[PATCH 1/2] shmem: Support for registration of Driver/file owner specific ops

[PATCH 1/2] shmem: Support for registration of Driver/file owner specific ops

[akash.goel]

From: Chris Wilson <chris@chris-wilson.co.uk>

This provides support for the Drivers or shmem file owners to register
a set of callbacks, which can be invoked from the address space operations
methods implemented by shmem.
This allow the file owners to hook into the shmem address space operations
to do some extra/custom operations in addition to the default ones.

The private_data field of address_space struct is used to store the pointer
to driver specific ops.
Currently only one ops field is defined, which is migratepage, but can be
extended on need basis.

The need for driver specific operations arises since some of the operations
(like migratepage) may not be handled completely within shmem, so as to be
effective, and would need some driver specific handling also.

Specifically, i915.ko would like to participate in migratepage().
i915.ko uses shmemfs to provide swappable backing storage for its user
objects, but when those objects are in use by the GPU it must pin the entire
object until the GPU is idle. As a result, large chunks of memory can be
arbitrarily withdrawn from page migration, resulting in premature
out-of-memory due to fragmentation. However, if i915.ko can receive the
migratepage() request, it can then flush the object from the GPU, remove
its pin and thus enable the migration.

Since Gfx allocations are one of the major consumer of system memory, its
imperative to have such a mechanism to effectively deal with fragmentation.
And therefore the need for such a provision for initiating driver specific
actions during address space operations.

Cc: Hugh Dickins <hughd@google.com>
Cc: linux-mm@kvack.org
Signed-off-by: Sourab Gupta <sourab.gupta@intel.com>
Signed-off-by: Akash Goel <akash.goel@intel.com>
---
 include/linux/shmem_fs.h | 17 +++++++++++++++++
 mm/shmem.c               | 17 ++++++++++++++++-
 2 files changed, 33 insertions(+), 1 deletion(-)

diff --git a/include/linux/shmem_fs.h b/include/linux/shmem_fs.h
index 4d4780c..6cfa76a 100644
--- a/include/linux/shmem_fs.h
+++ b/include/linux/shmem_fs.h
@@ -34,11 +34,28 @@ struct shmem_sb_info {
 	struct mempolicy *mpol;     /* default memory policy for mappings */
 };
 
+struct shmem_dev_info {
+	void *dev_private_data;
+	int (*dev_migratepage)(struct address_space *mapping,
+			       struct page *newpage, struct page *page,
+			       enum migrate_mode mode, void *dev_priv_data);
+};
+
 static inline struct shmem_inode_info *SHMEM_I(struct inode *inode)
 {
 	return container_of(inode, struct shmem_inode_info, vfs_inode);
 }
 
+static inline int shmem_set_device_ops(struct address_space *mapping,
+				struct shmem_dev_info *info)
+{
+	if (mapping->private_data != NULL)
+		return -EEXIST;
+
+	mapping->private_data = info;
+	return 0;
+}
+
 /*
  * Functions in mm/shmem.c called directly from elsewhere:
  */
diff --git a/mm/shmem.c b/mm/shmem.c
index 440e2a7..f8625c4 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -952,6 +952,21 @@ redirty:
 	return 0;
 }
 
+#ifdef CONFIG_MIGRATION
+static int shmem_migratepage(struct address_space *mapping,
+			     struct page *newpage, struct page *page,
+			     enum migrate_mode mode)
+{
+	struct shmem_dev_info *dev_info = mapping->private_data;
+
+	if (dev_info && dev_info->dev_migratepage)
+		return dev_info->dev_migratepage(mapping, newpage, page,
+				mode, dev_info->dev_private_data);
+
+	return migrate_page(mapping, newpage, page, mode);
+}
+#endif
+
 #ifdef CONFIG_NUMA
 #ifdef CONFIG_TMPFS
 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
@@ -3168,7 +3183,7 @@ static const struct address_space_operations shmem_aops = {
 	.write_end	= shmem_write_end,
 #endif
 #ifdef CONFIG_MIGRATION
-	.migratepage	= migrate_page,
+	.migratepage	= shmem_migratepage,
 #endif
 	.error_remove_page = generic_error_remove_page,
 };
-- 
1.9.2

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

[PATCH 2/2] drm/i915: Make pages of GFX allocations movable

[akash.goel]

From: Chris Wilson <chris@chris-wilson.co.uk>

On a long run of more than 2-3 days, physical memory tends to get fragmented
severely, which considerably slows down the system. In such a scenario,
Shrinker is also unable to help as lack of memory is not the actual problem,
since it has been observed that there are enough free pages of 0 order.

To address the issue of external fragementation, kernel does a compaction
(which involves migration of pages) but it's efficacy depends upon how many
pages are marked as MOVABLE, as only those pages can be migrated.

Currently the backing pages for GFX buffers are allocated from shmemfs
with GFP_RECLAIMABLE flag, in units of 4KB pages.
In the case of limited Swap space, it may not be possible always to reclaim
or swap-out pages of all the inactive objects, to make way for free space
allowing formation of higher order groups of physically-contiguous pages
on compaction.

Just marking the GFX pages as MOVABLE will not suffice, as i915 Driver
has to pin the pages if they are in use by GPU, which will prevent their
migration. So the migratepage callback in shmem is also hooked up to get a
notification when kernel initiates the page migration. On the notification,
i915 Driver appropriately unpin the pages.
With this Driver can effectively mark the GFX pages as MOVABLE and hence
mitigate the fragmentation problem.

Cc: Hugh Dickins <hughd@google.com>
Cc: linux-mm@kvack.org
Signed-off-by: Sourab Gupta <sourab.gupta@intel.com>
Signed-off-by: Akash Goel <akash.goel@intel.com>
---
 drivers/gpu/drm/i915/i915_drv.h |   3 +
 drivers/gpu/drm/i915/i915_gem.c | 128 +++++++++++++++++++++++++++++++++++++++-
 2 files changed, 128 insertions(+), 3 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index f330a53..28e50c7 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -52,6 +52,7 @@
 #include <linux/intel-iommu.h>
 #include <linux/kref.h>
 #include <linux/pm_qos.h>
+#include <linux/shmem_fs.h>
 #include "intel_guc.h"
 #include "intel_dpll_mgr.h"
 
@@ -1966,6 +1967,8 @@ struct drm_i915_private {
 
 	struct intel_encoder *dig_port_map[I915_MAX_PORTS];
 
+	struct shmem_dev_info  migrate_info;
+
 	/*
 	 * NOTE: This is the dri1/ums dungeon, don't add stuff here. Your patch
 	 * will be rejected. Instead look for a better place.
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index 8588c83..a4af5b6 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -33,6 +33,7 @@
 #include "i915_trace.h"
 #include "intel_drv.h"
 #include <linux/shmem_fs.h>
+#include <linux/migrate.h>
 #include <linux/slab.h>
 #include <linux/swap.h>
 #include <linux/pci.h>
@@ -2204,6 +2205,7 @@ i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
 		if (obj->madv == I915_MADV_WILLNEED)
 			mark_page_accessed(page);
 
+		set_page_private(page, 0);
 		page_cache_release(page);
 	}
 	obj->dirty = 0;
@@ -2318,6 +2320,7 @@ i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
 			sg->length += PAGE_SIZE;
 		}
 		last_pfn = page_to_pfn(page);
+		set_page_private(page, (unsigned long)obj);
 
 		/* Check that the i965g/gm workaround works. */
 		WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL));
@@ -2343,8 +2346,11 @@ i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
 
 err_pages:
 	sg_mark_end(sg);
-	for_each_sg_page(st->sgl, &sg_iter, st->nents, 0)
-		page_cache_release(sg_page_iter_page(&sg_iter));
+	for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
+		page = sg_page_iter_page(&sg_iter);
+		set_page_private(page, 0);
+		page_cache_release(page);
+	}
 	sg_free_table(st);
 	kfree(st);
 
@@ -4465,9 +4471,116 @@ static const struct drm_i915_gem_object_ops i915_gem_object_ops = {
 	.put_pages = i915_gem_object_put_pages_gtt,
 };
 
+#ifdef CONFIG_MIGRATION
+static int i915_migratepage(struct address_space *mapping,
+			    struct page *newpage, struct page *page,
+			    enum migrate_mode mode, void *dev_priv_data)
+{
+	struct drm_i915_private *dev_priv = dev_priv_data;
+	struct drm_device *dev = dev_priv->dev;
+	struct drm_i915_gem_object *obj;
+	unsigned long timeout = msecs_to_jiffies(10) + 1;
+	int ret = 0;
+
+	WARN((page_count(newpage) != 1), "Unexpected ref count for newpage\n");
+
+	/*
+	 * Clear the private field of the new target page as it could have a
+	 * stale value in the private field. Otherwise later on if this page
+	 * itself gets migrated, without getting referred by the Driver
+	 * in between, the stale value would cause the i915_migratepage
+	 * function to go for a toss as object pointer is derived from it.
+	 * This should be safe since at the time of migration, private field
+	 * of the new page (which is actually an independent free 4KB page now)
+	 * should be like a don't care for the kernel.
+	 */
+	set_page_private(newpage, 0);
+
+	/*
+	 * Check the page count, if Driver also has a reference then it should
+	 * be more than 2, as shmem will have one reference and one reference
+	 * would have been taken by the migration path itself. So if reference
+	 * is <=2, we can directly invoke the migration function.
+	 */
+	if (page_count(page) <= 2)
+		goto migrate;
+
+	/*
+	 * Use trylock here, with a timeout, for struct_mutex as
+	 * otherwise there is a possibility of deadlock due to lock
+	 * inversion. This path, which tries to migrate a particular
+	 * page after locking that page, can race with a path which
+	 * truncate/purge pages of the corresponding object (after
+	 * acquiring struct_mutex). Since page truncation will also
+	 * try to lock the page, a scenario of deadlock can arise.
+	 */
+	while (!mutex_trylock(&dev->struct_mutex) && --timeout)
+		schedule_timeout_killable(1);
+	if (timeout == 0) {
+		DRM_DEBUG_DRIVER("Unable to acquire device mutex.\n");
+		return -EBUSY;
+	}
+
+	obj = (struct drm_i915_gem_object *)page_private(page);
+
+	if (!PageSwapCache(page) && obj) {
+		/*
+		 * Avoid the migration of pages if they are being actively used
+		 * by GPU or pinned for Display.
+		 * Also skip the migration for purgeable objects otherwise there
+		 * will be a deadlock when shmem will try to lock the page for
+		 * truncation, which is already locked by the caller before
+		 * migration.
+		 * Also HW access would be required for a bound object for which
+		 * device has to be kept runtime active. But a deadlock scenario
+		 * can arise if the attempt is made to resume the device, when
+		 * either a suspend or a resume operation is already happening
+		 * concurrently from some other path and that only actually
+		 * triggered the compaction. So only unbind if the device is
+		 * currently runtime active.
+		 */
+		if (obj->active || obj->pin_display ||
+		    obj->madv == I915_MADV_DONTNEED ||
+		    !intel_runtime_pm_get_if_in_use(dev_priv)) {
+			ret = -EBUSY;
+		} else {
+			ret = drop_pages(obj);
+			intel_runtime_pm_put(dev_priv);
+		}
+
+		BUG_ON(!ret && page_private(page));
+	}
+
+	mutex_unlock(&dev->struct_mutex);
+	if (ret)
+		return ret;
+
+	/*
+	 * Ideally here we don't expect the page count to be > 2, as driver
+	 * would have dropped its reference, but occasionally it has been seen
+	 * coming as 3 & 4. This leads to a situation of unexpected page count,
+	 * causing migration failure, with -EGAIN error. This then leads to
+	 * multiple attempts by the kernel to migrate the same set of pages.
+	 * And sometimes the repeated attempts proves detrimental for stability.
+	 * Also since we don't know who is the other owner, and for how long its
+	 * gonna keep the reference, its better to return -EBUSY.
+	 */
+	if (page_count(page) > 2)
+		return -EBUSY;
+
+migrate:
+	ret = migrate_page(mapping, newpage, page, mode);
+	if (ret)
+		DRM_DEBUG_DRIVER("page=%p migration returned %d\n", page, ret);
+
+	return ret;
+}
+#endif
+
 struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
 						  size_t size)
 {
+	struct drm_i915_private *dev_priv = dev->dev_private;
 	struct drm_i915_gem_object *obj;
 	struct address_space *mapping;
 	gfp_t mask;
@@ -4481,7 +4594,7 @@ struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
 		return NULL;
 	}
 
-	mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
+	mask = GFP_HIGHUSER_MOVABLE;
 	if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) {
 		/* 965gm cannot relocate objects above 4GiB. */
 		mask &= ~__GFP_HIGHMEM;
@@ -4491,6 +4604,10 @@ struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
 	mapping = file_inode(obj->base.filp)->i_mapping;
 	mapping_set_gfp_mask(mapping, mask);
 
+#ifdef CONFIG_MIGRATION
+	shmem_set_device_ops(mapping, &dev_priv->migrate_info);
+#endif
+
 	i915_gem_object_init(obj, &i915_gem_object_ops);
 
 	obj->base.write_domain = I915_GEM_DOMAIN_CPU;
@@ -4993,6 +5110,11 @@ int i915_gem_init(struct drm_device *dev)
 		ret = 0;
 	}
 
+#ifdef CONFIG_MIGRATION
+	dev_priv->migrate_info.dev_private_data = dev_priv;
+	dev_priv->migrate_info.dev_migratepage = i915_migratepage;
+#endif
+
 out_unlock:
 	intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
 	mutex_unlock(&dev->struct_mutex);
-- 
1.9.2

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH 2/2] drm/i915: Make pages of GFX allocations movable

[Chris Wilson]

On Wed, Mar 23, 2016 at 11:39:44AM +0530, akash.goel@intel.com wrote:
> +#ifdef CONFIG_MIGRATION
> +static int i915_migratepage(struct address_space *mapping,
> +			    struct page *newpage, struct page *page,
> +			    enum migrate_mode mode, void *dev_priv_data)

If we move this to i915_gem_shrink_migratepage (i.e. i915_gem_shrink),
we can

> +	/*
> +	 * Use trylock here, with a timeout, for struct_mutex as
> +	 * otherwise there is a possibility of deadlock due to lock
> +	 * inversion. This path, which tries to migrate a particular
> +	 * page after locking that page, can race with a path which
> +	 * truncate/purge pages of the corresponding object (after
> +	 * acquiring struct_mutex). Since page truncation will also
> +	 * try to lock the page, a scenario of deadlock can arise.
> +	 */
> +	while (!mutex_trylock(&dev->struct_mutex) && --timeout)
> +		schedule_timeout_killable(1);

replace this with i915_gem_shrinker_lock() and like constructs with the
other shrinkers. Any reason for dropping the early
if (!page_private(obj)) skip?

Similarly there are other patterns here that would benefit from
integration with existing shrinker logic. However, things like tidying
up the pin_display, unbinding, rpm lock inversion are still only on
list.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH 2/2] drm/i915: Make pages of GFX allocations movable

[Goel, Akash]

On 3/23/2016 1:28 PM, Chris Wilson wrote:
> On Wed, Mar 23, 2016 at 11:39:44AM +0530, akash.goel@intel.com wrote:
>> +#ifdef CONFIG_MIGRATION
>> +static int i915_migratepage(struct address_space *mapping,
>> +			    struct page *newpage, struct page *page,
>> +			    enum migrate_mode mode, void *dev_priv_data)
>
> If we move this to i915_gem_shrink_migratepage (i.e. i915_gem_shrink),
> we can
>
>> +	/*
>> +	 * Use trylock here, with a timeout, for struct_mutex as
>> +	 * otherwise there is a possibility of deadlock due to lock
>> +	 * inversion. This path, which tries to migrate a particular
>> +	 * page after locking that page, can race with a path which
>> +	 * truncate/purge pages of the corresponding object (after
>> +	 * acquiring struct_mutex). Since page truncation will also
>> +	 * try to lock the page, a scenario of deadlock can arise.
>> +	 */
>> +	while (!mutex_trylock(&dev->struct_mutex) && --timeout)
>> +		schedule_timeout_killable(1);
>
> replace this with i915_gem_shrinker_lock() and like constructs with the
> other shrinkers.

fine, will rename the function to gem_shrink_migratepage, move it inside 
the gem_shrinker.c file, and use the existing constructs.

 > Any reason for dropping the early
 > if (!page_private(obj)) skip?
 >

Would this sequence be fine ?

	if (!page_private(page))
		goto migrate; /*skip */

	Loop for locking mutex

	obj = (struct drm_i915_gem_object *)page_private(page);

	if (!PageSwapCache(page) && obj) {


> Similarly there are other patterns here that would benefit from
> integration with existing shrinker logic. However, things like tidying
> up the pin_display, unbinding, rpm lock inversion are still only on
> list.

Tidying, like split that one single if condition into multiple if, else 
if blocks ?

Best regards
Akash
> -Chris
>

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH 2/2] drm/i915: Make pages of GFX allocations movable

[Chris Wilson]

On Wed, Mar 23, 2016 at 01:55:14PM +0530, Goel, Akash wrote:
> 
> 
> On 3/23/2016 1:28 PM, Chris Wilson wrote:
> >On Wed, Mar 23, 2016 at 11:39:44AM +0530, akash.goel@intel.com wrote:
> >>+#ifdef CONFIG_MIGRATION
> >>+static int i915_migratepage(struct address_space *mapping,
> >>+			    struct page *newpage, struct page *page,
> >>+			    enum migrate_mode mode, void *dev_priv_data)
> >
> >If we move this to i915_gem_shrink_migratepage (i.e. i915_gem_shrink),
> >we can
> >
> >>+	/*
> >>+	 * Use trylock here, with a timeout, for struct_mutex as
> >>+	 * otherwise there is a possibility of deadlock due to lock
> >>+	 * inversion. This path, which tries to migrate a particular
> >>+	 * page after locking that page, can race with a path which
> >>+	 * truncate/purge pages of the corresponding object (after
> >>+	 * acquiring struct_mutex). Since page truncation will also
> >>+	 * try to lock the page, a scenario of deadlock can arise.
> >>+	 */
> >>+	while (!mutex_trylock(&dev->struct_mutex) && --timeout)
> >>+		schedule_timeout_killable(1);
> >
> >replace this with i915_gem_shrinker_lock() and like constructs with the
> >other shrinkers.
> 
> fine, will rename the function to gem_shrink_migratepage, move it
> inside the gem_shrinker.c file, and use the existing constructs.
> 
> > Any reason for dropping the early
> > if (!page_private(obj)) skip?
> >
> 
> Would this sequence be fine ?
> 
> 	if (!page_private(page))
> 		goto migrate; /*skip */
> 
> 	Loop for locking mutex
> 
> 	obj = (struct drm_i915_gem_object *)page_private(page);
> 
> 	if (!PageSwapCache(page) && obj) {

Yes.

> >Similarly there are other patterns here that would benefit from
> >integration with existing shrinker logic. However, things like tidying
> >up the pin_display, unbinding, rpm lock inversion are still only on
> >list.
> 
> Tidying, like split that one single if condition into multiple if,
> else if blocks ?

Just outstanding patches that simplify the condition and work we have to
do here.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

[PATCH v2 2/2] drm/i915: Make pages of GFX allocations movable

[akash.goel]

From: Chris Wilson <chris@chris-wilson.co.uk>

On a long run of more than 2-3 days, physical memory tends to get fragmented
severely, which considerably slows down the system. In such a scenario,
Shrinker is also unable to help as lack of memory is not the actual problem,
since it has been observed that there are enough free pages of 0 order.

To address the issue of external fragementation, kernel does a compaction
(which involves migration of pages) but it's efficacy depends upon how many
pages are marked as MOVABLE, as only those pages can be migrated.

Currently the backing pages for GFX buffers are allocated from shmemfs
with GFP_RECLAIMABLE flag, in units of 4KB pages.
In the case of limited Swap space, it may not be possible always to reclaim
or swap-out pages of all the inactive objects, to make way for free space
allowing formation of higher order groups of physically-contiguous pages
on compaction.

Just marking the GFX pages as MOVABLE will not suffice, as i915 Driver
has to pin the pages if they are in use by GPU, which will prevent their
migration. So the migratepage callback in shmem is also hooked up to get a
notification when kernel initiates the page migration. On the notification,
i915 Driver appropriately unpin the pages.
With this Driver can effectively mark the GFX pages as MOVABLE and hence
mitigate the fragmentation problem.

v2:
 - Rename the migration routine to gem_shrink_migratepage, move it to the
   shrinker file, and use the existing constructs (Chris)
 - To cleanup, add a new helper function to encapsulate all page migration
   skip conditions (Chris)
 - Add a new local helper function in shrinker file, for dropping the
   backing pages, and call the same from gem_shrink() also (Chris)

Cc: Hugh Dickins <hughd@google.com>
Cc: linux-mm@kvack.org
Signed-off-by: Sourab Gupta <sourab.gupta@intel.com>
Signed-off-by: Akash Goel <akash.goel@intel.com>
---
 drivers/gpu/drm/i915/i915_drv.h          |   3 +
 drivers/gpu/drm/i915/i915_gem.c          |  16 ++-
 drivers/gpu/drm/i915/i915_gem_shrinker.c | 173 ++++++++++++++++++++++++++++---
 3 files changed, 177 insertions(+), 15 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index f330a53..28e50c7 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -52,6 +52,7 @@
 #include <linux/intel-iommu.h>
 #include <linux/kref.h>
 #include <linux/pm_qos.h>
+#include <linux/shmem_fs.h>
 #include "intel_guc.h"
 #include "intel_dpll_mgr.h"
 
@@ -1966,6 +1967,8 @@ struct drm_i915_private {
 
 	struct intel_encoder *dig_port_map[I915_MAX_PORTS];
 
+	struct shmem_dev_info  migrate_info;
+
 	/*
 	 * NOTE: This is the dri1/ums dungeon, don't add stuff here. Your patch
 	 * will be rejected. Instead look for a better place.
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index 8588c83..59ab58a 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -2204,6 +2204,7 @@ i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
 		if (obj->madv == I915_MADV_WILLNEED)
 			mark_page_accessed(page);
 
+		set_page_private(page, 0);
 		page_cache_release(page);
 	}
 	obj->dirty = 0;
@@ -2318,6 +2319,7 @@ i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
 			sg->length += PAGE_SIZE;
 		}
 		last_pfn = page_to_pfn(page);
+		set_page_private(page, (unsigned long)obj);
 
 		/* Check that the i965g/gm workaround works. */
 		WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL));
@@ -2343,8 +2345,11 @@ i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
 
 err_pages:
 	sg_mark_end(sg);
-	for_each_sg_page(st->sgl, &sg_iter, st->nents, 0)
-		page_cache_release(sg_page_iter_page(&sg_iter));
+	for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
+		page = sg_page_iter_page(&sg_iter);
+		set_page_private(page, 0);
+		page_cache_release(page);
+	}
 	sg_free_table(st);
 	kfree(st);
 
@@ -4468,6 +4473,7 @@ static const struct drm_i915_gem_object_ops i915_gem_object_ops = {
 struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
 						  size_t size)
 {
+	struct drm_i915_private *dev_priv = dev->dev_private;
 	struct drm_i915_gem_object *obj;
 	struct address_space *mapping;
 	gfp_t mask;
@@ -4481,7 +4487,7 @@ struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
 		return NULL;
 	}
 
-	mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
+	mask = GFP_HIGHUSER_MOVABLE;
 	if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) {
 		/* 965gm cannot relocate objects above 4GiB. */
 		mask &= ~__GFP_HIGHMEM;
@@ -4491,6 +4497,10 @@ struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
 	mapping = file_inode(obj->base.filp)->i_mapping;
 	mapping_set_gfp_mask(mapping, mask);
 
+#ifdef CONFIG_MIGRATION
+	shmem_set_device_ops(mapping, &dev_priv->migrate_info);
+#endif
+
 	i915_gem_object_init(obj, &i915_gem_object_ops);
 
 	obj->base.write_domain = I915_GEM_DOMAIN_CPU;
diff --git a/drivers/gpu/drm/i915/i915_gem_shrinker.c b/drivers/gpu/drm/i915/i915_gem_shrinker.c
index d3c473f..f6c2c20 100644
--- a/drivers/gpu/drm/i915/i915_gem_shrinker.c
+++ b/drivers/gpu/drm/i915/i915_gem_shrinker.c
@@ -24,6 +24,7 @@
 
 #include <linux/oom.h>
 #include <linux/shmem_fs.h>
+#include <linux/migrate.h>
 #include <linux/slab.h>
 #include <linux/swap.h>
 #include <linux/pci.h>
@@ -87,6 +88,71 @@ static bool can_release_pages(struct drm_i915_gem_object *obj)
 	return swap_available() || obj->madv == I915_MADV_DONTNEED;
 }
 
+static bool can_migrate_page(struct drm_i915_gem_object *obj)
+{
+	/* Avoid the migration of page if being actively used by GPU */
+	if (obj->active)
+		return false;
+
+	/* Skip the migration for purgeable objects otherwise there
+	 * will be a deadlock when shmem will try to lock the page for
+	 * truncation, which is already locked by the caller before
+	 * migration.
+	 */
+	if (obj->madv == I915_MADV_DONTNEED)
+		return false;
+
+	/* Skip the migration for a pinned object */
+	if (obj->pages_pin_count != num_vma_bound(obj))
+		return false;
+
+	return true;
+}
+
+static int
+unsafe_drop_pages(struct drm_i915_gem_object *obj)
+{
+	struct i915_vma *vma, *next;
+	int ret;
+
+	drm_gem_object_reference(&obj->base);
+	list_for_each_entry_safe(vma, next, &obj->vma_list, obj_link)
+		if (i915_vma_unbind(vma))
+			break;
+
+	ret = i915_gem_object_put_pages(obj);
+	drm_gem_object_unreference(&obj->base);
+
+	return ret;
+}
+
+static int
+do_migrate_page(struct drm_i915_gem_object *obj)
+{
+	struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
+	int ret = 0;
+
+	if (!can_migrate_page(obj))
+		return -EBUSY;
+
+	/* HW access would be required for a bound object for which
+	 * device has to be kept runtime active. But a deadlock scenario
+	 * can arise if the attempt is made to resume the device, when
+	 * either a suspend or a resume operation is already happening
+	 * concurrently from some other path and that only actually
+	 * triggered the compaction. So only unbind if the device is
+	 * currently runtime active.
+	 */
+	if (!intel_runtime_pm_get_if_in_use(dev_priv))
+		return -EBUSY;
+
+	if (!unsafe_drop_pages(obj))
+		ret = -EBUSY;
+
+	intel_runtime_pm_put(dev_priv);
+	return ret;
+}
+
 /**
  * i915_gem_shrink - Shrink buffer object caches
  * @dev_priv: i915 device
@@ -156,7 +222,6 @@ i915_gem_shrink(struct drm_i915_private *dev_priv,
 		INIT_LIST_HEAD(&still_in_list);
 		while (count < target && !list_empty(phase->list)) {
 			struct drm_i915_gem_object *obj;
-			struct i915_vma *vma, *v;
 
 			obj = list_first_entry(phase->list,
 					       typeof(*obj), global_list);
@@ -172,18 +237,8 @@ i915_gem_shrink(struct drm_i915_private *dev_priv,
 			if (!can_release_pages(obj))
 				continue;
 
-			drm_gem_object_reference(&obj->base);
-
-			/* For the unbound phase, this should be a no-op! */
-			list_for_each_entry_safe(vma, v,
-						 &obj->vma_list, obj_link)
-				if (i915_vma_unbind(vma))
-					break;
-
-			if (i915_gem_object_put_pages(obj) == 0)
+			if (unsafe_drop_pages(obj) == 0)
 				count += obj->base.size >> PAGE_SHIFT;
-
-			drm_gem_object_unreference(&obj->base);
 		}
 		list_splice(&still_in_list, phase->list);
 	}
@@ -356,6 +411,95 @@ i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
 	return NOTIFY_DONE;
 }
 
+#ifdef CONFIG_MIGRATION
+static int i915_gem_shrink_migratepage(struct address_space *mapping,
+			    struct page *newpage, struct page *page,
+			    enum migrate_mode mode, void *dev_priv_data)
+{
+	struct drm_i915_private *dev_priv = dev_priv_data;
+	struct drm_device *dev = dev_priv->dev;
+	struct drm_i915_gem_object *obj;
+	unsigned long timeout = msecs_to_jiffies(10) + 1;
+	bool unlock;
+	int ret = 0;
+
+	WARN((page_count(newpage) != 1), "Unexpected ref count for newpage\n");
+
+	/*
+	 * Clear the private field of the new target page as it could have a
+	 * stale value in the private field. Otherwise later on if this page
+	 * itself gets migrated, without getting referred by the Driver
+	 * in between, the stale value would cause the i915_migratepage
+	 * function to go for a toss as object pointer is derived from it.
+	 * This should be safe since at the time of migration, private field
+	 * of the new page (which is actually an independent free 4KB page now)
+	 * should be like a don't care for the kernel.
+	 */
+	set_page_private(newpage, 0);
+
+	if (!page_private(page))
+		goto migrate;
+
+	/*
+	 * Check the page count, if Driver also has a reference then it should
+	 * be more than 2, as shmem will have one reference and one reference
+	 * would have been taken by the migration path itself. So if reference
+	 * is <=2, we can directly invoke the migration function.
+	 */
+	if (page_count(page) <= 2)
+		goto migrate;
+
+	/*
+	 * Use trylock here, with a timeout, for struct_mutex as
+	 * otherwise there is a possibility of deadlock due to lock
+	 * inversion. This path, which tries to migrate a particular
+	 * page after locking that page, can race with a path which
+	 * truncate/purge pages of the corresponding object (after
+	 * acquiring struct_mutex). Since page truncation will also
+	 * try to lock the page, a scenario of deadlock can arise.
+	 */
+	while (!i915_gem_shrinker_lock(dev, &unlock) && --timeout)
+		schedule_timeout_killable(1);
+	if (timeout == 0) {
+		DRM_DEBUG_DRIVER("Unable to acquire device mutex.\n");
+		return -EBUSY;
+	}
+
+	obj = (struct drm_i915_gem_object *)page_private(page);
+
+	if (!PageSwapCache(page) && obj) {
+		ret = do_migrate_page(obj);
+		BUG_ON(!ret && page_private(page));
+	}
+
+	if (unlock)
+		mutex_unlock(&dev->struct_mutex);
+
+	if (ret)
+		return ret;
+
+	/*
+	 * Ideally here we don't expect the page count to be > 2, as driver
+	 * would have dropped its reference, but occasionally it has been seen
+	 * coming as 3 & 4. This leads to a situation of unexpected page count,
+	 * causing migration failure, with -EGAIN error. This then leads to
+	 * multiple attempts by the kernel to migrate the same set of pages.
+	 * And sometimes the repeated attempts proves detrimental for stability.
+	 * Also since we don't know who is the other owner, and for how long its
+	 * gonna keep the reference, its better to return -EBUSY.
+	 */
+	if (page_count(page) > 2)
+		return -EBUSY;
+
+migrate:
+	ret = migrate_page(mapping, newpage, page, mode);
+	if (ret)
+		DRM_DEBUG_DRIVER("page=%p migration returned %d\n", page, ret);
+
+	return ret;
+}
+#endif
+
 /**
  * i915_gem_shrinker_init - Initialize i915 shrinker
  * @dev_priv: i915 device
@@ -371,6 +515,11 @@ void i915_gem_shrinker_init(struct drm_i915_private *dev_priv)
 
 	dev_priv->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
 	WARN_ON(register_oom_notifier(&dev_priv->mm.oom_notifier));
+
+#ifdef CONFIG_MIGRATION
+	dev_priv->migrate_info.dev_private_data = dev_priv;
+	dev_priv->migrate_info.dev_migratepage = i915_gem_shrink_migratepage;
+#endif
 }
 
 /**
-- 
1.9.2

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH v2 2/2] drm/i915: Make pages of GFX allocations movable

[Chris Wilson]

On Thu, Mar 24, 2016 at 11:52:59PM +0530, akash.goel@intel.com wrote:
> +static int
> +unsafe_drop_pages(struct drm_i915_gem_object *obj)
> +{
> +	struct i915_vma *vma, *next;
> +	int ret;
> +
> +	drm_gem_object_reference(&obj->base);
> +	list_for_each_entry_safe(vma, next, &obj->vma_list, obj_link)
> +		if (i915_vma_unbind(vma))
> +			break;
> +
> +	ret = i915_gem_object_put_pages(obj);
> +	drm_gem_object_unreference(&obj->base);
> +
> +	return ret;
> +}
> +
> +static int
> +do_migrate_page(struct drm_i915_gem_object *obj)
> +{
> +	struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
> +	int ret = 0;
> +
> +	if (!can_migrate_page(obj))
> +		return -EBUSY;
> +
> +	/* HW access would be required for a bound object for which
> +	 * device has to be kept runtime active. But a deadlock scenario
> +	 * can arise if the attempt is made to resume the device, when
> +	 * either a suspend or a resume operation is already happening
> +	 * concurrently from some other path and that only actually
> +	 * triggered the compaction. So only unbind if the device is
> +	 * currently runtime active.
> +	 */
> +	if (!intel_runtime_pm_get_if_in_use(dev_priv))
> +		return -EBUSY;
> +
> +	if (!unsafe_drop_pages(obj))
> +		ret = -EBUSY;

Reversed!

> +
> +	intel_runtime_pm_put(dev_priv);
> +	return ret;
> +}
> +
>  /**
>   * i915_gem_shrink - Shrink buffer object caches
>   * @dev_priv: i915 device
> @@ -156,7 +222,6 @@ i915_gem_shrink(struct drm_i915_private *dev_priv,
>  		INIT_LIST_HEAD(&still_in_list);
>  		while (count < target && !list_empty(phase->list)) {
>  			struct drm_i915_gem_object *obj;
> -			struct i915_vma *vma, *v;
>  
>  			obj = list_first_entry(phase->list,
>  					       typeof(*obj), global_list);
> @@ -172,18 +237,8 @@ i915_gem_shrink(struct drm_i915_private *dev_priv,
>  			if (!can_release_pages(obj))
>  				continue;
>  
> -			drm_gem_object_reference(&obj->base);
> -
> -			/* For the unbound phase, this should be a no-op! */
> -			list_for_each_entry_safe(vma, v,
> -						 &obj->vma_list, obj_link)
> -				if (i915_vma_unbind(vma))
> -					break;
> -
> -			if (i915_gem_object_put_pages(obj) == 0)
> +			if (unsafe_drop_pages(obj) == 0)
>  				count += obj->base.size >> PAGE_SHIFT;

But correct here :)
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

[PATCH v3 2/2] drm/i915: Make pages of GFX allocations movable

[akash.goel]

From: Chris Wilson <chris@chris-wilson.co.uk>

On a long run of more than 2-3 days, physical memory tends to get fragmented
severely, which considerably slows down the system. In such a scenario,
Shrinker is also unable to help as lack of memory is not the actual problem,
since it has been observed that there are enough free pages of 0 order.

To address the issue of external fragementation, kernel does a compaction
(which involves migration of pages) but it's efficacy depends upon how many
pages are marked as MOVABLE, as only those pages can be migrated.

Currently the backing pages for GFX buffers are allocated from shmemfs
with GFP_RECLAIMABLE flag, in units of 4KB pages.
In the case of limited Swap space, it may not be possible always to reclaim
or swap-out pages of all the inactive objects, to make way for free space
allowing formation of higher order groups of physically-contiguous pages
on compaction.

Just marking the GFX pages as MOVABLE will not suffice, as i915 Driver
has to pin the pages if they are in use by GPU, which will prevent their
migration. So the migratepage callback in shmem is also hooked up to get a
notification when kernel initiates the page migration. On the notification,
i915 Driver appropriately unpin the pages.
With this Driver can effectively mark the GFX pages as MOVABLE and hence
mitigate the fragmentation problem.

v2:
 - Rename the migration routine to gem_shrink_migratepage, move it to the
   shrinker file, and use the existing constructs (Chris)
 - To cleanup, add a new helper function to encapsulate all page migration
   skip conditions (Chris)
 - Add a new local helper function in shrinker file, for dropping the
   backing pages, and call the same from gem_shrink() also (Chris)

v3:
 - Fix/invert the check on the return value of unsafe_drop_pages (Chris)

Cc: Hugh Dickins <hughd@google.com>
Cc: linux-mm@kvack.org
Signed-off-by: Sourab Gupta <sourab.gupta@intel.com>
Signed-off-by: Akash Goel <akash.goel@intel.com>
---
 drivers/gpu/drm/i915/i915_drv.h          |   3 +
 drivers/gpu/drm/i915/i915_gem.c          |  16 ++-
 drivers/gpu/drm/i915/i915_gem_shrinker.c | 173 ++++++++++++++++++++++++++++---
 3 files changed, 177 insertions(+), 15 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index dd18772..83415f0 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -52,6 +52,7 @@
 #include <linux/intel-iommu.h>
 #include <linux/kref.h>
 #include <linux/pm_qos.h>
+#include <linux/shmem_fs.h>
 #include "intel_guc.h"
 #include "intel_dpll_mgr.h"
 
@@ -1979,6 +1980,8 @@ struct drm_i915_private {
 
 	struct intel_encoder *dig_port_map[I915_MAX_PORTS];
 
+	struct shmem_dev_info  migrate_info;
+
 	/*
 	 * NOTE: This is the dri1/ums dungeon, don't add stuff here. Your patch
 	 * will be rejected. Instead look for a better place.
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index ca96fc1..88b717c 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -2206,6 +2206,7 @@ i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
 		if (obj->madv == I915_MADV_WILLNEED)
 			mark_page_accessed(page);
 
+		set_page_private(page, 0);
 		page_cache_release(page);
 	}
 	obj->dirty = 0;
@@ -2320,6 +2321,7 @@ i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
 			sg->length += PAGE_SIZE;
 		}
 		last_pfn = page_to_pfn(page);
+		set_page_private(page, (unsigned long)obj);
 
 		/* Check that the i965g/gm workaround works. */
 		WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL));
@@ -2345,8 +2347,11 @@ i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj)
 
 err_pages:
 	sg_mark_end(sg);
-	for_each_sg_page(st->sgl, &sg_iter, st->nents, 0)
-		page_cache_release(sg_page_iter_page(&sg_iter));
+	for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
+		page = sg_page_iter_page(&sg_iter);
+		set_page_private(page, 0);
+		page_cache_release(page);
+	}
 	sg_free_table(st);
 	kfree(st);
 
@@ -4468,6 +4473,7 @@ static const struct drm_i915_gem_object_ops i915_gem_object_ops = {
 struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
 						  size_t size)
 {
+	struct drm_i915_private *dev_priv = dev->dev_private;
 	struct drm_i915_gem_object *obj;
 	struct address_space *mapping;
 	gfp_t mask;
@@ -4481,7 +4487,7 @@ struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
 		return NULL;
 	}
 
-	mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
+	mask = GFP_HIGHUSER_MOVABLE;
 	if (IS_CRESTLINE(dev) || IS_BROADWATER(dev)) {
 		/* 965gm cannot relocate objects above 4GiB. */
 		mask &= ~__GFP_HIGHMEM;
@@ -4491,6 +4497,10 @@ struct drm_i915_gem_object *i915_gem_alloc_object(struct drm_device *dev,
 	mapping = file_inode(obj->base.filp)->i_mapping;
 	mapping_set_gfp_mask(mapping, mask);
 
+#ifdef CONFIG_MIGRATION
+	shmem_set_device_ops(mapping, &dev_priv->migrate_info);
+#endif
+
 	i915_gem_object_init(obj, &i915_gem_object_ops);
 
 	obj->base.write_domain = I915_GEM_DOMAIN_CPU;
diff --git a/drivers/gpu/drm/i915/i915_gem_shrinker.c b/drivers/gpu/drm/i915/i915_gem_shrinker.c
index d3c473f..220481e 100644
--- a/drivers/gpu/drm/i915/i915_gem_shrinker.c
+++ b/drivers/gpu/drm/i915/i915_gem_shrinker.c
@@ -24,6 +24,7 @@
 
 #include <linux/oom.h>
 #include <linux/shmem_fs.h>
+#include <linux/migrate.h>
 #include <linux/slab.h>
 #include <linux/swap.h>
 #include <linux/pci.h>
@@ -87,6 +88,71 @@ static bool can_release_pages(struct drm_i915_gem_object *obj)
 	return swap_available() || obj->madv == I915_MADV_DONTNEED;
 }
 
+static bool can_migrate_page(struct drm_i915_gem_object *obj)
+{
+	/* Avoid the migration of page if being actively used by GPU */
+	if (obj->active)
+		return false;
+
+	/* Skip the migration for purgeable objects otherwise there
+	 * will be a deadlock when shmem will try to lock the page for
+	 * truncation, which is already locked by the caller before
+	 * migration.
+	 */
+	if (obj->madv == I915_MADV_DONTNEED)
+		return false;
+
+	/* Skip the migration for a pinned object */
+	if (obj->pages_pin_count != num_vma_bound(obj))
+		return false;
+
+	return true;
+}
+
+static int
+unsafe_drop_pages(struct drm_i915_gem_object *obj)
+{
+	struct i915_vma *vma, *next;
+	int ret;
+
+	drm_gem_object_reference(&obj->base);
+	list_for_each_entry_safe(vma, next, &obj->vma_list, obj_link)
+		if (i915_vma_unbind(vma))
+			break;
+
+	ret = i915_gem_object_put_pages(obj);
+	drm_gem_object_unreference(&obj->base);
+
+	return ret;
+}
+
+static int
+do_migrate_page(struct drm_i915_gem_object *obj)
+{
+	struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
+	int ret = 0;
+
+	if (!can_migrate_page(obj))
+		return -EBUSY;
+
+	/* HW access would be required for a bound object for which
+	 * device has to be kept runtime active. But a deadlock scenario
+	 * can arise if the attempt is made to resume the device, when
+	 * either a suspend or a resume operation is already happening
+	 * concurrently from some other path and that only actually
+	 * triggered the compaction. So only unbind if the device is
+	 * currently runtime active.
+	 */
+	if (!intel_runtime_pm_get_if_in_use(dev_priv))
+		return -EBUSY;
+
+	if (unsafe_drop_pages(obj))
+		ret = -EBUSY;
+
+	intel_runtime_pm_put(dev_priv);
+	return ret;
+}
+
 /**
  * i915_gem_shrink - Shrink buffer object caches
  * @dev_priv: i915 device
@@ -156,7 +222,6 @@ i915_gem_shrink(struct drm_i915_private *dev_priv,
 		INIT_LIST_HEAD(&still_in_list);
 		while (count < target && !list_empty(phase->list)) {
 			struct drm_i915_gem_object *obj;
-			struct i915_vma *vma, *v;
 
 			obj = list_first_entry(phase->list,
 					       typeof(*obj), global_list);
@@ -172,18 +237,8 @@ i915_gem_shrink(struct drm_i915_private *dev_priv,
 			if (!can_release_pages(obj))
 				continue;
 
-			drm_gem_object_reference(&obj->base);
-
-			/* For the unbound phase, this should be a no-op! */
-			list_for_each_entry_safe(vma, v,
-						 &obj->vma_list, obj_link)
-				if (i915_vma_unbind(vma))
-					break;
-
-			if (i915_gem_object_put_pages(obj) == 0)
+			if (unsafe_drop_pages(obj) == 0)
 				count += obj->base.size >> PAGE_SHIFT;
-
-			drm_gem_object_unreference(&obj->base);
 		}
 		list_splice(&still_in_list, phase->list);
 	}
@@ -356,6 +411,95 @@ i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr)
 	return NOTIFY_DONE;
 }
 
+#ifdef CONFIG_MIGRATION
+static int i915_gem_shrink_migratepage(struct address_space *mapping,
+			    struct page *newpage, struct page *page,
+			    enum migrate_mode mode, void *dev_priv_data)
+{
+	struct drm_i915_private *dev_priv = dev_priv_data;
+	struct drm_device *dev = dev_priv->dev;
+	struct drm_i915_gem_object *obj;
+	unsigned long timeout = msecs_to_jiffies(10) + 1;
+	bool unlock;
+	int ret = 0;
+
+	WARN((page_count(newpage) != 1), "Unexpected ref count for newpage\n");
+
+	/*
+	 * Clear the private field of the new target page as it could have a
+	 * stale value in the private field. Otherwise later on if this page
+	 * itself gets migrated, without getting referred by the Driver
+	 * in between, the stale value would cause the i915_migratepage
+	 * function to go for a toss as object pointer is derived from it.
+	 * This should be safe since at the time of migration, private field
+	 * of the new page (which is actually an independent free 4KB page now)
+	 * should be like a don't care for the kernel.
+	 */
+	set_page_private(newpage, 0);
+
+	if (!page_private(page))
+		goto migrate;
+
+	/*
+	 * Check the page count, if Driver also has a reference then it should
+	 * be more than 2, as shmem will have one reference and one reference
+	 * would have been taken by the migration path itself. So if reference
+	 * is <=2, we can directly invoke the migration function.
+	 */
+	if (page_count(page) <= 2)
+		goto migrate;
+
+	/*
+	 * Use trylock here, with a timeout, for struct_mutex as
+	 * otherwise there is a possibility of deadlock due to lock
+	 * inversion. This path, which tries to migrate a particular
+	 * page after locking that page, can race with a path which
+	 * truncate/purge pages of the corresponding object (after
+	 * acquiring struct_mutex). Since page truncation will also
+	 * try to lock the page, a scenario of deadlock can arise.
+	 */
+	while (!i915_gem_shrinker_lock(dev, &unlock) && --timeout)
+		schedule_timeout_killable(1);
+	if (timeout == 0) {
+		DRM_DEBUG_DRIVER("Unable to acquire device mutex.\n");
+		return -EBUSY;
+	}
+
+	obj = (struct drm_i915_gem_object *)page_private(page);
+
+	if (!PageSwapCache(page) && obj) {
+		ret = do_migrate_page(obj);
+		BUG_ON(!ret && page_private(page));
+	}
+
+	if (unlock)
+		mutex_unlock(&dev->struct_mutex);
+
+	if (ret)
+		return ret;
+
+	/*
+	 * Ideally here we don't expect the page count to be > 2, as driver
+	 * would have dropped its reference, but occasionally it has been seen
+	 * coming as 3 & 4. This leads to a situation of unexpected page count,
+	 * causing migration failure, with -EGAIN error. This then leads to
+	 * multiple attempts by the kernel to migrate the same set of pages.
+	 * And sometimes the repeated attempts proves detrimental for stability.
+	 * Also since we don't know who is the other owner, and for how long its
+	 * gonna keep the reference, its better to return -EBUSY.
+	 */
+	if (page_count(page) > 2)
+		return -EBUSY;
+
+migrate:
+	ret = migrate_page(mapping, newpage, page, mode);
+	if (ret)
+		DRM_DEBUG_DRIVER("page=%p migration returned %d\n", page, ret);
+
+	return ret;
+}
+#endif
+
 /**
  * i915_gem_shrinker_init - Initialize i915 shrinker
  * @dev_priv: i915 device
@@ -371,6 +515,11 @@ void i915_gem_shrinker_init(struct drm_i915_private *dev_priv)
 
 	dev_priv->mm.oom_notifier.notifier_call = i915_gem_shrinker_oom;
 	WARN_ON(register_oom_notifier(&dev_priv->mm.oom_notifier));
+
+#ifdef CONFIG_MIGRATION
+	dev_priv->migrate_info.dev_private_data = dev_priv;
+	dev_priv->migrate_info.dev_migratepage = i915_gem_shrink_migratepage;
+#endif
 }
 
 /**
-- 
1.9.2

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [Intel-gfx] [PATCH 1/2] shmem: Support for registration of Driver/file owner specific ops

[Joonas Lahtinen]

On ke, 2016-03-23 at 11:39 +0530, akash.goel@intel.com wrote:
> From: Chris Wilson <chris@chris-wilson.co.uk>
> 
> This provides support for the Drivers or shmem file owners to register
> a set of callbacks, which can be invoked from the address space operations
> methods implemented by shmem.
> This allow the file owners to hook into the shmem address space operations
> to do some extra/custom operations in addition to the default ones.
> 
> The private_data field of address_space struct is used to store the pointer
> to driver specific ops.
> Currently only one ops field is defined, which is migratepage, but can be
> extended on need basis.
> 
> The need for driver specific operations arises since some of the operations
> (like migratepage) may not be handled completely within shmem, so as to be
> effective, and would need some driver specific handling also.
> 
> Specifically, i915.ko would like to participate in migratepage().
> i915.ko uses shmemfs to provide swappable backing storage for its user
> objects, but when those objects are in use by the GPU it must pin the entire
> object until the GPU is idle. As a result, large chunks of memory can be
> arbitrarily withdrawn from page migration, resulting in premature
> out-of-memory due to fragmentation. However, if i915.ko can receive the
> migratepage() request, it can then flush the object from the GPU, remove
> its pin and thus enable the migration.
> 
> Since Gfx allocations are one of the major consumer of system memory, its
> imperative to have such a mechanism to effectively deal with fragmentation.
> And therefore the need for such a provision for initiating driver specific
> actions during address space operations.
> 
> Cc: Hugh Dickins <hughd@google.com>
> Cc: linux-mm@kvack.org
> Signed-off-by: Sourab Gupta <sourab.gupta@intel.com>
> Signed-off-by: Akash Goel <akash.goel@intel.com>
> ---
> A include/linux/shmem_fs.h | 17 +++++++++++++++++
> A mm/shmem.cA A A A A A A A A A A A A A A | 17 ++++++++++++++++-
> A 2 files changed, 33 insertions(+), 1 deletion(-)
> 
> diff --git a/include/linux/shmem_fs.h b/include/linux/shmem_fs.h
> index 4d4780c..6cfa76a 100644
> --- a/include/linux/shmem_fs.h
> +++ b/include/linux/shmem_fs.h
> @@ -34,11 +34,28 @@ struct shmem_sb_info {
> A 	struct mempolicy *mpol;A A A A A /* default memory policy for mappings */
> A };
> A 
> +struct shmem_dev_info {
> +	void *dev_private_data;
> +	int (*dev_migratepage)(struct address_space *mapping,
> +			A A A A A A A struct page *newpage, struct page *page,
> +			A A A A A A A enum migrate_mode mode, void *dev_priv_data);

One might want to have a separate shmem_dev_operations struct or
similar.

> +};
> +
> A static inline struct shmem_inode_info *SHMEM_I(struct inode *inode)
> A {
> A 	return container_of(inode, struct shmem_inode_info, vfs_inode);
> A }
> A 
> +static inline int shmem_set_device_ops(struct address_space *mapping,
> +				struct shmem_dev_info *info)
> +{
> +	if (mapping->private_data != NULL)
> +		return -EEXIST;
> +

I did a quick random peek and most set functions are just void and
override existing data. I'd suggest the same.

> +	mapping->private_data = info;

Also, doesn't this kinda steal the mapping->private_data, might that be
unexpected for the user? I notice currently it's not being touched at
all.

> +	return 0;
> +}
> +
> A /*
> A  * Functions in mm/shmem.c called directly from elsewhere:
> A  */
> diff --git a/mm/shmem.c b/mm/shmem.c
> index 440e2a7..f8625c4 100644
> --- a/mm/shmem.c
> +++ b/mm/shmem.c
> @@ -952,6 +952,21 @@ redirty:
> A 	return 0;
> A }
> A 
> +#ifdef CONFIG_MIGRATION
> +static int shmem_migratepage(struct address_space *mapping,
> +			A A A A A struct page *newpage, struct page *page,
> +			A A A A A enum migrate_mode mode)
> +{
> +	struct shmem_dev_info *dev_info = mapping->private_data;
> +
> +	if (dev_info && dev_info->dev_migratepage)
> +		return dev_info->dev_migratepage(mapping, newpage, page,
> +				mode, dev_info->dev_private_data);
> +
> +	return migrate_page(mapping, newpage, page, mode);
> +}
> +#endif
> +
> A #ifdef CONFIG_NUMA
> A #ifdef CONFIG_TMPFS
> A static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
> @@ -3168,7 +3183,7 @@ static const struct address_space_operations shmem_aops = {
> A 	.write_end	= shmem_write_end,
> A #endif
> A #ifdef CONFIG_MIGRATION
> -	.migratepage	= migrate_page,
> +	.migratepage	= shmem_migratepage,
> A #endif
> A 	.error_remove_page = generic_error_remove_page,
> A };
-- 
Joonas Lahtinen
Open Source Technology Center
Intel Corporation

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [Intel-gfx] [PATCH 1/2] shmem: Support for registration of Driver/file owner specific ops

[akash goel]

On Thu, Mar 24, 2016 at 5:41 PM, Joonas Lahtinen
<joonas.lahtinen@linux.intel.com> wrote:
> On ke, 2016-03-23 at 11:39 +0530, akash.goel@intel.com wrote:
>> From: Chris Wilson <chris@chris-wilson.co.uk>
>>
>> This provides support for the Drivers or shmem file owners to register
>> a set of callbacks, which can be invoked from the address space operations
>> methods implemented by shmem.
>> This allow the file owners to hook into the shmem address space operations
>> to do some extra/custom operations in addition to the default ones.
>>
>> The private_data field of address_space struct is used to store the pointer
>> to driver specific ops.
>> Currently only one ops field is defined, which is migratepage, but can be
>> extended on need basis.
>>
>> The need for driver specific operations arises since some of the operations
>> (like migratepage) may not be handled completely within shmem, so as to be
>> effective, and would need some driver specific handling also.
>>
>> Specifically, i915.ko would like to participate in migratepage().
>> i915.ko uses shmemfs to provide swappable backing storage for its user
>> objects, but when those objects are in use by the GPU it must pin the entire
>> object until the GPU is idle. As a result, large chunks of memory can be
>> arbitrarily withdrawn from page migration, resulting in premature
>> out-of-memory due to fragmentation. However, if i915.ko can receive the
>> migratepage() request, it can then flush the object from the GPU, remove
>> its pin and thus enable the migration.
>>
>> Since Gfx allocations are one of the major consumer of system memory, its
>> imperative to have such a mechanism to effectively deal with fragmentation.
>> And therefore the need for such a provision for initiating driver specific
>> actions during address space operations.
>>
>> Cc: Hugh Dickins <hughd@google.com>
>> Cc: linux-mm@kvack.org
>> Signed-off-by: Sourab Gupta <sourab.gupta@intel.com>
>> Signed-off-by: Akash Goel <akash.goel@intel.com>
>> ---
>>  include/linux/shmem_fs.h | 17 +++++++++++++++++
>>  mm/shmem.c               | 17 ++++++++++++++++-
>>  2 files changed, 33 insertions(+), 1 deletion(-)
>>
>> diff --git a/include/linux/shmem_fs.h b/include/linux/shmem_fs.h
>> index 4d4780c..6cfa76a 100644
>> --- a/include/linux/shmem_fs.h
>> +++ b/include/linux/shmem_fs.h
>> @@ -34,11 +34,28 @@ struct shmem_sb_info {
>>       struct mempolicy *mpol;     /* default memory policy for mappings */
>>  };
>>
>> +struct shmem_dev_info {
>> +     void *dev_private_data;
>> +     int (*dev_migratepage)(struct address_space *mapping,
>> +                            struct page *newpage, struct page *page,
>> +                            enum migrate_mode mode, void *dev_priv_data);
>
> One might want to have a separate shmem_dev_operations struct or
> similar.
>
Sorry for the very late turnaround.

Sorry couldn't get your point here. Are you suggesting to rename the
structure to shmem_dev_operations ?

>> +};
>> +
>>  static inline struct shmem_inode_info *SHMEM_I(struct inode *inode)
>>  {
>>       return container_of(inode, struct shmem_inode_info, vfs_inode);
>>  }
>>
>> +static inline int shmem_set_device_ops(struct address_space *mapping,
>> +                             struct shmem_dev_info *info)
>> +{
>> +     if (mapping->private_data != NULL)
>> +             return -EEXIST;
>> +
>
> I did a quick random peek and most set functions are just void and
> override existing data. I'd suggest the same.
>
>> +     mapping->private_data = info;
>
Fine will change the return type to void and remove the check.

> Also, doesn't this kinda steal the mapping->private_data, might that be
> unexpected for the user? I notice currently it's not being touched at
> all.
>
Sorry by User do you mean the shmem client who called shmem_file_setup() ?
It seems clients are not expected to touch mapping->private_data and
so shmemfs can safely use it.

Best regards
Akash

>> +     return 0;
>> +}
>> +
>>  /*
>>   * Functions in mm/shmem.c called directly from elsewhere:
>>   */
>> diff --git a/mm/shmem.c b/mm/shmem.c
>> index 440e2a7..f8625c4 100644
>> --- a/mm/shmem.c
>> +++ b/mm/shmem.c
>> @@ -952,6 +952,21 @@ redirty:
>>       return 0;
>>  }
>>
>> +#ifdef CONFIG_MIGRATION
>> +static int shmem_migratepage(struct address_space *mapping,
>> +                          struct page *newpage, struct page *page,
>> +                          enum migrate_mode mode)
>> +{
>> +     struct shmem_dev_info *dev_info = mapping->private_data;
>> +
>> +     if (dev_info && dev_info->dev_migratepage)
>> +             return dev_info->dev_migratepage(mapping, newpage, page,
>> +                             mode, dev_info->dev_private_data);
>> +
>> +     return migrate_page(mapping, newpage, page, mode);
>> +}
>> +#endif
>> +
>>  #ifdef CONFIG_NUMA
>>  #ifdef CONFIG_TMPFS
>>  static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
>> @@ -3168,7 +3183,7 @@ static const struct address_space_operations shmem_aops = {
>>       .write_end      = shmem_write_end,
>>  #endif
>>  #ifdef CONFIG_MIGRATION
>> -     .migratepage    = migrate_page,
>> +     .migratepage    = shmem_migratepage,
>>  #endif
>>       .error_remove_page = generic_error_remove_page,
>>  };
> --
> Joonas Lahtinen
> Open Source Technology Center
> Intel Corporation
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/intel-gfx

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [Intel-gfx] [PATCH 1/2] shmem: Support for registration of Driver/file owner specific ops

[Joonas Lahtinen]

On ke, 2016-10-19 at 20:41 +0530, akash goel wrote:
> On Thu, Mar 24, 2016 at 5:41 PM, Joonas Lahtinen
> > <joonas.lahtinen@linux.intel.com> wrote:
> > On ke, 2016-03-23 at 11:39 +0530, akash.goel@intel.com wrote:
> > > @@ -34,11 +34,28 @@ struct shmem_sb_info {
> > > A A A A A A struct mempolicy *mpol;A A A A A /* default memory policy for mappings */
> > > A };
> > > 
> > > +struct shmem_dev_info {
> > > +A A A A A void *dev_private_data;
> > > +A A A A A int (*dev_migratepage)(struct address_space *mapping,
> > > +A A A A A A A A A A A A A A A A A A A A A A A A A A A A struct page *newpage, struct page *page,
> > > +A A A A A A A A A A A A A A A A A A A A A A A A A A A A enum migrate_mode mode, void *dev_priv_data);
> > 
> > One might want to have a separate shmem_dev_operations struct or
> > similar.
> > 
> Sorry for the very late turnaround.
> 
> Sorry couldn't get your point here. Are you suggesting to rename the
> structure to shmem_dev_operations ?

I'm pretty sure I was after putting migratepage function pointer in
shmem_dev_operations struct, but I think that can be done once there
are more functions.

s/dev_private_data/private_data/ and s/dev_priv_data/private_data/
might be in order, too. I should be obvious from context.

> > > +};
> > > +
> > > A static inline struct shmem_inode_info *SHMEM_I(struct inode *inode)
> > > A {
> > > A A A A A A return container_of(inode, struct shmem_inode_info, vfs_inode);
> > > A }
> > > 
> > > +static inline int shmem_set_device_ops(struct address_space *mapping,
> > > +A A A A A A A A A A A A A A A A A A A A A A A A A A A A A struct shmem_dev_info *info)
> > > +{

This name could be shmem_set_dev_info, if there will be separate _ops
struct in future.

> > > +A A A A A if (mapping->private_data != NULL)
> > > +A A A A A A A A A A A A A return -EEXIST;
> > > +
> > 
> > I did a quick random peek and most set functions are just void and
> > override existing data. I'd suggest the same.
> > 
> > > 
> > > +A A A A A mapping->private_data = info;
> > 
> Fine will change the return type to void and remove the check.
> 
> > 
> > Also, doesn't this kinda steal the mapping->private_data, might that be
> > unexpected for the user? I notice currently it's not being touched at
> > all.
> > 
> Sorry by User do you mean the shmem client who called shmem_file_setup() ?
> It seems clients are not expected to touch mapping->private_data and
> so shmemfs can safely use it.

If it's not used by others, should be fine. Not sure if WARN would be
in place, Chris?

Regards, Joonas
-- 
Joonas Lahtinen
Open Source Technology Center
Intel Corporation

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

[PATCH 1/2] shmem: Support for registration of driver/file owner specific ops

[akash.goel]

From: Chris Wilson <chris@chris-wilson.co.uk>

This provides support for the drivers or shmem file owners to register
a set of callbacks, which can be invoked from the address space
operations methods implemented by shmem.  This allow the file owners to
hook into the shmem address space operations to do some extra/custom
operations in addition to the default ones.

The private_data field of address_space struct is used to store the
pointer to driver specific ops.  Currently only one ops field is defined,
which is migratepage, but can be extended on an as-needed basis.

The need for driver specific operations arises since some of the
operations (like migratepage) may not be handled completely within shmem,
so as to be effective, and would need some driver specific handling also.
Specifically, i915.ko would like to participate in migratepage().
i915.ko uses shmemfs to provide swappable backing storage for its user
objects, but when those objects are in use by the GPU it must pin the
entire object until the GPU is idle.  As a result, large chunks of memory
can be arbitrarily withdrawn from page migration, resulting in premature
out-of-memory due to fragmentation.  However, if i915.ko can receive the
migratepage() request, it can then flush the object from the GPU, remove
its pin and thus enable the migration.

Since gfx allocations are one of the major consumer of system memory, its
imperative to have such a mechanism to effectively deal with
fragmentation.  And therefore the need for such a provision for initiating
driver specific actions during address space operations.

v2:
- Drop dev_ prefix from the members of shmem_dev_info structure. (Joonas)
- Change the return type of shmem_set_device_op() to void and remove the
  check for pre-existing data. (Joonas)
- Rename shmem_set_device_op() to shmem_set_dev_info() to be consistent
  with shmem_dev_info structure. (Joonas)

Cc: Hugh Dickins <hughd@google.com>
Cc: linux-mm@kvack.org
Cc: linux-kernel@vger.linux.org
Signed-off-by: Sourab Gupta <sourab.gupta@intel.com>
Signed-off-by: Akash Goel <akash.goel@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
---
 include/linux/shmem_fs.h | 13 +++++++++++++
 mm/shmem.c               | 17 ++++++++++++++++-
 2 files changed, 29 insertions(+), 1 deletion(-)

diff --git a/include/linux/shmem_fs.h b/include/linux/shmem_fs.h
index ff078e7..22796a0 100644
--- a/include/linux/shmem_fs.h
+++ b/include/linux/shmem_fs.h
@@ -39,11 +39,24 @@ struct shmem_sb_info {
 	unsigned long shrinklist_len; /* Length of shrinklist */
 };
 
+struct shmem_dev_info {
+	void *private_data;
+	int (*migratepage)(struct address_space *mapping,
+			   struct page *newpage, struct page *page,
+			   enum migrate_mode mode, void *dev_priv_data);
+};
+
 static inline struct shmem_inode_info *SHMEM_I(struct inode *inode)
 {
 	return container_of(inode, struct shmem_inode_info, vfs_inode);
 }
 
+static inline void shmem_set_dev_info(struct address_space *mapping,
+					 struct shmem_dev_info *info)
+{
+	mapping->private_data = info;
+}
+
 /*
  * Functions in mm/shmem.c called directly from elsewhere:
  */
diff --git a/mm/shmem.c b/mm/shmem.c
index ad7813d..bf71ddd 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -1290,6 +1290,21 @@ static int shmem_writepage(struct page *page, struct writeback_control *wbc)
 	return 0;
 }
 
+#ifdef CONFIG_MIGRATION
+static int shmem_migratepage(struct address_space *mapping,
+			     struct page *newpage, struct page *page,
+			     enum migrate_mode mode)
+{
+	struct shmem_dev_info *dev_info = mapping->private_data;
+
+	if (dev_info && dev_info->migratepage)
+		return dev_info->migratepage(mapping, newpage, page,
+				mode, dev_info->private_data);
+
+	return migrate_page(mapping, newpage, page, mode);
+}
+#endif
+
 #if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
 {
@@ -3654,7 +3669,7 @@ static void shmem_destroy_inodecache(void)
 	.write_end	= shmem_write_end,
 #endif
 #ifdef CONFIG_MIGRATION
-	.migratepage	= migrate_page,
+	.migratepage	= shmem_migratepage,
 #endif
 	.error_remove_page = generic_error_remove_page,
 };
-- 
1.9.2

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

[PATCH 2/2] drm/i915: Make GPU pages movable

[akash.goel]

From: Chris Wilson <chris@chris-wilson.co.uk>

On a long run of more than 2-3 days, physical memory tends to get
fragmented severely, which considerably slows down the system. In such a
scenario, the shrinker is also unable to help as lack of memory is not
the actual problem, since it has been observed that there are enough free
pages of 0 order. This also manifests itself when an indiviual zone in
the mm runs out of pages and if we cannot migrate pages between zones,
the kernel hits an out-of-memory even though there are free pages (and
often all of swap) available.

To address the issue of external fragementation, kernel does a compaction
(which involves migration of pages) but it's efficacy depends upon how
many pages are marked as MOVABLE, as only those pages can be migrated.

Currently the backing pages for GPU buffers are allocated from shmemfs
with GFP_RECLAIMABLE flag, in units of 4KB pages.  In the case of limited
swap space, it may not be possible always to reclaim or swap-out pages of
all the inactive objects, to make way for free space allowing formation
of higher order groups of physically-contiguous pages on compaction.

Just marking the GPU pages as MOVABLE will not suffice, as i915.ko has to
pin the pages if they are in use by GPU, which will prevent their
migration. So the migratepage callback in shmem is also hooked up to get
a notification when kernel initiates the page migration. On the
notification, i915.ko appropriately unpin the pages.  With this we can
effectively mark the GPU pages as MOVABLE and hence mitigate the
fragmentation problem.

v2:
 - Rename the migration routine to gem_shrink_migratepage, move it to the
   shrinker file, and use the existing constructs (Chris)
 - To cleanup, add a new helper function to encapsulate all page migration
   skip conditions (Chris)
 - Add a new local helper function in shrinker file, for dropping the
   backing pages, and call the same from gem_shrink() also (Chris)

v3:
 - Fix/invert the check on the return value of unsafe_drop_pages (Chris)

v4:
 - Minor tidy

v5:
 - Fix unsafe usage of unsafe_drop_pages()
 - Rebase onto vmap-notifier

Testcase: igt/gem_shrink
Bugzilla: (e.g.) https://bugs.freedesktop.org/show_bug.cgi?id=90254
Cc: Hugh Dickins <hughd@google.com>
Cc: linux-mm@kvack.org
Signed-off-by: Sourab Gupta <sourab.gupta@intel.com>
Signed-off-by: Akash Goel <akash.goel@intel.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
---
 drivers/gpu/drm/i915/i915_drv.h          |   2 +
 drivers/gpu/drm/i915/i915_gem.c          |   9 ++-
 drivers/gpu/drm/i915/i915_gem_shrinker.c | 134 +++++++++++++++++++++++++++++++
 3 files changed, 144 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index 4735b417..7f2717b 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -1357,6 +1357,8 @@ struct intel_l3_parity {
 };
 
 struct i915_gem_mm {
+	struct shmem_dev_info shmem_info;
+
 	/** Memory allocator for GTT stolen memory */
 	struct drm_mm stolen;
 	/** Protects the usage of the GTT stolen memory allocator. This is
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index 1f995ce..f0d4ce7 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -2164,6 +2164,7 @@ void __i915_gem_object_invalidate(struct drm_i915_gem_object *obj)
 		if (obj->mm.madv == I915_MADV_WILLNEED)
 			mark_page_accessed(page);
 
+		set_page_private(page, 0);
 		put_page(page);
 	}
 	obj->mm.dirty = false;
@@ -2310,6 +2311,7 @@ static unsigned int swiotlb_max_size(void)
 			sg->length += PAGE_SIZE;
 		}
 		last_pfn = page_to_pfn(page);
+		set_page_private(page, (unsigned long)obj);
 
 		/* Check that the i965g/gm workaround works. */
 		WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL));
@@ -2334,8 +2336,10 @@ static unsigned int swiotlb_max_size(void)
 
 err_pages:
 	sg_mark_end(sg);
-	for_each_sgt_page(page, sgt_iter, st)
+	for_each_sgt_page(page, sgt_iter, st) {
+		set_page_private(page, 0);
 		put_page(page);
+	}
 	sg_free_table(st);
 	kfree(st);
 
@@ -4185,6 +4189,8 @@ struct drm_i915_gem_object *
 		goto fail;
 
 	mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
+	if (IS_ENABLED(MIGRATION))
+		mask |= __GFP_MOVABLE;
 	if (IS_CRESTLINE(dev_priv) || IS_BROADWATER(dev_priv)) {
 		/* 965gm cannot relocate objects above 4GiB. */
 		mask &= ~__GFP_HIGHMEM;
@@ -4193,6 +4199,7 @@ struct drm_i915_gem_object *
 
 	mapping = obj->base.filp->f_mapping;
 	mapping_set_gfp_mask(mapping, mask);
+	shmem_set_dev_info(mapping, &dev_priv->mm.shmem_info);
 
 	i915_gem_object_init(obj, &i915_gem_object_ops);
 
diff --git a/drivers/gpu/drm/i915/i915_gem_shrinker.c b/drivers/gpu/drm/i915/i915_gem_shrinker.c
index a6fc1bd..2862523 100644
--- a/drivers/gpu/drm/i915/i915_gem_shrinker.c
+++ b/drivers/gpu/drm/i915/i915_gem_shrinker.c
@@ -24,6 +24,7 @@
 
 #include <linux/oom.h>
 #include <linux/shmem_fs.h>
+#include <linux/migrate.h>
 #include <linux/slab.h>
 #include <linux/swap.h>
 #include <linux/pci.h>
@@ -473,6 +474,134 @@ struct shrinker_lock_uninterruptible {
 	return NOTIFY_DONE;
 }
 
+#ifdef CONFIG_MIGRATION
+static bool can_migrate_page(struct drm_i915_gem_object *obj)
+{
+	/* Avoid the migration of page if being actively used by GPU */
+	if (i915_gem_object_is_active(obj))
+		return false;
+
+	/* Skip the migration for purgeable objects otherwise there
+	 * will be a deadlock when shmem will try to lock the page for
+	 * truncation, which is already locked by the caller before
+	 * migration.
+	 */
+	if (obj->mm.madv == I915_MADV_DONTNEED)
+		return false;
+
+	/* Skip the migration for a pinned object */
+	if (atomic_read(&obj->mm.pages_pin_count) > obj->bind_count)
+		return false;
+
+	if (any_vma_pinned(obj))
+		return false;
+
+	return true;
+}
+
+static int do_migrate_page(struct drm_i915_gem_object *obj)
+{
+	struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
+	int ret = 0;
+
+	if (!can_migrate_page(obj))
+		return -EBUSY;
+
+	/* HW access would be required for a GGTT bound object, for which
+	 * device has to be kept awake. But a deadlock scenario can arise if
+	 * the attempt is made to resume the device, when either a suspend
+	 * or a resume operation is already happening concurrently from some
+	 * other path and that only also triggers compaction. So only unbind
+	 * if the device is currently awake.
+	 */
+	if (!intel_runtime_pm_get_if_in_use(dev_priv))
+		return -EBUSY;
+
+	i915_gem_object_get(obj);
+	if (!unsafe_drop_pages(obj))
+		ret = -EBUSY;
+	i915_gem_object_put(obj);
+
+	intel_runtime_pm_put(dev_priv);
+	return ret;
+}
+
+static int i915_gem_shrinker_migratepage(struct address_space *mapping,
+					 struct page *newpage,
+					 struct page *page,
+					 enum migrate_mode mode,
+					 void *dev_priv_data)
+{
+	struct drm_i915_private *dev_priv = dev_priv_data;
+	struct shrinker_lock_uninterruptible slu;
+	int ret;
+
+	/*
+	 * Clear the private field of the new target page as it could have a
+	 * stale value in the private field. Otherwise later on if this page
+	 * itself gets migrated, without getting referred by the Driver
+	 * in between, the stale value would cause the i915_migratepage
+	 * function to go for a toss as object pointer is derived from it.
+	 * This should be safe since at the time of migration, private field
+	 * of the new page (which is actually an independent free 4KB page now)
+	 * should be like a don't care for the kernel.
+	 */
+	set_page_private(newpage, 0);
+
+	if (!page_private(page))
+		goto migrate;
+
+	/*
+	 * Check the page count, if Driver also has a reference then it should
+	 * be more than 2, as shmem will have one reference and one reference
+	 * would have been taken by the migration path itself. So if reference
+	 * is <=2, we can directly invoke the migration function.
+	 */
+	if (page_count(page) <= 2)
+		goto migrate;
+
+	/*
+	 * Use trylock here, with a timeout, for struct_mutex as
+	 * otherwise there is a possibility of deadlock due to lock
+	 * inversion. This path, which tries to migrate a particular
+	 * page after locking that page, can race with a path which
+	 * truncate/purge pages of the corresponding object (after
+	 * acquiring struct_mutex). Since page truncation will also
+	 * try to lock the page, a scenario of deadlock can arise.
+	 */
+	if (!i915_gem_shrinker_lock_uninterruptible(dev_priv, &slu, 10))
+		return -EBUSY;
+
+	ret = 0;
+	if (!PageSwapCache(page) && page_private(page)) {
+		struct drm_i915_gem_object *obj =
+			(struct drm_i915_gem_object *)page_private(page);
+
+		ret = do_migrate_page(obj);
+	}
+
+	i915_gem_shrinker_unlock_uninterruptible(dev_priv, &slu);
+	if (ret)
+		return ret;
+
+	/*
+	 * Ideally here we don't expect the page count to be > 2, as driver
+	 * would have dropped its reference, but occasionally it has been seen
+	 * coming as 3 & 4. This leads to a situation of unexpected page count,
+	 * causing migration failure, with -EGAIN error. This then leads to
+	 * multiple attempts by the kernel to migrate the same set of pages.
+	 * And sometimes the repeated attempts proves detrimental for stability.
+	 * Also since we don't know who is the other owner, and for how long its
+	 * gonna keep the reference, its better to return -EBUSY.
+	 */
+	if (page_count(page) > 2)
+		return -EBUSY;
+
+migrate:
+	return migrate_page(mapping, newpage, page, mode);
+}
+#endif
+
 /**
  * i915_gem_shrinker_init - Initialize i915 shrinker
  * @dev_priv: i915 device
@@ -491,6 +620,11 @@ void i915_gem_shrinker_init(struct drm_i915_private *dev_priv)
 
 	dev_priv->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
 	WARN_ON(register_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
+
+	dev_priv->mm.shmem_info.private_data = dev_priv;
+#ifdef CONFIG_MIGRATION
+	dev_priv->mm.shmem_info.migratepage = i915_gem_shrinker_migratepage;
+#endif
 }
 
 /**
-- 
1.9.2

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH 2/2] drm/i915: Make GPU pages movable

[Chris Wilson]

Best if we send these as a new series to unconfuse CI.

On Fri, Nov 04, 2016 at 06:18:26PM +0530, akash.goel@intel.com wrote:
> +static int do_migrate_page(struct drm_i915_gem_object *obj)
> +{
> +	struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
> +	int ret = 0;
> +
> +	if (!can_migrate_page(obj))
> +		return -EBUSY;
> +
> +	/* HW access would be required for a GGTT bound object, for which
> +	 * device has to be kept awake. But a deadlock scenario can arise if
> +	 * the attempt is made to resume the device, when either a suspend
> +	 * or a resume operation is already happening concurrently from some
> +	 * other path and that only also triggers compaction. So only unbind
> +	 * if the device is currently awake.
> +	 */
> +	if (!intel_runtime_pm_get_if_in_use(dev_priv))
> +		return -EBUSY;
> +
> +	i915_gem_object_get(obj);
> +	if (!unsafe_drop_pages(obj))
> +		ret = -EBUSY;
> +	i915_gem_object_put(obj);

Since the object release changes, we can now do this without the
i915_gem_object_get / i915_gem_object_put (as we are guarded by the BKL
struct_mutex).
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH 2/2] drm/i915: Make GPU pages movable

[Goel, Akash]

On 11/4/2016 7:07 PM, Chris Wilson wrote:
> Best if we send these as a new series to unconfuse CI.
>
Okay will send as a new series.

> On Fri, Nov 04, 2016 at 06:18:26PM +0530, akash.goel@intel.com wrote:
>> +static int do_migrate_page(struct drm_i915_gem_object *obj)
>> +{
>> +	struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
>> +	int ret = 0;
>> +
>> +	if (!can_migrate_page(obj))
>> +		return -EBUSY;
>> +
>> +	/* HW access would be required for a GGTT bound object, for which
>> +	 * device has to be kept awake. But a deadlock scenario can arise if
>> +	 * the attempt is made to resume the device, when either a suspend
>> +	 * or a resume operation is already happening concurrently from some
>> +	 * other path and that only also triggers compaction. So only unbind
>> +	 * if the device is currently awake.
>> +	 */
>> +	if (!intel_runtime_pm_get_if_in_use(dev_priv))
>> +		return -EBUSY;
>> +
>> +	i915_gem_object_get(obj);
>> +	if (!unsafe_drop_pages(obj))
>> +		ret = -EBUSY;
>> +	i915_gem_object_put(obj);
>
> Since the object release changes, we can now do this without the
> i915_gem_object_get / i915_gem_object_put (as we are guarded by the BKL
> struct_mutex).
Fine will remove object_get/put as with struct_mutex protection object 
can't disappear across unsafe_drop_pages().

Best regards
Akash


> -Chris
>

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

[PATCH 1/2] shmem: Support for registration of driver/file owner specific ops

[akash.goel]

From: Chris Wilson <chris@chris-wilson.co.uk>

This provides support for the drivers or shmem file owners to register
a set of callbacks, which can be invoked from the address space
operations methods implemented by shmem.  This allow the file owners to
hook into the shmem address space operations to do some extra/custom
operations in addition to the default ones.

The private_data field of address_space struct is used to store the
pointer to driver specific ops.  Currently only one ops field is defined,
which is migratepage, but can be extended on an as-needed basis.

The need for driver specific operations arises since some of the
operations (like migratepage) may not be handled completely within shmem,
so as to be effective, and would need some driver specific handling also.
Specifically, i915.ko would like to participate in migratepage().
i915.ko uses shmemfs to provide swappable backing storage for its user
objects, but when those objects are in use by the GPU it must pin the
entire object until the GPU is idle.  As a result, large chunks of memory
can be arbitrarily withdrawn from page migration, resulting in premature
out-of-memory due to fragmentation.  However, if i915.ko can receive the
migratepage() request, it can then flush the object from the GPU, remove
its pin and thus enable the migration.

Since gfx allocations are one of the major consumer of system memory, its
imperative to have such a mechanism to effectively deal with
fragmentation.  And therefore the need for such a provision for initiating
driver specific actions during address space operations.

v2:
- Drop dev_ prefix from the members of shmem_dev_info structure. (Joonas)
- Change the return type of shmem_set_device_op() to void and remove the
  check for pre-existing data. (Joonas)
- Rename shmem_set_device_op() to shmem_set_dev_info() to be consistent
  with shmem_dev_info structure. (Joonas)

Cc: Hugh Dickins <hughd@google.com>
Cc: linux-mm@kvack.org
Cc: linux-kernel@vger.linux.org
Signed-off-by: Sourab Gupta <sourab.gupta@intel.com>
Signed-off-by: Akash Goel <akash.goel@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
---
 include/linux/shmem_fs.h | 13 +++++++++++++
 mm/shmem.c               | 17 ++++++++++++++++-
 2 files changed, 29 insertions(+), 1 deletion(-)

diff --git a/include/linux/shmem_fs.h b/include/linux/shmem_fs.h
index ff078e7..454c3ba 100644
--- a/include/linux/shmem_fs.h
+++ b/include/linux/shmem_fs.h
@@ -39,11 +39,24 @@ struct shmem_sb_info {
 	unsigned long shrinklist_len; /* Length of shrinklist */
 };
 
+struct shmem_dev_info {
+	void *private_data;
+	int (*migratepage)(struct address_space *mapping,
+			   struct page *newpage, struct page *page,
+			   enum migrate_mode mode, void *dev_priv_data);
+};
+
 static inline struct shmem_inode_info *SHMEM_I(struct inode *inode)
 {
 	return container_of(inode, struct shmem_inode_info, vfs_inode);
 }
 
+static inline void shmem_set_dev_info(struct address_space *mapping,
+				      struct shmem_dev_info *info)
+{
+	mapping->private_data = info;
+}
+
 /*
  * Functions in mm/shmem.c called directly from elsewhere:
  */
diff --git a/mm/shmem.c b/mm/shmem.c
index ad7813d..fce8de3 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -1290,6 +1290,21 @@ static int shmem_writepage(struct page *page, struct writeback_control *wbc)
 	return 0;
 }
 
+#ifdef CONFIG_MIGRATION
+static int shmem_migratepage(struct address_space *mapping,
+			     struct page *newpage, struct page *page,
+			     enum migrate_mode mode)
+{
+	struct shmem_dev_info *dev_info = mapping->private_data;
+
+	if (dev_info && dev_info->migratepage)
+		return dev_info->migratepage(mapping, newpage, page,
+					     mode, dev_info->private_data);
+
+	return migrate_page(mapping, newpage, page, mode);
+}
+#endif
+
 #if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
 {
@@ -3654,7 +3669,7 @@ static void shmem_destroy_inodecache(void)
 	.write_end	= shmem_write_end,
 #endif
 #ifdef CONFIG_MIGRATION
-	.migratepage	= migrate_page,
+	.migratepage	= shmem_migratepage,
 #endif
 	.error_remove_page = generic_error_remove_page,
 };
-- 
1.9.2

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

[PATCH 2/2] drm/i915: Make GPU pages movable

[akash.goel]

From: Chris Wilson <chris@chris-wilson.co.uk>

On a long run of more than 2-3 days, physical memory tends to get
fragmented severely, which considerably slows down the system. In such a
scenario, the shrinker is also unable to help as lack of memory is not
the actual problem, since it has been observed that there are enough free
pages of 0 order. This also manifests itself when an indiviual zone in
the mm runs out of pages and if we cannot migrate pages between zones,
the kernel hits an out-of-memory even though there are free pages (and
often all of swap) available.

To address the issue of external fragementation, kernel does a compaction
(which involves migration of pages) but it's efficacy depends upon how
many pages are marked as MOVABLE, as only those pages can be migrated.

Currently the backing pages for GPU buffers are allocated from shmemfs
with GFP_RECLAIMABLE flag, in units of 4KB pages.  In the case of limited
swap space, it may not be possible always to reclaim or swap-out pages of
all the inactive objects, to make way for free space allowing formation
of higher order groups of physically-contiguous pages on compaction.

Just marking the GPU pages as MOVABLE will not suffice, as i915.ko has to
pin the pages if they are in use by GPU, which will prevent their
migration. So the migratepage callback in shmem is also hooked up to get
a notification when kernel initiates the page migration. On the
notification, i915.ko appropriately unpin the pages.  With this we can
effectively mark the GPU pages as MOVABLE and hence mitigate the
fragmentation problem.

v2:
 - Rename the migration routine to gem_shrink_migratepage, move it to the
   shrinker file, and use the existing constructs (Chris)
 - To cleanup, add a new helper function to encapsulate all page migration
   skip conditions (Chris)
 - Add a new local helper function in shrinker file, for dropping the
   backing pages, and call the same from gem_shrink() also (Chris)

v3:
 - Fix/invert the check on the return value of unsafe_drop_pages (Chris)

v4:
 - Minor tidy

v5:
 - Fix unsafe usage of unsafe_drop_pages()
 - Rebase onto vmap-notifier

v6:
- Remove i915_gem_object_get/put across unsafe_drop_pages() as with
  struct_mutex protection object can't disappear. (Chris)

Testcase: igt/gem_shrink
Bugzilla: (e.g.) https://bugs.freedesktop.org/show_bug.cgi?id=90254
Cc: Hugh Dickins <hughd@google.com>
Cc: linux-mm@kvack.org
Signed-off-by: Sourab Gupta <sourab.gupta@intel.com>
Signed-off-by: Akash Goel <akash.goel@intel.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
---
 drivers/gpu/drm/i915/i915_drv.h          |   2 +
 drivers/gpu/drm/i915/i915_gem.c          |   9 ++-
 drivers/gpu/drm/i915/i915_gem_shrinker.c | 132 +++++++++++++++++++++++++++++++
 3 files changed, 142 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index 4735b417..7f2717b 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -1357,6 +1357,8 @@ struct intel_l3_parity {
 };
 
 struct i915_gem_mm {
+	struct shmem_dev_info shmem_info;
+
 	/** Memory allocator for GTT stolen memory */
 	struct drm_mm stolen;
 	/** Protects the usage of the GTT stolen memory allocator. This is
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index 1f995ce..f0d4ce7 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -2164,6 +2164,7 @@ void __i915_gem_object_invalidate(struct drm_i915_gem_object *obj)
 		if (obj->mm.madv == I915_MADV_WILLNEED)
 			mark_page_accessed(page);
 
+		set_page_private(page, 0);
 		put_page(page);
 	}
 	obj->mm.dirty = false;
@@ -2310,6 +2311,7 @@ static unsigned int swiotlb_max_size(void)
 			sg->length += PAGE_SIZE;
 		}
 		last_pfn = page_to_pfn(page);
+		set_page_private(page, (unsigned long)obj);
 
 		/* Check that the i965g/gm workaround works. */
 		WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL));
@@ -2334,8 +2336,10 @@ static unsigned int swiotlb_max_size(void)
 
 err_pages:
 	sg_mark_end(sg);
-	for_each_sgt_page(page, sgt_iter, st)
+	for_each_sgt_page(page, sgt_iter, st) {
+		set_page_private(page, 0);
 		put_page(page);
+	}
 	sg_free_table(st);
 	kfree(st);
 
@@ -4185,6 +4189,8 @@ struct drm_i915_gem_object *
 		goto fail;
 
 	mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
+	if (IS_ENABLED(MIGRATION))
+		mask |= __GFP_MOVABLE;
 	if (IS_CRESTLINE(dev_priv) || IS_BROADWATER(dev_priv)) {
 		/* 965gm cannot relocate objects above 4GiB. */
 		mask &= ~__GFP_HIGHMEM;
@@ -4193,6 +4199,7 @@ struct drm_i915_gem_object *
 
 	mapping = obj->base.filp->f_mapping;
 	mapping_set_gfp_mask(mapping, mask);
+	shmem_set_dev_info(mapping, &dev_priv->mm.shmem_info);
 
 	i915_gem_object_init(obj, &i915_gem_object_ops);
 
diff --git a/drivers/gpu/drm/i915/i915_gem_shrinker.c b/drivers/gpu/drm/i915/i915_gem_shrinker.c
index a6fc1bd..051135ac 100644
--- a/drivers/gpu/drm/i915/i915_gem_shrinker.c
+++ b/drivers/gpu/drm/i915/i915_gem_shrinker.c
@@ -24,6 +24,7 @@
 
 #include <linux/oom.h>
 #include <linux/shmem_fs.h>
+#include <linux/migrate.h>
 #include <linux/slab.h>
 #include <linux/swap.h>
 #include <linux/pci.h>
@@ -473,6 +474,132 @@ struct shrinker_lock_uninterruptible {
 	return NOTIFY_DONE;
 }
 
+#ifdef CONFIG_MIGRATION
+static bool can_migrate_page(struct drm_i915_gem_object *obj)
+{
+	/* Avoid the migration of page if being actively used by GPU */
+	if (i915_gem_object_is_active(obj))
+		return false;
+
+	/* Skip the migration for purgeable objects otherwise there
+	 * will be a deadlock when shmem will try to lock the page for
+	 * truncation, which is already locked by the caller before
+	 * migration.
+	 */
+	if (obj->mm.madv == I915_MADV_DONTNEED)
+		return false;
+
+	/* Skip the migration for a pinned object */
+	if (atomic_read(&obj->mm.pages_pin_count) > obj->bind_count)
+		return false;
+
+	if (any_vma_pinned(obj))
+		return false;
+
+	return true;
+}
+
+static int do_migrate_page(struct drm_i915_gem_object *obj)
+{
+	struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
+	int ret = 0;
+
+	if (!can_migrate_page(obj))
+		return -EBUSY;
+
+	/* HW access would be required for a GGTT bound object, for which
+	 * device has to be kept awake. But a deadlock scenario can arise if
+	 * the attempt is made to resume the device, when either a suspend
+	 * or a resume operation is already happening concurrently from some
+	 * other path and that only also triggers compaction. So only unbind
+	 * if the device is currently awake.
+	 */
+	if (!intel_runtime_pm_get_if_in_use(dev_priv))
+		return -EBUSY;
+
+	if (!unsafe_drop_pages(obj))
+		ret = -EBUSY;
+
+	intel_runtime_pm_put(dev_priv);
+	return ret;
+}
+
+static int i915_gem_shrinker_migratepage(struct address_space *mapping,
+					 struct page *newpage,
+					 struct page *page,
+					 enum migrate_mode mode,
+					 void *dev_priv_data)
+{
+	struct drm_i915_private *dev_priv = dev_priv_data;
+	struct shrinker_lock_uninterruptible slu;
+	int ret;
+
+	/*
+	 * Clear the private field of the new target page as it could have a
+	 * stale value in the private field. Otherwise later on if this page
+	 * itself gets migrated, without getting referred by the Driver
+	 * in between, the stale value would cause the i915_migratepage
+	 * function to go for a toss as object pointer is derived from it.
+	 * This should be safe since at the time of migration, private field
+	 * of the new page (which is actually an independent free 4KB page now)
+	 * should be like a don't care for the kernel.
+	 */
+	set_page_private(newpage, 0);
+
+	if (!page_private(page))
+		goto migrate;
+
+	/*
+	 * Check the page count, if Driver also has a reference then it should
+	 * be more than 2, as shmem will have one reference and one reference
+	 * would have been taken by the migration path itself. So if reference
+	 * is <=2, we can directly invoke the migration function.
+	 */
+	if (page_count(page) <= 2)
+		goto migrate;
+
+	/*
+	 * Use trylock here, with a timeout, for struct_mutex as
+	 * otherwise there is a possibility of deadlock due to lock
+	 * inversion. This path, which tries to migrate a particular
+	 * page after locking that page, can race with a path which
+	 * truncate/purge pages of the corresponding object (after
+	 * acquiring struct_mutex). Since page truncation will also
+	 * try to lock the page, a scenario of deadlock can arise.
+	 */
+	if (!i915_gem_shrinker_lock_uninterruptible(dev_priv, &slu, 10))
+		return -EBUSY;
+
+	ret = 0;
+	if (!PageSwapCache(page) && page_private(page)) {
+		struct drm_i915_gem_object *obj =
+			(struct drm_i915_gem_object *)page_private(page);
+
+		ret = do_migrate_page(obj);
+	}
+
+	i915_gem_shrinker_unlock_uninterruptible(dev_priv, &slu);
+	if (ret)
+		return ret;
+
+	/*
+	 * Ideally here we don't expect the page count to be > 2, as driver
+	 * would have dropped its reference, but occasionally it has been seen
+	 * coming as 3 & 4. This leads to a situation of unexpected page count,
+	 * causing migration failure, with -EGAIN error. This then leads to
+	 * multiple attempts by the kernel to migrate the same set of pages.
+	 * And sometimes the repeated attempts proves detrimental for stability.
+	 * Also since we don't know who is the other owner, and for how long its
+	 * gonna keep the reference, its better to return -EBUSY.
+	 */
+	if (page_count(page) > 2)
+		return -EBUSY;
+
+migrate:
+	return migrate_page(mapping, newpage, page, mode);
+}
+#endif
+
 /**
  * i915_gem_shrinker_init - Initialize i915 shrinker
  * @dev_priv: i915 device
@@ -491,6 +618,11 @@ void i915_gem_shrinker_init(struct drm_i915_private *dev_priv)
 
 	dev_priv->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
 	WARN_ON(register_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
+
+	dev_priv->mm.shmem_info.private_data = dev_priv;
+#ifdef CONFIG_MIGRATION
+	dev_priv->mm.shmem_info.migratepage = i915_gem_shrinker_migratepage;
+#endif
 }
 
 /**
-- 
1.9.2

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH 2/2] drm/i915: Make GPU pages movable

[Daniel Vetter]

On Fri, Nov 04, 2016 at 08:32:56PM +0530, akash.goel@intel.com wrote:
> From: Chris Wilson <chris@chris-wilson.co.uk>
> 
> On a long run of more than 2-3 days, physical memory tends to get
> fragmented severely, which considerably slows down the system. In such a
> scenario, the shrinker is also unable to help as lack of memory is not
> the actual problem, since it has been observed that there are enough free
> pages of 0 order. This also manifests itself when an indiviual zone in
> the mm runs out of pages and if we cannot migrate pages between zones,
> the kernel hits an out-of-memory even though there are free pages (and
> often all of swap) available.
> 
> To address the issue of external fragementation, kernel does a compaction
> (which involves migration of pages) but it's efficacy depends upon how
> many pages are marked as MOVABLE, as only those pages can be migrated.
> 
> Currently the backing pages for GPU buffers are allocated from shmemfs
> with GFP_RECLAIMABLE flag, in units of 4KB pages.  In the case of limited
> swap space, it may not be possible always to reclaim or swap-out pages of
> all the inactive objects, to make way for free space allowing formation
> of higher order groups of physically-contiguous pages on compaction.
> 
> Just marking the GPU pages as MOVABLE will not suffice, as i915.ko has to
> pin the pages if they are in use by GPU, which will prevent their
> migration. So the migratepage callback in shmem is also hooked up to get
> a notification when kernel initiates the page migration. On the
> notification, i915.ko appropriately unpin the pages.  With this we can
> effectively mark the GPU pages as MOVABLE and hence mitigate the
> fragmentation problem.
> 
> v2:
>  - Rename the migration routine to gem_shrink_migratepage, move it to the
>    shrinker file, and use the existing constructs (Chris)
>  - To cleanup, add a new helper function to encapsulate all page migration
>    skip conditions (Chris)
>  - Add a new local helper function in shrinker file, for dropping the
>    backing pages, and call the same from gem_shrink() also (Chris)
> 
> v3:
>  - Fix/invert the check on the return value of unsafe_drop_pages (Chris)
> 
> v4:
>  - Minor tidy
> 
> v5:
>  - Fix unsafe usage of unsafe_drop_pages()
>  - Rebase onto vmap-notifier
> 
> v6:
> - Remove i915_gem_object_get/put across unsafe_drop_pages() as with
>   struct_mutex protection object can't disappear. (Chris)
> 
> Testcase: igt/gem_shrink
> Bugzilla: (e.g.) https://bugs.freedesktop.org/show_bug.cgi?id=90254
> Cc: Hugh Dickins <hughd@google.com>
> Cc: linux-mm@kvack.org
> Signed-off-by: Sourab Gupta <sourab.gupta@intel.com>
> Signed-off-by: Akash Goel <akash.goel@intel.com>
> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
> Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>

Hi all -mm folks!

Any feedback on these two? It's kinda an intermediate step towards a
full-blown gemfs, and I think useful for that. Or do we need to go
directly to our own backing storage thing? Aside from ack/nack from -mm I
think this is ready for merging.

Thanks, Daniel

> ---
>  drivers/gpu/drm/i915/i915_drv.h          |   2 +
>  drivers/gpu/drm/i915/i915_gem.c          |   9 ++-
>  drivers/gpu/drm/i915/i915_gem_shrinker.c | 132 +++++++++++++++++++++++++++++++
>  3 files changed, 142 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> index 4735b417..7f2717b 100644
> --- a/drivers/gpu/drm/i915/i915_drv.h
> +++ b/drivers/gpu/drm/i915/i915_drv.h
> @@ -1357,6 +1357,8 @@ struct intel_l3_parity {
>  };
>  
>  struct i915_gem_mm {
> +	struct shmem_dev_info shmem_info;
> +
>  	/** Memory allocator for GTT stolen memory */
>  	struct drm_mm stolen;
>  	/** Protects the usage of the GTT stolen memory allocator. This is
> diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> index 1f995ce..f0d4ce7 100644
> --- a/drivers/gpu/drm/i915/i915_gem.c
> +++ b/drivers/gpu/drm/i915/i915_gem.c
> @@ -2164,6 +2164,7 @@ void __i915_gem_object_invalidate(struct drm_i915_gem_object *obj)
>  		if (obj->mm.madv == I915_MADV_WILLNEED)
>  			mark_page_accessed(page);
>  
> +		set_page_private(page, 0);
>  		put_page(page);
>  	}
>  	obj->mm.dirty = false;
> @@ -2310,6 +2311,7 @@ static unsigned int swiotlb_max_size(void)
>  			sg->length += PAGE_SIZE;
>  		}
>  		last_pfn = page_to_pfn(page);
> +		set_page_private(page, (unsigned long)obj);
>  
>  		/* Check that the i965g/gm workaround works. */
>  		WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL));
> @@ -2334,8 +2336,10 @@ static unsigned int swiotlb_max_size(void)
>  
>  err_pages:
>  	sg_mark_end(sg);
> -	for_each_sgt_page(page, sgt_iter, st)
> +	for_each_sgt_page(page, sgt_iter, st) {
> +		set_page_private(page, 0);
>  		put_page(page);
> +	}
>  	sg_free_table(st);
>  	kfree(st);
>  
> @@ -4185,6 +4189,8 @@ struct drm_i915_gem_object *
>  		goto fail;
>  
>  	mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
> +	if (IS_ENABLED(MIGRATION))
> +		mask |= __GFP_MOVABLE;
>  	if (IS_CRESTLINE(dev_priv) || IS_BROADWATER(dev_priv)) {
>  		/* 965gm cannot relocate objects above 4GiB. */
>  		mask &= ~__GFP_HIGHMEM;
> @@ -4193,6 +4199,7 @@ struct drm_i915_gem_object *
>  
>  	mapping = obj->base.filp->f_mapping;
>  	mapping_set_gfp_mask(mapping, mask);
> +	shmem_set_dev_info(mapping, &dev_priv->mm.shmem_info);
>  
>  	i915_gem_object_init(obj, &i915_gem_object_ops);
>  
> diff --git a/drivers/gpu/drm/i915/i915_gem_shrinker.c b/drivers/gpu/drm/i915/i915_gem_shrinker.c
> index a6fc1bd..051135ac 100644
> --- a/drivers/gpu/drm/i915/i915_gem_shrinker.c
> +++ b/drivers/gpu/drm/i915/i915_gem_shrinker.c
> @@ -24,6 +24,7 @@
>  
>  #include <linux/oom.h>
>  #include <linux/shmem_fs.h>
> +#include <linux/migrate.h>
>  #include <linux/slab.h>
>  #include <linux/swap.h>
>  #include <linux/pci.h>
> @@ -473,6 +474,132 @@ struct shrinker_lock_uninterruptible {
>  	return NOTIFY_DONE;
>  }
>  
> +#ifdef CONFIG_MIGRATION
> +static bool can_migrate_page(struct drm_i915_gem_object *obj)
> +{
> +	/* Avoid the migration of page if being actively used by GPU */
> +	if (i915_gem_object_is_active(obj))
> +		return false;
> +
> +	/* Skip the migration for purgeable objects otherwise there
> +	 * will be a deadlock when shmem will try to lock the page for
> +	 * truncation, which is already locked by the caller before
> +	 * migration.
> +	 */
> +	if (obj->mm.madv == I915_MADV_DONTNEED)
> +		return false;
> +
> +	/* Skip the migration for a pinned object */
> +	if (atomic_read(&obj->mm.pages_pin_count) > obj->bind_count)
> +		return false;
> +
> +	if (any_vma_pinned(obj))
> +		return false;
> +
> +	return true;
> +}
> +
> +static int do_migrate_page(struct drm_i915_gem_object *obj)
> +{
> +	struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
> +	int ret = 0;
> +
> +	if (!can_migrate_page(obj))
> +		return -EBUSY;
> +
> +	/* HW access would be required for a GGTT bound object, for which
> +	 * device has to be kept awake. But a deadlock scenario can arise if
> +	 * the attempt is made to resume the device, when either a suspend
> +	 * or a resume operation is already happening concurrently from some
> +	 * other path and that only also triggers compaction. So only unbind
> +	 * if the device is currently awake.
> +	 */
> +	if (!intel_runtime_pm_get_if_in_use(dev_priv))
> +		return -EBUSY;
> +
> +	if (!unsafe_drop_pages(obj))
> +		ret = -EBUSY;
> +
> +	intel_runtime_pm_put(dev_priv);
> +	return ret;
> +}
> +
> +static int i915_gem_shrinker_migratepage(struct address_space *mapping,
> +					 struct page *newpage,
> +					 struct page *page,
> +					 enum migrate_mode mode,
> +					 void *dev_priv_data)
> +{
> +	struct drm_i915_private *dev_priv = dev_priv_data;
> +	struct shrinker_lock_uninterruptible slu;
> +	int ret;
> +
> +	/*
> +	 * Clear the private field of the new target page as it could have a
> +	 * stale value in the private field. Otherwise later on if this page
> +	 * itself gets migrated, without getting referred by the Driver
> +	 * in between, the stale value would cause the i915_migratepage
> +	 * function to go for a toss as object pointer is derived from it.
> +	 * This should be safe since at the time of migration, private field
> +	 * of the new page (which is actually an independent free 4KB page now)
> +	 * should be like a don't care for the kernel.
> +	 */
> +	set_page_private(newpage, 0);
> +
> +	if (!page_private(page))
> +		goto migrate;
> +
> +	/*
> +	 * Check the page count, if Driver also has a reference then it should
> +	 * be more than 2, as shmem will have one reference and one reference
> +	 * would have been taken by the migration path itself. So if reference
> +	 * is <=2, we can directly invoke the migration function.
> +	 */
> +	if (page_count(page) <= 2)
> +		goto migrate;
> +
> +	/*
> +	 * Use trylock here, with a timeout, for struct_mutex as
> +	 * otherwise there is a possibility of deadlock due to lock
> +	 * inversion. This path, which tries to migrate a particular
> +	 * page after locking that page, can race with a path which
> +	 * truncate/purge pages of the corresponding object (after
> +	 * acquiring struct_mutex). Since page truncation will also
> +	 * try to lock the page, a scenario of deadlock can arise.
> +	 */
> +	if (!i915_gem_shrinker_lock_uninterruptible(dev_priv, &slu, 10))
> +		return -EBUSY;
> +
> +	ret = 0;
> +	if (!PageSwapCache(page) && page_private(page)) {
> +		struct drm_i915_gem_object *obj =
> +			(struct drm_i915_gem_object *)page_private(page);
> +
> +		ret = do_migrate_page(obj);
> +	}
> +
> +	i915_gem_shrinker_unlock_uninterruptible(dev_priv, &slu);
> +	if (ret)
> +		return ret;
> +
> +	/*
> +	 * Ideally here we don't expect the page count to be > 2, as driver
> +	 * would have dropped its reference, but occasionally it has been seen
> +	 * coming as 3 & 4. This leads to a situation of unexpected page count,
> +	 * causing migration failure, with -EGAIN error. This then leads to
> +	 * multiple attempts by the kernel to migrate the same set of pages.
> +	 * And sometimes the repeated attempts proves detrimental for stability.
> +	 * Also since we don't know who is the other owner, and for how long its
> +	 * gonna keep the reference, its better to return -EBUSY.
> +	 */
> +	if (page_count(page) > 2)
> +		return -EBUSY;
> +
> +migrate:
> +	return migrate_page(mapping, newpage, page, mode);
> +}
> +#endif
> +
>  /**
>   * i915_gem_shrinker_init - Initialize i915 shrinker
>   * @dev_priv: i915 device
> @@ -491,6 +618,11 @@ void i915_gem_shrinker_init(struct drm_i915_private *dev_priv)
>  
>  	dev_priv->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
>  	WARN_ON(register_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
> +
> +	dev_priv->mm.shmem_info.private_data = dev_priv;
> +#ifdef CONFIG_MIGRATION
> +	dev_priv->mm.shmem_info.migratepage = i915_gem_shrinker_migratepage;
> +#endif
>  }
>  
>  /**
> -- 
> 1.9.2
> 
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@kvack.org.  For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH 2/2] drm/i915: Make GPU pages movable

[Hugh Dickins]

On Wed, 9 Nov 2016, Daniel Vetter wrote:
> 
> Hi all -mm folks!
> 
> Any feedback on these two? It's kinda an intermediate step towards a
> full-blown gemfs, and I think useful for that. Or do we need to go
> directly to our own backing storage thing? Aside from ack/nack from -mm I
> think this is ready for merging.

I'm currently considering them at last: will report back later.

Full-blown gemfs does not come in here, of course; but let me
fire a warning shot since you mention it: if it's going to use swap,
then we shall probably have to nak it in favour of continuing to use 
infrastructure from mm/shmem.c.  I very much understand why you would
love to avoid that dependence, but I doubt it can be safely bypassed.

Hugh

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH 2/2] drm/i915: Make GPU pages movable

[Hugh Dickins]

On Fri, 4 Nov 2016, akash.goel@intel.com wrote:
> From: Chris Wilson <chris@chris-wilson.co.uk>
> 
> On a long run of more than 2-3 days, physical memory tends to get
> fragmented severely, which considerably slows down the system. In such a
> scenario, the shrinker is also unable to help as lack of memory is not
> the actual problem, since it has been observed that there are enough free
> pages of 0 order. This also manifests itself when an indiviual zone in
> the mm runs out of pages and if we cannot migrate pages between zones,
> the kernel hits an out-of-memory even though there are free pages (and
> often all of swap) available.
> 
> To address the issue of external fragementation, kernel does a compaction
> (which involves migration of pages) but it's efficacy depends upon how
> many pages are marked as MOVABLE, as only those pages can be migrated.
> 
> Currently the backing pages for GPU buffers are allocated from shmemfs
> with GFP_RECLAIMABLE flag, in units of 4KB pages.  In the case of limited
> swap space, it may not be possible always to reclaim or swap-out pages of
> all the inactive objects, to make way for free space allowing formation
> of higher order groups of physically-contiguous pages on compaction.
> 
> Just marking the GPU pages as MOVABLE will not suffice, as i915.ko has to
> pin the pages if they are in use by GPU, which will prevent their
> migration. So the migratepage callback in shmem is also hooked up to get
> a notification when kernel initiates the page migration. On the
> notification, i915.ko appropriately unpin the pages.  With this we can
> effectively mark the GPU pages as MOVABLE and hence mitigate the
> fragmentation problem.
> 
> v2:
>  - Rename the migration routine to gem_shrink_migratepage, move it to the
>    shrinker file, and use the existing constructs (Chris)
>  - To cleanup, add a new helper function to encapsulate all page migration
>    skip conditions (Chris)
>  - Add a new local helper function in shrinker file, for dropping the
>    backing pages, and call the same from gem_shrink() also (Chris)
> 
> v3:
>  - Fix/invert the check on the return value of unsafe_drop_pages (Chris)
> 
> v4:
>  - Minor tidy
> 
> v5:
>  - Fix unsafe usage of unsafe_drop_pages()
>  - Rebase onto vmap-notifier
> 
> v6:
> - Remove i915_gem_object_get/put across unsafe_drop_pages() as with
>   struct_mutex protection object can't disappear. (Chris)
> 
> Testcase: igt/gem_shrink
> Bugzilla: (e.g.) https://bugs.freedesktop.org/show_bug.cgi?id=90254
> Cc: Hugh Dickins <hughd@google.com>
> Cc: linux-mm@kvack.org
> Signed-off-by: Sourab Gupta <sourab.gupta@intel.com>
> Signed-off-by: Akash Goel <akash.goel@intel.com>
> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
> Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>

I'm confused!  But perhaps it's gone around and around between you all,
I'm not sure what the rules are then.  I think this sequence implies
that Sourab wrote it originally, then Akash and Chris passed it on
with refinements - but then Chris wouldn't add Reviewed-by.

> ---
>  drivers/gpu/drm/i915/i915_drv.h          |   2 +
>  drivers/gpu/drm/i915/i915_gem.c          |   9 ++-
>  drivers/gpu/drm/i915/i915_gem_shrinker.c | 132 +++++++++++++++++++++++++++++++
>  3 files changed, 142 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> index 4735b417..7f2717b 100644
> --- a/drivers/gpu/drm/i915/i915_drv.h
> +++ b/drivers/gpu/drm/i915/i915_drv.h
> @@ -1357,6 +1357,8 @@ struct intel_l3_parity {
>  };
>  
>  struct i915_gem_mm {
> +	struct shmem_dev_info shmem_info;
> +
>  	/** Memory allocator for GTT stolen memory */
>  	struct drm_mm stolen;
>  	/** Protects the usage of the GTT stolen memory allocator. This is
> diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> index 1f995ce..f0d4ce7 100644
> --- a/drivers/gpu/drm/i915/i915_gem.c
> +++ b/drivers/gpu/drm/i915/i915_gem.c
> @@ -2164,6 +2164,7 @@ void __i915_gem_object_invalidate(struct drm_i915_gem_object *obj)
>  		if (obj->mm.madv == I915_MADV_WILLNEED)
>  			mark_page_accessed(page);
>  
> +		set_page_private(page, 0);
>  		put_page(page);
>  	}
>  	obj->mm.dirty = false;
> @@ -2310,6 +2311,7 @@ static unsigned int swiotlb_max_size(void)
>  			sg->length += PAGE_SIZE;
>  		}
>  		last_pfn = page_to_pfn(page);
> +		set_page_private(page, (unsigned long)obj);
>  
>  		/* Check that the i965g/gm workaround works. */
>  		WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL));
> @@ -2334,8 +2336,10 @@ static unsigned int swiotlb_max_size(void)
>  
>  err_pages:
>  	sg_mark_end(sg);
> -	for_each_sgt_page(page, sgt_iter, st)
> +	for_each_sgt_page(page, sgt_iter, st) {
> +		set_page_private(page, 0);
>  		put_page(page);
> +	}
>  	sg_free_table(st);
>  	kfree(st);
>  

I think your page_private() games there and below are correct (and
I suspect took a few iterations to get right); and we don't currently
have a use for page_private() in mm/shmem.c that conflicts with what
you're doing here (of course it's used for swap, but you're already
careful to exclude that case).

But I'd nonetheless prefer not to give it away to you: you're welcome
to use mapping->private_data as you have, but I'd rather keep the more
valuable page_private() available for mm or shmem use.

Would it be reasonable to ask you to rework this with the shmem_dev_info
in dev_priv replaced by shmem_obj_info in drm_i915_gem_object?  Then,
IIUC, you can access both it and the object which contains it from
the page->mapping pointer, without needing page->private.

It that's unreasonable to ask of you, would it be reasonable if I
added a third patch to make that change myself?

> @@ -4185,6 +4189,8 @@ struct drm_i915_gem_object *
>  		goto fail;
>  
>  	mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
> +	if (IS_ENABLED(MIGRATION))
> +		mask |= __GFP_MOVABLE;

I was going to suggest just make that unconditional,
        mask = GFP_HIGHUSER_MOVABLE | __GFP_RECLAIMABLE;

But then I wondered what that __GFP_RECLAIMABLE actually achieves?
These pages are already __GFP_RECLAIM (inside GFP_HIGHUSER) and on
the LRU.  It affects gfpflags_to_migratetype(), but I'm not familiar
with what that different migratetype will end up doing.

>  	if (IS_CRESTLINE(dev_priv) || IS_BROADWATER(dev_priv)) {
>  		/* 965gm cannot relocate objects above 4GiB. */
>  		mask &= ~__GFP_HIGHMEM;
> @@ -4193,6 +4199,7 @@ struct drm_i915_gem_object *
>  
>  	mapping = obj->base.filp->f_mapping;
>  	mapping_set_gfp_mask(mapping, mask);
> +	shmem_set_dev_info(mapping, &dev_priv->mm.shmem_info);
>  
>  	i915_gem_object_init(obj, &i915_gem_object_ops);
>  
> diff --git a/drivers/gpu/drm/i915/i915_gem_shrinker.c b/drivers/gpu/drm/i915/i915_gem_shrinker.c
> index a6fc1bd..051135ac 100644
> --- a/drivers/gpu/drm/i915/i915_gem_shrinker.c
> +++ b/drivers/gpu/drm/i915/i915_gem_shrinker.c
> @@ -24,6 +24,7 @@
>  
>  #include <linux/oom.h>
>  #include <linux/shmem_fs.h>
> +#include <linux/migrate.h>
>  #include <linux/slab.h>
>  #include <linux/swap.h>
>  #include <linux/pci.h>
> @@ -473,6 +474,132 @@ struct shrinker_lock_uninterruptible {
>  	return NOTIFY_DONE;
>  }
>  
> +#ifdef CONFIG_MIGRATION
> +static bool can_migrate_page(struct drm_i915_gem_object *obj)
> +{
> +	/* Avoid the migration of page if being actively used by GPU */
> +	if (i915_gem_object_is_active(obj))
> +		return false;
> +
> +	/* Skip the migration for purgeable objects otherwise there
> +	 * will be a deadlock when shmem will try to lock the page for
> +	 * truncation, which is already locked by the caller before
> +	 * migration.
> +	 */
> +	if (obj->mm.madv == I915_MADV_DONTNEED)
> +		return false;
> +
> +	/* Skip the migration for a pinned object */
> +	if (atomic_read(&obj->mm.pages_pin_count) > obj->bind_count)
> +		return false;
> +
> +	if (any_vma_pinned(obj))
> +		return false;
> +
> +	return true;
> +}
> +
> +static int do_migrate_page(struct drm_i915_gem_object *obj)

do_migrate_page()?  But it does not.  Maybe prepare_for_migrate_page()?

> +{
> +	struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
> +	int ret = 0;
> +
> +	if (!can_migrate_page(obj))
> +		return -EBUSY;
> +
> +	/* HW access would be required for a GGTT bound object, for which
> +	 * device has to be kept awake. But a deadlock scenario can arise if
> +	 * the attempt is made to resume the device, when either a suspend
> +	 * or a resume operation is already happening concurrently from some
> +	 * other path and that only also triggers compaction. So only unbind
> +	 * if the device is currently awake.
> +	 */
> +	if (!intel_runtime_pm_get_if_in_use(dev_priv))
> +		return -EBUSY;
> +
> +	if (!unsafe_drop_pages(obj))
> +		ret = -EBUSY;
> +
> +	intel_runtime_pm_put(dev_priv);
> +	return ret;
> +}
> +
> +static int i915_gem_shrinker_migratepage(struct address_space *mapping,
> +					 struct page *newpage,
> +					 struct page *page,
> +					 enum migrate_mode mode,
> +					 void *dev_priv_data)
> +{
> +	struct drm_i915_private *dev_priv = dev_priv_data;
> +	struct shrinker_lock_uninterruptible slu;
> +	int ret;
> +
> +	/*
> +	 * Clear the private field of the new target page as it could have a
> +	 * stale value in the private field. Otherwise later on if this page
> +	 * itself gets migrated, without getting referred by the Driver
> +	 * in between, the stale value would cause the i915_migratepage
> +	 * function to go for a toss as object pointer is derived from it.
> +	 * This should be safe since at the time of migration, private field
> +	 * of the new page (which is actually an independent free 4KB page now)
> +	 * should be like a don't care for the kernel.
> +	 */
> +	set_page_private(newpage, 0);
> +
> +	if (!page_private(page))
> +		goto migrate;
> +
> +	/*
> +	 * Check the page count, if Driver also has a reference then it should
> +	 * be more than 2, as shmem will have one reference and one reference
> +	 * would have been taken by the migration path itself. So if reference
> +	 * is <=2, we can directly invoke the migration function.
> +	 */
> +	if (page_count(page) <= 2)
> +		goto migrate;
> +
> +	/*
> +	 * Use trylock here, with a timeout, for struct_mutex as
> +	 * otherwise there is a possibility of deadlock due to lock
> +	 * inversion. This path, which tries to migrate a particular
> +	 * page after locking that page, can race with a path which
> +	 * truncate/purge pages of the corresponding object (after
> +	 * acquiring struct_mutex). Since page truncation will also
> +	 * try to lock the page, a scenario of deadlock can arise.
> +	 */
> +	if (!i915_gem_shrinker_lock_uninterruptible(dev_priv, &slu, 10))
> +		return -EBUSY;
> +
> +	ret = 0;
> +	if (!PageSwapCache(page) && page_private(page)) {
> +		struct drm_i915_gem_object *obj =
> +			(struct drm_i915_gem_object *)page_private(page);
> +
> +		ret = do_migrate_page(obj);
> +	}
> +
> +	i915_gem_shrinker_unlock_uninterruptible(dev_priv, &slu);
> +	if (ret)
> +		return ret;
> +
> +	/*
> +	 * Ideally here we don't expect the page count to be > 2, as driver
> +	 * would have dropped its reference, but occasionally it has been seen
> +	 * coming as 3 & 4. This leads to a situation of unexpected page count,
> +	 * causing migration failure, with -EGAIN error. This then leads to

s/EGAIN/EAGAIN/

> +	 * multiple attempts by the kernel to migrate the same set of pages.
> +	 * And sometimes the repeated attempts proves detrimental for stability.
> +	 * Also since we don't know who is the other owner, and for how long its
> +	 * gonna keep the reference, its better to return -EBUSY.

Fair enough, I know those 10 repeats can be quite a waste.  And you've
got a potential 10ms timeout above, which you don't want to multiply by 10.
I can't get too sniffy about your timeout, we have other sources of delay
in there, but it is always sad to add latency to MIGRATE_ASYNC mode.

> +	 */
> +	if (page_count(page) > 2)
> +		return -EBUSY;
> +
> +migrate:
> +	return migrate_page(mapping, newpage, page, mode);
> +}
> +#endif
> +
>  /**
>   * i915_gem_shrinker_init - Initialize i915 shrinker
>   * @dev_priv: i915 device
> @@ -491,6 +618,11 @@ void i915_gem_shrinker_init(struct drm_i915_private *dev_priv)
>  
>  	dev_priv->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
>  	WARN_ON(register_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
> +
> +	dev_priv->mm.shmem_info.private_data = dev_priv;
> +#ifdef CONFIG_MIGRATION
> +	dev_priv->mm.shmem_info.migratepage = i915_gem_shrinker_migratepage;
> +#endif

If we avoid playing with page_private(), this initialization would go
away, but the equivalent be done near the call to i915_gem_object_init().

>  }
>  
>  /**
> -- 
> 1.9.2

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH 2/2] drm/i915: Make GPU pages movable

[Goel, Akash]

On 11/10/2016 12:09 PM, Hugh Dickins wrote:
> On Fri, 4 Nov 2016, akash.goel@intel.com wrote:
>> From: Chris Wilson <chris@chris-wilson.co.uk>
>>
>> On a long run of more than 2-3 days, physical memory tends to get
>> fragmented severely, which considerably slows down the system. In such a
>> scenario, the shrinker is also unable to help as lack of memory is not
>> the actual problem, since it has been observed that there are enough free
>> pages of 0 order. This also manifests itself when an indiviual zone in
>> the mm runs out of pages and if we cannot migrate pages between zones,
>> the kernel hits an out-of-memory even though there are free pages (and
>> often all of swap) available.
>>
>> To address the issue of external fragementation, kernel does a compaction
>> (which involves migration of pages) but it's efficacy depends upon how
>> many pages are marked as MOVABLE, as only those pages can be migrated.
>>
>> Currently the backing pages for GPU buffers are allocated from shmemfs
>> with GFP_RECLAIMABLE flag, in units of 4KB pages.  In the case of limited
>> swap space, it may not be possible always to reclaim or swap-out pages of
>> all the inactive objects, to make way for free space allowing formation
>> of higher order groups of physically-contiguous pages on compaction.
>>
>> Just marking the GPU pages as MOVABLE will not suffice, as i915.ko has to
>> pin the pages if they are in use by GPU, which will prevent their
>> migration. So the migratepage callback in shmem is also hooked up to get
>> a notification when kernel initiates the page migration. On the
>> notification, i915.ko appropriately unpin the pages.  With this we can
>> effectively mark the GPU pages as MOVABLE and hence mitigate the
>> fragmentation problem.
>>
>> v2:
>>  - Rename the migration routine to gem_shrink_migratepage, move it to the
>>    shrinker file, and use the existing constructs (Chris)
>>  - To cleanup, add a new helper function to encapsulate all page migration
>>    skip conditions (Chris)
>>  - Add a new local helper function in shrinker file, for dropping the
>>    backing pages, and call the same from gem_shrink() also (Chris)
>>
>> v3:
>>  - Fix/invert the check on the return value of unsafe_drop_pages (Chris)
>>
>> v4:
>>  - Minor tidy
>>
>> v5:
>>  - Fix unsafe usage of unsafe_drop_pages()
>>  - Rebase onto vmap-notifier
>>
>> v6:
>> - Remove i915_gem_object_get/put across unsafe_drop_pages() as with
>>   struct_mutex protection object can't disappear. (Chris)
>>
>> Testcase: igt/gem_shrink
>> Bugzilla: (e.g.) https://bugs.freedesktop.org/show_bug.cgi?id=90254
>> Cc: Hugh Dickins <hughd@google.com>
>> Cc: linux-mm@kvack.org
>> Signed-off-by: Sourab Gupta <sourab.gupta@intel.com>
>> Signed-off-by: Akash Goel <akash.goel@intel.com>
>> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
>> Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
>> Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
>
> I'm confused!  But perhaps it's gone around and around between you all,
> I'm not sure what the rules are then.  I think this sequence implies
> that Sourab wrote it originally, then Akash and Chris passed it on
> with refinements - but then Chris wouldn't add Reviewed-by.
>
Thank you very much for the review and sorry for all the needless confusion.

Chris actually conceived the patches and prepared an initial version of 
them (hence he is the Author).
I & Sourab did the further refinements and fixed issues (all those 
page_private stuff).
Chris then reviewed the final patch and also recently did a rebase for it.

>> ---
>>  drivers/gpu/drm/i915/i915_drv.h          |   2 +
>>  drivers/gpu/drm/i915/i915_gem.c          |   9 ++-
>>  drivers/gpu/drm/i915/i915_gem_shrinker.c | 132 +++++++++++++++++++++++++++++++
>>  3 files changed, 142 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
>> index 4735b417..7f2717b 100644
>> --- a/drivers/gpu/drm/i915/i915_drv.h
>> +++ b/drivers/gpu/drm/i915/i915_drv.h
>> @@ -1357,6 +1357,8 @@ struct intel_l3_parity {
>>  };
>>
>>  struct i915_gem_mm {
>> +	struct shmem_dev_info shmem_info;
>> +
>>  	/** Memory allocator for GTT stolen memory */
>>  	struct drm_mm stolen;
>>  	/** Protects the usage of the GTT stolen memory allocator. This is
>> diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
>> index 1f995ce..f0d4ce7 100644
>> --- a/drivers/gpu/drm/i915/i915_gem.c
>> +++ b/drivers/gpu/drm/i915/i915_gem.c
>> @@ -2164,6 +2164,7 @@ void __i915_gem_object_invalidate(struct drm_i915_gem_object *obj)
>>  		if (obj->mm.madv == I915_MADV_WILLNEED)
>>  			mark_page_accessed(page);
>>
>> +		set_page_private(page, 0);
>>  		put_page(page);
>>  	}
>>  	obj->mm.dirty = false;
>> @@ -2310,6 +2311,7 @@ static unsigned int swiotlb_max_size(void)
>>  			sg->length += PAGE_SIZE;
>>  		}
>>  		last_pfn = page_to_pfn(page);
>> +		set_page_private(page, (unsigned long)obj);
>>
>>  		/* Check that the i965g/gm workaround works. */
>>  		WARN_ON((gfp & __GFP_DMA32) && (last_pfn >= 0x00100000UL));
>> @@ -2334,8 +2336,10 @@ static unsigned int swiotlb_max_size(void)
>>
>>  err_pages:
>>  	sg_mark_end(sg);
>> -	for_each_sgt_page(page, sgt_iter, st)
>> +	for_each_sgt_page(page, sgt_iter, st) {
>> +		set_page_private(page, 0);
>>  		put_page(page);
>> +	}
>>  	sg_free_table(st);
>>  	kfree(st);
>>
>
> I think your page_private() games there and below are correct (and
> I suspect took a few iterations to get right); and we don't currently
> have a use for page_private() in mm/shmem.c that conflicts with what
> you're doing here (of course it's used for swap, but you're already
> careful to exclude that case).
>
> But I'd nonetheless prefer not to give it away to you: you're welcome
> to use mapping->private_data as you have, but I'd rather keep the more
> valuable page_private() available for mm or shmem use.
>
> Would it be reasonable to ask you to rework this with the shmem_dev_info
> in dev_priv replaced by shmem_obj_info in drm_i915_gem_object?  Then,
> IIUC, you can access both it and the object which contains it from
> the page->mapping pointer, without needing page->private.
>
If I understood your suggestion correctly, instead of page_private the 
object pointer can be derived from mapping->private (by having 
shmem_obj_info embedded inside drm_i915_gem_object instead of dev_priv).

Will definitely try to rework the patch as per your suggestions.

> It that's unreasonable to ask of you, would it be reasonable if I
> added a third patch to make that change myself?
>
>> @@ -4185,6 +4189,8 @@ struct drm_i915_gem_object *
>>  		goto fail;
>>
>>  	mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
>> +	if (IS_ENABLED(MIGRATION))
>> +		mask |= __GFP_MOVABLE;
>
> I was going to suggest just make that unconditional,
>         mask = GFP_HIGHUSER_MOVABLE | __GFP_RECLAIMABLE;
>
> But then I wondered what that __GFP_RECLAIMABLE actually achieves?
> These pages are already __GFP_RECLAIM (inside GFP_HIGHUSER) and on
> the LRU.  It affects gfpflags_to_migratetype(), but I'm not familiar
> with what that different migratetype will end up doing.
>

Will check for this.

>>  	if (IS_CRESTLINE(dev_priv) || IS_BROADWATER(dev_priv)) {
>>  		/* 965gm cannot relocate objects above 4GiB. */
>>  		mask &= ~__GFP_HIGHMEM;
>> @@ -4193,6 +4199,7 @@ struct drm_i915_gem_object *
>>
>>  	mapping = obj->base.filp->f_mapping;
>>  	mapping_set_gfp_mask(mapping, mask);
>> +	shmem_set_dev_info(mapping, &dev_priv->mm.shmem_info);
>>
>>  	i915_gem_object_init(obj, &i915_gem_object_ops);
>>
>> diff --git a/drivers/gpu/drm/i915/i915_gem_shrinker.c b/drivers/gpu/drm/i915/i915_gem_shrinker.c
>> index a6fc1bd..051135ac 100644
>> --- a/drivers/gpu/drm/i915/i915_gem_shrinker.c
>> +++ b/drivers/gpu/drm/i915/i915_gem_shrinker.c
>> @@ -24,6 +24,7 @@
>>
>>  #include <linux/oom.h>
>>  #include <linux/shmem_fs.h>
>> +#include <linux/migrate.h>
>>  #include <linux/slab.h>
>>  #include <linux/swap.h>
>>  #include <linux/pci.h>
>> @@ -473,6 +474,132 @@ struct shrinker_lock_uninterruptible {
>>  	return NOTIFY_DONE;
>>  }
>>
>> +#ifdef CONFIG_MIGRATION
>> +static bool can_migrate_page(struct drm_i915_gem_object *obj)
>> +{
>> +	/* Avoid the migration of page if being actively used by GPU */
>> +	if (i915_gem_object_is_active(obj))
>> +		return false;
>> +
>> +	/* Skip the migration for purgeable objects otherwise there
>> +	 * will be a deadlock when shmem will try to lock the page for
>> +	 * truncation, which is already locked by the caller before
>> +	 * migration.
>> +	 */
>> +	if (obj->mm.madv == I915_MADV_DONTNEED)
>> +		return false;
>> +
>> +	/* Skip the migration for a pinned object */
>> +	if (atomic_read(&obj->mm.pages_pin_count) > obj->bind_count)
>> +		return false;
>> +
>> +	if (any_vma_pinned(obj))
>> +		return false;
>> +
>> +	return true;
>> +}
>> +
>> +static int do_migrate_page(struct drm_i915_gem_object *obj)
>
> do_migrate_page()?  But it does not.  Maybe prepare_for_migrate_page()?
>
fine..

>> +{
>> +	struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
>> +	int ret = 0;
>> +
>> +	if (!can_migrate_page(obj))
>> +		return -EBUSY;
>> +
>> +	/* HW access would be required for a GGTT bound object, for which
>> +	 * device has to be kept awake. But a deadlock scenario can arise if
>> +	 * the attempt is made to resume the device, when either a suspend
>> +	 * or a resume operation is already happening concurrently from some
>> +	 * other path and that only also triggers compaction. So only unbind
>> +	 * if the device is currently awake.
>> +	 */
>> +	if (!intel_runtime_pm_get_if_in_use(dev_priv))
>> +		return -EBUSY;
>> +
>> +	if (!unsafe_drop_pages(obj))
>> +		ret = -EBUSY;
>> +
>> +	intel_runtime_pm_put(dev_priv);
>> +	return ret;
>> +}
>> +
>> +static int i915_gem_shrinker_migratepage(struct address_space *mapping,
>> +					 struct page *newpage,
>> +					 struct page *page,
>> +					 enum migrate_mode mode,
>> +					 void *dev_priv_data)
>> +{
>> +	struct drm_i915_private *dev_priv = dev_priv_data;
>> +	struct shrinker_lock_uninterruptible slu;
>> +	int ret;
>> +
>> +	/*
>> +	 * Clear the private field of the new target page as it could have a
>> +	 * stale value in the private field. Otherwise later on if this page
>> +	 * itself gets migrated, without getting referred by the Driver
>> +	 * in between, the stale value would cause the i915_migratepage
>> +	 * function to go for a toss as object pointer is derived from it.
>> +	 * This should be safe since at the time of migration, private field
>> +	 * of the new page (which is actually an independent free 4KB page now)
>> +	 * should be like a don't care for the kernel.
>> +	 */
>> +	set_page_private(newpage, 0);
>> +
>> +	if (!page_private(page))
>> +		goto migrate;
>> +
>> +	/*
>> +	 * Check the page count, if Driver also has a reference then it should
>> +	 * be more than 2, as shmem will have one reference and one reference
>> +	 * would have been taken by the migration path itself. So if reference
>> +	 * is <=2, we can directly invoke the migration function.
>> +	 */
>> +	if (page_count(page) <= 2)
>> +		goto migrate;
>> +
>> +	/*
>> +	 * Use trylock here, with a timeout, for struct_mutex as
>> +	 * otherwise there is a possibility of deadlock due to lock
>> +	 * inversion. This path, which tries to migrate a particular
>> +	 * page after locking that page, can race with a path which
>> +	 * truncate/purge pages of the corresponding object (after
>> +	 * acquiring struct_mutex). Since page truncation will also
>> +	 * try to lock the page, a scenario of deadlock can arise.
>> +	 */
>> +	if (!i915_gem_shrinker_lock_uninterruptible(dev_priv, &slu, 10))
>> +		return -EBUSY;
>> +
>> +	ret = 0;
>> +	if (!PageSwapCache(page) && page_private(page)) {
>> +		struct drm_i915_gem_object *obj =
>> +			(struct drm_i915_gem_object *)page_private(page);
>> +
>> +		ret = do_migrate_page(obj);
>> +	}
>> +
>> +	i915_gem_shrinker_unlock_uninterruptible(dev_priv, &slu);
>> +	if (ret)
>> +		return ret;
>> +
>> +	/*
>> +	 * Ideally here we don't expect the page count to be > 2, as driver
>> +	 * would have dropped its reference, but occasionally it has been seen
>> +	 * coming as 3 & 4. This leads to a situation of unexpected page count,
>> +	 * causing migration failure, with -EGAIN error. This then leads to
>
> s/EGAIN/EAGAIN/
Fine.
>
>> +	 * multiple attempts by the kernel to migrate the same set of pages.
>> +	 * And sometimes the repeated attempts proves detrimental for stability.
>> +	 * Also since we don't know who is the other owner, and for how long its
>> +	 * gonna keep the reference, its better to return -EBUSY.
>
> Fair enough, I know those 10 repeats can be quite a waste.  And you've
> got a potential 10ms timeout above, which you don't want to multiply by 10.
> I can't get too sniffy about your timeout, we have other sources of delay
> in there, but it is always sad to add latency to MIGRATE_ASYNC mode.
>
>> +	 */
>> +	if (page_count(page) > 2)
>> +		return -EBUSY;
>> +
>> +migrate:
>> +	return migrate_page(mapping, newpage, page, mode);
>> +}
>> +#endif
>> +
>>  /**
>>   * i915_gem_shrinker_init - Initialize i915 shrinker
>>   * @dev_priv: i915 device
>> @@ -491,6 +618,11 @@ void i915_gem_shrinker_init(struct drm_i915_private *dev_priv)
>>
>>  	dev_priv->mm.vmap_notifier.notifier_call = i915_gem_shrinker_vmap;
>>  	WARN_ON(register_vmap_purge_notifier(&dev_priv->mm.vmap_notifier));
>> +
>> +	dev_priv->mm.shmem_info.private_data = dev_priv;
>> +#ifdef CONFIG_MIGRATION
>> +	dev_priv->mm.shmem_info.migratepage = i915_gem_shrinker_migratepage;
>> +#endif
>
> If we avoid playing with page_private(), this initialization would go
> away, but the equivalent be done near the call to i915_gem_object_init().
>
Agree.

Best regards
Akash

>>  }
>>
>>  /**
>> --
>> 1.9.2

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH 2/2] drm/i915: Make GPU pages movable

[akash goel]

On Thu, Nov 10, 2016 at 1:00 PM, Goel, Akash <akash.goel@intel.com> wrote:
>
>
> On 11/10/2016 12:09 PM, Hugh Dickins wrote:
>>
>> On Fri, 4 Nov 2016, akash.goel@intel.com wrote:
>>>
>>> From: Chris Wilson <chris@chris-wilson.co.uk>
>>>
>>> On a long run of more than 2-3 days, physical memory tends to get
>>> fragmented severely, which considerably slows down the system. In such a
>>> scenario, the shrinker is also unable to help as lack of memory is not
>>> the actual problem, since it has been observed that there are enough free
>>> pages of 0 order. This also manifests itself when an indiviual zone in
>>> the mm runs out of pages and if we cannot migrate pages between zones,
>>> the kernel hits an out-of-memory even though there are free pages (and
>>> often all of swap) available.
>>>
>>> To address the issue of external fragementation, kernel does a compaction
>>> (which involves migration of pages) but it's efficacy depends upon how
>>> many pages are marked as MOVABLE, as only those pages can be migrated.
>>>
>>> Currently the backing pages for GPU buffers are allocated from shmemfs
>>> with GFP_RECLAIMABLE flag, in units of 4KB pages.  In the case of limited
>>> swap space, it may not be possible always to reclaim or swap-out pages of
>>> all the inactive objects, to make way for free space allowing formation
>>> of higher order groups of physically-contiguous pages on compaction.
>>>
>>> Just marking the GPU pages as MOVABLE will not suffice, as i915.ko has to
>>> pin the pages if they are in use by GPU, which will prevent their
>>> migration. So the migratepage callback in shmem is also hooked up to get
>>> a notification when kernel initiates the page migration. On the
>>> notification, i915.ko appropriately unpin the pages.  With this we can
>>> effectively mark the GPU pages as MOVABLE and hence mitigate the
>>> fragmentation problem.
>>>
>>> v2:
>>>  - Rename the migration routine to gem_shrink_migratepage, move it to the
>>>    shrinker file, and use the existing constructs (Chris)
>>>  - To cleanup, add a new helper function to encapsulate all page
>>> migration
>>>    skip conditions (Chris)
>>>  - Add a new local helper function in shrinker file, for dropping the
>>>    backing pages, and call the same from gem_shrink() also (Chris)
>>>
>>> v3:
>>>  - Fix/invert the check on the return value of unsafe_drop_pages (Chris)
>>>
>>> v4:
>>>  - Minor tidy
>>>
>>> v5:
>>>  - Fix unsafe usage of unsafe_drop_pages()
>>>  - Rebase onto vmap-notifier
>>>
>>> v6:
>>> - Remove i915_gem_object_get/put across unsafe_drop_pages() as with
>>>   struct_mutex protection object can't disappear. (Chris)
>>>
>>> Testcase: igt/gem_shrink
>>> Bugzilla: (e.g.) https://bugs.freedesktop.org/show_bug.cgi?id=90254
>>> Cc: Hugh Dickins <hughd@google.com>
>>> Cc: linux-mm@kvack.org
>>> Signed-off-by: Sourab Gupta <sourab.gupta@intel.com>
>>> Signed-off-by: Akash Goel <akash.goel@intel.com>
>>> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
>>> Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
>>> Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
>>
>>
>> I'm confused!  But perhaps it's gone around and around between you all,
>> I'm not sure what the rules are then.  I think this sequence implies
>> that Sourab wrote it originally, then Akash and Chris passed it on
>> with refinements - but then Chris wouldn't add Reviewed-by.
>>
> Thank you very much for the review and sorry for all the needless confusion.
>
> Chris actually conceived the patches and prepared an initial version of them
> (hence he is the Author).
> I & Sourab did the further refinements and fixed issues (all those
> page_private stuff).
> Chris then reviewed the final patch and also recently did a rebase for it.
>
>
>>> ---
>>>  drivers/gpu/drm/i915/i915_drv.h          |   2 +
>>>  drivers/gpu/drm/i915/i915_gem.c          |   9 ++-
>>>  drivers/gpu/drm/i915/i915_gem_shrinker.c | 132
>>> +++++++++++++++++++++++++++++++
>>>  3 files changed, 142 insertions(+), 1 deletion(-)
>>>
snip
>>
>>> @@ -4185,6 +4189,8 @@ struct drm_i915_gem_object *
>>>                 goto fail;
>>>
>>>         mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
>>> +       if (IS_ENABLED(MIGRATION))
>>> +               mask |= __GFP_MOVABLE;
>>
>>
>> I was going to suggest just make that unconditional,
>>         mask = GFP_HIGHUSER_MOVABLE | __GFP_RECLAIMABLE;
>>
>> But then I wondered what that __GFP_RECLAIMABLE actually achieves?
>> These pages are already __GFP_RECLAIM (inside GFP_HIGHUSER) and on
>> the LRU.  It affects gfpflags_to_migratetype(), but I'm not familiar
>> with what that different migratetype will end up doing.
>>
>
> Will check for this.
>

The anti-fragmentation technique used by kernel is based on the idea
of grouping pages with identical mobility (UNMOVABLE, RECLAIMABLE,
MOVABLE) together.
__GFP_RECLAIMABLE, like  __GFP_MOVABLE, specifies the
mobility/migration type of the page and serves a different purpose
than __GFP_RECLAIM.

Also as per the below snippet from gfpflags_to_migratetype(), looks
like __GFP_MOVABLE &  __GFP_RECLAIMABLE can't be used together, which
makes sense.
/* Convert GFP flags to their corresponding migrate type */
#define GFP_MOVABLE_MASK (__GFP_RECLAIMABLE | __GFP_MOVABLE)
static inline int gfpflags_to_migratetype(const gfp_t gfp_flags)
{
        VM_WARN_ON((gfp_flags & GFP_MOVABLE_MASK) == GFP_MOVABLE_MASK);
.....

So probably would need to update the mask like this,
       mask = GFP_HIGHUSER;
       if (IS_ENABLED(MIGRATION))
             mask |= __GFP_MOVABLE;
       else
             mask |=  __GFP_RECLAIMABLE;

Please kindly let us know if this looks fine to you or not.

Best regards
Akash

>
>>>         if (IS_CRESTLINE(dev_priv) || IS_BROADWATER(dev_priv)) {
>>>                 /* 965gm cannot relocate objects above 4GiB. */
>>>                 mask &= ~__GFP_HIGHMEM;
>>> @@ -4193,6 +4199,7 @@ struct drm_i915_gem_object *
>>>
>>>         mapping = obj->base.filp->f_mapping;
>>>         mapping_set_gfp_mask(mapping, mask);
>>> +       shmem_set_dev_info(mapping, &dev_priv->mm.shmem_info);
>>>
>>>         i915_gem_object_init(obj, &i915_gem_object_ops);
>>>
>>>  }
>>>
>>>  /**
>>> --
>>> 1.9.2

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH 2/2] drm/i915: Make GPU pages movable

[Hugh Dickins]

On Mon, 14 Nov 2016, akash goel wrote:
> On Thu, Nov 10, 2016 at 1:00 PM, Goel, Akash <akash.goel@intel.com> wrote:
> > On 11/10/2016 12:09 PM, Hugh Dickins wrote:
> >> On Fri, 4 Nov 2016, akash.goel@intel.com wrote:
> >>> @@ -4185,6 +4189,8 @@ struct drm_i915_gem_object *
> >>>
> >>>         mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
> >>> +       if (IS_ENABLED(MIGRATION))

Oh, I knew I'd seen a line like that recently, and it was bugging me
that I ought to search my mailboxes for it; but now I'm glad to find
it again.  If that condition stays, it would really need to say
              if (IS_ENABLED(CONFIG_MIGRATION))
wouldn't it?

> >>> +               mask |= __GFP_MOVABLE;
> >>
> >>
> >> I was going to suggest just make that unconditional,
> >>         mask = GFP_HIGHUSER_MOVABLE | __GFP_RECLAIMABLE;
> >>
> >> But then I wondered what that __GFP_RECLAIMABLE actually achieves?
> >> These pages are already __GFP_RECLAIM (inside GFP_HIGHUSER) and on
> >> the LRU.  It affects gfpflags_to_migratetype(), but I'm not familiar
> >> with what that different migratetype will end up doing.
> >>
> >
> > Will check for this.
> >
> 
> The anti-fragmentation technique used by kernel is based on the idea
> of grouping pages with identical mobility (UNMOVABLE, RECLAIMABLE,
> MOVABLE) together.

Yes.

> __GFP_RECLAIMABLE, like  __GFP_MOVABLE, specifies the
> mobility/migration type of the page and serves a different purpose
> than __GFP_RECLAIM.

Yes, I was wrong to mention __GFP_RECLAIM above: it describes what
to do when in difficulty allocating a page, but says nothing at all
about the nature of the page to be allocated.

> 
> Also as per the below snippet from gfpflags_to_migratetype(), looks
> like __GFP_MOVABLE &  __GFP_RECLAIMABLE can't be used together, which
> makes sense.
> /* Convert GFP flags to their corresponding migrate type */
> #define GFP_MOVABLE_MASK (__GFP_RECLAIMABLE | __GFP_MOVABLE)
> static inline int gfpflags_to_migratetype(const gfp_t gfp_flags)
> {
>         VM_WARN_ON((gfp_flags & GFP_MOVABLE_MASK) == GFP_MOVABLE_MASK);
> .....

You're right, that does exclude them from being used together.  And it
makes sense inasmuch as they're expected to be appled to quite different
uses of a page (lru pages versus slab pages).

The comment on __GFP_MOVABLE says "or can be reclaimed"; and
the comment on __GFP_RECLAIMABLE says "used for slab allocations...".
Though it does not say "used for allocations not put on a reclaimable
lru", I think that is the intention; whereas shmem allocations are put
on a reclaimable lru (though they might need your shrinker to unpin them).

> 
> So probably would need to update the mask like this,
>        mask = GFP_HIGHUSER;
>        if (IS_ENABLED(MIGRATION))
>              mask |= __GFP_MOVABLE;
>        else
>              mask |=  __GFP_RECLAIMABLE;
> 
> Please kindly let us know if this looks fine to you or not.

Thanks for looking into it more deeply.  You leave me thinking that
it should simply say

        mask = GFP_HIGHUSER_MOVABLE;

Which is the default anyway, but it then has the Crestline+Broadwater
condition to modify the mask further, so it's probably clearest to
leave the mask = GFP_HIGHUSER_MOVABLE explicit.

GFP_HIGHUSER_MOVABLE is used in many places, and includes __GFP_MOVABLE
without any condition on CONFIG_MIGRATION - because the migratetype is
irrelevant if there is no migration, I presume.

Would you lose something by not or'ing in __GFP_RECLAIMABLE when
CONFIG_MIGRATION=n?  No, because __GFP_RECLAIMABLE is not used for
anything but the migratetype, and the migratetype is then irrelevant.
(I didn't study the code closely enough to say whether the grouping
can still happen even when migration is disabled, but even if it
does still happen, I can't see that it would have any benefit.)

Hugh

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH 2/2] drm/i915: Make GPU pages movable

[akash goel]

On Wed, Nov 16, 2016 at 6:55 AM, Hugh Dickins <hughd@google.com> wrote:
> On Mon, 14 Nov 2016, akash goel wrote:
>> On Thu, Nov 10, 2016 at 1:00 PM, Goel, Akash <akash.goel@intel.com> wrote:
>> > On 11/10/2016 12:09 PM, Hugh Dickins wrote:
>> >> On Fri, 4 Nov 2016, akash.goel@intel.com wrote:
>> >>> @@ -4185,6 +4189,8 @@ struct drm_i915_gem_object *
>> >>>
>> >>>         mask = GFP_HIGHUSER | __GFP_RECLAIMABLE;
>> >>> +       if (IS_ENABLED(MIGRATION))
>
> Oh, I knew I'd seen a line like that recently, and it was bugging me
> that I ought to search my mailboxes for it; but now I'm glad to find
> it again.  If that condition stays, it would really need to say
>               if (IS_ENABLED(CONFIG_MIGRATION))
> wouldn't it?
>
Sorry this was a blooper, should have been
             if (IS_ENABLED(CONFIG_MIGRATION))

>> >>> +               mask |= __GFP_MOVABLE;
>> >>
>> >>
>> >> I was going to suggest just make that unconditional,
>> >>         mask = GFP_HIGHUSER_MOVABLE | __GFP_RECLAIMABLE;
>> >>
>> >> But then I wondered what that __GFP_RECLAIMABLE actually achieves?
>> >> These pages are already __GFP_RECLAIM (inside GFP_HIGHUSER) and on
>> >> the LRU.  It affects gfpflags_to_migratetype(), but I'm not familiar
>> >> with what that different migratetype will end up doing.
>> >>
>> >
>> > Will check for this.
>> >
>>
>> The anti-fragmentation technique used by kernel is based on the idea
>> of grouping pages with identical mobility (UNMOVABLE, RECLAIMABLE,
>> MOVABLE) together.
>
> Yes.
>
>> __GFP_RECLAIMABLE, like  __GFP_MOVABLE, specifies the
>> mobility/migration type of the page and serves a different purpose
>> than __GFP_RECLAIM.
>
> Yes, I was wrong to mention __GFP_RECLAIM above: it describes what
> to do when in difficulty allocating a page, but says nothing at all
> about the nature of the page to be allocated.
>
Right, nicely phrased, thanks.

>>
>> Also as per the below snippet from gfpflags_to_migratetype(), looks
>> like __GFP_MOVABLE &  __GFP_RECLAIMABLE can't be used together, which
>> makes sense.
>> /* Convert GFP flags to their corresponding migrate type */
>> #define GFP_MOVABLE_MASK (__GFP_RECLAIMABLE | __GFP_MOVABLE)
>> static inline int gfpflags_to_migratetype(const gfp_t gfp_flags)
>> {
>>         VM_WARN_ON((gfp_flags & GFP_MOVABLE_MASK) == GFP_MOVABLE_MASK);
>> .....
>
> You're right, that does exclude them from being used together.  And it
> makes sense inasmuch as they're expected to be appled to quite different
> uses of a page (lru pages versus slab pages).
>
> The comment on __GFP_MOVABLE says "or can be reclaimed"; and
> the comment on __GFP_RECLAIMABLE says "used for slab allocations...".
> Though it does not say "used for allocations not put on a reclaimable
> lru", I think that is the intention; whereas shmem allocations are put
> on a reclaimable lru (though they might need your shrinker to unpin them).
>

As per my understanding both  __GFP_MOVABLE & __GFP_RECLAIMABLE type
pages would get added to the LRU list for reclaiming.
Irrespective of whether a shmem page is allocated as __GFP_MOVABLE
type or  __GFP_RECLAIMABLE type, it will be added to the LRU list.

>>
>> So probably would need to update the mask like this,
>>        mask = GFP_HIGHUSER;
>>        if (IS_ENABLED(MIGRATION))
>>              mask |= __GFP_MOVABLE;
>>        else
>>              mask |=  __GFP_RECLAIMABLE;
>>
>> Please kindly let us know if this looks fine to you or not.
>
> Thanks for looking into it more deeply.  You leave me thinking that
> it should simply say
>
>         mask = GFP_HIGHUSER_MOVABLE;
>
> Which is the default anyway, but it then has the Crestline+Broadwater
> condition to modify the mask further, so it's probably clearest to
> leave the mask = GFP_HIGHUSER_MOVABLE explicit.
>
> GFP_HIGHUSER_MOVABLE is used in many places, and includes __GFP_MOVABLE
> without any condition on CONFIG_MIGRATION - because the migratetype is
> irrelevant if there is no migration, I presume.
>
> Would you lose something by not or'ing in __GFP_RECLAIMABLE when
> CONFIG_MIGRATION=n?  No, because __GFP_RECLAIMABLE is not used for
> anything but the migratetype, and the migratetype is then irrelevant.
> (I didn't study the code closely enough to say whether the grouping
> can still happen even when migration is disabled, but even if it
> does still happen, I can't see that it would have any benefit.)
>
The freelist, for a particular order,  in buddy allocator is always
organized based on the migrate type.
from <mmzone.h>
struct free_area {
       struct list_head free_list[MIGRATE_TYPES];
        unsigned long nr_free;
};
And the page grouping, based on migrate type, is also always done by the kernel.

Its just that when CONFIG_MIGRATION=n, the actual migration does not
take place, leveraging the page grouping, so no compaction effectively
happens.
I think even without the migration, the page grouping could still be
beneficial in limiting the fragmentation to some extent, as at least
the non-movable pages will not be scattered/dispersed randomly.
But yes when CONFIG_MIGRATION=n,  __GFP_MOVABLE & __GFP_RECLAIMABLE
would be equivalent of each other, as then reclaiming is the only way
to facilitate the formation of higher order free page.
So as you suggested looks like we can simply set
              mask = GFP_HIGHUSER_MOVABLE;

Many thanks for all your inputs.

Best regards
Akash

> Hugh

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>