Thread (21 messages) flat view 21 messages, 5 authors, 2012-07-21

Re: [PATCH] b44: add 64 bit stats

From: Eric Dumazet <hidden>
Date: 2012-07-21 05:09:42
Also in: lvs-devel

On Fri, 2012-07-20 at 22:22 -0400, Kevin Groeneveld wrote:
On Fri, Jul 20, 2012 at 2:56 PM, Kevin Groeneveld [off-list ref] wrote:
quoted
quoted
In fact all network drivers should use the _bh version.
Could you send a patch for all of them, based on net-next tree ?
Sure, I can work on that.  It should be a relatively easy thing to
update.  I can probably send a patch within the next couple days.
As I have been working on the patch I have been trying convince myself
that each case I change actually needs the _bh version of the
functions instead of blindly changing them.  So far I have found the
following where the change seems to make sense:

drivers/net/dummy.c
drivers/net/ethernet/neterion/vxge/vxge-main.c
drivers/net/loopback.c
drivers/net/virtio_net.c
net/bridge/br_device.c
Thats right.
The only two other places in the networking code that use
u64_stats_fetch_begin/u64_stats_fetch_retry are:

net/l2tp/l2tp_netlink.c
This one is completely buggy, dont waste your time on it.

My plan for this one : dont try to have 64bit stats on 32bit arches, and
use plain "unsigned long" counters (if they are percpu), or
atomic_long_t (if they are shared by all cpus)

The writer sides might be run concurrently by several cpus, so
u64_stats_update_begin(&sstats->syncp); are racy : a reader can
be trapped forever.
net/netfilter/ipvs/ip_vs_est.c
Same problem for this one, I think.

I CCed ipvs maintainers so that they can take a look.
Do these need to be updated as well?  Looking at these files quickly
and with my limited knowledge of the kernel I am not sure if they
update the stats in a BH context or not.


Kevin
Thanks !
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help