block: fix pgmap handling for zone device pages in bio merge paths

When zone device memory is registered in multiple chunks, each chunk gets its own dev_pagemap. A single bio can contain bvecs from different pgmaps -- iov_iter_extract_bvecs() breaks at pgmap boundaries but the outer loop in bio_iov_iter_get_pages() continues filling the same bio.

There are two problems with the current code:

1. biovec_phys_mergeable() has no pgmap check, so the request merge, DMA mapping, and integrity merge paths can coalesce physically contiguous bvec segments from different pgmaps. This makes it impossible to recover the correct pgmap for the merged segment via page_pgmap().

2. bio_add_page() and bio_integrity_add_page() reject pages from a different pgmap entirely (returning 0), rather than just skipping the merge and adding them as new bvec entries. This forces callers to start a new bio unnecessarily.

Patch 1 fixes the merge-path gap by adding a pgmap check to biovec_phys_mergeable().

Patch 2 introduces zone_device_pages_compatible() which replaces the blanket zone_device_pages_have_same_pgmap() rejection in bio_add_page() and bio_integrity_add_page(). Pages that are safe to coexist as separate bvec entries (e.g. MEMORY_DEVICE_GENERIC from different pgmaps) are now accepted, while P2PDMA pages from different pgmaps or mixed P2PDMA and non-P2PDMA pages are still rejected, since the DMA iterator caches the P2PDMA mapping state from the first segment. zone_device_pages_have_same_pgmap() is kept as a merge guard so pages from different pgmaps are not coalesced into the same bvec segment.

Changes since v1:
https://lore.kernel.org/all/[email protected]/
- Reworked patch 2 to introduce zone_device_pages_compatible() which rejects P2PDMA pages from different pgmaps at the bio-building level, not just at merge time. The previous version only moved the pgmap check into the merge conditional without preventing incompatible pages from being added as separate bvec entries. (Christoph Hellwig)

Naman Jain (2):
block: add pgmap check to biovec_phys_mergeable
block: relax pgmap check in bio_add_page for compatible zone device
pages

 block/bio-integrity.c |  6 +++---
 block/bio.c           |  6 +++---
 block/blk.h           | 21 +++++++++++++++++++++
 3 files changed, 27 insertions(+), 6 deletions(-)

2.43.0

biovec_phys_mergeable() is used by the request merge, DMA mapping, and integrity merge paths to decide if two physically contiguous bvec segments can be coalesced into one. It currently has no check for whether the segments belong to different dev_pagemaps.

When zone device memory is registered in multiple chunks, each chunk gets its own dev_pagemap. A single bio can legitimately contain bvecs from different pgmaps -- iov_iter_extract_bvecs() breaks at pgmap boundaries but the outer loop in bio_iov_iter_get_pages() continues filling the same bio. If such bvecs are physically contiguous, biovec_phys_mergeable() will coalesce them, making it impossible to recover the correct pgmap for the merged segment via page_pgmap().

Add a zone_device_pages_have_same_pgmap() check to prevent merging bvec segments that span different pgmaps.

---
 block/blk.h | 2 ++
 1 file changed, 2 insertions(+)

index ec4674cdf2ead..50a41db039133 100644
@@ -127,6 +127,8 @@ static inline bool biovec_phys_mergeable(struct request_queue *q,
 
 	if (addr1 + vec1->bv_len != addr2)
 		return false;
+	if (!zone_device_pages_have_same_pgmap(vec1->bv_page, vec2->bv_page))
+		return false;
 	if (xen_domain() && !xen_biovec_phys_mergeable(vec1, vec2->bv_page))
 		return false;
 	if ((addr1 | mask) != ((addr2 + vec2->bv_len - 1) | mask))

2.43.0

bio_add_page() and bio_integrity_add_page() reject pages from different dev_pagemaps entirely, returning 0 even when those pages have compatible DMA mapping requirements. This forces callers to start a new bio when buffers span pgmap boundaries, even though the pages could safely coexist as separate bvec entries.

This matters for guests where memory is registered through devm_memremap_pages() with MEMORY_DEVICE_GENERIC in multiple calls, creating separate dev_pagemaps for each chunk. When a direct I/O buffer spans two such chunks, bio_add_page() rejects the second page, forcing an unnecessary bio split or I/O failure.

Introduce zone_device_pages_compatible() in blk.h to check whether two pages can coexist in the same bio as separate bvec entries. The block DMA iterator (blk_dma_map_iter_start) caches the P2PDMA mapping state from the first segment and applies it to all others, so P2PDMA pages from different pgmaps must not be mixed, and neither must P2PDMA and non-P2PDMA pages. All other combinations (MEMORY_DEVICE_GENERIC pages from different pgmaps, or MEMORY_DEVICE_GENERIC with normal RAM) use the same dma_map_phys path and are safe.

Replace the blanket zone_device_pages_have_same_pgmap() rejection with zone_device_pages_compatible(), while keeping zone_device_pages_have_same_pgmap() as a merge guard.
Pages from different pgmaps can be added as separate bvec entries but must not be coalesced into the same segment, as that would make it impossible to recover the correct pgmap via page_pgmap().

---
 block/bio-integrity.c |  6 +++---
 block/bio.c           |  6 +++---
 block/blk.h           | 19 +++++++++++++++++++
 3 files changed, 25 insertions(+), 6 deletions(-)

index e79eaf0477943..e54c6e06e1cbb 100644
@@ -231,10 +231,10 @@ int bio_integrity_add_page(struct bio *bio, struct page *page,
 	if (bip->bip_vcnt > 0) {
 		struct bio_vec *bv = &bip->bip_vec[bip->bip_vcnt - 1];
 
-		if (!zone_device_pages_have_same_pgmap(bv->bv_page, page))
+		if (!zone_device_pages_compatible(bv->bv_page, page))
 			return 0;
-
-		if (bvec_try_merge_hw_page(q, bv, page, len, offset)) {
+		if (zone_device_pages_have_same_pgmap(bv->bv_page, page) &&
+		    bvec_try_merge_hw_page(q, bv, page, len, offset)) {
 			bip->bip_iter.bi_size += len;
 			return len;
 		}

index 641ef0928d735..c52a0bd1e8993 100644
@@ -1048,10 +1048,10 @@ int bio_add_page(struct bio *bio, struct page *page,
 	if (bio->bi_vcnt > 0) {
 		struct bio_vec *bv = &bio->bi_io_vec[bio->bi_vcnt - 1];
 
-		if (!zone_device_pages_have_same_pgmap(bv->bv_page, page))
+		if (!zone_device_pages_compatible(bv->bv_page, page))
 			return 0;
-
-		if (bvec_try_merge_page(bv, page, len, offset)) {
+		if (zone_device_pages_have_same_pgmap(bv->bv_page, page) &&
+		    bvec_try_merge_page(bv, page, len, offset)) {
 			bio->bi_iter.bi_size += len;
 			return len;
 		}

index 50a41db039133..b998a7761faf3 100644
@@ -136,6 +136,25 @@ static inline bool biovec_phys_mergeable(struct request_queue *q,
 	return true;
 }
 
+/*
+ * Check if two pages from potentially different zone device pgmaps can
+ * coexist as separate bvec entries in the same bio.
+ *
+ * The block DMA iterator (blk_dma_map_iter_start) caches the P2PDMA mapping
+ * state from the first segment and applies it to all subsequent segments, so
+ * P2PDMA pages from different pgmaps must not be mixed in the same bio.
+ *
+ * Other zone device types (FS_DAX, GENERIC) use the same dma_map_phys() path
+ * as normal RAM.  PRIVATE and COHERENT pages never appear in bios.
+ */
+static inline bool zone_device_pages_compatible(const struct page *a,
+						const struct page *b)
+{
+	if (is_pci_p2pdma_page(a) || is_pci_p2pdma_page(b))
+		return zone_device_pages_have_same_pgmap(a, b);
+	return true;
+}
+
 static inline bool __bvec_gap_to_prev(const struct queue_limits *lim,
 		struct bio_vec *bprv, unsigned int offset)
 {

2.43.0

Looks good:

Applied, thanks!

[1/2] block: add pgmap check to biovec_phys_mergeable

      commit: 13920e4b7b784b40cf4519ff1f0f3e513476a499
[2/2] block: relax pgmap check in bio_add_page for compatible zone device pages
      commit: 41c665aae2b5dbecddddcc8ace344caf630cc7a4

Best regards,

Jens Axboe