[PATCH 0/2 v6] add ioctl/sysfs to donate file-backed pages

[PATCH 0/2 v6] add ioctl/sysfs to donate file-backed pages

[Jaegeuk Kim]

If users clearly know which file-backed pages to reclaim in system view, they
can use this ioctl() to register in advance and reclaim all at once later.

To MM and others,

I'd like to propose this API in F2FS only, since
1) the use-case is quite limited in Android at the moment. Once it's generall
accepted with more use-cases, happy to propose a generic API such as fadvise.
Please chime in, if there's any needs.

2) it's file-backed pages which requires to maintain the list of inode objects.
I'm not sure this fits in MM tho, also happy to listen to any feedback.

Jaegeuk Kim (2):
  f2fs: register inodes which is able to donate pages
  f2fs: add a sysfs entry to request donate file-backed pages

 Documentation/ABI/testing/sysfs-fs-f2fs |  7 +++
 fs/f2fs/debug.c                         |  3 ++
 fs/f2fs/f2fs.h                          | 14 +++++-
 fs/f2fs/file.c                          | 65 +++++++++++++++++++++++++
 fs/f2fs/inode.c                         | 14 ++++++
 fs/f2fs/shrinker.c                      | 27 ++++++++++
 fs/f2fs/super.c                         |  1 +
 fs/f2fs/sysfs.c                         |  8 +++
 include/uapi/linux/f2fs.h               |  7 +++
 9 files changed, 145 insertions(+), 1 deletion(-)

-- 
2.48.0.rc2.279.g1de40edade-goog

Re: [PATCH 0/2 v6] add ioctl/sysfs to donate file-backed pages

[Matthew Wilcox]

On Fri, Jan 17, 2025 at 04:41:16PM +0000, Jaegeuk Kim wrote:
> If users clearly know which file-backed pages to reclaim in system view, they
> can use this ioctl() to register in advance and reclaim all at once later.
> 
> To MM and others,
> 
> I'd like to propose this API in F2FS only, since
> 1) the use-case is quite limited in Android at the moment. Once it's generall
> accepted with more use-cases, happy to propose a generic API such as fadvise.
> Please chime in, if there's any needs.
> 
> 2) it's file-backed pages which requires to maintain the list of inode objects.
> I'm not sure this fits in MM tho, also happy to listen to any feedback.

You didn't cc the patches to linux-mm, so that's a bad start.

I don't understand how this is different from MADV_COLD.  Please
explain.

Re: [PATCH 0/2 v6] add ioctl/sysfs to donate file-backed pages

[Jaegeuk Kim]

On 01/17, Matthew Wilcox wrote:
> On Fri, Jan 17, 2025 at 04:41:16PM +0000, Jaegeuk Kim wrote:
> > If users clearly know which file-backed pages to reclaim in system view, they
> > can use this ioctl() to register in advance and reclaim all at once later.
> > 
> > To MM and others,
> > 
> > I'd like to propose this API in F2FS only, since
> > 1) the use-case is quite limited in Android at the moment. Once it's generall
> > accepted with more use-cases, happy to propose a generic API such as fadvise.
> > Please chime in, if there's any needs.
> > 
> > 2) it's file-backed pages which requires to maintain the list of inode objects.
> > I'm not sure this fits in MM tho, also happy to listen to any feedback.
> 
> You didn't cc the patches to linux-mm, so that's a bad start.

Because #1.

> 
> I don't understand how this is different from MADV_COLD.  Please
> explain.

MADV_COLD is a vma range, while this is a file range. So, it's more close to
fadvise(POSIX_FADV_DONTNEED) which tries to reclaim the file-backed pages
at the time when it's called. The idea is to keep the hints only, and try to
reclaim all later when admin expects system memory pressure soon.

Re: [PATCH 0/2 v6] add ioctl/sysfs to donate file-backed pages

[Matthew Wilcox]

On Fri, Jan 17, 2025 at 06:48:55PM +0000, Jaegeuk Kim wrote:
> > I don't understand how this is different from MADV_COLD.  Please
> > explain.
> 
> MADV_COLD is a vma range, while this is a file range. So, it's more close to
> fadvise(POSIX_FADV_DONTNEED) which tries to reclaim the file-backed pages
> at the time when it's called. The idea is to keep the hints only, and try to
> reclaim all later when admin expects system memory pressure soon.

So you're saying you want POSIX_FADV_COLD?

Re: [PATCH 0/2 v6] add ioctl/sysfs to donate file-backed pages

[Jaegeuk Kim]

On 01/17, Matthew Wilcox wrote:
> On Fri, Jan 17, 2025 at 06:48:55PM +0000, Jaegeuk Kim wrote:
> > > I don't understand how this is different from MADV_COLD.  Please
> > > explain.
> > 
> > MADV_COLD is a vma range, while this is a file range. So, it's more close to
> > fadvise(POSIX_FADV_DONTNEED) which tries to reclaim the file-backed pages
> > at the time when it's called. The idea is to keep the hints only, and try to
> > reclaim all later when admin expects system memory pressure soon.
> 
> So you're saying you want POSIX_FADV_COLD?

Yeah, the intention looks similar like marking it cold and paging out later.

Re: [PATCH 0/2 v6] add ioctl/sysfs to donate file-backed pages

[Jaegeuk Kim]

On 01/17, Jaegeuk Kim wrote:
> On 01/17, Matthew Wilcox wrote:
> > On Fri, Jan 17, 2025 at 06:48:55PM +0000, Jaegeuk Kim wrote:
> > > > I don't understand how this is different from MADV_COLD.  Please
> > > > explain.
> > > 
> > > MADV_COLD is a vma range, while this is a file range. So, it's more close to
> > > fadvise(POSIX_FADV_DONTNEED) which tries to reclaim the file-backed pages
> > > at the time when it's called. The idea is to keep the hints only, and try to
> > > reclaim all later when admin expects system memory pressure soon.
> > 
> > So you're saying you want POSIX_FADV_COLD?
> 
> Yeah, the intention looks similar like marking it cold and paging out later.

Kindly ping, for the feedback on the direction. If there's demand for something
generalized api, I'm happy to explore.

Re: [PATCH 0/2 v6] add ioctl/sysfs to donate file-backed pages

[Jaegeuk Kim]

On 02/04, Jaegeuk Kim wrote:
> On 01/17, Jaegeuk Kim wrote:
> > On 01/17, Matthew Wilcox wrote:
> > > On Fri, Jan 17, 2025 at 06:48:55PM +0000, Jaegeuk Kim wrote:
> > > > > I don't understand how this is different from MADV_COLD.  Please
> > > > > explain.
> > > > 
> > > > MADV_COLD is a vma range, while this is a file range. So, it's more close to
> > > > fadvise(POSIX_FADV_DONTNEED) which tries to reclaim the file-backed pages
> > > > at the time when it's called. The idea is to keep the hints only, and try to
> > > > reclaim all later when admin expects system memory pressure soon.
> > > 
> > > So you're saying you want POSIX_FADV_COLD?
> > 
> > Yeah, the intention looks similar like marking it cold and paging out later.
> 
> Kindly ping, for the feedback on the direction. If there's demand for something
> generalized api, I'm happy to explore.

If there's no objection, let me push the change in f2fs and keep an eye on
who more will need this in general.

Re: [PATCH 0/2 v6] add ioctl/sysfs to donate file-backed pages

[Matthew Wilcox]

On Mon, Feb 10, 2025 at 05:00:47PM +0000, Jaegeuk Kim wrote:
> On 02/04, Jaegeuk Kim wrote:
> > On 01/17, Jaegeuk Kim wrote:
> > > On 01/17, Matthew Wilcox wrote:
> > > > On Fri, Jan 17, 2025 at 06:48:55PM +0000, Jaegeuk Kim wrote:
> > > > > > I don't understand how this is different from MADV_COLD.  Please
> > > > > > explain.
> > > > > 
> > > > > MADV_COLD is a vma range, while this is a file range. So, it's more close to
> > > > > fadvise(POSIX_FADV_DONTNEED) which tries to reclaim the file-backed pages
> > > > > at the time when it's called. The idea is to keep the hints only, and try to
> > > > > reclaim all later when admin expects system memory pressure soon.
> > > > 
> > > > So you're saying you want POSIX_FADV_COLD?
> > > 
> > > Yeah, the intention looks similar like marking it cold and paging out later.
> > 
> > Kindly ping, for the feedback on the direction. If there's demand for something
> > generalized api, I'm happy to explore.
> 
> If there's no objection, let me push the change in f2fs and keep an eye on
> who more will need this in general.

I don't know why you're asking for direction.  I gave my direction: use
fadvise().

Putting this directly in f2fs is a horrible idea.  NAK.

Re: [PATCH 0/2 v6] add ioctl/sysfs to donate file-backed pages

[Jaegeuk Kim]

On 02/10, Matthew Wilcox wrote:
> On Mon, Feb 10, 2025 at 05:00:47PM +0000, Jaegeuk Kim wrote:
> > On 02/04, Jaegeuk Kim wrote:
> > > On 01/17, Jaegeuk Kim wrote:
> > > > On 01/17, Matthew Wilcox wrote:
> > > > > On Fri, Jan 17, 2025 at 06:48:55PM +0000, Jaegeuk Kim wrote:
> > > > > > > I don't understand how this is different from MADV_COLD.  Please
> > > > > > > explain.
> > > > > > 
> > > > > > MADV_COLD is a vma range, while this is a file range. So, it's more close to
> > > > > > fadvise(POSIX_FADV_DONTNEED) which tries to reclaim the file-backed pages
> > > > > > at the time when it's called. The idea is to keep the hints only, and try to
> > > > > > reclaim all later when admin expects system memory pressure soon.
> > > > > 
> > > > > So you're saying you want POSIX_FADV_COLD?
> > > > 
> > > > Yeah, the intention looks similar like marking it cold and paging out later.
> > > 
> > > Kindly ping, for the feedback on the direction. If there's demand for something
> > > generalized api, I'm happy to explore.
> > 
> > If there's no objection, let me push the change in f2fs and keep an eye on
> > who more will need this in general.
> 
> I don't know why you're asking for direction.  I gave my direction: use
> fadvise().

Funny, that single question didn't mean like this at all. Will take a look
how the patch looks like.

> 
> Putting this directly in f2fs is a horrible idea.  NAK.

Re: [PATCH 0/2 v6] add ioctl/sysfs to donate file-backed pages

[Jaegeuk Kim]

On 02/10, Jaegeuk Kim wrote:
> On 02/10, Matthew Wilcox wrote:
> > On Mon, Feb 10, 2025 at 05:00:47PM +0000, Jaegeuk Kim wrote:
> > > On 02/04, Jaegeuk Kim wrote:
> > > > On 01/17, Jaegeuk Kim wrote:
> > > > > On 01/17, Matthew Wilcox wrote:
> > > > > > On Fri, Jan 17, 2025 at 06:48:55PM +0000, Jaegeuk Kim wrote:
> > > > > > > > I don't understand how this is different from MADV_COLD.  Please
> > > > > > > > explain.
> > > > > > > 
> > > > > > > MADV_COLD is a vma range, while this is a file range. So, it's more close to
> > > > > > > fadvise(POSIX_FADV_DONTNEED) which tries to reclaim the file-backed pages
> > > > > > > at the time when it's called. The idea is to keep the hints only, and try to
> > > > > > > reclaim all later when admin expects system memory pressure soon.
> > > > > > 
> > > > > > So you're saying you want POSIX_FADV_COLD?
> > > > > 
> > > > > Yeah, the intention looks similar like marking it cold and paging out later.
> > > > 
> > > > Kindly ping, for the feedback on the direction. If there's demand for something
> > > > generalized api, I'm happy to explore.
> > > 
> > > If there's no objection, let me push the change in f2fs and keep an eye on
> > > who more will need this in general.
> > 
> > I don't know why you're asking for direction.  I gave my direction: use
> > fadvise().
> 
> Funny, that single question didn't mean like this at all. Will take a look
> how the patch looks like.

Ok, it seems we can get this hint via POSIX_FADV_NOREUSE. I'll take that
instead of adding a new API. Thanks.

> 
> > 
> > Putting this directly in f2fs is a horrible idea.  NAK.