Fix token leak if client performs non-blocking poll #5

flow commented 2026年05月22日 20:36:51 +02:00

If a client performs a non-blocking poll while no tokens were
available, FUSE/CUSE passes a nullptr as the pollhandle argument
to steve_poll(). Steve would previously push this nullptr into the
waiters queue.

Later, when a token became available, the server would wake this
"ghost" waiter, permanently reserving a token for a client that was
not actually waiting for it. This effectively leaks the token until
the client process exited.

Fix this by only queuing a waiter request if the pollhandle is
non-null (representing an actual sleeping client).

Signed-off-by: Florian Schmaus flow@gentoo.org

If a client performs a non-blocking poll while no tokens were available, FUSE/CUSE passes a nullptr as the pollhandle argument to steve_poll(). Steve would previously push this nullptr into the waiters queue. Later, when a token became available, the server would wake this "ghost" waiter, permanently reserving a token for a client that was not actually waiting for it. This effectively leaks the token until the client process exited. Fix this by only queuing a waiter request if the pollhandle is non-null (representing an actual sleeping client). Signed-off-by: Florian Schmaus <flow@gentoo.org>

thesamesam commented 2026年05月22日 22:51:54 +02:00

Out of interest, do you have an example client?

Out of interest, do you have an example client?

flow commented 2026年05月23日 08:05:34 +02:00

@thesamesam wrote in #5 (comment):

Out of interest, do you have an example client?

No. The possibility for this came up while working on #2

I am not sure why a client should poll with a zero timeout nor if there a clients that actually do so. So, we might not see such clients, but it can't hurt for Steve to be prepared for this case.

@thesamesam wrote in https://codeberg.org/gentoo/steve/pulls/5#issuecomment-15684110: > Out of interest, do you have an example client? No. The possibility for this came up while working on https://codeberg.org/gentoo/steve/pulls/2 I am not sure why a client should poll with a zero timeout nor if there a clients that actually do so. So, we might not see such clients, but it can't hurt for Steve to be prepared for this case.

flow commented 2026年05月23日 08:34:39 +02:00

I am not sure why a client should poll with a zero timeout

Just to add: event loop reactor designs where timeout it dynamically calculated may end up with timeout being zero (intentionally or unintentionally). But I did not (yet) look at the code of the typical candidates if such a case exists.

> I am not sure why a client should poll with a zero timeout Just to add: event loop reactor designs where timeout it dynamically calculated may end up with timeout being zero (intentionally or unintentionally). But I did not (yet) look at the code of the typical candidates if such a case exists.