Merge pull request open-mpi#12823 from bosilca/topic/fix_ob1_segmentation

Topic/fix ob1 segmentation with UCT BTL

Author: bosilca

ompi/mca/pml/ob1/pml_ob1_isend.c

@@ -143,7 +143,7 @@ static inline int mca_pml_ob1_send_inline (const void *buf, size_t count,
     }
 
     if (OPAL_UNLIKELY(OMPI_SUCCESS != rc)) {
-	return rc;
+        return rc;
     }
 
     return (int) size;

opal/datatype/opal_datatype_internal.h

@@ -539,7 +539,7 @@ struct opal_datatype_t;
 #    define OPAL_DATATYPE_SAFEGUARD_POINTER(ACTPTR, LENGTH, INITPTR, PDATA, COUNT)                \
         {                                                                                         \
             unsigned char *__lower_bound = (INITPTR), *__upper_bound;                             \
-            assert(((LENGTH) != 0) && ((COUNT) != 0));                                            \
+            assert( (COUNT) != 0 );                                                               \
             __lower_bound += (PDATA)->true_lb;                                                    \
             __upper_bound = (INITPTR) + (PDATA)->true_ub +                                        \
                             ((PDATA)->ub - (PDATA)->lb) * ((COUNT) -1);                           \

opal/datatype/opal_datatype_position.c

@@ -66,8 +66,8 @@ static inline void position_single_block(opal_convertor_t *CONVERTOR, unsigned c
 }
 
 /**
- * Advance the convertors' position according. Update the pointer and the remaining space
- * accordingly.
+ * Advance the convertors' position according to account for *COUNT elements. Update
+ * the pointer and the remaining space accordingly.
  */
 static inline void position_predefined_data(opal_convertor_t *CONVERTOR, dt_elem_desc_t *ELEM,
                                             size_t *COUNT, unsigned char **POINTER, size_t *SPACE)
@@ -82,7 +82,8 @@ static inline void position_predefined_data(opal_convertor_t *CONVERTOR, dt_elem
 
     if (cando_count > *(COUNT)) {
         cando_count = *(COUNT);
-    }
+    } else if( 0 == cando_count )
+        return;
 
     if (1 == _elem->blocklen) {
         DO_DEBUG(opal_output(0,

opal/mca/btl/sm/btl_sm_send.c

@@ -73,18 +73,4 @@ int mca_btl_sm_send(struct mca_btl_base_module_t *btl, struct mca_btl_base_endpo
     }
 
     return OPAL_SUCCESS;
-
-#if 0
-    if (((frag->hdr->flags & MCA_BTL_SM_FLAG_SINGLE_COPY) ||
-        !(frag->base.des_flags & MCA_BTL_DES_FLAGS_BTL_OWNERSHIP)) &&
-        frag->base.des_cbfunc) {
-        frag->base.des_flags |= MCA_BTL_DES_SEND_ALWAYS_CALLBACK;
-
-        return OPAL_SUCCESS;
-    }
-
-    /* data is gone (from the pml's perspective). frag callback/release will
-       happen later */
-    return 1;
-#endif
 }

opal/mca/btl/uct/btl_uct_am.c

@@ -51,7 +51,7 @@ mca_btl_base_descriptor_t *mca_btl_uct_alloc(mca_btl_base_module_t *btl,
 }
 
 static inline void _mca_btl_uct_send_pack(void *data, void *header, size_t header_size,
-                                          opal_convertor_t *convertor, size_t payload_size)
+                                          opal_convertor_t *convertor, size_t* payload_size)
 {
     uint32_t iov_count = 1;
     struct iovec iov;
@@ -64,11 +64,9 @@ static inline void _mca_btl_uct_send_pack(void *data, void *header, size_t heade
 
     /* pack the data into the supplied buffer */
     iov.iov_base = (IOVBASE_TYPE *) ((intptr_t) data + header_size);
-    iov.iov_len = length = payload_size;
+    iov.iov_len = *payload_size;
 
-    (void) opal_convertor_pack(convertor, &iov, &iov_count, &length);
-
-    assert(length == payload_size);
+    (void) opal_convertor_pack(convertor, &iov, &iov_count, payload_size);
 }
 
 struct mca_btl_base_descriptor_t *mca_btl_uct_prepare_src(mca_btl_base_module_t *btl,
@@ -92,7 +90,10 @@ struct mca_btl_base_descriptor_t *mca_btl_uct_prepare_src(mca_btl_base_module_t
         }
 
         _mca_btl_uct_send_pack((void *) ((intptr_t) frag->uct_iov.buffer + reserve), NULL, 0,
-                               convertor, *size);
+                               convertor, size);
+        /* update the length of the fragment according to the convertor packed data */
+        frag->segments[0].seg_len = reserve + *size;
+        frag->uct_iov.length = frag->segments[0].seg_len;
     } else {
         opal_convertor_get_current_pointer(convertor, &data_ptr);
         assert(NULL != data_ptr);
@@ -286,7 +287,7 @@ static size_t mca_btl_uct_sendi_pack(void *data, void *arg)
 
     am_header->value = args->am_header;
     _mca_btl_uct_send_pack((void *) ((intptr_t) data + 8), args->header, args->header_size,
-                           args->convertor, args->payload_size);
+                           args->convertor, &args->payload_size);
     return args->header_size + args->payload_size + 8;
 }
 
@@ -329,9 +330,18 @@ int mca_btl_uct_sendi(mca_btl_base_module_t *btl, mca_btl_base_endpoint_t *endpo
     } else if (msg_size < (size_t) MCA_BTL_UCT_TL_ATTR(uct_btl->am_tl, context->context_id)
                               .cap.am.max_short) {
         int8_t *data = alloca(total_size);
-        _mca_btl_uct_send_pack(data, header, header_size, convertor, payload_size);
-        ucs_status = uct_ep_am_short(ep_handle, MCA_BTL_UCT_FRAG, am_header.value, data,
-                                     total_size);
+        size_t packed_payload_size = payload_size;
+        _mca_btl_uct_send_pack(data, header, header_size, convertor, &packed_payload_size);
+        if (packed_payload_size != payload_size) {
+            /* This should never happen as the packed data should go in a single pack. But
+               in case it does, fallback onto a descriptor allocation and let the caller
+               send the data.
+             */
+            ucs_status = UCS_ERR_NO_RESOURCE;
+        } else {
+            ucs_status = uct_ep_am_short(ep_handle, MCA_BTL_UCT_FRAG, am_header.value, data,
+                                         total_size);
+        }
     } else {
         ssize_t size;