btree_path_get_locks, on failure, shouldn't unlock if we're not issuing
a transaction restart: we might drop locks we're not supposed to (if
path->should_be_locked is set).
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Don't put btree locking asserts behind CONFIG_BCACHEFS_DEBUG, put them
behind a module parameter.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
New helper for dropping all write locks; which is distinct from the
helper the transaction commit path uses, which is faster and only
touches updates.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
We rely on the trans->locked to know if a trans has nodes locked for
assertions about deadlocks; there can't be more than one trans in the
same process that is locked.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
This adds lockdep tracking for held btree locks with a single dep_map in
btree_trans, i.e. tracking all held btree locks as one object.
This is more practical and more useful than having lockdep track held
btree locks individually, because
- we can take more locks than lockdep can track (unbounded, now that we
have dynamically resizable btree paths)
- there's no lock ordering between btree locks for lockdep to track (we
do cycle detection)
- and this makes it easy to teach lockdep that btree locks are not safe
to hold while invoking memory reclaim.
The last rule is one that lockdep would never learn, because we only do
trylock() from within shrinkers - but we very much do not want to be
invoking memory reclaim while holding btree node locks.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
In the key cache fill path, we use path_upgrade() on a path that isn't
uptodate yet but should be locked.
This change makes bch2_btree_path_upgrade() slightly looser so we can
use it in key cache upgrade, instead of the __ version.
Also, make the related assert - that path->uptodate implies nodes_locked
- slightly clearer.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
path->idx is now a code smell: we should be using path_idx_t, since it's
stable across btree path reallocation.
This is also a bit faster, using the same loop counter vs. fetching
path->idx from each path we iterate over.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
We should only be downgrading locks on success - otherwise, our
transaction restarts won't be getting the correct locks and we'll
livelock.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
This changes mark_btree_node_locked() to take an enum
btree_node_locked_type, not a six_lock_type, since BTREE_NODE_UNLOCKED
is -1 which may cause problems converting back and forth to
six_lock_type if short enums are in use.
With this change, we never store BTREE_NODE_UNLOCKED in a six_lock_type
enum.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
clang had a few more warnings about enum conversion, and also didn't
like the opts.c initializer.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
- endianness fixes
- mark some things static
- fix a few __percpu annotations
- fix silent enum conversions
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
- Expanded and revamped overview documentation in six.h, giving an
overview of all features
- docbook-comments for all external interfaces
- Rename some functions for simplicity, i.e.
six_lock_ip_type() -> six_lock_ip()
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
As suggested by Linus, this drops the six_lock_state union in favor of
raw bitmasks.
On the one hand, bitfields give more type-level structure to the code.
However, a significant amount of the code was working with
six_lock_state as a u64/atomic64_t, and the conversions from the
bitfields to the u64 were deemed a bit too out-there.
More significantly, because bitfield order is poorly defined (#ifdef
__LITTLE_ENDIAN_BITFIELD can be used, but is gross), incrementing the
sequence number would overflow into the rest of the bitfield if the
compiler didn't put the sequence number at the high end of the word.
The new code is a bit saner when we're on an architecture without real
atomic64_t support - all accesses to lock->state now go through
atomic64_*() operations.
On architectures with real atomic64_t support, we additionally use
atomic bit ops for setting/clearing individual bits.
Text size: 7467 bytes -> 4649 bytes - compilers still suck at
bitfields.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
six_lock_pcpu_alloc() is an unsafe interface: it's not safe to allocate
or free the percpu reader count on an existing lock that's in use, the
only safe time to allocate percpu readers is when the lock is first
being initialized.
This patch adds a flags parameter to six_lock_init(), and instead of
six_lock_pcpu_free() we now expose six_lock_exit(), which does the same
thing but is less likely to be misused.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
This fixes some confusion in the lockdep code due to initializing btree
node/key cache locks with the same lockdep key, but different names.
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>
local_clock() isn't always completely accurate - e.g. on machines with
TSC drift - but ktime_get_ns() overhead is too high, unfortunately.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Most of the node_relock_fail trace events are generated from
bch2_btree_path_verify_level(), when debugcheck_iterators is enabled -
but we're not interested in these trace events, they don't indicate that
we're in a slowpath.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
In order for bch2_btree_node_lock_write_nofail() to never produce a
deadlock, we must ensure we're never holding read locks when using it.
Fortunately, it's only used from code paths where any read locks may be
safely dropped.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
In the event that we're not finished debugging the cycle detector, this
adds a new file to debugfs that shows what the cycle detector finds, if
anything. By comparing this with btree_transactions, which shows held
locks for every btree_transaction, we'll be able to determine if it's
the cycle detector that's buggy or something else.
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>
Centralizing the transaction restart/tracepoint in
bch2_btree_path_upgrade() lets us improve the tracepoint - now it emits
old and new locks_want.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Ideally, all the code in btree_locking.c should be converted, but then
we'd want to convert btree_path to point to btree_key_cached_common too,
and then we'd be in for a much bigger cleanup - but a bit of incremental
cleanup will still be helpful for the next patches.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Taking a write lock will be able to fail, with the new cycle detector -
unless we pass it nofail, which is possible but not preferred.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
In the future, with the new deadlock cycle detector, we won't be using
bare six_lock_* anymore: lock wait entries will all be embedded in
btree_trans, and we will need a btree_trans context whenever locking a
btree node.
This patch plumbs a btree_trans to the few places that need it, and adds
two new locking functions
- btree_node_lock_nopath, which may fail returning a transaction
restart, and
- btree_node_lock_nopath_nofail, to be used in places where we know we
cannot deadlock (i.e. because we're holding no other locks).
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
six locks are unfair: while a thread is blocked trying to take a write
lock, new read locks will fail. The new deadlock cycle detector makes
use of our existing lock tracing, so we need to tell it we're holding a
write lock before we take the lock for it to work correctly.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Since we've now got time_stats for lock hold times (per btree
transaction), we don't need this anymore.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
