modules/swcore: bugfix to allocator preventing from using twice the same page when memory gets full

The problem occured in swc_page_alloc_ram_bug.vhd when there were no more free pages (out_nomem_d1='1') and input block was requesting allocation and setting usecnt (alloc FSM in S_PCKSTART_SET_AND_REQ). The operation was _done_ because there is always one free page read in advance, prepared to be allocated to a requesting input block and set usecnt does not depend on out_nomem_d1. However, reading of the next free page from the memory was done only on alloc_req, only if there were still some free pages (q_read_p0 <= alloc_req & !out_nomem_d1). That means allocator was holding address of the page just allocated to one of the input blocks, and was giving the same address to the next one requesting allocation (if in the meantime out_nomem_d1 went to '0'). To fix that I've added one register keeping information if the page read from the memory was already given to any requestor. If it was then I do _read_ of the new page imediately after there are some new pages (out_nomem_d1 = '0') so that it is ready when next time somebody asks.

modules/swcore: bugfix to allocator preventing from using twice the same page when memory gets full
The problem occured in swc_page_alloc_ram_bug.vhd when there were no more free pages (out_nomem_d1='1') and input block was requesting allocation and setting usecnt (alloc FSM in S_PCKSTART_SET_AND_REQ). The operation was _done_ because there is always one free page read in advance, prepared to be allocated to a requesting input block and set usecnt does not depend on out_nomem_d1. However, reading of the next free page from the memory was done only on alloc_req, only if there were still some free pages (q_read_p0 <= alloc_req & !out_nomem_d1). That means allocator was holding address of the page just allocated to one of the input blocks, and was giving the same address to the next one requesting allocation (if in the meantime out_nomem_d1 went to '0'). To fix that I've added one register keeping information if the page read from the memory was already given to any requestor. If it was then I do _read_ of the new page imediately after there are some new pages (out_nomem_d1 = '0') so that it is ready when next time somebody asks.
c6783f6a · Grzegorz Daniluk · dd0dadd6 · c6783f6a
Commit c6783f6a authored Apr 23, 2015 by Grzegorz Daniluk
Hide whitespace changes
Inline Side-by-side

Showing with 15 additions and 1 deletion

swc_page_alloc_ram_bug.vhd ...swcore/optimized_new_allocator/swc_page_alloc_ram_bug.vhd +15 -1

No files found.
--- a/modules/wrsw_swcore/optimized_new_allocator/swc_page_alloc_ram_bug.vhd
+++ b/modules/wrsw_swcore/optimized_new_allocator/swc_page_alloc_ram_bug.vhd
@@ -273,6 +273,7 @@ architecture syn of swc_page_allocator_new is
  signal res_almost_full         : std_logic_vector(g_resource_num      -1 downto 0);
  -----------------------------
  
+  signal pg_adv_valid : std_logic;

  type t_alloc_req is record
    alloc                 : std_logic;
@@ -339,6 +340,19 @@ begin  -- syn
    end if;  
  
  end process;
+
+  -- keeping the information if free page read from memory was already allocated
+  -- to any of the ports
+  process(clk_i)
+  begin
+    if rising_edge(clk_i) then
+      if (rst_n_i = '0' or alloc_req_in.alloc = '1') then
+        pg_adv_valid <= '0';
+      elsif(out_nomem_d1 = '0') then
+        pg_adv_valid <= '1';
+      end if;
+    end if;
+  end process;
  
  -- write queue when freeing and not initializing (because when initializing we use other
  -- port of this memory)
@@ -356,7 +370,7 @@ begin  -- syn
  -- nomem_d0  : _______|-------
  -- alloc_d0  " _______|-|____  <= this need to be handled
  -- nomem_d1  : ________|-------
-  q_read_p0  <= '1' when (alloc_req_d0.alloc = '1') and (out_nomem_d1 = '0') else '0'; 
+  q_read_p0 <= '1' when (pg_adv_valid = '0' and out_nomem_d1 = '0') else '0';
  
  -- address of page stored in the memory (queue)
  q_input_addr_p1 <= std_logic_vector(wr_ptr_p1) when initializing = '1' else alloc_req_d1.pgaddr_free;