[PATCH v6 5/6] drm/etnaviv: add driver support for the PCI devices

[Sui Jingfeng]

Hi,

On 2023/5/31 03:02, Bjorn Helgaas wrote:
> On Wed, May 31, 2023 at 12:06:42AM +0800, Sui Jingfeng wrote:
>> This patch adds PCI driver support on top of what already have. Take the
>> GC1000 in LS7A1000/LS2K1000 as the first instance of the PCI device driver.
>> There is only one GPU core for the GC1000 in the LS7A1000 and LS2K1000.
>> Therefore, component frameworks can be avoided. Because we want to bind the
>> DRM driver service to the PCI driver manually.
>> +	 * Loongson Mips and LoongArch CPU(ls3a5000, ls3a4000, ls2k1000la)
>> +	 * maintain cache coherency by hardware
>> +	 */
>> +	if (IS_ENABLED(CONFIG_CPU_LOONGSON64) || IS_ENABLED(CONFIG_LOONGARCH))
>> +		priv->has_cached_coherent = true;
> This looks like something that should be a runtime check, not a
> compile-time check.
>
> If it's possible to build a single kernel image that runs on Loongson
> MIPS or LoongArch CPU and, in addition, runs on other platforms, you
> cannot assume that all the others maintain this cache coherency.

Nice catch! I don't even realize this!


LS3A4000 is mips64r2 with MSA SIMD, while LS3A5000 is LoongArch,

instruction set, compiler, and binary interface are totally changed.

Therefore, it's impossible to build a single kernel image that runs on 
all Loongson CPUs.

Currently, I can guarantee that this works on the Loongson platform.

My initial intent here is to let priv->has_cached_coherent be *true* on 
the Loongson platform (both mips and loongarch).

I do know there are some other vendors who bought GPU IP from Vivante.

say GC7000, and integrate it into their discrete GPU product.

But it is also a PCI device, but this is another story; it deserves 
another patch.

I don't know if Etnaviv folk find some similar hardware on Arm Arch,

Some Arm CPUs do not maintain cached coherency on hardware.

The has_cached_coherent member can be set to false on such hardware.

For us, it seems that there is no need to do runtime checking,

because they are all cached coherent by default.


Can I improve this in the future, currently I don't have a good idea.

>> +static struct etnaviv_drm_private *etna_private_s;
> A static pointer looks wrong because it probably limits you to a
> single instance of something.

This structure is shared by all GPU cores.

Originally, Etnaviv was a component-based driver.

It's one driver wrangler for all GPU cores. (One piece of driver rules 
them all.)

This structure is shared by all GPU cores and does not have a copy per 
GPU core.

>> @@ -727,6 +756,12 @@ static int __init etnaviv_init(void)
>>   	if (ret != 0)
>>   		goto unregister_gpu_driver;
>>   
>> +#ifdef CONFIG_DRM_ETNAVIV_PCI_DRIVER
>> +	ret = pci_register_driver(&etnaviv_pci_driver);
>> +#endif
>> +	if (ret != 0)
>> +		goto unregister_platform_driver;
> Why is this outside the #ifdef?  If CONFIG_DRM_ETNAVIV_PCI_DRIVER is
> not set, you already tested "ret != 0" above and will never take this
> goto.

On arch/platform without CONFIG_PCI enabled,

CONFIG_DRM_ETNAVIV_PCI_DRIVER config option will not be enabled.


On such cases, GCC complains when compile with  W=1:


drivers/gpu/drm/etnaviv/etnaviv_drv.c: In function ‘etnaviv_init’:
drivers/gpu/drm/etnaviv/etnaviv_drv.c:787:1: warning: label 
‘unregister_platform_driver’ defined but not used [-Wunused-label]
   787 | unregister_platform_driver:
       | ^~~~~~~~~~~~~~~~~~~~~~~~~~

This is the pain that #ifdefs and #endif bring to us.

We want to zero out compile warnings.

Otherwise, Intel's compiler test robot will come here and warn you!

Yet another test for "ret != 0" doesn't hurt much, probably could be 
optimized out by the compiler.

We are afraid of warnings.

>> +static int etnaviv_gpu_plat_drv_init(struct etnaviv_gpu *gpu, bool component)
>> +{
>> +	struct device *dev = gpu->dev;
>> +	struct platform_device *pdev = to_platform_device(dev);
>> +	int err;
>> +
>> +	/* Map registers: */
>> +	gpu->mmio = devm_platform_ioremap_resource(pdev, 0);
>> +	if (IS_ERR(gpu->mmio))
>> +		return PTR_ERR(gpu->mmio);
>> +
>> +	if (component) {
>> +		err = component_add(dev, &gpu_ops);
>> +		if (err < 0) {
>> +			dev_err(dev, "failed to register component: %d\n", err);
>> +			return err;
>> +		}
>> +	}
>> +
>> +	return 0;
>> +}
> All this platform driver rearrangement looks like it should be a
> separate patch so adding PCI support only adds PCI-related stuff.
>
Indeed, this is acceptable.
>> +++ b/drivers/gpu/drm/etnaviv/etnaviv_pci_drv.c
>> @@ -0,0 +1,87 @@
>> +// SPDX-License-Identifier: GPL-2.0
>> +
>> +#include <linux/pci.h>
>> +
>> +#include "etnaviv_drv.h"
>> +#include "etnaviv_gpu.h"
>> +#include "etnaviv_pci_drv.h"
>> +
>> +enum etnaviv_pci_gpu_family {
>> +	GC1000_IN_LS7A1000 = 0,
>> +	GC1000_IN_LS2K1000 = 1,
> Seems unused; why is this here?

Want to use it to provide device-specific information.

For example, LS2K1000 is an SoC, and LS7A1000 is a bridge chipset.

The GC1000 in those chips is a 32-bit GPU. even though the host is a 
64-bit system.

This GPU can only access memory below 4 GB.

This can be used to provide information known at compile time.

Attach the has_cached_coherent member to the chip model defined here？

>> +static int etnaviv_pci_probe(struct pci_dev *pdev,
>> +			     const struct pci_device_id *ent)
>> +{
>> +	struct device *dev = &pdev->dev;
>> +	int ret;
>> +
>> +	ret = pcim_enable_device(pdev);
>> +	if (ret) {
>> +		dev_err(&pdev->dev, "failed to enable\n");
> Use "dev", no need for "&pdev->dev" since you already looked it up
> above.  Also below for dma_set_mask_and_coherent().

Ok, acceptable.  You are a really serious man and quite professional.

>
>> +		return ret;
>> +	}
>> +
>> +	pci_set_master(pdev);
>> +
>> +	ret = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32));
>> +static const struct pci_device_id etnaviv_pci_id_lists[] = {
>> +	{0x0014, 0x7a15, PCI_ANY_ID, PCI_ANY_ID, 0, 0, GC1000_IN_LS7A1000},
>> +	{0x0014, 0x7a05, PCI_ANY_ID, PCI_ANY_ID, 0, 0, GC1000_IN_LS2K1000},
>> +	{0, 0, 0, 0, 0, 0, 0}
> Should probably use PCI_DEVICE_DATA().  Use PCI_VENDOR_ID_LOONGSON.

Originally, I wanted to keep them on the same line, at the same time, 
avoiding complaints from checkpatch.pl.

PCI_VENDOR_ID_LOONGSON is too long. No problem, will be fixed in the next version.

> Only "{ }" required to terminate.
>
>> +++ b/drivers/gpu/drm/etnaviv/etnaviv_pci_drv.h
>> @@ -0,0 +1,12 @@
>> +/* SPDX-License-Identifier: GPL-2.0 */
>> +
>> +#ifndef __ETNAVIV_PCI_DRV_H__
>> +#define __ETNAVIV_PCI_DRV_H__
>> +
>> +#include <linux/pci.h>
> This #include isn't required by this file.

Let me give you a reason to do this:

My initial idea is that other source files only need to include 
"etnaviv_pci_drv.h",

OK, new.  will be fixed in the next version.

>> +#ifdef CONFIG_DRM_ETNAVIV_PCI_DRIVER
>> +extern struct pci_driver etnaviv_pci_driver;
>> +#endif
>> +
>> +#endif

-- 
Jingfeng