NA: add NA_IO return code for generic I/O errors

HG: add HG_IO return code and reserve space for additional NA codes Update OFI and UCX plugins to use new code
mercury-hpc · Oct 24, 2024 · 928034b · 928034b
1 parent 5ee197c
commit 928034b
Show file tree

Hide file tree

Showing 4 changed files with 63 additions and 37 deletions.
diff --git a/src/mercury_core_types.h b/src/mercury_core_types.h
@@ -119,39 +119,40 @@ struct hg_init_info {
     unsigned int multi_recv_copy_threshold;
 };
 
+/* Keep offset to keep room for additional NA error codes */
+#define HG_NA_ERRNO_OFFSET 64
+
 /* Error return codes:
  * Functions return 0 for success or corresponding return code */
-#define HG_RETURN_VALUES                                                       \
-    X(HG_SUCCESS)        /*!< operation succeeded */                           \
-    X(HG_PERMISSION)     /*!< operation not permitted */                       \
-    X(HG_NOENTRY)        /*!< no such file or directory */                     \
-    X(HG_INTERRUPT)      /*!< operation interrupted */                         \
-    X(HG_AGAIN)          /*!< operation must be retried */                     \
-    X(HG_NOMEM)          /*!< out of memory */                                 \
-    X(HG_ACCESS)         /*!< permission denied */                             \
-    X(HG_FAULT)          /*!< bad address */                                   \
-    X(HG_BUSY)           /*!< device or resource busy */                       \
-    X(HG_EXIST)          /*!< entry already exists */                          \
-    X(HG_NODEV)          /*!< no such device */                                \
-    X(HG_INVALID_ARG)    /*!< invalid argument */                              \
-    X(HG_PROTOCOL_ERROR) /*!< protocol error */                                \
-    X(HG_OVERFLOW)       /*!< value too large */                               \
-    X(HG_MSGSIZE)        /*!< message size too long */                         \
-    X(HG_PROTONOSUPPORT) /*!< protocol not supported */                        \
-    X(HG_OPNOTSUPPORTED) /*!< operation not supported on endpoint */           \
-    X(HG_ADDRINUSE)      /*!< address already in use */                        \
-    X(HG_ADDRNOTAVAIL)   /*!< cannot assign requested address */               \
-    X(HG_HOSTUNREACH)    /*!< cannot reach host during operation */            \
-    X(HG_TIMEOUT)        /*!< operation reached timeout */                     \
-    X(HG_CANCELED)       /*!< operation canceled */                            \
-    X(HG_CHECKSUM_ERROR) /*!< checksum error */                                \
-    X(HG_NA_ERROR)       /*!< generic NA error */                              \
-    X(HG_OTHER_ERROR)    /*!< generic HG error */                              \
-    X(HG_RETURN_MAX)
-
-#define X(a) a,
-typedef enum hg_return { HG_RETURN_VALUES } hg_return_t;
-#undef X
+typedef enum hg_return {
+    HG_SUCCESS,        /*!< operation succeeded */
+    HG_PERMISSION,     /*!< operation not permitted */
+    HG_NOENTRY,        /*!< no such file or directory */
+    HG_INTERRUPT,      /*!< operation interrupted */
+    HG_AGAIN,          /*!< operation must be retried */
+    HG_NOMEM,          /*!< out of memory */
+    HG_ACCESS,         /*!< permission denied */
+    HG_FAULT,          /*!< bad address */
+    HG_BUSY,           /*!< device or resource busy */
+    HG_EXIST,          /*!< entry already exists */
+    HG_NODEV,          /*!< no such device */
+    HG_INVALID_ARG,    /*!< invalid argument */
+    HG_PROTOCOL_ERROR, /*!< protocol error */
+    HG_OVERFLOW,       /*!< value too large */
+    HG_MSGSIZE,        /*!< message size too long */
+    HG_PROTONOSUPPORT, /*!< protocol not supported */
+    HG_OPNOTSUPPORTED, /*!< operation not supported on endpoint */
+    HG_ADDRINUSE,      /*!< address already in use */
+    HG_ADDRNOTAVAIL,   /*!< cannot assign requested address */
+    HG_HOSTUNREACH,    /*!< cannot reach host during operation */
+    HG_TIMEOUT,        /*!< operation reached timeout */
+    HG_CANCELED,       /*!< operation canceled */
+    HG_IO,             /*!< I/O error */
+    HG_CHECKSUM_ERROR = HG_NA_ERRNO_OFFSET, /*!< checksum error */
+    HG_NA_ERROR,                            /*!< generic NA error */
+    HG_OTHER_ERROR,                         /*!< generic HG error */
+    HG_RETURN_MAX
+} hg_return_t;
 
 /* Compat return codes */
 #define HG_INVALID_PARAM HG_INVALID_ARG

diff --git a/src/na/na_ofi.c b/src/na/na_ofi.c
@@ -2169,13 +2169,28 @@ na_ofi_errno_to_na(int rc)
         case FI_EINTR:
             ret = NA_INTERRUPT;
             break;
+        case FI_EIO:
+#if !defined(__APPLE__)
+        case FI_EREMOTEIO:
+#endif
+            ret = NA_IO;
+            break;
         case FI_EAGAIN:
+#ifdef _WIN32
+        case FI_EWOULDBLOCK:
+#endif
             ret = NA_AGAIN;
             break;
         case FI_ENOMEM:
+        case FI_EMFILE:
+        case FI_ENOSPC:
+        case FI_ENOBUFS:
             ret = NA_NOMEM;
             break;
         case FI_EACCES:
+#if !defined(_WIN32) && !defined(__APPLE__)
+        case FI_EKEYREJECTED:
+#endif
             ret = NA_ACCESS;
             break;
         case FI_EFAULT:
@@ -2187,6 +2202,8 @@ na_ofi_errno_to_na(int rc)
         case FI_ENODEV:
             ret = NA_NODEV;
             break;
+        case FI_E2BIG:
+        case FI_EBADF:
         case FI_EINVAL:
             ret = NA_INVALID_ARG;
             break;
@@ -2197,6 +2214,7 @@ na_ofi_errno_to_na(int rc)
             ret = NA_MSGSIZE;
             break;
         case FI_ENOPROTOOPT:
+        case FI_ENOSYS:
             ret = NA_PROTONOSUPPORT;
             break;
         case FI_EOPNOTSUPP:
@@ -2210,14 +2228,12 @@ na_ofi_errno_to_na(int rc)
             break;
         case FI_ENETDOWN:
         case FI_ENETUNREACH:
-        case FI_ENOTCONN:
         case FI_ECONNABORTED:
-        case FI_ECONNREFUSED:
         case FI_ECONNRESET:
-#ifndef _WIN32
+        case FI_ENOTCONN:
         case FI_ESHUTDOWN:
+        case FI_ECONNREFUSED:
         case FI_EHOSTDOWN:
-#endif
         case FI_EHOSTUNREACH:
             ret = NA_HOSTUNREACH;
             break;
@@ -2227,6 +2243,11 @@ na_ofi_errno_to_na(int rc)
         case FI_ECANCELED:
             ret = NA_CANCELED;
             break;
+        case FI_ENOMSG:
+        case FI_ENODATA:
+        case FI_EISCONN:
+        case FI_EALREADY:
+        case FI_EINPROGRESS:
         default:
             ret = NA_PROTOCOL_ERROR;
             break;

diff --git a/src/na/na_types.h b/src/na/na_types.h
@@ -151,6 +151,7 @@ struct na_protocol_info {
     X(NA_HOSTUNREACH)    /*!< cannot reach host during operation */            \
     X(NA_TIMEOUT)        /*!< operation reached timeout */                     \
     X(NA_CANCELED)       /*!< operation canceled */                            \
+    X(NA_IO)             /*!< I/O error */                                     \
     X(NA_RETURN_MAX)
 
 #define X(a) a,

diff --git a/src/na/na_ucx.c b/src/na/na_ucx.c
@@ -1045,7 +1045,6 @@ na_ucs_status_to_na(ucs_status_t status)
             ret = NA_ADDRNOTAVAIL;
             break;
 
-        case UCS_ERR_SOME_CONNECTS_FAILED:
         case UCS_ERR_UNREACHABLE:
         case UCS_ERR_CONNECTION_RESET:
         case UCS_ERR_NOT_CONNECTED:
@@ -1062,8 +1061,12 @@ na_ucs_status_to_na(ucs_status_t status)
             ret = NA_CANCELED;
             break;
 
-        case UCS_ERR_NO_MESSAGE:
+        case UCS_ERR_SOME_CONNECTS_FAILED:
         case UCS_ERR_IO_ERROR:
+            ret = NA_IO;
+            break;
+
+        case UCS_ERR_NO_MESSAGE:
         case UCS_ERR_SHMEM_SEGMENT:
         default:
             ret = NA_PROTOCOL_ERROR;