Re: [PATCH v4 14/15] cxl/pci: Add support to assign and clear pci_driver::cxl_err_handlers

["Bowman, Terry" <terry.bowman@amd.com>]

On 12/24/2024 12:50 PM, Jonathan Cameron wrote:
> On Wed, 11 Dec 2024 17:40:01 -0600
> Terry Bowman <terry.bowman@amd.com> wrote:
>
>> pci_driver::cxl_err_handlers are not currently assigned handler callbacks.
>> The handlers can't be set in the pci_driver static definition because the
>> CXL PCIe Port devices are bound to the portdrv driver which is not CXL
>> driver aware.
>>
>> Add cxl_assign_port_error_handlers() in the cxl_core module. This
>> function will assign the default handlers for a CXL PCIe Port device.
>>
>> When the CXL Port (cxl_port or cxl_dport) is destroyed the device's
>> pci_driver::cxl_err_handlers must be set to NULL indicating they should no
>> longer be used.
>>
>> Create cxl_clear_port_error_handlers() and register it to be called
>> when the CXL Port device (cxl_port or cxl_dport) is destroyed.
>>
>> Signed-off-by: Terry Bowman <terry.bowman@amd.com>
>> ---
>>  drivers/cxl/core/pci.c | 40 ++++++++++++++++++++++++++++++++++++++++
>>  1 file changed, 40 insertions(+)
>>
>> diff --git a/drivers/cxl/core/pci.c b/drivers/cxl/core/pci.c
>> index 3294ad5ff28f..9734a4c55b29 100644
>> --- a/drivers/cxl/core/pci.c
>> +++ b/drivers/cxl/core/pci.c
>> @@ -841,8 +841,38 @@ static bool cxl_port_error_detected(struct pci_dev *pdev)
>>  	return __cxl_handle_ras(&pdev->dev, ras_base);
>>  }
>>  
>> +static const struct cxl_error_handlers cxl_port_error_handlers = {
>> +	.error_detected	= cxl_port_error_detected,
>> +	.cor_error_detected = cxl_port_cor_error_detected,
>> +};
>> +
>> +static void cxl_assign_port_error_handlers(struct pci_dev *pdev)
>> +{
>> +	struct pci_driver *pdrv;
>> +
>> +	if (!pdev || !pdev->driver)
>> +		return;
>> +
>> +	pdrv = pdev->driver;
> What stops a race here?  It's fiddly to remove that driver but
> it can be done.  At least I think we are messing withe portdrv
> but this is such a fiddly stack I'm not 100% sure.
>
>> +	pdrv->cxl_err_handler = &cxl_port_error_handlers;
>> +}
>> +
>> +static void cxl_clear_port_error_handlers(void *data)
>> +{
>> +	struct pci_dev *pdev = data;
>> +	struct pci_driver *pdrv;
>> +
>> +	if (!pdev || !pdev->driver)
>> +		return;
>> +
>> +	pdrv = pdev->driver;
> Likewise. Smells like a possible race.
>
>> +	pdrv->cxl_err_handler = NULL;
>> +}
>> +

I can add a get_device()/put_device() for both cxl_clear_port_error_handlers() and cxl_assign_port_error_handlers() to prevent operating on a recently destroyed pci_dev. Is that sufficient? Regards, Terry
>>  void cxl_uport_init_ras_reporting(struct cxl_port *port)
>>  {
>> +	struct pci_dev *pdev = to_pci_dev(port->uport_dev);
>> +
>>  	/* uport may have more than 1 downstream EP. Check if already mapped. */
>>  	if (port->uport_regs.ras)
>>  		return;
>> @@ -853,6 +883,9 @@ void cxl_uport_init_ras_reporting(struct cxl_port *port)
>>  		dev_err(&port->dev, "Failed to map RAS capability.\n");
>>  		return;
>>  	}
>> +
>> +	cxl_assign_port_error_handlers(pdev);
>> +	devm_add_action_or_reset(port->uport_dev, cxl_clear_port_error_handlers, pdev);
>>  }
>>  EXPORT_SYMBOL_NS_GPL(cxl_uport_init_ras_reporting, CXL);
>>  
>> @@ -864,6 +897,7 @@ void cxl_dport_init_ras_reporting(struct cxl_dport *dport)
>>  {
>>  	struct device *dport_dev = dport->dport_dev;
>>  	struct pci_host_bridge *host_bridge = to_pci_host_bridge(dport_dev);
>> +	struct pci_dev *pdev = to_pci_dev(dport_dev);
>>  
>>  	dport->reg_map.host = dport_dev;
>>  	if (dport->rch && host_bridge->native_aer) {
>> @@ -880,6 +914,12 @@ void cxl_dport_init_ras_reporting(struct cxl_dport *dport)
>>  		dev_err(dport_dev, "Failed to map RAS capability.\n");
>>  		return;
>>  	}
>> +
>> +	if (dport->rch)
>> +		return;
>> +
>> +	cxl_assign_port_error_handlers(pdev);
>> +	devm_add_action_or_reset(dport_dev, cxl_clear_port_error_handlers, pdev);
>>  }
>>  EXPORT_SYMBOL_NS_GPL(cxl_dport_init_ras_reporting, CXL);
>>