iris/bufmgr: round up vma reservations on DG2

For DG2 we can't share PDEs with different sized PTEs. To ensure 4K PTE buffers don't try to share a PDE with 64K PTE buffers we reserve VA up to the next alignment.

See https://patchwork.freedesktop.org/series/98996 for further details

Signed-off-by: Robert Beckett bob.beckett@collabora.com

mentioned in merge request !12706 (merged)

We round the VMA base address for every buffer used to 2MiB. This should prevent any two buffers from sharing VMA space within the same 2MiB block. Does rounding the size make a difference?

Yeah, it certainly seems like bufmgr->vma_min_align in vma_alloc ought to take care of this problem already...or is there some case we missed?

added iris label

agreed. Sorry for the noise. I was mirroring what I did in the kernel, without paying enough attention to what mesa does here.

closed

binder_realloc currently assumes consecutive 64K offsets. upon chunk overflow, it tries to place the next buffer within the same 2mb region as the previous, which breaks on DG2.

Just to update this with our irc conversation, hopefully @pzanoni's !12028 (d3794e07) patch will help with this.