Re: Solving address deletion bottleneck in SCTP

On Mon, Oct 28, 2024 at 8:49 AM Gilad Naaman [off-list ref] wrote:

Hello,

We've noticed that when a namespace has a large amount of IP addresses,
the list `net->sctp.local_addr_list` gets obscenely long.

This list contains both IPv4 and IPv6 addresses, of all scopes, and it is
a single long list, instead of a hashtable.

In our case we had 12K interfaces, each with an IPv4 and 2 IPv6 addresses
(GUA+LLA), which made deletion of a single address pretty expensive, since
it requires a linear search through 36K addresses.

Internally we solved it pretty naively by turning the list into hashmap, which
helped us avoid this bottleneck:

    + #define SCTP_ADDR_HSIZE_SHIFT     8
    + #define SCTP_ADDR_HSIZE           (1 << SCTP_ADDR_HSIZE_SHIFT)

    -   struct list_head local_addr_list;
    +   struct list_head local_addr_list[SCTP_ADDR_HSIZE];


I've used the same factor used by the IPv6 & IPv4 address tables.

I am not entirely sure this patch solves a big enough problem for the greater
general kernel community to warrant the increased memory usage (~2KiB-p-netns),
so I'll avoid sending it.

Recently, though, both IPv4 and IPv6 tables were namespacified, which makes
me think that maybe local_addr_list is no longer necessary, enabling us to
them directly instead of maintaining a separate list.

As far as I could tell, the only field of `struct sctp_sockaddr_entry` that
are used for items of this list, aside from the address itself, is the `valid`
bit, which can probably be folded into `struct in_ifaddr` and `struct inet6_ifaddr`.

What I'm suggesting, in short is:
 - Represent `valid` inside the original address structs.
 - Replace iteration of `local_addr_list` with iteration of ns addr tables
 - Eliminate `local_addr_list`

Is this a reasonable proposal?

This would simplify sctp_inet6addr_event() and sctp_inetaddr_event(),
but complicate sctp_copy_laddrs() and sctp_copy_local_addr_list().

Would you like to create a patch for this and let's see how it looks?

Note I don't think that that 'valid' bit is useful:

               if (addr->a.sa.sa_family == AF_INET &&
                               addr->a.v4.sin_addr.s_addr ==
                               ifa->ifa_local) {
                       sctp_addr_wq_mgmt(net, addr, SCTP_ADDR_DEL);
                       found = 1;
                                      <-------- [1]
                       addr->valid = 0;
                       list_del_rcu(&addr->list);
                       break;
               }

'addr' can be copied before "addr->valid = 0;" with addr->valid =1 in
another thread anyway. I think you can ignore this 'valid' bit.

Thanks.