[Top][All Lists]

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

Re: [Qemu-devel] [PATCH 42/51] ram: Pass RAMBlock to bitmap_sync

From: Juan Quintela
Subject: Re: [Qemu-devel] [PATCH 42/51] ram: Pass RAMBlock to bitmap_sync
Date: Fri, 24 Mar 2017 11:05:45 +0100
User-agent: Gnus/5.13 (Gnus v5.13) Emacs/25.1 (gnu/linux)

Yang Hongyang <address@hidden> wrote:
> Hi Juan,
> On 2017/3/24 16:29, Juan Quintela wrote:
>> Yang Hongyang <address@hidden> wrote:
>>> On 2017/3/24 4:45, Juan Quintela wrote:
>>>> We change the meaning of start to be the offset from the beggining of
>>>> the block.
>>>> @@ -701,7 +701,7 @@ static void migration_bitmap_sync(RAMState *rs)
>>>>      qemu_mutex_lock(&rs->bitmap_mutex);
>>>>      rcu_read_lock();
>>>>      QLIST_FOREACH_RCU(block, &ram_list.blocks, next) {
>>>> -        migration_bitmap_sync_range(rs, block->offset, 
>>>> block->used_length);
>>>> +        migration_bitmap_sync_range(rs, block, 0, block->used_length);
>>> Since RAMBlock been passed to bitmap_sync, could we remove
>>> param 'block->used_length' either?
>> Hi
>> good catch.
>> I had that removed, and then realized that I want to synchronize parts
>> of the bitmap, not the whole one.  That part of the series is still not
>> done.
>> Right now we do something like (I have simplified a lot of details):
>> while(true) {
>>             foreach(block)
>>                 bitmap_sync(block)
>>             foreach(page)
>>                 if(dirty(page))
>>                    page_send(page)
>> }
>> If you have several terabytes of RAM that is too ineficient, because
>> when we arrive to the page_send(page), it is possible that it is already
>> dirty again, and we have to send it twice.  So, the idea is to change to
>> something like:
>> while(true) {
>>             foreach(block)
>>                 bitmap_sync(block)
> Do you mean sync with KVM here?
>>             foreach(block)
>>                 foreach(64pages)
>>                     bitmap_sync(64pages)
> Then here, we will sync with KVM too. For huge MEM,
> it will generates lots of ioctl()...
> Bitmap in KVM is per Memory region IIRC. KVM module currently
> haven't the ability to sync parts of the bitmap. A sync have
> to sync the whole mr. So if we want to do small sync, we might
> need to modify KVM also, but that still won't solve the preblem
> of increased ioctls.

And why I remembered incorrectly that we could sync part of the bitmaps.
Yes, we could have more ioctls, but less pages written twice, it is a
tradeoff, at some point it makes sense to change it.

Problem is that now this is going to be more difficult that I thought to
test for it.

Thanks, Juan.

reply via email to

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