Re: [PATCH v2] drm/i915: Fix page cleanup on DMA remap failure
On Wed, Jan 22, 2025 at 10:07 PM Srinivas, Vidya
wrote:
>
> Hello Brian, Many thanks for the fix. I am adding my tested-by.
> Tested-by: Vidya Srinivas
Thanks for testing Vidya.
Can we get a maintainer to take a look?
>
>
> > -Original Message-
> > From: Brian Geffon
> > Sent: 16 January 2025 21:24
> > To: intel-...@lists.freedesktop.org
> > Cc: Wilson, Chris P ; Saarinen, Jani
> > ; Mistat, Tomasz ;
> > Srinivas, Vidya ; ville.syrj...@linux.intel.com;
> > jani.nik...@linux.intel.com; linux-ker...@vger.kernel.org; dri-
> > de...@lists.freedesktop.org; Joonas Lahtinen
> > ; Brian Geffon ;
> > sta...@vger.kernel.org; Tomasz Figa
> > Subject: [PATCH v2] drm/i915: Fix page cleanup on DMA remap failure
> >
> > When converting to folios the cleanup path of shmem_get_pages() was
> > missed. When a DMA remap fails and the max segment size is greater than
> > PAGE_SIZE it will attempt to retry the remap with a PAGE_SIZEd segment size.
> > The cleanup code isn't properly using the folio apis and as a result isn't
> > handling compound pages correctly.
> >
> > v1 -> v2:
> > (Ville) Fixed locations where we were not clearing mapping unevictable.
> >
> > Cc: sta...@vger.kernel.org
> > Cc: Ville Syrjala
> > Cc: Vidya Srinivas
> > Link: https://gitlab.freedesktop.org/drm/i915/kernel/-/issues/13487
> > Link: https://lore.kernel.org/lkml/20250116135636.410164-1-
> > bgef...@google.com/
> > Fixes: 0b62af28f249 ("i915: convert shmem_sg_free_table() to use a
> > folio_batch")
> > Signed-off-by: Brian Geffon
> > Suggested-by: Tomasz Figa
> > ---
> > drivers/gpu/drm/i915/gem/i915_gem_object.h | 3 +--
> > drivers/gpu/drm/i915/gem/i915_gem_shmem.c | 23 +-
> > drivers/gpu/drm/i915/gem/i915_gem_ttm.c| 7 ---
> > 3 files changed, 14 insertions(+), 19 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.h
> > b/drivers/gpu/drm/i915/gem/i915_gem_object.h
> > index 3dc61cbd2e11..0f122a12d4a5 100644
> > --- a/drivers/gpu/drm/i915/gem/i915_gem_object.h
> > +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.h
> > @@ -843,8 +843,7 @@ int shmem_sg_alloc_table(struct drm_i915_private
> > *i915, struct sg_table *st,
> >size_t size, struct intel_memory_region *mr,
> >struct address_space *mapping,
> >unsigned int max_segment);
> > -void shmem_sg_free_table(struct sg_table *st, struct address_space
> > *mapping,
> > - bool dirty, bool backup);
> > +void shmem_sg_free_table(struct sg_table *st, bool dirty, bool backup);
> > void __shmem_writeback(size_t size, struct address_space *mapping);
> >
> > #ifdef CONFIG_MMU_NOTIFIER
> > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > index fe69f2c8527d..b320d9dfd6d3 100644
> > --- a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > +++ b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > @@ -29,16 +29,13 @@ static void check_release_folio_batch(struct
> > folio_batch *fbatch)
> > cond_resched();
> > }
> >
> > -void shmem_sg_free_table(struct sg_table *st, struct address_space
> > *mapping,
> > - bool dirty, bool backup)
> > +void shmem_sg_free_table(struct sg_table *st, bool dirty, bool backup)
> > {
> > struct sgt_iter sgt_iter;
> > struct folio_batch fbatch;
> > struct folio *last = NULL;
> > struct page *page;
> >
> > - mapping_clear_unevictable(mapping);
> > -
> > folio_batch_init(&fbatch);
> > for_each_sgt_page(page, sgt_iter, st) {
> > struct folio *folio = page_folio(page); @@ -180,10 +177,10
> > @@ int shmem_sg_alloc_table(struct drm_i915_private *i915, struct sg_table
> > *st,
> > return 0;
> > err_sg:
> > sg_mark_end(sg);
> > + mapping_clear_unevictable(mapping);
> > if (sg != st->sgl) {
> > - shmem_sg_free_table(st, mapping, false, false);
> > + shmem_sg_free_table(st, false, false);
> > } else {
> > - mapping_clear_unevictable(mapping);
> > sg_free_table(st);
> > }
> >
> > @@ -209,8 +206,6 @@ static int shmem_get_pages(struct
> > drm_i915_gem_object *obj)
> > struct address_space *mapping = obj->base.filp->f_mapping;
> > unsigned int m
Re: [PATCH v2] drm/i915: Fix page cleanup on DMA remap failure
On Mon, Jan 27, 2025 at 1:47 PM Ville Syrjälä
wrote:
>
> On Thu, Jan 16, 2025 at 10:53:40AM -0500, Brian Geffon wrote:
> > When converting to folios the cleanup path of shmem_get_pages() was
> > missed. When a DMA remap fails and the max segment size is greater than
> > PAGE_SIZE it will attempt to retry the remap with a PAGE_SIZEd segment
> > size. The cleanup code isn't properly using the folio apis and as a
> > result isn't handling compound pages correctly.
> >
> > v1 -> v2:
> > (Ville) Fixed locations where we were not clearing mapping unevictable.
> >
> > Cc: sta...@vger.kernel.org
> > Cc: Ville Syrjala
> > Cc: Vidya Srinivas
> > Link: https://gitlab.freedesktop.org/drm/i915/kernel/-/issues/13487
> > Link:
> > https://lore.kernel.org/lkml/20250116135636.410164-1-bgef...@google.com/
> > Fixes: 0b62af28f249 ("i915: convert shmem_sg_free_table() to use a
> > folio_batch")
> > Signed-off-by: Brian Geffon
> > Suggested-by: Tomasz Figa
> > ---
> > drivers/gpu/drm/i915/gem/i915_gem_object.h | 3 +--
> > drivers/gpu/drm/i915/gem/i915_gem_shmem.c | 23 +-
> > drivers/gpu/drm/i915/gem/i915_gem_ttm.c| 7 ---
> > 3 files changed, 14 insertions(+), 19 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.h
> > b/drivers/gpu/drm/i915/gem/i915_gem_object.h
> > index 3dc61cbd2e11..0f122a12d4a5 100644
> > --- a/drivers/gpu/drm/i915/gem/i915_gem_object.h
> > +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.h
> > @@ -843,8 +843,7 @@ int shmem_sg_alloc_table(struct drm_i915_private *i915,
> > struct sg_table *st,
> >size_t size, struct intel_memory_region *mr,
> >struct address_space *mapping,
> >unsigned int max_segment);
> > -void shmem_sg_free_table(struct sg_table *st, struct address_space
> > *mapping,
> > - bool dirty, bool backup);
> > +void shmem_sg_free_table(struct sg_table *st, bool dirty, bool backup);
> > void __shmem_writeback(size_t size, struct address_space *mapping);
> >
> > #ifdef CONFIG_MMU_NOTIFIER
> > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > index fe69f2c8527d..b320d9dfd6d3 100644
> > --- a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > +++ b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > @@ -29,16 +29,13 @@ static void check_release_folio_batch(struct
> > folio_batch *fbatch)
> > cond_resched();
> > }
> >
> > -void shmem_sg_free_table(struct sg_table *st, struct address_space
> > *mapping,
> > - bool dirty, bool backup)
> > +void shmem_sg_free_table(struct sg_table *st, bool dirty, bool backup)
>
> This still makes the alloc vs. free completely asymmetric.
> This is not what we want because it just makes it very easy
> to make it mistake in the caller.
>
> I think the correct fix is to simply call the current
> shmem_sg_free_table() from the now broken failure path.
> mapping_{set,clear}_unevictable() just seems to be some
> bit operation so AFAICS the slight ping-pong should be
> inconsequential.
Ok, I guess that's true, it'll be set unevictable again in the call to
shmem_sg_alloc_table() after it jumps back to rebuild_st. I'll send a
v3 which should then become a 1 line change.
>
> > {
> > struct sgt_iter sgt_iter;
> > struct folio_batch fbatch;
> > struct folio *last = NULL;
> > struct page *page;
> >
> > - mapping_clear_unevictable(mapping);
> > -
> > folio_batch_init(&fbatch);
> > for_each_sgt_page(page, sgt_iter, st) {
> > struct folio *folio = page_folio(page);
> > @@ -180,10 +177,10 @@ int shmem_sg_alloc_table(struct drm_i915_private
> > *i915, struct sg_table *st,
> > return 0;
> > err_sg:
> > sg_mark_end(sg);
> > + mapping_clear_unevictable(mapping);
> > if (sg != st->sgl) {
> > - shmem_sg_free_table(st, mapping, false, false);
> > + shmem_sg_free_table(st, false, false);
> > } else {
> > - mapping_clear_unevictable(mapping);
> > sg_free_table(st);
> > }
> >
> > @@ -209,8 +206,6 @@ static int shmem_get_pages(struct drm_i915_gem_object
> > *obj)
> > struct address_space *mapping = obj->base.filp->f_mapping;
> > unsigned int max_segment = i915_sg_segment_size(i915-&
[PATCH v3] drm/i915: Fix page cleanup on DMA remap failure
When converting to folios the cleanup path of shmem_get_pages() was
missed. When a DMA remap fails and the max segment size is greater than
PAGE_SIZE it will attempt to retry the remap with a PAGE_SIZEd segment
size. The cleanup code isn't properly using the folio apis and as a
result isn't handling compound pages correctly.
v2 -> v3:
(Ville) Just use shmem_sg_free_table() as-is in the failure path of
shmem_get_pages(). shmem_sg_free_table() will clear mapping unevictable
but it will be reset when it retries in shmem_sg_alloc_table().
v1 -> v2:
(Ville) Fixed locations where we were not clearing mapping unevictable.
Cc: sta...@vger.kernel.org
Cc: Ville Syrjala
Cc: Vidya Srinivas
Link: https://gitlab.freedesktop.org/drm/i915/kernel/-/issues/13487
Link: https://lore.kernel.org/lkml/20250116135636.410164-1-bgef...@google.com/
Fixes: 0b62af28f249 ("i915: convert shmem_sg_free_table() to use a folio_batch")
Signed-off-by: Brian Geffon
Suggested-by: Tomasz Figa
---
drivers/gpu/drm/i915/gem/i915_gem_shmem.c | 6 +-
1 file changed, 1 insertion(+), 5 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
index fe69f2c8527d..ae3343c81a64 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
@@ -209,8 +209,6 @@ static int shmem_get_pages(struct drm_i915_gem_object *obj)
struct address_space *mapping = obj->base.filp->f_mapping;
unsigned int max_segment = i915_sg_segment_size(i915->drm.dev);
struct sg_table *st;
- struct sgt_iter sgt_iter;
- struct page *page;
int ret;
/*
@@ -239,9 +237,7 @@ static int shmem_get_pages(struct drm_i915_gem_object *obj)
* for PAGE_SIZE chunks instead may be helpful.
*/
if (max_segment > PAGE_SIZE) {
- for_each_sgt_page(page, sgt_iter, st)
- put_page(page);
- sg_free_table(st);
+ shmem_sg_free_table(st, mapping, false, false);
kfree(st);
max_segment = PAGE_SIZE;
--
2.48.1.262.g85cc9f2d1e-goog
[PATCH v2] drm/i915: Fix page cleanup on DMA remap failure
When converting to folios the cleanup path of shmem_get_pages() was
missed. When a DMA remap fails and the max segment size is greater than
PAGE_SIZE it will attempt to retry the remap with a PAGE_SIZEd segment
size. The cleanup code isn't properly using the folio apis and as a
result isn't handling compound pages correctly.
v1 -> v2:
(Ville) Fixed locations where we were not clearing mapping unevictable.
Cc: sta...@vger.kernel.org
Cc: Ville Syrjala
Cc: Vidya Srinivas
Link: https://gitlab.freedesktop.org/drm/i915/kernel/-/issues/13487
Link: https://lore.kernel.org/lkml/20250116135636.410164-1-bgef...@google.com/
Fixes: 0b62af28f249 ("i915: convert shmem_sg_free_table() to use a folio_batch")
Signed-off-by: Brian Geffon
Suggested-by: Tomasz Figa
---
drivers/gpu/drm/i915/gem/i915_gem_object.h | 3 +--
drivers/gpu/drm/i915/gem/i915_gem_shmem.c | 23 +-
drivers/gpu/drm/i915/gem/i915_gem_ttm.c| 7 ---
3 files changed, 14 insertions(+), 19 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.h
b/drivers/gpu/drm/i915/gem/i915_gem_object.h
index 3dc61cbd2e11..0f122a12d4a5 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_object.h
+++ b/drivers/gpu/drm/i915/gem/i915_gem_object.h
@@ -843,8 +843,7 @@ int shmem_sg_alloc_table(struct drm_i915_private *i915,
struct sg_table *st,
size_t size, struct intel_memory_region *mr,
struct address_space *mapping,
unsigned int max_segment);
-void shmem_sg_free_table(struct sg_table *st, struct address_space *mapping,
-bool dirty, bool backup);
+void shmem_sg_free_table(struct sg_table *st, bool dirty, bool backup);
void __shmem_writeback(size_t size, struct address_space *mapping);
#ifdef CONFIG_MMU_NOTIFIER
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
index fe69f2c8527d..b320d9dfd6d3 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
@@ -29,16 +29,13 @@ static void check_release_folio_batch(struct folio_batch
*fbatch)
cond_resched();
}
-void shmem_sg_free_table(struct sg_table *st, struct address_space *mapping,
-bool dirty, bool backup)
+void shmem_sg_free_table(struct sg_table *st, bool dirty, bool backup)
{
struct sgt_iter sgt_iter;
struct folio_batch fbatch;
struct folio *last = NULL;
struct page *page;
- mapping_clear_unevictable(mapping);
-
folio_batch_init(&fbatch);
for_each_sgt_page(page, sgt_iter, st) {
struct folio *folio = page_folio(page);
@@ -180,10 +177,10 @@ int shmem_sg_alloc_table(struct drm_i915_private *i915,
struct sg_table *st,
return 0;
err_sg:
sg_mark_end(sg);
+ mapping_clear_unevictable(mapping);
if (sg != st->sgl) {
- shmem_sg_free_table(st, mapping, false, false);
+ shmem_sg_free_table(st, false, false);
} else {
- mapping_clear_unevictable(mapping);
sg_free_table(st);
}
@@ -209,8 +206,6 @@ static int shmem_get_pages(struct drm_i915_gem_object *obj)
struct address_space *mapping = obj->base.filp->f_mapping;
unsigned int max_segment = i915_sg_segment_size(i915->drm.dev);
struct sg_table *st;
- struct sgt_iter sgt_iter;
- struct page *page;
int ret;
/*
@@ -239,9 +234,8 @@ static int shmem_get_pages(struct drm_i915_gem_object *obj)
* for PAGE_SIZE chunks instead may be helpful.
*/
if (max_segment > PAGE_SIZE) {
- for_each_sgt_page(page, sgt_iter, st)
- put_page(page);
- sg_free_table(st);
+ /* Leave the mapping unevictable while we retry */
+ shmem_sg_free_table(st, false, false);
kfree(st);
max_segment = PAGE_SIZE;
@@ -265,7 +259,8 @@ static int shmem_get_pages(struct drm_i915_gem_object *obj)
return 0;
err_pages:
- shmem_sg_free_table(st, mapping, false, false);
+ mapping_clear_unevictable(mapping);
+ shmem_sg_free_table(st, false, false);
/*
* shmemfs first checks if there is enough memory to allocate the page
* and reports ENOSPC should there be insufficient, along with the usual
@@ -402,8 +397,8 @@ void i915_gem_object_put_pages_shmem(struct
drm_i915_gem_object *obj, struct sg_
if (i915_gem_object_needs_bit17_swizzle(obj))
i915_gem_object_save_bit_17_swizzle(obj, pages);
- shmem_sg_free_table(pages, file_inode(obj->base.filp)->i_mapping,
- obj->mm.dirty, obj->mm.madv == I915_MADV_WILLNEED);
+ mapping_clear_unevictable(file_inode(obj->base.
Re: [PATCH] drm/i915: Fix page cleanup on DMA remap failure
On Thu, Jan 16, 2025 at 9:24 AM Ville Syrjälä
wrote:
>
> On Thu, Jan 16, 2025 at 08:56:36AM -0500, Brian Geffon wrote:
> > When converting to folios the cleanup path of shmem_get_pages() was
> > missed. When a DMA remap fails and the max segment size is greater than
> > PAGE_SIZE it will attempt to retry the remap with a PAGE_SIZEd segment
> > size. The cleanup code isn't properly using the folio apis and as a
> > result isn't handling compound pages correctly.
> >
> > Link: https://gitlab.freedesktop.org/drm/i915/kernel/-/issues/13487
> > Fixes: 0b62af28f249 ("i915: convert shmem_sg_free_table() to use a
> > folio_batch")
> > Signed-off-by: Brian Geffon
> > Suggested-by: Tomasz Figa
> > ---
> > drivers/gpu/drm/i915/gem/i915_gem_shmem.c | 13 +
> > 1 file changed, 5 insertions(+), 8 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > index fe69f2c8527d..02ddab5bf5c0 100644
> > --- a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > +++ b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > @@ -37,8 +37,6 @@ void shmem_sg_free_table(struct sg_table *st, struct
> > address_space *mapping,
> > struct folio *last = NULL;
> > struct page *page;
> >
> > - mapping_clear_unevictable(mapping);
> > -
>
> The assymmetry here between the alloc vs. free is a bit annoying.
> Maybe we can just keep this here?
If you want, I think this can also be fixed by something like the
following I believe.
Ultimately we don't want to put page on non-head pages in a compound
page. What do you think? If you like this better I can test and mail a v2.
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
index fe69f2c8527d..b79cd396e878 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
@@ -239,8 +239,14 @@ static int shmem_get_pages(struct drm_i915_gem_object *obj)
* for PAGE_SIZE chunks instead may be helpful.
*/
if (max_segment > PAGE_SIZE) {
- for_each_sgt_page(page, sgt_iter, st)
+ struct folio *last = NULL;
+ for_each_sgt_page(page, sgt_iter, st) {
+ struct folio *folio = page_folio(page);
+ if (folio == last)
+ continue;
+ last = folio;
put_page(page);
+ }
sg_free_table(st);
kfree(st);
--
2.48.0.rc2.279.g1de40edade-goog
>
> Or if avoiding the ping-pong actually mattes in the gtt prepare
> error case, then maybe we should rename this guy into
> __shmem_sg_free_table() without the mapping_clear_unevictable()
> and wrap it in a higher level shmem_sg_free_table() that does
> everything?
>
> > folio_batch_init(&fbatch);
> > for_each_sgt_page(page, sgt_iter, st) {
> > struct folio *folio = page_folio(page);
> > @@ -180,10 +178,10 @@ int shmem_sg_alloc_table(struct drm_i915_private
> > *i915, struct sg_table *st,
> > return 0;
> > err_sg:
> > sg_mark_end(sg);
> > + mapping_clear_unevictable(mapping);
> > if (sg != st->sgl) {
> > shmem_sg_free_table(st, mapping, false, false);
> > } else {
> > - mapping_clear_unevictable(mapping);
> > sg_free_table(st);
> > }
> >
> > @@ -209,8 +207,6 @@ static int shmem_get_pages(struct drm_i915_gem_object
> > *obj)
> > struct address_space *mapping = obj->base.filp->f_mapping;
> > unsigned int max_segment = i915_sg_segment_size(i915->drm.dev);
> > struct sg_table *st;
> > - struct sgt_iter sgt_iter;
> > - struct page *page;
> > int ret;
> >
> > /*
> > @@ -239,9 +235,8 @@ static int shmem_get_pages(struct drm_i915_gem_object
> > *obj)
> >* for PAGE_SIZE chunks instead may be helpful.
> >*/
> > if (max_segment > PAGE_SIZE) {
> > - for_each_sgt_page(page, sgt_iter, st)
> > - put_page(page);
> > - sg_free_table(st);
> > + /* Leave the mapping unevictable while we retry */
> > + shmem_sg_free_table(st, mapping, false, false);
> > kfree(st);
> >
> >
Re: [PATCH] drm/i915: Fix page cleanup on DMA remap failure
On Thu, Jan 16, 2025 at 9:24 AM Ville Syrjälä
wrote:
>
> On Thu, Jan 16, 2025 at 08:56:36AM -0500, Brian Geffon wrote:
> > When converting to folios the cleanup path of shmem_get_pages() was
> > missed. When a DMA remap fails and the max segment size is greater than
> > PAGE_SIZE it will attempt to retry the remap with a PAGE_SIZEd segment
> > size. The cleanup code isn't properly using the folio apis and as a
> > result isn't handling compound pages correctly.
> >
> > Link: https://gitlab.freedesktop.org/drm/i915/kernel/-/issues/13487
> > Fixes: 0b62af28f249 ("i915: convert shmem_sg_free_table() to use a
> > folio_batch")
> > Signed-off-by: Brian Geffon
> > Suggested-by: Tomasz Figa
> > ---
> > drivers/gpu/drm/i915/gem/i915_gem_shmem.c | 13 +
> > 1 file changed, 5 insertions(+), 8 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > index fe69f2c8527d..02ddab5bf5c0 100644
> > --- a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > +++ b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > @@ -37,8 +37,6 @@ void shmem_sg_free_table(struct sg_table *st, struct
> > address_space *mapping,
> > struct folio *last = NULL;
> > struct page *page;
> >
> > - mapping_clear_unevictable(mapping);
> > -
>
> The assymmetry here between the alloc vs. free is a bit annoying.
> Maybe we can just keep this here?
My thought on this was that freeing the sg table is orthogonal to
making the mapping evictable, and shmem_sg_free_table() can be
simplified to not even take the mapping as a parameter.
>
> Or if avoiding the ping-pong actually mattes in the gtt prepare
> error case, then maybe we should rename this guy into
> __shmem_sg_free_table() without the mapping_clear_unevictable()
> and wrap it in a higher level shmem_sg_free_table() that does
> everything?
>
> > folio_batch_init(&fbatch);
> > for_each_sgt_page(page, sgt_iter, st) {
> > struct folio *folio = page_folio(page);
> > @@ -180,10 +178,10 @@ int shmem_sg_alloc_table(struct drm_i915_private
> > *i915, struct sg_table *st,
> > return 0;
> > err_sg:
> > sg_mark_end(sg);
> > + mapping_clear_unevictable(mapping);
> > if (sg != st->sgl) {
> > shmem_sg_free_table(st, mapping, false, false);
> > } else {
> > - mapping_clear_unevictable(mapping);
> > sg_free_table(st);
> > }
> >
> > @@ -209,8 +207,6 @@ static int shmem_get_pages(struct drm_i915_gem_object
> > *obj)
> > struct address_space *mapping = obj->base.filp->f_mapping;
> > unsigned int max_segment = i915_sg_segment_size(i915->drm.dev);
> > struct sg_table *st;
> > - struct sgt_iter sgt_iter;
> > - struct page *page;
> > int ret;
> >
> > /*
> > @@ -239,9 +235,8 @@ static int shmem_get_pages(struct drm_i915_gem_object
> > *obj)
> >* for PAGE_SIZE chunks instead may be helpful.
> >*/
> > if (max_segment > PAGE_SIZE) {
> > - for_each_sgt_page(page, sgt_iter, st)
> > - put_page(page);
> > - sg_free_table(st);
> > + /* Leave the mapping unevictable while we retry */
> > + shmem_sg_free_table(st, mapping, false, false);
> > kfree(st);
> >
> > max_segment = PAGE_SIZE;
> > @@ -265,6 +260,7 @@ static int shmem_get_pages(struct drm_i915_gem_object
> > *obj)
> > return 0;
> >
> > err_pages:
> > + mapping_clear_unevictable(mapping);
> > shmem_sg_free_table(st, mapping, false, false);
> > /*
> >* shmemfs first checks if there is enough memory to allocate the page
> > @@ -402,6 +398,7 @@ void i915_gem_object_put_pages_shmem(struct
> > drm_i915_gem_object *obj, struct sg_
> > if (i915_gem_object_needs_bit17_swizzle(obj))
> > i915_gem_object_save_bit_17_swizzle(obj, pages);
> >
> > + mapping_clear_unevictable(file_inode(obj->base.filp)->i_mapping);
> > shmem_sg_free_table(pages, file_inode(obj->base.filp)->i_mapping,
> > obj->mm.dirty, obj->mm.madv ==
> > I915_MADV_WILLNEED);
> > kfree(pages);
> > --
> > 2.48.0.rc2.279.g1de40edade-goog
>
> --
> Ville Syrjälä
> Intel
Re: [PATCH] drm/i915: Fix page cleanup on DMA remap failure
On Thu, Jan 16, 2025 at 9:38 AM Ville Syrjälä
wrote:
>
> On Thu, Jan 16, 2025 at 04:24:26PM +0200, Ville Syrjälä wrote:
> > On Thu, Jan 16, 2025 at 08:56:36AM -0500, Brian Geffon wrote:
> > > When converting to folios the cleanup path of shmem_get_pages() was
> > > missed. When a DMA remap fails and the max segment size is greater than
> > > PAGE_SIZE it will attempt to retry the remap with a PAGE_SIZEd segment
> > > size. The cleanup code isn't properly using the folio apis and as a
> > > result isn't handling compound pages correctly.
> > >
> > > Link: https://gitlab.freedesktop.org/drm/i915/kernel/-/issues/13487
> > > Fixes: 0b62af28f249 ("i915: convert shmem_sg_free_table() to use a
> > > folio_batch")
> > > Signed-off-by: Brian Geffon
> > > Suggested-by: Tomasz Figa
> > > ---
> > > drivers/gpu/drm/i915/gem/i915_gem_shmem.c | 13 +
> > > 1 file changed, 5 insertions(+), 8 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > > b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > > index fe69f2c8527d..02ddab5bf5c0 100644
> > > --- a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > > +++ b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > > @@ -37,8 +37,6 @@ void shmem_sg_free_table(struct sg_table *st, struct
> > > address_space *mapping,
> > > struct folio *last = NULL;
> > > struct page *page;
> > >
> > > - mapping_clear_unevictable(mapping);
> > > -
> >
> > The assymmetry here between the alloc vs. free is a bit annoying.
> > Maybe we can just keep this here?
>
> Actually, I guess it's a bit more than just annoying since
> shmem_sg_free_table() is actually used from outside i915_gem_shmem.c
> as well.
You're correct, this was a bad oversight on my part. This patch is
going to require a v2 regardless, how do you feel about just changing
shmem_sg_free_table() to not accept a mapping given that managing the
mapping is really orthogonal to cleaning up the sg table.
>
> >
> > Or if avoiding the ping-pong actually mattes in the gtt prepare
> > error case, then maybe we should rename this guy into
> > __shmem_sg_free_table() without the mapping_clear_unevictable()
> > and wrap it in a higher level shmem_sg_free_table() that does
> > everything?
> >
> > > folio_batch_init(&fbatch);
> > > for_each_sgt_page(page, sgt_iter, st) {
> > > struct folio *folio = page_folio(page);
> > > @@ -180,10 +178,10 @@ int shmem_sg_alloc_table(struct drm_i915_private
> > > *i915, struct sg_table *st,
> > > return 0;
> > > err_sg:
> > > sg_mark_end(sg);
> > > + mapping_clear_unevictable(mapping);
> > > if (sg != st->sgl) {
> > > shmem_sg_free_table(st, mapping, false, false);
> > > } else {
> > > - mapping_clear_unevictable(mapping);
> > > sg_free_table(st);
> > > }
> > >
> > > @@ -209,8 +207,6 @@ static int shmem_get_pages(struct drm_i915_gem_object
> > > *obj)
> > > struct address_space *mapping = obj->base.filp->f_mapping;
> > > unsigned int max_segment = i915_sg_segment_size(i915->drm.dev);
> > > struct sg_table *st;
> > > - struct sgt_iter sgt_iter;
> > > - struct page *page;
> > > int ret;
> > >
> > > /*
> > > @@ -239,9 +235,8 @@ static int shmem_get_pages(struct drm_i915_gem_object
> > > *obj)
> > > * for PAGE_SIZE chunks instead may be helpful.
> > > */
> > > if (max_segment > PAGE_SIZE) {
> > > - for_each_sgt_page(page, sgt_iter, st)
> > > - put_page(page);
> > > - sg_free_table(st);
> > > + /* Leave the mapping unevictable while we retry */
> > > + shmem_sg_free_table(st, mapping, false, false);
> > > kfree(st);
> > >
> > > max_segment = PAGE_SIZE;
> > > @@ -265,6 +260,7 @@ static int shmem_get_pages(struct drm_i915_gem_object
> > > *obj)
> > > return 0;
> > >
> > > err_pages:
> > > + mapping_clear_unevictable(mapping);
> > > shmem_sg_free_table(st, mapping, false, false);
> > > /*
> > > * shmemfs first checks if there is enough memory to allocate the page
> > > @@ -402,6 +398,7 @@ void i915_gem_object_put_pages_shmem(struct
> > > drm_i915_gem_object *obj, struct sg_
> > > if (i915_gem_object_needs_bit17_swizzle(obj))
> > > i915_gem_object_save_bit_17_swizzle(obj, pages);
> > >
> > > + mapping_clear_unevictable(file_inode(obj->base.filp)->i_mapping);
> > > shmem_sg_free_table(pages, file_inode(obj->base.filp)->i_mapping,
> > > obj->mm.dirty, obj->mm.madv ==
> > > I915_MADV_WILLNEED);
> > > kfree(pages);
> > > --
> > > 2.48.0.rc2.279.g1de40edade-goog
> >
> > --
> > Ville Syrjälä
> > Intel
>
> --
> Ville Syrjälä
> Intel
[PATCH] drm/i915: Fix page cleanup on DMA remap failure
When converting to folios the cleanup path of shmem_get_pages() was
missed. When a DMA remap fails and the max segment size is greater than
PAGE_SIZE it will attempt to retry the remap with a PAGE_SIZEd segment
size. The cleanup code isn't properly using the folio apis and as a
result isn't handling compound pages correctly.
Link: https://gitlab.freedesktop.org/drm/i915/kernel/-/issues/13487
Fixes: 0b62af28f249 ("i915: convert shmem_sg_free_table() to use a folio_batch")
Signed-off-by: Brian Geffon
Suggested-by: Tomasz Figa
---
drivers/gpu/drm/i915/gem/i915_gem_shmem.c | 13 +
1 file changed, 5 insertions(+), 8 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
index fe69f2c8527d..02ddab5bf5c0 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
@@ -37,8 +37,6 @@ void shmem_sg_free_table(struct sg_table *st, struct
address_space *mapping,
struct folio *last = NULL;
struct page *page;
- mapping_clear_unevictable(mapping);
-
folio_batch_init(&fbatch);
for_each_sgt_page(page, sgt_iter, st) {
struct folio *folio = page_folio(page);
@@ -180,10 +178,10 @@ int shmem_sg_alloc_table(struct drm_i915_private *i915,
struct sg_table *st,
return 0;
err_sg:
sg_mark_end(sg);
+ mapping_clear_unevictable(mapping);
if (sg != st->sgl) {
shmem_sg_free_table(st, mapping, false, false);
} else {
- mapping_clear_unevictable(mapping);
sg_free_table(st);
}
@@ -209,8 +207,6 @@ static int shmem_get_pages(struct drm_i915_gem_object *obj)
struct address_space *mapping = obj->base.filp->f_mapping;
unsigned int max_segment = i915_sg_segment_size(i915->drm.dev);
struct sg_table *st;
- struct sgt_iter sgt_iter;
- struct page *page;
int ret;
/*
@@ -239,9 +235,8 @@ static int shmem_get_pages(struct drm_i915_gem_object *obj)
* for PAGE_SIZE chunks instead may be helpful.
*/
if (max_segment > PAGE_SIZE) {
- for_each_sgt_page(page, sgt_iter, st)
- put_page(page);
- sg_free_table(st);
+ /* Leave the mapping unevictable while we retry */
+ shmem_sg_free_table(st, mapping, false, false);
kfree(st);
max_segment = PAGE_SIZE;
@@ -265,6 +260,7 @@ static int shmem_get_pages(struct drm_i915_gem_object *obj)
return 0;
err_pages:
+ mapping_clear_unevictable(mapping);
shmem_sg_free_table(st, mapping, false, false);
/*
* shmemfs first checks if there is enough memory to allocate the page
@@ -402,6 +398,7 @@ void i915_gem_object_put_pages_shmem(struct
drm_i915_gem_object *obj, struct sg_
if (i915_gem_object_needs_bit17_swizzle(obj))
i915_gem_object_save_bit_17_swizzle(obj, pages);
+ mapping_clear_unevictable(file_inode(obj->base.filp)->i_mapping);
shmem_sg_free_table(pages, file_inode(obj->base.filp)->i_mapping,
obj->mm.dirty, obj->mm.madv == I915_MADV_WILLNEED);
kfree(pages);
--
2.48.0.rc2.279.g1de40edade-goog
