When we fail to allocate because of insufficient open buckets, we don't
want to retry from the full set of devices - we just want to retry in
blocking mode.

But if the retry in blocking mode fails with a different error code, we
end up squashing the -BCH_ERR_open_buckets_empty error with an error
that makes us thing we won't be able to allocate (insufficient_devices)
- which is incorrect when we didn't try to allocate from the full set of
devices, and causes the write to fail.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

When allocating from devices with different durability, we might end up
with more replicas than required; this changes
bch2_alloc_sectors_start() to check for this, and drop replicas that
aren't needed to hit the number of replicas requested.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

We still have disk space accounting changes coming for erasure coding,
and the changes won't be as strictly backwards compatible as they'd
ought to be - specifically, we need to start accounting striped data
under a separate counter in bch_alloc (which describes buckets).

A fsck will suffice for upgrading/downgrading, but since erasure coding
is the most incomplete major feature of bcachefs it still makes sense to
put behind a separate kconfig option, so that users are fully aware.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

A recent bug report uncovered a scenario where a filesystem never
runs with freespace_initialized, and therefore the user observes
significantly degraded write performance by virtue of running the
early bucket allocator. The associated bug aside, the primary cause
of the performance drop in this particular instance is that the
early bucket allocator does not update the allocation cursor. This
means that every allocation walks the alloc btree from the first
bucket of the associated device looking for a bucket marked as free
space.

Update the early allocator code to set the alloc cursor to the last
processed position in the tree, similar to how the freelist
allocator behaves. With the alloc_cursor being updated, the retry
logic also needs to be updated to restart from the beginning of the
device when a free bucket is not available between the cursor and
the end of the device. Track the restart position in a first_bucket
variable to make the code a bit more easily readable and consistent
with the freelist allocator.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs had a transient bug where freespace_initialized was not
properly being set, which lead to unexpected use of the early bucket
allocator at runtime. This issue has been fixed, but the existence
of it uncovered a coherency issue in the early bucket allocation
code that is somewhat related to how uncached iterators deal with
the key cache.

The problem itself manifests as occasional failure of generic/113
due to corruption, often seen as a duplicate backpointer or multiple
data types per-bucket error. The immediate cause of the error is a
racing bucket allocation along the lines of the following sequence:

- Task 1 selects key A in bch2_bucket_alloc_early() and schedules.
- Task 2 selects the same key A, but proceeds to complete the
  allocation and associated I/O, after which it releases the
  open_bucket.
- Task 1 resumes with key A, but does not recognize the bucket is
  now allocated because the open_bucket has been removed
  from the hash when it was released in the previous step.

This generally shouldn't happen because the allocating task updates
the alloc btree key before releasing the bucket. This is not
sufficient in this particular instance, however, because an uncached
iterator for a cached btree doesn't actually lock the key cache slot
when no key exists for a given slot in the cache. Thus the fact that
the allocation side updates the cached key means that multiple
uncached iters can stumble across the same alloc key and duplicate
the bucket allocation as described above.

This is something that probably needs a longer term fix in the
iterator code. As a short term fix, close the race through explicit
use of a cached iterator for likely allocation candidates. We don't
want to scan the btree with a cached iterator because that would
unnecessarily pollute the cache. This mitigates cache pollution by
primarily scanning the tree with an uncached iterator, but closes
the race by creating a key cache entry for any prospective slot
prior to the bucket allocation attempt (also similar to how
_alloc_freelist() works via try_alloc_bucket()). This survives many
iterations of generic/113 on a kernel hacked to always use the early
bucket allocator.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

We're using more stack than we'd like in a number of functions, and
btree_trans is the biggest object that we stack allocate.

But we have to do a heap allocatation to initialize it anyways, so
there's no real downside to heap allocating the entire thing.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

More reorganization, this splits up io.c into
 - io_read.c
 - io_misc.c - fallocate, fpunch, truncate
 - io_write.c

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

Now we also print the open_buckets owned by each write_point - this is
to help with debugging a shutdown hang.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>