Le ven. 6 nov. 2026 à 01:10, Paul Moore [off-list ref] a écrit :
On Wed, Nov 4, 2026 at 10:03 AM Eric Dumazet [off-list ref] wrote:
quoted
calipso_req_setattr() and calipso_req_delattr() allocate ireq->ipv6_opt
with ipv6_renew_options(), which charges the listener sk_omem_alloc
through sock_kmalloc().
tcp_v6_reqsk_destructor() frees ireq->ipv6_opt without uncharging it,
so each request socket leaks its options size from the listener
budget. Once sysctl_optmem_max is reached, sock_kmalloc() fails,
calipso_req_setattr() returns -ENOBUFS, and tcp_v6_route_req() drops
all subsequent SYN packets. __sk_destruct() also reports an optmem
leakage when the listener is destroyed.
Uncharging in tcp_v6_reqsk_destructor() would not be correct: after a
migration, req->rsk_listener is no longer the socket that was charged.
These options belong to the request socket, there is no reason to
charge them to the listener. Change ipv6_renew_options() to use
kmalloc() when its sk argument is NULL, and use this from
calipso_req_setattr() and calipso_req_delattr(),
like cipso_v4_req_setattr() which uses kzalloc().
Note:
After this patch, calipso_req_setattr() and calipso_req_delattr()
no longer need the listener. Commit 10876da918fa ("calipso: Fix
null-ptr-deref in calipso_req_{set,del}attr().") could be reverted
in net-next (also removing the then unused sk variables), so that
CALIPSO works with SYN cookies.
That would be great, are you planning to submit a follow-up patch for
that once this lands? Or is that left for others? Either answer is
fine, just trying to get ahead of any duplicated effort.
I will provide the follow-up patch, thanks !
Otherwise, this looks good to me, thanks Eric.
Acked-by: Paul Moore <paul@paul-moore.com>