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>
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>
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>
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>
This is a fairly straightforward conversion, mostly reshuffling -
op->lookup_done goes away, replaced by MAP_DONE/MAP_CONTINUE. And the
code for handling cache hits and misses wasn't really btree code, so it
gets moved to request.c.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
With the new btree_map() functions, we don't need to export the stuff
needed for traversing the btree anymore.
Signed-off-by: Kent 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: Kent 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: Kent 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: Kent 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: Kent Overstreet <kmo@daterainc.com>
Slowly working on pruning struct btree_op - the aim is for it to only
contain things that are actually necessary for traversing the btree.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
Making things less asynchronous that don't need to be - bch_journal()
only has to block when the journal or journal entry is full, which is
emphatically not a fast path. So make it a normal function that just
returns when it finishes, to make the code and control flow easier to
follow.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
Try to improve some of the naming a bit to be more consistent, and also
improve the flow of control in request_write() a bit.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
Some refactoring - better to explicitly pass stuff around instead of
having it all in the "big bag of state", struct btree_op. Going to prune
struct btree_op quite a bit over time.
Signed-off-by: Kent 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: Kent Overstreet <kmo@daterainc.com>
We'll often end up with a list of adjacent keys to insert -
because bch_data_insert() may have to fragment the data it writes.
Originally, to simplify things and avoid having to deal with corner
cases bch_btree_insert() would pass keys from this list one at a time to
btree_insert_recurse() - mainly because the list of keys might span leaf
nodes, so it was easier this way.
With the btree_insert_node() refactoring, it's now a lot easier to just
pass down the whole list and have btree_insert_recurse() iterate over
leaf nodes until it's done.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
The flow of control in the old btree insertion code was rather -
backwards; we'd recurse down the btree (in btree_insert_recurse()), and
then if we needed to split the keys to be inserted into the parent node
would be effectively returned up to btree_insert_recurse(), which would
notice there was more work to do and finish the insertion.
The main problem with this was that the full logic for btree insertion
could only be used by calling btree_insert_recurse; if you'd gotten to a
btree leaf some other way and had a key to insert, if it turned out that
node needed to be split you were SOL.
This inverts the flow of control so btree_insert_node() does _full_
btree insertion, including splitting - and takes a (leaf) btree node to
insert into as a parameter.
This means we can now _correctly_ handle cache misses - for cache
misses, we need to insert a fake "check" key into the btree when we
discover we have a cache miss - while we still have the btree locked.
Previously, if the btree node was full inserting a cache miss would just
fail.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
This is prep work for the reworked btree insertion code.
The way we set b->parent is ugly and hacky... the problem is, when
btree_split() or garbage collection splits or rewrites a btree node, the
parent changes for all its (potentially already cached) children.
I may change this later and add some code to look through the btree node
cache and find all our cached child nodes and change the parent pointer
then...
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
Checking i->seq was redundant, because since ages ago we always
initialize the new bset when advancing b->written
Signed-off-by: Kent 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: Kent 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: Kent Overstreet <kmo@daterainc.com>
bch_keybuf_del() takes a spinlock that can't be taken in interrupt context -
whoops. Fortunately, this code isn't enabled by default (you have to toggle a
sysfs thing).
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.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
Cc: linux-stable <stable@vger.kernel.org> # >= v3.10
Someone cut and pasted md's md_trim_bio() into xen-blkfront.c. Come on,
we should know better than this.
Signed-off-by: Kent Overstreet <kmo@daterainc.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Neil Brown <neilb@suse.de>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Jeremy Fitzhardinge <jeremy@goop.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
SCSI discard will damage discard stripe bio setting, eg, some fields are
changed. If the stripe is reused very soon, we have wrong bios setting. We
remove discard stripe from hash list, so next time the strip will be fully
initialized.
Suitable for backport to 3.7+.
Cc: <stable@vger.kernel.org> (3.7+)
Signed-off-by: Shaohua Li <shli@fusionio.com>
Signed-off-by: NeilBrown <neilb@suse.de>
SCSI layer will add new payload for discard request. If two bios are merged
to one, the second bio has bi_vcnt 1 which is set in raid5. This will confuse
SCSI and cause oops.
Suitable for backport to 3.7+
Cc: stable@vger.kernel.org (v3.7+)
Reported-by: Jes Sorensen <Jes.Sorensen@redhat.com>
Signed-off-by: Shaohua Li <shli@fusionio.com>
Signed-off-by: NeilBrown <neilb@suse.de>
Acked-by: Martin K. Petersen <martin.petersen@oracle.com>
Since:
commit 7ceb17e87b
md: Allow devices to be re-added to a read-only array.
spares are activated on a read-only array. In case of raid1 and raid10
personalities it causes that not-in-sync devices are marked in-sync
without checking if recovery has been finished.
If a read-only array is degraded and one of its devices is not in-sync
(because the array has been only partially recovered) recovery will be skipped.
This patch adds checking if recovery has been finished before marking a device
in-sync for raid1 and raid10 personalities. In case of raid5 personality
such condition is already present (at raid5.c:6029).
Bug was introduced in 3.10 and causes data corruption.
Cc: stable@vger.kernel.org
Signed-off-by: Pawel Baldysiak <pawel.baldysiak@intel.com>
Signed-off-by: Lukasz Dorau <lukasz.dorau@intel.com>
Signed-off-by: NeilBrown <neilb@suse.de>
