The send buffer of the coredump socket is subject to the limit in net.core.wmem_default. af_unix uses sk_sndbuf / 2 - 64 bytes for a single skb. That means a send buffer below that would split a page-sized write into multiple skbs. Raise the send buffer to leave room for a page-sized write plus the header. The default value is well above that. So we really change it when it's below our minimum.
Signed-off-by: Christian Brauner (Amutable) <[email protected]> --- fs/coredump.c | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/fs/coredump.c b/fs/coredump.c index ac4e922c9cb9..24a6405ea242 100644 --- a/fs/coredump.c +++ b/fs/coredump.c @@ -665,6 +665,9 @@ static int umh_coredump_setup(struct subprocess_info *info, struct cred *new) } #ifdef CONFIG_UNIX +/* af_unix halves the send buffer to size a single skb. */ +#define COREDUMP_SOCK_SNDBUF_MIN (3 * PAGE_SIZE) + static bool coredump_sock_connect(struct core_name *cn, struct coredump_params *cprm) { struct file *file __free(fput) = NULL; @@ -690,6 +693,10 @@ static bool coredump_sock_connect(struct core_name *cn, struct coredump_params * if (retval < 0) return false; + /* Don't let a page-sized write split into several skbs. */ + socket->sk->sk_sndbuf = max_t(int, socket->sk->sk_sndbuf, + COREDUMP_SOCK_SNDBUF_MIN); + file = sock_alloc_file(socket, 0, NULL); if (IS_ERR(file)) return false; -- 2.53.0
