On Mon, Oct 01, 2018 at 05:50:13AM -0700, Christoph Hellwig wrote: > On Sat, Sep 29, 2018 at 09:21:17AM -0700, Matthew Wilcox wrote: > > > being slow to pick it up. It looks like there are several patterns, and > > > we have to support both set_page_dirty() and set_page_dirty_lock(). So > > > the best combination looks to be adding a few variations of > > > release_user_pages*(), but leaving put_user_page() alone, because it's > > > the "do it yourself" basic one. Scatter-gather will be stuck with that. > > > > I think our current interfaces are wrong. We should really have a > > get_user_sg() / put_user_sg() function that will set up / destroy an > > SG list appropriate for that range of user memory. This is almost > > orthogonal to the original intent here, so please don't see this as a > > "must do first" kind of argument that might derail the whole thing. > > The SG list really is the wrong interface, as it mixes up information > about the pages/phys addr range and a potential dma mapping. I think > the right interface is an array of bio_vecs. In fact I've recently > been looking into a get_user_pages variant that does fill bio_vecs, > as it fundamentally is the right thing for doing I/O on large pages, > and will really help with direct I/O performance in that case.
I don't think the bio_vec is really a big improvement; it's just a (page, offset, length) tuple. Not to mention that due to the annoying divergence between block and networking [1] this is actually a less useful interface. I don't understand the dislike of the sg list. Other than for special cases which we should't be optimising for (ramfs, brd, loopback filesystems), when we get a page to do I/O, we're going to want a dma mapping for them. It makes sense to already allocate space to store the mapping at the outset. [1] Can we ever admit that the bio_vec and the skb_frag_t are actually the same thing?
