Add support for dynamic UAR usage

index 9e2d6a872e5a..4f2d71298937 100644
@@ -1417,9 +1417,18 @@ static void print_lib_caps(struct mlx5_ib_dev *dev, u64 caps)
 		    caps & MLX5_LIB_CAP_4K_UAR ? "y" : "n");
 }
 
+static u16 calc_dynamic_bfregs(int uars_per_sys_page)
+{
+	/* Large page with non 4k uar support might limit the dynamic size */
+	if (uars_per_sys_page == 1  && PAGE_SIZE > 4096)
+		return MLX5_MIN_DYN_BFREGS;
+
+	return MLX5_MAX_DYN_BFREGS;
+}
+
 static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
 			     struct mlx5_ib_alloc_ucontext_req_v2 *req,
-			     u32 *num_sys_pages)
+			     struct mlx5_bfreg_info *bfregi)
 {
 	int uars_per_sys_page;
 	int bfregs_per_sys_page;

@@ -1436,16 +1445,21 @@ static int calc_total_bfregs(struct mlx5_ib_dev *dev, bool lib_uar_4k,
 
 	uars_per_sys_page = get_uars_per_sys_page(dev, lib_uar_4k);
 	bfregs_per_sys_page = uars_per_sys_page * MLX5_NON_FP_BFREGS_PER_UAR;
+	/* This holds the required static allocation asked by the user */
 	req->total_num_bfregs = ALIGN(req->total_num_bfregs, bfregs_per_sys_page);
-	*num_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
-
 	if (req->num_low_latency_bfregs > req->total_num_bfregs - 1)
 		return -EINVAL;
 
-	mlx5_ib_dbg(dev, "uar_4k: fw support %s, lib support %s, user requested %d bfregs, allocated %d, using %d sys pages\n",
+	bfregi->num_static_sys_pages = req->total_num_bfregs / bfregs_per_sys_page;
+	bfregi->num_dyn_bfregs = ALIGN(calc_dynamic_bfregs(uars_per_sys_page), bfregs_per_sys_page);
+	bfregi->total_num_bfregs = req->total_num_bfregs + bfregi->num_dyn_bfregs;
+	bfregi->num_sys_pages = bfregi->total_num_bfregs / bfregs_per_sys_page;
+
+	mlx5_ib_dbg(dev, "uar_4k: fw support %s, lib support %s, user requested %d bfregs, allocated %d, total bfregs %d, using %d sys pages\n",
 		    MLX5_CAP_GEN(dev->mdev, uar_4k) ? "yes" : "no",
 		    lib_uar_4k ? "yes" : "no", ref_bfregs,
-		    req->total_num_bfregs, *num_sys_pages);
+		    req->total_num_bfregs, bfregi->total_num_bfregs,
+		    bfregi->num_sys_pages);
 
 	return 0;
 }

@@ -1457,7 +1471,7 @@ static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *conte
 	int i;
 
 	bfregi = &context->bfregi;
-	for (i = 0; i < bfregi->num_sys_pages; i++) {
+	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
 		err = mlx5_cmd_alloc_uar(dev->mdev, &bfregi->sys_pages[i]);
 		if (err)
 			goto error;

@@ -1481,7 +1495,7 @@ static int deallocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *con
 	int i;
 
 	bfregi = &context->bfregi;
-	for (i = 0; i < bfregi->num_sys_pages; i++) {
+	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
 		err = mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
 		if (err) {
 			mlx5_ib_warn(dev, "failed to free uar %d\n", i);

@@ -1596,13 +1610,13 @@ static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
 	bfregi = &context->bfregi;
 
 	/* updates req->total_num_bfregs */
-	err = calc_total_bfregs(dev, lib_uar_4k, &req, &bfregi->num_sys_pages);
+	err = calc_total_bfregs(dev, lib_uar_4k, &req, bfregi);
 	if (err)
 		goto out_ctx;
 
 	mutex_init(&bfregi->lock);
 	bfregi->lib_uar_4k = lib_uar_4k;
-	bfregi->count = kcalloc(req.total_num_bfregs, sizeof(*bfregi->count),
+	bfregi->count = kcalloc(bfregi->total_num_bfregs, sizeof(*bfregi->count),
 				GFP_KERNEL);
 	if (!bfregi->count) {
 		err = -ENOMEM;

@@ -1685,6 +1699,11 @@ static struct ib_ucontext *mlx5_ib_alloc_ucontext(struct ib_device *ibdev,
 	if (field_avail(typeof(resp), num_uars_per_page, udata->outlen))
 		resp.response_length += sizeof(resp.num_uars_per_page);
 
+	if (field_avail(typeof(resp), num_dyn_bfregs, udata->outlen)) {
+		resp.num_dyn_bfregs = bfregi->num_dyn_bfregs;
+		resp.response_length += sizeof(resp.num_dyn_bfregs);
+	}
+
 	err = ib_copy_to_udata(udata, &resp, resp.response_length);
 	if (err)
 		goto out_td;

index 196b55b90c60..32b317d16280 100644
@@ -1177,10 +1177,10 @@ static inline int get_uars_per_sys_page(struct mlx5_ib_dev *dev, bool lib_suppor
 				MLX5_UARS_IN_PAGE : 1;
 }
 
-static inline int get_num_uars(struct mlx5_ib_dev *dev,
-			       struct mlx5_bfreg_info *bfregi)
+static inline int get_num_static_uars(struct mlx5_ib_dev *dev,
+				      struct mlx5_bfreg_info *bfregi)
 {
-	return get_uars_per_sys_page(dev, bfregi->lib_uar_4k) * bfregi->num_sys_pages;
+	return get_uars_per_sys_page(dev, bfregi->lib_uar_4k) * bfregi->num_static_sys_pages;
 }
 
 #endif /* MLX5_IB_H */

index 33b132a9a0fc..166b37daa794 100644
@@ -493,7 +493,7 @@ enum {
 
 static int max_bfregs(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi)
 {
-	return get_num_uars(dev, bfregi) * MLX5_NON_FP_BFREGS_PER_UAR;
+	return get_num_static_uars(dev, bfregi) * MLX5_NON_FP_BFREGS_PER_UAR;
 }
 
 static int num_med_bfreg(struct mlx5_ib_dev *dev,

index 18c041966ab8..2ff9f48ebec3 100644
@@ -249,6 +249,8 @@ enum {
 					  MLX5_NON_FP_BFREGS_PER_UAR,
 	MLX5_UARS_IN_PAGE		= PAGE_SIZE / MLX5_ADAPTER_PAGE_SIZE,
 	MLX5_NON_FP_BFREGS_IN_PAGE	= MLX5_NON_FP_BFREGS_PER_UAR * MLX5_UARS_IN_PAGE,
+	MLX5_MIN_DYN_BFREGS		= 512,
+	MLX5_MAX_DYN_BFREGS		= 1024,
 };
 
 enum {

index bc494ddefcf8..4d1ed35fdbb4 100644
@@ -230,6 +230,9 @@ struct mlx5_bfreg_info {
 	u32			ver;
 	bool			lib_uar_4k;
 	u32			num_sys_pages;
+	u32			num_static_sys_pages;
+	u32			total_num_bfregs;
+	u32			num_dyn_bfregs;
 };
 
 struct mlx5_cmd_first {

index 062d14f07b61..12c523a1dd18 100644
@@ -125,6 +125,8 @@ struct mlx5_ib_alloc_ucontext_resp {
 	__u64	hca_core_clock_offset;
 	__u32	log_uar_size;
 	__u32	num_uars_per_page;
+	__u32	num_dyn_bfregs;
+	__u32	reserved3;
 };
 
 struct mlx5_ib_alloc_pd_resp {

index 4f2d71298937..9c737cc65503 100644
@@ -1478,6 +1478,10 @@ static int allocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *conte
 
 		mlx5_ib_dbg(dev, "allocated uar %d\n", bfregi->sys_pages[i]);
 	}
+
+	for (i = bfregi->num_static_sys_pages; i < bfregi->num_sys_pages; i++)
+		bfregi->sys_pages[i] = MLX5_IB_INVALID_UAR_INDEX;
+
 	return 0;
 
 error:

@@ -1495,13 +1499,17 @@ static int deallocate_uars(struct mlx5_ib_dev *dev, struct mlx5_ib_ucontext *con
 	int i;
 
 	bfregi = &context->bfregi;
-	for (i = 0; i < bfregi->num_static_sys_pages; i++) {
-		err = mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
-		if (err) {
-			mlx5_ib_warn(dev, "failed to free uar %d\n", i);
-			return err;
+	for (i = 0; i < bfregi->num_sys_pages; i++) {
+		if (i < bfregi->num_static_sys_pages ||
+		    bfregi->sys_pages[i] != MLX5_IB_INVALID_UAR_INDEX) {
+			err = mlx5_cmd_free_uar(dev->mdev, bfregi->sys_pages[i]);
+			if (err) {
+				mlx5_ib_warn(dev, "failed to free uar %d, err=%d\n", i, err);
+				return err;
+			}
 		}
 	}
+
 	return 0;
 }
 

@@ -1758,15 +1766,13 @@ static int mlx5_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
 }
 
 static phys_addr_t uar_index2pfn(struct mlx5_ib_dev *dev,
-				 struct mlx5_bfreg_info *bfregi,
-				 int idx)
+				 int uar_idx)
 {
 	int fw_uars_per_page;
 
 	fw_uars_per_page = MLX5_CAP_GEN(dev->mdev, uar_4k) ? MLX5_UARS_IN_PAGE : 1;
 
-	return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) +
-			bfregi->sys_pages[idx] / fw_uars_per_page;
+	return (pci_resource_start(dev->mdev->pdev, 0) >> PAGE_SHIFT) + uar_idx / fw_uars_per_page;
 }
 
 static int get_command(unsigned long offset)

@@ -1784,6 +1790,12 @@ static int get_index(unsigned long offset)
 	return get_arg(offset);
 }
 
+/* Index resides in an extra byte to enable larger values than 255 */
+static int get_extended_index(unsigned long offset)
+{
+	return get_arg(offset) | ((offset >> 16) & 0xff) << 8;
+}
+
 static void  mlx5_ib_vma_open(struct vm_area_struct *area)
 {
 	/* vma_open is called when a new VMA is created on top of our VMA.  This

@@ -1927,21 +1939,29 @@ static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
 	unsigned long idx;
 	phys_addr_t pfn, pa;
 	pgprot_t prot;
-	int uars_per_page;
+	u32 bfreg_dyn_idx = 0;
+	u32 uar_index;
+	int dyn_uar = (cmd == MLX5_IB_MMAP_ALLOC_WC);
+	int max_valid_idx = dyn_uar ? bfregi->num_sys_pages :
+				bfregi->num_static_sys_pages;
 
 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
 		return -EINVAL;
 
-	uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
-	idx = get_index(vma->vm_pgoff);
-	if (idx % uars_per_page ||
-	    idx * uars_per_page >= bfregi->num_sys_pages) {
-		mlx5_ib_warn(dev, "invalid uar index %lu\n", idx);
+	if (dyn_uar)
+		idx = get_extended_index(vma->vm_pgoff) + bfregi->num_static_sys_pages;
+	else
+		idx = get_index(vma->vm_pgoff);
+
+	if (idx >= max_valid_idx) {
+		mlx5_ib_warn(dev, "invalid uar index %lu, max=%d\n",
+			     idx, max_valid_idx);
 		return -EINVAL;
 	}
 
 	switch (cmd) {
 	case MLX5_IB_MMAP_WC_PAGE:
+	case MLX5_IB_MMAP_ALLOC_WC:
 /* Some architectures don't support WC memory */
 #if defined(CONFIG_X86)
 		if (!pat_enabled())

@@ -1961,7 +1981,40 @@ static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
 		return -EINVAL;
 	}
 
-	pfn = uar_index2pfn(dev, bfregi, idx);
+	if (dyn_uar) {
+		int uars_per_page;
+
+		uars_per_page = get_uars_per_sys_page(dev, bfregi->lib_uar_4k);
+		bfreg_dyn_idx = idx * (uars_per_page * MLX5_NON_FP_BFREGS_PER_UAR);
+		if (bfreg_dyn_idx >= bfregi->total_num_bfregs) {
+			mlx5_ib_warn(dev, "invalid bfreg_dyn_idx %u, max=%u\n",
+				     bfreg_dyn_idx, bfregi->total_num_bfregs);
+			return -EINVAL;
+		}
+
+		mutex_lock(&bfregi->lock);
+		/* Fail if uar already allocated, first bfreg index of each
+		 * page holds its count.
+		 */
+		if (bfregi->count[bfreg_dyn_idx]) {
+			mlx5_ib_warn(dev, "wrong offset, idx %lu is busy, bfregn=%u\n", idx, bfreg_dyn_idx);
+			mutex_unlock(&bfregi->lock);
+			return -EINVAL;
+		}
+
+		bfregi->count[bfreg_dyn_idx]++;
+		mutex_unlock(&bfregi->lock);
+
+		err = mlx5_cmd_alloc_uar(dev->mdev, &uar_index);
+		if (err) {
+			mlx5_ib_warn(dev, "UAR alloc failed\n");
+			goto free_bfreg;
+		}
+	} else {
+		uar_index = bfregi->sys_pages[idx];
+	}
+
+	pfn = uar_index2pfn(dev, uar_index);
 	mlx5_ib_dbg(dev, "uar idx 0x%lx, pfn %pa\n", idx, &pfn);
 
 	vma->vm_page_prot = prot;

@@ -1970,14 +2023,32 @@ static int uar_mmap(struct mlx5_ib_dev *dev, enum mlx5_ib_mmap_cmd cmd,
 	if (err) {
 		mlx5_ib_err(dev, "io_remap_pfn_range failed with error=%d, vm_start=0x%lx, pfn=%pa, mmap_cmd=%s\n",
 			    err, vma->vm_start, &pfn, mmap_cmd2str(cmd));
-		return -EAGAIN;
+		err = -EAGAIN;
+		goto err;
 	}
 
 	pa = pfn << PAGE_SHIFT;
 	mlx5_ib_dbg(dev, "mapped %s at 0x%lx, PA %pa\n", mmap_cmd2str(cmd),
 		    vma->vm_start, &pa);
 
-	return mlx5_ib_set_vma_data(vma, context);
+	err = mlx5_ib_set_vma_data(vma, context);
+	if (err)
+		goto err;
+
+	if (dyn_uar)
+		bfregi->sys_pages[idx] = uar_index;
+	return 0;
+
+err:
+	if (!dyn_uar)
+		return err;
+
+	mlx5_cmd_free_uar(dev->mdev, idx);
+
+free_bfreg:
+	mlx5_ib_free_bfreg(dev, bfregi, bfreg_dyn_idx);
+
+	return err;
 }
 
 static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)

@@ -1992,6 +2063,7 @@ static int mlx5_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vm
 	case MLX5_IB_MMAP_WC_PAGE:
 	case MLX5_IB_MMAP_NC_PAGE:
 	case MLX5_IB_MMAP_REGULAR_PAGE:
+	case MLX5_IB_MMAP_ALLOC_WC:
 		return uar_mmap(dev, command, vma, context);
 
 	case MLX5_IB_MMAP_GET_CONTIGUOUS_PAGES:

index 32b317d16280..c471e11b720c 100644
@@ -77,6 +77,7 @@ enum mlx5_ib_mmap_cmd {
 	MLX5_IB_MMAP_NC_PAGE			= 3,
 	/* 5 is chosen in order to be compatible with old versions of libmlx5 */
 	MLX5_IB_MMAP_CORE_CLOCK			= 5,
+	MLX5_IB_MMAP_ALLOC_WC			= 6,
 };
 
 enum {

@@ -112,6 +113,10 @@ enum {
 	MLX5_TM_MAX_SGE			= 1,
 };
 
+enum {
+	MLX5_IB_INVALID_UAR_INDEX	= BIT(31),
+};
+
 struct mlx5_ib_vma_private_data {
 	struct list_head list;
 	struct vm_area_struct *vma;

@@ -1085,6 +1090,7 @@ struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *dev,
 void mlx5_ib_put_native_port_mdev(struct mlx5_ib_dev *dev,
 				  u8 port_num);
 
+void mlx5_ib_free_bfreg(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi, int bfregn);
 static inline void init_query_mad(struct ib_smp *mad)
 {
 	mad->base_version  = 1;

index 166b37daa794..0f299dfd8ae0 100644
@@ -581,7 +581,7 @@ static int alloc_bfreg(struct mlx5_ib_dev *dev,
 	return bfregn;
 }
 
-static void free_bfreg(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi, int bfregn)
+void mlx5_ib_free_bfreg(struct mlx5_ib_dev *dev, struct mlx5_bfreg_info *bfregi, int bfregn)
 {
 	mutex_lock(&bfregi->lock);
 	bfregi->count[bfregn]--;

@@ -874,7 +874,7 @@ static int create_user_qp(struct mlx5_ib_dev *dev, struct ib_pd *pd,
 		ib_umem_release(ubuffer->umem);
 
 err_bfreg:
-	free_bfreg(dev, &context->bfregi, bfregn);
+	mlx5_ib_free_bfreg(dev, &context->bfregi, bfregn);
 	return err;
 }
 

@@ -887,7 +887,7 @@ static void destroy_qp_user(struct mlx5_ib_dev *dev, struct ib_pd *pd,
 	mlx5_ib_db_unmap_user(context, &qp->db);
 	if (base->ubuffer.umem)
 		ib_umem_release(base->ubuffer.umem);
-	free_bfreg(dev, &context->bfregi, qp->bfregn);
+	mlx5_ib_free_bfreg(dev, &context->bfregi, qp->bfregn);
 }
 
 static int create_kernel_qp(struct mlx5_ib_dev *dev,

index c471e11b720c..88b89ed15d1c 100644
@@ -115,6 +115,7 @@ enum {
 
 enum {
 	MLX5_IB_INVALID_UAR_INDEX	= BIT(31),
+	MLX5_IB_INVALID_BFREG		= BIT(31),
 };
 
 struct mlx5_ib_vma_private_data {

103 lines not on this page · raw message

index 12c523a1dd18..0f7e45680ce5 100644
@@ -41,6 +41,7 @@ enum {
 	MLX5_QP_FLAG_SIGNATURE		= 1 << 0,
 	MLX5_QP_FLAG_SCATTER_CQE	= 1 << 1,
 	MLX5_QP_FLAG_TUNNEL_OFFLOADS	= 1 << 2,
+	MLX5_QP_FLAG_BFREG_INDEX	= 1 << 3,
 };
 
 enum {

@@ -282,7 +283,7 @@ struct mlx5_ib_create_qp {
 	__u32	rq_wqe_shift;
 	__u32	flags;
 	__u32	uidx;
-	__u32	reserved0;
+	__u32	bfreg_index;
 	__u64	sq_buf_addr;
 };