This fixes a crash that would occasionally occur when visiting profiles.
NdbTxn objects were being deinitialized on different threads from their
initialization, causing incorrect reference count decrements in thread-local
transaction dictionaries. This led to premature destruction of shared ndb_txn
C objects still in use by other tasks, resulting in use-after-free crashes.
The root cause is that Swift does not guarantee tasks resume on the same
thread after await suspension points, while NdbTxn's init/deinit rely on
thread-local storage to track inherited transaction reference counts.
This means that `NdbTxn` objects cannot be used in async functions, as
that may cause the garbage collector to deinitialize `NdbTxn` at the end
of such function, which may be running on a different thread at that
point, causing the issue explained above.
The fix in this case is to eliminate the `async` version of the
`NdbNoteLender.borrow` method, and update usages to utilize other
available methods.
Note: This is a rewrite of the fix in https://github.com/damus-io/damus/pull/3329
Note 2: This relates to the fix of an unreleased feature, so therefore no
changelog is needed.
Changelog-None
Co-authored-by: alltheseas <64376233+alltheseas@users.noreply.github.com>
Closes: https://github.com/damus-io/damus/issues/3327
Signed-off-by: Daniel D’Aquino <daniel@daquino.me>
A race condition was identified where notes would get dropped if they
get indexed in the time window between when a query is made and the subscription is made.
The issue was fixed by making the subscribe call before making the query
call, to ensure we get all notes from that time when we perform the
query.
This dropped the failure rate for ndb subscription tests from about 20%
down to about 4%.
Local relay model issue was not publicly released, which is why the
changelog entry is "none".
Changelog-None
Signed-off-by: Daniel D’Aquino <daniel@daquino.me>
This attempts to reduce race conditions coming from Ndb streaming
functions that could lead to lost notes or crashes.
It does so by making two improvements:
1. Instead of callbacks, now the callback handler uses async streams,
which reduces the chances of a callback being called before the last
item was processed by the consumer.
2. The callback handler will now queue up received notes if there are
no listeners yet. This is helpful because we need to issue the
subscribe call to nostrdb before getting the subscription id and
setting up a listener, but in between that time nostrdb may still
send notes which would effectively get dropped without this queuing
mechanism.
Changelog-Fixed: Improved robustness in the part of the code that streams notes from nostrdb
Signed-off-by: Daniel D’Aquino <daniel@daquino.me>
This commit improves the loading speed for the home timeline (and likely
other areas of the app) by employing various techniques and changes:
- Network EOSE timeout reduced from 10 seconds down to 5 seconds
- Network EOSE does not wait on relays with broken connections
- Offload HomeModel handler event processing to separate tasks to
avoid a large backlog
- Give SubscriptionManager streamers more fine-grained EOSE signals for
local optimization
- Only wait for Ndb EOSE on the home timeline for faster loading
- Add logging with time elapsed measurements for easier identification of
loading problems
Signed-off-by: Daniel D’Aquino <daniel@daquino.me>
This commit takes a step back from the full local relay model by
treating NostrDB as one of the many relays streamed from, instead of the
one exclusive relay that other classes rely on.
This was done to reduce regression risk from the local relay model
migration, without discarding the migration work already done.
The full "local relay model" behavior (exclusive NDB streaming) was
hidden behind a feature flag for easy migration later on.
Closes: https://github.com/damus-io/damus/issues/3225
Signed-off-by: Daniel D’Aquino <daniel@daquino.me>
This commit improves NostrNetworkManager interfaces to be easier to use,
and with more options on how to read data from the Nostr network
This reduces the amount of duplicate logic in handling streams, and also
prevents possible common mistakes when using the standard subscribe method.
This fixes an issue with the mute list manager (which prompted for this
interface improvement, as the root cause is similar to other similar
issues).
Closes: https://github.com/damus-io/damus/issues/3221
Signed-off-by: Daniel D’Aquino <daniel@daquino.me>
This commit implements nostr network subscriptions that survive between
sessions, as well as improved handling of RelayPool opening/closing with
respect to the app lifecycle.
This prevents stale data after users swap out and back into Damus.
Signed-off-by: Daniel D’Aquino <daniel@daquino.me>
Previously, HomeModel could listen to all subscriptions throughout the
app, and it would handle reaction and repost counting.
Once moved to the local relay model, HomeModel no longer had access to
all subscriptions, causing those counts to disappear.
The issue was fixed by doing the counting from ThreadModel itself, which
better isolates concerns throughout the app.
Signed-off-by: Daniel D’Aquino <daniel@daquino.me>
