On Mon, Feb 02, 2015 at 10:48:00AM +0900, Sergey Senozhatsky wrote:
> Hello Minchan,
>
> On (02/02/15 10:30), Minchan Kim wrote:
> > > > static inline int init_done(struct zram *zram)
> > > > {
> > > > - return zram->meta != NULL;
> > > > + return zram->disksize != 0;
> > >
> > > we don't set ->disksize to 0 when create device. and I think
> > > it's better to use refcount here, but set it to 0 during device creation.
> > > (see the patch below)
> >
> > There was a reason I didn't use refcount there.
> > I should have written down it.
> >
> > We need something to prevent further I/O handling on other CPUs.
> > Otherwise, it's livelock. For example, new 'A' I/O rw path on CPU 1
> > can see non-zero refcount if another CPU is going on rw.
> > Then, another new 'B' I/O rw path on CPU 2 can see non-zero refcount
> > if A I/O is going on. Then, another new 'C' I/O rw path on CPU 3 can
> > see non-zero refcount if B I/O is going on. Finally, 'A' IO is done
> > on CPU 1 and next I/O 'D' on CPU 1 can see non-zero refcount because
> > 'C' on CPU 3 is going on. Infinite loop.
>
> sure, I did think about this. and I actually didn't find any reason not
> to use ->refcount there. if user wants to reset the device, he first
> should umount it to make bdev->bd_holders check happy. and that's where
> IOs will be failed. so it makes sense to switch to ->refcount there, IMHO.
If we use zram as block device itself(not a fs or swap) and open the
block device as !FMODE_EXCL, bd_holders will be void.
Another topic: As I didn't see enough fs/block_dev.c bd_holders in zram
would be mess. I guess we need to study hotplug of device and implement
it for zram reset rather than strange own konb. It should go TODO. :(
>
>
> > > here and later:
> > > we can't take zram_meta_get() first and then check for init_done(zram),
> > > because ->meta can be NULL, so it fill be ->NULL->refcount.
> >
> > True.
> > Actually, it was totally RFC I forgot adding the tag in the night but I can't
> > escape from my shame with the escuse. Thanks!
>
> no problem at all. you were throwing solutions all week long.
>
> >
> > >
> > > let's keep ->completion and ->refcount in zram and rename zram_meta_[get|put]
> > > to zram_[get|put].
> >
> > Good idea but still want to name it as zram_meta_get/put because zram_get naming
> > might confuse struct zram's refcount rather than zram_meta. :)
>
> no objections. but I assume we agreed to keep ->io_done completion
> and ->refcount in zram.
>
> -ss
--
Kind regards,
Minchan Kim
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>