[dpdk-dev] [RFC v20.20] mbuf: introduce pktmbuf pool with pinned external buffers

[Shahaf Shuler <shahafs@mellanox.com>]

Today's pktmbuf pool contains only mbufs with no external buffers.
This means data buffer for the mbuf should be placed right after the
mbuf structure (+ the private data when enabled).

On some cases, the application would want to have the buffers allocated
from a different device in the platform. This is in order to do zero
copy for the packet directly to the device memory. Examples for such
devices can be GPU or storage device. For such cases the native pktmbuf
pool does not fit since each mbuf would need to point to external
buffer.

To support above, the pktmbuf pool will be populated with mbuf pointing
to the device buffers using the mbuf external buffer feature.
The PMD will populate its receive queues with those buffer, so that
every packet received will be scattered directly to the device memory.
on the other direction, embedding the buffer pointer to the transmit
queues of the NIC, will make the DMA to fetch device memory
using peer to peer communication.

Such mbuf with external buffer should be handled with care when mbuf is
freed. Mainly The external buffer should not be detached, so that it can
be reused for the next packet receive.

This patch introduce a new flag on the rte_pktmbuf_pool_private
structure to specify this mempool is for mbuf with pinned external
buffer. Upon detach this flag is validated and buffer is not detached.
A new mempool create wrapper is also introduced to help application to
create and populate such mempool.

Signed-off-by: Shahaf Shuler <shahafs@mellanox.com>
---
 lib/librte_mbuf/rte_mbuf.h | 75 ++++++++++++++++++++++++++++++++++++++++++----
 1 file changed, 69 insertions(+), 6 deletions(-)

diff --git a/lib/librte_mbuf/rte_mbuf.h b/lib/librte_mbuf/rte_mbuf.h
index 92d81972ab..e631dfff30 100644
--- a/lib/librte_mbuf/rte_mbuf.h
+++ b/lib/librte_mbuf/rte_mbuf.h
@@ -295,6 +295,13 @@ rte_mbuf_to_priv(struct rte_mbuf *m)
 }
 
 /**
+ * When set pktmbuf mempool will hold only mbufs with pinned external buffer.
+ * The external buffer will be attached on the mbuf creation and will not be
+ * detached by the mbuf free calls.
+ * mbuf should not contain any room for data after the mbuf structure.
+ */
+#define RTE_PKTMBUF_POOL_F_PINNED_EXT_BUF (1 << 0)
+/**
  * Private data in case of pktmbuf pool.
  *
  * A structure that contains some pktmbuf_pool-specific data that are
@@ -303,6 +310,7 @@ rte_mbuf_to_priv(struct rte_mbuf *m)
 struct rte_pktmbuf_pool_private {
 	uint16_t mbuf_data_room_size; /**< Size of data space in each mbuf. */
 	uint16_t mbuf_priv_size;      /**< Size of private area in each mbuf. */
+	uint32_t flags;		      /**< Use RTE_PKTMMBUF_POOL_F_*. */
 };
 
 #ifdef RTE_LIBRTE_MBUF_DEBUG
@@ -660,6 +668,50 @@ rte_pktmbuf_pool_create(const char *name, unsigned n,
 	int socket_id);
 
 /**
+ * Create a mbuf pool with pinned external buffers.
+ *
+ * This function creates and initializes a packet mbuf pool that contains
+ * only mbufs with external buffer. It is a wrapper to rte_mempool functions.
+ *
+ * @param name
+ *   The name of the mbuf pool.
+ * @param n
+ *   The number of elements in the mbuf pool. The optimum size (in terms
+ *   of memory usage) for a mempool is when n is a power of two minus one:
+ *   n = (2^q - 1).
+ * @param cache_size
+ *   Size of the per-core object cache. See rte_mempool_create() for
+ *   details.
+ * @param priv_size
+ *   Size of application private are between the rte_mbuf structure
+ *   and the data buffer. This value must be aligned to RTE_MBUF_PRIV_ALIGN.
+ * @param socket_id
+ *   The socket identifier where the mempool memory should be allocated. The
+ *   value can be *SOCKET_ID_ANY* if there is no NUMA constraint for the
+ *   reserved zone.
+ * @param buffers
+ *   Array of buffers to be attached to the mbufs in the pool.
+ *   Array size should be n.
+ * @param buffers_len
+ *   Array of buffer length. buffers_len[i] describes the length of a buffer
+ *   pointed by buffer[i].
+ * @return
+ *   The pointer to the new allocated mempool, on success. NULL on error
+ *   with rte_errno set appropriately. Possible rte_errno values include:
+ *    - E_RTE_NO_CONFIG - function could not get pointer to rte_config structure
+ *    - E_RTE_SECONDARY - function was called from a secondary process instance
+ *    - EINVAL - cache size provided is too large, or priv_size is not aligned.
+ *    - ENOSPC - the maximum number of memzones has already been allocated
+ *    - EEXIST - a memzone with the same name already exists
+ *    - ENOMEM - no appropriate memory area found in which to create memzone
+ */
+struct rte_mempool *
+rte_pktmbuf_ext_buffer_pool_create(const char *name, unsigned n,
+				   unsigned cache_size, uint16_t priv_size,
+				   int socket_id, void **buffers,
+				   uint16_t *buffer_len);
+
+/**
  * Create a mbuf pool with a given mempool ops name
  *
  * This function creates and initializes a packet mbuf pool. It is
@@ -1137,25 +1189,36 @@ __rte_pktmbuf_free_direct(struct rte_mbuf *m)
 static inline void rte_pktmbuf_detach(struct rte_mbuf *m)
 {
 	struct rte_mempool *mp = m->pool;
+	struct rte_pktmbuf_pool_private *priv =
+		(struct rte_pktmbuf_pool_private *)rte_mempool_get_priv(mp);
+	uint8_t pinned_ext_mbuf = priv->flags &
+				  RTE_PKTMBUF_POOL_F_PINNED_EXT_BUF;
 	uint32_t mbuf_size, buf_len;
 	uint16_t priv_size;
 
-	if (RTE_MBUF_HAS_EXTBUF(m))
-		__rte_pktmbuf_free_extbuf(m);
-	else
+	if (RTE_MBUF_HAS_EXTBUF(m)) {
+		if (pinned_ext_mbuf) {
+			m->ol_flags = EXT_ATTACHED_MBUF;
+			goto reset_data;
+		} else {
+			__rte_pktmbuf_free_extbuf(m);
+		}
+	} else {
 		__rte_pktmbuf_free_direct(m);
+	}
 
-	priv_size = rte_pktmbuf_priv_size(mp);
+	priv_size = priv->mbuf_priv_size;
 	mbuf_size = (uint32_t)(sizeof(struct rte_mbuf) + priv_size);
-	buf_len = rte_pktmbuf_data_room_size(mp);
+	buf_len = priv->mbuf_data_room_size;
 
 	m->priv_size = priv_size;
 	m->buf_addr = (char *)m + mbuf_size;
 	m->buf_iova = rte_mempool_virt2iova(m) + mbuf_size;
 	m->buf_len = (uint16_t)buf_len;
+	m->ol_flags = 0;
+reset_data:
 	rte_pktmbuf_reset_headroom(m);
 	m->data_len = 0;
-	m->ol_flags = 0;
 }
 
 /**
-- 
2.12.0