Merge branch 'zero-copy-tx-20' of github.com:cloudius-systems/seastar-dev

dpdk zero-copy tx, from Vlad:

"This patch series introduces zero-copy Tx with DPDK networking backend:
 - Split the dpdk_qp mempool into separate pools for Rx and Tx queues.
 - Configure the dpdk_qp mempools to use external memory buffer when we
   can ensure pinning and virt2phys translation (currently only when
   running on top of hugetlbfs).
 - Properly divide the memory between seastar and DPDK when running on
   top of hugetlbfs.
 - Tx zero-copy itself. See more details in the PATCH7 description."

configure.py

@@ -229,7 +229,7 @@ if args.with_osv:
 if args.dpdk_target:
     args.user_cflags = (args.user_cflags +
         ' -DHAVE_DPDK -I' +
-        args.dpdk_target + '/include -Wno-error=literal-suffix -Wno-literal-suffix')
+        args.dpdk_target + '/include -Wno-error=literal-suffix -Wno-literal-suffix -Wno-invalid-offsetof')
     libs += (' -L' + args.dpdk_target + '/lib ' +
         '-Wl,--whole-archive -lrte_pmd_bond -lrte_pmd_vmxnet3_uio -lrte_pmd_virtio_uio -lrte_pmd_i40e -lrte_pmd_ixgbe -lrte_pmd_e1000 -lrte_pmd_ring -Wl,--no-whole-archive -lrte_distributor -lrte_kni -lrte_pipeline -lrte_table -lrte_port -lrte_timer -lrte_hash -lrte_lpm -lrte_power -lrte_acl -lrte_meter -lrte_sched -lrte_kvargs -lrte_mbuf -lrte_ip_frag -lethdev -lrte_eal -lrte_malloc -lrte_mempool -lrte_ring -lrte_cmdline -lrte_cfgfile -lrt -lm -ldl')
 

core/dpdk_rte.cc

@@ -1,5 +1,6 @@
 #ifdef HAVE_DPDK
 
+#include "net/dpdk.hh"
 #include "core/dpdk_rte.hh"
 #include "util/conversions.hh"
 #include <experimental/optional>
@@ -35,6 +36,18 @@ void eal::init(cpuset cpus, boost::program_options::variables_map opts)
     if (hugepages_path) {
         args.push_back(string2vector("--huge-dir"));
         args.push_back(string2vector(hugepages_path.value()));
+
+        //
+        // We don't know what is going to be our networking configuration so we
+        // assume there is going to be a queue per-CPU. Plus we'll give a DPDK
+        // 64MB for "other stuff".
+        //
+        size_t size_MB = mem_size(cpus.count()) >> 20;
+        std::stringstream size_MB_str;
+        size_MB_str << size_MB;
+
+        args.push_back(string2vector("-m"));
+        args.push_back(string2vector(size_MB_str.str()));
     } else if (!opts.count("dpdk-pmd")) {
         args.push_back(string2vector("--no-huge"));
     }
@@ -61,6 +74,23 @@ void eal::init(cpuset cpus, boost::program_options::variables_map opts)
     initialized = true;
 }
 
+size_t eal::mem_size(int num_cpus)
+{
+    size_t memsize = 0;
+    //
+    // PMD mempool memory:
+    //
+    // We don't know what is going to be our networking configuration so we
+    // assume there is going to be a queue per-CPU.
+    //
+    memsize += num_cpus * qp_mempool_obj_size();
+
+    // Plus we'll give a DPDK 64MB for "other stuff".
+    memsize += (64UL << 20);
+
+    return memsize;
+}
+
 } // namespace dpdk
 
 #endif // HAVE_DPDK

core/dpdk_rte.hh

@@ -28,6 +28,9 @@
 #define rte_mbuf_nb_segs(m)  ((m)->pkt.nb_segs)
 #define rte_mbuf_l2_len(m)   ((m)->pkt.vlan_macip.f.l2_len)
 #define rte_mbuf_l3_len(m)   ((m)->pkt.vlan_macip.f.l3_len)
+#define rte_mbuf_buf_addr(m) ((m)->pkt.buf_addr)
+#define rte_mbuf_buf_physaddr(m) ((m)->pkt.buf_physaddr)
+#define rte_mbuf_data_off(m) ((m)->pkt.data_off)
 
 #else
 
@@ -39,6 +42,9 @@
 #define rte_mbuf_nb_segs(m)  ((m)->nb_segs)
 #define rte_mbuf_l2_len(m)   ((m)->l2_len)
 #define rte_mbuf_l3_len(m)   ((m)->l3_len)
+#define rte_mbuf_buf_addr(m) ((m)->buf_addr)
+#define rte_mbuf_buf_physaddr(m) ((m)->buf_physaddr)
+#define rte_mbuf_data_off(m) ((m)->data_off)
 
 #endif
 
@@ -52,6 +58,13 @@ public:
     using cpuset = std::bitset<RTE_MAX_LCORE>;
 
     static void init(cpuset cpus, boost::program_options::variables_map opts);
+    /**
+     * Returns the amount of memory needed for DPDK
+     * @param num_cpus Number of CPUs the application is going to use
+     *
+     * @return
+     */
+    static size_t mem_size(int num_cpus);
     static bool initialized;
 };
 

core/memory.cc

@@ -56,9 +56,6 @@
 
 namespace memory {
 
-static constexpr const size_t page_bits = 12;
-static constexpr const size_t page_size = 1 << page_bits;
-static constexpr const size_t huge_page_size = 512 * page_size;
 static constexpr const unsigned cpu_id_shift = 36; // FIXME: make dynamic
 static constexpr const unsigned max_cpus = 256;
 static constexpr const size_t cache_line_size = 64;

core/memory.hh

@@ -12,6 +12,11 @@
 
 namespace memory {
 
+// TODO: Use getpagesize() in order to learn a size of a system PAGE.
+static constexpr const size_t page_bits = 12;
+static constexpr const size_t page_size = 1 << page_bits;       // 4K
+static constexpr const size_t huge_page_size = 512 * page_size; // 2M
+
 void configure(std::vector<resource::memory> m,
         std::experimental::optional<std::string> hugetlbfs_path = {});
 

core/reactor.cc

@@ -1268,6 +1268,26 @@ void smp::configure(boost::program_options::variables_map configuration)
     resource::configuration rc;
     if (configuration.count("memory")) {
         rc.total_memory = parse_memory_size(configuration["memory"].as<std::string>());
+#ifdef HAVE_DPDK
+        if (configuration.count("hugepages") &&
+            !configuration["network-stack"].as<std::string>().compare("native") &&
+            configuration.count("dpdk-pmd")) {
+            size_t dpdk_memory = dpdk::eal::mem_size(smp::count);
+
+            if (dpdk_memory >= rc.total_memory) {
+                std::cerr<<"Can't run with the given amount of memory: ";
+                std::cerr<<configuration["memory"].as<std::string>();
+                std::cerr<<". Consider giving more."<<std::endl;
+                exit(1);
+            }
+
+            //
+            // Subtract the memory we are about to give to DPDK from the total
+            // amount of memory we are allowed to use.
+            //
+            rc.total_memory.value() -= dpdk_memory;
+        }
+#endif
     }
     if (configuration.count("reserve-memory")) {
         rc.reserve_memory = parse_memory_size(configuration["reserve-memory"].as<std::string>());

net/dpdk.hh

@@ -17,6 +17,13 @@ std::unique_ptr<net::device> create_dpdk_net_device(
 
 boost::program_options::options_description get_dpdk_net_options_description();
 
+namespace dpdk {
+/**
+ * @return Number of bytes needed for mempool objects of each QP.
+ */
+uint32_t qp_mempool_obj_size();
+}
+
 #endif // _SEASTAR_DPDK_DEV_H
 
 #endif // HAVE_DPDK