On Wed, Jul 10, 2024 at 8:08 AM Lei Liu <liulei.rjpt@xxxxxxxx> wrote: > > > on 2024/7/10 22:48, Christian König wrote: > > Am 10.07.24 um 16:35 schrieb Lei Liu: > >> > >> on 2024/7/10 22:14, Christian König wrote: > >>> Am 10.07.24 um 15:57 schrieb Lei Liu: > >>>> Use vm_insert_page to establish a mapping for the memory allocated > >>>> by dmabuf, thus supporting direct I/O read and write; and fix the > >>>> issue of incorrect memory statistics after mapping dmabuf memory. > >>> > >>> Well big NAK to that! Direct I/O is intentionally disabled on DMA-bufs. > >> > >> Hello! Could you explain why direct_io is disabled on DMABUF? Is > >> there any historical reason for this? > > > > It's basically one of the most fundamental design decision of DMA-Buf. > > The attachment/map/fence model DMA-buf uses is not really compatible > > with direct I/O on the underlying pages. > > Thank you! Is there any related documentation on this? I would like to > understand and learn more about the fundamental reasons for the lack of > support. Hi Lei and Christian, This is now the third request I've seen from three different companies who are interested in this, but the others are not for reasons of read performance that you mention in the commit message on your first patch. Someone else at Google ran a comparison between a normal read() and a direct I/O read() into a preallocated user buffer and found that with large readahead (16 MB) the throughput can actually be slightly higher than direct I/O. If you have concerns about read performance, have you tried increasing the readahead size? The other motivation is to load a gajillion byte file from disk into a dmabuf without evicting the entire contents of pagecache while doing so. Something like this (which does not currently work because read() tries to GUP on the dmabuf memory as you mention): static int dmabuf_heap_alloc(int heap_fd, size_t len) { struct dma_heap_allocation_data data = { .len = len, .fd = 0, .fd_flags = O_RDWR | O_CLOEXEC, .heap_flags = 0, }; int ret = ioctl(heap_fd, DMA_HEAP_IOCTL_ALLOC, &data); if (ret < 0) return ret; return data.fd; } int main(int, char **argv) { const char *file_path = argv[1]; printf("File: %s\n", file_path); int file_fd = open(file_path, O_RDONLY | O_DIRECT); struct stat st; stat(file_path, &st); ssize_t file_size = st.st_size; ssize_t aligned_size = (file_size + 4095) & ~4095; printf("File size: %zd Aligned size: %zd\n", file_size, aligned_size); int heap_fd = open("/dev/dma_heap/system", O_RDONLY); int dmabuf_fd = dmabuf_heap_alloc(heap_fd, aligned_size); void *vm = mmap(nullptr, aligned_size, PROT_READ | PROT_WRITE, MAP_SHARED, dmabuf_fd, 0); printf("VM at 0x%lx\n", (unsigned long)vm); dma_buf_sync sync_flags { DMA_BUF_SYNC_START | DMA_BUF_SYNC_READ | DMA_BUF_SYNC_WRITE }; ioctl(dmabuf_fd, DMA_BUF_IOCTL_SYNC, &sync_flags); ssize_t rc = read(file_fd, vm, file_size); printf("Read: %zd %s\n", rc, rc < 0 ? strerror(errno) : ""); sync_flags.flags = DMA_BUF_SYNC_END | DMA_BUF_SYNC_READ | DMA_BUF_SYNC_WRITE; ioctl(dmabuf_fd, DMA_BUF_IOCTL_SYNC, &sync_flags); } Or replace the mmap() + read() with sendfile(). So I would also like to see the above code (or something else similar) be able to work and I understand some of the reasons why it currently does not, but I don't understand why we should actively prevent this type of behavior entirely. Best, T.J. > > > >>> > >>> We already discussed enforcing that in the DMA-buf framework and > >>> this patch probably means that we should really do that. > >>> > >>> Regards, > >>> Christian. > >> > >> Thank you for your response. With the application of AI large model > >> edgeification, we urgently need support for direct_io on DMABUF to > >> read some very large files. Do you have any new solutions or plans > >> for this? > > > > We have seen similar projects over the years and all of those turned > > out to be complete shipwrecks. > > > > There is currently a patch set under discussion to give the network > > subsystem DMA-buf support. If you are interest in network direct I/O > > that could help. > > Is there a related introduction link for this patch? > > > > > Additional to that a lot of GPU drivers support userptr usages, e.g. > > to import malloced memory into the GPU driver. You can then also do > > direct I/O on that malloced memory and the kernel will enforce correct > > handling with the GPU driver through MMU notifiers. > > > > But as far as I know a general DMA-buf based solution isn't possible. > > 1.The reason we need to use DMABUF memory here is that we need to share > memory between the CPU and APU. Currently, only DMABUF memory is > suitable for this purpose. Additionally, we need to read very large files. > > 2. Are there any other solutions for this? Also, do you have any plans > to support direct_io for DMABUF memory in the future? > > > > > Regards, > > Christian. > > > >> > >> Regards, > >> Lei Liu. > >> > >>> > >>>> > >>>> Lei Liu (2): > >>>> mm: dmabuf_direct_io: Support direct_io for memory allocated by > >>>> dmabuf > >>>> mm: dmabuf_direct_io: Fix memory statistics error for dmabuf > >>>> allocated > >>>> memory with direct_io support > >>>> > >>>> drivers/dma-buf/heaps/system_heap.c | 5 +++-- > >>>> fs/proc/task_mmu.c | 8 +++++++- > >>>> include/linux/mm.h | 1 + > >>>> mm/memory.c | 15 ++++++++++----- > >>>> mm/rmap.c | 9 +++++---- > >>>> 5 files changed, 26 insertions(+), 12 deletions(-) > >>>> > >>> > >