On Mon, Apr 25, 2011 at 05:31:36PM -0500, Scott Duplichan wrote:
Peter Stuge wrote:
]Still it's not nice to write outside the callers buffer. Another OS ]might call same function and SeaBIOS would end up corrupting some ]variable. Ungood. I guess memmove() is the only choice?
]//Peter
I had a couple of ideas for a more sound solution.
This came up for DMA on IDE too. What I did there was just fallback to PIO if the buffer was not aligned.
Out of curiosity, can you see what happens if you return DISK_RET_EBOUNDARY in the unaligned case?
Debugging them is a challenge for me because I am new to seabios development and its method for writing code that works correctly in both 16-bit and 32-bit mode. Anyway, the first method I tried was to allocate a temporary I/O buffer that has the proper alignment. While that creates overhead, it would only be invoked in the rare case of unaligned request. But then there is the possibility that the allocation function will not be able to satisfy the request.
It's not currently possible to dynamically allocate memory during runtime. However, it is possible to allocate a buffer at init and keep it around for later. How big of a buffer would you need?
Since several devices require DMA buffers, I had thought of adding code for a low-memory dma pool. I never got around to it though.
Another possibility is splitting the request. The caller's buffer could handle the bigger part, and a stack buffer could be used for the remaining part.
The stack is only 512 bytes (for disk ops) so you don't have that much space.
Yet another possibility is to backup the byte that gets overwritten by the current patch method and restore it afterwards. While that is an unreliable method for paged code, it might work here. The only danger I can think of is if the caller's buffer is at the end of a physical DRAM range.
That seems fragile.
Thanks, -Kevin