[PATCH v2 0/8] RDMA: Enable operation with DMA debug enabled

[PATCH v2 0/8] RDMA: Enable operation with DMA debug enabled

[Leon Romanovsky]

Add a new DMA_ATTR_REQUIRE_COHERENT attribute to the DMA API to mark
mappings that must run on a DMA‑coherent system. Such buffers cannot
use the SWIOTLB path, may overlap with CPU caches, and do not depend on
explicit cache flushing.

Mappings using this attribute are rejected on systems where cache
side‑effects could lead to data corruption, and therefore do not need
the cache‑overlap debugging logic. This series also includes fixes for
DMA_ATTR_CPU_CACHE_CLEAN handling.
Thanks.

---
Changes in v2:
- Added DMA_ATTR_REQUIRE_COHERENT attribute
- Added HMM patch which needs this attribute as well
- Renamed DMA_ATTR_CPU_CACHE_CLEAN to be DMA_ATTR_DEBUGGING_IGNORE_CACHELINES
- Link to v1: https://patch.msgid.link/20260307-dma-debug-overlap-v1-0-c034c38872af@nvidia.com

---
Leon Romanovsky (8):
      dma-debug: Allow multiple invocations of overlapping entries
      dma-mapping: handle DMA_ATTR_CPU_CACHE_CLEAN in trace output
      dma-mapping: Clarify valid conditions for CPU cache line overlap
      dma-mapping: Introduce DMA require coherency attribute
      dma-direct: prevent SWIOTLB path when DMA_ATTR_REQUIRE_COHERENT is set
      iommu/dma: add support for DMA_ATTR_REQUIRE_COHERENT attribute
      RDMA/umem: Tell DMA mapping that UMEM requires coherency
      mm/hmm: Indicate that HMM requires DMA coherency

 Documentation/core-api/dma-attributes.rst | 34 +++++++++++++++++++++++--------
 drivers/infiniband/core/umem.c            |  5 +++--
 drivers/iommu/dma-iommu.c                 | 21 +++++++++++++++----
 drivers/virtio/virtio_ring.c              | 10 ++++-----
 include/linux/dma-mapping.h               | 15 ++++++++++----
 include/trace/events/dma.h                |  4 +++-
 kernel/dma/debug.c                        |  9 ++++----
 kernel/dma/direct.h                       |  7 ++++---
 kernel/dma/mapping.c                      |  6 ++++++
 mm/hmm.c                                  |  4 ++--
 10 files changed, 82 insertions(+), 33 deletions(-)
---
base-commit: 11439c4635edd669ae435eec308f4ab8a0804808
change-id: 20260305-dma-debug-overlap-21487c3fa02c

Best regards,
--  
Leon Romanovsky <leonro@nvidia.com>

[PATCH v2 1/8] dma-debug: Allow multiple invocations of overlapping entries

[Leon Romanovsky]

From: Leon Romanovsky <leonro@nvidia.com>

Repeated DMA mappings with DMA_ATTR_CPU_CACHE_CLEAN trigger the
following splat. This prevents using the attribute in cases where a DMA
region is shared and reused more than seven times.

 ------------[ cut here ]------------
 DMA-API: exceeded 7 overlapping mappings of cacheline 0x000000000438c440
 WARNING: kernel/dma/debug.c:467 at add_dma_entry+0x219/0x280, CPU#4: ibv_rc_pingpong/1644
 Modules linked in: xt_conntrack xt_MASQUERADE nf_conntrack_netlink nfnetlink iptable_nat nf_nat xt_addrtype br_netfilter rpcsec_gss_krb5 auth_rpcgss oid_registry overlay mlx5_fwctl zram zsmalloc mlx5_ib fuse rpcrdma rdma_ucm ib_uverbs ib_iser libiscsi scsi_transport_iscsi ib_umad rdma_cm ib_ipoib iw_cm ib_cm mlx5_core ib_core
 CPU: 4 UID: 2733 PID: 1644 Comm: ibv_rc_pingpong Not tainted 6.19.0+ #129 PREEMPT
 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.13.0-0-gf21b5a4aeb02-prebuilt.qemu.org 04/01/2014
 RIP: 0010:add_dma_entry+0x221/0x280
 Code: c0 0f 84 f2 fe ff ff 83 e8 01 89 05 6d 99 11 01 e9 e4 fe ff ff 0f 8e 1f ff ff ff 48 8d 3d 07 ef 2d 01 be 07 00 00 00 48 89 e2 <67> 48 0f b9 3a e9 06 ff ff ff 48 c7 c7 98 05 2b 82 c6 05 72 92 28
 RSP: 0018:ff1100010e657970 EFLAGS: 00010002
 RAX: 0000000000000007 RBX: ff1100010234eb00 RCX: 0000000000000000
 RDX: ff1100010e657970 RSI: 0000000000000007 RDI: ffffffff82678660
 RBP: 000000000438c440 R08: 0000000000000228 R09: 0000000000000000
 R10: 00000000000001be R11: 000000000000089d R12: 0000000000000800
 R13: 00000000ffffffef R14: 0000000000000202 R15: ff1100010234eb00
 FS:  00007fb15f3f6740(0000) GS:ff110008dcc19000(0000) knlGS:0000000000000000
 CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
 CR2: 00007fb15f32d3a0 CR3: 0000000116f59001 CR4: 0000000000373eb0
 Call Trace:
  <TASK>
  debug_dma_map_sg+0x1b4/0x390
  __dma_map_sg_attrs+0x6d/0x1a0
  dma_map_sgtable+0x19/0x30
  ib_umem_get+0x284/0x3b0 [ib_uverbs]
  mlx5_ib_reg_user_mr+0x68/0x2a0 [mlx5_ib]
  ib_uverbs_reg_mr+0x17f/0x2a0 [ib_uverbs]
  ib_uverbs_handler_UVERBS_METHOD_INVOKE_WRITE+0xc2/0x130 [ib_uverbs]
  ib_uverbs_cmd_verbs+0xa0b/0xae0 [ib_uverbs]
  ? ib_uverbs_handler_UVERBS_METHOD_QUERY_PORT_SPEED+0xe0/0xe0 [ib_uverbs]
  ? mmap_region+0x7a/0xb0
  ? do_mmap+0x3b8/0x5c0
  ib_uverbs_ioctl+0xa7/0x110 [ib_uverbs]
  __x64_sys_ioctl+0x14f/0x8b0
  ? ksys_mmap_pgoff+0xc5/0x190
  do_syscall_64+0x8c/0xbf0
  entry_SYSCALL_64_after_hwframe+0x4b/0x53
 RIP: 0033:0x7fb15f5e4eed
 Code: 04 25 28 00 00 00 48 89 45 c8 31 c0 48 8d 45 10 c7 45 b0 10 00 00 00 48 89 45 b8 48 8d 45 d0 48 89 45 c0 b8 10 00 00 00 0f 05 <89> c2 3d 00 f0 ff ff 77 1a 48 8b 45 c8 64 48 2b 04 25 28 00 00 00
 RSP: 002b:00007ffe09a5c540 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
 RAX: ffffffffffffffda RBX: 00007ffe09a5c5d0 RCX: 00007fb15f5e4eed
 RDX: 00007ffe09a5c5f0 RSI: 00000000c0181b01 RDI: 0000000000000003
 RBP: 00007ffe09a5c590 R08: 0000000000000028 R09: 00007ffe09a5c794
 R10: 0000000000000001 R11: 0000000000000246 R12: 00007ffe09a5c794
 R13: 000000000000000c R14: 0000000025a49170 R15: 000000000000000c
  </TASK>
 ---[ end trace 0000000000000000 ]---

Fixes: 61868dc55a11 ("dma-mapping: add DMA_ATTR_CPU_CACHE_CLEAN")
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
 kernel/dma/debug.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/kernel/dma/debug.c b/kernel/dma/debug.c
index 86f87e43438c3..be207be749968 100644
--- a/kernel/dma/debug.c
+++ b/kernel/dma/debug.c
@@ -453,7 +453,7 @@ static int active_cacheline_set_overlap(phys_addr_t cln, int overlap)
 	return overlap;
 }
 
-static void active_cacheline_inc_overlap(phys_addr_t cln)
+static void active_cacheline_inc_overlap(phys_addr_t cln, bool is_cache_clean)
 {
 	int overlap = active_cacheline_read_overlap(cln);
 
@@ -462,7 +462,7 @@ static void active_cacheline_inc_overlap(phys_addr_t cln)
 	/* If we overflowed the overlap counter then we're potentially
 	 * leaking dma-mappings.
 	 */
-	WARN_ONCE(overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
+	WARN_ONCE(!is_cache_clean && overlap > ACTIVE_CACHELINE_MAX_OVERLAP,
 		  pr_fmt("exceeded %d overlapping mappings of cacheline %pa\n"),
 		  ACTIVE_CACHELINE_MAX_OVERLAP, &cln);
 }
@@ -495,7 +495,7 @@ static int active_cacheline_insert(struct dma_debug_entry *entry,
 	if (rc == -EEXIST) {
 		struct dma_debug_entry *existing;
 
-		active_cacheline_inc_overlap(cln);
+		active_cacheline_inc_overlap(cln, entry->is_cache_clean);
 		existing = radix_tree_lookup(&dma_active_cacheline, cln);
 		/* A lookup failure here after we got -EEXIST is unexpected. */
 		WARN_ON(!existing);

-- 
2.53.0

[PATCH v2 2/8] dma-mapping: handle DMA_ATTR_CPU_CACHE_CLEAN in trace output

[Leon Romanovsky]

From: Leon Romanovsky <leonro@nvidia.com>

Tracing prints decoded DMA attribute flags, but it does not yet
include the recently added DMA_ATTR_CPU_CACHE_CLEAN. Add support
for decoding and displaying this attribute in the trace output.

Fixes: 61868dc55a11 ("dma-mapping: add DMA_ATTR_CPU_CACHE_CLEAN")
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
 include/trace/events/dma.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/include/trace/events/dma.h b/include/trace/events/dma.h
index 33e99e792f1aa..69cb3805ee81c 100644
--- a/include/trace/events/dma.h
+++ b/include/trace/events/dma.h
@@ -32,7 +32,8 @@ TRACE_DEFINE_ENUM(DMA_NONE);
 		{ DMA_ATTR_ALLOC_SINGLE_PAGES, "ALLOC_SINGLE_PAGES" }, \
 		{ DMA_ATTR_NO_WARN, "NO_WARN" }, \
 		{ DMA_ATTR_PRIVILEGED, "PRIVILEGED" }, \
-		{ DMA_ATTR_MMIO, "MMIO" })
+		{ DMA_ATTR_MMIO, "MMIO" }, \
+		{ DMA_ATTR_CPU_CACHE_CLEAN, "CACHE_CLEAN" })
 
 DECLARE_EVENT_CLASS(dma_map,
 	TP_PROTO(struct device *dev, phys_addr_t phys_addr, dma_addr_t dma_addr,

-- 
2.53.0

[PATCH v2 3/8] dma-mapping: Clarify valid conditions for CPU cache line overlap

[Leon Romanovsky]

From: Leon Romanovsky <leonro@nvidia.com>

Rename the DMA_ATTR_CPU_CACHE_CLEAN attribute to better reflect that it
is debugging aid to inform DMA core code that CPU cache line overlaps are
allowed, and refine the documentation describing its use.

Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
 Documentation/core-api/dma-attributes.rst | 22 ++++++++++++++--------
 drivers/virtio/virtio_ring.c              | 10 +++++-----
 include/linux/dma-mapping.h               |  8 ++++----
 include/trace/events/dma.h                |  2 +-
 kernel/dma/debug.c                        |  2 +-
 5 files changed, 25 insertions(+), 19 deletions(-)

diff --git a/Documentation/core-api/dma-attributes.rst b/Documentation/core-api/dma-attributes.rst
index 1d7bfad73b1c7..48cfe86cc06d7 100644
--- a/Documentation/core-api/dma-attributes.rst
+++ b/Documentation/core-api/dma-attributes.rst
@@ -149,11 +149,17 @@ For architectures that require cache flushing for DMA coherence
 DMA_ATTR_MMIO will not perform any cache flushing. The address
 provided must never be mapped cacheable into the CPU.
 
-DMA_ATTR_CPU_CACHE_CLEAN
-------------------------
-
-This attribute indicates the CPU will not dirty any cacheline overlapping this
-DMA_FROM_DEVICE/DMA_BIDIRECTIONAL buffer while it is mapped. This allows
-multiple small buffers to safely share a cacheline without risk of data
-corruption, suppressing DMA debug warnings about overlapping mappings.
-All mappings sharing a cacheline should have this attribute.
+DMA_ATTR_DEBUGGING_IGNORE_CACHELINES
+------------------------------------
+
+This attribute indicates that CPU cache lines may overlap for buffers mapped
+with DMA_FROM_DEVICE or DMA_BIDIRECTIONAL.
+
+Such overlap may occur when callers map multiple small buffers that reside
+within the same cache line. In this case, callers must guarantee that the CPU
+will not dirty these cache lines after the mappings are established. When this
+condition is met, multiple buffers can safely share a cache line without risking
+data corruption.
+
+All mappings that share a cache line must set this attribute to suppress DMA
+debug warnings about overlapping mappings.
diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
index 335692d41617a..fbca7ce1c6bf0 100644
--- a/drivers/virtio/virtio_ring.c
+++ b/drivers/virtio/virtio_ring.c
@@ -2912,10 +2912,10 @@ EXPORT_SYMBOL_GPL(virtqueue_add_inbuf);
  * @data: the token identifying the buffer.
  * @gfp: how to do memory allocations (if necessary).
  *
- * Same as virtqueue_add_inbuf but passes DMA_ATTR_CPU_CACHE_CLEAN to indicate
- * that the CPU will not dirty any cacheline overlapping this buffer while it
- * is available, and to suppress overlapping cacheline warnings in DMA debug
- * builds.
+ * Same as virtqueue_add_inbuf but passes DMA_ATTR_DEBUGGING_IGNORE_CACHELINES
+ * to indicate that the CPU will not dirty any cacheline overlapping this buffer
+ * while it is available, and to suppress overlapping cacheline warnings in DMA
+ * debug builds.
  *
  * Caller must ensure we don't call this with other virtqueue operations
  * at the same time (except where noted).
@@ -2928,7 +2928,7 @@ int virtqueue_add_inbuf_cache_clean(struct virtqueue *vq,
 				    gfp_t gfp)
 {
 	return virtqueue_add(vq, &sg, num, 0, 1, data, NULL, false, gfp,
-			     DMA_ATTR_CPU_CACHE_CLEAN);
+			     DMA_ATTR_DEBUGGING_IGNORE_CACHELINES);
 }
 EXPORT_SYMBOL_GPL(virtqueue_add_inbuf_cache_clean);
 
diff --git a/include/linux/dma-mapping.h b/include/linux/dma-mapping.h
index 29973baa05816..da44394b3a1a7 100644
--- a/include/linux/dma-mapping.h
+++ b/include/linux/dma-mapping.h
@@ -80,11 +80,11 @@
 #define DMA_ATTR_MMIO		(1UL << 10)
 
 /*
- * DMA_ATTR_CPU_CACHE_CLEAN: Indicates the CPU will not dirty any cacheline
- * overlapping this buffer while it is mapped for DMA. All mappings sharing
- * a cacheline must have this attribute for this to be considered safe.
+ * DMA_ATTR_DEBUGGING_IGNORE_CACHELINES: Indicates the CPU cache line can be
+ * overlapped. All mappings sharing a cacheline must have this attribute for
+ * this to be considered safe.
  */
-#define DMA_ATTR_CPU_CACHE_CLEAN	(1UL << 11)
+#define DMA_ATTR_DEBUGGING_IGNORE_CACHELINES	(1UL << 11)
 
 /*
  * A dma_addr_t can hold any valid DMA or bus address for the platform.  It can
diff --git a/include/trace/events/dma.h b/include/trace/events/dma.h
index 69cb3805ee81c..8c64bc0721fe4 100644
--- a/include/trace/events/dma.h
+++ b/include/trace/events/dma.h
@@ -33,7 +33,7 @@ TRACE_DEFINE_ENUM(DMA_NONE);
 		{ DMA_ATTR_NO_WARN, "NO_WARN" }, \
 		{ DMA_ATTR_PRIVILEGED, "PRIVILEGED" }, \
 		{ DMA_ATTR_MMIO, "MMIO" }, \
-		{ DMA_ATTR_CPU_CACHE_CLEAN, "CACHE_CLEAN" })
+		{ DMA_ATTR_DEBUGGING_IGNORE_CACHELINES, "CACHELINES_OVERLAP" })
 
 DECLARE_EVENT_CLASS(dma_map,
 	TP_PROTO(struct device *dev, phys_addr_t phys_addr, dma_addr_t dma_addr,
diff --git a/kernel/dma/debug.c b/kernel/dma/debug.c
index be207be749968..83e1cfe05f08d 100644
--- a/kernel/dma/debug.c
+++ b/kernel/dma/debug.c
@@ -601,7 +601,7 @@ static void add_dma_entry(struct dma_debug_entry *entry, unsigned long attrs)
 	unsigned long flags;
 	int rc;
 
-	entry->is_cache_clean = !!(attrs & DMA_ATTR_CPU_CACHE_CLEAN);
+	entry->is_cache_clean = attrs & DMA_ATTR_DEBUGGING_IGNORE_CACHELINES;
 
 	bucket = get_hash_bucket(entry, &flags);
 	hash_bucket_add(bucket, entry);

-- 
2.53.0

[PATCH v2 4/8] dma-mapping: Introduce DMA require coherency attribute

[Leon Romanovsky]

From: Leon Romanovsky <leonro@nvidia.com>

The mapping buffers which carry this attribute require DMA coherent system.
This means that they can't take SWIOTLB path, can perform CPU cache overlap
and doesn't perform cache flushing.

Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
 Documentation/core-api/dma-attributes.rst | 12 ++++++++++++
 include/linux/dma-mapping.h               |  7 +++++++
 include/trace/events/dma.h                |  3 ++-
 kernel/dma/debug.c                        |  3 ++-
 kernel/dma/mapping.c                      |  6 ++++++
 5 files changed, 29 insertions(+), 2 deletions(-)

diff --git a/Documentation/core-api/dma-attributes.rst b/Documentation/core-api/dma-attributes.rst
index 48cfe86cc06d7..69d094f144c70 100644
--- a/Documentation/core-api/dma-attributes.rst
+++ b/Documentation/core-api/dma-attributes.rst
@@ -163,3 +163,15 @@ data corruption.
 
 All mappings that share a cache line must set this attribute to suppress DMA
 debug warnings about overlapping mappings.
+
+DMA_ATTR_REQUIRE_COHERENT
+-------------------------
+
+The mapping buffers which carry this attribute require DMA coherent system. This means
+that they can't take SWIOTLB path, can perform CPU cache overlap and doesn't perform
+cache flushing.
+
+If the mapping has this attribute then it is prevented from running on systems
+where these cache artifacts can cause corruption, and as such doesn't need
+cache overlapping debugging code (same behavior as for
+DMA_ATTR_DEBUGGING_IGNORE_CACHELINES).
diff --git a/include/linux/dma-mapping.h b/include/linux/dma-mapping.h
index da44394b3a1a7..482b919f040f7 100644
--- a/include/linux/dma-mapping.h
+++ b/include/linux/dma-mapping.h
@@ -86,6 +86,13 @@
  */
 #define DMA_ATTR_DEBUGGING_IGNORE_CACHELINES	(1UL << 11)
 
+/*
+ * DMA_ATTR_REQUIRE_COHERENT: Indicates that DMA coherency is required.
+ * All mappings that carry this attribute can't work with SWIOTLB and cache
+ * flushing.
+ */
+#define DMA_ATTR_REQUIRE_COHERENT	(1UL << 12)
+
 /*
  * A dma_addr_t can hold any valid DMA or bus address for the platform.  It can
  * be given to a device to use as a DMA source or target.  It is specific to a
diff --git a/include/trace/events/dma.h b/include/trace/events/dma.h
index 8c64bc0721fe4..63597b0044247 100644
--- a/include/trace/events/dma.h
+++ b/include/trace/events/dma.h
@@ -33,7 +33,8 @@ TRACE_DEFINE_ENUM(DMA_NONE);
 		{ DMA_ATTR_NO_WARN, "NO_WARN" }, \
 		{ DMA_ATTR_PRIVILEGED, "PRIVILEGED" }, \
 		{ DMA_ATTR_MMIO, "MMIO" }, \
-		{ DMA_ATTR_DEBUGGING_IGNORE_CACHELINES, "CACHELINES_OVERLAP" })
+		{ DMA_ATTR_DEBUGGING_IGNORE_CACHELINES, "CACHELINES_OVERLAP" }, \
+		{ DMA_ATTR_REQUIRE_COHERENT, "REQUIRE_COHERENT" })
 
 DECLARE_EVENT_CLASS(dma_map,
 	TP_PROTO(struct device *dev, phys_addr_t phys_addr, dma_addr_t dma_addr,
diff --git a/kernel/dma/debug.c b/kernel/dma/debug.c
index 83e1cfe05f08d..0677918f06a80 100644
--- a/kernel/dma/debug.c
+++ b/kernel/dma/debug.c
@@ -601,7 +601,8 @@ static void add_dma_entry(struct dma_debug_entry *entry, unsigned long attrs)
 	unsigned long flags;
 	int rc;
 
-	entry->is_cache_clean = attrs & DMA_ATTR_DEBUGGING_IGNORE_CACHELINES;
+	entry->is_cache_clean = attrs & (DMA_ATTR_DEBUGGING_IGNORE_CACHELINES |
+					 DMA_ATTR_REQUIRE_COHERENT);
 
 	bucket = get_hash_bucket(entry, &flags);
 	hash_bucket_add(bucket, entry);
diff --git a/kernel/dma/mapping.c b/kernel/dma/mapping.c
index 3928a509c44c2..6d3dd0bd3a886 100644
--- a/kernel/dma/mapping.c
+++ b/kernel/dma/mapping.c
@@ -164,6 +164,9 @@ dma_addr_t dma_map_phys(struct device *dev, phys_addr_t phys, size_t size,
 	if (WARN_ON_ONCE(!dev->dma_mask))
 		return DMA_MAPPING_ERROR;
 
+	if (!dev_is_dma_coherent(dev) && (attrs & DMA_ATTR_REQUIRE_COHERENT))
+		return DMA_MAPPING_ERROR;
+
 	if (dma_map_direct(dev, ops) ||
 	    (!is_mmio && arch_dma_map_phys_direct(dev, phys + size)))
 		addr = dma_direct_map_phys(dev, phys, size, dir, attrs);
@@ -235,6 +238,9 @@ static int __dma_map_sg_attrs(struct device *dev, struct scatterlist *sg,
 
 	BUG_ON(!valid_dma_direction(dir));
 
+	if (!dev_is_dma_coherent(dev) && (attrs & DMA_ATTR_REQUIRE_COHERENT))
+		return -EOPNOTSUPP;
+
 	if (WARN_ON_ONCE(!dev->dma_mask))
 		return 0;
 

-- 
2.53.0

[PATCH v2 5/8] dma-direct: prevent SWIOTLB path when DMA_ATTR_REQUIRE_COHERENT is set

[Leon Romanovsky]

From: Leon Romanovsky <leonro@nvidia.com>

DMA_ATTR_REQUIRE_COHERENT indicates that SWIOTLB must not be used.
Ensure the SWIOTLB path is declined whenever the DMA direct path is
selected.

Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
 kernel/dma/direct.h | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/kernel/dma/direct.h b/kernel/dma/direct.h
index e89f175e9c2d0..6184ff303f080 100644
--- a/kernel/dma/direct.h
+++ b/kernel/dma/direct.h
@@ -84,7 +84,7 @@ static inline dma_addr_t dma_direct_map_phys(struct device *dev,
 	dma_addr_t dma_addr;
 
 	if (is_swiotlb_force_bounce(dev)) {
-		if (attrs & DMA_ATTR_MMIO)
+		if (attrs & (DMA_ATTR_MMIO | DMA_ATTR_REQUIRE_COHERENT))
 			return DMA_MAPPING_ERROR;
 
 		return swiotlb_map(dev, phys, size, dir, attrs);
@@ -98,7 +98,8 @@ static inline dma_addr_t dma_direct_map_phys(struct device *dev,
 		dma_addr = phys_to_dma(dev, phys);
 		if (unlikely(!dma_capable(dev, dma_addr, size, true)) ||
 		    dma_kmalloc_needs_bounce(dev, size, dir)) {
-			if (is_swiotlb_active(dev))
+			if (is_swiotlb_active(dev) &&
+			    !(attrs & DMA_ATTR_REQUIRE_COHERENT))
 				return swiotlb_map(dev, phys, size, dir, attrs);
 
 			goto err_overflow;
@@ -123,7 +124,7 @@ static inline void dma_direct_unmap_phys(struct device *dev, dma_addr_t addr,
 {
 	phys_addr_t phys;
 
-	if (attrs & DMA_ATTR_MMIO)
+	if (attrs & (DMA_ATTR_MMIO | DMA_ATTR_REQUIRE_COHERENT))
 		/* nothing to do: uncached and no swiotlb */
 		return;
 

-- 
2.53.0

[PATCH v2 6/8] iommu/dma: add support for DMA_ATTR_REQUIRE_COHERENT attribute

[Leon Romanovsky]

From: Leon Romanovsky <leonro@nvidia.com>

Add support for the DMA_ATTR_REQUIRE_COHERENT attribute to the exported
functions. This attribute indicates that the SWIOTLB path must not be
used and that no sync operations should be performed.

Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
 drivers/iommu/dma-iommu.c | 21 +++++++++++++++++----
 1 file changed, 17 insertions(+), 4 deletions(-)

diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
index 5dac64be61bb2..94d5141696424 100644
--- a/drivers/iommu/dma-iommu.c
+++ b/drivers/iommu/dma-iommu.c
@@ -1211,7 +1211,7 @@ dma_addr_t iommu_dma_map_phys(struct device *dev, phys_addr_t phys, size_t size,
 	 */
 	if (dev_use_swiotlb(dev, size, dir) &&
 	    iova_unaligned(iovad, phys, size)) {
-		if (attrs & DMA_ATTR_MMIO)
+		if (attrs & (DMA_ATTR_MMIO | DMA_ATTR_REQUIRE_COHERENT))
 			return DMA_MAPPING_ERROR;
 
 		phys = iommu_dma_map_swiotlb(dev, phys, size, dir, attrs);
@@ -1223,7 +1223,8 @@ dma_addr_t iommu_dma_map_phys(struct device *dev, phys_addr_t phys, size_t size,
 		arch_sync_dma_for_device(phys, size, dir);
 
 	iova = __iommu_dma_map(dev, phys, size, prot, dma_mask);
-	if (iova == DMA_MAPPING_ERROR && !(attrs & DMA_ATTR_MMIO))
+	if (iova == DMA_MAPPING_ERROR &&
+	    !(attrs & (DMA_ATTR_MMIO | DMA_ATTR_REQUIRE_COHERENT)))
 		swiotlb_tbl_unmap_single(dev, phys, size, dir, attrs);
 	return iova;
 }
@@ -1233,7 +1234,7 @@ void iommu_dma_unmap_phys(struct device *dev, dma_addr_t dma_handle,
 {
 	phys_addr_t phys;
 
-	if (attrs & DMA_ATTR_MMIO) {
+	if (attrs & (DMA_ATTR_MMIO | DMA_ATTR_REQUIRE_COHERENT)) {
 		__iommu_dma_unmap(dev, dma_handle, size);
 		return;
 	}
@@ -1945,9 +1946,21 @@ int dma_iova_link(struct device *dev, struct dma_iova_state *state,
 	if (WARN_ON_ONCE(iova_start_pad && offset > 0))
 		return -EIO;
 
+	/*
+	 * DMA_IOVA_USE_SWIOTLB is set on state after some entry
+	 * took SWIOTLB path, which we were supposed to prevent
+	 * for DMA_ATTR_REQUIRE_COHERENT attribute.
+	 */
+	if (WARN_ON_ONCE((state->__size & DMA_IOVA_USE_SWIOTLB) &&
+			 (attrs & DMA_ATTR_REQUIRE_COHERENT)))
+		return -EOPNOTSUPP;
+
+	if (!dev_is_dma_coherent(dev) && (attrs & DMA_ATTR_REQUIRE_COHERENT))
+		return -EOPNOTSUPP;
+
 	if (dev_use_swiotlb(dev, size, dir) &&
 	    iova_unaligned(iovad, phys, size)) {
-		if (attrs & DMA_ATTR_MMIO)
+		if (attrs & (DMA_ATTR_MMIO | DMA_ATTR_REQUIRE_COHERENT))
 			return -EPERM;
 
 		return iommu_dma_iova_link_swiotlb(dev, state, phys, offset,

-- 
2.53.0

[PATCH v2 7/8] RDMA/umem: Tell DMA mapping that UMEM requires coherency

[Leon Romanovsky]

From: Leon Romanovsky <leonro@nvidia.com>

The RDMA subsystem exposes DMA regions through the verbs interface, which
assumes a coherent system. Use the DMA_ATTR_REQUIRE_COHERENCE attribute to
ensure coherency and avoid taking the SWIOTLB path.

In addition, a given region may be exported multiple times, which will trigger
warnings about cacheline overlaps. These warnings are suppressed when using
the new attribute.

infiniband rocep8s0f0: mlx5_ib_reg_user_mr:1592:(pid 5812): start 0x2b28c000, iova 0x2b28c000, length 0x1000, access_flags 0x1
infiniband rocep8s0f0: mlx5_ib_reg_user_mr:1592:(pid 5812): start 0x2b28c001, iova 0x2b28c001, length 0xfff, access_flags 0x1
 ------------[ cut here ]------------
 DMA-API: mlx5_core 0000:08:00.0: cacheline tracking EEXIST, overlapping mappings aren't supported
 WARNING: kernel/dma/debug.c:620 at add_dma_entry+0x1bb/0x280, CPU#6: ibv_rc_pingpong/5812
 Modules linked in: veth xt_conntrack xt_MASQUERADE nf_conntrack_netlink nfnetlink iptable_nat nf_nat xt_addrtype br_netfilter rpcsec_gss_krb5 auth_rpcgss oid_registry overlay mlx5_fwctl zram zsmalloc mlx5_ib fuse rpcrdma rdma_ucm ib_uverbs ib_iser libiscsi scsi_transport_iscsi ib_umad rdma_cm ib_ipoib iw_cm ib_cm mlx5_core ib_core
 CPU: 6 UID: 2733 PID: 5812 Comm: ibv_rc_pingpong Tainted: G        W           6.19.0+ #129 PREEMPT
 Tainted: [W]=WARN
 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.13.0-0-gf21b5a4aeb02-prebuilt.qemu.org 04/01/2014
 RIP: 0010:add_dma_entry+0x1be/0x280
 Code: 8b 7b 10 48 85 ff 0f 84 c3 00 00 00 48 8b 6f 50 48 85 ed 75 03 48 8b 2f e8 ff 8e 6a 00 48 89 c6 48 8d 3d 55 ef 2d 01 48 89 ea <67> 48 0f b9 3a 48 85 db 74 1a 48 c7 c7 b0 00 2b 82 e8 9c 25 fd ff
 RSP: 0018:ff11000138717978 EFLAGS: 00010286
 RAX: ffffffffa02d7831 RBX: ff1100010246de00 RCX: 0000000000000000
 RDX: ff110001036fac30 RSI: ffffffffa02d7831 RDI: ffffffff82678650
 RBP: ff110001036fac30 R08: ff11000110dcb4a0 R09: ff11000110dcb478
 R10: 0000000000000000 R11: ffffffff824b30a8 R12: 0000000000000000
 R13: 00000000ffffffef R14: 0000000000000202 R15: ff1100010246de00
 FS:  00007f59b411c740(0000) GS:ff110008dcc99000(0000) knlGS:0000000000000000
 CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
 CR2: 00007ffe538f7000 CR3: 000000010e066005 CR4: 0000000000373eb0
 Call Trace:
  <TASK>
  debug_dma_map_sg+0x1b4/0x390
  __dma_map_sg_attrs+0x6d/0x1a0
  dma_map_sgtable+0x19/0x30
  ib_umem_get+0x254/0x380 [ib_uverbs]
  mlx5_ib_reg_user_mr+0x68/0x2a0 [mlx5_ib]
  ib_uverbs_reg_mr+0x17f/0x2a0 [ib_uverbs]
  ib_uverbs_handler_UVERBS_METHOD_INVOKE_WRITE+0xc2/0x130 [ib_uverbs]
  ib_uverbs_cmd_verbs+0xa0b/0xae0 [ib_uverbs]
  ? ib_uverbs_handler_UVERBS_METHOD_QUERY_PORT_SPEED+0xe0/0xe0 [ib_uverbs]
  ? mmap_region+0x7a/0xb0
  ? do_mmap+0x3b8/0x5c0
  ib_uverbs_ioctl+0xa7/0x110 [ib_uverbs]
  __x64_sys_ioctl+0x14f/0x8b0
  ? ksys_mmap_pgoff+0xc5/0x190
  do_syscall_64+0x8c/0xbf0
  entry_SYSCALL_64_after_hwframe+0x4b/0x53
 RIP: 0033:0x7f59b430aeed
 Code: 04 25 28 00 00 00 48 89 45 c8 31 c0 48 8d 45 10 c7 45 b0 10 00 00 00 48 89 45 b8 48 8d 45 d0 48 89 45 c0 b8 10 00 00 00 0f 05 <89> c2 3d 00 f0 ff ff 77 1a 48 8b 45 c8 64 48 2b 04 25 28 00 00 00
 RSP: 002b:00007ffe538f9430 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
 RAX: ffffffffffffffda RBX: 00007ffe538f94c0 RCX: 00007f59b430aeed
 RDX: 00007ffe538f94e0 RSI: 00000000c0181b01 RDI: 0000000000000003
 RBP: 00007ffe538f9480 R08: 0000000000000028 R09: 00007ffe538f9684
 R10: 0000000000000001 R11: 0000000000000246 R12: 00007ffe538f9684
 R13: 000000000000000c R14: 000000002b28d170 R15: 000000000000000c
  </TASK>
 ---[ end trace 0000000000000000 ]---

Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
 drivers/infiniband/core/umem.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/drivers/infiniband/core/umem.c b/drivers/infiniband/core/umem.c
index cff4fcca2c345..edc34c69f0f23 100644
--- a/drivers/infiniband/core/umem.c
+++ b/drivers/infiniband/core/umem.c
@@ -55,7 +55,8 @@ static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int d
 
 	if (dirty)
 		ib_dma_unmap_sgtable_attrs(dev, &umem->sgt_append.sgt,
-					   DMA_BIDIRECTIONAL, 0);
+					   DMA_BIDIRECTIONAL,
+					   DMA_ATTR_REQUIRE_COHERENT);
 
 	for_each_sgtable_sg(&umem->sgt_append.sgt, sg, i) {
 		unpin_user_page_range_dirty_lock(sg_page(sg),
@@ -169,7 +170,7 @@ struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 	unsigned long lock_limit;
 	unsigned long new_pinned;
 	unsigned long cur_base;
-	unsigned long dma_attr = 0;
+	unsigned long dma_attr = DMA_ATTR_REQUIRE_COHERENT;
 	struct mm_struct *mm;
 	unsigned long npages;
 	int pinned, ret;

-- 
2.53.0

[PATCH v2 8/8] mm/hmm: Indicate that HMM requires DMA coherency

[Leon Romanovsky]

From: Leon Romanovsky <leonro@nvidia.com>

HMM mirroring can work on coherent systems without SWIOTLB path only.
Until introduction of DMA_ATTR_REQUIRE_COHERENT, there was no reliable
way to indicate that and various approximation was done:

int hmm_dma_map_alloc(struct device *dev, struct hmm_dma_map *map,
                      size_t nr_entries, size_t dma_entry_size)
{
<...>
        /*
         * The HMM API violates our normal DMA buffer ownership rules and can't
         * transfer buffer ownership.  The dma_addressing_limited() check is a
         * best approximation to ensure no swiotlb buffering happens.
         */
        dma_need_sync = !dev->dma_skip_sync;
        if (dma_need_sync || dma_addressing_limited(dev))
                return -EOPNOTSUPP;

So let's mark mapped buffers with DMA_ATTR_REQUIRE_COHERENT attribute
to prevent DMA debugging warnings for cache overlapped entries.

Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
 mm/hmm.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/mm/hmm.c b/mm/hmm.c
index f6c4ddff4bd61..5955f2f0c83db 100644
--- a/mm/hmm.c
+++ b/mm/hmm.c
@@ -778,7 +778,7 @@ dma_addr_t hmm_dma_map_pfn(struct device *dev, struct hmm_dma_map *map,
 	struct page *page = hmm_pfn_to_page(pfns[idx]);
 	phys_addr_t paddr = hmm_pfn_to_phys(pfns[idx]);
 	size_t offset = idx * map->dma_entry_size;
-	unsigned long attrs = 0;
+	unsigned long attrs = DMA_ATTR_REQUIRE_COHERENT;
 	dma_addr_t dma_addr;
 	int ret;
 
@@ -871,7 +871,7 @@ bool hmm_dma_unmap_pfn(struct device *dev, struct hmm_dma_map *map, size_t idx)
 	struct dma_iova_state *state = &map->state;
 	dma_addr_t *dma_addrs = map->dma_list;
 	unsigned long *pfns = map->pfn_list;
-	unsigned long attrs = 0;
+	unsigned long attrs = DMA_ATTR_REQUIRE_COHERENT;
 
 	if ((pfns[idx] & valid_dma) != valid_dma)
 		return false;

-- 
2.53.0

Re: [PATCH v2 4/8] dma-mapping: Introduce DMA require coherency attribute

[Jason Gunthorpe]

On Wed, Mar 11, 2026 at 09:08:47PM +0200, Leon Romanovsky wrote:
> From: Leon Romanovsky <leonro@nvidia.com>
> 
> The mapping buffers which carry this attribute require DMA coherent system.
> This means that they can't take SWIOTLB path, can perform CPU cache overlap
> and doesn't perform cache flushing.
> 
> Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
> ---
>  Documentation/core-api/dma-attributes.rst | 12 ++++++++++++
>  include/linux/dma-mapping.h               |  7 +++++++
>  include/trace/events/dma.h                |  3 ++-
>  kernel/dma/debug.c                        |  3 ++-
>  kernel/dma/mapping.c                      |  6 ++++++
>  5 files changed, 29 insertions(+), 2 deletions(-)
> 
> diff --git a/Documentation/core-api/dma-attributes.rst b/Documentation/core-api/dma-attributes.rst
> index 48cfe86cc06d7..69d094f144c70 100644
> --- a/Documentation/core-api/dma-attributes.rst
> +++ b/Documentation/core-api/dma-attributes.rst
> @@ -163,3 +163,15 @@ data corruption.
>  
>  All mappings that share a cache line must set this attribute to suppress DMA
>  debug warnings about overlapping mappings.
> +
> +DMA_ATTR_REQUIRE_COHERENT
> +-------------------------
> +
> +The mapping buffers which carry this attribute require DMA coherent system. This means
> +that they can't take SWIOTLB path, can perform CPU cache overlap and doesn't perform
> +cache flushing.

DMA mapping requests with the DMA_ATTR_REQUIRE_COHERENT fail on any
system where SWIOTLB or cache management is required. This should only
be used to support uAPI designs that require continuous HW DMA
coherence with userspace processes, for example RDMA and DRM. At a
minimum the memory being mapped must be userspace memory from
pin_user_pages() or similar.

Drivers should consider using dma_mmap_pages() instead of this
interface when building their uAPIs, when possible.

It must never be used in an in-kernel driver that only works with
kernal memory.

> @@ -164,6 +164,9 @@ dma_addr_t dma_map_phys(struct device *dev, phys_addr_t phys, size_t size,
>  	if (WARN_ON_ONCE(!dev->dma_mask))
>  		return DMA_MAPPING_ERROR;
>  
> +	if (!dev_is_dma_coherent(dev) && (attrs & DMA_ATTR_REQUIRE_COHERENT))
> +		return DMA_MAPPING_ERROR;

This doesn't capture enough conditions.. is_swiotlb_force_bounce(),
dma_kmalloc_needs_bounce(), dma_capable(), etc all need to be blocked
too

So check it inside swiotlb_map() too, and maybe shift the above
into the existing branches:

        if (!dev_is_dma_coherent(dev) &&
            !(attrs & (DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_MMIO)))
                arch_sync_dma_for_device(phys, size, dir);

Jason

Re: [PATCH v2 5/8] dma-direct: prevent SWIOTLB path when DMA_ATTR_REQUIRE_COHERENT is set

[Jason Gunthorpe]

On Wed, Mar 11, 2026 at 09:08:48PM +0200, Leon Romanovsky wrote:
> From: Leon Romanovsky <leonro@nvidia.com>
> 
> DMA_ATTR_REQUIRE_COHERENT indicates that SWIOTLB must not be used.
> Ensure the SWIOTLB path is declined whenever the DMA direct path is
> selected.
> 
> Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
> ---
>  kernel/dma/direct.h | 7 ++++---
>  1 file changed, 4 insertions(+), 3 deletions(-)
> 
> diff --git a/kernel/dma/direct.h b/kernel/dma/direct.h
> index e89f175e9c2d0..6184ff303f080 100644
> --- a/kernel/dma/direct.h
> +++ b/kernel/dma/direct.h
> @@ -84,7 +84,7 @@ static inline dma_addr_t dma_direct_map_phys(struct device *dev,
>  	dma_addr_t dma_addr;
>  
>  	if (is_swiotlb_force_bounce(dev)) {
> -		if (attrs & DMA_ATTR_MMIO)
> +		if (attrs & (DMA_ATTR_MMIO | DMA_ATTR_REQUIRE_COHERENT))
>  			return DMA_MAPPING_ERROR;
>  
>  		return swiotlb_map(dev, phys, size, dir, attrs);

Oh here it is, still maybe it is better to put it in swiotlb_map() ?

Jason

Re: [PATCH v2 7/8] RDMA/umem: Tell DMA mapping that UMEM requires coherency

[Jason Gunthorpe]

On Wed, Mar 11, 2026 at 09:08:50PM +0200, Leon Romanovsky wrote:
> From: Leon Romanovsky <leonro@nvidia.com>
> 
> The RDMA subsystem exposes DMA regions through the verbs interface, which
> assumes a coherent system. Use the DMA_ATTR_REQUIRE_COHERENCE attribute to
> ensure coherency and avoid taking the SWIOTLB path.
> 
> In addition, a given region may be exported multiple times, which will trigger
> warnings about cacheline overlaps. These warnings are suppressed when using
> the new attribute.
 
> Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
> ---
>  drivers/infiniband/core/umem.c | 5 +++--
>  1 file changed, 3 insertions(+), 2 deletions(-)

Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>

Jason

Re: [PATCH v2 8/8] mm/hmm: Indicate that HMM requires DMA coherency

[Jason Gunthorpe]

On Wed, Mar 11, 2026 at 09:08:51PM +0200, Leon Romanovsky wrote:
> From: Leon Romanovsky <leonro@nvidia.com>
> 
> HMM mirroring can work on coherent systems without SWIOTLB path only.
> Until introduction of DMA_ATTR_REQUIRE_COHERENT, there was no reliable
> way to indicate that and various approximation was done:

HMM is fundamentally about allowing a sophisticated device to
independently DMA to a process's memory concurrently with the CPU
accessing the same memory. It is similar to SVA but does not rely on
IOMMU support. Since the entire use model is concurrent access to the
same memory it becomes fatally broken as a uAPI if SWIOTLB is
replacing the memory, or the CPU caches are incoherent with DMA.

Till now there was no reliable way to indicate that and various
approximation was done:

> int hmm_dma_map_alloc(struct device *dev, struct hmm_dma_map *map,
>                       size_t nr_entries, size_t dma_entry_size)
> {
> <...>
>         /*
>          * The HMM API violates our normal DMA buffer ownership rules and can't
>          * transfer buffer ownership.  The dma_addressing_limited() check is a
>          * best approximation to ensure no swiotlb buffering happens.
>          */
>         dma_need_sync = !dev->dma_skip_sync;
>         if (dma_need_sync || dma_addressing_limited(dev))
>                 return -EOPNOTSUPP;

Can it get dropped now then?

> So let's mark mapped buffers with DMA_ATTR_REQUIRE_COHERENT attribute
> to prevent DMA debugging warnings for cache overlapped entries.

Well, that isn't the main motivation, this prevents silent data
corruption if someone tries to use hmm in a system with swiotlb or
incoherent DMA,

Looks OK otherwise

Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>

Jason

Re: [PATCH v2 7/8] RDMA/umem: Tell DMA mapping that UMEM requires coherency

[Jason Gunthorpe]

On Wed, Mar 11, 2026 at 09:08:50PM +0200, Leon Romanovsky wrote:
> From: Leon Romanovsky <leonro@nvidia.com>
> 
> The RDMA subsystem exposes DMA regions through the verbs interface, which
> assumes a coherent system. Use the DMA_ATTR_REQUIRE_COHERENCE attribute to
> ensure coherency and avoid taking the SWIOTLB path.

Lets elaborate a bit more so people understand why verbs is like
this:

The RDMA verbs programming model is like HMM and assumes concurrent DMA and
CPU access to userspace memory in a process. The HW device and
programming model has so-called "one-sided" operations which are
initiated over the network by a remote CPU without notification or
involvement of the local CPU. These include things like ATOMIC
compare/swap, READ, and WRITE. Using these operations a remote CPU can
traverse data structures, form locks, and so on without awareness of
the host CPU. Having SWIOTLB substitute the memory or the DMA be cache
incoherent completely breaks these use cases.

RDMA in-kernel is OK with incoherence because none of the kernel use
cases make use of one-sided operations that would cause problems.

Jason

Re: [PATCH v2 4/8] dma-mapping: Introduce DMA require coherency attribute

[Leon Romanovsky]

On Thu, Mar 12, 2026 at 09:19:37AM -0300, Jason Gunthorpe wrote:
> On Wed, Mar 11, 2026 at 09:08:47PM +0200, Leon Romanovsky wrote:
> > From: Leon Romanovsky <leonro@nvidia.com>
> > 
> > The mapping buffers which carry this attribute require DMA coherent system.
> > This means that they can't take SWIOTLB path, can perform CPU cache overlap
> > and doesn't perform cache flushing.
> > 
> > Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
> > ---
> >  Documentation/core-api/dma-attributes.rst | 12 ++++++++++++
> >  include/linux/dma-mapping.h               |  7 +++++++
> >  include/trace/events/dma.h                |  3 ++-
> >  kernel/dma/debug.c                        |  3 ++-
> >  kernel/dma/mapping.c                      |  6 ++++++
> >  5 files changed, 29 insertions(+), 2 deletions(-)
> > 
> > diff --git a/Documentation/core-api/dma-attributes.rst b/Documentation/core-api/dma-attributes.rst
> > index 48cfe86cc06d7..69d094f144c70 100644
> > --- a/Documentation/core-api/dma-attributes.rst
> > +++ b/Documentation/core-api/dma-attributes.rst
> > @@ -163,3 +163,15 @@ data corruption.
> >  
> >  All mappings that share a cache line must set this attribute to suppress DMA
> >  debug warnings about overlapping mappings.
> > +
> > +DMA_ATTR_REQUIRE_COHERENT
> > +-------------------------
> > +
> > +The mapping buffers which carry this attribute require DMA coherent system. This means
> > +that they can't take SWIOTLB path, can perform CPU cache overlap and doesn't perform
> > +cache flushing.
> 
> DMA mapping requests with the DMA_ATTR_REQUIRE_COHERENT fail on any
> system where SWIOTLB or cache management is required. This should only
> be used to support uAPI designs that require continuous HW DMA
> coherence with userspace processes, for example RDMA and DRM. At a
> minimum the memory being mapped must be userspace memory from
> pin_user_pages() or similar.
> 
> Drivers should consider using dma_mmap_pages() instead of this
> interface when building their uAPIs, when possible.
> 
> It must never be used in an in-kernel driver that only works with
> kernal memory.
> 
> > @@ -164,6 +164,9 @@ dma_addr_t dma_map_phys(struct device *dev, phys_addr_t phys, size_t size,
> >  	if (WARN_ON_ONCE(!dev->dma_mask))
> >  		return DMA_MAPPING_ERROR;
> >  
> > +	if (!dev_is_dma_coherent(dev) && (attrs & DMA_ATTR_REQUIRE_COHERENT))
> > +		return DMA_MAPPING_ERROR;
> 
> This doesn't capture enough conditions.. is_swiotlb_force_bounce(),
> dma_kmalloc_needs_bounce(), dma_capable(), etc all need to be blocked
> too

These checks exist in dma_direct_map_phys() and here is the common check
between direct and IOMMU modes.

Thanks

> 
> So check it inside swiotlb_map() too, and maybe shift the above
> into the existing branches:
> 
>         if (!dev_is_dma_coherent(dev) &&
>             !(attrs & (DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_MMIO)))
>                 arch_sync_dma_for_device(phys, size, dir);
> 
> Jason

Re: [PATCH v2 5/8] dma-direct: prevent SWIOTLB path when DMA_ATTR_REQUIRE_COHERENT is set

[Leon Romanovsky]

On Thu, Mar 12, 2026 at 09:20:58AM -0300, Jason Gunthorpe wrote:
> On Wed, Mar 11, 2026 at 09:08:48PM +0200, Leon Romanovsky wrote:
> > From: Leon Romanovsky <leonro@nvidia.com>
> > 
> > DMA_ATTR_REQUIRE_COHERENT indicates that SWIOTLB must not be used.
> > Ensure the SWIOTLB path is declined whenever the DMA direct path is
> > selected.
> > 
> > Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
> > ---
> >  kernel/dma/direct.h | 7 ++++---
> >  1 file changed, 4 insertions(+), 3 deletions(-)
> > 
> > diff --git a/kernel/dma/direct.h b/kernel/dma/direct.h
> > index e89f175e9c2d0..6184ff303f080 100644
> > --- a/kernel/dma/direct.h
> > +++ b/kernel/dma/direct.h
> > @@ -84,7 +84,7 @@ static inline dma_addr_t dma_direct_map_phys(struct device *dev,
> >  	dma_addr_t dma_addr;
> >  
> >  	if (is_swiotlb_force_bounce(dev)) {
> > -		if (attrs & DMA_ATTR_MMIO)
> > +		if (attrs & (DMA_ATTR_MMIO | DMA_ATTR_REQUIRE_COHERENT))
> >  			return DMA_MAPPING_ERROR;
> >  
> >  		return swiotlb_map(dev, phys, size, dir, attrs);
> 
> Oh here it is, still maybe it is better to put it in swiotlb_map() ?

It is better do not call function which is not going to work. We have
same flow for DMA_ATTR_REQUIRE_COHERENT and DMA_ATTR_MMIO, which both
don't work with SWIOTLB.

Thanks

> 
> Jason
> 

Re: [PATCH v2 8/8] mm/hmm: Indicate that HMM requires DMA coherency

[Leon Romanovsky]

On Thu, Mar 12, 2026 at 09:26:45AM -0300, Jason Gunthorpe wrote:
> On Wed, Mar 11, 2026 at 09:08:51PM +0200, Leon Romanovsky wrote:
> > From: Leon Romanovsky <leonro@nvidia.com>
> > 
> > HMM mirroring can work on coherent systems without SWIOTLB path only.
> > Until introduction of DMA_ATTR_REQUIRE_COHERENT, there was no reliable
> > way to indicate that and various approximation was done:
> 
> HMM is fundamentally about allowing a sophisticated device to
> independently DMA to a process's memory concurrently with the CPU
> accessing the same memory. It is similar to SVA but does not rely on
> IOMMU support. Since the entire use model is concurrent access to the
> same memory it becomes fatally broken as a uAPI if SWIOTLB is
> replacing the memory, or the CPU caches are incoherent with DMA.
> 
> Till now there was no reliable way to indicate that and various
> approximation was done:
> 
> > int hmm_dma_map_alloc(struct device *dev, struct hmm_dma_map *map,
> >                       size_t nr_entries, size_t dma_entry_size)
> > {
> > <...>
> >         /*
> >          * The HMM API violates our normal DMA buffer ownership rules and can't
> >          * transfer buffer ownership.  The dma_addressing_limited() check is a
> >          * best approximation to ensure no swiotlb buffering happens.
> >          */
> >         dma_need_sync = !dev->dma_skip_sync;
> >         if (dma_need_sync || dma_addressing_limited(dev))
> >                 return -EOPNOTSUPP;
> 
> Can it get dropped now then?

Better not, it allows us to reject caller much earlier than DMA mapping
flow. It is much saner to fail during UMEM ODP creation than start to
fail for ODP pagefaults.

> 
> > So let's mark mapped buffers with DMA_ATTR_REQUIRE_COHERENT attribute
> > to prevent DMA debugging warnings for cache overlapped entries.
> 
> Well, that isn't the main motivation, this prevents silent data
> corruption if someone tries to use hmm in a system with swiotlb or
> incoherent DMA,
> 
> Looks OK otherwise
> 
> Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
> 
> Jason
>