On Wed, Jul 28, 2010 at 11:46:24AM -0700, Joe Eykholt wrote:
> On 7/28/10 11:42 AM, Chris Leech wrote:
> > If the link state is changing while fipvlan is running, it could result in a
> > packet socket being added to the poll list more than once. When that
> > happened,
> > recvmsg would be called multiple times for a single FIP response.
> >
> > Previously this caused a hang, which the change to MSG_DONTWAIT worked
> > around,
> > but an error message is still printed every time the condition occurs. This
> > change fixes the poll list managment to correctly avoid the error.
> >
> > Signed-off-by: Chris Leech<[email protected]>
> > ---
> >
> > fipvlan.c | 7 +++++++
> > 1 files changed, 7 insertions(+), 0 deletions(-)
> >
> > diff --git a/fipvlan.c b/fipvlan.c
> > index eddffbb..b400427 100644
> > --- a/fipvlan.c
> > +++ b/fipvlan.c
> > @@ -80,6 +80,11 @@ static int pfd_len = 0;
> > void pfd_add(int fd)
> > {
> > struct pollfd *npfd;
> > + int i;
> > +
> > + for (i = 0; i< pfd_len; i++)
> > + if (pfd[i].fd == fd)
> > + return;
> >
> > npfd = realloc(pfd, (pfd_len + 1) * sizeof(struct pollfd));
> > if (!npfd) {
> > @@ -319,6 +324,8 @@ void rtnl_recv_newlink(struct nlmsghdr *nh)
> > iff->running = (ifm->ifi_flags& IFF_RUNNING) == IFF_RUNNING;
> > if (iff->running)
> > pfd_add(iff->ps);
> > + else
> > + pfd_remove(iff->ps);
> > return;
> > }
> >
> >
>
> Is it possible the caller of pfd_add would later want to remove the socket
> from the list and succeed even though it wasn't the last user?
>
> I didn't look into this, but thought I'd mention the issue anyway.
> Maybe it can't happen.
Or (I just re-read an interpreted your email differently), or you asking
about the pfd_add return added here preventing what would essentially be
an attempt to reference count the fd?
I'll admit that the two chunks of this patch are both addressing parts
of same bug in different ways; it's both removing the fd on link down
and making sure that if there are multiple netlink change events that
all show link up the fd will only be added once.
- Chris
_______________________________________________
devel mailing list
[email protected]
http://www.open-fcoe.org/mailman/listinfo/devel
