[PATCH bpf-next 0/6] nfp: bpf: improve efficiency of offloaded perf events

[PATCH bpf-next 0/6] nfp: bpf: improve efficiency of offloaded perf events

[Jakub Kicinski]

Hi!

This set is focused on improving the performance of perf events
reported from BPF offload.  Perf events can now be received on
packet data queues, which significantly improves the performance
(from total of 0.5 Msps to 5Msps per core).  To get to this
performance we need a fast path for control messages which will
operate on raw buffers and recycle them immediately.

Patch 5 replaces the map pointers for perf maps with map IDs.
We look the pointers up in a hashtable, anyway, to validate they
are correct, so there is no performance difference.  Map IDs
have the advantage of being easier to understand for users in
case of errors (we no longer print raw pointers to the logs).

Last patch improves info messages about map offload.

Jakub Kicinski (6):
  nfp: move repr handling on RX path
  nfp: allow control message reception on data queues
  nfp: bpf: pass raw data buffer to nfp_bpf_event_output()
  nfp: bpf: allow receiving perf events on data queues
  nfp: bpf: remember maps by ID
  nfp: bpf: improve map offload info messages

 drivers/net/ethernet/netronome/nfp/bpf/cmsg.c | 25 +++++++-
 drivers/net/ethernet/netronome/nfp/bpf/jit.c  | 12 ++--
 drivers/net/ethernet/netronome/nfp/bpf/main.c |  5 +-
 drivers/net/ethernet/netronome/nfp/bpf/main.h |  9 ++-
 .../net/ethernet/netronome/nfp/bpf/offload.c  | 63 +++++++++++--------
 drivers/net/ethernet/netronome/nfp/nfp_app.c  |  2 +
 drivers/net/ethernet/netronome/nfp/nfp_app.h  | 17 +++++
 .../ethernet/netronome/nfp/nfp_net_common.c   | 40 +++++++-----
 .../net/ethernet/netronome/nfp/nfp_net_ctrl.h |  1 +
 9 files changed, 125 insertions(+), 49 deletions(-)

-- 
2.17.1

[PATCH bpf-next 1/6] nfp: move repr handling on RX path

[Jakub Kicinski]

Representor packets are received on PF queues with special metadata tag
for demux.  There is no reason to resolve the representor ID -> netdev
after the skb has been allocated.  Move the code, this will allow us to
handle special FW messages without SKB allocation overhead.

Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
---
 .../ethernet/netronome/nfp/nfp_net_common.c   | 29 ++++++++++---------
 1 file changed, 15 insertions(+), 14 deletions(-)

diff --git a/drivers/net/ethernet/netronome/nfp/nfp_net_common.c b/drivers/net/ethernet/netronome/nfp/nfp_net_common.c
index cf1704e972b7..cdfbd1e8bf4b 100644
--- a/drivers/net/ethernet/netronome/nfp/nfp_net_common.c
+++ b/drivers/net/ethernet/netronome/nfp/nfp_net_common.c
@@ -1757,6 +1757,21 @@ static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget)
 			}
 		}
 
+		if (likely(!meta.portid)) {
+			netdev = dp->netdev;
+		} else {
+			struct nfp_net *nn;
+
+			nn = netdev_priv(dp->netdev);
+			netdev = nfp_app_repr_get(nn->app, meta.portid);
+			if (unlikely(!netdev)) {
+				nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf,
+						NULL);
+				continue;
+			}
+			nfp_repr_inc_rx_stats(netdev, pkt_len);
+		}
+
 		skb = build_skb(rxbuf->frag, true_bufsz);
 		if (unlikely(!skb)) {
 			nfp_net_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
@@ -1772,20 +1787,6 @@ static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget)
 
 		nfp_net_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
 
-		if (likely(!meta.portid)) {
-			netdev = dp->netdev;
-		} else {
-			struct nfp_net *nn;
-
-			nn = netdev_priv(dp->netdev);
-			netdev = nfp_app_repr_get(nn->app, meta.portid);
-			if (unlikely(!netdev)) {
-				nfp_net_rx_drop(dp, r_vec, rx_ring, NULL, skb);
-				continue;
-			}
-			nfp_repr_inc_rx_stats(netdev, pkt_len);
-		}
-
 		skb_reserve(skb, pkt_off);
 		skb_put(skb, pkt_len);
 
-- 
2.17.1

[PATCH bpf-next 2/6] nfp: allow control message reception on data queues

[Jakub Kicinski]

Port id 0xffffffff is reserved for control messages.  Allow reception
of messages with this id on data queues.  Hand off a raw buffer to
the higher layer code, without allocating SKB for max efficiency.
The RX handle can't modify or keep the buffer, after it returns
buffer is handed back over to the NIC RX free buffer list.

Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
---
 drivers/net/ethernet/netronome/nfp/nfp_app.c    |  2 ++
 drivers/net/ethernet/netronome/nfp/nfp_app.h    | 17 +++++++++++++++++
 .../net/ethernet/netronome/nfp/nfp_net_common.c | 11 +++++++++++
 .../net/ethernet/netronome/nfp/nfp_net_ctrl.h   |  1 +
 4 files changed, 31 insertions(+)

diff --git a/drivers/net/ethernet/netronome/nfp/nfp_app.c b/drivers/net/ethernet/netronome/nfp/nfp_app.c
index 69d4ae7a61f3..8607d09ab732 100644
--- a/drivers/net/ethernet/netronome/nfp/nfp_app.c
+++ b/drivers/net/ethernet/netronome/nfp/nfp_app.c
@@ -172,6 +172,8 @@ struct nfp_app *nfp_app_alloc(struct nfp_pf *pf, enum nfp_app_id id)
 
 	if (WARN_ON(!apps[id]->name || !apps[id]->vnic_alloc))
 		return ERR_PTR(-EINVAL);
+	if (WARN_ON(!apps[id]->ctrl_msg_rx && apps[id]->ctrl_msg_rx_raw))
+		return ERR_PTR(-EINVAL);
 
 	app = kzalloc(sizeof(*app), GFP_KERNEL);
 	if (!app)
diff --git a/drivers/net/ethernet/netronome/nfp/nfp_app.h b/drivers/net/ethernet/netronome/nfp/nfp_app.h
index afbc19aa66a8..ccb244cf6c30 100644
--- a/drivers/net/ethernet/netronome/nfp/nfp_app.h
+++ b/drivers/net/ethernet/netronome/nfp/nfp_app.h
@@ -98,6 +98,7 @@ extern const struct nfp_app_type app_abm;
  * @start:	start application logic
  * @stop:	stop application logic
  * @ctrl_msg_rx:    control message handler
+ * @ctrl_msg_rx_raw:	handler for control messages from data queues
  * @setup_tc:	setup TC ndo
  * @bpf:	BPF ndo offload-related calls
  * @xdp_offload:    offload an XDP program
@@ -150,6 +151,8 @@ struct nfp_app_type {
 	void (*stop)(struct nfp_app *app);
 
 	void (*ctrl_msg_rx)(struct nfp_app *app, struct sk_buff *skb);
+	void (*ctrl_msg_rx_raw)(struct nfp_app *app, const void *data,
+				unsigned int len);
 
 	int (*setup_tc)(struct nfp_app *app, struct net_device *netdev,
 			enum tc_setup_type type, void *type_data);
@@ -318,6 +321,11 @@ static inline bool nfp_app_ctrl_has_meta(struct nfp_app *app)
 	return app->type->ctrl_has_meta;
 }
 
+static inline bool nfp_app_ctrl_uses_data_vnics(struct nfp_app *app)
+{
+	return app && app->type->ctrl_msg_rx_raw;
+}
+
 static inline const char *nfp_app_extra_cap(struct nfp_app *app,
 					    struct nfp_net *nn)
 {
@@ -381,6 +389,15 @@ static inline void nfp_app_ctrl_rx(struct nfp_app *app, struct sk_buff *skb)
 	app->type->ctrl_msg_rx(app, skb);
 }
 
+static inline void
+nfp_app_ctrl_rx_raw(struct nfp_app *app, const void *data, unsigned int len)
+{
+	trace_devlink_hwmsg(priv_to_devlink(app->pf), true, 0, data, len);
+
+	if (app && app->type->ctrl_msg_rx_raw)
+		app->type->ctrl_msg_rx_raw(app, data, len);
+}
+
 static inline int nfp_app_eswitch_mode_get(struct nfp_app *app, u16 *mode)
 {
 	if (!app->type->eswitch_mode_get)
diff --git a/drivers/net/ethernet/netronome/nfp/nfp_net_common.c b/drivers/net/ethernet/netronome/nfp/nfp_net_common.c
index cdfbd1e8bf4b..ca42f7da4f50 100644
--- a/drivers/net/ethernet/netronome/nfp/nfp_net_common.c
+++ b/drivers/net/ethernet/netronome/nfp/nfp_net_common.c
@@ -1759,6 +1759,14 @@ static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget)
 
 		if (likely(!meta.portid)) {
 			netdev = dp->netdev;
+		} else if (meta.portid == NFP_META_PORT_ID_CTRL) {
+			struct nfp_net *nn = netdev_priv(dp->netdev);
+
+			nfp_app_ctrl_rx_raw(nn->app, rxbuf->frag + pkt_off,
+					    pkt_len);
+			nfp_net_rx_give_one(dp, rx_ring, rxbuf->frag,
+					    rxbuf->dma_addr);
+			continue;
 		} else {
 			struct nfp_net *nn;
 
@@ -3857,6 +3865,9 @@ int nfp_net_init(struct nfp_net *nn)
 		nn->dp.mtu = NFP_NET_DEFAULT_MTU;
 	nn->dp.fl_bufsz = nfp_net_calc_fl_bufsz(&nn->dp);
 
+	if (nfp_app_ctrl_uses_data_vnics(nn->app))
+		nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_CMSG_DATA;
+
 	if (nn->cap & NFP_NET_CFG_CTRL_RSS_ANY) {
 		nfp_net_rss_init(nn);
 		nn->dp.ctrl |= nn->cap & NFP_NET_CFG_CTRL_RSS2 ?:
diff --git a/drivers/net/ethernet/netronome/nfp/nfp_net_ctrl.h b/drivers/net/ethernet/netronome/nfp/nfp_net_ctrl.h
index bb63c115537d..44d3ea75d043 100644
--- a/drivers/net/ethernet/netronome/nfp/nfp_net_ctrl.h
+++ b/drivers/net/ethernet/netronome/nfp/nfp_net_ctrl.h
@@ -127,6 +127,7 @@
 #define   NFP_NET_CFG_CTRL_GATHER	  (0x1 <<  9) /* Gather DMA */
 #define   NFP_NET_CFG_CTRL_LSO		  (0x1 << 10) /* LSO/TSO (version 1) */
 #define   NFP_NET_CFG_CTRL_CTAG_FILTER	  (0x1 << 11) /* VLAN CTAG filtering */
+#define   NFP_NET_CFG_CTRL_CMSG_DATA	  (0x1 << 12) /* RX cmsgs on data Qs */
 #define   NFP_NET_CFG_CTRL_RINGCFG	  (0x1 << 16) /* Ring runtime changes */
 #define   NFP_NET_CFG_CTRL_RSS		  (0x1 << 17) /* RSS (version 1) */
 #define   NFP_NET_CFG_CTRL_IRQMOD	  (0x1 << 18) /* Interrupt moderation */
-- 
2.17.1

[PATCH bpf-next 3/6] nfp: bpf: pass raw data buffer to nfp_bpf_event_output()

[Jakub Kicinski]

In preparation for SKB-less perf event handling make
nfp_bpf_event_output() take buffer address and length,
not SKB as parameters.

Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
---
 drivers/net/ethernet/netronome/nfp/bpf/cmsg.c |  5 ++++-
 drivers/net/ethernet/netronome/nfp/bpf/main.h |  3 ++-
 .../net/ethernet/netronome/nfp/bpf/offload.c  | 21 ++++++++-----------
 3 files changed, 15 insertions(+), 14 deletions(-)

diff --git a/drivers/net/ethernet/netronome/nfp/bpf/cmsg.c b/drivers/net/ethernet/netronome/nfp/bpf/cmsg.c
index cb87fccb9f6a..0a89b53962aa 100644
--- a/drivers/net/ethernet/netronome/nfp/bpf/cmsg.c
+++ b/drivers/net/ethernet/netronome/nfp/bpf/cmsg.c
@@ -441,7 +441,10 @@ void nfp_bpf_ctrl_msg_rx(struct nfp_app *app, struct sk_buff *skb)
 	}
 
 	if (nfp_bpf_cmsg_get_type(skb) == CMSG_TYPE_BPF_EVENT) {
-		nfp_bpf_event_output(bpf, skb);
+		if (!nfp_bpf_event_output(bpf, skb->data, skb->len))
+			dev_consume_skb_any(skb);
+		else
+			dev_kfree_skb_any(skb);
 		return;
 	}
 
diff --git a/drivers/net/ethernet/netronome/nfp/bpf/main.h b/drivers/net/ethernet/netronome/nfp/bpf/main.h
index bec935468f90..e25d3c0c7e43 100644
--- a/drivers/net/ethernet/netronome/nfp/bpf/main.h
+++ b/drivers/net/ethernet/netronome/nfp/bpf/main.h
@@ -501,7 +501,8 @@ int nfp_bpf_ctrl_lookup_entry(struct bpf_offloaded_map *offmap,
 int nfp_bpf_ctrl_getnext_entry(struct bpf_offloaded_map *offmap,
 			       void *key, void *next_key);
 
-int nfp_bpf_event_output(struct nfp_app_bpf *bpf, struct sk_buff *skb);
+int nfp_bpf_event_output(struct nfp_app_bpf *bpf, const void *data,
+			 unsigned int len);
 
 void nfp_bpf_ctrl_msg_rx(struct nfp_app *app, struct sk_buff *skb);
 #endif
diff --git a/drivers/net/ethernet/netronome/nfp/bpf/offload.c b/drivers/net/ethernet/netronome/nfp/bpf/offload.c
index 49b03f7dbf46..293dda84818f 100644
--- a/drivers/net/ethernet/netronome/nfp/bpf/offload.c
+++ b/drivers/net/ethernet/netronome/nfp/bpf/offload.c
@@ -453,23 +453,24 @@ nfp_bpf_perf_event_copy(void *dst, const void *src,
 	return 0;
 }
 
-int nfp_bpf_event_output(struct nfp_app_bpf *bpf, struct sk_buff *skb)
+int nfp_bpf_event_output(struct nfp_app_bpf *bpf, const void *data,
+			 unsigned int len)
 {
-	struct cmsg_bpf_event *cbe = (void *)skb->data;
+	struct cmsg_bpf_event *cbe = (void *)data;
 	u32 pkt_size, data_size;
 	struct bpf_map *map;
 
-	if (skb->len < sizeof(struct cmsg_bpf_event))
-		goto err_drop;
+	if (len < sizeof(struct cmsg_bpf_event))
+		return -EINVAL;
 
 	pkt_size = be32_to_cpu(cbe->pkt_size);
 	data_size = be32_to_cpu(cbe->data_size);
 	map = (void *)(unsigned long)be64_to_cpu(cbe->map_ptr);
 
-	if (skb->len < sizeof(struct cmsg_bpf_event) + pkt_size + data_size)
-		goto err_drop;
+	if (len < sizeof(struct cmsg_bpf_event) + pkt_size + data_size)
+		return -EINVAL;
 	if (cbe->hdr.ver != CMSG_MAP_ABI_VERSION)
-		goto err_drop;
+		return -EINVAL;
 
 	rcu_read_lock();
 	if (!rhashtable_lookup_fast(&bpf->maps_neutral, &map,
@@ -477,7 +478,7 @@ int nfp_bpf_event_output(struct nfp_app_bpf *bpf, struct sk_buff *skb)
 		rcu_read_unlock();
 		pr_warn("perf event: dest map pointer %px not recognized, dropping event\n",
 			map);
-		goto err_drop;
+		return -EINVAL;
 	}
 
 	bpf_event_output(map, be32_to_cpu(cbe->cpu_id),
@@ -485,11 +486,7 @@ int nfp_bpf_event_output(struct nfp_app_bpf *bpf, struct sk_buff *skb)
 			 cbe->data, pkt_size, nfp_bpf_perf_event_copy);
 	rcu_read_unlock();
 
-	dev_consume_skb_any(skb);
 	return 0;
-err_drop:
-	dev_kfree_skb_any(skb);
-	return -EINVAL;
 }
 
 static int
-- 
2.17.1

[PATCH bpf-next 4/6] nfp: bpf: allow receiving perf events on data queues

[Jakub Kicinski]

Control queue is fairly low latency, and requires SKB allocations,
which means we can't even reach 0.5Msps with perf events.  Allow
perf events to be delivered to data queues.  This allows us to not
only use multiple queues, but also receive and deliver to user space
more than 5Msps per queue (Xeon E5-2630 v4 2.20GHz, no retpolines).

Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
---
 drivers/net/ethernet/netronome/nfp/bpf/cmsg.c | 18 ++++++++++++++++++
 drivers/net/ethernet/netronome/nfp/bpf/main.c |  1 +
 drivers/net/ethernet/netronome/nfp/bpf/main.h |  3 +++
 3 files changed, 22 insertions(+)

diff --git a/drivers/net/ethernet/netronome/nfp/bpf/cmsg.c b/drivers/net/ethernet/netronome/nfp/bpf/cmsg.c
index 0a89b53962aa..1946291bf4fd 100644
--- a/drivers/net/ethernet/netronome/nfp/bpf/cmsg.c
+++ b/drivers/net/ethernet/netronome/nfp/bpf/cmsg.c
@@ -468,3 +468,21 @@ void nfp_bpf_ctrl_msg_rx(struct nfp_app *app, struct sk_buff *skb)
 err_free:
 	dev_kfree_skb_any(skb);
 }
+
+void
+nfp_bpf_ctrl_msg_rx_raw(struct nfp_app *app, const void *data, unsigned int len)
+{
+	struct nfp_app_bpf *bpf = app->priv;
+	const struct cmsg_hdr *hdr = data;
+
+	if (unlikely(len < sizeof(struct cmsg_reply_map_simple))) {
+		cmsg_warn(bpf, "cmsg drop - too short %d!\n", len);
+		return;
+	}
+
+	if (hdr->type == CMSG_TYPE_BPF_EVENT)
+		nfp_bpf_event_output(bpf, data, len);
+	else
+		cmsg_warn(bpf, "cmsg drop - msg type %d with raw buffer!\n",
+			  hdr->type);
+}
diff --git a/drivers/net/ethernet/netronome/nfp/bpf/main.c b/drivers/net/ethernet/netronome/nfp/bpf/main.c
index 994d2b756fe1..192e88981fb2 100644
--- a/drivers/net/ethernet/netronome/nfp/bpf/main.c
+++ b/drivers/net/ethernet/netronome/nfp/bpf/main.c
@@ -490,6 +490,7 @@ const struct nfp_app_type app_bpf = {
 	.vnic_free	= nfp_bpf_vnic_free,
 
 	.ctrl_msg_rx	= nfp_bpf_ctrl_msg_rx,
+	.ctrl_msg_rx_raw	= nfp_bpf_ctrl_msg_rx_raw,
 
 	.setup_tc	= nfp_bpf_setup_tc,
 	.bpf		= nfp_ndo_bpf,
diff --git a/drivers/net/ethernet/netronome/nfp/bpf/main.h b/drivers/net/ethernet/netronome/nfp/bpf/main.h
index e25d3c0c7e43..017e0ae5e736 100644
--- a/drivers/net/ethernet/netronome/nfp/bpf/main.h
+++ b/drivers/net/ethernet/netronome/nfp/bpf/main.h
@@ -505,4 +505,7 @@ int nfp_bpf_event_output(struct nfp_app_bpf *bpf, const void *data,
 			 unsigned int len);
 
 void nfp_bpf_ctrl_msg_rx(struct nfp_app *app, struct sk_buff *skb);
+void
+nfp_bpf_ctrl_msg_rx_raw(struct nfp_app *app, const void *data,
+			unsigned int len);
 #endif
-- 
2.17.1

[PATCH bpf-next 5/6] nfp: bpf: remember maps by ID

[Jakub Kicinski]

Record perf maps by map ID, not raw kernel pointer.  This helps
with debug messages, because printing pointers to logs is frowned
upon, and makes debug easier for the users, as map ID is something
they should be more familiar with.  Note that perf maps are offload
neutral, therefore IDs won't be orphaned.

While at it use a rate limited print helper for the error message.

Reported-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
---
CC: Kees Cook <keescook@chromium.org>

 drivers/net/ethernet/netronome/nfp/bpf/cmsg.c |  2 --
 drivers/net/ethernet/netronome/nfp/bpf/jit.c  | 12 ++++++----
 drivers/net/ethernet/netronome/nfp/bpf/main.c |  4 ++--
 drivers/net/ethernet/netronome/nfp/bpf/main.h |  3 +++
 .../net/ethernet/netronome/nfp/bpf/offload.c  | 22 +++++++++++--------
 5 files changed, 26 insertions(+), 17 deletions(-)

diff --git a/drivers/net/ethernet/netronome/nfp/bpf/cmsg.c b/drivers/net/ethernet/netronome/nfp/bpf/cmsg.c
index 1946291bf4fd..2572a4b91c7c 100644
--- a/drivers/net/ethernet/netronome/nfp/bpf/cmsg.c
+++ b/drivers/net/ethernet/netronome/nfp/bpf/cmsg.c
@@ -43,8 +43,6 @@
 #include "fw.h"
 #include "main.h"
 
-#define cmsg_warn(bpf, msg...)	nn_dp_warn(&(bpf)->app->ctrl->dp, msg)
-
 #define NFP_BPF_TAG_ALLOC_SPAN	(U16_MAX / 4)
 
 static bool nfp_bpf_all_tags_busy(struct nfp_app_bpf *bpf)
diff --git a/drivers/net/ethernet/netronome/nfp/bpf/jit.c b/drivers/net/ethernet/netronome/nfp/bpf/jit.c
index 1d9e36835404..3c22d27de9da 100644
--- a/drivers/net/ethernet/netronome/nfp/bpf/jit.c
+++ b/drivers/net/ethernet/netronome/nfp/bpf/jit.c
@@ -3883,6 +3883,7 @@ static int nfp_bpf_replace_map_ptrs(struct nfp_prog *nfp_prog)
 	struct nfp_insn_meta *meta1, *meta2;
 	struct nfp_bpf_map *nfp_map;
 	struct bpf_map *map;
+	u32 id;
 
 	nfp_for_each_insn_walk2(nfp_prog, meta1, meta2) {
 		if (meta1->skip || meta2->skip)
@@ -3894,11 +3895,14 @@ static int nfp_bpf_replace_map_ptrs(struct nfp_prog *nfp_prog)
 
 		map = (void *)(unsigned long)((u32)meta1->insn.imm |
 					      (u64)meta2->insn.imm << 32);
-		if (bpf_map_offload_neutral(map))
-			continue;
-		nfp_map = map_to_offmap(map)->dev_priv;
+		if (bpf_map_offload_neutral(map)) {
+			id = map->id;
+		} else {
+			nfp_map = map_to_offmap(map)->dev_priv;
+			id = nfp_map->tid;
+		}
 
-		meta1->insn.imm = nfp_map->tid;
+		meta1->insn.imm = id;
 		meta2->insn.imm = 0;
 	}
 
diff --git a/drivers/net/ethernet/netronome/nfp/bpf/main.c b/drivers/net/ethernet/netronome/nfp/bpf/main.c
index 192e88981fb2..cce1d2945a32 100644
--- a/drivers/net/ethernet/netronome/nfp/bpf/main.c
+++ b/drivers/net/ethernet/netronome/nfp/bpf/main.c
@@ -45,8 +45,8 @@
 
 const struct rhashtable_params nfp_bpf_maps_neutral_params = {
 	.nelem_hint		= 4,
-	.key_len		= FIELD_SIZEOF(struct nfp_bpf_neutral_map, ptr),
-	.key_offset		= offsetof(struct nfp_bpf_neutral_map, ptr),
+	.key_len		= FIELD_SIZEOF(struct bpf_map, id),
+	.key_offset		= offsetof(struct nfp_bpf_neutral_map, map_id),
 	.head_offset		= offsetof(struct nfp_bpf_neutral_map, l),
 	.automatic_shrinking	= true,
 };
diff --git a/drivers/net/ethernet/netronome/nfp/bpf/main.h b/drivers/net/ethernet/netronome/nfp/bpf/main.h
index 017e0ae5e736..57573bfa8c03 100644
--- a/drivers/net/ethernet/netronome/nfp/bpf/main.h
+++ b/drivers/net/ethernet/netronome/nfp/bpf/main.h
@@ -47,6 +47,8 @@
 #include "../nfp_asm.h"
 #include "fw.h"
 
+#define cmsg_warn(bpf, msg...)	nn_dp_warn(&(bpf)->app->ctrl->dp, msg)
+
 /* For relocation logic use up-most byte of branch instruction as scratch
  * area.  Remember to clear this before sending instructions to HW!
  */
@@ -221,6 +223,7 @@ struct nfp_bpf_map {
 struct nfp_bpf_neutral_map {
 	struct rhash_head l;
 	struct bpf_map *ptr;
+	u32 map_id;
 	u32 count;
 };
 
diff --git a/drivers/net/ethernet/netronome/nfp/bpf/offload.c b/drivers/net/ethernet/netronome/nfp/bpf/offload.c
index 293dda84818f..b1fbb3babc7f 100644
--- a/drivers/net/ethernet/netronome/nfp/bpf/offload.c
+++ b/drivers/net/ethernet/netronome/nfp/bpf/offload.c
@@ -67,7 +67,7 @@ nfp_map_ptr_record(struct nfp_app_bpf *bpf, struct nfp_prog *nfp_prog,
 	ASSERT_RTNL();
 
 	/* Reuse path - other offloaded program is already tracking this map. */
-	record = rhashtable_lookup_fast(&bpf->maps_neutral, &map,
+	record = rhashtable_lookup_fast(&bpf->maps_neutral, &map->id,
 					nfp_bpf_maps_neutral_params);
 	if (record) {
 		nfp_prog->map_records[nfp_prog->map_records_cnt++] = record;
@@ -89,6 +89,7 @@ nfp_map_ptr_record(struct nfp_app_bpf *bpf, struct nfp_prog *nfp_prog,
 	}
 
 	record->ptr = map;
+	record->map_id = map->id;
 	record->count = 1;
 
 	err = rhashtable_insert_fast(&bpf->maps_neutral, &record->l,
@@ -457,15 +458,17 @@ int nfp_bpf_event_output(struct nfp_app_bpf *bpf, const void *data,
 			 unsigned int len)
 {
 	struct cmsg_bpf_event *cbe = (void *)data;
-	u32 pkt_size, data_size;
-	struct bpf_map *map;
+	struct nfp_bpf_neutral_map *record;
+	u32 pkt_size, data_size, map_id;
+	u64 map_id_full;
 
 	if (len < sizeof(struct cmsg_bpf_event))
 		return -EINVAL;
 
 	pkt_size = be32_to_cpu(cbe->pkt_size);
 	data_size = be32_to_cpu(cbe->data_size);
-	map = (void *)(unsigned long)be64_to_cpu(cbe->map_ptr);
+	map_id_full = be64_to_cpu(cbe->map_ptr);
+	map_id = map_id_full;
 
 	if (len < sizeof(struct cmsg_bpf_event) + pkt_size + data_size)
 		return -EINVAL;
@@ -473,15 +476,16 @@ int nfp_bpf_event_output(struct nfp_app_bpf *bpf, const void *data,
 		return -EINVAL;
 
 	rcu_read_lock();
-	if (!rhashtable_lookup_fast(&bpf->maps_neutral, &map,
-				    nfp_bpf_maps_neutral_params)) {
+	record = rhashtable_lookup_fast(&bpf->maps_neutral, &map_id,
+					nfp_bpf_maps_neutral_params);
+	if (!record || map_id_full > U32_MAX) {
 		rcu_read_unlock();
-		pr_warn("perf event: dest map pointer %px not recognized, dropping event\n",
-			map);
+		cmsg_warn(bpf, "perf event: map id %lld (0x%llx) not recognized, dropping event\n",
+			  map_id_full, map_id_full);
 		return -EINVAL;
 	}
 
-	bpf_event_output(map, be32_to_cpu(cbe->cpu_id),
+	bpf_event_output(record->ptr, be32_to_cpu(cbe->cpu_id),
 			 &cbe->data[round_up(pkt_size, 4)], data_size,
 			 cbe->data, pkt_size, nfp_bpf_perf_event_copy);
 	rcu_read_unlock();
-- 
2.17.1

[PATCH bpf-next 6/6] nfp: bpf: improve map offload info messages

[Jakub Kicinski]

FW can put constraints on map element size to maximize resource
use and efficiency.  When user attempts offload of a map which
does not fit into those constraints an informational message is
printed to kernel logs to inform user about the reason offload
failed.  Map offload does not have access to any advanced error
reporting like verifier log or extack.  There is also currently
no way for us to nicely expose the FW capabilities to user
space.  Given all those constraints we should make sure log
messages are as informative as possible.  Improve them.

Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
---
 .../net/ethernet/netronome/nfp/bpf/offload.c  | 20 +++++++++++++++----
 1 file changed, 16 insertions(+), 4 deletions(-)

diff --git a/drivers/net/ethernet/netronome/nfp/bpf/offload.c b/drivers/net/ethernet/netronome/nfp/bpf/offload.c
index b1fbb3babc7f..1ccd6371a15b 100644
--- a/drivers/net/ethernet/netronome/nfp/bpf/offload.c
+++ b/drivers/net/ethernet/netronome/nfp/bpf/offload.c
@@ -380,11 +380,23 @@ nfp_bpf_map_alloc(struct nfp_app_bpf *bpf, struct bpf_offloaded_map *offmap)
 			bpf->maps.max_elems - bpf->map_elems_in_use);
 		return -ENOMEM;
 	}
-	if (offmap->map.key_size > bpf->maps.max_key_sz ||
-	    offmap->map.value_size > bpf->maps.max_val_sz ||
-	    round_up(offmap->map.key_size, 8) +
+
+	if (round_up(offmap->map.key_size, 8) +
 	    round_up(offmap->map.value_size, 8) > bpf->maps.max_elem_sz) {
-		pr_info("elements don't fit in device constraints\n");
+		pr_info("map elements too large: %u, FW max element size (key+value): %u\n",
+			round_up(offmap->map.key_size, 8) +
+			round_up(offmap->map.value_size, 8),
+			bpf->maps.max_elem_sz);
+		return -ENOMEM;
+	}
+	if (offmap->map.key_size > bpf->maps.max_key_sz) {
+		pr_info("map key size %u, FW max is %u\n",
+			offmap->map.key_size, bpf->maps.max_key_sz);
+		return -ENOMEM;
+	}
+	if (offmap->map.value_size > bpf->maps.max_val_sz) {
+		pr_info("map value size %u, FW max is %u\n",
+			offmap->map.value_size, bpf->maps.max_val_sz);
 		return -ENOMEM;
 	}
 
-- 
2.17.1

Re: [PATCH bpf-next 0/6] nfp: bpf: improve efficiency of offloaded perf events

[Daniel Borkmann]

On 07/26/2018 04:53 AM, Jakub Kicinski wrote:
> Hi!
> 
> This set is focused on improving the performance of perf events
> reported from BPF offload.  Perf events can now be received on
> packet data queues, which significantly improves the performance
> (from total of 0.5 Msps to 5Msps per core).  To get to this
> performance we need a fast path for control messages which will
> operate on raw buffers and recycle them immediately.
> 
> Patch 5 replaces the map pointers for perf maps with map IDs.
> We look the pointers up in a hashtable, anyway, to validate they
> are correct, so there is no performance difference.  Map IDs
> have the advantage of being easier to understand for users in
> case of errors (we no longer print raw pointers to the logs).
> 
> Last patch improves info messages about map offload.
> 
> Jakub Kicinski (6):
>   nfp: move repr handling on RX path
>   nfp: allow control message reception on data queues
>   nfp: bpf: pass raw data buffer to nfp_bpf_event_output()
>   nfp: bpf: allow receiving perf events on data queues
>   nfp: bpf: remember maps by ID
>   nfp: bpf: improve map offload info messages
> 
>  drivers/net/ethernet/netronome/nfp/bpf/cmsg.c | 25 +++++++-
>  drivers/net/ethernet/netronome/nfp/bpf/jit.c  | 12 ++--
>  drivers/net/ethernet/netronome/nfp/bpf/main.c |  5 +-
>  drivers/net/ethernet/netronome/nfp/bpf/main.h |  9 ++-
>  .../net/ethernet/netronome/nfp/bpf/offload.c  | 63 +++++++++++--------
>  drivers/net/ethernet/netronome/nfp/nfp_app.c  |  2 +
>  drivers/net/ethernet/netronome/nfp/nfp_app.h  | 17 +++++
>  .../ethernet/netronome/nfp/nfp_net_common.c   | 40 +++++++-----
>  .../net/ethernet/netronome/nfp/nfp_net_ctrl.h |  1 +
>  9 files changed, 125 insertions(+), 49 deletions(-)
> 

Applied to bpf-next, thanks Jakub!