prov/cxi: ignore FLT_OVERFLOW and FLT_INVALID errors

Don't treat CXIP_COLL_RC_FLT_OVERFLOW and CXIP_COLL_RC_FLT_INVALID as errors. Add logging for when one of these two or CXIP_COLL_RC_FLT_INEXACT is encountered. Signed-off-by: John Heemstra <[email protected]>
pmodels · Nov 5, 2024 · a03c5a7 · a03c5a7
1 parent 5c701e5
commit a03c5a7
Showing 1 changed file with 15 additions and 6 deletions.
diff --git a/prov/cxi/src/cxip_coll.c b/prov/cxi/src/cxip_coll.c
@@ -1820,16 +1820,25 @@ static void _post_coll_complete(struct cxip_coll_reduction *reduction)
 
 	/* convert Rosetta return codes to CXIP return codes */
 	if (reduction->accum.red_rc == CXIP_COLL_RC_SUCCESS ||
-	    reduction->accum.red_rc == CXIP_COLL_RC_FLT_INEXACT) {
-		ret = cxip_cq_req_complete(req);
-	} else {
+	    reduction->accum.red_rc == CXIP_COLL_RC_FLT_INEXACT ||
+	    reduction->accum.red_rc == CXIP_COLL_RC_FLT_INVALID ||
+	    reduction->accum.red_rc == CXIP_COLL_RC_FLT_OVERFLOW) {
 		switch (reduction->accum.red_rc) {
-		case CXIP_COLL_RC_FLT_OVERFLOW:
-			prov = FI_CXI_ERRNO_RED_FLT_OVERFLOW;
+		case CXIP_COLL_RC_FLT_INEXACT:
+			CXIP_WARN("coll reduce FLT result was rounded\n");
 			break;
 		case CXIP_COLL_RC_FLT_INVALID:
-			prov = FI_CXI_ERRNO_RED_FLT_INVALID;
+			CXIP_WARN("coll reduce FLT invalid\n");
 			break;
+		case CXIP_COLL_RC_FLT_OVERFLOW:
+			CXIP_WARN("coll reduce FLT overflow\n");
+			break;
+		default:
+			break;
+		}
+		ret = cxip_cq_req_complete(req);
+	} else {
+		switch (reduction->accum.red_rc) {
 		case CXIP_COLL_RC_INT_OVERFLOW:
 			prov = FI_CXI_ERRNO_RED_INT_OVERFLOW;
 			break;