Re: [Qemu-devel] [PATCH V6 07/10] migration: add bitmap for copied page

[Peter Xu]

On Wed, May 24, 2017 at 03:16:23PM +0300, Alexey Perevalov wrote:
> On 05/24/2017 03:01 PM, Peter Xu wrote:
> >On Wed, May 24, 2017 at 10:56:37AM +0300, Alexey wrote:
> >>On Wed, May 24, 2017 at 02:57:36PM +0800, Peter Xu wrote:
> >>>On Tue, May 23, 2017 at 02:31:08PM +0300, Alexey Perevalov wrote:
> >>>>This patch adds ability to track down already copied
> >>>>pages, it's necessary for calculation vCPU block time in
> >>>>postcopy migration feature and maybe for restore after
> >>>>postcopy migration failure.
> >>>>
> >>>>Functions which work with RAMBlock are placed into ram.c,
> >>>>due to TARGET_WORDS_BIGENDIAN is poisoned int postcopy-ram.c -
> >>>>hardware independed code.
> >>>>
> >>>>Signed-off-by: Alexey Perevalov <address@hidden>
> >>>>---
> >>>>  include/migration/migration.h | 16 +++++++++++
> >>>>  migration/postcopy-ram.c      | 22 ++++++++++++---
> >>>>  migration/ram.c               | 63 
> >>>> +++++++++++++++++++++++++++++++++++++++++++
> >>>>  3 files changed, 97 insertions(+), 4 deletions(-)
> >>>>
> >>>>diff --git a/include/migration/migration.h b/include/migration/migration.h
> >>>>index 449cb07..4e05c83 100644
> >>>>--- a/include/migration/migration.h
> >>>>+++ b/include/migration/migration.h
> >>>>@@ -101,6 +101,20 @@ struct MigrationIncomingState {
> >>>>      LoadStateEntry_Head loadvm_handlers;
> >>>>      /*
> >>>>+     * bitmap indicates whether page copied,
> >>>>+     * based on ramblock offset
> >>>>+     * now it is using only for blocktime calculation in
> >>>>+     * postcopy migration, so livetime of this entry:
> >>>>+     * since user requested blocktime calculation,
> >>>>+     * till the end of postcopy migration
> >>>>+     * as an example it could represend following memory map
> >>>>+     * ___________________________________
> >>>>+     * |4k pages | hugepages | 4k pages
> >>>Can we really do this?
> >>>The problem is AFAIU migration stream is sending pages only in target
> >>>page size, even for huge pages... so even for huge pages we should
> >>>still need per TARGET_PAGE_SIZE bitmap?
> >>sending maybe, but copying to userfault fd is doing in hugepage size
> >Yes you are right. :)
> >
> >>in case of hugetlbfs memory backend.
> >>>(Please refer to ram_state_init() on init of RAMBlock.bmap)
> >>I thought to use bitmap per ramblock, but I decided to not do it,
> >>because of following reasons:
> >>    1. There is only 4k ramblocks, for such ramblock it's not
> >>    optimal
> >Could you explain what do you mean by "there is only 4k ramblocks"?
> sorry, could be ramblocks with 4k size,
> as example, your's qemu hmp info ramblock shows
>  Block Name    PSize              Offset Used              Total
>             /objects/mem    2 MiB  0x0000000000000000 0x0000000020000000
> 0x0000000020000000
>                 vga.vram    4 KiB  0x0000000020060000 0x0000000001000000
> 0x0000000001000000
>     /address@hidden/acpi/tables    4 KiB  0x0000000021130000 
> 0x0000000000020000
> 0x0000000000200000
>                  pc.bios    4 KiB  0x0000000020000000 0x0000000000040000
> 0x0000000000040000
>   0000:00:03.0/e1000.rom    4 KiB  0x0000000021070000 0x0000000000040000
> 0x0000000000040000
>   0000:00:04.0/e1000.rom    4 KiB  0x00000000210b0000 0x0000000000040000
> 0x0000000000040000
>   0000:00:05.0/e1000.rom    4 KiB  0x00000000210f0000 0x0000000000040000
> 0x0000000000040000
>                   pc.rom    4 KiB  0x0000000020040000 0x0000000000020000
> 0x0000000000020000
>     0000:00:02.0/vga.rom    4 KiB  0x0000000021060000 0x0000000000010000
> 0x0000000000010000
>    /address@hidden/table-loader    4 KiB  0x0000000021330000 
> 0x0000000000001000
> 0x0000000000001000
>       /address@hidden/acpi/rsdp    4 KiB  0x0000000021331000 
> 0x0000000000001000
> 0x0000000000001000
> 
> /address@hidden/table-loader and /address@hidden/acpi/rsdp has only one 4K 
> page.

I see. Thanks for explaining this.

A 4k sized ramblock means the bitmap would be only one unsigned long
width (via bitmap_new(1)). I still don't see why it's not good... :-)

-- 
Peter Xu