[PATCH v2 0/6] Add DMA attributes tracking

[PATCH v2 0/6] Add DMA attributes tracking

[Leon Romanovsky]

The DMA API expects that callers use the same attributes when mapping
and unmapping. Add tracking to verify this and catch mismatches.

Thanks

---
Changes in v2:
- Split original patch to be a series
- Added all unmap flows
- Skipped DMA attributes which exist only in map path
- Added ntb patch
- Link to v1: https://patch.msgid.link/20260323-dma-attrs-debug-v1-1-6275228ca300@nvidia.com

---
Leon Romanovsky (6):
      ntb: Store original DMA address for future release
      ntb: Use consistent DMA attributes when freeing DMA mappings
      dma-debug: Remove unused DMA attribute parameter
      dma-debug: Record DMA attributes in debug entry
      dma-debug: Feed DMA attribute for unmapping flows too
      dma-debug: Ensure mappings are created and released with matching attributes

 drivers/ntb/ntb_transport.c | 10 +++---
 kernel/dma/debug.c          | 87 +++++++++++++++++++++++++++++++--------------
 kernel/dma/debug.h          | 25 ++++++-------
 kernel/dma/mapping.c        | 10 +++---
 4 files changed, 85 insertions(+), 47 deletions(-)
---
base-commit: 254f49634ee16a731174d2ae34bc50bd5f45e731
change-id: 20260323-dma-attrs-debug-85e282d6f3bb

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

[PATCH v2 1/6] ntb: Store original DMA address for future release

[Leon Romanovsky]

From: Leon Romanovsky <leonro@nvidia.com>

The DMA API requires that dma_free_attrs receive the exact dma_handle
originally returned by the allocation function. Do not modify it.

Fixes: fc5d1829f9bf ("NTB: transport: Try harder to alloc an aligned MW buffer")
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
 drivers/ntb/ntb_transport.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/drivers/ntb/ntb_transport.c b/drivers/ntb/ntb_transport.c
index 7cabc82305d6b..771eb7d9f4356 100644
--- a/drivers/ntb/ntb_transport.c
+++ b/drivers/ntb/ntb_transport.c
@@ -223,6 +223,7 @@ struct ntb_transport_mw {
 	void *alloc_addr;
 	void *virt_addr;
 	dma_addr_t dma_addr;
+	dma_addr_t original_dma_addr;
 };
 
 struct ntb_transport_client_dev {
@@ -766,7 +767,7 @@ static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
 
 	ntb_mw_clear_trans(nt->ndev, PIDX, num_mw);
 	dma_free_coherent(dma_dev, mw->alloc_size,
-			  mw->alloc_addr, mw->dma_addr);
+			  mw->alloc_addr, mw->original_dma_addr);
 	mw->xlat_size = 0;
 	mw->buff_size = 0;
 	mw->alloc_size = 0;
@@ -802,6 +803,7 @@ static int ntb_alloc_mw_buffer(struct ntb_transport_mw *mw,
 		return -ENOMEM;
 	}
 	virt_addr = alloc_addr;
+	mw->original_dma_addr = dma_addr;
 
 	/*
 	 * we must ensure that the memory address allocated is BAR size

-- 
2.53.0

[PATCH v2 2/6] ntb: Use consistent DMA attributes when freeing DMA mappings

[Leon Romanovsky]

From: Leon Romanovsky <leonro@nvidia.com>

The counterpart of dma_alloc_attrs() is dma_free_attrs(), which must
receive the same DMA attributes used during allocation. The code
previously used dma_free_coherent(), which does not accept or apply any
DMA attributes.

Fixes: 061a785a114f ("ntb: Force physically contiguous allocation of rx ring buffers")
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
 drivers/ntb/ntb_transport.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/drivers/ntb/ntb_transport.c b/drivers/ntb/ntb_transport.c
index 771eb7d9f4356..f59f926d4bfaa 100644
--- a/drivers/ntb/ntb_transport.c
+++ b/drivers/ntb/ntb_transport.c
@@ -766,8 +766,8 @@ static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
 		return;
 
 	ntb_mw_clear_trans(nt->ndev, PIDX, num_mw);
-	dma_free_coherent(dma_dev, mw->alloc_size,
-			  mw->alloc_addr, mw->original_dma_addr);
+	dma_free_attrs(dma_dev, mw->alloc_size, mw->alloc_addr,
+		       mw->original_dma_addr, DMA_ATTR_FORCE_CONTIGUOUS);
 	mw->xlat_size = 0;
 	mw->buff_size = 0;
 	mw->alloc_size = 0;
@@ -828,8 +828,8 @@ static int ntb_alloc_mw_buffer(struct ntb_transport_mw *mw,
 	return 0;
 
 err:
-	dma_free_coherent(ntb_dev, mw->alloc_size, alloc_addr, dma_addr);
-
+	dma_free_attrs(ntb_dev, mw->alloc_size, alloc_addr, dma_addr,
+		       DMA_ATTR_FORCE_CONTIGUOUS);
 	return rc;
 }
 

-- 
2.53.0

[PATCH v2 3/6] dma-debug: Remove unused DMA attribute parameter

[Leon Romanovsky]

From: Leon Romanovsky <leonro@nvidia.com>

debug_dma_alloc_pages() always receives a DMA attribute value of 0,
because dma_alloc_pages() never receives any attributes from its callers.
As preparation for upcoming patches, remove this unused attribute from
the debug routine.

Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
 kernel/dma/debug.c   | 5 ++---
 kernel/dma/debug.h   | 6 ++----
 kernel/dma/mapping.c | 2 +-
 3 files changed, 5 insertions(+), 8 deletions(-)

diff --git a/kernel/dma/debug.c b/kernel/dma/debug.c
index 1a725edbbbf6a..3b53495337f5c 100644
--- a/kernel/dma/debug.c
+++ b/kernel/dma/debug.c
@@ -1567,8 +1567,7 @@ void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
 
 void debug_dma_alloc_pages(struct device *dev, struct page *page,
 			   size_t size, int direction,
-			   dma_addr_t dma_addr,
-			   unsigned long attrs)
+			   dma_addr_t dma_addr)
 {
 	struct dma_debug_entry *entry;
 
@@ -1586,7 +1585,7 @@ void debug_dma_alloc_pages(struct device *dev, struct page *page,
 	entry->dev_addr  = dma_addr;
 	entry->direction = direction;
 
-	add_dma_entry(entry, attrs);
+	add_dma_entry(entry, 0);
 }
 
 void debug_dma_free_pages(struct device *dev, struct page *page,
diff --git a/kernel/dma/debug.h b/kernel/dma/debug.h
index da7be0bddcf67..24b8610850fbd 100644
--- a/kernel/dma/debug.h
+++ b/kernel/dma/debug.h
@@ -47,8 +47,7 @@ extern void debug_dma_sync_sg_for_device(struct device *dev,
 					 int nelems, int direction);
 extern void debug_dma_alloc_pages(struct device *dev, struct page *page,
 				  size_t size, int direction,
-				  dma_addr_t dma_addr,
-				  unsigned long attrs);
+				  dma_addr_t dma_addr);
 extern void debug_dma_free_pages(struct device *dev, struct page *page,
 				 size_t size, int direction,
 				 dma_addr_t dma_addr);
@@ -113,8 +112,7 @@ static inline void debug_dma_sync_sg_for_device(struct device *dev,
 
 static inline void debug_dma_alloc_pages(struct device *dev, struct page *page,
 					 size_t size, int direction,
-					 dma_addr_t dma_addr,
-					 unsigned long attrs)
+					 dma_addr_t dma_addr)
 {
 }
 
diff --git a/kernel/dma/mapping.c b/kernel/dma/mapping.c
index 23ed8eb9233e5..6cbefbd4158c8 100644
--- a/kernel/dma/mapping.c
+++ b/kernel/dma/mapping.c
@@ -733,7 +733,7 @@ struct page *dma_alloc_pages(struct device *dev, size_t size,
 	if (page) {
 		trace_dma_alloc_pages(dev, page_to_virt(page), *dma_handle,
 				      size, dir, gfp, 0);
-		debug_dma_alloc_pages(dev, page, size, dir, *dma_handle, 0);
+		debug_dma_alloc_pages(dev, page, size, dir, *dma_handle);
 	} else {
 		trace_dma_alloc_pages(dev, NULL, 0, size, dir, gfp, 0);
 	}

-- 
2.53.0

[PATCH v2 4/6] dma-debug: Record DMA attributes in debug entry

[Leon Romanovsky]

From: Leon Romanovsky <leonro@nvidia.com>

To enable reliable comparison of DMA attributes between map and
unmap operations, store the attribute value in dma_debug_entry.

Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
 kernel/dma/debug.c | 48 +++++++++++++++++++++++++++++-------------------
 1 file changed, 29 insertions(+), 19 deletions(-)

diff --git a/kernel/dma/debug.c b/kernel/dma/debug.c
index 3b53495337f5c..f07e6a1e9fbab 100644
--- a/kernel/dma/debug.c
+++ b/kernel/dma/debug.c
@@ -63,7 +63,7 @@ enum map_err_types {
  * @sg_mapped_ents: 'mapped_ents' from dma_map_sg
  * @paddr: physical start address of the mapping
  * @map_err_type: track whether dma_mapping_error() was checked
- * @is_cache_clean: driver promises not to write to buffer while mapped
+ * @attrs: dma attributes
  * @stack_len: number of backtrace entries in @stack_entries
  * @stack_entries: stack of backtrace history
  */
@@ -78,7 +78,7 @@ struct dma_debug_entry {
 	int		 sg_mapped_ents;
 	phys_addr_t	 paddr;
 	enum map_err_types map_err_type;
-	bool		 is_cache_clean;
+	unsigned long	 attrs;
 #ifdef CONFIG_STACKTRACE
 	unsigned int	stack_len;
 	unsigned long	stack_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
@@ -478,6 +478,9 @@ static int active_cacheline_insert(struct dma_debug_entry *entry,
 				   bool *overlap_cache_clean)
 {
 	phys_addr_t cln = to_cacheline_number(entry);
+	bool is_cache_clean = entry->attrs &
+			      (DMA_ATTR_DEBUGGING_IGNORE_CACHELINES |
+			       DMA_ATTR_REQUIRE_COHERENT);
 	unsigned long flags;
 	int rc;
 
@@ -495,12 +498,15 @@ static int active_cacheline_insert(struct dma_debug_entry *entry,
 	if (rc == -EEXIST) {
 		struct dma_debug_entry *existing;
 
-		active_cacheline_inc_overlap(cln, entry->is_cache_clean);
+		active_cacheline_inc_overlap(cln, is_cache_clean);
 		existing = radix_tree_lookup(&dma_active_cacheline, cln);
 		/* A lookup failure here after we got -EEXIST is unexpected. */
 		WARN_ON(!existing);
 		if (existing)
-			*overlap_cache_clean = existing->is_cache_clean;
+			*overlap_cache_clean =
+				existing->attrs &
+				(DMA_ATTR_DEBUGGING_IGNORE_CACHELINES |
+				 DMA_ATTR_REQUIRE_COHERENT);
 	}
 	spin_unlock_irqrestore(&radix_lock, flags);
 
@@ -544,12 +550,13 @@ void debug_dma_dump_mappings(struct device *dev)
 			if (!dev || dev == entry->dev) {
 				cln = to_cacheline_number(entry);
 				dev_info(entry->dev,
-					 "%s idx %d P=%pa D=%llx L=%llx cln=%pa %s %s\n",
+					 "%s idx %d P=%pa D=%llx L=%llx cln=%pa %s %s attrs=0x%lx\n",
 					 type2name[entry->type], idx,
 					 &entry->paddr, entry->dev_addr,
 					 entry->size, &cln,
 					 dir2name[entry->direction],
-					 maperr2str[entry->map_err_type]);
+					 maperr2str[entry->map_err_type],
+					 entry->attrs);
 			}
 		}
 		spin_unlock_irqrestore(&bucket->lock, flags);
@@ -575,14 +582,15 @@ static int dump_show(struct seq_file *seq, void *v)
 		list_for_each_entry(entry, &bucket->list, list) {
 			cln = to_cacheline_number(entry);
 			seq_printf(seq,
-				   "%s %s %s idx %d P=%pa D=%llx L=%llx cln=%pa %s %s\n",
+				   "%s %s %s idx %d P=%pa D=%llx L=%llx cln=%pa %s %s attrs=0x%lx\n",
 				   dev_driver_string(entry->dev),
 				   dev_name(entry->dev),
 				   type2name[entry->type], idx,
 				   &entry->paddr, entry->dev_addr,
 				   entry->size, &cln,
 				   dir2name[entry->direction],
-				   maperr2str[entry->map_err_type]);
+				   maperr2str[entry->map_err_type],
+				   entry->attrs);
 		}
 		spin_unlock_irqrestore(&bucket->lock, flags);
 	}
@@ -594,16 +602,14 @@ DEFINE_SHOW_ATTRIBUTE(dump);
  * Wrapper function for adding an entry to the hash.
  * This function takes care of locking itself.
  */
-static void add_dma_entry(struct dma_debug_entry *entry, unsigned long attrs)
+static void add_dma_entry(struct dma_debug_entry *entry)
 {
+	unsigned long attrs = entry->attrs;
 	bool overlap_cache_clean;
 	struct hash_bucket *bucket;
 	unsigned long flags;
 	int rc;
 
-	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);
 	put_hash_bucket(bucket, flags);
@@ -612,9 +618,10 @@ static void add_dma_entry(struct dma_debug_entry *entry, unsigned long attrs)
 	if (rc == -ENOMEM) {
 		pr_err_once("cacheline tracking ENOMEM, dma-debug disabled\n");
 		global_disable = true;
-	} else if (rc == -EEXIST &&
-		   !(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
-		   !(entry->is_cache_clean && overlap_cache_clean) &&
+	} else if (rc == -EEXIST && !(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
+		   !(attrs & (DMA_ATTR_DEBUGGING_IGNORE_CACHELINES |
+			      DMA_ATTR_REQUIRE_COHERENT) &&
+		     overlap_cache_clean) &&
 		   dma_get_cache_alignment() >= L1_CACHE_BYTES &&
 		   !(IS_ENABLED(CONFIG_DMA_BOUNCE_UNALIGNED_KMALLOC) &&
 		     is_swiotlb_active(entry->dev))) {
@@ -1250,6 +1257,7 @@ void debug_dma_map_phys(struct device *dev, phys_addr_t phys, size_t size,
 	entry->size      = size;
 	entry->direction = direction;
 	entry->map_err_type = MAP_ERR_NOT_CHECKED;
+	entry->attrs     = attrs;
 
 	if (!(attrs & DMA_ATTR_MMIO)) {
 		check_for_stack(dev, phys);
@@ -1258,7 +1266,7 @@ void debug_dma_map_phys(struct device *dev, phys_addr_t phys, size_t size,
 			check_for_illegal_area(dev, phys_to_virt(phys), size);
 	}
 
-	add_dma_entry(entry, attrs);
+	add_dma_entry(entry);
 }
 
 void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
@@ -1345,10 +1353,11 @@ void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
 		entry->direction      = direction;
 		entry->sg_call_ents   = nents;
 		entry->sg_mapped_ents = mapped_ents;
+		entry->attrs          = attrs;
 
 		check_sg_segment(dev, s);
 
-		add_dma_entry(entry, attrs);
+		add_dma_entry(entry);
 	}
 }
 
@@ -1440,8 +1449,9 @@ void debug_dma_alloc_coherent(struct device *dev, size_t size,
 	entry->size      = size;
 	entry->dev_addr  = dma_addr;
 	entry->direction = DMA_BIDIRECTIONAL;
+	entry->attrs     = attrs;
 
-	add_dma_entry(entry, attrs);
+	add_dma_entry(entry);
 }
 
 void debug_dma_free_coherent(struct device *dev, size_t size,
@@ -1585,7 +1595,7 @@ void debug_dma_alloc_pages(struct device *dev, struct page *page,
 	entry->dev_addr  = dma_addr;
 	entry->direction = direction;
 
-	add_dma_entry(entry, 0);
+	add_dma_entry(entry);
 }
 
 void debug_dma_free_pages(struct device *dev, struct page *page,

-- 
2.53.0

[PATCH v2 5/6] dma-debug: Feed DMA attribute for unmapping flows too

[Leon Romanovsky]

From: Leon Romanovsky <leonro@nvidia.com>

There are multiple unmapping flows which didn't provide DMA attributes,
which limited DMA debug code to compare the mapping and unmapping
attributes. Let's fix it.

Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
 kernel/dma/debug.c   | 13 ++++++++-----
 kernel/dma/debug.h   | 19 +++++++++++--------
 kernel/dma/mapping.c |  8 ++++----
 3 files changed, 23 insertions(+), 17 deletions(-)

diff --git a/kernel/dma/debug.c b/kernel/dma/debug.c
index f07e6a1e9fbab..3dfed51c3d9aa 100644
--- a/kernel/dma/debug.c
+++ b/kernel/dma/debug.c
@@ -1307,8 +1307,8 @@ void debug_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
 }
 EXPORT_SYMBOL(debug_dma_mapping_error);
 
-void debug_dma_unmap_phys(struct device *dev, dma_addr_t dma_addr,
-			  size_t size, int direction)
+void debug_dma_unmap_phys(struct device *dev, dma_addr_t dma_addr, size_t size,
+			  int direction, unsigned long attrs)
 {
 	struct dma_debug_entry ref = {
 		.type           = dma_debug_phy,
@@ -1316,6 +1316,7 @@ void debug_dma_unmap_phys(struct device *dev, dma_addr_t dma_addr,
 		.dev_addr       = dma_addr,
 		.size           = size,
 		.direction      = direction,
+		.attrs          = attrs,
 	};
 
 	if (unlikely(dma_debug_disabled()))
@@ -1381,7 +1382,7 @@ static int get_nr_mapped_entries(struct device *dev,
 }
 
 void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
-			int nelems, int dir)
+			int nelems, int dir, unsigned long attrs)
 {
 	struct scatterlist *s;
 	int mapped_ents = 0, i;
@@ -1399,6 +1400,7 @@ void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
 			.size           = sg_dma_len(s),
 			.direction      = dir,
 			.sg_call_ents   = nelems,
+			.attrs          = attrs,
 		};
 
 		if (mapped_ents && i >= mapped_ents)
@@ -1454,8 +1456,8 @@ void debug_dma_alloc_coherent(struct device *dev, size_t size,
 	add_dma_entry(entry);
 }
 
-void debug_dma_free_coherent(struct device *dev, size_t size,
-			 void *virt, dma_addr_t dma_addr)
+void debug_dma_free_coherent(struct device *dev, size_t size, void *virt,
+			     dma_addr_t dma_addr, unsigned long attrs)
 {
 	struct dma_debug_entry ref = {
 		.type           = dma_debug_coherent,
@@ -1463,6 +1465,7 @@ void debug_dma_free_coherent(struct device *dev, size_t size,
 		.dev_addr       = dma_addr,
 		.size           = size,
 		.direction      = DMA_BIDIRECTIONAL,
+		.attrs          = attrs,
 	};
 
 	/* handle vmalloc and linear addresses */
diff --git a/kernel/dma/debug.h b/kernel/dma/debug.h
index 24b8610850fbd..13e384633c32a 100644
--- a/kernel/dma/debug.h
+++ b/kernel/dma/debug.h
@@ -14,21 +14,22 @@ extern void debug_dma_map_phys(struct device *dev, phys_addr_t phys,
 			       unsigned long attrs);
 
 extern void debug_dma_unmap_phys(struct device *dev, dma_addr_t addr,
-				 size_t size, int direction);
+				 size_t size, int direction,
+				 unsigned long attrs);
 
 extern void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
 			     int nents, int mapped_ents, int direction,
 			     unsigned long attrs);
 
 extern void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
-			       int nelems, int dir);
+			       int nelems, int dir, unsigned long attrs);
 
 extern void debug_dma_alloc_coherent(struct device *dev, size_t size,
 				     dma_addr_t dma_addr, void *virt,
 				     unsigned long attrs);
 
-extern void debug_dma_free_coherent(struct device *dev, size_t size,
-				    void *virt, dma_addr_t addr);
+extern void debug_dma_free_coherent(struct device *dev, size_t size, void *virt,
+				    dma_addr_t addr, unsigned long attrs);
 
 extern void debug_dma_sync_single_for_cpu(struct device *dev,
 					  dma_addr_t dma_handle, size_t size,
@@ -59,7 +60,8 @@ static inline void debug_dma_map_phys(struct device *dev, phys_addr_t phys,
 }
 
 static inline void debug_dma_unmap_phys(struct device *dev, dma_addr_t addr,
-					size_t size, int direction)
+					size_t size, int direction,
+					unsigned long attrs)
 {
 }
 
@@ -70,8 +72,8 @@ static inline void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
 }
 
 static inline void debug_dma_unmap_sg(struct device *dev,
-				      struct scatterlist *sglist,
-				      int nelems, int dir)
+				      struct scatterlist *sglist, int nelems,
+				      int dir, unsigned long attrs)
 {
 }
 
@@ -82,7 +84,8 @@ static inline void debug_dma_alloc_coherent(struct device *dev, size_t size,
 }
 
 static inline void debug_dma_free_coherent(struct device *dev, size_t size,
-					   void *virt, dma_addr_t addr)
+					   void *virt, dma_addr_t addr,
+					   unsigned long attrs)
 {
 }
 
diff --git a/kernel/dma/mapping.c b/kernel/dma/mapping.c
index 6cbefbd4158c8..f010b3cc0ece4 100644
--- a/kernel/dma/mapping.c
+++ b/kernel/dma/mapping.c
@@ -225,7 +225,7 @@ void dma_unmap_phys(struct device *dev, dma_addr_t addr, size_t size,
 	else if (ops->unmap_phys)
 		ops->unmap_phys(dev, addr, size, dir, attrs);
 	trace_dma_unmap_phys(dev, addr, size, dir, attrs);
-	debug_dma_unmap_phys(dev, addr, size, dir);
+	debug_dma_unmap_phys(dev, addr, size, dir, attrs);
 }
 EXPORT_SYMBOL_GPL(dma_unmap_phys);
 
@@ -351,7 +351,7 @@ void dma_unmap_sg_attrs(struct device *dev, struct scatterlist *sg,
 
 	BUG_ON(!valid_dma_direction(dir));
 	trace_dma_unmap_sg(dev, sg, nents, dir, attrs);
-	debug_dma_unmap_sg(dev, sg, nents, dir);
+	debug_dma_unmap_sg(dev, sg, nents, dir, attrs);
 	if (dma_map_direct(dev, ops) ||
 	    arch_dma_unmap_sg_direct(dev, sg, nents))
 		dma_direct_unmap_sg(dev, sg, nents, dir, attrs);
@@ -693,7 +693,7 @@ void dma_free_attrs(struct device *dev, size_t size, void *cpu_addr,
 	if (!cpu_addr)
 		return;
 
-	debug_dma_free_coherent(dev, size, cpu_addr, dma_handle);
+	debug_dma_free_coherent(dev, size, cpu_addr, dma_handle, attrs);
 	if (dma_alloc_direct(dev, ops) || arch_dma_free_direct(dev, dma_handle))
 		dma_direct_free(dev, size, cpu_addr, dma_handle, attrs);
 	else if (use_dma_iommu(dev))
@@ -840,7 +840,7 @@ void dma_free_noncontiguous(struct device *dev, size_t size,
 		struct sg_table *sgt, enum dma_data_direction dir)
 {
 	trace_dma_free_sgt(dev, sgt, size, dir);
-	debug_dma_unmap_sg(dev, sgt->sgl, sgt->orig_nents, dir);
+	debug_dma_unmap_sg(dev, sgt->sgl, sgt->orig_nents, dir, 0);
 
 	if (use_dma_iommu(dev))
 		iommu_dma_free_noncontiguous(dev, size, sgt, dir);

-- 
2.53.0

[PATCH v2 6/6] dma-debug: Ensure mappings are created and released with matching attributes

[Leon Romanovsky]

From: Leon Romanovsky <leonro@nvidia.com>

The DMA API expects that callers use the same attributes when mapping
and unmapping. Add tracking to verify this and catch mismatches.

Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
 kernel/dma/debug.c | 23 +++++++++++++++++++++++
 1 file changed, 23 insertions(+)

diff --git a/kernel/dma/debug.c b/kernel/dma/debug.c
index 3dfed51c3d9aa..c38efc1ac8d6c 100644
--- a/kernel/dma/debug.c
+++ b/kernel/dma/debug.c
@@ -1074,6 +1074,29 @@ static void check_unmap(struct dma_debug_entry *ref)
 			   type2name[entry->type]);
 	}
 
+	/*
+	 * This may be no bug in reality - but DMA API still expects
+	 * that entry is unmapped with same attributes as it was mapped.
+	 *
+	 * DMA_ATTR_UNMAP_VALID lists the attributes that must be identical
+	 * between map and unmap. Any attribute outside this set (e.g.
+	 * DMA_ATTR_NO_WARN, DMA_ATTR_SKIP_CPU_SYNC) is allowed to differ.
+	 */
+#define DMA_ATTR_UNMAP_VALID                                               \
+	(DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_FORCE_CONTIGUOUS |          \
+	 DMA_ATTR_MMIO | DMA_ATTR_REQUIRE_COHERENT | DMA_ATTR_PRIVILEGED | \
+	 DMA_ATTR_CC_SHARED)
+	if ((ref->attrs & DMA_ATTR_UNMAP_VALID) !=
+	    (entry->attrs & DMA_ATTR_UNMAP_VALID)) {
+		err_printk(ref->dev, entry,
+			   "device driver frees "
+			   "DMA memory with different attributes "
+			   "[device address=0x%016llx] [size=%llu bytes] "
+			   "[mapped with 0x%lx] [unmapped with 0x%lx]\n",
+			   ref->dev_addr, ref->size, entry->attrs, ref->attrs);
+	}
+#undef DMA_ATTR_UNMAP_VALID
+
 	hash_bucket_del(entry);
 	put_hash_bucket(bucket, flags);
 

-- 
2.53.0

Re: [PATCH v2 1/6] ntb: Store original DMA address for future release

[Dave Jiang]

On 4/30/26 11:35 PM, Leon Romanovsky wrote:
> From: Leon Romanovsky <leonro@nvidia.com>
> 
> The DMA API requires that dma_free_attrs receive the exact dma_handle
> originally returned by the allocation function. Do not modify it.
> 
> Fixes: fc5d1829f9bf ("NTB: transport: Try harder to alloc an aligned MW buffer")
> Signed-off-by: Leon Romanovsky <leonro@nvidia.com>

Reviewed-by: Dave Jiang <dave.jiang@intel.com>

> ---
>  drivers/ntb/ntb_transport.c | 4 +++-
>  1 file changed, 3 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/ntb/ntb_transport.c b/drivers/ntb/ntb_transport.c
> index 7cabc82305d6b..771eb7d9f4356 100644
> --- a/drivers/ntb/ntb_transport.c
> +++ b/drivers/ntb/ntb_transport.c
> @@ -223,6 +223,7 @@ struct ntb_transport_mw {
>  	void *alloc_addr;
>  	void *virt_addr;
>  	dma_addr_t dma_addr;
> +	dma_addr_t original_dma_addr;
>  };
>  
>  struct ntb_transport_client_dev {
> @@ -766,7 +767,7 @@ static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
>  
>  	ntb_mw_clear_trans(nt->ndev, PIDX, num_mw);
>  	dma_free_coherent(dma_dev, mw->alloc_size,
> -			  mw->alloc_addr, mw->dma_addr);
> +			  mw->alloc_addr, mw->original_dma_addr);
>  	mw->xlat_size = 0;
>  	mw->buff_size = 0;
>  	mw->alloc_size = 0;
> @@ -802,6 +803,7 @@ static int ntb_alloc_mw_buffer(struct ntb_transport_mw *mw,
>  		return -ENOMEM;
>  	}
>  	virt_addr = alloc_addr;
> +	mw->original_dma_addr = dma_addr;
>  
>  	/*
>  	 * we must ensure that the memory address allocated is BAR size
> 

Re: [PATCH v2 2/6] ntb: Use consistent DMA attributes when freeing DMA mappings

[Dave Jiang]

On 4/30/26 11:35 PM, Leon Romanovsky wrote:
> From: Leon Romanovsky <leonro@nvidia.com>
> 
> The counterpart of dma_alloc_attrs() is dma_free_attrs(), which must
> receive the same DMA attributes used during allocation. The code
> previously used dma_free_coherent(), which does not accept or apply any
> DMA attributes.
> 
> Fixes: 061a785a114f ("ntb: Force physically contiguous allocation of rx ring buffers")
> Signed-off-by: Leon Romanovsky <leonro@nvidia.com>

Reviewed-by: Dave Jiang <dave.jiang@intel.com>


> ---
>  drivers/ntb/ntb_transport.c | 8 ++++----
>  1 file changed, 4 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/ntb/ntb_transport.c b/drivers/ntb/ntb_transport.c
> index 771eb7d9f4356..f59f926d4bfaa 100644
> --- a/drivers/ntb/ntb_transport.c
> +++ b/drivers/ntb/ntb_transport.c
> @@ -766,8 +766,8 @@ static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
>  		return;
>  
>  	ntb_mw_clear_trans(nt->ndev, PIDX, num_mw);
> -	dma_free_coherent(dma_dev, mw->alloc_size,
> -			  mw->alloc_addr, mw->original_dma_addr);
> +	dma_free_attrs(dma_dev, mw->alloc_size, mw->alloc_addr,
> +		       mw->original_dma_addr, DMA_ATTR_FORCE_CONTIGUOUS);
>  	mw->xlat_size = 0;
>  	mw->buff_size = 0;
>  	mw->alloc_size = 0;
> @@ -828,8 +828,8 @@ static int ntb_alloc_mw_buffer(struct ntb_transport_mw *mw,
>  	return 0;
>  
>  err:
> -	dma_free_coherent(ntb_dev, mw->alloc_size, alloc_addr, dma_addr);
> -
> +	dma_free_attrs(ntb_dev, mw->alloc_size, alloc_addr, dma_addr,
> +		       DMA_ATTR_FORCE_CONTIGUOUS);
>  	return rc;
>  }
>  
>