Re: [Bug 201085] New: Kernel allows mlock() on pages in CMA without migrating pages out of CMA first

["Aneesh Kumar K.V" <aneesh.kumar@linux.ibm.com>]

On 9/13/18 3:49 AM, Mike Kravetz wrote:
> On 09/12/2018 12:47 PM, Andrew Morton wrote:
>>
>> (switched to email.  Please respond via emailed reply-to-all, not via the
>> bugzilla web interface).
>>
>> On Tue, 11 Sep 2018 03:59:11 +0000 bugzilla-daemon@bugzilla.kernel.org wrote:
>>
>>> https://bugzilla.kernel.org/show_bug.cgi?id=201085
>>>
>>>              Bug ID: 201085
>>>             Summary: Kernel allows mlock() on pages in CMA without
>>>                      migrating pages out of CMA first
>>>             Product: Memory Management
>>>             Version: 2.5
>>>      Kernel Version: 4.18
>>>            Hardware: All
>>>                  OS: Linux
>>>                Tree: Mainline
>>>              Status: NEW
>>>            Severity: normal
>>>            Priority: P1
>>>           Component: Page Allocator
>>>            Assignee: akpm@linux-foundation.org
>>>            Reporter: tpearson@raptorengineering.com
>>>          Regression: No
>>>
>>> Pages allocated in CMA are not migrated out of CMA when non-CMA memory is
>>> available and locking is attempted via mlock().  This can result in rapid
>>> exhaustion of the CMA pool if memory locking is used by an application with
>>> large memory requirements such as QEMU.
>>>
>>> To reproduce, on a dual-CPU (NUMA) POWER9 host try to launch a VM with mlock=on
>>> and 1/2 or more of physical memory allocated to the guest.  Observe full CMA
>>> pool depletion occurs despite plenty of normal free RAM available.
>>>
>>> -- 
>>> You are receiving this mail because:
>>> You are the assignee for the bug.
> 
> IIRC, Aneesh is working on some powerpc IOMMU patches for a similar issue
> (long term pinning of cma pages).  Added him on Cc:
> https://lkml.kernel.org/r/20180906054342.25094-2-aneesh.kumar@linux.ibm.com
> 
> This report seems to be suggesting a more general solution/change.  Wondering
> if there is any overlap with this and Aneesh's work.
> 

This is a related issue. I am looking at doing something similar to what 
I did with IOMMU patches. That is migrate pages out of CMA region bfore 
mlock.

The problem mentioned is similar to vfio. With VFIO we do pin the guest 
pages and that is similar with -realtime mlock=on option of Qemu.

We can endup backing guest RAM with pages from CMA area and these are 
different qemu options that do pin these guest pages for the lifetime of 
the guest.

-aneesh