On 2026-02-20 09:36:58 [+0100], Maarten Lankhorst wrote:
> New version, new approach!
>
> A lot of the problems with PREEMPT_RT happen because of the current vblank
> handling. The real solution is making it more deterministic.
>
…
I applied this on top of drm-tip because nothing else I had around
applied cleanly. I had to tell the selftests to use
dma_fence_lock_irqsave() instead of spin_lock_irq() but I guess you are
aware of it.
I've been using it a bit and haven't noticed any big spikes while doing
this and that. The help message for CONFIG_DRM_I915_SELFTEST refers to
i915.selftest but this seems to have been replaced by
{mock|live|perf}_selftests. Using mock_selftest I get
| i915: Running i915_sw_fence_mock_selftests/test_timer
| BUG: sleeping function called from invalid context at
kernel/locking/spinlock_rt.c:48
| in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 4667, name: modprobe
| preempt_count: 1, expected: 0
| RCU nest depth: 0, expected: 0
| 1 lock held by modprobe/4667:
| #0: ffffd06703d0ba58 (&tf->fence/1){+.+.}-{2:2}, at:
__i915_sw_fence_wake_up_all+0x122/0x1d0 [i915]
| Preemption disabled at:
| [<ffffffffc1a1990e>] test_timer+0x2e/0x230 [i915]
^
| CPU: 2 UID: 0 PID: 4667 Comm: modprobe Tainted: G U
7.0.0-rc1+ #1 PREEMPT_{RT,(lazy)}
| Tainted: [U]=USER
| Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./Z68 Pro3-M, BIOS
P2.30 06/29/2012
| Call Trace:
| <TASK>
| dump_stack_lvl+0x68/0x90
| __might_resched.cold+0xf0/0x12b
| rt_spin_lock_nested+0x5d/0x200
| __i915_sw_fence_wake_up_all+0x122/0x1d0 [i915]
| i915_sw_fence_complete+0x3e/0x60 [i915]
| test_timer+0x42/0x230 [i915]
^ has a preempt_disable()
| __i915_subtests.cold+0x22/0x71 [i915]
| __run_selftests.cold+0x89/0xb8 [i915]
| i915_mock_selftests+0x30/0x70 [i915]
| i915_init+0x22/0x80 [i915]
|
| i915: Running i915_sw_fence_mock_selftests/test_dma_fence
| Asynchronous wait on fence mock:mock:0 timed out (hint:fence_notify [i915])
| BUG: sleeping function called from invalid context at
kernel/locking/spinlock_rt.c:48
| in_atomic(): 0, irqs_disabled(): 1, non_block: 0, pid: 86, name: ktimers/7
| preempt_count: 0, expected: 0
| RCU nest depth: 2, expected: 2
| 6 locks held by ktimers/7/86:
| #0: ffffffff8f881700 (local_bh){.+.+}-{1:2}, at:
__local_bh_disable_ip+0x23/0x230
| #1: ffffffff8f909440 (rcu_read_lock){....}-{1:2}, at:
__local_bh_disable_ip+0x124/0x230
| #2: ffff8acf9751fce0 (&base->expiry_lock){+...}-{2:2}, at:
timer_expire_remote+0x37/0x60
| #3: ffffffff8f909440 (rcu_read_lock){....}-{1:2}, at: rt_spin_lock+0xf6/0x200
| #4: ffffd0670031fc58 ((&timer->timer)){....}-{0:0}, at:
call_timer_fn+0x7e/0x280
| #5: ffff8acf857eb7d8 (fence/1){+.+.}-{2:2}, at:
__i915_sw_fence_wake_up_all+0x122/0x1d0 [i915]
| irq event stamp: 729273
| hardirqs last enabled at (729272): [<ffffffff8f0df4dc>]
_raw_spin_unlock_irqrestore+0x4c/0x70
| hardirqs last disabled at (729273): [<ffffffff8f0df162>]
_raw_spin_lock_irq+0x52/0x60
I guess this could be improved…
| softirqs last enabled at (729252): [<ffffffff8e50da42>] run_ktimerd+0x72/0xb0
| softirqs last disabled at (729256): [<ffffffff8e50d9db>] run_ktimerd+0xb/0xb0
| CPU: 7 UID: 0 PID: 86 Comm: ktimers/7 Tainted: G U W
7.0.0-rc1+ #1 PREEMPT_{RT,(lazy)}
| Tainted: [U]=USER, [W]=WARN
| Hardware name: To Be Filled By O.E.M. To Be Filled By O.E.M./Z68 Pro3-M, BIOS
P2.30 06/29/2012
| Call Trace:
| <TASK>
| dump_stack_lvl+0x68/0x90
| __might_resched.cold+0xf0/0x12b
| rt_spin_lock_nested+0x5d/0x200
| __i915_sw_fence_wake_up_all+0x122/0x1d0 [i915]
| i915_sw_fence_complete+0x3e/0x60 [i915]
| call_timer_fn+0xaa/0x280
| __run_timers+0x1e8/0x340
| timer_expire_remote+0x47/0x60
| tmigr_handle_remote+0x381/0x500
| handle_softirqs.isra.0+0xc0/0x3f0
| run_ktimerd+0x50/0xb0
| smpboot_thread_fn+0x12d/0x2e0
| i915: Running scatterlist
| i915: Running scatterlist_mock_selftests/igt_sg_alloc
| sg_alloc_table timed out
| i915: Running scatterlist_mock_selftests/igt_sg_trim
| i915_sg_trim timed out
…
There is nothing else, that popped up. So far no objections from my side
;) Thank you.
Sebastian
