On 14.07.2020 19:14, Guenter Roeck wrote:
> On 7/14/20 7:11 AM, Maciej S. Szmigiero wrote:
>> On 14.07.2020 16:02, Guenter Roeck wrote:
>>> On 7/11/20 1:41 PM, Maciej S. Szmigiero wrote:
>>>> It has been observed that Toshiba DT01ACA family drives have
>>>> WRITE FPDMA QUEUED command timeouts and sometimes just freeze until
>>>> power-cycled under heavy write loads when their temperature is getting
>>>> polled in SCT mode. The SMART mode seems to be fine, though.
>>>>
>>>> Let's make sure we don't use SCT mode for these drives then.
>>>>
>>>> While only the 3 TB model was actually caught exhibiting the problem let's
>>>> play safe here to avoid data corruption and extend the ban to the whole
>>>> family.
>>>>
>>>> Fixes: 5b46903d8bf3 ("hwmon: Driver for disk and solid state drives with temperature sensors")
>>>> Cc: stable@xxxxxxxxxxxxxxx
>>>> Signed-off-by: Maciej S. Szmigiero <mail@xxxxxxxxxxxxxxxxxxxxx>
>>>> ---
>>>>
>>>> Notes:
>>>> This behavior was observed on two different DT01ACA3 drives.
>>>>
>>>> Usually, a series of queued WRITE FPDMA QUEUED commands just time out,
>>>> but sometimes the whole drive freezes. Merely disconnecting and
>>>> reconnecting SATA interface cable then does not unfreeze the drive.
>>>>
>>>> One has to disconnect and reconnect the drive power connector for the
>>>> drive to be detected again (suggesting the drive firmware itself has
>>>> crashed).
>>>>
>>>> This only happens when the drive temperature is polled very often (like
>>>> every second), so occasional SCT usage via smartmontools is probably
>>>> safe.
>>>>
>>>> Changes from v1:
>>>> 'SCT blacklist' -> 'SCT avoid models'
>>>>
>>>> drivers/hwmon/drivetemp.c | 37 +++++++++++++++++++++++++++++++++++++
>>>> 1 file changed, 37 insertions(+)
>>>>
>>>> diff --git a/drivers/hwmon/drivetemp.c b/drivers/hwmon/drivetemp.c
>>>> index 0d4f3d97ffc6..da9cfcbecc96 100644
>>>> --- a/drivers/hwmon/drivetemp.c
>>>> +++ b/drivers/hwmon/drivetemp.c
>>>> @@ -285,6 +285,36 @@ static int drivetemp_get_scttemp(struct drivetemp_data *st, u32 attr, long *val)
>>>> return err;
>>>> }
>>>>
>>>> +static const char * const sct_avoid_models[] = {
>>>> +/*
>>>> + * These drives will have WRITE FPDMA QUEUED command timeouts and sometimes just
>>>> + * freeze until power-cycled under heavy write loads when their temperature is
>>>> + * getting polled in SCT mode. The SMART mode seems to be fine, though.
>>>> + *
>>>> + * While only the 3 TB model was actually caught exhibiting the problem
>>>> + * let's play safe here to avoid data corruption and ban the whole family.
>>>> + */
>>>> + "TOSHIBA DT01ACA0",
>>>> + "TOSHIBA DT01ACA1",
>>>> + "TOSHIBA DT01ACA2",
>>>> + "TOSHIBA DT01ACA3",
>>>> +};
>>>> +
>>>> +static bool drivetemp_sct_avoid(struct drivetemp_data *st)
>>>> +{
>>>> + struct scsi_device *sdev = st->sdev;
>>>> + unsigned int ctr;
>>>> +
>>>> + if (!sdev->model)
>>>> + return false;
>>>> +
>>>> + for (ctr = 0; ctr < ARRAY_SIZE(sct_avoid_models); ctr++)
>>>> + if (strncmp(sdev->model, sct_avoid_models[ctr], 16) == 0)
>>>
>>> Why strncmp, and why length 16 ? Both strings are, as far as I can see,
>>> 0 terminated. A fixed length only asks for trouble later on as more models
>>> are added to the list.
>>
>> The first 16 bytes of sdev->model contain the model number, the rest
>> seems to be the drive serial number.
>> There is no NULL separator between them.
>>
>
> If the "16" is based on some SCSI standard, there should be
> a define for it somewhere.
The "model" field seems to contain just the raw SCSI INQUIRY
"product identification" field, which is space-filled to 16 bytes, but
is not NULL-terminated.
The SCSI layer seems to just open-code the number 16 everywhere, can't
see any specific define for this.
> Otherwise, the comparison should
> use strlen(sct_avoid_models[ctr]) and explain the reason.
> In the latter case, it might possibly make sense to match
> "TOSHIBA DT01ACA" to also catch later models.
Will use strlen() for prefix matching and match on the
"TOSHIBA DT01ACA" prefix then.
> Thanks,
> Guenter
Thanks,
Maciej
