Re: [Qemu-devel] [PATCH v2 4/8] mem/nvdimm: ensure write persistence to PMEM in label emulation

[Haozhong Zhang]

On 02/09/18 14:27 +0000, Stefan Hajnoczi wrote:
> On Wed, Feb 07, 2018 at 03:33:27PM +0800, Haozhong Zhang wrote:
> > @@ -156,11 +157,17 @@ static void nvdimm_write_label_data(NVDIMMDevice 
> > *nvdimm, const void *buf,
> >  {
> >      MemoryRegion *mr;
> >      PCDIMMDevice *dimm = PC_DIMM(nvdimm);
> > +    bool is_pmem = object_property_get_bool(OBJECT(dimm->hostmem),
> > +                                            "pmem", NULL);
> >      uint64_t backend_offset;
> >  
> >      nvdimm_validate_rw_label_data(nvdimm, size, offset);
> >  
> > -    memcpy(nvdimm->label_data + offset, buf, size);
> > +    if (!is_pmem) {
> > +        memcpy(nvdimm->label_data + offset, buf, size);
> > +    } else {
> > +        pmem_memcpy_persist(nvdimm->label_data + offset, buf, size);
> > +    }
> 
> Is this enough to prevent label corruption in case of power failure?
> 
> pmem_memcpy_persist() is not atomic.  Power failure can result in a mix
> of the old and new label data.
> 
> If we want this operation to be 100% safe there needs to be some kind of
> update protocol that makes the change atomic, like a Label A and Label B
> area with a single Label Index field that can be updated atomically to
> point to the active Label A/B area.

All this patch series is to guarantee: if the guest is still alive and
running, all its previous writes to pmem, which were performed by
QEMU, will be still persistent on pmem.

If a power failure happens before QEMU returns to the guest, e.g., in
the middle of above pmem_memcpy_persist(), yes, the guest label data
may be in an inconsistent state, but the guest also has no chance to
progress.  And, that is what could happen in the non-virtualization
environment as well, and it's the responsibility of the (guest) SW to
defend such failures, e.g., by the protocol you mentioned.

Haozhong