[LU-15207] ASSERTION( !cfs_hash_is_rehashing(hs)

• Type: Bug
• Resolution: Fixed
• Priority: Minor
• Fix Version/s: Lustre 2.16.0, Lustre 2.15.5
• Affects Version/s: Upstream
• Labels: LTS15
• Severity: 3
• Assignee: Alex Zhuravlev
• Reporter: Alex Zhuravlev

Description

cfs_hash_rehash(struct cfs_hash *hs, int do_rehash)
{
...
	hs->hs_rehash_bits = rc;
	if (!do_rehash) {
		/* launch and return */
		queue_work(cfs_rehash_wq, &hs->hs_rehash_work);
		cfs_hash_unlock(hs, 1);
		return;
}

cfs_hash_rehash_cancel(struct cfs_hash *hs)
{
	LASSERT(cfs_hash_with_rehash(hs));
	cancel_work_sync(&hs->hs_rehash_work);
}

cfs_hash_for_each_enter(struct cfs_hash *hs)
{
...
	if (cfs_hash_is_rehashing(hs))
		cfs_hash_rehash_cancel(hs);
}

if we enter iteration (cfs_hash_for_each_enter()) and find rehashing scheduled or in progress (hs_rehash_bits != 0), then we want to cancel that work using cancel_work_sync().
but if the work hasn't started yet then who would reset hs_rehash_bits back to 0?

a trivial test demonstrates this:

static struct workqueue_struct *test_rehash_wq;
struct work_struct test_rehash_work;
static volatile int test_var = 0;
static void test_worker(struct work_struct *work)
{
       test_var = 0;
}
void work_test(void)
{
       test_rehash_wq = alloc_workqueue("test", WQ_SYSFS, 4);
       INIT_WORK(&test_rehash_work, test_worker);
       test_var = 1;
       queue_work(test_rehash_wq, &test_rehash_work);
       cancel_work_sync(&test_rehash_work);
       LASSERT(test_var == 0);
       destroy_workqueue(test_rehash_wq);
}