sfc: ARFS filter IDs

Associate an arbitrary ID with each ARFS filter, allowing to properly query
 for expiry.  The association is maintained in a hash table, which is
 protected by a spinlock.

v3: fix build warnings when CONFIG_RFS_ACCEL is disabled (thanks lkp-robot).
v2: fixed uninitialised variable (thanks davem and lkp-robot).

Fixes: 3af0f342 ("sfc: replace asynchronous filter operations")
Signed-off-by: Edward Cree <ecree@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/sfc/ef10.c

@@ -3999,29 +3999,6 @@ static void efx_ef10_prepare_flr(struct efx_nic *efx)
 	atomic_set(&efx->active_queues, 0);
 }
 
-static bool efx_ef10_filter_equal(const struct efx_filter_spec *left,
-				  const struct efx_filter_spec *right)
-{
-	if ((left->match_flags ^ right->match_flags) |
-	    ((left->flags ^ right->flags) &
-	     (EFX_FILTER_FLAG_RX | EFX_FILTER_FLAG_TX)))
-		return false;
-
-	return memcmp(&left->outer_vid, &right->outer_vid,
-		      sizeof(struct efx_filter_spec) -
-		      offsetof(struct efx_filter_spec, outer_vid)) == 0;
-}
-
-static unsigned int efx_ef10_filter_hash(const struct efx_filter_spec *spec)
-{
-	BUILD_BUG_ON(offsetof(struct efx_filter_spec, outer_vid) & 3);
-	return jhash2((const u32 *)&spec->outer_vid,
-		      (sizeof(struct efx_filter_spec) -
-		       offsetof(struct efx_filter_spec, outer_vid)) / 4,
-		      0);
-	/* XXX should we randomise the initval? */
-}
-
 /* Decide whether a filter should be exclusive or else should allow
  * delivery to additional recipients.  Currently we decide that
  * filters for specific local unicast MAC and IP addresses are
@@ -4346,7 +4323,7 @@ static s32 efx_ef10_filter_insert(struct efx_nic *efx,
 		goto out_unlock;
 	match_pri = rc;
 
-	hash = efx_ef10_filter_hash(spec);
+	hash = efx_filter_spec_hash(spec);
 	is_mc_recip = efx_filter_is_mc_recipient(spec);
 	if (is_mc_recip)
 		bitmap_zero(mc_rem_map, EFX_EF10_FILTER_SEARCH_LIMIT);
@@ -4378,7 +4355,7 @@ static s32 efx_ef10_filter_insert(struct efx_nic *efx,
 		if (!saved_spec) {
 			if (ins_index < 0)
 				ins_index = i;
-		} else if (efx_ef10_filter_equal(spec, saved_spec)) {
+		} else if (efx_filter_spec_equal(spec, saved_spec)) {
 			if (spec->priority < saved_spec->priority &&
 			    spec->priority != EFX_FILTER_PRI_AUTO) {
 				rc = -EPERM;
@@ -4762,27 +4739,62 @@ static s32 efx_ef10_filter_get_rx_ids(struct efx_nic *efx,
 static bool efx_ef10_filter_rfs_expire_one(struct efx_nic *efx, u32 flow_id,
 					   unsigned int filter_idx)
 {
+	struct efx_filter_spec *spec, saved_spec;
 	struct efx_ef10_filter_table *table;
-	struct efx_filter_spec *spec;
-	bool ret;
+	struct efx_arfs_rule *rule = NULL;
+	bool ret = true, force = false;
+	u16 arfs_id;
 
 	down_read(&efx->filter_sem);
 	table = efx->filter_state;
 	down_write(&table->lock);
 	spec = efx_ef10_filter_entry_spec(table, filter_idx);
 
-	if (!spec || spec->priority != EFX_FILTER_PRI_HINT) {
-		ret = true;
+	if (!spec || spec->priority != EFX_FILTER_PRI_HINT)
 		goto out_unlock;
-	}
 
-	if (!rps_may_expire_flow(efx->net_dev, spec->dmaq_id, flow_id, 0)) {
-		ret = false;
+	spin_lock_bh(&efx->rps_hash_lock);
+	if (!efx->rps_hash_table) {
+		/* In the absence of the table, we always return 0 to ARFS. */
+		arfs_id = 0;
+	} else {
+		rule = efx_rps_hash_find(efx, spec);
+		if (!rule)
+			/* ARFS table doesn't know of this filter, so remove it */
+			goto expire;
+		arfs_id = rule->arfs_id;
+		ret = efx_rps_check_rule(rule, filter_idx, &force);
+		if (force)
+			goto expire;
+		if (!ret) {
+			spin_unlock_bh(&efx->rps_hash_lock);
 			goto out_unlock;
 		}
-
+	}
+	if (!rps_may_expire_flow(efx->net_dev, spec->dmaq_id, flow_id, arfs_id))
+		ret = false;
+	else if (rule)
+		rule->filter_id = EFX_ARFS_FILTER_ID_REMOVING;
+expire:
+	saved_spec = *spec; /* remove operation will kfree spec */
+	spin_unlock_bh(&efx->rps_hash_lock);
+	/* At this point (since we dropped the lock), another thread might queue
+	 * up a fresh insertion request (but the actual insertion will be held
+	 * up by our possession of the filter table lock).  In that case, it
+	 * will set rule->filter_id to EFX_ARFS_FILTER_ID_PENDING, meaning that
+	 * the rule is not removed by efx_rps_hash_del() below.
+	 */
 	ret = efx_ef10_filter_remove_internal(efx, 1U << spec->priority,
 					      filter_idx, true) == 0;
+	/* While we can't safely dereference rule (we dropped the lock), we can
+	 * still test it for NULL.
+	 */
+	if (ret && rule) {
+		/* Expiring, so remove entry from ARFS table */
+		spin_lock_bh(&efx->rps_hash_lock);
+		efx_rps_hash_del(efx, &saved_spec);
+		spin_unlock_bh(&efx->rps_hash_lock);
+	}
 out_unlock:
 	up_write(&table->lock);
 	up_read(&efx->filter_sem);

drivers/net/ethernet/sfc/efx.c

@@ -3027,6 +3027,10 @@ static int efx_init_struct(struct efx_nic *efx,
 	mutex_init(&efx->mac_lock);
 #ifdef CONFIG_RFS_ACCEL
 	mutex_init(&efx->rps_mutex);
+	spin_lock_init(&efx->rps_hash_lock);
+	/* Failure to allocate is not fatal, but may degrade ARFS performance */
+	efx->rps_hash_table = kcalloc(EFX_ARFS_HASH_TABLE_SIZE,
+				      sizeof(*efx->rps_hash_table), GFP_KERNEL);
 #endif
 	efx->phy_op = &efx_dummy_phy_operations;
 	efx->mdio.dev = net_dev;
@@ -3070,6 +3074,10 @@ static void efx_fini_struct(struct efx_nic *efx)
 {
 	int i;
 
+#ifdef CONFIG_RFS_ACCEL
+	kfree(efx->rps_hash_table);
+#endif
+
 	for (i = 0; i < EFX_MAX_CHANNELS; i++)
 		kfree(efx->channel[i]);
 
@@ -3092,6 +3100,141 @@ void efx_update_sw_stats(struct efx_nic *efx, u64 *stats)
 	stats[GENERIC_STAT_rx_noskb_drops] = atomic_read(&efx->n_rx_noskb_drops);
 }
 
+bool efx_filter_spec_equal(const struct efx_filter_spec *left,
+			   const struct efx_filter_spec *right)
+{
+	if ((left->match_flags ^ right->match_flags) |
+	    ((left->flags ^ right->flags) &
+	     (EFX_FILTER_FLAG_RX | EFX_FILTER_FLAG_TX)))
+		return false;
+
+	return memcmp(&left->outer_vid, &right->outer_vid,
+		      sizeof(struct efx_filter_spec) -
+		      offsetof(struct efx_filter_spec, outer_vid)) == 0;
+}
+
+u32 efx_filter_spec_hash(const struct efx_filter_spec *spec)
+{
+	BUILD_BUG_ON(offsetof(struct efx_filter_spec, outer_vid) & 3);
+	return jhash2((const u32 *)&spec->outer_vid,
+		      (sizeof(struct efx_filter_spec) -
+		       offsetof(struct efx_filter_spec, outer_vid)) / 4,
+		      0);
+}
+
+#ifdef CONFIG_RFS_ACCEL
+bool efx_rps_check_rule(struct efx_arfs_rule *rule, unsigned int filter_idx,
+			bool *force)
+{
+	if (rule->filter_id == EFX_ARFS_FILTER_ID_PENDING) {
+		/* ARFS is currently updating this entry, leave it */
+		return false;
+	}
+	if (rule->filter_id == EFX_ARFS_FILTER_ID_ERROR) {
+		/* ARFS tried and failed to update this, so it's probably out
+		 * of date.  Remove the filter and the ARFS rule entry.
+		 */
+		rule->filter_id = EFX_ARFS_FILTER_ID_REMOVING;
+		*force = true;
+		return true;
+	} else if (WARN_ON(rule->filter_id != filter_idx)) { /* can't happen */
+		/* ARFS has moved on, so old filter is not needed.  Since we did
+		 * not mark the rule with EFX_ARFS_FILTER_ID_REMOVING, it will
+		 * not be removed by efx_rps_hash_del() subsequently.
+		 */
+		*force = true;
+		return true;
+	}
+	/* Remove it iff ARFS wants to. */
+	return true;
+}
+
+struct hlist_head *efx_rps_hash_bucket(struct efx_nic *efx,
+				       const struct efx_filter_spec *spec)
+{
+	u32 hash = efx_filter_spec_hash(spec);
+
+	WARN_ON(!spin_is_locked(&efx->rps_hash_lock));
+	if (!efx->rps_hash_table)
+		return NULL;
+	return &efx->rps_hash_table[hash % EFX_ARFS_HASH_TABLE_SIZE];
+}
+
+struct efx_arfs_rule *efx_rps_hash_find(struct efx_nic *efx,
+					const struct efx_filter_spec *spec)
+{
+	struct efx_arfs_rule *rule;
+	struct hlist_head *head;
+	struct hlist_node *node;
+
+	head = efx_rps_hash_bucket(efx, spec);
+	if (!head)
+		return NULL;
+	hlist_for_each(node, head) {
+		rule = container_of(node, struct efx_arfs_rule, node);
+		if (efx_filter_spec_equal(spec, &rule->spec))
+			return rule;
+	}
+	return NULL;
+}
+
+struct efx_arfs_rule *efx_rps_hash_add(struct efx_nic *efx,
+				       const struct efx_filter_spec *spec,
+				       bool *new)
+{
+	struct efx_arfs_rule *rule;
+	struct hlist_head *head;
+	struct hlist_node *node;
+
+	head = efx_rps_hash_bucket(efx, spec);
+	if (!head)
+		return NULL;
+	hlist_for_each(node, head) {
+		rule = container_of(node, struct efx_arfs_rule, node);
+		if (efx_filter_spec_equal(spec, &rule->spec)) {
+			*new = false;
+			return rule;
+		}
+	}
+	rule = kmalloc(sizeof(*rule), GFP_ATOMIC);
+	*new = true;
+	if (rule) {
+		memcpy(&rule->spec, spec, sizeof(rule->spec));
+		hlist_add_head(&rule->node, head);
+	}
+	return rule;
+}
+
+void efx_rps_hash_del(struct efx_nic *efx, const struct efx_filter_spec *spec)
+{
+	struct efx_arfs_rule *rule;
+	struct hlist_head *head;
+	struct hlist_node *node;
+
+	head = efx_rps_hash_bucket(efx, spec);
+	if (WARN_ON(!head))
+		return;
+	hlist_for_each(node, head) {
+		rule = container_of(node, struct efx_arfs_rule, node);
+		if (efx_filter_spec_equal(spec, &rule->spec)) {
+			/* Someone already reused the entry.  We know that if
+			 * this check doesn't fire (i.e. filter_id == REMOVING)
+			 * then the REMOVING mark was put there by our caller,
+			 * because caller is holding a lock on filter table and
+			 * only holders of that lock set REMOVING.
+			 */
+			if (rule->filter_id != EFX_ARFS_FILTER_ID_REMOVING)
+				return;
+			hlist_del(node);
+			kfree(rule);
+			return;
+		}
+	}
+	/* We didn't find it. */
+	WARN_ON(1);
+}
+#endif
+
 /* RSS contexts.  We're using linked lists and crappy O(n) algorithms, because
  * (a) this is an infrequent control-plane operation and (b) n is small (max 64)
  */

drivers/net/ethernet/sfc/efx.h

@@ -186,6 +186,27 @@ static inline void efx_filter_rfs_expire(struct work_struct *data) {}
 #endif
 bool efx_filter_is_mc_recipient(const struct efx_filter_spec *spec);
 
+bool efx_filter_spec_equal(const struct efx_filter_spec *left,
+			   const struct efx_filter_spec *right);
+u32 efx_filter_spec_hash(const struct efx_filter_spec *spec);
+
+#ifdef CONFIG_RFS_ACCEL
+bool efx_rps_check_rule(struct efx_arfs_rule *rule, unsigned int filter_idx,
+			bool *force);
+
+struct efx_arfs_rule *efx_rps_hash_find(struct efx_nic *efx,
+					const struct efx_filter_spec *spec);
+
+/* @new is written to indicate if entry was newly added (true) or if an old
+ * entry was found and returned (false).
+ */
+struct efx_arfs_rule *efx_rps_hash_add(struct efx_nic *efx,
+				       const struct efx_filter_spec *spec,
+				       bool *new);
+
+void efx_rps_hash_del(struct efx_nic *efx, const struct efx_filter_spec *spec);
+#endif
+
 /* RSS contexts */
 struct efx_rss_context *efx_alloc_rss_context_entry(struct efx_nic *efx);
 struct efx_rss_context *efx_find_rss_context_entry(struct efx_nic *efx, u32 id);

drivers/net/ethernet/sfc/farch.c

@@ -2905,18 +2905,45 @@ bool efx_farch_filter_rfs_expire_one(struct efx_nic *efx, u32 flow_id,
 {
 	struct efx_farch_filter_state *state = efx->filter_state;
 	struct efx_farch_filter_table *table;
-	bool ret = false;
+	bool ret = false, force = false;
+	u16 arfs_id;
 
 	down_write(&state->lock);
+	spin_lock_bh(&efx->rps_hash_lock);
 	table = &state->table[EFX_FARCH_FILTER_TABLE_RX_IP];
 	if (test_bit(index, table->used_bitmap) &&
-	    table->spec[index].priority == EFX_FILTER_PRI_HINT &&
-	    rps_may_expire_flow(efx->net_dev, table->spec[index].dmaq_id,
-				flow_id, 0)) {
+	    table->spec[index].priority == EFX_FILTER_PRI_HINT) {
+		struct efx_arfs_rule *rule = NULL;
+		struct efx_filter_spec spec;
+
+		efx_farch_filter_to_gen_spec(&spec, &table->spec[index]);
+		if (!efx->rps_hash_table) {
+			/* In the absence of the table, we always returned 0 to
+			 * ARFS, so use the same to query it.
+			 */
+			arfs_id = 0;
+		} else {
+			rule = efx_rps_hash_find(efx, &spec);
+			if (!rule) {
+				/* ARFS table doesn't know of this filter, remove it */
+				force = true;
+			} else {
+				arfs_id = rule->arfs_id;
+				if (!efx_rps_check_rule(rule, index, &force))
+					goto out_unlock;
+			}
+		}
+		if (force || rps_may_expire_flow(efx->net_dev, spec.dmaq_id,
+						 flow_id, arfs_id)) {
+			if (rule)
+				rule->filter_id = EFX_ARFS_FILTER_ID_REMOVING;
+			efx_rps_hash_del(efx, &spec);
 			efx_farch_filter_table_clear_entry(efx, table, index);
 			ret = true;
 		}
-
+	}
+out_unlock:
+	spin_unlock_bh(&efx->rps_hash_lock);
 	up_write(&state->lock);
 	return ret;
 }

drivers/net/ethernet/sfc/net_driver.h

@@ -734,6 +734,35 @@ struct efx_rss_context {
 };
 
 #ifdef CONFIG_RFS_ACCEL
+/* Order of these is important, since filter_id >= %EFX_ARFS_FILTER_ID_PENDING
+ * is used to test if filter does or will exist.
+ */
+#define EFX_ARFS_FILTER_ID_PENDING	-1
+#define EFX_ARFS_FILTER_ID_ERROR	-2
+#define EFX_ARFS_FILTER_ID_REMOVING	-3
+/**
+ * struct efx_arfs_rule - record of an ARFS filter and its IDs
+ * @node: linkage into hash table
+ * @spec: details of the filter (used as key for hash table).  Use efx->type to
+ *	determine which member to use.
+ * @rxq_index: channel to which the filter will steer traffic.
+ * @arfs_id: filter ID which was returned to ARFS
+ * @filter_id: index in software filter table.  May be
+ *	%EFX_ARFS_FILTER_ID_PENDING if filter was not inserted yet,
+ *	%EFX_ARFS_FILTER_ID_ERROR if filter insertion failed, or
+ *	%EFX_ARFS_FILTER_ID_REMOVING if expiry is currently removing the filter.
+ */
+struct efx_arfs_rule {
+	struct hlist_node node;
+	struct efx_filter_spec spec;
+	u16 rxq_index;
+	u16 arfs_id;
+	s32 filter_id;
+};
+
+/* Size chosen so that the table is one page (4kB) */
+#define EFX_ARFS_HASH_TABLE_SIZE	512
+
 /**
  * struct efx_async_filter_insertion - Request to asynchronously insert a filter
  * @net_dev: Reference to the netdevice
@@ -873,6 +902,10 @@ struct efx_async_filter_insertion {
  *	@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
+ *	@rps_next_id).
+ * @rps_hash_table: Mapping between ARFS filters and their various IDs
+ * @rps_next_id: next arfs_id for an ARFS filter
  * @active_queues: Count of RX and TX queues that haven't been flushed and drained.
  * @rxq_flush_pending: Count of number of receive queues that need to be flushed.
  *	Decremented when the efx_flush_rx_queue() is called.
@@ -1029,6 +1062,9 @@ struct efx_nic {
 	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;
+	struct hlist_head *rps_hash_table;
+	u32 rps_next_id;
 #endif
 
 	atomic_t active_queues;

drivers/net/ethernet/sfc/rx.c

@@ -834,9 +834,29 @@ static void efx_filter_rfs_work(struct work_struct *data)
 	struct efx_nic *efx = netdev_priv(req->net_dev);
 	struct efx_channel *channel = efx_get_channel(efx, req->rxq_index);
 	int slot_idx = req - efx->rps_slot;
+	struct efx_arfs_rule *rule;
+	u16 arfs_id = 0;
 	int rc;
 
 	rc = efx->type->filter_insert(efx, &req->spec, true);
+	if (efx->rps_hash_table) {
+		spin_lock_bh(&efx->rps_hash_lock);
+		rule = efx_rps_hash_find(efx, &req->spec);
+		/* The rule might have already gone, if someone else's request
+		 * for the same spec was already worked and then expired before
+		 * we got around to our work.  In that case we have nothing
+		 * tying us to an arfs_id, meaning that as soon as the filter
+		 * is considered for expiry it will be removed.
+		 */
+		if (rule) {
+			if (rc < 0)
+				rule->filter_id = EFX_ARFS_FILTER_ID_ERROR;
+			else
+				rule->filter_id = rc;
+			arfs_id = rule->arfs_id;
+		}
+		spin_unlock_bh(&efx->rps_hash_lock);
+	}
 	if (rc >= 0) {
 		/* Remember this so we can check whether to expire the filter
 		 * later.
@@ -848,18 +868,18 @@ static void efx_filter_rfs_work(struct work_struct *data)
 
 		if (req->spec.ether_type == htons(ETH_P_IP))
 			netif_info(efx, rx_status, efx->net_dev,
-				   "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d]\n",
+				   "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %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);
+				   req->rxq_index, req->flow_id, rc, arfs_id);
 		else
 			netif_info(efx, rx_status, efx->net_dev,
-				   "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d]\n",
+				   "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %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);
+				   req->rxq_index, req->flow_id, rc, arfs_id);
 	}
 
 	/* Release references */
@@ -872,8 +892,10 @@ int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
 {
 	struct efx_nic *efx = netdev_priv(net_dev);
 	struct efx_async_filter_insertion *req;
+	struct efx_arfs_rule *rule;
 	struct flow_keys fk;
 	int slot_idx;
+	bool new;
 	int rc;
 
 	/* find a free slot */
@@ -926,12 +948,42 @@ int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
 	req->spec.rem_port = fk.ports.src;
 	req->spec.loc_port = fk.ports.dst;
 
+	if (efx->rps_hash_table) {
+		/* Add it to ARFS hash table */
+		spin_lock(&efx->rps_hash_lock);
+		rule = efx_rps_hash_add(efx, &req->spec, &new);
+		if (!rule) {
+			rc = -ENOMEM;
+			goto out_unlock;
+		}
+		if (new)
+			rule->arfs_id = efx->rps_next_id++ % RPS_NO_FILTER;
+		rc = rule->arfs_id;
+		/* Skip if existing or pending filter already does the right thing */
+		if (!new && rule->rxq_index == rxq_index &&
+		    rule->filter_id >= EFX_ARFS_FILTER_ID_PENDING)
+			goto out_unlock;
+		rule->rxq_index = rxq_index;
+		rule->filter_id = EFX_ARFS_FILTER_ID_PENDING;
+		spin_unlock(&efx->rps_hash_lock);
+	} else {
+		/* Without an ARFS hash table, we just use arfs_id 0 for all
+		 * filters.  This means if multiple flows hash to the same
+		 * flow_id, all but the most recently touched will be eligible
+		 * for expiry.
+		 */
+		rc = 0;
+	}
+
+	/* Queue the request */
 	dev_hold(req->net_dev = net_dev);
 	INIT_WORK(&req->work, efx_filter_rfs_work);
 	req->rxq_index = rxq_index;
 	req->flow_id = flow_id;
 	schedule_work(&req->work);
-	return 0;
+	return rc;
+out_unlock:
+	spin_unlock(&efx->rps_hash_lock);
 out_clear:
 	clear_bit(slot_idx, &efx->rps_slot_map);
 	return rc;