, which would have been bad.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>

If we stopped a bcache device when we were already detaching (or
something like that), bcache_device_unlink() would try to remove a
symlink from sysfs that was already gone because the bcache dev kobject
had already been removed from sysfs.

So keep track of whether we've removed stuff from sysfs.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>
Cc: linux-stable <stable@vger.kernel.org> # >= v3.10

Now that we're tracking dirty data per stripe, we can add two
optimizations for raid5/6:

 * If a stripe is already dirty, force writes to that stripe to
   writeback mode - to help build up full stripes of dirty data

 * When flushing dirty data, preferentially write out full stripes first
   if there are any.
Signed-off-by: NKent Overstreet <koverstreet@google.com>

To make background writeback aware of raid5/6 stripes, we first need to
track the amount of dirty data within each stripe - we do this by
breaking up the existing sectors_dirty into per stripe atomic_ts
Signed-off-by: NKent Overstreet <koverstreet@google.com>