svn commit: r366958 - head/sys/dev/cxgbe
Navdeep Parhar
np at FreeBSD.org
Fri Oct 23 01:36:55 UTC 2020
Author: np
Date: Fri Oct 23 01:36:54 2020
New Revision: 366958
URL: https://svnweb.freebsd.org/changeset/base/366958
Log:
cxgbe(4): refine the values reported in if_ratelimit_query.
- Get the number of classes from chip_params.
- Get the number of ethofld tids from the firmware.
- Do not let tcp_ratelimit allocate all traffic classes.
Sponsored by: Chelsio Communications
Modified:
head/sys/dev/cxgbe/t4_sched.c
Modified: head/sys/dev/cxgbe/t4_sched.c
==============================================================================
--- head/sys/dev/cxgbe/t4_sched.c Fri Oct 23 01:06:42 2020 (r366957)
+++ head/sys/dev/cxgbe/t4_sched.c Fri Oct 23 01:36:54 2020 (r366958)
@@ -903,34 +903,46 @@ cxgbe_rate_tag_free(struct m_snd_tag *mst)
mtx_unlock(&cst->lock);
}
-#define CXGBE_MAX_FLOWS 4000 /* Testing show so far thats all this adapter can do */
-#define CXGBE_UNIQUE_RATE_COUNT 16 /* Number of unique rates that can be setup */
-
void
-cxgbe_ratelimit_query(struct ifnet *ifp __unused,
- struct if_ratelimit_query_results *q)
+cxgbe_ratelimit_query(struct ifnet *ifp, struct if_ratelimit_query_results *q)
{
- /*
- * This is a skeleton and needs future work
- * by the driver supporters. It should be
- * enhanced to look at the specific type of
- * interface and select approprate values
- * for these settings. This example goes
- * with an earlier card (t5), it has a maximum
- * number of 16 rates that the first guys in
- * select (thus the flags value RT_IS_SELECTABLE).
- * If it was a fixed table then we would setup a
- * const array (example mlx5). Note the card tested
- * can only support reasonably 4000 flows before
- * the adapter has issues with sending so here
- * we limit the number of flows using hardware
- * pacing to that number, other cards may
- * be able to raise or eliminate this limit.
- */
+ struct vi_info *vi = ifp->if_softc;
+ struct adapter *sc = vi->adapter;
+
q->rate_table = NULL;
q->flags = RT_IS_SELECTABLE;
- q->max_flows = CXGBE_MAX_FLOWS;
- q->number_of_rates = CXGBE_UNIQUE_RATE_COUNT;
- q->min_segment_burst = 4; /* Driver emits 4 in a burst */
+ /*
+ * Absolute max limits from the firmware configuration. Practical
+ * limits depend on the burstsize, pktsize (ifp->if_mtu ultimately) and
+ * the card's cclk.
+ */
+ q->max_flows = sc->tids.netids;
+ q->number_of_rates = sc->chip_params->nsched_cls;
+ q->min_segment_burst = 4; /* matches PKTSCHED_BURST in the firmware. */
+
+#if 1
+ if (chip_id(sc) < CHELSIO_T6) {
+ /* Based on testing by rrs@ with a T580 at burstsize = 4. */
+ MPASS(q->min_segment_burst == 4);
+ q->max_flows = max(4000, q->max_flows);
+ } else {
+ /* XXX: TBD, carried forward from T5 for now. */
+ q->max_flows = max(4000, q->max_flows);
+ }
+
+ /*
+ * XXX: tcp_ratelimit.c grabs all available rates on link-up before it
+ * even knows whether hw pacing will be used or not. This prevents
+ * other consumers like SO_MAX_PACING_RATE or those using cxgbetool or
+ * the private ioctls from using any of traffic classes.
+ *
+ * Underreport the number of rates to tcp_ratelimit so that it doesn't
+ * hog all of them. This can be removed if/when tcp_ratelimit switches
+ * to making its allocations on first-use rather than link-up. There is
+ * nothing wrong with one particular consumer reserving all the classes
+ * but it should do so only if it'll actually use hw rate limiting.
+ */
+ q->number_of_rates /= 4;
+#endif
}
#endif
More information about the svn-src-all
mailing list