Ensure cancellation safety #56

GTwhy · 2022-04-22T01:14:24Z

Panic or undefined behavior will happen if the future dropped(cancelled) during the execution of ibv_post_send. #47

The text was updated successfully, but these errors were encountered:

Let `event_listener` holds the `Arc`s of the `LocalMrInner`s that are being used by RDMA ops to ensure cancellation safety. `LocalMr` was replaced with New struct `LocalMrInner`. Because every struct that can use APIs should hold an `Arc` of the mr's metadata, but previous `LocalMr` can't hold itself. Fixes: datenlord#47 datenlord#56

Let `event_listener` holds the `Arc`s of the `LocalMrInner`s that are being used by RDMA ops to ensure cancellation safety. `LocalMr` was replaced with New struct `LocalMrInner`. Because every struct that can use APIs should hold an `Arc` of the mr's metadata, but previous `LocalMr` can't hold itself. Add `MrState` to avoid potential race condition When the RDMA operations future were canceled but the related mr is still being used by netdev or kernel. This is a potential race condition. So we add `MrState` to `LocalMrInner`, to make sure we can get slice or mutable slice only if the mr is readable or writeable. Fixes: datenlord#47 datenlord#56

Let `event_listener` holds the `Arc`s of the `LocalMrInner`s that are being used by RDMA ops to ensure cancellation safety. `LocalMr` was replaced with New struct `LocalMrInner`. Because every struct that can use APIs should hold an `Arc` of the mr's metadata, but previous `LocalMr` can't hold itself. Add `RwLock` to avoid potential race condition When the RDMA operations future were canceled but the related mr is still being used by netdev or kernel. This is a potential race condition. So we add `Rwlock` to `LocalMrInner`, to make sure we can get slice or mutable slice only if the mr is readable or writeable. Replace hashmap to reduce time overhead and ensure guards drop after remove immediately to unlock the rwlock.`std::collections::HashMap` has less time overhead than `LockFreeCuckooHash` in the low concurrency situation. `LockFreeCuckooHash` can't drop guards immediately after remove them and that led to a deadlock. Add some unchecked methods without lock, which should be put into unsafe trait later. Integrate common `HashMap` methods in hashmap_utilities. Fixes: datenlord#47 datenlord#56

Let `event_listener` holds the `Arc`s of the `LocalMrInner`s that are being used by RDMA ops to ensure cancellation safety. `LocalMr` was replaced with New struct `LocalMrInner`. Because every struct that can use APIs should hold an `Arc` of the mr's metadata, but previous `LocalMr` can't hold itself. Add `RwLock` to avoid potential race condition When the RDMA operations future were canceled but the related mr is still being used by netdev or kernel. This is a potential race condition. So we add `Rwlock` to `LocalMrInner`, to make sure we can get slice or mutable slice only if the mr is readable or writeable. Replace hashmap to reduce time overhead and ensure guards drop after remove immediately to unlock the rwlock.`std::collections::HashMap` has less time overhead than `LockFreeCuckooHash` in the low concurrency situation. `LockFreeCuckooHash` can't drop guards immediately after remove them and that led to a deadlock. Add some unchecked methods without lock, which should be put into unsafe trait later. Integrate common `HashMap` methods in hashmap_utilities. Fixes: #47 #56

GTwhy closed this as completed May 26, 2022

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Ensure cancellation safety #56

Ensure cancellation safety #56

GTwhy commented Apr 22, 2022

Ensure cancellation safety #56

Ensure cancellation safety #56

Comments

GTwhy commented Apr 22, 2022