element-hq / element-web

A glossy Matrix collaboration client for the web.
https://element.io
GNU Affero General Public License v3.0
11.13k stars 1.98k forks source link

Element-R: hang during login #25779

Open richvdh opened 1 year ago

richvdh commented 1 year ago

Sometimes, when attempting to log in, you get faced with a spinner of doom:

image

I think it's something indexedDB related. From the logs:

16:00:02.449 Removing indexeddb instance: matrix-js-sdk:crypto [rageshake.ts:74:27](webpack:///src/rageshake/rageshake.ts)
16:00:02.450 Removing IndexedDB instance matrix-js-sdk::matrix-sdk-crypto [rageshake.ts:74:27](webpack:///src/rageshake/rageshake.ts)
16:00:02.451 Removed indexeddb instance: matrix-js-sdk:crypto [rageshake.ts:74:27](webpack:///src/rageshake/rageshake.ts)
16:00:02.452 cannot yet remove IndexedDB instance matrix-js-sdk::matrix-sdk-crypto [rageshake.ts:74:27](webpack:///src/rageshake/rageshake.ts)
16:00:02.551 IndexedDB worker is ready [rageshake.ts:74:27](webpack:///src/rageshake/rageshake.ts)
16:00:02.551 Removing indexeddb instance: matrix-js-sdk:riot-web-sync [959eaaf02cce2e15d078.worker.js:5127:33](https://pr10080--matrix-react-sdk.netlify.app/959eaaf02cce2e15d078.worker.js)
16:00:02.552 Removed indexeddb instance: matrix-js-sdk:riot-web-sync [959eaaf02cce2e15d078.worker.js:5127:33](https://pr10080--matrix-react-sdk.netlify.app/959eaaf02cce2e15d078.worker.js)
16:00:02.553 Deleted indexeddb data. [rageshake.ts:74:27](webpack:///src/rageshake/rageshake.ts)

​ note cannot yet remove IndexedDB instance in particular.

It's fine after a reload.

richvdh commented 1 year ago

This is back

richvdh commented 1 year ago

There is a test that is supposed to catch this (MatrixClient.clearStores > should clear the indexeddbs in matrix-js-sdk/spec/integ/crypto/rust-crypto.spec.ts). Once we find the leak, we should extend the test to reproduce the problem area.

richvdh commented 12 months ago

I now can't reproduce this.

richvdh commented 3 months ago

We're still seeing this intermittently

richvdh commented 3 months ago

The root cause here is failures to clean up objects on the rust side during logout, which means we end up holding onto a connection to the IndexedDB. That means that we can't delete the IndexedDB, and when we log in again, attempts to open the IndexedDB block indefinitely.

The reason that the Rust objects aren't being cleaned up is that we're relying on javascript's FinalizationRegistry to call the rust-side destructors. As the MDN page makes clear, that mechanism is unreliable.

We could attempt to explicitly call .free on every reference to a Rust object that is ever passed to the JS layer (as we did for a specific case with https://github.com/matrix-org/matrix-js-sdk/pull/3610), but that is likely to add a lot of boilerplate, and be very brittle.

A more plausible solution is to update the matrix-rust-sdk IndexeddbCryptoStore so that it doesn't maintain a connection to the Indexeddb throughout its lifetime, but rather opens a new one each time an operation is performed. Opening the Indexeddb is quick enough that this shouldn't have a significant performance impact.