On 2023/06/26 20:35, Michal Hocko wrote:
> On Mon 26-06-23 20:26:02, Tetsuo Handa wrote:
>> On 2023/06/26 19:48, Peter Zijlstra wrote:
>>> On Mon, Jun 26, 2023 at 06:25:56PM +0900, Tetsuo Handa wrote:
>>>> On 2023/06/26 17:12, Sebastian Andrzej Siewior wrote:
>>>>> On 2023-06-24 15:54:12 [+0900], Tetsuo Handa wrote:
>>>>>> Why not to do the same on the end side?
>>>>>>
>>>>>> static inline void do_write_seqcount_end(seqcount_t *s)
>>>>>> {
>>>>>> - seqcount_release(&s->dep_map, _RET_IP_);
>>>>>> do_raw_write_seqcount_end(s);
>>>>>> + seqcount_release(&s->dep_map, _RET_IP_);
>>>>>> }
>>>>>
>>>>> I don't have a compelling argument for doing it. It is probably better
>>>>> to release the lock from lockdep's point of view and then really release
>>>>> it (so it can't be acquired before it is released).
>>>>
>>>> We must do it because this is a source of possible printk() deadlock.
>>>> Otherwise, I will nack on PATCH 2/2.
>>>
>>> Don't be like that... just hate on prink like the rest of us. In fact,
>>> i've been patching out the actual printk code for years because its
>>> unusable garbage.
>>>
>>> Will this actually still be a problem once all the fancy printk stuff
>>> lands? That shouldn't do synchronous prints except to 'atomic' consoles
>>> by default IIRC.
>>
>> Commit 1007843a9190 ("mm/page_alloc: fix potential deadlock on zonelist_update_seq
>> seqlock") was applied to 4.14-stable trees, and CONFIG_PREEMPT_RT is available
>> since 5.3. Thus, we want a fix which can be applied to 5.4-stable and later.
>> This means that we can't count on all the fancy printk stuff being available.
>
> Is there any reason to backport RT specific fixup to stable trees? I
> mean seriously, is there any actual memory hotplug user using
> PREEMPT_RT? I would be more than curious to hear the usecase.
Even if we don't backport RT specific fixup to stable trees, [PATCH 2/2] requires
that [PATCH 1/2] guarantees that synchronous printk() never happens (for whatever
reasons) between write_seqlock_irqsave(&zonelist_update_seq, flags) and
write_sequnlock_irqrestore(&zonelist_update_seq, flags).
If [PATCH 1/2] cannot guarantee it, [PATCH 2/2] will be automatically rejected.
If [PATCH 2/2] cannot be applied, we have several alternatives.
Alternative 1:
Revert both commit 3d36424b3b58 ("mm/page_alloc: fix race condition between build_all_zonelists and page allocation")
and commit 1007843a9190 ("mm/page_alloc: fix potential deadlock on zonelist_update_seq seqlock").
I don't think this will happen, for nobody will be happy.
Alternative 2:
Revert commit 1007843a9190 ("mm/page_alloc: fix potential deadlock on zonelist_update_seq seqlock")
and apply "mm/page_alloc: don't check zonelist_update_seq from atomic allocations" at
https://lkml.kernel.org/r/dfdb9da6-ca8f-7a81-bfdd-d74b4c401f11@xxxxxxxxxxxxxxxxxxx .
I think this is reasonable, for this reduces locking dependency. But Michal Hocko did not like it.
Alternative 3:
Somehow preserve printk_deferred_enter() => write_seqlock(&zonelist_update_seq) and
write_sequnlock(&zonelist_update_seq) => printk_deferred_exit() pattern. Something like below?
----------------------------------------
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 47421bedc12b..ded3ac3856e7 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -5805,6 +5805,7 @@ static void __build_all_zonelists(void *data)
int nid;
int __maybe_unused cpu;
pg_data_t *self = data;
+#ifndef CONFIG_PREEMPT_RT
unsigned long flags;
/*
@@ -5813,6 +5814,9 @@ static void __build_all_zonelists(void *data)
* (e.g. GFP_ATOMIC) that could hit zonelist_iter_begin and livelock.
*/
local_irq_save(flags);
+#else
+ migrate_disable();
+#endif
/*
* Explicitly disable this CPU's synchronous printk() before taking
* seqlock to prevent any printk() from trying to hold port->lock, for
@@ -5859,7 +5863,11 @@ static void __build_all_zonelists(void *data)
write_sequnlock(&zonelist_update_seq);
printk_deferred_exit();
+#ifndef CONFIG_PREEMPT_RT
local_irq_restore(flags);
+#else
+ migrate_enable();
+#endif
}
static noinline void __init
----------------------------------------
