poh: fix leader ticking race

Author: yufeng-jump

src/discof/poh/fd_poh.c

@@ -256,6 +256,11 @@ fd_poh_hashing_to_leader_slot( fd_poh_t const * poh ) {
   return hashing && poh->slot<poh->next_leader_slot;
 }
 
+int
+fd_poh_must_tick( fd_poh_t const * poh ) {
+  return poh->state==STATE_LEADER && (poh->hashcnt%poh->hashcnt_per_tick)==(poh->hashcnt_per_tick-1UL);
+}
+
 void
 fd_poh_done_packing( fd_poh_t * poh,
                      ulong      microblocks_in_slot ) {
@@ -650,6 +655,7 @@ fd_poh1_mixin( fd_poh_t *          poh,
   if( FD_UNLIKELY( slot!=poh->next_leader_slot || slot!=poh->slot ) ) {
     FD_LOG_ERR(( "packed too early or late slot=%lu, current_slot=%lu", slot, poh->slot ));
   }
+  if( FD_UNLIKELY( (poh->hashcnt%poh->hashcnt_per_tick)==(poh->hashcnt_per_tick-1UL) ) ) FD_LOG_CRIT(( "a tick will be skipped due to hashcnt %lu hashcnt_per_tick %lu", poh->hashcnt, poh->hashcnt_per_tick ));
 
   FD_TEST( poh->state==STATE_LEADER );
   FD_TEST( poh->microblocks_lower_bound<poh->max_microblocks_per_slot );

src/discof/poh/fd_poh.h

@@ -483,6 +483,9 @@ fd_poh_have_leader_bank( fd_poh_t const * poh );
 int
 fd_poh_hashing_to_leader_slot( fd_poh_t const * poh );
 
+int
+fd_poh_must_tick( fd_poh_t const * poh );
+
 void
 fd_poh_begin_leader( fd_poh_t * poh,
                      ulong      slot,

src/discof/poh/fd_poh_tile.c

@@ -67,7 +67,7 @@ after_credit( fd_poh_tile_t *     ctx,
               int *               opt_poll_in,
               int *               charge_busy ) {
   ctx->idle_cnt++;
-  if( FD_UNLIKELY( ctx->idle_cnt>=2UL*ctx->in_cnt ) ) {
+  if( FD_LIKELY( ctx->idle_cnt>=2UL*ctx->in_cnt || fd_poh_must_tick( ctx->poh ) ) ) {
     /* We would like to fully drain input links to the best of our
        knowledge, before we spend cycles on hashing.  That is, we would
        like to assert that all input links have stayed empty since the
@@ -77,7 +77,12 @@ after_credit( fd_poh_tile_t *     ctx,
        in_cnt-1 in the subsequent input link shuffle.  So strictly
        speaking we will need to have observed 2*in_cnt-1 consecutive
        empty in links to be able to assert that link L has been empty
-       since the last time we polled it. */
+       since the last time we polled it.
+
+       Except that when we are leader and the hashcnt is right before a
+       tick boundary, poh must advance to the tick boundary and produce
+       the tick.  Otherwise, a tick will be skipped if a microblock
+       mixin happens. */
     fd_poh_advance( ctx->poh, stem, opt_poll_in, charge_busy );
     ctx->idle_cnt = 0UL;
   }