On 10/26/18, Hannes Reinecke <hare@xxxxxxxx> wrote:
> On 10/26/18 10:16 AM, Arnd Bergmann wrote:
>> On Wed, Oct 24, 2018 at 1:00 PM James Bottomley
>> <James.Bottomley@xxxxxxxxxxxxxxxxxxxxx> wrote:
>>>
>>> James Bottomley (1):
>>> scsi: myrs: fix build failure on 32 bit
>>
>> Hi James and Hannes,
>>
>> Since James mentioned 32-bit compiles during the kernel summit,
>> I'd like to confirm that I hit this on my randconfig builder now,
>> with some latency since the last linux-next tree I tested before
>> flying to Edinburgh did not have the bug, and the latest
>> linux-next tree that is available now (dated last Friday) does, and
>> I see your tree is fixed. During normal times, I should catch these
>> within a short time of the patch getting into scsi-next.
>>
>> However, while looking at this bug, I found two more issues related
>> to the specific computation:
>>
>> percent_complete = ldev_info->rbld_lba * 100 / ldev_info->cfg_devsize;
>>
>> I see that both rbld_lba and cfg_devsize are reported by the
>> device, but only the former is 64 bit but the latter is 32 bit and
>> also intended to be the larger of the two. I suspect this is a
>> bug, and the same is also present in the old DAC960.c.
>> cfg_devsize is followed by four reserved bytes in the header,
>> so I suppose it was meant to be 64-bit?
>> If you divide two 64-bit numbers, you also have to use div_u64_64()
>> instead of do_div().
>>
>> On top of that, I see we get those values from the device but
>> never do any endianess conversion on them. It seems likely
>> that they are all little-endian and require a le32_to_cpu()
>> conversion to also work on big-endian kernel builds. Alternatively
>> we could make the Kconfig symbol as
>> 'depends on !CPU_BIG_ENDIAN || COMPILE_TEST'.
>>
> It _really_ is questionable if these device ever work on big-endian
> machines, as they rely on the BIOS to start up the RAID engine; I've had
> a hard enough time getting them to work on normal machines :-)
Shall we add the Kconfig dependency then? I can send a patch for that.
> Plus the firmware refused to handle any drive larger than 4GB (!), so
> it's really a purely theoretical issue whether 'cfgsize' was meant to be
> 64 bit ...
I see. It was likely meant as a possible extension for future firmware
version then, which may or may not have been developed or released.
I suppose the do_div() could be replaced with 32-bit arithmetic, but
it's not in a fast path and James' version should be correct as well, so
I won't send another patch for that one.
Arnd
