-
Notifications
You must be signed in to change notification settings - Fork 20
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
RFC: use atomic reference count to track external in-memory cache record usage #779
Comments
The RFC LGTM. Let me put the background of this RFC here (@MrCroxx correct me if I am wrong):
|
A drawback of using atomic add and atomic sub is that we are not aware of the value of the variable when the operation is actually applied on it. To resolve this drawback, a feasible solution can be using CAS. With CAS, we can be aware of the value before the operation is applied on the atomic variable. A general implementation can be:
|
After discussion, we found that there are several situations that may lead to entry clone:
Therefore, clone is not always initiated by the caller (3), but can also come from 1 and 2, so I'm in favour of adopting the current RFCs to address all the scenarios involved. Since the frequency of locks has been greatly reduced since the RFC, using CAS / Lock is acceptable to me for both. Thanks for the efforts |
I think this optimization does not solve the problem. If I understand incorrectly, please correct me. When the releaser finds itself responsible for releasing the object, it still needs to acquire the mutex lock of the shard. During the gap between the lock guard is actually acquired, there is a chance that another thread calls "get/fetch" on the same key and increases the refs again. |
Atomic Reference Count Management
[
RawCache
] uses an atomic reference count to management the release of an entry.The atomic reference count represents the external references of a cache record.
When the reference count drops to zero, the related cache shard is locked to release to cache record.
It is important to guarantee the correctness of the usage of the atomic reference count. Especially when triggering
the release of a record. Without any other synchronize mechanism, there would be dangling pointers or double frees:
Thankfully, we can prevent it from happening with the usage of the shard lock:
The only ops that will increase the atomic reference count are:
RawCache
] and get external entries. (locked)The op 1 is always guarded by a mutex/rwlock, which means it is impossible to happen while releasing a record with
the shard is locked.
The op 2 is lock-free, but only happens when there is at least 1 external reference. So, it cannot happen while
releasing a record. Because when releasing is happening, there must be no external reference.
So, this case will never happen:
When starting to release a record, after locking the shard, it's still required to check the atomic reference count.
Because this cause still exist:
Although the op1 requires to be locked, but the release operation can be delayed after that. So the release
operation can be ignored after checking the atomic reference count is not zero.
There is no need to be afraid of leaking, there is always to be a release operation following the increasing when
the atomic reference count drops to zero again.
Related issues & PRs
#778
The text was updated successfully, but these errors were encountered: