There was an idea of making the necessary mid/rid adjustments to grm in
lfsr_mdir_commt implicitly.
Explored this, but:
1. It looked like the result would increase code size, though only by
a small (~12 byte) amount.
2. It wouldn't actually work, because lfr_mkdir needs to create a grm
for an mid/rid that doesn't actually exist at the time of commit.
Such a grm can't be created and survive any implicit mid/rid
adjustment.
So scratching that idea for now.
We never encode/decode the grm to/from disk and we always know the
buffer size statically.
Even when we calculate the size for the grm tag, we ignore the encoded
size and optimistically scan for the number of trailing zeros, giving us
a potentially smaller gdelta.
This change drops the encoded length completely in grm encoding/decoding
functions, assuming all related buffers are statically sized and padded
with zeros.
This also means you can't forget to zero the buffer when encoding, which
was already overlooked several times, leading to internal garbage on
disk. So that's nice.
This is entirely a pragmatic change, lfsr_mdir_commit already does
several hairy things with grm tags, decoding, fixing, reencoding, etc,
so it makes sense to move all the encoding logic into lfsr_mdir_commit.
This leads to a couple optimizations:
- We don't need to decode the grm to apply any last minute fixes.
- By allowing the grm arugment to be mutated (they are just sitting on
the stack anyways, we need a copy in case we back out of change due to
error), we can apply and save any grm fixes in the grm argument
itself.
This means we only need to fix the grm at most once, after any mtree
modifications.
Which in turn saves some code and stack cost:
code stack
before: 22930 2392
after: 22706 (-1.0%) 2344 (-2.0%)
Unfortunately the previous attempt to fix the dir seek system didn't
really work. Using a packed mid/rid integer for the offset is tempting,
but since mid/rid can change with any metadata id change in the
filesystem, dir tell offsets would become invalidated if you modified
files in unrelated directories, which isn't great and likely to catch
users by surprise.
This solution builds on the previous dir offset design, which tracks the
dstart-relative position independently from the current mid/rid in our
directory. To update this correctly when there are unrelated changes to
the filesystem, we need to know if metadata id changes are in the range
between our directories dstart and current mid/rid. This in turn means
we need to track our dstart. So our opened directories need three
separate pointers we need to update on every mdir commit:
dir->pos
|
.-------+-------.
a b c d e f g h i j k l m n o p
^ ^
| |
dir->dstart dir->mdir
This has quite a few moving parts, which I was hoping to avoid.
Fortunately we don't need a second mdir, so the RAM cost is pretty
small.
We can also drop dir->did, since the dstart mid/rid render it redundant,
which is interesting.
This is an attempt to fix issues with dir seeking in a filesystem
undergoing changes. The problem with the previous dstart-relative
position encoding is that if we deleted/created new entries outside of
our current directory, we didn't if they were inside or outside of the
current directory, so we couldn't always update our position correctly.
Instead of using a dstart-relative position, this solution crams both
the mid and rid into a single 31-bit integer. Things get a bit tight
here, so we use the current block_size as a heuristic for how many
possible rids we can ever have in a single mdir. The idea is the larger
the rid encoding needs to be, the smaller the mid encoding needs to be,
and we should, _roughly_, approach the same encoding limitation we would
have to dstart-relative position anyways.
Making some assumptions about the maximum possible number of rids in a
block gives us at most ~block_size/8 rids per mdir.
So for 4096 byte blocks (note the exact encoding is dynamic):
sbbbbbbb bbbbbbbb bbbbbbbr rrrrrrrr
^'-----------+----------''----+---'
'------------|----------------|----- sign bit (used for errors)
'----------------|----- 22-bit metadata bid
'----- 9-bit metadata rid
Note this introduced as new, significant limitation on the number of
total mdirs in the system. Normally I would be against this solution for
that reason, however if we adopt this encoding elsewhere in the system it
may improve some RAM cost and in general simplify things by being able to
store any mid in a single integer. More work needs to be done here...
This approach needs some fleshing out and has its own issues (the
offset returned by tell quickly becomes out of date if the filesystem
is modified, but is that really a problem?), but it improves over the
previous implementation by making tell always correct at that moment.
The idea here is to combine the current mtree size with the theoretical
upper bound on the number of directories in a single mdir, assuming our
block size, to give us a heuristic for did truncation that does not
require any extra state.
- Each directory needs 1 name tag, 1 did tag, and 1 dstart
- Each tag needs ~2 alts with our current compaction strategy
- Each tag/alt encodes to a minimum of 4 bytes
- We can also assume ~1/2 block utilization due to our split threshold
This gives us ~3*3*4*2 or ~72 bytes per directory at minimum, or
rounding down, ~block_size/32 directories per mdir.
This is a nice number because for common NOR flash geometry,
4096/32 = 128, so a filesystem with a single mdir encodes dids in a
single byte.
The biggest benefit though is being able to drop the mlimit state from
the lfs_t struct.
---
Unfortunately, this change revealed several bugs.
It turns out __builtin_clz in GCC is undefined at 0, which caused our
lfs_nlog2 function to return incorrect values at 1. This was causing
our dids to all collide when the mtree was inlined, which was resolved
by the linear scanning that resolves dids, but was severely limiting
what exactly our tests covered.
Now that this is fixed (with a simple if statement in lfs_nlog2,
lfs_nlog2 now always has defined behavior, even at 0), several bugs
needed fixing:
- We update the rid based on attrs in lfsr_mdir_commit before updating
the mdir. If we have multiple attrs this causes the assert on
rid-in-bounds to trigger incorrectly. Just removed that assert for now.
- We needed to adjust second grms if they are affected by the fixing
of the first grm.
- Directory position updates are incorrectly updated if an unrelated
weight change occurs before an opened directory, but is not a part of
that opened directory.
This is NOT fixed yet, the current implementation is just broken
enough that I've just ripped it out for now (it was causing the
read_with_rms test to fail because pos backed up into the "."/".."
entries).
This needs some thinking to fix.
Because of that last, unfixed bug, tests are not all passing at the
moment. To pass testing -DSEEK=0 is needed to disable the failing tests.
- Prevented removing and renaming of the root directory. This is done by
repurposing the INVAL error in lfsr_mtree_lookup to indicate the
found entry is the root.
The root entry has special behavior in almost every function, owing to
the fact it doesn't really have an mid/rid. So I think this is a
reasonable approach.
- Added support for lfsr_stat of the root directory.
- Fixed off-by-two in lfsr_dir_seek thanks to the "." and ".." entries.
Humorously there is a comment noting this but the code didn't
actually match the comment.
Unfortunately the powerloss testing risks being a big time sink.
Figuring out the best scale of powerloss testing during normal testing
is probably going to be a constant balancing act.
With a bit of color, this is very useful for debugging and finding
incorrect dstart/grm situations.
This was used to find and fix the bugs in the previous commit.
Mainly trying to match the tests over mkdir/rm, which seem to have a
good amount of coverage.
- Fixed issue where move's desination rid wasn't updated correctly if
the destination split.
- Prevented renaming into nonexistant directories.
- Fixed neighboring rid adjustment in rename (+1 not -1 silly).
- Fixed erronously updating the grm's rid during lfsr_fs_fixgrm. In the
"I can't believe this ever worked" category, it seems this usually
didn't cause issues since mid was often marked as removed, making the
erronously updated rid ignored.
Only simple tests right now, but the theory is sound.
This mainly required the addition of the fancy in-device move attribute,
which copies all tags associated with an rid from one rbyd to another in
a single transaction.
This is a carryover from the previous littlefs implementation, though it
is easier to implement here since it is effectively a range query on the
rbyd tree, which trees are really good at. This was intentional.
Oh and I suppose this also required implementing lfsr_rename, which has
a few corner cases to watch out for.
It is nice that both lfsr_remove and lfsr_rename can rely on
lfsr_fs_fixgrm to finish all of the removes, which wasn't previously
reasonable due to the overhead of deorphaning.
Ugh. I overlooked a weird corner case in rename's behavior that requires
changes to the grm to support.
POSIX's rename, which lfsr_rename is trying to match, supports renaming
files over existing files, effectively removing the previous file during
the rename.
This is supported, even if the files are directories, but with the
additional requirement that the previous directory is empty (matching
the behavior of lfsr_remove).
This creates a weird situation for littlefs. In order to remove
directories in littlefs, we need to atomically remove both the dstart
entry that reserves the directory's did and the directories entry in its
parent. This is made possible by using the grm to mark one entry as
pending removed while removing the other.
But in order to rename atomically, we need to use the grm to mark the
source of the rename as removed while creating/replacing the destination
of the rename.
So we end up needing two grms simultaneously.
This is extra annoying because the niche case of renaming a directory
over another empty directory is the only case where we need two grms,
but this requirement almost doubles the grm size both in-ram and
reserved in every mdir, from 11 bytes to 21 bytes, and increases the
lfs_t size by 28 bytes.
---
Anyways, this commit extends the grm to support up to two pending removes.
Fortunately the implementation was simple since we already have a type
field that can be extended, and grm operations just needed to be
changed from if statements to for loops.
Hopefully third times the charm.
The previous solution pretty bluntly did not work outside of the
recursive remove case, because the moment we mark the rid as deleted,
the directory positions no longer get updates. It's not possible to
update the directory position because we don't know how it maps into our
mtree without a full seek from the dstart.
After staring at it a bit, I think this solution should work:
1. Instead of marking the mid/rid as removed when dropping an mdir, we
set the weight to zero and the trunk to zero, causing mdir lookups to
return NOENT without actually going to disk.
This is very important since later mdirs could be allocated on the
same block, and going to disk can result in a corrupted lookup.
2. Eagerly seek to the next mid/rid after every lfsr_dir_read call. This
puts us in a position where rid can be >= the current mdir weight
without issues, and avoids degenerate cases that may be caused by
recursive removes.
3. If we remove an opened dir, instead of marking the mdir as deleted,
move the rid to the next rid. If the mdir was dropped, this leaves us
with rid == mdir weight, and the mdir trunk == 0.
The rid == mdir weight also occurs when we are creating a new file, so
we have a bit of common behavior we can rely on. We just need to make
sure that mdir updates respect the rid == mdir weight situation.
4. On each lfsr_dir_read call, we do an mtree seek of zero. This just
serves to fix our mdir if our rid == mdir weight, without much
additional code (yay for code reuse).
The use of weight=0, trunk=0, for a dropped mdir here is key, and makes
me wonder if this is a better indicator of a dropped mdir than another
reserved mid value. This probably deserves some investigation later.
Recursive removes is proving more challenging to implement than I
expected. The problem with the previous approach is that it moved the
mid into a potentially non-sensical position with the expectation it
would be updated in lfsr_dir_read because the rid overflows the current
weight (since dropping mdirs always set the weight to zero).
But if an unrelated mdir commit followed that happened to touch that
nonsense mid, the mdir would incorrectly be updated to the previous
block, causing problems for the dir's read state.
---
The solution here is to toss all of that out and rely solely on directory
position updates, which are a bit simpler.
So in lfsr_dir_read, if our mid/rid is deleted, we perform a full
rewind+seek to the new position. This can be more costly, but since the
most common case, recursive removal, leaves us with all mid/rids < pos
deleted, it should only add a single mtree lookup per lfsr_dir_read.
Also added prototypes for dir seek/tell/rewind, since we're using
they're logic for this. Though these aren't yet tested. These are built
on the new function lfsr_mtree_seek which captures the common logic of
seek over multiple mdirs in the mtree efficiently, and skips unnecessary
rid lookups where possible.
"Recursion" here just refers to the ability to remove entries in a
directory while iterating over it. This is very useful when you just
want a directory gone, and can be extended to a "true" recursive remove
straightforwardly. This mainly tests that mid/rid updates in opened
mdirs are correct.
To make this work, we need to update opened dirs differently than files,
since opened dirs do not get marked as removed when its rid is removed
and contain an additional position in the dir that needs to be updated.
To keep track of the different types, littlefs now contains 2
linked-lists for opened mdirs. Maybe these should be correctly typed,
but by hiding the specific types behind an array of mdir linked-lists,
we can more efficiently iterate over both lists when necessary.
We should probably compare this approach to the type-tagged approach in
the previous littlefs implementation, but I think the idea of an array
of type-hidden linked-lists just didn't come to me then. There was also
a bit more room in the mdir structs to hide a 1-bit type field. The mdir
structs here are getting pretty squeezed since they are used everywhere.
These mirror the lfsr_mkdir tests, but backwards.
It's interesting to note the rm powerloss testing is much slower than
mkdir powerloss testing. This is because the rm tests can make
significant backwards progress if power is lost (these tests both make
and remove dirs), but mkdir tests always make forward progress (by only
making dirs).
In theory this is pretty much the same as lfsr_mkdir, but backwards.
The main work was making the interactions between removing mids/rids and
the grm correct. This ends up meaning we just need to update the grm on
any mid/rid update the same way we update the list of opened mdirs.
On the plus side, it turned out to be possible to deduplicate the mdir
uninlining route a bit, by adding range argument to lfsr_mdir_commit_
and changing the write of the newly uninlined mtree/mdir to marking
mtree as dirty and then joining the common path.
This lets us move the pre-commit round of grm updates into a single
location in lfsr_mdir_commit, removing and extra function definition and
the related state marshalling while also simplifying the control-flow.
This also raises the question, can more lfsr_mdir_commit be deduplicated
more? Uninlining is a infrequent operation we don't really need to
optimize for.
---
Testing lfsr_remove also found a bug related to incorrect propagation of
when the mroot becomes "unerased" (when rbyd overflows). This raises the
concern that we're not propagating unerased-states very rigorously, and
unexpected errors may not allow the filesystem to resume.
This has never been in a very good place for littlefs, but would be
worth improving in the future.
Especially with partial builds of tests (TESTS=tests/t1_rbyd.toml)
becoming more useful, these warning have little value and hide other,
actually-useful warnings.
Instead of truncating to exactly 28-bits for nice leb128 alignment, we
now truncate to ~the number of metadata entries, which must be >= ~2x
the number dids since each did needs a dir entry and dstart entry.
This has the downside of needing to actually keep track of an estimate
of the number of metadata entries, which is made a bit difficult due to
integer overflow issues (we can have more than 2^32 metadata entries),
but has the upside of allowing a full 2^32 number of dids worst case.
This is really unlikely, but it's nice to not need another configuration
option to control the did limit.
Another option would be to scale the hashes based on the number dids,
which would be a more direct solution. Unfortunately determining the
number of dids during mount requires a O(m*log(m)) scan of each rbyd
to find either dir entries or dstart entries. This solution can easily
end up with an overestimate, but only needs to weight of each rbyd which
can be (and already is) found in O(m).
Instead of iterating over a number of seeds in the test itself, the
seeds are now permuted as a part of normal test defines.
This lets each seed take advantage of other test features, mainly the
ability to test powerlosses heuristically.
This is probably how it should have been done in the first place, but
the permutation tests can't do this since the number of permutations
changes as the size of the test input changes. The test define system
can't handle that very well.
The tradeoffs here are:
- We can't do cross-fuzz checks, such as the balance checks in the rbyd
tests, though those really should be moved to benchmarks anyways.
- The large number of cheap fuzz permutations skews the total
permutation count, though I'm not sure this matters.
before: 3083 permutations (-Gnor)
after: 409893 permutations (-Gnor)
To help with this, added TEST_PL, which is set to true when powerloss
testing. This way tests can check for stronger conditions (no EEXIST)
when not powerloss testing.
With TEST_PL, there's really no reason every test in t5_dirs shouldn't
be reentrant, and this gives us a huge improvement of test coverage very
cheaply.
---
The increased test coverage caught a bug, which is that gstate wasn't
being consumed properly when mtree uninlining. Humorously, this went
unnoticed because the most common form of mtree uninlining, mdir splitting,
ended up incorrectly consuming the gstate twice, which canceled itself
out since the consume operation is basically just xor.
Also added support for printing dstarts to dbglfs.py, to help debugging.
The grm bugs were mostly issues with:
1. Not maintaining the on-disk grm state in RAM (lfs->grm) correctly,
this needs to be updated correctly after every commit or littlefs
gets a confused.
2. lfsr_fs_fixgrm got a bit confused when it was missed when changing
the no-rm encoding from 0 to -2. Added some inline functions to help
avoid this in the future.
3. Leaking information due to mixing fixed sized and variable sized
encodings of the grm delta in places. This is a bit tricky to write
an assert for as we don't parse the full grm when we see a no-rm grm.
This makes it easier to read the output, at a cost of these scripts not
terminating if the underlying call sctucture contains loops.
Previously these scripts would not terminate, but at least output the
call tree as they visit each function. This was hard to read, and wasn't
really that useful? If you hit a case with infinite recursion, you can
limit the output size explicitly with -Z.
Note this also drops --tree in stack.py. Since we get more readable
output, this flag is less useful. This simplifies the script a bit.
- Changed how names are rendered in dbgbtree.py/dbgmtree.py to be
consistent with non-names. The special rendering isn't really worth it
now that names aren't just ascii/utf8.
- Changed the ordering of raw/device/human rendering of btree entries to
be more consistent with rendering of other entries (don't attempt to
group btree entries).
- Changed dbgmtree.py header to show information about the mtree.
This implementation is in theory correct, but of course, being untested,
who knows?
Though this does come with remounting added to all of the directory
tests. This effectively tests that all of the directory creation tests
we have so far maintain grm=0 after each unmount-mount cycle. Which is
valuable.
This has, in theory, global-removes (grm) being written out as a part of
of directory creation, but they aren't used in any form and so may not
be being written correctly.
But it did require quite a bit of problem solving to get to this point
(the interactions between mtree splitsand grms is really annoying), so
it's worth a commit.
This bug was just overlooked in testing the mtree, fortunately dir
fuzzing found it. Though since this depends on neighboring mdirs, it
probably would have been found quicker with smaller block sizes. At the
moment I am only testing on NOR-liked geometry (4KiB blocks).
The fix is easy, we can use the difference in the mtree size to
determine if a split or drop happened in mdir commit, since at most one
of these can happen on any mdir commit.
Also added an explicit test for mid updates when splitting and dropping.
lfsr_stat is really a directory operation underneath, so it's good to
add to our testing while we are building up the dir tests.
It's interesting to note lfsr_stat and lfsr_dir_read are less
deduplicatable than their previous versions, since lfsr_stat can get
most of it's info from lfsr_mtree_pathlookup. Though there will probably
need to be some code sharing when we get to files with sizes.
- Checksum collisions
- Collisions with root did
- Collisions needing wraparound
- Possible leb128 encoding issues
Sure enough the last one caught an off-by-one error in our calculation
of the leb128 encoded size. I sort of expected a bug there, since it's
rather nuanced math, so it's good to have test coverage now.
The main issues:
- The addition of the root's dstart entry during lfsr_format throws off
our mtree tests. It's a bit of a hack, but for now I am just manually
deleting the root's dstart entry at the beginning of each tests.
It might be possible to make the mtree tests work around the root's
dstart, but it seems to cause problems for when exactly the mtree
splits.
- btree dnamelookup and mdir dnamelookup need different things from
the rbyd dnamelookup when the dname is not found. The btree lookup
needs the largest branch smaller than the dname, since this is the
"bucket" containing our dname, while the mdir dnamelookup needs
the id that _follows_ the id smaller than the dname, since insertion
causes all ids >= the inserting id to shift up.
The solution here is to make rbyd dnamelookup behave as expected by
btree dnamelookup. btree needs more info about the branch (weight
mostly), so this avoids more issues. mdir dnamelookup adjusts the
id as needed, which costs a bit of code, but makes things work.
Fortunately, mdir dnamelookup can assume the weight is 1, which
simplifies things a bit.
This makes it now possible to create directories in the new system.
The new system now uses a single global "mtree" to store all metadata
entries in the filesystem. In this system, a directory is simply a range
of metadata entries. This has a number of benefits, but does come with
its own problems:
1. We need to indicate which directory each file belongs to. To do this
the file's name entry has been changed to a tuple of leb128-encoded
directory-id + actual file name:
01 66 69 6c 65 2e 74 78 74 .file.txt
^ '----------+----------'
'------------|------------ leb128 directory-id
'------------ ascii/utf8 name
If we include the directory-id as part of filename comparison, files
should naturally be next to other files in the same directory.
2. We need a way allocate directory-ids for new directories. This turns
out to be a bit more tricky than I expected.
We can't use any mid/bid/rid inherent to the mtree, because these
change on any file creation/deletion. And since we commit the did
into the tree, that's not acceptable.
Initially I though you could just find the largest did and increment,
but this gives you no way to reclaim deleted dids. And sure, deleted
dids have no storage consumption, but eventually you will overflow
the did integer. Since this can suddenly happen in a filesystem
that's been in a steady-state for years, that's pretty unnacceptable.
One solution is to do a simple linear search over the mtree for an
unused did. But with a runtime of O(n^2 log(n)), this raises
performance concerns.
Sidenote: It's interesting to note that the Linux kernel's allocation
of process-ids, a very similar problem, is surprisingly complex and
relies on a radix-tree of bitmaps (struct idr). This suggests I'm not
missing an obvious solution somewhere.
The solution I settled on here is to instead treat the set of dids as
a sort of hash table:
1. Hash the full directory path into a did.
2. Perform a linear search until we have no collision.
leb128(truncate28(crc32c("dir")))
.--------'
v
9e cd c8 30 66 69 6c 65 2e 74 78 74 ...0file.txt
'----+----' '----------+----------'
'-----------------|------------ leb128 directory-id
'------------ ascii/utf8 name
Worst case, this can still exhibit the worst case O(n^2 log(n))
performance when we are close to full dids. However that seems
unlikely to happen in practice, since we don't truncate our hashes,
unlike normal hash tables. An additional 32-bit word for each file
is a small price to pay for a low-chance of collisions.
In the current implementation, I do truncate the hash to 28-bits.
Since we encode the hash with leb128, and hashes are statistically
random, this gives us better usage of the leb128 encoding. However
it does limit a 32-bit littlefs to 256 Mi directories.
Maybe this should be a configurable limit in the future.
But that highlights another benefit of this scheme. It's easy to
change in the future without disk changes.
3. We need a way to know if a directory-id is allocated, even if the
directory is empty.
For this we just introduce a new tag: LFSR_TAG_DSTART, which
is an empty file entry that indicates the directory at the given did
in the mtree is allocated.
To create/delete these atomically with the reference in our parent
directory, we can use the GRM system for atomic renames.
Note this isn't implemented yet.
This is also the first time we finally get around to testing all of the
dname lookup functions, so this did find a few bugs, mostly around
reporting the root correctly.
The plan is that names in littlefs now include a directory-id prefixed
as a single leb128.
01 66 69 6c 65 2e 74 78 74 .file.txt
^ '----------+----------'
'------------|------------ leb128 directory-id
'------------ ascii/utf8 name
Unfortunately, while this is easy for read/compare operations to implement,
it creates a bit of a problem for writes. We can't allocate a new buffer
for each name, so we need some sort of extra mechanism.
The solution here is to just add a did member to lfsr_data_t that is
written when non-negative. This works, though it does introduce some
complexity.
Fortunately, did in lfsr_data_t is somewhat free when
sizeof(void*) == sizeof(lfs_size_t), due to the union with disk
references.
It doesn't make sense to test more complex logic, such as t2_btree.toml,
when the logic it is built on, t1_rbyd.toml, does not past testing. The
test runner already guarantees a consistent lexicographic order, so all
we need to do is renamed these from test_* -> tn_*.
Note, if we every have more than 10 tests, we will need to bump up the
number of digits for all tests, so t1_rbyd.toml -> t01_rbyd.toml. This
is the main downside of lexicographic ordering. But we'll cross that
bridge when we get to it.
Took the opportunity to make some allocator tweaks:
- Renamed lfs.free -> lfs.lookahead, it's previous name did cause some
confusion.
- Renamed lfs.free.off -> lfs.lookahead.start
- Renamed lfs.free.i -> lfs.lookahead.next
- Renamed lfs.free.ack -> lfs.lookahead.acked
- Changed bitmap from using 32-bit words to using 8-bit bytes, dropping
the alignment requirement. One of the reasons for 32-bit alignment was
an attempt at future proofing for some sort of free-list.
This never landed, and if it did, it could have been provided without
breaking backwards compatiblity via an additional config option, at a
minor RAM cost.
We never used ffs/clz instructions for this bitmap, so I don't think
using 32-bit words offers much advantage. It just creates another
potential issue for users if their lookahead buffer is unaligned.
These changes should probably also be upstreamed to the current version.
They don't depend on anything rbyd specific.
Note, at some point lfs_alloc will need to be extended to mark block tags,
etc, as in-use during traversal.
- Renamed mpair -> mptr, may have >2 blocks in the future.
- Renamed branch -> bptr for consistency.
- Renamed other_block -> redund_rbyd.
- Changed comparison functions to use -1, 0, +1, even for unordered
types.
- Added lfs_cmp function for unioning comparisons with signed errors.
Now that tree rebalancing is implemented and needed a null terminator
anyways, I think it's clear that the benefit of the alt-always pointers
as trunk terminator has pretty limited value.
Now a null or other tag is needed for every trunk, which simplifies
checks for end-of-trunk.
Alt-always tags are still emitted for deletes, etc, but there their
behavior is implicit, so no special checks are needed. Alt-always tags
are naturally cleaned up as a part of rbyd pruning.
- Since both trunks emit altle tags now, reworked the trunk merging to
reuse more code.
- Changed lfsr_mdir_fetch to rely on trunk=0 to detect the no-commit
state. This is purely for consistency.
This actually broke some tests that committed nothing, resulting in
trunk-less rbyds, which is a bit concerning, but I don't think
trunk-less rbyds will ever be valid in our system?
- Simplified lfsr_rbyd_estimate calculation, merged lfsr_rbyd_bisect
since this is almost always needed after an estimated failure, and
that way dependent function have to call fewer things to implement
rbyd splitting.
- Dropped vestigial names for now, though need to revisit this later.
After these changes the code size difference between rebalancing and
appending is a bit smaller at ~392 bytes: 16208 -> 16600 (+2.4%). It's
interesting to note this is mostly because the conservative overhead
calculation is easier with rebalancing.
In theory this also saves some stack usage, but since I'm measuring
maximum stack usage it doesn't show up since it's not on the deepest
path.
This isn't actually for performance reasons, but to reduce storage
overhead of the rbyd metadata tree, which was showing signs of being
problematic for small block sizes.
Originally, the plan for compaction was to rely on the self-balancing
rbyd append algorithm and simply append each tag to a new tree.
Unfortunately, since each append requires a rewrite of the trunk
(current search path), this introduces ~n*log(n) alts but only uses ~n alts
for the final tree. This really starts to put pressure on small blocks,
where the exponential-ness of the log doesn't kick in and overhead
limits are already tight.
Measuring lfsr_mdir_commit code size, this shows a ~556 byte cost on
thumb: 16416 -> 16972 (+3.4%). Though there are still some optimizations
on the table, this implementation needs a cleanup pass.
alt overhead code cost
rebalance: <= 28*n 16972
append: <= 24*n*log(n) 16416
Note these all assume worst case alt overhead, but we _need_ to assume
worst case for our rbyd estimations, or else the filesystem can get
stuck in unrecoverable compaction states.
Because of the code cost I'm not sure if rebalancing will stay, be
optional, or replace append-compaction completely yet.
Some implementation notes:
- Most tree balancing algorithms rely on true recursion, I suspect
recursion may be a hard requirement in general, but it's hard to find
bounded-ram algorithms.
This solution gets around the ram requirement by leveraging the fact
that our tags exist in a log to build up each layer in the tree
tail-recursively. It's interesting to note that this is a special
case of having little ram but lots of storage.
- Humorously this shouldn't result in a performance improvement. Rbyd
trees result in a worst case 2*log(n) height, and rebalancing gives us
a perfect worst case log(n) height, but, since we need an additional
alt pointer for each node in our tree, things bump back up to 2*log(n).
- Originally the plan was to terminate each node with an alt-always tag,
but during implementation I realized there was no easy way to get the
key that splits the children with awkward tree lookups. As a
workaround each node is terminated with an altle tag that contains the
key followed by an unreachable null tag. This is redundant information,
but makes the algorithm easier to implement.
Fortunately null tags use the smallest tag encoding, which isn't that
small, but that means this wastes at most 4*n bytes.
- Note this preserves the first-tag-always-ends-up-at-off=0x4 rule, which
is necessary for the littlefs magic to end up in a consistent place.
- I've dropped dropping vestigial names for now, which means vestigial
names can remain in btrees indefinitely. Need to revisit this.
This really helps just make the mess that is lfsr_mdir_commit readable,
though seems to only save ~200 bytes. The number of arguments that need
to be set up in order to call lfsr_mdir_commit seem to be offsetting
code savings.
It's interesting to note more code could probably be saved if
lfsr_mtree_split_ was inlined into lfsr_mdir_commit, with one of the
two invocations code using a goto both to jump in and jump out of the
common split logic. But I'm not about to go down that sort of hellish
path.
- Finally figured out how to test multiple mroot extensions without an
allocator, though hopefully forcing PROG_SIZE doesn't break test
framework things at some point...
- Added tests that magic string is always in the same place. This isn't
strictly required for littlefs to work, but is a nice feature to have.
Of course, the new tests found a bug, but it was in a surprisingly
place. Accidentally allowed the revision count to be uninitialized when
compacting the mroot. At least there's a test that covers this now.
Now we read the revision count on-demand, trading off some extra reads
for a smaller lfsr_rbyd_t struct.
I believe this is worth it because:
1. We're created a lot of lfsr_rbyd_t structs as a part of the relatively
complicated mdir/btree commit logic in order to safely fallback on errors.
2. We don't really need the revision count for our Cow btrees, so we
only need to read the revision count on mdir fetch (which we were
already reading too many times), on mdir compact, and on rbyd fetch
as a part of checksum calculation.
This really only adds a O(1) cost when we are compacting, which is rather
small.
Current measurements:
code: 8980 -> 9036 (+0.6%)
stack: 1024 -> 1000 (-2.3%)
Though note this is currently without any mdir/btree commit code being
dragged in.
More code reuse => smaller code size generally. Though this adds another
special mid value.
Note that the range in this compact excludes all tags, we really only
want the revision count. It's tempting to implicitly copy the
magic/config via the compact range, but this risks included user
attributes and other things that we really don't want cluttering up our
mroot chain.
This should have been done as a part of the earlier tag reencoding work,
since having the block at the end was what allowed us to move the
redund-count out of the tag encoding.
New encoding:
[-- 32-bit csum --]
[-- leb128 weight --]
[-- leb128 trunk --]
[-- leb128 block --]
Note that since our tags have an explicit size, we can store a variable
number of blocks. The plan is to use this to eventually store redundant
copies for error correction:
[-- 32-bit csum --]
[-- leb128 weight --]
[-- leb128 trunk --]
[-- leb128 block --] -.
[-- leb128 block --] +- n redundant blocks
[-- leb128 block --] |
... -'
This does have a significant tradeoff, we need to know the checksum size
to access the btree structure. This doesn't seem like a big deal, but
with the possibility of different checksum types may be an annoying
issue.
Note that FCRC was also flipped for consistency.
Wide tags are a happy accident that fell out of the realization that we
can view all subtypes of a given tag suptype as a range in our rbyd.
Combining this with how natural it is to operate on ranges in an rbyd
allows us to perform operations on an entire range of subtypes as though
it were a single tag.
- lookup wide tag => find the smallest tag with this tag's suptype, O(log(n))
- remove wide tag => remove all tags with this tag's suptype, O(log(n))
- append wide tag => remove all tags with this tag's suptype, and then
append our tag, O(log(n))
This is very useful for littlefs, where we've already been using tag's
subtypes to hold extra type info, and have had to rely on awkward
alternatives such as deleting existing subtypes before writing our new
subtype.
For example, when committing file metadata (not yet implemented), we can
append a wide struct tag to update the metadata while also clearing out any
lingering struct tags from previous commits, all in one rbyd append
operation.
This uses another mode bit in-device to change the behavior of
lfsr_rbyd_commit, of which we have a couple:
vwgrtttt 0TTTTTTT
^^^^---^--------^- valid bit (currently unused, maybe errors?)
'||---|--------|- wide bit, ignores subtype (in-device)
'|---|--------|- grow bit, don't create new id (in-device)
'---|--------|- rm bit, remove this tag (in-device)
'--------|- 4-bit suptype
'- leb128 subtype
Christopher Haster