[PATCH] i2c: cadence: Fix regression with bus recovery

[Carsten Haitzler]

On 11/28/22 06:24, Datta, Shubhrajyoti wrote:
> [AMD Official Use Only - General]
> 
> 
> Hi ,
> Thanks for the patch
> 
>> -----Original Message-----
>> From: carsten.haitzler at foss.arm.com <carsten.haitzler at foss.arm.com>
>> Sent: Friday, November 25, 2022 5:21 PM
>> To: michal.simek at xilinx.com; shubhrajyoti.datta at xilinx.com; linux-arm-
>> kernel at lists.infradead.org; linux-i2c at vger.kernel.org; linux-
>> kernel at vger.kernel.org
>> Cc: torvalds at linux-foundation.org
>> Subject: [PATCH] i2c: cadence: Fix regression with bus recovery
>>
>> CAUTION: This message has originated from an External Source. Please use
>> proper judgment and caution when opening attachments, clicking links, or
>> responding to this email.
>>
>>
>> From: Carsten Haitzler <carsten.haitzler at arm.com>
>>
>> Commit "i2c: cadence: Add standard bus recovery support" breaks for i2c
>> devices that have no pinctrl defined. There is no requirement for this to exist
>> in the DT. This has worked perfectly well without this before in at least 1 real
>> usage case on hardware (Mali Komeda DPU, Cadence i2c to talk to a tda99xx
>> phy). Adding the requirement to have pinctrl set up in the device tree (or
>> otherwise be found) is a regression where the whole i2c device is lost
>> entirely (in this case dropping entire devices which then leads to the drm
>> display stack unable to find the phy for display output, thus having no drm
>> display device and so on down the chain).
>>
>> This converts the above commit to an enhancement if pinctrl can be found
>> for the i2c device, providing a timeout on read with recovery, but if not, do
>> what used to be done rather than a fatal loss of a device.
>>
>> This restores the mentioned display devices to their working state again.
>>
>> Fixes: 58b924241d0a ("i2c: cadence: Add standard bus recovery support")
>> Signed-off-by: Carsten Haitzler <carsten.haitzler at arm.com>
>> ---
>> Note: This issue was discovered during the porting of the linux kernel on
>> Morello [1].
>>
>> [1]
>> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgit.
>> morello-
>> project.org%2Fmorello%2Fkernel%2Flinux&data=05%7C01%7Cshubhraj
>> yoti.datta%40amd.com%7C651e141296fb497b6c2e08dacedb65ad%7C3dd896
>> 1fe4884e608e11a82d994e183d%7C0%7C0%7C638049738949855658%7CUnkn
>> own%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik
>> 1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=ny08zTdgJKxXC
>> dwYaCWFS3mwd65LDqFAIYfjao6gO8Y%3D&reserved=0
>> ---
>>   drivers/i2c/busses/i2c-cadence.c | 10 +++++-----
>>   1 file changed, 5 insertions(+), 5 deletions(-)
>>
>> diff --git a/drivers/i2c/busses/i2c-cadence.c b/drivers/i2c/busses/i2c-
>> cadence.c
>> index fe0cd205502d..40d0cc6bb996 100644
>> --- a/drivers/i2c/busses/i2c-cadence.c
>> +++ b/drivers/i2c/busses/i2c-cadence.c
>> @@ -852,7 +852,8 @@ static int cdns_i2c_master_xfer(struct i2c_adapter
>> *adap, struct i2c_msg *msgs,
>>                                           CDNS_I2C_POLL_US, CDNS_I2C_TIMEOUT_US);
>>          if (ret) {
>>                  ret = -EAGAIN;
>> -               i2c_recover_bus(adap);
>> +               if (id->adap.bus_recovery_info)
>> +                       i2c_recover_bus(adap);
>>                  goto out;
>>          }
>>
>> @@ -1262,10 +1263,10 @@ static int cdns_i2c_probe(struct platform_device
>> *pdev)
>>          }
>>
>>          id->rinfo.pinctrl = devm_pinctrl_get(&pdev->dev);
>> -       if (IS_ERR(id->rinfo.pinctrl)) {
>> +       if (IS_ERR(id->rinfo.pinctrl))
>>                  dev_info(&pdev->dev, "can't get pinctrl, bus recovery not
>> supported\n");
>> -               return PTR_ERR(id->rinfo.pinctrl);
>> -       }
>> +       else
>> +               id->adap.bus_recovery_info = &id->rinfo;
> 
> May be we could differentiate between  ENODEV and  other errors.
> As we want to do this for the case when it is not provided.

We could, but just keep in mind this patch is essentially restoring 
behavior on any pinctrl error to previous behavior - so trying to undo 
the regression as strictly as possible.

Certainly checking specifically for ENODEV does fix it for me which is 
the case of no pinctrl. I'll send a v2.

>>          id->membase = devm_platform_get_and_ioremap_resource(pdev, 0,
>> &r_mem);
>>          if (IS_ERR(id->membase))
>> @@ -1283,7 +1284,6 @@ static int cdns_i2c_probe(struct platform_device
>> *pdev)
>>          id->adap.retries = 3;           /* Default retry value. */
>>          id->adap.algo_data = id;
>>          id->adap.dev.parent = &pdev->dev;
>> -       id->adap.bus_recovery_info = &id->rinfo;
>>          init_completion(&id->xfer_done);
>>          snprintf(id->adap.name, sizeof(id->adap.name),
>>                   "Cadence I2C at %08lx", (unsigned long)r_mem->start);
>> --
>> 2.32.0