Re: [PATCH] iproute2 flush: handle larger tables and deleted entries

[Stephen Hemminger <shemminger@vyatta.com>]

On Wed, 15 Jul 2009 10:50:57 -0700
Gautam Kachroo <gk@aristanetworks.com> wrote:

> On Wed, Jul 15, 2009 at 8:19 AM, Patrick McHardy<kaber@trash.net> wrote:
> > Gautam Kachroo wrote:
> >> On Tue, Jul 14, 2009 at 2:38 AM, Patrick McHardy<kaber@trash.net> wrote:
> >>> Gautam Kachroo wrote:
> >>>> use a new netlink socket when sending flush messages to avoid reading
> >>>> any pending data on the existing netlink socket.
> >>>>
> >>>> read all of the response from the netlink request -- this response can
> >>>> be split over multiple recv calls, pretty much one per netlink request
> >>>> message. ENOENT errors, which correspond to attempts to delete an
> >>>> already deleted entry, are ignored. Other errors are not ignored.
> >>>
> >>> In which case would there be any pending data? From what I can see,
> >>> this can only happen when using batching, but in that case the
> >>> previous command should continue reading until it has received all
> >>> responses (which the netlink functions appear to be doing properly).
> >>
> >> What is the "previous command"?
> >
> > The last command before the one executing when using batching.
> 
> This is independent of batching (I assume you're referring to the
> -batch option to the ip command).
> It happens when running a command like "ip neigh flush to 0.0.0.0/0"
> if there are many neighbor entries.
> 
> The implementation of flush commands, e.g. ip neigh flush, sends a
> dump request, e.g. RTM_GETNEIGH, and then sends requests, e.g.
> RTM_DELNEIGH, *while* there can be unread data from the dump request.
> There would be unread data if the response to the dump request was
> split over multiple calls to recvmsg.
> 
> >> Are you referring to rtnl_dump_filter? If rtnl_send_check comes across
> >> a failure, rtnl_dump_filter will not continue reading.
> >>
> >> Here's the situation that I'm referring to:
> >>
> >> If rtnl_send_check detects an error, it returns -1. rtnl_send_check is
> >> called from flush_update. The multiple implementations of flush_update
> >> (e.g. in ipneigh.c, ipaddress.c) propagate this return value to their
> >> caller, e.g. print_neigh or print_addrinfo.
> >>
> >> print_neigh, print_addrinfo, etc. are called from rtnl_dump_filter.
> >> rtnl_dump_filter sits in a loop calling recvmsg on the netlink socket.
> >> However, it returns the error value if the filter function (e.g.
> >> print_neigh) returns an error. In this case, rtnl_dump_filter can
> >> return before it's read all the responses.
> >> The error return from rtnl_dump_filter causes the program to exit.
> >
> > Yes, and I agree with your patch so far. My question is why you
> > need another socket.
> >
> >> use a new netlink socket when sending flush messages to avoid reading
> >> any pending data on the existing netlink socket.
> >
> > Under what circumstances would there be pending data when
> > performing a new iproute operation?
> 
> As above, it's not that there is pending data when performing a new
> iproute operation, it's that there can be pending data while
> performing a single iproute operation, namely ip <object> flush.
> The benefit of a new socket is that it won't have any data from the
> dump request waiting for it.

I posted a better fix (using MSG_PEEK).