mm/madvise: set ra_pages as device max request size during ADV_POPULATE_READ

  madvise(MADV_POPULATE_READ) tries to populate all page tables in the
specific range, so it is usually sequential IO if VMA is backed by
file.

Set ra_pages as device max request size for the involved readahead in
the ADV_POPULATE_READ, this way reduces latency of madvise(MADV_POPULATE_READ)
to 1/10 when running madvise(MADV_POPULATE_READ) over one 1GB file with
usual(default) 128KB of read_ahead_kb.

Cc: David Hildenbrand <david@redhat.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Don Dutile <ddutile@redhat.com>
Cc: Rafael Aquini <raquini@redhat.com>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Mike Snitzer <snitzer@kernel.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
---
 mm/madvise.c | 52 +++++++++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 51 insertions(+), 1 deletion(-)
  

On Fri, Feb 02, 2024 at 10:20:29AM +0800, Ming Lei wrote:
> +static struct file *madvise_override_ra_win(struct file *f,
> +		unsigned long start, unsigned long end,
> +		unsigned int *old_ra_pages)
> +{
> +	unsigned int io_pages;
> +
> +	if (!f || !f->f_mapping || !f->f_mapping->host)
> +		return NULL;

How can ->f_mapping be NULL?  How can f_mapping->host be NULL?
  

On Thu, Feb 01 2024 at  9:20P -0500,
Ming Lei <ming.lei@redhat.com> wrote:

> madvise(MADV_POPULATE_READ) tries to populate all page tables in the
> specific range, so it is usually sequential IO if VMA is backed by
> file.
> 
> Set ra_pages as device max request size for the involved readahead in
> the ADV_POPULATE_READ, this way reduces latency of madvise(MADV_POPULATE_READ)
> to 1/10 when running madvise(MADV_POPULATE_READ) over one 1GB file with
> usual(default) 128KB of read_ahead_kb.
> 
> Cc: David Hildenbrand <david@redhat.com>
> Cc: Matthew Wilcox <willy@infradead.org>
> Cc: Alexander Viro <viro@zeniv.linux.org.uk>
> Cc: Christian Brauner <brauner@kernel.org>
> Cc: Don Dutile <ddutile@redhat.com>
> Cc: Rafael Aquini <raquini@redhat.com>
> Cc: Dave Chinner <david@fromorbit.com>
> Cc: Mike Snitzer <snitzer@kernel.org>
> Cc: Andrew Morton <akpm@linux-foundation.org>
> Signed-off-by: Ming Lei <ming.lei@redhat.com>
> ---
>  mm/madvise.c | 52 +++++++++++++++++++++++++++++++++++++++++++++++++++-
>  1 file changed, 51 insertions(+), 1 deletion(-)
> 
> diff --git a/mm/madvise.c b/mm/madvise.c
> index 912155a94ed5..db5452c8abdd 100644
> --- a/mm/madvise.c
> +++ b/mm/madvise.c
> @@ -900,6 +900,37 @@ static long madvise_dontneed_free(struct vm_area_struct *vma,
>  		return -EINVAL;
>  }
>  
> +static void madvise_restore_ra_win(struct file **file, unsigned int ra_pages)
> +{
> +	if (*file) {
> +		struct file *f = *file;
> +
> +		f->f_ra.ra_pages = ra_pages;
> +		fput(f);
> +		*file = NULL;
> +	}
> +}
> +
> +static struct file *madvise_override_ra_win(struct file *f,
> +		unsigned long start, unsigned long end,
> +		unsigned int *old_ra_pages)
> +{
> +	unsigned int io_pages;
> +
> +	if (!f || !f->f_mapping || !f->f_mapping->host)
> +		return NULL;
> +
> +	io_pages = inode_to_bdi(f->f_mapping->host)->io_pages;
> +	if (((end - start) >> PAGE_SHIFT) < io_pages)
> +		return NULL;
> +
> +	f = get_file(f);
> +	*old_ra_pages = f->f_ra.ra_pages;
> +	f->f_ra.ra_pages = io_pages;
> +
> +	return f;
> +}
> +

Does this override imply that madvise_populate resorts to calling
filemap_fault() and here you're just arming it to use the larger
->io_pages for the duration of all associated faulting?

Wouldn't it be better to avoid faulting and build up larger page
vectors that get sent down to the block layer in one go and let the
block layer split using the device's limits? (like happens with
force_page_cache_ra)

I'm concerned that madvise_populate isn't so efficient with filemap
due to excessive faulting (*BUT* I haven't traced to know, I'm just
inferring that is why twiddling f->f_ra.ra_pages helps improve
madvise_populate by having it issue larger IO. Apologies if I'm way
off base)

Mike
  

On Fri, Feb 02, 2024 at 04:15:39AM +0000, Matthew Wilcox wrote:
> On Fri, Feb 02, 2024 at 10:20:29AM +0800, Ming Lei wrote:
> > +static struct file *madvise_override_ra_win(struct file *f,
> > +		unsigned long start, unsigned long end,
> > +		unsigned int *old_ra_pages)
> > +{
> > +	unsigned int io_pages;
> > +
> > +	if (!f || !f->f_mapping || !f->f_mapping->host)
> > +		return NULL;
> 
> How can ->f_mapping be NULL?  How can f_mapping->host be NULL?

You are right, the two checks can be removed because both two won't
be NULL for opened file, and .f_ra is initialized with
f->f_mapping->host->i_mapping directly too. I will drop the checks
in next version.

BTW, looks the same check in madvise_remove() can removed too.

Thanks,
Ming
  

On Thu, Feb 01, 2024 at 11:43:11PM -0500, Mike Snitzer wrote:
> On Thu, Feb 01 2024 at  9:20P -0500,
> Ming Lei <ming.lei@redhat.com> wrote:
> 
> > madvise(MADV_POPULATE_READ) tries to populate all page tables in the
> > specific range, so it is usually sequential IO if VMA is backed by
> > file.
> > 
> > Set ra_pages as device max request size for the involved readahead in
> > the ADV_POPULATE_READ, this way reduces latency of madvise(MADV_POPULATE_READ)
> > to 1/10 when running madvise(MADV_POPULATE_READ) over one 1GB file with
> > usual(default) 128KB of read_ahead_kb.
> > 
> > Cc: David Hildenbrand <david@redhat.com>
> > Cc: Matthew Wilcox <willy@infradead.org>
> > Cc: Alexander Viro <viro@zeniv.linux.org.uk>
> > Cc: Christian Brauner <brauner@kernel.org>
> > Cc: Don Dutile <ddutile@redhat.com>
> > Cc: Rafael Aquini <raquini@redhat.com>
> > Cc: Dave Chinner <david@fromorbit.com>
> > Cc: Mike Snitzer <snitzer@kernel.org>
> > Cc: Andrew Morton <akpm@linux-foundation.org>
> > Signed-off-by: Ming Lei <ming.lei@redhat.com>
> > ---
> >  mm/madvise.c | 52 +++++++++++++++++++++++++++++++++++++++++++++++++++-
> >  1 file changed, 51 insertions(+), 1 deletion(-)
> > 
> > diff --git a/mm/madvise.c b/mm/madvise.c
> > index 912155a94ed5..db5452c8abdd 100644
> > --- a/mm/madvise.c
> > +++ b/mm/madvise.c
> > @@ -900,6 +900,37 @@ static long madvise_dontneed_free(struct vm_area_struct *vma,
> >  		return -EINVAL;
> >  }
> >  
> > +static void madvise_restore_ra_win(struct file **file, unsigned int ra_pages)
> > +{
> > +	if (*file) {
> > +		struct file *f = *file;
> > +
> > +		f->f_ra.ra_pages = ra_pages;
> > +		fput(f);
> > +		*file = NULL;
> > +	}
> > +}
> > +
> > +static struct file *madvise_override_ra_win(struct file *f,
> > +		unsigned long start, unsigned long end,
> > +		unsigned int *old_ra_pages)
> > +{
> > +	unsigned int io_pages;
> > +
> > +	if (!f || !f->f_mapping || !f->f_mapping->host)
> > +		return NULL;
> > +
> > +	io_pages = inode_to_bdi(f->f_mapping->host)->io_pages;
> > +	if (((end - start) >> PAGE_SHIFT) < io_pages)
> > +		return NULL;
> > +
> > +	f = get_file(f);
> > +	*old_ra_pages = f->f_ra.ra_pages;
> > +	f->f_ra.ra_pages = io_pages;
> > +
> > +	return f;
> > +}
> > +
> 
> Does this override imply that madvise_populate resorts to calling
> filemap_fault() and here you're just arming it to use the larger
> ->io_pages for the duration of all associated faulting?

Yes.

> 
> Wouldn't it be better to avoid faulting and build up larger page

How can we avoid the fault handling? which is needed to build VA->PA mapping.

> vectors that get sent down to the block layer in one go and let the

filemap_fault() already tries to allocate folio in big size(max order
is MAX_PAGECACHE_ORDER), see page_cache_ra_order() and ra_alloc_folio().

> block layer split using the device's limits? (like happens with
> force_page_cache_ra)

Here filemap code won't deal with block directly because there is VFS &
FS and io mapping is required, and it just calls aops->readahead() or
aops->read_folio(), but block plug & readahead_control are applied for
handling everything in batch.

> 
> I'm concerned that madvise_populate isn't so efficient with filemap

That is why this patch increases readahead window, then
madvise_populate() performance can be improved by X10 in big file-backed
popluate read.

> due to excessive faulting (*BUT* I haven't traced to know, I'm just
> inferring that is why twiddling f->f_ra.ra_pages helps improve
> madvise_populate by having it issue larger IO. Apologies if I'm way
> off base)

As mentioned, fault handling can't be avoided, but we can improve
involved readahead IO perf.



Thanks,
Ming
  

On Fri, Feb 02 2024 at  5:52P -0500,
Ming Lei <ming.lei@redhat.com> wrote:

> On Thu, Feb 01, 2024 at 11:43:11PM -0500, Mike Snitzer wrote:
> > On Thu, Feb 01 2024 at  9:20P -0500,
> > Ming Lei <ming.lei@redhat.com> wrote:
> > 
> > > madvise(MADV_POPULATE_READ) tries to populate all page tables in the
> > > specific range, so it is usually sequential IO if VMA is backed by
> > > file.
> > > 
> > > Set ra_pages as device max request size for the involved readahead in
> > > the ADV_POPULATE_READ, this way reduces latency of madvise(MADV_POPULATE_READ)
> > > to 1/10 when running madvise(MADV_POPULATE_READ) over one 1GB file with
> > > usual(default) 128KB of read_ahead_kb.
> > > 
> > > Cc: David Hildenbrand <david@redhat.com>
> > > Cc: Matthew Wilcox <willy@infradead.org>
> > > Cc: Alexander Viro <viro@zeniv.linux.org.uk>
> > > Cc: Christian Brauner <brauner@kernel.org>
> > > Cc: Don Dutile <ddutile@redhat.com>
> > > Cc: Rafael Aquini <raquini@redhat.com>
> > > Cc: Dave Chinner <david@fromorbit.com>
> > > Cc: Mike Snitzer <snitzer@kernel.org>
> > > Cc: Andrew Morton <akpm@linux-foundation.org>
> > > Signed-off-by: Ming Lei <ming.lei@redhat.com>
> > > ---
> > >  mm/madvise.c | 52 +++++++++++++++++++++++++++++++++++++++++++++++++++-
> > >  1 file changed, 51 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/mm/madvise.c b/mm/madvise.c
> > > index 912155a94ed5..db5452c8abdd 100644
> > > --- a/mm/madvise.c
> > > +++ b/mm/madvise.c
> > > @@ -900,6 +900,37 @@ static long madvise_dontneed_free(struct vm_area_struct *vma,
> > >  		return -EINVAL;
> > >  }
> > >  
> > > +static void madvise_restore_ra_win(struct file **file, unsigned int ra_pages)
> > > +{
> > > +	if (*file) {
> > > +		struct file *f = *file;
> > > +
> > > +		f->f_ra.ra_pages = ra_pages;
> > > +		fput(f);
> > > +		*file = NULL;
> > > +	}
> > > +}
> > > +
> > > +static struct file *madvise_override_ra_win(struct file *f,
> > > +		unsigned long start, unsigned long end,
> > > +		unsigned int *old_ra_pages)
> > > +{
> > > +	unsigned int io_pages;
> > > +
> > > +	if (!f || !f->f_mapping || !f->f_mapping->host)
> > > +		return NULL;
> > > +
> > > +	io_pages = inode_to_bdi(f->f_mapping->host)->io_pages;
> > > +	if (((end - start) >> PAGE_SHIFT) < io_pages)
> > > +		return NULL;
> > > +
> > > +	f = get_file(f);
> > > +	*old_ra_pages = f->f_ra.ra_pages;
> > > +	f->f_ra.ra_pages = io_pages;
> > > +
> > > +	return f;
> > > +}
> > > +
> > 
> > Does this override imply that madvise_populate resorts to calling
> > filemap_fault() and here you're just arming it to use the larger
> > ->io_pages for the duration of all associated faulting?
> 
> Yes.
> 
> > 
> > Wouldn't it be better to avoid faulting and build up larger page
> 
> How can we avoid the fault handling? which is needed to build VA->PA mapping.

I was wondering if it made sense to add fadvise_populate -- but given
my lack of experience with MM I then get handwavvy quick -- I have
more work ahead to round out my MM understanding so that I'm more
informed.
 
> > vectors that get sent down to the block layer in one go and let the
> 
> filemap_fault() already tries to allocate folio in big size(max order
> is MAX_PAGECACHE_ORDER), see page_cache_ra_order() and ra_alloc_folio().
> 
> > block layer split using the device's limits? (like happens with
> > force_page_cache_ra)
> 
> Here filemap code won't deal with block directly because there is VFS &
> FS and io mapping is required, and it just calls aops->readahead() or
> aops->read_folio(), but block plug & readahead_control are applied for
> handling everything in batch.
> 
> > 
> > I'm concerned that madvise_populate isn't so efficient with filemap
> 
> That is why this patch increases readahead window, then
> madvise_populate() performance can be improved by X10 in big file-backed
> popluate read.

Right, as you know I've tested your patch, the larger readahead window
certainly did provide the much more desirable performance.  I'll reply
to your v2 (with reduced negative checks) with my Reviewed-by and
Tested-by.

I was just wondering if there an opportunity to plumb in more a
specific (and potentially better) fadvise_populate for dealing with
file backed pages.

> > due to excessive faulting (*BUT* I haven't traced to know, I'm just
> > inferring that is why twiddling f->f_ra.ra_pages helps improve
> > madvise_populate by having it issue larger IO. Apologies if I'm way
> > off base)
> 
> As mentioned, fault handling can't be avoided, but we can improve
> involved readahead IO perf.

Thanks, and sorry for asking such a naive question (put more pressure
on you to educate than I should have).

Mike
  

On Fri, Feb 02, 2024 at 10:20:29AM +0800, Ming Lei wrote:
> madvise(MADV_POPULATE_READ) tries to populate all page tables in the
> specific range, so it is usually sequential IO if VMA is backed by
> file.
> 
> Set ra_pages as device max request size for the involved readahead in
> the ADV_POPULATE_READ, this way reduces latency of madvise(MADV_POPULATE_READ)
> to 1/10 when running madvise(MADV_POPULATE_READ) over one 1GB file with
> usual(default) 128KB of read_ahead_kb.
> 
> Cc: David Hildenbrand <david@redhat.com>
> Cc: Matthew Wilcox <willy@infradead.org>
> Cc: Alexander Viro <viro@zeniv.linux.org.uk>
> Cc: Christian Brauner <brauner@kernel.org>
> Cc: Don Dutile <ddutile@redhat.com>
> Cc: Rafael Aquini <raquini@redhat.com>
> Cc: Dave Chinner <david@fromorbit.com>
> Cc: Mike Snitzer <snitzer@kernel.org>
> Cc: Andrew Morton <akpm@linux-foundation.org>
> Signed-off-by: Ming Lei <ming.lei@redhat.com>
> ---
>  mm/madvise.c | 52 +++++++++++++++++++++++++++++++++++++++++++++++++++-
>  1 file changed, 51 insertions(+), 1 deletion(-)
> 
> diff --git a/mm/madvise.c b/mm/madvise.c
> index 912155a94ed5..db5452c8abdd 100644
> --- a/mm/madvise.c
> +++ b/mm/madvise.c
> @@ -900,6 +900,37 @@ static long madvise_dontneed_free(struct vm_area_struct *vma,
>  		return -EINVAL;
>  }
>  
> +static void madvise_restore_ra_win(struct file **file, unsigned int ra_pages)
> +{
> +	if (*file) {
> +		struct file *f = *file;
> +
> +		f->f_ra.ra_pages = ra_pages;
> +		fput(f);
> +		*file = NULL;
> +	}
> +}
> +
> +static struct file *madvise_override_ra_win(struct file *f,
> +		unsigned long start, unsigned long end,
> +		unsigned int *old_ra_pages)
> +{
> +	unsigned int io_pages;
> +
> +	if (!f || !f->f_mapping || !f->f_mapping->host)
> +		return NULL;
> +
> +	io_pages = inode_to_bdi(f->f_mapping->host)->io_pages;
> +	if (((end - start) >> PAGE_SHIFT) < io_pages)
> +		return NULL;
> +
> +	f = get_file(f);
> +	*old_ra_pages = f->f_ra.ra_pages;
> +	f->f_ra.ra_pages = io_pages;
> +
> +	return f;
> +}

This won't do what you think if the file has been marked
FMODE_RANDOM before this populate call.

IOWs, I don't think madvise should be digging in the struct file
readahead stuff here. It should call vfs_fadvise(FADV_SEQUENTIAL) to
do the set the readahead mode, rather that try to duplicate
FADV_SEQUENTIAL (badly).  We already do this for WILLNEED to make it
do the right thing, we should be doing the same thing here.

Also, AFAICT, there is no need for get_file()/fput() here - the vma
already has a reference to the struct file, and the vma should not
be going away whilst the madvise() operation is in progress.

-Dave.
  

On Mon, Feb 05, 2024 at 10:34:47AM +1100, Dave Chinner wrote:
> On Fri, Feb 02, 2024 at 10:20:29AM +0800, Ming Lei wrote:
> > madvise(MADV_POPULATE_READ) tries to populate all page tables in the
> > specific range, so it is usually sequential IO if VMA is backed by
> > file.
> > 
> > Set ra_pages as device max request size for the involved readahead in
> > the ADV_POPULATE_READ, this way reduces latency of madvise(MADV_POPULATE_READ)
> > to 1/10 when running madvise(MADV_POPULATE_READ) over one 1GB file with
> > usual(default) 128KB of read_ahead_kb.
> > 
> > Cc: David Hildenbrand <david@redhat.com>
> > Cc: Matthew Wilcox <willy@infradead.org>
> > Cc: Alexander Viro <viro@zeniv.linux.org.uk>
> > Cc: Christian Brauner <brauner@kernel.org>
> > Cc: Don Dutile <ddutile@redhat.com>
> > Cc: Rafael Aquini <raquini@redhat.com>
> > Cc: Dave Chinner <david@fromorbit.com>
> > Cc: Mike Snitzer <snitzer@kernel.org>
> > Cc: Andrew Morton <akpm@linux-foundation.org>
> > Signed-off-by: Ming Lei <ming.lei@redhat.com>
> > ---
> >  mm/madvise.c | 52 +++++++++++++++++++++++++++++++++++++++++++++++++++-
> >  1 file changed, 51 insertions(+), 1 deletion(-)
> > 
> > diff --git a/mm/madvise.c b/mm/madvise.c
> > index 912155a94ed5..db5452c8abdd 100644
> > --- a/mm/madvise.c
> > +++ b/mm/madvise.c
> > @@ -900,6 +900,37 @@ static long madvise_dontneed_free(struct vm_area_struct *vma,
> >  		return -EINVAL;
> >  }
> >  
> > +static void madvise_restore_ra_win(struct file **file, unsigned int ra_pages)
> > +{
> > +	if (*file) {
> > +		struct file *f = *file;
> > +
> > +		f->f_ra.ra_pages = ra_pages;
> > +		fput(f);
> > +		*file = NULL;
> > +	}
> > +}
> > +
> > +static struct file *madvise_override_ra_win(struct file *f,
> > +		unsigned long start, unsigned long end,
> > +		unsigned int *old_ra_pages)
> > +{
> > +	unsigned int io_pages;
> > +
> > +	if (!f || !f->f_mapping || !f->f_mapping->host)
> > +		return NULL;
> > +
> > +	io_pages = inode_to_bdi(f->f_mapping->host)->io_pages;
> > +	if (((end - start) >> PAGE_SHIFT) < io_pages)
> > +		return NULL;
> > +
> > +	f = get_file(f);
> > +	*old_ra_pages = f->f_ra.ra_pages;
> > +	f->f_ra.ra_pages = io_pages;
> > +
> > +	return f;
> > +}
> 
> This won't do what you think if the file has been marked
> FMODE_RANDOM before this populate call.

Yeah.

But madvise(POPULATE_READ) is actually one action,
so userspace can call fadvise(POSIX_FADV_NORMAL) or fadvise(POSIX_FADV_SEQUENTIAL)
before madvise(POPULATE_READ), and set RANDOM advise back after
madvise(POPULATE_READ) returns, so looks not big issue in reality.

> 
> IOWs, I don't think madvise should be digging in the struct file
> readahead stuff here. It should call vfs_fadvise(FADV_SEQUENTIAL) to
> do the set the readahead mode, rather that try to duplicate
> FADV_SEQUENTIAL (badly).  We already do this for WILLNEED to make it
> do the right thing, we should be doing the same thing here.

FADV_SEQUENTIAL doubles current readahead window, which is far from
enough to get top performance, such as, latency of doubling (default) ra
window is still 2X of setting ra windows as bdi->io_pages.

If application sets small 'bdi/read_ahead_kb' just like this report, the
gap can be very big.

Or can we add one API/helper in fs code to set file readahead ra_pages for
this use case?

> 
> Also, AFAICT, there is no need for get_file()/fput() here - the vma
> already has a reference to the struct file, and the vma should not
> be going away whilst the madvise() operation is in progress.

You are right, get_file() is only needed in case of dropping mm lock.


Thanks,
Ming