10:11 <robmur01> bbrezillon: TBH I think I'd prefer it to look like a loop in the caller :P

10:12 <robmur01> (which there still has to be anyway, since arm_lpae_map_pages() can only handle a single table worth of PTEs per call)

10:22 <robmur01> if and when it can be shown that page-by-page mapping has a significant performance impact (I can't imagine that sort of large VMA initialisation is a frequent operation?), then I'd probably lean towards having a separate map-with-non-incrementing-PA callback rather than complicate the existing map_pages interface any further

10:59 <bbrezillon> if it has to be page-by-page internally, not sure it's worth teaching io_pgtable_ops about the repeated thing...

11:03 <bbrezillon> I thought things could be optmized by preparing a full page table of dummy mappings that could be re-used instead of duplicated, but then you need to refcount those pages, oh well

11:25 <robmur01> yeah, sharing the same *table* across multiple VAs gets nasty when you then have to un-share it again to change anything at the leaf level.

16:23 <robclark> bbrezillon: I do it page by page in msm

16:32 <robclark> https://patchwork.freedesktop.org/patch/643687/?series=142263&rev=2

17:17 <cwabbott> also of note, the dummy page mapping by itself is not super useful, games also rely on the ability to zero it *and* the ability to tell in the shader if it's a zeroed page when texture sampling