bnxt_en: add support for HWRM request slices

Slices are a mechanism for suballocating DMA mapped regions from the
request buffer. Such regions can be used for indirect command data
instead of creating new mappings with dma_alloc_coherent().

The advantage of using a slice is that the lifetime of the slice is
bound to the request and will be automatically unmapped when the
request is consumed.

A single external region is also supported. This allows for regions
that will not fit inside the spare request buffer space such that
the same API can be used consistently even for larger mappings.
Signed-off-by: Edwin Peer <edwin.peer@broadcom.com>
Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/broadcom/bnxt/bnxt_hwrm.c

@@ -91,6 +91,9 @@ int __hwrm_req_init(struct bnxt *bp, void **req, u16 req_type, u32 req_len)
 	ctx->dma_handle = dma_handle;
 	ctx->flags = 0; /* __GFP_ZERO, but be explicit regarding ownership */
 	ctx->timeout = bp->hwrm_cmd_timeout ?: DFLT_HWRM_CMD_TIMEOUT;
+	ctx->allocated = BNXT_HWRM_DMA_SIZE - BNXT_HWRM_CTX_OFFSET;
+	ctx->gfp = GFP_KERNEL;
+	ctx->slice_addr = NULL;
 
 	/* initialize common request fields */
 	ctx->req->req_type = cpu_to_le16(req_type);
@@ -147,6 +150,29 @@ void hwrm_req_timeout(struct bnxt *bp, void *req, unsigned int timeout)
 		ctx->timeout = timeout;
 }
 
+/**
+ * hwrm_req_alloc_flags() - Sets GFP allocation flags for slices.
+ * @bp: The driver context.
+ * @req: The request for which calls to hwrm_req_dma_slice() will have altered
+ *	allocation flags.
+ * @flags: A bitmask of GFP flags. These flags are passed to
+ *	dma_alloc_coherent() whenever it is used to allocate backing memory
+ *	for slices. Note that calls to hwrm_req_dma_slice() will not always
+ *	result in new allocations, however, memory suballocated from the
+ *	request buffer is already __GFP_ZERO.
+ *
+ * Sets the GFP allocation flags associated with the request for subsequent
+ * calls to hwrm_req_dma_slice(). This can be useful for specifying __GFP_ZERO
+ * for slice allocations.
+ */
+void hwrm_req_alloc_flags(struct bnxt *bp, void *req, gfp_t gfp)
+{
+	struct bnxt_hwrm_ctx *ctx = __hwrm_ctx(bp, req);
+
+	if (ctx)
+		ctx->gfp = gfp;
+}
+
 /**
  * hwrm_req_replace() - Replace request data.
  * @bp: The driver context.
@@ -166,7 +192,8 @@ void hwrm_req_timeout(struct bnxt *bp, void *req, unsigned int timeout)
  * reference the new request and use it in lieu of req during subsequent
  * calls to hwrm_req_send(). The resource management is associated with
  * req and is independent of and does not apply to new_req. The caller must
- * ensure that the lifetime of new_req is least as long as req.
+ * ensure that the lifetime of new_req is least as long as req. Any slices
+ * that may have been associated with the original request are released.
  *
  * Return: zero on success, negative error code otherwise:
  *     E2BIG: Request is too large.
@@ -184,6 +211,15 @@ int hwrm_req_replace(struct bnxt *bp, void *req, void *new_req, u32 len)
 	if (len > BNXT_HWRM_CTX_OFFSET)
 		return -E2BIG;
 
+	/* free any existing slices */
+	ctx->allocated = BNXT_HWRM_DMA_SIZE - BNXT_HWRM_CTX_OFFSET;
+	if (ctx->slice_addr) {
+		dma_free_coherent(&bp->pdev->dev, ctx->slice_size,
+				  ctx->slice_addr, ctx->slice_handle);
+		ctx->slice_addr = NULL;
+	}
+	ctx->gfp = GFP_KERNEL;
+
 	if ((bp->fw_cap & BNXT_FW_CAP_SHORT_CMD) || len > BNXT_HWRM_MAX_REQ_LEN) {
 		memcpy(internal_req, new_req, len);
 	} else {
@@ -274,6 +310,11 @@ static void __hwrm_ctx_drop(struct bnxt *bp, struct bnxt_hwrm_ctx *ctx)
 	void *addr = ((u8 *)ctx) - BNXT_HWRM_CTX_OFFSET;
 	dma_addr_t dma_handle = ctx->dma_handle; /* save before invalidate */
 
+	/* unmap any auxiliary DMA slice */
+	if (ctx->slice_addr)
+		dma_free_coherent(&bp->pdev->dev, ctx->slice_size,
+				  ctx->slice_addr, ctx->slice_handle);
+
 	/* invalidate, ensure ownership, sentinel and dma_handle are cleared */
 	memset(ctx, 0, sizeof(struct bnxt_hwrm_ctx));
 
@@ -286,7 +327,8 @@ static void __hwrm_ctx_drop(struct bnxt *bp, struct bnxt_hwrm_ctx *ctx)
  * hwrm_req_drop() - Release all resources associated with the request.
  * @bp: The driver context.
  * @req: The request to consume, releasing the associated resources. The
- *	request object and its associated response are no longer valid.
+ *	request object, any slices, and its associated response are no
+ *	longer valid.
  *
  * It is legal to call hwrm_req_drop() on an unowned request, provided it
  * has not already been consumed by hwrm_req_send() (for example, to release
@@ -671,3 +713,72 @@ int hwrm_req_send_silent(struct bnxt *bp, void *req)
 	hwrm_req_flags(bp, req, BNXT_HWRM_CTX_SILENT);
 	return hwrm_req_send(bp, req);
 }
+
+/**
+ * hwrm_req_dma_slice() - Allocate a slice of DMA mapped memory.
+ * @bp: The driver context.
+ * @req: The request for which indirect data will be associated.
+ * @size: The size of the allocation.
+ * @dma: The bus address associated with the allocation. The HWRM API has no
+ *	knowledge about the type of the request and so cannot infer how the
+ *	caller intends to use the indirect data. Thus, the caller is
+ *	responsible for configuring the request object appropriately to
+ *	point to the associated indirect memory. Note, DMA handle has the
+ *	same definition as it does in dma_alloc_coherent(), the caller is
+ *	responsible for endian conversions via cpu_to_le64() before assigning
+ *	this address.
+ *
+ * Allocates DMA mapped memory for indirect data related to a request. The
+ * lifetime of the DMA resources will be bound to that of the request (ie.
+ * they will be automatically released when the request is either consumed by
+ * hwrm_req_send() or dropped by hwrm_req_drop()). Small allocations are
+ * efficiently suballocated out of the request buffer space, hence the name
+ * slice, while larger requests are satisfied via an underlying call to
+ * dma_alloc_coherent(). Multiple suballocations are supported, however, only
+ * one externally mapped region is.
+ *
+ * Return: The kernel virtual address of the DMA mapping.
+ */
+void *
+hwrm_req_dma_slice(struct bnxt *bp, void *req, u32 size, dma_addr_t *dma_handle)
+{
+	struct bnxt_hwrm_ctx *ctx = __hwrm_ctx(bp, req);
+	u8 *end = ((u8 *)req) + BNXT_HWRM_DMA_SIZE;
+	struct input *input = req;
+	u8 *addr, *req_addr = req;
+	u32 max_offset, offset;
+
+	if (!ctx)
+		return NULL;
+
+	max_offset = BNXT_HWRM_DMA_SIZE - ctx->allocated;
+	offset = max_offset - size;
+	offset = ALIGN_DOWN(offset, BNXT_HWRM_DMA_ALIGN);
+	addr = req_addr + offset;
+
+	if (addr < req_addr + max_offset && req_addr + ctx->req_len <= addr) {
+		ctx->allocated = end - addr;
+		*dma_handle = ctx->dma_handle + offset;
+		return addr;
+	}
+
+	/* could not suballocate from ctx buffer, try create a new mapping */
+	if (ctx->slice_addr) {
+		/* if one exists, can only be due to software bug, be loud */
+		netdev_err(bp->dev, "HWRM refusing to reallocate DMA slice, req_type = %u\n",
+			   (u32)le16_to_cpu(input->req_type));
+		dump_stack();
+		return NULL;
+	}
+
+	addr = dma_alloc_coherent(&bp->pdev->dev, size, dma_handle, ctx->gfp);
+
+	if (!addr)
+		return NULL;
+
+	ctx->slice_addr = addr;
+	ctx->slice_size = size;
+	ctx->slice_handle = *dma_handle;
+
+	return addr;
+}

drivers/net/ethernet/broadcom/bnxt/bnxt_hwrm.h

@@ -27,9 +27,14 @@ struct bnxt_hwrm_ctx {
 	dma_addr_t dma_handle;
 	struct output *resp;
 	struct input *req;
+	dma_addr_t slice_handle;
+	void *slice_addr;
+	u32 slice_size;
 	u32 req_len;
 	enum bnxt_hwrm_ctx_flags flags;
 	unsigned int timeout;
+	u32 allocated;
+	gfp_t gfp;
 };
 
 #define BNXT_HWRM_MAX_REQ_LEN		(bp->hwrm_max_req_len)
@@ -140,4 +145,6 @@ void hwrm_req_timeout(struct bnxt *bp, void *req, unsigned int timeout);
 int hwrm_req_send(struct bnxt *bp, void *req);
 int hwrm_req_send_silent(struct bnxt *bp, void *req);
 int hwrm_req_replace(struct bnxt *bp, void *req, void *new_req, u32 len);
+void hwrm_req_alloc_flags(struct bnxt *bp, void *req, gfp_t flags);
+void *hwrm_req_dma_slice(struct bnxt *bp, void *req, u32 size, dma_addr_t *dma);
 #endif