SMM based flash storage driver

This documents the API exposed by the x86 system management based storage driver. The current protocol is the former SMMSTORE version 2 API; the legacy version 1 key/value protocol has been removed.

SMMSTORE

SMMSTORE is a SMM mediated driver to read from, write to and erase a predefined region in flash. It can be enabled by setting CONFIG_SMMSTORE=y in menuconfig.

This can be used by the OS or the payload to implement persistent storage to hold for instance configuration data, without needing to implement a (platform specific) storage driver in the payload itself.

Storage size and alignment

SMMSTORE uses a configurable logical block size (CONFIG_SMMSTORE_BLOCK_SIZE, default 64 KiB). Platforms may select a smaller size such as 4 KiB when that matches flash erase geometry. Not having to perform read-modify-write operations is desired, as it reduces complexity and potential for bugs.

This can be used by a FTW (FaultTolerantWrite) implementation that uses at least two regions in an A/B update scheme. The FTW implementation in edk2 uses three different regions in the store:

  • The variable store

  • The FTW spare block

  • The FTW working block

All regions must be block-aligned, and the FTW spare size must be larger than that of the variable store. FTW working block can be much smaller. With 64 KiB as block size, the minimum size of the FTW-enabled store is:

  • The variable store: 1 block = 64 KiB

  • The FTW spare block: 2 blocks = 2 * 64 KiB

  • The FTW working block: 1 block = 64 KiB

Therefore, the minimum size for edk2 FTW is 4 blocks, or 256 KiB using 64 KiB blocks.

API

The API provides read and write access to an unformatted block storage.

Storage region

By default SMMSTORE will operate on a separate FMAP region called SMMSTORE. The default generated FMAP will include such a region. On systems with a locked FMAP, e.g. in an existing vboot setup with a locked RO region, the option exists to add a cbfsfile called smm_store in the RW_LEGACY (if CHROMEOS) or in the COREBOOT FMAP regions. It is recommended for new builds using a handcrafted FMD that intend to make use of SMMSTORE to include a sufficiently large SMMSTORE FMAP region. The region must be aligned to the logical block size and should be aligned to the largest flash erase block. The FMAP region must be at least 64 KiB even when the logical block size is smaller.

When a default generated FMAP is used, the size of the FMAP region is equal to CONFIG_SMMSTORE_SIZE. UEFI payloads expect at least 64 KiB. To support a fault tolerant write mechanism, at least a multiple of this size is recommended.

Communication buffer

To prevent malicious ring0 code to access arbitrary memory locations, SMMSTORE uses a communication buffer in CBMEM/HOB for all transfers. This buffer is sized to match the logical block size and must be installed before calling any of the SMMSTORE read or write operations. Usually, coreboot will install this buffer to transfer data between ring0 and the SMM handler.

In order to get the communication buffer address, the payload or OS has to read the coreboot table with tag 0x0039, containing:

struct lb_smmstorev2 {
	uint32_t tag;
	uint32_t size;
	uint32_t num_blocks;		/* Number of writable blocks in SMM */
	uint32_t block_size;		/* Size of a block in bytes. Default: 64 KiB */
	uint32_t mmap_addr_deprecated;	/* 32-bit MMIO address of the store for read only access.
					   Prefer 'mmap_addr' for new software.
					   Zero when the address won't fit into 32-bits. */
	uint32_t com_buffer;		/* Physical address of the communication buffer */
	uint32_t com_buffer_size;	/* Size of the communication buffer in bytes */
	uint8_t apm_cmd;		/* The command byte to write to the APM I/O port */
	uint8_t unused[3];		/* Set to zero */
	uint64_t mmap_addr;		/* 64-bit MMIO address of the store for read only access.
					   Introduced after the initial implementation. Users of
					   this table must check the 'size' field to detect if its
					   written out by coreboot. */
};

The absence of this coreboot table entry indicates that there’s no SMMSTORE support. The table and type name retain the historical smmstorev2 spelling for ABI compatibility.

mmap_addr is an optional field added after the initial implementation. Users of this table must check the size field to know if it’s written by coreboot. In case it’s not present ‘mmap_addr_deprecated’ is to be used as the SPI ROM MMIO address and it must be below 4 GiB.

Blocks

SMMSTORE splits the SMMSTORE FMAP partition into smaller chunks called blocks. The logical block size defaults to 64 KiB and is advertised in the coreboot table as block_size. A payload or OS must use that value and make no further assumptions about the block or communication buffer size.

Generating the SMI

SMMSTORE is called via an SMI, which is generated via a write to the IO port defined in the smi_cmd entry of the FADT ACPI table. %al contains APM_CNT_SMMSTORE=0xed and is written to the smi_cmd IO port. %ah contains the SMMSTORE command. %ebx contains the parameter buffer to the SMMSTORE command.

Return values

If a command succeeds, SMMSTORE will return with SMMSTORE_RET_SUCCESS=0 in %eax. On failure SMMSTORE will return SMMSTORE_RET_FAILURE=1. For unsupported SMMSTORE commands SMMSTORE_RET_UNSUPPORTED=2 is returned.

NOTE 1: The caller must check the return value and should make no assumption on the returned data if %eax does not contain SMMSTORE_RET_SUCCESS.

NOTE 2: If the SMI returns without changing %ax, it can be assumed that the SMMSTORE feature is not installed.

Calling arguments

SMMSTORE supports the following subcommands that are passed via %ah; the additional calling arguments are passed via %ebx.

NOTE: The size of the struct entries are in the native word size of smihandler. This means 32 bits in almost all cases.

- SMMSTORE_CMD_INIT_DEPRECATED = 4

Unused, returns SMMSTORE_RET_UNSUPPORTED.

- SMMSTORE_CMD_RAW_READ = 5

SMMSTORE allows reading arbitrary data. It is up to the caller to initialize the store with meaningful data before using it.

The additional parameter buffer %ebx contains a pointer to the following struct:

struct smmstore_params_raw_read {
	uint32_t bufsize;
	uint32_t bufoffset;
	uint32_t block_id;
} __packed;

INPUT:

  • bufsize: Size of data to read within the communication buffer

  • bufoffset: Offset within the communication buffer

  • block_id: Block to read from

- SMMSTORE_CMD_RAW_WRITE = 6

SMMSTORE allows writing arbitrary data. It is up to the caller to erase a block before writing it.

The additional parameter buffer %ebx contains a pointer to the following struct:

struct smmstore_params_raw_write {
        uint32_t bufsize;
        uint32_t bufoffset;
        uint32_t block_id;
} __packed;

INPUT:

  • bufsize: Size of data to write within the communication buffer

  • bufoffset: Offset within the communication buffer

  • block_id: Block to write to

- SMMSTORE_CMD_RAW_CLEAR = 7

SMMSTORE allows clearing blocks. A cleared block will read as 0xff. By providing multiple blocks the caller can implement a fault tolerant write mechanism. It is up to the caller to clear blocks before writing to them.

struct smmstore_params_raw_clear {
	uint32_t block_id;
} __packed;

INPUT:

  • block_id: Block to erase

Security

Pointers provided by the payload or OS are checked to not overlap with SMM. This protects the SMM handler from being compromised.

As all information is exchanged using the communication buffer and coreboot tables, there’s no risk that a malicious application capable of issuing SMIs could extract arbitrary data or modify the currently running kernel.

Capsule update API

Availability of this command is tied to CONFIG_DRIVERS_EFI_UPDATE_CAPSULES.

To allow updating full flash content (except if locked at hardware level), few new calls were added. They reuse communication buffer, SMI command, return values and calling arguments of SMMSTORE commands listed above, with the exception of subcommand passed via %ah. If the subcommand is to operate on full flash size, it has the highest bit set, e.g. it is 0x85 for SMMSTORE_CMD_RAW_READ and 0x86 for SMMSTORE_CMD_RAW_WRITE. Every block_id describes block relative to the beginning of a flash, maximum value depends on its size.

Attempts to write the protected memory regions can lead to undesired consequences ranging from system instability to bricking and security vulnerabilities. When this feature is used, care must be taken to temporarily lift protections for the duration of an update when the whole flash is rewritten or the update must be constrained to affect only writable portions of the flash (e.g., “BIOS” region).

There is one new subcommand that must be called before any other subcommands with highest bit set can be used.

- SMMSTORE_CMD_USE_FULL_FLASH = 0x80

This command can only be executed once and is done by the firmware. Calling this function at runtime has no effect. It takes one additional parameter that, contrary to other commands, isn’t a pointer. Instead, %ebx indicates requested state of full flash access. If it equals 0, commands for accessing full flash are permanently disabled, otherwise they are permanently enabled until the next boot.

The assumption is that if capsule updates are enabled at build time and whole flash access is enabled at runtime, a UEFI payload (highly likely EDK2 or its derivative) won’t allow a regular OS to boot if the handler is enabled without rebooting first. There could be a way of deactivating the handler, but coreboot, having no way of enforcing its usage, might as well permit access until a reboot and rely on the payload to do the right thing.