Skip to contentSkip to conversation

reduce open coded skb->next access for gso segment walking

PATCH In review Jason A. Donenfeld 3 participants 12 messages Jan 8, 2020 – Feb 18, 2020

Subsystems (9)
lore
Share & tools
lore

Jason A. Donenfeld

author maintainer
#0 lore raw

cover reduce open coded skb->next access for gso segment walking

This patchset introduces the skb_list_walk_safe helper macro, in order to add some sanity to the myrid ways drivers have of walking through gso segments. The goal is to reduce future bugs commonly caused by open coding these sorts of things, and to in the future make it easier to swap out the underlying list representation.

This first patch series addresses the easy uses of drivers iterating over the returned list of skb_gso_segments, for drivers that live in drivers/net/*. There are still other use cases to tackle later for net/*, and after these low-hanging fruits are taken care of, I imagine there are more subtle cases of gso segment walking that isn't just a direct return value from skb_gso_segments, and eventually this will have to be tackled. This series is the first in that direction.

Jason A. Donenfeld (8):
net: introduce skb_list_walk_safe for skb segment walking
net: tap: use skb_list_walk_safe helper for gso segments
net: r8152: use skb_list_walk_safe helper for gso segments
net: tg3: use skb_list_walk_safe helper for gso segments
net: sunvnet: use skb_list_walk_safe helper for gso segments
net: sfc: use skb_list_walk_safe helper for gso segments
net: myri10ge: use skb_list_walk_safe helper for gso segments
net: iwlwifi: use skb_list_walk_safe helper for gso segments

 drivers/net/ethernet/broadcom/tg3.c              | 12 +++++-------
 drivers/net/ethernet/myricom/myri10ge/myri10ge.c |  8 +++-----
 drivers/net/ethernet/sfc/tx.c                    |  7 ++-----
 drivers/net/ethernet/sun/sunvnet_common.c        |  9 +++------
 drivers/net/tap.c                                | 14 ++++++--------
 drivers/net/usb/r8152.c                          | 12 +++++-------
 drivers/net/wireguard/device.h                   |  8 --------
 drivers/net/wireless/intel/iwlwifi/mvm/tx.c      |  9 ++-------
 include/linux/skbuff.h                           |  5 +++++
 9 files changed, 31 insertions(+), 53 deletions(-)
Signature · 1 line
2.24.1
author maintainer
#1 lore raw

PATCH 1/8 net: introduce skb_list_walk_safe for skb segment walking

As part of the continual effort to remove direct usage of skb->next and skb->prev, this patch adds a helper for iterating through the singly-linked variant of skb lists, which are used for lists of GSO packet. The name "skb_list_..." has been chosen to match the existing function, "kfree_skb_list, which also operates on these singly-linked lists, and the "..._walk_safe" part is the same idiom as elsewhere in the kernel.

This patch removes the helper from wireguard and puts it into linux/skbuff.h, while making it a bit more robust for general usage. In particular, parenthesis are added around the macro argument usage, and it now accounts for trying to iterate through an already-null skb pointer, which will simply run the iteration zero times. This latter enhancement means it can be used to replace both do { ... } while and while (...) open-coded idioms.

This should take care of these three possible usages, which match all current methods of iterations.

skb_list_walk_safe(segs, skb, next) { ... }
skb_list_walk_safe(skb, skb, next) { ... }
skb_list_walk_safe(segs, skb, segs) { ... }

Gcc appears to generate efficient code for each of these.

---
 drivers/net/wireguard/device.h | 8 --------
 include/linux/skbuff.h         | 5 +++++
 2 files changed, 5 insertions(+), 8 deletions(-)
drivers/net/wireguard/device.h +0 -8
index c91f3051c5c7..b15a8be9d816 100644
@@ -62,12 +62,4 @@ struct wg_device {
 int wg_device_init(void);
 void wg_device_uninit(void);
 
-/* Later after the dust settles, this can be moved into include/linux/skbuff.h,
- * where virtually all code that deals with GSO segs can benefit, around ~30
- * drivers as of writing.
- */
-#define skb_list_walk_safe(first, skb, next)                                   \
-	for (skb = first, next = skb->next; skb;                               \
-	     skb = next, next = skb ? skb->next : NULL)
-
 #endif /* _WG_DEVICE_H */
include/linux/skbuff.h +5 -0
index e9133bcf0544..64e5b1be9ff5 100644
@@ -1478,6 +1478,11 @@ static inline void skb_mark_not_on_list(struct sk_buff *skb)
 	skb->next = NULL;
 }
 
+/* Iterate through singly-linked GSO fragments of an skb. */
+#define skb_list_walk_safe(first, skb, next)                                   \
+	for ((skb) = (first), (next) = (skb) ? (skb)->next : NULL; (skb);      \
+	     (skb) = (next), (next) = (skb) ? (skb)->next : NULL)
+
 static inline void skb_list_del_init(struct sk_buff *skb)
 {
 	__list_del_entry(&skb->list);
Signature · 1 line
2.24.1
author maintainer
#2 lore raw

PATCH 2/8 net: tap: use skb_list_walk_safe helper for gso segments

In reply to Jason A. Donenfeld

This is a straight-forward conversion case for the new function, and while we're at it, we can remove a null write to skb->next by replacing it with skb_mark_not_on_list.

---
 drivers/net/tap.c | 14 ++++++--------
 1 file changed, 6 insertions(+), 8 deletions(-)
drivers/net/tap.c +6 -8
index a6d63665ad03..1f4bdd94407a 100644
@@ -341,6 +341,7 @@ rx_handler_result_t tap_handle_frame(struct sk_buff **pskb)
 		features |= tap->tap_features;
 	if (netif_needs_gso(skb, features)) {
 		struct sk_buff *segs = __skb_gso_segment(skb, features, false);
+		struct sk_buff *next;
 
 		if (IS_ERR(segs))
 			goto drop;
@@ -352,16 +353,13 @@ rx_handler_result_t tap_handle_frame(struct sk_buff **pskb)
 		}
 
 		consume_skb(skb);
-		while (segs) {
-			struct sk_buff *nskb = segs->next;
-
-			segs->next = NULL;
-			if (ptr_ring_produce(&q->ring, segs)) {
-				kfree_skb(segs);
-				kfree_skb_list(nskb);
+		skb_list_walk_safe(segs, skb, next) {
+			skb_mark_not_on_list(skb);
+			if (ptr_ring_produce(&q->ring, skb)) {
+				kfree_skb(skb);
+				kfree_skb_list(next);
 				break;
 			}
-			segs = nskb;
 		}
 	} else {
 		/* If we receive a partial checksum and the tap side
Signature · 1 line
2.24.1
author maintainer
#3 lore raw

PATCH 3/8 net: r8152: use skb_list_walk_safe helper for gso segments

In reply to Jason A. Donenfeld

This is a straight-forward conversion case for the new function, and while we're at it, we can remove a null write to skb->next by replacing it with skb_mark_not_on_list.

---
 drivers/net/usb/r8152.c | 12 +++++-------
 1 file changed, 5 insertions(+), 7 deletions(-)
drivers/net/usb/r8152.c +5 -7
index 9ec1da429514..fe22a582373b 100644
@@ -1897,8 +1897,8 @@ static void r8152_csum_workaround(struct r8152 *tp, struct sk_buff *skb,
 {
 	if (skb_shinfo(skb)->gso_size) {
 		netdev_features_t features = tp->netdev->features;
+		struct sk_buff *segs, *seg, *next;
 		struct sk_buff_head seg_list;
-		struct sk_buff *segs, *nskb;
 
 		features &= ~(NETIF_F_SG | NETIF_F_IPV6_CSUM | NETIF_F_TSO6);
 		segs = skb_gso_segment(skb, features);
@@ -1907,12 +1907,10 @@ static void r8152_csum_workaround(struct r8152 *tp, struct sk_buff *skb,
 
 		__skb_queue_head_init(&seg_list);
 
-		do {
-			nskb = segs;
-			segs = segs->next;
-			nskb->next = NULL;
-			__skb_queue_tail(&seg_list, nskb);
-		} while (segs);
+		skb_list_walk_safe(segs, seg, next) {
+			skb_mark_not_on_list(seg);
+			__skb_queue_tail(&seg_list, seg);
+		}
 
 		skb_queue_splice(&seg_list, list);
 		dev_kfree_skb(skb);
Signature · 1 line
2.24.1
author maintainer
#4 lore raw

PATCH 4/8 net: tg3: use skb_list_walk_safe helper for gso segments

In reply to Jason A. Donenfeld

This is a straight-forward conversion case for the new function, and while we're at it, we can remove a null write to skb->next by replacing it with skb_mark_not_on_list.

---
 drivers/net/ethernet/broadcom/tg3.c | 12 +++++-------
 1 file changed, 5 insertions(+), 7 deletions(-)
drivers/net/ethernet/broadcom/tg3.c +5 -7
index 460b4992914a..88466255bf66 100644
@@ -7874,8 +7874,8 @@ static netdev_tx_t tg3_start_xmit(struct sk_buff *, struct net_device *);
 static int tg3_tso_bug(struct tg3 *tp, struct tg3_napi *tnapi,
 		       struct netdev_queue *txq, struct sk_buff *skb)
 {
-	struct sk_buff *segs, *nskb;
 	u32 frag_cnt_est = skb_shinfo(skb)->gso_segs * 3;
+	struct sk_buff *segs, *seg, *next;
 
 	/* Estimate the number of fragments in the worst case */
 	if (unlikely(tg3_tx_avail(tnapi) <= frag_cnt_est)) {
@@ -7898,12 +7898,10 @@ static int tg3_tso_bug(struct tg3 *tp, struct tg3_napi *tnapi,
 	if (IS_ERR(segs) || !segs)
 		goto tg3_tso_bug_end;
 
-	do {
-		nskb = segs;
-		segs = segs->next;
-		nskb->next = NULL;
-		tg3_start_xmit(nskb, tp->dev);
-	} while (segs);
+	skb_list_walk_safe(segs, seg, next) {
+		skb_mark_not_on_list(seg);
+		tg3_start_xmit(seg, tp->dev);
+	}
 
 tg3_tso_bug_end:
 	dev_consume_skb_any(skb);
Signature · 1 line
2.24.1
author maintainer
#5 lore raw

PATCH 5/8 net: sunvnet: use skb_list_walk_safe helper for gso segments

In reply to Jason A. Donenfeld

This is a straight-forward conversion case for the new function, and while we're at it, we can remove a null write to skb->next by replacing it with skb_mark_not_on_list.

---
 drivers/net/ethernet/sun/sunvnet_common.c | 9 +++------
 1 file changed, 3 insertions(+), 6 deletions(-)
drivers/net/ethernet/sun/sunvnet_common.c +3 -6
index a601a306f9a5..c23ce838ff63 100644
@@ -1223,7 +1223,7 @@ vnet_handle_offloads(struct vnet_port *port, struct sk_buff *skb,
 {
 	struct net_device *dev = VNET_PORT_TO_NET_DEVICE(port);
 	struct vio_dring_state *dr = &port->vio.drings[VIO_DRIVER_TX_RING];
-	struct sk_buff *segs;
+	struct sk_buff *segs, *curr, *next;
 	int maclen, datalen;
 	int status;
 	int gso_size, gso_type, gso_segs;
@@ -1282,11 +1282,8 @@ vnet_handle_offloads(struct vnet_port *port, struct sk_buff *skb,
 	skb_reset_mac_header(skb);
 
 	status = 0;
-	while (segs) {
-		struct sk_buff *curr = segs;
-
-		segs = segs->next;
-		curr->next = NULL;
+	skb_list_walk_safe(segs, curr, next) {
+		skb_mark_not_on_list(curr);
 		if (port->tso && curr->len > dev->mtu) {
 			skb_shinfo(curr)->gso_size = gso_size;
 			skb_shinfo(curr)->gso_type = gso_type;
Signature · 1 line
2.24.1
author maintainer
#6 lore raw

PATCH 6/8 net: sfc: use skb_list_walk_safe helper for gso segments

In reply to Jason A. Donenfeld

This is a straight-forward conversion case for the new function, and while we're at it, we can remove a null write to skb->next by replacing it with skb_mark_not_on_list.

---
 drivers/net/ethernet/sfc/tx.c | 7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)
drivers/net/ethernet/sfc/tx.c +2 -5
index 00c1c4402451..547692b33b4d 100644
@@ -473,12 +473,9 @@ static int efx_tx_tso_fallback(struct efx_tx_queue *tx_queue,
 	dev_consume_skb_any(skb);
 	skb = segments;
 
-	while (skb) {
-		next = skb->next;
-		skb->next = NULL;
-
+	skb_list_walk_safe(skb, skb, next) {
+		skb_mark_not_on_list(skb);
 		efx_enqueue_skb(tx_queue, skb);
-		skb = next;
 	}
 
 	return 0;
Signature · 1 line
2.24.1
#7 lore raw

Re: [PATCH 6/8] net: sfc: use skb_list_walk_safe helper for gso segments

Quoted 14 lines · Jason A. Donenfeld

This is a straight-forward conversion case for the new function, and while we're at it, we can remove a null write to skb->next by replacing it with skb_mark_not_on_list.

---
 drivers/net/ethernet/sfc/tx.c | 7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)
drivers/net/ethernet/sfc/tx.c +0 -0
index 00c1c4402451..547692b33b4d 100644

drivers/net/ethernet/sfc/tx.c static int efx_tx_tso_fallback(struct efx_tx_queue *tx_queue, · line 476

Quoted 8 lines
drivers/net/ethernet/sfc/tx.c +1 -4
@@ -473,12 +473,9 @@ static int efx_tx_tso_fallback(struct efx_tx_queue *tx_queue,
 	dev_consume_skb_any(skb);
 	skb = segments;
 
-	while (skb) {
-		next = skb->next;
-		skb->next = NULL;
-
+	skb_list_walk_safe(skb, skb, next) {

Could this be replaced with skb_list_walk_safe(segments, skb, next) { and elide the assignment just above?
Or is there some reason I'm missing not to do that? -ed

Quoted 6 lines
drivers/net/ethernet/sfc/tx.c +1 -1
+		skb_mark_not_on_list(skb);
 		efx_enqueue_skb(tx_queue, skb);
-		skb = next;
 	}
 
 	return 0;
author maintainer
#8 lore raw

Re: [PATCH 6/8] net: sfc: use skb_list_walk_safe helper for gso segments

Quoted 25 lines · Edward Cree
Quoted 14 lines · Jason A. Donenfeld

This is a straight-forward conversion case for the new function, and while we're at it, we can remove a null write to skb->next by replacing it with skb_mark_not_on_list.

---
 drivers/net/ethernet/sfc/tx.c | 7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)
drivers/net/ethernet/sfc/tx.c +0 -0
index 00c1c4402451..547692b33b4d 100644

drivers/net/ethernet/sfc/tx.c static int efx_tx_tso_fallback(struct efx_tx_queue *tx_queue, · line 476

Quoted 8 lines
drivers/net/ethernet/sfc/tx.c +1 -4
@@ -473,12 +473,9 @@ static int efx_tx_tso_fallback(struct efx_tx_queue *tx_queue,
      dev_consume_skb_any(skb);
      skb = segments;
 
-     while (skb) {
-             next = skb->next;
-             skb->next = NULL;
-
+     skb_list_walk_safe(skb, skb, next) {

Could this be replaced with

Quoted 3 lines
    skb_list_walk_safe(segments, skb, next) {

and elide the assignment just above?
Or is there some reason I'm missing not to do that?

Yes that's probably correct.

author maintainer
#9 lore raw

PATCH 7/8 net: myri10ge: use skb_list_walk_safe helper for gso segments

In reply to Jason A. Donenfeld

This is a straight-forward conversion case for the new function, and while we're at it, we can remove a null write to skb->next by replacing it with skb_mark_not_on_list.

---
 drivers/net/ethernet/myricom/myri10ge/myri10ge.c | 8 +++-----
 1 file changed, 3 insertions(+), 5 deletions(-)
drivers/net/ethernet/myricom/myri10ge/myri10ge.c +3 -5
index c979f38a2e0c..2ee0d0be113a 100644
@@ -2892,7 +2892,7 @@ static netdev_tx_t myri10ge_xmit(struct sk_buff *skb,
 static netdev_tx_t myri10ge_sw_tso(struct sk_buff *skb,
 					 struct net_device *dev)
 {
-	struct sk_buff *segs, *curr;
+	struct sk_buff *segs, *curr, *next;
 	struct myri10ge_priv *mgp = netdev_priv(dev);
 	struct myri10ge_slice_state *ss;
 	netdev_tx_t status;
@@ -2901,10 +2901,8 @@ static netdev_tx_t myri10ge_sw_tso(struct sk_buff *skb,
 	if (IS_ERR(segs))
 		goto drop;
 
-	while (segs) {
-		curr = segs;
-		segs = segs->next;
-		curr->next = NULL;
+	skb_list_walk_safe(segs, curr, next) {
+		skb_mark_not_on_list(curr);
 		status = myri10ge_xmit(curr, dev);
 		if (status != 0) {
 			dev_kfree_skb_any(curr);
Signature · 1 line
2.24.1
author maintainer
#10 lore raw

PATCH 8/8 net: iwlwifi: use skb_list_walk_safe helper for gso segments

In reply to Jason A. Donenfeld

This is a straight-forward conversion case for the new function, and while we're at it, we can remove a null write to skb->next by replacing it with skb_mark_not_on_list.

---
 drivers/net/wireless/intel/iwlwifi/mvm/tx.c | 9 ++-------
 1 file changed, 2 insertions(+), 7 deletions(-)
drivers/net/wireless/intel/iwlwifi/mvm/tx.c +2 -7
index dc5c02fbc65a..6a241d37a057 100644
@@ -847,10 +847,7 @@ iwl_mvm_tx_tso_segment(struct sk_buff *skb, unsigned int num_subframes,
 	else if (next)
 		consume_skb(skb);
 
-	while (next) {
-		tmp = next;
-		next = tmp->next;
-
+	skb_list_walk_safe(next, tmp, next) {
 		memcpy(tmp->cb, cb, sizeof(tmp->cb));
 		/*
 		 * Compute the length of all the data added for the A-MSDU.
@@ -880,9 +877,7 @@ iwl_mvm_tx_tso_segment(struct sk_buff *skb, unsigned int num_subframes,
 			skb_shinfo(tmp)->gso_size = 0;
 		}
 
-		tmp->prev = NULL;
-		tmp->next = NULL;
-
+		skb_mark_not_on_list(tmp);
 		__skb_queue_tail(mpdus_skb, tmp);
 		i++;
 	}
Signature · 1 line
2.24.1
maintainer
#11 lore raw
In reply to Jason A. Donenfeld

From: "Jason A. Donenfeld" <[email protected]>
Date: Wed, 8 Jan 2020 16:59:01 -0500

Quoted 6 lines

This patchset introduces the skb_list_walk_safe helper macro, in order to add some sanity to the myrid ways drivers have of walking through gso segments. The goal is to reduce future bugs commonly caused by open coding these sorts of things, and to in the future make it easier to swap out the underlying list representation.

Quoted 7 lines

This first patch series addresses the easy uses of drivers iterating over the returned list of skb_gso_segments, for drivers that live in drivers/net/*. There are still other use cases to tackle later for net/*, and after these low-hanging fruits are taken care of, I imagine there are more subtle cases of gso segment walking that isn't just a direct return value from skb_gso_segments, and eventually this will have to be tackled. This series is the first in that direction.

I like this, applied to net-next and build testing. Let's see where this goes.

In the iwlwifi case, the skb_mark_not_on_list() is really redundant because deep down inside the skb queue tail insert, __skb_insert() will unconditionally always write both ->next and ->prev and there are no debugging checks along the way which would trigger if skb->next was non-NULL.

I guess you could argue for defensive programming here, so there's that.

Anyways, thanks.