[PATCH v2 00/21] ipmi: Allow raw access to KCS devices

Zev Weiss zweiss at equinix.com
Thu Apr 1 18:24:05 AEDT 2021


On Fri, Mar 19, 2021 at 01:19:30AM CDT, Andrew Jeffery wrote:
>Hello,
>
>This series is a bit of a mix of things, but its primary purpose is to
>expose BMC KCS IPMI devices to userspace in a way that enables userspace
>to talk to host firmware using protocols that are not IPMI.
>
>v1 can be found here:
>
>https://lore.kernel.org/openbmc/20210219142523.3464540-1-andrew@aj.id.au/
>
>Changes in v2 include:
>
>* A rebase onto v5.12-rc2
>* Incorporation of off-list feedback on SerIRQ configuration from
>  Chiawei
>* Further validation on hardware for ASPEED KCS devices 2, 3 and 4
>* Lifting the existing single-open constraint of the IPMI chardev
>* Fixes addressing Rob's feedback on the conversion of the ASPEED KCS
>  binding to dt-schema
>* Fixes addressing Rob's feedback on the new aspeed,lpc-interrupts
>  property definition for the ASPEED KCS binding
>
>A new chardev device is added whose implementation exposes the Input
>Data Register (IDR), Output Data Register (ODR) and Status Register
>(STR) via read() and write(), and implements poll() for event
>monitoring.
>
>The existing /dev/ipmi-kcs* chardev interface exposes the KCS devices in
>a way which encoded the IPMI protocol in its behaviour. However, as
>LPC[0] KCS devices give us bi-directional interrupts between the host
>and a BMC with both a data and status byte, they are useful for purposes
>beyond IPMI.
>
>As a concrete example, libmctp[1] implements a vendor-defined MCTP[2]
>binding using a combination of LPC Firmware cycles for bulk data
>transfer and a KCS device via LPC IO cycles for out-of-band protocol
>control messages[3]. This gives a throughput improvement over the
>standard KCS binding[4] while continuing to exploit the ease of setup of
>the LPC bus for early boot firmware on the host processor.
>
>The series takes a bit of a winding path to achieve its aim:
>
>1. It begins with patches 1-5 put together by Chia-Wei, which I've
>rebased on v5.12-rc2. These fix the ASPEED LPC bindings and other
>non-KCS LPC-related ASPEED device drivers in a way that enables the
>SerIRQ patches at the end of the series. With Joel's review I'm hoping
>these 5 can go through the aspeed tree, and that the rest can go through
>the IPMI tree.
>
>2. Next, patches 6-13 fairly heavily refactor the KCS support in the
>IPMI part of the tree, re-architecting things such that it's possible to
>support multiple chardev implementations sitting on top of the ASPEED
>and Nuvoton device drivers. However, the KCS code didn't really have
>great separation of concerns as it stood, so even if we disregard the
>multiple-chardev support I think the cleanups are worthwhile.
>
>3. Patch 14 adds some interrupt management capabilities to the KCS
>device drivers in preparation for patch 16, which introduces the new
>"raw" KCS device interface. I'm not stoked about the device name/path,
>so if people are looking to bikeshed something then feel free to lay
>into that.
>
>4. The remaining patches switch the ASPEED KCS devicetree binding to
>dt-schema, add a new interrupt property to describe the SerIRQ behaviour
>of the device and finally clean up Serial IRQ support in the ASPEED KCS
>driver.
>
>Rob: The dt-binding patches still come before the relevant driver
>changes, I tried to keep the two close together in the series, hence the
>bindings changes not being patches 1 and 2.
>
>I've exercised the series under qemu with the rainier-bmc machine plus
>additional patches for KCS support[5]. I've also substituted this series in
>place of a hacky out-of-tree driver that we've been using for the
>libmctp stack and successfully booted the host processor under our
>internal full-platform simulation tools for a Rainier system.
>
>Note that this work touches the Nuvoton driver as well as ASPEED's, but
>I don't have the capability to test those changes or the IPMI chardev
>path. Tested-by tags would be much appreciated if you can exercise one
>or both.
>
>Please review!
>
>Andrew
>

After rebasing the series onto the OpenBMC dev-5.10 kernel (with only a
tiny conflict for the addition of the ast2600 entry in
ast_kcs_bmc_match) and enabling CONFIG_IPMI_KCS_BMC_CDEV_IPMI, my
e3c246d4i system booted healthily and handled some basic ipmitool
operations as expected.

Tested-by: Zev Weiss <zweiss at equinix.com>


Zev



More information about the Linux-aspeed mailing list