[lm-sensors] AMD Phenom temperature driver
r.marek at assembler.cz
Fri Jul 18 23:08:13 CEST 2008
I think I put all parties up in CC. Thanks to Andreas, I have received missing
information about the TControlMax temperature. It means we can get further to a
I did some experimental driver which works fine (attached as k8temp.c) Kryzstof
Went even further and made changes to original k8temp driver.
I will go for vacation soonish (middle of next week) and I'm very busy, but at
least we can start. I will write a short list of things which needs to be
discussed in the meanwhile:
1) driver name
AMD proposes the change to amdtemp (from k8temp) I think we can merge the
drivers together, it should be still simple.
Jean, please do you remember how autoloading/driver name change is handled in
the kernel? I have some feeling that it should be simple to handle. Please confirm.
The name change can be done if we know that old name will work, so we dont break
too much things around.
2) Maximum temperature.
The new Phenom temperature is non-physical it means we must know the limit so we
know when we cross it.
Therefore our new driver must export either temp1_max or temp1_crit. The
TcontrolMax is for all fam 10h CPUs 70 degrees.
The Intel driver has temp1_max as maximum temperature from which all fans must
start cooling. The temp1_crit is when the CPU will fail.
I think the thremtrip_l is asserted for much higher temperatures, I think we can
go for temp1_max.
We can put the TcontrolMax into a variable, because for future revisions we may
need cpuid to put there right value.
3) Errata #319
I think all CPU revisions still may have problem with the temp sensor inside the
CPU (not the diode interface which is analog and works fine).
Please can someone test mine k8temp.c attached, and report back the values you
got? If we detect a CPU with the errata we should print a warning to syslog.
4) Kryzstof driver.
Kryzstof already posted to me some changes for driver merger. Please can you
repost them in form of patch here to mailing list? Please avoid white space
changes. AMD suggest to call K10 fam10 because it is in fact family 10h
and not K10. From CPUID perspective it is more model then family to make it worse ;)
5) merged driver architecture
For the different code conditions I would propose some amdfam variable with
model code (0xe) 0xf 0x10 and some
if data->amdfam = FAM_E ...
Now the neat trick how the PCI driver teach about CPUID. There is a register
with CPUID inside function 3 (offset FC), we can read the CPUID from there, and
get the model for 10h and fh model eh and older do not have this register.
We cannot use cpuid_eax, because we may have two processors with different
erratas and we need to know which is which PCI device. The old K8temp driver
simply ignored this - the module code may be executed on different CPU other
then PCI device we are operating on (multi CPU system not multicore).
Therefore we can do:
1) for PCI ID of "old" family put amdfam with 0xf
2) for new fill 10h and save CPUID from the register FC
3) check the errata, print warning
4) future versions may need the cpuid info for different TcontolMax
5) for fam10 create new get temps funs, new formula
6) enhancements for older k8temp
this should be done in separate patch.
6a) print warning about sensors inaccurancy for all model 0xf
6b) extend scale to .00 .25 .50 .75
6c) the automatic core detecting may detect single core cpu which were in fact
dualcore but with disabled core. Typically the second core has -49. For some
reason it gets printed on mine sempron G2 fam 0fh
Of course this are just mine ideas, feel free to come with even better ideas ;)
The code is very welcomed too, I can try to at least review it when I will have
some free time. I can code this too but free time window is two/three weeks far.
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 4652 bytes
Desc: not available
More information about the lm-sensors