qemu-devel
[Top][All Lists]
Advanced

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

Re: [PATCH 1/2] accel: kvm: Fix memory waste under mismatch page size


From: Peter Xu
Subject: Re: [PATCH 1/2] accel: kvm: Fix memory waste under mismatch page size
Date: Tue, 15 Dec 2020 12:57:19 -0500

On Tue, Dec 15, 2020 at 03:19:47PM +0800, Keqian Zhu wrote:
> When handle dirty log, we face qemu_real_host_page_size and
> TARGET_PAGE_SIZE. The first one is the granule of KVM dirty
> bitmap, and the second one is the granule of QEMU dirty bitmap.
> 
> Generally speaking, qemu_real_host_page_size >= TARGET_PAGE_SIZE,
> so misuse TARGET_PAGE_SIZE to init kvmslot dirty_bmap may waste
> memory. For example, when qemu_real_host_page_size is 64K and
> TARGET_PAGE_SIZE is 4K, this bugfix can save 93.75% memory.
> 
> Signed-off-by: Keqian Zhu <zhukeqian1@huawei.com>
> ---
>  accel/kvm/kvm-all.c | 6 +++++-
>  1 file changed, 5 insertions(+), 1 deletion(-)
> 
> diff --git a/accel/kvm/kvm-all.c b/accel/kvm/kvm-all.c
> index baaa54249d..c5e06288eb 100644
> --- a/accel/kvm/kvm-all.c
> +++ b/accel/kvm/kvm-all.c
> @@ -620,8 +620,12 @@ static void kvm_memslot_init_dirty_bitmap(KVMSlot *mem)
>       * too, in most cases).
>       * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
>       * a hope that sizeof(long) won't become >8 any time soon.
> +     *
> +     * Note: the granule of kvm dirty log is qemu_real_host_page_size.
> +     * And mem->memory_size is aligned to it (otherwise this mem can't
> +     * be registered to KVM).
>       */
> -    hwaddr bitmap_size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
> +    hwaddr bitmap_size = ALIGN(mem->memory_size / qemu_real_host_page_size,
>                                          /*HOST_LONG_BITS*/ 64) / 8;

Yes I think this is correct.  Thanks.

So one thing I failed to notice is cpu_physical_memory_set_dirty_lebitmap()
will "amplify" the host dirty pages into guest ones; seems we're all good then.

Reviewed-by: Peter Xu <peterx@redhat.com>

-- 
Peter Xu




reply via email to

[Prev in Thread] Current Thread [Next in Thread]