RE: [PATCH v4 1/2] ACPI: APEI: Add support to notify the vendor specific HW errors

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Hi James,

>-----Original Message-----
>From: James Morse [mailto:james.morse@xxxxxxx]
>Sent: 13 March 2020 15:17
>To: Shiju Jose <shiju.jose@xxxxxxxxxx>
>Cc: linux-acpi@xxxxxxxxxxxxxxx; linux-pci@xxxxxxxxxxxxxxx; linux-
>kernel@xxxxxxxxxxxxxxx; rjw@xxxxxxxxxxxxx; helgaas@xxxxxxxxxx;
>lenb@xxxxxxxxxx; bp@xxxxxxxxx; tony.luck@xxxxxxxxx;
>gregkh@xxxxxxxxxxxxxxxxxxx; zhangliguang@xxxxxxxxxxxxxxxxx;
>tglx@xxxxxxxxxxxxx; Linuxarm <linuxarm@xxxxxxxxxx>; Jonathan Cameron
><jonathan.cameron@xxxxxxxxxx>; tanxiaofei <tanxiaofei@xxxxxxxxxx>;
>yangyicong <yangyicong@xxxxxxxxxx>
>Subject: Re: [PATCH v4 1/2] ACPI: APEI: Add support to notify the vendor
>specific HW errors
>
>Hi Shiju,
>
>On 3/12/20 12:10 PM, Shiju Jose wrote:
>>> On 07/02/2020 10:31, Shiju Jose wrote:
>>>> Presently APEI does not support reporting the vendor specific HW
>>>> errors, received in the vendor defined table entries, to the vendor
>>>> drivers for any recovery.
>>>>
>>>> This patch adds the support to register and unregister the error
>>>> handling function for the vendor specific HW errors and notify the
>>>> registered kernel driver.
>>>
>>> Is it possible to use the kernel's existing
>>> atomic_notifier_chain_register() API for this?
>>>
>>> The one thing that can't be done in the same way is the GUID filtering in
>ghes.c.
>>> Each driver would need to check if the call matched a GUID they knew
>>> about, and return NOTIFY_DONE if they "don't care".
>>>
>>> I think this patch would be a lot smaller if it was tweaked to be
>>> able to use the existing API. If there is a reason not to use it, it
>>> would be good to know what it is.
>
>> I think when using atomic_notifier_chain_register we have following
>limitations,
>> 1. All the registered error handlers would get called, though an error is not
>related to those handlers.
>
>The notifier chain provides NOTIFY_STOP_MASK, so that one of the callers can
>say the work is done. We only expect a handful of these, so I don't think there is
>going to be a scalability problem.
Ok. I will check the error reporting by using atomic_notifier_chain and test.

>
>
>>     Also this may lead to mishandling of the error information if a handler does
>not
>>     implement GUID checking etc.
>
>Which would be a bug we can fix.
>There is no point worrying about bugs in out of tree code.
Ok.

>
>
>> 2. atomic_notifier_chain_register (notifier_chain_register) looks like does not
>support
>>     pass the handler's private data during the registration which supposed to
>>     passed later in the call back function *notifier_fn_t(... ,void *data) to the
>handler.
>
>The callback is provided with the struct notifier_block. A bit of
>container_of() magic will give you whatever structure you embedded it in!
Ok. I will check this.
 
>
>
>> 3. Also got difficulty in passing the ghes error data(acpi_hest_generic_data),
>GUID
>>     for the error received to the handler through the notifier_chain  callback
>interface.
>
>Here you've lost me. Because you need to pass more than one thing? Can't we
>have a struct for that?
>
>But, isn't it all in struct acpi_hest_generic_data already? That is where the guid
>and severity come from.
Ok.  right. 
 
>
>
>>>> diff --git a/drivers/acpi/apei/ghes.c b/drivers/acpi/apei/ghes.c
>>>> index
>>>> 103acbb..69e18d7 100644
>>>> --- a/drivers/acpi/apei/ghes.c
>>>> +++ b/drivers/acpi/apei/ghes.c
>>>> @@ -490,6 +490,109 @@ static void ghes_handle_aer(struct
>>>> acpi_hest_generic_data *gdata)
>>>
>>>> +/**
>>>> + * ghes_unregister_event_handler - unregister the previously
>>>> + * registered event handling function.
>>>> + * @sec_type: sec_type of the corresponding CPER.
>>>> + * @data: driver specific data to distinguish devices.
>>>> + */
>>>> +void ghes_unregister_event_handler(guid_t sec_type, void *data) {
>>>> +	struct ghes_event_notify *event_notify;
>>>> +	bool found = false;
>>>> +
>>>> +	mutex_lock(&ghes_event_notify_mutex);
>>>> +	rcu_read_lock();
>>>> +	list_for_each_entry_rcu(event_notify,
>>>> +				&ghes_event_handler_list, list) {
>>>> +		if (guid_equal(&event_notify->sec_type, &sec_type)) {
>>>
>>>> +			if (data != event_notify->data)
>>>
>>> It looks like you need multiple drivers to handle the same GUID
>>> because of multiple root ports. Can't the handler lookup the right device?
>
>> This check was because GUID is shared among multiple devices with one
>> driver as seen in the B2889FC9 driver (pcie-hisi-error.c).
>
>(we should stop calling it by its guid ... does it have a name?!)
>
>
>This must be some kind of error collector for a bus right?
>
>I agree we may need to have multiple drivers register to handle vendor events,
>but it looks like you are registering the same handler multiple times, with
>different private structures.
>
>Can't it find the affected device from the error description?
Yes. We already have the code in the PCIe error handling driver to identify the right device
from the error information.

>
>
>>>> @@ -525,11 +628,14 @@ static void ghes_do_proc(struct ghes *ghes,
>>>>
>>>>  			log_arm_hw_error(err);
>>>>  		} else {
>>>> -			void *err = acpi_hest_get_payload(gdata);
>>>> -
>>>> -			log_non_standard_event(sec_type, fru_id, fru_text,
>>>> -					       sec_sev, err,
>>>> -					       gdata->error_data_length);
>>>> +			if (!ghes_handle_non_standard_event(sec_type, gdata,
>>>> +							    sev)) {
>>>> +				void *err = acpi_hest_get_payload(gdata);
>>>> +
>>>> +				log_non_standard_event(sec_type, fru_id,
>>>> +						       fru_text, sec_sev, err,
>>>> +						       gdata->error_data_length);
>>>> +			}
>>>
>>> So, a side effect of the kernel handling these is they no longer get
>>> logged out of trace points?
>>>
>>> I guess the driver the claims this logs some more accurate
>>> information. Are there expected to be any user-space programs doing
>>> something useful with B2889FC9... today?
>
>> The B2889FC9 driver does not expect any corresponding user space
>programs.
>> The driver mainly for the error recovery and basic error decoding and logging.
>
>> Previously we added the error logging for the B2889FC9 in the rasdaemon.
>
>So this series would break the error logging in rasdaemon.
It does not affect the logging information to the user for the HiSilicon PCIe controller errors
because the level of logging information is the same both in the rasdaemon and in the
newly adding HiSilicon PCIe controller error handling driver.
>
>User-space would need to be upgraded to receive the trace information from
>the specific driver instead. (how does it know?!)
>
>Could we log_non_standard_event() unconditionally, maybe adding a field to
>indicate that a driver claimed it, so there may be more data somewhere else...
sure, I will check the possibility of adding the field to indicate driver claimed it and
calling log_non_standard_event() always.
>
>
>Thanks,
>
>James

Thanks,
Shiju




[Index of Archives]     [Linux IBM ACPI]     [Linux Power Management]     [Linux Kernel]     [Linux Laptop]     [Kernel Newbies]     [Share Photos]     [Security]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Video 4 Linux]     [Device Mapper]     [Linux Resources]

  Powered by Linux