Re: [PATCH v2] PCI: endpoint: Use blocking notifier instead of atomic

[Kishon Vijay Abraham I]

Hi Mani,

On 28/02/22 11:58 am, Manivannan Sadhasivam wrote:
> Hi,
> 
> On Mon, Feb 28, 2022 at 11:46:52AM +0530, Kishon Vijay Abraham I wrote:
>> Hi Manivannan,
>>
>> On 28/02/22 11:22 am, Manivannan Sadhasivam wrote:
>>> The use of atomic notifier causes sleeping in atomic context bug when
>>> the EPC core functions are used in the notifier chain. This is due to the
>>> use of epc->lock (mutex) in core functions protecting the concurrent use of
>>> EPC.
>>
>> The notification from the controller to the function driver is used for
>> propagating interrupts to function driver and should be in interrupt context.
>> How it should be handled maybe left to the function driver. I don't prefer
>> moving everything to blocking notifier.
>>
> 
> I agree that we need to handle it quick enough but I don't see any other valid
> options to get rid of the issue. EPF driver may use a non-atomic notifier but
> that seems to be an overkill workaround for something that could be fixed in the
> EPC core.
> 
> And propagating interrupts is not going to work or needed all the time. Do you
> forsee any issue with blocking notifier?

I think any interrupt to the EP should be delivered to the function driver in
interrupt context, it could be function level reset interrupt, hot reset
interrupt, link state interrupt etc., These are right now not supported but it
will use the same notification mechanism to propagate interrupt from controller
driver to function driver.

Thanks,
Kishon

> 
>> I'm wondering how other users for CORE_INIT didn't see this issue.
> 
> This can be triggered with EPF test or NTB if CONFIG_DEBUG_ATOMIC_SLEEP is
> enabled.
> 
> Thanks,
> Mani
> 
>>
>> Thanks,
>> Kishon
>>
>>>
>>> So switch to blocking notifier for getting rid of the bug as it runs in
>>> non-atomic context and allows sleeping in notifier chain.
>>>
>>> Signed-off-by: Manivannan Sadhasivam <manivannan.sadhasivam@xxxxxxxxxx>
>>> ---
>>>
>>> Changes in v2:
>>>
>>> * Removed the changes related to non-upstreamed patches
>>>
>>>  drivers/pci/endpoint/pci-epc-core.c | 6 +++---
>>>  include/linux/pci-epc.h             | 4 ++--
>>>  2 files changed, 5 insertions(+), 5 deletions(-)
>>>
>>> diff --git a/drivers/pci/endpoint/pci-epc-core.c b/drivers/pci/endpoint/pci-epc-core.c
>>> index 3bc9273d0a08..c4347f472618 100644
>>> --- a/drivers/pci/endpoint/pci-epc-core.c
>>> +++ b/drivers/pci/endpoint/pci-epc-core.c
>>> @@ -693,7 +693,7 @@ void pci_epc_linkup(struct pci_epc *epc)
>>>  	if (!epc || IS_ERR(epc))
>>>  		return;
>>>  
>>> -	atomic_notifier_call_chain(&epc->notifier, LINK_UP, NULL);
>>> +	blocking_notifier_call_chain(&epc->notifier, LINK_UP, NULL);
>>>  }
>>>  EXPORT_SYMBOL_GPL(pci_epc_linkup);
>>>  
>>> @@ -710,7 +710,7 @@ void pci_epc_init_notify(struct pci_epc *epc)
>>>  	if (!epc || IS_ERR(epc))
>>>  		return;
>>>  
>>> -	atomic_notifier_call_chain(&epc->notifier, CORE_INIT, NULL);
>>> +	blocking_notifier_call_chain(&epc->notifier, CORE_INIT, NULL);
>>>  }
>>>  EXPORT_SYMBOL_GPL(pci_epc_init_notify);
>>>  
>>> @@ -774,7 +774,7 @@ __pci_epc_create(struct device *dev, const struct pci_epc_ops *ops,
>>>  
>>>  	mutex_init(&epc->lock);
>>>  	INIT_LIST_HEAD(&epc->pci_epf);
>>> -	ATOMIC_INIT_NOTIFIER_HEAD(&epc->notifier);
>>> +	BLOCKING_INIT_NOTIFIER_HEAD(&epc->notifier);
>>>  
>>>  	device_initialize(&epc->dev);
>>>  	epc->dev.class = pci_epc_class;
>>> diff --git a/include/linux/pci-epc.h b/include/linux/pci-epc.h
>>> index a48778e1a4ee..04a2e74aed63 100644
>>> --- a/include/linux/pci-epc.h
>>> +++ b/include/linux/pci-epc.h
>>> @@ -149,7 +149,7 @@ struct pci_epc {
>>>  	/* mutex to protect against concurrent access of EP controller */
>>>  	struct mutex			lock;
>>>  	unsigned long			function_num_map;
>>> -	struct atomic_notifier_head	notifier;
>>> +	struct blocking_notifier_head	notifier;
>>>  };
>>>  
>>>  /**
>>> @@ -195,7 +195,7 @@ static inline void *epc_get_drvdata(struct pci_epc *epc)
>>>  static inline int
>>>  pci_epc_register_notifier(struct pci_epc *epc, struct notifier_block *nb)
>>>  {
>>> -	return atomic_notifier_chain_register(&epc->notifier, nb);
>>> +	return blocking_notifier_chain_register(&epc->notifier, nb);
>>>  }
>>>  
>>>  struct pci_epc *
>>>