On Thu, Jul 22, 2021 at 10:00:40AM -0300, Jason Gunthorpe wrote:

this is better:

   struct sg_append_table state;

   sg_append_init(&state, sgt, gfp_mask);

   while (..)
     ret = sg_append_pages(&state, pages, n_pages, ..)
     if (ret)
	 sg_append_abort(&state); // Frees the sgt and puts it to NULL
   sg_append_complete(&state)

Which allows sg_alloc_table_from_pages() to be written as

   struct sg_append_table state;
   sg_append_init(&state, sgt, gfp_mask);
   ret = sg_append_pages(&state,pages, n_pages, offset, size, UINT_MAX)
   if (ret) {
      sg_append_abort(&state);
      return ret;
   }
   sg_append_complete(&state);
   return 0;

And then the API can manage all of this in some sane and
understandable way.

That would be a lot easier to use for sure.  Not sure how invasive the
changes would be, though.