Would this be within the scope of the eSDK?

[alexrp]

Hi!

There was recently a discussion on the Technical Q&A forum about physical memory allocation for Epiphany under Linux: viewtopic.php?f=23&t=668

The TL;DR is that it is not possible to allocate physical memory from Linux user space and pass it down to the Epiphany safely, because Linux may decide to swap out the memory at any point.

It would, however, be perfectly possible to write a kernel driver that allocates a bunch of physical memory and hands the address to the user space caller. Such a module would be 'dangerous' in that it could easily exhaust physical memory without the kernel being able to do anything about it, but then, it's not really any more unsafe than the Epiphany cores being able to access any and all physical memory.

So, would a kernel module that does this be within the scope of the eSDK? I figure I can't be the only one who thinks the 32 kB per core is somewhat small for some tasks...

[ysapir]

@alexrp - I am a little confused by your question. You mention that the 32KB per-core memory is a limitation - but it is so regardless of the amount of external memory available.

The kernel mode mechanism you described certainly can work for statically allocating shared memory. In fact, this is, in a way, what we ended up doing in the Parallella. In order to prevent Linux from moving buffers around while they are accessed by the Epiphany, we configured the system to allocate 32MB of SDRAM outside of Linux' space. This way, the allocation is static and the memory is accessible by both processors.

Note that the amount of External Memory allocated can be reconfigured to any amount you need for your application.

[shodruk]

[alexrp]

[alexrp]

[over9000]

[over9000]

[alexrp]

@over9000 thanks for the testing, but the problem I see with this scheme is that there is no way to actually obtain the physical address of a locked memory region. Or am I missing something?

[over9000]

[shodruk]