Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Subject: Re: [PATCH] firmware: arm_sdei: pass sdei_api_event_register right
 parameters
To:     =?UTF-8?B?5Lmx55+z?= <zhangliguang@linux.alibaba.com>
Cc:     linux-arm-kernel@lists.infradead.org, linux-kernel@vger.kernel.org
References: <20210910040117.98736-1-zhangliguang@linux.alibaba.com>
 <3fb354d1-bdc1-8aa2-aa90-4fd92e9a2e9a@arm.com>
 <b93cf74a-ec1a-dcfc-990b-d3dbc4b55c3d@linux.alibaba.com>
From:   James Morse <james.morse@arm.com>
Message-ID: <f983a17a-27d4-d43b-e6e6-9f03aa5508b0@arm.com>
Date:   Mon, 18 Oct 2021 18:32:33 +0100
User-Agent: Mozilla/5.0 (X11; Linux aarch64; rv:78.0) Gecko/20100101
 Thunderbird/78.12.0
MIME-Version: 1.0
In-Reply-To: <b93cf74a-ec1a-dcfc-990b-d3dbc4b55c3d@linux.alibaba.com>
Content-Type: text/plain; charset=utf-8
Content-Language: en-GB
Content-Transfer-Encoding: 8bit
Precedence: bulk

Hi Liguang,

On 11/10/2021 06:40, 乱石 wrote:
> 在 2021/10/9 1:39, James Morse 写道:
>> On 10/09/2021 05:01, Liguang Zhang wrote:
>>> Function _local_event_enable is used for private sdei event
>>> registeration called by sdei_event_register. We should pass
>> (registration)

>>> sdei_api_event_register right flag and mpidr parameters, otherwise atf
>>> may trigger assert errors.
>>> diff --git a/drivers/firmware/arm_sdei.c b/drivers/firmware/arm_sdei.c
>>> index a7e762c352f9..0736752dadde 100644
>>> --- a/drivers/firmware/arm_sdei.c
>>> +++ b/drivers/firmware/arm_sdei.c
>>> @@ -558,14 +558,16 @@ static int sdei_api_event_register(u32 event_num, unsigned long
>>> entry_point,
>>>   static void _local_event_register(void *data)
>>>   {
>>>       int err;
>>> +    u64 mpidr;
>>>       struct sdei_registered_event *reg;
>>>       struct sdei_crosscall_args *arg = data;
>>>         WARN_ON(preemptible());
>>>   +    mpidr = read_cpuid_mpidr();
>>>       reg = per_cpu_ptr(arg->event->private_registered, smp_processor_id());
>>>       err = sdei_api_event_register(arg->event->event_num, sdei_entry_point,
>>> -                      reg, 0, 0);
>>> +                      reg, SDEI_EVENT_REGISTER_RM_PE, mpidr);

>> Hmmm, this looks like a bug in TFA.
>>
>> 5.1.2.2 "Parameters" of DEN 0054B has:
>> | Routing mode is valid only for a shared event. For a private event, the routing mode is
>> | ignored.
>>
>> Worse, the mpidr field has:
>> | Currently the format is defined only when the selected routing mode is RM_PE.


> For a private event, we route SDEI_EVENT_REGISTER_RM_PE and mpidr parameters may be more
> rationable.

You are making this call from Linux?

This isn't valid for private events. Private events are private to the CPU - they can only
be reset, register and taken on that CPU. The specification for SDEI_EVENT_ROUTING_SET has
this:
| This call is used to change the routing information of a shared event.

To borrow the GIC's terms: Private events are like PPI, Shared events are like SPI.


>> Over in trusted firmware land:
>>
>> https://git.trustedfirmware.org/TF-A/trusted-firmware-a.git/tree/services/std_svc/sdei/sdei_main.c?h=v2.5#n361
>>
>>
>> | static int64_t sdei_event_register(int ev_num,
>> |                 uint64_t ep,
>> |                 uint64_t arg,
>> |                 uint64_t flags,
>> |                uint64_t mpidr)
>> | {
>>
>> |    /* Private events always target the PE */
>> |    if (is_event_private(map))
>> |        flags = SDEI_REGF_RM_PE;
>>
>> It looks like this re-uses the 'caller specified the routing' code, but didn't update the
>> mpidr.
>>
>>
>> You mention TFA takes an assert failure, I assume that brings the machine down.
>> (Presumably you don't have a CPU with an affinity of zero.)

> Yes, that brings the machine down. In opensource ATF, CPU with an affinity of zero.
> 
> The problem backaround:
> 
> we use local secure arch timer as sdei watchdog timer 

Is that an SPI? If so, you should really be generating a shared event.


> for hardlockup detection, in  os
> panic ,we mask sdei event, then trigger the assert

> if (se->reg_flags == SDEI_REGF_RM_PE)
>
>     assert(se->affinity == my_mpidr);


I'm not sure where this code in TFA is, but RM_PE for a private event is going to hit this
on all but one CPU. You shouldn't be able to set RM_PE for a private event.


I assume this is the TFA side of the problem from your colleague:
https://review.trustedfirmware.org/c/TF-A/trusted-firmware-a/+/11393


Does the problem occur with this TFA patch applied, and without any attempt to mess with
the routing of per-cpu/private events?


Thanks,

James