move strcut FlagcxSlice from cc to h

Author: leoda1

flagcx/adaptor/net/ibrc_p2p_adaptor.cc

@@ -27,8 +27,6 @@
 #include <unistd.h>
 #include <vector>
 
-struct FlagcxSlice;
-
 extern struct ibv_cq *flagcxP2pPoolGetSharedCq(int ibDevN,
                                                struct ibv_context *ctx);
 extern void flagcxP2pPoolRegisterQp(int ibDevN, void *sendComm,
@@ -41,45 +39,6 @@ extern flagcxResult_t flagcxP2pPoolSubmit(int ibDevN, void *sendComm,
 /*  Internal structs                                                   */
 /* ------------------------------------------------------------------ */
 
-struct FlagcxTransferTask {
-  std::atomic<uint64_t> sliceCount{0};
-  std::atomic<uint64_t> doneSliceCount{0};
-  std::vector<FlagcxSlice *> sliceList;
-
-  bool isAllDone() const {
-    auto total = sliceCount.load(std::memory_order_acquire);
-    auto done = doneSliceCount.load(std::memory_order_acquire);
-    return total > 0 && done >= total;
-  }
-};
-
-enum FlagcxSliceOp : uint8_t {
-  FLAGCX_SLICE_OP_WRITE = 0,
-  FLAGCX_SLICE_OP_READ = 1,
-};
-
-struct FlagcxSlice {
-  uint64_t srcVa;
-  uint64_t dstVa;
-  uint32_t length;
-  uint32_t lkey;
-  uint32_t rkey;
-  uint8_t opcode;
-  std::string peerNicPath;
-  FlagcxTransferTask *task;
-  volatile int *qpDepth;
-
-  inline void markSuccess() {
-    if (task)
-      task->doneSliceCount.fetch_add(1, std::memory_order_release);
-  }
-
-  inline void markFailed() {
-    if (task)
-      task->doneSliceCount.fetch_add(1, std::memory_order_release);
-  }
-};
-
 // Per-device context — created at init, holds eagerly allocated PD.
 // Passed as the `comm` parameter to regMr/deregMr when no connection exists.
 // ibDevN MUST be the first field so regMr can cast any comm pointer to extract

flagcx/core/flagcx_p2p.cc

@@ -44,8 +44,6 @@
 
 extern struct flagcxNetAdaptor flagcxNetIbP2p;
 
-struct FlagcxSlice;
-
 extern "C" flagcxResult_t flagcxP2pSliceBatch(void *sendComm,
                                               struct ibv_qp *qp, int count,
                                               FlagcxSlice **slices);
@@ -301,47 +299,6 @@ static std::unordered_map<uint64_t, FlagcxP2pXfer> gXferMap;
 static std::mutex gXferMutex;
 static uint64_t gNextXferId = 1;
 
-struct FlagcxTransferTask {
-  std::atomic<uint64_t> sliceCount{0};
-  std::atomic<uint64_t> doneSliceCount{0};
-  std::vector<FlagcxSlice *> sliceList;
-
-  bool isAllDone() const {
-    auto total = sliceCount.load(std::memory_order_acquire);
-    auto done = doneSliceCount.load(std::memory_order_acquire);
-    return total > 0 && done >= total;
-  }
-};
-
-enum FlagcxSliceOp : uint8_t {
-  FLAGCX_SLICE_OP_WRITE = 0,
-  FLAGCX_SLICE_OP_READ = 1,
-};
-
-struct FlagcxSlice {
-  // WRITE: local source VA; READ: local destination VA.
-  uint64_t srcVa = 0;
-  // WRITE: remote destination VA; READ: remote source VA.
-  uint64_t dstVa;
-  uint32_t length;
-  uint32_t lkey;
-  uint32_t rkey;
-  uint8_t opcode;
-  std::string peerNicPath;
-  FlagcxTransferTask *task;
-  volatile int *qpDepth;
-
-  inline void markSuccess() {
-    if (task)
-      task->doneSliceCount.fetch_add(1, std::memory_order_release);
-  }
-
-  inline void markFailed() {
-    if (task)
-      task->doneSliceCount.fetch_add(1, std::memory_order_release);
-  }
-};
-
 struct FlagcxNixlSlicePolicy {
   static constexpr bool kFurtherCut = false;
   static constexpr size_t kBlockSize = SIZE_MAX;

flagcx/include/flagcx_p2p.h

@@ -13,9 +13,11 @@
 #ifndef FLAGCX_P2P_H_
 #define FLAGCX_P2P_H_
 
+#include <atomic>
 #include <cstring>
 #include <stddef.h>
 #include <stdint.h>
+#include <string>
 #include <vector>
 
 /* ------------------------------------------------------------------ */
@@ -89,6 +91,51 @@ inline void flagcxP2pDeserializeRdmaDesc(const char *buf,
   std::memcpy(desc->padding, buf + 32, sizeof(desc->padding));
 }
 
+/* ------------------------------------------------------------------ */
+/*  Slice / transfer-task types (shared by engine and adaptor)        */
+/* ------------------------------------------------------------------ */
+
+struct FlagcxTransferTask {
+  std::atomic<uint64_t> sliceCount{0};
+  std::atomic<uint64_t> doneSliceCount{0};
+  std::vector<struct FlagcxSlice *> sliceList;
+
+  bool isAllDone() const {
+    auto total = sliceCount.load(std::memory_order_acquire);
+    auto done = doneSliceCount.load(std::memory_order_acquire);
+    return total > 0 && done >= total;
+  }
+};
+
+enum FlagcxSliceOp : uint8_t {
+  FLAGCX_SLICE_OP_WRITE = 0,
+  FLAGCX_SLICE_OP_READ = 1,
+};
+
+struct FlagcxSlice {
+  // WRITE: local source VA; READ: local destination VA.
+  uint64_t srcVa = 0;
+  // WRITE: remote destination VA; READ: remote source VA.
+  uint64_t dstVa;
+  uint32_t length;
+  uint32_t lkey;
+  uint32_t rkey;
+  uint8_t opcode;
+  std::string peerNicPath;
+  FlagcxTransferTask *task;
+  volatile int *qpDepth;
+
+  inline void markSuccess() {
+    if (task)
+      task->doneSliceCount.fetch_add(1, std::memory_order_release);
+  }
+
+  inline void markFailed() {
+    if (task)
+      task->doneSliceCount.fetch_add(1, std::memory_order_release);
+  }
+};
+
 /* ------------------------------------------------------------------ */
 /*  Notification message                                              */
 /* ------------------------------------------------------------------ */
@@ -472,11 +519,4 @@ const FlagcxP2pGlobalConfig &flagcxP2pGlobalConfig();
    flagcxP2pGlobalConfig() call. */
 void flagcxP2pDumpGlobalConfig();
 
-/* Clamp size-limited fields against ibv_query_device() results — call
-   once from the adaptor's init path after IB attributes are known. The
-   four uint32 inputs are the obvious ibv_device_attr counterparts; we
-   take plain ints to keep verbs out of this header. */
-void flagcxP2pClampToDeviceLimits(uint32_t maxQpWr, uint32_t maxSge,
-                                  uint32_t maxCqe, uint32_t maxQp);
-
 #endif /* FLAGCX_P2P_H_ */