feat: Allow reusing same shared IpcSharedMem for transfers

[sagudev]

Simple "solution" to #126, it's not exactly zero-copy, but it allows to modify inner buffer of IpcSharedMem and then send it back. This is all unsafe but we can uphold all safety variants for GPUBuffer (request-response msgs).

Example

before:

const SIZE: usize = 50;
let (tx1, rx1) = ipc::channel().unwrap();
let (tx2, rx2) = ipc::channel().unwrap();
std::thread::spawn(move || {
    let mut data = rx1.recv().unwrap() // allocation 1
                                      .to_vec(); // allocation 2
    for i in 0..=SIZE {
        *data[i] = i as u8;
    }
    tx2.send(IpcSharedMemory::from_bytes(&data)).unwrap(); // allocation 3
});
tx1.send(IpcSharedMemory::from_byte(0, SIZE)).unwrap();
rx2.recv().unwrap();

after:

const SIZE: usize = 50;
let (tx1, rx1) = ipc::channel().unwrap();
let (tx2, rx2) = ipc::channel().unwrap();
std::thread::spawn(move || {
    // SAFETY: This is safe because IpcSharedMemory is constructed in place from main thread (will only be accessed by send)
    let mut ism = unsafe{ rx1.recv().unwrap() }; // allocation 1
    {
        let data = ism.deref_mut();
        for i in 0..=SIZE {
            *data[i] = i as u8;
        }
    }
    tx2.send(ism).unwrap();
});
tx1.send(IpcSharedMemory::from_byte(0, SIZE)).unwrap();
rx2.recv().unwrap();

[sagudev]

Benchmarking ping_pong_shared_mem_mut_4194304_100: Collecting 100 samples in estimated 5.0000 s (
ping_pong_shared_mem_mut_4194304_100
                        time:   [0.0000 ps 0.0000 ps 0.0000 ps]
Found 13 outliers among 100 measurements (13.00%)
  5 (5.00%) high mild
  8 (8.00%) high severe

Benchmarking ping_pong_shared_mem_4194304_100: Collecting 100 samples in estimated 5.1199 s (200 
ping_pong_shared_mem_4194304_100
                        time:   [225.63 ms 226.61 ms 227.80 ms]
Found 4 outliers among 100 measurements (4.00%)
  1 (1.00%) low mild
  2 (2.00%) high mild
  1 (1.00%) high severe

Benchmarking ping_pong_shared_mem_mut_4194304_125: Collecting 100 samples in estimated 5.0000 s (
ping_pong_shared_mem_mut_4194304_125
                        time:   [0.0192 ps 0.0205 ps 0.0222 ps]
Found 12 outliers among 100 measurements (12.00%)
  4 (4.00%) high mild
  8 (8.00%) high severe

Benchmarking ping_pong_shared_mem_4194304_125: Warming up for 3.0000 s
Warning: Unable to complete 100 samples in 5.0s. You may wish to increase target time to 5.5s, or reduce sample count to 90.
Benchmarking ping_pong_shared_mem_4194304_125: Collecting 100 samples in estimated 5.5099 s (100 
ping_pong_shared_mem_4194304_125
                        time:   [558.99 ms 562.31 ms 566.19 ms]
Found 7 outliers among 100 measurements (7.00%)
  3 (3.00%) high mild
  4 (4.00%) high severe

[sagudev]

Benches from CI: https://github.com/servo/ipc-channel/actions/runs/10472511882

[jdm]

This seems reasonable as a fully-unsafe API. It would be nice if we could detect invalid usage at runtime similar to Arc::get_mut, but I don't have ideas for how to do that soundly.

[sagudev]

This seems reasonable as a fully-unsafe API. It would be nice if we could detect invalid usage at runtime similar to Arc::get_mut, but I don't have ideas for how to do that soundly.

The main problem is we would need to do cross-process locking which is costly, but we might want to implement this in the future to also fulfill other usecases.