Date:   Mon, 5 Mar 2018 16:57:40 -0800
From:   Mike Kravetz <mike.kravetz@...cle.com>
To:     Andrew Morton <akpm@...ux-foundation.org>
Cc:     linux-mm@...ck.org, linux-kernel@...r.kernel.org,
        KAMEZAWA Hiroyuki <kamezawa.hiroyu@...fujitsu.com>,
        Luiz Capitulino <lcapitulino@...hat.com>,
        Michal Nazarewicz <mina86@...a86.com>,
        Michal Hocko <mhocko@...nel.org>,
        Vlastimil Babka <vbabka@...e.cz>,
        Mel Gorman <mgorman@...hsingularity.net>,
        Johannes Weiner <hannes@...xchg.org>
Subject: Re: [PATCH 1/1] mm: make start_isolate_page_range() fail if already
 isolated

On 03/02/2018 05:39 PM, Mike Kravetz wrote:
> On 03/02/2018 04:56 PM, Andrew Morton wrote:
>> On Fri, 2 Mar 2018 16:38:33 -0800 Mike Kravetz <mike.kravetz@...cle.com> wrote:
>>
>>> On 03/02/2018 04:06 PM, Andrew Morton wrote:
>>>> On Mon, 26 Feb 2018 11:10:54 -0800 Mike Kravetz <mike.kravetz@...cle.com> wrote:
>>>>
>>>>> start_isolate_page_range() is used to set the migrate type of a
>>>>> set of page blocks to MIGRATE_ISOLATE while attempting to start
>>>>> a migration operation.  It assumes that only one thread is
>>>>> calling it for the specified range.  This routine is used by
>>>>> CMA, memory hotplug and gigantic huge pages.  Each of these users
>>>>> synchronize access to the range within their subsystem.  However,
>>>>> two subsystems (CMA and gigantic huge pages for example) could
>>>>> attempt operations on the same range.  If this happens, page
>>>>> blocks may be incorrectly left marked as MIGRATE_ISOLATE and
>>>>> therefore not available for page allocation.
>>>>>
>>>>> Without 'locking code' there is no easy way to synchronize access
>>>>> to the range of page blocks passed to start_isolate_page_range.
>>>>> However, if two threads are working on the same set of page blocks
>>>>> one will stumble upon blocks set to MIGRATE_ISOLATE by the other.
>>>>> In such conditions, make the thread noticing MIGRATE_ISOLATE
>>>>> clean up as normal and return -EBUSY to the caller.
>>>>>
>>>>> This will allow start_isolate_page_range to serve as a
>>>>> synchronization mechanism and will allow for more general use
>>>>> of callers making use of these interfaces.  So, update comments
>>>>> in alloc_contig_range to reflect this new functionality.
>>>>>
>>>>> ...
>>>>>
>>>>> --- a/mm/page_isolation.c
>>>>> +++ b/mm/page_isolation.c
>>>>> @@ -28,6 +28,13 @@ static int set_migratetype_isolate(struct page *page, int migratetype,
>>>>>  
>>>>>  	spin_lock_irqsave(&zone->lock, flags);
>>>>>  
>>>>> +	/*
>>>>> +	 * We assume we are the only ones trying to isolate this block.
>>>>> +	 * If MIGRATE_ISOLATE already set, return -EBUSY
>>>>> +	 */
>>>>> +	if (is_migrate_isolate_page(page))
>>>>> +		goto out;
>>>>> +
>>>>>  	pfn = page_to_pfn(page);
>>>>>  	arg.start_pfn = pfn;
>>>>>  	arg.nr_pages = pageblock_nr_pages;
>>>>
>>>> Seems a bit ugly and I'm not sure that it's correct.  If the loop in
>>>> start_isolate_page_range() gets partway through a number of pages then
>>>> we hit the race, start_isolate_page_range() will then go and "undo" the
>>>> work being done by the thread which it is racing against?
>>>
>>> I agree that it is a bit ugly.  However, when a thread hits the above
>>> condition it will only undo what it has done.  Only one thread is able
>>> to set migrate state to isolate (under the zone lock).  So, a thread
>>> will only undo what it has done.
>>
>> I don't get it.  That would make sense if start_isolate_page_range()
>> held zone->lock across the entire loop, but it doesn't do that.
>>
> 
> It works because all threads set migrate isolate on page blocks going
> from pfn low to pfn high.  When they encounter a conflict, they know
> exactly which blocks they set and only undo those blocks.  Perhaps, I
> am missing something, but it does not matter because ...
> 
>>> The exact problem of one thread undoing what another thread has done
>>> is possible with the code today and is what this patch is attempting
>>> to address.
>>>
>>>> Even if that can't happen, blundering through a whole bunch of pages
>>>> then saying whoops then undoing everything is unpleasing.
>>>>
>>>> Should we be looking at preventing these races at a higher level?
>>>
>>> I could not immediately come up with a good idea here.  The zone lock
>>> would be the obvious choice, but I don't think we want to hold it while
>>> examining each of the page blocks.  Perhaps a new lock or semaphore
>>> associated with the zone?  I'm open to suggestions.
>>
>> Yes, I think it would need a new lock.  Hopefully a mutex.
> 
> I'll look into adding an 'isolate' mutex to the zone structure and reworking
> this patch.

I went back and examined the 'isolation functionality' with an eye on perhaps
adding a mutex for some higher level synchronization.  However, there does
not appear to be a straight forward solution.

What we really need is some way of preventing two threads from operating on
the same set of page blocks concurrently.  We do not want a big mutex, as
we do want two threads to run in parallel if operating on separate
non-overlapping ranges (CMA does this today).  If we did this, I think we
would need a new data structure to represent page blocks within a zone.
start_isolate_page_range() would then then check the new data structure for
conflicts, and if none found mark the range it is operating on as 'in use'.
undo_isolate_page_range() would clear the entries for the range in the new
data structure.  Such information would hang off the zone and be protected
by the zone lock.  The new data structure could be static (like a bit map),
or dynamic.  It certainly is doable, but ...

The more I think about it, the more I like my original proposal.  The
comment "blundering through a whole bunch of pages then saying whoops
then undoing everything is unpleasing" is certainly true.  But do note
that after isolating the page blocks, we will then attempt to migrate
pages within those blocks.  There is a more than a minimal chance that
we will not be able to migrate something within the set of page blocks.
In that case we again say whoops and undo even more work.

I am relatively new to this area of code.  Therefore, it would be good to
get comments from some of the original authors.
-- 
Mike Kravetz

Powered by blists - more mailing lists