Merge branch 'bpf-nfp-map-op-cache'

Jakub Kicinski says:

====================
This set adds a small batching and cache mechanism to the driver.
Map dumps require two operations per element - get next, and
lookup. Each of those needs a round trip to the device, and on
a loaded system scheduling out and in of the dumping process.
This set makes the driver request a number of entries at the same
time, and if no operation which would modify the map happens
from the host side those entries are used to serve lookup
requests for up to 250us, at which point they are considered
stale.

This set has been measured to provide almost 4x dumping speed
improvement, Jaco says:

OLD dump times
    500 000 elements: 26.1s
  1 000 000 elements: 54.5s

NEW dump times
    500 000 elements: 7.6s
  1 000 000 elements: 16.5s
====================
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

drivers/net/ethernet/netronome/nfp/bpf/cmsg.c

@@ -6,6 +6,7 @@
 #include <linux/bug.h>
 #include <linux/jiffies.h>
 #include <linux/skbuff.h>
+#include <linux/timekeeping.h>
 
 #include "../ccm.h"
 #include "../nfp_app.h"
@@ -175,29 +176,151 @@ nfp_bpf_ctrl_reply_val(struct nfp_app_bpf *bpf, struct cmsg_reply_map_op *reply,
 	return &reply->data[bpf->cmsg_key_sz * (n + 1) + bpf->cmsg_val_sz * n];
 }
 
+static bool nfp_bpf_ctrl_op_cache_invalidate(enum nfp_ccm_type op)
+{
+	return op == NFP_CCM_TYPE_BPF_MAP_UPDATE ||
+	       op == NFP_CCM_TYPE_BPF_MAP_DELETE;
+}
+
+static bool nfp_bpf_ctrl_op_cache_capable(enum nfp_ccm_type op)
+{
+	return op == NFP_CCM_TYPE_BPF_MAP_LOOKUP ||
+	       op == NFP_CCM_TYPE_BPF_MAP_GETNEXT;
+}
+
+static bool nfp_bpf_ctrl_op_cache_fill(enum nfp_ccm_type op)
+{
+	return op == NFP_CCM_TYPE_BPF_MAP_GETFIRST ||
+	       op == NFP_CCM_TYPE_BPF_MAP_GETNEXT;
+}
+
+static unsigned int
+nfp_bpf_ctrl_op_cache_get(struct nfp_bpf_map *nfp_map, enum nfp_ccm_type op,
+			  const u8 *key, u8 *out_key, u8 *out_value,
+			  u32 *cache_gen)
+{
+	struct bpf_map *map = &nfp_map->offmap->map;
+	struct nfp_app_bpf *bpf = nfp_map->bpf;
+	unsigned int i, count, n_entries;
+	struct cmsg_reply_map_op *reply;
+
+	n_entries = nfp_bpf_ctrl_op_cache_fill(op) ? bpf->cmsg_cache_cnt : 1;
+
+	spin_lock(&nfp_map->cache_lock);
+	*cache_gen = nfp_map->cache_gen;
+	if (nfp_map->cache_blockers)
+		n_entries = 1;
+
+	if (nfp_bpf_ctrl_op_cache_invalidate(op))
+		goto exit_block;
+	if (!nfp_bpf_ctrl_op_cache_capable(op))
+		goto exit_unlock;
+
+	if (!nfp_map->cache)
+		goto exit_unlock;
+	if (nfp_map->cache_to < ktime_get_ns())
+		goto exit_invalidate;
+
+	reply = (void *)nfp_map->cache->data;
+	count = be32_to_cpu(reply->count);
+
+	for (i = 0; i < count; i++) {
+		void *cached_key;
+
+		cached_key = nfp_bpf_ctrl_reply_key(bpf, reply, i);
+		if (memcmp(cached_key, key, map->key_size))
+			continue;
+
+		if (op == NFP_CCM_TYPE_BPF_MAP_LOOKUP)
+			memcpy(out_value, nfp_bpf_ctrl_reply_val(bpf, reply, i),
+			       map->value_size);
+		if (op == NFP_CCM_TYPE_BPF_MAP_GETNEXT) {
+			if (i + 1 == count)
+				break;
+
+			memcpy(out_key,
+			       nfp_bpf_ctrl_reply_key(bpf, reply, i + 1),
+			       map->key_size);
+		}
+
+		n_entries = 0;
+		goto exit_unlock;
+	}
+	goto exit_unlock;
+
+exit_block:
+	nfp_map->cache_blockers++;
+exit_invalidate:
+	dev_consume_skb_any(nfp_map->cache);
+	nfp_map->cache = NULL;
+exit_unlock:
+	spin_unlock(&nfp_map->cache_lock);
+	return n_entries;
+}
+
+static void
+nfp_bpf_ctrl_op_cache_put(struct nfp_bpf_map *nfp_map, enum nfp_ccm_type op,
+			  struct sk_buff *skb, u32 cache_gen)
+{
+	bool blocker, filler;
+
+	blocker = nfp_bpf_ctrl_op_cache_invalidate(op);
+	filler = nfp_bpf_ctrl_op_cache_fill(op);
+	if (blocker || filler) {
+		u64 to = 0;
+
+		if (filler)
+			to = ktime_get_ns() + NFP_BPF_MAP_CACHE_TIME_NS;
+
+		spin_lock(&nfp_map->cache_lock);
+		if (blocker) {
+			nfp_map->cache_blockers--;
+			nfp_map->cache_gen++;
+		}
+		if (filler && !nfp_map->cache_blockers &&
+		    nfp_map->cache_gen == cache_gen) {
+			nfp_map->cache_to = to;
+			swap(nfp_map->cache, skb);
+		}
+		spin_unlock(&nfp_map->cache_lock);
+	}
+
+	dev_consume_skb_any(skb);
+}
+
 static int
 nfp_bpf_ctrl_entry_op(struct bpf_offloaded_map *offmap, enum nfp_ccm_type op,
 		      u8 *key, u8 *value, u64 flags, u8 *out_key, u8 *out_value)
 {
 	struct nfp_bpf_map *nfp_map = offmap->dev_priv;
+	unsigned int n_entries, reply_entries, count;
 	struct nfp_app_bpf *bpf = nfp_map->bpf;
 	struct bpf_map *map = &offmap->map;
 	struct cmsg_reply_map_op *reply;
 	struct cmsg_req_map_op *req;
 	struct sk_buff *skb;
+	u32 cache_gen;
 	int err;
 
 	/* FW messages have no space for more than 32 bits of flags */
 	if (flags >> 32)
 		return -EOPNOTSUPP;
 
+	/* Handle op cache */
+	n_entries = nfp_bpf_ctrl_op_cache_get(nfp_map, op, key, out_key,
+					      out_value, &cache_gen);
+	if (!n_entries)
+		return 0;
+
 	skb = nfp_bpf_cmsg_map_req_alloc(bpf, 1);
-	if (!skb)
-		return -ENOMEM;
+	if (!skb) {
+		err = -ENOMEM;
+		goto err_cache_put;
+	}
 
 	req = (void *)skb->data;
 	req->tid = cpu_to_be32(nfp_map->tid);
-	req->count = cpu_to_be32(1);
+	req->count = cpu_to_be32(n_entries);
 	req->flags = cpu_to_be32(flags);
 
 	/* Copy inputs */
@@ -207,16 +330,38 @@ nfp_bpf_ctrl_entry_op(struct bpf_offloaded_map *offmap, enum nfp_ccm_type op,
 		memcpy(nfp_bpf_ctrl_req_val(bpf, req, 0), value,
 		       map->value_size);
 
-	skb = nfp_ccm_communicate(&bpf->ccm, skb, op,
-				  nfp_bpf_cmsg_map_reply_size(bpf, 1));
-	if (IS_ERR(skb))
-		return PTR_ERR(skb);
+	skb = nfp_ccm_communicate(&bpf->ccm, skb, op, 0);
+	if (IS_ERR(skb)) {
+		err = PTR_ERR(skb);
+		goto err_cache_put;
+	}
+
+	if (skb->len < sizeof(*reply)) {
+		cmsg_warn(bpf, "cmsg drop - type 0x%02x too short %d!\n",
+			  op, skb->len);
+		err = -EIO;
+		goto err_free;
+	}
 
 	reply = (void *)skb->data;
+	count = be32_to_cpu(reply->count);
 	err = nfp_bpf_ctrl_rc_to_errno(bpf, &reply->reply_hdr);
+	/* FW responds with message sized to hold the good entries,
+	 * plus one extra entry if there was an error.
+	 */
+	reply_entries = count + !!err;
+	if (n_entries > 1 && count)
+		err = 0;
 	if (err)
 		goto err_free;
 
+	if (skb->len != nfp_bpf_cmsg_map_reply_size(bpf, reply_entries)) {
+		cmsg_warn(bpf, "cmsg drop - type 0x%02x too short %d for %d entries!\n",
+			  op, skb->len, reply_entries);
+		err = -EIO;
+		goto err_free;
+	}
+
 	/* Copy outputs */
 	if (out_key)
 		memcpy(out_key, nfp_bpf_ctrl_reply_key(bpf, reply, 0),
@@ -225,11 +370,13 @@ nfp_bpf_ctrl_entry_op(struct bpf_offloaded_map *offmap, enum nfp_ccm_type op,
 		memcpy(out_value, nfp_bpf_ctrl_reply_val(bpf, reply, 0),
 		       map->value_size);
 
-	dev_consume_skb_any(skb);
+	nfp_bpf_ctrl_op_cache_put(nfp_map, op, skb, cache_gen);
 
 	return 0;
 err_free:
 	dev_kfree_skb_any(skb);
+err_cache_put:
+	nfp_bpf_ctrl_op_cache_put(nfp_map, op, NULL, cache_gen);
 	return err;
 }
 
@@ -267,13 +414,31 @@ int nfp_bpf_ctrl_getnext_entry(struct bpf_offloaded_map *offmap,
 				     key, NULL, 0, next_key, NULL);
 }
 
-unsigned int nfp_bpf_ctrl_cmsg_mtu(struct nfp_app_bpf *bpf)
+unsigned int nfp_bpf_ctrl_cmsg_min_mtu(struct nfp_app_bpf *bpf)
 {
-	return max3((unsigned int)NFP_NET_DEFAULT_MTU,
-		    nfp_bpf_cmsg_map_req_size(bpf, 1),
+	return max(nfp_bpf_cmsg_map_req_size(bpf, 1),
 		   nfp_bpf_cmsg_map_reply_size(bpf, 1));
 }
 
+unsigned int nfp_bpf_ctrl_cmsg_mtu(struct nfp_app_bpf *bpf)
+{
+	return max3(NFP_NET_DEFAULT_MTU,
+		    nfp_bpf_cmsg_map_req_size(bpf, NFP_BPF_MAP_CACHE_CNT),
+		    nfp_bpf_cmsg_map_reply_size(bpf, NFP_BPF_MAP_CACHE_CNT));
+}
+
+unsigned int nfp_bpf_ctrl_cmsg_cache_cnt(struct nfp_app_bpf *bpf)
+{
+	unsigned int mtu, req_max, reply_max, entry_sz;
+
+	mtu = bpf->app->ctrl->dp.mtu;
+	entry_sz = bpf->cmsg_key_sz + bpf->cmsg_val_sz;
+	req_max = (mtu - sizeof(struct cmsg_req_map_op)) / entry_sz;
+	reply_max = (mtu - sizeof(struct cmsg_reply_map_op)) / entry_sz;
+
+	return min3(req_max, reply_max, NFP_BPF_MAP_CACHE_CNT);
+}
+
 void nfp_bpf_ctrl_msg_rx(struct nfp_app *app, struct sk_buff *skb)
 {
 	struct nfp_app_bpf *bpf = app->priv;

drivers/net/ethernet/netronome/nfp/bpf/fw.h

@@ -24,6 +24,7 @@ enum bpf_cap_tlv_type {
 	NFP_BPF_CAP_TYPE_QUEUE_SELECT	= 5,
 	NFP_BPF_CAP_TYPE_ADJUST_TAIL	= 6,
 	NFP_BPF_CAP_TYPE_ABI_VERSION	= 7,
+	NFP_BPF_CAP_TYPE_CMSG_MULTI_ENT	= 8,
 };
 
 struct nfp_bpf_cap_tlv_func {

drivers/net/ethernet/netronome/nfp/bpf/main.c

@@ -299,6 +299,14 @@ nfp_bpf_parse_cap_adjust_tail(struct nfp_app_bpf *bpf, void __iomem *value,
 	return 0;
 }
 
+static int
+nfp_bpf_parse_cap_cmsg_multi_ent(struct nfp_app_bpf *bpf, void __iomem *value,
+				 u32 length)
+{
+	bpf->cmsg_multi_ent = true;
+	return 0;
+}
+
 static int
 nfp_bpf_parse_cap_abi_version(struct nfp_app_bpf *bpf, void __iomem *value,
 			      u32 length)
@@ -375,6 +383,11 @@ static int nfp_bpf_parse_capabilities(struct nfp_app *app)
 							  length))
 				goto err_release_free;
 			break;
+		case NFP_BPF_CAP_TYPE_CMSG_MULTI_ENT:
+			if (nfp_bpf_parse_cap_cmsg_multi_ent(app->priv, value,
+							     length))
+				goto err_release_free;
+			break;
 		default:
 			nfp_dbg(cpp, "unknown BPF capability: %d\n", type);
 			break;
@@ -415,6 +428,25 @@ static void nfp_bpf_ndo_uninit(struct nfp_app *app, struct net_device *netdev)
 	bpf_offload_dev_netdev_unregister(bpf->bpf_dev, netdev);
 }
 
+static int nfp_bpf_start(struct nfp_app *app)
+{
+	struct nfp_app_bpf *bpf = app->priv;
+
+	if (app->ctrl->dp.mtu < nfp_bpf_ctrl_cmsg_min_mtu(bpf)) {
+		nfp_err(bpf->app->cpp,
+			"ctrl channel MTU below min required %u < %u\n",
+			app->ctrl->dp.mtu, nfp_bpf_ctrl_cmsg_min_mtu(bpf));
+		return -EINVAL;
+	}
+
+	if (bpf->cmsg_multi_ent)
+		bpf->cmsg_cache_cnt = nfp_bpf_ctrl_cmsg_cache_cnt(bpf);
+	else
+		bpf->cmsg_cache_cnt = 1;
+
+	return 0;
+}
+
 static int nfp_bpf_init(struct nfp_app *app)
 {
 	struct nfp_app_bpf *bpf;
@@ -488,6 +520,7 @@ const struct nfp_app_type app_bpf = {
 
 	.init		= nfp_bpf_init,
 	.clean		= nfp_bpf_clean,
+	.start		= nfp_bpf_start,
 
 	.check_mtu	= nfp_bpf_check_mtu,
 

drivers/net/ethernet/netronome/nfp/bpf/main.h

@@ -99,6 +99,7 @@ enum pkt_vec {
  * @maps_neutral:	hash table of offload-neutral maps (on pointer)
  *
  * @abi_version:	global BPF ABI version
+ * @cmsg_cache_cnt:	number of entries to read for caching
  *
  * @adjust_head:	adjust head capability
  * @adjust_head.flags:		extra flags for adjust head
@@ -124,6 +125,7 @@ enum pkt_vec {
  * @pseudo_random:	FW initialized the pseudo-random machinery (CSRs)
  * @queue_select:	BPF can set the RX queue ID in packet vector
  * @adjust_tail:	BPF can simply trunc packet size for adjust tail
+ * @cmsg_multi_ent:	FW can pack multiple map entries in a single cmsg
  */
 struct nfp_app_bpf {
 	struct nfp_app *app;
@@ -134,6 +136,8 @@ struct nfp_app_bpf {
 	unsigned int cmsg_key_sz;
 	unsigned int cmsg_val_sz;
 
+	unsigned int cmsg_cache_cnt;
+
 	struct list_head map_list;
 	unsigned int maps_in_use;
 	unsigned int map_elems_in_use;
@@ -169,6 +173,7 @@ struct nfp_app_bpf {
 	bool pseudo_random;
 	bool queue_select;
 	bool adjust_tail;
+	bool cmsg_multi_ent;
 };
 
 enum nfp_bpf_map_use {
@@ -183,11 +188,21 @@ struct nfp_bpf_map_word {
 	unsigned char non_zero_update	:1;
 };
 
+#define NFP_BPF_MAP_CACHE_CNT		4U
+#define NFP_BPF_MAP_CACHE_TIME_NS	(250 * 1000)
+
 /**
  * struct nfp_bpf_map - private per-map data attached to BPF maps for offload
  * @offmap:	pointer to the offloaded BPF map
  * @bpf:	back pointer to bpf app private structure
  * @tid:	table id identifying map on datapath
+ *
+ * @cache_lock:	protects @cache_blockers, @cache_to, @cache
+ * @cache_blockers:	number of ops in flight which block caching
+ * @cache_gen:	counter incremented by every blocker on exit
+ * @cache_to:	time when cache will no longer be valid (ns)
+ * @cache:	skb with cached response
+ *
  * @l:		link on the nfp_app_bpf->map_list list
  * @use_map:	map of how the value is used (in 4B chunks)
  */
@@ -195,6 +210,13 @@ struct nfp_bpf_map {
 	struct bpf_offloaded_map *offmap;
 	struct nfp_app_bpf *bpf;
 	u32 tid;
+
+	spinlock_t cache_lock;
+	u32 cache_blockers;
+	u32 cache_gen;
+	u64 cache_to;
+	struct sk_buff *cache;
+
 	struct list_head l;
 	struct nfp_bpf_map_word use_map[];
 };
@@ -564,7 +586,9 @@ nfp_bpf_goto_meta(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
 
 void *nfp_bpf_relo_for_vnic(struct nfp_prog *nfp_prog, struct nfp_bpf_vnic *bv);
 
+unsigned int nfp_bpf_ctrl_cmsg_min_mtu(struct nfp_app_bpf *bpf);
 unsigned int nfp_bpf_ctrl_cmsg_mtu(struct nfp_app_bpf *bpf);
+unsigned int nfp_bpf_ctrl_cmsg_cache_cnt(struct nfp_app_bpf *bpf);
 long long int
 nfp_bpf_ctrl_alloc_map(struct nfp_app_bpf *bpf, struct bpf_map *map);
 void

drivers/net/ethernet/netronome/nfp/bpf/offload.c

@@ -385,6 +385,7 @@ nfp_bpf_map_alloc(struct nfp_app_bpf *bpf, struct bpf_offloaded_map *offmap)
 	offmap->dev_priv = nfp_map;
 	nfp_map->offmap = offmap;
 	nfp_map->bpf = bpf;
+	spin_lock_init(&nfp_map->cache_lock);
 
 	res = nfp_bpf_ctrl_alloc_map(bpf, &offmap->map);
 	if (res < 0) {
@@ -407,6 +408,8 @@ nfp_bpf_map_free(struct nfp_app_bpf *bpf, struct bpf_offloaded_map *offmap)
 	struct nfp_bpf_map *nfp_map = offmap->dev_priv;
 
 	nfp_bpf_ctrl_free_map(bpf, nfp_map);
+	dev_consume_skb_any(nfp_map->cache);
+	WARN_ON_ONCE(nfp_map->cache_blockers);
 	list_del_init(&nfp_map->l);
 	bpf->map_elems_in_use -= offmap->map.max_entries;
 	bpf->maps_in_use--;

drivers/net/ethernet/netronome/nfp/nfp_net.h

@@ -66,7 +66,7 @@
 #define NFP_NET_MAX_DMA_BITS	40
 
 /* Default size for MTU and freelist buffer sizes */
-#define NFP_NET_DEFAULT_MTU		1500
+#define NFP_NET_DEFAULT_MTU		1500U
 
 /* Maximum number of bytes prepended to a packet */
 #define NFP_NET_MAX_PREPEND		64

drivers/net/ethernet/netronome/nfp/nfp_net_common.c

@@ -4116,14 +4116,7 @@ int nfp_net_init(struct nfp_net *nn)
 
 	/* Set default MTU and Freelist buffer size */
 	if (!nfp_net_is_data_vnic(nn) && nn->app->ctrl_mtu) {
-		if (nn->app->ctrl_mtu <= nn->max_mtu) {
-			nn->dp.mtu = nn->app->ctrl_mtu;
-		} else {
-			if (nn->app->ctrl_mtu != NFP_APP_CTRL_MTU_MAX)
-				nn_warn(nn, "app requested MTU above max supported %u > %u\n",
-					nn->app->ctrl_mtu, nn->max_mtu);
-			nn->dp.mtu = nn->max_mtu;
-		}
+		nn->dp.mtu = min(nn->app->ctrl_mtu, nn->max_mtu);
 	} else if (nn->max_mtu < NFP_NET_DEFAULT_MTU) {
 		nn->dp.mtu = nn->max_mtu;
 	} else {