RCCL Multinode DMA Buffer crash fix (#1682)

This commit handles DMABUF initialization and call appropriate handling function. This fixes crash in OS with no peermem support and relying on only DMABUF.

* Initial test commit
* Handling Dmabuf_fd opening and closing
* Cleanup
* Use DMABuff or Peermem as needed
* Using user input for ibDmaBufSupportInitOnce
* Revert all changes to rocmwrap.cc
* Revert all changes to rocmwrap.cc
* Changing to func definition braces
* Reverting line removal in utils.h
* useDmaBuf to calculate  flushEnabled

src/transport/net_ib.cc

@@ -793,7 +793,8 @@ static void ibDmaBufSupportInitOnce(){
   ncclIbDev* ibDev = ncclIbDevs + mergedDev->vProps.devs[0];
   struct ibv_pd* pd;
   struct ibv_context* ctx = ibDev->context;
-  rocmLibraryInit();
+  res = rocmLibraryInit();
+  if (res != ncclSuccess) goto failure;
   NCCLCHECKGOTO(wrap_ibv_alloc_pd(&pd, ctx), res, failure);
   // Test kernel DMA-BUF support with a dummy call (fd=-1)
   (void)wrap_direct_ibv_reg_dmabuf_mr(pd, 0ULL /*offset*/, 0ULL /*len*/, 0ULL /*iova*/, -1 /*fd*/, 0 /*flags*/);
@@ -1071,6 +1072,7 @@ struct ncclIbGpuFlush {
   int* gpuFlushGpuMem;
   struct ibv_sge sge;
   struct ncclIbQp qp;
+  int dmabuf_fd;
 };
 
 struct ncclIbRemFifo {
@@ -1620,6 +1622,7 @@ ib_recv:
   struct ncclIbRecvCommDev* rCommDev;
   struct ncclIbDevInfo* remDevInfo;
   struct ncclIbQp* qp;
+  bool useDmaBuf; 
 
   mergedDev = ncclIbMergedDevs + lComm->dev;
   rComm->base.nRemDevs = remMeta.ndevs;
@@ -1685,9 +1688,9 @@ ib_recv:
     NCCLCHECKGOTO(ncclIbRtsQp(qp->qp), ret, fail);
   }
 
-  rComm->flushEnabled = ((ncclIbGdrSupport() == ncclSuccess || ncclIbDmaBufSupport(lComm->dev) == ncclSuccess)
-                            && (ncclParamIbGdrFlushDisable() == 0)) ? 1 : 0;
-
+  useDmaBuf  = (ncclIbDmaBufSupport(lComm->dev) == ncclSuccess);
+  rComm->flushEnabled = ((ncclIbGdrSupport() == ncclSuccess || useDmaBuf)
+                            && (ncclParamIbGdrFlushDisable() == 0)) ? 1 : 0;              
   for (int i = 0; i < rComm->base.vProps.ndevs; i++) {
     rCommDev = rComm->devs + i;
     ibDev = ncclIbDevs + rCommDev->base.ibDevN;
@@ -1706,10 +1709,29 @@ ib_recv:
 #else
         NCCLCHECKGOTO(ncclCudaCalloc(&rCommDev->gpuFlush.gpuFlushGpuMem, sizeof(int), nullptr, hipDeviceMallocFinegrained), ret, fail);
 #endif
-        NCCLCHECKGOTO(wrap_ibv_reg_mr(&rCommDev->gpuFlush.gpuMr, rCommDev->base.pd, rCommDev->gpuFlush.gpuFlushGpuMem, sizeof(int), IBV_ACCESS_LOCAL_WRITE|IBV_ACCESS_REMOTE_WRITE|IBV_ACCESS_REMOTE_READ), ret, fail);
+        if (useDmaBuf)
+        {
+          uint64_t export_offset = 0;
+          void *aligned_ptr = NULL;
+          size_t aligned_size = 0;
+          get_aligned_ptr_and_size(rCommDev->gpuFlush.gpuFlushGpuMem, sizeof(int) /*devicebuffersize*/, &aligned_ptr, &aligned_size);
+          hsa_status_t export_status = pfn_hsa_amd_portable_export_dmabuf(aligned_ptr, aligned_size, &rCommDev->gpuFlush.dmabuf_fd, &export_offset);
+          if (rCommDev->gpuFlush.dmabuf_fd < 0 || export_status != HSA_STATUS_SUCCESS)
+          {
+            WARN("Failed to export DMA BUF");
+            goto fail;
+          }
+          NCCLCHECKGOTO(wrap_ibv_reg_dmabuf_mr(&rCommDev->gpuFlush.gpuMr, rCommDev->base.pd, export_offset, sizeof(int), (uint64_t)rCommDev->gpuFlush.gpuFlushGpuMem /*iova*/, rCommDev->gpuFlush.dmabuf_fd, IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE | IBV_ACCESS_REMOTE_READ), ret, fail);
+        }
+        else
+        {
+          rCommDev->gpuFlush.dmabuf_fd = -1;
+          NCCLCHECKGOTO(wrap_ibv_reg_mr(&rCommDev->gpuFlush.gpuMr, rCommDev->base.pd, rCommDev->gpuFlush.gpuFlushGpuMem, sizeof(int), IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE | IBV_ACCESS_REMOTE_READ), ret, fail);
+        }
       } else {
         rCommDev->gpuFlush.gpuFlushGpuMem = nullptr;
         rCommDev->gpuFlush.gpuMr = nullptr;
+        rCommDev->gpuFlush.dmabuf_fd = -1;
       }
       NCCLCHECKGOTO(wrap_ibv_reg_mr(&rCommDev->gpuFlush.hostMr, rCommDev->base.pd, &rComm->gpuFlushHostMem, sizeof(int), IBV_ACCESS_LOCAL_WRITE), ret, fail);
       rCommDev->gpuFlush.sge.addr = (uint64_t)&rComm->gpuFlushHostMem;
@@ -2446,6 +2468,7 @@ ncclResult_t ncclIbCloseRecv(void* recvComm) {
           commDev->gpuFlush.gpuFlushGpuMem = nullptr;
           if (commDev->gpuFlush.gpuMr != nullptr) NCCLCHECK(wrap_ibv_dereg_mr(commDev->gpuFlush.gpuMr));
           commDev->gpuFlush.gpuMr = nullptr;
+          if(commDev->gpuFlush.dmabuf_fd > 0) { close(commDev->gpuFlush.dmabuf_fd);}
         }
         if (commDev->gpuFlush.qp.qp != NULL) NCCLCHECK(wrap_ibv_destroy_qp(commDev->gpuFlush.qp.qp));
         if (commDev->gpuFlush.hostMr != NULL) NCCLCHECK(wrap_ibv_dereg_mr(commDev->gpuFlush.hostMr));