From: Mengmeng Zhao <[email protected]> This series adds dma-buf read/write support for ext4. The dma-buf direct I/O infrastructure allows dma-buf to be handed to file I/O as an iov_iter, but currently it only supports raw block devices. To better serve the GPU-NVMe direct access through normal filesystems, prior work for XFS has been discussed at [1]. We are also interested in this field and working on it [2].
The series is based on v7.4-rc2 (7475ea8b5861), which already contains the common dma-buf direct I/O infrastructure, so it no longer depends on a separate base tree. Patches 1-2 carry the block and iomap prerequisites from Nitesh Shetty and Anuj Gupta: - bdev_init_dma_buf_io_ctx in patch 1 was moved to block/bdev.c as suggested by Christoph. - iomap_file_init_dma_buf_io_ctx from that series is not carried. ext4 calls bdev_init_dma_buf_io_ctx directly, and the capability check is handled by FMODE_CAN_DMABUF_IO. The design is: - FMODE_CAN_DMABUF_IO marks files that can be the target of dma-buf I/O; a filesystem sets it at open time for inodes whose data can be reached without the page cache. - dma_buf_io_ctx_create() refuses files that lack the flag and checks the current f_flags for O_DIRECT at registration. - As F_SETFL can clear O_DIRECT afterwards, the paths that would fall back to buffered or copied I/O still need per-I/O checks in the callers. - DAX is rechecked per I/O because it can be enabled after open. - iomap must not ask for a buffered fallback on behalf of a dma-buf iterator: -ENOTBLK would be refused by the callers and would turn a failure into a successful transfer of zero bytes. Patch 3 keeps the buffered fallback out of dma-buf I/O, patches 4-7 add the refusals on the common paths and the capability flag, and patch 8 enables ext4. Known limitation: 1. A dma-buf read that reaches a hole or an unwritten extent fails as a whole, even when an earlier mapped extent was already transferred into the buffer, so the caller does not learn how much of the dma-buf was filled. Reporting the prefix instead would return a short read, which callers cannot tell apart from the end of the file, and both the buffered and the direct path normally hide holes by zero-filling. Probing the mapping before submitting anything would fix it, but that belongs to the common iomap series. 2. A refused dma-buf I/O reaches the caller as -EOPNOTSUPP through iomap_dio_set_error(), which also emits an fserror notification even though no I/O was submitted. The report could be limited to requests that did submit a bio, since iomap_dio_bio_iter() accounts dio->size at submission time, but that changes what fserror users see for every iomap filesystem, so this version leaves it as it is and asks the fserror maintainers. Testing: - io_uring dma-buf suite (udmabuf + REGISTER_BUFFERS_UPDATE + READ_FIXED / WRITE_FIXED) against ext4 on a PCIe NVMe namespace, all passed. - xfstests -g quick on ext4 devices: 689 tests ran, 11 reported failures, all of them environment related, no failure is attributable to this series. [1] https://lore.kernel.org/linux-xfs/[email protected]/ [2] https://lore.kernel.org/all/[email protected]/ Changelog: ========== RFC v2: - Rebased onto v7.4-rc2 and dropped the dependency on a separate base tree; the block and iomap prerequisites are patches 1-2. - Added patch 3, which reports why a dma-buf direct I/O could not be done instead of asking for a buffered fallback that the callers refuse. - Split the block changes: patch 4 only factors the integrity check into a helper, and patch 5 refuses a dma-buf bio in the inline encryption fallback, which is where the data pages are walked instead of in blk_mq_submit_bio(). (Christoph) - Dropped the mm/filemap and iomap_file_buffered_write() guards; the refusals now happen in the callers that would fall back to buffered or copied I/O. (Christoph) - Added the FMODE_CAN_DMABUF_IO flag and use it to check whether a bdev or a filesystem can do dma-buf I/O. (Christoph) RFC v1: - https://lore.kernel.org/all/[email protected]/ Mengmeng Zhao (6): iomap: don't ask for a buffered fallback for dma-buf iterators blk-mq: factor out a dma-buf bio refusal helper blk-crypto: refuse dma-buf bios in the fallback path fs: add FMODE_CAN_DMABUF_IO and require it for dma-buf I/O contexts dax: refuse dma-buf iterators in dax_iomap_rw() ext4: support dma-buf registered buffers for direct I/O Nitesh Shetty (2): block: factor out bdev_init_dma_buf_io_ctx() iomap: add dma-buf direct I/O guards .../filesystems/iomap/operations.rst | 3 ++ block/bdev.c | 14 ++++++ block/blk-crypto.c | 8 ++++ block/blk-mq.c | 24 ++++++---- block/fops.c | 12 +---- drivers/dma-buf/dma-buf-io.c | 13 ++++-- fs/dax.c | 4 ++ fs/ext4/ext4.h | 11 +++++ fs/ext4/file.c | 44 +++++++++++++++++-- fs/iomap/direct-io.c | 41 +++++++++++++++-- include/linux/blkdev.h | 8 ++++ include/linux/fs.h | 8 +++- 12 files changed, 159 insertions(+), 31 deletions(-) base-commit: 7475ea8b58616b33c04afb976cf0fbaad9dd63fb -- 2.43.0
