On an io_uring O_DIRECT write, bio_integrity_map_user() can attach the caller's protection information (PI) buffer directly to the bio. The block layer pins the pages for reading, but reference-tag remapping can write to those same pages. A caller with write access to a block device whose integrity profile enables reference-tag remapping can therefore modify a page it can only map for reading. In testing, remapping changed a tag in the page cache of a read-only file. With a PI buffer backed by the shared zero page, the kernel faulted during remapping. Use the existing bounce-buffer path for REQ_OP_WRITE when the queue's integrity profile has BLK_INTEGRITY_REF_TAG set. This copies the PI buffer into kernel memory before remapping, so the caller's buffer retains its original reference tags. I found this issue during a source review assisted by an LLM. Fixes: 3d8b5a22d404 ("block: add support to pass user meta buffer") Cc: stable@vger.kernel.org Suggested-by: Christoph Hellwig Signed-off-by: Zakaria Belkadi --- v2: - Bounce the user PI buffer instead of pinning it for writing (Christoph). - Dropped Caleb's Reviewed-by, since the patch changed. - Corrected what the caller controls (Caleb): it chooses the address of the PI buffer, the seed and the device offset, so both where the tag is stored and its value. v1: https://lore.kernel.org/linux-block/20261118154053.13218-1-zakaria.belkadi@inria.fr/ Open question on stacked devices: the test uses the profile of the top-level queue. By my reading of the source, a device-mapper target with members of different profiles can get a top-level profile without BLK_INTEGRITY_REF_TAG while a member has it, because dm_set_device_limits() ignores a stacking mismatch. This condition would then not force a bounce. I have not tested whether such a configuration can accept user PI and reach reference-tag remapping on the underlying device. Would you prefer bouncing all user PI buffers on writes to cover this case? Tested on v7.4-rc3 in a QEMU guest, emulated NVMe namespace with T10-DIF type 1, as an unprivileged user who owns only the device node: - PI buffer = read-only mapping of a root-owned file: without the patch the reference tag in the file changes; with it the write succeeds and the file is unchanged. - PI buffer = shared zero page: without the patch the kernel oopses in blk_integrity_remap(); with it the write succeeds with no fault. - Writable PI buffers still work, and the kernel no longer modifies them. - Stored tags (as root): a write with a seed different from the block number succeeds with reference-tag checking enabled. A subsequent read using the block number as the seed returns the expected remapped tags, with identical results before and after the patch. block/bio-integrity.c | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/block/bio-integrity.c b/block/bio-integrity.c index d3df726e0..aaea5802e 100644 --- a/block/bio-integrity.c +++ b/block/bio-integrity.c @@ -394,6 +394,11 @@ int bio_integrity_map_user(struct bio *bio, struct iov_iter *iter) copy = iov_iter_alignment(iter) & blk_lim_dma_alignment_and_pad(&q->limits); + /* Reference-tag remapping must operate on a private copy of user PI. */ + if (bio_op(bio) == REQ_OP_WRITE && + (q->limits.integrity.flags & BLK_INTEGRITY_REF_TAG)) + copy = true; + if (blk_queue_pci_p2pdma(q)) extraction_flags |= ITER_ALLOW_P2PDMA; base-commit: aebb7274476403040d582c61478df8f4a7bf9cd3 -- 2.39.5