[RFC PATCH 1/2] libnvdimm: Add prctl control for disabling synchronous fault support.

[Dan Williams]

On Sat, May 30, 2020 at 12:18 AM Aneesh Kumar K.V
<aneesh.kumar at linux.ibm.com> wrote:
>
> On 5/30/20 12:52 AM, Dan Williams wrote:
> > On Fri, May 29, 2020 at 3:55 AM Aneesh Kumar K.V
> > <aneesh.kumar at linux.ibm.com> wrote:
> >>
> >> On 5/29/20 3:22 PM, Jan Kara wrote:
> >>> Hi!
> >>>
> >>> On Fri 29-05-20 15:07:31, Aneesh Kumar K.V wrote:
> >>>> Thanks Michal. I also missed Jeff in this email thread.
> >>>
> >>> And I think you'll also need some of the sched maintainers for the prctl
> >>> bits...
> >>>
> >>>> On 5/29/20 3:03 PM, Michal Suchánek wrote:
> >>>>> Adding Jan
> >>>>>
> >>>>> On Fri, May 29, 2020 at 11:11:39AM +0530, Aneesh Kumar K.V wrote:
> >>>>>> With POWER10, architecture is adding new pmem flush and sync instructions.
> >>>>>> The kernel should prevent the usage of MAP_SYNC if applications are not using
> >>>>>> the new instructions on newer hardware.
> >>>>>>
> >>>>>> This patch adds a prctl option MAP_SYNC_ENABLE that can be used to enable
> >>>>>> the usage of MAP_SYNC. The kernel config option is added to allow the user
> >>>>>> to control whether MAP_SYNC should be enabled by default or not.
> >>>>>>
> >>>>>> Signed-off-by: Aneesh Kumar K.V <aneesh.kumar at linux.ibm.com>
> >>> ...
> >>>>>> diff --git a/kernel/fork.c b/kernel/fork.c
> >>>>>> index 8c700f881d92..d5a9a363e81e 100644
> >>>>>> --- a/kernel/fork.c
> >>>>>> +++ b/kernel/fork.c
> >>>>>> @@ -963,6 +963,12 @@ __cacheline_aligned_in_smp DEFINE_SPINLOCK(mmlist_lock);
> >>>>>>     static unsigned long default_dump_filter = MMF_DUMP_FILTER_DEFAULT;
> >>>>>> +#ifdef CONFIG_ARCH_MAP_SYNC_DISABLE
> >>>>>> +unsigned long default_map_sync_mask = MMF_DISABLE_MAP_SYNC_MASK;
> >>>>>> +#else
> >>>>>> +unsigned long default_map_sync_mask = 0;
> >>>>>> +#endif
> >>>>>> +
> >>>
> >>> I'm not sure CONFIG is really the right approach here. For a distro that would
> >>> basically mean to disable MAP_SYNC for all PPC kernels unless application
> >>> explicitly uses the right prctl. Shouldn't we rather initialize
> >>> default_map_sync_mask on boot based on whether the CPU we run on requires
> >>> new flush instructions or not? Otherwise the patch looks sensible.
> >>>
> >>
> >> yes that is correct. We ideally want to deny MAP_SYNC only w.r.t
> >> POWER10. But on a virtualized platform there is no easy way to detect
> >> that. We could ideally hook this into the nvdimm driver where we look at
> >> the new compat string ibm,persistent-memory-v2 and then disable MAP_SYNC
> >> if we find a device with the specific value.
> >>
> >> BTW with the recent changes I posted for the nvdimm driver, older kernel
> >> won't initialize persistent memory device on newer hardware. Newer
> >> hardware will present the device to OS with a different device tree
> >> compat string.
> >>
> >> My expectation  w.r.t this patch was, Distro would want to  mark
> >> CONFIG_ARCH_MAP_SYNC_DISABLE=n based on the different application
> >> certification.  Otherwise application will have to end up calling the
> >> prctl(MMF_DISABLE_MAP_SYNC, 0) any way. If that is the case, should this
> >> be dependent on P10?
> >>
> >> With that I am wondering should we even have this patch? Can we expect
> >> userspace get updated to use new instruction?.
> >>
> >> With ppc64 we never had a real persistent memory device available for
> >> end user to try. The available persistent memory stack was using vPMEM
> >> which was presented as a volatile memory region for which there is no
> >> need to use any of the flush instructions. We could safely assume that
> >> as we get applications certified/verified for working with pmem device
> >> on ppc64, they would all be using the new instructions?
> >
> > I think prctl is the wrong interface for this. I was thinking a sysfs
> > interface along the same lines as /sys/block/pmemX/dax/write_cache.
> > That attribute is toggling DAXDEV_WRITE_CACHE for the determination of
> > whether the platform or the kernel needs to handle cache flushing
> > relative to power loss. A similar attribute can be established for
> > DAXDEV_SYNC, it would simply default to off based on a configuration
> > time policy, but be dynamically changeable at runtime via sysfs.
> >
> > These flags are device properties that affect the kernel and
> > userspace's handling of persistence.
> >
>
> That will not handle the scenario with multiple applications using the
> same fsdax mount point where one is updated to use the new instruction
> and the other is not.

Right, it needs to be a global setting / flag day to switch from one
regime to another. Per-process control is a recipe for disaster.