Add cache line size padding to RWLock

[mj10021]

Fixing issue #117470, this PR just copies the cfg_attr repr from the cache padding crate to stop RWLocks from sharing cache lines.

[rustbot]

r? @joshtriplett

(rustbot has picked a reviewer for you, use r? to override)

[mj10021]

@bors try @rust-timer queue

[rust-timer]

Insufficient permissions to issue commands to rust-timer.

[bors]

@mj10021: 🔑 Insufficient privileges: not in try users

[Kobzol]

@bors try @rust-timer queue

[bors]

⌛ Trying commit 1e43257 with merge e08cb3e...

[bors]

☀️ Try build successful - checks-actions
Build commit: e08cb3e (e08cb3ef876cf8045096a682db4e7fe0baafc4db)

[cuviper]

We have a CachePadded<T> in mpmc::utils that also does this, which you could wrap around the inner field. That should keep it out of the docs too.

[cuviper]

Alternatively, it might make more sense for the original reporter to wrap their own CachePadded<RwLock<T>> for their use case, without inflating the size for all std users.

[mj10021]

Alternatively, it might make more sense for the original reporter to wrap their own CachePadded<RwLock<T>> for their use case, without inflating the size for all std users.

Yeah that makes sense to me since its a pretty simple optimization on a case by case basis. Would it make sense to mention in the RWLock docs that padding to cache line size can be a known speed optomization?

[quininer]

Going one step further, maybe the std should provide the CachePadded? this is a very commonly used tool in concurrent programming.

[pingzhaozz]

repr(align(n)) doesn't work here. Inside the structure it requires something like:

struct RwLock<T> {
   inner : CachePadded<RwLock>,
   ...
}

But std seems not support such kind yet.

[Noratrieb]

As others have already commented, making these locks this big for everyone seems like a big pessimization if the locks aren't in a really hot path next to each other.
I think it makes sense to close this and suggest moving forwards with adding the cache padding outside the lock, either with a type in std with the ACP or not. Maybe it would make sense to also mention this in the documentation.
Thank you for the PR anyways!