Re: [Qemu-devel] [PATCH] mlock: fix bug when mlockall called before mbind

[zhanghailiang <zhang.zhanghailiang@huawei.com>]

On 2014/8/13 19:50, Michael S. Tsirkin wrote:
> On Wed, Aug 13, 2014 at 07:21:57PM +0800, zhanghailiang wrote:
>> If we configure qemu with realtime-mlock-on and memory-node-bind at the same time,
>> Qemu will fail to start, and mbind() fails with message "Input/output error".
>>
>> > From man page:
>> int mbind(void *addr, unsigned long len, int mode,
>>                   unsigned long *nodemask, unsigned long maxnode,
>>                   unsigned flags);
>> The *MPOL_BIND* mode specifies a strict policy that restricts memory allocation
>> to the nodes specified in nodemask.
>> If *MPOL_MF_STRICT* is passed in flags and policy is not MPOL_DEFAULT(In qemu
>> here is MPOL_BIND), then the call will fail with the error EIO if the existing
>> pages in  the memory range don't follow the policy.
>>
>> The memory locked ahead by mlockall can not guarantee to follow the policy above,
>> And if that happens, it will result in an EIO error.
>>
>> So we should call mlock after mbind, here we adjust the place where called mlock,
>> Move it to function pc_memory_init.
>>
>> Signed-off-by: xiexiangyou<xiexiangyou@huawei.com>
>> Signed-off-by: zhanghailiang<zhang.zhanghailiang@huawei.com>
>
> OK but won't this still fail in case of memory hotplug?
> We set MCL_FUTURE so the same will apply?
> Maybe it's enough to set MPOL_MF_MOVE?
> Does the following work for you?
>

Hi Michael,

I have tested memory hotplug, use virsh command like
'virsh setmem redhat-6.4 6388608 --config --live', it is OK, and
it will not call mbind when do such memory hotplug.
but i don't know if there is command like 'memory-node hotplug' ?

MPOL_MF_MOVE can work, it is more simple, but it is not perfect. It 
consumes more time to *move the memory*(i guess will reconstruct pages 
and copy memory)which has been locked by mlockall. The result is VM will 
start slower than the above scenario.

BTW, i think the follow process is clearer and more logical:
Allocate memory--->Set memory policy--->Lock memory.

So what's your opinion? Thanks very much.


> -->
>
> hostmem: set MPOL_MF_MOVE
>
> When memory is allocated on a wrong node, MPOL_MF_STRICT
> doesn't move it - it just fails the allocation.
> A simple way to reproduce the failure is with mlock=on
> realtime feature.
>
> The code comment actually says: "ensure policy won't be ignored"
> so setting MPOL_MF_MOVE seems like a better way to do this.
>
> Signed-off-by: Michael S. Tsirkin<mst@redhat.com>
>
> ---
>
> diff --git a/backends/hostmem.c b/backends/hostmem.c
> index ca10c51..a9905c0 100644
> --- a/backends/hostmem.c
> +++ b/backends/hostmem.c
> @@ -304,7 +304,7 @@ host_memory_backend_memory_complete(UserCreatable *uc, Error **errp)
>           /* ensure policy won't be ignored in case memory is preallocated
>            * before mbind(). note: MPOL_MF_STRICT is ignored on hugepages so
>            * this doesn't catch hugepage case. */
> -        unsigned flags = MPOL_MF_STRICT;
> +        unsigned flags = MPOL_MF_STRICT | MPOL_MF_MOVE;
>
>           /* check for invalid host-nodes and policies and give more verbose
>            * error messages than mbind(). */
>