2021-08-27 14:42:52

by Li Zhijian

[permalink] [raw]
Subject: [PATCH] mm/hmm: bypass devmap pte when all pfn requested flags are fulfilled

Previously, we noticed the one rpma example was failed[1] since 36f30e486d,
where it will use ODP feature to do RDMA WRITE between fsdax files.

After digging into the code, we found hmm_vma_handle_pte() will still
return EFAULT even though all the its requesting flags has been
fulfilled. That's because a DAX page will be marked as
(_PAGE_SPECIAL | PAGE_DEVMAP) by pte_mkdevmap().

[1]: https://github.com/pmem/rpma/issues/1142

CC: [email protected]
Signed-off-by: Li Zhijian <[email protected]>
---
mm/hmm.c | 7 +++++++
1 file changed, 7 insertions(+)

diff --git a/mm/hmm.c b/mm/hmm.c
index fad6be2bf072..4766bdefb6c3 100644
--- a/mm/hmm.c
+++ b/mm/hmm.c
@@ -294,6 +294,12 @@ static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
if (required_fault)
goto fault;

+ /*
+ * just bypass devmap pte such as DAX page when all pfn requested
+ * flags(pfn_req_flags) are fulfilled.
+ */
+ if (pte_devmap(pte))
+ goto out;
/*
* Since each architecture defines a struct page for the zero page, just
* fall through and treat it like a normal page.
@@ -307,6 +313,7 @@ static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
return 0;
}

+out:
*hmm_pfn = pte_pfn(pte) | cpu_flags;
return 0;

--
2.31.1




2021-08-27 16:31:09

by Jason Gunthorpe

[permalink] [raw]
Subject: Re: [PATCH] mm/hmm: bypass devmap pte when all pfn requested flags are fulfilled

On Fri, Aug 27, 2021 at 10:45:00PM +0800, Li Zhijian wrote:
> Previously, we noticed the one rpma example was failed[1] since 36f30e486d,
> where it will use ODP feature to do RDMA WRITE between fsdax files.
>
> After digging into the code, we found hmm_vma_handle_pte() will still
> return EFAULT even though all the its requesting flags has been
> fulfilled. That's because a DAX page will be marked as
> (_PAGE_SPECIAL | PAGE_DEVMAP) by pte_mkdevmap().
>
> [1]: https://github.com/pmem/rpma/issues/1142
>
> CC: [email protected]
> Signed-off-by: Li Zhijian <[email protected]>

You need to add a

Fixes: 405506274922 ("mm/hmm: add missing call to hmm_pte_need_fault in HMM_PFN_SPECIAL handling")

> diff --git a/mm/hmm.c b/mm/hmm.c
> index fad6be2bf072..4766bdefb6c3 100644
> +++ b/mm/hmm.c
> @@ -294,6 +294,12 @@ static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
> if (required_fault)
> goto fault;
>
> + /*
> + * just bypass devmap pte such as DAX page when all pfn requested
> + * flags(pfn_req_flags) are fulfilled.
> + */
> + if (pte_devmap(pte))
> + goto out;

I liked your ealier version better where this was added to the
pte_special test - logically this is about disambiguating the
pte_special and the devmap case as they are different things.

Jason

2021-08-27 17:48:45

by Christoph Hellwig

[permalink] [raw]
Subject: Re: [PATCH] mm/hmm: bypass devmap pte when all pfn requested flags are fulfilled

On Fri, Aug 27, 2021 at 01:28:52PM -0300, Jason Gunthorpe wrote:
> > + /*
> > + * just bypass devmap pte such as DAX page when all pfn requested
> > + * flags(pfn_req_flags) are fulfilled.
> > + */
> > + if (pte_devmap(pte))
> > + goto out;
>
> I liked your ealier version better where this was added to the
> pte_special test - logically this is about disambiguating the
> pte_special and the devmap case as they are different things.

Yes, I think that is much more logical. Also please capitalize the
first word in multi-line comments.