On 6/28/21 6:28 PM, Matthew Auld wrote:

On 28/06/2021 15:46, Thomas Hellström wrote:

...

Introduce an interface to migrate objects between regions. This is primarily intended to migrate objects to LMEM for display and to SYSTEM for dma-buf, but might be reused in one form or another for performance-based migration.

v2:

• Verify that the memory region given as an id really exists.

(Reported by Matthew Auld)

• Call i915_gem_object_{init,release}_memory_region() when switching

region    to handle also switching region lists. (Reported by Matthew Auld) v3:

• Fix i915_gem_object_can_migrate() to return true if object is

already in    the correct region, even if the object ops doesn't have a migrate()    callback.

• Update typo in commit message.
• Fix kerneldoc of i915_gem_object_wait_migration().

Reported-by: kernel test robot lkp@intel.com Signed-off-by: Thomas Hellström thomas.hellstrom@linux.intel.com

---

drivers/gpu/drm/i915/gem/i915_gem_object.c    | 96 +++++++++++++++++++   drivers/gpu/drm/i915/gem/i915_gem_object.h    | 12 +++   .../gpu/drm/i915/gem/i915_gem_object_types.h  |  9 ++   drivers/gpu/drm/i915/gem/i915_gem_ttm.c       | 69 +++++++++----   drivers/gpu/drm/i915/gem/i915_gem_wait.c      | 19 ++++   5 files changed, 188 insertions(+), 17 deletions(-)

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.c b/drivers/gpu/drm/i915/gem/i915_gem_object.c index 07e8ff9a8aae..1c18be067b58 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.c @@ -513,6 +513,102 @@ bool i915_gem_object_has_iomem(const struct drm_i915_gem_object *obj)       return obj->mem_flags & I915_BO_FLAG_IOMEM;   }   +/**

• • i915_gem_object_can_migrate - Whether an object likely can be

migrated

• • @obj: The object to migrate
• • @id: The region intended to migrate to
• • Check whether the object backend supports migration to the
• • given region. Note that pinning may affect the ability to migrate.
• • Return: true if migration is possible, false otherwise.
• */

+bool i915_gem_object_can_migrate(struct drm_i915_gem_object *obj, +                 enum intel_region_id id) +{ +    struct drm_i915_private *i915 = to_i915(obj->base.dev); +    unsigned int num_allowed = obj->mm.n_placements; +    struct intel_memory_region *mr; +    unsigned int i;

+    GEM_BUG_ON(id >= INTEL_REGION_UNKNOWN); +    GEM_BUG_ON(obj->mm.madv != I915_MADV_WILLNEED);

+    mr = i915->mm.regions[id]; +    if (!mr) +        return false;

+    if (obj->mm.region == mr) +        return true;

+    if (!i915_gem_object_evictable(obj)) +        return false;

+    if (!obj->ops->migrate) +        return false;

+    if (!(obj->flags & I915_BO_ALLOC_USER)) +        return true;

+    if (num_allowed == 0) +        return false;

+    for (i = 0; i < num_allowed; ++i) { +        if (mr == obj->mm.placements[i]) +            return true; +    }

+    return false; +}

+/**

• • i915_gem_object_migrate - Migrate an object to the desired region id
• • @obj: The object to migrate.
• • @ww: An optional struct i915_gem_ww_ctx. If NULL, the backend may
• • not be successful in evicting other objects to make room for this

object.

• • @id: The region id to migrate to.
• • Attempt to migrate the object to the desired memory region. The
• • object backend must support migration and the object may not be
• • pinned, (explicitly pinned pages or pinned vmas). The object must
• • be locked.
• • On successful completion, the object will have pages pointing to
• • memory in the new region, but an async migration task may not have
• • completed yet, and to accomplish that,

i915_gem_object_wait_migration()

• • must be called.
• • Return: 0 on success. Negative error code on failure. In

particular may

• • return -ENXIO on lack of region space, -EDEADLK for deadlock

avoidance

• • if @ww is set, -EINTR or -ERESTARTSYS if signal pending, and
• • -EBUSY if the object is pinned.
• */

+int i915_gem_object_migrate(struct drm_i915_gem_object *obj, +                struct i915_gem_ww_ctx *ww, +                enum intel_region_id id) +{ +    struct drm_i915_private *i915 = to_i915(obj->base.dev); +    struct intel_memory_region *mr;

+    GEM_BUG_ON(id >= INTEL_REGION_UNKNOWN); +    GEM_BUG_ON(obj->mm.madv != I915_MADV_WILLNEED); +    assert_object_held(obj);

+    mr = i915->mm.regions[id]; +    GEM_BUG_ON(!mr);

+    if (obj->mm.region == mr) +        return 0;

+    if (!i915_gem_object_evictable(obj)) +        return -EBUSY;

+    if (!obj->ops->migrate) +        return -EOPNOTSUPP;

+    return obj->ops->migrate(obj, mr); +}

void i915_gem_init__objects(struct drm_i915_private *i915)   {       INIT_WORK(&i915->mm.free_work, __i915_gem_free_work); diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.h b/drivers/gpu/drm/i915/gem/i915_gem_object.h index ea3224a480c4..8cbd7a5334e2 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.h +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.h @@ -17,6 +17,8 @@   #include "i915_gem_ww.h"   #include "i915_vma_types.h"   +enum intel_region_id;

/*    * XXX: There is a prevalence of the assumption that we fit the    * object's page count inside a 32bit _signed_ variable. Let's document @@ -597,6 +599,16 @@ bool i915_gem_object_migratable(struct drm_i915_gem_object *obj);     bool i915_gem_object_validates_to_lmem(struct drm_i915_gem_object *obj);   +int i915_gem_object_migrate(struct drm_i915_gem_object *obj, +                struct i915_gem_ww_ctx *ww, +                enum intel_region_id id);

+bool i915_gem_object_can_migrate(struct drm_i915_gem_object *obj, +                 enum intel_region_id id);

+int i915_gem_object_wait_migration(struct drm_i915_gem_object *obj, +                   unsigned int flags);

#ifdef CONFIG_MMU_NOTIFIER   static inline bool   i915_gem_object_is_userptr(struct drm_i915_gem_object *obj) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object_types.h b/drivers/gpu/drm/i915/gem/i915_gem_object_types.h index 441f913c87e6..ef3de2ae9723 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object_types.h +++ b/drivers/gpu/drm/i915/gem/i915_gem_object_types.h @@ -18,6 +18,7 @@     struct drm_i915_gem_object;   struct intel_fronbuffer; +struct intel_memory_region;     /*    * struct i915_lut_handle tracks the fast lookups from handle to vma used @@ -77,6 +78,14 @@ struct drm_i915_gem_object_ops {        * delayed_free - Override the default delayed free implementation        */       void (*delayed_free)(struct drm_i915_gem_object *obj);

+    /** +     * migrate - Migrate object to a different region either for +     * pinning or for as long as the object lock is held. +     */ +    int (*migrate)(struct drm_i915_gem_object *obj, +               struct intel_memory_region *mr);

void (*release)(struct drm_i915_gem_object *obj);         const struct vm_operations_struct *mmap_ops; diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c index c39d982c4fa6..8f89185b6507 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c @@ -617,7 +617,8 @@ struct ttm_device_funcs *i915_ttm_driver(void)       return &i915_ttm_bo_driver;   }   -static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) +static int __i915_ttm_get_pages(struct drm_i915_gem_object *obj, +                struct ttm_placement *placement)   {       struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);       struct ttm_operation_ctx ctx = { @@ -625,19 +626,12 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj)           .no_wait_gpu = false,       };       struct sg_table *st; -    struct ttm_place requested, busy[I915_TTM_MAX_PLACEMENTS]; -    struct ttm_placement placement;       int real_num_busy;       int ret;   -    GEM_BUG_ON(obj->mm.n_placements > I915_TTM_MAX_PLACEMENTS);

-    /* Move to the requested placement. */ -    i915_ttm_placement_from_obj(obj, &requested, busy, &placement);

/* First try only the requested placement. No eviction. */ -    real_num_busy = fetch_and_zero(&placement.num_busy_placement); -    ret = ttm_bo_validate(bo, &placement, &ctx); +    real_num_busy = fetch_and_zero(&placement->num_busy_placement); +    ret = ttm_bo_validate(bo, placement, &ctx);       if (ret) {           ret = i915_ttm_err_to_gem(ret);           /* @@ -652,8 +646,8 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj)            * If the initial attempt fails, allow all accepted placements,            * evicting if necessary.            */ -        placement.num_busy_placement = real_num_busy; -        ret = ttm_bo_validate(bo, &placement, &ctx); +        placement->num_busy_placement = real_num_busy; +        ret = ttm_bo_validate(bo, placement, &ctx);           if (ret)               return i915_ttm_err_to_gem(ret);       } @@ -668,16 +662,56 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj)           i915_ttm_adjust_gem_after_move(obj);       }   -    /* Object either has a page vector or is an iomem object */ -    st = bo->ttm ? i915_ttm_tt_get_st(bo->ttm) : obj->ttm.cached_io_st; -    if (IS_ERR(st)) -        return PTR_ERR(st); +    if (!obj->mm.pages) {

Is this for the migration case? Where are we updating the mm.pages, assuming the migration step has to take care of pre-existing pages? What am I missing?

Hmm, yes this is for a hypotetical migration case where ttm_bo_validate() decided to not change placement of the object, meaning we migrate to the very same region that we're already in. That should be blocked at the gem level, so I guess we could replace this with a GEM_BUG_ON().

...

+        /* Object either has a page vector or is an iomem object */ +        st = bo->ttm ? i915_ttm_tt_get_st(bo->ttm) : obj->ttm.cached_io_st; +        if (IS_ERR(st)) +            return PTR_ERR(st);   -    __i915_gem_object_set_pages(obj, st, i915_sg_dma_sizes(st->sgl)); +        __i915_gem_object_set_pages(obj, st, i915_sg_dma_sizes(st->sgl)); +    }         return ret;   }   +static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) +{ +    struct ttm_place requested, busy[I915_TTM_MAX_PLACEMENTS]; +    struct ttm_placement placement;

+    GEM_BUG_ON(obj->mm.n_placements > I915_TTM_MAX_PLACEMENTS);

+    /* Move to the requested placement. */ +    i915_ttm_placement_from_obj(obj, &requested, busy, &placement);

+    return __i915_ttm_get_pages(obj, &placement); +}

+static int i915_ttm_migrate(struct drm_i915_gem_object *obj, +                struct intel_memory_region *mr) +{ +    struct ttm_place requested; +    struct ttm_placement placement; +    int ret;

+    i915_ttm_place_from_region(mr, &requested, obj->flags); +    placement.num_placement = 1; +    placement.num_busy_placement = 1; +    placement.placement = &requested; +    placement.busy_placement = &requested;

+    ret = __i915_ttm_get_pages(obj, &placement); +    if (ret) +        return ret;

+    if (obj->mm.region != mr) { +        i915_gem_object_release_memory_region(obj); +        i915_gem_object_init_memory_region(obj, mr);

Hmm, so is this not already handled by adjust_gem_after_move()?

No, not always. adjust_gem_after_move() only considers an object migrated and flips region if the new region is in one of the allowed placements. Otherwise it's considered evicted and the next get_pages() will revalidate into the preferred placement.

the gem_migrate interface is more flexible, and offers a possibility to migrate to whatever it's told, so as long as we want that behaviour, we need the above.

/Thomas

...

+    }

+    return 0; +}

static void i915_ttm_put_pages(struct drm_i915_gem_object *obj,                      struct sg_table *st)   { @@ -814,6 +848,7 @@ static const struct drm_i915_gem_object_ops i915_gem_ttm_obj_ops = {       .truncate = i915_ttm_purge,       .adjust_lru = i915_ttm_adjust_lru,       .delayed_free = i915_ttm_delayed_free, +    .migrate = i915_ttm_migrate,       .mmap_offset = i915_ttm_mmap_offset,       .mmap_ops = &vm_ops_ttm,   }; diff --git a/drivers/gpu/drm/i915/gem/i915_gem_wait.c b/drivers/gpu/drm/i915/gem/i915_gem_wait.c index 1070d3afdce7..f909aaa09d9c 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_wait.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_wait.c @@ -290,3 +290,22 @@ i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)       i915_gem_object_put(obj);       return ret;   }

+/**

• • i915_gem_object_wait_migration - Sync an accelerated migration

operation

• • @obj: The migrating object.
• • @flags: waiting flags. Currently supports only

I915_WAIT_INTERRUPTIBLE.

• • Wait for any pending async migration operation on the object,
• • whether it's explicitly (i915_gem_object_migrate()) or implicitly
• • (swapin, initial clearing) initiated.
• • Return: 0 if successful, -ERESTARTSYS if a signal was hit during

waiting.

• */

+int i915_gem_object_wait_migration(struct drm_i915_gem_object *obj, +                   unsigned int flags) +{ +    might_sleep(); +    /* NOP for now. */ +    return 0; +}

On 6/28/21 8:11 PM, Ruhl, Michael J wrote:

...

-----Original Message----- From: dri-devel dri-devel-bounces@lists.freedesktop.org On Behalf Of Thomas Hellström Sent: Monday, June 28, 2021 10:46 AM To: intel-gfx@lists.freedesktop.org; dri-devel@lists.freedesktop.org Cc: Thomas Hellström thomas.hellstrom@linux.intel.com; Auld, Matthew matthew.auld@intel.com; lkp lkp@intel.com Subject: [PATCH v3 1/5] drm/i915/gem: Implement object migration

Introduce an interface to migrate objects between regions. This is primarily intended to migrate objects to LMEM for display and to SYSTEM for dma-buf, but might be reused in one form or another for performance-based migration.

v2:

• Verify that the memory region given as an id really exists. (Reported by Matthew Auld)
• Call i915_gem_object_{init,release}_memory_region() when switching

region to handle also switching region lists. (Reported by Matthew Auld) v3:

• Fix i915_gem_object_can_migrate() to return true if object is already in the correct region, even if the object ops doesn't have a migrate() callback.
• Update typo in commit message.
• Fix kerneldoc of i915_gem_object_wait_migration().

Reported-by: kernel test robot lkp@intel.com Signed-off-by: Thomas Hellström thomas.hellstrom@linux.intel.com

---

drivers/gpu/drm/i915/gem/i915_gem_object.c | 96 +++++++++++++++++++ drivers/gpu/drm/i915/gem/i915_gem_object.h | 12 +++ .../gpu/drm/i915/gem/i915_gem_object_types.h | 9 ++ drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 69 +++++++++---- drivers/gpu/drm/i915/gem/i915_gem_wait.c | 19 ++++ 5 files changed, 188 insertions(+), 17 deletions(-)

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.c b/drivers/gpu/drm/i915/gem/i915_gem_object.c index 07e8ff9a8aae..1c18be067b58 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.c @@ -513,6 +513,102 @@ bool i915_gem_object_has_iomem(const struct drm_i915_gem_object *obj) return obj->mem_flags & I915_BO_FLAG_IOMEM; }

+/**

• • i915_gem_object_can_migrate - Whether an object likely can be migrated
• • @obj: The object to migrate
• • @id: The region intended to migrate to
• • Check whether the object backend supports migration to the
• • given region. Note that pinning may affect the ability to migrate.
• • Return: true if migration is possible, false otherwise.
• */

+bool i915_gem_object_can_migrate(struct drm_i915_gem_object *obj,

• 		 enum intel_region_id id)
  

+{

• struct drm_i915_private *i915 = to_i915(obj->base.dev);
• unsigned int num_allowed = obj->mm.n_placements;
• struct intel_memory_region *mr;
• unsigned int i;
• GEM_BUG_ON(id >= INTEL_REGION_UNKNOWN);
• GEM_BUG_ON(obj->mm.madv != I915_MADV_WILLNEED);
• mr = i915->mm.regions[id];
• if (!mr)

• return false;
  

• if (obj->mm.region == mr)

• return true;
  

• if (!i915_gem_object_evictable(obj))

• return false;
  

• if (!obj->ops->migrate)

• return false;
  

• if (!(obj->flags & I915_BO_ALLOC_USER))

• return true;
  

• if (num_allowed == 0)

• return false;
  

• for (i = 0; i < num_allowed; ++i) {

• if (mr == obj->mm.placements[i])
  

• 	return true;
  

• }

Hi Thomas,

I am a little confused over the difference in checks between this function and i915_gem_object_migrate().

Why is the lack of an mr a BUG_ON in _object_migrate(), but here it just false?

So that means that under certain circumstances, you could not have a mr?

If that is the case, when?

Would that be when the I915_BO_ALLOC_USER is set?

If so, should there be a check for "non" user vs user?

Or maybe document this function pointing out why there are differences and why?

Yes, I'll give it some more documentation. The basic idea is that the above function also could be used to validate user supplied data, whereas there might be cases where we want to use the gem_object_migrate() function and override the above.

...

• return false;

+}

+/**

• • i915_gem_object_migrate - Migrate an object to the desired region id
• • @obj: The object to migrate.
• • @ww: An optional struct i915_gem_ww_ctx. If NULL, the backend may
• • not be successful in evicting other objects to make room for this object.

Is the ww for future consideration? (I don't see any use of it in the patch).

Yes, but it will remain optional.

...

• • @id: The region id to migrate to.
• • Attempt to migrate the object to the desired memory region. The
• • object backend must support migration and the object may not be
• • pinned, (explicitly pinned pages or pinned vmas). The object must
• • be locked.
• • On successful completion, the object will have pages pointing to
• • memory in the new region, but an async migration task may not have
• • completed yet, and to accomplish that,

i915_gem_object_wait_migration()

• • must be called.
• • Return: 0 on success. Negative error code on failure. In particular may
• • return -ENXIO on lack of region space, -EDEADLK for deadlock avoidance
• • if @ww is set, -EINTR or -ERESTARTSYS if signal pending, and
• • -EBUSY if the object is pinned.
• */

+int i915_gem_object_migrate(struct drm_i915_gem_object *obj,

• 	    struct i915_gem_ww_ctx *ww,
  

• 	    enum intel_region_id id)
  

+{

• struct drm_i915_private *i915 = to_i915(obj->base.dev);
• struct intel_memory_region *mr;
• GEM_BUG_ON(id >= INTEL_REGION_UNKNOWN);
• GEM_BUG_ON(obj->mm.madv != I915_MADV_WILLNEED);
• assert_object_held(obj);
• mr = i915->mm.regions[id];
• GEM_BUG_ON(!mr);
• if (obj->mm.region == mr)

• return 0;
  

• if (!i915_gem_object_evictable(obj))

• return -EBUSY;
  

• if (!obj->ops->migrate)

• return -EOPNOTSUPP;
  

Why aren't you using _can_migrate here?

It's just in case we want to override. I'll add some more comments about this.

...

• return obj->ops->migrate(obj, mr);

+}

void i915_gem_init__objects(struct drm_i915_private *i915) { INIT_WORK(&i915->mm.free_work, __i915_gem_free_work); diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.h b/drivers/gpu/drm/i915/gem/i915_gem_object.h index ea3224a480c4..8cbd7a5334e2 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.h +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.h @@ -17,6 +17,8 @@ #include "i915_gem_ww.h" #include "i915_vma_types.h"

+enum intel_region_id;

/*

• XXX: There is a prevalence of the assumption that we fit the
• object's page count inside a 32bit _signed_ variable. Let's document

@@ -597,6 +599,16 @@ bool i915_gem_object_migratable(struct drm_i915_gem_object *obj);

bool i915_gem_object_validates_to_lmem(struct drm_i915_gem_object *obj);

+int i915_gem_object_migrate(struct drm_i915_gem_object *obj,

• 	    struct i915_gem_ww_ctx *ww,
  

• 	    enum intel_region_id id);
  

+bool i915_gem_object_can_migrate(struct drm_i915_gem_object *obj,

• 		 enum intel_region_id id);
  

+int i915_gem_object_wait_migration(struct drm_i915_gem_object *obj,

• 		   unsigned int flags);
  

#ifdef CONFIG_MMU_NOTIFIER static inline bool i915_gem_object_is_userptr(struct drm_i915_gem_object *obj) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object_types.h b/drivers/gpu/drm/i915/gem/i915_gem_object_types.h index 441f913c87e6..ef3de2ae9723 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object_types.h +++ b/drivers/gpu/drm/i915/gem/i915_gem_object_types.h @@ -18,6 +18,7 @@

struct drm_i915_gem_object; struct intel_fronbuffer; +struct intel_memory_region;

/*

• struct i915_lut_handle tracks the fast lookups from handle to vma used

@@ -77,6 +78,14 @@ struct drm_i915_gem_object_ops { * delayed_free - Override the default delayed free implementation */ void (*delayed_free)(struct drm_i915_gem_object *obj);

• /**

• * migrate - Migrate object to a different region either for
  

• * pinning or for as long as the object lock is held.
  

• */
  

• int (*migrate)(struct drm_i915_gem_object *obj,

•        struct intel_memory_region *mr);
  

• void (*release)(struct drm_i915_gem_object *obj);

  const struct vm_operations_struct *mmap_ops;

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c index c39d982c4fa6..8f89185b6507 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c @@ -617,7 +617,8 @@ struct ttm_device_funcs *i915_ttm_driver(void) return &i915_ttm_bo_driver; }

-static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) +static int __i915_ttm_get_pages(struct drm_i915_gem_object *obj,

• 		struct ttm_placement *placement)
  

{ struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); struct ttm_operation_ctx ctx = { @@ -625,19 +626,12 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) .no_wait_gpu = false, }; struct sg_table *st;

• struct ttm_place requested, busy[I915_TTM_MAX_PLACEMENTS];

• struct ttm_placement placement; int real_num_busy; int ret;

• GEM_BUG_ON(obj->mm.n_placements >

I915_TTM_MAX_PLACEMENTS);

• /* Move to the requested placement. */
• i915_ttm_placement_from_obj(obj, &requested, busy, &placement);
• /* First try only the requested placement. No eviction. */
• real_num_busy =

fetch_and_zero(&placement.num_busy_placement);

• ret = ttm_bo_validate(bo, &placement, &ctx);

• real_num_busy = fetch_and_zero(&placement-

...

num_busy_placement);

• ret = ttm_bo_validate(bo, placement, &ctx); if (ret) { ret = i915_ttm_err_to_gem(ret); /*

@@ -652,8 +646,8 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) * If the initial attempt fails, allow all accepted placements, * evicting if necessary. */

• placement.num_busy_placement = real_num_busy;
  

• ret = ttm_bo_validate(bo, &placement, &ctx);
  

• placement->num_busy_placement = real_num_busy;
  

• ret = ttm_bo_validate(bo, placement, &ctx);
  

  if (ret) return i915_ttm_err_to_gem(ret); }

@@ -668,16 +662,56 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) i915_ttm_adjust_gem_after_move(obj); }

• /* Object either has a page vector or is an iomem object */
• st = bo->ttm ? i915_ttm_tt_get_st(bo->ttm) : obj->ttm.cached_io_st;
• if (IS_ERR(st))

• return PTR_ERR(st);
  

• if (!obj->mm.pages) {

• /* Object either has a page vector or is an iomem object */
  

• st = bo->ttm ? i915_ttm_tt_get_st(bo->ttm) : obj-
  

...

ttm.cached_io_st;

• if (IS_ERR(st))
  

• 	return PTR_ERR(st);
  
  

• __i915_gem_object_set_pages(obj, st, i915_sg_dma_sizes(st->sgl));

• __i915_gem_object_set_pages(obj, st,
  

i915_sg_dma_sizes(st->sgl));

• }

  return ret;

}

+static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) +{

• struct ttm_place requested, busy[I915_TTM_MAX_PLACEMENTS];
• struct ttm_placement placement;
• GEM_BUG_ON(obj->mm.n_placements >

I915_TTM_MAX_PLACEMENTS);

• /* Move to the requested placement. */
• i915_ttm_placement_from_obj(obj, &requested, busy, &placement);
• return __i915_ttm_get_pages(obj, &placement);

+}

+static int i915_ttm_migrate(struct drm_i915_gem_object *obj,

• 	    struct intel_memory_region *mr)
  

+{

• struct ttm_place requested;
• struct ttm_placement placement;
• int ret;
• i915_ttm_place_from_region(mr, &requested, obj->flags);
• placement.num_placement = 1;
• placement.num_busy_placement = 1;
• placement.placement = &requested;
• placement.busy_placement = &requested;
• ret = __i915_ttm_get_pages(obj, &placement);
• if (ret)

• return ret;
  

• if (obj->mm.region != mr) {

• i915_gem_object_release_memory_region(obj);
  

• i915_gem_object_init_memory_region(obj, mr);
  

• }

Perhaps a minor nit:

Doing this after we have done the _get_pages() just doesn't seem right.

I.e. we do work on the object, and then we init some portion of it.

Do we need to do this incase the migration/placement fails? If so, maybe a comment to that effect?

This is simply switching memory region under the lock, and to also move to another memory region list. Is it the naming _release_ and _init_ that is confusing?

/Thomas

Thanks,

Mike

...

• return 0;

+}

static void i915_ttm_put_pages(struct drm_i915_gem_object *obj, struct sg_table *st) { @@ -814,6 +848,7 @@ static const struct drm_i915_gem_object_ops i915_gem_ttm_obj_ops = { .truncate = i915_ttm_purge, .adjust_lru = i915_ttm_adjust_lru, .delayed_free = i915_ttm_delayed_free,

• .migrate = i915_ttm_migrate, .mmap_offset = i915_ttm_mmap_offset, .mmap_ops = &vm_ops_ttm,

}; diff --git a/drivers/gpu/drm/i915/gem/i915_gem_wait.c b/drivers/gpu/drm/i915/gem/i915_gem_wait.c index 1070d3afdce7..f909aaa09d9c 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_wait.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_wait.c @@ -290,3 +290,22 @@ i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file) i915_gem_object_put(obj); return ret; }

+/**

• • i915_gem_object_wait_migration - Sync an accelerated migration

operation

• • @obj: The migrating object.
• • @flags: waiting flags. Currently supports only I915_WAIT_INTERRUPTIBLE.
• • Wait for any pending async migration operation on the object,
• • whether it's explicitly (i915_gem_object_migrate()) or implicitly
• • (swapin, initial clearing) initiated.
• • Return: 0 if successful, -ERESTARTSYS if a signal was hit during waiting.
• */

+int i915_gem_object_wait_migration(struct drm_i915_gem_object *obj,

• 		   unsigned int flags)
  

+{

• might_sleep();
• /* NOP for now. */
• return 0;

+}

2.31.1

On 6/28/21 9:50 PM, Ruhl, Michael J wrote:

...

-----Original Message----- From: Thomas Hellström thomas.hellstrom@linux.intel.com Sent: Monday, June 28, 2021 3:03 PM To: Ruhl, Michael J michael.j.ruhl@intel.com; intel- gfx@lists.freedesktop.org; dri-devel@lists.freedesktop.org Cc: Auld, Matthew matthew.auld@intel.com; lkp lkp@intel.com Subject: Re: [PATCH v3 1/5] drm/i915/gem: Implement object migration

On 6/28/21 8:11 PM, Ruhl, Michael J wrote:

...

...

-----Original Message----- From: dri-devel dri-devel-bounces@lists.freedesktop.org On Behalf Of Thomas Hellström Sent: Monday, June 28, 2021 10:46 AM To: intel-gfx@lists.freedesktop.org; dri-devel@lists.freedesktop.org Cc: Thomas Hellström thomas.hellstrom@linux.intel.com; Auld,

Matthew

...

...

matthew.auld@intel.com; lkp lkp@intel.com Subject: [PATCH v3 1/5] drm/i915/gem: Implement object migration

Introduce an interface to migrate objects between regions. This is primarily intended to migrate objects to LMEM for display and to SYSTEM for dma-buf, but might be reused in one form or another for performance-based migration.

v2:

• Verify that the memory region given as an id really exists. (Reported by Matthew Auld)
• Call i915_gem_object_{init,release}_memory_region() when switching

region to handle also switching region lists. (Reported by Matthew Auld) v3:

• Fix i915_gem_object_can_migrate() to return true if object is already in the correct region, even if the object ops doesn't have a migrate() callback.
• Update typo in commit message.
• Fix kerneldoc of i915_gem_object_wait_migration().

Reported-by: kernel test robot lkp@intel.com Signed-off-by: Thomas Hellström thomas.hellstrom@linux.intel.com

---

drivers/gpu/drm/i915/gem/i915_gem_object.c | 96 +++++++++++++++++++ drivers/gpu/drm/i915/gem/i915_gem_object.h | 12 +++ .../gpu/drm/i915/gem/i915_gem_object_types.h | 9 ++ drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 69 +++++++++---- drivers/gpu/drm/i915/gem/i915_gem_wait.c | 19 ++++ 5 files changed, 188 insertions(+), 17 deletions(-)

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.c b/drivers/gpu/drm/i915/gem/i915_gem_object.c index 07e8ff9a8aae..1c18be067b58 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.c @@ -513,6 +513,102 @@ bool i915_gem_object_has_iomem(const struct drm_i915_gem_object *obj) return obj->mem_flags & I915_BO_FLAG_IOMEM; }

+/**

• • i915_gem_object_can_migrate - Whether an object likely can be

migrated

...

...

• • @obj: The object to migrate
• • @id: The region intended to migrate to
• • Check whether the object backend supports migration to the
• • given region. Note that pinning may affect the ability to migrate.
• • Return: true if migration is possible, false otherwise.
• */

+bool i915_gem_object_can_migrate(struct drm_i915_gem_object *obj,

• 		 enum intel_region_id id)
  

+{

• struct drm_i915_private *i915 = to_i915(obj->base.dev);
• unsigned int num_allowed = obj->mm.n_placements;
• struct intel_memory_region *mr;
• unsigned int i;
• GEM_BUG_ON(id >= INTEL_REGION_UNKNOWN);
• GEM_BUG_ON(obj->mm.madv != I915_MADV_WILLNEED);
• mr = i915->mm.regions[id];
• if (!mr)

• return false;
  

• if (obj->mm.region == mr)

• return true;
  

• if (!i915_gem_object_evictable(obj))

• return false;
  

• if (!obj->ops->migrate)

• return false;
  

• if (!(obj->flags & I915_BO_ALLOC_USER))

• return true;
  

• if (num_allowed == 0)

• return false;
  

• for (i = 0; i < num_allowed; ++i) {

• if (mr == obj->mm.placements[i])
  

• 	return true;
  

• }

Hi Thomas,

I am a little confused over the difference in checks between this function and i915_gem_object_migrate().

Why is the lack of an mr a BUG_ON in _object_migrate(), but here it just false?

So that means that under certain circumstances, you could not have a mr?

If that is the case, when?

Would that be when the I915_BO_ALLOC_USER is set?

If so, should there be a check for "non" user vs user?

Or maybe document this function pointing out why there are differences and why?

Yes, I'll give it some more documentation. The basic idea is that the above function also could be used to validate user supplied data, whereas there might be cases where we want to use the gem_object_migrate() function and override the above.

...

...

• return false;

+}

+/**

• • i915_gem_object_migrate - Migrate an object to the desired region id
• • @obj: The object to migrate.
• • @ww: An optional struct i915_gem_ww_ctx. If NULL, the backend may
• • not be successful in evicting other objects to make room for this object.

Is the ww for future consideration? (I don't see any use of it in the patch).

Yes, but it will remain optional.

...

...

• • @id: The region id to migrate to.
• • Attempt to migrate the object to the desired memory region. The
• • object backend must support migration and the object may not be
• • pinned, (explicitly pinned pages or pinned vmas). The object must
• • be locked.
• • On successful completion, the object will have pages pointing to
• • memory in the new region, but an async migration task may not have
• • completed yet, and to accomplish that,

i915_gem_object_wait_migration()

• • must be called.
• • Return: 0 on success. Negative error code on failure. In particular may
• • return -ENXIO on lack of region space, -EDEADLK for deadlock

avoidance

...

...

• • if @ww is set, -EINTR or -ERESTARTSYS if signal pending, and
• • -EBUSY if the object is pinned.
• */

+int i915_gem_object_migrate(struct drm_i915_gem_object *obj,

• 	    struct i915_gem_ww_ctx *ww,
  

• 	    enum intel_region_id id)
  

+{

• struct drm_i915_private *i915 = to_i915(obj->base.dev);
• struct intel_memory_region *mr;
• GEM_BUG_ON(id >= INTEL_REGION_UNKNOWN);
• GEM_BUG_ON(obj->mm.madv != I915_MADV_WILLNEED);
• assert_object_held(obj);
• mr = i915->mm.regions[id];
• GEM_BUG_ON(!mr);
• if (obj->mm.region == mr)

• return 0;
  

• if (!i915_gem_object_evictable(obj))

• return -EBUSY;
  

• if (!obj->ops->migrate)

• return -EOPNOTSUPP;
  

Why aren't you using _can_migrate here?

It's just in case we want to override. I'll add some more comments about this.

...

...

• return obj->ops->migrate(obj, mr);

+}

void i915_gem_init__objects(struct drm_i915_private *i915) { INIT_WORK(&i915->mm.free_work, __i915_gem_free_work); diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.h b/drivers/gpu/drm/i915/gem/i915_gem_object.h index ea3224a480c4..8cbd7a5334e2 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.h +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.h @@ -17,6 +17,8 @@ #include "i915_gem_ww.h" #include "i915_vma_types.h"

+enum intel_region_id;

/*

• XXX: There is a prevalence of the assumption that we fit the
• object's page count inside a 32bit _signed_ variable. Let's document

@@ -597,6 +599,16 @@ bool i915_gem_object_migratable(struct drm_i915_gem_object *obj);

bool i915_gem_object_validates_to_lmem(struct drm_i915_gem_object *obj);

+int i915_gem_object_migrate(struct drm_i915_gem_object *obj,

• 	    struct i915_gem_ww_ctx *ww,
  

• 	    enum intel_region_id id);
  

+bool i915_gem_object_can_migrate(struct drm_i915_gem_object *obj,

• 		 enum intel_region_id id);
  

+int i915_gem_object_wait_migration(struct drm_i915_gem_object *obj,

• 		   unsigned int flags);
  

#ifdef CONFIG_MMU_NOTIFIER static inline bool i915_gem_object_is_userptr(struct drm_i915_gem_object *obj) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object_types.h b/drivers/gpu/drm/i915/gem/i915_gem_object_types.h index 441f913c87e6..ef3de2ae9723 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object_types.h +++ b/drivers/gpu/drm/i915/gem/i915_gem_object_types.h @@ -18,6 +18,7 @@

struct drm_i915_gem_object; struct intel_fronbuffer; +struct intel_memory_region;

/*

• struct i915_lut_handle tracks the fast lookups from handle to vma used

@@ -77,6 +78,14 @@ struct drm_i915_gem_object_ops { * delayed_free - Override the default delayed free implementation */ void (*delayed_free)(struct drm_i915_gem_object *obj);

• /**

• * migrate - Migrate object to a different region either for
  

• * pinning or for as long as the object lock is held.
  

• */
  

• int (*migrate)(struct drm_i915_gem_object *obj,

•        struct intel_memory_region *mr);
  

• void (*release)(struct drm_i915_gem_object *obj);

  const struct vm_operations_struct *mmap_ops;

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c index c39d982c4fa6..8f89185b6507 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c @@ -617,7 +617,8 @@ struct ttm_device_funcs *i915_ttm_driver(void) return &i915_ttm_bo_driver; }

-static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) +static int __i915_ttm_get_pages(struct drm_i915_gem_object *obj,

• 		struct ttm_placement *placement)
  

{ struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); struct ttm_operation_ctx ctx = { @@ -625,19 +626,12 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) .no_wait_gpu = false, }; struct sg_table *st;

• struct ttm_place requested, busy[I915_TTM_MAX_PLACEMENTS];

• struct ttm_placement placement; int real_num_busy; int ret;

• GEM_BUG_ON(obj->mm.n_placements >

I915_TTM_MAX_PLACEMENTS);

• /* Move to the requested placement. */
• i915_ttm_placement_from_obj(obj, &requested, busy, &placement);
• /* First try only the requested placement. No eviction. */
• real_num_busy =

fetch_and_zero(&placement.num_busy_placement);

• ret = ttm_bo_validate(bo, &placement, &ctx);

• real_num_busy = fetch_and_zero(&placement-

...

num_busy_placement);

• ret = ttm_bo_validate(bo, placement, &ctx); if (ret) { ret = i915_ttm_err_to_gem(ret); /*

@@ -652,8 +646,8 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) * If the initial attempt fails, allow all accepted placements, * evicting if necessary. */

• placement.num_busy_placement = real_num_busy;
  

• ret = ttm_bo_validate(bo, &placement, &ctx);
  

• placement->num_busy_placement = real_num_busy;
  

• ret = ttm_bo_validate(bo, placement, &ctx);
  

  if (ret) return i915_ttm_err_to_gem(ret); }

@@ -668,16 +662,56 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) i915_ttm_adjust_gem_after_move(obj); }

• /* Object either has a page vector or is an iomem object */
• st = bo->ttm ? i915_ttm_tt_get_st(bo->ttm) : obj->ttm.cached_io_st;
• if (IS_ERR(st))

• return PTR_ERR(st);
  

• if (!obj->mm.pages) {

• /* Object either has a page vector or is an iomem object */
  

• st = bo->ttm ? i915_ttm_tt_get_st(bo->ttm) : obj-
  

...

ttm.cached_io_st;

• if (IS_ERR(st))
  

• 	return PTR_ERR(st);
  
  

• __i915_gem_object_set_pages(obj, st, i915_sg_dma_sizes(st->sgl));

• __i915_gem_object_set_pages(obj, st,
  

i915_sg_dma_sizes(st->sgl));

• }

  return ret;

}

+static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) +{

• struct ttm_place requested, busy[I915_TTM_MAX_PLACEMENTS];
• struct ttm_placement placement;
• GEM_BUG_ON(obj->mm.n_placements >

I915_TTM_MAX_PLACEMENTS);

• /* Move to the requested placement. */
• i915_ttm_placement_from_obj(obj, &requested, busy, &placement);
• return __i915_ttm_get_pages(obj, &placement);

+}

+static int i915_ttm_migrate(struct drm_i915_gem_object *obj,

• 	    struct intel_memory_region *mr)
  

+{

• struct ttm_place requested;
• struct ttm_placement placement;
• int ret;
• i915_ttm_place_from_region(mr, &requested, obj->flags);
• placement.num_placement = 1;
• placement.num_busy_placement = 1;
• placement.placement = &requested;
• placement.busy_placement = &requested;
• ret = __i915_ttm_get_pages(obj, &placement);
• if (ret)

• return ret;
  

• if (obj->mm.region != mr) {

• i915_gem_object_release_memory_region(obj);
  

• i915_gem_object_init_memory_region(obj, mr);
  

• }

Perhaps a minor nit:

Doing this after we have done the _get_pages() just doesn't seem right.

I.e. we do work on the object, and then we init some portion of it.

Do we need to do this incase the migration/placement fails? If so, maybe a comment to that effect?

This is simply switching memory region under the lock, and to also move to another memory region list. Is it the naming _release_ and _init_ that is confusing?

Hmm, re-reading my question, I am wondering if I was unclear.

My "real" question was, can the release/init occur before the _get_pages()?

But looking at this some more, I can see answer is probably no.

I was going to suggest calling _init_ _set_, but when I looked at it, it was doing init things as well as setting things

Maybe just a comment like:

/* Complete the migration by updating the memory region info. */ if (object->mm...)

Sure, I'll add that. Note that another thread would have to take the object lock anyway to inspect both the region and the pages, so it shouldn't be confused by the order here.

/Thomas

On Mon, Jun 28, 2021 at 04:46:22PM +0200, Thomas Hellström wrote:

Introduce an interface to migrate objects between regions. This is primarily intended to migrate objects to LMEM for display and to SYSTEM for dma-buf, but might be reused in one form or another for performance-based migration.

v2:

• Verify that the memory region given as an id really exists. (Reported by Matthew Auld)
• Call i915_gem_object_{init,release}_memory_region() when switching region to handle also switching region lists. (Reported by Matthew Auld)

v3:

• Fix i915_gem_object_can_migrate() to return true if object is already in the correct region, even if the object ops doesn't have a migrate() callback.
• Update typo in commit message.
• Fix kerneldoc of i915_gem_object_wait_migration().

Reported-by: kernel test robot lkp@intel.com Signed-off-by: Thomas Hellström thomas.hellstrom@linux.intel.com

Looks reasonable to me as the main interface. I'm a bit on the fence on hiding everything behind obj->ops, since we're not going to have another implementation of migrate than the ttm one. But also while everything is in-flight it's probably good to at least try and maintain some boundaries, for otherwise the messiness might become unmanageable

Acked-by: Daniel Vetter daniel.vetter@ffwll.ch

But don't count this as real review :-)

One suggestion at the very bottom.

---

drivers/gpu/drm/i915/gem/i915_gem_object.c | 96 +++++++++++++++++++ drivers/gpu/drm/i915/gem/i915_gem_object.h | 12 +++ .../gpu/drm/i915/gem/i915_gem_object_types.h | 9 ++ drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 69 +++++++++---- drivers/gpu/drm/i915/gem/i915_gem_wait.c | 19 ++++ 5 files changed, 188 insertions(+), 17 deletions(-)

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.c b/drivers/gpu/drm/i915/gem/i915_gem_object.c index 07e8ff9a8aae..1c18be067b58 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.c @@ -513,6 +513,102 @@ bool i915_gem_object_has_iomem(const struct drm_i915_gem_object *obj) return obj->mem_flags & I915_BO_FLAG_IOMEM; }

+/**

• • i915_gem_object_can_migrate - Whether an object likely can be migrated
• • @obj: The object to migrate
• • @id: The region intended to migrate to
• • Check whether the object backend supports migration to the
• • given region. Note that pinning may affect the ability to migrate.
• • Return: true if migration is possible, false otherwise.
• */

+bool i915_gem_object_can_migrate(struct drm_i915_gem_object *obj,

• 		 enum intel_region_id id)
  

+{

• struct drm_i915_private *i915 = to_i915(obj->base.dev);
• unsigned int num_allowed = obj->mm.n_placements;
• struct intel_memory_region *mr;
• unsigned int i;
• GEM_BUG_ON(id >= INTEL_REGION_UNKNOWN);
• GEM_BUG_ON(obj->mm.madv != I915_MADV_WILLNEED);
• mr = i915->mm.regions[id];
• if (!mr)

• return false;
  

• if (obj->mm.region == mr)

• return true;
  

• if (!i915_gem_object_evictable(obj))

• return false;
  

• if (!obj->ops->migrate)

• return false;
  

• if (!(obj->flags & I915_BO_ALLOC_USER))

• return true;
  

• if (num_allowed == 0)

• return false;
  

• for (i = 0; i < num_allowed; ++i) {

• if (mr == obj->mm.placements[i])
  

• 	return true;
  

• }
• return false;

+}

+/**

• • i915_gem_object_migrate - Migrate an object to the desired region id
• • @obj: The object to migrate.
• • @ww: An optional struct i915_gem_ww_ctx. If NULL, the backend may
• • not be successful in evicting other objects to make room for this object.
• • @id: The region id to migrate to.
• • Attempt to migrate the object to the desired memory region. The
• • object backend must support migration and the object may not be
• • pinned, (explicitly pinned pages or pinned vmas). The object must
• • be locked.
• • On successful completion, the object will have pages pointing to
• • memory in the new region, but an async migration task may not have
• • completed yet, and to accomplish that, i915_gem_object_wait_migration()
• • must be called.
• • Return: 0 on success. Negative error code on failure. In particular may
• • return -ENXIO on lack of region space, -EDEADLK for deadlock avoidance
• • if @ww is set, -EINTR or -ERESTARTSYS if signal pending, and
• • -EBUSY if the object is pinned.
• */

+int i915_gem_object_migrate(struct drm_i915_gem_object *obj,

• 	    struct i915_gem_ww_ctx *ww,
  

• 	    enum intel_region_id id)
  

+{

• struct drm_i915_private *i915 = to_i915(obj->base.dev);
• struct intel_memory_region *mr;
• GEM_BUG_ON(id >= INTEL_REGION_UNKNOWN);
• GEM_BUG_ON(obj->mm.madv != I915_MADV_WILLNEED);
• assert_object_held(obj);
• mr = i915->mm.regions[id];
• GEM_BUG_ON(!mr);
• if (obj->mm.region == mr)

• return 0;
  

• if (!i915_gem_object_evictable(obj))

• return -EBUSY;
  

• if (!obj->ops->migrate)

• return -EOPNOTSUPP;
  

• return obj->ops->migrate(obj, mr);

+}

void i915_gem_init__objects(struct drm_i915_private *i915) { INIT_WORK(&i915->mm.free_work, __i915_gem_free_work); diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.h b/drivers/gpu/drm/i915/gem/i915_gem_object.h index ea3224a480c4..8cbd7a5334e2 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.h +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.h @@ -17,6 +17,8 @@ #include "i915_gem_ww.h" #include "i915_vma_types.h"

+enum intel_region_id;

/*

• XXX: There is a prevalence of the assumption that we fit the
• object's page count inside a 32bit _signed_ variable. Let's document

@@ -597,6 +599,16 @@ bool i915_gem_object_migratable(struct drm_i915_gem_object *obj);

bool i915_gem_object_validates_to_lmem(struct drm_i915_gem_object *obj);

+int i915_gem_object_migrate(struct drm_i915_gem_object *obj,

• 	    struct i915_gem_ww_ctx *ww,
  

• 	    enum intel_region_id id);
  

+bool i915_gem_object_can_migrate(struct drm_i915_gem_object *obj,

• 		 enum intel_region_id id);
  

+int i915_gem_object_wait_migration(struct drm_i915_gem_object *obj,

• 		   unsigned int flags);
  

#ifdef CONFIG_MMU_NOTIFIER static inline bool i915_gem_object_is_userptr(struct drm_i915_gem_object *obj) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object_types.h b/drivers/gpu/drm/i915/gem/i915_gem_object_types.h index 441f913c87e6..ef3de2ae9723 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object_types.h +++ b/drivers/gpu/drm/i915/gem/i915_gem_object_types.h @@ -18,6 +18,7 @@

struct drm_i915_gem_object; struct intel_fronbuffer; +struct intel_memory_region;

/*

• struct i915_lut_handle tracks the fast lookups from handle to vma used

@@ -77,6 +78,14 @@ struct drm_i915_gem_object_ops { * delayed_free - Override the default delayed free implementation */ void (*delayed_free)(struct drm_i915_gem_object *obj);

• /**

• * migrate - Migrate object to a different region either for
  

• * pinning or for as long as the object lock is held.
  

• */
  

• int (*migrate)(struct drm_i915_gem_object *obj,

•        struct intel_memory_region *mr);
  

• void (*release)(struct drm_i915_gem_object *obj);

  const struct vm_operations_struct *mmap_ops;

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c index c39d982c4fa6..8f89185b6507 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c @@ -617,7 +617,8 @@ struct ttm_device_funcs *i915_ttm_driver(void) return &i915_ttm_bo_driver; }

-static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) +static int __i915_ttm_get_pages(struct drm_i915_gem_object *obj,

• 		struct ttm_placement *placement)
  

{ struct ttm_buffer_object *bo = i915_gem_to_ttm(obj); struct ttm_operation_ctx ctx = { @@ -625,19 +626,12 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) .no_wait_gpu = false, }; struct sg_table *st;

• struct ttm_place requested, busy[I915_TTM_MAX_PLACEMENTS];

• struct ttm_placement placement; int real_num_busy; int ret;

• GEM_BUG_ON(obj->mm.n_placements > I915_TTM_MAX_PLACEMENTS);

• /* Move to the requested placement. */

• i915_ttm_placement_from_obj(obj, &requested, busy, &placement);

• /* First try only the requested placement. No eviction. */

• real_num_busy = fetch_and_zero(&placement.num_busy_placement);

• ret = ttm_bo_validate(bo, &placement, &ctx);

• real_num_busy = fetch_and_zero(&placement->num_busy_placement);
• ret = ttm_bo_validate(bo, placement, &ctx); if (ret) { ret = i915_ttm_err_to_gem(ret); /*

@@ -652,8 +646,8 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) * If the initial attempt fails, allow all accepted placements, * evicting if necessary. */

• placement.num_busy_placement = real_num_busy;
  

• ret = ttm_bo_validate(bo, &placement, &ctx);
  

• placement->num_busy_placement = real_num_busy;
  

• ret = ttm_bo_validate(bo, placement, &ctx);
  

  if (ret) return i915_ttm_err_to_gem(ret); }

@@ -668,16 +662,56 @@ static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) i915_ttm_adjust_gem_after_move(obj); }

• /* Object either has a page vector or is an iomem object */
• st = bo->ttm ? i915_ttm_tt_get_st(bo->ttm) : obj->ttm.cached_io_st;
• if (IS_ERR(st))

• return PTR_ERR(st);
  

• if (!obj->mm.pages) {

• /* Object either has a page vector or is an iomem object */
  

• st = bo->ttm ? i915_ttm_tt_get_st(bo->ttm) : obj->ttm.cached_io_st;
  

• if (IS_ERR(st))
  

• 	return PTR_ERR(st);
  
  

• __i915_gem_object_set_pages(obj, st, i915_sg_dma_sizes(st->sgl));

• __i915_gem_object_set_pages(obj, st, i915_sg_dma_sizes(st->sgl));
  

• }

  return ret;

}

+static int i915_ttm_get_pages(struct drm_i915_gem_object *obj) +{

• struct ttm_place requested, busy[I915_TTM_MAX_PLACEMENTS];
• struct ttm_placement placement;
• GEM_BUG_ON(obj->mm.n_placements > I915_TTM_MAX_PLACEMENTS);
• /* Move to the requested placement. */
• i915_ttm_placement_from_obj(obj, &requested, busy, &placement);
• return __i915_ttm_get_pages(obj, &placement);

+}

+static int i915_ttm_migrate(struct drm_i915_gem_object *obj,

• 	    struct intel_memory_region *mr)
  

+{

• struct ttm_place requested;
• struct ttm_placement placement;
• int ret;
• i915_ttm_place_from_region(mr, &requested, obj->flags);
• placement.num_placement = 1;
• placement.num_busy_placement = 1;
• placement.placement = &requested;
• placement.busy_placement = &requested;
• ret = __i915_ttm_get_pages(obj, &placement);
• if (ret)

• return ret;
  

• if (obj->mm.region != mr) {

• i915_gem_object_release_memory_region(obj);
  

• i915_gem_object_init_memory_region(obj, mr);
  

• }
• return 0;

+}

static void i915_ttm_put_pages(struct drm_i915_gem_object *obj, struct sg_table *st) { @@ -814,6 +848,7 @@ static const struct drm_i915_gem_object_ops i915_gem_ttm_obj_ops = { .truncate = i915_ttm_purge, .adjust_lru = i915_ttm_adjust_lru, .delayed_free = i915_ttm_delayed_free,

• .migrate = i915_ttm_migrate, .mmap_offset = i915_ttm_mmap_offset, .mmap_ops = &vm_ops_ttm,

}; diff --git a/drivers/gpu/drm/i915/gem/i915_gem_wait.c b/drivers/gpu/drm/i915/gem/i915_gem_wait.c index 1070d3afdce7..f909aaa09d9c 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_wait.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_wait.c @@ -290,3 +290,22 @@ i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file) i915_gem_object_put(obj); return ret; }

+/**

• • i915_gem_object_wait_migration - Sync an accelerated migration operation
• • @obj: The migrating object.
• • @flags: waiting flags. Currently supports only I915_WAIT_INTERRUPTIBLE.
• • Wait for any pending async migration operation on the object,
• • whether it's explicitly (i915_gem_object_migrate()) or implicitly
• • (swapin, initial clearing) initiated.
• • Return: 0 if successful, -ERESTARTSYS if a signal was hit during waiting.
• */

+int i915_gem_object_wait_migration(struct drm_i915_gem_object *obj,

• 		   unsigned int flags)
  

+{

• might_sleep();
• /* NOP for now. */

Maybe dma_fence_might_wait() here, since that's a notch more nasty?

Cheers, Daniel

• return 0;

+}

2.31.1

---

Intel-gfx mailing list Intel-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/intel-gfx

On 28/06/2021 15:46, Thomas Hellström wrote:

From: Matthew Auld matthew.auld@intel.com

A selftest for the gem object migrate functionality. Slightly adapted from the original by Matthew to the new interface and new fill blit code.

Co-developed-by: Thomas Hellström thomas.hellstrom@linux.intel.com Signed-off-by: Thomas Hellström thomas.hellstrom@linux.intel.com Signed-off-by: Matthew Auld matthew.auld@intel.com

---

drivers/gpu/drm/i915/gem/i915_gem_object.c | 1 + .../drm/i915/gem/selftests/i915_gem_migrate.c | 237 ++++++++++++++++++ .../drm/i915/selftests/i915_live_selftests.h | 1 + 3 files changed, 239 insertions(+) create mode 100644 drivers/gpu/drm/i915/gem/selftests/i915_gem_migrate.c

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.c b/drivers/gpu/drm/i915/gem/i915_gem_object.c index 1c18be067b58..ff147fd59874 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.c @@ -649,6 +649,7 @@ static const struct drm_gem_object_funcs i915_gem_object_funcs = { #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) #include "selftests/huge_gem_object.c" #include "selftests/huge_pages.c" +#include "selftests/i915_gem_migrate.c" #include "selftests/i915_gem_object.c" #include "selftests/i915_gem_coherency.c" #endif diff --git a/drivers/gpu/drm/i915/gem/selftests/i915_gem_migrate.c b/drivers/gpu/drm/i915/gem/selftests/i915_gem_migrate.c new file mode 100644 index 000000000000..a437b66f64d9 --- /dev/null +++ b/drivers/gpu/drm/i915/gem/selftests/i915_gem_migrate.c @@ -0,0 +1,237 @@ +// SPDX-License-Identifier: MIT +/*

• • Copyright © 2020-2021 Intel Corporation
• */

+#include "gt/intel_migrate.h"

+static int igt_smem_create_migrate(void *arg) +{

• struct intel_gt *gt = arg;
• struct drm_i915_private *i915 = gt->i915;
• struct drm_i915_gem_object *obj;
• struct i915_gem_ww_ctx ww;
• int err = 0;
• /* Switch object backing-store on create */
• obj = i915_gem_object_create_lmem(i915, PAGE_SIZE, 0);
• if (IS_ERR(obj))

• return PTR_ERR(obj);
  

• for_i915_gem_ww(&ww, err, true) {

• err = i915_gem_object_lock(obj, &ww);
  

• if (err)
  

• 	continue;
  

• if (!i915_gem_object_can_migrate(obj, INTEL_REGION_SMEM)) {
  

• 	err = -EINVAL;
  

• 	continue;
  

• }
  

• err = i915_gem_object_migrate(obj, &ww, INTEL_REGION_SMEM);
  

• if (err)
  

• 	continue;
  

• err = i915_gem_object_pin_pages(obj);
  

• if (err)
  

• 	continue;
  

• if (i915_gem_object_can_migrate(obj, INTEL_REGION_LMEM))
  

• 	err = -EINVAL;
  

• i915_gem_object_unpin_pages(obj);
  

• }
• i915_gem_object_put(obj);
• return err;

+}

+static int igt_lmem_create_migrate(void *arg) +{

• struct intel_gt *gt = arg;
• struct drm_i915_private *i915 = gt->i915;
• struct drm_i915_gem_object *obj;
• struct i915_gem_ww_ctx ww;
• int err = 0;
• /* Switch object backing-store on create */
• obj = i915_gem_object_create_shmem(i915, PAGE_SIZE);
• if (IS_ERR(obj))

• return PTR_ERR(obj);
  

• for_i915_gem_ww(&ww, err, true) {

• err = i915_gem_object_lock(obj, &ww);
  

• if (err)
  

• 	continue;
  

• if (!i915_gem_object_can_migrate(obj, INTEL_REGION_LMEM)) {
  

• 	err = -EINVAL;
  

• 	continue;
  

• }
  

• err = i915_gem_object_migrate(obj, &ww, INTEL_REGION_LMEM);
  

• if (err)
  

• 	continue;
  

• err = i915_gem_object_pin_pages(obj);
  

• if (err)
  

• 	continue;
  

• if (i915_gem_object_can_migrate(obj, INTEL_REGION_SMEM))
  

• 	err = -EINVAL;
  

• i915_gem_object_unpin_pages(obj);
  

• }
• i915_gem_object_put(obj);
• return err;

+}

+static int lmem_pages_migrate_one(struct i915_gem_ww_ctx *ww,

• 		  struct drm_i915_gem_object *obj)
  

+{

• int err;
• err = i915_gem_object_lock(obj, ww);
• if (err)

• return err;
  

• err = i915_gem_object_wait(obj,

• 		   I915_WAIT_INTERRUPTIBLE |
  

• 		   I915_WAIT_PRIORITY |
  

• 		   I915_WAIT_ALL,
  

• 		   MAX_SCHEDULE_TIMEOUT);
  

• if (err)

• return err;
  

• if (i915_gem_object_is_lmem(obj)) {

• if (!i915_gem_object_can_migrate(obj, INTEL_REGION_SMEM)) {
  

• 	pr_err("object can't migrate to smem.\n");
  

• 	return -EINVAL;
  

• }
  

• err = i915_gem_object_migrate(obj, ww, INTEL_REGION_SMEM);
  

• if (err) {
  

• 	pr_err("Object failed migration to smem\n");
  

• 	if (err)
  

• 		return err;
  

• }
  

• if (i915_gem_object_is_lmem(obj)) {
  

• 	pr_err("object still backed by lmem\n");
  

• 	err = -EINVAL;
  

• }
  

• if (!i915_gem_object_has_struct_page(obj)) {
  

• 	pr_err("object not backed by struct page\n");
  

• 	err = -EINVAL;
  

• }
  

• } else {

• if (!i915_gem_object_can_migrate(obj, INTEL_REGION_LMEM)) {
  

• 	pr_err("object can't migrate to lmem.\n");
  

• 	return -EINVAL;
  

• }
  

• err = i915_gem_object_migrate(obj, ww, INTEL_REGION_LMEM);
  

• if (err) {
  

• 	pr_err("Object failed migration to lmem\n");
  

• 	if (err)
  

• 		return err;
  

• }
  

• if (i915_gem_object_has_struct_page(obj)) {
  

• 	pr_err("object still backed by struct page\n");
  

• 	err = -EINVAL;
  

• }
  

• if (!i915_gem_object_is_lmem(obj)) {
  

• 	pr_err("object not backed by lmem\n");
  

• 	err = -EINVAL;
  

• }
  

• }
• return err;

+}

+static int igt_lmem_pages_migrate(void *arg) +{

• struct intel_gt *gt = arg;
• struct drm_i915_private *i915 = gt->i915;
• struct drm_i915_gem_object *obj;
• struct i915_gem_ww_ctx ww;
• struct i915_request *rq;
• int err;
• int i;
• /* From LMEM to shmem and back again */
• obj = i915_gem_object_create_lmem(i915, SZ_2M, 0);
• if (IS_ERR(obj))

• return PTR_ERR(obj);
  

• err = i915_gem_object_lock(obj, NULL);
• if (err)

• goto out_put;
  

• err = ____i915_gem_object_get_pages(obj);
• if (err) {

• i915_gem_object_unlock(obj);
  

• goto out_put;
  

• }
• err = intel_context_migrate_clear(gt->migrate.context, NULL,

• 			  obj->mm.pages->sgl, obj->cache_level,
  

• 			  i915_gem_object_is_lmem(obj),
  

• 			  0, &rq);
  

• if (rq) {

• dma_resv_add_excl_fence(obj->base.resv, &rq->fence);
  

• i915_request_put(rq);
  

• }
• i915_gem_object_unlock(obj);
• if (err)

• goto out_put;
  

• for (i = 1; i <= 4; ++i) {

• for_i915_gem_ww(&ww, err, true) {
  

• 	err = lmem_pages_migrate_one(&ww, obj);
  

• 	if (err)
  

• 		continue;
  

• 	err = i915_gem_object_wait_migration(obj, true);
  

• 	if (err)
  

• 		continue;
  

• 	err = intel_migrate_clear(&gt->migrate, &ww, NULL,
  

• 				  obj->mm.pages->sgl,
  

• 				  obj->cache_level,
  

• 				  i915_gem_object_is_lmem(obj),
  

• 				  0xdeadbeaf, &rq);
  

• 	if (rq) {
  

• 		dma_resv_add_excl_fence(obj->base.resv,
  

• 					&rq->fence);
  

• 		i915_request_put(rq);
  

• 	}
  

• }
  

• if (err)
  

• 	break;
  

• }

+out_put:

• i915_gem_object_put(obj);
• return err;

+}

+int i915_gem_migrate_live_selftests(struct drm_i915_private *i915) +{

• static const struct i915_subtest tests[] = {

• SUBTEST(igt_smem_create_migrate),
  

• SUBTEST(igt_lmem_create_migrate),
  

• SUBTEST(igt_lmem_pages_migrate),
  

Might be good to dirty the pages before the migration, and then verify they still match at the end?

• };
• if (!HAS_LMEM(i915))

• return 0;
  

• return intel_gt_live_subtests(tests, &i915->gt);

+} diff --git a/drivers/gpu/drm/i915/selftests/i915_live_selftests.h b/drivers/gpu/drm/i915/selftests/i915_live_selftests.h index a68197cf1044..e2fd1b61af71 100644 --- a/drivers/gpu/drm/i915/selftests/i915_live_selftests.h +++ b/drivers/gpu/drm/i915/selftests/i915_live_selftests.h @@ -40,6 +40,7 @@ selftest(hugepages, i915_gem_huge_page_live_selftests) selftest(gem_contexts, i915_gem_context_live_selftests) selftest(gem_execbuf, i915_gem_execbuffer_live_selftests) selftest(client, i915_gem_client_blt_live_selftests) +selftest(gem_migrate, i915_gem_migrate_live_selftests) selftest(reset, intel_reset_live_selftests) selftest(memory_region, intel_memory_region_live_selftests) selftest(hangcheck, intel_hangcheck_live_selftests)

On Mon, 2021-06-28 at 18:53 +0000, Ruhl, Michael J wrote:

...

-----Original Message----- From: Intel-gfx intel-gfx-bounces@lists.freedesktop.org On Behalf Of Thomas Hellström Sent: Monday, June 28, 2021 10:46 AM To: intel-gfx@lists.freedesktop.org; dri-devel@lists.freedesktop.org Cc: Thomas Hellström thomas.hellstrom@linux.intel.com; Auld, Matthew matthew.auld@intel.com Subject: [Intel-gfx] [PATCH v3 2/5] drm/i915/gem: Introduce a selftest for the gem object migrate functionality

From: Matthew Auld matthew.auld@intel.com

A selftest for the gem object migrate functionality. Slightly adapted from the original by Matthew to the new interface and new fill blit code.

Co-developed-by: Thomas Hellström thomas.hellstrom@linux.intel.com Signed-off-by: Thomas Hellström thomas.hellstrom@linux.intel.com Signed-off-by: Matthew Auld matthew.auld@intel.com

---

drivers/gpu/drm/i915/gem/i915_gem_object.c    |   1 + .../drm/i915/gem/selftests/i915_gem_migrate.c | 237 ++++++++++++++++++ .../drm/i915/selftests/i915_live_selftests.h  |   1 + 3 files changed, 239 insertions(+) create mode 100644 drivers/gpu/drm/i915/gem/selftests/i915_gem_migrate.c

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.c b/drivers/gpu/drm/i915/gem/i915_gem_object.c index 1c18be067b58..ff147fd59874 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.c @@ -649,6 +649,7 @@ static const struct drm_gem_object_funcs i915_gem_object_funcs = { #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) #include "selftests/huge_gem_object.c" #include "selftests/huge_pages.c" +#include "selftests/i915_gem_migrate.c" #include "selftests/i915_gem_object.c" #include "selftests/i915_gem_coherency.c" #endif diff --git a/drivers/gpu/drm/i915/gem/selftests/i915_gem_migrate.c b/drivers/gpu/drm/i915/gem/selftests/i915_gem_migrate.c new file mode 100644 index 000000000000..a437b66f64d9 --- /dev/null +++ b/drivers/gpu/drm/i915/gem/selftests/i915_gem_migrate.c @@ -0,0 +1,237 @@ +// SPDX-License-Identifier: MIT +/*

• • Copyright © 2020-2021 Intel Corporation
• */

+#include "gt/intel_migrate.h"

+static int igt_smem_create_migrate(void *arg) +{ +       struct intel_gt *gt = arg; +       struct drm_i915_private *i915 = gt->i915; +       struct drm_i915_gem_object *obj; +       struct i915_gem_ww_ctx ww; +       int err = 0;

+       /* Switch object backing-store on create */ +       obj = i915_gem_object_create_lmem(i915, PAGE_SIZE, 0); +       if (IS_ERR(obj)) +               return PTR_ERR(obj);

+       for_i915_gem_ww(&ww, err, true) { +               err = i915_gem_object_lock(obj, &ww); +               if (err) +                       continue;

+               if (!i915_gem_object_can_migrate(obj, INTEL_REGION_SMEM)) { +                       err = -EINVAL; +                       continue; +               }

+               err = i915_gem_object_migrate(obj, &ww, INTEL_REGION_SMEM); +               if (err) +                       continue;

+               err = i915_gem_object_pin_pages(obj); +               if (err) +                       continue;

+               if (i915_gem_object_can_migrate(obj, INTEL_REGION_LMEM)) +                       err = -EINVAL;

+               i915_gem_object_unpin_pages(obj); +       } +       i915_gem_object_put(obj);

+       return err; +}

+static int igt_lmem_create_migrate(void *arg) +{ +       struct intel_gt *gt = arg; +       struct drm_i915_private *i915 = gt->i915; +       struct drm_i915_gem_object *obj; +       struct i915_gem_ww_ctx ww; +       int err = 0;

+       /* Switch object backing-store on create */ +       obj = i915_gem_object_create_shmem(i915, PAGE_SIZE); +       if (IS_ERR(obj)) +               return PTR_ERR(obj);

+       for_i915_gem_ww(&ww, err, true) { +               err = i915_gem_object_lock(obj, &ww); +               if (err) +                       continue;

+               if (!i915_gem_object_can_migrate(obj, INTEL_REGION_LMEM)) { +                       err = -EINVAL; +                       continue; +               }

+               err = i915_gem_object_migrate(obj, &ww, INTEL_REGION_LMEM); +               if (err) +                       continue;

+               err = i915_gem_object_pin_pages(obj); +               if (err) +                       continue;

+               if (i915_gem_object_can_migrate(obj, INTEL_REGION_SMEM)) +                       err = -EINVAL;

+               i915_gem_object_unpin_pages(obj); +       } +       i915_gem_object_put(obj);

+       return err; +}

+static int lmem_pages_migrate_one(struct i915_gem_ww_ctx *ww, +                                 struct drm_i915_gem_object *obj) +{ +       int err;

+       err = i915_gem_object_lock(obj, ww); +       if (err) +               return err;

+       err = i915_gem_object_wait(obj, +                                  I915_WAIT_INTERRUPTIBLE | +                                  I915_WAIT_PRIORITY | +                                  I915_WAIT_ALL, +                                  MAX_SCHEDULE_TIMEOUT); +       if (err) +               return err;

+       if (i915_gem_object_is_lmem(obj)) { +               if (!i915_gem_object_can_migrate(obj, INTEL_REGION_SMEM)) {

I don't see any  testing of the parameter num_allowed.

Is that done somewhere else?

Mike

That's a user-space submitted parameter only, dictating what region the object is allowed in when bound to the GPU. It's not exercised in this selftest.

Thanks, Thomas

On 6/28/21 9:27 PM, Ruhl, Michael J wrote:

...

-----Original Message----- From: Thomas Hellström thomas.hellstrom@linux.intel.com Sent: Monday, June 28, 2021 3:15 PM To: Ruhl, Michael J michael.j.ruhl@intel.com; intel- gfx@lists.freedesktop.org; dri-devel@lists.freedesktop.org Cc: Auld, Matthew matthew.auld@intel.com Subject: Re: [Intel-gfx] [PATCH v3 2/5] drm/i915/gem: Introduce a selftest for the gem object migrate functionality

On Mon, 2021-06-28 at 18:53 +0000, Ruhl, Michael J wrote:

...

...

-----Original Message----- From: Intel-gfx intel-gfx-bounces@lists.freedesktop.org On Behalf Of Thomas Hellström Sent: Monday, June 28, 2021 10:46 AM To: intel-gfx@lists.freedesktop.org; dri-devel@lists.freedesktop.org Cc: Thomas Hellström thomas.hellstrom@linux.intel.com; Auld, Matthew matthew.auld@intel.com Subject: [Intel-gfx] [PATCH v3 2/5] drm/i915/gem: Introduce a selftest for the gem object migrate functionality

From: Matthew Auld matthew.auld@intel.com

A selftest for the gem object migrate functionality. Slightly adapted from the original by Matthew to the new interface and new fill blit code.

Co-developed-by: Thomas Hellström thomas.hellstrom@linux.intel.com Signed-off-by: Thomas Hellström thomas.hellstrom@linux.intel.com Signed-off-by: Matthew Auld matthew.auld@intel.com

---

drivers/gpu/drm/i915/gem/i915_gem_object.c    |   1 + .../drm/i915/gem/selftests/i915_gem_migrate.c | 237 ++++++++++++++++++ .../drm/i915/selftests/i915_live_selftests.h  |   1 + 3 files changed, 239 insertions(+) create mode 100644 drivers/gpu/drm/i915/gem/selftests/i915_gem_migrate.c

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.c b/drivers/gpu/drm/i915/gem/i915_gem_object.c index 1c18be067b58..ff147fd59874 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.c @@ -649,6 +649,7 @@ static const struct drm_gem_object_funcs i915_gem_object_funcs = { #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) #include "selftests/huge_gem_object.c" #include "selftests/huge_pages.c" +#include "selftests/i915_gem_migrate.c" #include "selftests/i915_gem_object.c" #include "selftests/i915_gem_coherency.c" #endif diff --git a/drivers/gpu/drm/i915/gem/selftests/i915_gem_migrate.c b/drivers/gpu/drm/i915/gem/selftests/i915_gem_migrate.c new file mode 100644 index 000000000000..a437b66f64d9 --- /dev/null +++ b/drivers/gpu/drm/i915/gem/selftests/i915_gem_migrate.c @@ -0,0 +1,237 @@ +// SPDX-License-Identifier: MIT +/*

• • Copyright © 2020-2021 Intel Corporation
• */

+#include "gt/intel_migrate.h"

+static int igt_smem_create_migrate(void *arg) +{ +       struct intel_gt *gt = arg; +       struct drm_i915_private *i915 = gt->i915; +       struct drm_i915_gem_object *obj; +       struct i915_gem_ww_ctx ww; +       int err = 0;

+       /* Switch object backing-store on create */ +       obj = i915_gem_object_create_lmem(i915, PAGE_SIZE, 0); +       if (IS_ERR(obj)) +               return PTR_ERR(obj);

+       for_i915_gem_ww(&ww, err, true) { +               err = i915_gem_object_lock(obj, &ww); +               if (err) +                       continue;

+               if (!i915_gem_object_can_migrate(obj, INTEL_REGION_SMEM)) { +                       err = -EINVAL; +                       continue; +               }

+               err = i915_gem_object_migrate(obj, &ww, INTEL_REGION_SMEM); +               if (err) +                       continue;

+               err = i915_gem_object_pin_pages(obj); +               if (err) +                       continue;

+               if (i915_gem_object_can_migrate(obj, INTEL_REGION_LMEM)) +                       err = -EINVAL;

+               i915_gem_object_unpin_pages(obj); +       } +       i915_gem_object_put(obj);

+       return err; +}

+static int igt_lmem_create_migrate(void *arg) +{ +       struct intel_gt *gt = arg; +       struct drm_i915_private *i915 = gt->i915; +       struct drm_i915_gem_object *obj; +       struct i915_gem_ww_ctx ww; +       int err = 0;

+       /* Switch object backing-store on create */ +       obj = i915_gem_object_create_shmem(i915, PAGE_SIZE); +       if (IS_ERR(obj)) +               return PTR_ERR(obj);

+       for_i915_gem_ww(&ww, err, true) { +               err = i915_gem_object_lock(obj, &ww); +               if (err) +                       continue;

+               if (!i915_gem_object_can_migrate(obj, INTEL_REGION_LMEM)) { +                       err = -EINVAL; +                       continue; +               }

+               err = i915_gem_object_migrate(obj, &ww, INTEL_REGION_LMEM); +               if (err) +                       continue;

+               err = i915_gem_object_pin_pages(obj); +               if (err) +                       continue;

+               if (i915_gem_object_can_migrate(obj, INTEL_REGION_SMEM)) +                       err = -EINVAL;

+               i915_gem_object_unpin_pages(obj); +       } +       i915_gem_object_put(obj);

+       return err; +}

+static int lmem_pages_migrate_one(struct i915_gem_ww_ctx *ww, +                                 struct drm_i915_gem_object *obj) +{ +       int err;

+       err = i915_gem_object_lock(obj, ww); +       if (err) +               return err;

+       err = i915_gem_object_wait(obj, +                                  I915_WAIT_INTERRUPTIBLE | +                                  I915_WAIT_PRIORITY | +                                  I915_WAIT_ALL, +                                  MAX_SCHEDULE_TIMEOUT); +       if (err) +               return err;

+       if (i915_gem_object_is_lmem(obj)) { +               if (!i915_gem_object_can_migrate(obj, INTEL_REGION_SMEM)) {

I don't see any  testing of the parameter num_allowed.

Is that done somewhere else?

Mike

That's a user-space submitted parameter only, dictating what region the object is allowed in when bound to the GPU. It's not exercised in this selftest.

Got it.

Reviewed-by: Michael J. Ruhl michael.j.ruhl@intel.com

M

Thanks!

Thomas

...

Thanks, Thomas

-----Original Message----- From: Thomas Hellström thomas.hellstrom@linux.intel.com Sent: Monday, June 28, 2021 10:46 AM To: intel-gfx@lists.freedesktop.org; dri-devel@lists.freedesktop.org Cc: Auld, Matthew matthew.auld@intel.com; maarten.lankhorst@linux.intel.com; Thomas Hellström thomas.hellstrom@linux.intel.com; Ruhl; Ruhl, Michael J michael.j.ruhl@intel.com Subject: [PATCH v3 4/5] drm/i915/gem: Fix same-driver-another-instance dma-buf export

If our exported dma-bufs are imported by another instance of our driver, that instance will typically have the imported dma-bufs locked during map_attachment(). But the exporter also locks the same reservation object in the map_dma_buf() callback, which leads to recursive locking.

Add a live selftest to catch this case, and as a workaround until we fully support dynamic import and export, declare the exporter dynamic by providing NOP pin() and unpin() functions. This means our map_dma_buf() callback will *always* get called locked, and by pinning unconditionally in i915_gem_map_dma_buf() we make sure we don't need to use the move_notify() functionality which is not yet implemented.

An interesting abuse of the interface, but it seems reasonable (for now) to me.

Reviewed-by: Michael J. Ruhl michael.j.ruhl@intel.com

Mike

Reported-by: Ruhl, Michael J michael.j.ruhl@intel.com Cc: Ruhl, Michael J michael.j.ruhl@intel.com Signed-off-by: Thomas Hellström thomas.hellstrom@linux.intel.com

---

drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c | 31 ++++++- .../drm/i915/gem/selftests/i915_gem_dmabuf.c | 81 ++++++++++++++++++- 2 files changed, 108 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c b/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c index 616c3a2f1baf..1d1eeb167d28 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c @@ -12,6 +12,8 @@ #include "i915_gem_object.h" #include "i915_scatterlist.h"

+I915_SELFTEST_DECLARE(static bool force_different_devices;)

static struct drm_i915_gem_object *dma_buf_to_obj(struct dma_buf *buf) { return to_intel_bo(buf->priv); @@ -25,7 +27,9 @@ static struct sg_table *i915_gem_map_dma_buf(struct dma_buf_attachment *attachme struct scatterlist *src, *dst; int ret, i;

• ret = i915_gem_object_pin_pages_unlocked(obj);

• assert_object_held(obj);
• ret = i915_gem_object_pin_pages(obj); if (ret) goto err;

@@ -168,6 +172,26 @@ static int i915_gem_end_cpu_access(struct dma_buf *dma_buf, enum dma_data_direct return err; }

+/*

• • As a workaround until we fully support dynamic import and export,
• • declare the exporter dynamic by providing NOP pin() and unpin()

functions.

• • This means our i915_gem_map_dma_buf() callback will *always* get

called

• • locked, and by pinning unconditionally in i915_gem_map_dma_buf() we

make

• • sure we don't need to use the move_notify() functionality which is
• • not yet implemented. Typically for the same-driver-another-instance case,
• • i915_gem_map_dma_buf() will be called at importer attach time and the
• • mapped sg_list will be cached by the dma-buf core for the
• • duration of the attachment.
• */

+static int i915_gem_dmabuf_pin(struct dma_buf_attachment *attach) +{

• return 0;

+}

+static void i915_gem_dmabuf_unpin(struct dma_buf_attachment *attach) +{ +}

static const struct dma_buf_ops i915_dmabuf_ops = { .map_dma_buf = i915_gem_map_dma_buf, .unmap_dma_buf = i915_gem_unmap_dma_buf, @@ -177,6 +201,8 @@ static const struct dma_buf_ops i915_dmabuf_ops = { .vunmap = i915_gem_dmabuf_vunmap, .begin_cpu_access = i915_gem_begin_cpu_access, .end_cpu_access = i915_gem_end_cpu_access,

• .pin = i915_gem_dmabuf_pin,
• .unpin = i915_gem_dmabuf_unpin,

};

struct dma_buf *i915_gem_prime_export(struct drm_gem_object *gem_obj, int flags) @@ -241,7 +267,8 @@ struct drm_gem_object *i915_gem_prime_import(struct drm_device *dev, if (dma_buf->ops == &i915_dmabuf_ops) { obj = dma_buf_to_obj(dma_buf); /* is it from our device? */

• if (obj->base.dev == dev) {
  

• if (obj->base.dev == dev &&
  

•     !I915_SELFTEST_ONLY(force_different_devices)) {
  /*
   * Importing dmabuf exported from out own gem
  

increases * refcount on gem itself instead of f_count of dmabuf. diff --git a/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c b/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c index dd74bc09ec88..24735d6c12a2 100644 --- a/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c +++ b/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c @@ -35,7 +35,7 @@ static int igt_dmabuf_export(void *arg) static int igt_dmabuf_import_self(void *arg) { struct drm_i915_private *i915 = arg;

• struct drm_i915_gem_object *obj;

• struct drm_i915_gem_object *obj, *import_obj; struct drm_gem_object *import; struct dma_buf *dmabuf; int err;

@@ -65,14 +65,90 @@ static int igt_dmabuf_import_self(void *arg) err = -EINVAL; goto out_import; }

• import_obj = to_intel_bo(import);

• i915_gem_object_lock(import_obj, NULL);

• err = ____i915_gem_object_get_pages(import_obj);

• i915_gem_object_unlock(import_obj);

• if (err) {

• pr_err("Same object dma-buf get_pages failed!\n");
  

• goto out_import;
  

• }

  err = 0;

out_import:

• i915_gem_object_put(to_intel_bo(import));

• i915_gem_object_put(import_obj);

+out_dmabuf:

• dma_buf_put(dmabuf);

+out:

• i915_gem_object_put(obj);
• return err;

+}

+static int igt_dmabuf_import_same_driver(void *arg) +{

• struct drm_i915_private *i915 = arg;
• struct drm_i915_gem_object *obj, *import_obj;
• struct drm_gem_object *import;
• struct dma_buf *dmabuf;
• int err;
• force_different_devices = true;
• obj = i915_gem_object_create_shmem(i915, PAGE_SIZE);
• if (IS_ERR(obj))

• goto out_ret;
  

• dmabuf = i915_gem_prime_export(&obj->base, 0);
• if (IS_ERR(dmabuf)) {

• pr_err("i915_gem_prime_export failed with err=%d\n",
  

•        (int)PTR_ERR(dmabuf));
  

• err = PTR_ERR(dmabuf);
  

• goto out;
  

• }
• import = i915_gem_prime_import(&i915->drm, dmabuf);
• if (IS_ERR(import)) {

• pr_err("i915_gem_prime_import failed with err=%d\n",
  

•        (int)PTR_ERR(import));
  

• err = PTR_ERR(import);
  

• goto out_dmabuf;
  

• }
• if (import == &obj->base) {

• pr_err("i915_gem_prime_import reused gem object!\n");
  

• err = -EINVAL;
  

• goto out_import;
  

• }
• import_obj = to_intel_bo(import);
• i915_gem_object_lock(import_obj, NULL);
• err = ____i915_gem_object_get_pages(import_obj);
• if (err) {

• pr_err("Different objects dma-buf get_pages failed!\n");
  

• i915_gem_object_unlock(import_obj);
  

• goto out_import;
  

• }
• /*

• * If the exported object is not in system memory, something
  

• * weird is going on. TODO: When p2p is supported, this is no
  

• * longer considered weird.
  

• */
  

• if (obj->mm.region != i915->mm.regions[INTEL_REGION_SMEM]) {

• pr_err("Exported dma-buf is not in system memory\n");
  

• err = -EINVAL;
  

• }
• i915_gem_object_unlock(import_obj);

+out_import:

• i915_gem_object_put(import_obj);

out_dmabuf: dma_buf_put(dmabuf); out: i915_gem_object_put(obj); +out_ret:

• force_different_devices = false; return err;

}

@@ -286,6 +362,7 @@ int i915_gem_dmabuf_live_selftests(struct drm_i915_private *i915) { static const struct i915_subtest tests[] = { SUBTEST(igt_dmabuf_export),

• SUBTEST(igt_dmabuf_import_same_driver),
  

  };

  return i915_subtests(tests, i915);

-- 2.31.1

On 6/29/21 10:43 AM, Daniel Vetter wrote:

On Mon, Jun 28, 2021 at 07:45:31PM +0000, Ruhl, Michael J wrote:

...

...

-----Original Message----- From: Thomas Hellström thomas.hellstrom@linux.intel.com Sent: Monday, June 28, 2021 10:46 AM To: intel-gfx@lists.freedesktop.org; dri-devel@lists.freedesktop.org Cc: Auld, Matthew matthew.auld@intel.com; maarten.lankhorst@linux.intel.com; Thomas Hellström thomas.hellstrom@linux.intel.com; Ruhl; Ruhl, Michael J michael.j.ruhl@intel.com Subject: [PATCH v3 4/5] drm/i915/gem: Fix same-driver-another-instance dma-buf export

If our exported dma-bufs are imported by another instance of our driver, that instance will typically have the imported dma-bufs locked during map_attachment(). But the exporter also locks the same reservation object in the map_dma_buf() callback, which leads to recursive locking.

Add a live selftest to catch this case, and as a workaround until we fully support dynamic import and export, declare the exporter dynamic by providing NOP pin() and unpin() functions. This means our map_dma_buf() callback will *always* get called locked, and by pinning unconditionally in i915_gem_map_dma_buf() we make sure we don't need to use the move_notify() functionality which is not yet implemented.

An interesting abuse of the interface, but it seems reasonable (for now) to me.

Hm I'm not sure this is the best interface abuse, since if we combine this with amdgpu it goes boom. Also I thought the dynamic stuff is optional (or is that only for the importer).

I'm not sure what would go wrong here when combined with amdgpu? I figure an amdgpu importer being dynamic, would expect to get notified using move_notify() on move, but that never happens since the exported bo is pinned. If it matters for interface abuse then I could add real implementations of pin() and unpin(). But choosing to not evict a mapped dma-buf must remain at the exporter's discretion and is not an interface abuse IMO.

Could you point me to a case that would not work with this code?

What I discussed a bit with Maarten on irc is to essentially emulate the rules of what a dynamic exporter would end up with with a non-dynamic importer: pin/unpin the buffer at attach/detach time. We could fake this in our attach/detach callbacks.

Yes, but that would only reimplement what's already in the dma-buf core? Since we're about to add a real and correct implementation of this, that sounds like a waste of time IMHO.

At least I don't think it's the locking changes that saves us here, but the caching of the sgt list in attach/detach.

Yes that saves us for the case of a non-locking non-dynamic importer, but for same-driver-another-instance, it's indeed the locking changes.

As long as we hand-roll that we should be fine. So hand-rolling that feels like the best option to make sure we're not making this worse, as long as we haven't fully validated the true dynamic importer _and_ exporter case.

/Thomas

Cheers, Daniel

...

Reviewed-by: Michael J. Ruhl michael.j.ruhl@intel.com

Mike

...

Reported-by: Ruhl, Michael J michael.j.ruhl@intel.com Cc: Ruhl, Michael J michael.j.ruhl@intel.com Signed-off-by: Thomas Hellström thomas.hellstrom@linux.intel.com

---

drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c | 31 ++++++- .../drm/i915/gem/selftests/i915_gem_dmabuf.c | 81 ++++++++++++++++++- 2 files changed, 108 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c b/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c index 616c3a2f1baf..1d1eeb167d28 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c @@ -12,6 +12,8 @@ #include "i915_gem_object.h" #include "i915_scatterlist.h"

+I915_SELFTEST_DECLARE(static bool force_different_devices;)

static struct drm_i915_gem_object *dma_buf_to_obj(struct dma_buf *buf) { return to_intel_bo(buf->priv); @@ -25,7 +27,9 @@ static struct sg_table *i915_gem_map_dma_buf(struct dma_buf_attachment *attachme struct scatterlist *src, *dst; int ret, i;

• ret = i915_gem_object_pin_pages_unlocked(obj);

• assert_object_held(obj);
• ret = i915_gem_object_pin_pages(obj); if (ret) goto err;

@@ -168,6 +172,26 @@ static int i915_gem_end_cpu_access(struct dma_buf *dma_buf, enum dma_data_direct return err; }

+/*

• • As a workaround until we fully support dynamic import and export,
• • declare the exporter dynamic by providing NOP pin() and unpin()

functions.

• • This means our i915_gem_map_dma_buf() callback will *always* get

called

• • locked, and by pinning unconditionally in i915_gem_map_dma_buf() we

make

• • sure we don't need to use the move_notify() functionality which is
• • not yet implemented. Typically for the same-driver-another-instance case,
• • i915_gem_map_dma_buf() will be called at importer attach time and the
• • mapped sg_list will be cached by the dma-buf core for the
• • duration of the attachment.
• */

+static int i915_gem_dmabuf_pin(struct dma_buf_attachment *attach) +{

• return 0;

+}

+static void i915_gem_dmabuf_unpin(struct dma_buf_attachment *attach) +{ +}

static const struct dma_buf_ops i915_dmabuf_ops = { .map_dma_buf = i915_gem_map_dma_buf, .unmap_dma_buf = i915_gem_unmap_dma_buf, @@ -177,6 +201,8 @@ static const struct dma_buf_ops i915_dmabuf_ops = { .vunmap = i915_gem_dmabuf_vunmap, .begin_cpu_access = i915_gem_begin_cpu_access, .end_cpu_access = i915_gem_end_cpu_access,

• .pin = i915_gem_dmabuf_pin,
• .unpin = i915_gem_dmabuf_unpin,

};

struct dma_buf *i915_gem_prime_export(struct drm_gem_object *gem_obj, int flags) @@ -241,7 +267,8 @@ struct drm_gem_object *i915_gem_prime_import(struct drm_device *dev, if (dma_buf->ops == &i915_dmabuf_ops) { obj = dma_buf_to_obj(dma_buf); /* is it from our device? */

• if (obj->base.dev == dev) {
  

• if (obj->base.dev == dev &&
  

•     !I915_SELFTEST_ONLY(force_different_devices)) {
  /*
   * Importing dmabuf exported from out own gem
  

increases * refcount on gem itself instead of f_count of dmabuf. diff --git a/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c b/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c index dd74bc09ec88..24735d6c12a2 100644 --- a/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c +++ b/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c @@ -35,7 +35,7 @@ static int igt_dmabuf_export(void *arg) static int igt_dmabuf_import_self(void *arg) { struct drm_i915_private *i915 = arg;

• struct drm_i915_gem_object *obj;

• struct drm_i915_gem_object *obj, *import_obj; struct drm_gem_object *import; struct dma_buf *dmabuf; int err;

@@ -65,14 +65,90 @@ static int igt_dmabuf_import_self(void *arg) err = -EINVAL; goto out_import; }

• import_obj = to_intel_bo(import);

• i915_gem_object_lock(import_obj, NULL);

• err = ____i915_gem_object_get_pages(import_obj);

• i915_gem_object_unlock(import_obj);

• if (err) {

• pr_err("Same object dma-buf get_pages failed!\n");
  

• goto out_import;
  

• }

  err = 0;

out_import:

• i915_gem_object_put(to_intel_bo(import));

• i915_gem_object_put(import_obj);

+out_dmabuf:

• dma_buf_put(dmabuf);

+out:

• i915_gem_object_put(obj);
• return err;

+}

+static int igt_dmabuf_import_same_driver(void *arg) +{

• struct drm_i915_private *i915 = arg;
• struct drm_i915_gem_object *obj, *import_obj;
• struct drm_gem_object *import;
• struct dma_buf *dmabuf;
• int err;
• force_different_devices = true;
• obj = i915_gem_object_create_shmem(i915, PAGE_SIZE);
• if (IS_ERR(obj))

• goto out_ret;
  

• dmabuf = i915_gem_prime_export(&obj->base, 0);
• if (IS_ERR(dmabuf)) {

• pr_err("i915_gem_prime_export failed with err=%d\n",
  

•        (int)PTR_ERR(dmabuf));
  

• err = PTR_ERR(dmabuf);
  

• goto out;
  

• }
• import = i915_gem_prime_import(&i915->drm, dmabuf);
• if (IS_ERR(import)) {

• pr_err("i915_gem_prime_import failed with err=%d\n",
  

•        (int)PTR_ERR(import));
  

• err = PTR_ERR(import);
  

• goto out_dmabuf;
  

• }
• if (import == &obj->base) {

• pr_err("i915_gem_prime_import reused gem object!\n");
  

• err = -EINVAL;
  

• goto out_import;
  

• }
• import_obj = to_intel_bo(import);
• i915_gem_object_lock(import_obj, NULL);
• err = ____i915_gem_object_get_pages(import_obj);
• if (err) {

• pr_err("Different objects dma-buf get_pages failed!\n");
  

• i915_gem_object_unlock(import_obj);
  

• goto out_import;
  

• }
• /*

• * If the exported object is not in system memory, something
  

• * weird is going on. TODO: When p2p is supported, this is no
  

• * longer considered weird.
  

• */
  

• if (obj->mm.region != i915->mm.regions[INTEL_REGION_SMEM]) {

• pr_err("Exported dma-buf is not in system memory\n");
  

• err = -EINVAL;
  

• }
• i915_gem_object_unlock(import_obj);

+out_import:

• i915_gem_object_put(import_obj);

out_dmabuf: dma_buf_put(dmabuf); out: i915_gem_object_put(obj); +out_ret:

• force_different_devices = false; return err;

}

@@ -286,6 +362,7 @@ int i915_gem_dmabuf_live_selftests(struct drm_i915_private *i915) { static const struct i915_subtest tests[] = { SUBTEST(igt_dmabuf_export),

• SUBTEST(igt_dmabuf_import_same_driver),
  

  };

  return i915_subtests(tests, i915);

-- 2.31.1

-----Original Message----- From: dri-devel dri-devel-bounces@lists.freedesktop.org On Behalf Of Thomas Hellström Sent: Monday, June 28, 2021 10:46 AM To: intel-gfx@lists.freedesktop.org; dri-devel@lists.freedesktop.org Cc: Thomas Hellström thomas.hellstrom@linux.intel.com; Auld, Matthew matthew.auld@intel.com Subject: [PATCH v3 5/5] drm/i915/gem: Migrate to system at dma-buf map time

Until we support p2p dma or as a complement to that, migrate data to system memory at dma-buf map time if possible.

v2:

• Rebase on dynamic exporter. Update the igt_dmabuf_import_same_driver

selftest to migrate if we are LMEM capable.

Signed-off-by: Thomas Hellström thomas.hellstrom@linux.intel.com

drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c | 7 ++++++- drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c | 4 +++- 2 files changed, 9 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c b/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c index 1d1eeb167d28..5207aa3af009 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_dmabuf.c @@ -29,7 +29,12 @@ static struct sg_table *i915_gem_map_dma_buf(struct dma_buf_attachment *attachme

assert_object_held(obj);

• ret = i915_gem_object_pin_pages(obj);

• if (!i915_gem_object_can_migrate(obj, INTEL_REGION_SMEM))

• return ERR_PTR(-EOPNOTSUPP);
  

• ret = i915_gem_object_migrate(obj, NULL, INTEL_REGION_SMEM);
• if (!ret)

• ret = i915_gem_object_pin_pages(obj);
  

  if (ret) goto err;

With the caveats of the previous patches this looks good.

Reviewed-by: Michael J. Ruhl michael.j.ruhl@intel.com

Mike

diff --git a/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c b/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c index 24735d6c12a2..b6278dba98d0 100644 --- a/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c +++ b/drivers/gpu/drm/i915/gem/selftests/i915_gem_dmabuf.c @@ -94,7 +94,9 @@ static int igt_dmabuf_import_same_driver(void *arg) int err;

force_different_devices = true;

• obj = i915_gem_object_create_shmem(i915, PAGE_SIZE);

• obj = i915_gem_object_create_lmem(i915, PAGE_SIZE, 0);
• if (IS_ERR(obj))

• obj = i915_gem_object_create_shmem(i915, PAGE_SIZE);
  

  if (IS_ERR(obj)) goto out_ret;

-- 2.31.1