qed: Add unaligned and packed packet processing
authorMichal Kalderon <Michal.Kalderon@cavium.com>
Mon, 9 Oct 2017 09:37:51 +0000 (12:37 +0300)
committerDavid S. Miller <davem@davemloft.net>
Mon, 9 Oct 2017 17:21:26 +0000 (10:21 -0700)
The fpdu data structure is preallocated per connection.
Each connection stores the current status of the connection:
either nothing pending, or there is a partial fpdu that is waiting for
the rest of the fpdu (incomplete bytes != 0).
The same structure is also used for splitting a packet when there are
packed fpdus. The structure is initialized with all data required
for sending the fpdu back to the FW. A fpdu will always be spanned across
a maximum of 3 tx bds. One for the header, one for the partial fdpu
received and one for the remainder (unaligned) packet.
In case of packed fpdu's, two fragments are used, one for the header
and one for the data.
Corner cases are not handled in the patch for clarity, and will be added
as a separate patch.

Signed-off-by: Michal Kalderon <Michal.Kalderon@cavium.com>
Signed-off-by: Ariel Elior <Ariel.Elior@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/qlogic/qed/qed_iwarp.c
drivers/net/ethernet/qlogic/qed/qed_iwarp.h

index efd4861c72e21ee02bc7584562e756f3d4e74ba0..83b147fdacdeb4eb5877b2817714a53e78130922 100644 (file)
@@ -1419,6 +1419,7 @@ void qed_iwarp_resc_free(struct qed_hwfn *p_hwfn)
 
        qed_rdma_bmap_free(p_hwfn, &p_hwfn->p_rdma_info->tcp_cid_map, 1);
        kfree(iwarp_info->mpa_bufs);
+       kfree(iwarp_info->partial_fpdus);
 }
 
 int qed_iwarp_accept(void *rdma_cxt, struct qed_iwarp_accept_in *iparams)
@@ -1716,8 +1717,170 @@ qed_iwarp_parse_rx_pkt(struct qed_hwfn *p_hwfn,
        return 0;
 }
 
+static struct qed_iwarp_fpdu *qed_iwarp_get_curr_fpdu(struct qed_hwfn *p_hwfn,
+                                                     u16 cid)
+{
+       struct qed_iwarp_info *iwarp_info = &p_hwfn->p_rdma_info->iwarp;
+       struct qed_iwarp_fpdu *partial_fpdu;
+       u32 idx;
+
+       idx = cid - qed_cxt_get_proto_cid_start(p_hwfn, PROTOCOLID_IWARP);
+       if (idx >= iwarp_info->max_num_partial_fpdus) {
+               DP_ERR(p_hwfn, "Invalid cid %x max_num_partial_fpdus=%x\n", cid,
+                      iwarp_info->max_num_partial_fpdus);
+               return NULL;
+       }
+
+       partial_fpdu = &iwarp_info->partial_fpdus[idx];
+
+       return partial_fpdu;
+}
+
+enum qed_iwarp_mpa_pkt_type {
+       QED_IWARP_MPA_PKT_PACKED,
+       QED_IWARP_MPA_PKT_PARTIAL,
+       QED_IWARP_MPA_PKT_UNALIGNED
+};
+
+#define QED_IWARP_MPA_FPDU_LENGTH_SIZE (2)
+#define QED_IWARP_MPA_CRC32_DIGEST_SIZE (4)
+
+/* Pad to multiple of 4 */
+#define QED_IWARP_PDU_DATA_LEN_WITH_PAD(data_len) ALIGN(data_len, 4)
+#define QED_IWARP_FPDU_LEN_WITH_PAD(_mpa_len)                             \
+       (QED_IWARP_PDU_DATA_LEN_WITH_PAD((_mpa_len) +                      \
+                                        QED_IWARP_MPA_FPDU_LENGTH_SIZE) + \
+                                        QED_IWARP_MPA_CRC32_DIGEST_SIZE)
+
 /* fpdu can be fragmented over maximum 3 bds: header, partial mpa, unaligned */
 #define QED_IWARP_MAX_BDS_PER_FPDU 3
+
+char *pkt_type_str[] = {
+       "QED_IWARP_MPA_PKT_PACKED",
+       "QED_IWARP_MPA_PKT_PARTIAL",
+       "QED_IWARP_MPA_PKT_UNALIGNED"
+};
+
+static enum qed_iwarp_mpa_pkt_type
+qed_iwarp_mpa_classify(struct qed_hwfn *p_hwfn,
+                      struct qed_iwarp_fpdu *fpdu,
+                      u16 tcp_payload_len, u8 *mpa_data)
+{
+       enum qed_iwarp_mpa_pkt_type pkt_type;
+       u16 mpa_len;
+
+       if (fpdu->incomplete_bytes) {
+               pkt_type = QED_IWARP_MPA_PKT_UNALIGNED;
+               goto out;
+       }
+
+       mpa_len = ntohs(*((u16 *)(mpa_data)));
+       fpdu->fpdu_length = QED_IWARP_FPDU_LEN_WITH_PAD(mpa_len);
+
+       if (fpdu->fpdu_length <= tcp_payload_len)
+               pkt_type = QED_IWARP_MPA_PKT_PACKED;
+       else
+               pkt_type = QED_IWARP_MPA_PKT_PARTIAL;
+
+out:
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+                  "MPA_ALIGN: %s: fpdu_length=0x%x tcp_payload_len:0x%x\n",
+                  pkt_type_str[pkt_type], fpdu->fpdu_length, tcp_payload_len);
+
+       return pkt_type;
+}
+
+static void
+qed_iwarp_init_fpdu(struct qed_iwarp_ll2_buff *buf,
+                   struct qed_iwarp_fpdu *fpdu,
+                   struct unaligned_opaque_data *pkt_data,
+                   u16 tcp_payload_size, u8 placement_offset)
+{
+       fpdu->mpa_buf = buf;
+       fpdu->pkt_hdr = buf->data_phys_addr + placement_offset;
+       fpdu->pkt_hdr_size = pkt_data->tcp_payload_offset;
+       fpdu->mpa_frag = buf->data_phys_addr + pkt_data->first_mpa_offset;
+       fpdu->mpa_frag_virt = (u8 *)(buf->data) + pkt_data->first_mpa_offset;
+
+       if (tcp_payload_size < fpdu->fpdu_length)
+               fpdu->incomplete_bytes = fpdu->fpdu_length - tcp_payload_size;
+       else
+               fpdu->incomplete_bytes = 0;     /* complete fpdu */
+
+       fpdu->mpa_frag_len = fpdu->fpdu_length - fpdu->incomplete_bytes;
+}
+
+static int
+qed_iwarp_send_fpdu(struct qed_hwfn *p_hwfn,
+                   struct qed_iwarp_fpdu *fpdu,
+                   struct unaligned_opaque_data *curr_pkt,
+                   struct qed_iwarp_ll2_buff *buf,
+                   u16 tcp_payload_size, enum qed_iwarp_mpa_pkt_type pkt_type)
+{
+       struct qed_ll2_tx_pkt_info tx_pkt;
+       u8 ll2_handle;
+       int rc;
+
+       memset(&tx_pkt, 0, sizeof(tx_pkt));
+
+       /* An unaligned packet means it's split over two tcp segments. So the
+        * complete packet requires 3 bds, one for the header, one for the
+        * part of the fpdu of the first tcp segment, and the last fragment
+        * will point to the remainder of the fpdu. A packed pdu, requires only
+        * two bds, one for the header and one for the data.
+        */
+       tx_pkt.num_of_bds = (pkt_type == QED_IWARP_MPA_PKT_UNALIGNED) ? 3 : 2;
+       tx_pkt.tx_dest = QED_LL2_TX_DEST_LB;
+       tx_pkt.l4_hdr_offset_w = fpdu->pkt_hdr_size >> 2; /* offset in words */
+
+       /* Send the mpa_buf only with the last fpdu (in case of packed) */
+       if (pkt_type == QED_IWARP_MPA_PKT_UNALIGNED ||
+           tcp_payload_size <= fpdu->fpdu_length)
+               tx_pkt.cookie = fpdu->mpa_buf;
+
+       tx_pkt.first_frag = fpdu->pkt_hdr;
+       tx_pkt.first_frag_len = fpdu->pkt_hdr_size;
+       tx_pkt.enable_ip_cksum = true;
+       tx_pkt.enable_l4_cksum = true;
+       tx_pkt.calc_ip_len = true;
+       /* vlan overload with enum iwarp_ll2_tx_queues */
+       tx_pkt.vlan = IWARP_LL2_ALIGNED_TX_QUEUE;
+
+       ll2_handle = p_hwfn->p_rdma_info->iwarp.ll2_mpa_handle;
+
+       /* Set first fragment to header */
+       rc = qed_ll2_prepare_tx_packet(p_hwfn, ll2_handle, &tx_pkt, true);
+       if (rc)
+               goto out;
+
+       /* Set second fragment to first part of packet */
+       rc = qed_ll2_set_fragment_of_tx_packet(p_hwfn, ll2_handle,
+                                              fpdu->mpa_frag,
+                                              fpdu->mpa_frag_len);
+       if (rc)
+               goto out;
+
+       if (!fpdu->incomplete_bytes)
+               goto out;
+
+       /* Set third fragment to second part of the packet */
+       rc = qed_ll2_set_fragment_of_tx_packet(p_hwfn,
+                                              ll2_handle,
+                                              buf->data_phys_addr +
+                                              curr_pkt->first_mpa_offset,
+                                              fpdu->incomplete_bytes);
+out:
+       DP_VERBOSE(p_hwfn,
+                  QED_MSG_RDMA,
+                  "MPA_ALIGN: Sent FPDU num_bds=%d first_frag_len=%x, mpa_frag_len=0x%x, incomplete_bytes:0x%x rc=%d\n",
+                  tx_pkt.num_of_bds,
+                  tx_pkt.first_frag_len,
+                  fpdu->mpa_frag_len,
+                  fpdu->incomplete_bytes, rc);
+
+       return rc;
+}
+
 static void
 qed_iwarp_mpa_get_data(struct qed_hwfn *p_hwfn,
                       struct unaligned_opaque_data *curr_pkt,
@@ -1741,9 +1904,79 @@ static int
 qed_iwarp_process_mpa_pkt(struct qed_hwfn *p_hwfn,
                          struct qed_iwarp_ll2_mpa_buf *mpa_buf)
 {
+       struct unaligned_opaque_data *curr_pkt = &mpa_buf->data;
        struct qed_iwarp_ll2_buff *buf = mpa_buf->ll2_buf;
+       enum qed_iwarp_mpa_pkt_type pkt_type;
+       struct qed_iwarp_fpdu *fpdu;
        int rc = -EINVAL;
+       u8 *mpa_data;
+
+       fpdu = qed_iwarp_get_curr_fpdu(p_hwfn, curr_pkt->cid & 0xffff);
+       if (!fpdu) { /* something corrupt with cid, post rx back */
+               DP_ERR(p_hwfn, "Invalid cid, drop and post back to rx cid=%x\n",
+                      curr_pkt->cid);
+               goto err;
+       }
 
+       do {
+               mpa_data = ((u8 *)(buf->data) + curr_pkt->first_mpa_offset);
+
+               pkt_type = qed_iwarp_mpa_classify(p_hwfn, fpdu,
+                                                 mpa_buf->tcp_payload_len,
+                                                 mpa_data);
+
+               switch (pkt_type) {
+               case QED_IWARP_MPA_PKT_PARTIAL:
+                       qed_iwarp_init_fpdu(buf, fpdu,
+                                           curr_pkt,
+                                           mpa_buf->tcp_payload_len,
+                                           mpa_buf->placement_offset);
+
+                       mpa_buf->tcp_payload_len = 0;
+                       break;
+               case QED_IWARP_MPA_PKT_PACKED:
+                       qed_iwarp_init_fpdu(buf, fpdu,
+                                           curr_pkt,
+                                           mpa_buf->tcp_payload_len,
+                                           mpa_buf->placement_offset);
+
+                       rc = qed_iwarp_send_fpdu(p_hwfn, fpdu, curr_pkt, buf,
+                                                mpa_buf->tcp_payload_len,
+                                                pkt_type);
+                       if (rc) {
+                               DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+                                          "Can't send FPDU:reset rc=%d\n", rc);
+                               memset(fpdu, 0, sizeof(*fpdu));
+                               break;
+                       }
+
+                       mpa_buf->tcp_payload_len -= fpdu->fpdu_length;
+                       curr_pkt->first_mpa_offset += fpdu->fpdu_length;
+                       break;
+               case QED_IWARP_MPA_PKT_UNALIGNED:
+                       rc = qed_iwarp_send_fpdu(p_hwfn, fpdu, curr_pkt, buf,
+                                                mpa_buf->tcp_payload_len,
+                                                pkt_type);
+                       if (rc) {
+                               DP_VERBOSE(p_hwfn, QED_MSG_RDMA,
+                                          "Can't send FPDU:delay rc=%d\n", rc);
+                               /* don't reset fpdu -> we need it for next
+                                * classify
+                                */
+                               break;
+                       }
+
+                       mpa_buf->tcp_payload_len -= fpdu->incomplete_bytes;
+                       curr_pkt->first_mpa_offset += fpdu->incomplete_bytes;
+                       /* The framed PDU was sent - no more incomplete bytes */
+                       fpdu->incomplete_bytes = 0;
+                       break;
+               }
+       } while (mpa_buf->tcp_payload_len && !rc);
+
+       return rc;
+
+err:
        qed_iwarp_ll2_post_rx(p_hwfn,
                              buf,
                              p_hwfn->p_rdma_info->iwarp.ll2_mpa_handle);
@@ -1989,11 +2222,27 @@ static void qed_iwarp_ll2_rel_tx_pkt(void *cxt, u8 connection_handle,
        kfree(buffer);
 }
 
+/* The only slowpath for iwarp ll2 is unalign flush. When this completion
+ * is received, need to reset the FPDU.
+ */
 void
 qed_iwarp_ll2_slowpath(void *cxt,
                       u8 connection_handle,
                       u32 opaque_data_0, u32 opaque_data_1)
 {
+       struct unaligned_opaque_data unalign_data;
+       struct qed_hwfn *p_hwfn = cxt;
+       struct qed_iwarp_fpdu *fpdu;
+
+       qed_iwarp_mpa_get_data(p_hwfn, &unalign_data,
+                              opaque_data_0, opaque_data_1);
+
+       DP_VERBOSE(p_hwfn, QED_MSG_RDMA, "(0x%x) Flush fpdu\n",
+                  unalign_data.cid);
+
+       fpdu = qed_iwarp_get_curr_fpdu(p_hwfn, (u16)unalign_data.cid);
+       if (fpdu)
+               memset(fpdu, 0, sizeof(*fpdu));
 }
 
 static int qed_iwarp_ll2_stop(struct qed_hwfn *p_hwfn, struct qed_ptt *p_ptt)
@@ -2194,6 +2443,14 @@ qed_iwarp_ll2_start(struct qed_hwfn *p_hwfn,
                                         iwarp_info->ll2_mpa_handle);
        if (rc)
                goto err;
+
+       iwarp_info->partial_fpdus = kcalloc((u16)p_hwfn->p_rdma_info->num_qps,
+                                           sizeof(*iwarp_info->partial_fpdus),
+                                           GFP_KERNEL);
+       if (!iwarp_info->partial_fpdus)
+               goto err;
+
+       iwarp_info->max_num_partial_fpdus = (u16)p_hwfn->p_rdma_info->num_qps;
        /* The mpa_bufs array serves for pending RX packets received on the
         * mpa ll2 that don't have place on the tx ring and require later
         * processing. We can't fail on allocation of such a struct therefore
index 2c53fe46345cee113092af2f2dfdd32e4d62b93d..858755cafd2bcc69ec824ecb24330dea0fbc28a5 100644 (file)
@@ -68,6 +68,17 @@ struct qed_iwarp_ll2_mpa_buf {
        u8 placement_offset;
 };
 
+struct qed_iwarp_fpdu {
+       struct qed_iwarp_ll2_buff *mpa_buf;
+       void *mpa_frag_virt;
+       dma_addr_t mpa_frag;
+       dma_addr_t pkt_hdr;
+       u16 mpa_frag_len;
+       u16 fpdu_length;
+       u16 incomplete_bytes;
+       u8 pkt_hdr_size;
+};
+
 struct qed_iwarp_info {
        struct list_head listen_list;   /* qed_iwarp_listener */
        struct list_head ep_list;       /* qed_iwarp_ep */
@@ -87,7 +98,9 @@ struct qed_iwarp_info {
        u8 peer2peer;
        enum mpa_negotiation_mode mpa_rev;
        enum mpa_rtr_type rtr_type;
+       struct qed_iwarp_fpdu *partial_fpdus;
        struct qed_iwarp_ll2_mpa_buf *mpa_bufs;
+       u16 max_num_partial_fpdus;
 };
 
 enum qed_iwarp_ep_state {