Re: [PATCH] mm/page_owner.c: allow page_owner with given start_pfn/count

[Kassey Li <quic_yingangl@quicinc.com>]

On 7/27/2022 6:59 PM, Vlastimil Babka (SUSE) wrote:
> On 7/27/22 09:44, Kassey Li wrote:
>>
>>
>> On 7/26/2022 10:03 PM, Vlastimil Babka (SUSE) wrote:
>>> On 7/25/22 10:39, Kassey Li wrote:
>>>> hi, Matthew:
>>>>       sorry for the delay, I just started to learn how to upstream patch, and
>>>> setup my Thunderbird with plain text only.
>>>>       you are right, two users will cause problem here.
>>>>       the uses case is dump CMA area to understand the page usage in a given
>>>> cma pool. 2nd, dump whole memory page owner is very time cost, mostly our
>>>> android device has 8G memory now.
>>>>       I will research and check again, if you have more idea on this , please
>>>> kindly to share.
>>>
>>> You could try employing lseek() to specify the start pfn, and as for end
>>> pfn, the process can just stop reading and close when it has seen enough?
>>
>> lseek is a good idea.
>> read_page_owner start with below
>>      pfn = min_low_pfn + *ppos;
>> so we need to export the min_low_pfn to user then decide the ppos to seek.
>> (my_cma.base_pfn - min_low_pfn) is the ppos we want to set.
> 
> Hm could we just pfn = *ppos and then anything below min_low_pfn is skipped
> internally? So we don't need to teach userspace min_low_pfn.
	that makes sense.
	I send out a new path "mm/page_owner.c: add llseek for page_owner" 
according your suggest, please help to review again, thanks.
> 
>> is there concern to export min_low_pfn  ?
>> or use a mutex lock for my previous debugfs version patch ?
>>
>>>
>>>> BR
>>>> Kassey
>>>>
>>>> On 7/22/2022 11:38 PM, Matthew Wilcox wrote:
>>>>> On Fri, Jul 22, 2022 at 11:08:10PM +0800, Kassey Li wrote:
>>>>>> by default, page_owner iterates all page from min_low_pfn to
>>>>>> max_pfn, this cost too much time if we want an alternative pfn range.
>>>>>>
>>>>>> with this patch it allows user to set pfn range to dump the page_onwer.
>>>>>
>>>>> This is a really bad UI.  If two users try to do different ranges at the
>>>>> same time, it'll go wrong.  What use cases are you actually trying to
>>>>> solve?
>>>>
>>>
>