Date:   Fri, 27 Nov 2020 15:53:14 +0530
From:   Charan Teja Kalla <charante@...eaurora.org>
To:     Michal Hocko <mhocko@...e.com>
Cc:     akpm@...ux-foundation.org, david@...hat.com, linux-mm@...ck.org,
        linux-kernel@...r.kernel.org,
        "vinmenon@...eaurora.org" <vinmenon@...eaurora.org>
Subject: Re: [PATCH] mm: memory_hotplug: put migration failure information
 under DEBUG_VM

Thanks Michal!!

On 11/26/2020 2:48 PM, Michal Hocko wrote:
> On Wed 25-11-20 16:18:06, Charan Teja Kalla wrote:
>>
>>
>> On 11/24/2020 1:11 PM, Michal Hocko wrote:
>>> On Mon 23-11-20 20:40:40, Charan Teja Kalla wrote:
>>>>
>>>> Thanks Michal!
>>>> On 11/23/2020 7:43 PM, Michal Hocko wrote:
>>>>> On Mon 23-11-20 19:33:16, Charan Teja Reddy wrote:
>>>>>> When the pages are failed to get isolate or migrate, the page owner
>>>>>> information along with page info is dumped. If there are continuous
>>>>>> failures in migration(say page is pinned) or isolation, the log buffer
>>>>>> is simply getting flooded with the page owner information. As most of
>>>>>> the times page info is sufficient to know the causes for failures of
>>>>>> migration or isolation, place the page owner information under DEBUG_VM.
>>>>>
>>>>> I do not see why this path is any different from others that call
>>>>> dump_page. Page owner can add a very valuable information to debug
>>>>> the underlying reasons for failures here. It is an opt-in debugging
>>>>> feature which needs to be enabled explicitly. So I would argue users
>>>>> are ready to accept a lot of data in the kernel log.
>>>>
>>>> Just thinking how frequently failures can happen in those paths. In the
>>>> memory hotplug path, we can flood the page owner logs just by making one
>>>> page pinned.
>>>
>>> If you are operating on a movable zone then pages shouldn't be pinned
>>> for unbound amount of time. Yeah there are some ways to break this
>>> fundamental assumption but this is a bigger problem that needs a
>>> solution.
>>>
>>>> Say If it is anonymous page, the page owner information
>>>> shows is something like below, which is not really telling anything
>>>> other than how the pinned page is allocated.
>>>
>>> Well you can tell an anonymous page from __dump_page, all right, but
>>> this is not true universally.
>>>
>>>> page last allocated via order 0, migratetype Movable, gfp_mask
>>>> 0x100dca(GFP_HIGHUSER_MOVABLE|__GFP_ZERO)
>>>>   prep_new_page+0x7c/0x1a4
>>>>   get_page_from_freelist+0x1ac/0x1c4
>>>>   __alloc_pages_nodemask+0x12c/0x378
>>>>   do_anonymous_page+0xac/0x3b4
>>>>   handle_pte_fault+0x2a4/0x3bc
>>>>   __handle_speculative_fault+0x208/0x3c0
>>>>   do_page_fault+0x280/0x508
>>>>   do_translation_fault+0x3c/0x54
>>>>   do_mem_abort+0x64/0xf4
>>>>   el0_da+0x1c/0x20
>>>>  page last free stack trace:
>>>>   free_pcp_prepare+0x320/0x454
>>>>   free_unref_page_list+0x9c/0x2a4
>>>>   release_pages+0x370/0x3c8
>>>>   free_pages_and_swap_cache+0xdc/0x10c
>>>>   tlb_flush_mmu+0x110/0x134
>>>>   tlb_finish_mmu+0x48/0xc0
>>>>   unmap_region+0x104/0x138
>>>>   __do_munmap+0x2ec/0x3b4
>>>>   __arm64_sys_munmap+0x80/0xd8
>>>>
>>>> I see at some places in the kernel where they put the dump_page under
>>>> DEBUG_VM, but in the end I agree that it is up to the users need. Then
>>>> there are some users who don't care for these page owner logs.
>>>
>>> Well, as I've said page_owner requires an explicit enabling and I would
>>> expect that if somebody enables this tracking then it is expected to see
>>> the information when we dump a page state.
>>>
>>>> And an issue on Embedded systems with these continuous logs being
>>>> printed to the console is the watchdog timeouts, because console logging
>>>> happens by disabling the interrupts.
>>>
>>> Are you enabling page_owner on those systems unconditionally?
>>>
>>
>> Yes, We do always enable the page owner on just the internal debug
>> builds for memory analysis, But never on the production kernels. And on
>> these builds excessive logging, at times because of a pinned page,
>> causing the watchdog timeouts, is the problem.
> 
> OK, I see but I still believe that the debugging might be useful
> especially when the owner is not really obvious from the page state.
> I also agree that if the output is swapping the logs then the situation
> is not really great either. Would something like the below work for your
> situation?
> 
> MAGIC_NUMBER would need to be somehow figured but I would start with 10
> or so. 
> 
> diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
> index b44d4c7ba73b..3da5c434fb77 100644
> --- a/mm/memory_hotplug.c
> +++ b/mm/memory_hotplug.c
> @@ -1299,6 +1299,8 @@ do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
>  	LIST_HEAD(source);
>  
>  	for (pfn = start_pfn; pfn < end_pfn; pfn++) {
> +		int dumped_page = MAGIC_NUMBER;
> +
>  		if (!pfn_valid(pfn))
>  			continue;
>  		page = pfn_to_page(pfn);
> @@ -1344,7 +1346,10 @@ do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
>  
>  		} else {
>  			pr_warn("failed to isolate pfn %lx\n", pfn);
> -			dump_page(page, "isolation failed");
> +			if (dumped_page--) {
> +				dump_page(page, "isolation failed");
> +				dumped_page = true;
> +			}
>  		}
>  		put_page(page);
>  	}
> @@ -1372,10 +1377,14 @@ do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
>  		ret = migrate_pages(&source, alloc_migration_target, NULL,
>  			(unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_HOTPLUG);
>  		if (ret) {
> +			int dumped_page = MAGIC_NUMBER;
> +
>  			list_for_each_entry(page, &source, lru) {
>  				pr_warn("migrating pfn %lx failed ret:%d ",
>  				       page_to_pfn(page), ret);
> -				dump_page(page, "migration failure");
> +				if (dumped_page--) {
> +					dump_page(page, "migration failure");
> +				}
>  			}
>  			putback_movable_pages(&source);
>  		}
> 

These are working. Rate limiting these logs with default rate limit
interval and burst also helping me.

Thanks!!

-- 
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora
Forum, a Linux Foundation Collaborative Project

Powered by blists - more mailing lists