Author: Patrick Venture <venture!>
Primary assignee: Patrick Venture
Created: 2018-10-18
The BMC needs a mechanism for receiving a new firmware image from the host through a variety of mechanisms. This can best be served with one protocol into which multiple approaches can be routed.
BMC hardware provides at a minimum some interface for sending and receiving IPMI messages. This hardware may also provide regions that can be memory mapped for higher speed communication between the BMC and the host. Certain infrastructures do not provide network access to the BMC, therefore it is required to provide an update mechanism that can be done in-band between the host and the BMC.
In-band here refers to a communications channel that is directly connected between the host and BMC.
Please read the IPMI BLOB protocol design as primer here.
The following statements are reflective of the initial requirements.
Any update mechanism must provide support for UBI tarballs and legacy flash images. Leveraging the BLOB protocol allows a system to provide support for any image type simply by implementing a mechanism for handling it.
Any update mechanism must allow for triggering an image verification step before the image is used.
Any update mechanism must allow implementing the data staging via different in-band mechanisms.
Any update mechanism must provide a handshake or equivalent protocol for coordinating the data transfer. For instance, whether the BMC should enable the P2A bridge and what region to use or whether to turn on the LPC memory map bridge.
Any update mechanism must attempt to maintain security, insomuch as not leaving a memory region open by default. For example, before starting the verification step, access to the staged firmware image must not be still accessible from the host.
OpenBMC supports a BLOB protocol that provides primitives. These primitives allow a variety of different "handlers" to exist that implement those primitives for specific "blobs." A blob in this context is a file path that is strictly unique.
Sending the firmware image over the BLOB protocol will be done via routing the phosphor-ipmi-flash design through a BLOB handler. This is meant to supplant phosphor-ipmi-flash's current approach to centralize on one flexible handler.
The BLOB protocol allows a handler to specify a list of blob ids. This list will be leveraged to specify whether the platform supports either the legacy or the UBI mechanism, or both. The flags provided to the open command identify the mechanism selected by the client-side. The stat command will return the list of supported mechanisms for the blob.
The blob ids for the mechanisms will be as follows:
| Flash Blob Id | Type |
|---|---|
| /flash/image | Legacy |
| /flash/tarball | UBI |
The flash handler will determine what commands it should expect to receive and responses it will return given the blob opened, based on the flags provided to open.
The following blob ids are defined for storing the hash for the image:
| Hash Blob | Id Mechanism |
|---|---|
| /flash/hash | Legacy |
The flash handler will only allow one open file at a time, such that if the host attempts to send a firmware image down over IPMI BlockTransfer, it won't allow the host to start a PCI send until the BlockTransfer file is closed.
There is only one hash "file" mechanism. The exact hash used will only be important to your verification service. The value provided will be written to a known place.
Similarly to the OEM IPMI Flash protocol, the flash image will be staged in a compile-time configured location.
Other mechanisms can readily be added by adding more blob_ids or flags to the handler.
The update mechanism will expect a specific sequence of commands depending on the transport mechanism selected. Some mechanisms require a handshake.
If an image update process is started but goes stale there are multiple mechanisms in place to ensure cleanup. If a session is left open after the blob timeout period it'll be closed. Because expiration is not the same action as closing, the cache will be flushed and any staged pieces deleted.
The image itself, in legacy mode will be placed and named in such a way that it will disappear if the BMC reboots. In the UBI case, the file will be stored in /tmp and deleted accordingly.
At any point during the upload process, one can abort by closing the open blobs and deleting them by name.
The update mechanism will implement the Blob primitives as follows.
The blob open primitive allows supplying blob specific flags. These flags are used for specifying the transport mechanism. To obtain the list of supported mechanisms on a platform, see the Stat command below.
enum OpenFlags
{
read = (1 << 0),
write = (1 << 1),
};
enum FirmwareUpdateFlags
{
bt = (1 << 8), /* Expect to send contents over IPMI BlockTransfer. */
p2c = (1 << 9), /* Expect to send contents over P2A bridge. */
lpc = (1 << 10), /* Expect to send contents over LPC bridge. */
};
An open request must specify that it is opening for writing and one transport mechanism, otherwise it is rejected. If the request is also set for reading, this is not rejected but currently provides no additional value.
Once opened a new file will appear in the blob_id list (for both the image and hash) indicating they are in progress. The name will be flash/{item}/0 and flash/hash/0 which has no meaning beyond representing the current update in progress. Closing the file does not delete the staged images. Only delete will.
This will initially not perform any function and will return success with 0 bytes.
The write command's contents will depend on the transport mechanism. This command must not return until it has copied the data out of the mapped region into either a staging buffer or written down to a staging file. How the command reads from the mapped region is beyond the scope of this design.
The data section of the payload is only data.
The data section of the payload is the following structure:
struct ExtChunkHdr
{
uint32_t length; /* Length of the data queued (little endian). */
};
The data section of the payload is the following structure:
struct ExtChunkHdr
{
uint32_t length; /* Length of the data queued (little endian). */
};
If this command is called on the session of the firmware image itself, nothing will happen at present. It will return a no-op success.
If this command is called on the session for the hash image, it'll trigger a systemd service verify_image.service to attempt to verify the image. Before doing this, if the transport mechanism is not IPMI BT, it'll shut down the mechanism used for transport preventing the host from updating anything.
When this is started, only the BmcBlobSessionStat command will respond. Details on that response are below under BmcBlobSessionStat.
Close must be called on the firmware image before triggering verification via commit. Once the verification is complete, you can then close the hash file.
If the verify_image.service returned success, closing the hash file will have a specific behavior depending on the update. If it's UBI, it'll perform the install. If it's legacy, it'll do nothing. The verify_image service in the legacy case is responsible for placing the file in the correct staging position. A BMC warm reset command will initiate the firmware update process.
If the image verification fails, it will automatically delete any files associated with the update.
Note: During development testing, a developer will want to upload files that are not signed. Therefore, an additional bit will be added to the flags to change this behavior.
Aborts any update that's in progress:
In the event the update is already in progress, such as the tarball mechanism is used and in the middle of updating the files, it cannot be aborted.
Blob stat on a blob_id (not SessionStat) will return the capabilities of the blob_id handler.
struct BmcBlobStatRx {
uint16_t crc16;
/* This will have the bits set from the FirmwareUpdateFlags enum. */
uint16_t blob_state;
uint32_t size; /* 0 - it's set to zero when there's no session */
uint8_t metadata_len; /* 0 */
};
If called pre-commit, it'll return the following information:
struct BmcBlobStatRx {
uint16_t crc16;
uint16_t blob_state; /* OpenFlags::write | (one of the interfaces) */
uint32_t size; /* Size in bytes so far written */
uint8_t metadata_len; /* 0. */
};
If called post-commit on the hash file session, it'll return:
struct BmcBlobStatRx {
uint16_t crc16;
uint16_t blob_state; /* OPEN_W | (one of the interfaces) */
uint32_t size; /* Size in bytes so far written */
uint8_t metadata_len; /* 1. */
uint8_t verify_response; /* one byte from the below enum */
};
enum VerifyCheckResponses
{
VerifyRunning = 0x00,
VerifySuccess = 0x01,
VerifyFailed = 0x02,
VerifyOther = 0x03,
};
The write metadata command is meant to allow the host to provide specific configuration data to the BMC for the in-band update. Currently that is only aimed at LPC which needs to be told the memory address so it can configure the window.
The write meta command's blob will be this structure:
struct LpcRegion
{
uint32_t address; /* Host LPC address where the chunk is to be mapped. */
uint32_t length; /* Size of the chunk to be mapped. */
};
There is currently another implementation in-use by Google that leverages the same mechanisms, however, it's not as flexible because every command is a custom piece. Mapping it into blobs primitives allows for easier future modification while maintaining backwards compatibility (without simply adding a separate OEM library to handle a new process, etc).
This impacts security because it can leverage the memory mapped windows. There is not an expected performance impact, as the blob handler existing only generates a couple extra entries during the blob enumerate command's response.
Where possible (nearly everywhere), mockable interfaces will be used such that the entire process has individual unit-tests that verify flags are checked, as well as states and sequences.
A required functional test is one whereby an image is sent down to the BMC, however the signature is invalid for that image. The expected result is that the verification step will return failure and the files will be deleted from the BMC without user intervention.
A required functional test is one whereby an image is sent down to the BMC with a valid signature. The expected result is that the verification step will return success.