On Thu, Aug 28, 2025 at 04:18:20PM -0300, Jason Gunthorpe wrote:

On Thu, Aug 28, 2025 at 01:10:32PM -0600, Keith Busch wrote:

quoted

Data and metadata are mapped as separate operations. They're just
different parts of one blk-mq request.

In that case the new bit leon proposes should only be used for the
unmap of the data pages and the metadata unmap should always be
unmapped as CPU?

The common path uses host allocated memory to attach integrity metadata,
but that isn't the only path. A user can attach their own metadata with
nvme passthrough or the recent io_uring application metadata, and that
could have been allocated from anywhere.

In truth though, I hadn't tried p2p metadata before today, and it looks
like bio_integrity_map_user() is missing the P2P extraction flags to
make that work. Just added this patch below, now I can set p2p or host
memory independently for data and integrity payloads:

---

diff --git a/block/bio-integrity.c b/block/bio-integrity.c
index 6b077ca937f6b..cf45603e378d5 100644
--- a/block/bio-integrity.c
+++ b/block/bio-integrity.c

@@ -265,6 +265,7 @@ int bio_integrity_map_user(struct bio *bio, struct iov_iter *iter)
 	unsigned int align = blk_lim_dma_alignment_and_pad(&q->limits);
 	struct page *stack_pages[UIO_FASTIOV], **pages = stack_pages;
 	struct bio_vec stack_vec[UIO_FASTIOV], *bvec = stack_vec;
+	iov_iter_extraction_t extraction_flags = 0;
 	size_t offset, bytes = iter->count;
 	unsigned int nr_bvecs;
 	int ret, nr_vecs;

@@ -286,7 +287,12 @@ int bio_integrity_map_user(struct bio *bio, struct iov_iter *iter)
 	}
 
 	copy = !iov_iter_is_aligned(iter, align, align);
-	ret = iov_iter_extract_pages(iter, &pages, bytes, nr_vecs, 0, &offset);
+
+	if (blk_queue_pci_p2pdma(q))
+		extraction_flags |= ITER_ALLOW_P2PDMA;
+
+	ret = iov_iter_extract_pages(iter, &pages, bytes, nr_vecs,
+					extraction_flags, &offset);
 	if (unlikely(ret < 0))
 		goto free_bvec;
 
--