Re: [-next] system hangs likely due to "modules: Only return -EEXIST for modules that have finished loading"

[Heiko Carstens]

On Fri, Apr 26, 2019 at 08:20:52PM -0400, Prarit Bhargava wrote:
> Heiko and Jessica,
> 
> The issue doesn't appear to be with my patch AFAICT.  The s390_trng fails to
> load and then the kernel occasionally hangs (as Heiko mentioned) calling
> synchronize_rcu().
> 
> The call sequence is
> 
> module_load()
> 	do_init_module()
> 		do_one_initcall(mod->init)
> 
> which fails.
> 
> The failure path in do_one_initcall() is entered and we start executing code at
> kernel/module.c:3541
> 
> fail_free_freeinit:
>         kfree(freeinit);
> fail:
>         /* Try to protect us from buggy refcounters. */
>         mod->state = MODULE_STATE_GOING;
>         synchronize_rcu();
> 
> ^^^ the kernel hangs here.  Sometimes it's very short and other times it seems
> to hang.   I've left systems that appear to be hung and come back after 10
> minutes to find that they've somehow made it through this call.
> 
> Is there a known issue with RCU on s390 that is making this occur?

No there is no known issue with RCU on s390. The reason that
synchronize_rcu() doesn't finish is because a different cpu is within
an endless loop in add_unformed_module() just like Jessica suspected.

Note: the kernel is compiled with CONFIG_PREEMPT off - there is no
kernel preemption that will make the looping cpu ever go over schedule
and subsequently let synchronize_rcu() finish.

To confirm Jessicas theory - looking into the dump we have:

crash> bt 742
PID: 742    TASK: 1efa6c000         CPU: 7   COMMAND: "systemd-udevd"
 #0 [3e0043aba30] __schedule at abb25e
 #1 [3e0043abaa0] schedule at abb6a2
 #2 [3e0043abac8] schedule_timeout at abf49a
 #3 [3e0043abb60] wait_for_common at abc396
 #4 [3e0043abbf0] __wait_rcu_gp at 1c0136
 #5 [3e0043abc48] synchronize_rcu at 1c72ea
 #6 [3e0043abc98] do_init_module at 1f10be
 #7 [3e0043abcf0] load_module at 1f3594
 #8 [3e0043abdd0] __se_sys_init_module at 1f3af0
 #9 [3e0043abea8] system_call at ac0766

Which is the process waiting for synchronize_rcu to finish. Wading
through the stack frames gives me this struct module:

struct module {
  state = MODULE_STATE_GOING,
  list = {
    next = 0x3ff80394508,
    prev = 0xe25090 <modules>
  },
  name = "s390_trng\000...
...

Then we have the looping task/cpu:

PID: 731    TASK: 1e79ba000         CPU: 7   COMMAND: "systemd-udevd"
 LOWCORE INFO:
  -psw      : 0x0704c00180000000 0x0000000000ab666a
  -function : memcmp at ab666a
...
  -general registers:
     0x0000000000000009 0x0000000000000009
     0x000003ff80347321 000000000000000000
     0x000003ff8034f321 000000000000000000
     0x000000000000001e 0x000003ff8c592708
     0x000003e0047da900 0x000003ff8034f318
     0x0000000000000001 0x0000000000000009
     0x000003ff80347300 0x0000000000ad81b8
     0x00000000001ee062 0x000003e004357cb0
 #0 [3e004357cf0] load_module at 1f1eb0
 #1 [3e004357dd0] __se_sys_init_module at 1f3af0
 #2 [3e004357ea8] system_call at ac0766

which is find_module_all() calling memcmp with this string:

3ff80347318:  733339305f74726e 6700000000000000   s390_trng.......

So it all seems to fit. A simple cond_resched() call, which enforces
an RCU quiescent state for the calliung cpu, fixes the problem for me
(patch on top of linux-next 20190424 -- c392798a85ab):

diff --git a/kernel/module.c b/kernel/module.c
index 410eeb7e4f1d..48748cfec991 100644
--- a/kernel/module.c
+++ b/kernel/module.c
@@ -3585,6 +3585,7 @@ again:
 					       finished_loading(mod->name));
 			if (err)
 				goto out_unlocked;
+			cond_resched();
 			goto again;
 		}
 		err = -EEXIST;