From: chunmei Liu <chunmei.liu@intel.com>
Date: Sat, 16 Dec 2017 18:45:32 +0000 (-0800)
Subject: osd: fix dpdk runtime issue based on spdk/dpdk libarary
X-Git-Tag: v13.0.2~654^2
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=c9051fbad38f3ce31cad41e3cadb9447e395630a;p=ceph.git

osd: fix dpdk runtime issue based on spdk/dpdk libarary

1.misunderstand mbuf_overhead before, so correct it.
2.spdk/dpdk library will set refcnt, so needn't set it here.
3.avoid dpdk library modification. when create memory pool set elt_size as
mbuf_overhead + mbuf_data_size to avoid dpdk library check size assert.
4. call rte_pktmbuf_prefree_seg to set mbuf->next = null avoid dpdk/lib
modification .
5. use memzon to allocate mbuf data part, so these data buf can be
processes the same way as mbuf allocated by mempool create.

Signed-off-by: chunmei Liu <chunmei.liu@intel.com>
---

diff --git a/src/msg/async/dpdk/DPDK.cc b/src/msg/async/dpdk/DPDK.cc
index 42da69af8444..e7a4ceddafed 100644
--- a/src/msg/async/dpdk/DPDK.cc
+++ b/src/msg/async/dpdk/DPDK.cc
@@ -98,7 +98,7 @@ static constexpr uint16_t mbuf_cache_size        = 512;
 static constexpr size_t mbuf_data_size = 4096;
 
 static constexpr uint16_t mbuf_overhead          =
-sizeof(struct rte_mbuf) + mbuf_data_size + RTE_PKTMBUF_HEADROOM;
+                          sizeof(struct rte_mbuf) + RTE_PKTMBUF_HEADROOM;
 //
 // We'll allocate 2K data buffers for an inline case because this would require
 // a single page per mbuf. If we used 4K data buffers here it would require 2
@@ -496,6 +496,9 @@ bool DPDKQueuePair::init_rx_mbuf_pool()
 {
   std::string name = std::string(pktmbuf_pool_name) + std::to_string(_qid) + "_rx";
 
+  int bufs_count =  cct->_conf->ms_dpdk_rx_buffer_count_per_core - mbufs_per_queue_rx;
+  int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
+  char mz_name[RTE_MEMZONE_NAMESIZE];
   // reserve the memory for Rx buffers containers
   _rx_free_pkts.reserve(mbufs_per_queue_rx);
   _rx_free_bufs.reserve(mbufs_per_queue_rx);
@@ -513,7 +516,7 @@ bool DPDKQueuePair::init_rx_mbuf_pool()
     roomsz.mbuf_data_room_size = mbuf_data_size + RTE_PKTMBUF_HEADROOM;
     _pktmbuf_pool_rx = rte_mempool_create(
         name.c_str(),
-        mbufs_per_queue_rx, mbuf_overhead,
+        mbufs_per_queue_rx, mbuf_overhead + mbuf_data_size,
         mbuf_cache_size,
         sizeof(struct rte_pktmbuf_pool_private),
         rte_pktmbuf_pool_init, as_cookie(roomsz),
@@ -529,31 +532,21 @@ bool DPDKQueuePair::init_rx_mbuf_pool()
     // 2) Bind data buffers to each of them.
     // 3) Return them back to the pool.
     //
-    for (int i = 0; i < mbufs_per_queue_rx; i++) {
-      rte_mbuf* m = rte_pktmbuf_alloc(_pktmbuf_pool_rx);
-      assert(m);
-      rte_mbuf_refcnt_set(m, 1);
-      _rx_free_bufs.push_back(m);
+    int ret = snprintf(mz_name, sizeof(mz_name),
+         "%s",  "rx_buffer_data" + std::to_string(_qid));
+    if (ret < 0 || ret >= (int)sizeof(mz_name)) {
+      return false;
     }
-
-    for (int i = 0; i < cct->_conf->ms_dpdk_rx_buffer_count_per_core; i++) {
-      void* m = rte_malloc(NULL, mbuf_data_size, mbuf_data_size);
+    const struct rte_memzone *mz = rte_memzone_reserve_aligned(mz_name, mbuf_data_size*bufs_count,
+                                   _pktmbuf_pool_rx->socket_id, mz_flags, mbuf_data_size);
+    assert(mz);
+    void* m = mz->addr;
+    for (int i = 0; i < bufs_count; i++) {
       assert(m);
       _alloc_bufs.push_back(m);
+      m += mbuf_data_size;
     }
 
-    for (auto&& m : _rx_free_bufs) {
-      if (!init_noninline_rx_mbuf(m, mbuf_data_size, _alloc_bufs)) {
-        lderr(cct) << __func__ << " Failed to allocate data buffers for Rx ring. "
-                   "Consider increasing the amount of memory." << dendl;
-        return false;
-      }
-    }
-
-    rte_mempool_put_bulk(_pktmbuf_pool_rx, (void**)_rx_free_bufs.data(),
-                         _rx_free_bufs.size());
-
-    _rx_free_bufs.clear();
     if (rte_eth_rx_queue_setup(_dev_port_idx, _qid, default_ring_size,
                                rte_eth_dev_socket_id(_dev_port_idx),
                                _dev->def_rx_conf(), _pktmbuf_pool_rx) < 0) {
@@ -800,6 +793,9 @@ bool DPDKQueuePair::rx_gc(bool force)
         break;
       }
     }
+    for (auto&& m : _rx_free_bufs) {
+      rte_pktmbuf_prefree_seg(m);
+    }
 
     if (_rx_free_bufs.size()) {
       rte_mempool_put_bulk(_pktmbuf_pool_rx,
diff --git a/src/msg/async/dpdk/DPDK.h b/src/msg/async/dpdk/DPDK.h
index 3c5c924eaa30..bf87c7be7e3f 100644
--- a/src/msg/async/dpdk/DPDK.h
+++ b/src/msg/async/dpdk/DPDK.h
@@ -626,18 +626,7 @@ class DPDKQueuePair {
     // actual data buffer.
     //
     m->buf_addr      = (char*)data - RTE_PKTMBUF_HEADROOM;
-    m->buf_physaddr  = rte_malloc_virt2iova(data) - RTE_PKTMBUF_HEADROOM;
-    return true;
-  }
-
-  static bool init_noninline_rx_mbuf(rte_mbuf* m, size_t size,
-                                     std::vector<void*> &datas) {
-    if (!refill_rx_mbuf(m, size, datas)) {
-      return false;
-    }
-    // The below fields stay constant during the execution.
-    m->buf_len       = size + RTE_PKTMBUF_HEADROOM;
-    m->data_off      = RTE_PKTMBUF_HEADROOM;
+    m->buf_physaddr  = rte_mem_virt2phy(data) - RTE_PKTMBUF_HEADROOM;
     return true;
   }