Re: [PATCH v3 4/6] drm/xe/cper: Make software CPER emission atomic-safe

["Mallesh, Koujalagi" <mallesh.koujalagi@intel.com>]

On 10-11-2026 08:31 pm, Nilawar, Badal wrote:
>
> On 05-11-2026 14:23, Mallesh Koujalagi wrote:
>> Use GFP_ATOMIC for software CPER records, as software
>> SIGID logging may occur in atomic context.
>>
>> Keep hardware CPER handling unchanged.
>>
>> Cc: Badal Nilawar <badal.nilawar@intel.com>
>> Signed-off-by: Mallesh Koujalagi <mallesh.koujalagi@intel.com>
>> ---
>> v3:
>> - Handle atomic context when SIGIDs are logged. (Sashiko)
>> ---
>>   drivers/gpu/drm/xe/xe_cper.c | 10 ++++++----
>>   1 file changed, 6 insertions(+), 4 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/xe/xe_cper.c b/drivers/gpu/drm/xe/xe_cper.c
>> index 9510dcd1754c..8c683e320194 100644
>> --- a/drivers/gpu/drm/xe/xe_cper.c
>> +++ b/drivers/gpu/drm/xe/xe_cper.c
>> @@ -173,7 +173,7 @@ static void cper_init_section_desc(struct 
>> xe_device *xe, u8 severity,
>>   static void cper_record_emit(struct xe_device *xe, u8 severity,
>>                    const guid_t *notification_type,
>>                    const struct xe_cper_intel_hdr *ihdr,
>> -                 const void *einfo, u32 einfo_len)
>> +                 const void *einfo, u32 einfo_len, gfp_t gfp)
> May be bool can_sleep or bool atomic?
>>   {
>>       struct cper_section_descriptor *sdesc;
>>       struct xe_cper_nonstd_record *rec;
>> @@ -181,7 +181,7 @@ static void cper_record_emit(struct xe_device 
>> *xe, u8 severity,
>>       u32 total_len;
>>         total_len = sizeof(struct xe_cper_nonstd_record) + einfo_len;
>> -    rec = kzalloc(total_len, GFP_KERNEL);
>> +    rec = kzalloc(total_len, gfp);
>>       if (!rec)
>>           return;
>>   @@ -228,7 +228,8 @@ void xe_cper_emit_counter(struct xe_device *xe, 
>> int cper_sev, enum xe_sigid sigi
>>         cper_init_intel_hdr(xe, &error_class, response->timestamp,
>>                   sigid, response->value, &ihdr);
>> -    cper_record_emit(xe, cper_sev, &INTEL_CPER_NOTIFY_GPU_ERROR, 
>> &ihdr, NULL, 0);
>> +    cper_record_emit(xe, cper_sev, &INTEL_CPER_NOTIFY_GPU_ERROR, 
>> &ihdr, NULL, 0,
>> +             GFP_KERNEL);
>>   }
>>     static void cper_emit_hw_err(struct xe_device *xe, int cper_sev, 
>> enum xe_sigid sigid,
>> @@ -275,7 +276,8 @@ static void cper_emit_sw_err(struct xe_device 
>> *xe, int cper_sev, enum xe_sigid s
>>       msg_len = cper_format_sw_msg(vaf, msg, sizeof(msg));
>>         cper_init_intel_hdr(xe, &error_class, timestamp, sigid, 1, 
>> &ihdr);
>> -    cper_record_emit(xe, cper_sev, &INTEL_CPER_NOTIFY_DRV_ERROR, 
>> &ihdr, msg, msg_len);
>> +    cper_record_emit(xe, cper_sev, &INTEL_CPER_NOTIFY_DRV_ERROR, 
>> &ihdr, msg, msg_len,
>> +             GFP_ATOMIC | __GFP_NOWARN);
>
> Why __GFP_NOWARN needed here?

If allocation failed in atomic context, it's going to logged in dmesg. 
__GFP_NOWARN help to avoid

filling dmesg with additional allocator warning (reduce dmesg noise).


Thanks,

-/Mallesh

>
> Thanks,
> Badal
>
>>   }
>>     /**