The root of the problem was the notorious Python quirk with mutable
default parameters. The default defines for the TestSuite class ended
up being mutated as the class determined the permutations to test,
corrupting other test's defines.
However, the only define that was mutated this way was the CACHE_SIZE
config in test_entries.
The crazy thing was how this small innocuous change would cause
"./scripts/test.py -nr test_relocations" and "./scripts/test.py -nr"
to drift out of sync only after a commit spanning the different cache
sizes would be written out with a different number of prog calls. This
offset the power-cycle counter enough to cause one case to make it to
an erase, and the other to not.
Normally, the difference between a successful/unsuccessful erase
wouldn't change the result of a test, but in this case it offset the
revision count used for wear-leveling, causing one run run expand the
superblock and the other to not.
This change to the filesystem would then propogate through the rest of
the test, making it difficult to reproduce test failures.
Fortunately the fix was to just make a copy of the default define
dictionary. This should also prevent accidently mutating of dicts
belonging to our caller.
Oh, also fixed a buffer overflow in test_files.
The power-cycled-relocation test with random renames has been the most
aggressive test applied to littlefs so far, with:
- Random nested directory creation
- Random nested directory removal
- Random nested directory renames (this could make the
threaded linked-list very interesting)
- Relocating blocks every write (maximum wear-leveling)
- Incrementally cycling power every write
Also added a couple other tests to test_orphans and test_relocations.
The good news is the added testing worked well, it found quite a number
of complex and subtle bugs that have been difficult to find.
1. It's actually possible for our parent to be relocated and go out of
sync in lfs_mkdir. This can happen if our predecessor's predecessor
is our parent as we are threading ourselves into the filesystem's
threaded list. (note this doesn't happen if our predecessor _is_ our
parent, as we then update our parent in a single commit).
This is annoying because it only happens if our parent is a long (>1
pair) directory, otherwise we wouldn't need to catch relocations.
Fortunately we can reuse the internal open file/dir linked-list to
catch relocations easily, as long as we're careful to unhook our
parent whenever lfs_mkdir returns.
2. Even more surprising, it's possible for the child in lfs_remove
to be relocated while we delete the entry from our parent. This
can happen if we are our own parent's predecessor, since we need
to be updated then if our parent relocates.
Fortunately we can also hook into the open linked-list here.
Note this same issue was present in lfs_rename.
Fortunately, this means now all fetched dirs are hooked into the
open linked-list if they are needed across a commit. This means
we shouldn't need assumptions about tree movement for correctness.
3. lfs_rename("deja/vu", "deja/vu") with the same source and destination
was broken and tried to delete the entry twice.
4. Managing gstate deltas when we lose power during relocations was
broken. And unfortunately complicated.
The issue happens when we lose power during a relocation while
removing a directory.
When we remove a directory, we need to move the contents of its
gstate delta to another directory or we'll corrupt littlefs gstate.
(gstate is an xor of all deltas on the filesystem). We used to just
xor the gstate into our parent's gstate, however this isn't correct.
The gstate isn't built out of the directory tree, but rather out of
the threaded linked-list (which exists to make collecting this
gstate efficient).
Because we have to remove our dir in two operations, there's a point
were both the updated parent and child can exist in threaded
linked-list and duplicate the child's gstate delta.
.--------.
->| parent |-.
| gstate | |
.-| a |-'
| '--------'
| X <- child is orphaned
| .--------.
'>| child |->
| gstate |
| a |
'--------'
What we need to do is save our child's gstate and only give it to our
predecessor, since this finalizes the removal of the child.
However we still need to make valid updates to the gstate to mark
that we've created an orphan when we start removing the child.
This led to a small rework of how the gstate is handled. Now we have
a separation of the gpending state that should be written out ASAP
and the gdelta state that is collected from orphans awaiting
deletion.
5. lfs_deorphan wasn't actually able to handle deorphaning/desyncing
more than one orphan after a power-cycle. Having more than one orphan
is very rare, but of course very possible. Fortunately this was just
a mistake with using a break the in the deorphan, perhaps left from
v1 where multiple orphans weren't possible?
Note that we use a continue to force a refetch of the orphaned block.
This is needed in the case of a half-orphan, since the fetched
half-orphan may have an outdated tail pointer.
The root of the problem was some assumptions about what tags could be
sent to lfs_dir_commit.
- The first assumption is that there could be only one splice (create/delete)
tag at a time, which is trivially broken by the core commit in lfs_rename.
- The second assumption is that there is at most one create and one delete in
a single commit. This is less obvious but turns out to not be true in
the case that we rename a file such that it overwrites another file in
the same directory (1 delete for source file, 1 delete for destination).
- The third assumption was that there was an ordering to the
delete/creates passed to lfs_dir_commit. It may be possible to force all
deletes to follow creates by rearranging the tags in lfs_rename, but
this risks overflowing tag ids.
The way the lfs_dir_commit first collected the "deletetag" and "createtag"
broke all three of these assumptions. And because we lose the ordering
information we can no longer apply the directory changes to open files
correctly. The file ids may be shifted in a way that doesn't reflect the
actual operations on disk.
These problems were made worst by lfs_dir_commit cleaning up moves
implicitly, which also creates deletes implicitly. While cleaning up moves
in lfs_dir_commit may save some code size, it makes the commit logic much more
difficult to implement correctly.
This bug turned into pulling out a dead tree stump, roots and all.
I ended up reworking how lfs_dir_commit updates open files so that it
has less assumptions, now it just traverses the commit tags multiple
times in order to update file ids after a successful commit in the
correct order.
This also got rid of the dir copy by carefully updating split dirs
after all files have an up-to-date copy of the original dir.
I also just removed the implicit move cleanup. It turns out the only
commits that can occur before we have cleaned up the move is in
lfs_fs_relocate, so it was simple enough to explicitly handle this case
when we update our parent and pred during a relocate.
Cases where we may need to fix moves:
- In lfs_rename when we move a file/dir
- In lfs_demove if we lose power
- In lfs_fs_relocate if we have to relocate our parent and we find it
had a pending move (or else the move will be outdated)
- In lfs_fs_relocate if we have to relocate our predecessor and we find it
had a pending move (or else the move will be outdated)
Note the two cases in lfs_fs_relocate may be recursive. But
lfs_fs_relocate can only trigger other lfs_fs_relocates so it's not
possible for pending moves to spill out into other filesystem commits
And of couse, I added several tests to cover these situations. Hopefully
the rename-with-open-files logic should be fairly locked down now.
found with initial fix by eastmoutain
Also fixed a bug in dir splitting when there's a large number of open
files, which was the main reason I was trying to make it easier to debug
disk images.
One part of the recent test changes was to move away from the
file-per-block emubd and instead simulate storage with a single
contiguous file. The file-per-block format was marginally useful
at the beginning, but as the remaining bugs get more subtle, it
becomes more useful to inspect littlefs through scripts that
make the underlying metadata more human-readable.
The key benefit of switching to a contiguous file is these same
scripts can be reused for real disk images and can even read through
/dev/sdb or similar.
- ./scripts/readblock.py disk block_size block
off data
00000000: 71 01 00 00 f0 0f ff f7 6c 69 74 74 6c 65 66 73 q.......littlefs
00000010: 2f e0 00 10 00 00 02 00 00 02 00 00 00 04 00 00 /...............
00000020: ff 00 00 00 ff ff ff 7f fe 03 00 00 20 00 04 19 ...............
00000030: 61 00 00 0c 00 62 20 30 0c 09 a0 01 00 00 64 00 a....b 0......d.
...
readblock.py prints a hex dump of a given block on disk. It's basically
just "dd if=disk bs=block_size count=1 skip=block | xxd -g1 -" but with
less typing.
- ./scripts/readmdir.py disk block_size block1 block2
off tag type id len data (truncated)
0000003b: 0020000a dir 0 10 63 6f 6c 64 63 6f 66 66 coldcoff
00000049: 20000008 dirstruct 0 8 02 02 00 00 03 02 00 00 ........
00000008: 00200409 dir 1 9 68 6f 74 63 6f 66 66 65 hotcoffe
00000015: 20000408 dirstruct 1 8 fe 01 00 00 ff 01 00 00 ........
readmdir.py prints info about the tags in a metadata pair on disk. It
can print the currently active tags as well as the raw log of the
metadata pair.
- ./scripts/readtree.py disk block_size
superblock "littlefs"
version v2.0
block_size 512
block_count 1024
name_max 255
file_max 2147483647
attr_max 1022
gstate 0x000000000000000000000000
dir "/"
mdir {0x0, 0x1} rev 3
v id 0 superblock "littlefs" inline size 24
mdir {0x77, 0x78} rev 1
id 0 dir "coffee" dir {0x1fc, 0x1fd}
dir "/coffee"
mdir {0x1fd, 0x1fc} rev 2
id 0 dir "coldcoffee" dir {0x202, 0x203}
id 1 dir "hotcoffee" dir {0x1fe, 0x1ff}
dir "/coffee/coldcoffee"
mdir {0x202, 0x203} rev 1
dir "/coffee/warmcoffee"
mdir {0x200, 0x201} rev 1
readtree.py parses the littlefs tree and prints info about the
semantics of what's on disk. This includes the superblock,
global-state, and directories/metadata-pairs. It doesn't print
the filesystem tree though, that could be a different tool.
Also finished migrating tests with test_relocations and test_exhaustion.
The issue I was running into when migrating these tests was a lack of
flexibility with what you could do with the block devices. It was
possible to hack in some hooks for things like bad blocks and power
loss, but it wasn't clean or easily extendable.
The solution here was to just put all of these test extensions into a
third block device, testbd, that uses the other two example block
devices internally.
testbd has several useful features for testing. Note this makes it a
pretty terrible block device _example_ since these hooks look more
complicated than a block device needs to be.
- testbd can simulate different erase values, supporting 1s, 0s, other byte
patterns, or no erases at all (which can cause surprising bugs). This
actually depends on the simulated erase values in ramdb and filebd.
I did try to move this out of rambd/filebd, but it's not possible to
simulate erases in testbd without buffering entire blocks and creating
an excessive amount of extra write operations.
- testbd also helps simulate power-loss by containing a "power cycles"
counter that is decremented every write operation until it calls exit.
This is notably faster than the previous gdb approach, which is
valuable since the reentrant tests tend to take a while to resolve.
- testbd also tracks wear, which can be manually set and read. This is
very useful for testing things like bad block handling, wear leveling,
or even changing the effective size of the block device at runtime.
Even with adding better reentrance testing, the bad-block tests are
still very useful at isolating the block eviction logic.
This also required rewriting a bit of the internal testing wirework
to allow custom block devices which opens up quite a bit more straegies
for testing.
Both test_move and test_orphan needed internal knowledge which comes
with the addition of the "in" attribute. This was in the plan for the
test-revamp from the beginning as it really opens up the ability to
write more unit-style-tests using internal knowledge of how littlefs
works. More unit-style-tests should help _fix_ bugs by limiting the
scope of the test and where the bug could be hiding.
The "in" attribute effectively runs tests _inside_ the .c file
specified, giving the test access to all static members without
needed to change their visibility.
This involved some minor tweaks for the various types of tests, added
predicates to the test framework (necessary for test_entries and
test_alloc), and cleaned up some of the testing semantics such as
reporting how many tests are filtered, showing permutation config on
the result screen, and properly inheriting suite config in cases.
Aside from reworking the internals of test_.py to work well with
inherited TestCase classes, this also provides the two main features
that were the main reason for revamping the test framework
1. ./scripts/test_.py --reentrant
Runs reentrant tests (tests with reentrant=true in the .toml
configuration) under gdb such that the program is killed on every
call to lfs_emubd_prog or lfs_emubd_erase.
Currently this just increments a number of prog/erases to skip, which
means it doesn't necessarily check every possible branch of the test,
but this should still provide a good coverage of power-loss tests.
2. ./scripts/test_.py --gdb
Run the tests and if a failure is hit, drop into GDB. In theory this
will be very useful for reproducing and debugging test failures.
Note this can be combined with --reentrant to drop into GDB on the
exact cycle of power-loss where the tests fail.
This is the start of reworking littlefs's testing framework based on
lessons learned from the initial testing framework.
1. The testing framework needs to be _flexible_. It was hacky, which by
itself isn't a downside, but it wasn't _flexible_. This limited what
could be done with the tests and there ended up being many
workarounds just to reproduce bugs.
The idea behind this revamped framework is to separate the
description of tests (tests/test_dirs.toml) and the running of tests
(scripts/test.py).
Now, with the logic moved entirely to python, it's possible to run
the test under varying environments. In addition to the "just don't
assert" run, I'm also looking to run the tests in valgrind for memory
checking, and an environment with simulated power-loss.
The test description can also contain abstract attributes that help
control how tests can be ran, such as "leaky" to identify tests where
memory leaks are expected. This keeps test limitations at a minimum
without limiting how the tests can be ran.
2. Multi-stage-process tests didn't really add value and limited what
the testing environment.
Unmounting + mounting can be done in a single process to test the
same logic. It would be really difficult to make this fail only
when memory is zeroed, though that can still be caught by
power-resilient tests.
Requiring every test to be a single process adds several options
for test execution, such as using a RAM-backed block device for
speed, or even running the tests on a device.
3. Added fancy assert interception. This wasn't really a requirement,
but something I've been wanting to experiment with for a while.
During testing, scripts/explode_asserts.py is added to the build
process. This is a custom C-preprocessor that parses out assert
statements and replaces them with _very_ verbose asserts that
wouldn't normally be possible with just C macros.
It even goes as far as to report the arguments to strcmp, since the
lack of visibility here was very annoying.
tests_/test_dirs.toml:186:assert: assert failed with "..", expected eq "..."
assert(strcmp(info.name, "...") == 0);
One downside is that simply parsing C in python is slower than the
entire rest of the compilation, but fortunately this can be
alleviated by parallelizing the test builds through make.
Other neat bits:
- All generated files are a suffix of the test description, this helps
cleanup and means it's (theoretically) possible to parallelize the
tests.
- The generated test.c is shoved base64 into an ad-hoc Makefile, this
means it doesn't force a rebuild of tests all the time.
- Test parameterizing is now easier.
- Hopefully this framework can be repurposed also for benchmarks in the
future.
