[Letux-kernel] LetuxOS: Kernel: m-stable 5.4.77 successful

[H. Nikolaus Schaller]

Hi Andreas,

> Am 20.11.2020 um 22:37 schrieb Andreas Kemnade <andreas at kemnade.info>:
> 
> On Fri, 20 Nov 2020 22:26:16 +0100
> Andreas Kemnade <andreas at kemnade.info> wrote:
> 
>> On Mon, 16 Nov 2020 15:20:24 +0100 (CET)
>> hns at goldelico.com (H. Nikolaus Schaller) wrote:
>> 
>>> KVERSION: v5.4.77 - http://www.kernel.org
>>> CONFIGS:  arch/arm/configs/letux_defconfig arch/arm/configs/letux_lpae_defconfig arch/mips/configs/letux_defconfig
>>> 
>> does not compile here
>>  CC [M]  drivers/gpu/drm/omapdrm/omap_gem.o
>> /home/andi/kernel/drivers/gpu/drm/omapdrm/omap_gem.c: In function ‘_omap_gem_unpin’:
>> /home/andi/kernel/drivers/gpu/drm/omapdrm/omap_gem.c:886:28: error: passing argument 1 of ‘refcount_dec_and_test_checked’ from incompatible pointer type [-Werror=incompatible-pointer-types]
>>  if (refcount_dec_and_test(&omap_obj->dma_addr_cnt)) {
>>                            ^~~~~~~~~~~~~~~~~~~~~~~
>> In file included from /home/andi/kernel/include/linux/kref.h:17,
>>                 from /home/andi/kernel/include/linux/kobject.h:23,
>>                 from /home/andi/kernel/include/linux/device.h:16,
>>                 from /home/andi/kernel/include/linux/dma-mapping.h:7,
>>                 from /home/andi/kernel/drivers/gpu/drm/omapdrm/omap_gem.c:7:
>> /home/andi/kernel/include/linux/refcount.h:54:68: note: expected ‘refcount_t *’ {aka ‘struct refcount_struct *’} but argument is of type ‘u32 *’ {aka ‘unsigned int *’}
>> extern __must_check bool refcount_dec_and_test_checked(refcount_t *r);
>>                                                        ~~~~~~~~~~~~^
>> cc1: some warnings being treated as errors

Yes, my compiler also complains but does not treat it as error:

https://download.goldelico.com/letux-kernel/letux-5.4.77/src/warnings.log

>> 
>> git blame:
>> commit a882ed5f130ce0d467944a5d0a4d1e7259d815af
>> Author: Matthijs van Duin <matthijsvanduin at gmail.com>
>> Date:   Fri Jun 19 16:34:41 2020 +0100
>> 
>>    omapdrm: Use sg_table for dmabufs to enable width != 8192
>> 
>>    This fixes cases other than width=8k/stride=32k by using a proper
>>    sg_table with multiple entries for dmabufs, mapping each line
>>    to its correct TILER address.
>> 
>>    Original-by: Matthijs van Duin <matthijsvanduin at gmail.com>
>>    Signed-off-by: David Shah <dave at ds0.me>
>> 
>> Hmm, letux-5.9 compiles, so any dangerous cherry-picking here?!

AFAIR 5.4 did have a lot of trouble and I am not 100% sure if the tiler rotation works
properly at all. 5.6 was the best kernel for a while and 5.9 should be best.

>> 
> the logic looks ill:
>               return;
>        if (omap_obj->dma_addr_cnt--)
>                return;
>        if (WARN_ON(!omap_obj->block))
>                return;
> 
>        if (refcount_dec_and_test(&omap_obj->dma_addr_cnt)) {
> 
> That looks a bit duplicated here (besides of the type problem).
> dma_addr_cnt will always be 0 if that if is reached.
> And then the 0 is decremented and checked if it is zero?!

Yes, it may be a cherry-pick with bad automatic or manual conflict resolution.
It may come from a patch backported from 5.6. AFAIR, 5.6 was the best Pyra kernel
for a while and is currently used by PyraOS. But it is no longer maintained.

I haven't looked into 5.9 code - does it look better? Then we should
backport/fix the 5.4 version as needed.

BR,
Nikolaus