mirror of
https://github.com/torvalds/linux.git
synced 2024-11-16 00:52:01 +00:00
5ecd3ea188
Under certain conditions, the SPI-FSM Controller can be left in a state where the data FIFO is not entirely empty. This can lead to problems where subsequent data transfers appear to have been shifted by a number of unidentified bytes. One simple example would be an errant FSM sequence which loaded more data to the FIFO than was read by the host. Another more interesting case results from an obscure artefact in the FSM Controller. When switching from data transfers in x4 or x2 mode to data transfers in x1 mode, extraneous bytes will appear in the FIFO, unless the previous data transfer was a multiple of 32 cycles (i.e. 8 bytes for x2, and 16 bytes for x4). This applies equally whether FSM is being operated directly by a S/W driver, or by the SPI boot-controller in FSM-Boot mode. Furthermore, data in the FIFO not only survive a transition between FSM-Boot and FSM, but also a S/W reset of IP block [1]. By taking certain precautions, it is possible to prevent the driver from causing this type of problem (e.g. ensuring that the host and programmed sequence agree on the transfer size, and restricting transfer sizes to multiples of 32-cycles [2]). However, at the point the driver is loaded, no assumptions can be made regarding the state of the FIFO. Even if previous S/W drivers have behaved correctly, it is impossible to control the number of transactions serviced by the controller operating in FSM-Boot. To address this problem, we ensure the FIFO is cleared during initialisation, before performing any FSM operations. Previously, the fsm_clear_fifo() code was capable of detecting and clearing any unwanted 32-bit words from the FIFO. This patch extends the capability to handle an arbitrary number of bytes present in the FIFO [3]. Now that the issue is better understood, we also remove the calls to fsm_clear_fifo() following the fsm_read() and fsm_write() operations. The process of actually clearing the FIFO deserves a mention. While the FIFO may contain any number of bytes, the SPI_FAST_SEQ_STA register only reports the number of complete 32-bit words present. Furthermore, data can only be drained from the FIFO by reading complete 32-bit words. With this in mind, a two stage process is used to the clear the FIFO: 1. Read any complete 32-bit words from the FIFO, as reported by the SPI_FAST_SEQ_STA register. 2. Mop up any remaining bytes. At this point, it is not known if there are 0, 1, 2, or 3 bytes in the FIFO. To handle all cases, a dummy FSM sequence is used to load one byte at a time, until a complete 32-bit word is formed; at most, 4 bytes will need to be loaded. [1] Although this issue has existed since early versions of the SPI-FSM controller, its full extent only emerged recently as a consequence of the targetpacks starting to use FSM-Boot(x4) as the default configuration. [2] The requirement to restrict transfers to multiples of 32 cycles was found empirically back when DUAL and QUAD mode support was added. The current analysis now gives a satisfactory explanation for this requirement. [3] Theoretically, it is possible for the FIFO to contain an arbitrary number of bits. However, since there are no known use-cases that leave incomplete bytes in the FIFO, only words and bytes are considered here. Signed-off-by: Angus Clark <angus.clark@st.com> Signed-off-by: Lee Jones <lee.jones@linaro.org> Signed-off-by: Brian Norris <computersforpeace@gmail.com> |
||
---|---|---|
.. | ||
chips | ||
devices | ||
lpddr | ||
maps | ||
nand | ||
onenand | ||
spi-nor | ||
tests | ||
ubi | ||
afs.c | ||
ar7part.c | ||
bcm47xxpart.c | ||
bcm63xxpart.c | ||
cmdlinepart.c | ||
ftl.c | ||
inftlcore.c | ||
inftlmount.c | ||
Kconfig | ||
Makefile | ||
mtd_blkdevs.c | ||
mtdblock_ro.c | ||
mtdblock.c | ||
mtdchar.c | ||
mtdconcat.c | ||
mtdcore.c | ||
mtdcore.h | ||
mtdoops.c | ||
mtdpart.c | ||
mtdsuper.c | ||
mtdswap.c | ||
nftlcore.c | ||
nftlmount.c | ||
ofpart.c | ||
redboot.c | ||
rfd_ftl.c | ||
sm_ftl.c | ||
sm_ftl.h | ||
ssfdc.c |