rtnetlink: do_setlink: Use true struct sockaddr

PATCH In review Kees Cook 3 participants 6 messages Dec 17, 2024

Subsystems (2)

Networking Network drivers

lore

Share & tools

lore

Kees Cook

author

2024-12-17 02:04 UTC #0 lore raw

PATCH 1/1 rtnetlink: do_setlink: Use true struct sockaddr

Instead of a heap allocation use a stack allocated struct sockaddr, as dev_set_mac_address_user() is the consumer (which uses a classic struct sockaddr). Cap the copy to the minimum address size between the incoming address and the traditional sa_data field itself.

Putting "sa" on the stack means it will get a reused stack slot since it is smaller than other existing single-scope stack variables (like the vfinfo array).

Signed-off-by Kees Cook <[email protected]>

---
Cc: Eric Dumazet <[email protected]>
Cc: "David S. Miller" <[email protected]>
Cc: Jakub Kicinski <[email protected]>
Cc: Paolo Abeni <[email protected]>
Cc: Ido Schimmel <[email protected]>
Cc: Petr Machata <[email protected]>
Cc: [email protected]

---
 net/core/rtnetlink.c | 22 +++++++---------------
 1 file changed, 7 insertions(+), 15 deletions(-)

net/core/rtnetlink.c +7 ~~-15~~

index ab5f201bf0ab..6da0edc0870d 100644
@@ -3048,21 +3048,13 @@ static int do_setlink(const struct sk_buff *skb, struct net_device *dev,
 	}
 
 	if (tb[IFLA_ADDRESS]) {
-		struct sockaddr *sa;
-		int len;
-
-		len = sizeof(sa_family_t) + max_t(size_t, dev->addr_len,
-						  sizeof(*sa));
-		sa = kmalloc(len, GFP_KERNEL);
-		if (!sa) {
-			err = -ENOMEM;
-			goto errout;
-		}
-		sa->sa_family = dev->type;
-		memcpy(sa->sa_data, nla_data(tb[IFLA_ADDRESS]),
-		       dev->addr_len);
-		err = dev_set_mac_address_user(dev, sa, extack);
-		kfree(sa);
+		struct sockaddr sa = { };
+
+		/* dev_set_mac_address_user() uses a true struct sockaddr. */
+		sa.sa_family = dev->type;
+		memcpy(sa.sa_data, nla_data(tb[IFLA_ADDRESS]),
+		       min(dev->addr_len, sizeof(sa.sa_data_min)));
+		err = dev_set_mac_address_user(dev, &sa, extack);
 		if (err)
 			goto errout;
 		status |= DO_SETLINK_MODIFIED;

Signature · 1 line

2.34.1

Kuniyuki Iwashima

2024-12-17 02:42 UTC #1 lore raw

From: Kees Cook <[email protected]>
Date: Mon, 16 Dec 2024 18:04:45 -0800

Quoted 3 lines

Instead of a heap allocation use a stack allocated struct sockaddr, as dev_set_mac_address_user() is the consumer (which uses a classic struct sockaddr).

I remember Eric's feedback was to keep using heap instead of stack because rtnl_newlink() path already uses too much on stack.

Quoted 55 lines

Cap the copy to the minimum address size between
the incoming address and the traditional sa_data field itself.
Putting "sa" on the stack means it will get a reused stack slot since it is smaller than other existing single-scope stack variables (like the vfinfo array).
Signed-off-by Kees Cook <[email protected]>
---
Cc: Eric Dumazet <[email protected]>
Cc: "David S. Miller" <[email protected]>
Cc: Jakub Kicinski <[email protected]>
Cc: Paolo Abeni <[email protected]>
Cc: Ido Schimmel <[email protected]>
Cc: Petr Machata <[email protected]>
Cc: [email protected]
---
 net/core/rtnetlink.c | 22 +++++++---------------
 1 file changed, 7 insertions(+), 15 deletions(-)
net/core/rtnetlink.c +7 ~~-15~~
index ab5f201bf0ab..6da0edc0870d 100644
@@ -3048,21 +3048,13 @@ static int do_setlink(const struct sk_buff *skb, struct net_device *dev,
 	}
 
 	if (tb[IFLA_ADDRESS]) {
-		struct sockaddr *sa;
-		int len;
-
-		len = sizeof(sa_family_t) + max_t(size_t, dev->addr_len,
-						  sizeof(*sa));
-		sa = kmalloc(len, GFP_KERNEL);
-		if (!sa) {
-			err = -ENOMEM;
-			goto errout;
-		}
-		sa->sa_family = dev->type;
-		memcpy(sa->sa_data, nla_data(tb[IFLA_ADDRESS]),
-		       dev->addr_len);
-		err = dev_set_mac_address_user(dev, sa, extack);
-		kfree(sa);
+		struct sockaddr sa = { };
+
+		/* dev_set_mac_address_user() uses a true struct sockaddr. */
+		sa.sa_family = dev->type;
+		memcpy(sa.sa_data, nla_data(tb[IFLA_ADDRESS]),
+		       min(dev->addr_len, sizeof(sa.sa_data_min)));
+		err = dev_set_mac_address_user(dev, &sa, extack);
 		if (err)
 			goto errout;
 		status |= DO_SETLINK_MODIFIED;
Signature · 1 line
2.34.1

Kees Cook

author

2024-12-17 07:53 UTC #2 lore raw

Quoted 8 lines · Kuniyuki Iwashima

From: Kees Cook <[email protected]>
Date: Mon, 16 Dec 2024 18:04:45 -0800
Quoted 3 lines
Instead of a heap allocation use a stack allocated struct sockaddr, as dev_set_mac_address_user() is the consumer (which uses a classic struct sockaddr).
I remember Eric's feedback was to keep using heap instead of stack because rtnl_newlink() path already uses too much on stack.

See below...

Quoted 8 lines

Quoted 8 lines
Cap the copy to the minimum address size between
the incoming address and the traditional sa_data field itself.
Putting "sa" on the stack means it will get a reused stack slot since it is smaller than other existing single-scope stack variables (like the vfinfo array).

That's why I included the rationale above. (I.e. stack usage does not grow with this patch.)

-Kees

Quoted 50 lines

Quoted 49 lines

Signed-off-by Kees Cook <[email protected]>

---
 net/core/rtnetlink.c | 22 +++++++---------------
 1 file changed, 7 insertions(+), 15 deletions(-)

net/core/rtnetlink.c +7 ~~-15~~

index ab5f201bf0ab..6da0edc0870d 100644
@@ -3048,21 +3048,13 @@ static int do_setlink(const struct sk_buff *skb, struct net_device *dev,
 	}
 
 	if (tb[IFLA_ADDRESS]) {
-		struct sockaddr *sa;
-		int len;
-
-		len = sizeof(sa_family_t) + max_t(size_t, dev->addr_len,
-						  sizeof(*sa));
-		sa = kmalloc(len, GFP_KERNEL);
-		if (!sa) {
-			err = -ENOMEM;
-			goto errout;
-		}
-		sa->sa_family = dev->type;
-		memcpy(sa->sa_data, nla_data(tb[IFLA_ADDRESS]),
-		       dev->addr_len);
-		err = dev_set_mac_address_user(dev, sa, extack);
-		kfree(sa);
+		struct sockaddr sa = { };
+
+		/* dev_set_mac_address_user() uses a true struct sockaddr. */
+		sa.sa_family = dev->type;
+		memcpy(sa.sa_data, nla_data(tb[IFLA_ADDRESS]),
+		       min(dev->addr_len, sizeof(sa.sa_data_min)));
+		err = dev_set_mac_address_user(dev, &sa, extack);
 		if (err)
 			goto errout;
 		status |= DO_SETLINK_MODIFIED;

Signature · 1 line

2.34.1

Signature · 1 line

Kees Cook

ericnetdev dumazet

2024-12-17 08:03 UTC #5 lore raw

In reply to Kees Cook

Quoted 5 lines · Kees Cook

Instead of a heap allocation use a stack allocated struct sockaddr, as dev_set_mac_address_user() is the consumer (which uses a classic struct sockaddr). Cap the copy to the minimum address size between the incoming address and the traditional sa_data field itself.

Not sure what is a 'classic sockaddr'

Quoted 38 lines

Putting "sa" on the stack means it will get a reused stack slot since it is smaller than other existing single-scope stack variables (like the vfinfo array).
Signed-off-by Kees Cook <[email protected]>
---
Cc: Eric Dumazet <[email protected]>
Cc: "David S. Miller" <[email protected]>
Cc: Jakub Kicinski <[email protected]>
Cc: Paolo Abeni <[email protected]>
Cc: Ido Schimmel <[email protected]>
Cc: Petr Machata <[email protected]>
Cc: [email protected]
---
 net/core/rtnetlink.c | 22 +++++++---------------
 1 file changed, 7 insertions(+), 15 deletions(-)
net/core/rtnetlink.c +0 ~~-12~~
index ab5f201bf0ab..6da0edc0870d 100644
@@ -3048,21 +3048,13 @@ static int do_setlink(const struct sk_buff *skb, struct net_device *dev,
        }
 
        if (tb[IFLA_ADDRESS]) {
-               struct sockaddr *sa;
-               int len;
-
-               len = sizeof(sa_family_t) + max_t(size_t, dev->addr_len,
-                                                 sizeof(*sa));
-               sa = kmalloc(len, GFP_KERNEL);
-               if (!sa) {
-                       err = -ENOMEM;
-                       goto errout;
-               }
-               sa->sa_family = dev->type;
-               memcpy(sa->sa_data, nla_data(tb[IFLA_ADDRESS]),

net/core/rtnetlink.c static int do_setlink(const struct sk_buff *skb, struct net_device *dev, · line 3057

Quoted 10 lines

net/core/rtnetlink.c +7 -3

-                      dev->addr_len);
-               err = dev_set_mac_address_user(dev, sa, extack);
-               kfree(sa);
+               struct sockaddr sa = { };
+
+               /* dev_set_mac_address_user() uses a true struct sockaddr. */
+               sa.sa_family = dev->type;
+               memcpy(sa.sa_data, nla_data(tb[IFLA_ADDRESS]),
+                      min(dev->addr_len, sizeof(sa.sa_data_min)));
+               err = dev_set_mac_address_user(dev, &sa, extack);

Have you added debug checks in dev_set_mac_address_user() to make sure dev->addr_len is always smaller than 14 ?

I think we support devices with bigger addresses.

Kuniyuki Iwashima

2024-12-17 08:03 UTC #3 lore raw

In reply to Kees Cook

From: Kees Cook <[email protected]>
Date: Mon, 16 Dec 2024 23:53:46 -0800

Quoted 14 lines · Kuniyuki Iwashima

Quoted 8 lines
From: Kees Cook <[email protected]>
Date: Mon, 16 Dec 2024 18:04:45 -0800
Quoted 3 lines
Instead of a heap allocation use a stack allocated struct sockaddr, as dev_set_mac_address_user() is the consumer (which uses a classic struct sockaddr).
I remember Eric's feedback was to keep using heap instead of stack because rtnl_newlink() path already uses too much on stack.
See below...
Quoted 2 lines

Quoted 8 lines

Quoted 6 lines
Quoted 6 lines
Cap the copy to the minimum address size between
the incoming address and the traditional sa_data field itself.
Putting "sa" on the stack means it will get a reused stack slot since it is smaller than other existing single-scope stack variables (like the vfinfo array).
That's why I included the rationale above. (I.e. stack usage does not grow with this patch.)

Ah okay, but I think we can't cap the address size to 14 bytes. MAX_ADDR_LEN is 32.

Also, dev_set_mac_address_user() still uses dev->addr_len.

Quoted 56 lines

-Kees

Quoted 50 lines

Quoted 49 lines

Signed-off-by Kees Cook <[email protected]>

---
 net/core/rtnetlink.c | 22 +++++++---------------
 1 file changed, 7 insertions(+), 15 deletions(-)

net/core/rtnetlink.c +7 ~~-15~~

index ab5f201bf0ab..6da0edc0870d 100644
@@ -3048,21 +3048,13 @@ static int do_setlink(const struct sk_buff *skb, struct net_device *dev,
 	}
 
 	if (tb[IFLA_ADDRESS]) {
-		struct sockaddr *sa;
-		int len;
-
-		len = sizeof(sa_family_t) + max_t(size_t, dev->addr_len,
-						  sizeof(*sa));
-		sa = kmalloc(len, GFP_KERNEL);
-		if (!sa) {
-			err = -ENOMEM;
-			goto errout;
-		}
-		sa->sa_family = dev->type;
-		memcpy(sa->sa_data, nla_data(tb[IFLA_ADDRESS]),
-		       dev->addr_len);
-		err = dev_set_mac_address_user(dev, sa, extack);
-		kfree(sa);
+		struct sockaddr sa = { };
+
+		/* dev_set_mac_address_user() uses a true struct sockaddr. */
+		sa.sa_family = dev->type;
+		memcpy(sa.sa_data, nla_data(tb[IFLA_ADDRESS]),
+		       min(dev->addr_len, sizeof(sa.sa_data_min)));
+		err = dev_set_mac_address_user(dev, &sa, extack);
 		if (err)
 			goto errout;
 		status |= DO_SETLINK_MODIFIED;

Signature · 1 line

2.34.1

Signature · 1 line

Kees Cook

Kees Cook

author

2024-12-17 09:33 UTC #4 lore raw

Quoted 23 lines · Kuniyuki Iwashima

From: Kees Cook <[email protected]>
Date: Mon, 16 Dec 2024 23:53:46 -0800
Quoted 21 lines · Kuniyuki Iwashima
Quoted 8 lines
From: Kees Cook <[email protected]>
Date: Mon, 16 Dec 2024 18:04:45 -0800
Quoted 3 lines
Instead of a heap allocation use a stack allocated struct sockaddr, as dev_set_mac_address_user() is the consumer (which uses a classic struct sockaddr).
I remember Eric's feedback was to keep using heap instead of stack because rtnl_newlink() path already uses too much on stack.
See below...
Quoted 8 lines
Quoted 8 lines
Cap the copy to the minimum address size between
the incoming address and the traditional sa_data field itself.
Putting "sa" on the stack means it will get a reused stack slot since it is smaller than other existing single-scope stack variables (like the vfinfo array).

Quoted 6 lines

Quoted 1 line
That's why I included the rationale above. (I.e. stack usage does not grow with this patch.)
Ah okay, but I think we can't cap the address size to 14 bytes. MAX_ADDR_LEN is 32.
Also, dev_set_mac_address_user() still uses dev->addr_len.

Oh, hrm, yes, that's true. I had audited callers of dev_set_mac_address(), but I think I must have missed callers of dev_set_mac_address_user(). Ugh. :( Let me take another look at this...

-Kees

Quoted 58 lines

-Kees

Quoted 50 lines

Quoted 49 lines

Signed-off-by Kees Cook <[email protected]>

---
 net/core/rtnetlink.c | 22 +++++++---------------
 1 file changed, 7 insertions(+), 15 deletions(-)

net/core/rtnetlink.c +7 ~~-15~~

index ab5f201bf0ab..6da0edc0870d 100644
@@ -3048,21 +3048,13 @@ static int do_setlink(const struct sk_buff *skb, struct net_device *dev,
 	}
 
 	if (tb[IFLA_ADDRESS]) {
-		struct sockaddr *sa;
-		int len;
-
-		len = sizeof(sa_family_t) + max_t(size_t, dev->addr_len,
-						  sizeof(*sa));
-		sa = kmalloc(len, GFP_KERNEL);
-		if (!sa) {
-			err = -ENOMEM;
-			goto errout;
-		}
-		sa->sa_family = dev->type;
-		memcpy(sa->sa_data, nla_data(tb[IFLA_ADDRESS]),
-		       dev->addr_len);
-		err = dev_set_mac_address_user(dev, sa, extack);
-		kfree(sa);
+		struct sockaddr sa = { };
+
+		/* dev_set_mac_address_user() uses a true struct sockaddr. */
+		sa.sa_family = dev->type;
+		memcpy(sa.sa_data, nla_data(tb[IFLA_ADDRESS]),
+		       min(dev->addr_len, sizeof(sa.sa_data_min)));
+		err = dev_set_mac_address_user(dev, &sa, extack);
 		if (err)
 			goto errout;
 		status |= DO_SETLINK_MODIFIED;

Signature · 1 line

2.34.1

Signature · 1 line

Kees Cook

Signature · 1 line

Kees Cook