[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-devel] [PATCH 3/8] block: implement reference count for BlockD
From: |
Fam Zheng |
Subject: |
Re: [Qemu-devel] [PATCH 3/8] block: implement reference count for BlockDriverState |
Date: |
Fri, 26 Jul 2013 09:56:13 +0800 |
User-agent: |
Mutt/1.5.21 (2010-09-15) |
On Thu, 07/25 21:50, Jeff Cody wrote:
> On Fri, Jul 26, 2013 at 09:13:32AM +0800, Fam Zheng wrote:
> > On Thu, 07/25 09:15, Jeff Cody wrote:
> > > On Thu, Jul 25, 2013 at 05:01:41PM +0800, Fam Zheng wrote:
> > > > Introduce bdrv_ref/bdrv_unref to manage the lifecycle of
> > > > BlockDriverState. They are unused for now but will used to replace
> > > > bdrv_delete() later.
> > > >
> > > > Signed-off-by: Fam Zheng <address@hidden>
> > > > ---
> > > > block.c | 22 ++++++++++++++++++++++
> > > > include/block/block.h | 2 ++
> > > > include/block/block_int.h | 1 +
> > > > 3 files changed, 25 insertions(+)
> > > >
> > > > diff --git a/block.c b/block.c
> > > > index 6cd39fa..6f7ad7f 100644
> > > > --- a/block.c
> > > > +++ b/block.c
> > > > @@ -306,6 +306,7 @@ BlockDriverState *bdrv_new(const char *device_name)
> > > > bdrv_iostatus_disable(bs);
> > > > notifier_list_init(&bs->close_notifiers);
> > > > notifier_with_return_list_init(&bs->before_write_notifiers);
> > > > + bs->refcnt = 1;
> > > >
> > > > return bs;
> > > > }
> > > > @@ -1511,6 +1512,9 @@ static void
> > > > bdrv_move_feature_fields(BlockDriverState *bs_dest,
> > > > /* dirty bitmap */
> > > > bs_dest->dirty_bitmap = bs_src->dirty_bitmap;
> > > >
> > > > + /* reference count */
> > > > + bs_dest->refcnt = bs_src->refcnt;
> > > > +
> > > > /* job */
> > > > bs_dest->in_use = bs_src->in_use;
> > > > bs_dest->job = bs_src->job;
> > > > @@ -4385,6 +4389,24 @@ int64_t bdrv_get_dirty_count(BlockDriverState
> > > > *bs)
> > > > }
> > > > }
> > > >
> > > > +/* Get a reference to bs */
> > > > +void bdrv_ref(BlockDriverState *bs)
> > > > +{
> > > > + bs->refcnt++;
> > > > +}
> > > > +
> > > > +/* Release a previously grabbed reference to bs.
> > > > + * If after releasing, reference count is zero, the BlockDriverState is
> > > > + * deleted. */
> > > > +void bdrv_unref(BlockDriverState *bs)
> > > > +{
> > > > + assert(bs->refcnt > 0);
> > > > + if (--bs->refcnt == 0) {
> > > > + bdrv_close(bs);
> > > > + bdrv_delete(bs);
> > > > + }
> > > > +}
> > >
> > > The problem with this is that a caller to bdrv_unref() has no
> > > way of knowing after calling bdrv_unref() if bs is still valid. We
> > > can't just set bs to NULL after calling bdrv_unref() as with
> > > bdrv_delete(), because now it may not have been freed.
> > >
> > By calling bdrv_unref, it means the caller is not going to use bs any
> > more. In other words, bdrv_unref() is a bdrv_delete() as seen by the
> > caller, if bs is still valid pointer after unref, it's no longer safe
> > for the caller: it can be freed by other code, in any time, but the
> > caller can't know.
>
> OK, I can go with that. I can't think off the top of my head where
> this would cause a problem if it is just used in place of current
> bdrv_delete(), so long as everyone refs it when they should.
>
> But then going with a bdrv_delete() equivalency model, I don't know if
> it is appropriate to call bdrv_unref() in bdrv_detach_dev() (in patch
> 5/8). Maybe all that is really needed there is some more cleanup in
> the places that call bdrv_detach_dev(), and then it would be OK.
Yes, you're right. I need to audit the device code yet. Thanks for
pointing out.
> >
> > > Maybe bdrv_unref should either return the current bs pointer, or
> > > alternatively accept as its argument a pointer to the BDS pointer:
> > >
> > > void bdrv_unref(BlockDriverState **bs)
> > > {
> > > assert(*bs->refcnt > 0);
> > > if (--*bs->refcnt == 0) {
> > > bdrv_close(*bs);
> > > bdrv_delete(*bs);
> > > *bs = NULL;
> > > }
> > > }
> > >
> > > Of course, all callers would need to then check for NULL.
> > >
> > > Also, do we need to call bdrv_close() in here? In bdrv_delete(),
> > > bdrv_close() is called prior to the free.
> > >
> > Yes, it can be omited.
> >
> > > > +
> > > > void bdrv_set_in_use(BlockDriverState *bs, int in_use)
> > > > {
> > > > assert(bs->in_use != in_use);
> > > > diff --git a/include/block/block.h b/include/block/block.h
> > > > index 742fce5..b33ef62 100644
> > > > --- a/include/block/block.h
> > > > +++ b/include/block/block.h
> > > > @@ -356,6 +356,8 @@ int64_t bdrv_get_dirty_count(BlockDriverState *bs);
> > > > void bdrv_enable_copy_on_read(BlockDriverState *bs);
> > > > void bdrv_disable_copy_on_read(BlockDriverState *bs);
> > > >
> > > > +void bdrv_ref(BlockDriverState *bs);
> > > > +void bdrv_unref(BlockDriverState *bs);
> > > > void bdrv_set_in_use(BlockDriverState *bs, int in_use);
> > > > int bdrv_in_use(BlockDriverState *bs);
> > > >
> > > > diff --git a/include/block/block_int.h b/include/block/block_int.h
> > > > index c6ac871..a282d56 100644
> > > > --- a/include/block/block_int.h
> > > > +++ b/include/block/block_int.h
> > > > @@ -294,6 +294,7 @@ struct BlockDriverState {
> > > > BlockDeviceIoStatus iostatus;
> > > > char device_name[32];
> > > > HBitmap *dirty_bitmap;
> > > > + int refcnt;
> > > > int in_use; /* users other than guest access, eg. block migration
> > > > */
> > > > QTAILQ_ENTRY(BlockDriverState) list;
> > > >
> > > > --
> > > > 1.8.3.2
> > > >
> > > >
> >
> > --
> > Fam
--
Fam
[Qemu-devel] [PATCH 2/8] iscsi: use bdrv_new() instead of stack structure, Fam Zheng, 2013/07/25
[Qemu-devel] [PATCH 4/8] block: make bdrv_delete() static, Fam Zheng, 2013/07/25
[Qemu-devel] [PATCH 5/8] block: use BlockDriverState refcnt for device attach/detach, Fam Zheng, 2013/07/25
[Qemu-devel] [PATCH 6/8] migration: omit drive ref as we have bdrv_ref now, Fam Zheng, 2013/07/25
[Qemu-devel] [PATCH 7/8] xen_disk: simplify blk_disconnect with refcnt, Fam Zheng, 2013/07/25
[Qemu-devel] [PATCH 8/8] nbd: use BlockDriverState refcnt, Fam Zheng, 2013/07/25