Add item functions for deleting items that we know to be dirty and add a
user in another function that deletes many items without leaving parial
deletions behind in the case of errors.
Signed-off-by: Zach Brown <zab@versity.com>
Restore unlink functionality by converting unlink and orphan item
processing from the old btree interface to the new item cache interface.
Signed-off-by: Zach Brown <zab@versity.com>
Add scoutfs_item_next_same() which requires that the key lengths be
identical but which allows any values, including no value by way of a
null kvec.
Signed-off-by: Zach Brown <zab@versity.com>
Some callers know that it's safe to increment their partial keys. Let
them do so without trying to expand the keys to full precision and
triggering warnings that their buffers aren't large enough.
Signed-off-by: Zach Brown <zab@versity.com>
So far we were only able to add items to the segments. To support
deletion we have to insert deletion items and then remove them and the
item they reference when their segments are compacted.
As callers attempt to delete items from the item cache we replace the
existing item with a deletion marker with the key but no value. Now
that there are deletion items in the cache we have to teach the other
item cache operations to skip them. There's some noise in the patch
from moving functions around so that item insertion can free a deletion
item it finds.
The deletion items are written out to the segment as usual except now
the in-segment item struct has a flag to mark a deletion item and the
deletion item is removed from the cache once its written to the segment.
Item reading knows to skip deletion items and not add them back into
the cache.
Compaction proceeds as usual for most of the levels with the deletion
item clobbering any older higher level items with the same key.
Eventually the deletion item itself is removed by skipping over it when
compacting to the largest final level. We support this by adding a
little call that describes the max level of the tree at the time the
compaction starts so that compaction can tell when it should skip
copying the deletion item to the final lower level.
All of this is for deletion of items with a precise key. In the future
we'll expand the deletion items so that they can reference a contiguous
range of keys.
Signed-off-by: Zach Brown <zab@versity.com>
The key and value offsets and lengths were aggressively packed into the
item structs in the segments. This saved a few bytes per item but
didn't leave any room left for expansion without growing the item. We
want to add a deletion item flag so let's just grow the item struct. It
now has room for full precision offsets and lengths that we can access
natively so we can get rid fo the packing and unpacking functions.
Signed-off-by: Zach Brown <zab@versity.com>
The BUILD_BUG_ON() to test that the start of the items in the segment
header is naturally aligned had a typo that masked the length instead of
checking the remainder of division by the length.
Signed-off-by: Zach Brown <zab@versity.com>
After segments have finished IO and while they're in the rbtree we track
them with an LRU. Under memory pressure we can remove the oldest
segments from the rbtree and free them.
Signed-off-by: Zach Brown <zab@versity.com>
Using kvecs for keys seemed like a good idea because there were a few
uses that had keys in fragmented memory: dirent keys made up of an
on-stack struct and the file name in the dentry, and keys straddling the
pages that make up a cached segment.
But it hasn't worked out very well. The code to perform ops on keys
by iterating over vectors is pretty fiddly. And the raw kvecs only
describe the actively referenced key, they know nothing about the total
size of the buffer that the key resides in. Some ops can't check that
they're not clobbering things, they're relying on callers not to mess
up.
And critically, the kvec iteration's become a bottleneck. It turns out
that comparing keys is a very hot path in the item cache. All the code
to initialize and iterate over two key vectors adds up when each high
level fs operation is a few tree descents and each tree descent is a
bunch of compares.
So let's back off and have a specific struct for tracking keys that are
stored in contiguous memory regions. Users ensure that keys are
contiguous. The code ends up being a lot clearer, code now can see how
big the full key buffer is, and the rbtree node comparison fast path is
now just a memcmp.
Almost all of the changes in the patch are mechanical semantic changes
involving types, function names, args, and occasionaly slightly
different return conventions.
A slightly more involved change is that now dirent key users have to
manage an allocated contiguous key with a copy of the path from the
dentry.
Item reading is now a little more clever about calculating the greatest
range it can cache by initially walking all the segments instead of
trying to do it as it runs out of items in each segment.
The largest meaningful change is that now keys can't straddle page
boundaries in memory which means they can't cross block boundaries in
the segment. We align key offsets to the next block as we write keys to
segments that would have straddled a block.
We then also have to account for that padding when building segments.
We add a helper that calculates if a given number of items will fit in a
segment which is used by item dirtying, segment writing, and compaction.
I left the tracepoint formatting for another patch.
Signed-off-by: Zach Brown <zab@versity.com>
Previously the only clever compaction avoidance we'd try was in the
manifest walk. If we found that there were no overlapping segments in
the next level we'd just move the entry down a level and skip compaction
entirely.
But that's just one specific instance of the general case: either of the
lower or upper segments don't overlap with each other. There can be
many lower level segments that intersect with the full range of keys in
the upper level segment but which don't actually intersect with any
items in the upper segment.
So we refactor the compaction to notice this case. We get the first and
last keys and use them to skip each segment as we first start to iterate
through it.
We don't want to read segments that we never actually have to copy items
from so we read each segment on demand instead of concurrently as the
compaction starts. This means that item iteration can now have to read
a segment and can now return errors.
Signed-off-by: Zach Brown <zab@versity.com>
Compaction looks for the next entry at a given level to compact. It
only tested for not finding a next entry when it needs to wrap the key
and start over in the level, it missed the case where the next entry is
at a greater level.
Signed-off-by: Zach Brown <zab@versity.com>
Compaction is going to want to get additional references on a segment.
It could just "read" it again while holding a reference but this is more
clear.
Signed-off-by: Zach Brown <zab@versity.com>
Some drives don't set bi_size so just track the number of IOs. (And the
size argument to end_io has been removed in recent kernels.)
Signed-off-by: Zach Brown <zab@versity.com>
The first pass at compaction just kicked a thread any time we added a
segment that brought its level's count over the limit. Tasks could
create dirty items and write level0 segments regardless of the progress
of compaction.
This ties the writing rate to compaction. Writers have to wait to hold
a transaction until the dirty item count is under a segment and there's
no level0 segments. Usualy more level0 segments are allowed but we're
aggressively pushing compaction, we'll relax this later.
This also more forcefully ensures that compaction makes forward
progress. We kick the compaction thread if we exceed the level count,
wait for level0 to drain, or successfully complete a compaction. We
tweak scoutfs_manifest_next_compact() to return 0 if there's no
compaction work to do so the the compaction thread can exit without
triggering another.
For clarity we also kick off a sync after compaction so that we don't
sit around with a dirty manifest until the next sync. This may not be
wise.
Signed-off-by: Zach Brown <zab@versity.com>
A previous refactoring messed up and had scoutfs_trans_write_func()
always write the dirty super even when nothing was dirty and there was
nothing for the sync attempt to do. This was very confusing and made it
look like the segment and treap writes were being lost when in fact it
was the super write that shouldn't have happened.
Signed-off-by: Zach Brown <zab@versity.com>
The transaction writing thread tests if the manifest and alloc treaps
are dirty. It did this by testing if there were any dirty nodes in the
treap.
But this misses the case where the treap has been modified and all nodes
have been removed. In that case the root references no dirty nodes but
needs to be written.
Instead let's specifically mark the treap dirty when it's modified.
From then on sync will always try to write it out. We also integrate
updating the persistent root as part of writing the dirty nodes to the
persistent ring. It's required and every caller did it so it was silly
to make it a separate step.
Signed-off-by: Zach Brown <zab@versity.com>
Some recent refactoring accidentally set the trans task to null instead
of the current task. It's not used but until it's removed it should be
correct.
Signed-off-by: Zach Brown <zab@versity.com>
We were manually manipulating the level counts in the super in a bunch
of places under the manifest rwsem. This refactors them into simple get
and add functions. We protect them with a seqcount so that we'll be
able read them without blocking (from trans hold attempts). We also add
a helper for testing that a level is full because we already used
different comparisons in two call sites.
Signed-off-by: Zach Brown <zab@versity.com>
This is the first draft of compaction which has the core mechanics.
Add segment functions to free a segment's segno and to delete the entry
that refers to the given segment.
Add manifest functions that lock the manifest and dirty and delete
manifest entries. These are used by the compaction thread to atomically
modify the manfiest with the result of a compaction.
Sort the level 0 entries in the manifest by their sequence. This lets
compaction use the first oldest entry and reading can walk them
backwards to get them in order and not have to sort. We also more
carefully use the sequence field in the manifest search key to
differentiate between finding high level entries that overlap and
finding specific entries identified by their seq.
Add some fields to the per-super compact_info struct which support
compaction. We need to know the limit on the number of segments per
level and we record keys per level which tell us which segment to use
next time that level is compacted.
We kick a compaction thread when we add a manifest entry and that brings
the level count over the limit.
scoutfs_manifest_next_compact() is the first meaty function. The
compaction thread uses this to get all the segments involved in a
compaction. It does a quick manifest update if the next manifest
candidate doesn't overlap with any sgements in the next level.
The compaction operation itself is a pretty straight forward
read-modify-write operation. It asks the manifest to give it references
to the segments it'll need, reads them in, iterates over them to count
and copies items in order to output segments, and atomically updates the
manifest.
Now that the manifest can be dirty without any dirty segments we need to
fix the transaction writing function's assumption that everything flows
from dirty segments. It also has to now lock and unlock the manifest as
it adds the entry for its level 0 segment.
Signed-off-by: Zach Brown <zab@versity.com>
_lookup() and _lookup_next() each had nearly identical loops that took a
dirty boolean. We combine them into one walker with flags for dirty and
next and add a prev prev as well, giving us all the exported functions
with combinations of the flags.
We also add _last() to match _first() and _prev() to match _next().
Signed-off-by: Zach Brown <zab@versity.com>
While walking up parents looking for the next node we were comparing the
child with the wrong parent pointer. This is easily verified by
glancing at rb_next() :).
Signed-off-by: Zach Brown <zab@versity.com>
Treap deletion was pretty messed up. It forgot to reset parent and ref
for the swapped node before using them to finally delete. And it didn't
get all the weird cases right where the child node to swap is the direct
child of the node. In that case we can't just swap the parent pointers
and node pointers, they need to be special cased.
So nuts to all that. We'll just rotate the node down until it doesn't
have both children. They result in pretty similar patterns and the
rotation mechanism is much simpler to understand.
Signed-off-by: Zach Brown <zab@versity.com>
We were derefing the null parent when deleting a single node in a tree.
There's no need to use parent_ref() here, we know that there's no node
and we can just clear the root's aug bits.
Signed-off-by: Zach Brown <zab@versity.com>
The code was using parent_ref() to set the parent ref's node pointer.
But parent_ref() uses the parent's left node pointer to determine which
ref points to the node. If we were setting the left it would return the
right because the left isn't set yet. This messed up the tree shape and
all hell broke loose.
Just set it through the ref, we have it anyway.
Signed-off-by: Zach Brown <zab@versity.com>
The segment writing loop was assuming that the currently dirty items
will fit in a segment. That's not true.
Signed-off-by: Zach Brown <zab@versity.com>
We forgot to or in a node's children's augmentation bits when setting
the augmentation bits up in the parent's ref. This stopped ring
dirtying from finding all the dirty nodes in the treap.
Signed-off-by: Zach Brown <zab@versity.com>
We hadn't yet assigned real sequence numbers to the segments.
Let's track the next sequence in the super block and assign it to
segments as we write the first new item in each.
Signed-off-by: Zach Brown <zab@versity.com>
The first pass manifest and allocator storage used a simple ring log
that was entirely replayed into memory to be used. That risked the
manifest being too large to fit in memory, especially with large keys
and large volumes.
So we move to using an indexed persistent structure that can be read on
demand and cached. We use a treap of byte referenced nodoes stored in a
circular ring.
The code interface is modeled a bit on the in-memory rbtree interface.
Except that we can get IO errors and manage allocation so we return data
pointers to the item payload istead of item structs and we can return
errors.
The manifest and allocator are converted over and the old ring code is
removed entirely.
Signed-off-by: Zach Brown <zab@versity.com>
We have use of this little u64 comparison function in a few more places
so let's make it a proper usable inline function in a header.
Signed-off-by: Zach Brown <zab@versity.com>
The item cache only knew about present items in the rbtree. Attempts to
read items that didn't exist would always trigger expensive manfifest
and segment searches.
This reworks the item cache and item reading code to support the notion
of cached ranges of keys. When we read items we also communicate the
range of keys that we searched. This lets the cache return negative lookups
for key values in the search that don't have items.
The item cache gets an rbtree of key ranges. Each item lookup method
now uses it to determine if a missing item needs to trigger a read.
Item reading is now performed in batches instead of one at a time. This
lets us specify the cache range along with the batch and apply them all
atomically under the lock.
The item range code is much more robust now that it has to track
the range of keys that it searches. The read items call now takes a range.
It knows to look for all level0 segments that interesect that range, not
just the first key. The manifest segment references now include the
min and max keys for the segment so we can use those to define the
item search range.
Since the refs now include keys we no longer have them as a dumb allocated
array but instead have a list of alloced ref structs.
Signed-off-by: Zach Brown <zab@versity.com>
Add a suite of simple kvec functions to work with kvecs that point to
file system keys.
The ones worth mentioning format keys into strings. They're used to
add formatted strings for the keys to tracepoints. They're still a
little rough but this is a functional first step.
Signed-off-by: Zach Brown <zab@versity.com>
Some parts of the ring reading were still using the old 'nr' for the
number of blocks to read, but it's now the total number of blocks
in the ring. Use part instead.
Signed-off-by: Zach Brown <zab@versity.com>
When iterating over items the manifest would always insert
whatever values it found at the caller's key, instead of the
key that it found in the segment.
Signed-off-by: Zach Brown <zab@versity.com>
The initial kvec code was a bit wobbly. It had raw loops, some weird
constructs, and had more elements than we need.
Add some iterator helpers that make it less likely that we'll screw up
iterating over different length vectors. Get rid of reliance on a
tailing null pointer and always use the count of elements to stop
iterating. With that in place we can shrink the number of elements to
just the greatest user.
Signed-off-by: Zach Brown <zab@versity.com>
The estimate for the number of dirty segment bytes was wildly over
calculating the number of segment headers by confusing the length of the
segment header with the length of segments.
Signed-off-by: Zach Brown <zab@versity.com>
The space calculation didn't include the terminating zero entry. That
ensured that the space for the netry would never be consumed. But the
remaining space was used to zero the end of the block so the final entry
wasn't being zeroed.
So have the space remaining include the terminating entry and factor
that into the space consumption of each entry being appended.
Signed-off-by: Zach Brown <zab@versity.com>
The comparisons were a bit wrong when comparing overlaping kvec
endpoints. We want to compare the starts and ends with the ends and
starts, respectively.
Signed-off-by: Zach Brown <zab@versity.com>
We were trying to propagate dirty bits from a node itself when its dirty
bit is set. But it's bits are consistent so it stops immediately. We
need to propagate from the parent of the node that changed.
Signed-off-by: Zach Brown <zab@versity.com>
We went to the trouble of allocating a work queue with one work in
flight but then didn't use it. We could have concurrent trans write
func execution.
Signed-off-by: Zach Brown <zab@versity.com>
A reader that hits an allocated segment would wait on IO forever.
Setting the end_io bit lets readers use written segments.
Signed-off-by: Zach Brown <zab@versity.com>
Specifying the ring blocks with a head and tail index lead to pretty
confusing code to figure out how many blocks to read and if we had
passed the tail.
Instead specify the ring with a starting index and number of blocks.
The code to read and write the ring blocks naturally falls out and is a
lot more clear.
Signed-off-by: Zach Brown <zab@versity.com>
The ring appending next entry header cursor assignment was pointing past
the caller's src header, not at the next header to write to in the
block.
The writing block index and blkno calculations were just bad. Pretend
they never happened.
And finally we need to point the dirty super at the ring index for the
commit and we need to reset the append state for the next commit.
Signed-off-by: Zach Brown <zab@versity.com>
