to save the preserved space(1KB) and also avoid the heap overhead
Signed-off-by: Xiang Xiao <xiaoxiang@xiaomi.com>
Change-Id: I694073f68e1bd63960cedeea1ddec441437be025
This fixes the following 3 issues:
1. Wait for send to complete in exchange
Before shutting down the SPI, we have to wait for send to complete; not only
DMA, since DMA just puts data to the SPI fifo. It is not yet out of SPI.
When doing exchange with both send & receive this is not an issue because when
receive dma has completed, it is certain that also the send is.
This can be accomplished by completing the transfer in SPI TXC interrupt
instead of DMA callback.
2. Fix TXDMAEN and RXDMAEN placement
According to the spec, the RXDMAEN must be enabled before
enabling DMA requests for Tx and Rx in DMA registers, and TXDMAEN
after that.
Cleaner place to do this is in spi_dmarxstart and spi_dmatxstart, where
also the dma requests are enabled. This also handles properly the
simplex modes.
3. Remove bus signal glitches when shutting off SPI block
Use AFCNTR to avoid glitches in SPI lines while turning SPI block
on/off during calls to exchange.
Signed-off-by: Jukka Laitinen <jukka.laitinen@intel.com>
574b25 broke the internal DMA buffers usage that solved
the following problem: The DMA capable interface does
not know the buffers extent. It calulates it from size.
The user may need to transfer less than a cachline bytes,
but STILL have a DMA cabable buffer. The user is therefore
foreced to transfer more data then needed to "trick" the
DMA cabable function. This is a wast of bus bandwith and
may not work will all devices. The internal buffer, solve
this issue.
stm32h7:stm32_spi review changes
Added sugestion from jlaitine to support RX only.
This is a work in progress, and now only serves as
DMA enabled simplex RX-only mode bus controller
Signed-off-by: Jukka Laitinen <jukka.laitinen@intel.com>
For SDMMC1, IDMA cannot access SRAM123 or SRAM4. Refer to ST AN5200 for
details. This patch makes stm32_dmapreflight check the buffer address and
return an error when the buffer is located in a invalid address space.
This does not fix the hardware limitation but at least makes it visible.
This simplifies the sdmmc driver when the IDMA is in use. There is no need to mix
IDMA and interrupt based transfers; instead, when making unaligned data tranfers,
just make IDMA into an internal aligned buffer and then copy the data.
Signed-off-by: Jukka Laitinen <jukka.laitinen@intel.com>
It should not be an error to clean cache beyond the dma source buffer
boundaries. It would just prematurely push some unrelated data from
cache to memory.
The only case where it would corrupt memory is that there is a dma
destination buffer overlapping the same cache line with the source
buffer. But this can't happen, because a destination buffer must always
be cache-line aligned when using write-back cache.
This patch enables doing dma tx-only transfer from unaligned source
buffer when using write-back cache.
Signed-off-by: Jukka Laitinen <jukka.laitinen@intel.com>
Correct flash write and erase functions, they inherit some
broken code from other platforms. Also fix the confusion between
eraseblock(sector) and page sizes.
Signed-off-by: Jari Nippula <jari.nippula@intel.com>
If the flash option register was locked before modifying it, return
it to the locked state after modify.
Signed-off-by: Jukka Laitinen <jukka.laitinen@intel.com>
First, configure the dmacfg in spi_dmarxsetup and spi_dmatxsetup. Then,
check for dmacapable, and only after that set up the dma.
This way the dmacapable actually works, and we don't need to initialize
the dmacfg structures twice.
Signed-off-by: Jukka Laitinen <jukka.laitinen@intel.com>
When starting dma transfer, the dcache for the TX buffer should be cleaned.
"flush" performs also invalidate, which is unnecessary. The TX buffer
can be unaligned to the cahche line in some(most) cases, whereas RX buffer
can never be.
The cache for the receive buffer can be dirty and valid before call to exchange.
Thus another memory access (hitting the same cache line) may corrupt receive data
while waiting for transfer to complete. So the receive buffer should be
invalidated before the transfer
Signed-off-by: Jukka Laitinen <jukka.laitinen@intel.com>
Modify some comments and debug assertions, which inherit from previous versions
and make no sense. Also add a few nerr printouts to make it easier to debug
running out of buffers
Signed-off-by: Jukka Laitinen <jukka.laitinen@intel.com>
