lkml.org 
[lkml]   [2018]   [Apr]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [RFC v2 2/2] pmem: device flush over VIRTIO
On Thu, Apr 26, 2018 at 9:40 AM, Pankaj Gupta <pagupta@redhat.com> wrote:
>
>>
>> On Wed, Apr 25, 2018 at 04:54:14PM +0530, Pankaj Gupta wrote:
>> > This patch adds functionality to perform
>> > flush from guest to hosy over VIRTIO
>> > when 'ND_REGION_VIRTIO'flag is set on
>> > nd_negion. Flag is set by 'virtio-pmem'
>> > driver.
>> >
>> > Signed-off-by: Pankaj Gupta <pagupta@redhat.com>
>> > ---
>> > drivers/nvdimm/region_devs.c | 7 +++++++
>> > 1 file changed, 7 insertions(+)
>> >
>> > diff --git a/drivers/nvdimm/region_devs.c b/drivers/nvdimm/region_devs.c
>> > index a612be6..6c6454e 100644
>> > --- a/drivers/nvdimm/region_devs.c
>> > +++ b/drivers/nvdimm/region_devs.c
>> > @@ -20,6 +20,7 @@
>> > #include <linux/nd.h>
>> > #include "nd-core.h"
>> > #include "nd.h"
>> > +#include <linux/virtio_pmem.h>
>> >
>> > /*
>> > * For readq() and writeq() on 32-bit builds, the hi-lo, lo-hi order is
>> > @@ -1074,6 +1075,12 @@ void nvdimm_flush(struct nd_region *nd_region)
>> > struct nd_region_data *ndrd = dev_get_drvdata(&nd_region->dev);
>> > int i, idx;
>> >
>> > + /* call PV device flush */
>> > + if (test_bit(ND_REGION_VIRTIO, &nd_region->flags)) {
>> > + virtio_pmem_flush(&nd_region->dev);
>> > + return;
>> > + }
>>
>> How does libnvdimm know when flush has completed?
>>
>> Callers expect the flush to be finished when nvdimm_flush() returns but
>> the virtio driver has only queued the request, it hasn't waited for
>> completion!
>
> I tried to implement what nvdimm does right now. It just writes to
> flush hint address to make sure data persists.

nvdimm_flush() is currently expected to be synchronous. Currently it
is sfence(); write to special address; sfence(). By the time the
second sfence returns the data is flushed. So you would need to make
this virtio flush interface synchronous as well, but that appears
problematic to stop the guest for unbounded amounts of time. Instead,
you need to rework nvdimm_flush() and the pmem driver to make these
flush requests asynchronous and add the plumbing for completion
callbacks via bio_endio().

> I just did not want to block guest write requests till host side
> fsync completes.

You must complete the flush before bio_endio(), otherwise you're
violating the expectations of the guest filesystem/block-layer.

>
> be worse for operations on different guest files because all these operations would happen
> ultimately on same file at host.
>
> I think with current way, we can achieve an asynchronous queuing mechanism on cost of not
> 100% sure when fsync would complete but it is assured it will happen. Also, its entire block
> flush.

No, again, that's broken. We need to add the plumbing for
communicating the fsync() completion relative the WRITE_{FLUSH,FUA}
bio in the guest.

\
 
 \ /
  Last update: 2018-04-26 18:59    [W:0.092 / U:2.888 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site