Recently a severe performance regression was discovered, which bisected
to
a6548c8b5e bcachefs: Avoid flushing the journal in the discard path
It turns out the old behaviour, which issued excessive journal flushes,
worked around a performance issue where queueing delays would cause the
journal to not be able to write quickly enough and stall.
The journal flushes masked the issue because they periodically flushed
the device write cache, reducing write latency for non flushes.
This patch reworks the journalling code to allow more than one
(non-flush) write to be in flight at a time. With this patch, doing 4k
random writes and an iodepth of 128, we are now able to hit 560k iops to
a Samsung 970 EVO Plus - previously, we were stuck in the ~200k range.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Previosuly, the transaction commit path would have to add keys to the
btree write buffer as a separate operation, requiring additional global
synchronization.
This patch introduces a new journal entry type, which indicates that the
keys need to be copied into the btree write buffer prior to being
written out. We switch the journal entry type back to
JSET_ENTRY_btree_keys prior to write, so this is not an on disk format
change.
Flushing the btree write buffer may require pulling keys out of journal
entries yet to be written, and quiescing outstanding journal
reservations; we previously added journal->buf_lock for synchronization
with the journal write path.
We also can't put strict bounds on the number of keys in the journal
destined for the write buffer, which means we might overflow the size of
the preallocated buffer and have to reallocate - this introduces a
potentially fatal memory allocation failure. This is something we'll
have to watch for, if it becomes an issue in practice we can do
additional mitigation.
The transaction commit path no longer has to explicitly check if the
write buffer is full and wait on flushing; this is another performance
optimization. Instead, when the btree write buffer is close to full we
change the journal watermark, so that only reservations for journal
reclaim are allowed.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Since outstanding journal buffers hold a journal pin, when flushing all
pins we need to close the current journal entry if necessary so its pin
can be released.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
journal_cur_seq() can legitimately be used outside of the journal lock,
where this assert can race
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
This deletes the complicated and somewhat expensive journal
pre-reservation machinery in favor of just using journal watermarks:
when the journal is more than half full, we run journal reclaim more
aggressively, and when the journal is more than 3/4s full we only allow
journal reclaim to get new journal reservations.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
We can't mark device superblocks or allocate journal on a device that
isn't online.
That means we may need to do this on every mount, because we may have
formatted a new filesystem and then done the first mount
(bch2_fs_initialize()) in degraded mode.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
bcachefs freeze testing via fstests generic/390 occasionally
reproduces the following BUG from bch2_fs_read_only():
BUG_ON(atomic_long_read(&c->btree_key_cache.nr_dirty));
This indicates that one or more dirty key cache keys still exist
after the attempt to flush and quiesce the fs. The sequence that
leads to this problem actually occurs on unfreeze (ro->rw), and
looks something like the following:
- Task A begins a transaction commit and acquires journal_res for
the current seq. This transaction intends to perform key cache
insertion.
- Task B begins a bch2_journal_flush() via bch2_sync_fs(). This ends
up in journal_entry_want_write(), which closes the current journal
entry and drops the reference to the pin list created on entry open.
The pin put pops the front of the journal via fast reclaim since the
reference count has dropped to 0.
- Task A attempts to set the journal pin for the associated cached
key, but bch2_journal_pin_set() skips the pin insert because the
seq of the transaction reservation is behind the front of the pin
list fifo.
The end result is that the pin associated with the cached key is not
added, which prevents a subsequent reclaim from processing the key
and thus leaves it dangling at freeze time. The fundamental cause of
this problem is that the front of the journal is allowed to pop
before a transaction with outstanding reservation on the associated
journal seq is able to add a pin. The count for the pin list
associated with the seq drops to zero and is prematurely reclaimed
as a result.
The logical fix for this problem lies in how the journal buffer is
managed in similar scenarios where the entry might have been closed
before a transaction with outstanding reservations happens to be
committed.
When a journal entry is opened, the current sequence number is
bumped, the associated pin list is initialized with a reference
count of 1, and the journal buffer reference count is bumped (via
journal_state_inc()). When a journal reservation is acquired, the
reservation also acquires a reference on the associated buffer. If
the journal entry is closed in the meantime, it drops both the pin
and buffer references held by the open entry, but the buffer still
has references held by outstanding reservation. After the associated
transaction commits, the reservation release drops the associated
buffer references and the buffer is written out once the reference
count has dropped to zero.
The fundamental problem here is that the lifecycle of the pin list
reference held by an open journal entry is too short to cover the
processing of transactions with outstanding reservations. The
simplest way to address this is to expand the pin list reference to
the lifecycle of the buffer vs. the shorter lifecycle of the open
journal entry. This ensures the pin list for a seq with outstanding
reservation cannot be popped and reclaimed before all outstanding
reservations have been released, even if the associated journal
entry has been closed for further reservations.
Move the pin put from journal entry close to where final processing
of the journal buffer occurs. Create a duplicate helper to cover the
case where the caller doesn't already hold the journal lock. This
allows generic/390 to pass reliably.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
bcachefs freeze testing has uncovered some raciness between journal
entry open/close and pin list reference count management. The
details of the problem are described in a separate patch. In
preparation for the associated fix, refactor the journal buffer put
path a bit to allow it to eventually handle dropping the pin list
reference currently held by an open journal entry.
Retain the journal write dispatch helper since the closure code is
inlined and we don't want to increase the amount of inline code in
the transaction commit path, but rename the function to reflect
the purpose of final processing of the journal buffer.
Signed-off-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
This unifies JOURNAL_WATERMARK with BCH_WATERMARK; we're working towards
specifying watermarks once in the transaction commit path.
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>
This takes copying the payload out of bch2_journal_add_entry(), which
means we can use it for journal_transaction_name() - also prep work for
journalling overwrites.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Previously, the journal read path used a linked list for storing the
journal entries we read from disk. But there's been a bug that's been
causing journal_flush_delay to incorrectly be set to 0, leading to far
more journal entries than is normal being written out, which then means
filesystems are no longer able to start due to the O(n^2) behaviour of
inserting into/searching that linked list.
Fix this by switching to a radix tree.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Since journal reclaim -> btree key cache flushing may require the
allocation of new btree nodes, it has an implicit dependency on copygc
in order to make forward progress - so we should avoid blocking copygc
unless the journal is really close to full.
This introduces watermarks to replace our single MAY_GET_UNRESERVED bit
in the journal, and adds a watermark for copygc and plumbs it through.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
This adds bch2_journal_log_msg(), which just logs a message to the
journal, and uses it to mark startup and when journal replay finishes.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Previous patch just moved responsibility for incrementing the journal
sequence number and initializing the new journal entry from
__journal_entry_close() to journal_entry_open(); this patch makes the
analagous change for journal reservation state, incrementing the index
into array of journal_bufs at open time.
This means that __journal_entry_close() never fails to close an open
journal entry, which is important for the next patch that will change
our emergency shutdown behaviour.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
It makes the code more readable if we work off of sequence numbers,
instead of direct indexes into the array of journal buffers.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
In sysfs, files can only output at most PAGE_SIZE. This is a problem for
debug info that needs to list an arbitrary number of times, and because
of this limit some of our debug info has been terser and harder to read
than we'd like.
This patch moves info about journal pins and cached btree nodes to
debugfs, and greatly expands and improves the output we return.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
This patch was originally to work around the journal geting stuck in
nochanges mode - but that was just a hack, we needed to fix the actual
bug. It should be fixed now, so revert it.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
The journal can get stuck if we need to get a journal reservation for
something we have a pre-reservation for, but aren't able to reclaim
space, or if the pin fifo is full - it's impractical to resize the pin
fifo at runtime.
Previously, we reserved 8 entries in the pin fifo for pre-reservations,
but that seems small - we're seeing the journal occasionally get stuck.
Let's reserve a quarter of it.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Add bch2_journal_noflush_seq(), for telling the journal that entries
before a given sequence number should not be flushes - to be used by an
upcoming allocator optimization.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
This was used for recording which inodes have been modified by in flight
journal writes, but was broken and has been superceded.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
This tweaks the journal code to always act as if there's space available
in nochanges mode, when we're not going to be doing any writes. This
helps in recovering filesystems that won't mount because they need
journal replay and the journal has gotten stuck.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Using the normal transaction commit path to insert and journal updates
to interior nodes hadn't been done before this repair code was written,
not surprising that there was a bug.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
JOURNAL_RES_GET_RESERVED should only be used for updatse that need to be
done to free up space in the journal. In particular, when we're flushing
keys from the key cache, if we're flushing them out of order we
shouldn't be using it, since we're using up our remaining space in the
journal without dropping a pin that will let us make forward progress.
With this patch, BTREE_INSERT_JOURNAL_RECLAIM without
BTREE_INSERT_JOURNAL_RESERVED may return -EAGAIN - we can't wait on
journal reclaim if we're already in journal reclaim.
This means we need to propagate these errors up to journal reclaim,
indicating that flushing a journal pin should be retried in the future.
This is prep work for a patch to change the way journal reclaim works,
to split out flushing key cache keys because the btree key cache is too
dirty from journal reclaim because we need space in the journal.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
All writes prior to a journal write need to be flushed before the
journal write itself happens. On single device filesystems, it suffices
to mark the write with REQ_PREFLUSH|REQ_FUA, but on multi device
filesystems we need to issue flushes to every device - and wait for them
to complete - before issuing the journal writes. Previously, we were
issuing flushes to every device, but we weren't waiting for them to
complete before issuing the journal writes.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
- Try to always keep 1/8th of the journal free, on top of
pre-reservations
- Move the check for whether the journal is stuck to
bch2_journal_space_available, and make it only fire when there aren't
any journal writes in flight (that might free up space by updating
last_seq)
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
This patch adds a flag to journal entries which, if set, indicates that
they weren't done as flush/fua writes.
- non flush/fua journal writes don't update last_seq (i.e. they don't
free up space in the journal), thus the journal free space
calculations now check whether nonflush journal writes are currently
allowed (i.e. are we low on free space, or would doing a flush write
free up a lot of space in the journal)
- write_delay_ms, the user configurable option for when open journal
entries are automatically written, is now interpreted as the max
delay between flush journal writes (default 1 second).
- bch2_journal_flush_seq_async is changed to ensure a flush write >=
the requested sequence number has happened
- journal read/replay must now ignore, and blacklist, any journal
entries newer than the most recent flush entry in the journal. Also,
the way the read_entire_journal option is handled has been improved;
struct journal_replay now has an entry, 'ignore', for entries that
were read but should not be used.
- assorted refactoring and improvements related to journal read in
journal_io.c and recovery.c
Previously, we'd have to issue a flush/fua write every time we
accumulated a full journal entry - typically the bucket size. Now we
need to issue them much less frequently: when an fsync is requested, or
it's been more than write_delay_ms since the last flush, or when we need
to free up space in the journal. This is a significant performance
improvement on many write heavy workloads.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
This patch increases the maximum journal buffers in flight from 2 to 4 -
this will be particularly helpful when in the future we stop requiring
flush+fua for every journal write.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Improved the way we track various state by adding j->err_seq, which
records the first journal sequence number that encountered an error
being written, and j->last_empty_seq, which records the most recent
journal entry that was completely empty.
Also, use the low bits of the journal sequence number to index the
corresponding journal_buf.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Usage of the journal has gotten somewhat simpler over time - neat.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
We also need to update the journal's bloom filter of inode numbers that
each journal write has upudates for - in case the inode gets evicted
before it gets fsynced.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
printbufs know how big the buffer is that was allocated, so we can get
rid of the random PAGE_SIZEs all over the place.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
This is better than skipping the journal pre-reservation if we already
have one - we should still acount for the journal reservation we're
going to have to get.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
We now update the alloc info (bucket sector counts) atomically with
journalling the update to the interior btree nodes, and we also set new
btree roots atomically with the journalled part of the btree update.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
bch2_journal_res_get() in nonblocking mode is equivalent to a trylock.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Now, we store blacklisted journal sequence numbers in the superblock,
not the journal: this helps to greatly simplify the code, and more
importantly it's now implemented in a way that doesn't require all btree
nodes to be visited before starting the journal - instead, we
unconditionally blacklist the next 4 journal sequence numbers after an
unclean shutdown.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
