[Devel] [PATCH RFC VZ10 2/2] ve/sunrpc: Implement a per-ve sunrpc killer

[Vladimir Riabchun]

We currently have  /proc/<pid>/net/rpc/kill-tasks feature for
aborting pending RPC tasks. To fast-stop container, we need
to abort all RPC tasks, so we must iterate over all VE PIDs.

There are several problems with this approach:
1. In real life with cgroups-v2 we just process container
   init PID, all other network namespaces are ignored.
2. There is a risk of breaking unrelated NFS connections if a PID,
   that was a container process, is reused while we are
   stopping container.
3. This may be slow - container might have a lot of processes
   and a few network namespaces.

To fix this all, create a per-ve interface for aborting
RPC requests in all VE network namespaces.

https://virtuozzo.atlassian.net/browse/VSTOR-126316

Feature: improve kill-tasks
Signed-off-by: Vladimir Riabchun <vladimir.riabchun at virtuozzo.com>
---
 include/linux/sunrpc/clnt.h |  2 ++
 include/linux/ve.h          |  4 ++++
 kernel/ve/ve.c              | 44 +++++++++++++++++++++++++++++++++++++
 net/sunrpc/clnt.c           |  3 ++-
 net/sunrpc/sunrpc_syms.c    |  3 +++
 5 files changed, 55 insertions(+), 1 deletion(-)

diff --git a/include/linux/sunrpc/clnt.h b/include/linux/sunrpc/clnt.h
index 0133aeba248a..0e7c7c9107a2 100644
--- a/include/linux/sunrpc/clnt.h
+++ b/include/linux/sunrpc/clnt.h
@@ -253,6 +253,8 @@ void		rpc_clnt_probe_trunked_xprts(struct rpc_clnt *,
 
 const char *rpc_proc_name(const struct rpc_task *task);
 
+
+void rpc_kill_tasks(struct net *net);
 int rpc_task_kill_proc_init(struct net *net);
 void rpc_task_kill_proc_fini(struct net *net);
 
diff --git a/include/linux/ve.h b/include/linux/ve.h
index 224acf012821..95a83c7bc7de 100644
--- a/include/linux/ve.h
+++ b/include/linux/ve.h
@@ -279,6 +279,8 @@ extern bool is_ve_init_net(const struct net *net);
 
 void ve_setup_task(struct task_struct *p, struct ve_struct *ve);
 
+void ve_set_rpc_kill_fn(void (*fn)(struct net *));
+
 #else	/* CONFIG_VE */
 #include <linux/init_task.h>
 #define get_ve(ve)	((void)(ve), NULL)
@@ -336,6 +338,8 @@ static inline int vz_security_protocol_check(struct net *net, int protocol) { re
 
 static inline void ve_setup_task(struct task_struct *p, struct ve_struct *ve) { }
 
+static inline void ve_set_rpc_kill_fn(void (*fn)(struct net *)) { }
+
 #endif	/* CONFIG_VE */
 
 struct seq_file;
diff --git a/kernel/ve/ve.c b/kernel/ve/ve.c
index 198c82f010cc..856d7afb9d1c 100644
--- a/kernel/ve/ve.c
+++ b/kernel/ve/ve.c
@@ -113,6 +113,8 @@ EXPORT_SYMBOL(nr_ve);
 
 static DEFINE_IDR(ve_idr);
 
+static void (*rpc_kill_net_fn)(struct net *);
+
 struct ve_struct *get_ve(struct ve_struct *ve)
 {
 	if (ve)
@@ -1713,6 +1715,43 @@ static ssize_t ve_write_ctty(struct kernfs_open_file *of, char *buf,
 	return ret;
 }
 
+void ve_set_rpc_kill_fn(void (*fn)(struct net *))
+{
+	WRITE_ONCE(rpc_kill_net_fn, fn);
+	/* ve_rpc_kill_write might be using old function.
+	 * rpc_kill_tasks -> NULL is dangerous, block sunrpc exit while we use
+	 * functions from this module
+	 */
+	synchronize_rcu();
+}
+EXPORT_SYMBOL(ve_set_rpc_kill_fn);
+
+static int ve_rpc_kill_write(struct cgroup_subsys_state *css,
+				 struct cftype *cft, u64 val)
+{
+	struct net *net;
+	struct ve_struct *ve = css_to_ve(css);
+	void (*fn)(struct net *net);
+
+	guard(rwsem_read)(&net_rwsem);	/* for_each_net protection */
+	guard(rcu)();			/* Begin rpc_kill_net_fn usage section */
+
+	fn = READ_ONCE(rpc_kill_net_fn);
+	if (!fn) {
+		pr_info_ratelimited("SUNRPC module is not loaded.\n");
+		return 0;
+	}
+
+	for_each_net(net) {
+		if (net->owner_ve != ve)
+			continue;
+		/* rpc_kill_tasks is atomic. */
+		fn(net);
+	}
+
+	return 0;
+}
+
 static struct cftype ve_cftypes[] = {
 
 	{
@@ -1808,6 +1847,11 @@ static struct cftype ve_cftypes[] = {
 		.flags			= CFTYPE_ONLY_ON_ROOT,
 		.write			= ve_write_ctty,
 	},
+	{
+		.name			= "rpc_kill",
+		.flags			= CFTYPE_NOT_ON_ROOT,
+		.write_u64		= ve_rpc_kill_write,
+	},
 	{ }
 };
 
diff --git a/net/sunrpc/clnt.c b/net/sunrpc/clnt.c
index 10ef5d9c2696..d40d59145bec 100644
--- a/net/sunrpc/clnt.c
+++ b/net/sunrpc/clnt.c
@@ -3406,8 +3406,9 @@ rpc_clnt_swap_deactivate(struct rpc_clnt *clnt)
 EXPORT_SYMBOL_GPL(rpc_clnt_swap_deactivate);
 #endif /* CONFIG_SUNRPC_SWAP */
 
-static void rpc_kill_tasks(struct net *net)
+void rpc_kill_tasks(struct net *net)
 {
+	/* Note: function must be atomic, used under RCU read-lock. */
 	struct rpc_clnt *clnt;
 	struct sunrpc_net *sn = net_generic(net, sunrpc_net_id);
 
diff --git a/net/sunrpc/sunrpc_syms.c b/net/sunrpc/sunrpc_syms.c
index bab6cab29405..0cb280569135 100644
--- a/net/sunrpc/sunrpc_syms.c
+++ b/net/sunrpc/sunrpc_syms.c
@@ -13,6 +13,7 @@
 #include <linux/uio.h>
 #include <linux/unistd.h>
 #include <linux/init.h>
+#include <linux/ve.h>
 
 #include <linux/sunrpc/sched.h>
 #include <linux/sunrpc/clnt.h>
@@ -114,6 +115,7 @@ init_sunrpc(void)
 #endif
 	svc_init_xprt_sock();	/* svc sock transport */
 	init_socket_xprt();	/* clnt sock transport */
+	ve_set_rpc_kill_fn(rpc_kill_tasks);
 	return 0;
 
 out5:
@@ -131,6 +133,7 @@ init_sunrpc(void)
 static void __exit
 cleanup_sunrpc(void)
 {
+	ve_set_rpc_kill_fn(NULL);
 	rpc_sysfs_exit();
 	rpc_cleanup_clids();
 	xprt_cleanup_ids();
-- 
2.47.1