Re: [PATCH v2 2/4] libnvdimm/namespace: Enforce memremap_compat_align()

From: Dan Williams <dan.j.williams@intel.com>
To: Jeff Moyer <jmoyer@redhat.com>
Cc: linux-nvdimm <linux-nvdimm@lists.01.org>,
	"Aneesh Kumar K.V" <aneesh.kumar@linux.ibm.com>,
	Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
	linuxppc-dev <linuxppc-dev@lists.ozlabs.org>
Subject: Re: [PATCH v2 2/4] libnvdimm/namespace: Enforce memremap_compat_align()
Date: Thu, 13 Feb 2020 14:43:28 -0800	[thread overview]
Message-ID: <CAPcyv4hQouRNBcJ4uZ2mysr_aKstLhvUf66gRQ_3QoQNyOy72g@mail.gmail.com> (raw)
In-Reply-To: <x49k14q5ezs.fsf@segfault.boston.devel.redhat.com>

On Thu, Feb 13, 2020 at 1:55 PM Jeff Moyer <jmoyer@redhat.com> wrote:
>
> Dan Williams <dan.j.williams@intel.com> writes:
>
> > The pmem driver on PowerPC crashes with the following signature when
> > instantiating misaligned namespaces that map their capacity via
> > memremap_pages().
> >
> >     BUG: Unable to handle kernel data access at 0xc001000406000000
> >     Faulting instruction address: 0xc000000000090790
> >     NIP [c000000000090790] arch_add_memory+0xc0/0x130
> >     LR [c000000000090744] arch_add_memory+0x74/0x130
> >     Call Trace:
> >      arch_add_memory+0x74/0x130 (unreliable)
> >      memremap_pages+0x74c/0xa30
> >      devm_memremap_pages+0x3c/0xa0
> >      pmem_attach_disk+0x188/0x770
> >      nvdimm_bus_probe+0xd8/0x470
> >
> > With the assumption that only memremap_pages() has alignment
> > constraints, enforce memremap_compat_align() for
> > pmem_should_map_pages(), nd_pfn, or nd_dax cases.
> >
> > Reported-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
> > Cc: Jeff Moyer <jmoyer@redhat.com>
> > Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
> > Link: https://lore.kernel.org/r/158041477336.3889308.4581652885008605170.stgit@dwillia2-desk3.amr.corp.intel.com
> > Signed-off-by: Dan Williams <dan.j.williams@intel.com>
> > ---
> >  drivers/nvdimm/namespace_devs.c |   10 ++++++++++
> >  1 file changed, 10 insertions(+)
> >
> > diff --git a/drivers/nvdimm/namespace_devs.c b/drivers/nvdimm/namespace_devs.c
> > index 032dc61725ff..aff1f32fdb4f 100644
> > --- a/drivers/nvdimm/namespace_devs.c
> > +++ b/drivers/nvdimm/namespace_devs.c
> > @@ -1739,6 +1739,16 @@ struct nd_namespace_common *nvdimm_namespace_common_probe(struct device *dev)
> >               return ERR_PTR(-ENODEV);
> >       }
> >
> > +     if (pmem_should_map_pages(dev) || nd_pfn || nd_dax) {
> > +             struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
> > +             resource_size_t start = nsio->res.start;
> > +
> > +             if (!IS_ALIGNED(start | size, memremap_compat_align())) {
> > +                     dev_dbg(&ndns->dev, "misaligned, unable to map\n");
> > +                     return ERR_PTR(-EOPNOTSUPP);
> > +             }
> > +     }
> > +
> >       if (is_namespace_pmem(&ndns->dev)) {
> >               struct nd_namespace_pmem *nspm;
> >
>
> Actually, I take back my ack.  :) This prevents a previously working
> namespace from being successfully probed/setup.

Do you have a test case handy? I can see a potential gap with a
namespace that used internal padding to fix up the alignment. The goal
of this check is to catch cases that are just going to fail
devm_memremap_pages(), and the expectation is that it could not have
worked before unless it was ported from another platform, or someone
flipped the page-size switch on PowerPC.

> I thought we were only
> going to enforce the alignment for a newly created namespace?  This should
> only check whether the alignment works for the current platform.

The model is a new default 16MB alignment is enforced at creation
time, but if you need to support previously created namespaces then
you can manually trim that alignment requirement to no less than
memremap_compat_align() because that's the point at which
devm_memremap_pages() will start failing or crashing.
_______________________________________________
Linux-nvdimm mailing list -- linux-nvdimm@lists.01.org
To unsubscribe send an email to linux-nvdimm-leave@lists.01.org