Skip to content

W
White Rabbit Switch - Gateware
Commit 9199cf32 authored by Grzegorz Daniluk's avatar Grzegorz Daniluk
Browse files
Options

modules/swcore: allow force_free only when all pages for the frame are stored to LL 

That's how it should be done. Otherwise we always have the race
condition. Especially on high traffic rate on many ports LL fsm has to
wait sometimes few clock cycles for the write request to be served. If
a force_free request was already generated, freeing of pages may always
go ahead of LL stored pages resulting in free fsm stuck in R_NEXT.
parent 4b4e4c85
Hide whitespace changes
Inline Side-by-side
Showing with 59 additions and 40 deletions
modules/wrsw_swcore/xswc_input_block.vhd
View file @ 9199cf32
......@@ -375,6 +375,8 @@ architecture syn of xswc_input_block is
signal current_pckstart_pageaddr : std_logic_vector(g_page_addr_width - 1 downto 0);
-- we remember what was it's usecant at the time
signal current_pckstart_usecnt : std_logic_vector(g_usecount_width - 1 downto 0);
-- last pckstart stored in the LL
signal current_pckstart_ll_stored : std_logic_vector(g_page_addr_width - 1 downto 0);
-- signals sent to Pck Transfer Unit
signal pta_transfer_pck : std_logic;
......@@ -672,9 +674,11 @@ constant c_force_usecnt : boolean := TRUE;
signal tp_transfer_wait_sof : std_logic;
signal tp_transfer_set_usecnt : std_logic;
signal ll_pckstart_stored : std_logic;
signal ll_stored_eof : std_logic;
signal rcv_pckstart_new : std_logic;
signal pcknew_reg : std_logic;
signal ffree_mask : std_logic;
signal ffree_pre_mask : std_logic;
signal dbg_dropped_on_res_full : std_logic;
......@@ -2116,6 +2120,11 @@ begin
if(ll_wr_req = '1' and ll_wr_done_i = '1') then -- written
ll_wr_req <= '0';
-- remember the last startpage stored to the LL
if(current_pckstart_pageaddr = ll_entry.addr and ll_entry.valid='1') then
current_pckstart_ll_stored <= current_pckstart_pageaddr;
end if;
if(mpm_dlast_reg = '1') then
mpm_dlast_reg <= '0';
ll_wr_req <= '1';
......@@ -2187,6 +2196,10 @@ begin
--===========================================================================================
-- if (EOF or error or drop decision) happens on writing
if(ll_wr_req = '1' and ll_wr_done_i = '1') then -- written
-- remember the last startpage stored to the LL
if(current_pckstart_pageaddr = ll_entry.addr and ll_entry.valid='1') then
current_pckstart_ll_stored <= current_pckstart_pageaddr;
end if;
ll_wr_req <= '1';
ll_entry <= ll_entry_tmp;
--- not used because we write size and stuff
......@@ -2205,6 +2218,10 @@ begin
--===========================================================================================
in_pck_sof_reg_ack <= '1';
if(ll_wr_req = '1' and ll_wr_done_i = '1') then -- written
-- remember the last startpage stored to the LL
if(current_pckstart_pageaddr = ll_entry.addr and ll_entry.valid='1') then
current_pckstart_ll_stored <= current_pckstart_pageaddr;
end if;
--ll_wr_req <= '0';
--ll_entry <= c_zero_ll_entry;
if(mpm_dlast_reg = '1') then
......@@ -2350,53 +2367,55 @@ rp_in_pck_error <= '1' when (rp_in_pck_err = '1' or in_pck_err = '1') else '0';
begin
if rising_edge(clk_i) then
if (rst_n_i = '0') then
ll_pckstart_stored <= '0';
ffree_mask <= '1';
elsif(ffree_mask = '1' and rcv_pckstart_new = '1' and mmu_force_free_req = '1') or
-- if there is an ongoing FF req, then we cannot clear the mask because
-- it won't be done, never.. However, we remember the
-- rcv_pckstart_new='1' by setting ll_pckstart_stored to '0'.
(ffree_mask = '1' and rcv_pckstart_new = '1' and s_ll_write = S_SOF_ON_WR) then
-- if we get rcv_pckstart_new while we're in SOF_ON_WR, this means we
-- have a new frame, but the previous one might still need to be
-- force-freed. That's why we postpone driving ffree_mask to '0'.
ffree_mask <= '1';
ll_pckstart_stored <= '0';
elsif(ffree_mask = '1' and rcv_pckstart_new = '1') then
-- that is the most "clean" situation, there is no FF request and we're
-- receiving new frame, therefore we clear ffree_mask and
-- ll_pckstart_stored.
ffree_mask <= '0';
ll_pckstart_stored <= '0';
elsif(ffree_mask = '1' and ll_pckstart_stored = '0' and mmu_force_free_req = '0' and s_ll_write /= S_SOF_ON_WR) then
-- in case the first condition took place, (and in the meantime the
-- start page was not saved to linked list) clear the ffree_mask when
-- FF req is no more active.
ll_stored_eof <= '0';
ffree_mask <= '0';
ffree_pre_mask <= '0';
pcknew_reg <= '0';
else
if(rcv_pckstart_new='1' and ll_wr_req='1' and ll_entry.valid='1') then
pcknew_reg <= '1';
end if;
if(ffree_pre_mask='1' and (rcv_pckstart_new='1' or pcknew_reg='1') and mmu_force_free_req = '1') then-- or
-- if there is an ongoing FF req, then we cannot clear the mask because
-- it won't be done, never.. However, we remember the
-- rcv_pckstart_new='1' by setting ll_stored_eof to '0'.
ffree_pre_mask <= '1';
ll_stored_eof <= '0';
pcknew_reg <= '0';
elsif(ffree_pre_mask='1' and (rcv_pckstart_new='1' or pcknew_reg='1')) then
-- that is the most "clean" situation, there is no FF request and we're
-- receiving new frame, therefore we clear ffree_mask and ll_stored_eof.
ffree_pre_mask <= '0';
ll_stored_eof <= '0';
pcknew_reg <= '0';
elsif(ffree_pre_mask = '1' and ll_stored_eof = '0' and mmu_force_free_req = '0' and
(ll_wr_done_i='0' or ll_entry.valid='0' or ll_entry.eof='0')) then
-- in case the first condition took place, (and in the meantime the
-- start page was not saved to linked list) clear the ffree_mask when
-- FF req is no more active.
ffree_pre_mask <= '0';
elsif(ffree_pre_mask = '0' and ll_wr_done_i = '1' and ll_entry.valid='1' and ll_entry.eof='1') then
-- we allow force-free only if everything for the
-- frame is stored in the LL. Otherwise (and I've seen this for 18p
-- snake test) we have a race condition.
ll_stored_eof <= '1';
ffree_pre_mask <= '1';
end if;
elsif(ffree_mask = '0' and ll_entry.addr = current_pckstart_pageaddr and ll_wr_done_i = '1' and
ll_entry.valid='1' and (ll_entry.next_page_valid='1' or ll_entry.eof='1')) then
-- if we get ll_wr_done_i while LL FSM is in EOF_ON_WR, that means start
-- page will be written one more time to LL but this time with EOF bit
-- set. We have to wait for another ll_wr_done, therefore in that case
-- we don't yet set ffree_mask to '1', but only remember the situation
-- by setting ll_pckstart_stored to '1'.
-- Allow Force Free requests only if valid start page is written to LL.
-- Valid means either with next page pointer or EOF.
ll_pckstart_stored <= '1';
if(s_ll_write /= S_EOF_ON_WR) then
if (ffree_pre_mask='1' and mmu_force_free_req='1' and
(mmu_force_free_addr=current_pckstart_ll_stored or (mmu_force_free_addr=ll_entry.addr and ll_wr_done_i='1'))) then
ffree_mask <= '1';
elsif(ffree_pre_mask='0' or rcv_pckstart_new='1' or pcknew_reg='1' or
(ll_stored_eof='0' and mmu_force_free_req='0' and (ll_wr_done_i='0' or ll_entry.valid='0' or ll_entry.eof='0'))) then
ffree_mask <= '0';
end if;
elsif(ffree_mask = '0' and ll_pckstart_stored='1' and ll_wr_done_i='1') then
-- here is the continuation of the above condition. When we detect
-- second write after EOF_ON_WR we set ffree_mask to '1' marking that
-- the start page was stored to LL.
ffree_mask <= '1';
end if;
end if;
end process;
--================================================================================================
--------------------------------------------------------------------------------------------------
-------------------------------- resource management --------------------------------
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment