Re: [RFC PATCH rdma-next 01/10] RDMA: mr: Introduce is_pmem

[Dan Williams <dan.j.williams@intel.com>]

On Tue, Jan 18, 2022 at 12:55 AM lizhijian@fujitsu.com
<lizhijian@fujitsu.com> wrote:
>
>
>
> On 17/01/2022 02:11, Dan Williams wrote:
> > On Wed, Jan 5, 2022 at 10:13 PM lizhijian@fujitsu.com
> > <lizhijian@fujitsu.com> wrote:
> >>
> >> Add Dan to the party :)
> >>
> >> May i know whether there is any existing APIs to check whether
> >> a va/page backs to a nvdimm/pmem ?
> >>
> >>
> >>
> >> On 06/01/2022 08:21, Jason Gunthorpe wrote:
> >>> On Tue, Dec 28, 2021 at 04:07:08PM +0800, Li Zhijian wrote:
> >>>> We can use it to indicate whether the registering mr is associated with
> >>>> a pmem/nvdimm or not.
> >>>>
> >>>> Currently, we only assign it in rxe driver, for other device/drivers,
> >>>> they should implement it if needed.
> >>>>
> >>>> RDMA FLUSH will support the persistence feature for a pmem/nvdimm.
> >>>>
> >>>> Signed-off-by: Li Zhijian <lizhijian@cn.fujitsu.com>
> >>>>    drivers/infiniband/sw/rxe/rxe_mr.c | 47 ++++++++++++++++++++++++++++++
> >>>>    include/rdma/ib_verbs.h            |  1 +
> >>>>    2 files changed, 48 insertions(+)
> >>>>
> >>>> diff --git a/drivers/infiniband/sw/rxe/rxe_mr.c b/drivers/infiniband/sw/rxe/rxe_mr.c
> >>>> index 7c4cd19a9db2..bcd5e7afa475 100644
> >>>> +++ b/drivers/infiniband/sw/rxe/rxe_mr.c
> >>>> @@ -162,6 +162,50 @@ void rxe_mr_init_dma(struct rxe_pd *pd, int access, struct rxe_mr *mr)
> >>>>       mr->type = IB_MR_TYPE_DMA;
> >>>>    }
> >>>>
> >>>> +// XXX: the logic is similar with mm/memory-failure.c
> >>>> +static bool page_in_dev_pagemap(struct page *page)
> >>>> +{
> >>>> +    unsigned long pfn;
> >>>> +    struct page *p;
> >>>> +    struct dev_pagemap *pgmap = NULL;
> >>>> +
> >>>> +    pfn = page_to_pfn(page);
> >>>> +    if (!pfn) {
> >>>> +            pr_err("no such pfn for page %p\n", page);
> >>>> +            return false;
> >>>> +    }
> >>>> +
> >>>> +    p = pfn_to_online_page(pfn);
> >>>> +    if (!p) {
> >>>> +            if (pfn_valid(pfn)) {
> >>>> +                    pgmap = get_dev_pagemap(pfn, NULL);
> >>>> +                    if (pgmap)
> >>>> +                            put_dev_pagemap(pgmap);
> >>>> +            }
> >>>> +    }
> >>>> +
> >>>> +    return !!pgmap;
> >>> You need to get Dan to check this out, but I'm pretty sure this should
> >>> be more like this:
> >>>
> >>> if (is_zone_device_page(page) && page->pgmap->type == MEMORY_DEVICE_FS_DAX)
> >> Great, i have added him.
> >>
> >>
> >>
> >>>
> >>>> +static bool iova_in_pmem(struct rxe_mr *mr, u64 iova, int length)
> >>>> +{
> >>>> +    struct page *page = NULL;
> >>>> +    char *vaddr = iova_to_vaddr(mr, iova, length);
> >>>> +
> >>>> +    if (!vaddr) {
> >>>> +            pr_err("not a valid iova %llu\n", iova);
> >>>> +            return false;
> >>>> +    }
> >>>> +
> >>>> +    page = virt_to_page(vaddr);
> >>> And obviously this isn't uniform for the entire umem, so I don't even
> >>> know what this is supposed to mean.
> >> My intention is to check if a memory region belongs to a nvdimm/pmem.
> >> The approach is like that:
> >> iova(user space)-+                     +-> page -> page_in_dev_pagemap()
> >>                    |                     |
> >>                    +-> va(kernel space) -+
> >> Since current MR's va is associated with map_set where it record the relations
> >> between iova and va and page. Do do you mean we should travel map_set to
> >> get its page ? or by any other ways.
> > Apologies for the delay in responding.
> >
> > The Subject line of this patch is confusing, if you want to know if a
> > pfn is in persistent memory the only mechanism for that is:
> >
> > region_intersects(addr, length, IORESOURCE_MEM, IORES_DESC_PERSISTENT_MEMORY)
> >
> > ...there is otherwise nothing pmem specific about the dev_pagemap
> > infrastructure. Yes, pmem is the primary user, but it is also used for
> > mapping "soft-reserved" memory (See: the EFI_MEMORY_SP) attribute, and
> > other users.
> >
> > Can you clarify the intent? I am missing some context.
>
> thanks for your help @Dan
>
> I'm going to implement a new ibvers called RDMA FLUSH, it will support global visibility
> and persistence placement type.
>
> In my first design, only pmem can support persistence placement type, so i need to introduce
> new attribute is_pmem to RDMA memory region where it associates to a user space address iova
> so that i can reject a persistence placement type to DRAM(non-pmem).

Ok, I think for that case you are better served using the
IORES_DESC_PERSISTENT_MEMORY designation as the gate for attempting
the flush. The dev_pagemap is otherwise only indicating the presence
of device-backed memory pages, not persistence.