The test case below fails on 5.17 (and a bunch of older kernels)
with unshare getting -ENOSPC, because the rate at which ipc_namespace
structures can be freed is limited by each item on the to-free list
waiting in synchronize_rcu.
Making kern_unmount use queue_rcu_work gets rid of that slowdown,
allowing a batch of vfsmount structures to be freed after each
RCU grace period has expired.
That, in turn, allows us to just get rid of the workqueue in
ipc/namespace.c completely.
With these two changes the test case reliably succeeds at
calling unshare a million times, even with max_ipc_namespaces
reduced to 1000 :)
#define _GNU_SOURCE
#include <sched.h>
#include <error.h>
#include <errno.h>
#include <stdlib.h>
int main()
{
int i;
for (i = 0; i < 1000000; i++) {
if (unshare(CLONE_NEWIPC) < 0)
error(EXIT_FAILURE, errno, "unshare");
}
}
Rik van Riel (2):
vfs: free vfsmount through rcu work from kern_unmount
ipc: get rid of free_ipc_work workqueue
fs/namespace.c | 11 +++++++++--
include/linux/ipc_namespace.h | 2 --
include/linux/mount.h | 2 ++
ipc/namespace.c | 21 +--------------------
4 files changed, 12 insertions(+), 24 deletions(-)
--
2.34.1
