DAOS-16971 utils: two patches from Jerome to fix progress hang

Two patches from Jerome to fix progress hang Signed-off-by: Di Wang <[email protected]>
daos-stack · Jan 23, 2025 · 64ae3f6 · 64ae3f6
1 parent 8ddde42
commit 64ae3f6
Show file tree

Hide file tree

Showing 4 changed files with 89 additions and 1 deletion.
diff --git a/site_scons/prereq_tools/base.py b/site_scons/prereq_tools/base.py
@@ -1091,6 +1091,7 @@ def _resolve_patches(self):
             if "^" in raw:
                 (patch_subdir, raw) = raw.split("^")
             if "https://" not in raw:
+                raw = os.path.join(Dir('#').abspath, raw)
                 patches[raw] = patch_subdir
                 continue
             patch_name = f"{self.name}_{self._sanitize_patch_path(raw)}_{patchnum:d}"

diff --git a/utils/build.config b/utils/build.config
@@ -28,5 +28,6 @@ ucx=https://github.com/openucx/ucx.git
 [patch_versions]
 spdk=https://github.com/spdk/spdk/commit/b0aba3fcd5aceceea530a702922153bc75664978.diff,https://github.com/spdk/spdk/commit/445a4c808badbad3942696ecf16fa60e8129a747.diff
 fuse=https://github.com/libfuse/libfuse/commit/c9905341ea34ff9acbc11b3c53ba8bcea35eeed8.diff
-mercury=https://raw.githubusercontent.com/daos-stack/mercury/f3dc286fb40ec1a3a38a2e17c45497bc2aa6290d/na_ucx.patch
+mercury=https://raw.githubusercontent.com/daos-stack/mercury/f3dc286fb40ec1a3a38a2e17c45497bc2aa6290d/na_ucx.patch,utils/mrecv_err.patch
 pmdk=https://github.com/pmem/pmdk/commit/2abe15ac0b4eed894b6768cd82a3b0a7c4336284.diff
+ofi=utils/ofi_tcp.patch
diff --git a/utils/mrecv_err.patch b/utils/mrecv_err.patch
@@ -0,0 +1,60 @@
+diff --git a/src/mercury_core.c b/src/mercury_core.c
+index 3c177630..c6150852 100644
+--- a/src/mercury_core.c
++++ b/src/mercury_core.c
+@@ -4727,11 +4727,13 @@ hg_core_multi_recv_input_cb(const struct na_cb_info *callback_info)
+     } else if (callback_info->ret == NA_CANCELED) {
+         HG_LOG_SUBSYS_DEBUG(
+             rpc, "NA_CANCELED event on multi-recv op %d", multi_recv_op->id);
+-        hg_atomic_decr32(&context->multi_recv_op_count);
++        if (na_cb_info_multi_recv_unexpected->last)
++            hg_atomic_decr32(&context->multi_recv_op_count);
+     } else {
+         HG_LOG_SUBSYS_ERROR(rpc, "NA callback returned error (%s)",
+             NA_Error_to_string(callback_info->ret));
+-        hg_atomic_decr32(&context->multi_recv_op_count);
++        if (na_cb_info_multi_recv_unexpected->last)
++            hg_atomic_decr32(&context->multi_recv_op_count);
+         /* TODO can an unexpected multi-recv operation ever fail? */
+     }
+
+diff --git a/src/na/na_ofi.c b/src/na/na_ofi.c
+index 92ddacc4..ea1187be 100644
+--- a/src/na/na_ofi.c
++++ b/src/na/na_ofi.c
+@@ -6558,9 +6558,21 @@ na_ofi_cq_readerr(struct fid_cq *cq, struct fi_cq_tagged_entry *cq_event,
+                 !(hg_atomic_get32(&na_ofi_op_id->status) & NA_OFI_OP_CANCELED),
+                 "Operation ID was not canceled by user");
+             */
++            // if (na_ofi_op_id->type == NA_CB_MULTI_RECV_UNEXPECTED) {
++            //     if (cq_err.flags & FI_MULTI_RECV)
++            //         NA_LOG_SUBSYS_WARNING(
++            //             op, "FI_ECANCELED reported on multi-recv (completed)");
++            //     else
++            //         NA_LOG_SUBSYS_WARNING(op,
++            //             "FI_ECANCELED reported on multi-recv (not completed)");
++            // }
+
+             /* Complete operation in canceled state */
+-            na_ofi_op_id->complete(na_ofi_op_id, true, NA_CANCELED);
++            na_ofi_op_id->complete(na_ofi_op_id,
++                (na_ofi_op_id->type == NA_CB_MULTI_RECV_UNEXPECTED)
++                    ? cq_err.flags & FI_MULTI_RECV
++                    : true,
++                NA_CANCELED);
+         } break;
+
+         case FI_EADDRNOTAVAIL:
+@@ -6623,7 +6635,11 @@ na_ofi_cq_readerr(struct fid_cq *cq, struct fi_cq_tagged_entry *cq_event,
+                         na_ofi_op_id->addr->fi_addr, NA_HOSTUNREACH);
+
+                 /* Complete operation in error state */
+-                na_ofi_op_id->complete(na_ofi_op_id, true, na_ret);
++                na_ofi_op_id->complete(na_ofi_op_id,
++                    (na_ofi_op_id->type == NA_CB_MULTI_RECV_UNEXPECTED)
++                        ? cq_err.flags & FI_MULTI_RECV
++                        : true,
++                    na_ret);
+             }
+             break;
+     }
diff --git a/utils/ofi_tcp.patch b/utils/ofi_tcp.patch
@@ -0,0 +1,26 @@
+diff --git a/prov/tcp/src/xnet_cq.c b/prov/tcp/src/xnet_cq.c
+index 03ea97537..2090bdf71 100644
+--- a/prov/tcp/src/xnet_cq.c
++++ b/prov/tcp/src/xnet_cq.c
+@@ -202,13 +202,15 @@ void xnet_report_error(struct xnet_xfer_entry *xfer_entry, int err)
+
+ 	err_entry.flags = xfer_entry->cq_flags & ~FI_COMPLETION;
+ 	if (err_entry.flags & FI_RECV) {
+-		if (xfer_entry->ctrl_flags & XNET_MULTI_RECV &&
+-		    xfer_entry->mrecv) {
+-			xfer_entry->mrecv->ref_cnt--;
+-			if (!xfer_entry->mrecv->ref_cnt) {
++		if (xfer_entry->ctrl_flags & XNET_MULTI_RECV) {
++			if (xfer_entry->mrecv) {
++				xfer_entry->mrecv->ref_cnt--;
++				if (!xfer_entry->mrecv->ref_cnt) {
++					err_entry.flags |= FI_MULTI_RECV;
++					free(xfer_entry->mrecv);
++				}
++			} else
+ 				err_entry.flags |= FI_MULTI_RECV;
+-				free(xfer_entry->mrecv);
+-			}
+ 		}
+ 		xnet_get_cq_info(xfer_entry, &err_entry.flags, &err_entry.data,
+ 				 &err_entry.tag);