Merge branch 'sfc-ARFS-expiry-improvements'

Edward Cree says:

====================
A series of changes to how we check filters for expiry, manage how much
 of that work to do & when, etc.
Prompted by some pathological behaviour under heavy load, which was
Reported-by: David Ahern <dahern@digitalocean.com>
====================

Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>

6 files changed, 94 insertions, 41 deletions

diff --git a/drivers/net/ethernet/sfc/ef10.c b/drivers/net/ethernet/sfc/ef10.c
index ad68eb0cb8fd..4d9bbccc6f89 100644
--- a/drivers/net/ethernet/sfc/ef10.c
+++ b/drivers/net/ethernet/sfc/ef10.c
@@ -4202,11 +4202,15 @@ static int efx_ef10_filter_push(struct efx_nic *efx,
 {
 	MCDI_DECLARE_BUF(inbuf, MC_CMD_FILTER_OP_EXT_IN_LEN);
 	MCDI_DECLARE_BUF(outbuf, MC_CMD_FILTER_OP_EXT_OUT_LEN);
+	size_t outlen;
 	int rc;
 
 	efx_ef10_filter_push_prep(efx, spec, inbuf, *handle, ctx, replacing);
-	rc = efx_mcdi_rpc(efx, MC_CMD_FILTER_OP, inbuf, sizeof(inbuf),
-			  outbuf, sizeof(outbuf), NULL);
+	rc = efx_mcdi_rpc_quiet(efx, MC_CMD_FILTER_OP, inbuf, sizeof(inbuf),
+				outbuf, sizeof(outbuf), &outlen);
+	if (rc && spec->priority != EFX_FILTER_PRI_HINT)
+		efx_mcdi_display_error(efx, MC_CMD_FILTER_OP, sizeof(inbuf),
+				       outbuf, outlen, rc);
 	if (rc == 0)
 		*handle = MCDI_QWORD(outbuf, FILTER_OP_OUT_HANDLE);
 	if (rc == -ENOSPC)
diff --git a/drivers/net/ethernet/sfc/efx.c b/drivers/net/ethernet/sfc/efx.c
index 0fa9972027db..992c773620ec 100644
--- a/drivers/net/ethernet/sfc/efx.c
+++ b/drivers/net/ethernet/sfc/efx.c
@@ -355,7 +355,7 @@ static int efx_poll(struct napi_struct *napi, int budget)
 
 #ifdef CONFIG_RFS_ACCEL
 		/* Perhaps expire some ARFS filters */
-		schedule_work(&channel->filter_work);
+		mod_delayed_work(system_wq, &channel->filter_work, 0);
 #endif
 
 		/* There is no race here; although napi_disable() will
@@ -487,7 +487,7 @@ efx_alloc_channel(struct efx_nic *efx, int i, struct efx_channel *old_channel)
 	}
 
 #ifdef CONFIG_RFS_ACCEL
-	INIT_WORK(&channel->filter_work, efx_filter_rfs_expire);
+	INIT_DELAYED_WORK(&channel->filter_work, efx_filter_rfs_expire);
 #endif
 
 	rx_queue = &channel->rx_queue;
@@ -533,7 +533,7 @@ efx_copy_channel(const struct efx_channel *old_channel)
 	memset(&rx_queue->rxd, 0, sizeof(rx_queue->rxd));
 	timer_setup(&rx_queue->slow_fill, efx_rx_slow_fill, 0);
 #ifdef CONFIG_RFS_ACCEL
-	INIT_WORK(&channel->filter_work, efx_filter_rfs_expire);
+	INIT_DELAYED_WORK(&channel->filter_work, efx_filter_rfs_expire);
 #endif
 
 	return channel;
@@ -1969,6 +1969,8 @@ static int efx_probe_filters(struct efx_nic *efx)
 				     ++i)
 					channel->rps_flow_id[i] =
 						RPS_FLOW_ID_INVALID;
+			channel->rfs_expire_index = 0;
+			channel->rfs_filter_count = 0;
 		}
 
 		if (!success) {
@@ -1978,8 +1980,6 @@ static int efx_probe_filters(struct efx_nic *efx)
 			rc = -ENOMEM;
 			goto out_unlock;
 		}
-
-		efx->rps_expire_index = efx->rps_expire_channel = 0;
 	}
 #endif
 out_unlock:
@@ -1993,8 +1993,10 @@ static void efx_remove_filters(struct efx_nic *efx)
 #ifdef CONFIG_RFS_ACCEL
 	struct efx_channel *channel;
 
-	efx_for_each_channel(channel, efx)
+	efx_for_each_channel(channel, efx) {
+		cancel_delayed_work_sync(&channel->filter_work);
 		kfree(channel->rps_flow_id);
+	}
 #endif
 	down_write(&efx->filter_sem);
 	efx->type->filter_table_remove(efx);
diff --git a/drivers/net/ethernet/sfc/efx.h b/drivers/net/ethernet/sfc/efx.h
index 45c7ae4114ec..2dd8d5002315 100644
--- a/drivers/net/ethernet/sfc/efx.h
+++ b/drivers/net/ethernet/sfc/efx.h
@@ -166,15 +166,20 @@ static inline s32 efx_filter_get_rx_ids(struct efx_nic *efx,
 #ifdef CONFIG_RFS_ACCEL
 int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
 		   u16 rxq_index, u32 flow_id);
-bool __efx_filter_rfs_expire(struct efx_nic *efx, unsigned quota);
+bool __efx_filter_rfs_expire(struct efx_channel *channel, unsigned int quota);
 static inline void efx_filter_rfs_expire(struct work_struct *data)
 {
-	struct efx_channel *channel = container_of(data, struct efx_channel,
-						   filter_work);
-
-	if (channel->rfs_filters_added >= 60 &&
-	    __efx_filter_rfs_expire(channel->efx, 100))
-		channel->rfs_filters_added -= 60;
+	struct delayed_work *dwork = to_delayed_work(data);
+	struct efx_channel *channel;
+	unsigned int time, quota;
+
+	channel = container_of(dwork, struct efx_channel, filter_work);
+	time = jiffies - channel->rfs_last_expiry;
+	quota = channel->rfs_filter_count * time / (30 * HZ);
+	if (quota > 20 && __efx_filter_rfs_expire(channel, min(channel->rfs_filter_count, quota)))
+		channel->rfs_last_expiry += time;
+	/* Ensure we do more work eventually even if NAPI poll is not happening */
+	schedule_delayed_work(dwork, 30 * HZ);
 }
 #define efx_filter_rfs_enabled() 1
 #else
diff --git a/drivers/net/ethernet/sfc/ethtool.c b/drivers/net/ethernet/sfc/ethtool.c
index 8db593fb9699..6a9347cd67f3 100644
--- a/drivers/net/ethernet/sfc/ethtool.c
+++ b/drivers/net/ethernet/sfc/ethtool.c
@@ -56,6 +56,9 @@ static u64 efx_get_atomic_stat(void *field)
 #define EFX_ETHTOOL_UINT_CHANNEL_STAT(field)			\
 	EFX_ETHTOOL_STAT(field, channel, n_##field,		\
 			 unsigned int, efx_get_uint_stat)
+#define EFX_ETHTOOL_UINT_CHANNEL_STAT_NO_N(field)		\
+	EFX_ETHTOOL_STAT(field, channel, field,			\
+			 unsigned int, efx_get_uint_stat)
 
 #define EFX_ETHTOOL_UINT_TXQ_STAT(field)			\
 	EFX_ETHTOOL_STAT(tx_##field, tx_queue, field,		\
@@ -87,6 +90,9 @@ static const struct efx_sw_stat_desc efx_sw_stat_desc[] = {
 	EFX_ETHTOOL_UINT_CHANNEL_STAT(rx_xdp_bad_drops),
 	EFX_ETHTOOL_UINT_CHANNEL_STAT(rx_xdp_tx),
 	EFX_ETHTOOL_UINT_CHANNEL_STAT(rx_xdp_redirect),
+	EFX_ETHTOOL_UINT_CHANNEL_STAT_NO_N(rfs_filter_count),
+	EFX_ETHTOOL_UINT_CHANNEL_STAT(rfs_succeeded),
+	EFX_ETHTOOL_UINT_CHANNEL_STAT(rfs_failed),
 };
 
 #define EFX_ETHTOOL_SW_STAT_COUNT ARRAY_SIZE(efx_sw_stat_desc)
diff --git a/drivers/net/ethernet/sfc/net_driver.h b/drivers/net/ethernet/sfc/net_driver.h
index 04e49eac7327..1f88212be085 100644
--- a/drivers/net/ethernet/sfc/net_driver.h
+++ b/drivers/net/ethernet/sfc/net_driver.h
@@ -439,6 +439,13 @@ enum efx_sync_events_state {
  * @event_test_cpu: Last CPU to handle interrupt or test event for this channel
  * @irq_count: Number of IRQs since last adaptive moderation decision
  * @irq_mod_score: IRQ moderation score
+ * @rfs_filter_count: number of accelerated RFS filters currently in place;
+ *	equals the count of @rps_flow_id slots filled
+ * @rfs_last_expiry: value of jiffies last time some accelerated RFS filters
+ *	were checked for expiry
+ * @rfs_expire_index: next accelerated RFS filter ID to check for expiry
+ * @n_rfs_succeeded: number of successful accelerated RFS filter insertions
+ * @n_rfs_failed; number of failed accelerated RFS filter insertions
  * @filter_work: Work item for efx_filter_rfs_expire()
  * @rps_flow_id: Flow IDs of filters allocated for accelerated RFS,
  *      indexed by filter ID
@@ -489,8 +496,12 @@ struct efx_channel {
 	unsigned int irq_count;
 	unsigned int irq_mod_score;
 #ifdef CONFIG_RFS_ACCEL
-	unsigned int rfs_filters_added;
-	struct work_struct filter_work;
+	unsigned int rfs_filter_count;
+	unsigned int rfs_last_expiry;
+	unsigned int rfs_expire_index;
+	unsigned int n_rfs_succeeded;
+	unsigned int n_rfs_failed;
+	struct delayed_work filter_work;
 #define RPS_FLOW_ID_INVALID 0xFFFFFFFF
 	u32 *rps_flow_id;
 #endif
@@ -923,9 +934,6 @@ struct efx_async_filter_insertion {
  * @filter_sem: Filter table rw_semaphore, protects existence of @filter_state
  * @filter_state: Architecture-dependent filter table state
  * @rps_mutex: Protects RPS state of all channels
- * @rps_expire_channel: Next channel to check for expiry
- * @rps_expire_index: Next index to check for expiry in
- *	@rps_expire_channel's @rps_flow_id
  * @rps_slot_map: bitmap of in-flight entries in @rps_slot
  * @rps_slot: array of ARFS insertion requests for efx_filter_rfs_work()
  * @rps_hash_lock: Protects ARFS filter mapping state (@rps_hash_table and
@@ -1096,8 +1104,6 @@ struct efx_nic {
 	void *filter_state;
 #ifdef CONFIG_RFS_ACCEL
 	struct mutex rps_mutex;
-	unsigned int rps_expire_channel;
-	unsigned int rps_expire_index;
 	unsigned long rps_slot_map;
 	struct efx_async_filter_insertion rps_slot[EFX_RPS_MAX_IN_FLIGHT];
 	spinlock_t rps_hash_lock;
diff --git a/drivers/net/ethernet/sfc/rx.c b/drivers/net/ethernet/sfc/rx.c
index bec261905530..ef52b24ad9e7 100644
--- a/drivers/net/ethernet/sfc/rx.c
+++ b/drivers/net/ethernet/sfc/rx.c
@@ -988,6 +988,7 @@ static void efx_filter_rfs_work(struct work_struct *data)
 
 	rc = efx->type->filter_insert(efx, &req->spec, true);
 	if (rc >= 0)
+		/* Discard 'priority' part of EF10+ filter ID (mcdi_filters) */
 		rc %= efx->type->max_rx_ip_filters;
 	if (efx->rps_hash_table) {
 		spin_lock_bh(&efx->rps_hash_lock);
@@ -1012,8 +1013,9 @@ static void efx_filter_rfs_work(struct work_struct *data)
 		 * later.
 		 */
 		mutex_lock(&efx->rps_mutex);
+		if (channel->rps_flow_id[rc] == RPS_FLOW_ID_INVALID)
+			channel->rfs_filter_count++;
 		channel->rps_flow_id[rc] = req->flow_id;
-		++channel->rfs_filters_added;
 		mutex_unlock(&efx->rps_mutex);
 
 		if (req->spec.ether_type == htons(ETH_P_IP))
@@ -1030,6 +1032,28 @@ static void efx_filter_rfs_work(struct work_struct *data)
 				   req->spec.rem_host, ntohs(req->spec.rem_port),
 				   req->spec.loc_host, ntohs(req->spec.loc_port),
 				   req->rxq_index, req->flow_id, rc, arfs_id);
+		channel->n_rfs_succeeded++;
+	} else {
+		if (req->spec.ether_type == htons(ETH_P_IP))
+			netif_dbg(efx, rx_status, efx->net_dev,
+				  "failed to steer %s %pI4:%u:%pI4:%u to queue %u [flow %u rc %d id %u]\n",
+				  (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
+				  req->spec.rem_host, ntohs(req->spec.rem_port),
+				  req->spec.loc_host, ntohs(req->spec.loc_port),
+				  req->rxq_index, req->flow_id, rc, arfs_id);
+		else
+			netif_dbg(efx, rx_status, efx->net_dev,
+				  "failed to steer %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u rc %d id %u]\n",
+				  (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
+				  req->spec.rem_host, ntohs(req->spec.rem_port),
+				  req->spec.loc_host, ntohs(req->spec.loc_port),
+				  req->rxq_index, req->flow_id, rc, arfs_id);
+		channel->n_rfs_failed++;
+		/* We're overloading the NIC's filter tables, so let's do a
+		 * chunk of extra expiry work.
+		 */
+		__efx_filter_rfs_expire(channel, min(channel->rfs_filter_count,
+						     100u));
 	}
 
 	/* Release references */
@@ -1139,38 +1163,44 @@ out_clear:
 	return rc;
 }
 
-bool __efx_filter_rfs_expire(struct efx_nic *efx, unsigned int quota)
+bool __efx_filter_rfs_expire(struct efx_channel *channel, unsigned int quota)
 {
 	bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index);
-	unsigned int channel_idx, index, size;
+	struct efx_nic *efx = channel->efx;
+	unsigned int index, size, start;
 	u32 flow_id;
 
 	if (!mutex_trylock(&efx->rps_mutex))
 		return false;
 	expire_one = efx->type->filter_rfs_expire_one;
-	channel_idx = efx->rps_expire_channel;
-	index = efx->rps_expire_index;
+	index = channel->rfs_expire_index;
+	start = index;
 	size = efx->type->max_rx_ip_filters;
-	while (quota--) {
-		struct efx_channel *channel = efx_get_channel(efx, channel_idx);
+	while (quota) {
 		flow_id = channel->rps_flow_id[index];
 
-		if (flow_id != RPS_FLOW_ID_INVALID &&
-		    expire_one(efx, flow_id, index)) {
-			netif_info(efx, rx_status, efx->net_dev,
-				   "expired filter %d [queue %u flow %u]\n",
-				   index, channel_idx, flow_id);
-			channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID;
+		if (flow_id != RPS_FLOW_ID_INVALID) {
+			quota--;
+			if (expire_one(efx, flow_id, index)) {
+				netif_info(efx, rx_status, efx->net_dev,
+					   "expired filter %d [channel %u flow %u]\n",
+					   index, channel->channel, flow_id);
+				channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID;
+				channel->rfs_filter_count--;
+			}
 		}
-		if (++index == size) {
-			if (++channel_idx == efx->n_channels)
-				channel_idx = 0;
+		if (++index == size)
 			index = 0;
-		}
+		/* If we were called with a quota that exceeds the total number
+		 * of filters in the table (which shouldn't happen, but could
+		 * if two callers race), ensure that we don't loop forever -
+		 * stop when we've examined every row of the table.
+		 */
+		if (index == start)
+			break;
 	}
-	efx->rps_expire_channel = channel_idx;
-	efx->rps_expire_index = index;
 
+	channel->rfs_expire_index = index;
 	mutex_unlock(&efx->rps_mutex);
 	return true;
 }