[RFC Patch v2 0/3] I2C statistics as sysfs attributes
Sui Chen
suichen at google.com
Wed Dec 8 03:00:00 AEDT 2021
On 12/3/21 8:37 AM, Wolfram Sang wrote:
> On Thu, Dec 02, 2021 at 06:37:25PM -0800, Sui Chen wrote:
>> Add I2C statistics such as Bus Error counts and NACK counts as sysfs
>> attributes so they don't need to live in debugfs.
>
> What has changed since v1?
>
> From a glimpse, none of my questions to v1 have been answered or
> addressed?
>
Apologies for missing the first email, I double-checked my mailbox
and saw it. Difference in v2 is not much, with just the
tx_complete_cnt added.
Before answering the questions please let me give some
background info:
The motivation starts with monitoring the operation of BMCs and then
applying what we learned to monitoring BMC health at scale. The BMCs
we're interested in run the OpenBMC distribution, which is characterized
by its extensive use of DBus APIs in the sensor stack and everywhere
else in the system. We had been working on some tools centered around
DBus and had discovered issues on certain systems by looking at related
metrics.
A snapshot of the data we look into on a running OpenBMC system may
look like the following (a screenshot from one of the tools we're
working on):
+-------dbus-top v0.xx-----------------------------------------+
|Message Type | msg/s |Method Call Time (us) Histogram|
|Method Call 282.91 | 1387-: |
|Method Return 282.91 | 227-::. |
|Signal 56.98 | 37-:::::: :: |
|Error 0.00 | 6-::::::: ..::....::..::: |
|Total 622.79 |1%-99% 57.00 23899.00 |
+-------------------------------------------------------+------+
|History (Total msg/s) |
|- -700 |
|- :-525 |
|- . : :-350 |
|- . :: : : :-175 |
|- .:::::::::::::::::::::-0 |
+-------------------------------------------------------+---+
| Columns 1-3 of 6 200 sensors |
| mobo_pch....... Core_xx_CPUX... Core_xx_CPUX > |
| mobo_pch....... Core_xx_CPUX... Core_xx_CPUX > |
| mobo_pch....... Core_xx_CPUX Core_XX_CPUX > |
| mobo_pch....... Core_xx_CPUX Core_XX_CPUX .. > |
| cpuX_tem....... Core_xx_CPUX Core_XX_CPUX > |
*************************************************************
* Destination Sender I2C Tx/s Sender CMD *
* systemd n/a /usr/bin/cpusensor *
* :1.78 n/a (unknown) *
* systemd 172.44 (unknown) *
* systemd n/a /usr/bin/externalsenso*
* systemd 658.28 /usr/bin/fansensor *
* systemd 167.94 /usr/bin/hwmontempsens*
* systemd 10345.07 /usr/bin/psusensor *
* :1.15 n/a ipmid *
* :1.59 n/a ipmid *
* systemd n/a ipmid *
* :1.67 n/a /usr/libexec/kcsbridg *
* :1.26166 n/a (unknown) *
* systemd n/a (unknown) *
*************************************************************
As one can see from the figure above, there are a few hundred DBus
messages flowing through different DBus peers on the system at
any given moment, with a few daemons reading sensor data via I2C and
publishing them on DBus. We observe I2C to be a large chunk of the work
the BMC is constantly doing, and actually it has been the source of a
few problems we had seen and fixed.
By the time we started this tool, we have had enough experience
and confidence to say that I2C would be interesting for both development
and monitoring at-scale and is worth investing in.
The requirement for a stable API/ABI is mostly relevant to the
"at-scale" monitoring part. Here are the answers to the questions:
> 1) Why do you need this information? I don't think values like NACK
> count describe the health of a system. NACKs are perfectly OK in
> regular I2C communication. So, for now, I think debugfs is the better
> place. An exception might be the bus speed. Some people already had an
> interest in this.
- Because we're targeting at-scale monitoring of fleets of machines
in large numbers with identical hardware configuration, the
I2C counters (including NACK) can be used for apples-to-apples
comparison, for detecting anomaly, etc., while this may not be
applicable to a single machine.
> 2) Even when this information is kept in debugfs, we can still add
> some core helper to have a standardized structure for the created
> files. This is independent from sysfs. I don't think I want this a
> standardized ABI, currently. Unless you explain good reasons to me
- The monitoring is done in a production environment, so it is not a
"debug" usage. Android 11 removed support for debugfs [1] due to
unstable and undocumented API, code quality and vulnerability issues.
We would like to follow a similar rationale for the I2C counters.
- debugfs paths may be vendor-specific, so a monitoring framework would
need different code paths for different vendors without a stable
method for obtaining those counters. This may be resolved by the core
helpers, but our intention is to not use debugfs if possible.
- A monitoring system includes not only lower levels such as the kernel,
but also higher-level standards such as Redfish, and also the OpenBMC-
wide DBus interfaces and C++ bindings. Redfish has already committed
to creating a new Schema that contains similar metrics; the Schemas
is expected to be stable for a considerable amount of time. OpenBMC is
considering adding its system-wide DBus interfaces for the I2C
definitions to reflect the relatively stable Redfish bindings too.
Therefore we kindly hope that the kernel interface for I2Cs can
be similarly stable just like the other two.
For the errors found in the testing, and other potential errors in the
code: we will look into them and correct them.
Hope this may be useful,
Thanks
Sui
[1] https://source.android.com/setup/start/android-11-release#debugfs
More information about the openbmc
mailing list