net: dqs: add NIC stall detector based on BQL

softnet_data->time_squeeze is sometimes used as a proxy for
host overload or indication of scheduling problems. In practice
this statistic is very noisy and has hard to grasp units -
e.g. is 10 squeezes a second to be expected, or high?

Delaying network (NAPI) processing leads to drops on NIC queues
but also RTT bloat, impacting pacing and CA decisions.
Stalls are a little hard to detect on the Rx side, because
there may simply have not been any packets received in given
period of time. Packet timestamps help a little bit, but
again we don't know if packets are stale because we're
not keeping up or because someone (*cough* cgroups)
disabled IRQs for a long time.

We can, however, use Tx as a proxy for Rx stalls. Most drivers
use combined Rx+Tx NAPIs so if Tx gets starved so will Rx.
On the Tx side we know exactly when packets get queued,
and completed, so there is no uncertainty.

This patch adds stall checks to BQL. Why BQL? Because
it's a convenient place to add such checks, already
called by most drivers, and it has copious free space
in its structures (this patch adds no extra cache
references or dirtying to the fast path).

The algorithm takes one parameter - max delay AKA stall
threshold and increments a counter whenever NAPI got delayed
for at least that amount of time. It also records the length
of the longest stall.

To be precise every time NAPI has not polled for at least
stall thrs we check if there were any Tx packets queued
between last NAPI run and now - stall_thrs/2.

Unlike the classic Tx watchdog this mechanism does not
ignore stalls caused by Tx being disabled, or loss of link.
I don't think the check is worth the complexity, and
stall is a stall, whether due to host overload, flow
control, link down... doesn't matter much to the application.

We have been running this detector in production at Meta
for 2 years, with the threshold of 8ms. It's the lowest
value where false positives become rare. There's still
a constant stream of reported stalls (especially without
the ksoftirqd deferral patches reverted), those who like
their stall metrics to be 0 may prefer higher value.

Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Breno Leitao <leitao@debian.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

1 files changed, 74 insertions, 0 deletions

diff --git a/lib/dynamic_queue_limits.c b/lib/dynamic_queue_limits.c
index fde0aa244148..a1389db1c30a 100644
--- a/lib/dynamic_queue_limits.c
+++ b/lib/dynamic_queue_limits.c
@@ -10,10 +10,77 @@
 #include <linux/dynamic_queue_limits.h>
 #include <linux/compiler.h>
 #include <linux/export.h>
+#include <trace/events/napi.h>
 
 #define POSDIFF(A, B) ((int)((A) - (B)) > 0 ? (A) - (B) : 0)
 #define AFTER_EQ(A, B) ((int)((A) - (B)) >= 0)
 
+static void dql_check_stall(struct dql *dql)
+{
+	unsigned short stall_thrs;
+	unsigned long now;
+
+	stall_thrs = READ_ONCE(dql->stall_thrs);
+	if (!stall_thrs)
+		return;
+
+	now = jiffies;
+	/* Check for a potential stall */
+	if (time_after_eq(now, dql->last_reap + stall_thrs)) {
+		unsigned long hist_head, t, start, end;
+
+		/* We are trying to detect a period of at least @stall_thrs
+		 * jiffies without any Tx completions, but during first half
+		 * of which some Tx was posted.
+		 */
+dqs_again:
+		hist_head = READ_ONCE(dql->history_head);
+		/* pairs with smp_wmb() in dql_queued() */
+		smp_rmb();
+
+		/* Get the previous entry in the ring buffer, which is the
+		 * oldest sample.
+		 */
+		start = (hist_head - DQL_HIST_LEN + 1) * BITS_PER_LONG;
+
+		/* Advance start to continue from the last reap time */
+		if (time_before(start, dql->last_reap + 1))
+			start = dql->last_reap + 1;
+
+		/* Newest sample we should have already seen a completion for */
+		end = hist_head * BITS_PER_LONG + (BITS_PER_LONG - 1);
+
+		/* Shrink the search space to [start, (now - start_thrs/2)] if
+		 * `end` is beyond the stall zone
+		 */
+		if (time_before(now, end + stall_thrs / 2))
+			end = now - stall_thrs / 2;
+
+		/* Search for the queued time in [t, end] */
+		for (t = start; time_before_eq(t, end); t++)
+			if (test_bit(t % (DQL_HIST_LEN * BITS_PER_LONG),
+				     dql->history))
+				break;
+
+		/* Variable t contains the time of the queue */
+		if (!time_before_eq(t, end))
+			goto no_stall;
+
+		/* The ring buffer was modified in the meantime, retry */
+		if (hist_head != READ_ONCE(dql->history_head))
+			goto dqs_again;
+
+		dql->stall_cnt++;
+		dql->stall_max = max_t(unsigned short, dql->stall_max, now - t);
+
+		trace_dql_stall_detected(dql->stall_thrs, now - t,
+					 dql->last_reap, dql->history_head,
+					 now, dql->history);
+	}
+no_stall:
+	dql->last_reap = now;
+}
+
 /* Records completed count and recalculates the queue limit */
 void dql_completed(struct dql *dql, unsigned int count)
 {
@@ -110,6 +177,8 @@ void dql_completed(struct dql *dql, unsigned int count)
 	dql->prev_last_obj_cnt = dql->last_obj_cnt;
 	dql->num_completed = completed;
 	dql->prev_num_queued = num_queued;
+
+	dql_check_stall(dql);
 }
 EXPORT_SYMBOL(dql_completed);
 
@@ -125,6 +194,10 @@ void dql_reset(struct dql *dql)
 	dql->prev_ovlimit = 0;
 	dql->lowest_slack = UINT_MAX;
 	dql->slack_start_time = jiffies;
+
+	dql->last_reap = jiffies;
+	dql->history_head = jiffies / BITS_PER_LONG;
+	memset(dql->history, 0, sizeof(dql->history));
 }
 EXPORT_SYMBOL(dql_reset);
 
@@ -133,6 +206,7 @@ void dql_init(struct dql *dql, unsigned int hold_time)
 	dql->max_limit = DQL_MAX_LIMIT;
 	dql->min_limit = 0;
 	dql->slack_hold_time = hold_time;
+	dql->stall_thrs = 0;
 	dql_reset(dql);
 }
 EXPORT_SYMBOL(dql_init);