[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [libvirt] [PATCH] qemu: Don't use -mem-prealloc among with .prealloc=yes



On 11/5/18 3:49 PM, Michal Privoznik wrote:
> https://bugzilla.redhat.com/show_bug.cgi?id=1624223
> 
> There are two ways to request memory preallocation on cmd line:
> -mem-prealloc and .prealloc attribute to memory-backend-file.
> However, as it turns out it's not safe to use both at the same
> time. Prefer -mem-prealloc as it is more backward compatible
> compared to switching to "-numa node,memdev=  + -object
> memory-backend-file".
> 
> Signed-off-by: Michal Privoznik <mprivozn redhat com>
> ---
>  src/qemu/qemu_command.c                       | 37 +++++++++++++------
>  src/qemu/qemu_command.h                       |  1 +
>  src/qemu/qemu_domain.c                        |  2 +
>  src/qemu/qemu_domain.h                        |  3 ++
>  src/qemu/qemu_hotplug.c                       |  3 +-
>  .../hugepages-numa-default-dimm.args          |  2 +-
>  6 files changed, 35 insertions(+), 13 deletions(-)

Ping?

After some discussion with QEMU devels, I found out that this might be
harmful/suboptimal. Thing is, if -mem-prealloc is used then qemu will
fully allocate the memory (this is done by actually touching every page
that has been allocated). Then, if .prealloc=yes is specified,
mbind(flags = MPOL_MF_STRICT | MPOL_MF_MOVE) is called which:

a) has to (possibly) move the memory to a different NUMA node,
b) can have no effect when hugepages are in play (thus ignoring user
request to place memory on desired NUMA nodes).

While I could live with a), I couldn't with b). The patch is still valid.

Michal


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]