diff mbox series

[for-next,4/5] RDMA/rxe: refactor iova_to_vaddr

Message ID 1668141030-2-5-git-send-email-lizhijian@fujitsu.com (mailing list archive)
State Changes Requested
Delegated to: Jason Gunthorpe
Headers show
Series iova_to_vaddr refactor | expand

Commit Message

Li Zhijian Nov. 11, 2022, 4:30 a.m. UTC
For IB_MR_TYPE_USER MR, iova_to_vaddr() will call kmap_local_page() to
map page.

Signed-off-by: Li Zhijian <lizhijian@fujitsu.com>
---
 drivers/infiniband/sw/rxe/rxe_loc.h   |  1 +
 drivers/infiniband/sw/rxe/rxe_mr.c    | 38 +++++++++++++++------------
 drivers/infiniband/sw/rxe/rxe_resp.c  |  1 +
 drivers/infiniband/sw/rxe/rxe_verbs.h |  5 +++-
 4 files changed, 27 insertions(+), 18 deletions(-)

Comments

Fabio M. De Francesco Nov. 16, 2022, 12:37 p.m. UTC | #1
On venerdì 11 novembre 2022 05:30:29 CET Li Zhijian wrote:
> For IB_MR_TYPE_USER MR, iova_to_vaddr() will call kmap_local_page() to
> map page.
>
> Signed-off-by: Li Zhijian <lizhijian@fujitsu.com>
> ---
>  drivers/infiniband/sw/rxe/rxe_loc.h   |  1 +
>  drivers/infiniband/sw/rxe/rxe_mr.c    | 38 +++++++++++++++------------
>  drivers/infiniband/sw/rxe/rxe_resp.c  |  1 +
>  drivers/infiniband/sw/rxe/rxe_verbs.h |  5 +++-
>  4 files changed, 27 insertions(+), 18 deletions(-)
>
> diff --git a/drivers/infiniband/sw/rxe/rxe_loc.h
> b/drivers/infiniband/sw/rxe/rxe_loc.h index c2a5c8814a48..22a8c44d39c8
100644
> --- a/drivers/infiniband/sw/rxe/rxe_loc.h
> +++ b/drivers/infiniband/sw/rxe/rxe_loc.h
> @@ -73,6 +73,7 @@ int rxe_mr_copy(struct rxe_mr *mr, u64 iova, void *addr,
int
> length, int copy_data(struct rxe_pd *pd, int access, struct rxe_dma_info
> *dma, void *addr, int length, enum rxe_mr_copy_dir dir);
>  void *iova_to_vaddr(struct rxe_mr *mr, u64 iova, int length);
> +void rxe_unmap_vaddr(struct rxe_mr *mr, void *vaddr);
>  struct rxe_mr *lookup_mr(struct rxe_pd *pd, int access, u32 key,
>                        enum rxe_mr_lookup_type type);
>  int mr_check_range(struct rxe_mr *mr, u64 iova, size_t length);
> diff --git a/drivers/infiniband/sw/rxe/rxe_mr.c
> b/drivers/infiniband/sw/rxe/rxe_mr.c index a4e786b657b7..d26a4a33119c 100644
> --- a/drivers/infiniband/sw/rxe/rxe_mr.c
> +++ b/drivers/infiniband/sw/rxe/rxe_mr.c
> @@ -120,9 +120,7 @@ int rxe_mr_init_user(struct rxe_dev *rxe, u64 start, u64
> length, u64 iova, struct ib_umem              *umem;
>       struct sg_page_iter     sg_iter;
>       int                     num_buf;
> -     void                    *vaddr;
>       int err;
> -     int i;
>
>       umem = ib_umem_get(&rxe->ib_dev, start, length, access);
>       if (IS_ERR(umem)) {
> @@ -159,18 +157,9 @@ int rxe_mr_init_user(struct rxe_dev *rxe, u64 start,
u64
> length, u64 iova, num_buf = 0;
>                       }
>
> -                     vaddr =
page_address(sg_page_iter_page(&sg_iter));
> -                     if (!vaddr) {
> -                             pr_warn("%s: Unable to get virtual
address\n",
> -                                             __func__);
> -                             err = -ENOMEM;
> -                             goto err_cleanup_map;
> -                     }
> -
> -                     buf->addr = (uintptr_t)vaddr;
> +                     buf->page = sg_page_iter_page(&sg_iter);
>                       num_buf++;
>                       buf++;
> -
>               }
>       }
>
> @@ -182,10 +171,6 @@ int rxe_mr_init_user(struct rxe_dev *rxe, u64 start,
u64
> length, u64 iova,
>
>       return 0;
>
> -err_cleanup_map:
> -     for (i = 0; i < mr->num_map; i++)
> -             kfree(mr->map[i]);
> -     kfree(mr->map);
>  err_release_umem:
>       ib_umem_release(umem);
>  err_out:
> @@ -246,6 +231,12 @@ static void lookup_iova(struct rxe_mr *mr, u64 iova,
int
> *m_out, int *n_out, }
>  }
>
> +void rxe_unmap_vaddr(struct rxe_mr *mr, void *vaddr)
> +{
> +     if (mr->ibmr.type == IB_MR_TYPE_USER)
> +             kunmap_local(vaddr);
> +}
> +
>  static void *__iova_to_vaddr(struct rxe_mr *mr, u64 iova, int length)
>  {
>       size_t offset;
> @@ -258,9 +249,21 @@ static void *__iova_to_vaddr(struct rxe_mr *mr, u64
iova,
> int length) return NULL;
>       }
>
> -     return (void *)(uintptr_t)mr->map[m]->buf[n].addr + offset;
> +     if (mr->ibmr.type == IB_MR_TYPE_USER) {
> +             char *paddr;
> +             struct page *pg = mr->map[m]->buf[n].page;
> +
> +             paddr = kmap_local_page(pg);
> +             if (paddr == NULL) {
> +                     pr_warn("Failed to map page");
> +                     return NULL;
> +             }

I know nothing about this code but I am here as a result of regular checks for
changes to HIGHMEM mappings across the entire kernel. So please forgive me if
I'm objecting to the correct changes.

1) It looks like this code had a call to page_address() and you converted it
to mapping with kmap_local_page().

If page_address() is related and it used to work properly, the page you are
mapping cannot come from ZONE_HIGHMEM. Therefore, kmap_local_page() looks like
an overkill.

I'm probably missing something...

> +             return paddr + offset;
> +     } else
> +             return (void *)(uintptr_t)mr->map[m]->buf[n].addr +
offset;
>  }
>
> +/* must call rxe_unmap_vaddr to unmap vaddr */
>  void *iova_to_vaddr(struct rxe_mr *mr, u64 iova, int length)
>  {
>       if (mr->state != RXE_MR_STATE_VALID) {
> @@ -326,6 +329,7 @@ int rxe_mr_copy(struct rxe_mr *mr, u64 iova, void *addr,
> int length, dest = (dir == RXE_TO_MR_OBJ) ? va : addr;
>
>               memcpy(dest, src, bytes);
> +             rxe_unmap_vaddr(mr, va);
>
>               length  -= bytes;
>               addr    += bytes;
> diff --git a/drivers/infiniband/sw/rxe/rxe_resp.c
> b/drivers/infiniband/sw/rxe/rxe_resp.c index 483043dc4e89..765cb9f8538a
> 100644
> --- a/drivers/infiniband/sw/rxe/rxe_resp.c
> +++ b/drivers/infiniband/sw/rxe/rxe_resp.c
> @@ -636,6 +636,7 @@ static enum resp_states atomic_reply(struct rxe_qp *qp,
>
>               *vaddr = value;
>               spin_unlock_bh(&atomic_ops_lock);
> +             rxe_unmap_vaddr(mr, vaddr);
>
>               qp->resp.msn++;
>
> diff --git a/drivers/infiniband/sw/rxe/rxe_verbs.h
> b/drivers/infiniband/sw/rxe/rxe_verbs.h index acab785ba7e2..6080a4b32f09
> 100644
> --- a/drivers/infiniband/sw/rxe/rxe_verbs.h
> +++ b/drivers/infiniband/sw/rxe/rxe_verbs.h
> @@ -280,7 +280,10 @@ enum rxe_mr_lookup_type {
>  #define RXE_BUF_PER_MAP              (PAGE_SIZE / sizeof(struct
rxe_phys_buf))
>
>  struct rxe_phys_buf {
> -     u64      addr;
> +     union {
> +             u64 addr; /* IB_MR_TYPE_MEM_REG */
> +             struct page *page; /* IB_MR_TYPE_USER */
> +     };
>  };
>
>  struct rxe_map {
> --
> 2.31.1
Fabio M. De Francesco Nov. 16, 2022, 12:49 p.m. UTC | #2
On venerdì 11 novembre 2022 05:30:29 CET Li Zhijian wrote:
> For IB_MR_TYPE_USER MR, iova_to_vaddr() will call kmap_local_page() to
> map page.
>
> Signed-off-by: Li Zhijian <lizhijian@fujitsu.com>
> ---
>  drivers/infiniband/sw/rxe/rxe_loc.h   |  1 +
>  drivers/infiniband/sw/rxe/rxe_mr.c    | 38 +++++++++++++++------------
>  drivers/infiniband/sw/rxe/rxe_resp.c  |  1 +
>  drivers/infiniband/sw/rxe/rxe_verbs.h |  5 +++-
>  4 files changed, 27 insertions(+), 18 deletions(-)
>
> diff --git a/drivers/infiniband/sw/rxe/rxe_loc.h
> b/drivers/infiniband/sw/rxe/rxe_loc.h index c2a5c8814a48..22a8c44d39c8
100644
> --- a/drivers/infiniband/sw/rxe/rxe_loc.h
> +++ b/drivers/infiniband/sw/rxe/rxe_loc.h
> @@ -73,6 +73,7 @@ int rxe_mr_copy(struct rxe_mr *mr, u64 iova, void *addr,
int
> length, int copy_data(struct rxe_pd *pd, int access, struct rxe_dma_info
> *dma, void *addr, int length, enum rxe_mr_copy_dir dir);
>  void *iova_to_vaddr(struct rxe_mr *mr, u64 iova, int length);
> +void rxe_unmap_vaddr(struct rxe_mr *mr, void *vaddr);
>  struct rxe_mr *lookup_mr(struct rxe_pd *pd, int access, u32 key,
>                        enum rxe_mr_lookup_type type);
>  int mr_check_range(struct rxe_mr *mr, u64 iova, size_t length);
> diff --git a/drivers/infiniband/sw/rxe/rxe_mr.c
> b/drivers/infiniband/sw/rxe/rxe_mr.c index a4e786b657b7..d26a4a33119c 100644
> --- a/drivers/infiniband/sw/rxe/rxe_mr.c
> +++ b/drivers/infiniband/sw/rxe/rxe_mr.c
> @@ -120,9 +120,7 @@ int rxe_mr_init_user(struct rxe_dev *rxe, u64 start, u64
> length, u64 iova, struct ib_umem              *umem;
>       struct sg_page_iter     sg_iter;
>       int                     num_buf;
> -     void                    *vaddr;
>       int err;
> -     int i;
>
>       umem = ib_umem_get(&rxe->ib_dev, start, length, access);
>       if (IS_ERR(umem)) {
> @@ -159,18 +157,9 @@ int rxe_mr_init_user(struct rxe_dev *rxe, u64 start,
u64
> length, u64 iova, num_buf = 0;
>                       }
>
> -                     vaddr =
page_address(sg_page_iter_page(&sg_iter));
> -                     if (!vaddr) {
> -                             pr_warn("%s: Unable to get virtual
address\n",
> -                                             __func__);
> -                             err = -ENOMEM;
> -                             goto err_cleanup_map;
> -                     }
> -
> -                     buf->addr = (uintptr_t)vaddr;
> +                     buf->page = sg_page_iter_page(&sg_iter);
>                       num_buf++;
>                       buf++;
> -
>               }
>       }
>
> @@ -182,10 +171,6 @@ int rxe_mr_init_user(struct rxe_dev *rxe, u64 start,
u64
> length, u64 iova,
>
>       return 0;
>
> -err_cleanup_map:
> -     for (i = 0; i < mr->num_map; i++)
> -             kfree(mr->map[i]);
> -     kfree(mr->map);
>  err_release_umem:
>       ib_umem_release(umem);
>  err_out:
> @@ -246,6 +231,12 @@ static void lookup_iova(struct rxe_mr *mr, u64 iova,
int
> *m_out, int *n_out, }
>  }
>
> +void rxe_unmap_vaddr(struct rxe_mr *mr, void *vaddr)
> +{
> +     if (mr->ibmr.type == IB_MR_TYPE_USER)
> +             kunmap_local(vaddr);
> +}
> +
>  static void *__iova_to_vaddr(struct rxe_mr *mr, u64 iova, int length)
>  {
>       size_t offset;
> @@ -258,9 +249,21 @@ static void *__iova_to_vaddr(struct rxe_mr *mr, u64
iova,
> int length) return NULL;
>       }
>
> -     return (void *)(uintptr_t)mr->map[m]->buf[n].addr + offset;
> +     if (mr->ibmr.type == IB_MR_TYPE_USER) {
> +             char *paddr;
> +             struct page *pg = mr->map[m]->buf[n].page;
> +
> +             paddr = kmap_local_page(pg);
> +             if (paddr == NULL) {
> +                     pr_warn("Failed to map page");
> +                     return NULL;
> +             }

I know nothing about this code but I am here as a result of regular checks for
changes to HIGHMEM mappings across the entire kernel. So please forgive me if
I'm objecting on correct changes.

1) It looks like this code had a call to page_address() and you converted it
to mapping with kmap_local_page().

If page_address() is related and it used to work properly, the page you are
mapping cannot come from ZONE_HIGHMEM. Therefore, kmap_local_page() looks like
an overkill.

I'm probably missing something...

2) What made you think that kmap_local_page() might fail and return NULL?
AFAIK, kmap_local_page() won't ever return NULL, therefore there is no need to
check.

Regards,

Fabio

P.S.: I just noticed that the second entry in my list was missing from the
other email which should be discarded.



> +             return paddr + offset;
> +     } else
> +             return (void *)(uintptr_t)mr->map[m]->buf[n].addr +
offset;
>  }
>
> +/* must call rxe_unmap_vaddr to unmap vaddr */
>  void *iova_to_vaddr(struct rxe_mr *mr, u64 iova, int length)
>  {
>       if (mr->state != RXE_MR_STATE_VALID) {
> @@ -326,6 +329,7 @@ int rxe_mr_copy(struct rxe_mr *mr, u64 iova, void *addr,
> int length, dest = (dir == RXE_TO_MR_OBJ) ? va : addr;
>
>               memcpy(dest, src, bytes);
> +             rxe_unmap_vaddr(mr, va);
>
>               length  -= bytes;
>               addr    += bytes;
> diff --git a/drivers/infiniband/sw/rxe/rxe_resp.c
> b/drivers/infiniband/sw/rxe/rxe_resp.c index 483043dc4e89..765cb9f8538a
> 100644
> --- a/drivers/infiniband/sw/rxe/rxe_resp.c
> +++ b/drivers/infiniband/sw/rxe/rxe_resp.c
> @@ -636,6 +636,7 @@ static enum resp_states atomic_reply(struct rxe_qp *qp,
>
>               *vaddr = value;
>               spin_unlock_bh(&atomic_ops_lock);
> +             rxe_unmap_vaddr(mr, vaddr);
>
>               qp->resp.msn++;
>
> diff --git a/drivers/infiniband/sw/rxe/rxe_verbs.h
> b/drivers/infiniband/sw/rxe/rxe_verbs.h index acab785ba7e2..6080a4b32f09
> 100644
> --- a/drivers/infiniband/sw/rxe/rxe_verbs.h
> +++ b/drivers/infiniband/sw/rxe/rxe_verbs.h
> @@ -280,7 +280,10 @@ enum rxe_mr_lookup_type {
>  #define RXE_BUF_PER_MAP              (PAGE_SIZE / sizeof(struct
rxe_phys_buf))
>
>  struct rxe_phys_buf {
> -     u64      addr;
> +     union {
> +             u64 addr; /* IB_MR_TYPE_MEM_REG */
> +             struct page *page; /* IB_MR_TYPE_USER */
> +     };
>  };
>
>  struct rxe_map {
> --
> 2.31.1
Li Zhijian Nov. 18, 2022, 1:32 a.m. UTC | #3
On 16/11/2022 20:37, Fabio M. De Francesco wrote:
>> -     return (void *)(uintptr_t)mr->map[m]->buf[n].addr + offset;
>> +     if (mr->ibmr.type == IB_MR_TYPE_USER) {
>> +             char *paddr;
>> +             struct page *pg = mr->map[m]->buf[n].page;
>> +
>> +             paddr = kmap_local_page(pg);
>> +             if (paddr == NULL) {
>> +                     pr_warn("Failed to map page");
>> +                     return NULL;
>> +             }
> I know nothing about this code but I am here as a result of regular checks for
> changes to HIGHMEM mappings across the entire kernel. So please forgive me if
> I'm objecting to the correct changes.
> 
> 1) It looks like this code had a call to page_address() and you converted it
> to mapping with kmap_local_page().
> 
> If page_address() is related and it used to work properly, the page you are
> mapping cannot come from ZONE_HIGHMEM. 

Yes, you are totally right.


Therefore, kmap_local_page() looks like
> an overkill.


The confusion about the page_address() here has been raised for a long 
time[1][2].

https://www.spinics.net/lists/linux-rdma/msg113206.html
https://lore.kernel.org/all/20220121160654.GC773547@iweiny-DESK2.sc.intel.com/


Thanks
Zhijian

> 
> I'm probably missing something...
>
Jason Gunthorpe Nov. 18, 2022, 5:33 p.m. UTC | #4
On Fri, Nov 11, 2022 at 04:30:29AM +0000, Li Zhijian wrote:

> diff --git a/drivers/infiniband/sw/rxe/rxe_verbs.h b/drivers/infiniband/sw/rxe/rxe_verbs.h
> index acab785ba7e2..6080a4b32f09 100644
> --- a/drivers/infiniband/sw/rxe/rxe_verbs.h
> +++ b/drivers/infiniband/sw/rxe/rxe_verbs.h
> @@ -280,7 +280,10 @@ enum rxe_mr_lookup_type {
>  #define RXE_BUF_PER_MAP		(PAGE_SIZE / sizeof(struct rxe_phys_buf))
>  
>  struct rxe_phys_buf {
> -	u64      addr;
> +	union {
> +		u64 addr; /* IB_MR_TYPE_MEM_REG */
> +		struct page *page; /* IB_MR_TYPE_USER */
> +	};

Everything rxe handles is a struct page, even the kernel
registrations.

Jason
Jason Gunthorpe Nov. 21, 2022, 6:53 p.m. UTC | #5
On Wed, Nov 16, 2022 at 01:37:14PM +0100, Fabio M. De Francesco wrote:
> > -     return (void *)(uintptr_t)mr->map[m]->buf[n].addr + offset;
> > +     if (mr->ibmr.type == IB_MR_TYPE_USER) {
> > +             char *paddr;
> > +             struct page *pg = mr->map[m]->buf[n].page;
> > +
> > +             paddr = kmap_local_page(pg);
> > +             if (paddr == NULL) {
> > +                     pr_warn("Failed to map page");
> > +                     return NULL;
> > +             }
> 
> I know nothing about this code but I am here as a result of regular checks for
> changes to HIGHMEM mappings across the entire kernel. So please forgive me if
> I'm objecting to the correct changes.
> 
> 1) It looks like this code had a call to page_address() and you converted it
> to mapping with kmap_local_page().

It only worked properly because the kconfig is blocking CONFIG_HIGHMEM
so ZONE_HIGHMEM doesn't exist. These pages are obtained from GUP and
could be anything.

This is done indirectly through

config INFINIBAND_VIRT_DMA
        def_bool !HIGHMEM

But this patch is undoing parts of what are causing that restriction
by using the proper APIs. Though I'm not sure if we can eventually get
to remove it for RXE completely.

Jason
Fabio M. De Francesco Nov. 22, 2022, 11:24 p.m. UTC | #6
On lunedì 21 novembre 2022 19:53:54 CET Jason Gunthorpe wrote:
> On Wed, Nov 16, 2022 at 01:37:14PM +0100, Fabio M. De Francesco wrote:
> > > -     return (void *)(uintptr_t)mr->map[m]->buf[n].addr + offset;
> > > +     if (mr->ibmr.type == IB_MR_TYPE_USER) {
> > > +             char *paddr;
> > > +             struct page *pg = mr->map[m]->buf[n].page;
> > > +
> > > +             paddr = kmap_local_page(pg);
> > > +             if (paddr == NULL) {
> > > +                     pr_warn("Failed to map page");
> > > +                     return NULL;
> > > +             }
> > 
> > I know nothing about this code but I am here as a result of regular checks
> > for changes to HIGHMEM mappings across the entire kernel. So please 
forgive
> > me if I'm objecting to the correct changes.
> > 
> > 1) It looks like this code had a call to page_address() and you converted 
it
> > to mapping with kmap_local_page().
> 
> It only worked properly because the kconfig is blocking CONFIG_HIGHMEM
> so ZONE_HIGHMEM doesn't exist. These pages are obtained from GUP and
> could be anything.
> 
> This is done indirectly through
> 
> config INFINIBAND_VIRT_DMA
>         def_bool !HIGHMEM
> But this patch is undoing parts of what are causing that restriction
> by using the proper APIs. Though I'm not sure if we can eventually get
> to remove it for RXE completely.
> Jason

Ah, OK. This code was for !HIGHMEM kernels...

I understand but, FWIW I doubt that the use of page_address(), instead of 
kmapping, should ever be made on the bases that kconfig is blocking HIGHMEM.  

However, if I understand correctly, that restriction (!HIGHMEM) is going to be 
removed. Therefore, page_address() will break on HIGHMEM and Jason is 
correctly converting to kmap_local_page().

There is only one thing left... I think that he missed another mail from me. 
The one you responded to was missing the final paragraph, so I sent another 
message few minutes later.

kmap_local_page() returns always valid pointers to kernel virtual addresses. I 
can't see any ways for kmap_local_page() to return NULL. Therefore, I was 
asking for the reason to check for NULL.

Thanks,

Fabio
Fabio M. De Francesco Nov. 22, 2022, 11:55 p.m. UTC | #7
On mercoledì 23 novembre 2022 00:24:26 CET Fabio M. De Francesco wrote:
> On lunedì 21 novembre 2022 19:53:54 CET Jason Gunthorpe wrote:
> > On Wed, Nov 16, 2022 at 01:37:14PM +0100, Fabio M. De Francesco wrote:
> > > > -     return (void *)(uintptr_t)mr->map[m]->buf[n].addr + offset;
> > > > +     if (mr->ibmr.type == IB_MR_TYPE_USER) {
> > > > +             char *paddr;
> > > > +             struct page *pg = mr->map[m]->buf[n].page;
> > > > +
> > > > +             paddr = kmap_local_page(pg);
> > > > +             if (paddr == NULL) {
> > > > +                     pr_warn("Failed to map page");
> > > > +                     return NULL;
> > > > +             }
> > > 
> > > I know nothing about this code but I am here as a result of regular 
checks
> > > for changes to HIGHMEM mappings across the entire kernel. So please
> 
> forgive
> 
> > > me if I'm objecting to the correct changes.
> > > 
> > > 1) It looks like this code had a call to page_address() and you 
converted
> 
> it
> 
> > > to mapping with kmap_local_page().
> > 
> > It only worked properly because the kconfig is blocking CONFIG_HIGHMEM
> > so ZONE_HIGHMEM doesn't exist. These pages are obtained from GUP and
> > could be anything.
> > 
> > This is done indirectly through
> > 
> > config INFINIBAND_VIRT_DMA
> > 
> >         def_bool !HIGHMEM
> > 
> > But this patch is undoing parts of what are causing that restriction
> > by using the proper APIs. Though I'm not sure if we can eventually get
> > to remove it for RXE completely.
> > Jason
> 
> Ah, OK. This code was for !HIGHMEM kernels...
> 
> I understand but, FWIW I doubt that the use of page_address(), instead of
> kmapping, should ever be made on the bases that kconfig is blocking HIGHMEM.
> 
> However, if I understand correctly, that restriction (!HIGHMEM) is going to 
be
> removed. Therefore, page_address() will break on HIGHMEM and Jason is
> correctly converting to kmap_local_page().

Jason, I'm sorry for the erroneous attribution :-(

Fabio
 
> There is only one thing left... I think that he missed another mail from me.
> The one you responded to was missing the final paragraph, so I sent another
> message few minutes later.
> 
> kmap_local_page() returns always valid pointers to kernel virtual addresses. 
I
> can't see any ways for kmap_local_page() to return NULL. Therefore, I was
> asking for the reason to check for NULL.
> 
> Thanks,
> 
> Fabio
diff mbox series

Patch

diff --git a/drivers/infiniband/sw/rxe/rxe_loc.h b/drivers/infiniband/sw/rxe/rxe_loc.h
index c2a5c8814a48..22a8c44d39c8 100644
--- a/drivers/infiniband/sw/rxe/rxe_loc.h
+++ b/drivers/infiniband/sw/rxe/rxe_loc.h
@@ -73,6 +73,7 @@  int rxe_mr_copy(struct rxe_mr *mr, u64 iova, void *addr, int length,
 int copy_data(struct rxe_pd *pd, int access, struct rxe_dma_info *dma,
 	      void *addr, int length, enum rxe_mr_copy_dir dir);
 void *iova_to_vaddr(struct rxe_mr *mr, u64 iova, int length);
+void rxe_unmap_vaddr(struct rxe_mr *mr, void *vaddr);
 struct rxe_mr *lookup_mr(struct rxe_pd *pd, int access, u32 key,
 			 enum rxe_mr_lookup_type type);
 int mr_check_range(struct rxe_mr *mr, u64 iova, size_t length);
diff --git a/drivers/infiniband/sw/rxe/rxe_mr.c b/drivers/infiniband/sw/rxe/rxe_mr.c
index a4e786b657b7..d26a4a33119c 100644
--- a/drivers/infiniband/sw/rxe/rxe_mr.c
+++ b/drivers/infiniband/sw/rxe/rxe_mr.c
@@ -120,9 +120,7 @@  int rxe_mr_init_user(struct rxe_dev *rxe, u64 start, u64 length, u64 iova,
 	struct ib_umem		*umem;
 	struct sg_page_iter	sg_iter;
 	int			num_buf;
-	void			*vaddr;
 	int err;
-	int i;
 
 	umem = ib_umem_get(&rxe->ib_dev, start, length, access);
 	if (IS_ERR(umem)) {
@@ -159,18 +157,9 @@  int rxe_mr_init_user(struct rxe_dev *rxe, u64 start, u64 length, u64 iova,
 				num_buf = 0;
 			}
 
-			vaddr = page_address(sg_page_iter_page(&sg_iter));
-			if (!vaddr) {
-				pr_warn("%s: Unable to get virtual address\n",
-						__func__);
-				err = -ENOMEM;
-				goto err_cleanup_map;
-			}
-
-			buf->addr = (uintptr_t)vaddr;
+			buf->page = sg_page_iter_page(&sg_iter);
 			num_buf++;
 			buf++;
-
 		}
 	}
 
@@ -182,10 +171,6 @@  int rxe_mr_init_user(struct rxe_dev *rxe, u64 start, u64 length, u64 iova,
 
 	return 0;
 
-err_cleanup_map:
-	for (i = 0; i < mr->num_map; i++)
-		kfree(mr->map[i]);
-	kfree(mr->map);
 err_release_umem:
 	ib_umem_release(umem);
 err_out:
@@ -246,6 +231,12 @@  static void lookup_iova(struct rxe_mr *mr, u64 iova, int *m_out, int *n_out,
 	}
 }
 
+void rxe_unmap_vaddr(struct rxe_mr *mr, void *vaddr)
+{
+	if (mr->ibmr.type == IB_MR_TYPE_USER)
+		kunmap_local(vaddr);
+}
+
 static void *__iova_to_vaddr(struct rxe_mr *mr, u64 iova, int length)
 {
 	size_t offset;
@@ -258,9 +249,21 @@  static void *__iova_to_vaddr(struct rxe_mr *mr, u64 iova, int length)
 		return NULL;
 	}
 
-	return (void *)(uintptr_t)mr->map[m]->buf[n].addr + offset;
+	if (mr->ibmr.type == IB_MR_TYPE_USER) {
+		char *paddr;
+		struct page *pg = mr->map[m]->buf[n].page;
+
+		paddr = kmap_local_page(pg);
+		if (paddr == NULL) {
+			pr_warn("Failed to map page");
+			return NULL;
+		}
+		return paddr + offset;
+	} else
+		return (void *)(uintptr_t)mr->map[m]->buf[n].addr + offset;
 }
 
+/* must call rxe_unmap_vaddr to unmap vaddr */
 void *iova_to_vaddr(struct rxe_mr *mr, u64 iova, int length)
 {
 	if (mr->state != RXE_MR_STATE_VALID) {
@@ -326,6 +329,7 @@  int rxe_mr_copy(struct rxe_mr *mr, u64 iova, void *addr, int length,
 		dest = (dir == RXE_TO_MR_OBJ) ? va : addr;
 
 		memcpy(dest, src, bytes);
+		rxe_unmap_vaddr(mr, va);
 
 		length	-= bytes;
 		addr	+= bytes;
diff --git a/drivers/infiniband/sw/rxe/rxe_resp.c b/drivers/infiniband/sw/rxe/rxe_resp.c
index 483043dc4e89..765cb9f8538a 100644
--- a/drivers/infiniband/sw/rxe/rxe_resp.c
+++ b/drivers/infiniband/sw/rxe/rxe_resp.c
@@ -636,6 +636,7 @@  static enum resp_states atomic_reply(struct rxe_qp *qp,
 
 		*vaddr = value;
 		spin_unlock_bh(&atomic_ops_lock);
+		rxe_unmap_vaddr(mr, vaddr);
 
 		qp->resp.msn++;
 
diff --git a/drivers/infiniband/sw/rxe/rxe_verbs.h b/drivers/infiniband/sw/rxe/rxe_verbs.h
index acab785ba7e2..6080a4b32f09 100644
--- a/drivers/infiniband/sw/rxe/rxe_verbs.h
+++ b/drivers/infiniband/sw/rxe/rxe_verbs.h
@@ -280,7 +280,10 @@  enum rxe_mr_lookup_type {
 #define RXE_BUF_PER_MAP		(PAGE_SIZE / sizeof(struct rxe_phys_buf))
 
 struct rxe_phys_buf {
-	u64      addr;
+	union {
+		u64 addr; /* IB_MR_TYPE_MEM_REG */
+		struct page *page; /* IB_MR_TYPE_USER */
+	};
 };
 
 struct rxe_map {