This library provides a convenient way to store information as key-value pairs in non-volatile storage.
- Supports any storage which can be modeled as a block device, including internal flash or external flash (e.g. via QSPI).
- Allows partitioning storage by instantiating multiple instances of the library.
- Designed to be resilient to power failures.
- Designed to promote even wear of the storage.
The utility operates on the assumption that the underlying storage is a block device which may not
have the same program and erase sizes. The application is required to provide an implementation for
a block device interface (mtb_kvstore_bd_t
) that describes the characteristics of the underlying
storage and defines basic operations (read, write, erase). See mtb_kvstore_init
for a description
of the constraints that the storage must satisfy. Code snippets using the serial-flash library
(External Flash) and HAL flash driver (Internal Flash) are provided as example implementations of
the storage interface in the Getting Started section in the API Reference Guide.
Keys are ASCII strings (Null terminated). The maximum key length is defined by MTB_KVSTORE_MAX_KEY_SIZE
.
This can be overridden by the application by adding DEFINES+=MTB_KVSTORE_MAX_KEY_SIZE=<value>
with the
desired value to application Makefile.
Values are arbitrary length binary data. By default it is assumed that each value (plus the header info of
~20 bytes) is less than the mtb_kvstore_bd_t
erase_size
. If this is not true, a define
MTB_KVSTORE_MAX_VALUE_SIZE
must be defined with the actual max value size.
NOTE: This define is not currently used, but will be in a future update. Setting it now will help ensure
the application works properly with newer versions of the kv-store library.
In an RTOS environment, the library can be made thread safe by adding the RTOS_AWARE
component
(COMPONENTS+=RTOS_AWARE) or by defining the CY_RTOS_AWARE
macro (DEFINES+=CY_RTOS_AWARE). This
causes all API to be protected by a mutex to serialize access to underlying storage device. The
kvstore library must be initialized after the RTOS kernel has started for the mutex to be initialized
safely. The default timeout for the mutex is defined by MTB_KVSTORE_MUTEX_TIMEOUT_MS
and can be
overridden by specifying DEFINES+=MTB_KVSTORE_MUTEX_TIMEOUT_MS=<value>
with the application Makefile.
When determining a suitable timeout, consider that the execution time for KVStore modifying operations is impacted by several factors:
- The size of the key and value being written
- If garbage collection is required (see "Garbage collection" section for details) to complete a
modification, this operation must erase half of the storage which was provided for kv-store to use.
This is a potentially lengthy operation. The
mtb_kvstore_ensure_capacity
function can be used to trigger garbage collection, if necessary, in less timing sensitive contexts. - All operations are impacted by the write performance of the underlying storage device. For more details, see the datasheet of the selected MCU (for internal flash) or the external memory device.
The key-value pairs are stored sequentially as records. Each operation appends a new record to the next available program block. If multiple records exist for the same key, the record at the highest address takes precedence over those that came before it. In an update operation, a new record with the key and updated value is appended, overriding the previous value. In a delete operation, a new record with the key and a "deleted" flag is appended.
The address space provided by the application is divided into two areas. At any given time, one area is designated as active and the other as swap. All key/value pair records are stored in the active area. Each time garbage collection is performed, the active and swap area designations are reversed. The first record in the active area is an area header used to identify the current active area during initialization. This means that one half of the provided storage space is available for key value storage.
active area swap area
+-----------------------------+ +-----------------------------+
| area header record | | |
+-----------------------------+ | |
| key value record 1 | | |
+-----------------------------+ | |
| key value record 2 | | |
| | | |
+-----------------------------+ | |
| key value record 3 | | |
+-----------------------------+ | |
| free space | | |
| | | |
+-----------------------------+ +-----------------------------+
Each record contains a record header (_mtb_kvstore_record_header_t
) that contains metadata including
key/value sizes and a CRC. This is followed by the key and value data. The record is padded to the program
size.
+---------------------+-------------------------+--------------------------------+---------------+
| Record Header | Key | Data | prog size pad |
+---------------------+-------------------------+--------------------------------+---------------+
A table that contains a hash for every key and the corresponding offset of the latest record for that key in storage is maintained in RAM. This table is built from storage during initialization and is updated on every subsequent operation. The key is verified by matching the key in the record in the storage. If it does not match the next entry with the same hash in the RAM table is checked. This allows for the possibility that multiple distinct keys may hash to the same value.
The garbage collection operation copies all of the non-obsolete records (i.e. all of those listed in the RAM table) into the swap area. The swap area is then marked as the new active area by programming the area header at the start. The former active area is erased and becomes the new swap area. Garbage collection is performed in the following scenarios:
- The active area does not have sufficient space remaining to perform a requested modification (add, update, delete) and the active area contains obsolete records.
- A corrupted record is encountered during initialization. This may happen if a power failure occurs while the a record is being appended during a modifying operation. The garbage collection operation will copy all non-obsolete records preceding the corrupted record.
- The
mtb_kvstore_ensure_capacity
function is called and the active area does not contain the requested amount of space available for immediate usage.
NOTE: Due to the garbage collection operation, write and delete operations may consume significantly more time than typical when the active area becomes full. Hence, they must not be called from timing critical code.
- abstraction-rtos library if the
CY_RTOS_AWARE
macro is defined in the Makefile
- API Reference Guide
- Cypress Semiconductor, an Infineon Technologies Company
- Infineon GitHub
- ModusToolbox™
© Cypress Semiconductor Corporation (an Infineon company) or an affiliate of Cypress Semiconductor Corporation, 2021.