Re: [PATCH 3/4] amba: support masking data abort during identification

[Ahmad Fatoum <a.fatoum@pengutronix.de>]

Hello Sascha,

On 3/18/25 12:04, Sascha Hauer wrote:
> On Tue, Mar 18, 2025 at 09:25:06AM +0100, Ahmad Fatoum wrote:
>> Registration of an AMBA device requires reading its pid and cid
>> registers after turning on the device. This is the first I/O done to the
>> device and may end up crashing if the device tree doesn't describe
>> the correct clock and power domains or if a previous boot stage
>> should've enabled them, but didn't.
>>
>> This is not a likely issue, but when it happens it crashes the system
>> annoyingly early during device tree populating.
>>
>> Add a symbol that masks the data abort exception around the
>> identification code to turn the crash into an error message instead.
> 
> Do we need to make this optional? I could imagine I find out about the
> existence of this option only after having traced the issue down to the
> exact lines of code in which case I won't need the option anymore.

If it turns out to be needed for more platforms, we can consider
making it visible by default, but currently I think its of
rather limited use case.

Another way would be disabling the node in the barebox DT,
but I don't want to the Qemu virt DTs to diverge.

Cheers,
Ahmad

> 
> Sascha
> 
>>
>> Signed-off-by: Ahmad Fatoum <a.fatoum@pengutronix.de>
>> ---
>>  drivers/amba/Kconfig | 19 +++++++++++++++++++
>>  drivers/amba/bus.c   | 12 +++++++++++-
>>  2 files changed, 30 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/amba/Kconfig b/drivers/amba/Kconfig
>> index 444d4ce7435c..be5e053f4b84 100644
>> --- a/drivers/amba/Kconfig
>> +++ b/drivers/amba/Kconfig
>> @@ -1,3 +1,22 @@
>>  # SPDX-License-Identifier: GPL-2.0-only
>>  config ARM_AMBA
>>  	bool
>> +
>> +config ARM_AMBA_DABT_MASK
>> +	bool
>> +	prompt "Mask data abort while identifying AMBA devices" if COMPILE_TEST
>> +	depends on ARM_AMBA && ARCH_HAS_DATA_ABORT_MASK
>> +	help
>> +	  Registration of an AMBA device requires reading its pid and cid
>> +	  registers after turning on the device, which can involve enabling
>> +	  clocks, deasserting resets and powering on power domains.
>> +	  In some cases, these resources are not available to barebox running
>> +	  in the normal world and thus need to be setup by a previous
>> +	  stage bootloader running in the secure world.
>> +	  If such setup doesn't happen, accessing devices may lead to
>> +	  a data abort or complete system hang.
>> +
>> +	  As a workaround for the former case, this option can be enabled
>> +	  to gracefully handle data aborts during readout of the PID/CID
>> +	  registers. The proper solution is fixing your first stage
>> +	  bootloader or allow barebox access to the missing resources.
>> diff --git a/drivers/amba/bus.c b/drivers/amba/bus.c
>> index 6463366fafc5..5ed8336e6124 100644
>> --- a/drivers/amba/bus.c
>> +++ b/drivers/amba/bus.c
>> @@ -10,6 +10,7 @@
>>  #include <linux/amba/bus.h>
>>  #include <io.h>
>>  #include <init.h>
>> +#include <abort.h>
>>  
>>  #define to_amba_driver(d)	container_of(d, struct amba_driver, drv)
>>  
>> @@ -132,6 +133,9 @@ int amba_device_add(struct amba_device *dev)
>>  	if (ret == 0) {
>>  		u32 pid, cid;
>>  
>> +		if (IS_ENABLED(CONFIG_ARM_AMBA_DABT_MASK))
>> +			data_abort_mask();
>> +
>>  		/*
>>  		 * Read pid and cid based on size of resource
>>  		 * they are located at end of region
>> @@ -139,8 +143,14 @@ int amba_device_add(struct amba_device *dev)
>>  		pid = amba_device_get_pid(tmp, size);
>>  		cid = amba_device_get_cid(tmp, size);
>>  
>> -		if (cid == AMBA_CID)
>> +		if (IS_ENABLED(CONFIG_ARM_AMBA_DABT_MASK) && data_abort_unmask()) {
>> +			dev_err(amba_bustype.dev,
>> +				"data abort during MMIO read of PID/CID for %pOF\n",
>> +				dev->dev.of_node);
>> +			ret = -EFAULT;
>> +		} else if (cid == AMBA_CID) {
>>  			dev->periphid = pid;
>> +		}
>>  
>>  		if (!dev->periphid)
>>  			ret = -ENODEV;
>> -- 
>> 2.39.5
>>
>>
>>
>