| Message ID | 1350666204-8101-13-git-send-email-chris@chris-wilson.co.uk (mailing list archive) |
|---|---|
| State | Accepted |
| Delegated to: | Ben Widawsky |
| Headers | show |
On Fri, 19 Oct 2012 18:03:19 +0100 Chris Wilson <chris@chris-wilson.co.uk> wrote: > Allow for the creation of GEM objects backed by stolen memory. As these > are not backed by ordinary pages, we create a fake dma mapping and store > the address in the scatterlist rather than obj->pages. > > v2: Mark _i915_gem_object_create_stolen() as static, as noticed by Jesse > Barnes. > > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> > Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org> Deferring on an r-b for now until I understand the point of most of this patch. Some comments below however. > --- > drivers/gpu/drm/i915/i915_drv.h | 3 + > drivers/gpu/drm/i915/i915_gem.c | 1 + > drivers/gpu/drm/i915/i915_gem_stolen.c | 122 ++++++++++++++++++++++++++++++++ > 3 files changed, 126 insertions(+) > > diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h > index 6b893c7..5a0e0cd 100644 > --- a/drivers/gpu/drm/i915/i915_drv.h > +++ b/drivers/gpu/drm/i915/i915_drv.h > @@ -1526,6 +1526,9 @@ int i915_gem_init_stolen(struct drm_device *dev); > int i915_gem_stolen_setup_compression(struct drm_device *dev); > void i915_gem_stolen_cleanup_compression(struct drm_device *dev); > void i915_gem_cleanup_stolen(struct drm_device *dev); > +struct drm_i915_gem_object * > +i915_gem_object_create_stolen(struct drm_device *dev, u32 size); > +void i915_gem_object_release_stolen(struct drm_i915_gem_object *obj); > > /* i915_gem_tiling.c */ > void i915_gem_detect_bit_6_swizzle(struct drm_device *dev); > diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c > index d75c5c3..0349899 100644 > --- a/drivers/gpu/drm/i915/i915_gem.c > +++ b/drivers/gpu/drm/i915/i915_gem.c > @@ -3857,6 +3857,7 @@ void i915_gem_free_object(struct drm_gem_object *gem_obj) > obj->pages_pin_count = 0; > i915_gem_object_put_pages(obj); > i915_gem_object_free_mmap_offset(obj); > + i915_gem_object_release_stolen(obj); > > BUG_ON(obj->pages); > > diff --git a/drivers/gpu/drm/i915/i915_gem_stolen.c b/drivers/gpu/drm/i915/i915_gem_stolen.c > index 68ef22a..86c4af4 100644 > --- a/drivers/gpu/drm/i915/i915_gem_stolen.c > +++ b/drivers/gpu/drm/i915/i915_gem_stolen.c > @@ -204,3 +204,125 @@ int i915_gem_init_stolen(struct drm_device *dev) > > return 0; > } > + > +static struct sg_table * > +i915_pages_create_for_stolen(struct drm_device *dev, > + u32 offset, u32 size) > +{ > + struct drm_i915_private *dev_priv = dev->dev_private; > + struct sg_table *st; > + struct scatterlist *sg; > + BUG_ON(offset + size <= dev_priv->mm.gtt->stolen_size); > + /* We hide that we have no struct page backing our stolen object > + * by wrapping the contiguous physical allocation with a fake > + * dma mapping in a single scatterlist. > + */ > + > + st = kmalloc(sizeof(*st), GFP_KERNEL); > + if (st == NULL) > + return NULL; > + > + if (!sg_alloc_table(st, 1, GFP_KERNEL)) { > + kfree(st); > + return NULL; > + } > + > + sg = st->sgl; > + sg->offset = offset; > + sg->length = size; > + > + sg_dma_address(sg) = dev_priv->mm.stolen_base + offset; > + sg_dma_len(sg) = size; > + Do we want to make stolen_base a dma_addr_t (or at least typecast it)? > + return st; > +} > + > +static int i915_gem_object_get_pages_stolen(struct drm_i915_gem_object *obj) > +{ > + BUG(); > + return -EINVAL; > +} > + __noreturn, or maybe just make .get_pages = NULL, and do the check in the upper layer get_pages? > +static void i915_gem_object_put_pages_stolen(struct drm_i915_gem_object *obj) > +{ > + /* Should only be called during free */ > + sg_free_table(obj->pages); > + kfree(obj->pages); > +} > + > +static const struct drm_i915_gem_object_ops i915_gem_object_stolen_ops = { > + .get_pages = i915_gem_object_get_pages_stolen, > + .put_pages = i915_gem_object_put_pages_stolen, > +}; > + > +static struct drm_i915_gem_object * > +_i915_gem_object_create_stolen(struct drm_device *dev, > + struct drm_mm_node *stolen) > +{ > + struct drm_i915_gem_object *obj; > + > + obj = kzalloc(sizeof(*obj), GFP_KERNEL); > + if (obj == NULL) > + return NULL; > + > + if (drm_gem_private_object_init(dev, &obj->base, stolen->size)) > + goto cleanup; > + > + i915_gem_object_init(obj, &i915_gem_object_stolen_ops); > + > + obj->pages = i915_pages_create_for_stolen(dev, > + stolen->start, stolen->size); > + if (obj->pages == NULL) > + goto cleanup; > + > + obj->has_dma_mapping = true; > + obj->pages_pin_count = 1; > + obj->stolen = stolen; > + > + obj->base.write_domain = I915_GEM_DOMAIN_GTT; > + obj->base.read_domains = I915_GEM_DOMAIN_GTT; > + obj->cache_level = I915_CACHE_NONE; > + > + return obj; > + > +cleanup: > + kfree(obj); > + return NULL; > +} > + > +struct drm_i915_gem_object * > +i915_gem_object_create_stolen(struct drm_device *dev, u32 size) > +{ > + struct drm_i915_private *dev_priv = dev->dev_private; > + struct drm_i915_gem_object *obj; > + struct drm_mm_node *stolen; > + > + if (dev_priv->mm.stolen_base == 0) > + return 0; > + > + DRM_DEBUG_KMS("creating stolen object: size=%x\n", size); > + if (size == 0) > + return NULL; > + > + stolen = drm_mm_search_free(&dev_priv->mm.stolen, size, 4096, 0); > + if (stolen) > + stolen = drm_mm_get_block(stolen, size, 4096); > + if (stolen == NULL) > + return NULL; Could probably do slightly better here with ERR_PTR(-ENOMEM) but since we don't do that elsewhere, I guess it doesn't matter. > + > + obj = _i915_gem_object_create_stolen(dev, stolen); > + if (obj) > + return obj; > + > + drm_mm_put_block(stolen); > + return NULL; Similarly here maybe use PTR_ERR(obj) and return something meaningful in _i915_gem_object_create_stolen. Since just about everything is more or less, ENOMEM, maybe it doesn't matter. > +} > + > +void > +i915_gem_object_release_stolen(struct drm_i915_gem_object *obj) > +{ > + if (obj->stolen) { > + drm_mm_put_block(obj->stolen); > + obj->stolen = NULL; > + } > +}
On Mon, 5 Nov 2012 16:32:26 +0000, Ben Widawsky <ben@bwidawsk.net> wrote: > On Fri, 19 Oct 2012 18:03:19 +0100 > Chris Wilson <chris@chris-wilson.co.uk> wrote: > > > Allow for the creation of GEM objects backed by stolen memory. As these > > are not backed by ordinary pages, we create a fake dma mapping and store > > the address in the scatterlist rather than obj->pages. > > > > v2: Mark _i915_gem_object_create_stolen() as static, as noticed by Jesse > > Barnes. > > > > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> > > Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org> > > Deferring on an r-b for now until I understand the point of most of this > patch. The stolen support is a precursor for fastboot, where we need to wrap the allocations made by the BIOS from the stolen memory and reuse that for our own framebuffers. > > + struct scatterlist *sg; > > + > > BUG_ON(offset + size <= dev_priv->mm.gtt->stolen_size); Done with a minor amendment. > > + /* We hide that we have no struct page backing our stolen object > > + * by wrapping the contiguous physical allocation with a fake > > + * dma mapping in a single scatterlist. > > + */ > > + > > + st = kmalloc(sizeof(*st), GFP_KERNEL); > > + if (st == NULL) > > + return NULL; > > + > > + if (!sg_alloc_table(st, 1, GFP_KERNEL)) { Fixed. > > + kfree(st); > > + return NULL; > > + } > > + > > + sg = st->sgl; > > + sg->offset = offset; > > + sg->length = size; > > + > > + sg_dma_address(sg) = dev_priv->mm.stolen_base + offset; > > + sg_dma_len(sg) = size; > > + > > Do we want to make stolen_base a dma_addr_t (or at least typecast it)? Interesting enough, the current FBC registers are limited to only using 32bit addresses, so stolen_base atm is not technically a dma_addr_t. Maybe I'm picking hairs. :) > > + return st; > > +} > > + > > +static int i915_gem_object_get_pages_stolen(struct drm_i915_gem_object *obj) > > +{ > > + BUG(); > > + return -EINVAL; > > +} > > + > > __noreturn, or maybe just make .get_pages = NULL, and do the check in > the upper layer get_pages? I refer you to http://lwn.net/Articles/336262/ where the argument is put forth that default no-op functions are preferrable in most cases to interpretting special NULL vfuncs. We have adopted this elsewhere in i915.ko to good effect. > > + stolen = drm_mm_search_free(&dev_priv->mm.stolen, size, 4096, 0); > > + if (stolen) > > + stolen = drm_mm_get_block(stolen, size, 4096); > > + if (stolen == NULL) > > + return NULL; > > Could probably do slightly better here with ERR_PTR(-ENOMEM) but since > we don't do that elsewhere, I guess it doesn't matter. I was tempted - it would have just looked odd as being the only create routine to do so. :) -Chris
On Mon, 05 Nov 2012 16:59:44 +0000 Chris Wilson <chris@chris-wilson.co.uk> wrote: > On Mon, 5 Nov 2012 16:32:26 +0000, Ben Widawsky <ben@bwidawsk.net> wrote: > > On Fri, 19 Oct 2012 18:03:19 +0100 > > Chris Wilson <chris@chris-wilson.co.uk> wrote: > > > > > Allow for the creation of GEM objects backed by stolen memory. As these > > > are not backed by ordinary pages, we create a fake dma mapping and store > > > the address in the scatterlist rather than obj->pages. > > > > > > v2: Mark _i915_gem_object_create_stolen() as static, as noticed by Jesse > > > Barnes. > > > > > > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> > > > Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org> > > > > Deferring on an r-b for now until I understand the point of most of this > > patch. > > The stolen support is a precursor for fastboot, where we need to wrap > the allocations made by the BIOS from the stolen memory and reuse that > for our own framebuffers. > For some reason I thought this should be simpler than it is, but Jesse has successfully convinced me otherwise. Reviewed-by: Ben Widawsky <ben@bwidawsk.net> > > > + struct scatterlist *sg; > > > + > > > > BUG_ON(offset + size <= dev_priv->mm.gtt->stolen_size); > > Done with a minor amendment. > > > > + /* We hide that we have no struct page backing our stolen object > > > + * by wrapping the contiguous physical allocation with a fake > > > + * dma mapping in a single scatterlist. > > > + */ > > > + > > > + st = kmalloc(sizeof(*st), GFP_KERNEL); > > > + if (st == NULL) > > > + return NULL; > > > + > > > + if (!sg_alloc_table(st, 1, GFP_KERNEL)) { > Fixed. > > > > + kfree(st); > > > + return NULL; > > > + } > > > + > > > + sg = st->sgl; > > > + sg->offset = offset; > > > + sg->length = size; > > > + > > > + sg_dma_address(sg) = dev_priv->mm.stolen_base + offset; > > > + sg_dma_len(sg) = size; > > > + > > > > Do we want to make stolen_base a dma_addr_t (or at least typecast it)? > > Interesting enough, the current FBC registers are limited to only using > 32bit addresses, so stolen_base atm is not technically a dma_addr_t. > Maybe I'm picking hairs. :) > > > > + return st; > > > +} > > > + > > > +static int i915_gem_object_get_pages_stolen(struct drm_i915_gem_object *obj) > > > +{ > > > + BUG(); > > > + return -EINVAL; > > > +} > > > + > > > > __noreturn, or maybe just make .get_pages = NULL, and do the check in > > the upper layer get_pages? > > I refer you to http://lwn.net/Articles/336262/ where the argument is put > forth that default no-op functions are preferrable in most cases to > interpretting special NULL vfuncs. We have adopted this elsewhere in > i915.ko to good effect. > > > > + stolen = drm_mm_search_free(&dev_priv->mm.stolen, size, 4096, 0); > > > + if (stolen) > > > + stolen = drm_mm_get_block(stolen, size, 4096); > > > + if (stolen == NULL) > > > + return NULL; > > > > Could probably do slightly better here with ERR_PTR(-ENOMEM) but since > > we don't do that elsewhere, I guess it doesn't matter. > > I was tempted - it would have just looked odd as being the only create > routine to do so. :) > -Chris >
diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h index 6b893c7..5a0e0cd 100644 --- a/drivers/gpu/drm/i915/i915_drv.h +++ b/drivers/gpu/drm/i915/i915_drv.h @@ -1526,6 +1526,9 @@ int i915_gem_init_stolen(struct drm_device *dev); int i915_gem_stolen_setup_compression(struct drm_device *dev); void i915_gem_stolen_cleanup_compression(struct drm_device *dev); void i915_gem_cleanup_stolen(struct drm_device *dev); +struct drm_i915_gem_object * +i915_gem_object_create_stolen(struct drm_device *dev, u32 size); +void i915_gem_object_release_stolen(struct drm_i915_gem_object *obj); /* i915_gem_tiling.c */ void i915_gem_detect_bit_6_swizzle(struct drm_device *dev); diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c index d75c5c3..0349899 100644 --- a/drivers/gpu/drm/i915/i915_gem.c +++ b/drivers/gpu/drm/i915/i915_gem.c @@ -3857,6 +3857,7 @@ void i915_gem_free_object(struct drm_gem_object *gem_obj) obj->pages_pin_count = 0; i915_gem_object_put_pages(obj); i915_gem_object_free_mmap_offset(obj); + i915_gem_object_release_stolen(obj); BUG_ON(obj->pages); diff --git a/drivers/gpu/drm/i915/i915_gem_stolen.c b/drivers/gpu/drm/i915/i915_gem_stolen.c index 68ef22a..86c4af4 100644 --- a/drivers/gpu/drm/i915/i915_gem_stolen.c +++ b/drivers/gpu/drm/i915/i915_gem_stolen.c @@ -204,3 +204,125 @@ int i915_gem_init_stolen(struct drm_device *dev) return 0; } + +static struct sg_table * +i915_pages_create_for_stolen(struct drm_device *dev, + u32 offset, u32 size) +{ + struct drm_i915_private *dev_priv = dev->dev_private; + struct sg_table *st; + struct scatterlist *sg; + + /* We hide that we have no struct page backing our stolen object + * by wrapping the contiguous physical allocation with a fake + * dma mapping in a single scatterlist. + */ + + st = kmalloc(sizeof(*st), GFP_KERNEL); + if (st == NULL) + return NULL; + + if (!sg_alloc_table(st, 1, GFP_KERNEL)) { + kfree(st); + return NULL; + } + + sg = st->sgl; + sg->offset = offset; + sg->length = size; + + sg_dma_address(sg) = dev_priv->mm.stolen_base + offset; + sg_dma_len(sg) = size; + + return st; +} + +static int i915_gem_object_get_pages_stolen(struct drm_i915_gem_object *obj) +{ + BUG(); + return -EINVAL; +} + +static void i915_gem_object_put_pages_stolen(struct drm_i915_gem_object *obj) +{ + /* Should only be called during free */ + sg_free_table(obj->pages); + kfree(obj->pages); +} + +static const struct drm_i915_gem_object_ops i915_gem_object_stolen_ops = { + .get_pages = i915_gem_object_get_pages_stolen, + .put_pages = i915_gem_object_put_pages_stolen, +}; + +static struct drm_i915_gem_object * +_i915_gem_object_create_stolen(struct drm_device *dev, + struct drm_mm_node *stolen) +{ + struct drm_i915_gem_object *obj; + + obj = kzalloc(sizeof(*obj), GFP_KERNEL); + if (obj == NULL) + return NULL; + + if (drm_gem_private_object_init(dev, &obj->base, stolen->size)) + goto cleanup; + + i915_gem_object_init(obj, &i915_gem_object_stolen_ops); + + obj->pages = i915_pages_create_for_stolen(dev, + stolen->start, stolen->size); + if (obj->pages == NULL) + goto cleanup; + + obj->has_dma_mapping = true; + obj->pages_pin_count = 1; + obj->stolen = stolen; + + obj->base.write_domain = I915_GEM_DOMAIN_GTT; + obj->base.read_domains = I915_GEM_DOMAIN_GTT; + obj->cache_level = I915_CACHE_NONE; + + return obj; + +cleanup: + kfree(obj); + return NULL; +} + +struct drm_i915_gem_object * +i915_gem_object_create_stolen(struct drm_device *dev, u32 size) +{ + struct drm_i915_private *dev_priv = dev->dev_private; + struct drm_i915_gem_object *obj; + struct drm_mm_node *stolen; + + if (dev_priv->mm.stolen_base == 0) + return 0; + + DRM_DEBUG_KMS("creating stolen object: size=%x\n", size); + if (size == 0) + return NULL; + + stolen = drm_mm_search_free(&dev_priv->mm.stolen, size, 4096, 0); + if (stolen) + stolen = drm_mm_get_block(stolen, size, 4096); + if (stolen == NULL) + return NULL; + + obj = _i915_gem_object_create_stolen(dev, stolen); + if (obj) + return obj; + + drm_mm_put_block(stolen); + return NULL; +} + +void +i915_gem_object_release_stolen(struct drm_i915_gem_object *obj) +{ + if (obj->stolen) { + drm_mm_put_block(obj->stolen); + obj->stolen = NULL; + } +}