This adds a new method of doing btree updates - a straight write buffer,
implemented as a flat fixed size array.
This is only useful when we don't need to read from the btree in order
to do the update, and when reading is infrequent - perfect for the LRU
btree.
This will make LRU btree updates fast enough that we'll be able to use
it for persistently indexing buckets by fragmentation, which will be a
massive boost to copygc performance.
Changes:
- A new btree_insert_type enum, for btree_insert_entries. Specifies
btree, btree key cache, or btree write buffer.
- bch2_trans_update_buffered(): updates via the btree write buffer
don't need a btree path, so we need a new update path.
- Transaction commit path changes:
The update to the btree write buffer both mutates global, and can
fail if there isn't currently room. Therefore we do all write buffer
updates in the transaction all at once, and also if it fails we have
to revert filesystem usage counter changes.
If there isn't room we flush the write buffer in the transaction
commit error path and retry.
- A new persistent option, for specifying the number of entries in the
write buffer.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Recursive transaction commits are occasionally necessary - in
particular, for the upcoming btree write buffer's flush path.
This avoids bugs due to trans->flags being accidentally mutated
mid-commit, which can cause c->writes refcount leaks.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
When we unlock in order to submit IO, the next relock event is likely to
fail if submit_bio() blocked - we shouldn't those events in our _fail
stats, since those are expected events and shouldn't cause test
failures.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
This uses the new _ip() interface to six locks and hooks it up to
btree_path->ip_allocated, when available.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
In debug mode, we now track where btree iterators and paths are
initialized/allocated - helpful in tracking down btree path overflows.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
This patch
- Adds a mechanism for queuing up journal entries prior to the journal
being started, which will be used for early journal log messages
- Adds bch2_fs_log_msg() and improves bch2_trans_log_msg(), which now
take format strings. bch2_fs_log_msg() can be used before or after
the journal has been started, and will use the appropriate mechanism.
- Deletes the now obsolete bch2_journal_log_msg()
- And adds more log messages to the recovery path - messages for
journal/filesystem started, journal entries being blacklisted, and
journal replay starting/finishing.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
While a btree transaction is running, we hold a SRCU read lock on the
btree key cache that prevents btree key cache keys from being freed -
this is so that relock() operations won't access freed memory.
The downside of this is that long running btree transactions prevent
memory from being freed from the key cache. This adds a check in
bch2_trans_begin() - if the transaction has been running longer than 1
second, drop and retake the SRCU read lock and zero out pointers to
unlock key cache paths.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
This switches btree_key_cache_fill() to use a btree iterator, not a
btree path, so that it can search for keys in previous snapshots.
We also add another iterator flag, BTREE_ITER_KEY_CACHE_FILL, to avoid
recursion back into the key cache.
This will allow us to re-enable the key cache for inodes in the next
patch.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
When flags & btree_id are constants, we can constant fold the entire
calculation of the actual iterator flags - and the whole thing becomes
small enough to inline.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
b->write_type needs to be set atomically with setting the
btree_node_need_write flag, so move it into b->flags.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
This replaces sysfs btree_avg_write_size with btree_write_stats, which
now breaks out statistics by the source of the btree write.
Btree writes that are too small are a source of inefficiency, and
excessive btree resort overhead - this will let us see what's causing
them.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Long ago, bkey_unpack_key() was added to bset.h instead of bkey.h
because bkey.h didn't include btree_types.h, which it needs for the
compiled unpack function.
This patch finally moves it to the proper location.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
We've outgrown our own deadlock avoidance strategy.
The btree iterator API provides an interface where the user doesn't need
to concern themselves with lock ordering - different btree iterators can
be traversed in any order. Without special care, this will lead to
deadlocks.
Our previous strategy was to define a lock ordering internally, and
whenever we attempt to take a lock and trylock() fails, we'd check if
the current btree transaction is holding any locks that cause a lock
ordering violation. If so, we'd issue a transaction restart, and then
bch2_trans_begin() would re-traverse all previously used iterators, but
in the correct order.
That approach had some issues, though.
- Sometimes we'd issue transaction restarts unnecessarily, when no
deadlock would have actually occured. Lock ordering restarts have
become our primary cause of transaction restarts, on some workloads
totally 20% of actual transaction commits.
- To avoid deadlock or livelock, we'd often have to take intent locks
when we only wanted a read lock: with the lock ordering approach, it
is actually illegal to hold _any_ read lock while blocking on an intent
lock, and this has been causing us unnecessary lock contention.
- It was getting fragile - the various lock ordering rules are not
trivial, and we'd been seeing occasional livelock issues related to
this machinery.
So, since bcachefs is already a relational database masquerading as a
filesystem, we're stealing the next traditional database technique and
switching to a cycle detector for avoiding deadlocks.
When we block taking a btree lock, after adding ourself to the waitlist
but before sleeping, we do a DFS of btree transactions waiting on other
btree transactions, starting with the current transaction and walking
our held locks, and transactions blocking on our held locks.
If we find a cycle, we emit a transaction restart. Occasionally (e.g.
the btree split path) we can not allow the lock() operation to fail, so
if necessary we'll tell another transaction that it has to fail.
Result: trans_restart_would_deadlock events are reduced by a factor of
10 to 100, and we'll be able to delete a whole bunch of grotty, fragile
code.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Similar to "bcachefs: Fix usage of six lock's percpu mode", six locks
have a percpu mode, but we can't switch between percpu and non percpu
modes while a lock is in use: threads attempting to take a read lock may
race, and we'll end up with the read count permanently off.
Fixing this the "correct" way, in six_lock_pcpu_(alloc|free) would
require an RCU barrier, and we don't want to do that - instead, we have
to permanently segragate percpu and non percpu objects, including when
on freelists.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Add a type descriptor to btree_bkey_cached_common - there's no reason
not to since we've got padding that was otherwise unused, and this is a
nice cleanup (and helpful in later patches).
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
This patch
- tracks maximum bch2_trans_kmalloc() memory used in btree_transaction_stats
- makes it available in debugfs
- switches bch2_trans_init() to using that for the amount of memory to
preallocate, instead of the parameter passed in
This drastically reduces transaction restarts, and means we no longer
need to track this in the source code.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Previously, we used two different bit arrays for tracking held btree
node locks. This patch switches to an array of two bit integers, which
will let us track, in a future patch, when we hold a write lock.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Start to centralize some of the locking code in a new file; more locking
code will be moving here in the future.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
We need a way to check if the machinery for handling btree_paths with in
a transaction is behaving reasonably, as it often has not been - we've
had bugs with transaction path overflows caused by duplicate paths and
plenty of other things.
This patch tracks, per transaction fn, the most btree paths ever
allocated by that transaction and makes it available in debugfs.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Going to be adding more things to this in the next patch.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
These were used more prior to getting rid of the in-memory bucket arrays
- they don't serve much purpose anymore, and deleting them lets us write
better assertions.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
bch2_btree_trans_to_text() is used to print btree_transactions owned by
other threads; thus, it needs to be particularly careful. This fixes a
null ptr deref caused by racing with the owning thread changing
path->l[].b.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Now that we have error codes, with subtypes, we can switch to our own
error code for transaction restarts - and even better, a distinct error
code for each transaction restart reason: clearer code and better
debugging.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
This will help us improve nested transactions - we need to add
assertions that whenever an inner transaction handles a restart, it
still returns -EINTR to the outer transaction.
This also adds nested_lockrestart_do() and nested_commit_do() which use
the new counters to correctly return -EINTR when the transaction was
restarted.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
The new for_each_btree_key2() macro handles transaction retries,
allowing us to avoid nested transactions - which we want to avoid since
they're tricky to do completely correctly and upcoming assertions are
going to be checking for that.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
We now record the length of time btree locks are held and expose this in debugfs.
Enabled via CONFIG_BCACHEFS_LOCK_TIME_STATS.
Signed-off-by: Daniel Hill <daniel@gluo.nz>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
We try to ensure we never hold btree locks for too long - bcachefs tries
to be soft realtime. This adds a check when restarting a transaction,
where a transaction restart is cheap - if we've been holding locks for
too long, drop and retake them.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Originally, the btree key cache code would always allocate new entries
by reusing from the recently-freed list, if that list wasn't empty. But
that behaviour was dropped, for lock contention reasons.
But it seems that entries stranded on the freed list have been
contributing to some of our oom issues, because long running btree
transactions will prevent them from being freed.
This patch re-adds allocating from the freed list, but it also adds
percpu buffers to solve the lock contention issues - and the new percpu
freed lists will improve the evict paths, too.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
When inserting a key type that's not valid for a given btree, we should
print out which btree we were inserting into.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
With backpointers, alloc keys have gotten bigger, so we're needing more
memory here.
We're probably going to need to go with something more sophisticated
than a bump allocator, but - let's see if we can avoid doing that just
yet.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Previously, on every btree_iter_peek() operation we were searching the
journal keys, doing a full binary search - which was slow.
This patch fixes that by saving our position in the journal keys, so
that we only do a full binary search when moving our position backwards
or a large jump forwards.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
This adds bch2_btree_iter_peek_all_levels(), which returns keys from
every level of the btree - interior nodes included - in monotonically
increasing order, soon to be used by the backpointers check & repair
code.
- BTREE_ITER_ALL_LEVELS can now be passed to for_each_btree_key() to
iterate thusly, much like BTREE_ITER_SLOTS
- The existing algorithm in bch2_btree_iter_advance() doesn't work with
peek_all_levels(): we have to defer the actual advancing until the
next time we call peek, where we have the btree path traversed and
uptodate. So, we add an advanced bit to btree_iter; when
BTREE_ITER_ALL_LEVELS is set bch2_btree_iter_advanced() just marks
the iterator as advanced.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
For backpointers, we'll need the full key location - that means btree_id
and btree level. This patch plumbs it through.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
This adds two new btrees for the upcoming allocator rewrite: an extents
btree of free buckets, and a btree for buckets awaiting discards.
We also add a new trigger for alloc keys to keep the new btrees up to
date, and a compatibility path to initialize them on existing
filesystems.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
This introduces a new alloc key which doesn't use varints. Soon we'll be
adding backpointers and storing them in alloc keys, which means our
pack/unpack workflow for alloc keys won't really work - we'll need to be
mutating alloc keys in place.
Instead of bch2_alloc_unpack(), we now have bch2_alloc_to_v4() that
converts older types of alloc keys to v4 if needed.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Previously, we'd go into an infinite loop when attempting to cache a
bkey in the key cache larger than 128 u64s - since we were only using a
u8 for the size field, it'd get rounded up to 256 then truncated to 0.
Oops.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Six locks have a percpu mode, which we use for interior btree nodes, as
well as btree key cache keys for the subvolumes btree. We've been
switching locks back and forth between percpu and non percpu mode as
needed, but it turns out this is racy - when we're reusing an existing
node, other threads could be attempting to lock it while we're switching
it between modes.
This patch fixes this by never switching 'struct btree' between the two
modes, and instead segragating them between two different freed lists.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Checking btree_node_may_write() isn't atomic with the other btree flags,
dirty and need_write in particular. There was a rare race where we'd
unblock a node from writing while __btree_node_flush() was setting
need_write, and no thread would notice that the node was now both able
to write and needed to be written.
Fix this by adding btree node flags for will_make_reachable and
write_blocked that can be checked in the cmpxchg loop in
__bch2_btree_node_write.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
This helps with lock contention in the journalling code: instead of
updating our journal pin on every write, only get a journal pin if we
don't have one.
This means we can avoid hammering on journal locks nearly so much, at
the cost of carrying around a journal pin for an older entry than the
one we actually need. To handle that, if needed we update our journal
pin to the correct one when flushed by journal reclaim.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
The error code when we fail to allocate a node in the btree node cache
doesn't make it to bch2_btree_path_traverse_all(). Instead, we need to
stash a flag in btree_trans so we know we have to take the cannibalize
lock.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
This consolidates some of the btree node lock path, so that when we're
blocked taking a write lock on a node it shows up in
bch2_btree_trans_to_text(), along with intent and read locks.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>