According to documentation, there must be at least one cycle delay
between two consecutive writes of the DR register. Apparently, this is
true for reading too, though undocumented.
The read_dr() inline function is called as follows:
*(spi->rx_buf) = read_dr(info->regs);
which the compiler, with full optimizations on, turns it into:
ld_s r2,[r1,24] <- the spi->rx_buf
sr 0x80000000,[r4] <- this is the strobe bit write !!!
lr r4,[r4] <- this is the FIFO read!!!
stb r4,[r2] <- store the result
Unfortunately, the read from the FIFO is always 0 since the FIFO data is
not yet available.
During my investigations, I found that the following code works:
sys_out32(1 << 31, info->regs + 0xd);
*(spi->rx_buf) = sys_in32(info->regs + 0xd);
This does, basically, the same thing. But the compiler inserts an
instruction in between the write/read:
sr 0x80000000,[r4] <- write of the strobe bit!!!
ld_s r2,[r1,24] <- the pointer goes to r2
lr r4,[r4] <- read from FIFO!!!
stb r4,[r2] <- store the result
A single clock cycle between writing the register and reading seems
to be enough for the data to become available for reading.
This patch adds a nop in the read_dr() inline function.
Change-Id: I0c216d5738d5771835b1052e2e83363e8e3abf0c
Signed-off-by: Laurentiu Palcu <laurentiu.palcu@intel.com>
27a3f8bb10