Message-ID: <63102e0a-8a16-4b1d-b8f8-d9833079c924@bytedance.com>
Date: Fri, 1 Mar 2024 16:42:43 +0800
From: Abel Wu <wuyun.abel@...edance.com>
To: Chen Yu <yu.c.chen@...el.com>
Cc: Peter Zijlstra <peterz@...radead.org>, Ingo Molnar <mingo@...hat.com>,
 Vincent Guittot <vincent.guittot@...aro.org>,
 Juri Lelli <juri.lelli@...hat.com>, Tim Chen <tim.c.chen@...el.com>,
 Tiwei Bie <tiwei.btw@...group.com>,
 Honglei Wang <wanghonglei@...ichuxing.com>, Aaron Lu <aaron.lu@...el.com>,
 Chen Yu <yu.chen.surf@...il.com>, linux-kernel@...r.kernel.org,
 kernel test robot <oliver.sang@...el.com>
Subject: Re: [RFC PATCH] sched/eevdf: Return leftmost entity in pick_eevdf()
 if no eligible entity is found

On 3/1/24 3:07 PM, Chen Yu Wrote:
> On 2024-02-29 at 17:00:18 +0800, Abel Wu wrote:
>> Hi Chen, thanks for detailed analysis.
>>
>> The title of this patch sounds a little weird to me, since any
>> non-empty cfs_rq should have at least one eligible entity.
> 
> My understanding is that, current + cfs_rq's tree has at least one
> eligible entity, as the current running entity is not on the tree.
> In vruntime_eligible() check, we consider all the entities on the
> tree, plus the current running one. So, it could be possible that
> all the entities on the tree are not eligible, while the current is
> eligible, no?

Yes, and I didn't mean only rb-tree when I said "non-empty cfs_rq" :)

> 
>> Besides, choosing the leftmost entity which could be non-eligible can be
>> sub-optimal, anyway this is only a workaround..
>>
> 
> Yes, it is a workaround to show the possible scenario to bring NULL exception,
> and it was not root caused yet.
>   
>> On 2/26/24 4:23 PM, Chen Yu Wrote:
>>> There is occasional report from lkp that the kernel hits the NULL pointer
>>> exception:
>>>
>>> [  512.079810][ T8305] BUG: kernel NULL pointer dereference, address: 0000002c
>>> [  512.080897][ T8305] #PF: supervisor read access in kernel mode
>>> [  512.081636][ T8305] #PF: error_code(0x0000) - not-present page
>>> [  512.082337][ T8305] *pde = 00000000
>>> [  512.082829][ T8305] Oops: 0000 [#1] PREEMPT SMP
>>> [  512.083407][ T8305] CPU: 1 PID: 8305 Comm: watchdog Tainted: G        W
>>> [  512.086203][ T8305] EIP: set_next_entity (fair.c:?)
>>>
>>> This is caused by NULL candidate returned by pick_eevdf() as Abel analyzed.
>>> After
>>> commit 2227a957e1d5 ("sched/eevdf: Sort the rbtree by virtual deadline")
>>> the NULL candidate would trigger the NULL pointer exception. While before
>>> this commit, there would be warning.
>>>
>>> This NULL entity issue was always there before above commit. With debug
>>> patch to print the cfs_rq and all the entities in the tree, we have the
>>> information when the issue was reproduced:
>>>
>>> [  514.461242][ T8390] cfs_rq avg_vruntime:386638640128 avg_load:2048 min_vruntime:763383370431
>>> [  514.535935][ T8390] current on_rq se 0xc5851400, deadline:18435852013562231446
>>> 			min_vruntime:18437121115753667698 vruntime:18435852013561943404, load:629
>>> [  514.536772][ T8390] Traverse rb-tree from left to right
>>> [  514.537138][ T8390]  se 0xec1234e0 deadline:763384870431 min_vruntime:763383370431 vruntime:763383370431 non-eligible
>>> [  514.537835][ T8390]  se 0xec4fcf20 deadline:763762447228 min_vruntime:763760947228 vruntime:763760947228 non-eligible
>>> [  514.538539][ T8390] Traverse rb-tree from topdown
>>> [  514.538877][ T8390]  middle se 0xec1234e0 deadline:763384870431 min_vruntime:763383370431 vruntime:763383370431 non-eligible
>>> [  514.539605][ T8390]  middle se 0xec4fcf20 deadline:763762447228 min_vruntime:763760947228 vruntime:763760947228 non-eligible
>>> [  514.540340][ T8390] Found best:0x0
>>> [  514.540613][ T8390] BUG: kernel NULL pointer dereference, address: 00000074
>>>
>>> We can see that non of the entities in the tree are eligible, neither is
>>> the current entity on this cfs_rq. As a result, curr is set to NULL:
>>> if (curr && (!curr->on_rq || !entity_eligible(cfs_rq, curr)))
>>> 	curr = NULL;
>>>
>>> and the best is set to NULL, which caused the problem:
>>> if (!best || (curr && entity_before(curr, best)))
>>> 	best = curr;
>>>
>>> The cause is that, the curr is eligible, but vruntime_eligible()
>>> returns false. And the false negative is due to the following
>>> code in vruntime_eligible():
>>>
>>> return avg >= (s64)(vruntime - cfs_rq->min_vruntime) * load;
>>>
>>> According to the log, vruntime is 18435852013561943404, the
>>> cfs_rq->min_vruntime is 763383370431, the load is 629 + 2048 = 2677,
>>> thus:
>>> s64 delta = (s64)(18435852013561943404 - 763383370431) = -10892823530978643
>>>       delta * 2677 = 7733399554989275921
>>> that is to say, the multiply result overflow the s64, which turns the
>>> negative value into a positive value, thus eligible check fails.
>>
>> Indeed.
>>
>>>
>>> So where is this insane huge vruntime 18435852013561943404 coming from?
>>> My guess is that, it is because the initial value of cfs_rq->min_vruntime
>>> is set to (unsigned long)(-(1LL << 20)). If the task(watchdog in this case)
>>> seldom scheduled in, its vruntime might not move forward too much and
>>> remain its original value by previous place_entity().
>>
>> So why not just initialize to 0? The (unsigned long)(-(1LL << 20))
>> thing is dangerous as it can easily blow up lots of calculations in
>> lag, key, avg_vruntime and so on.
>>
> 
> Usually the min variable is initialized to a very large value, so later
> the real min can overwrite it easily. The 1LL << 20 was introduced decade ago
> in 67e9fb2a39a1, and I don't know why it is this value.

Yes, as the min_vruntime keeps moving forward given the way
update_min_vruntime() does, I don't think it would be a problem.

> 
> 
>> Say during this pre-life, which is about 1ms for 1024-weight entity,
>> there is only one entity running in this cfs_rq. Now another entity
>> with funny lag joins in, being placed somewhere at 0+ vruntime, so
>> cfs_rq->min_vruntime needs to be adjusted accordingly which leads to
>> the breakage of cfs_rq->curr's key as you showed above.
>>
> 
> Yes, this is possible.
> 
>>>
>>> The proper fix should deal with the overflow of entity_key() * load, but
>>> I don't have much clue on that, so propose this conservative method to
>>> restore the previous behavior before the mentioned commit.
>>
>> Inspired by Xuewen's proposal, will it work if limit the key?

After a second thought, this doesn't solve the insane huge vruntime
itself. Since the immediate problem is that 'huge' number actually is
a small one, is it possible to make vruntime s64 at definition?

Powered by blists - more mailing lists