Unaligned access to AXI read slave yields incorrect data
ru551n opened this issue · 4 comments
When doing multi-beat accesses to AXI read slave with an unaligned start address, every beat except the first one will yield wrong data due to the address offsetting
Line 116 in axi_read_slave.vhd
:
for j in 0 to burst.size-1 loop
idx := (address + j) mod self.data_size;
rdata(8*idx+7 downto 8*idx) <= std_logic_vector(to_unsigned(read_byte(axi_slave.p_memory, address+j), 8));
end loop;
Example:
Bus width: 32 bits
Address: 1
Burst length: 2 beats, 7 bytes
For the first beat, due to the modulo operation, the byte with index 1, 2 and 3 will be correctly, set. However the byte with index with index 0 will be fetched from memory index 4. For the first beat, this does not matter, since that address is "invalid" anyway.
But for the second beat and onwards, the same will occur, byte with address 4 (and index 0) will be fetched from address 8.
In the case of unallocated data this will fail the allocation check, but otherwise this will return the wrong data.
This should be relatively easily fixed by looping over address the address to the next aligned address and also aligning all subsequent addresses. This fix will also yield undefined data (if this is configured via generic) on the bytes of the first beat which are "below" the starting address, which is probably a good thing since those bytes are not of interest, should the user decide to configure the drive_invalid
generic.
I will post a patch later.
Is this the same issue as the one addressed in this PR? #936
Is this the same issue as the one addressed in this PR? #936
It is the same issue yes, but in that case reading from an unaligned address you don't really care if it is only one beat.
Reading multiple beats you will actually get the wrong data from the second beat onwards.
@LarsAsplund, I can check it when I get the time. The solution is similiar to what I had in mind.