[Adding Denis re: wmi discussion below...]
On Sat, Feb 26, 2022 at 03:54:42PM PST, Guenter Roeck wrote:
On 2/26/22 05:30, Zev Weiss wrote:
Hello,
This patch series augments the existing nct6775 driver with support
for the hardware's i2c interface.
Thus far the nct6775 driver has only supported the LPC interface,
which is the main interface by which the Super-I/O chip is typically
connected to the host (x86) processor.
However, these chips also provide an i2c interface, which can provide
a way for a BMC to also monitor sensor readings from them. On some
systems (such as the ASRock Rack ROMED8HM3 and X570-D4U) this may be
the only way for the BMC to monitor host CPU temperatures (e.g. to
indirectly access a TSI interface); this functionality is thus an
important component of enabling OpenBMC to support such systems.
In such an arrangement the Super-I/O chip is simultaneously controlled
by two independent processors (the host and the BMC) which typically
do not coordinate their accesses with each other. In order to avoid
conflicts between the two, the i2c driver avoids all writes to the
device, since the BMC's needs with the hardware are merely that it be
able to retrieve sensor readings. This allows the host processor to
remain ultimately in control of the chip and unaware of the BMC's use
of it at all.
The sole exception to the "no writes" rule for the i2c driver is for
the bank-select register -- while I haven't been able to find any
explicit statement in the Nuvoton datasheets guaranteeing this, all
experiments I've done have indicated that, as one might hope, the i2c
interface has its own bank-select register independent of the one used
by the LPC interface.
That will a more detailed confirmation. Please explain in detail
the experiments you have done.
Other chips (specifically those served by the it87 driver) have the
same problem, and there it was never really solved. That resulted
in random crashes. I don't want to end up in the same situation.
On an ASRock Rack ROMED8HM3, I used the following program to manually
peek and poke device registers from the host:
host# cat port.c
#include <stdlib.h>
#include <string.h>
#include <stdio.h>
#include <sys/io.h>
int main(int argc, char** argv)
{
unsigned long addr, val;
if (iopl(3)) {
perror("iopl");
exit(1);
}
if (argc == 3 && !strcmp(argv[1], "r")) {
addr = strtoul(argv[2], NULL, 0);
printf("%03lx: %02x\n", addr, inb(addr));
} else if (argc == 4 && !strcmp(argv[1], "w")) {
addr = strtoul(argv[2], NULL, 0);
val = strtoul(argv[3], NULL, 0);
printf("%03lx <- %02lx\n", addr, val);
outb(val, addr);
} else {
fprintf(stderr, "Usage: %s [rw] ADDR [VALUE]\n", argv[0]);
exit(1);
}
return 0;
}
host# cc -o port port.c
At the same time, I used i2cget & i2cset to do the same from the BMC
(the nct6779 is at address 0x2d on bus 1).
Interleaving some shell transcripts chronologically (no drivers bound on
either processor):
host# ./port w 0x295 0x4e && ./port r 0x296
295 <- 4e
296: 08
bmc# i2cget -y 1 0x2d 0x4e
0x02
bmc# i2cset -y 1 0x2d 0x4e 0x01
bmc# i2cget -y 1 0x2d 0x4e
0x01
host# ./port w 0x295 0x4e && ./port r 0x296
295 <- 4e
296: 08
So, starting from an initial state where the two sides were reporting
different values for the bank select register, a write to the bank
select register via the BMC's i2c interface is seen by a subsequent read
over that same interface, but the host still sees the same value it saw
initially.
Reversing the roles (host writing the bank-select register), the BMC is
likewise unaffected:
bmc# i2cget -y 1 0x2d 0x4e
0x01
host# ./port w 0x295 0x4e && ./port r 0x296
295 <- 4e
296: 08
host# ./port w 0x295 0x4e && ./port w 0x296 0x06
295 <- 4e
296 <- 06
host# ./port w 0x295 0x4e && ./port r 0x296
295 <- 4e
296: 06
bmc# i2cget -y 1 0x2d 0x4e
0x01
In contrast, looking at a different read/write register (the config
register at 0x40 in bank 0, for example), writes from each side are seen
by the other:
host# ./port w 0x295 0x4e && ./port w 0x296 0x00
295 <- 4e
296 <- 00
host# ./port w 0x295 0x40 && ./port r 0x296
295 <- 40
296: 03
bmc# i2cset -y 1 0x2d 0x4e 0x00
bmc# i2cget -y 1 0x2d 0x40
0x03
bmc# i2cset -y 1 0x2d 0x40 0x01
bmc# i2cget -y 1 0x2d 0x40
0x01
host# ./port w 0x295 0x40 && ./port r 0x296
295 <- 40
296: 01
host# ./port w 0x295 0x40 && ./port w 0x296 0x03
295 <- 40
296 <- 03
host# ./port w 0x295 0x40 && ./port r 0x296
295 <- 40
296: 03
bmc# i2cget -y 1 0x2d 0x40
0x03
To be cautious we could perhaps trim down the list of supported chips in
the i2c driver to include only the models that are known to pass tests
like the above; I can certainly make that change if that would be
preferred.
In terms of code structure, the approach taken in this series is to
split the LPC-specific parts of the driver out into a separate module
(called nct6775-platform), leaving the interface-independent parts in
a generic driver (called nct6775-core). The nct6775-i2c driver is
then added as an additional consumer of the nct6775-core module's
functionality.
How about wmi ? Shouldn't that be separated as well ?
I admittedly don't know much about wmi, but from what I could see in the
code it looked to me like it essentially just acts as an alternate
software/firmware path to the same underlying LPC interface, so I
figured it would make sense to leave it integrated with the rest of the
platform driver as it is currently. If there's a particular benefit to
be gained from splitting it out as well I can look into doing so, but
from my current understanding it seems like it would lead to either:
- duplication between the wmi and bare inb/outb versions of the
platform driver, or
- to eliminate the duplication, a third (intermediate) module
containing the code that was common to the two of them, leaving the
two platform driver modules as little else than the bare register I/O
routines (superio_[wmi_]inb, nct6775_[wmi_]read_value, etc.), i.e.
the differences that are already handled via the indirection of the
function pointers Denis added recently.
I suppose the hypothetical intermediate module in the second scenario
above could also be integrated back into nct6775-core, though that part
of it would then be needless dead weight for the typically more
resource-constrained systems (e.g. BMCs) that are only going to use the
i2c driver.
Anyway, none of the above possibilities seemed terribly appealing to me,
but if there's something I've misunderstood or factors I haven't taken
into account please let me know.
Thanks,
Zev
