Enable RCU #752
Enable RCU #752
Conversation
|
@LeslieKid Really excited to see this PR. Good job on submitting a working RCU 👍 Here is some initial feedbacks. I will add more when I have the time to look into the detail.
|
@tatetian Thanks for your feedbacks! I tried to use the RCU for the ramfs and it seems to run correctly. Do I need to create a new branch and then pull request? Or just commit to this branch? |
|
#748 is getting merged. So you should be able to add unit tests for RCU.
I am not sure whether ramfs is a good user or not. Does Linux use RCU for ramfs? If you have good confidence, submit the changes and we shall see.
The size of this PR is well under control. It won't be a problem to have more commits added to it. |
| /// # Non-Preemptible RCU | ||
| /// | ||
| /// In non-preemptible RCU, the read-side critical section is delimited using | ||
| /// a `PreemptGuard`. |
There was a problem hiding this comment.
Our current implementation is non-preemptible RCU. To avoid repeating "Non-Preemptible RCU" in each method, I think it should be moved to the type-level Rust doc.
| /// This method returns a `RcuReadGuard` which allows read-only access to the | ||
| /// underlying data protected by the RCU mechanism. | ||
| /// | ||
| /// This function has the semantics of _subscribe_ in RCU mechanism. |
There was a problem hiding this comment.
You mention "publish" and "subscribe" in the Rust docs. But I can't see the value of mentioning "publish" and "sub scribe". Where do these terms come from? I don't think Linux use these terms.
There was a problem hiding this comment.
These terms come from LWN. I think that the roles of get() and replace() can fit these terms. Should I delete it?
There was a problem hiding this comment.
Mentioning the two versions only complicates things and confuses the reader. Decide which one is better: get/replace or publish/subscribe. Choose one version and stick with it.
| @@ -56,6 +59,9 @@ pub(crate) fn get_idle_task_cx_ptr() -> *mut TaskContext { | |||
|
|
|||
| /// call this function to switch to other task by using GLOBAL_SCHEDULER | |||
| pub fn schedule() { | |||
| unsafe { | |||
There was a problem hiding this comment.
Every usage of unsafe should be justified with well-written comment. Inserting pass_quiescent_state is actually unsound. Consider the following code.
let read_guard = rcu.get();
use_rcu_data(read_guard);
schedule(); // Oops: RCU-protected data is accessed across a quiescent state!
use_rcu_data(read_guard);To prevent such situations, pass_quiescent_state must be called after checking whether any preemption guards exist.
LeslieKid
force-pushed
the
enable-rcu
branch
2 times, most recently
from
abeb197
to
9719fe7
Compare
| } | ||
|
|
||
| pub fn set_atime(&self, time: Duration) { | ||
| self.metadata.write().atime = time; | ||
| self.writer_lock.lock(); | ||
| let reclaimer = self.metadata.replace({ |
There was a problem hiding this comment.
Sorry, this is not a good use case for RCU. Yes, using RCU reduces the overhead of the getter methods. But the setter methods become much expensive. For file systems, updating the file attributes may be frequent. So making the setter methods significantly more expensive is highly undesirable.
Enable non-preemptible RCU to the pointer for aster-framework.