IB/umad: Add support for large RMPP transfers

Add support for sending and receiving large RMPP transfers.  The old
code supports transfers only as large as a single contiguous kernel
memory allocation.  This patch uses linked list of memory buffers when
sending and receiving data to avoid needing contiguous pages for
larger transfers.

  Receive side: copy the arriving MADs in chunks instead of coalescing
  to one large buffer in kernel space.

  Send side: split a multipacket MAD buffer to a list of segments,
  (multipacket_list) and send these using a gather list of size 2.
  Also, save pointer to last sent segment, and retrieve requested
  segments by walking list starting at last sent segment. Finally,
  save pointer to last-acked segment.  When retrying, retrieve
  segments for resending relative to this pointer.  When updating last
  ack, start at this pointer.
Signed-off-by: Jack Morgenstein <jackm@mellanox.co.il>
Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>

drivers/infiniband/core/mad.c

@@ -31,7 +31,7 @@
  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  *
- * $Id: mad.c 2817 2005-07-07 11:29:26Z halr $
+ * $Id: mad.c 5596 2006-03-03 01:00:07Z sean.hefty $
  */
 #include <linux/dma-mapping.h>
 
@@ -765,18 +765,67 @@ out:
 	return ret;
 }
 
-static int get_buf_length(int hdr_len, int data_len)
+static int get_pad_size(int hdr_len, int data_len)
 {
 	int seg_size, pad;
 
 	seg_size = sizeof(struct ib_mad) - hdr_len;
 	if (data_len && seg_size) {
 		pad = seg_size - data_len % seg_size;
-		if (pad == seg_size)
-			pad = 0;
+		return pad == seg_size ? 0 : pad;
 	} else
-		pad = seg_size;
-	return hdr_len + data_len + pad;
+		return seg_size;
+}
+
+static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
+{
+	struct ib_rmpp_segment *s, *t;
+
+	list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
+		list_del(&s->list);
+		kfree(s);
+	}
+}
+
+static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
+				gfp_t gfp_mask)
+{
+	struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
+	struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
+	struct ib_rmpp_segment *seg = NULL;
+	int left, seg_size, pad;
+
+	send_buf->seg_size = sizeof (struct ib_mad) - send_buf->hdr_len;
+	seg_size = send_buf->seg_size;
+	pad = send_wr->pad;
+
+	/* Allocate data segments. */
+	for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
+		seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
+		if (!seg) {
+			printk(KERN_ERR "alloc_send_rmpp_segs: RMPP mem "
+			       "alloc failed for len %zd, gfp %#x\n",
+			       sizeof (*seg) + seg_size, gfp_mask);
+			free_send_rmpp_list(send_wr);
+			return -ENOMEM;
+		}
+		seg->num = ++send_buf->seg_count;
+		list_add_tail(&seg->list, &send_wr->rmpp_list);
+	}
+
+	/* Zero any padding */
+	if (pad)
+		memset(seg->data + seg_size - pad, 0, pad);
+
+	rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
+					  agent.rmpp_version;
+	rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
+	ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
+
+	send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
+					struct ib_rmpp_segment, list);
+	send_wr->last_ack_seg = send_wr->cur_seg;
+	return 0;
 }
 
 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
@@ -787,32 +836,40 @@ struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
 {
 	struct ib_mad_agent_private *mad_agent_priv;
 	struct ib_mad_send_wr_private *mad_send_wr;
-	int buf_size;
+	int pad, message_size, ret, size;
 	void *buf;
 
 	mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
 				      agent);
-	buf_size = get_buf_length(hdr_len, data_len);
+	pad = get_pad_size(hdr_len, data_len);
+	message_size = hdr_len + data_len + pad;
 
 	if ((!mad_agent->rmpp_version &&
-	     (rmpp_active || buf_size > sizeof(struct ib_mad))) ||
-	    (!rmpp_active && buf_size > sizeof(struct ib_mad)))
+	     (rmpp_active || message_size > sizeof(struct ib_mad))) ||
+	    (!rmpp_active && message_size > sizeof(struct ib_mad)))
 		return ERR_PTR(-EINVAL);
 
-	buf = kzalloc(sizeof *mad_send_wr + buf_size, gfp_mask);
+	size = rmpp_active ? hdr_len : sizeof(struct ib_mad);
+	buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
 	if (!buf)
 		return ERR_PTR(-ENOMEM);
 
-	mad_send_wr = buf + buf_size;
+	mad_send_wr = buf + size;
+	INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
 	mad_send_wr->send_buf.mad = buf;
+	mad_send_wr->send_buf.hdr_len = hdr_len;
+	mad_send_wr->send_buf.data_len = data_len;
+	mad_send_wr->pad = pad;
 
 	mad_send_wr->mad_agent_priv = mad_agent_priv;
-	mad_send_wr->sg_list[0].length = buf_size;
+	mad_send_wr->sg_list[0].length = hdr_len;
 	mad_send_wr->sg_list[0].lkey = mad_agent->mr->lkey;
+	mad_send_wr->sg_list[1].length = sizeof(struct ib_mad) - hdr_len;
+	mad_send_wr->sg_list[1].lkey = mad_agent->mr->lkey;
 
 	mad_send_wr->send_wr.wr_id = (unsigned long) mad_send_wr;
 	mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
-	mad_send_wr->send_wr.num_sge = 1;
+	mad_send_wr->send_wr.num_sge = 2;
 	mad_send_wr->send_wr.opcode = IB_WR_SEND;
 	mad_send_wr->send_wr.send_flags = IB_SEND_SIGNALED;
 	mad_send_wr->send_wr.wr.ud.remote_qpn = remote_qpn;
@@ -820,13 +877,11 @@ struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
 	mad_send_wr->send_wr.wr.ud.pkey_index = pkey_index;
 
 	if (rmpp_active) {
-		struct ib_rmpp_mad *rmpp_mad = mad_send_wr->send_buf.mad;
-		rmpp_mad->rmpp_hdr.paylen_newwin = cpu_to_be32(hdr_len -
-						   IB_MGMT_RMPP_HDR + data_len);
-		rmpp_mad->rmpp_hdr.rmpp_version = mad_agent->rmpp_version;
-		rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
-		ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr,
-				  IB_MGMT_RMPP_FLAG_ACTIVE);
+		ret = alloc_send_rmpp_list(mad_send_wr, gfp_mask);
+		if (ret) {
+			kfree(buf);
+			return ERR_PTR(ret);
+		}
 	}
 
 	mad_send_wr->send_buf.mad_agent = mad_agent;
@@ -835,14 +890,50 @@ struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
 }
 EXPORT_SYMBOL(ib_create_send_mad);
 
+void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
+{
+	struct ib_mad_send_wr_private *mad_send_wr;
+	struct list_head *list;
+
+	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
+				   send_buf);
+	list = &mad_send_wr->cur_seg->list;
+
+	if (mad_send_wr->cur_seg->num < seg_num) {
+		list_for_each_entry(mad_send_wr->cur_seg, list, list)
+			if (mad_send_wr->cur_seg->num == seg_num)
+				break;
+	} else if (mad_send_wr->cur_seg->num > seg_num) {
+		list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
+			if (mad_send_wr->cur_seg->num == seg_num)
+				break;
+	}
+	return mad_send_wr->cur_seg->data;
+}
+EXPORT_SYMBOL(ib_get_rmpp_segment);
+
+static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
+{
+	if (mad_send_wr->send_buf.seg_count)
+		return ib_get_rmpp_segment(&mad_send_wr->send_buf,
+					   mad_send_wr->seg_num);
+	else
+		return mad_send_wr->send_buf.mad +
+		       mad_send_wr->send_buf.hdr_len;
+}
+
 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
 {
 	struct ib_mad_agent_private *mad_agent_priv;
+	struct ib_mad_send_wr_private *mad_send_wr;
 
 	mad_agent_priv = container_of(send_buf->mad_agent,
 				      struct ib_mad_agent_private, agent);
-	kfree(send_buf->mad);
+	mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
+				   send_buf);
 
+	free_send_rmpp_list(mad_send_wr);
+	kfree(send_buf->mad);
 	if (atomic_dec_and_test(&mad_agent_priv->refcount))
 		wake_up(&mad_agent_priv->wait);
 }
@@ -865,10 +956,17 @@ int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
 
 	mad_agent = mad_send_wr->send_buf.mad_agent;
 	sge = mad_send_wr->sg_list;
-	sge->addr = dma_map_single(mad_agent->device->dma_device,
-				   mad_send_wr->send_buf.mad, sge->length,
-				   DMA_TO_DEVICE);
-	pci_unmap_addr_set(mad_send_wr, mapping, sge->addr);
+	sge[0].addr = dma_map_single(mad_agent->device->dma_device,
+				     mad_send_wr->send_buf.mad,
+				     sge[0].length,
+				     DMA_TO_DEVICE);
+	pci_unmap_addr_set(mad_send_wr, header_mapping, sge[0].addr);
+
+	sge[1].addr = dma_map_single(mad_agent->device->dma_device,
+				     ib_get_payload(mad_send_wr),
+				     sge[1].length,
+				     DMA_TO_DEVICE);
+	pci_unmap_addr_set(mad_send_wr, payload_mapping, sge[1].addr);
 
 	spin_lock_irqsave(&qp_info->send_queue.lock, flags);
 	if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
@@ -885,11 +983,14 @@ int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
 		list_add_tail(&mad_send_wr->mad_list.list, list);
 	}
 	spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
-	if (ret)
+	if (ret) {
 		dma_unmap_single(mad_agent->device->dma_device,
-				 pci_unmap_addr(mad_send_wr, mapping),
-				 sge->length, DMA_TO_DEVICE);
-
+				 pci_unmap_addr(mad_send_wr, header_mapping),
+				 sge[0].length, DMA_TO_DEVICE);
+		dma_unmap_single(mad_agent->device->dma_device,
+				 pci_unmap_addr(mad_send_wr, payload_mapping),
+				 sge[1].length, DMA_TO_DEVICE);
+	}
 	return ret;
 }
 
@@ -1860,8 +1961,11 @@ static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
 
 retry:
 	dma_unmap_single(mad_send_wr->send_buf.mad_agent->device->dma_device,
-			 pci_unmap_addr(mad_send_wr, mapping),
+			 pci_unmap_addr(mad_send_wr, header_mapping),
 			 mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
+	dma_unmap_single(mad_send_wr->send_buf.mad_agent->device->dma_device,
+			 pci_unmap_addr(mad_send_wr, payload_mapping),
+			 mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
 	queued_send_wr = NULL;
 	spin_lock_irqsave(&send_queue->lock, flags);
 	list_del(&mad_list->list);

drivers/infiniband/core/mad_priv.h

@@ -31,7 +31,7 @@
  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  *
- * $Id: mad_priv.h 2730 2005-06-28 16:43:03Z sean.hefty $
+ * $Id: mad_priv.h 5596 2006-03-03 01:00:07Z sean.hefty $
  */
 
 #ifndef __IB_MAD_PRIV_H__
@@ -85,6 +85,12 @@ struct ib_mad_private {
 	} mad;
 } __attribute__ ((packed));
 
+struct ib_rmpp_segment {
+	struct list_head list;
+	u32 num;
+	u8 data[0];
+};
+
 struct ib_mad_agent_private {
 	struct list_head agent_list;
 	struct ib_mad_agent agent;
@@ -119,7 +125,8 @@ struct ib_mad_send_wr_private {
 	struct list_head agent_list;
 	struct ib_mad_agent_private *mad_agent_priv;
 	struct ib_mad_send_buf send_buf;
-	DECLARE_PCI_UNMAP_ADDR(mapping)
+	DECLARE_PCI_UNMAP_ADDR(header_mapping)
+	DECLARE_PCI_UNMAP_ADDR(payload_mapping)
 	struct ib_send_wr send_wr;
 	struct ib_sge sg_list[IB_MAD_SEND_REQ_MAX_SG];
 	__be64 tid;
@@ -130,11 +137,12 @@ struct ib_mad_send_wr_private {
 	enum ib_wc_status status;
 
 	/* RMPP control */
+	struct list_head rmpp_list;
+	struct ib_rmpp_segment *last_ack_seg;
+	struct ib_rmpp_segment *cur_seg;
 	int last_ack;
 	int seg_num;
 	int newwin;
-	int total_seg;
-	int data_offset;
 	int pad;
 };
 

include/rdma/ib_mad.h

@@ -33,7 +33,7 @@
  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  *
- * $Id: ib_mad.h 2775 2005-07-02 13:42:12Z halr $
+ * $Id: ib_mad.h 5596 2006-03-03 01:00:07Z sean.hefty $
  */
 
 #if !defined( IB_MAD_H )
@@ -208,15 +208,23 @@ struct ib_class_port_info
 /**
  * ib_mad_send_buf - MAD data buffer and work request for sends.
  * @next: A pointer used to chain together MADs for posting.
- * @mad: References an allocated MAD data buffer.
+ * @mad: References an allocated MAD data buffer for MADs that do not have
+ *   RMPP active.  For MADs using RMPP, references the common and management
+ *   class specific headers.
  * @mad_agent: MAD agent that allocated the buffer.
  * @ah: The address handle to use when sending the MAD.
  * @context: User-controlled context fields.
+ * @hdr_len: Indicates the size of the data header of the MAD.  This length
+ *   includes the common MAD, RMPP, and class specific headers.
+ * @data_len: Indicates the total size of user-transferred data.
+ * @seg_count: The number of RMPP segments allocated for this send.
+ * @seg_size: Size of each RMPP segment.
  * @timeout_ms: Time to wait for a response.
  * @retries: Number of times to retry a request for a response.
  *
  * Users are responsible for initializing the MAD buffer itself, with the
- * exception of specifying the payload length field in any RMPP MAD.
+ * exception of any RMPP header.  Additional segment buffer space allocated
+ * beyond data_len is padding.
  */
 struct ib_mad_send_buf {
 	struct ib_mad_send_buf	*next;
@@ -224,6 +232,10 @@ struct ib_mad_send_buf {
 	struct ib_mad_agent	*mad_agent;
 	struct ib_ah		*ah;
 	void			*context[2];
+	int			hdr_len;
+	int			data_len;
+	int			seg_count;
+	int			seg_size;
 	int			timeout_ms;
 	int			retries;
 };
@@ -299,7 +311,7 @@ typedef void (*ib_mad_snoop_handler)(struct ib_mad_agent *mad_agent,
  * @mad_recv_wc: Received work completion information on the received MAD.
  *
  * MADs received in response to a send request operation will be handed to
- * the user after the send operation completes.  All data buffers given
+ * the user before the send operation completes.  All data buffers given
  * to registered agents through this routine are owned by the receiving
  * client, except for snooping agents.  Clients snooping MADs should not
  * modify the data referenced by @mad_recv_wc.
@@ -485,17 +497,6 @@ int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent);
 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
 		     struct ib_mad_send_buf **bad_send_buf);
 
-/**
- * ib_coalesce_recv_mad - Coalesces received MAD data into a single buffer.
- * @mad_recv_wc: Work completion information for a received MAD.
- * @buf: User-provided data buffer to receive the coalesced buffers.  The
- *   referenced buffer should be at least the size of the mad_len specified
- *   by @mad_recv_wc.
- *
- * This call copies a chain of received MAD segments into a single data buffer,
- * removing duplicated headers.
- */
-void ib_coalesce_recv_mad(struct ib_mad_recv_wc *mad_recv_wc, void *buf);
 
 /**
  * ib_free_recv_mad - Returns data buffers used to receive a MAD.
@@ -590,9 +591,10 @@ int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
  * with an initialized work request structure.  Users may modify the returned
  * MAD data buffer before posting the send.
  *
- * The returned data buffer will be cleared.  Users are responsible for
- * initializing the common MAD and any class specific headers.  If @rmpp_active
- * is set, the RMPP header will be initialized for sending.
+ * The returned MAD header, class specific headers, and any padding will be
+ * cleared.  Users are responsible for initializing the common MAD header,
+ * any class specific header, and MAD data area.
+ * If @rmpp_active is set, the RMPP header will be initialized for sending.
  */
 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
 					    u32 remote_qpn, u16 pkey_index,
@@ -600,6 +602,16 @@ struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
 					    int hdr_len, int data_len,
 					    gfp_t gfp_mask);
 
+/**
+ * ib_get_rmpp_segment - returns the data buffer for a given RMPP segment.
+ * @send_buf: Previously allocated send data buffer.
+ * @seg_num: number of segment to return
+ *
+ * This routine returns a pointer to the data buffer of an RMPP MAD.
+ * Users must provide synchronization to @send_buf around this call.
+ */
+void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num);
+
 /**
  * ib_free_send_mad - Returns data buffers used to send a MAD.
  * @send_buf: Previously allocated send data buffer.