Re: [PATCH] qcow2: flush qcow2 l2 meta for new allocated clusters

[Kevin Wolf]

Am 07.08.2020 um 09:42 hat Ying Fang geschrieben:
> 
> 
> On 8/6/2020 5:13 PM, Kevin Wolf wrote:
> > Am 05.08.2020 um 04:38 hat Ying Fang geschrieben:
> > > From: fangying <fangying1@huawei.com>
> > > 
> > > When qemu or qemu-nbd process uses a qcow2 image and configured with
> > > 'cache = none', it will write to the qcow2 image with a cache to cache
> > > L2 tables, however the process will not use L2 tables without explicitly
> > > calling the flush command or closing the mirror flash into the disk.
> > > Which may cause the disk data inconsistent with the written data for
> > > a long time. If an abnormal process exit occurs here, the issued written
> > > data will be lost.
> > > 
> > > Therefore, in order to keep data consistency we need to flush the changes
> > > to the L2 entry to the disk in time for the newly allocated cluster.
> > > 
> > > Signed-off-by: Ying Fang <fangying1@huawei.com>
> > 
> > If you want to have data safely written to the disk after each write
> > request, you need to use cache=writethrough/directsync (in other words,
> > aliases that are equivalent to setting -device ...,write-cache=off).
> > Note that this will have a major impact on write performance.
> > 
> > cache=none means bypassing the kernel page cache (O_DIRECT), but not
> > flushing after each write request.
> 
> Well, IIUC, cache=none does not guarantee data safety and we should not
> expect that. Then this patch can be ignored.

Indeed, cache=none is a writeback cache mode with all of the
consequences. In practice, this is normally good enough because the
guest OS will send flush requests when needed (e.g. because a guest
application called fsync()), but if the guest doesn't do this, it may
suffer data loss. This behaviour is comparable to a volatile disk cache
on real hard disks and is a good default, but sometimes you need a
writethrough cache mode at the cost of a performance penalty.

Kevin