Re: [Qemu-devel] DIRTY_MEMORY_BLOCK_SIZE;

[Stefan Hajnoczi <stefanha@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 2339 bytes --]

On Wed, Jul 05, 2017 at 05:06:41PM +0100, Dr. David Alan Gilbert wrote:
> * Stefan Hajnoczi (stefanha@redhat.com) wrote:
> > On Fri, Jun 30, 2017 at 12:26:10PM +0100, Dr. David Alan Gilbert wrote:
> > > * ali saeedi (ali.saeedi56@gmail.com) wrote:
> > > > Hello
> > > > what does 'DIRTY_MEMORY_BLOCK_SIZE' mean?
> > > > is it the number of words in a block? or number of pages in a block? or
> > > > number of bytes in a block?
> > > > thanks a lot
> > > 
> > > (cc'ing Stefan)
> > > I think that DIRTY_MEMORY_BLOCK_SIZE is the number of TARGET_PAGEs
> > > within one DIRTY_MEMORY_BLOCK
> > > So with the common 4k target page that's 4k*256k*8=8GB/dirty memory
> > > block - note these are just the size of structure sin qemu, it's still
> > > got the ranularity ot mark individual target pages as dirty.
> > 
> > Right, the calculation is shown in the comment above the code:
> > 
> >  *   rcu_read_lock();
> >  *
> >  *   DirtyMemoryBlocks *blocks =
> >  *       atomic_rcu_read(&ram_list.dirty_memory[DIRTY_MEMORY_MIGRATION]);
> >  *
> >  *   ram_addr_t idx = (addr >> TARGET_PAGE_BITS) / DIRTY_MEMORY_BLOCK_SIZE;
> >  *   unsigned long *block = blocks.blocks[idx];
> >  *   ...access block bitmap...
> >  *
> >  *   rcu_read_unlock();
> > 
> > Rather than focussing on DIRTY_MEMORY_BLOCK_SIZE, make sure you
> > understand how DirtyMemoryBlocks works.  It is an array of bitmap
> > pointers.
> > 
> > Instead of directly indexing into a single huge dirty memory bitmap,
> > QEMU divides the dirty memory bitmap into fixed-sized chunks.  Each
> > chunk covers DIRTY_MEMORY_BLOCK_SIZE pages.
> > 
> > The reason for this layer of indirection is so that the dirty memory
> > bitmap can be accessed without taking a traditional lock (just RCU) and
> > also supports memory hotplug.
> > 
> > Without indirection it would be difficult to grow the bitmap while other
> > threads are writing to it.  Thanks to the indirection, it's possible to
> > allocate new chunks and continue using the old chunks when growth
> > occurs.
> 
> I guess this works like the old non-chunk version, in that there's no
> direct correspondence between DirtyMemoryBlocks and RAMBlock's - i.e.
> one RAMBlock might span two DirtyMemoryBlocks even if it's smaller
> than DIRTY_MEMORY_BLOCK_SIZE.

Yes.

Stefan

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 455 bytes --]