On 2021-10-20 17:50, Felix Kuehling wrote:
> On 2021-10-20 12:35 p.m., Kent Russell wrote:
>> Currently dmesg doesn't warn when the number of bad pages approaches the
>> threshold for page retirement. WARN when the number of bad pages
>> is at 90% or greater for easier checks and planning, instead of waiting
>> until the GPU is full of bad pages
>>
>> Cc: Luben Tuikov <luben.tuikov@xxxxxxx>
>> Cc: Mukul Joshi <Mukul.Joshi@xxxxxxx>
>> Signed-off-by: Kent Russell <kent.russell@xxxxxxx>
>> ---
>> drivers/gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.c | 17 +++++++++++++++++
>> 1 file changed, 17 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.c
>> index f4c05ff4b26c..1ede0f0d6f55 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.c
>> @@ -1071,12 +1071,29 @@ int amdgpu_ras_eeprom_init(struct amdgpu_ras_eeprom_control *control,
>> control->ras_fri = RAS_OFFSET_TO_INDEX(control, hdr->first_rec_offset);
>>
>> if (hdr->header == RAS_TABLE_HDR_VAL) {
>> + int threshold = 0;
> ras->bad_page_cnt_threshold is uint32_t. I'd recommend using the same
> type. Also add an empty line after the declaration to avoid a checkpatch
> warning.
>
>
>> DRM_DEBUG_DRIVER("Found existing EEPROM table with %d records",
>> control->ras_num_recs);
>> res = __verify_ras_table_checksum(control);
>> if (res)
>> DRM_ERROR("RAS table incorrect checksum or error:%d\n",
>> res);
>> +
>> + /* threshold = 0 means that page retirement is disabled, while
>> + * threshold = -1 means default behaviour
>> + */
>> + if (amdgpu_bad_page_threshold == -1)
>> + threshold = ras->bad_page_cnt_threshold;
>> + else if (amdgpu_bad_page_threshold > 0)
>> + threshold = amdgpu_bad_page_threshold;
>> +
>> + /* Since multiplcation is transitive, a = 9b/10 is the same
>> + * as 10a = 9b. Use this for our 90% limit to avoid rounding
>> + */
>> + if (threshold > 0 && ((control->ras_num_recs * 10) >= (threshold * 9)))
> Not sure how big these values can get, but you may need to cast to
> (uint64_t) before the multiplications to avoid overflows. Alternatively
> you could use (control->ras_num_recs / 9 >= threshold / 10). It'll
> round, but never overflow.
I sincerely hope that by the time those values overflow multiplication by 10, AI has taken over the planet. :-)
Avoiding rounding is preferable, since we deal with integers, and for small pages... could we get 0s after division? (if the page limit is 1 :-) )
(I think squashing numbers less than 10 to 0 is a bad idea, so as not to get false positives here on small numbers.)
Even for a 32-bit word size: can we be within 3 bits of 2^32? A value of 2^29? That's a lot of pages! Actually the number of pages of VRAM is *a lot smaller* than the number of pages we fit in a typical EEPROM we put with our parts. I think we're safe.
>
>
>> + DRM_WARN("RAS records:%u exceeds 90%% of threshold:%d",
> Nitpick: I'd add space after the two colons for readability. The
> threshold should use %u if you make it uint32_t. This can never be negative.
Aha, so that's exactly what I like "word:%q word:%q word:%q", so when I read it in the log, my eyes scan over it, on the spaces in between the word blocks. I prefer no spaces after the colon, so as to make scanning blocks (one has to squint to see it).
Regards,
Luben
>
> Regards,
> Felix
>
>
>> + control->ras_num_recs,
>> + threshold);
>> } else if (hdr->header == RAS_TABLE_HDR_BAD &&
>> amdgpu_bad_page_threshold != 0) {
>> res = __verify_ras_table_checksum(control);
