(In reply to Michel Dänzer from comment #1)
> (In reply to Shawn Starr from comment #0)
> > One solution discussed is to split up the transfer into smaller chunks in
> > radeon_ttm.
> 
> Specifically, here's how I think a fallback could be implemented in the
> kernel driver which can never fail because of fragmentation or resource
> starvation:
> 
> During initialization, reserve some pinned GTT memory for bounce buffers.
> When a BO can't be bound to GTT for eviction as in the case reported here,
> instead do the eviction directly from VRAM to CPU domain in one or several
> passes of:
> 1. Copy part of the BO from VRAM to one of the reserved bounce buffers in
> GTT using the GPU.
> 2. Copy that part of the BO from the bounce buffer to the BO's system RAM
> pages using the CPU.

if we can split a large BO to two parts, one is in VRAM, one is in GTT, seems
also to be helpful for this case.