rdma support polling mode

Author: yanglimingcn

docs/cn/rdma.md

@@ -47,6 +47,8 @@ RDMA要求数据收发所使用的内存空间必须被注册（memory register
 
 RDMA是硬件相关的通信技术，有很多独特的概念，比如device、port、GID、LID、MaxSge等。这些参数在初始化时会从对应的网卡中读取出来，并且做出默认的选择（参见src/brpc/rdma/rdma_helper.cpp）。有时默认的选择并非用户的期望，则可以通过flag参数方式指定。
 
+RDMA支持事件驱动和轮询两种模式，默认是事件驱动模式，通过设置rdma_use_polling可以开启轮询模式。轮询模式下还可以设置轮询器数目（rdma_poller_num），以及是否主动放弃CPU（rdma_poller_yield）。轮询模式下还可以设置一个回调函数，在每次轮询时调用，可以配合io_uring/spdk等使用。在配合使用spdk等驱动的时候，因为spdk只支持轮询模式，并且只能在单线程使用（或者叫Run To Completion模式上使用）执行一个任务过程中不允许被调度到别的线程上，所以这时候需要设置（rdma_enable_edisp_schedule）使事件驱动程序一直占用一个worker线程，不能调度别的任务。
+
 # 参数
 
 可配置参数说明：
@@ -68,3 +70,7 @@ RDMA是硬件相关的通信技术，有很多独特的概念，比如device、p
 * rdma_memory_pool_max_regions: 最大的内存池块数，默认16
 * rdma_memory_pool_buckets: 内存池中为避免竞争采用的bucket数目，默认为4
 * rdma_memory_pool_tls_cache_num: 内存池中thread local的缓存block数目，默认为128
+* rdma_use_polling: 是否使用RDMA的轮询模式，默认false
+* rdma_poller_num: 轮询模式下的poller数目，默认1
+* rdma_poller_yield: 轮询模式下的poller是否主动放弃CPU，默认是false
+* rdma_enable_edisp_schedule: 让事件驱动器可以被调度，默认是true

src/brpc/channel.cpp

@@ -180,6 +180,9 @@ int Channel::InitChannelOptions(const ChannelOptions* options) {
             return -1;
         }
         rdma::GlobalRdmaInitializeOrDie();
+        if (!rdma::InitPollingModeWithTag(bthread_self_tag())) {
+            return -1;
+        }
 #else
         LOG(WARNING) << "Cannot use rdma since brpc does not compile with rdma";
         return -1;

src/brpc/rdma/rdma_endpoint.cpp

@@ -31,6 +31,7 @@
 #include "brpc/rdma/rdma_helper.h"
 #include "brpc/rdma/rdma_endpoint.h"
 
+DECLARE_int32(task_group_ntags);
 
 namespace brpc {
 namespace rdma {
@@ -58,6 +59,11 @@ DEFINE_int32(rdma_prepared_qp_size, 128, "SQ and RQ size for prepared QP.");
 DEFINE_int32(rdma_prepared_qp_cnt, 1024, "Initial count of prepared QP.");
 DEFINE_bool(rdma_trace_verbose, false, "Print log message verbosely");
 BRPC_VALIDATE_GFLAG(rdma_trace_verbose, brpc::PassValidate);
+DEFINE_bool(rdma_use_polling, false, "Use polling mode for RDMA.");
+DEFINE_int32(rdma_poller_num, 1, "Poller number in RDMA polling mode.");
+DEFINE_bool(rdma_poller_yield, false, "Yield thread in RDMA polling mode.");
+DEFINE_bool(rdma_enable_edisp_schedule, true,
+            "Enable event dispatcher schedule in RDMA");
 
 static const size_t IOBUF_BLOCK_HEADER_LEN = 32; // implementation-dependent
 
@@ -699,7 +705,6 @@ void* RdmaEndpoint::ProcessHandshakeAtServer(void* arg) {
         LOG_IF(INFO, FLAGS_rdma_trace_verbose) 
             << "Handshake ends (use tcp) on " << s->description();
     }
- 
     ep->TryReadOnTcp();
 
     return NULL;
@@ -1041,26 +1046,36 @@ static RdmaResource* AllocateQpCq(uint16_t sq_size, uint16_t rq_size) {
         return NULL;
     }
 
-    res->comp_channel = IbvCreateCompChannel(GetRdmaContext());
-    if (!res->comp_channel) {
-        PLOG(WARNING) << "Fail to create comp channel for CQ";
-        delete res;
-        return NULL;
-    }
+    if (!FLAGS_rdma_use_polling) {
+        res->comp_channel = IbvCreateCompChannel(GetRdmaContext());
+        if (!res->comp_channel) {
+            PLOG(WARNING) << "Fail to create comp channel for CQ";
+            delete res;
+            return NULL;
+        }
 
-    butil::make_close_on_exec(res->comp_channel->fd);
-    if (butil::make_non_blocking(res->comp_channel->fd) < 0) {
-        PLOG(WARNING) << "Fail to set comp channel nonblocking";
-        delete res;
-        return NULL;
-    }
+        butil::make_close_on_exec(res->comp_channel->fd);
+        if (butil::make_non_blocking(res->comp_channel->fd) < 0) {
+            PLOG(WARNING) << "Fail to set comp channel nonblocking";
+            delete res;
+            return NULL;
+        }
 
-    res->cq = IbvCreateCq(GetRdmaContext(), 2 * FLAGS_rdma_prepared_qp_size,
-            NULL, res->comp_channel, GetRdmaCompVector());
-    if (!res->cq) {
-        PLOG(WARNING) << "Fail to create CQ";
-        delete res;
-        return NULL;
+        res->cq = IbvCreateCq(GetRdmaContext(), 2 * FLAGS_rdma_prepared_qp_size,
+                              NULL, res->comp_channel, GetRdmaCompVector());
+        if (!res->cq) {
+            PLOG(WARNING) << "Fail to create CQ";
+            delete res;
+            return NULL;
+        }
+    } else {
+        res->cq = IbvCreateCq(GetRdmaContext(), 2 * FLAGS_rdma_prepared_qp_size,
+                              NULL, NULL, 0);
+        if (!res->cq) {
+            PLOG(WARNING) << "Fail to create CQ";
+            delete res;
+            return NULL;
+        }
     }
 
     ibv_qp_init_attr attr;
@@ -1117,19 +1132,30 @@ int RdmaEndpoint::AllocateResources() {
         return -1;
     }
 
-    SocketOptions options;
-    options.user = this;
-    options.keytable_pool = _socket->_keytable_pool;
-    options.fd = _resource->comp_channel->fd;
-    options.on_edge_triggered_events = PollCq;
-    if (Socket::Create(options, &_cq_sid) < 0) {
-        PLOG(WARNING) << "Fail to create socket for cq";
-        return -1;
-    }
+    if (!FLAGS_rdma_use_polling) {
+        SocketOptions options;
+        options.user = this;
+        options.keytable_pool = _socket->_keytable_pool;
+        options.fd = _resource->comp_channel->fd;
+        options.on_edge_triggered_events = PollCq;
+        if (Socket::Create(options, &_cq_sid) < 0) {
+            PLOG(WARNING) << "Fail to create socket for cq";
+            return -1;
+        }
 
-    if (ibv_req_notify_cq(_resource->cq, 1) < 0) {
-        PLOG(WARNING) << "Fail to arm CQ comp channel";
-        return -1;
+        if (ibv_req_notify_cq(_resource->cq, 1) < 0) {
+            PLOG(WARNING) << "Fail to arm CQ comp channel";
+            return -1;
+        }
+    } else {
+        SocketOptions options;
+        options.user = this;
+        options.keytable_pool = _socket->_keytable_pool;
+        if (Socket::Create(options, &_cq_sid) < 0) {
+            PLOG(WARNING) << "Fail to create socket for cq";
+            return -1;
+        }
+        PollerAddCqSid();
     }
 
     _sbuf.resize(_sq_size - RESERVED_WR_NUM);
@@ -1227,6 +1253,9 @@ void RdmaEndpoint::DeallocateResources() {
     if (!_resource) {
         return;
     }
+    if (FLAGS_rdma_use_polling) {
+        PollerRemoveCqSid();
+    }
     bool move_to_rdma_resource_list = false;
     if (_sq_size <= FLAGS_rdma_prepared_qp_size &&
         _rq_size <= FLAGS_rdma_prepared_qp_size &&
@@ -1237,7 +1266,10 @@ void RdmaEndpoint::DeallocateResources() {
             move_to_rdma_resource_list = true;
         }
     }
-    int fd = _resource->comp_channel->fd;
+    int fd = -1;
+    if (_resource->comp_channel) {
+        fd = _resource->comp_channel->fd;
+    }
     if (!move_to_rdma_resource_list) {
         if (_resource->qp) {
             if (IbvDestroyQp(_resource->qp) < 0) {
@@ -1327,12 +1359,14 @@ void RdmaEndpoint::PollCq(Socket* m) {
     }
     CHECK(ep == s->_rdma_ep);
 
-    if (ep->GetAndAckEvents() < 0) {
-        const int saved_errno = errno;
-        PLOG(ERROR) << "Fail to get cq event: " << s->description();
-        s->SetFailed(saved_errno, "Fail to get cq event from %s: %s",
-                s->description().c_str(), berror(saved_errno));
-        return;
+    if (!FLAGS_rdma_use_polling) {
+        if (ep->GetAndAckEvents() < 0) {
+            const int saved_errno = errno;
+            PLOG(ERROR) << "Fail to get cq event: " << s->description();
+            s->SetFailed(saved_errno, "Fail to get cq event from %s: %s",
+                         s->description().c_str(), berror(saved_errno));
+            return;
+        }
     }
 
     int progress = Socket::PROGRESS_INIT;
@@ -1349,6 +1383,9 @@ void RdmaEndpoint::PollCq(Socket* m) {
             return;
         }
         if (cnt == 0) {
+            if (FLAGS_rdma_use_polling) {
+                return;
+            }
             if (!notified) {
                 // Since RDMA only provides one shot event, we have to call the
                 // notify function every time. Because there is a possibility
@@ -1370,7 +1407,7 @@ void RdmaEndpoint::PollCq(Socket* m) {
             }
             if (ep->GetAndAckEvents() < 0) {
                 s->SetFailed(errno, "Fail to ack CQ event on %s",
-                        s->description().c_str());
+                    s->description().c_str());
                 return;
             }
             notified = false;
@@ -1474,6 +1511,10 @@ int RdmaEndpoint::GlobalInitialize() {
         g_rdma_resource_list = res;
     }
 
+    if (FLAGS_rdma_use_polling) {
+        _poller_groups = std::vector<PollerGroup>(FLAGS_task_group_ntags);
+    }
+
     return 0;
 }
 
@@ -1486,6 +1527,120 @@ void RdmaEndpoint::GlobalRelease() {
             delete res;
         }
     }
+    // release polling mode at exit or call RdmaEndpoint::PollingModeRelease
+    // explicitly
+    if (FLAGS_rdma_use_polling) {
+        for (int i = 0; i < FLAGS_task_group_ntags; ++i) {
+            PollingModeRelease(i);
+        }
+    }
+}
+
+std::vector<RdmaEndpoint::PollerGroup> RdmaEndpoint::_poller_groups;
+
+void RdmaEndpoint::SetCallbackFn(std::function<void()> cb, bthread_tag_t tag) {
+    auto& group = _poller_groups[tag];
+    auto& pollers = group.pollers;
+    for (int i = 0; i < FLAGS_rdma_poller_num; ++i) {
+        auto& poller = pollers[i];
+        std::unique_lock<bthread::Mutex> lk(poller.callback_mutex);
+        poller.callback = cb;
+    }
+}
+
+int RdmaEndpoint::PollingModeInitialize(bthread_tag_t tag) {
+    if (!FLAGS_rdma_use_polling) {
+        LOG(ERROR) << "RDMA polling mode is not enabled";
+        return -1;
+    }
+    auto& group = _poller_groups[tag];
+    auto& pollers = group.pollers;
+    auto& running = group.running;
+    bool expected = false;
+    if (!running.compare_exchange_strong(expected, true)) {
+        return 0;
+    }
+    struct FnArgs {
+        Poller* poller;
+        std::atomic<bool>* running;
+    };
+    auto fn = [](void* p) -> void* {
+        std::unique_ptr<FnArgs> args(static_cast<FnArgs*>(p));
+        auto poller = args->poller;
+        auto running = args->running;
+        std::unordered_set<SocketId> cq_sids;
+        CqSidOp op;
+        while (running->load(std::memory_order_relaxed)) {
+            while (poller->op_queue.Dequeue(op)) {
+                if (op.type == CqSidOp::ADD) {
+                    cq_sids.emplace(op.sid);
+                } else if (op.type == CqSidOp::REMOVE) {
+                    cq_sids.erase(op.sid);
+                }
+            }
+            for (auto sid : cq_sids) {
+                SocketUniquePtr s;
+                if (Socket::Address(sid, &s) < 0) {
+                    continue;
+                }
+                PollCq(s.get());
+            }
+            {
+                std::unique_lock<bthread::Mutex> lk(poller->callback_mutex);
+                if (poller->callback) {
+                    poller->callback();
+                }
+            }
+            if (FLAGS_rdma_poller_yield) {
+                bthread_yield();
+            }
+        }
+        return nullptr;
+    };
+    for (int i = 0; i < FLAGS_rdma_poller_num; ++i) {
+        auto args = new FnArgs{&pollers[i], &running};
+        bthread_attr_t attr = BTHREAD_ATTR_NORMAL;
+        attr.tag = tag;
+        auto rc = bthread_start_background(&pollers[i].tid, &attr, fn, args);
+        if (rc != 0) {
+            LOG(ERROR) << "Fail to start rdma polling bthread";
+            return -1;
+        }
+    }
+    return 0;
+}
+
+void RdmaEndpoint::PollingModeRelease(bthread_tag_t tag) {
+    if (!FLAGS_rdma_use_polling) {
+        return;
+    }
+    auto& group = _poller_groups[tag];
+    auto& pollers = group.pollers;
+    auto& running = group.running;
+    running.store(false, std::memory_order_relaxed);
+    for (int i = 0; i < FLAGS_rdma_poller_num; ++i) {
+        bthread_join(pollers[i].tid, nullptr);
+    }
+}
+
+void RdmaEndpoint::PollerAddCqSid() {
+    auto index = butil::fmix32(_cq_sid) % FLAGS_rdma_poller_num;
+    auto& group = _poller_groups[bthread_self_tag()];
+    auto& pollers = group.pollers;
+    auto& poller = pollers[index];
+    if (_cq_sid != INVALID_SOCKET_ID) {
+        poller.op_queue.Enqueue(CqSidOp{_cq_sid, CqSidOp::ADD});
+    }
+}
+
+void RdmaEndpoint::PollerRemoveCqSid() {
+    auto index = butil::fmix32(_cq_sid) % FLAGS_rdma_poller_num;
+    auto& group = _poller_groups[bthread_self_tag()];
+    auto& pollers = group.pollers;
+    auto& poller = pollers[index];
+    if (_cq_sid != INVALID_SOCKET_ID) {
+        poller.op_queue.Enqueue(CqSidOp{_cq_sid, CqSidOp::REMOVE});
+    }
 }
 
 }  // namespace rdma