Wait for ueventd to create loop device on Android

[tiann]

Reference: https://cs.android.com/android/_/android/platform/external/toybox/+/51a43ad5225153582b40d3fd289701efc63c8f62

[tiann]

I will try it with inotify :)

[tiann]

We need this PR : notify-rs/notify#470 to be merged first

[flxo]

The loopdev crate has a neat small list of dependencies. Not sure if adding inotify (for target_os = "android") is the way to go. Using the libc interface directly would also be possible and the use case here is simple enough.

Sidenote: I use inotify-rs on Android in production.

[0xpr03]

If you only need inotify [the linux event API], and no other OS support, then notify [the crate] is probably overkill and inotify-rs could be enough.

[tiann]

And i did a benchmark:

Without inotify:

cost: 20.716512ms
cost: 35.034µs
cost: 20.799764ms
cost: 20.795166ms
cost: 20.71757ms
cost: 20.412802ms
cost: 20.288493ms
cost: 20.745402ms
cost: 20.602376ms
cost: 20.689046ms

Average: 18ms

With inotify:

cost: 11.391805ms
cost: 15.421µs
cost: 6.936767ms
cost: 2.65031ms
cost: 12.248576ms
cost: 8.989787ms
cost: 7.116781ms
cost: 9.271932ms
cost: 15.299µs
cost: 1.726196ms

Average: 6ms

Both are release build for arm64-v8a, It is about 12ms faster.

[tiann]

Maybe we change the interval wait time to 1ms without inotify, and it should also work

[flxo]

I'm unsure wether a fast poll with an initial may be a better idea.

There's a hardcoded 2s timeout that can bite.

Is WatchMask::LCREATE firing? Think there can be the case where the looped files are are already present, free and yielded by LOOP_CTL_GET_FREE.

[tiann]

What should we do now? I changed it to loop wait 1ms

[flxo]

hi @mdaffin. What's your opinion on that?

[tiann]

I've been using this fix for quite some time now: https://github.com/tiann/KernelSU/blob/main/userspace/ksud/Cargo.toml#L38, but I had to maintain my own branch. Would you consider merging it in? I'm open to any suggestions you may have.

[tiann]

@mdaffin Could you please check this?

[mdaffin]

I am not a huge fan of spinning on a 1ms delay for up to 2 seconds. It is all well and good if your tests are coming back in ~6ms, but then that raises the question as to what situation it could take to to 2 seconds to update. If we want a delay loop I think it might be better for an exponential back off instead - ie double the delay until a reasonable limit or the timeout is hit? Or does that sound like too much complexity?

Sorry for the delay in looking at this.