Skip to contentSkip to conversation

Fix TAP networking on host kernels without IFF_VNET_HDR support

PATCH In review Pierre Riteau 2 participants 8 messages Nov 24, 2009 – Nov 25, 2009 Networking

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Fix TAP networking on host kernels without IFF_VNET_HDR support

vnet_hdr is initialized at 1 by default. We need to reset it to 0 if the kernel doesn't support IFF_VNET_HDR.

---
 net/tap-linux.c |    2 ++
 1 files changed, 2 insertions(+), 0 deletions(-)
net/tap-linux.c +2 -0
index 0f621a2..e4f7e27 100644
@@ -52,6 +52,8 @@ int tap_open(char *ifname, int ifname_size, int *vnet_hdr, int vnet_hdr_required
             features & IFF_VNET_HDR) {
             *vnet_hdr = 1;
             ifr.ifr_flags |= IFF_VNET_HDR;
+        } else {
+            *vnet_hdr = 0;
         }
 
         if (vnet_hdr_required && !*vnet_hdr) {
Signature · 1 line
1.6.5
#1 lore raw

PATCH 1/1 net: initialize vnet_hdr in net_tap_init()

Hi Pierre,

Quoted 2 lines · Pierre Riteau

vnet_hdr is initialized at 1 by default. We need to reset it to 0 if the kernel doesn't support IFF_VNET_HDR.

Thanks for the patch, but I'd prefer us to make sure we catch all cases.

Does this work for you?

Thanks,
Mark.

From: Mark McLoughlin <[email protected]>
Subject: [PATCH] net: initialize vnet_hdr in net_tap_init()

Don't assume that all tap_open() implementations will set it to zero if VNET_HDR support isn't found.

Fixes tap networking on host kernels lacking IFF_VNET_HDR support.

---
 net/tap.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
net/tap.c +1 -1
index d34feec..7fb9e16 100644
@@ -378,7 +378,7 @@ static int net_tap_init(QemuOpts *opts, int *vnet_hdr)
 int net_init_tap(QemuOpts *opts, Monitor *mon, const char *name, VLANState *vlan)
 {
     TAPState *s;
-    int fd, vnet_hdr;
+    int fd, vnet_hdr = 0;
 
     if (qemu_opt_get(opts, "fd")) {
         if (qemu_opt_get(opts, "ifname") ||
Signature · 1 line
1.6.5.2
#2 lore raw

Re: [Qemu-devel] [PATCH] net: initialize vnet_hdr in net_tap_init()

Quoted 6 lines · Mark McLoughlin

Hi Pierre,

Quoted 2 lines · Pierre Riteau

vnet_hdr is initialized at 1 by default. We need to reset it to 0 if the kernel doesn't support IFF_VNET_HDR.

Quoted 6 lines

Thanks for the patch, but I'd prefer us to make sure we catch all cases.

Does this work for you?

Thanks,
Mark.

No it doesn't fix the problem, since vnet_hdr is then changed to 1 in net_tap_init by this line, just before calling tap_open:

    *vnet_hdr = qemu_opt_get_bool(opts, "vnet_hdr", 1);

I run qemu like this: sudo qemu -m 1024 -hda /mnt/hda.img -net nic,macaddr=DE:AD:BE:EF:69:25 -net tap,script=/home/priteau/qemu-ifup -boot c -vnc 0:0

Signature · 1 line
Pierre Riteau -- http://perso.univ-rennes1.fr/pierre.riteau/
#4 lore raw
Quoted 4 lines · Mark McLoughlin
Quoted 3 lines · Pierre Riteau

vnet_hdr is initialized at 1 by default. We need to reset it to 0 if the kernel doesn't support IFF_VNET_HDR.

Quoted 8 lines
Quoted 1 line

Thanks Pierre, I see why this is needed now

Cheers,
Mark.

Thanks for your rapid answer!

BTW, every time I run qemu I see this error message:

TUNSETOFFLOAD ioctl() failed: Invalid argument

It is caused by the piece of code at the end of net/tap-linux.c:

    if (ioctl(fd, TUNSETOFFLOAD, offload) != 0) {
        offload &= ~TUN_F_UFO;
        if (ioctl(fd, TUNSETOFFLOAD, offload) != 0) {
            fprintf(stderr, "TUNSETOFFLOAD ioctl() failed: %s\n",
                    strerror(errno));
        }
    }

Isn't there a way to detect whether the kernel supports the TUNSETOFFLOAD ioctl at all?

Signature · 1 line
Pierre Riteau -- http://perso.univ-rennes1.fr/pierre.riteau/
#5 lore raw
Quoted 32 lines · Pierre Riteau
Quoted 4 lines · Mark McLoughlin
Quoted 3 lines · Pierre Riteau

vnet_hdr is initialized at 1 by default. We need to reset it to 0 if the kernel doesn't support IFF_VNET_HDR.

Quoted 8 lines
Quoted 1 line

Thanks Pierre, I see why this is needed now

Cheers,
Mark.

Thanks for your rapid answer!

BTW, every time I run qemu I see this error message:

TUNSETOFFLOAD ioctl() failed: Invalid argument

It is caused by the piece of code at the end of net/tap-linux.c:

    if (ioctl(fd, TUNSETOFFLOAD, offload) != 0) {
        offload &= ~TUN_F_UFO;
        if (ioctl(fd, TUNSETOFFLOAD, offload) != 0) {
            fprintf(stderr, "TUNSETOFFLOAD ioctl() failed: %s\n",
                    strerror(errno));
        }
    }
Quoted 2 lines

Isn't there a way to detect whether the kernel supports the TUNSETOFFLOAD ioctl at all?

The kernel will set errno to EINVAL if TUNSETOFFLOAD isn't supported, so we could just ignore that case:

     if (ioctl(fd, TUNSETOFFLOAD, offload) != 0) {
          offload &= ~TUN_F_UFO;
          if (ioctl(fd, TUNSETOFFLOAD, offload) != 0 && errno != EINVAL) {
              fprintf(stderr, "TUNSETOFFLOAD ioctl() failed: %s\n",
                      strerror(errno));
          }
      }

The only concern is that we'll also miss out on an error message if EINVAL is set for another reason. Currently, the only other reason if we pass a offload flag not supported by the kernel, but that should never happen.

Feel free to send a patch with that change and I'll ack it

Thanks,
Mark.

#6 lore raw
Quoted 52 lines · Mark McLoughlin
Quoted 32 lines · Pierre Riteau
Quoted 4 lines · Mark McLoughlin
Quoted 3 lines · Pierre Riteau

vnet_hdr is initialized at 1 by default. We need to reset it to 0 if the kernel doesn't support IFF_VNET_HDR.

Quoted 8 lines
Quoted 1 line

Thanks Pierre, I see why this is needed now

Cheers,
Mark.

Thanks for your rapid answer!

BTW, every time I run qemu I see this error message:

TUNSETOFFLOAD ioctl() failed: Invalid argument

It is caused by the piece of code at the end of net/tap-linux.c:

   if (ioctl(fd, TUNSETOFFLOAD, offload) != 0) {
       offload &= ~TUN_F_UFO;
       if (ioctl(fd, TUNSETOFFLOAD, offload) != 0) {
           fprintf(stderr, "TUNSETOFFLOAD ioctl() failed: %s\n",
                   strerror(errno));
       }
   }
Quoted 2 lines

Isn't there a way to detect whether the kernel supports the TUNSETOFFLOAD ioctl at all?

The kernel will set errno to EINVAL if TUNSETOFFLOAD isn't supported, so we could just ignore that case:

    if (ioctl(fd, TUNSETOFFLOAD, offload) != 0) {
         offload &= ~TUN_F_UFO;
         if (ioctl(fd, TUNSETOFFLOAD, offload) != 0 && errno != EINVAL) {
             fprintf(stderr, "TUNSETOFFLOAD ioctl() failed: %s\n",
                     strerror(errno));
         }
     }

The only concern is that we'll also miss out on an error message if EINVAL is set for another reason. Currently, the only other reason if we pass a offload flag not supported by the kernel, but that should never happen.

Quoted 5 lines

Feel free to send a patch with that change and I'll ack it

Thanks,
Mark.

Couldn't we probe the kernel with a 0 offload value to check if it supports TUNSETOFFLOAD? I tried the following and it works on my 2.6.26 Debian kernel.
What do you think? I will send a proper patch if you agree.

net/tap-linux.c +5 -0
index 0f621a2..e038e1a 100644
@@ -129,6 +129,11 @@ void tap_fd_set_offload(int fd, int csum, int tso4,
 {
     unsigned int offload = 0;
 
+    /* Check if our kernel supports TUNSETOFFLOAD */
+    if (ioctl(fd, TUNSETOFFLOAD, 0) != 0 && errno == EINVAL) {
+        return;
+    }
+
     if (csum) {
         offload |= TUN_F_CSUM;
         if (tso4)
Signature · 1 line
Pierre Riteau -- http://perso.univ-rennes1.fr/pierre.riteau/
#7 lore raw
Quoted 24 lines · Pierre Riteau
Quoted 22 lines · Mark McLoughlin
Quoted 2 lines · Pierre Riteau

Isn't there a way to detect whether the kernel supports the TUNSETOFFLOAD ioctl at all?

The kernel will set errno to EINVAL if TUNSETOFFLOAD isn't supported, so we could just ignore that case:

    if (ioctl(fd, TUNSETOFFLOAD, offload) != 0) {
         offload &= ~TUN_F_UFO;
         if (ioctl(fd, TUNSETOFFLOAD, offload) != 0 && errno != EINVAL) {
             fprintf(stderr, "TUNSETOFFLOAD ioctl() failed: %s\n",
                     strerror(errno));
         }
     }

The only concern is that we'll also miss out on an error message if EINVAL is set for another reason. Currently, the only other reason if we pass a offload flag not supported by the kernel, but that should never happen.

Feel free to send a patch with that change and I'll ack it

....

Quoted 10 lines

Couldn't we probe the kernel with a 0 offload value to check if it supports TUNSETOFFLOAD? I tried the following and it works on my 2.6.26 Debian kernel.
What do you think? I will send a proper patch if you agree.

net/tap-linux.c +0 -0
index 0f621a2..e038e1a 100644

net/tap-linux.c void tap_fd_set_offload(int fd, int csum, int tso4, · line 136

Quoted 8 lines
net/tap-linux.c +5 -0
@@ -129,6 +129,11 @@ void tap_fd_set_offload(int fd, int csum, int tso4,
 {
     unsigned int offload = 0;
 
+    /* Check if our kernel supports TUNSETOFFLOAD */
+    if (ioctl(fd, TUNSETOFFLOAD, 0) != 0 && errno == EINVAL) {
+        return;
+    }
+

It's not ideal because a) it's another syscall and b) you're briefly disabling any flags that may have been set previously ... but in our case, neither is a real concern and it's a nice, simple solution.

So, sounds good to me :)

Cheers,
Mark.