fusefs: change how poll() on /dev/fuse works on unmount

[asomers]

Summary

If the file system has been unmounted or is marked as dead due to server
misbehavior:

• Don't return POLLIN or POLLRDNORM. We shouldn't, since a subsequent read() won't return any data.

• Do return POLLHUP since we'll never again return data on read().

• Don't return POLLOUT. According to poll(2), POLLHUP and POLLOUT should never be returned at the same time.

Don't MFC this change, as there's a small risk of breakage for fuse
servers that depend on the old behavior. Note that libfuse is not
affected; while it does use poll() it doesn't check revents. Nor are
any applications affect that use kqueue, select, or plain blocking
reads.

Reported by: arrowd
MFC after: Never
Sponsored by: ConnectWise

Test Plan

Test case added

[arrowd]

sys/fs/fuse/fuse_device.c

284

I think, to be in line with that FreeBSD does for other descriptor types, this should be

revents = POLLHUP;

See my test program from this post: https://lists.freebsd.org/archives/freebsd-hackers/2025-June/004569.html

In that code I don't even request POLLHUP in events, but FreeBSD returns its nevertheless.

[asomers]

sys/fs/fuse/fuse_device.c

284

Yes, good catch. We should output POLLHUP regardless of the input events.

[asomers]

• fixup: output POLLHUP regardless of the contents of events

[arrowd]

LGTM, but you're still ORing POLLHUP into revents, not setting it. As I mentioned in comment, it seems that the FreeBSD way to report EOF is to return POLLHUP only.

[asomers]

In D50826#1161127, @arrowd wrote:

LGTM, but you're still ORing POLLHUP into revents, not setting it. As I mentioned in comment, it seems that the FreeBSD way to report EOF is to return POLLHUP only.

How about if I change it to return POLLIN if there are still any events to read, and only return POLLHUP once there are no more events _and_ the mountpoint is dead?

[arrowd]

Sounds about right.