Allow setting dev->weight using ip(8)

Hello

For my own testing purposes, I often need to modify dev->weight. So I patched the kernel and iproute2 to be able to modify dev->weight using ip(8). After all, one can modify tx_queue_len so why not weight?

Two kernel patches attached: one using the ioctl framework and another using the rtnetlink framework. Warning: the attached rtnetlink patch is a superset of Thomas Graf's rtnetlink patch for setting mtu and tx_queue_len.

PS: both patches are against 2.6.5-rc3-ben0 but should apply to any 2.6.

--
Eric

Looks sane.
Not sure if you should encourage ioctl though ;->
Out of curiosity, what/when do you change weight? I think its valuable to do so, just curious. BTW, As Dave pointed out alwauIFLA_WEIGHT

missing user space/ip patch btw.

cheers,
jamal

LOL ;->
As Dave pointed out any new attributes go to the end of the enumeration. It may not matter in this case if Dave hasnt applied Thomas' patch. Otherwise IFLA_WEIGHT should be last.

cheers,
jamal

Right.

New patch attached.

Eric

Please make a patch against Linus's current tree. I put the rest of your rtnetlink device interface changes in, and your patches here duplicate all of that.

I agree also with Jamel wrt. ioctl(), let's not go there if possible. :-)

See below.

Ok.

I did also go the ioctl() way because that's what iproute2-2.6.8 uses.

[Dave: sorry for the sending you this twice]

--
Eric

===== include/linux/rtnetlink.h 1.41 vs edited =====

@@ -565,6 +565,8 @@
 #define IFLA_TXQLEN IFLA_TXQLEN
        IFLA_MAP,
 #define IFLA_MAP IFLA_MAP
+       IFLA_WEIGHT,
+#define IFLA_WEIGHT IFLA_WEIGHT
        __IFLA_MAX
 };
 

===== net/core/rtnetlink.c 1.23 vs edited =====

@@ -177,6 +177,11 @@
        }
 
        if (1) {
+               u32 weight = dev->weight;
+               RTA_PUT(skb, IFLA_WEIGHT, sizeof(weight), &weight);
+       }
+
+       if (1) {
                struct ifmap map = {
                        .mem_start   = dev->mem_start,
                        .mem_end     = dev->mem_end,

@@ -331,6 +336,13 @@
                        goto out;
 
                dev->tx_queue_len = *((u32 *) RTA_DATA(ida[IFLA_TXQLEN - 1]));
+       }
+
+       if (ida[IFLA_WEIGHT - 1]) {
+               if (ida[IFLA_WEIGHT - 1]->rta_len != RTA_LENGTH(sizeof(u32)))
+                       goto out;
+
+               dev->weight = *((u32 *) RTA_DATA(ida[IFLA_WEIGHT - 1]));
        }
 
        err = 0;

iproute still uses ioctl at that time, doesn't it?

With weight one can have the NIC fetch rx descriptors while the rx interrupt remains disabled. This further means that under high input load, one single rx interrupt can result in more than RX_RING_SIZE rx packets polled. Using this + tx polling, I could observe a system fully loaded by the network with no interrupts at all.

I attached a patch to iproute-2.6.8. To make the patch compile I had to tinker with the Makefile - it seems that KERNEL_INCLUDE does no longer serve any purpose, except for ATM maybe.

Eric

* Eric Lemoine <[email protected]> 2004-08-30 16:33

I have a patch for it and added support for your new changes. I'm currently holding it back until all issues with the 2.4 backport have been solved.