Description
There's a bug in std::sync::Once
that makes it so that, under certain conditions, a call to call_once
in a process which was forked from another Rust process can block forever.
How Once
works
Simplifying a bit (ignoring poisoning), the algorithm employed by Once
works as follows: The Once
can be in one of three states: INCOMPLETE
, COMPLETE
, and RUNNING
.
- The
Once
starts off in theINCOMPLETE
state. - When a call to
call_once
begins, theOnce
might be in any of the three states:- If the
Once
is in theINCOMPLETE
state, then it is transitioned to theRUNNING
state, and the function begins executing. - If the
Once
is in theRUNNING
state, then some other call tocall_once
is executing the function, so this call puts itself on a list of waiters and goes to sleep. It will be woken back up once the function is done executing in whatever thread is executing it. - If the
Once
is in theCOMPLETE
state, then the function has already been executed, socall_once
returns immediately without doing anything.
- If the
Finally, when the function's execution completes, the thread doing the execution transitions the Once
into the COMPLETE
state, and wakes up any waiters that accumulated while it was executing the function.
The issue
This algorithm is broken when forking. In particular, if a Once
is in the RUNNING
state at the point that the process forks, when the child's memory space (which, by default, is a copy-on-write copy of the parent's) is created, the Once
will still be in the RUNNING
state. However, in the child process, calls to call_once
will fail for two reasons:
- If the call happens while the function is still being executed, the waiter object that is enqueued will not actually be visible to the executor because it will only affect the child's memory space, not the parent's, and so the executor (a member of the parent thread) finishes, it will wake up all of the waiters in the parent process, blissfully ignorant that a thread from the child process is also waiting.
- If the function execution finishes first, the change of the
Once
's state fromRUNNING
toCOMPLETE
will not be reflected in the child's memory space. Thus, a future call tocall_once
will spuriously find theOnce
still in theRUNNING
state even though it isn't really in that state anymore.
These two problems can be seen in action in two proofs of concept that I wrote: This one demonstrates the first issue, while this one demonstrates the second.
A proposed fix
Joint credit for this proposal goes to @ezrosent.
The idea behind this fix is to record the process' PID when transitioning a Once
from INCOMPLETE
to RUNNING
, and having future accesses that find the Once
in the RUNNING
state verify that it wasn't transitioned by a parent process. Unfortunately, this doesn't quite work because PIDs can be re-used, so if process A spawns process B, then process A quits, then process B spawns process C, it's possible for A and C to have the same PID.
Instead, we introduce the idea of an "MPID" - a monotonically-increasing PID-like counter that is maintained by the process (e.g. . We increment it every time a process forks, and use it in the Once
objects to record which process transitioned an object from INCOMPLETE
to RUNNING
.
More concretely, here are the components of the proposed solution:
- There is a process-global MPID variable (could be either
usize
oru64
) that is initialized to 0 and is incremented immediately afterfork
. Note that this does not guarantee that no two processes anywhere in the tree of processes forked from a particular process have the same MPID. In fact, all processes forked by a given process will all have the same MPID. However, it does guarantee that a process will not share an MPID with any of its ancestors, and that is all we need. - The
Once
object is modified to have anothermpid
field that is initialized to 0. - Each waiter object is modified to have another
mpid
field that is initialized to the MPID of the current process when the object is created. - A modified algorithm for
call_once
looks roughly like this:- Loop:
- Load the current state. If it is
COMPLETE
, return. - If the state is
INCOMPLETE
, do then loadmpid
and:- If
mpid
is equal to the current MPID, then try to CAS the state fromINCOMPLETE
toRUNNING
. If it fails, retry the entire loop. If it succeeds, you're responsible for running the function, so do the original algorithm. - If
mpid
is not equal to the current MPID, then try to CAS it from its current value to the current MPID. If this succeeds, go to the previous step (wherempid
is equal to the current MPID), and if it fails, retry the entire loop.
- If
- If the state is
RUNNING
, then loadmpid
and:- If
mpid
is equal to the current MPID, then the thread that transitioned theOnce
into theRUNNING
state is in the current process, so do the normal algorithm: wait for it to be done (recording the current MPID in the waiter object). - If
mpid
is not equal to the current MPID, then the thread that transitioned theOnce
into theRUNNING
state is in an ancestor process. Thus, attempt to CASmpid
to the current MPID. If it fails, repeat the entire loop. If it succeeds, then it is your responsibility to run the function, so continue as if you had transitioned theOnce
into theRUNNING
state, with one exception: when waking up waiters, you need to check that they are not waiters in an ancestor process; do this by checking the waiter object'smpid
field, and only waking waiters with anmpid
field equal to the current MPID.
- If
- Load the current state. If it is
- Loop:
One thing to note: It is safe to try to CAS mpid
and then separately to transition into RUNNING
even though the value of mpid
needs to reflect the MPID of the thread that transitioned into RUNNING
- a thread that successfully transitions a Once
into the RUNNING
state will have previously verified that mpid
is correct, and thus it will not change forever in the future (at least, not in this process) since the MPID of a process never changes.
Open question
One open question is how to ensure that code is run just after fork
(to increment the global MPID variable) and, critically, before any other code runs (especially code that uses Once
). pthread
provides the pthread_atfork
function to register callbacks that run before and after fork
calls, but obviously this doesn't address Windows, and I also don't know if there's a good way to ensure that the necessary call to pthread_atfork
is made at process initialization time.
Prior art
There's some prior art here. In particular, jemalloc has acknowledged a similar issue, and has a partial fix for it.