[linux-sunxi] Re: Problems to Allwinner H3's eFUSE/SID

Thu Dec 22 02:41:01 PST 2016

Hi,

On 22-12-16 11:31, Chen-Yu Tsai wrote:
> On Tue, Dec 20, 2016 at 12:17 AM, Hans de Goede <hdegoede at redhat.com> wrote:
>> Hi,
>>
>>
>> On 19-12-16 17:06, Icenowy Zheng wrote:
>>>
>>>
>>>
>>> 19.12.2016, 23:30, "Hans de Goede" <hdegoede at redhat.com>:
>>>>
>>>> Hi,
>>>>
>>>> On 19-12-16 16:22, Icenowy Zheng wrote:
>>>>>
>>>>>  Hi everyone,
>>>>>
>>>>>  Today, I and KotCzarny on IRC of linux-sunxi found a problem in the SID
>>>>>  controller of H3 (incl. H2+).
>>>>>
>>>>>  See https://irclog.whitequark.org/linux-sunxi/2016-12-19 .
>>>>>
>>>>>  Two read method of the H3 eFUSE is used in the BSP: by register
>>>>> accessing, or
>>>>>  directly access 0x01c14200.
>>>>>
>>>>>  From http://linux-sunxi.org/SID_Register_Guide we can see a difference
>>>>> between
>>>>>  the H3 SIDs read out by sunxi-fel and the H3 SIDs read out by devmem2
>>>>> (in
>>>>>  legacy kernel).
>>>>>
>>>>>  According to the source of H2+ BSP[1], H2+ and H3 can be differed by
>>>>> the last
>>>>>  byte of the first word of SID. (0x42 and 0x83 is H2+, 0x00 and 0x81 is
>>>>> H3,
>>>>>  0x58 is H3D (currently not known SoC) )
>>>>>
>>>>>  However, all the SIDs retrieved by `sunxi-fel sid`, both H2+ and H3,
>>>>> start
>>>>>  with 0x02004620, which do not match this rule.
>>>>>
>>>>>  The readout by devmem2 is satisfying this rule: their first word is
>>>>>  0x02c00081, matches H3.
>>>>>
>>>>>  Then I found the SID-reading code from BSP U-Boot[2], which is based on
>>>>>  register operations. With this kind of code (I wrote one prototype in
>>>>>  userspace with /dev/mem), I got "02c00081 74004620 50358720 3c27048e"
>>>>> on
>>>>>  my Orange Pi One. ("02004620 74358720 5027048e 3c0000c3" with sunxi-fel
>>>>> sid)
>>>>>  And, after accessing to the SID by registers, the value of *0x01c14200
>>>>> become
>>>>>  also "02c00081".
>>>>>
>>>>>  With direct access to 0x01c14200 after boot with mainline kernel, I got
>>>>> also
>>>>>  "02004620".
>>>>>
>>>>>  Then I altered the program to do the register operations with
>>>>> sunxi-fel, the
>>>>>  result is also "02c00081", and changed `sunxi-fel sid` result to
>>>>> "02c00081".
>>>>>
>>>>>  Summary:
>>>>>
>>>>>  +-----------------------------------------------+----------------+
>>>>>  | Read situation | The first word |
>>>>>  +-----------------------------------------------+----------------+
>>>>>  | Direct read by sunxi-fel | 02004620 |
>>>>>  | Direct read in mainline /dev/mem | 02004620 |
>>>>>  | Direct read in legacy /dev/mem | 02c00081 |
>>>>>  | Register access in FEL | 02c00081 |
>>>>>  | Register access in mainline | 02c00081 |
>>>>>  | Direct read after register access in FEL | 02c00081 |
>>>>>  | Direct read after register access in mainline | 02c00081 |
>>>>>  +-----------------------------------------------+----------------+
>>>>>
>>>>>  According to some facts:
>>>>>  - The register based access to SID is weird: it needs ~5 register
>>>>>    operations per word of SID.
>>>>>  - Reading via register access will change the value when reading by
>>>>> accessing
>>>>>    0x01c14200.
>>>>>  - In the u-boot code[2] there's some functions which read out the SID
>>>>> by
>>>>>    registers and then abandoned the value.
>>>>>  - This mismatch do not exist on A64.
>>>>>
>>>>>  I think that: Allwinner designed a "cache" to the SID to make the
>>>>> simplify the
>>>>>  code to read it, and it automatically loaded the cache when booting;
>>>>> however,
>>>>>  when doing first cache on H3, some byte shifts occured, and the value
>>>>> become
>>>>>  wrong. A manual read on H3 can make the cache right again. This is a
>>>>> silicon
>>>>>  bug, and fixed in A64.
>>>>>
>>>>>  This raises a problem: currently many systems has used the misread SID
>>>>> value to
>>>>>  generated lots of MAC addresses, and workaround this SID bug will
>>>>> change them.
>>>>>
>>>>>  However, if this bug is not workarounded, the sun8i-ths driver won't
>>>>> work well
>>>>>  (as some calibartion value lies in eFUSE). I think some early user of
>>>>> this
>>>>>  driver has already experienced bad readout value.
>>>>>  (The calibration value differs on my opi1 and KotCzarny's opipc)
>>>>>
>>>>>  And many wrong SID values have been generated by `sunxi-fel sid`.
>>>>> (Although I
>>>>>  think sunxi-fel must have the workaround)
>>>>>
>>>>>  Note: in this email, "SID" and "eFUSE" both indicate the controller on
>>>>> H3/A64
>>>>>  at 0x01c14000, which is a OTP memory implemented by eFUSE technique.
>>>>>
>>>>>  Furthermore, A83T may also have this problem, testers are welcome!
>>>>>
>>>>>  [1]
>>>>> http://filez.zoobab.com/allwinner/h2/201609022/lichee/linux-3.4/arch/arm/mach-sunxi/sun8i.c
>>>>>  [2]
>>>>> http://filez.zoobab.com/allwinner/h2/201609022/lichee/brandy/u-boot-2011.09/arch/arm/cpu/armv7/sun8iw7/efuse.c
>>>>>
>>>>>  Experiments:
>>>>>  - https://gist.github.com/Icenowy/2f4859ab1bc05814522fc7445179a8c9
>>>>>    A SID readout shell script via FEL with register access.
>>>>>  - https://31.135.195.151:20281/d/efuse/
>>>>>    A SID readout program via /dev/mem with register access by KotCzarny.
>>>>>    (with statically compiled binary)
>>>>
>>>>
>>>> Good detective work!
>>>>
>>>> I believe this would best be fixed by making u-boot use the register
>>>> access
>>>> method to get the SID on affected chips, and make sure u-boot reads the
>>>> SID at-least once.
>>>
>>>
>>> Yes.
>>>
>>> However, what I considered is that fixing this bug will change H3 devices'
>>> MAC addresses, as they are derived from SID.
>>
>>
>> I know, but I think we will just need to accept this onetime change
>> of the fixed MAC addresses to fix this bug. I don't think this is
>> a big problem since the driver for the H3 ethernet has not been
>> merged into the mainline kernel yet.
>
> Do we still need to do the CRC32 across the SID values to generate
> the MAC addresses?

Yes this does not change that a single word has not enough
randomness in it.

Regards,

Hans