[CRIU] [PATCH v4 3/3] early_init: Raise nr_open limit on particular commands

[Cyrill Gorcunov]

For swrk, service, dump, restore we need to to raise
nr_file limit to be able to process containers with
huge number of files opened.

Signed-off-by: Cyrill Gorcunov <gorcunov at gmail.com>
---
v4: raise limits for restore as well

 criu/cr-restore.c | 32 ----------------------
 criu/crtools.c    | 69 +++++++++++++++++++++++++++++++++++++++++++++--
 2 files changed, 67 insertions(+), 34 deletions(-)

diff --git a/criu/cr-restore.c b/criu/cr-restore.c
index af2ca2921d00..a6cb8f1b3e07 100644
--- a/criu/cr-restore.c
+++ b/criu/cr-restore.c
@@ -2327,42 +2327,10 @@ int prepare_dummy_task_state(struct pstree_item *pi)
 	return 0;
 }
 
-static void rlimit_unlimit_nofile_self(void)
-{
-	struct rlimit new;
-
-	new.rlim_cur = kdat.sysctl_nr_open;
-	new.rlim_max = kdat.sysctl_nr_open;
-
-	if (prlimit(getpid(), RLIMIT_NOFILE, &new, NULL)) {
-		pr_perror("rlimir: Can't setup RLIMIT_NOFILE for self");
-		return;
-	} else
-		pr_debug("rlimit: RLIMIT_NOFILE unlimited for self\n");
-	service_fd_rlim_cur = kdat.sysctl_nr_open;
-}
-
 int cr_restore_tasks(void)
 {
 	int ret = -1;
 
-	/*
-	 * Service fd engine implies that file descriptors
-	 * used won't be borrowed by the rest of the code
-	 * and default 1024 limit is not enough for high
-	 * loaded test/containers. Thus use kdat engine
-	 * to fetch current system level limit for numbers
-	 * of files allowed to open up and lift up own
-	 * limits.
-	 *
-	 * Note we have to do it before the service fd
-	 * get inited and we dont exit with errors here
-	 * because in worst scenario where clash of fd
-	 * happen we simply exit with explicit error
-	 * during real action stage.
-	 */
-	rlimit_unlimit_nofile_self();
-
 	if (cr_plugin_init(CR_PLUGIN_STAGE__RESTORE))
 		return -1;
 
diff --git a/criu/crtools.c b/criu/crtools.c
index 3f64000c1e87..95d221b483e8 100644
--- a/criu/crtools.c
+++ b/criu/crtools.c
@@ -51,8 +51,73 @@
 #include "sysctl.h"
 #include "img-remote.h"
 
-static int early_init(void)
+static void rlimit_unlimit_nofile(void)
 {
+	struct rlimit new;
+
+	new.rlim_cur = kdat.sysctl_nr_open;
+	new.rlim_max = kdat.sysctl_nr_open;
+
+	if (prlimit(getpid(), RLIMIT_NOFILE, &new, NULL)) {
+		pr_perror("rlimit: Can't setup RLIMIT_NOFILE for self");
+		return;
+	} else
+		pr_debug("rlimit: RLIMIT_NOFILE unlimited for self\n");
+
+	service_fd_rlim_cur = kdat.sysctl_nr_open;
+}
+
+static int early_init(const char *cmd)
+{
+	static const char *nofile_cmds[] = {
+		"swrk", "service",
+		"dump", "pre-dump",
+		"restore",
+	};
+	size_t i;
+
+	/*
+	 * Service fd engine implies that file descriptors
+	 * used won't be borrowed by the rest of the code
+	 * and default 1024 limit is not enough for high
+	 * loaded test/containers. Thus use kdat engine
+	 * to fetch current system level limit for numbers
+	 * of files allowed to open up and lift up own
+	 * limits.
+	 *
+	 * Note we have to do it before the service fd
+	 * get inited and we dont exit with errors here
+	 * because in worst scenario where clash of fd
+	 * happen we simply exit with explicit error
+	 * during real action stage.
+	 *
+	 * Same time raising limits cause kernel fdtable
+	 * to bloat so we do this only on the @nofile_cmds:
+	 *
+	 *  - on dump criu needs additional files for sfd,
+	 *    thus if container already has many files opened
+	 *    we need to have at least not less space when
+	 *    fetching fds from a target process;
+	 *
+	 *  - on pre-dump we might need a lot of pipes to
+	 *    fetch huge number of pages to dump;
+	 *
+	 *  - on restore we still need to raise limits since
+	 *    there is no guarantee that on dump we've not
+	 *    been hitting fd limit already;
+	 *
+	 *  - swrk and service obtain requests on the fly,
+	 *    thus we don't know if on of above will be
+	 *    there thus raise limits.
+	 */
+	for (i = 0; i < ARRAY_SIZE(nofile_cmds); i++) {
+		if (strcmp(nofile_cmds[i], cmd))
+			continue;
+		if (!kerndat_files_stat(true))
+			rlimit_unlimit_nofile();
+		break;
+	}
+
 	if (init_service_fd())
 		return 1;
 
@@ -94,7 +159,7 @@ int main(int argc, char *argv[], char *envp[])
 
 	log_set_loglevel(opts.log_level);
 
-	if (early_init())
+	if (early_init(argv[optind]))
 		return -1;
 
 	if (!strcmp(argv[1], "swrk")) {
-- 
2.20.1