[PATCH] drm/amd/display: use GFP_ATOMIC for bounding box

[Arnd Bergmann]

On Tue, Jun 4, 2024, at 16:22, Christian König wrote:
> Am 04.06.24 um 15:50 schrieb Alex Deucher:
>> This can be called in atomic context.  Should fix:
>>
>> BUG: sleeping function called from invalid context at include/linux/sched/mm.h:306
>> in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 449, name: kworker/u64:8
>> preempt_count: 2, expected: 0
>> RCU nest depth: 0, expected: 0
>> Preemption disabled at:
>> ffffffffc0ce1580>] dc_fpu_begin+0x30/0xd0 [amdgpu]
>> CPU: 5 PID: 449 Comm: kworker/u64:8 Tainted: G        W          6.8.0+ #35
>> Hardware name: System manufacturer System Product Name/ROG STRIX X570-E GAMING WIFI II, BIOS 4204 02/24/2022
>> Workqueue: events_unbound async_run_entry_fn
>
> That most likely only papers over the real problem and is not a valid fix.
>
> The question is why is that an atomic context?  If the function is used 
> under a spinlock then this might indeed be the right fix.
>
> If it's because of floating point operation then that here won't work 
> either.

It looks like there is only one caller, and this turns on
floating point instructions just for the call:

        if (dc->res_pool->funcs->update_bw_bounding_box) {
                DC_FP_START();
                dc->res_pool->funcs->update_bw_bounding_box(dc, dc->clk_mgr->bw_params);
                DC_FP_END();
        }

but then they are enabled again inside of the function.

If we can drop the outer DC_FP_START(), that means
the GFP_KERNEL allocation works. On the other hand if
we actually have to enabled it before calling into
the function (e.g. because there is an architecture that
has incompatible function calling conventions when FP
is enabled), the inner one is redundant, but we can
potentially move the kmemdup() into the caller and
pass the copy by refernence.

      Arnd