lkml.org 
[lkml]   [2014]   [Dec]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [RFC 2/2] OOM, PM: make OOM detection in the freezer path raceless
On Tue 02-12-14 17:08:04, Tejun Heo wrote:
> Hello, sorry about the delay. Was on vacation.
>
> Generally looks good to me. Some comments below.
>
> > @@ -355,8 +355,10 @@ static struct sysrq_key_op sysrq_term_op = {
> >
> > static void moom_callback(struct work_struct *ignored)
> > {
> > - out_of_memory(node_zonelist(first_memory_node, GFP_KERNEL), GFP_KERNEL,
> > - 0, NULL, true);
> > + if (!out_of_memory(node_zonelist(first_memory_node, GFP_KERNEL),
> > + GFP_KERNEL, 0, NULL, true)) {
> > + printk(KERN_INFO "OOM request ignored because killer is disabled\n");
> > + }
> > }
>
> CodingStyle line 157 says "Do not unnecessarily use braces where a
> single statement will do.".

Sure. Fixed

> > +/**
> > + * oom_killer_disable - disable OOM killer
> > + *
> > + * Forces all page allocations to fail rather than trigger OOM killer.
> > + * Will block and wait until all OOM victims are dead.
> > + *
> > + * Returns true if successfull and false if the OOM killer cannot be
> > + * disabled.
> > + */
> > +extern bool oom_killer_disable(void);
>
> And function comments usually go where the function body is, not where
> the function is declared, no?

Fixed

> > @@ -157,27 +132,11 @@ int freeze_processes(void)
> > pm_wakeup_clear();
> > printk("Freezing user space processes ... ");
> > pm_freezing = true;
> > - oom_kills_saved = oom_kills_count();
> > error = try_to_freeze_tasks(true);
> > if (!error) {
> > __usermodehelper_set_disable_depth(UMH_DISABLED);
> > - oom_killer_disable();
> > -
> > - /*
> > - * There might have been an OOM kill while we were
> > - * freezing tasks and the killed task might be still
> > - * on the way out so we have to double check for race.
> > - */
> > - if (oom_kills_count() != oom_kills_saved &&
> > - !check_frozen_processes()) {
> > - __usermodehelper_set_disable_depth(UMH_ENABLED);
> > - printk("OOM in progress.");
> > - error = -EBUSY;
> > - } else {
> > - printk("done.");
> > - }
> > + printk("done.\n");
>
> A delta but shouldn't it be pr_cont()?

kernel/power/process.c doesn't use pr_* so I've stayed with what the
rest of the file is using. I can add a patch which transforms all of
them.

> ...
> > @@ -206,6 +165,18 @@ int freeze_kernel_threads(void)
> > printk("\n");
> > BUG_ON(in_atomic());
> >
> > + /*
> > + * Now that everything freezable is handled we need to disbale
> > + * the OOM killer to disallow any further interference with
> > + * killable tasks.
> > + */
> > + printk("Disabling OOM killer ... ");
> > + if (!oom_killer_disable()) {
> > + printk("failed.\n");
> > + error = -EAGAIN;
> > + } else
> > + printk("done.\n");
>
> Ditto on pr_cont() and
>
> CodingStyle line 169 says "This does not apply
> if only one branch of a conditional statement is a single statement;
> in the latter case use braces in both branches:"

Fixed

> > @@ -251,6 +220,9 @@ void thaw_kernel_threads(void)
> > {
> > struct task_struct *g, *p;
> >
> > + printk("Enabling OOM killer again.\n");
>
> Do we really need this printk? The same goes for Disabling OOM
> killer. For freezing it makes some sense because freezing may take a
> considerable amount of time and even occassionally fail due to
> timeout. We aren't really expecting those to happen for OOM victims.

I just considered them useful if there are follow up allocation failure
messages to know that they are due to OOM killer.
I can remove them.

> > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > index 302e0fc6d121..34bcbb053132 100644
> > --- a/mm/memcontrol.c
> > +++ b/mm/memcontrol.c
> > @@ -2128,6 +2128,8 @@ static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
> > current->memcg_oom.order = order;
> > }
> >
> > +extern bool oom_killer_disabled;
>
> Ugh... don't we wanna put this in a header file?

Who else would need the declaration? This is not something random code
should look at.

> > +void mark_tsk_oom_victim(struct task_struct *tsk)
> > {
> > - return atomic_read(&oom_kills);
> > + BUG_ON(oom_killer_disabled);
>
> WARN_ON_ONCE() is prolly a better option here?

Well, something fishy is going on when oom_killer_disabled is set and we
mark new OOM victim. This is a clear bug. Why would be warning and a
allow the follow up breakage?

> > + if (test_and_set_tsk_thread_flag(tsk, TIF_MEMDIE))
>
> Can a task actually be selected as an OOM victim multiple times?

AFAICS nothing prevents from global OOM and memcg OOM killers racing.

> > + return;
> > + atomic_inc(&oom_victims);
> > }
> >
> > -void note_oom_kill(void)
> > +void unmark_tsk_oom_victim(struct task_struct *tsk)
> > {
> > - atomic_inc(&oom_kills);
> > + int count;
> > +
> > + if (!test_and_clear_tsk_thread_flag(tsk, TIF_MEMDIE))
> > + return;
>
> Maybe test this inline in exit_mm()? e.g.
>
> if (test_thread_flag(TIF_MEMDIE))
> unmark_tsk_oom_victim(current);

Why do you think testing TIF_MEMDIE in exit_mm is better? I would like
to reduce the usage of the flag as much as possible.

> Also, can the function ever be called by someone other than current?
> If not, why would it take @task?

Changed to use current only. If there is anybody who needs that we can
change that later. I wanted to have it symmetric to mark_tsk_oom_victim
but that is not that important.

> > +
> > + down_read(&oom_sem);
> > + /*
> > + * There is no need to signal the lasst oom_victim if there
> > + * is nobody who cares.
> > + */
> > + if (!atomic_dec_return(&oom_victims) && oom_killer_disabled)
> > + complete(&oom_victims_wait);
>
> I don't think using completion this way is safe. Please read on.
>
> > + up_read(&oom_sem);
> > }
> >
> > -void mark_tsk_oom_victim(struct task_struct *tsk)
> > +bool oom_killer_disable(void)
> > {
> > - set_tsk_thread_flag(tsk, TIF_MEMDIE);
> > + /*
> > + * Make sure to not race with an ongoing OOM killer
> > + * and that the current is not the victim.
> > + */
> > + down_write(&oom_sem);
> > + if (!test_tsk_thread_flag(current, TIF_MEMDIE))
> > + oom_killer_disabled = true;
>
> Prolly "if (TIF_MEMDIE) { unlock; return; }" is easier to follow.

OK

> > +
> > + count = atomic_read(&oom_victims);
> > + up_write(&oom_sem);
> > +
> > + if (count && oom_killer_disabled)
> > + wait_for_completion(&oom_victims_wait);
>
> So, each complete() increments the done count and wait decs. The
> above code works iff the complete()'s and wait()'s are always balanced
> which usually isn't true in this type of wait code. Either use
> reinit_completion() / complete_all() combos or wait_event().

Hmm, I thought that only a single instance of freeze_kernel_threads
(which calls oom_killer_disable) can run at a time. But I am currently
not sure that all paths are called under lock_system_sleep.
I am not familiar with reinit_completion API. Is the following correct?
[...]
@@ -434,10 +434,23 @@ void unmark_tsk_oom_victim(struct task_struct *tsk)
* is nobody who cares.
*/
if (!atomic_dec_return(&oom_victims) && oom_killer_disabled)
- complete(&oom_victims_wait);
+ complete_all(&oom_victims_wait);
up_read(&oom_sem);
}
[...]
@@ -445,16 +458,23 @@ bool oom_killer_disable(void)
* and that the current is not the victim.
*/
down_write(&oom_sem);
- if (!test_tsk_thread_flag(current, TIF_MEMDIE))
- oom_killer_disabled = true;
+ if (test_thread_flag(TIF_MEMDIE)) {
+ up_write(&oom_sem);
+ return false;
+ }
+
+ /* unmark_tsk_oom_victim is calling complete_all */
+ if (!oom_killer_disable)
+ reinit_completion(&oom_victims_wait);

+ oom_killer_disabled = true;
count = atomic_read(&oom_victims);
up_write(&oom_sem);

- if (count && oom_killer_disabled)
+ if (count)
wait_for_completion(&oom_victims_wait);

- return oom_killer_disabled;
+ return true;
}

> > +
> > + return oom_killer_disabled;
>
> Maybe 0 / -errno is better choice as return values?

I do not have problem to change this if you feel strong about it but
true/false sounds easier to me and it allows the caller to decide what to
report. If there were multiple reasons to fail then sure but that is not
the case.
> > +/** out_of_memory - tries to invoke OOM killer.
>
> Formatting?

fixed

> > + * @zonelist: zonelist pointer
> > + * @gfp_mask: memory allocation flags
> > + * @order: amount of memory being requested as a power of 2
> > + * @nodemask: nodemask passed to page allocator
> > + * @force_kill: true if a task must be killed, even if others are exiting
> > + *
> > + * invokes __out_of_memory if the OOM is not disabled by oom_killer_disable()
> > + * when it returns false. Otherwise returns true.
> > + */
> > +bool out_of_memory(struct zonelist *zonelist, gfp_t gfp_mask,
> > + int order, nodemask_t *nodemask, bool force_kill)
> > +{
> > + bool ret = false;
> > +
> > + down_read(&oom_sem);
> > + if (!oom_killer_disabled) {
> > + __out_of_memory(zonelist, gfp_mask, order, nodemask, force_kill);
> > + ret = true;
> > + }
> > + up_read(&oom_sem);
> > +
> > + return ret;
>
> Ditto on return value. 0 / -EBUSY seem like a better choice to me.
>
> > @@ -712,12 +770,16 @@ void pagefault_out_of_memory(void)
> > {
> > struct zonelist *zonelist;
> >
> > + down_read(&oom_sem);
> > if (mem_cgroup_oom_synchronize(true))
> > - return;
> > + goto unlock;
> >
> > zonelist = node_zonelist(first_memory_node, GFP_KERNEL);
> > if (oom_zonelist_trylock(zonelist, GFP_KERNEL)) {
> > - out_of_memory(NULL, 0, 0, NULL, false);
> > + if (!oom_killer_disabled)
> > + __out_of_memory(NULL, 0, 0, NULL, false);
> > oom_zonelist_unlock(zonelist, GFP_KERNEL);
>
> Is this a condition which can happen and we can deal with? With
> userland fully frozen, there shouldn't be page faults which lead to
> memory allocation, right?

Except for racing OOM victims which were missed by try_to_freeze_tasks
because they didn't get cpu slice to wake up from the freezer. The task
would die on the way out from the page fault exception. I have updated
the changelog to be more verbose about this.

> Shouldn't we document how oom disable/enable is supposed to be used

Well the API shouldn't be used outside of the PM freezer IMO. This is not a
general API that other part of the kernel should be using. I can surely
add more documentation for the PM usage though. I have rewritten the
changelog:
"
As oom_killer_disable() is a full OOM barrier now we can postpone it in
the PM freezer to later after all freezable user tasks are considered
frozen (to freeze_kernel_threads).

Normally there wouldn't be any unfrozen user tasks at this moment so
the function will not block. But if there was an OOM killer racing with
try_to_freeze_tasks and the OOM victim didn't finish yet then we have to
wait for it. This should complete in a finite time, though, because
- the victim cannot loop in the page fault handler (it would die
on the way out from the exception)
- it cannot loop in the page allocator because all the further\
allocation would fail
- it shouldn't be blocked on any locks held by frozen tasks
(try_to_freeze expects lockless context) and kernel threads and
work queues are not frozen yet
"

And I've added:
+/**
+ * oom_killer_disable - disable OOM killer
+ *
+ * Forces all page allocations to fail rather than trigger OOM killer.
+ * Will block and wait until all OOM victims are dead.
+ *
+ * The function cannot be called when there are runnable user tasks because
+ * the userspace would see unexpected allocation failures as a result. Any
+ * new usage of this function should be consulted with MM people.
+ *
+ * Returns true if successful and false if the OOM killer cannot be
+ * disabled.
+ */
bool oom_killer_disable(void)
> (it only makes sense while the whole system is in quiescent state)
> and at least trigger WARN_ON_ONCE() if the above code path gets
> triggered while oom killer is disabled?

I can add a WARN_ON(!test_thread_flag(tsk, TIF_MEMDIE)).

Thanks for the review!
--
Michal Hocko
SUSE Labs


\
 
 \ /
  Last update: 2014-12-04 16:01    [W:0.239 / U:0.604 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site