Currently `TDBStore::offset_in_erase_unit()` and
`TDBStore::check_erase_before_write()` loop through erase units
one-by-one, until the entire range is covered. This is very inefficient
when the erase size is tiny, e.g. one-byte on a non-flash device for
which we use program as erase.
This commit reworks the algorithms, based on the fact that a block
device can erase or program as many units as needed in one go.
The "erased_set_get" test case deinits `FileSystemStore`, erases the
block device, and reinits `FileSystemStore`. This of course fails,
because `BlockDevice::erase()` removes all stored data including the
format of the `FileSystem` (middle layer), unless the particular type
of block device has a no-op erase implementation.
Note: Previously `HeapBlockDevice::erase()` was essentially a no-op
so this test case did not fail. We recently added the freeing of heap
memory and it uncovered the problem.
Mbed OS does not enable C++ exceptions, so we should call `new` with
`std::nothrow` which returns a C-style NULL pointer when allocation
fails to allow error handling.
For consistency of style within the same file, this commit also
replaces `malloc()` and `free()` to `new (std::nothrow)` and `delete`.
f_mount() is doing both initializing (via find_volume() -> disk_initialize() where disk_initialize api calls Mbed OS init() function)
underlying block device and register filesystem but in case of "unmount" calls this f_mount is doing deregister filesystem
and missed to a deinitializing block device. So added the Mbed OS "deint()" call in unmount API to deinitialize block device
as f_mount API doesn't have a way to call deinit().
Targets that inherit from this target will have the defines necessary to
place the WiFi firmware in external storage and enable use of the
external WiFi firmware reserved region block device.
Currently, the only target inheriting from this new target is
CY8CPROTO-062S3-4343W.
PR #13908 replaces the umbrella library for all KVStore from
"kvstore_global_api" to "kvstore". Now "kvstore_global_api"
still exists but is restricted to the global API part only.
By adding "kvstore" as a dependency of "kvstore_global_api",
existing bare metal projects (e.g. mbed-bootloader) that use
the latter library name to pull in the KVStore prototype class
will continue to work.
Previously, we always set the work buffer to 64 bytes, without
checking it was no less the actual program size. But we can't
simply switch to use the program size for the work buffer,
because if it's too small (e.g. 1 byte in some cases), we
will not be able to read the status header (24 bytes), and
small buffers means more underlying write operations and
lower efficiency.
This PR changes the work buffer size to be the program size,
or 64 bytes as an absolute minimum like before.
Previously, when writing a record header into a TDBStore, we
took the pointer to the record_header_t variable as the input,
but used a program-aligned chunk size which is larger.
As a result, garbage from the stack memory gets written.
This commit fixes that by buffering the record header.
Co-authored-by: Seppo Takalo <seppo.takalo@arm.com>
Add const to function declaration. Const objects cannot access const
safe functions, a compiler error is thrown if one tries to do so. This
helps ensure that the object calling the method will not be modified.
This is espescially desirable in functions where all we need to do is
query existing data and not make any alterations.
[Q/O/SPIFBlockDevice::erase() begin with an alignment check,
after which unaligned erases should not happen or be allowed.
If the erase address is not aligned to the value returned by
sfdp_iterate_next_largest_erase_type(), it indicates an
internal error in erase table parsing which should not be
hidden.
The test case set_add_data_set_key_value_five_Kbytes stores
5KB of value, and to allow for overheads, we allocate at least
8KB per area of TDBStore (so 16KB for two areas of a whole
TDBStore).
From the datasheet of the only OctaSPI flash we currently support
(http://www.mt-system.ru/sites/default/files/docs/Macronix/mx25lm51245g_3v_512mb_v0.01.pdf):
After program/erase command is issued, auto program/erase
algorithms which program/erase and verify the specified page
or sector/block locations will be executed. Program command is
executed on byte basis, or page (256 bytes) basis, or word basis.
Erase command is executed on sector (4K-byte), or block (64K-byte),
or whole chip basis.
So the minimum write size is one byte.
Prior to this PR, the minimum program size (QSPI_MIN_PROG_SIZE) of
QSPIFBlockDevice was 256 by default and 512 for some targets.
Those values were in fact page sizes, not program sizes.
Here's an explanation:
* Most QSPI flashes can be programmed to a granularity of a
single byte or a few bytes - no need to be a whole page.
This should be the value of QSPI_MIN_PROG_SIZE. Applications
need to align buffer sizes to this granularity when
programming QSPI flashes.
* Each sending of the underlying QSPI program signal requires
destination bytes to be located within the same page.
If a QSPIFBlockDevice::program() call crosses page boundaries,
this function breaks down the operation into multiple chunks,
so it's not a concern for the application.
So this PR changes the default program size to 1 (byte), and
for targets with a 4-byte (1-word) read size it overrides the
program size.
Note: No config is needed for the page size, as it comes from
the SFDP table parsed during initialisation.
Aside from the core mbed-os CMake target, a number of targets have been created so they can optionally be included by application executables that require them using `target_link_libraries()`.
Co-authored-by: Martin Kojtal <martin.kojtal@arm.com>
Co-authored-by: Rajkumar Kanagaraj <rajkumar.kanagaraj@arm.com>
Lingkai Dong
George Psimenos
jeromecoutant
rogeryou
Vikas Katariya
Martin Kojtal
Rajkumar Kanagaraj
Oliver Wildtgrube
Martino Facchin
Hugues Kamba
Matthew Macovsky
Harrison Mutai
Ahmet Alincak
Anna Bridge
Hugues Kamba
Evelyne Donnaes
pea-pod