iflib: leave only 1 receive descriptor unused

The pidx argument of isc_rxd_flush() indicates which is the last valid
receive descriptor to be used by the NIC. However, current code has
multiple issues:
  - Intel drivers write pidx to their RDT register, which means that
    NICs will only use the descriptors up to pidx-1 (modulo ring size N),
    and won't actually use the one pointed by pidx. This does not break
    reception, but it is anyway confusing and suboptimal (the NIC will
    actually see only N-2 descriptors as available, rather than N-1).
    Other drivers (if_vmx, if_bnxt, if_mgb) adhere to this semantic).
  - The semantic used by Intel (RDT is one descriptor past the last
     valid one) is used by most (if not all) NICs, and it is also used
     on the TX side (also in iflib). Since iflib is not currently
     using this semantic for RX, it must decrement fl->ifl_pidx
     (modulo N) before calling isc_rxd_flush(), and then the
     per-driver callback implementation must increment the index
     again (to match the real semantic). This is confusing and suboptimal.
  -  The iflib refill function is also called at initialization.
     However, in case the ring size is smaller than 128 (e.g. if_mgb),
     the refill function will actually prepare all the receive
     descriptors (N), without leaving one unused, as most of NICs assume
     (e.g. to avoid RDT to overrun RDH). I can speculate that the code
     looks like this right now because this issue showed up during
     testing (e.g. with if_mgb), and it was easy to workaround by
     decrementing pidx before isc_rxd_flush().

The goal of this change is to simplify the code (removing a bunch
of instructions from the RX fast path), and to make the semantic of
isc_rxd_flush() consistent across drivers. To achieve this, we:
  - change the semantics of the pidx argument to the usual one (that
    is the index one past the last valid one), so that both iflib and
    drivers avoid the decrement/increment dance.
  - fix the initialization code to prepare at most N-1 descriptors.

Reviewed by:	markj
MFC after:	2 weeks
Differential Revision:	https://reviews.freebsd.org/D26191

1 files changed, 24 insertions, 13 deletions

diff --git a/sys/net/iflib.c b/sys/net/iflib.c
index 8a1254f930d9..2901ad94a844 100644
--- a/sys/net/iflib.c
+++ b/sys/net/iflib.c
@@ -1959,7 +1959,8 @@ _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
  * @count: the number of new buffers to allocate
  *
  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
- * The caller must assure that @count does not exceed the queue's capacity.
+ * The caller must assure that @count does not exceed the queue's capacity
+ * minus one (since we always leave a descriptor unavailable).
  */
 static uint8_t
 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
@@ -1974,6 +1975,8 @@ iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
 	int err, frag_idx, i, idx, n, pidx;
 	qidx_t credits;
 
+	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
+
 	sd_m = fl->ifl_sds.ifsd_m;
 	sd_map = fl->ifl_sds.ifsd_map;
 	sd_cl = fl->ifl_sds.ifsd_cl;
@@ -2081,15 +2084,10 @@ iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
 			fl->ifl_credits = credits;
 		}
 		DBG_COUNTER_INC(rxd_flush);
-		if (fl->ifl_pidx == 0)
-			pidx = fl->ifl_size - 1;
-		else
-			pidx = fl->ifl_pidx - 1;
-
 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
-		    fl->ifl_id, pidx);
+		    fl->ifl_id, fl->ifl_pidx);
 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
 			fl->ifl_fragidx = frag_idx + 1;
 			if (fl->ifl_fragidx == fl->ifl_size)
@@ -2105,7 +2103,17 @@ iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
 static inline uint8_t
 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
 {
-	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
+	/*
+	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
+	 * This is important as it confuses most NICs. For instance,
+	 * Intel NICs have (per receive ring) RDH and RDT registers, where
+	 * RDH points to the next receive descriptor to be used by the NIC,
+	 * and RDT for the next receive descriptor to be published by the
+	 * driver to the NIC (RDT - 1 is thus the last valid one).
+	 * The condition RDH == RDT means no descriptors are available to
+	 * the NIC, and thus it would be ambiguous if it also meant that
+	 * all the descriptors are available to the NIC.
+	 */
 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
 #ifdef INVARIANTS
 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
@@ -2210,12 +2218,15 @@ iflib_fl_setup(iflib_fl_t fl)
 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
 
 
-	/* avoid pre-allocating zillions of clusters to an idle card
-	 * potentially speeding up attach
+	/*
+	 * Avoid pre-allocating zillions of clusters to an idle card
+	 * potentially speeding up attach. In any case make sure
+	 * to leave a descriptor unavailable. See the comment in
+	 * iflib_fl_refill_all().
 	 */
-	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
-	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
-	if (min(128, fl->ifl_size) != fl->ifl_credits)
+	MPASS(fl->ifl_size > 0);
+	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
+	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
 		return (ENOBUFS);
 	/*
 	 * handle failure