Add WASI support

[GregoryConrad]

Adds experimental WASI support. Requires nightly and file locking is currently a no-op, at least for the immediate future (needs to get implemented upstream).

Partially fixes #507

[GregoryConrad]

@cberner This is currently blocked by:

thread 'main' panicked at 'called `Result::unwrap()` on an `Err` value: Error { kind: Unsupported, message: "operation not supported on this platform" }', src/tree_store/page_store/page_manager.rs:451:30

Which is:

        let mut storage = PagedCachedFile::new(
            file.try_clone().unwrap(), // panics on the unwrap()
            page_size as u64,
            read_cache_size_bytes,
            write_cache_size_bytes,
        )?;

Any ideas on what can be done to bypass this try_clone(), as it is not supported on WASI? I might be able to dig into the source of std later to see if it is possible to implement but that may be a long time from now since I am a bit busy

[cberner]

I'm pretty sure it can just be removed: #584

I don't remember why I added that, but it looks like I refactored the code at some point and it's no longer necessary

[GregoryConrad]

I really hate this, but I don't think there is any other way to do benchmark specific dependencies (and you can't create a benching feature either because you can't have optional dev dependencies). If you have any ideas though, I am all ears.

(This divides test specific deps and benchmark specific deps, since essentially none of the benchmark deps compile to WASI)

[GregoryConrad]

This silences some unused import warnings that will go away once wasi-libc has flock

[GregoryConrad]

Currently running the wasi tests with:

cargo +nightly wasi test --features wasi -- --nocapture

Not sure how I'll be able to allow fs access with the above approach since simply adding --dir xyz doesn't work, but I'll cross that bridge when I get there.

For now, I'm getting

test db::test::crash_regression1 ... thread 'main' panicked at 'no filesystem on wasm', library/std/src/sys/wasi/os.rs:228:5

Guessing this is because of tempfile. Might see if I can patch in some dummy implementation that will work for WASI (i.e., hardcodes a tmp/ on WASI or something).

Edit: I was right; it originates from here

[GregoryConrad]

I've started to get tests passing! 🥳🎉

To get useful test output, I've needed cargo-wasi; here's the command for that:

CARGO_TARGET_WASM32_WASI_RUNNER="wasmtime --dir=." cargo +nightly wasi test --features=wasi -- --nocapture

To run tests without any extra dependencies, you can use this (but note this is near impossible to debug in case a test fails):

CARGO_TARGET_WASM32_WASI_RUNNER="wasmtime --dir=." cargo +nightly test --target wasm32-wasi --features=wasi

So I recommend we go with the cargo-wasi option for CI.

[GregoryConrad]

@cberner Left some comments above that I would appreciate some feedback on. Other than that, this PR is ready to go as far as I can tell

[GregoryConrad]

This file is the same as before (diff here isn't the clearest). Just wrapped it in a mod to avoid compiling/running any of it on WASI (no true support for multithreading atm)

[cberner]

Overall this is looking much simpler than I had expected it to be!

[GregoryConrad]

In addition to the tempfile test changes I'll make at some point later today, I wanted to mention some warnings you get while compiling with nightly:

warning: variable does not need to be mutable
   --> src/tree_store/page_store/page_manager.rs:824:13
    |
824 |         let mut secondary = state.header.secondary_slot_mut();
    |             ----^^^^^^^^^
    |             |
    |             help: remove this `mut`
    |
    = note: `#[warn(unused_mut)]` on by default

warning: variable does not need to be mutable
   --> src/tree_store/page_store/page_manager.rs:880:13
    |
880 |         let mut secondary = state.header.secondary_slot_mut();
    |             ----^^^^^^^^^
    |             |
    |             help: remove this `mut`

   Compiling librocksdb-sys v0.11.0+8.1.1
warning: using `.borrow()` on a double reference, which returns `&<K as RedbValue>::SelfType<'_>` instead of borrowing the inner type
  --> src/table.rs:49:62
   |
49 |             .map(|x| x.map(|(key, _)| K::as_bytes(key.value().borrow()).as_ref().to_vec()));
   |                                                              ^^^^^^^^^
   |
   = note: `#[warn(suspicious_double_ref_op)]` on by default

warning: using `.borrow()` on a double reference, which returns `&<K as RedbValue>::SelfType<'_>` instead of borrowing the inner type
  --> src/table.rs:68:62
   |
68 |             .map(|x| x.map(|(key, _)| K::as_bytes(key.value().borrow()).as_ref().to_vec()));
   |                                                              ^^^^^^^^^

warning: using `.borrow()` on a double reference, which returns `&<K as RedbValue>::SelfType<'_>` instead of borrowing the inner type
   --> src/tree_store/btree.rs:270:55
    |
270 |             assert!(operation.safe_delete(entry?.key().borrow())?.is_some());
    |                                                       ^^^^^^^^^

warning: using `.borrow()` on a double reference, which returns `&<K as RedbValue>::SelfType<'_>` instead of borrowing the inner type
   --> src/tree_store/btree.rs:307:58
    |
307 |                 assert!(operation.safe_delete(entry.key().borrow())?.is_some());
    |                                                          ^^^^^^^^^

I'd be happy to fix these in a new PR since I imagine they will trickle down to stable eventually.

Edit: nevermind; I need to fix them in this PR because wasi ci is failing due to them

[GregoryConrad]

@cberner Instead of adding that create_tempfile back in, how would you feel about all tests using a tmp/ folder within the project root that can be created & cleaned up via just? Even with the create_tempfile function, WASI would still pollute the project root whenever tests are run (and would probably need this tmp/ anyways because of that). Further, with a local tmp/, you can also inspect failed tests' temp files in the future more easily.

[cberner]

I don't think that's a good idea, because I want to support development workflows that do not use just. I do something similar in the benchmarks, where they create and try to cleanup their own tmpdir and it's kind of a mess. The only reason I do it there is that some OSs put /tmp in a tmpfs which messes up the benchmark timing.

What about reading the tmpdir from an environment variable? Would it be possible to create a directory in /tmp and then mount it via wasmtime --dir= in the justfile?

[GregoryConrad]

Should be all set now. Non-wasi platforms will act the same way they did before, and just test_wasi will use the system's $TMPDIR.

[GregoryConrad]

After this gets merged I can make 2 dummy commits to show the diffs that will be needed:

• Once wasi-libc gets flock (and after that change is propagated to Rust's libc)
• Once WASI no longer requires nightly

Just so those changes are less of a guessing game when the time comes. I'll link those two diffs in #507

[cberner]

Merged. Thanks!