-
-
Notifications
You must be signed in to change notification settings - Fork 1.7k
UNIX: ensure single reader and writer to system fd #16209
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
base: master
Are you sure you want to change the base?
UNIX: ensure single reader and writer to system fd #16209
Conversation
98d7872 to
904dd95
Compare
|
I split the fdlock in two different commits (refcount then serial R/W) that outline the different steps for merging as individual PRs. |
src/crystal/fd_lock.cr
Outdated
| # acquire spinlock + forward declare pending waiter | ||
| m, success = @m.compare_and_set(m, (m | xspin | xwait) & ~clear, :acquire, :relaxed) | ||
| if success | ||
| waiters.value.push(pointerof(waiter)) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Another missing trick from nsync: only new waiters shall be pushed to the end the queue, awoken fibers that still failed to lock shall be pushed to the start of the queue (another mechanism against starvation).
It's not implemented because Crystal::PointerLinkedList doesn't have an #unshift method (yet).
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Method added in #16287.
904dd95 to
6be2dd7
Compare
Serializes reads and writes so we can assume any IO object will only have at most one read op and one write op. The benefits are: 1. it avoids a race condition in the polling event loops: - Fiber 1 then Fiber 2 try to read from fd; - Since fd isn't ready so both are waiting; - When fd becomes ready then Fiber 1 is resumed; - Fiber 1 doesn't read everything and returns; - Fiber 2 won't be resumed because events are edge-triggered; 2. we can simplify the UNIX event loops (epoll, kqueue, io_uring) that are guaranteed to only have at most one reader and one writer at any time.
6be2dd7 to
b92814a
Compare
This patch extends the fdlock to serialize reads and writes by extending the reference counted lock with a read lock and a write lock, so taking a reference and locking acts as a single operation instead of two (1. acquire/release the lock; 2. take/return a reference). This avoids a race condition in the polling event loops:
fd;fdisn't ready, both fibers start waiting;fdbecomes ready then Fiber 1 is resumed;With the read lock, fiber 2 will wait on the lock then be resumed by fiber 1 when it returns. A concrete example is multiple fibers waiting to accept on a socket where fiber 1 would keep handling connections, while fiber 2 sits idle.
The other benefit is that it can help to simplify the evloops that will now only deal with a single reader + single writer per
IOand is required for the io_uring evloop (the MT version requires it).NOTE: While this patch only serializes reads/writes on UNIX at the
Crystal::System, which is where the bugs are, we may want to move it into stdlib for all targets at some point, for example to serialize reads and writes aroundIO::Buffered.Depends on #16288 and #16289.
Required by #16264.