kstat_hold() blocks forever

[lundman]

I had a deadlock with zpool import -d c:/src/diskimage.bin funk with the stack:

 	nt!KeWaitForMultipleObjects+0x287	C/C++/ASM
 	OpenZFS!spl_cv_wait+0x104 [C:\src\openzfs\module\os\windows\spl\spl-condvar.c @ 123]	C/C++/ASM
 	OpenZFS!kstat_hold+0xaf [C:\src\openzfs\module\os\windows\spl\spl-kstat.c @ 610]	C/C++/ASM
 	OpenZFS!kstat_hold_bykid+0x54 [C:\src\openzfs\module\os\windows\spl\spl-kstat.c @ 646]	C/C++/ASM
 	OpenZFS!kstat_delete+0xd2 [C:\src\openzfs\module\os\windows\spl\spl-kstat.c @ 1295]	C/C++/ASM
 	OpenZFS!procfs_list_destroy+0x5e [C:\src\openzfs\module\os\windows\spl\spl-proc_list.c @ 141]	C/C++/ASM
 	OpenZFS!spa_read_history_destroy+0x4b [C:\src\openzfs\module\zfs\spa_stats.c @ 146]	C/C++/ASM
 	OpenZFS!spa_stats_destroy+0x3b [C:\src\openzfs\module\zfs\spa_stats.c @ 1013]	C/C++/ASM
 	OpenZFS!spa_remove+0x3c3 [C:\src\openzfs\module\zfs\spa_misc.c @ 816]	C/C++/ASM
 	OpenZFS!spa_tryimport+0x5f7 [C:\src\openzfs\module\zfs\spa.c @ 6222]	C/C++/ASM
 	OpenZFS!zfs_ioc_pool_tryimport+0x76 [C:\src\openzfs\module\zfs\zfs_ioctl.c @ 1646]	C/C++/ASM

Where we are stuck here:

kstat_hold(avl_tree_t *t, ekstat_t *template)
{
	kstat_t *ksp;
	ekstat_t *e;

	mutex_enter(&kstat_chain_lock);
	for (;;) {
		ksp = avl_find(t, template, NULL);
		if (ksp == NULL)
			break;
		e = (ekstat_t *)ksp;
		if (e->e_owner == NULL) {
			e->e_owner = (void *)curthread;
			break;
		}
		cv_wait(&e->e_cv, &kstat_chain_lock);
	}

e->e_owner is the value of curthread already, so it is waiting for ourselves to wake us up. This could suggest a leak in kstat_hold somewhere,. perhaps from procfs_list_destroy - code we got from FreeBSD. Check and see if they have fixed anything in this area.

[datacore-rm]

Later commits done in spl-proc_list.c/procfs_list_destroy() could resolve this issue?

[lundman]

I think it is a bit racey, between mutex_exit and destroy. I've not come across it in quite a while, so hoping other things fixed it. Have you come across it?

[datacore-rm]

I have not come across this issue anytime. Thanks.

[datacore-rm]

In kstat_install() and kstat_delete() there is below check where there is no corresponding kstat_release(). Can there be any
chance that kstat_hold_bykid() returns some another not null ksp?

if (kstat_hold_bykid(ksp->ks_kid, zoneid) != ksp) {
cmn_err(CE_WARN, "kstat_delete(%p): does not exist",
(void *)ksp);
return;
}