Huge pmds and puds for graphics questions

["Thomas Hellström (VMware)" <thomas_os@shipmail.org>]

Hi!

I'm looking at supporting huge pmds and puds on VM_MIXEDMAPs for the TTM 
graphics memory manager. While huge puds might not be that common, huge 
pmds will probably be.

We would also want to support vrite-notify vmas, in which case we need 
to split the pud / pmds, and handle the write-notification on the pte 
level. Today we also support cow mappings, but I'm not sure anyone ever 
used it in this context. I figure in this case we just split the huge 
puds / pmds and we should be fine.

In any case, over to the questions:

In __handle_mm_fault() we call pmd_alloc() without considering puds 
being unstable (somewone faulting in a huge pud entry, or madvise() 
removing a huge pud entry). Is there some kind of special mechanism 
taking care of that potential race?

https://elixir.bootlin.com/linux/latest/source/mm/memory.c#L3931

In wp_huge_pmd() and wp_huge_pud(), if huge_fault() returns 
VM_FAULT_FALLBACK, and there is an existing read-only huge entry, we 
don't split that entry, causing the write fault to be endlessly retried. 
Is this an oversight or is the huge_fault handler responsible for 
splitting the huge entry?

https://elixir.bootlin.com/linux/latest/source/mm/memory.c#L3738

https://elixir.bootlin.com/linux/latest/source/mm/memory.c#L3769


Any insight or comments would be appreciated.

Thanks,

Thomas