On Thu, 17 Aug 2023 16:57:16 +0300 Ilias Apalodimas wrote:

Why should we care about this?  Even an architecture that's 32-bit and
has a 64bit DMA should be allowed to split the pages internally if it
decides to do so.  The trick that drivers usually do is elevate the
page refcnt and deal with that internally.

Can we assume the DMA mapping of page pool is page aligned? We should
be, right? That means we're storing 12 bits of 0 at the lower end.
So even with 32b of space we can easily store addresses for 32b+12b =>
16TB of memory. "Ought to be enough" to paraphrase Bill G, and the
problem is only in our heads?

Before we go that way - Mina, are the dma-buf "chunks" you're working
with going to be fragment-able? Or rather can driver and/or core take
multiple references on a single buffer?