design: Add bios-bmc-smm-error-logging.md
This design doc is for providing a shared buffer between the BMC and the
BIOS that allows the BIOS to send a very fast message without requiring
an ack from the BMC within SMM.
Signed-off-by: Brandon Kim <brandonkim@google.com>
Change-Id: Id76029a9f5c869a5e2a77fd53843c926b438b67d
diff --git a/designs/bios-bmc-smm-error-logging.md b/designs/bios-bmc-smm-error-logging.md
new file mode 100644
index 0000000..87fe24c
--- /dev/null
+++ b/designs/bios-bmc-smm-error-logging.md
@@ -0,0 +1,154 @@
+# BIOS->BMC SMM Error Logging Queue Daemon
+
+Author:
+* Brandon Kim / brandonkim@google.com / @brandonk
+
+Other contributors:
+* Marco Cruz-Heredia / mcruzheredia@google.com
+
+Created:
+ Mar 15, 2022
+
+## Problem Description
+
+We've identified use cases where the BIOS will go into System Management Mode
+(SMM) to provide error logs to the BMC, requiring messages to be sent as quickly
+as possible without a handshake / ack back from the BMC due to the time
+constraint that it's under. The goal of this daemon we are proposing is to
+implement a circular buffer over a shared BIOS->BMC buffer that the BIOS can
+fire-and-forget.
+
+## Background and References
+
+There are various ways of communicating between the BMC and the BIOS, but there
+are only a few that don't require a handshake and lets the data persist in
+shared memory. These will be listed in the "Alternatives Considered" section.
+
+Different BMC vendors support different methods such as Shared Memory (SHM, via
+LPC / eSPI) and P2A or PCI Mailbox, but the existing daemon that utilizes them
+do it over IPMI blob to communicate where and how much data has been
+transferred (see [phosphor-ipmi-flash](https://github.com/openbmc/phosphor-ipmi-flash)
+and [libmctp/astlpc](https://github.com/openbmc/libmctp/blob/master/docs/bindings/vendor-ibm-astlpc.md))
+
+## Requirements
+
+The fundamental requirements for this daemon are listed as follows:
+
+1. The BMC shall initialize the shared buffer in a way that the BIOS can
+ recognize when it can write to the buffer
+2. After initialization, the BIOS shall not have to wait for an ack back from
+ the BMC before any writes to the shared buffer (**no synchronization**)
+3. The BIOS shall be the main writer to the shared buffer, with the BMC mainly
+ reading the payloads, only writing to the buffer to update the header
+4. The BMC shall read new payloads from the shared buffer for further processing
+5. The BIOS must be able to write a payload (~1KB) to the buffer within 50µs
+
+The shared buffer will be as big as the protocol allows for a given BMC platform
+(for Nuvoton's PCI Mailbox for NPCM 7xx as an example, 16KB) and each of the
+payloads is estimated to be less than 1KB.
+
+This daemon assumes that no other traffic will communicate through the given
+protocol. The circular buffer and its header will provide some protection
+against corruption, but it should not be relied upon.
+
+## Proposed Design
+
+The implementation of interfacing with the shared buffer will very closely
+follow [phosphor-ipmi-flash](https://github.com/openbmc/phosphor-ipmi-flash). In
+the future, it may be wise to extract out the PCI Mailbox, P2A and LPC as
+separate libraries shared between `phosphor-ipmi-flash` and this daemon to
+reduce duplication of code.
+
+Taken from Marco's (mcruzheredia@google.com) internal design document for the
+circular buffer, the data structure of its header will look like the following:
+
+| Name | Size | Offset | Written by | Description |
+| ---------- | ----------- | ----------- | ------------ | ----------------- |
+| BMC Interface Version | 4 bytes | 0x0 | BMC at init | Allows the BIOS to determine if it is compatible with the BMC |
+| BIOS Interface Version | 4 bytes | 0x4 | BIOS at init | Allows the BMC to determine if it is compatible with the BIOS |
+| Magic Number | 16 bytes | 0x8 | BMC at init | Magic number to set the state of the queue as described below. Written by BMC once the memory region is ready to be used. Must be checked by BIOS before logging. BMC can change this number when it suspects data corruption to prevent BIOS from writing anything during reinitialization |
+| Queue size | 4 bytes | 0x18 | BMC at init | Indicates the size of the region allocated for the circular queue. Written by BMC on init only, should not change during runtime |
+| Uncorrectable Error region size | 2 bytes | 0x1c | BMC at init | Indicates the size of the region reserved for Uncorrectable Error (UE) logs. Written by BMC on init only, should not change during runtime |
+| BMC flags | 4 bytes | 0x1e | BMC | <ul><li>BIT0 - BMC UE reserved region “switch”<ul><li>Toggled when BMC reads a UE from the reserved region.</li></ul><li>BIT1 - Overflow<ul><li>Lets BIOS know BMC has seen the overflow incident</li><li>Toggled when BMC acks the overflow incident</li></ul><li>BIT2 - BMC_READY<ul><li>BMC sets this bit once it has received any initialization information it needs to get from the BIOS before it’s ready to receive logs.</li></ul> |
+| BMC read pointer | 4 bytes | 0x22 | BMC | Used to allow the BIOS to detect when the BMC was unable to read the previous error logs in time to prevent the circular buffer from overflowing. |
+| Padding | 2 bytes | 0x26 | Reserved | Padding for 8 byte alignment |
+| BIOS flags | 4 bytes | 0x28 | BIOS | <ul><li>BIT0 - BIOS UE reserved region “switch”<ul><li> Toggled when BIOS writes a UE to the reserved region.</li></ul><li>BIT1 - Overflow<ul><li>Lets the BMC know that it missed an error log</li><li>Toggled when BIOS sees overflow and not already overflowed</li></ul><li>BIT2 - Incomplete Initialization<ul><li>Set when BIOS has attempted to initialize but did not see BMC ack back with `BMC_READY` bit in BMC flags</li></ul>|
+| BIOS write pointer | 4 bytes | 0x2c | BIOS | Indicates where the next log will be written by BIOS. Used to tell BMC when it should read a new log |
+| Uncorrectable Error reserved region | TBD1 | 0x30 | BIOS | Reserved region only for UE logs. This region is only used if the rest of the buffer is going to overflow and there is no unread UE log already in the region. |
+| Error Logs from BIOS | Size of the Buffer - 0x30 - TBD1 | 0x30 + TBD1 | BIOS | Logs vary by type, so each log will self-describe with a header. This region will fill up the rest of the buffer |
+
+### Initialization
+
+This daemon will first initialize the shared buffer by writing zero to the whole
+buffer, then initializing the header's `BMC at init` fields before writing the
+`Magic Number`. Once the `Magic Number` is written to, the BIOS will assume that
+the shared buffer has been properly initialized, and will be able to start
+writing entries to it.
+
+If there are any further initialization between the BIOS and the BMC required,
+the BMC needs to set the `BMC_READY` bit in the BMC flags once the
+initialization completes. If the BIOS does not see the flag being set, the BIOS
+shall set the `Incomplete Initialization` flag to notify the BMC to reinitialize
+the buffer.
+
+### Reading and Processing
+
+This daemon will poll the buffer at a set interval (the exact number will be
+configurable as the processing time and performance of different platforms may
+require different polling rate) and once a new payload is detected, the payload
+will be processed by a library that can also be chosen and configured at
+compile-time.
+
+Note that the Uncorrectable Error logs have a reserved region as they contain
+critical information that we don't want to lose, and should be prioritized over
+normal error logs. This reserved region will be used to log a UE log only if an
+overflow of the normal error log queue is imminent and the BMC has acked that
+any preexisting UE log in this region has already been read using Bit0 of the
+`BMC flag`.
+
+An example of a processing library (and something we would like to push in our
+initial version of this daemon) would be an RDE decoder for processing a subset
+of Redfish Device Enablement (RDE) commands, and decoding its attached Binary
+Encoded JSON (BEJ) payloads.
+
+## Alternatives Considered
+
+* IPMI was considered, did not meet our speed requirement of writing 1KB entry
+ in about 50 microseconds.
+ * For reference, initial PCI Mailbox performance measurement showed 1KB entry
+ write took roughly 10 microseconds.
+* LPC / eSPI was also considered but our BMC's SHM buffer was limited to 4KB
+ which was not enough for our use case.
+* `libmctp` and MCTP PCIe VDM were considered.
+ * `libmctp`'s current implementation relies on LPC as the transport binding
+ and IPMI KCS for synchronization. LPC as discussed, does not fit our current
+ need and synchronization does not work.
+ * We may use MCTP PCIe VDM on our future platforms once we have more resources
+ with expertise both from the BMC and the BIOS side (which we currently lack)
+ for our current project timeline.
+
+## Impacts
+
+Reading from the buffer and processing it may hinder performance of the BMC,
+especially if the polling rate is set too high.
+
+### Organizational
+
+This design will require 2 repositories:
+- bios-bmc-smm-error-logger
+ - This repository will implement the daemon described in this document
+ - Proposed maintainer: wltu@google.com , brandonkim@google.com
+- libbej
+ - This repository will follow the PLDM RDE specification as much as possible
+ for RDE BEJ decoding (initially, encoding may come in the future) and will
+ host a library written in C
+ - Proposed maintainer: wltu@google.com , brandonkim@google.com
+
+## Testing
+
+Unit tests will cover each parts of the daemon, mainly:
+
+* Initialization
+* Circular buffer processing
+* Decoding / Processing library
+