Skip to contentSkip to conversation

SG fix together with update to RDMA umem

PATCH In review Leon Romanovsky 2 participants 7 messages Jun 22, 2021 – Jun 23, 2021

Subsystems (6)
lore
Share & tools
lore
#0 lore raw

cover SG fix together with update to RDMA umem

From: Leon Romanovsky <[email protected]>

Christoph, what do you think about first patch?

Thanks

Maor Gottlieb (2):
lib/scatterlist: Fix wrong update of orig_nents
RDMA: Use dma_map_sgtable for map umem pages

 drivers/infiniband/core/umem.c        | 29 +++++++++---------------
 drivers/infiniband/core/umem_dmabuf.c |  1 -
 drivers/infiniband/hw/mlx4/mr.c       |  4 ++--
 drivers/infiniband/hw/mlx5/mr.c       |  3 ++-
 drivers/infiniband/sw/rdmavt/mr.c     |  2 +-
 drivers/infiniband/sw/rxe/rxe_mr.c    |  3 ++-
 include/linux/scatterlist.h           |  8 +++++--
 include/rdma/ib_umem.h                |  5 ++---
 include/rdma/ib_verbs.h               | 28 +++++++++++++++++++++++
 lib/scatterlist.c                     | 32 +++++++--------------------
 10 files changed, 62 insertions(+), 53 deletions(-)
Signature · 1 line
2.31.1
#1 lore raw

PATCH 1/2 lib/scatterlist: Fix wrong update of orig_nents

From: Maor Gottlieb <[email protected]>

orig_nents should represent the number of entries with pages, but __sg_alloc_table_from_pages sets orig_nents as the number of total entries in the table. This is wrong when the API is used for dynamic allocation where not all the table entries are mapped with pages. It wasn't observed until now, since RDMA umem who uses this API in the dynamic form doesn't use orig_nents implicit or explicit by the scatterlist APIs.

Fix it by:
1. Set orig_nents as number of entries with pages also in __sg_alloc_table_from_pages.
2. Add a new field total_nents to reflect the total number of entries in the table. This is required for the release flow (sg_free_table). This filed should be used internally only by scatterlist.

Fixes 07da1223ec93 ("lib/scatterlist: Add support in dynamic allocation of SG table from pages")
---
 include/linux/scatterlist.h |  8 ++++++--
 lib/scatterlist.c           | 32 ++++++++------------------------
 2 files changed, 14 insertions(+), 26 deletions(-)
include/linux/scatterlist.h +6 -2
index 6f70572b2938..1c889141eb91 100644
@@ -35,8 +35,12 @@ struct scatterlist {
 
 struct sg_table {
 	struct scatterlist *sgl;	/* the list */
-	unsigned int nents;		/* number of mapped entries */
-	unsigned int orig_nents;	/* original size of list */
+	unsigned int nents;		/* number of DMA mapped entries */
+	unsigned int orig_nents;	/* number of CPU mapped entries */
+	/* The fields below should be used internally only by
+	 * scatterlist implementation.
+	 */
+	unsigned int total_nents;	/* number of total entries in the table */
 };
 
 /*
lib/scatterlist.c +8 -24
index a59778946404..6db70a1e7dd0 100644
@@ -192,33 +192,26 @@ static void sg_kfree(struct scatterlist *sg, unsigned int nents)
 void __sg_free_table(struct sg_table *table, unsigned int max_ents,
 		     unsigned int nents_first_chunk, sg_free_fn *free_fn)
 {
-	struct scatterlist *sgl, *next;
+	struct scatterlist *sgl, *next = NULL;
 	unsigned curr_max_ents = nents_first_chunk ?: max_ents;
 
 	if (unlikely(!table->sgl))
 		return;
 
 	sgl = table->sgl;
-	while (table->orig_nents) {
-		unsigned int alloc_size = table->orig_nents;
-		unsigned int sg_size;
+	while (table->total_nents) {
+		unsigned int alloc_size = table->total_nents;
 
 		/*
 		 * If we have more than max_ents segments left,
 		 * then assign 'next' to the sg table after the current one.
-		 * sg_size is then one less than alloc size, since the last
-		 * element is the chain pointer.
 		 */
 		if (alloc_size > curr_max_ents) {
 			next = sg_chain_ptr(&sgl[curr_max_ents - 1]);
 			alloc_size = curr_max_ents;
-			sg_size = alloc_size - 1;
-		} else {
-			sg_size = alloc_size;
-			next = NULL;
 		}
 
-		table->orig_nents -= sg_size;
+		table->total_nents -= alloc_size;
 		if (nents_first_chunk)
 			nents_first_chunk = 0;
 		else
@@ -301,20 +294,11 @@ int __sg_alloc_table(struct sg_table *table, unsigned int nents,
 		} else {
 			sg = alloc_fn(alloc_size, gfp_mask);
 		}
-		if (unlikely(!sg)) {
-			/*
-			 * Adjust entry count to reflect that the last
-			 * entry of the previous table won't be used for
-			 * linkage.  Without this, sg_kfree() may get
-			 * confused.
-			 */
-			if (prv)
-				table->nents = ++table->orig_nents;
-
+		if (unlikely(!sg))
 			return -ENOMEM;
-		}
 
 		sg_init_table(sg, alloc_size);
+		table->total_nents += alloc_size;
 		table->nents = table->orig_nents += sg_size;
 
 		/*
@@ -385,12 +369,11 @@ static struct scatterlist *get_next_sg(struct sg_table *table,
 	if (!new_sg)
 		return ERR_PTR(-ENOMEM);
 	sg_init_table(new_sg, alloc_size);
+	table->total_nents += alloc_size;
 	if (cur) {
 		__sg_chain(next_sg, new_sg);
-		table->orig_nents += alloc_size - 1;
 	} else {
 		table->sgl = new_sg;
-		table->orig_nents = alloc_size;
 		table->nents = 0;
 	}
 	return new_sg;
@@ -515,6 +498,7 @@ struct scatterlist *__sg_alloc_table_from_pages(struct sg_table *sgt,
 		cur_page = j;
 	}
 	sgt->nents += added_nents;
+	sgt->orig_nents = sgt->nents;
 out:
 	if (!left_pages)
 		sg_mark_end(s);
Signature · 1 line
2.31.1
#2 lore raw

PATCH 2/2 RDMA: Use dma_map_sgtable for map umem pages

In reply to Leon Romanovsky

From: Maor Gottlieb <[email protected]>

In order to avoid incorrect usage of sg_table fields, change umem to use dma_map_sgtable for map the pages for DMA. Since dma_map_sgtable update the nents field (number of DMA entries), there is no need anymore for nmap variable, hence do some cleanups accordingly.

---
 drivers/infiniband/core/umem.c        | 29 ++++++++++-----------------
 drivers/infiniband/core/umem_dmabuf.c |  1 -
 drivers/infiniband/hw/mlx4/mr.c       |  4 ++--
 drivers/infiniband/hw/mlx5/mr.c       |  3 ++-
 drivers/infiniband/sw/rdmavt/mr.c     |  2 +-
 drivers/infiniband/sw/rxe/rxe_mr.c    |  3 ++-
 include/rdma/ib_umem.h                |  5 ++---
 include/rdma/ib_verbs.h               | 28 ++++++++++++++++++++++++++
 8 files changed, 48 insertions(+), 27 deletions(-)
drivers/infiniband/core/umem.c +11 -18
index 0eb40025075f..a76ef6a6bac5 100644
@@ -51,11 +51,11 @@ static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int d
 	struct scatterlist *sg;
 	unsigned int i;
 
-	if (umem->nmap > 0)
-		ib_dma_unmap_sg(dev, umem->sg_head.sgl, umem->sg_nents,
-				DMA_BIDIRECTIONAL);
+	if (dirty)
+		ib_dma_unmap_sgtable_attrs(dev, &umem->sg_head,
+					   DMA_BIDIRECTIONAL, 0);
 
-	for_each_sg(umem->sg_head.sgl, sg, umem->sg_nents, i)
+	for_each_sgtable_dma_sg(&umem->sg_head, sg, i)
 		unpin_user_page_range_dirty_lock(sg_page(sg),
 			DIV_ROUND_UP(sg->length, PAGE_SIZE), make_dirty);
 
@@ -111,7 +111,7 @@ unsigned long ib_umem_find_best_pgsz(struct ib_umem *umem,
 	/* offset into first SGL */
 	pgoff = umem->address & ~PAGE_MASK;
 
-	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) {
+	for_each_sgtable_dma_sg(&umem->sg_head, sg, i) {
 		/* Walk SGL and reduce max page size if VA/PA bits differ
 		 * for any address.
 		 */
@@ -121,7 +121,7 @@ unsigned long ib_umem_find_best_pgsz(struct ib_umem *umem,
 		 * the maximum possible page size as the low bits of the iova
 		 * must be zero when starting the next chunk.
 		 */
-		if (i != (umem->nmap - 1))
+		if (i != (umem->sg_head.nents - 1))
 			mask |= va;
 		pgoff = 0;
 	}
@@ -230,7 +230,6 @@ struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 				0, ret << PAGE_SHIFT,
 				ib_dma_max_seg_size(device), sg, npages,
 				GFP_KERNEL);
-		umem->sg_nents = umem->sg_head.nents;
 		if (IS_ERR(sg)) {
 			unpin_user_pages_dirty_lock(page_list, ret, 0);
 			ret = PTR_ERR(sg);
@@ -241,16 +240,10 @@ struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 	if (access & IB_ACCESS_RELAXED_ORDERING)
 		dma_attr |= DMA_ATTR_WEAK_ORDERING;
 
-	umem->nmap =
-		ib_dma_map_sg_attrs(device, umem->sg_head.sgl, umem->sg_nents,
-				    DMA_BIDIRECTIONAL, dma_attr);
-
-	if (!umem->nmap) {
-		ret = -ENOMEM;
+	ret = ib_dma_map_sgtable_attrs(device, &umem->sg_head,
+				       DMA_BIDIRECTIONAL, dma_attr);
+	if (ret)
 		goto umem_release;
-	}
-
-	ret = 0;
 	goto out;
 
 umem_release:
@@ -310,8 +303,8 @@ int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
 		return -EINVAL;
 	}
 
-	ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->sg_nents, dst, length,
-				 offset + ib_umem_offset(umem));
+	ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->sg_head.orig_nents,
+				 dst, length, offset + ib_umem_offset(umem));
 
 	if (ret < 0)
 		return ret;
drivers/infiniband/core/umem_dmabuf.c +0 -1
index 0d65ce146fc4..cd2dd1f39aa7 100644
@@ -57,7 +57,6 @@ int ib_umem_dmabuf_map_pages(struct ib_umem_dmabuf *umem_dmabuf)
 
 	umem_dmabuf->umem.sg_head.sgl = umem_dmabuf->first_sg;
 	umem_dmabuf->umem.sg_head.nents = nmap;
-	umem_dmabuf->umem.nmap = nmap;
 	umem_dmabuf->sgt = sgt;
 
 wait_fence:
drivers/infiniband/hw/mlx4/mr.c +2 -2
index 50becc0e4b62..ab5dc8eac7f8 100644
@@ -200,7 +200,7 @@ int mlx4_ib_umem_write_mtt(struct mlx4_ib_dev *dev, struct mlx4_mtt *mtt,
 	mtt_shift = mtt->page_shift;
 	mtt_size = 1ULL << mtt_shift;
 
-	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) {
+	for_each_sgtable_dma_sg(&umem->sg_head, sg, i) {
 		if (cur_start_addr + len == sg_dma_address(sg)) {
 			/* still the same block */
 			len += sg_dma_len(sg);
@@ -273,7 +273,7 @@ int mlx4_ib_umem_calc_optimal_mtt_size(struct ib_umem *umem, u64 start_va,
 
 	*num_of_mtts = ib_umem_num_dma_blocks(umem, PAGE_SIZE);
 
-	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) {
+	for_each_sgtable_dma_sg(&umem->sg_head, sg, i) {
 		/*
 		 * Initialization - save the first chunk start as the
 		 * current_block_start - block means contiguous pages.
drivers/infiniband/hw/mlx5/mr.c +2 -1
index e288531fedc9..870209178a65 100644
@@ -1226,7 +1226,8 @@ int mlx5_ib_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags)
 	orig_sg_length = sg.length;
 
 	cur_mtt = mtt;
-	rdma_for_each_block (mr->umem->sg_head.sgl, &biter, mr->umem->nmap,
+	rdma_for_each_block (mr->umem->sg_head.sgl, &biter,
+			     mr->umem->sg_head.nents,
 			     BIT(mr->page_shift)) {
 		if (cur_mtt == (void *)mtt + sg.length) {
 			dma_sync_single_for_device(ddev, sg.addr, sg.length,
drivers/infiniband/sw/rdmavt/mr.c +1 -1
index 34b7af6ab9c2..d955c8c4acc4 100644
@@ -410,7 +410,7 @@ struct ib_mr *rvt_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 	mr->mr.page_shift = PAGE_SHIFT;
 	m = 0;
 	n = 0;
-	for_each_sg_page (umem->sg_head.sgl, &sg_iter, umem->nmap, 0) {
+	for_each_sg_page (umem->sg_head.sgl, &sg_iter, umem->sg_head.nents, 0) {
 		void *vaddr;
 
 		vaddr = page_address(sg_page_iter_page(&sg_iter));
drivers/infiniband/sw/rxe/rxe_mr.c +2 -1
index 6aabcb4de235..a269085e0946 100644
@@ -142,7 +142,8 @@ int rxe_mr_init_user(struct rxe_pd *pd, u64 start, u64 length, u64 iova,
 	if (length > 0) {
 		buf = map[0]->buf;
 
-		for_each_sg_page(umem->sg_head.sgl, &sg_iter, umem->nmap, 0) {
+		for_each_sg_page(umem->sg_head.sgl, &sg_iter,
+				 umem->sg_head.nents, 0) {
 			if (num_buf >= RXE_BUF_PER_MAP) {
 				map++;
 				buf = map[0]->buf;
include/rdma/ib_umem.h +2 -3
index 676c57f5ca80..c754b1a31cc9 100644
@@ -27,8 +27,6 @@ struct ib_umem {
 	u32 is_dmabuf : 1;
 	struct work_struct	work;
 	struct sg_table sg_head;
-	int             nmap;
-	unsigned int    sg_nents;
 };
 
 struct ib_umem_dmabuf {
@@ -77,7 +75,8 @@ static inline void __rdma_umem_block_iter_start(struct ib_block_iter *biter,
 						struct ib_umem *umem,
 						unsigned long pgsz)
 {
-	__rdma_block_iter_start(biter, umem->sg_head.sgl, umem->nmap, pgsz);
+	__rdma_block_iter_start(biter, umem->sg_head.sgl, umem->sg_head.nents,
+				pgsz);
 }
 
 /**
include/rdma/ib_verbs.h +28 -0
index 371df1c80aeb..2dba30849731 100644
@@ -4057,6 +4057,34 @@ static inline void ib_dma_unmap_sg_attrs(struct ib_device *dev,
 				   dma_attrs);
 }
 
+/**
+ * ib_dma_map_sgtable_attrs - Map a scatter/gather table to DMA addresses
+ * @dev: The device for which the DMA addresses are to be created
+ * @sg: The sg_table object describing the buffer
+ * @direction: The direction of the DMA
+ * @attrs: Optional DMA attributes for the map operation
+ */
+static inline int ib_dma_map_sgtable_attrs(struct ib_device *dev,
+					   struct sg_table *sgt,
+					   enum dma_data_direction direction,
+					   unsigned long dma_attrs)
+{
+	if (ib_uses_virt_dma(dev)) {
+		ib_dma_virt_map_sg(dev, sgt->sgl, sgt->orig_nents);
+		return 0;
+	}
+	return dma_map_sgtable(dev->dma_device, sgt, direction, dma_attrs);
+}
+
+static inline void ib_dma_unmap_sgtable_attrs(struct ib_device *dev,
+					      struct sg_table *sgt,
+					      enum dma_data_direction direction,
+					      unsigned long dma_attrs)
+{
+	if (!ib_uses_virt_dma(dev))
+		dma_unmap_sgtable(dev->dma_device, sgt, direction, dma_attrs);
+}
+
 /**
  * ib_dma_map_sg - Map a scatter/gather list to DMA addresses
  * @dev: The device for which the DMA addresses are to be created
Signature · 1 line
2.31.1
maintainer
#3 lore raw

Re: [PATCH rdma-next 2/2] RDMA: Use dma_map_sgtable for map umem pages

Quoted 8 lines · Leon Romanovsky
drivers/infiniband/core/umem.c +1 -1
index 0eb40025075f..a76ef6a6bac5 100644
+++ b/drivers/infiniband/core/umem.c
@@ -51,11 +51,11 @@ static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int d
 	struct scatterlist *sg;
 	unsigned int i;
 
-	if (umem->nmap > 0)

drivers/infiniband/core/umem.c static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int d · line 60

Quoted 10 lines
drivers/infiniband/core/umem.c +4 -3
-		ib_dma_unmap_sg(dev, umem->sg_head.sgl, umem->sg_nents,
-				DMA_BIDIRECTIONAL);
+	if (dirty)
+		ib_dma_unmap_sgtable_attrs(dev, &umem->sg_head,
+					   DMA_BIDIRECTIONAL, 0);
 
-	for_each_sg(umem->sg_head.sgl, sg, umem->sg_nents, i)
+	for_each_sgtable_dma_sg(&umem->sg_head, sg, i)
 		unpin_user_page_range_dirty_lock(sg_page(sg),
 			DIV_ROUND_UP(sg->length, PAGE_SIZE), make_dirty);

This isn't right, can't mix sg_page with a _dma_ API

Jason

#4 lore raw

Re: [PATCH rdma-next 2/2] RDMA: Use dma_map_sgtable for map umem pages

Quoted 16 lines · Jason Gunthorpe
Quoted 14 lines · Leon Romanovsky
drivers/infiniband/core/umem.c +4 -3
index 0eb40025075f..a76ef6a6bac5 100644
+++ b/drivers/infiniband/core/umem.c
@@ -51,11 +51,11 @@ static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int d
 	struct scatterlist *sg;
 	unsigned int i;
 
-	if (umem->nmap > 0)
-		ib_dma_unmap_sg(dev, umem->sg_head.sgl, umem->sg_nents,
-				DMA_BIDIRECTIONAL);
+	if (dirty)
+		ib_dma_unmap_sgtable_attrs(dev, &umem->sg_head,
+					   DMA_BIDIRECTIONAL, 0);
 
Quoted 6 lines

drivers/infiniband/core/umem.c static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int d · line 60

Quoted 4 lines
drivers/infiniband/core/umem.c +1 -1
-	for_each_sg(umem->sg_head.sgl, sg, umem->sg_nents, i)
+	for_each_sgtable_dma_sg(&umem->sg_head, sg, i)
 		unpin_user_page_range_dirty_lock(sg_page(sg),
 			DIV_ROUND_UP(sg->length, PAGE_SIZE), make_dirty);

This isn't right, can't mix sg_page with a _dma_ API

Jason, why is that?

We use same pages that were passed to __sg_alloc_table_from_pages() in __ib_umem_get().

Thanks

Quoted 2 lines

Jason

maintainer
#5 lore raw

Re: [PATCH rdma-next 2/2] RDMA: Use dma_map_sgtable for map umem pages

Quoted 24 lines · Leon Romanovsky
Quoted 16 lines · Jason Gunthorpe
Quoted 14 lines · Leon Romanovsky
drivers/infiniband/core/umem.c +4 -3
index 0eb40025075f..a76ef6a6bac5 100644
+++ b/drivers/infiniband/core/umem.c
@@ -51,11 +51,11 @@ static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int d
 	struct scatterlist *sg;
 	unsigned int i;
 
-	if (umem->nmap > 0)
-		ib_dma_unmap_sg(dev, umem->sg_head.sgl, umem->sg_nents,
-				DMA_BIDIRECTIONAL);
+	if (dirty)
+		ib_dma_unmap_sgtable_attrs(dev, &umem->sg_head,
+					   DMA_BIDIRECTIONAL, 0);
 
Quoted 6 lines

drivers/infiniband/core/umem.c static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int d · line 60

Quoted 4 lines
drivers/infiniband/core/umem.c +1 -1
-	for_each_sg(umem->sg_head.sgl, sg, umem->sg_nents, i)
+	for_each_sgtable_dma_sg(&umem->sg_head, sg, i)
 		unpin_user_page_range_dirty_lock(sg_page(sg),
 			DIV_ROUND_UP(sg->length, PAGE_SIZE), make_dirty);

This isn't right, can't mix sg_page with a _dma_ API

Quoted 3 lines

Jason, why is that?

We use same pages that were passed to __sg_alloc_table_from_pages() in __ib_umem_get().

A sgl has two lists inside it a 'dma' list and a 'page' list, they are not the same length and not interchangable.

If you use for_each_sgtable_dma_sg() then you iterate over the 'dma' list and have to use 'dma' accessors

If you use for_each_sgtable_sg() then you iterate over the 'page' list and have to use 'page' acessors

Mixing dma iteration with page accessors or vice-versa, like above, is always a bug.

You can see it alos because the old code used umem->sg_nents which is the CPU list length while this new code is using the dma list length.

Jason

#6 lore raw

Re: [PATCH rdma-next 2/2] RDMA: Use dma_map_sgtable for map umem pages

Quoted 35 lines · Jason Gunthorpe
Quoted 24 lines · Leon Romanovsky
Quoted 16 lines · Jason Gunthorpe
Quoted 14 lines · Leon Romanovsky
drivers/infiniband/core/umem.c +4 -3
index 0eb40025075f..a76ef6a6bac5 100644
+++ b/drivers/infiniband/core/umem.c
@@ -51,11 +51,11 @@ static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int d
 	struct scatterlist *sg;
 	unsigned int i;
 
-	if (umem->nmap > 0)
-		ib_dma_unmap_sg(dev, umem->sg_head.sgl, umem->sg_nents,
-				DMA_BIDIRECTIONAL);
+	if (dirty)
+		ib_dma_unmap_sgtable_attrs(dev, &umem->sg_head,
+					   DMA_BIDIRECTIONAL, 0);
 
Quoted 6 lines

drivers/infiniband/core/umem.c static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int d · line 60

Quoted 4 lines
drivers/infiniband/core/umem.c +1 -1
-	for_each_sg(umem->sg_head.sgl, sg, umem->sg_nents, i)
+	for_each_sgtable_dma_sg(&umem->sg_head, sg, i)
 		unpin_user_page_range_dirty_lock(sg_page(sg),
 			DIV_ROUND_UP(sg->length, PAGE_SIZE), make_dirty);

This isn't right, can't mix sg_page with a _dma_ API

Quoted 3 lines

Jason, why is that?

We use same pages that were passed to __sg_alloc_table_from_pages() in __ib_umem_get().

A sgl has two lists inside it a 'dma' list and a 'page' list, they are not the same length and not interchangable.

If you use for_each_sgtable_dma_sg() then you iterate over the 'dma' list and have to use 'dma' accessors

Quoted 8 lines

If you use for_each_sgtable_sg() then you iterate over the 'page' list and have to use 'page' acessors

Mixing dma iteration with page accessors or vice-versa, like above, is always a bug.

You can see it alos because the old code used umem->sg_nents which is the CPU list length while this new code is using the dma list length.

Ohh, I see difference between types now, thank you for the explanation, will consult with Maor once he returns next week to the office.

Thanks

Quoted 2 lines

Jason