* reset local start time at each recording boundary
* include the final frame's time when converting from wall to media time
Together at 30 fps these caused about 15 seconds of drift over 4 hours.
As noted in the changelog, this problem caused restarts to take a long
time, and the SQLite database to (slowly) grow without bound.
The problem was simple: the directory worker incorrectly aborted
`collect_garbage` if nothing was watching the result. For some
operations (e.g. reading from sample files) this is the correct
behavior, but in this case it contradicts the `Pool::collect_garbage`
contract.
`v0.7.26` introduced a busy loop error when the dir pool was shutting
down with multiple streams. This commit tests and fixes this scenario.
Also fixes a second problem in dropping a writer without an explicit
close would panic in debug mode about using (dropping) the blake3 hasher
while holding a lock. The check is meant to detect slow operations while
holding a lock. The panic wouldn't happen in release mode and probably
isn't a significant source of slowness, but fix it anyway.
This puts a lot of state into a new `LockedStream` that was formerly
guarded more coarsely by the recording lock or more granularly by
`RecordingToInsert` locks. And it restructures `LockedDatabase::flush`
to avoid holding this lock while performing SQLite operations.
Streamers write only into a recent frames buffer in RAM, which the
directory pool workers asynchronously write to disk approximately once
per GOP, where formerly there was a write call per frame.
Before, streamers could fall behind due to slow sample file directory
disk I/O calls or if writing got stuck in a retry loop, for example if
out of disk space. If Moonfire falls too far behind in receiving,
cameras tend to disconnect or (sadly) mis-frame messages because they do
not handle this case well. This removes a confusing potential cause of
such errors.
In the new model, streamers never wait for disk I/O. Frames are buffered
up to a limit. Recordings are aborted if the I/O pool falls too far
behind. There's no retry loop.
This roughly halves CPU usage as thread hand-offs and syscalls are less
common. It also enables several desirable future changes:
* audio recording (which radically increases the number of frames and
thus would otherwise significantly increase CPU usage).
* "monitor mode" (allowing live view without recording to disk),
* further CPU reductions for live view (by skipping all SQLite ops and
disk I/O).
In the new model, streamers also rarely wait for SQLite database
operations either—now only on video parameter changes. That might go
away in an upcoming change too.
The "syncer" (which synced files/dirs and kept track of what has to be
flushed into the SQLite database when) has mostly gone away too. The
directory pool workers handle the syncing now, and then they notify
a new "flusher" for the latter portion.
I'd like to go further still and have the database as a whole be owned
by a dedicated thread, the only one that uses the SQLite connection.
This prevents the tokio threads from blocking on SQLite operations.
I tried out doing so before making the streamers less dependent on the
db, but the extra thread handoffs were awful for performance.
Now, all IO happens on each directory's two-thread pool with threads
named `dir-<PATH>`. Formerly, IO could happen on the per-stream writer
threads, per-directory reader threads, and occasional on even the tokio
worker threads.
The recording path uses more CPU than before due to extra thread
hand-offs. That will change in the next commit, which will radically
change the approach to writing.
`slices::Slices::get_range` was too closely following the example of
`http_serve::Entity::get_range`:
The latter is once-per-request, so just boxing is low-cost and makes
sense to easily avoid monomorphization bloat when there are potentially
many types of entity streams in one program. In Moonfire, it's used with
different streams defined in the `moonfire_nvr::web::mp4`,
`moonfire_nvr::bundled_ui`, and `http_serve::file` modules. Putting them
all into a single boxless enum would be a pain. In particular, the last
one is not a nameable type today and would need more generic parameters
to implement the caller-demanded `Entity` definition.
The former is once-per-slice, there are tons of slices per request, and
it's easy to define a two-case enum right where it's needed. So the
trade-off is quite different.
Also fix up some out-of-date comments.
I think this is a big improvement in readability.
I removed the `lnav` config, which is a little sad, but I don't think it
supports this structured logging format well. Still seems worthwhile on
balance.
I saw this recently while working on new-schema. It was probably due
to some manual upgrade or downgrade I did rather than an actual bug.
Improve debuggability a little nonetheless.
Reading from the mmap()ed region in the tokio threads could cause
them to stall:
* That could affect UI serving when there were concurrent
UI requests (i.e., not just requests that needed the reads in
question anyway).
* If there's a faulty disk, it could cause the UI to totally hang.
Better to not mix disks between threads.
* Soon, I want to handle RTSP from the tokio threads (#37). Similarly,
we don't want RTSP streaming to block on operations from unrelated
disks.
I went with just one thread per disk which I think is sufficient.
But it'd be possible to do a fixed-size pool instead which might improve
latency when some pages are already cached.
I also dropped the memmap dependency. I had to compute the page
alignment anyway to get mremap to work, and Moonfire NVR already is
Unix-specific, so there wasn't much value from the memmap or memmap2
crates.
Fixes#88
Scott Lamb