Messages in this thread | | | Date | Tue, 19 Nov 2013 14:51:12 -0800 | From | Andy Lutomirski <> | Subject | Re: [PATCH RFC 0/3] Add dirty-tracking infrastructure for non-page-backed address spaces |
| |
On 11/19/2013 12:06 PM, Thomas Hellstrom wrote: > Hi! > > Before going any further with this I'd like to check whether this is an > acceptable way to go. > Background: > GPU buffer objects in general and vmware svga GPU buffers in > particular are mapped by user-space using MIXEDMAP or PFNMAP. Sometimes the > address space is backed by a set of pages, sometimes it's backed by PCI memory. > In the latter case in particular, there is no way to track dirty regions > using page_mkwrite() and page_mkclean(), other than allocating a bounce > buffer and perform dirty tracking on it, and then copy data to the real GPU > buffer. This comes with a big memory- and performance overhead. > > So I'd like to add the following infrastructure with a callback pfn_mkwrite() > and a function mkclean_mapping_range(). Typically we will be cleaning a range > of ptes rather than random ptes in a vma. > This comes with the extra benefit of being usable when the backing memory of > the GPU buffer is not coherent with the GPU itself, and where we either need > to flush caches or move data to synchronize. > > So this is a RFC for > 1) The API. Is it acceptable? Any other suggestions if not? > 2) Modifying apply_to_page_range(). Better to make a standalone > non-populating version? > 3) tlb- mmu- and cache-flushing calls. I've looked at unmap_mapping_range() > and page_mkclean_one() to try to get it right, but still unsure.
Most (all?) architectures have real dirty tracking -- you can mark a pte as "clean" and the hardware (or arch code) will mark it dirty when written, *without* a page fault.
I'm not convinced that it works completely correctly right now (I suspect that there are some TLB flushing issues on the dirty->clean transition), and it's likely prone to bit-rot, since the page cache doesn't rely on it.
That being said, using hardware dirty tracking should be *much* faster and less latency-inducing than doing it in software like this. It may be worth trying to get HW dirty tracking working before adding more page fault-based tracking.
(I think there's also some oddity on S/390. I don't know what that oddity is or whether you should care.)
--Andy
| |