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: Kent 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: Nicholas Swenson <nks@daterainc.com>
Signed-off-by: Kent 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: Nicholas Swenson <nks@daterainc.com>
Signed-off-by: Kent 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: Nicholas Swenson <nks@daterainc.com>
Signed-off-by: Kent Overstreet <kmo@daterainc.com>

Signed-off-by: Nicholas Swenson <nks@daterainc.com>
Signed-off-by: Kent Overstreet <kmo@daterainc.com>

Signed-off-by: Nicholas Swenson <nks@daterainc.com>
Signed-off-by: Kent Overstreet <kmo@daterainc.com>

Signed-off-by: Nicholas Swenson <nks@daterainc.com>
Signed-off-by: Kent 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: Kent 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: Kent 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: Kent Overstreet <kmo@daterainc.com>

Fixes the following sparse warning:

drivers/md/bcache/btree.c:2220:5: warning:
 symbol 'btree_insert_fn' was not declared. Should it be static?
Signed-off-by: Wei Yongjun <yongjun_wei@trendmicro.com.cn>
Signed-off-by: Kent Overstreet <kmo@daterainc.com>

Just to be safe, call the error reporting function with "%s" to avoid
any possible future format string leak.
Signed-off-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Kent Overstreet <kmo@daterainc.com>

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

Never saw a profile of bset_search_tree() where it wasn't bottlenecked
on memory until I got my new Haswell machine, but when I tried it there
it was suddenly burning 20% of the cpu in the inner loop on shrd...

Turns out, the version of shrd that takes 64 bit operands has a 9 cycle
latency. hah.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>

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

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

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

The flow control in btree_insert_node() was... fragile... before,
this'll use more stack (but since our btrees are never more than depth
1, that shouldn't matter) and it should be significantly clearer and
less fragile.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>

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

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

This dates from before the btree iterator, and now it's finally gone
Signed-off-by: Kent Overstreet <kmo@daterainc.com>

Not a complete fix - we could still deadlock if btree_insert_node() has
to split...
Signed-off-by: Kent 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: Kent Overstreet <kmo@daterainc.com>

Refactoring, prep work for incremental garbage collection.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>

More refactoring - mostly making the interfaces more explicit about what
we actually want to do.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>

btree_insert_key() was open coding this, this is just refactoring.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>

Trying to treat btree pointers and leaf node pointers the same way was a
mistake - going to start being more explicit about the type of
key/pointer we're dealing with. This is the first part of that
refactoring; this patch shouldn't change any actual behaviour.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>

The bucket refcount (dropped with bkey_put()) is only needed to prevent
the newly allocated bucket from being garbage collected until we've
added a pointer to it somewhere. But for btree node allocations, the
fact that we have btree nodes locked is enough to guard against races
with garbage collection.

Eventually the per bucket refcount is going to be replaced with
something specific to bch_alloc_sectors().
Signed-off-by: Kent 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: Kent Overstreet <kmo@daterainc.com>

Previously, bch_ptr_bad() could return false when there was a pointer to
a nonexistant device... it only filtered out keys with PTR_CHECK_DEV
pointers.

This behaviour was intended for multiple cache device support; for that,
just because the device for one of the pointers has gone away doesn't
mean we want to filter out the rest of the pointers.

But we don't yet explicitly filter/check individual pointers, so without
that this behaviour was wrong - a corrupt bkey with a bad device pointer
could cause us to deref a bad pointer. Doh.
Signed-off-by: Kent 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: Kent Overstreet <kmo@daterainc.com>

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

With all the recent refactoring around struct btree op struct search has
gotten rather large.

But we can now easily break it up in a different way - we break out
struct btree_insert_op which is for inserting data into the cache, and
that's now what the copying gc code uses - struct search is now specific
to request.c
Signed-off-by: Kent Overstreet <kmo@daterainc.com>

Last of the btree_map() conversions. Main visible effect is
bch_btree_insert() is no longer taking a struct btree_op as an argument
anymore - there's no fancy state machine stuff going on, it's just a
normal function.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>

When we convert bch_btree_insert() to bch_btree_map_leaf_nodes(), we
won't be passing struct btree_op to bch_btree_insert() anymore - so we
need a different way of returning whether there was a collision (really,
a replace collision).
Signed-off-by: Kent Overstreet <kmo@daterainc.com>

This is prep work for converting bch_btree_insert to
bch_btree_map_leaf_nodes() - we have to convert all its arguments to
actual arguments. Bunch of churn, but should be straightforward.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>

With a the recent bcache refactoring, some of the closure code isn't
needed anymore.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>

This isn't used for waiting asynchronously anymore - so this is a fairly
trivial refactoring.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>

Eventual goal is for struct btree_op to contain only what is necessary
for traversing the btree.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>

There was some looping in submit_partial_cache_hit() and
submit_partial_cache_hit() that isn't needed anymore - originally, we
wouldn't necessarily process the full hit or miss all at once because
when splitting the bio, we took into account the restrictions of the
device we were sending it to.

But, device bio size restrictions are now handled elsewhere, with a
wrapper around generic_make_request() - so that looping has been
unnecessary for awhile now and we can now do quite a bit of cleanup.

And if we trim the key we're reading from to match the subset we're
actually reading, we don't have to explicitly calculate bi_sector
anymore. Neat.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>