Have btls check for whether convertor involves device memory

ob1 might hand device memory to the btls, which then have to check
the convertor and make sure the data is packed if direct access
to device memory is not supported.

Adapted the following btls:
portals4, self, sm, tcp, ugni, usnic

Signed-off-by: Joseph Schuchart <schuchart@icl.utk.edu>

Author: devreal

ompi/mca/mtl/base/mtl_base_datatype.h

@@ -39,19 +39,19 @@ ompi_mtl_datatype_pack(struct opal_convertor_t *convertor,
 {
     struct iovec iov;
     uint32_t iov_count = 1;
+    bool is_accelerator = opal_convertor_on_device(convertor);
 #if !(OPAL_ENABLE_HETEROGENEOUS_SUPPORT)
     if (convertor->pDesc &&
 	!(convertor->flags & CONVERTOR_COMPLETED) &&
 	opal_datatype_is_contiguous_memory_layout(convertor->pDesc,
 						  convertor->count) &&
-        !(convertor->flags & CONVERTOR_ACCELERATOR)) {
+        !is_accelerator) {
 	    *free_after = false;
 	    *buffer = convertor->pBaseBuf + convertor->bConverted + convertor->pDesc->true_lb;
 	    *buffer_len = convertor->local_size;
 	    return OPAL_SUCCESS;
     }
 #endif
-    bool is_accelerator = convertor->flags & CONVERTOR_ACCELERATOR;
 
     opal_convertor_get_packed_size(convertor, buffer_len);
     *free_after = false;
@@ -94,7 +94,7 @@ ompi_mtl_datatype_recv_buf(struct opal_convertor_t *convertor,
         *buffer_len = 0;
         return OMPI_SUCCESS;
     }
-    bool is_accelerator = convertor->flags & CONVERTOR_ACCELERATOR;
+    bool is_accelerator = opal_convertor_on_device(convertor);;
 
     /* If we need buffers or we don't have accelerator support and it is a device buffer, we will need to copy */
     if (opal_convertor_need_buffers(convertor) || (is_accelerator && false == ompi_mtl_base_selected_component->accelerator_support)) {
@@ -122,7 +122,7 @@ ompi_mtl_datatype_unpack(struct opal_convertor_t *convertor,
 {
     struct iovec iov;
     uint32_t iov_count = 1;
-    bool is_accelerator = convertor->flags & CONVERTOR_ACCELERATOR;
+    bool is_accelerator = opal_convertor_on_device(convertor);
 
     /* If the buffer length is greater than 0 and we allocated buffers previously, we need to unpack them */
     if (buffer_len > 0 && (opal_convertor_need_buffers(convertor) || (is_accelerator && false == ompi_mtl_base_selected_component->accelerator_support))) {

opal/datatype/opal_convertor.h

@@ -187,6 +187,11 @@ static inline int32_t opal_convertor_need_buffers(const opal_convertor_t *pConve
     return 1;
 }
 
+static inline int32_t opal_convertor_on_device(const opal_convertor_t *pConvertor)
+{
+    return !!(pConvertor->flags & CONVERTOR_ACCELERATOR);
+}
+
 /**
  * Update the size of the remote datatype representation. The size will
  * depend on the configuration of the master convertor. In homogeneous

opal/mca/btl/portals4/btl_portals4.c

@@ -533,7 +533,9 @@ mca_btl_base_descriptor_t *mca_btl_portals4_prepare_src(struct mca_btl_base_modu
                          ((struct mca_btl_portals4_module_t *) btl_base)->interface_num,
                          reserve, *size, max_data));
 
-    if (0 != reserve || 0 != opal_convertor_need_buffers(convertor)) {
+    if (0 != reserve ||
+        0 != opal_convertor_need_buffers(convertor) ||
+        0 != opal_convertor_on_device(convertor)) {
         OPAL_OUTPUT_VERBOSE((90, opal_btl_base_framework.framework_output,
                              "mca_btl_portals4_prepare_src NEED BUFFERS or RESERVE\n"));
         frag = (mca_btl_portals4_frag_t *) mca_btl_portals4_alloc(btl_base, peer, MCA_BTL_NO_ORDER,

opal/mca/btl/self/btl_self.c

@@ -151,7 +151,7 @@ static struct mca_btl_base_descriptor_t *mca_btl_self_prepare_src(
     struct mca_btl_base_module_t *btl, struct mca_btl_base_endpoint_t *endpoint,
     struct opal_convertor_t *convertor, uint8_t order, size_t reserve, size_t *size, uint32_t flags)
 {
-    bool inline_send = !opal_convertor_need_buffers(convertor);
+    bool inline_send = !(opal_convertor_need_buffers(convertor) || opal_convertor_on_device(convertor));
     size_t buffer_len = reserve + (inline_send ? 0 : *size);
     mca_btl_self_frag_t *frag;
 
@@ -229,7 +229,9 @@ static int mca_btl_self_sendi(struct mca_btl_base_module_t *btl,
 {
     mca_btl_base_descriptor_t *frag;
 
-    if (!payload_size || !opal_convertor_need_buffers(convertor)) {
+    if (!payload_size ||
+        !(opal_convertor_need_buffers(convertor) ||
+          opal_convertor_on_device(convertor))) {
         void *data_ptr = NULL;
         if (payload_size) {
             opal_convertor_get_current_pointer(convertor, &data_ptr);

opal/mca/btl/sm/btl_sm_module.c

@@ -429,7 +429,8 @@ static struct mca_btl_base_descriptor_t *sm_prepare_src(struct mca_btl_base_modu
     assert(NULL != data_ptr);
 
     /* in place send fragment */
-    if (OPAL_UNLIKELY(opal_convertor_need_buffers(convertor))) {
+    if (OPAL_UNLIKELY(opal_convertor_need_buffers(convertor) ||
+                      opal_convertor_on_device(convertor))) {
         uint32_t iov_count = 1;
         struct iovec iov;
 

opal/mca/btl/sm/btl_sm_sendi.c

@@ -59,7 +59,7 @@ int mca_btl_sm_sendi(struct mca_btl_base_module_t *btl, struct mca_btl_base_endp
         opal_convertor_get_current_pointer(convertor, &data_ptr);
     }
 
-    if (!(payload_size && opal_convertor_need_buffers(convertor))
+    if (!(payload_size && (opal_convertor_need_buffers(convertor) || opal_convertor_on_device(convertor)))
         && mca_btl_sm_fbox_sendi(endpoint, tag, header, header_size, data_ptr, payload_size)) {
         return OPAL_SUCCESS;
     }

opal/mca/btl/tcp/btl_tcp.c

@@ -250,7 +250,7 @@ mca_btl_base_descriptor_t *mca_btl_tcp_prepare_src(struct mca_btl_base_module_t
     frag->segments[0].seg_len = reserve;
 
     frag->base.des_segment_count = 1;
-    if (opal_convertor_need_buffers(convertor)) {
+    if (opal_convertor_need_buffers(convertor) || opal_convertor_on_device(convertor)) {
 
         if (max_data + reserve > frag->size) {
             max_data = frag->size - reserve;

opal/mca/btl/ugni/btl_ugni_prepare.h

@@ -179,6 +179,7 @@ mca_btl_ugni_prepare_src_send(struct mca_btl_base_module_t *btl, mca_btl_base_en
 
     send_in_place = (btl->btl_flags & MCA_BTL_FLAGS_SEND_INPLACE)
                     && !(opal_convertor_need_buffers(convertor)
+                         || opal_convertor_on_device(convertor)
                          || (use_eager_get && ((uintptr_t) data_ptr & 3)));
 
     if (send_in_place) {

opal/mca/btl/usnic/btl_usnic_compat.c

@@ -91,7 +91,7 @@ prepare_src_small(struct opal_btl_usnic_module_t *module, struct mca_btl_base_en
      * we might still use INLINE for the send, and in that case we do not want
      * to copy the data at all.
      */
-    if (OPAL_UNLIKELY(opal_convertor_need_buffers(convertor))) {
+    if (OPAL_UNLIKELY(opal_convertor_need_buffers(convertor) || opal_convertor_on_device(convertor))) {
         /* put user data just after end of 1st seg (upper layer header) */
         assert(payload_len <= module->max_frag_payload);
         usnic_convertor_pack_simple(convertor,
@@ -227,7 +227,7 @@ static opal_btl_usnic_send_frag_t *prepare_src_large(struct opal_btl_usnic_modul
     /* make sure upper header small enough */
     assert(reserve <= sizeof(lfrag->lsf_ompi_header));
 
-    if (OPAL_UNLIKELY(opal_convertor_need_buffers(convertor))) {
+    if (OPAL_UNLIKELY(opal_convertor_need_buffers(convertor) || opal_convertor_on_device(convertor))) {
         /* threshold == -1 means always pack eagerly */
         if (mca_btl_usnic_component.pack_lazy_threshold >= 0
             && *size >= (size_t) mca_btl_usnic_component.pack_lazy_threshold) {