Re: [RFC PATCH v2 1/3] meminfo_extra: introduce meminfo extra

[Jaewon Kim]

On 2020ë 03ì 24ì 20:46, Greg KH wrote:
> On Tue, Mar 24, 2020 at 08:37:38PM +0900, Jaewon Kim wrote:
>>
>> On 2020ë 03ì 24ì 19:11, Greg KH wrote:
>>> On Tue, Mar 24, 2020 at 06:11:17PM +0900, Jaewon Kim wrote:
>>>> On 2020ë 03ì 23ì 18:53, Greg KH wrote:
>>>>>> +int register_meminfo_extra(atomic_long_t *val, int shift, const char *name)
>>>>>> +{
>>>>>> +	struct meminfo_extra *meminfo, *memtemp;
>>>>>> +	int len;
>>>>>> +	int error = 0;
>>>>>> +
>>>>>> +	meminfo = kzalloc(sizeof(*meminfo), GFP_KERNEL);
>>>>>> +	if (!meminfo) {
>>>>>> +		error = -ENOMEM;
>>>>>> +		goto out;
>>>>>> +	}
>>>>>> +
>>>>>> +	meminfo->val = val;
>>>>>> +	meminfo->shift_for_page = shift;
>>>>>> +	strncpy(meminfo->name, name, NAME_SIZE);
>>>>>> +	len = strlen(meminfo->name);
>>>>>> +	meminfo->name[len] = ':';
>>>>>> +	strncpy(meminfo->name_pad, meminfo->name, NAME_BUF_SIZE);
>>>>>> +	while (++len < NAME_BUF_SIZE - 1)
>>>>>> +		meminfo->name_pad[len] = ' ';
>>>>>> +
>>>>>> +	spin_lock(&meminfo_lock);
>>>>>> +	list_for_each_entry_rcu(memtemp, &meminfo_head, list) {
>>>>>> +		if (memtemp->val == val) {
>>>>>> +			error = -EINVAL;
>>>>>> +			break;
>>>>>> +		}
>>>>>> +	}
>>>>>> +	if (!error)
>>>>>> +		list_add_tail_rcu(&meminfo->list, &meminfo_head);
>>>>>> +	spin_unlock(&meminfo_lock);
>>>>> If you have a lock, why are you needing rcu?
>>>> I think _rcu should be removed out of list_for_each_entry_rcu.
>>>> But I'm confused about what you meant.
>>>> I used rcu_read_lock on __meminfo_extra,
>>>> and I think spin_lock is also needed for addition and deletion to handle multiple modifiers.
>>> If that's the case, then that's fine, it just didn't seem like that was
>>> needed.  Or I might have been reading your rcu logic incorrectly...
>>>
>>>>>> +	if (error)
>>>>>> +		kfree(meminfo);
>>>>>> +out:
>>>>>> +
>>>>>> +	return error;
>>>>>> +}
>>>>>> +EXPORT_SYMBOL(register_meminfo_extra);
>>>>> EXPORT_SYMBOL_GPL()?  I have to ask :)
>>>> I can use EXPORT_SYMBOL_GPL.
>>>>> thanks,
>>>>>
>>>>> greg k-h
>>>>>
>>>>>
>>>> Hello
>>>> Thank you for your comment.
>>>>
>>>> By the way there was not resolved discussion on v1 patch as I mentioned on cover page.
>>>> I'd like to hear your opinion on this /proc/meminfo_extra node.
>>> I think it is the propagation of an old and obsolete interface that you
>>> will have to support for the next 20+ years and yet not actually be
>>> useful :)
>>>
>>>> Do you think this is meaningful or cannot co-exist with other future
>>>> sysfs based API.
>>> What sysfs-based API?
>> Please refer to mail thread on v1 patch set - https://protect2.fireeye.com/url?k=16e3accc-4b2f6548-16e22783-0cc47aa8f5ba-935fe828ac2f6656&u=https://lkml.org/lkml/fancy/2020/3/10/2102
>> especially discussion with Leon Romanovsky on https://protect2.fireeye.com/url?k=74208ed9-29ec475d-74210596-0cc47aa8f5ba-0bd4ef48931fec95&u=https://lkml.org/lkml/fancy/2020/3/16/140
> I really do not understand what you are referring to here, sorry.   I do
> not see any sysfs-based code in that thread.
Sorry. I also did not see actual code.
Hello Leon Romanovsky, could you elaborate your plan regarding sysfs stuff?
>
> And try to use lore.kernel.org, lkml.org doesn't always work and we have
> no control over that :(
>
>>> I still don't know _why_ you want this.  The ION stuff is not needed as
>>> that code is about to be deleted, so who else wants this?  What is the
>>> use-case for it that is so desperately needed that parsing
>>> yet-another-proc file is going to solve the problem?
>> In my Android device, there are graphic driver memory, zsmalloc memory except ION.
> Ok, so what does Android have to do with this?
Some driver in Android platform may use my API to show its memory usage.
>
>> I don't know other cases in other platform.
>> Not desperately needed but I think we need one userspace knob to see overall hidden huge memory.
> Why?  Who wants that?  What would userspace do with that?  And what
> exactly do you want to show?
>
> Is this just a debugging thing?  Then use debugfs for that, not proc.
> Isn't that what the DRM developers are starting to do?
>
>> Additionally I'd like to see all those hidden memory in OutOfMemory log.
> How is anything hidden, can't you see it in the slab information?
>
Let me explain more.

0. slab
As I said in cover page, this is not for memory allocated by slab.
I'd like to know where huge memory has gone.
Those are directly allocated by alloc_pages instead of slab.
/proc/slabinfo does not show this information.

1. /proc/meminfo_extra
/proc/meminfo_extra could be debugging thing to see memory status at a certain time.
But it, I think, is also basic information rather than just for debugging.
It is similar with /proc/meminfo which is in procfs instead of debugfs.

2. oom log
oom log in show_mem is more than just debugging.
As existing oom log shows much memory information, I think we need the hidden memory info.
Without these information, we do NOT know oom reason because other traditional stats are not enough.
>> This is useful to get clue to find memory hogger.
>> i.e.) show_mem on oom
>> <6>[  420.856428]  Mem-Info:
>> <6>[  420.856433]  IonSystemHeap:32813kB ZsPages:44114kB GraphicDriver::13091kB
>> <6>[  420.856450]  active_anon:957205 inactive_anon:159383 isolated_anon:0
> So what does this show you?  That someone is takign a ton of ION memory
> for some unknown use?  What can you do with that?  What would you do
> with that?
We may not know exact memory owner. But we can narrow down.
Anyway I think this is meaningful instead of no clue.
>
> And memory is almost never assigned to a "driver", it is assigned to a
> "device" that uses it.  Drivers can handle multiple devices at the same
> time, so why would you break this down by drivers?  Are you assuming
> that a driver only talks to one piece of hardware?
Yes a driver may support several devices. I don't know if it same on an embedded device.
Anyway I think the idea works even for several devices, although the driver should
distinguish memory usage for each device and should register each memory stat.
>
> I think you need a much better use case for all of this other than
> "wouldn't it be nice to see some numbers", as that isn't going to help
> anyone out in the end.
Sorry. As of now, I do not know other better use case, but I still think
memory information should cover most of memory usage.
Huge memory consumed by driver or other core logic should be shown in OoM.
>
> thanks,
>
> greg k-h
>
>