Linux Virtualization - May 2019 - [Qemu-devel] [PATCH v8 3/6] libnvdimm: add dax_dev sync flag

On Fri, May 10, 2019 at 5:45 PM Pankaj Gupta <pagupta at redhat.com>
wrote:
>
>
>
> > >
> > > This patch adds 'DAXDEV_SYNC' flag which is set
> > > for nd_region doing synchronous flush. This later
> > > is used to disable MAP_SYNC functionality for
> > > ext4 & xfs filesystem for devices don't support
> > > synchronous flush.
> > >
> > > Signed-off-by: Pankaj Gupta <pagupta at redhat.com>
> > > ---
> > >  drivers/dax/bus.c            |  2 +-
> > >  drivers/dax/super.c          | 13 ++++++++++++-
> > >  drivers/md/dm.c              |  3 ++-
> > >  drivers/nvdimm/pmem.c        |  5 ++++-
> > >  drivers/nvdimm/region_devs.c |  7 +++++++
> > >  include/linux/dax.h          |  8 ++++++--
> > >  include/linux/libnvdimm.h    |  1 +
> > >  7 files changed, 33 insertions(+), 6 deletions(-)
> > [..]
> > > diff --git a/drivers/md/dm.c b/drivers/md/dm.c
> > > index 043f0761e4a0..ee007b75d9fd 100644
> > > --- a/drivers/md/dm.c
> > > +++ b/drivers/md/dm.c
> > > @@ -1969,7 +1969,8 @@ static struct mapped_device *alloc_dev(int
minor)
> > >         sprintf(md->disk->disk_name, "dm-%d",
minor);
> > >
> > >         if (IS_ENABLED(CONFIG_DAX_DRIVER)) {
> > > -               dax_dev = alloc_dax(md,
md->disk->disk_name, &dm_dax_ops);
> > > +               dax_dev = alloc_dax(md,
md->disk->disk_name, &dm_dax_ops,
> > > +                                                       
DAXDEV_F_SYNC);
> >
> > Apologies for not realizing this until now, but this is broken.
> > Imaging a device-mapper configuration composed of both 'async'
> > virtio-pmem and 'sync' pmem. The 'sync' flag needs to
be unified
> > across all members. I would change this argument to '0' and
then
> > arrange for it to be set at dm_table_supports_dax() time after
> > validating that all components support synchronous dax.
>
> o.k. Need to set 'DAXDEV_F_SYNC' flag after verifying all the
target
> components support synchronous DAX.
>
> Just a question, If device mapper configuration have composed of both
> virtio-pmem or pmem devices, we want to configure device mapper for async
flush?
If it's composed of both then, yes, it needs to be async flush at the
device-mapper level. Otherwise MAP_SYNC will succeed and fail to
trigger fsync on the host file when necessary for the virtio-pmem
backed portion of the device-mapper device.

> > > >
> > > > This patch adds 'DAXDEV_SYNC' flag which is set
> > > > for nd_region doing synchronous flush. This later
> > > > is used to disable MAP_SYNC functionality for
> > > > ext4 & xfs filesystem for devices don't support
> > > > synchronous flush.
> > > >
> > > > Signed-off-by: Pankaj Gupta <pagupta at redhat.com>
> > > > ---
> > > >  drivers/dax/bus.c            |  2 +-
> > > >  drivers/dax/super.c          | 13 ++++++++++++-
> > > >  drivers/md/dm.c              |  3 ++-
> > > >  drivers/nvdimm/pmem.c        |  5 ++++-
> > > >  drivers/nvdimm/region_devs.c |  7 +++++++
> > > >  include/linux/dax.h          |  8 ++++++--
> > > >  include/linux/libnvdimm.h    |  1 +
> > > >  7 files changed, 33 insertions(+), 6 deletions(-)
> > > [..]
> > > > diff --git a/drivers/md/dm.c b/drivers/md/dm.c
> > > > index 043f0761e4a0..ee007b75d9fd 100644
> > > > --- a/drivers/md/dm.c
> > > > +++ b/drivers/md/dm.c
> > > > @@ -1969,7 +1969,8 @@ static struct mapped_device
*alloc_dev(int minor)
> > > >         sprintf(md->disk->disk_name,
"dm-%d", minor);
> > > >
> > > >         if (IS_ENABLED(CONFIG_DAX_DRIVER)) {
> > > > -               dax_dev = alloc_dax(md,
md->disk->disk_name,
> > > > &dm_dax_ops);
> > > > +               dax_dev = alloc_dax(md,
md->disk->disk_name,
> > > > &dm_dax_ops,
> > > > +
> > > > DAXDEV_F_SYNC);
> > >
> > > Apologies for not realizing this until now, but this is broken.
> > > Imaging a device-mapper configuration composed of both
'async'
> > > virtio-pmem and 'sync' pmem. The 'sync' flag
needs to be unified
> > > across all members. I would change this argument to '0'
and then
> > > arrange for it to be set at dm_table_supports_dax() time after
> > > validating that all components support synchronous dax.
> >
> > o.k. Need to set 'DAXDEV_F_SYNC' flag after verifying all the
target
> > components support synchronous DAX.
> >
> > Just a question, If device mapper configuration have composed of both
> > virtio-pmem or pmem devices, we want to configure device mapper for
async
> > flush?
> 
> If it's composed of both then, yes, it needs to be async flush at the
> device-mapper level. Otherwise MAP_SYNC will succeed and fail to
> trigger fsync on the host file when necessary for the virtio-pmem
> backed portion of the device-mapper device.
o.k. Agree.

Thanks you,
Pankaj
> 
>

Hi Dan,

While testing device mapper with DAX, I faced a bug with the commit:

commit ad428cdb525a97d15c0349fdc80f3d58befb50df
Author: Dan Williams <dan.j.williams at intel.com>
Date:   Wed Feb 20 21:12:50 2019 -0800

When I reverted the condition to old code[1] it worked for me. I 
am thinking when we map two different devices (e.g with device mapper), will 
start & end pfn still point to same pgmap? Or there is something else which
I am missing here. 

Note: I tested only EXT4.

[1]

-               if (pgmap && pgmap->type == MEMORY_DEVICE_FS_DAX)
+               end_pgmap = get_dev_pagemap(pfn_t_to_pfn(end_pfn), NULL);
+               if (pgmap && pgmap == end_pgmap &&
pgmap->type == MEMORY_DEVICE_FS_DAX
+                               && pfn_t_to_page(pfn)->pgmap == pgmap
+                               && pfn_t_to_page(end_pfn)->pgmap ==
pgmap
+                               && pfn_t_to_pfn(pfn) ==
PHYS_PFN(__pa(kaddr))
+                               && pfn_t_to_pfn(end_pfn) ==
PHYS_PFN(__pa(end_kaddr)))
                        dax_enabled = true;
                put_dev_pagemap(pgmap);

Thanks,
Pankaj