bcachefs: Split extents if necessary in bch2_trans_update()

Currently, we handle multiple overlapping extents in the same
transaction commit by doing fixups in bch2_trans_update() - this patch
extents that to split updates when necessary. The next patch that
changes the reflink code to not fragment extents when making them
indirect will require this.

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>

bcachefs: Ratelimiting for writeback IOs

Writeback throttling is a kernel config option and not always enabled.
When it's not enabled we need a fallback, to avoid unbounded memory
pinning and work item backlogs.

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

bcachefs: statfs resports incorrect avail blocks

The current implementation of bch_statfs does not scale the number of
available blocks provided in f_bavail by the reserve factor. This causes
an allocation of a file of this size to fail.

Signed-off-by: Dan Robertson <dan@dlrobertson.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Fix for bch2_bkey_pack_pos() not initializing len/version fields

This bug led to push_whiteout() generating whiteouts that failed
bch2_bkey_invalid() due to nonzero length fields - oops.

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>

bcachefs: Fix a memcpy call

Not supposed to pass a null ptr to memcpy (even if the size is 0).

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>

bcachefs: Fix bch2_extent_can_insert() call

It was being skipped when hole punching, leading to problems when
splitting compressed extents.

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>

bcachefs: Make sure to pass a disk reservation to bch2_extent_update()

It's needed when we split an existing compressed extent - we get a null
ptr deref without it.

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>

bcachefs: made changes to support clang, fixed a couple bugs

fs/bcachefs/bset.c              edited prefetch macro to add clang support
fs/bcachefs/btree_iter.c        bugfix: initialize iter->real_pos in bch2_btree_iter_init for later use
fs/bcachefs/io.c                bugfix: eliminated undefined behavior (negative bitshift)
fs/bcachefs/buckets.c           bugfix: invert sign to handle 64bit abs()

Signed-off-by: Brett Holman <bpholman5@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Fix locking in __bch2_set_nr_journal_buckets()

We weren't holding mark_lock correctly - it's needed for the new_fs
path.

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

bcachefs: properly initialize used values

 - Ensure the second key value in bch_hash_info is initialized to zero
   if the info type is of type BCH_STR_HASH_SIPHASH.

 - Initialize the possibly returned value in bch2_inode_create. Assuming
   bch2_btree_iter_peek returns bkey_s_c_null, the uninitialized value
   of ret could be returned to the user as an error pointer.

 - Fix compiler warning in initialization of bkey_s_c_stripe

fs/bcachefs/buckets.c:1646:35: warning: suggest braces around initialization
of subobject [-Wmissing-braces]
        struct bkey_s_c_stripe new_s = { NULL };
                                         ^~~~

Signed-off-by: Dan Robertson <dan@dlrobertson.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Repair code for multiple types of data in same bucket

bch2_check_fix_ptrs() is awkward, we need to find a way to improve it.

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

bcachefs: Fix out of bounds read in fs usage ioctl

Fix a possible read out of bounds if bch2_ioctl_fs_usage is called when
replica_entries_bytes is set to a value that is smaller than the size
of bch_replicas_usage.

Signed-off-by: Dan Robertson <dan@dlrobertson.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Fix null deref in bch2_ioctl_read_super

Do not attempt to cleanup the returned value of bch2_device_lookup if
the returned value was an error pointer. We currently check to see if
the returned value is null and run the cleanup otherwise. As a result,
we attempt to run the cleanup on a error pointer.

Signed-off-by: Dan Robertson <dan@dlrobertson.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Fix possible null deref on mount

Ensure that the block device pointer in a superblock handle is not
null before dereferencing it in bch2_dev_to_fs. The block device pointer
may be null when mounting a new bcachefs filesystem given another mounted
bcachefs filesystem exists that has at least one device that is offline.

Signed-off-by: Dan Robertson <dan@dlrobertson.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Fix error in parsing of mount options

When parsing the mount options duplicate the given options. This is
required as the options are parsed twice and strsep is used in parsing.
The options will be modified into a possibly invalid options set for the
second round of parsing if the options are not duplicated before
parsing.

Signed-off-by: Dan Robertson <dan@dlrobertson.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: avoid out-of-bounds in split_devs

Calling mount with an empty source string causes an out-of-bounds error
in split_devs. Check the length of the source string to avoid this.

Signed-off-by: Stijn Tintel <stijn@linux-ipv6.be>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Make sure to use BTREE_ITER_PREFETCH in fsck

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

bcachefs: Fix bch2_btree_iter_peek_with_updates()

By not re-fetching the next update we were going into an infinite loop.

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

bcachefs: Fix reflink trigger

The trigger for reflink pointers wasn't always incrementing/decrementing
the refcounts correctly - this patch fixes that logic.

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

bcachefs: Fix some refcounting bugs

We really need debug mode assertions that ca->ref and ca->io_ref are
used correctly.

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

bcachefs: Fix oob write in __bch2_btree_node_write

Fix a possible out of bounds write in __bch2_btree_node_write when
the data buffer padding is cleared up to the block size. The out of
bounds write is possible if the data buffers size is not a multiple
of the block size.

Signed-off-by: Dan Robertson <dan@dlrobertson.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

bcachefs: Fix usage of last_seq + encryption

jset->last_seq is in the region that's encrypted - on journal write
completion, we were using it and getting garbage. This patch shadows it
to fix.

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

bcachefs: Clean up bch2_btree_and_journal_walk()

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

bcachefs: Mark newly allocated btree nodes as accessed

This was a major oversight - this means under memory pressure we can end
up reading in a btree node, then having it evicted before we get to use
it.

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

bcachefs: Fix time handling

There were some overflows in the time conversion functions - fix this by
converting tv_sec and tv_nsec separately. Also, set sb->time_min and
sb->time_max.

Fixes xfstest generic/258.

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

bcachefs: Add a tracepoint for when we block on journal reclaim

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

bcachefs: Make sure to initialize j->last_flushed

If the journal reclaim thread makes it to the timeout without ever
initializing j->last_flushed, we could end up sleeping for a very long
time.

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

bcachefs: Ensure that fpunch updates inode timestamps

Fixes xfstests generic/059

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

bcachefs: Change copygc wait amount to be min of per device waits

We're seeing a filesystem get stuck when all devices but one have no
more reclaimable buckets - because the copygc wait amount is curretly
filesystem wide.

This patch should fix that, possibly at the expensive of running too
much when only one or a few devices is full and the rebalance thread
needs to move data around.

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

bcachefs: Change bch2_btree_key_cache_count() to exclude dirty keys

We're seeing livelocks that appear to be due to
bch2_btree_key_cache_scan repeatedly scanning and blocking other tasks
from using the key cache lock - we probably shouldn't be reporting
objects that can't actually be freed yet.

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

bcachefs: Call bch2_inconsistent_error() on missing stripe/indirect extent

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

bcachefs: New tracepoint for bch2_trans_get_iter()

Trying to debug an issue where after traverse_all() we shouldn't have to
traverse any iterators... yet we are

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

bcachefs: Fix __bch2_trans_get_iter()

We need to also set iter->uptodate to indicate it needs to be traversed.

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

bcachefs: Evict btree nodes we're deleting

There was a bug that led to duplicate btree node pointers being inserted
at the wrong level. The new topology repair code can fix that, except
that the btree cache code gets confused when we read in a btree node
from the pointer that was at the wrong level. This patch evicts nodes
that we're deleting to, which nicely solves the problem.

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

bcachefs: New check_nlinks algorithm for snapshots

With snapshots, using a radix tree for the table of link counts won't
work anymore because we also need to distinguish between inodes with
different snapshot IDs. Instead, this patch builds up a sorted array of
inodes that have hardlinks that we can binary search on - taking
advantage of the fact that with inode backpointers, the check_nlinks()
pass _only_ needs to concern itself with inodes that have hardlinks now.

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

bcachefs: Fix a null ptr deref

Fix a few memory safety issues, found by asan in userspace.

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

bcachefs: New and improved topology repair code

This splits out btree topology repair into a separate pass, and makes
some improvements:
 - When we have to pick which of two overlapping nodes to drop keys
   from, we use the btree node header sequence number to preserve the
   newer node

 - the gc code has been changed so that it doesn't bail out if we're
   continuing/ignoring on fsck error - this way the dump tool can skip
   running the repair pass but still walk all reachable metadata

 - add a new superblock flag indicating when a filesystem is known to
   have btree topology issues, and the topology repair pass should be
   run

 - changing the start/end of a node might mean keys in that node have to
   be deleted: this patch handles that better by splitting it out into a
   separate function and running it explicitly in the topology repair
   code, previously those keys were only being dropped when the btree
   node was read in.

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

bcachefs: Fix key cache assertion

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

bcachefs: New helper __bch2_btree_insert_keys_interior()

Consolidate common parts of bch2_btree_insert_keys_interior() and
btree_split_insert_keys() - prep work for adding some new topology
assertions.

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

bcachefs: Rewrite btree nodes with errors

This patch adds self healing functionality for btree nodes - if we
notice a problem when reading a btree node, we just rewrite it.

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

bcachefs: Fix bch2_verify_keylist_sorted

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

bcachefs: Fix an out of bounds read

bch2_varint_decode() can read up to 7 bytes past the end of the buffer,
which means we need to allocate slightly larger key cache buffers.

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

bcachefs: Use mmap() instead of vmalloc_exec() in userspace

Calling mmap() directly is much better than malloc() then mprotect(), we
end up with much less address space fragmentation.

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

bcachefs: Don't BUG_ON() btree topology error

This replaces an assertion in the btree merge path with a
bch2_inconsistent_error() - fsck will fix it.

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

bcachefs: Fix repair leading to replicas not marked

bch2_check_fix_ptrs() was being called after checking if the replicas
set was marked - but repair could change which replicas set needed to be
marked. Oops.

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

bcachefs: Lookup/create lost+found lazily

This is prep work for subvolumes - each subvolume will have its own
lost+found.

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

bcachefs: Don't BUG() in update_replicas

Apparently, we have a bug where in mark and sweep while accounting for a
key, a replicas entry isn't found. Change the code to print out the key
we couldn't mark and halt instead of a BUG_ON().

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

bcachefs: Fix a deadlock on journal reclaim

Flushing the btree key cache needs to use allocation reserves - journal
reclaim depends on flushing the btree key cache for making forward
progress, and the allocator and copygc depend on journal reclaim making
forward progress.

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

bcachefs: Update bch2_btree_verify()

bch2_btree_verify() verifies that the btree node on disk matches what we
have in memory. This patch changes it to verify every replica, and also
fixes it for interior btree nodes - there's a mem_ptr field which is
used as a scratch space and needs to be zeroed out for comparing with
what's on disk.

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

bcachefs: Fix two btree iterator leaks

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

bcachefs: Punt btree writes to workqueue to submit

We don't want to be submitting IO with btree locks held, and btree
writes usually aren't latency sensitive.

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

bcachefs: Fix a use after free

Turns out, we weren't waiting on in flight btree writes when freeing
existing btree nodes. This lead to stray btree writes overwriting newly
allocated buckets, but only started showing itself with some of the
recent allocator work and another patch to move submitting of btree
writes to worqueues.

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

bcachefs: Fix for btree_gc repairing interior btree ptrs

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>

bcachefs: Preallocate trans mem in bch2_migrate_index_update()

This will help avoid transaction restarts.

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

bcachefs: Allocator refactoring

This uses the kthread_wait_freezable() macro to simplify a lot of the
allocator thread code, along with cleaning up bch2_invalidate_bucket2().

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

bcachefs: Always check for invalid bkeys in trans commit path

We check for this prior to metadata being written, but we're seeing some
strange bugs lately, and this will help catch those closer to where they
occur.

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

bcachefs: Check that keys are in the correct btrees

We've started seeing bug reports of pointers to btree nodes being
detected in leaf nodes. This should catch that before it's happened, and
it's something we should've been checking anyways.

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

bcachefs: Handle errors in bch2_trans_mark_update()

It's not actually the case that iterators are always checked here -
__bch2_trans_commit() checks for that after running triggers.

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

bcachefs: Allocator thread doesn't need gc_lock anymore

Even with runtime gc (which currently isn't supported), runtime gc no
longer clears/recalculates the main set of bucket marks - it allocates
and calculates another set, updating the primary at the end.

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

bcachefs: gc shouldn't care about owned_by_allocator

The owned_by_allocator field is a purely in memory thing, even if/when
we bring back GC at runtime there's no need for it to be recalculating
this field. This is prep work for pulling it out of struct bucket, and
eventually getting rid of the bucket array.

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

bcachefs: Refactor bchfs_fallocate() to not nest btree_trans on stack

Upcoming patch is going to disallow multiple btree_trans on the stack.

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

bcachefs: Fix an unused var warning in userspace

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

bcachefs: Fix some small memory leaks

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

bcachefs: Simplify fsck remove_dirent()

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

bcachefs: Fix transaction restarts due to upgrading of cloned iterators

This fixes a regression from
  52d86202fd bcachefs: Improve bch2_btree_iter_traverse_all()

We want to avoid mucking with other iterators in the btree transaction
in operations that are only supposed to be touching individual iterators
- that patch was a cleanup to move lock ordering handling to
bch2_btree_iter_traverse_all(). But it broke upgrading of cloned
iterators.

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

bcachefs: Fix journal reclaim loop

When dirty key cache keys were separated from other journal pins, we
broke the loop conditional in __bch2_journal_reclaim() - it's supposed
to keep looping as long as there's work to do.

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

bcachefs: Fix an RCU splat

Writepoints are never deallocated so the rcu_read_lock() isn't really
needed, but we are doing lockless list traversal.

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

bcachefs: Simplify bch2_set_nr_journal_buckets()

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

bcachefs: Fix bch2_trans_mark_dev_sb()

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

bcachefs: Improve trans_restart_mem_realloced tracepoint

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

bcachefs: Don't downgrade iterators in bch2_trans_get_iter()

This fixes a livelock with btree node splits.

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

bcachefs: Improve bch2_btree_iter_traverse_all()

By changing it to upgrade iterators to intent locks to avoid lock
restarts we can simplify __bch2_btree_node_lock() quite a bit - this
fixes a probable bug where it could potentially drop a lock on an
unrelated error but still succeed instead of causing a transaction
restart.

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

bcachefs: Fix journal_reclaim_wait_done()

Can't run arbitrary code inside a wait_event() conditional, due to
task state being weird...

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

bcachefs: Fix bch2_gc_done() error messages

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

bcachefs: Don't call bch2_btree_iter_traverse() unnecessarily

If we let bch2_trans_commit() do it, it'll traverse iterators in sorted
order which means we'll get fewer lock restarts.

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

bcachefs: Better iterator picking

Avoid cloning iterators if we don't have to.

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

bcachefs: Drop old style btree node coalescing

We have foreground btree node merging now, and any future btree node
merging improvements are going to be based off of that code.

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

bcachefs: Add a perf test for multiple updates per commit

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

bcachefs: Ensure bucket gen gc completes

We don't want it to block, if it can't allocate it should just continue
instead of possibly deadlocking.

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

bcachefs: Add the status of bucket gen gc to sysfs

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

bcachefs: Fix heap overrun in bch2_fs_usage_read() XXX squash

oops

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

bcachefs: BCH_BEATURE_atomic_nlink is obsolete

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

bcachefs: Improved check_directory_structure()

Now that we have inode backpointers, we can simplify checking directory
structure: instead of doing a DFS from the filesystem root and then
checking if we found everything, we can iterate over every inode and see
if we can go up until we get to the root.

This patch also has a number of fixes and simplifications for the inode
backpointer checks. Also, it turns out we don't actually need the
BCH_INODE_BACKPTR_UNTRUSTED flag.

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

bcachefs: Fix fsck to not use bch2_link_trans()

bch2_link_trans() uses the btree key cache for inode updates, and fsck
isn't supposed to - also, it's not really what we want for reattaching
unreachable inodes anyways.

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

bcachefs: Fix bch2_trans_relock()

The patch that changed bch2_trans_relock() to not look at iter->uptodate
also tried to add an optimization by only having it relock
btree_iter_key() iterators (iterators that are live or have been marked
as keep). But, this wasn't thought through - this pops internal iterator
assertions because on transaction restart, when we're traversing
iterators we traverse all iterators marked as linked, and having
bch2_trans_relock() skip some of those mean that it can skil the
iterator that bch2_btree_iter_traverse_one() is currently traversing.

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

bcachefs: Redo check_nlink fsck pass

Now that we have inode backpointers the check_nlink pass only is
concerned with files that have hardlinks, and can be simplified.

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

bcachefs: Inode backpointers are now required

This lets us simplify fsck quite a bit, which we need for making fsck
snapshot aware.

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

bcachefs: Simplify hash table checks

Very early on there was a period where we were accidentally generating
dirents with trailing garbage; we've since dropped support for
filesystems that old and the fsck code can be dropped.

Also, this patch switches to a simpler algorithm for checking hash
tables. It's less efficient on hash collision - but with 64 bit keys,
those are very rare.

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

bcachefs: Check inodes at start of fsck

This splits out checking inode nlinks from the rest of the inode checks
and moves most of the inode checks to the start of fsck, so that other
fsck passes can depend on it.

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

bcachefs: Fix BTREE_ITER_NOT_EXTENTS

bch2_btree_iter_peek() wasn't properly checking for
BTREE_ITER_IS_EXTENTS when updating iter->pos.

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

bcachefs: Fix bch2_gc_btree_gens()

Since we're using a NOT_EXTENTS iterator, we shouldn't be setting the
iter pos to the start of the extent.

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

bcachefs: Make sure to kick journal reclaim when we're waiting on it

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

bcachefs: Don't wait for ALLOC_SCAN_BATCH buckets in allocator

It used to be necessary for the allocator thread to batch up
invalidating buckets when possible - but since we added the btree key
cache that hasn't been a concern, and now it's causing the allocator
thread to livelock when the filesystem is nearly full.

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

bcachefs: Drop bch2_fsck_inode_nlink()

We've had BCH_FEATURE_atomic_nlink for quite some time, we can drop this
now.

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

bcachefs: Move some dirent checks to bch2_dirent_invalid()

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

bcachefs: Improve bset compaction

The previous patch that fixed btree nodes being written too aggressively
now meant that we weren't sorting btree node bsets optimally - this
patch fixes that.

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

bcachefs: Don't flush btree writes more aggressively because of btree key cache

We need to flush the btree key cache when it's too dirty, because
otherwise the shrinker won't be able to reclaim memory - this is done by
journal reclaim. But journal reclaim also kicks btree node writes: this
meant that btree node writes were getting kicked much too often just
because we needed to flush btree key cache keys.

This patch splits journal pins into two different lists, and teaches
journal reclaim to not flush btree node writes when it only needs to
flush key cache keys.

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

bcachefs: Eliminate more PAGE_SIZE uses

In userspace, we don't really have a well defined PAGE_SIZE and shouln't
be relying on it. This is some more incremental work to remove
references to it.

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

bcachefs: Increase BSET_CACHELINE to 256 bytes

Linear searches have gotten cheaper relative to binary searches on
modern hardware, due to better branch prediction behaviour.

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

bcachefs: Fix a startup race

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