Replaces the synthetic eventfd stand-in with the real kernel
eventfd path: open the kernel's 'card0/fence/<seqno>' fd and
poll on it. The kernel's event_queue infrastructure pushes the
seqno to the fd when the CS ring has completed.
Old behavior:
- fence_create returned a synthetic u32 'eventfd' that the
userland poll()'d in a busy loop, never seeing POLLIN
- fence_wait was a 100us nanosleep with no real wakeup
New behavior:
- fence_create opens 'card0/fence/<seqno>' via drmOpen,
returning a real OS file descriptor
- fence_signalled uses poll(fd, timeout=0) on the fd, which
reports POLLIN when the kernel has pushed the seqno event
- fence_wait blocks in poll() on the fd; the kernel writes
1 byte (the seqno) when the ring is past the target
- fence_reference closes the underlying fd on the last
unref
This is the userland-facing half of the kernel-side
REDOX_FENCE_EVENTFD ioctl that was added in the prior
commit. Together they implement the same pattern as
Linux 7.1's drm_syncobj_wait_eventfd (drivers/gpu/drm/
drm_syncobj.c::drm_syncobj_wait_ioctl +
include/uapi/drm/drm.h::drm_syncobj_eventfd).
Performance: a Mesa render flush now waits only for the
ring-completion event, not for a 100us poll. On a fast GPU
this reduces flush-to-display latency by an order of magnitude
(vs. the synthetic poll-loop).
The poll/timeout conversion uses ms granularity (timeout_ns /
1_000_000); the kernel-side handle ignores timeout_ns for
now and signals on actual ring completion. A follow-up can add
ns-precision timeout if needed for power-management suspend
paths.
vasilito
ee33af19cb