Re: [PATCH net] net: Add a gfp_t parameter in ip_fib_metrics_init to support atomic context

On 11/28/22 10:53 PM, Duoming Zhou wrote:

The ip_fib_metrics_init() do not support atomic context, because it
calls "kzalloc(..., GFP_KERNEL)". When ip_fib_metrics_init() is used
in atomic context, the sleep-in-atomic-context bug will happen.

Did you actually hit this sleep-in-atomic-context bug or is it theory
based on code analysis?

For example, the neigh_proxy_process() is a timer handler that is
used to process the proxy request that is timeout. But it could call
ip_fib_metrics_init(). As a result, the can_block flag in ipv6_add_addr()
and the gfp_flags in addrconf_f6i_alloc() and ip6_route_info_create()
are useless. The process is shown below.

    (atomic context)
neigh_proxy_process()
  pndisc_redo()
    ndisc_recv_ns()
      addrconf_dad_failure()
        ipv6_add_addr(..., bool can_block)
          addrconf_f6i_alloc(..., gfp_t gfp_flags)

	cfg has fc_mx == NULL.

            ip6_route_info_create(..., gfp_t gfp_flags)

	rt->fib6_metrics = ip_fib_metrics_init(net, cfg->fc_mx, cfg->fc_mx_len,
extack);

              ip_fib_metrics_init()

        if (!fc_mx)
                return (struct dst_metrics *)&dst_default_metrics;

                kzalloc(..., GFP_KERNEL) //may sleep