Merge pull request #6667 from hoopoepg/topic/alloc-with-hint-realloc-inplace

ALLOC_WITH_HINT: added inplace realloc

oshmem/mca/sshmem/ucx/sshmem_ucx.h

@@ -49,10 +49,19 @@ sshmem_ucx_shadow_allocator_t *sshmem_ucx_shadow_create(unsigned count);
 void sshmem_ucx_shadow_destroy(sshmem_ucx_shadow_allocator_t *allocator);
 int sshmem_ucx_shadow_alloc(sshmem_ucx_shadow_allocator_t *allocator,
                             unsigned count, unsigned *index);
+
+/* Reallocate existing allocated buffer. If possible - used inplace
+ * reallocation.
+ * Parameter 'inplace' - out, in case if zero - new buffer was allocated
+ * (inplace is not possible), user should remove original buffer after data
+ * is copied, else (if inplace == 0) - no additional action required */
+int sshmem_ucx_shadow_realloc(sshmem_ucx_shadow_allocator_t *allocator,
+                              unsigned count, unsigned old_index, unsigned *index,
+                              int *inplace);
 int sshmem_ucx_shadow_free(sshmem_ucx_shadow_allocator_t *allocator,
                            unsigned index);
-size_t sshmem_ucx_shadow_size(sshmem_ucx_shadow_allocator_t *allocator,
+unsigned sshmem_ucx_shadow_size(sshmem_ucx_shadow_allocator_t *allocator,
-                              unsigned index);
+                                unsigned index);
 
 END_C_DECLS
 

oshmem/mca/sshmem/ucx/sshmem_ucx_module.c

@@ -190,7 +190,6 @@ static uct_ib_device_mem_h alloc_device_mem(mca_spml_ucx_t *spml, size_t size,
     uct_md_h uct_md;
     void *address;
     size_t length;
-    int ret;
 
     uct_md = ucp_context_find_tl_md(spml->ucp_context, "mlx5");
     if (uct_md == NULL) {
@@ -353,8 +352,9 @@ static int sshmem_ucx_memheap_realloc(map_segment_t *s, size_t size,
                                       void* old_ptr, void** new_ptr)
 {
     mca_sshmem_ucx_segment_context_t *ctx = s->context;
-    unsigned alloc_count, index;
+    unsigned alloc_count, index, old_index, old_alloc_count;
     int res;
+    int inplace;
 
     if (size > s->seg_size) {
         return OSHMEM_ERR_OUT_OF_RESOURCE;
@@ -371,7 +371,15 @@ static int sshmem_ucx_memheap_realloc(map_segment_t *s, size_t size,
     /* Allocate new element. Zero-size allocation should still return a unique
      * pointer, so allocate 1 byte */
     alloc_count = max((size + ALLOC_ELEM_SIZE - 1) / ALLOC_ELEM_SIZE, 1);
-    res = sshmem_ucx_shadow_alloc(ctx->shadow_allocator, alloc_count, &index);
+
+    if (!old_ptr) {
+        res = sshmem_ucx_shadow_alloc(ctx->shadow_allocator, alloc_count, &index);
+    } else {
+        old_index = sshmem_ucx_memheap_ptr2index(s, old_ptr);
+        res       = sshmem_ucx_shadow_realloc(ctx->shadow_allocator, alloc_count,
+                                              old_index, &index, &inplace);
+    }
+
     if (res != OSHMEM_SUCCESS) {
         return res;
     }
@@ -379,10 +387,8 @@ static int sshmem_ucx_memheap_realloc(map_segment_t *s, size_t size,
     *new_ptr = sshmem_ucx_memheap_index2ptr(s, index);
 
     /* Copy to new segment and release old*/
-    if (old_ptr) {
+    if (old_ptr && !inplace) {
-        unsigned old_index       = sshmem_ucx_memheap_ptr2index(s, old_ptr);
+        old_alloc_count = sshmem_ucx_shadow_size(ctx->shadow_allocator, old_index);
-        unsigned old_alloc_count = sshmem_ucx_shadow_size(ctx->shadow_allocator,
-                                                          old_index);
         sshmem_ucx_memheap_wordcopy(*new_ptr, old_ptr,
                                     min(size, old_alloc_count * ALLOC_ELEM_SIZE));
         sshmem_ucx_shadow_free(ctx->shadow_allocator, old_index);

oshmem/mca/sshmem/ucx/sshmem_ucx_shadow.c

@@ -107,6 +107,68 @@ static void sshmem_ucx_shadow_merge_blocks(sshmem_ucx_shadow_allocator_t *alloca
     }
 }
 
+
+
+int sshmem_ucx_shadow_realloc(sshmem_ucx_shadow_allocator_t *allocator,
+                              unsigned count, unsigned old_index, unsigned *index,
+                              int *inplace)
+{
+    sshmem_ucx_shadow_alloc_elem_t *elem = &allocator->elems[old_index];
+    unsigned old_count                   = elem->block_size;
+    sshmem_ucx_shadow_alloc_elem_t *end;
+    sshmem_ucx_shadow_alloc_elem_t *next;
+
+    assert(count > 0);
+    assert(!sshmem_ucx_shadow_is_free(elem));
+
+    *inplace = 1;
+
+    if (count == old_count) {
+        *index = old_index;
+        return OSHMEM_SUCCESS;
+    }
+
+    if (count < old_count) {
+        /* requested block is shorter than allocated block
+         * then just cut current buffer */
+        sshmem_ucx_shadow_set_elem(elem + count,
+                                   SSHMEM_UCX_SHADOW_ELEM_FLAG_FREE,
+                                   elem->block_size - count);
+        elem->block_size = count;
+        *index           = old_index;
+        sshmem_ucx_shadow_merge_blocks(allocator);
+        return OSHMEM_SUCCESS;
+    }
+
+    assert(count > old_count);
+
+    end  = &allocator->elems[allocator->num_elems];
+    next = &elem[old_count];
+    /* try to check if next element is free & has enough length */
+    if ((next < end) &&                    /* non-last element? */
+        sshmem_ucx_shadow_is_free(next) && /* next is free */
+        (old_count + next->block_size >= count))
+    {
+        assert(elem < next);
+        assert(elem + count > next);
+        assert(elem + count <= end);
+        assert(next + next->block_size <= end);
+
+        if (old_count + next->block_size > count) {
+            sshmem_ucx_shadow_set_elem(elem + count, SSHMEM_UCX_SHADOW_ELEM_FLAG_FREE,
+                                       old_count + next->block_size - count);
+        }
+
+        sshmem_ucx_shadow_set_elem(next, 0, 0);
+        elem->block_size = count;
+        *index           = old_index;
+        return OSHMEM_SUCCESS;
+    }
+
+    *inplace = 0;
+    return sshmem_ucx_shadow_alloc(allocator, count, index);
+}
+
 int sshmem_ucx_shadow_free(sshmem_ucx_shadow_allocator_t *allocator,
                            unsigned index)
 {
@@ -117,8 +179,8 @@ int sshmem_ucx_shadow_free(sshmem_ucx_shadow_allocator_t *allocator,
     return OSHMEM_SUCCESS;
 }
 
-size_t sshmem_ucx_shadow_size(sshmem_ucx_shadow_allocator_t *allocator,
+unsigned sshmem_ucx_shadow_size(sshmem_ucx_shadow_allocator_t *allocator,
-                              unsigned index)
+                                unsigned index)
 {
     sshmem_ucx_shadow_alloc_elem_t *elem = &allocator->elems[index];