Skip to contentSkip to conversation

s390: qeth driver fixes [3/6]

PATCH In review Frank Pavlic 2 participants 3 messages Sep 15, 2006

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

PATCH 6/9 s390: qeth driver fixes [3/6]

[PATCH 6/9] s390: qeth driver fixes [3/6]

From: Frank Pavlic <[email protected]>
fixed kernel panic caused by qeth driver:
Using a bonding device qeth driver will realloc headroom for every skb coming from the bond device. Once this happens qeth frees the original skb and set the skb pointer to the new realloced skb. Under heavy transmit workload (e.g.UDP streams) through bond network device the qdio output queue might get full. In this case we return with EBUSY from qeth_send_packet. Returning to qeth_hard_start_xmit routine
the skb address on the stack still points to the old address, which has been freed before.
Returning from qeth_hard_start_xmit with EBUSY results in requeuing the skb. In this case it corrupts the qdisc queue and results in kernel panic.

---

 drivers/s390/net/qeth.h      |   33 +++----
 drivers/s390/net/qeth_main.c |  193 ++++++++++++++++++++++--------------------
 drivers/s390/net/qeth_tso.h  |    2 
 3 files changed, 113 insertions(+), 115 deletions(-)

80a8bd094c70579c027f15838501759ab8f3331d

drivers/s390/net/qeth.h +13 -20
index e8bd8c5..c04ee91 100644
@@ -859,23 +859,18 @@ qeth_get_ipa_adp_type(enum qeth_link_typ
 	}
 }
 
-static inline int
-qeth_realloc_headroom(struct qeth_card *card, struct sk_buff **skb, int size)
+static inline struct sk_buff *
+qeth_realloc_headroom(struct qeth_card *card, struct sk_buff *skb, int size)
 {
-	struct sk_buff *new_skb = NULL;
+	struct sk_buff *new_skb = skb;
 
-	if (skb_headroom(*skb) < size){
-		new_skb = skb_realloc_headroom(*skb, size);
-		if (!new_skb) {
-                        PRINT_ERR("qeth_prepare_skb: could "
-                                  "not realloc headroom for qeth_hdr "
-                                  "on interface %s", QETH_CARD_IFNAME(card));
-                        return -ENOMEM;
-                }
-		kfree_skb(*skb);
-                *skb = new_skb;
-	}
-	return 0;
+	if (skb_headroom(skb) >= size)
+		return skb;
+	new_skb = skb_realloc_headroom(skb, size);
+	if (!new_skb) 
+		PRINT_ERR("Could not realloc headroom for qeth_hdr "
+			  "on interface %s", QETH_CARD_IFNAME(card));
+	return new_skb;
 }
 
 static inline struct sk_buff *
@@ -885,16 +880,15 @@ qeth_pskb_unshare(struct sk_buff *skb, i
         if (!skb_cloned(skb))
                 return skb;
         nskb = skb_copy(skb, pri);
-        kfree_skb(skb); /* free our shared copy */
         return nskb;
 }
 
 static inline void *
-qeth_push_skb(struct qeth_card *card, struct sk_buff **skb, int size)
+qeth_push_skb(struct qeth_card *card, struct sk_buff *skb, int size)
 {
         void *hdr;
 
-	hdr = (void *) skb_push(*skb, size);
+	hdr = (void *) skb_push(skb, size);
         /*
          * sanity check, the Linux memory allocation scheme should
          * never present us cases like this one (the qdio header size plus
@@ -903,8 +897,7 @@ qeth_push_skb(struct qeth_card *card, st
         if ((((unsigned long) hdr) & (~(PAGE_SIZE - 1))) !=
             (((unsigned long) hdr + size +
               QETH_IP_HEADER_SIZE) & (~(PAGE_SIZE - 1)))) {
-                PRINT_ERR("qeth_prepare_skb: misaligned "
-                          "packet on interface %s. Discarded.",
+                PRINT_ERR("Misaligned packet on interface %s. Discarded.",
                           QETH_CARD_IFNAME(card));
                 return NULL;
         }
drivers/s390/net/qeth_main.c +99 -94

333 lines not on this page · raw message

drivers/s390/net/qeth_tso.h +1 -1
index 593f298..14504af 100644
@@ -24,7 +24,7 @@ static inline struct qeth_hdr_tso *
 qeth_tso_prepare_skb(struct qeth_card *card, struct sk_buff **skb)
 {
 	QETH_DBF_TEXT(trace, 5, "tsoprsk");
-	return qeth_push_skb(card, skb, sizeof(struct qeth_hdr_tso));
+	return qeth_push_skb(card, *skb, sizeof(struct qeth_hdr_tso));
 }
 
 /**
Signature · 1 line
1.2.4
#1 lore raw
Quoted 8 lines · Frank Pavlic

[PATCH 6/9] s390: qeth driver fixes [3/6]

From: Frank Pavlic <[email protected]>
fixed kernel panic caused by qeth driver:
Using a bonding device qeth driver will realloc headroom for every skb coming from the bond device. Once this happens qeth frees the original skb and set the skb pointer to the new realloced skb.

Frank, does the following patch to bonding (to track larger than
usual hard_header_len) resolve this problem without changing qeth? I believe this patch is queued for 2.6.19.

To: [email protected], [email protected]
Cc: Jeff Garzik <[email protected]>
Subject: [PATCH 5/7] bonding: Handle large hard_header_len
X-Mailer: MH-E 7.83; nmh 1.1-RC4; GNU Emacs 21.4.1
Date: Fri, 01 Sep 2006 15:12:44 -0700
From: Jay Vosburgh <fubar@death>

The bonding driver fails to adjust its hard_header_len when enslaving interfaces. Whenever an interface with a hard_header_len greater than the ETH_HLEN default is enslaved, the potential for an oops exists, and if the oops happens while responding to an arp request, for example, the system panics. GIANFAR devices may use an extended hard_header for VLAN or hardware checksumming. Enslaving such a device and then transmitting over it causes a kernel panic.

Patch modified from submitter's original, but submitter agreed with this patch in private email.

drivers/net/bonding/bond_main.c +6 -1
@@ -1211,10 +1211,14 @@ static int bond_compute_features(struct 
 	unsigned long features = BOND_INTERSECT_FEATURES;
 	struct slave *slave;
 	struct net_device *bond_dev = bond->dev;
+	unsigned short max_hard_header_len = ETH_HLEN;
 	int i;
 
-	bond_for_each_slave(bond, slave, i)
+	bond_for_each_slave(bond, slave, i) {
 		features &= (slave->dev->features & BOND_INTERSECT_FEATURES);
+		if (slave->dev->hard_header_len > max_hard_header_len)
+			max_hard_header_len = slave->dev->hard_header_len;
+	}
 
 	if ((features & NETIF_F_SG) && 
 	    !(features & NETIF_F_ALL_CSUM))
@@ -1232,6 +1236,7 @@ static int bond_compute_features(struct 
 
 	features |= (bond_dev->features & ~BOND_INTERSECT_FEATURES);
 	bond_dev->features = features;
+	bond_dev->hard_header_len = max_hard_header_len;
 
 	return 0;
 }
	-J

---
-Jay Vosburgh, IBM Linux Technology Center, [email protected]

#2 lore raw

yes it would but the problem in qeth is not only bonding related. We just have found it in a bonding scenario . But this kernel panic could also occur without bonding.

But thanks for reviewing it

Quoted 71 lines · Jay Vosburgh
Quoted 8 lines · Frank Pavlic

[PATCH 6/9] s390: qeth driver fixes [3/6]

From: Frank Pavlic <[email protected]>
fixed kernel panic caused by qeth driver:
Using a bonding device qeth driver will realloc headroom for every skb coming from the bond device. Once this happens qeth frees the original skb and set the skb pointer to the new realloced skb.

Frank, does the following patch to bonding (to track larger than
usual hard_header_len) resolve this problem without changing qeth? I believe this patch is queued for 2.6.19.

To: [email protected], [email protected]
Cc: Jeff Garzik <[email protected]>
Subject: [PATCH 5/7] bonding: Handle large hard_header_len
X-Mailer: MH-E 7.83; nmh 1.1-RC4; GNU Emacs 21.4.1
Date: Fri, 01 Sep 2006 15:12:44 -0700
From: Jay Vosburgh <fubar@death>

The bonding driver fails to adjust its hard_header_len when enslaving interfaces. Whenever an interface with a hard_header_len greater than the ETH_HLEN default is enslaved, the potential for an oops exists, and if the oops happens while responding to an arp request, for example, the system panics. GIANFAR devices may use an extended hard_header for VLAN or hardware checksumming. Enslaving such a device and then transmitting over it causes a kernel panic.

Patch modified from submitter's original, but submitter agreed with this patch in private email.

drivers/net/bonding/bond_main.c +6 -1
@@ -1211,10 +1211,14 @@ static int bond_compute_features(struct 
 	unsigned long features = BOND_INTERSECT_FEATURES;
 	struct slave *slave;
 	struct net_device *bond_dev = bond->dev;
+	unsigned short max_hard_header_len = ETH_HLEN;
 	int i;
 
-	bond_for_each_slave(bond, slave, i)
+	bond_for_each_slave(bond, slave, i) {
 		features &= (slave->dev->features & BOND_INTERSECT_FEATURES);
+		if (slave->dev->hard_header_len > max_hard_header_len)
+			max_hard_header_len = slave->dev->hard_header_len;
+	}
 
 	if ((features & NETIF_F_SG) && 
 	    !(features & NETIF_F_ALL_CSUM))
@@ -1232,6 +1236,7 @@ static int bond_compute_features(struct 
 
 	features |= (bond_dev->features & ~BOND_INTERSECT_FEATURES);
 	bond_dev->features = features;
+	bond_dev->hard_header_len = max_hard_header_len;
 
 	return 0;
 }


	-J

---
-Jay Vosburgh, IBM Linux Technology Center, [email protected]