Re: [PATCH nouveau 06/11] platform: complete the power up/down sequence

[Vince Hsu <vinceh@nvidia.com>]

On 01/05/2015 11:25 PM, Thierry Reding wrote:
> * PGP Signed by an unknown key
>
> On Thu, Dec 25, 2014 at 10:42:58AM +0800, Vince Hsu wrote:
>> On 12/24/2014 09:23 PM, Lucas Stach wrote:
>>> Am Dienstag, den 23.12.2014, 18:39 +0800 schrieb Vince Hsu:
>>>> This patch adds some missing pieces of the rail gaing/ungating sequence that
>>>> can improve the stability in theory.
>>>>
>>>> Signed-off-by: Vince Hsu <vinceh@nvidia.com>
>>>> ---
>>>>   drm/nouveau_platform.c | 42 ++++++++++++++++++++++++++++++++++++++++++
>>>>   drm/nouveau_platform.h |  3 +++
>>>>   2 files changed, 45 insertions(+)
>>>>
>>>> diff --git a/drm/nouveau_platform.c b/drm/nouveau_platform.c
>>>> index 68788b17a45c..527fe2358fc9 100644
>>>> --- a/drm/nouveau_platform.c
>>>> +++ b/drm/nouveau_platform.c
>>>> @@ -25,9 +25,11 @@
>>>>   #include <linux/module.h>
>>>>   #include <linux/platform_device.h>
>>>>   #include <linux/of.h>
>>>> +#include <linux/of_platform.h>
>>>>   #include <linux/reset.h>
>>>>   #include <linux/regulator/consumer.h>
>>>>   #include <soc/tegra/fuse.h>
>>>> +#include <soc/tegra/mc.h>
>>>>   #include <soc/tegra/pmc.h>
>>>>   #include "nouveau_drm.h"
>>>> @@ -61,6 +63,9 @@ static int nouveau_platform_power_up(struct nouveau_platform_gpu *gpu)
>>>>   	reset_control_deassert(gpu->rst);
>>>>   	udelay(10);
>>>> +	tegra_mc_flush(gpu->mc, gpu->swgroup, false);
>>>> +	udelay(10);
>>>> +
>>>>   	return 0;
>>>>   err_clamp:
>>>> @@ -77,6 +82,14 @@ static int nouveau_platform_power_down(struct nouveau_platform_gpu *gpu)
>>>>   {
>>>>   	int err;
>>>> +	tegra_mc_flush(gpu->mc, gpu->swgroup, true);
>>>> +	udelay(10);
>>>> +
>>>> +	err = tegra_powergate_gpu_set_clamping(true);
>>>> +	if (err)
>>>> +		return err;
>>>> +	udelay(10);
>>>> +
>>>>   	reset_control_assert(gpu->rst);
>>>>   	udelay(10);
>>>> @@ -91,6 +104,31 @@ static int nouveau_platform_power_down(struct nouveau_platform_gpu *gpu)
>>>>   	return 0;
>>>>   }
>>>> +static int nouveau_platform_get_mc(struct device *dev,
>>>> +		struct tegra_mc **mc, unsigned int *swgroup)
>>> Uhm, no. If this is needed this has to be a Tegra MC function and not
>>> burried into nouveau code. You are using knowledge about the internal
>>> workings of the MC driver here.
>>>
>>> Also this should probably only take the Dt node pointer as argument and
>>> return a something like a tegra_mc_client struct that contains both the
>>> MC device pointer and the swgroup so you can pass that to
>>> tegra_mc_flush().
>> Good idea. I will have something as below in V2 if there is no other
>> comments for this.
>>
>> tegra_mc_client *tegra_mc_find_client(struct device_node *node)
>> {
>>      ...
>>      ret = of_parse_phandle_with_args(node, "nvidia,memory-client", ...)
>>      ...
>> }
>>
>> There were some discussion about this few weeks ago. I'm not sure whether we
>> have some conclusion/implementation though. Thierry?
>>
>> http://lists.infradead.org/pipermail/linux-arm-kernel/2014-December/308703.html
> I don't think client is a good fit here. Flushing is done per SWGROUP
> (on all clients of the SWGROUP). So I think we'll want something like:
>
> 	gpu@0,57000000 {
> 		...
> 		nvidia,swgroup = <&mc TEGRA_SWGROUP_GPU>;
> 		...
> 	};
>
> In the DT and return a struct tegra_mc_swgroup along the lines of:
>
> 	struct tegra_mc_client {
> 		unsigned int id;
> 		unsigned int swgroup;
>
> 		struct list_head list;
> 	};
>
> 	struct tegra_mc_swgroup {
> 		struct list_head clients;
> 		unsigned int id;
> 	};
>
> Where tegra_mc_swgroup.clients is a list of struct tegra_mc_client
> structures, each representing a memory client pertaining to the
> SWGROUP.
Based on your suggestion above, I created a struct tegra_mc_swgroup:

struct tegra_mc_swgroup {
     unsigned int id;
     struct tegra_mc *mc;
     struct list_head head;
     struct list_head clients;
};

And added the list head in the struct tegra_mc_soc.

struct tegra_mc_soc {
     struct tegra_mc_client *clients;
     unsigned int num_clients;

     struct tegra_mc_hr *hr_clients;
     unsigned int num_hr_clients;

     struct list_head swgroups;
...

Created one function to build the swgroup list.

static int tegra_mc_build_swgroup(struct tegra_mc *mc)
{
     int i;

     for (i = 0; i < mc->soc->num_clients; i++) {
         struct tegra_mc_swgroup *sg;
         bool found = false;

         list_for_each_entry(sg, &mc->soc->swgroups, head) {
             if (sg->id == mc->soc->clients[i].swgroup) {
                 found = true;
                 break;
             }
         }

         if (!found) {
             sg = devm_kzalloc(mc->dev, sizeof(*sg), GFP_KERNEL);
             if (!sg)
                 return -ENOMEM;

             sg->id = mc->soc->clients[i].swgroup;
             sg->mc = mc;
             list_add_tail(&sg->head, &mc->soc->swgroups);
             INIT_LIST_HEAD(&sg->clients);
         }

         list_add_tail(&mc->soc->clients[i].head, &sg->clients);
     }

     return 0;
}

>
> We probably don't want to expose these structures publicly, an opaque
> type should be enough. Then you can use functions like:
>
> 	struct tegra_mc_swgroup *tegra_mc_find_swgroup(struct device_node *node);
And then I can use the tegra_find_swgroup() in GK20A driver to get the 
swgroup
and flush the memory clients by tegra_mc_flush(swgroup).

One problem is that the mc_soc and mc_clients are are defined as const. 
To build
the swgroup list dynamically, I have to discard the const. I guess you 
won't like that. :(

Thanks,
Vince

> At some point we may even need something like:
>
> 	struct tegra_mc_client *tegra_mc_find_client(struct device_node *node,
> 						     const char *name);
>
> And DT content like this:
>
> 	gpu@0,57000000 {
> 		...
> 		nvidia,memory-clients = <&mc 0x58>, <&mc 0x59>;
> 		nvidia,memory-client-names = "read", "write";
> 		...
> 	};
>
> This could be useful for latency allowance programming, but we can cross
> that bridge when we come to it.
>
> Thierry
>
> * Unknown Key
> * 0x7F3EB3A1