On Tue, Jul 16, 2024 at 06:14:48PM +0800, Huan Yang wrote: > > 在 2024/7/16 17:31, Daniel Vetter 写道: > > [你通常不会收到来自 daniel.vetter@xxxxxxxx 的电子邮件。请访问 https://aka.ms/LearnAboutSenderIdentification，以了解这一点为什么很重要;] > > > > On Tue, Jul 16, 2024 at 10:48:40AM +0800, Huan Yang wrote: > > > I just research the udmabuf, Please correct me if I'm wrong. > > > > > > 在 2024/7/15 20:32, Christian König 写道: > > > > Am 15.07.24 um 11:11 schrieb Daniel Vetter: > > > > > On Thu, Jul 11, 2024 at 11:00:02AM +0200, Christian König wrote: > > > > > > Am 11.07.24 um 09:42 schrieb Huan Yang: > > > > > > > Some user may need load file into dma-buf, current > > > > > > > way is: > > > > > > > 1. allocate a dma-buf, get dma-buf fd > > > > > > > 2. mmap dma-buf fd into vaddr > > > > > > > 3. read(file_fd, vaddr, fsz) > > > > > > > This is too heavy if fsz reached to GB. > > > > > > You need to describe a bit more why that is to heavy. I can only > > > > > > assume you > > > > > > need to save memory bandwidth and avoid the extra copy with the CPU. > > > > > > > > > > > > > This patch implement a feature called DMA_HEAP_IOCTL_ALLOC_READ_FILE. > > > > > > > User need to offer a file_fd which you want to load into > > > > > > > dma-buf, then, > > > > > > > it promise if you got a dma-buf fd, it will contains the file content. > > > > > > Interesting idea, that has at least more potential than trying > > > > > > to enable > > > > > > direct I/O on mmap()ed DMA-bufs. > > > > > > > > > > > > The approach with the new IOCTL might not work because it is a very > > > > > > specialized use case. > > > > > > > > > > > > But IIRC there was a copy_file_range callback in the file_operations > > > > > > structure you could use for that. I'm just not sure when and how > > > > > > that's used > > > > > > with the copy_file_range() system call. > > > > > I'm not sure any of those help, because internally they're all still > > > > > based > > > > > on struct page (or maybe in the future on folios). And that's the thing > > > > > dma-buf can't give you, at least without peaking behind the curtain. > > > > > > > > > > I think an entirely different option would be malloc+udmabuf. That > > > > > essentially handles the impendence-mismatch between direct I/O and > > > > > dma-buf > > > > > on the dma-buf side. The downside is that it'll make the permanently > > > > > pinned memory accounting and tracking issues even more apparent, but I > > > > > guess eventually we do need to sort that one out. > > > > Oh, very good idea! > > > > Just one minor correction: it's not malloc+udmabuf, but rather > > > > create_memfd()+udmabuf. > > Hm right, it's create_memfd() + mmap(memfd) + udmabuf > > > > > > And you need to complete your direct I/O before creating the udmabuf > > > > since that reference will prevent direct I/O from working. > > > udmabuf will pin all pages, so, if returned fd, can't trigger direct I/O > > > (same as dmabuf). So, must complete read before pin it. > > Why does pinning prevent direct I/O? I haven't tested, but I'd expect the > > rdma folks would be really annoyed if that's the case ... > > > > > But current way is use `memfd_pin_folios` to boost alloc and pin, so maybe > > > need suit it. > > > > > > > > > I currently doubt that the udmabuf solution is suitable for our > > > gigabyte-level read operations. > > > > > > 1. The current mmap operation uses faulting, so frequent page faults will be > > > triggered during reads, resulting in a lot of context switching overhead. > > > > > > 2. current udmabuf size limit is 64MB, even can change, maybe not good to > > > use in large size? > > Yeah that's just a figleaf so we don't have to bother about the accounting > > issue. > > > > > 3. The migration and adaptation of the driver is also a challenge, and > > > currently, we are unable to control it. > > Why does a udmabuf fd not work instead of any other dmabuf fd? That > > shouldn't matter for the consuming driver ... > > Hmm, our production's driver provider by other oem. I see many of they > implement > > their own dma_buf_ops. These may not be generic and may require them to > reimplement. Yeah, for exporting a buffer object allocated by that driver. But any competent gles/vk stack also supports importing dma-buf, and that should work with udmabuf exactly the same way as with a dma-buf allocated from the system heap. > > > Perhaps implementing `copy_file_range` would be more suitable for us. > > See my other mail, fundamentally these all rely on struct page being > > present, and dma-buf doesn't give you that. Which means you need to go > > below the dma-buf abstraction. And udmabuf is pretty much the thing for > > that, because it wraps normal struct page memory into a dmabuf. > Yes, udmabuf give this, I am very interested in whether the page provided by > udmabuf can trigger direct I/O. > > So, I'll give a test and report soon. > > > > And copy_file_range on the underlying memfd might already work, I haven't > > checked though. > > I have doubts. > > I recently tested and found that I need to modify many places in > vfs_copy_file_range in order to run the copy file range with DMA_BUF fd.(I > have managed to get it working, I'm talking about memfd, not dma-buf here. I think copy_file_range to dma-buf is as architecturally unsound as allowing O_DIRECT on the dma-buf mmap. Cheers, Sima > but I don't think the implementation is good enough, so I can't provide the > source code.) > > Maybe memfd can work or not, let's give it a test.:) > > Anyway, it's a good idea too. I currently need to focus on whether it can be > achieved, as well as the performance comparison. > > > > > Cheers, Sima > > -- > > Daniel Vetter > > Software Engineer, Intel Corporation > > http://blog.ffwll.ch/ -- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch
