2020-05-23 04:01:31

by John Hubbard

[permalink] [raw]
Subject: [PATCH v2] orangefs: convert get_user_pages() --> pin_user_pages()

This code was using get_user_pages*(), in a "Case 1" scenario
(Direct IO), using the categorization from [1]. That means that it's
time to convert the get_user_pages*() + put_page() calls to
pin_user_pages*() + unpin_user_pages() calls.

There is some helpful background in [2]: basically, this is a small
part of fixing a long-standing disconnect between pinning pages, and
file systems' use of those pages.

[1] Documentation/core-api/pin_user_pages.rst

[2] "Explicit pinning of user-space pages":
https://lwn.net/Articles/807108/

Cc: Mike Marshall <[email protected]>
Cc: Martin Brandenburg <[email protected]>
Cc: [email protected]
Cc: [email protected]
Signed-off-by: John Hubbard <[email protected]>
---

Hi,

Note that I have only compile-tested this patch, although that does
also include cross-compiling for a few other arches.

Changes since v1 [3]: correct the commit description, so that
it refers to "Case 1" instead of "Case 2".


[3] https://lore.kernel.org/r/[email protected]

thanks,
John Hubbard
NVIDIA

fs/orangefs/orangefs-bufmap.c | 9 +++------
1 file changed, 3 insertions(+), 6 deletions(-)

diff --git a/fs/orangefs/orangefs-bufmap.c b/fs/orangefs/orangefs-bufmap.c
index 2bb916d68576..538e839590ef 100644
--- a/fs/orangefs/orangefs-bufmap.c
+++ b/fs/orangefs/orangefs-bufmap.c
@@ -168,10 +168,7 @@ static DEFINE_SPINLOCK(orangefs_bufmap_lock);
static void
orangefs_bufmap_unmap(struct orangefs_bufmap *bufmap)
{
- int i;
-
- for (i = 0; i < bufmap->page_count; i++)
- put_page(bufmap->page_array[i]);
+ unpin_user_pages(bufmap->page_array, bufmap->page_count);
}

static void
@@ -268,7 +265,7 @@ orangefs_bufmap_map(struct orangefs_bufmap *bufmap,
int offset = 0, ret, i;

/* map the pages */
- ret = get_user_pages_fast((unsigned long)user_desc->ptr,
+ ret = pin_user_pages_fast((unsigned long)user_desc->ptr,
bufmap->page_count, FOLL_WRITE, bufmap->page_array);

if (ret < 0)
@@ -280,7 +277,7 @@ orangefs_bufmap_map(struct orangefs_bufmap *bufmap,

for (i = 0; i < ret; i++) {
SetPageError(bufmap->page_array[i]);
- put_page(bufmap->page_array[i]);
+ unpin_user_page(bufmap->page_array[i]);
}
return -ENOMEM;
}
--
2.26.2


2020-05-30 07:41:43

by John Hubbard

[permalink] [raw]
Subject: Re: [PATCH v2] orangefs: convert get_user_pages() --> pin_user_pages()

On 2020-05-22 20:59, John Hubbard wrote:
> This code was using get_user_pages*(), in a "Case 1" scenario
> (Direct IO), using the categorization from [1]. That means that it's
> time to convert the get_user_pages*() + put_page() calls to
> pin_user_pages*() + unpin_user_pages() calls.
>
> There is some helpful background in [2]: basically, this is a small
> part of fixing a long-standing disconnect between pinning pages, and
> file systems' use of those pages.
>
> [1] Documentation/core-api/pin_user_pages.rst
>
> [2] "Explicit pinning of user-space pages":
> https://lwn.net/Articles/807108/
>
> Cc: Mike Marshall <[email protected]>
> Cc: Martin Brandenburg <[email protected]>
> Cc: [email protected]
> Cc: [email protected]
> Signed-off-by: John Hubbard <[email protected]>
> ---
>
> Hi,
>
> Note that I have only compile-tested this patch, although that does
> also include cross-compiling for a few other arches.

An update on the run-time testing: Just now, I got basic orangefs tests
running in xfstests, with this patch applied, and it all looks normal.

thanks,
--
John Hubbard
NVIDIA

>
> Changes since v1 [3]: correct the commit description, so that
> it refers to "Case 1" instead of "Case 2".
>
>
> [3] https://lore.kernel.org/r/[email protected]
>
> thanks,
> John Hubbard
> NVIDIA
>
> fs/orangefs/orangefs-bufmap.c | 9 +++------
> 1 file changed, 3 insertions(+), 6 deletions(-)
>
> diff --git a/fs/orangefs/orangefs-bufmap.c b/fs/orangefs/orangefs-bufmap.c
> index 2bb916d68576..538e839590ef 100644
> --- a/fs/orangefs/orangefs-bufmap.c
> +++ b/fs/orangefs/orangefs-bufmap.c
> @@ -168,10 +168,7 @@ static DEFINE_SPINLOCK(orangefs_bufmap_lock);
> static void
> orangefs_bufmap_unmap(struct orangefs_bufmap *bufmap)
> {
> - int i;
> -
> - for (i = 0; i < bufmap->page_count; i++)
> - put_page(bufmap->page_array[i]);
> + unpin_user_pages(bufmap->page_array, bufmap->page_count);
> }
>
> static void
> @@ -268,7 +265,7 @@ orangefs_bufmap_map(struct orangefs_bufmap *bufmap,
> int offset = 0, ret, i;
>
> /* map the pages */
> - ret = get_user_pages_fast((unsigned long)user_desc->ptr,
> + ret = pin_user_pages_fast((unsigned long)user_desc->ptr,
> bufmap->page_count, FOLL_WRITE, bufmap->page_array);
>
> if (ret < 0)
> @@ -280,7 +277,7 @@ orangefs_bufmap_map(struct orangefs_bufmap *bufmap,
>
> for (i = 0; i < ret; i++) {
> SetPageError(bufmap->page_array[i]);
> - put_page(bufmap->page_array[i]);
> + unpin_user_page(bufmap->page_array[i]);
> }
> return -ENOMEM;
> }
>

2020-06-01 11:16:19

by Mike Marshall

[permalink] [raw]
Subject: Re: [PATCH v2] orangefs: convert get_user_pages() --> pin_user_pages()

Hi John.

Thanks for the patch, and the testing, and the reference to the
lwn article.

I have also applied your patch to 5.7-rc7 and have run xfstests.

I applied your patch to the orangefs for-next tree.

-Mike

On Sat, May 30, 2020 at 3:39 AM John Hubbard <[email protected]> wrote:
>
> On 2020-05-22 20:59, John Hubbard wrote:
> > This code was using get_user_pages*(), in a "Case 1" scenario
> > (Direct IO), using the categorization from [1]. That means that it's
> > time to convert the get_user_pages*() + put_page() calls to
> > pin_user_pages*() + unpin_user_pages() calls.
> >
> > There is some helpful background in [2]: basically, this is a small
> > part of fixing a long-standing disconnect between pinning pages, and
> > file systems' use of those pages.
> >
> > [1] Documentation/core-api/pin_user_pages.rst
> >
> > [2] "Explicit pinning of user-space pages":
> > https://lwn.net/Articles/807108/
> >
> > Cc: Mike Marshall <[email protected]>
> > Cc: Martin Brandenburg <[email protected]>
> > Cc: [email protected]
> > Cc: [email protected]
> > Signed-off-by: John Hubbard <[email protected]>
> > ---
> >
> > Hi,
> >
> > Note that I have only compile-tested this patch, although that does
> > also include cross-compiling for a few other arches.
>
> An update on the run-time testing: Just now, I got basic orangefs tests
> running in xfstests, with this patch applied, and it all looks normal.
>
> thanks,
> --
> John Hubbard
> NVIDIA
>
> >
> > Changes since v1 [3]: correct the commit description, so that
> > it refers to "Case 1" instead of "Case 2".
> >
> >
> > [3] https://lore.kernel.org/r/[email protected]
> >
> > thanks,
> > John Hubbard
> > NVIDIA
> >
> > fs/orangefs/orangefs-bufmap.c | 9 +++------
> > 1 file changed, 3 insertions(+), 6 deletions(-)
> >
> > diff --git a/fs/orangefs/orangefs-bufmap.c b/fs/orangefs/orangefs-bufmap.c
> > index 2bb916d68576..538e839590ef 100644
> > --- a/fs/orangefs/orangefs-bufmap.c
> > +++ b/fs/orangefs/orangefs-bufmap.c
> > @@ -168,10 +168,7 @@ static DEFINE_SPINLOCK(orangefs_bufmap_lock);
> > static void
> > orangefs_bufmap_unmap(struct orangefs_bufmap *bufmap)
> > {
> > - int i;
> > -
> > - for (i = 0; i < bufmap->page_count; i++)
> > - put_page(bufmap->page_array[i]);
> > + unpin_user_pages(bufmap->page_array, bufmap->page_count);
> > }
> >
> > static void
> > @@ -268,7 +265,7 @@ orangefs_bufmap_map(struct orangefs_bufmap *bufmap,
> > int offset = 0, ret, i;
> >
> > /* map the pages */
> > - ret = get_user_pages_fast((unsigned long)user_desc->ptr,
> > + ret = pin_user_pages_fast((unsigned long)user_desc->ptr,
> > bufmap->page_count, FOLL_WRITE, bufmap->page_array);
> >
> > if (ret < 0)
> > @@ -280,7 +277,7 @@ orangefs_bufmap_map(struct orangefs_bufmap *bufmap,
> >
> > for (i = 0; i < ret; i++) {
> > SetPageError(bufmap->page_array[i]);
> > - put_page(bufmap->page_array[i]);
> > + unpin_user_page(bufmap->page_array[i]);
> > }
> > return -ENOMEM;
> > }
> >
>