Patchwork [2/4] mm: writeback: distribute write pages across allowable zones

login
register
mail settings
Submitter Johannes Weiner
Date Sept. 20, 2011, 1:45 p.m.
Message ID <1316526315-16801-3-git-send-email-jweiner@redhat.com>
Download mbox | patch
Permalink /patch/115537/
State Not Applicable
Headers show

Comments

Johannes Weiner - Sept. 20, 2011, 1:45 p.m.
This patch allows allocators to pass __GFP_WRITE when they know in
advance that the allocated page will be written to and become dirty
soon.  The page allocator will then attempt to distribute those
allocations across zones, such that no single zone will end up full of
dirty, and thus more or less, unreclaimable pages.

The global dirty limits are put in proportion to the respective zone's
amount of dirtyable memory and allocations diverted to other zones
when the limit is reached.

For now, the problem remains for NUMA configurations where the zones
allowed for allocation are in sum not big enough to trigger the global
dirty limits, but a future approach to solve this can reuse the
per-zone dirty limit infrastructure laid out in this patch to have
dirty throttling and the flusher threads consider individual zones.

Signed-off-by: Johannes Weiner <jweiner@redhat.com>
---
 include/linux/gfp.h       |    4 ++-
 include/linux/writeback.h |    1 +
 mm/page-writeback.c       |   66 +++++++++++++++++++++++++++++++++++++-------
 mm/page_alloc.c           |   22 ++++++++++++++-
 4 files changed, 80 insertions(+), 13 deletions(-)
Rik van Riel - Sept. 20, 2011, 6:36 p.m.
On 09/20/2011 09:45 AM, Johannes Weiner wrote:
> This patch allows allocators to pass __GFP_WRITE when they know in
> advance that the allocated page will be written to and become dirty
> soon.  The page allocator will then attempt to distribute those
> allocations across zones, such that no single zone will end up full of
> dirty, and thus more or less, unreclaimable pages.
>
> The global dirty limits are put in proportion to the respective zone's
> amount of dirtyable memory and allocations diverted to other zones
> when the limit is reached.
>
> For now, the problem remains for NUMA configurations where the zones
> allowed for allocation are in sum not big enough to trigger the global
> dirty limits, but a future approach to solve this can reuse the
> per-zone dirty limit infrastructure laid out in this patch to have
> dirty throttling and the flusher threads consider individual zones.
>
> Signed-off-by: Johannes Weiner<jweiner@redhat.com>

Reviewed-by: Rik van Riel <riel@redhat.com>

The amount of work done in a __GFP_WRITE allocation looks
a little daunting, but doing that a million times probably
outweighs waiting on the disk even once, so...
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Shaohua Li - Sept. 21, 2011, 11:04 a.m.
On Tue, 2011-09-20 at 21:45 +0800, Johannes Weiner wrote:
> This patch allows allocators to pass __GFP_WRITE when they know in
> advance that the allocated page will be written to and become dirty
> soon.  The page allocator will then attempt to distribute those
> allocations across zones, such that no single zone will end up full of
> dirty, and thus more or less, unreclaimable pages.
> 
> The global dirty limits are put in proportion to the respective zone's
> amount of dirtyable memory and allocations diverted to other zones
> when the limit is reached.
> 
> For now, the problem remains for NUMA configurations where the zones
> allowed for allocation are in sum not big enough to trigger the global
> dirty limits, but a future approach to solve this can reuse the
> per-zone dirty limit infrastructure laid out in this patch to have
> dirty throttling and the flusher threads consider individual zones.
> 
> Signed-off-by: Johannes Weiner <jweiner@redhat.com>
> ---
>  include/linux/gfp.h       |    4 ++-
>  include/linux/writeback.h |    1 +
>  mm/page-writeback.c       |   66 +++++++++++++++++++++++++++++++++++++-------
>  mm/page_alloc.c           |   22 ++++++++++++++-
>  4 files changed, 80 insertions(+), 13 deletions(-)
> 
> diff --git a/include/linux/gfp.h b/include/linux/gfp.h
> index 3a76faf..50efc7e 100644
> --- a/include/linux/gfp.h
> +++ b/include/linux/gfp.h
> @@ -36,6 +36,7 @@ struct vm_area_struct;
>  #endif
>  #define ___GFP_NO_KSWAPD	0x400000u
>  #define ___GFP_OTHER_NODE	0x800000u
> +#define ___GFP_WRITE		0x1000000u
>  
>  /*
>   * GFP bitmasks..
> @@ -85,6 +86,7 @@ struct vm_area_struct;
>  
>  #define __GFP_NO_KSWAPD	((__force gfp_t)___GFP_NO_KSWAPD)
>  #define __GFP_OTHER_NODE ((__force gfp_t)___GFP_OTHER_NODE) /* On behalf of other node */
> +#define __GFP_WRITE	((__force gfp_t)___GFP_WRITE)	/* Allocator intends to dirty page */
>  
>  /*
>   * This may seem redundant, but it's a way of annotating false positives vs.
> @@ -92,7 +94,7 @@ struct vm_area_struct;
>   */
>  #define __GFP_NOTRACK_FALSE_POSITIVE (__GFP_NOTRACK)
>  
> -#define __GFP_BITS_SHIFT 24	/* Room for N __GFP_FOO bits */
> +#define __GFP_BITS_SHIFT 25	/* Room for N __GFP_FOO bits */
>  #define __GFP_BITS_MASK ((__force gfp_t)((1 << __GFP_BITS_SHIFT) - 1))
>  
>  /* This equals 0, but use constants in case they ever change */
> diff --git a/include/linux/writeback.h b/include/linux/writeback.h
> index a5f495f..c96ee0c 100644
> --- a/include/linux/writeback.h
> +++ b/include/linux/writeback.h
> @@ -104,6 +104,7 @@ void laptop_mode_timer_fn(unsigned long data);
>  static inline void laptop_sync_completion(void) { }
>  #endif
>  void throttle_vm_writeout(gfp_t gfp_mask);
> +bool zone_dirty_ok(struct zone *zone);
>  
>  extern unsigned long global_dirty_limit;
>  
> diff --git a/mm/page-writeback.c b/mm/page-writeback.c
> index 9f896db..1fc714c 100644
> --- a/mm/page-writeback.c
> +++ b/mm/page-writeback.c
> @@ -142,6 +142,22 @@ unsigned long global_dirty_limit;
>  static struct prop_descriptor vm_completions;
>  static struct prop_descriptor vm_dirties;
>  
> +static unsigned long zone_dirtyable_memory(struct zone *zone)
> +{
> +	unsigned long x;
> +	/*
> +	 * To keep a reasonable ratio between dirty memory and lowmem,
> +	 * highmem is not considered dirtyable on a global level.
> +	 *
> +	 * But we allow individual highmem zones to hold a potentially
> +	 * bigger share of that global amount of dirty pages as long
> +	 * as they have enough free or reclaimable pages around.
> +	 */
> +	x = zone_page_state(zone, NR_FREE_PAGES) - zone->totalreserve_pages;
> +	x += zone_reclaimable_pages(zone);
> +	return x;
> +}
> +
>  /*
>   * Work out the current dirty-memory clamping and background writeout
>   * thresholds.
> @@ -417,7 +433,7 @@ static unsigned long hard_dirty_limit(unsigned long thresh)
>  }
>  
>  /*
> - * global_dirty_limits - background-writeback and dirty-throttling thresholds
> + * dirty_limits - background-writeback and dirty-throttling thresholds
>   *
>   * Calculate the dirty thresholds based on sysctl parameters
>   * - vm.dirty_background_ratio  or  vm.dirty_background_bytes
> @@ -425,24 +441,35 @@ static unsigned long hard_dirty_limit(unsigned long thresh)
>   * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
>   * real-time tasks.
>   */
> -void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
> +static void dirty_limits(struct zone *zone,
> +			 unsigned long *pbackground,
> +			 unsigned long *pdirty)
>  {
> +	unsigned long uninitialized_var(zone_memory);
> +	unsigned long available_memory;
> +	unsigned long global_memory;
>  	unsigned long background;
> -	unsigned long dirty;
> -	unsigned long uninitialized_var(available_memory);
>  	struct task_struct *tsk;
> +	unsigned long dirty;
>  
> -	if (!vm_dirty_bytes || !dirty_background_bytes)
> -		available_memory = determine_dirtyable_memory();
> +	global_memory = determine_dirtyable_memory();
> +	if (zone)
> +		available_memory = zone_memory = zone_dirtyable_memory(zone);
> +	else
> +		available_memory = global_memory;
>  
> -	if (vm_dirty_bytes)
> +	if (vm_dirty_bytes) {
>  		dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
> -	else
> +		if (zone)
> +			dirty = dirty * zone_memory / global_memory;
> +	} else
>  		dirty = (vm_dirty_ratio * available_memory) / 100;
>  
> -	if (dirty_background_bytes)
> +	if (dirty_background_bytes) {
>  		background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
> -	else
> +		if (zone)
> +			background = background * zone_memory / global_memory;
> +	} else
>  		background = (dirty_background_ratio * available_memory) / 100;
>  
>  	if (background >= dirty)
> @@ -452,9 +479,15 @@ void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
>  		background += background / 4;
>  		dirty += dirty / 4;
>  	}
> +	if (!zone)
> +		trace_global_dirty_state(background, dirty);
>  	*pbackground = background;
>  	*pdirty = dirty;
> -	trace_global_dirty_state(background, dirty);
> +}
> +
> +void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
> +{
> +	dirty_limits(NULL, pbackground, pdirty);
>  }
>  
>  /**
> @@ -875,6 +908,17 @@ void throttle_vm_writeout(gfp_t gfp_mask)
>          }
>  }
>  
> +bool zone_dirty_ok(struct zone *zone)
> +{
> +	unsigned long background_thresh, dirty_thresh;
> +
> +	dirty_limits(zone, &background_thresh, &dirty_thresh);
> +
> +	return zone_page_state(zone, NR_FILE_DIRTY) +
> +		zone_page_state(zone, NR_UNSTABLE_NFS) +
> +		zone_page_state(zone, NR_WRITEBACK) <= dirty_thresh;
> +}
> +
>  /*
>   * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
>   */
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 7e8e2ee..3cca043 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -1368,6 +1368,7 @@ failed:
>  #define ALLOC_HARDER		0x10 /* try to alloc harder */
>  #define ALLOC_HIGH		0x20 /* __GFP_HIGH set */
>  #define ALLOC_CPUSET		0x40 /* check for correct cpuset */
> +#define ALLOC_SLOWPATH		0x80 /* allocator retrying */
>  
>  #ifdef CONFIG_FAIL_PAGE_ALLOC
>  
> @@ -1667,6 +1668,25 @@ zonelist_scan:
>  		if ((alloc_flags & ALLOC_CPUSET) &&
>  			!cpuset_zone_allowed_softwall(zone, gfp_mask))
>  				continue;
> +		/*
> +		 * This may look like it would increase pressure on
> +		 * lower zones by failing allocations in higher zones
> +		 * before they are full.  But once they are full, the
> +		 * allocations fall back to lower zones anyway, and
> +		 * then this check actually protects the lower zones
> +		 * from a flood of dirty page allocations.
if increasing pressure on lower zones isn't a problem since higher zones
will eventually be full, how about a workload without too many writes,
so higher zones will not be full. In such case, increasing low zone
pressure sounds not good.

Thanks,
Shaohua


--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Johannes Weiner - Sept. 21, 2011, 1:35 p.m.
On Wed, Sep 21, 2011 at 07:04:28PM +0800, Shaohua Li wrote:
> On Tue, 2011-09-20 at 21:45 +0800, Johannes Weiner wrote:
> > This patch allows allocators to pass __GFP_WRITE when they know in
> > advance that the allocated page will be written to and become dirty
> > soon.  The page allocator will then attempt to distribute those
> > allocations across zones, such that no single zone will end up full of
> > dirty, and thus more or less, unreclaimable pages.
> > 
> > The global dirty limits are put in proportion to the respective zone's
> > amount of dirtyable memory and allocations diverted to other zones
> > when the limit is reached.
> > 
> > For now, the problem remains for NUMA configurations where the zones
> > allowed for allocation are in sum not big enough to trigger the global
> > dirty limits, but a future approach to solve this can reuse the
> > per-zone dirty limit infrastructure laid out in this patch to have
> > dirty throttling and the flusher threads consider individual zones.
> > 
> > Signed-off-by: Johannes Weiner <jweiner@redhat.com>
> > ---
> >  include/linux/gfp.h       |    4 ++-
> >  include/linux/writeback.h |    1 +
> >  mm/page-writeback.c       |   66 +++++++++++++++++++++++++++++++++++++-------
> >  mm/page_alloc.c           |   22 ++++++++++++++-
> >  4 files changed, 80 insertions(+), 13 deletions(-)
> > 
> > diff --git a/include/linux/gfp.h b/include/linux/gfp.h
> > index 3a76faf..50efc7e 100644
> > --- a/include/linux/gfp.h
> > +++ b/include/linux/gfp.h
> > @@ -36,6 +36,7 @@ struct vm_area_struct;
> >  #endif
> >  #define ___GFP_NO_KSWAPD	0x400000u
> >  #define ___GFP_OTHER_NODE	0x800000u
> > +#define ___GFP_WRITE		0x1000000u
> >  
> >  /*
> >   * GFP bitmasks..
> > @@ -85,6 +86,7 @@ struct vm_area_struct;
> >  
> >  #define __GFP_NO_KSWAPD	((__force gfp_t)___GFP_NO_KSWAPD)
> >  #define __GFP_OTHER_NODE ((__force gfp_t)___GFP_OTHER_NODE) /* On behalf of other node */
> > +#define __GFP_WRITE	((__force gfp_t)___GFP_WRITE)	/* Allocator intends to dirty page */
> >  
> >  /*
> >   * This may seem redundant, but it's a way of annotating false positives vs.
> > @@ -92,7 +94,7 @@ struct vm_area_struct;
> >   */
> >  #define __GFP_NOTRACK_FALSE_POSITIVE (__GFP_NOTRACK)
> >  
> > -#define __GFP_BITS_SHIFT 24	/* Room for N __GFP_FOO bits */
> > +#define __GFP_BITS_SHIFT 25	/* Room for N __GFP_FOO bits */
> >  #define __GFP_BITS_MASK ((__force gfp_t)((1 << __GFP_BITS_SHIFT) - 1))
> >  
> >  /* This equals 0, but use constants in case they ever change */
> > diff --git a/include/linux/writeback.h b/include/linux/writeback.h
> > index a5f495f..c96ee0c 100644
> > --- a/include/linux/writeback.h
> > +++ b/include/linux/writeback.h
> > @@ -104,6 +104,7 @@ void laptop_mode_timer_fn(unsigned long data);
> >  static inline void laptop_sync_completion(void) { }
> >  #endif
> >  void throttle_vm_writeout(gfp_t gfp_mask);
> > +bool zone_dirty_ok(struct zone *zone);
> >  
> >  extern unsigned long global_dirty_limit;
> >  
> > diff --git a/mm/page-writeback.c b/mm/page-writeback.c
> > index 9f896db..1fc714c 100644
> > --- a/mm/page-writeback.c
> > +++ b/mm/page-writeback.c
> > @@ -142,6 +142,22 @@ unsigned long global_dirty_limit;
> >  static struct prop_descriptor vm_completions;
> >  static struct prop_descriptor vm_dirties;
> >  
> > +static unsigned long zone_dirtyable_memory(struct zone *zone)
> > +{
> > +	unsigned long x;
> > +	/*
> > +	 * To keep a reasonable ratio between dirty memory and lowmem,
> > +	 * highmem is not considered dirtyable on a global level.
> > +	 *
> > +	 * But we allow individual highmem zones to hold a potentially
> > +	 * bigger share of that global amount of dirty pages as long
> > +	 * as they have enough free or reclaimable pages around.
> > +	 */
> > +	x = zone_page_state(zone, NR_FREE_PAGES) - zone->totalreserve_pages;
> > +	x += zone_reclaimable_pages(zone);
> > +	return x;
> > +}
> > +
> >  /*
> >   * Work out the current dirty-memory clamping and background writeout
> >   * thresholds.
> > @@ -417,7 +433,7 @@ static unsigned long hard_dirty_limit(unsigned long thresh)
> >  }
> >  
> >  /*
> > - * global_dirty_limits - background-writeback and dirty-throttling thresholds
> > + * dirty_limits - background-writeback and dirty-throttling thresholds
> >   *
> >   * Calculate the dirty thresholds based on sysctl parameters
> >   * - vm.dirty_background_ratio  or  vm.dirty_background_bytes
> > @@ -425,24 +441,35 @@ static unsigned long hard_dirty_limit(unsigned long thresh)
> >   * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
> >   * real-time tasks.
> >   */
> > -void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
> > +static void dirty_limits(struct zone *zone,
> > +			 unsigned long *pbackground,
> > +			 unsigned long *pdirty)
> >  {
> > +	unsigned long uninitialized_var(zone_memory);
> > +	unsigned long available_memory;
> > +	unsigned long global_memory;
> >  	unsigned long background;
> > -	unsigned long dirty;
> > -	unsigned long uninitialized_var(available_memory);
> >  	struct task_struct *tsk;
> > +	unsigned long dirty;
> >  
> > -	if (!vm_dirty_bytes || !dirty_background_bytes)
> > -		available_memory = determine_dirtyable_memory();
> > +	global_memory = determine_dirtyable_memory();
> > +	if (zone)
> > +		available_memory = zone_memory = zone_dirtyable_memory(zone);
> > +	else
> > +		available_memory = global_memory;
> >  
> > -	if (vm_dirty_bytes)
> > +	if (vm_dirty_bytes) {
> >  		dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
> > -	else
> > +		if (zone)
> > +			dirty = dirty * zone_memory / global_memory;
> > +	} else
> >  		dirty = (vm_dirty_ratio * available_memory) / 100;
> >  
> > -	if (dirty_background_bytes)
> > +	if (dirty_background_bytes) {
> >  		background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
> > -	else
> > +		if (zone)
> > +			background = background * zone_memory / global_memory;
> > +	} else
> >  		background = (dirty_background_ratio * available_memory) / 100;
> >  
> >  	if (background >= dirty)
> > @@ -452,9 +479,15 @@ void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
> >  		background += background / 4;
> >  		dirty += dirty / 4;
> >  	}
> > +	if (!zone)
> > +		trace_global_dirty_state(background, dirty);
> >  	*pbackground = background;
> >  	*pdirty = dirty;
> > -	trace_global_dirty_state(background, dirty);
> > +}
> > +
> > +void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
> > +{
> > +	dirty_limits(NULL, pbackground, pdirty);
> >  }
> >  
> >  /**
> > @@ -875,6 +908,17 @@ void throttle_vm_writeout(gfp_t gfp_mask)
> >          }
> >  }
> >  
> > +bool zone_dirty_ok(struct zone *zone)
> > +{
> > +	unsigned long background_thresh, dirty_thresh;
> > +
> > +	dirty_limits(zone, &background_thresh, &dirty_thresh);
> > +
> > +	return zone_page_state(zone, NR_FILE_DIRTY) +
> > +		zone_page_state(zone, NR_UNSTABLE_NFS) +
> > +		zone_page_state(zone, NR_WRITEBACK) <= dirty_thresh;
> > +}
> > +
> >  /*
> >   * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
> >   */
> > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > index 7e8e2ee..3cca043 100644
> > --- a/mm/page_alloc.c
> > +++ b/mm/page_alloc.c
> > @@ -1368,6 +1368,7 @@ failed:
> >  #define ALLOC_HARDER		0x10 /* try to alloc harder */
> >  #define ALLOC_HIGH		0x20 /* __GFP_HIGH set */
> >  #define ALLOC_CPUSET		0x40 /* check for correct cpuset */
> > +#define ALLOC_SLOWPATH		0x80 /* allocator retrying */
> >  
> >  #ifdef CONFIG_FAIL_PAGE_ALLOC
> >  
> > @@ -1667,6 +1668,25 @@ zonelist_scan:
> >  		if ((alloc_flags & ALLOC_CPUSET) &&
> >  			!cpuset_zone_allowed_softwall(zone, gfp_mask))
> >  				continue;
> > +		/*
> > +		 * This may look like it would increase pressure on
> > +		 * lower zones by failing allocations in higher zones
> > +		 * before they are full.  But once they are full, the
> > +		 * allocations fall back to lower zones anyway, and
> > +		 * then this check actually protects the lower zones
> > +		 * from a flood of dirty page allocations.
> if increasing pressure on lower zones isn't a problem since higher zones
> will eventually be full, how about a workload without too many writes,
> so higher zones will not be full. In such case, increasing low zone
> pressure sounds not good.

While there is a shift of dirty pages possible in workloads that were
able to completely fit those pages in the highest zone, the extent of
that shift is limited, which should prevent it from becoming a
practical burden for the lower zones.

Because the number of dirtyable pages does include neither a zone's
lowmem reserves, nor the watermarks, nor kernel allocations, a lower
zone does not receive a bigger share than it can afford when the
allocations are diverted from the higher zones.

To put it into perspective: with these patches there could be an
increased allocation latency for a workload with a writer of fixed
size fitting into the Normal zone and an allocator that suddenly
requires more than 3G (~ DMA32 size minus the 20% allowable dirty
pages) DMA32 memory.  That sounds a bit artificial, to me at least.
And without these patches, we encounter exactly those allocation
latencies on a regular basis when writing files larger than memory.

	Hannes
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Mel Gorman - Sept. 21, 2011, 2:30 p.m.
On Tue, Sep 20, 2011 at 03:45:13PM +0200, Johannes Weiner wrote:
> This patch allows allocators to pass __GFP_WRITE when they know in
> advance that the allocated page will be written to and become dirty
> soon.  The page allocator will then attempt to distribute those
> allocations across zones, such that no single zone will end up full of
> dirty, and thus more or less, unreclaimable pages.
> 

I know this came up the last time but an explanation why lowmem
pressure is not expected to be a problem should be in the changelog.

"At first glance, it would appear that there is a lowmem pressure risk
but it is not the case. Highmem is not considered dirtyable memory.
Hence, if highmem is very large, the global amount of dirty memory
will fit in the highmem zone without falling back to the lower zones
and causing lowmem pressure. If highmem is small then the amount of
pages that 'spill over' to lower zones is limited and no likely to
significantly increase the risk of lowmem pressure due to things like
pagetable page allocations for example. In other words, the timing of
when lowmem pressure happens changes but overall the pressure is roughly
the same".

or something.

> The global dirty limits are put in proportion to the respective zone's
> amount of dirtyable memory and allocations diverted to other zones
> when the limit is reached.
> 
> For now, the problem remains for NUMA configurations where the zones
> allowed for allocation are in sum not big enough to trigger the global
> dirty limits, but a future approach to solve this can reuse the
> per-zone dirty limit infrastructure laid out in this patch to have
> dirty throttling and the flusher threads consider individual zones.
> 

While I think this particular point is important, I don't think it
should be a show stopped for the series.

I'm going to steal Andrew's line here as well - you explain what you are
doing in the patch leader but not why.

> Signed-off-by: Johannes Weiner <jweiner@redhat.com>
> ---
>  include/linux/gfp.h       |    4 ++-
>  include/linux/writeback.h |    1 +
>  mm/page-writeback.c       |   66 +++++++++++++++++++++++++++++++++++++-------
>  mm/page_alloc.c           |   22 ++++++++++++++-
>  4 files changed, 80 insertions(+), 13 deletions(-)
> 
> diff --git a/include/linux/gfp.h b/include/linux/gfp.h
> index 3a76faf..50efc7e 100644
> --- a/include/linux/gfp.h
> +++ b/include/linux/gfp.h
> @@ -36,6 +36,7 @@ struct vm_area_struct;
>  #endif
>  #define ___GFP_NO_KSWAPD	0x400000u
>  #define ___GFP_OTHER_NODE	0x800000u
> +#define ___GFP_WRITE		0x1000000u
>  
>  /*
>   * GFP bitmasks..
> @@ -85,6 +86,7 @@ struct vm_area_struct;
>  
>  #define __GFP_NO_KSWAPD	((__force gfp_t)___GFP_NO_KSWAPD)
>  #define __GFP_OTHER_NODE ((__force gfp_t)___GFP_OTHER_NODE) /* On behalf of other node */
> +#define __GFP_WRITE	((__force gfp_t)___GFP_WRITE)	/* Allocator intends to dirty page */
>  
>  /*
>   * This may seem redundant, but it's a way of annotating false positives vs.
> @@ -92,7 +94,7 @@ struct vm_area_struct;
>   */
>  #define __GFP_NOTRACK_FALSE_POSITIVE (__GFP_NOTRACK)
>  
> -#define __GFP_BITS_SHIFT 24	/* Room for N __GFP_FOO bits */
> +#define __GFP_BITS_SHIFT 25	/* Room for N __GFP_FOO bits */
>  #define __GFP_BITS_MASK ((__force gfp_t)((1 << __GFP_BITS_SHIFT) - 1))
>  
>  /* This equals 0, but use constants in case they ever change */
> diff --git a/include/linux/writeback.h b/include/linux/writeback.h
> index a5f495f..c96ee0c 100644
> --- a/include/linux/writeback.h
> +++ b/include/linux/writeback.h
> @@ -104,6 +104,7 @@ void laptop_mode_timer_fn(unsigned long data);
>  static inline void laptop_sync_completion(void) { }
>  #endif
>  void throttle_vm_writeout(gfp_t gfp_mask);
> +bool zone_dirty_ok(struct zone *zone);
>  
>  extern unsigned long global_dirty_limit;
>  
> diff --git a/mm/page-writeback.c b/mm/page-writeback.c
> index 9f896db..1fc714c 100644
> --- a/mm/page-writeback.c
> +++ b/mm/page-writeback.c
> @@ -142,6 +142,22 @@ unsigned long global_dirty_limit;
>  static struct prop_descriptor vm_completions;
>  static struct prop_descriptor vm_dirties;
>  
> +static unsigned long zone_dirtyable_memory(struct zone *zone)
> +{
> +	unsigned long x;
> +	/*
> +	 * To keep a reasonable ratio between dirty memory and lowmem,
> +	 * highmem is not considered dirtyable on a global level.
> +	 *
> +	 * But we allow individual highmem zones to hold a potentially
> +	 * bigger share of that global amount of dirty pages as long
> +	 * as they have enough free or reclaimable pages around.
> +	 */
> +	x = zone_page_state(zone, NR_FREE_PAGES) - zone->totalreserve_pages;
> +	x += zone_reclaimable_pages(zone);
> +	return x;
> +}
> +
>  /*
>   * Work out the current dirty-memory clamping and background writeout
>   * thresholds.
> @@ -417,7 +433,7 @@ static unsigned long hard_dirty_limit(unsigned long thresh)
>  }
>  
>  /*
> - * global_dirty_limits - background-writeback and dirty-throttling thresholds
> + * dirty_limits - background-writeback and dirty-throttling thresholds
>   *
>   * Calculate the dirty thresholds based on sysctl parameters
>   * - vm.dirty_background_ratio  or  vm.dirty_background_bytes
> @@ -425,24 +441,35 @@ static unsigned long hard_dirty_limit(unsigned long thresh)
>   * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
>   * real-time tasks.
>   */
> -void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
> +static void dirty_limits(struct zone *zone,
> +			 unsigned long *pbackground,
> +			 unsigned long *pdirty)
>  {
> +	unsigned long uninitialized_var(zone_memory);
> +	unsigned long available_memory;
> +	unsigned long global_memory;
>  	unsigned long background;
> -	unsigned long dirty;
> -	unsigned long uninitialized_var(available_memory);
>  	struct task_struct *tsk;
> +	unsigned long dirty;
>  
> -	if (!vm_dirty_bytes || !dirty_background_bytes)
> -		available_memory = determine_dirtyable_memory();
> +	global_memory = determine_dirtyable_memory();
> +	if (zone)
> +		available_memory = zone_memory = zone_dirtyable_memory(zone);
> +	else
> +		available_memory = global_memory;
>  
> -	if (vm_dirty_bytes)
> +	if (vm_dirty_bytes) {
>  		dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
> -	else
> +		if (zone)
> +			dirty = dirty * zone_memory / global_memory;
> +	} else
>  		dirty = (vm_dirty_ratio * available_memory) / 100;
>  
> -	if (dirty_background_bytes)
> +	if (dirty_background_bytes) {
>  		background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
> -	else
> +		if (zone)
> +			background = background * zone_memory / global_memory;
> +	} else
>  		background = (dirty_background_ratio * available_memory) / 100;
>  
>  	if (background >= dirty)
> @@ -452,9 +479,15 @@ void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
>  		background += background / 4;
>  		dirty += dirty / 4;
>  	}
> +	if (!zone)
> +		trace_global_dirty_state(background, dirty);
>  	*pbackground = background;
>  	*pdirty = dirty;
> -	trace_global_dirty_state(background, dirty);
> +}
> +
> +void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
> +{
> +	dirty_limits(NULL, pbackground, pdirty);
>  }
>  
>  /**
> @@ -875,6 +908,17 @@ void throttle_vm_writeout(gfp_t gfp_mask)
>          }
>  }
>  
> +bool zone_dirty_ok(struct zone *zone)
> +{
> +	unsigned long background_thresh, dirty_thresh;
> +
> +	dirty_limits(zone, &background_thresh, &dirty_thresh);
> +
> +	return zone_page_state(zone, NR_FILE_DIRTY) +
> +		zone_page_state(zone, NR_UNSTABLE_NFS) +
> +		zone_page_state(zone, NR_WRITEBACK) <= dirty_thresh;
> +}
> +
>  /*
>   * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
>   */
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 7e8e2ee..3cca043 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -1368,6 +1368,7 @@ failed:
>  #define ALLOC_HARDER		0x10 /* try to alloc harder */
>  #define ALLOC_HIGH		0x20 /* __GFP_HIGH set */
>  #define ALLOC_CPUSET		0x40 /* check for correct cpuset */
> +#define ALLOC_SLOWPATH		0x80 /* allocator retrying */
>  
>  #ifdef CONFIG_FAIL_PAGE_ALLOC
>  
> @@ -1667,6 +1668,25 @@ zonelist_scan:
>  		if ((alloc_flags & ALLOC_CPUSET) &&
>  			!cpuset_zone_allowed_softwall(zone, gfp_mask))
>  				continue;
> +		/*
> +		 * This may look like it would increase pressure on
> +		 * lower zones by failing allocations in higher zones
> +		 * before they are full.  But once they are full, the
> +		 * allocations fall back to lower zones anyway, and
> +		 * then this check actually protects the lower zones
> +		 * from a flood of dirty page allocations.
> +		 *
> +		 * XXX: Allow allocations to potentially exceed the
> +		 * per-zone dirty limit in the slowpath before going
> +		 * into reclaim, which is important when NUMA nodes
> +		 * are not big enough to reach the global limit.  The
> +		 * proper fix on these setups will require awareness
> +		 * of zones in the dirty-throttling and the flusher
> +		 * threads.
> +		 */

Here would be a good reason to explain why we sometimes allow
__GFP_WRITE pages to fall back to lower zones. As it is, the reader
is required to remember that this affects LRU ordering and when/if
reclaim tries to write back the page.

> +		if (!(alloc_flags & ALLOC_SLOWPATH) &&
> +		    (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
> +			goto this_zone_full;
>  
>  		BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
>  		if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
> @@ -2111,7 +2131,7 @@ restart:
>  	 * reclaim. Now things get more complex, so set up alloc_flags according
>  	 * to how we want to proceed.
>  	 */
> -	alloc_flags = gfp_to_alloc_flags(gfp_mask);
> +	alloc_flags = gfp_to_alloc_flags(gfp_mask) | ALLOC_SLOWPATH;
>  

Instead of adding ALLOC_SLOWPATH, check for ALLOC_WMARK_LOW which is
only set in the fast path.

>  	/*
>  	 * Find the true preferred zone if the allocation is unconstrained by

Functionally, I did not find a problem with the patch.
Andrew Morton - Sept. 21, 2011, 11:02 p.m.
On Tue, 20 Sep 2011 15:45:13 +0200
Johannes Weiner <jweiner@redhat.com> wrote:

> This patch allows allocators to pass __GFP_WRITE when they know in
> advance that the allocated page will be written to and become dirty
> soon.  The page allocator will then attempt to distribute those
> allocations across zones, such that no single zone will end up full of
> dirty, and thus more or less, unreclaimable pages.

Across all zones, or across the zones within the node or what?  Some
more description of how all this plays with NUMA is needed, please.

> The global dirty limits are put in proportion to the respective zone's
> amount of dirtyable memory

I don't know what this means.  How can a global limit be controlled by
what is happening within each single zone?  Please describe this design
concept fully.

> and allocations diverted to other zones
> when the limit is reached.

hm.

> For now, the problem remains for NUMA configurations where the zones
> allowed for allocation are in sum not big enough to trigger the global
> dirty limits, but a future approach to solve this can reuse the
> per-zone dirty limit infrastructure laid out in this patch to have
> dirty throttling and the flusher threads consider individual zones.
> 
> ...
>
> --- a/include/linux/gfp.h
> +++ b/include/linux/gfp.h
> @@ -36,6 +36,7 @@ struct vm_area_struct;
>  #endif
>  #define ___GFP_NO_KSWAPD	0x400000u
>  #define ___GFP_OTHER_NODE	0x800000u
> +#define ___GFP_WRITE		0x1000000u
>  
>  /*
>   * GFP bitmasks..
> 
> ...
>
> +static unsigned long zone_dirtyable_memory(struct zone *zone)

Appears to return the number of pages in a particular zone which are
considered "dirtyable".  Some discussion of how this decision is made
would be illuminating.

> +{
> +	unsigned long x;
> +	/*
> +	 * To keep a reasonable ratio between dirty memory and lowmem,
> +	 * highmem is not considered dirtyable on a global level.

Whereabouts in the kernel is this policy implemented? 
determine_dirtyable_memory()?  It does (or can) consider highmem
pages?  Comment seems wrong?

Should we rename determine_dirtyable_memory() to
global_dirtyable_memory(), to get some sense of its relationship with
zone_dirtyable_memory()?

> +	 * But we allow individual highmem zones to hold a potentially
> +	 * bigger share of that global amount of dirty pages as long
> +	 * as they have enough free or reclaimable pages around.
> +	 */
> +	x = zone_page_state(zone, NR_FREE_PAGES) - zone->totalreserve_pages;
> +	x += zone_reclaimable_pages(zone);
> +	return x;
> +}
> +
> 
> ...
>
> -void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
> +static void dirty_limits(struct zone *zone,
> +			 unsigned long *pbackground,
> +			 unsigned long *pdirty)
>  {
> +	unsigned long uninitialized_var(zone_memory);
> +	unsigned long available_memory;
> +	unsigned long global_memory;
>  	unsigned long background;
> -	unsigned long dirty;
> -	unsigned long uninitialized_var(available_memory);
>  	struct task_struct *tsk;
> +	unsigned long dirty;
>  
> -	if (!vm_dirty_bytes || !dirty_background_bytes)
> -		available_memory = determine_dirtyable_memory();
> +	global_memory = determine_dirtyable_memory();
> +	if (zone)
> +		available_memory = zone_memory = zone_dirtyable_memory(zone);
> +	else
> +		available_memory = global_memory;
>  
> -	if (vm_dirty_bytes)
> +	if (vm_dirty_bytes) {
>  		dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
> -	else
> +		if (zone)

So passing zone==NULL alters dirty_limits()'s behaviour.  Seems that it
flips the function between global_dirty_limits and zone_dirty_limits?

Would it be better if we actually had separate global_dirty_limits()
and zone_dirty_limits() rather than a magical mode?

> +			dirty = dirty * zone_memory / global_memory;
> +	} else
>  		dirty = (vm_dirty_ratio * available_memory) / 100;
>  
> -	if (dirty_background_bytes)
> +	if (dirty_background_bytes) {
>  		background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
> -	else
> +		if (zone)
> +			background = background * zone_memory / global_memory;
> +	} else
>  		background = (dirty_background_ratio * available_memory) / 100;
>  
>  	if (background >= dirty)
> 
> ...
>
> +bool zone_dirty_ok(struct zone *zone)

Full description of the return value, please.

> +{
> +	unsigned long background_thresh, dirty_thresh;
> +
> +	dirty_limits(zone, &background_thresh, &dirty_thresh);
> +
> +	return zone_page_state(zone, NR_FILE_DIRTY) +
> +		zone_page_state(zone, NR_UNSTABLE_NFS) +
> +		zone_page_state(zone, NR_WRITEBACK) <= dirty_thresh;
> +}

We never needed to calculate &background_thresh,.  I wonder if that
matters.

> 
> ...
>

--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Johannes Weiner - Sept. 22, 2011, 8:52 a.m.
On Wed, Sep 21, 2011 at 04:02:26PM -0700, Andrew Morton wrote:
> On Tue, 20 Sep 2011 15:45:13 +0200
> Johannes Weiner <jweiner@redhat.com> wrote:
> 
> > This patch allows allocators to pass __GFP_WRITE when they know in
> > advance that the allocated page will be written to and become dirty
> > soon.  The page allocator will then attempt to distribute those
> > allocations across zones, such that no single zone will end up full of
> > dirty, and thus more or less, unreclaimable pages.
> 
> Across all zones, or across the zones within the node or what?  Some
> more description of how all this plays with NUMA is needed, please.

Across the zones the allocator considers for allocation, which on NUMA
is determined by the allocating task's NUMA memory policy.

> > The global dirty limits are put in proportion to the respective zone's
> > amount of dirtyable memory
> 
> I don't know what this means.  How can a global limit be controlled by
> what is happening within each single zone?  Please describe this design
> concept fully.

Yikes, it's mein English.

A zone's dirty limit is to the zone's contribution of dirtyable memory
what the global dirty limit is to the global amount of dirtyable
memory.

As a result, the sum of the dirty limits of all existing zones equals
the global dirty limit, such that no single zone receives more than
its fair share of the globally allowable dirty pages.

When the allocator tries to allocate from the list of allowable zones,
it skips those that have reached their maximum share of dirty pages.

> > For now, the problem remains for NUMA configurations where the zones
> > allowed for allocation are in sum not big enough to trigger the global
> > dirty limits, but a future approach to solve this can reuse the
> > per-zone dirty limit infrastructure laid out in this patch to have
> > dirty throttling and the flusher threads consider individual zones.

> > +static unsigned long zone_dirtyable_memory(struct zone *zone)
> 
> Appears to return the number of pages in a particular zone which are
> considered "dirtyable".  Some discussion of how this decision is made
> would be illuminating.

Is the proportional relationship between zones and the global level a
satisfactory explanation?

Because I am looking for a central place to explain all this.

> > +{
> > +	unsigned long x;
> > +	/*
> > +	 * To keep a reasonable ratio between dirty memory and lowmem,
> > +	 * highmem is not considered dirtyable on a global level.
> 
> Whereabouts in the kernel is this policy implemented? 
> determine_dirtyable_memory()?  It does (or can) consider highmem
> pages?  Comment seems wrong?

Yes, in determine_dirtyable_memory().

It is possible to configure an unreasonable ratio between dirty memory
and lowmem with the vm_highmem_is_dirtyable sysctl.  The point is that
even though highmem is subtracted from the effective amount of global
dirtyable memory again (which is strictly a big-picture measure), we
only care about the individual zone here and so highmem can very much
always hold dirty pages up to its dirty limit.

> Should we rename determine_dirtyable_memory() to
> global_dirtyable_memory(), to get some sense of its relationship with
> zone_dirtyable_memory()?

Sounds good.

> > +	 * But we allow individual highmem zones to hold a potentially
> > +	 * bigger share of that global amount of dirty pages as long
> > +	 * as they have enough free or reclaimable pages around.
> > +	 */
> > +	x = zone_page_state(zone, NR_FREE_PAGES) - zone->totalreserve_pages;
> > +	x += zone_reclaimable_pages(zone);
> > +	return x;
> > +}
> > +
> > 
> > ...
> >
> > -void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
> > +static void dirty_limits(struct zone *zone,
> > +			 unsigned long *pbackground,
> > +			 unsigned long *pdirty)
> >  {
> > +	unsigned long uninitialized_var(zone_memory);
> > +	unsigned long available_memory;
> > +	unsigned long global_memory;
> >  	unsigned long background;
> > -	unsigned long dirty;
> > -	unsigned long uninitialized_var(available_memory);
> >  	struct task_struct *tsk;
> > +	unsigned long dirty;
> >  
> > -	if (!vm_dirty_bytes || !dirty_background_bytes)
> > -		available_memory = determine_dirtyable_memory();
> > +	global_memory = determine_dirtyable_memory();
> > +	if (zone)
> > +		available_memory = zone_memory = zone_dirtyable_memory(zone);
> > +	else
> > +		available_memory = global_memory;
> >  
> > -	if (vm_dirty_bytes)
> > +	if (vm_dirty_bytes) {
> >  		dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
> > -	else
> > +		if (zone)
> 
> So passing zone==NULL alters dirty_limits()'s behaviour.  Seems that it
> flips the function between global_dirty_limits and zone_dirty_limits?

Yes.

> Would it be better if we actually had separate global_dirty_limits()
> and zone_dirty_limits() rather than a magical mode?

I did that the first time around, but Mel raised the valid point that
this will be bad for maintainability.

The global dirty limit and the per-zone dirty limit are not only
incidentally calculated the same way, they are intentionally similar
in the geometrical sense (modulo workarounds for not having fp
arithmetic), so it would be good to keep this stuff together.

But the same applies to determine_dirtyable_memory() and
zone_dirtyable_memory(), so they should be done the same way and I
don't care too much which that would be.

If noone complains, I would structure the code such that
global_dirtyable_memory() and zone_dirtyable_memory(), as well as
global_dirty_limits() and zone_dirty_limits() are separate functions
next to each other with a big fat comment above that block explaining
the per-zone dirty limits and the proportional relationship to the
global parameters.

> > +			dirty = dirty * zone_memory / global_memory;
> > +	} else
> >  		dirty = (vm_dirty_ratio * available_memory) / 100;
> >  
> > -	if (dirty_background_bytes)
> > +	if (dirty_background_bytes) {
> >  		background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
> > -	else
> > +		if (zone)
> > +			background = background * zone_memory / global_memory;
> > +	} else
> >  		background = (dirty_background_ratio * available_memory) / 100;
> >  
> >  	if (background >= dirty)
> > 
> > ...
> >
> > +bool zone_dirty_ok(struct zone *zone)
> 
> Full description of the return value, please.

Returns false when the zone has reached its maximum share of the
global allowed dirty pages, true otherwise.

> 
> > +{
> > +	unsigned long background_thresh, dirty_thresh;
> > +
> > +	dirty_limits(zone, &background_thresh, &dirty_thresh);
> > +
> > +	return zone_page_state(zone, NR_FILE_DIRTY) +
> > +		zone_page_state(zone, NR_UNSTABLE_NFS) +
> > +		zone_page_state(zone, NR_WRITEBACK) <= dirty_thresh;
> > +}
> 
> We never needed to calculate &background_thresh,.  I wonder if that
> matters.

I didn't think dirty_limits() could take another branch, but if I
split up the function I will drop it.  It's not rocket science and can
be easily added on demand.
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch

diff --git a/include/linux/gfp.h b/include/linux/gfp.h
index 3a76faf..50efc7e 100644
--- a/include/linux/gfp.h
+++ b/include/linux/gfp.h
@@ -36,6 +36,7 @@  struct vm_area_struct;
 #endif
 #define ___GFP_NO_KSWAPD	0x400000u
 #define ___GFP_OTHER_NODE	0x800000u
+#define ___GFP_WRITE		0x1000000u
 
 /*
  * GFP bitmasks..
@@ -85,6 +86,7 @@  struct vm_area_struct;
 
 #define __GFP_NO_KSWAPD	((__force gfp_t)___GFP_NO_KSWAPD)
 #define __GFP_OTHER_NODE ((__force gfp_t)___GFP_OTHER_NODE) /* On behalf of other node */
+#define __GFP_WRITE	((__force gfp_t)___GFP_WRITE)	/* Allocator intends to dirty page */
 
 /*
  * This may seem redundant, but it's a way of annotating false positives vs.
@@ -92,7 +94,7 @@  struct vm_area_struct;
  */
 #define __GFP_NOTRACK_FALSE_POSITIVE (__GFP_NOTRACK)
 
-#define __GFP_BITS_SHIFT 24	/* Room for N __GFP_FOO bits */
+#define __GFP_BITS_SHIFT 25	/* Room for N __GFP_FOO bits */
 #define __GFP_BITS_MASK ((__force gfp_t)((1 << __GFP_BITS_SHIFT) - 1))
 
 /* This equals 0, but use constants in case they ever change */
diff --git a/include/linux/writeback.h b/include/linux/writeback.h
index a5f495f..c96ee0c 100644
--- a/include/linux/writeback.h
+++ b/include/linux/writeback.h
@@ -104,6 +104,7 @@  void laptop_mode_timer_fn(unsigned long data);
 static inline void laptop_sync_completion(void) { }
 #endif
 void throttle_vm_writeout(gfp_t gfp_mask);
+bool zone_dirty_ok(struct zone *zone);
 
 extern unsigned long global_dirty_limit;
 
diff --git a/mm/page-writeback.c b/mm/page-writeback.c
index 9f896db..1fc714c 100644
--- a/mm/page-writeback.c
+++ b/mm/page-writeback.c
@@ -142,6 +142,22 @@  unsigned long global_dirty_limit;
 static struct prop_descriptor vm_completions;
 static struct prop_descriptor vm_dirties;
 
+static unsigned long zone_dirtyable_memory(struct zone *zone)
+{
+	unsigned long x;
+	/*
+	 * To keep a reasonable ratio between dirty memory and lowmem,
+	 * highmem is not considered dirtyable on a global level.
+	 *
+	 * But we allow individual highmem zones to hold a potentially
+	 * bigger share of that global amount of dirty pages as long
+	 * as they have enough free or reclaimable pages around.
+	 */
+	x = zone_page_state(zone, NR_FREE_PAGES) - zone->totalreserve_pages;
+	x += zone_reclaimable_pages(zone);
+	return x;
+}
+
 /*
  * Work out the current dirty-memory clamping and background writeout
  * thresholds.
@@ -417,7 +433,7 @@  static unsigned long hard_dirty_limit(unsigned long thresh)
 }
 
 /*
- * global_dirty_limits - background-writeback and dirty-throttling thresholds
+ * dirty_limits - background-writeback and dirty-throttling thresholds
  *
  * Calculate the dirty thresholds based on sysctl parameters
  * - vm.dirty_background_ratio  or  vm.dirty_background_bytes
@@ -425,24 +441,35 @@  static unsigned long hard_dirty_limit(unsigned long thresh)
  * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
  * real-time tasks.
  */
-void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
+static void dirty_limits(struct zone *zone,
+			 unsigned long *pbackground,
+			 unsigned long *pdirty)
 {
+	unsigned long uninitialized_var(zone_memory);
+	unsigned long available_memory;
+	unsigned long global_memory;
 	unsigned long background;
-	unsigned long dirty;
-	unsigned long uninitialized_var(available_memory);
 	struct task_struct *tsk;
+	unsigned long dirty;
 
-	if (!vm_dirty_bytes || !dirty_background_bytes)
-		available_memory = determine_dirtyable_memory();
+	global_memory = determine_dirtyable_memory();
+	if (zone)
+		available_memory = zone_memory = zone_dirtyable_memory(zone);
+	else
+		available_memory = global_memory;
 
-	if (vm_dirty_bytes)
+	if (vm_dirty_bytes) {
 		dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
-	else
+		if (zone)
+			dirty = dirty * zone_memory / global_memory;
+	} else
 		dirty = (vm_dirty_ratio * available_memory) / 100;
 
-	if (dirty_background_bytes)
+	if (dirty_background_bytes) {
 		background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
-	else
+		if (zone)
+			background = background * zone_memory / global_memory;
+	} else
 		background = (dirty_background_ratio * available_memory) / 100;
 
 	if (background >= dirty)
@@ -452,9 +479,15 @@  void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
 		background += background / 4;
 		dirty += dirty / 4;
 	}
+	if (!zone)
+		trace_global_dirty_state(background, dirty);
 	*pbackground = background;
 	*pdirty = dirty;
-	trace_global_dirty_state(background, dirty);
+}
+
+void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
+{
+	dirty_limits(NULL, pbackground, pdirty);
 }
 
 /**
@@ -875,6 +908,17 @@  void throttle_vm_writeout(gfp_t gfp_mask)
         }
 }
 
+bool zone_dirty_ok(struct zone *zone)
+{
+	unsigned long background_thresh, dirty_thresh;
+
+	dirty_limits(zone, &background_thresh, &dirty_thresh);
+
+	return zone_page_state(zone, NR_FILE_DIRTY) +
+		zone_page_state(zone, NR_UNSTABLE_NFS) +
+		zone_page_state(zone, NR_WRITEBACK) <= dirty_thresh;
+}
+
 /*
  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
  */
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 7e8e2ee..3cca043 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -1368,6 +1368,7 @@  failed:
 #define ALLOC_HARDER		0x10 /* try to alloc harder */
 #define ALLOC_HIGH		0x20 /* __GFP_HIGH set */
 #define ALLOC_CPUSET		0x40 /* check for correct cpuset */
+#define ALLOC_SLOWPATH		0x80 /* allocator retrying */
 
 #ifdef CONFIG_FAIL_PAGE_ALLOC
 
@@ -1667,6 +1668,25 @@  zonelist_scan:
 		if ((alloc_flags & ALLOC_CPUSET) &&
 			!cpuset_zone_allowed_softwall(zone, gfp_mask))
 				continue;
+		/*
+		 * This may look like it would increase pressure on
+		 * lower zones by failing allocations in higher zones
+		 * before they are full.  But once they are full, the
+		 * allocations fall back to lower zones anyway, and
+		 * then this check actually protects the lower zones
+		 * from a flood of dirty page allocations.
+		 *
+		 * XXX: Allow allocations to potentially exceed the
+		 * per-zone dirty limit in the slowpath before going
+		 * into reclaim, which is important when NUMA nodes
+		 * are not big enough to reach the global limit.  The
+		 * proper fix on these setups will require awareness
+		 * of zones in the dirty-throttling and the flusher
+		 * threads.
+		 */
+		if (!(alloc_flags & ALLOC_SLOWPATH) &&
+		    (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
+			goto this_zone_full;
 
 		BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
 		if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
@@ -2111,7 +2131,7 @@  restart:
 	 * reclaim. Now things get more complex, so set up alloc_flags according
 	 * to how we want to proceed.
 	 */
-	alloc_flags = gfp_to_alloc_flags(gfp_mask);
+	alloc_flags = gfp_to_alloc_flags(gfp_mask) | ALLOC_SLOWPATH;
 
 	/*
 	 * Find the true preferred zone if the allocation is unconstrained by