Hi Piotr, Tudor,

On Wed, 27 Jun 2018 15:16:03 +0200
Piotr Bugalski [off-list ref] wrote:

Hello,

Atmel SAMA5D2 is equipped with two QSPI interfaces. These interfaces can
work as in SPI-compatible mode or use two / four lines to improve
communication speed. At the moment there is QSPI driver strongly tied to
NOR-flash memory and MTD subsystem.
Intention of this change is to provide new driver which will not be tied
to MTD and allows using QSPI with NAND-flash memory or other peripherals
New spi-mem API provides abstraction layer which can disconnect QSPI
from MTD. This driver doesn't support regular SPI interface, it should
be used with spi-mem interface only.
Unfortunately SAMA5D2 hardware by default supports only NOR-flash
memory. It allows 24- and 32-bit addressing while NAND-flash requires
16-bit long. To workaround hardware limitation driver is a bit more
complicated.

Request to spi-mem contains three fiels: opcode (command), address,
dummy bytes. SAMA5D2 QSPI hardware supports opcode, address, dummy and
option byte where address field can only be 24- or 32- bytes long.
Handling 8-bits long addresses is done using option field. For 16-bits
address behaviour depends of number of requested dummy bits. If there
are 8 or more dummy cycles, address is shifted and sent with first dummy
byte. Otherwise opcode is disabled and first byte of address contains
command opcode (works only if opcode and address use the same buswidth).
The limitation is when 16-bit address is used without enough dummy
cycles and opcode is using different buswidth than address. Other modes
are supported with described workaround.

It looks like hardware has some limitation in performance. The same issue
exists in current QSPI driver (MTD/nor-flash) and soft-pack (bare-metal
library from Atmel). Without using DMA read speed is much worse than
maximum bandwidth (efficiency 30-40%). Any help with performance
improvement is highly welcome, especially for NAND-flash memories which
offers higher capacity than NOR-flash used with previous driver.

Best Regards,
Piotr

v2 changes:
 - driver is now replacement of existing atmel-quadspi
 - code was re-written to follow original code structure
 - deinitialization order fixed
 - empty atmel_qspi_adjust_op_size function removed
 - code formatting fixes
 - use spi_device->max_speed_hz to get spi speed
 - spi freqency set in spi_controller->setup() hook
 - address range checkng for 4-bytes addressing
 - use timeout to avoid infinite waiting

IIRC, this version was almost ready to be merged except for the missing
->get_name() implementation which should be easy to add.

Can one of you post a new version of this patchset?

Thanks,

Boris