mvneta: access skb_shared_info only on last frag

PATCH In review Lorenzo Bianconi 4 participants 11 messages Nov 20, 2020 – Nov 24, 2020

Subsystems (3)

Networking Network drivers Marvell mvneta Ethernet driver

lore

Share & tools

lore

Lorenzo Bianconi

author

2020-11-20 17:06 UTC #0 lore raw

cover mvneta: access skb_shared_info only on last frag

Build skb_shared_info on mvneta_rx_swbm stack and sync it to xdp_buff skb_shared_info area only on the last fragment.
Avoid avoid unnecessary xdp_buff initialization in mvneta_rx_swbm routine. This a preliminary series to complete xdp multi-buff in mvneta driver.

Lorenzo Bianconi (3):
net: mvneta: avoid unnecessary xdp_buff initialization
net: mvneta: move skb_shared_info in mvneta_xdp_put_buff
net: mvneta: alloc skb_shared_info on the mvneta_rx_swbm stack

 drivers/net/ethernet/marvell/mvneta.c | 55 +++++++++++++++++----------
 1 file changed, 35 insertions(+), 20 deletions(-)

Signature · 1 line

2.26.2

Lorenzo Bianconi

author

2020-11-20 17:06 UTC #1 lore raw

PATCH 1/3 net: mvneta: avoid unnecessary xdp_buff initialization

Explicitly initialize mandatory fields defining xdp_buff struct in mvneta_rx_swbm routine

Signed-off-by Lorenzo Bianconi <[email protected]>

---
 drivers/net/ethernet/marvell/mvneta.c | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

drivers/net/ethernet/marvell/mvneta.c +5 -4

index 5d559117f78c..60a285f3a55f 100644
@@ -2347,13 +2347,14 @@ static int mvneta_rx_swbm(struct napi_struct *napi,
 {
 	int rx_proc = 0, rx_todo, refill, size = 0;
 	struct net_device *dev = pp->dev;
-	struct xdp_buff xdp_buf = {
-		.frame_sz = PAGE_SIZE,
-		.rxq = &rxq->xdp_rxq,
-	};
 	struct mvneta_stats ps = {};
 	struct bpf_prog *xdp_prog;
 	u32 desc_status, frame_sz;
+	struct xdp_buff xdp_buf;
+
+	xdp_buf.data_hard_start = NULL;
+	xdp_buf.frame_sz = PAGE_SIZE;
+	xdp_buf.rxq = &rxq->xdp_rxq;
 
 	/* Get number of received packets */
 	rx_todo = mvneta_rxq_busy_desc_num_get(pp, rxq);

Signature · 1 line

2.26.2

Lorenzo Bianconi

author

2020-11-20 17:06 UTC #2 lore raw

PATCH 2/3 net: mvneta: move skb_shared_info in mvneta_xdp_put_buff caller

In reply to Lorenzo Bianconi

Pass skb_shared_info pointer from mvneta_xdp_put_buff caller. This is a preliminary patch to reduce accesses to skb_shared_info area and reduce cache misses.
Remove napi parameter in mvneta_xdp_put_buff signature since it is always run in NAPI context

Signed-off-by Lorenzo Bianconi <[email protected]>

---
 drivers/net/ethernet/marvell/mvneta.c | 32 ++++++++++++++++++---------
 1 file changed, 21 insertions(+), 11 deletions(-)

drivers/net/ethernet/marvell/mvneta.c +21 ~~-11~~

index 60a285f3a55f..17c446b1cb94 100644
@@ -2033,16 +2033,16 @@ int mvneta_rx_refill_queue(struct mvneta_port *pp, struct mvneta_rx_queue *rxq)
 
 static void
 mvneta_xdp_put_buff(struct mvneta_port *pp, struct mvneta_rx_queue *rxq,
-		    struct xdp_buff *xdp, int sync_len, bool napi)
+		    struct xdp_buff *xdp, struct skb_shared_info *sinfo,
+		    int sync_len)
 {
-	struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
 	int i;
 
 	for (i = 0; i < sinfo->nr_frags; i++)
 		page_pool_put_full_page(rxq->page_pool,
-					skb_frag_page(&sinfo->frags[i]), napi);
+					skb_frag_page(&sinfo->frags[i]), true);
 	page_pool_put_page(rxq->page_pool, virt_to_head_page(xdp->data),
-			   sync_len, napi);
+			   sync_len, true);
 }
 
 static int

@@ -2179,6 +2179,7 @@ mvneta_run_xdp(struct mvneta_port *pp, struct mvneta_rx_queue *rxq,
 	       struct bpf_prog *prog, struct xdp_buff *xdp,
 	       u32 frame_sz, struct mvneta_stats *stats)
 {
+	struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
 	unsigned int len, data_len, sync;
 	u32 ret, act;

@@ -2199,7 +2200,7 @@ mvneta_run_xdp(struct mvneta_port *pp, struct mvneta_rx_queue *rxq,
 
 		err = xdp_do_redirect(pp->dev, xdp, prog);
 		if (unlikely(err)) {
-			mvneta_xdp_put_buff(pp, rxq, xdp, sync, true);
+			mvneta_xdp_put_buff(pp, rxq, xdp, sinfo, sync);
 			ret = MVNETA_XDP_DROPPED;
 		} else {
 			ret = MVNETA_XDP_REDIR;

@@ -2210,7 +2211,7 @@ mvneta_run_xdp(struct mvneta_port *pp, struct mvneta_rx_queue *rxq,
 	case XDP_TX:
 		ret = mvneta_xdp_xmit_back(pp, xdp);
 		if (ret != MVNETA_XDP_TX)
-			mvneta_xdp_put_buff(pp, rxq, xdp, sync, true);
+			mvneta_xdp_put_buff(pp, rxq, xdp, sinfo, sync);
 		break;
 	default:
 		bpf_warn_invalid_xdp_action(act);

@@ -2219,7 +2220,7 @@ mvneta_run_xdp(struct mvneta_port *pp, struct mvneta_rx_queue *rxq,
 		trace_xdp_exception(pp->dev, prog, act);
 		fallthrough;
 	case XDP_DROP:
-		mvneta_xdp_put_buff(pp, rxq, xdp, sync, true);
+		mvneta_xdp_put_buff(pp, rxq, xdp, sinfo, sync);
 		ret = MVNETA_XDP_DROPPED;
 		stats->xdp_drop++;
 		break;

@@ -2406,7 +2407,10 @@ static int mvneta_rx_swbm(struct napi_struct *napi,
 			continue;
 
 		if (size) {
-			mvneta_xdp_put_buff(pp, rxq, &xdp_buf, -1, true);
+			struct skb_shared_info *sinfo;
+
+			sinfo = xdp_get_shared_info_from_buff(&xdp_buf);
+			mvneta_xdp_put_buff(pp, rxq, &xdp_buf, sinfo, -1);
 			goto next;
 		}

@@ -2417,8 +2421,10 @@ static int mvneta_rx_swbm(struct napi_struct *napi,
 		skb = mvneta_swbm_build_skb(pp, rxq, &xdp_buf, desc_status);
 		if (IS_ERR(skb)) {
 			struct mvneta_pcpu_stats *stats = this_cpu_ptr(pp->stats);
+			struct skb_shared_info *sinfo;
 
-			mvneta_xdp_put_buff(pp, rxq, &xdp_buf, -1, true);
+			sinfo = xdp_get_shared_info_from_buff(&xdp_buf);
+			mvneta_xdp_put_buff(pp, rxq, &xdp_buf, sinfo, -1);
 
 			u64_stats_update_begin(&stats->syncp);
 			stats->es.skb_alloc_error++;

@@ -2438,8 +2444,12 @@ static int mvneta_rx_swbm(struct napi_struct *napi,
 	}
 	rcu_read_unlock();
 
-	if (xdp_buf.data_hard_start)
-		mvneta_xdp_put_buff(pp, rxq, &xdp_buf, -1, true);
+	if (xdp_buf.data_hard_start) {
+		struct skb_shared_info *sinfo;
+
+		sinfo = xdp_get_shared_info_from_buff(&xdp_buf);
+		mvneta_xdp_put_buff(pp, rxq, &xdp_buf, sinfo, -1);
+	}
 
 	if (ps.xdp_redirect)
 		xdp_do_flush_map();

Signature · 1 line

2.26.2

Lorenzo Bianconi

author

2020-11-20 17:06 UTC #3 lore raw

PATCH 3/3 net: mvneta: alloc skb_shared_info on the mvneta_rx_swbm stack

In reply to Lorenzo Bianconi

Build skb_shared_info on mvneta_rx_swbm stack and sync it to xdp_buff skb_shared_info area only on the last fragment. Leftover cache miss in mvneta_swbm_rx_frame will be addressed introducing mb bit in xdp_buff/xdp_frame struct

Signed-off-by Lorenzo Bianconi <[email protected]>

---
 drivers/net/ethernet/marvell/mvneta.c | 42 +++++++++++++++------------
 1 file changed, 23 insertions(+), 19 deletions(-)

drivers/net/ethernet/marvell/mvneta.c +23 ~~-19~~

index 17c446b1cb94..b40804e421a7 100644
@@ -2278,9 +2278,9 @@ mvneta_swbm_add_rx_fragment(struct mvneta_port *pp,
 			    struct mvneta_rx_desc *rx_desc,
 			    struct mvneta_rx_queue *rxq,
 			    struct xdp_buff *xdp, int *size,
+			    struct skb_shared_info *xdp_sinfo,
 			    struct page *page)
 {
-	struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
 	struct net_device *dev = pp->dev;
 	enum dma_data_direction dma_dir;
 	int data_len, len;

@@ -2298,13 +2298,22 @@ mvneta_swbm_add_rx_fragment(struct mvneta_port *pp,
 				len, dma_dir);
 	rx_desc->buf_phys_addr = 0;
 
-	if (data_len > 0 && sinfo->nr_frags < MAX_SKB_FRAGS) {
-		skb_frag_t *frag = &sinfo->frags[sinfo->nr_frags];
+	if (data_len > 0 && xdp_sinfo->nr_frags < MAX_SKB_FRAGS) {
+		skb_frag_t *frag = &xdp_sinfo->frags[xdp_sinfo->nr_frags++];
 
 		skb_frag_off_set(frag, pp->rx_offset_correction);
 		skb_frag_size_set(frag, data_len);
 		__skb_frag_set_page(frag, page);
-		sinfo->nr_frags++;
+
+		/* last fragment */
+		if (len == *size) {
+			struct skb_shared_info *sinfo;
+
+			sinfo = xdp_get_shared_info_from_buff(xdp);
+			sinfo->nr_frags = xdp_sinfo->nr_frags;
+			memcpy(sinfo->frags, xdp_sinfo->frags,
+			       sinfo->nr_frags * sizeof(skb_frag_t));
+		}
 	} else {
 		page_pool_put_full_page(rxq->page_pool, page, true);
 	}

@@ -2348,6 +2357,7 @@ static int mvneta_rx_swbm(struct napi_struct *napi,
 {
 	int rx_proc = 0, rx_todo, refill, size = 0;
 	struct net_device *dev = pp->dev;
+	struct skb_shared_info sinfo;
 	struct mvneta_stats ps = {};
 	struct bpf_prog *xdp_prog;
 	u32 desc_status, frame_sz;

@@ -2357,6 +2367,8 @@ static int mvneta_rx_swbm(struct napi_struct *napi,
 	xdp_buf.frame_sz = PAGE_SIZE;
 	xdp_buf.rxq = &rxq->xdp_rxq;
 
+	sinfo.nr_frags = 0;
+
 	/* Get number of received packets */
 	rx_todo = mvneta_rxq_busy_desc_num_get(pp, rxq);

@@ -2395,11 +2407,11 @@ static int mvneta_rx_swbm(struct napi_struct *napi,
 				rx_desc->buf_phys_addr = 0;
 				page_pool_put_full_page(rxq->page_pool, page,
 							true);
-				continue;
+				goto next;
 			}
 
 			mvneta_swbm_add_rx_fragment(pp, rx_desc, rxq, &xdp_buf,
-						    &size, page);
+						    &size, &sinfo, page);
 		} /* Middle or Last descriptor */
 
 		if (!(rx_status & MVNETA_RXD_LAST_DESC))

@@ -2407,10 +2419,7 @@ static int mvneta_rx_swbm(struct napi_struct *napi,
 			continue;
 
 		if (size) {
-			struct skb_shared_info *sinfo;
-
-			sinfo = xdp_get_shared_info_from_buff(&xdp_buf);
-			mvneta_xdp_put_buff(pp, rxq, &xdp_buf, sinfo, -1);
+			mvneta_xdp_put_buff(pp, rxq, &xdp_buf, &sinfo, -1);
 			goto next;
 		}

@@ -2421,10 +2430,8 @@ static int mvneta_rx_swbm(struct napi_struct *napi,
 		skb = mvneta_swbm_build_skb(pp, rxq, &xdp_buf, desc_status);
 		if (IS_ERR(skb)) {
 			struct mvneta_pcpu_stats *stats = this_cpu_ptr(pp->stats);
-			struct skb_shared_info *sinfo;
 
-			sinfo = xdp_get_shared_info_from_buff(&xdp_buf);
-			mvneta_xdp_put_buff(pp, rxq, &xdp_buf, sinfo, -1);
+			mvneta_xdp_put_buff(pp, rxq, &xdp_buf, &sinfo, -1);
 
 			u64_stats_update_begin(&stats->syncp);
 			stats->es.skb_alloc_error++;

@@ -2441,15 +2448,12 @@ static int mvneta_rx_swbm(struct napi_struct *napi,
 		napi_gro_receive(napi, skb);
 next:
 		xdp_buf.data_hard_start = NULL;
+		sinfo.nr_frags = 0;
 	}
 	rcu_read_unlock();
 
-	if (xdp_buf.data_hard_start) {
-		struct skb_shared_info *sinfo;
-
-		sinfo = xdp_get_shared_info_from_buff(&xdp_buf);
-		mvneta_xdp_put_buff(pp, rxq, &xdp_buf, sinfo, -1);
-	}
+	if (xdp_buf.data_hard_start)
+		mvneta_xdp_put_buff(pp, rxq, &xdp_buf, &sinfo, -1);
 
 	if (ps.xdp_redirect)
 		xdp_do_flush_map();

Signature · 1 line

2.26.2

Jesse Brandeburg

Reviewed-by on the v1 series

2020-11-20 23:10 UTC #4 lore raw

In reply to Lorenzo Bianconi

Quoted 5 lines · Lorenzo Bianconi

Build skb_shared_info on mvneta_rx_swbm stack and sync it to xdp_buff skb_shared_info area only on the last fragment.
Avoid avoid unnecessary xdp_buff initialization in mvneta_rx_swbm routine. This a preliminary series to complete xdp multi-buff in mvneta driver.

Quoted 8 lines

Lorenzo Bianconi (3):
net: mvneta: avoid unnecessary xdp_buff initialization
net: mvneta: move skb_shared_info in mvneta_xdp_put_buff
net: mvneta: alloc skb_shared_info on the mvneta_rx_swbm stack
 drivers/net/ethernet/marvell/mvneta.c | 55 +++++++++++++++++----------
 1 file changed, 35 insertions(+), 20 deletions(-)

For the series:

Reviewed-by Jesse Brandeburg <[email protected]>

Jakub Kicinski

maintainer Reviewed-by on the v1 series

2020-11-24 20:27 UTC #5 lore raw

In reply to Lorenzo Bianconi

Quoted 4 lines · On Fri, 20 Nov 2020 18:05:41 +0100 Lorenzo Bianconi wrote:

Build skb_shared_info on mvneta_rx_swbm stack and sync it to xdp_buff skb_shared_info area only on the last fragment.
Avoid avoid unnecessary xdp_buff initialization in mvneta_rx_swbm routine. This a preliminary series to complete xdp multi-buff in mvneta driver.

Looks fine, but since you need this for XDP multi-buff it should probably go via bpf-next, right?

Reviewed-by Jakub Kicinski <[email protected]>

Lorenzo Bianconi

author

2020-11-24 22:19 UTC #6 lore raw

Quoted 10 lines · On Fri, 20 Nov 2020 18:05:41 +0100 Lorenzo Bianconi wrote:

Quoted 4 lines
Build skb_shared_info on mvneta_rx_swbm stack and sync it to xdp_buff skb_shared_info area only on the last fragment.
Avoid avoid unnecessary xdp_buff initialization in mvneta_rx_swbm routine. This a preliminary series to complete xdp multi-buff in mvneta driver.
Looks fine, but since you need this for XDP multi-buff it should probably go via bpf-next, right?
Reviewed-by Jakub Kicinski <[email protected]>

Hi Jakub,

thx for the review. Since the series changes networking-only bits I sent it for net-next, but I agree bpf-next is better.

@Alexei, Daniel: is it fine to merge the series in bpf-next?

Regards,
Lorenzo

Daniel Borkmann

2020-11-24 22:26 UTC #7 lore raw

Quoted 9 lines · Lorenzo Bianconi

Quoted 9 lines · On Fri, 20 Nov 2020 18:05:41 +0100 Lorenzo Bianconi wrote:
Quoted 4 lines
Build skb_shared_info on mvneta_rx_swbm stack and sync it to xdp_buff skb_shared_info area only on the last fragment.
Avoid avoid unnecessary xdp_buff initialization in mvneta_rx_swbm routine. This a preliminary series to complete xdp multi-buff in mvneta driver.
Looks fine, but since you need this for XDP multi-buff it should probably go via bpf-next, right?

Quoted 8 lines

Quoted 1 line
Reviewed-by Jakub Kicinski <[email protected]>
Hi Jakub,
thx for the review. Since the series changes networking-only bits I sent it for net-next, but I agree bpf-next is better.
@Alexei, Daniel: is it fine to merge the series in bpf-next?

Yeah totally fine, will take it into bpf-next in a bit.

Thanks,
Daniel

Jakub Kicinski

maintainer

2020-11-24 22:31 UTC #8 lore raw

Quoted 12 lines · On Tue, 24 Nov 2020 23:25:11 +0100 Daniel Borkmann wrote:

Quoted 11 lines · Lorenzo Bianconi
Quoted 10 lines · On Fri, 20 Nov 2020 18:05:41 +0100 Lorenzo Bianconi wrote:
Quoted 4 lines
Build skb_shared_info on mvneta_rx_swbm stack and sync it to xdp_buff skb_shared_info area only on the last fragment.
Avoid avoid unnecessary xdp_buff initialization in mvneta_rx_swbm routine. This a preliminary series to complete xdp multi-buff in mvneta driver.
Looks fine, but since you need this for XDP multi-buff it should probably go via bpf-next, right?
Reviewed-by Jakub Kicinski <[email protected]>

Quoted 8 lines

Quoted 6 lines
Hi Jakub,
thx for the review. Since the series changes networking-only bits I sent it for net-next, but I agree bpf-next is better.
@Alexei, Daniel: is it fine to merge the series in bpf-next?
Yeah totally fine, will take it into bpf-next in a bit.

FWIW watch out with the Link:s, it wasn't CCed to bpf@vger.

Daniel Borkmann

2020-11-24 23:00 UTC #9 lore raw

Quoted 15 lines · Jakub Kicinski

Quoted 14 lines · On Tue, 24 Nov 2020 23:25:11 +0100 Daniel Borkmann wrote:
Quoted 13 lines · Lorenzo Bianconi
Quoted 10 lines · On Fri, 20 Nov 2020 18:05:41 +0100 Lorenzo Bianconi wrote:
Quoted 4 lines
Build skb_shared_info on mvneta_rx_swbm stack and sync it to xdp_buff skb_shared_info area only on the last fragment.
Avoid avoid unnecessary xdp_buff initialization in mvneta_rx_swbm routine. This a preliminary series to complete xdp multi-buff in mvneta driver.
Looks fine, but since you need this for XDP multi-buff it should probably go via bpf-next, right?
Reviewed-by Jakub Kicinski <[email protected]>
Hi Jakub,

Quoted 8 lines

Quoted 6 lines
Quoted 4 lines
thx for the review. Since the series changes networking-only bits I sent it for net-next, but I agree bpf-next is better.
@Alexei, Daniel: is it fine to merge the series in bpf-next?
Yeah totally fine, will take it into bpf-next in a bit.
FWIW watch out with the Link:s, it wasn't CCed to bpf@vger.

@Jakub, I think it's less hassle if you take the series in. Looking closer, net-next has commit 9c79a8ab5f12 ("net: mvneta: fix possible memory leak in mvneta_swbm_add_rx_fragment") which bpf-next is currently lacking, and this series here is touching the part of this code, so it will create unnecessary merge conflicts. I'll likely flush out bpf-next PR on Thurs/Fri at latest, so bpf-next will then have everything needed once we sync back from net-next after merge.

Thanks,
Daniel

Jakub Kicinski

maintainer

2020-11-24 23:10 UTC #10 lore raw

Quoted 23 lines · On Wed, 25 Nov 2020 00:00:33 +0100 Daniel Borkmann wrote:

Quoted 22 lines · Jakub Kicinski
Quoted 20 lines · On Tue, 24 Nov 2020 23:25:11 +0100 Daniel Borkmann wrote:
Quoted 9 lines · Lorenzo Bianconi
Quoted 9 lines · On Fri, 20 Nov 2020 18:05:41 +0100 Lorenzo Bianconi wrote:
Quoted 4 lines
Build skb_shared_info on mvneta_rx_swbm stack and sync it to xdp_buff skb_shared_info area only on the last fragment.
Avoid avoid unnecessary xdp_buff initialization in mvneta_rx_swbm routine. This a preliminary series to complete xdp multi-buff in mvneta driver.
Looks fine, but since you need this for XDP multi-buff it should probably go via bpf-next, right?
Quoted 8 lines
Quoted 1 line
Reviewed-by Jakub Kicinski <[email protected]>
Hi Jakub,
thx for the review. Since the series changes networking-only bits I sent it for net-next, but I agree bpf-next is better.
@Alexei, Daniel: is it fine to merge the series in bpf-next?
Yeah totally fine, will take it into bpf-next in a bit.

Quoted 8 lines

Quoted 1 line
FWIW watch out with the Link:s, it wasn't CCed to bpf@vger.
@Jakub, I think it's less hassle if you take the series in. Looking closer, net-next has commit 9c79a8ab5f12 ("net: mvneta: fix possible memory leak in mvneta_swbm_add_rx_fragment") which bpf-next is currently lacking, and this series here is touching the part of this code, so it will create unnecessary merge conflicts. I'll likely flush out bpf-next PR on Thurs/Fri at latest, so bpf-next will then have everything needed once we sync back from net-next after merge.

I see, applied to net-next then. Thanks!