Hi,
On 27-03-15 05:30, Rob Herring wrote:
On Thu, Mar 26, 2015 at 4:11 AM, Paul Kocialkowski <contact@xxxxxxxx> wrote:
Le jeudi 26 mars 2015 à 09:53 +0100, Hans de Goede a écrit :
Hi,
On 25-03-15 23:35, Paul Kocialkowski wrote:
Le mardi 24 mars 2015 à 09:01 +0100, Hans de Goede a écrit :
Hi,
On 24-03-15 00:12, Rob Herring wrote:
On Mon, Mar 23, 2015 at 6:30 AM, Hans de Goede <hdegoede@xxxxxxxxxx> wrote:
Hi,
On 22-03-15 22:01, Rob Herring wrote:
<snip>
There is already "serial-number" (a string) which exists for
OpenFirmware. Also, "copyright" corresponds to vendor/manufacturer
string. Both of these are supported by lshw already.
Ok, so if I understand you correctly then you're saying that we
should set a "serial-number" string property at the dt root level
and that this may contain pretty much anything, e.g. in the
sunxi case the full 128 bit SID in hex.
Right.
Is the use of the "serial-number" string property already documented
somewhere? If not I'll submit a kernel patch to document it.
Not that I'm aware of. It is something that predates our documentation
requirements. It could be in OpenFirmware specs. Documenting it in the
DT bindings does not hurt.
Ok.
And for older kernels we should not set any serial atag (u-boot
always sets it, so this leaves it at 0) and old kernel users are
out of luck wrt getting to the serial ?
If there is sufficient reason to support this on old kernels you could.
One problem with supporting this for older kernels is that if a non 0
serial gets shown in /proc/cpuinfo with older atag booted kernels, we
should really show the same number in /proc/cpuinfo which means adding
code to the kernel to get the devicetree "serial-number" string property
and somehow put that into the 64 bits which we have in /proc/cpuinfo,
but given that the "serial-number" string could be hex or decimal or
what ever and > 64 bits that will likely require a platform specific
solution. All doable, but the question then becomes is this worth the
effort ?
After investigating a bit more, I found out that the USB serial number
is expected to be a string of 32 bytes, so a 128 bit numeric serial
doesn't fit (it takes 32 bytes for the hex representation of 128 bits,
so there is no room left for the terminating null byte), hence it makes
sense to keep a 64 bit limitation for the serial number, if users are
going to rely on it as USB serial string. Moreover, it seems that
Android devices are mostly used 64 bit numbers for serial numbers/
I was initially going to suggest that we set it in stone that serial
must be 64 numeric bits (as it was in the ATAGs days) and that
bootloaders would pass it that way to the kernel through device tree
(with two 32 bits numeric integers), but Hans talked me out of it.
I just want to expose the situation (especially the USB and Android
thing) here to double-check that everyone still is convinced that a
string approach in device tree is best (which is fine with me).
There are already existing users of the serial-number property in devicetree,
and these already use a free-format string, so AFAICT we have no choice
but to do the same as the existing users.
But Rob is the expert here, so lets see what Rob has to say.
This way, users that still want to use the serial passed through device
tree as a USB serial number will have to use a string of 32 bits,
including the null terminating byte (which is what I'll suggest for
sunxi by using only 64 bits for the serial number).
Also, I suggest that we show that serial-number string as-is in cpuinfo
as well
We cannot do that because we must guarantee that the serial shown
in cpu info is a 64 bits / 16 hex values (0 padded) number, anything
else would break the kernel <-> userspace API and potentially break
userspace apps. So we must do the devicetree -> serialnumber low/high
-> /proc/cpinfo version to guarantee that this format does not change.
And as discussed before if you want a non 0 serial in cpuinfo then
the devicetree -> serialnumber low/high should be done in sunxi
specific kernel code, as on sunxi we will know that the string in
devicetree will be a hex value, but we've no such guarantee for
other platforms, so we cannot simply have a generic function to
populate erialnumber low/high from the devicetree serial-number
string.
> and instead make a string out of the serial ATAG in the kernel
prior to showing it in cpuinfo (as opposed to translating the string
coming from device tree to a numeric value that cpuinfo will end up
showing as a string at the end of the day). Thus, the serial number
coming from device tree will still be shown in cpuinfo as well and no
ABI gets broken.
You're forgetting the userspace <-> kernel ABI here, the serial line
in /proc/cpuinfo is not a free form string it is a 64 bit int shown
as 0 padded hex, and we cannot change that as changing that would be
an ABI break.
IMHO this really is all about interpretation. If you consider that the
serial is already a *string* and not a hex-representation of a number
(which it is when using ATAGs, but has no reason to be in general), then
my suggestion will introduce no ABI break.
Generally speaking, I found no documentation that indicates that the
serial has to be in that format. It just happens to be the case when
using ATAGs.
Also, I found an email from Rob suggesting he would be fine with wiring
the dts serial-number string to cpuinfo:
http://lkml.iu.edu/hypermail/linux/kernel/1412.0/02975.html
I think it's the most flexible solution and we can think of it as an
extension of the current scheme: the serial string will no longer be
limited to a hex representation of a number but can become any string.
Now I would appreciate it if Rob could weigh-in and state whether he
changed his mind on this or not.
The only ABI is on platforms that used ATAGS as any DT only platform
has so far had a serial# in cpuinfo of 16 0's. With a variable length
string we can still have a fixed 16 char hex serial number that is
compatible with ATAGS. I can't imagine that we have userspace that
cares about the length and yet doesn't care the value is always 0's
since converting to DT. As long as we keep 16 0's as the default I
don't see an issue with allowing other lengths.
There are 2 problems with playing around with the serial line in
/proc/cpuinfo:
1) Length, currently it is always 16 chars, making it either shorter
or longer may break some userspace parsing code for it.
2) Contents, currently it is always in hex, so we cannot simply
copy a string from devicetree in there without verifying that that
string contains hex.
Now if we were to add code to:
1) Verify that the serial from devicetree is hex
2) Truncate it to 16 chars max
That would work, the easiest way to get there is to simply store
the serial from devicetree in the existing serial low and high
32 bit ints.
It is only an ABI if someone notices and if we find an ABI issue later
we can always fix it.
Erm, no it is an ABI once it has shipped, and we cannot always fix it later
because in the mean time other apps may have started depending on the new
ABI and then we've a big problem.
There are plenty of examples of "innocent" changes to files in /proc
breaking stuff, so we really really should be careful here.
Also I don't see why this is that important at all, we can limit
things to not break the ABI and apps who want the full serial in
the original string format can always get it from
/sys/firmware/devicetree/base/serial-number
Regards,
Hans
Rob
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
