drm_gem_object_funcs->vm_ops alone can't handle
everything mmap() needs. Add a new callback for it.
Signed-off-by: Gerd Hoffmann <[email protected]>
---
include/drm/drm_gem.h | 9 +++++++++
drivers/gpu/drm/drm_gem.c | 6 ++++++
2 files changed, 15 insertions(+)
diff --git a/include/drm/drm_gem.h b/include/drm/drm_gem.h
index ae693c0666cd..ee3c4ad742c6 100644
--- a/include/drm/drm_gem.h
+++ b/include/drm/drm_gem.h
@@ -150,6 +150,15 @@ struct drm_gem_object_funcs {
*/
void (*vunmap)(struct drm_gem_object *obj, void *vaddr);
+ /**
+ * @mmap:
+ *
+ * Called by drm_gem_mmap() for additional checks/setup.
+ *
+ * This callback is optional.
+ */
+ int (*mmap)(struct drm_gem_object *obj, struct vm_area_struct *vma);
+
/**
* @vm_ops:
*
diff --git a/drivers/gpu/drm/drm_gem.c b/drivers/gpu/drm/drm_gem.c
index afc38cece3f5..84db8de217e1 100644
--- a/drivers/gpu/drm/drm_gem.c
+++ b/drivers/gpu/drm/drm_gem.c
@@ -1105,6 +1105,8 @@ int drm_gem_mmap_obj(struct drm_gem_object *obj, unsigned long obj_size,
vma->vm_ops = obj->funcs->vm_ops;
else if (dev->driver->gem_vm_ops)
vma->vm_ops = dev->driver->gem_vm_ops;
+ else if (obj->funcs && obj->funcs->mmap)
+ /* obj->funcs->mmap must set vma->vm_ops */;
else
return -EINVAL;
@@ -1192,6 +1194,10 @@ int drm_gem_mmap(struct file *filp, struct vm_area_struct *vma)
ret = drm_gem_mmap_obj(obj, drm_vma_node_size(node) << PAGE_SHIFT,
vma);
+ if (ret == 0)
+ if (obj->funcs->mmap)
+ ret = obj->funcs->mmap(obj, vma);
+
drm_gem_object_put_unlocked(obj);
return ret;
--
2.18.1
On Thu, Aug 08, 2019 at 03:44:05PM +0200, Gerd Hoffmann wrote:
> drm_gem_object_funcs->vm_ops alone can't handle
> everything mmap() needs. Add a new callback for it.
>
> Signed-off-by: Gerd Hoffmann <[email protected]>
> ---
> include/drm/drm_gem.h | 9 +++++++++
> drivers/gpu/drm/drm_gem.c | 6 ++++++
> 2 files changed, 15 insertions(+)
>
> diff --git a/include/drm/drm_gem.h b/include/drm/drm_gem.h
> index ae693c0666cd..ee3c4ad742c6 100644
> --- a/include/drm/drm_gem.h
> +++ b/include/drm/drm_gem.h
> @@ -150,6 +150,15 @@ struct drm_gem_object_funcs {
> */
> void (*vunmap)(struct drm_gem_object *obj, void *vaddr);
>
> + /**
> + * @mmap:
> + *
> + * Called by drm_gem_mmap() for additional checks/setup.
> + *
> + * This callback is optional.
> + */
> + int (*mmap)(struct drm_gem_object *obj, struct vm_area_struct *vma);
I think if we do an mmap callback, it should replace all the mmap handling
(except the drm_gem_object_get) that drm_gem_mmap_obj does. So maybe
something like the below:
diff --git a/drivers/gpu/drm/drm_gem.c b/drivers/gpu/drm/drm_gem.c
index 6854f5867d51..e8b7779633dd 100644
--- a/drivers/gpu/drm/drm_gem.c
+++ b/drivers/gpu/drm/drm_gem.c
@@ -1104,17 +1104,22 @@ int drm_gem_mmap_obj(struct drm_gem_object *obj, unsigned long obj_size,
if (obj_size < vma->vm_end - vma->vm_start)
return -EINVAL;
- if (obj->funcs && obj->funcs->vm_ops)
- vma->vm_ops = obj->funcs->vm_ops;
- else if (dev->driver->gem_vm_ops)
- vma->vm_ops = dev->driver->gem_vm_ops;
- else
- return -EINVAL;
-
- vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
vma->vm_private_data = obj;
- vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
- vma->vm_page_prot = pgprot_decrypted(vma->vm_page_prot);
+
+ if (obj->funcs && obj->funcs->mmap)
+ obj->funcs->mmap(obj, vma);
+ else {
+ if (obj->funcs && obj->funcs->vm_ops)
+ vma->vm_ops = obj->funcs->vm_ops;
+ else if (dev->driver->gem_vm_ops)
+ vma->vm_ops = dev->driver->gem_vm_ops;
+ else
+ return -EINVAL;
+
+ vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
+ vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
+ vma->vm_page_prot = pgprot_decrypted(vma->vm_page_prot);
+ }
/* Take a ref for this mapping of the object, so that the fault
* handler can dereference the mmap offset's pointer to the object.
Since I remember quite a few discussions where the default vma flag
wrangling we're doing is seriously getting in the way of things too.
I think even better would be if this new ->mmap hook could also be used
directly by the dma-buf mmap code, without having to jump through hoops
creating a fake file and fake vma offset and everything. I think with that
we'd have a really solid case to add this ->mmap hook.
-Daniel
> +
> /**
> * @vm_ops:
> *
> diff --git a/drivers/gpu/drm/drm_gem.c b/drivers/gpu/drm/drm_gem.c
> index afc38cece3f5..84db8de217e1 100644
> --- a/drivers/gpu/drm/drm_gem.c
> +++ b/drivers/gpu/drm/drm_gem.c
> @@ -1105,6 +1105,8 @@ int drm_gem_mmap_obj(struct drm_gem_object *obj, unsigned long obj_size,
> vma->vm_ops = obj->funcs->vm_ops;
> else if (dev->driver->gem_vm_ops)
> vma->vm_ops = dev->driver->gem_vm_ops;
> + else if (obj->funcs && obj->funcs->mmap)
> + /* obj->funcs->mmap must set vma->vm_ops */;
> else
> return -EINVAL;
>
> @@ -1192,6 +1194,10 @@ int drm_gem_mmap(struct file *filp, struct vm_area_struct *vma)
> ret = drm_gem_mmap_obj(obj, drm_vma_node_size(node) << PAGE_SHIFT,
> vma);
>
> + if (ret == 0)
> + if (obj->funcs->mmap)
> + ret = obj->funcs->mmap(obj, vma);
> +
> drm_gem_object_put_unlocked(obj);
>
> return ret;
> --
> 2.18.1
>
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
Hi,
> I think if we do an mmap callback, it should replace all the mmap handling
> (except the drm_gem_object_get) that drm_gem_mmap_obj does. So maybe
> something like the below:
[ snip ]
> Since I remember quite a few discussions where the default vma flag
> wrangling we're doing is seriously getting in the way of things too.
Yep, makes sense.
> I think even better would be if this new ->mmap hook could also be used
> directly by the dma-buf mmap code, without having to jump through hoops
> creating a fake file and fake vma offset and everything. I think with that
> we'd have a really solid case to add this ->mmap hook.
Looks easy on a quick glance. So something like the patch below (quick
sketch, not tested yet other than compiling it)?
cheers,
Gerd
From c13b30d776fb593a03473fa3bc93baf470cba728 Mon Sep 17 00:00:00 2001
From: Gerd Hoffmann <[email protected]>
Date: Wed, 19 Jun 2019 14:26:51 +0200
Subject: [PATCH] drm: add mmap() to drm_gem_object_funcs
drm_gem_object_funcs->vm_ops alone can't handle everything which needs
to be done for mmap(), tweaking vm_flags for example. So add a new
mmap() callback to drm_gem_object_funcs where this code can go to.
Note that the vm_ops field is not used in case the mmap callback is
presnt, it is expected that the callback sets vma->vm_ops instead.
drm_gem_mmap_obj() will use the new callback for object specific mmap
setup. With this in place the need for driver-speific fops->mmap
callbacks goes away, drm_gem_mmap can be hooked instead.
drm_gem_prime_mmap() will use the new callback too to just mmap gem
objects directly instead of jumping though loops to make
drm_gem_object_lookup() and fops->mmap work.
Signed-off-by: Gerd Hoffmann <[email protected]>
---
include/drm/drm_gem.h | 14 ++++++++++++++
drivers/gpu/drm/drm_gem.c | 26 +++++++++++++++++---------
drivers/gpu/drm/drm_prime.c | 9 +++++++++
3 files changed, 40 insertions(+), 9 deletions(-)
diff --git a/include/drm/drm_gem.h b/include/drm/drm_gem.h
index 6aaba14f5972..e71f75a2ab57 100644
--- a/include/drm/drm_gem.h
+++ b/include/drm/drm_gem.h
@@ -150,6 +150,20 @@ struct drm_gem_object_funcs {
*/
void (*vunmap)(struct drm_gem_object *obj, void *vaddr);
+ /**
+ * @mmap:
+ *
+ * Handle mmap() of the gem object, setup vma accordingly.
+ *
+ * This callback is optional.
+ *
+ * The callback is used by by both drm_gem_mmap_obj() and
+ * drm_gem_prime_mmap(). When @mmap is present @vm_ops is not
+ * used, the @mmap callback must set vma->vm_ops instead.
+ *
+ */
+ int (*mmap)(struct drm_gem_object *obj, struct vm_area_struct *vma);
+
/**
* @vm_ops:
*
diff --git a/drivers/gpu/drm/drm_gem.c b/drivers/gpu/drm/drm_gem.c
index 6854f5867d51..aabde003ac4a 100644
--- a/drivers/gpu/drm/drm_gem.c
+++ b/drivers/gpu/drm/drm_gem.c
@@ -1099,22 +1099,30 @@ int drm_gem_mmap_obj(struct drm_gem_object *obj, unsigned long obj_size,
struct vm_area_struct *vma)
{
struct drm_device *dev = obj->dev;
+ int ret;
/* Check for valid size. */
if (obj_size < vma->vm_end - vma->vm_start)
return -EINVAL;
- if (obj->funcs && obj->funcs->vm_ops)
- vma->vm_ops = obj->funcs->vm_ops;
- else if (dev->driver->gem_vm_ops)
- vma->vm_ops = dev->driver->gem_vm_ops;
- else
- return -EINVAL;
+ if (obj->funcs && obj->funcs->mmap) {
+ ret = obj->funcs->mmap(obj, vma);
+ if (ret)
+ return ret;
+ } else {
+ if (obj->funcs && obj->funcs->vm_ops)
+ vma->vm_ops = obj->funcs->vm_ops;
+ else if (dev->driver->gem_vm_ops)
+ vma->vm_ops = dev->driver->gem_vm_ops;
+ else
+ return -EINVAL;
+
+ vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
+ vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
+ vma->vm_page_prot = pgprot_decrypted(vma->vm_page_prot);
+ }
- vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
vma->vm_private_data = obj;
- vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
- vma->vm_page_prot = pgprot_decrypted(vma->vm_page_prot);
/* Take a ref for this mapping of the object, so that the fault
* handler can dereference the mmap offset's pointer to the object.
diff --git a/drivers/gpu/drm/drm_prime.c b/drivers/gpu/drm/drm_prime.c
index 0a2316e0e812..0814211b0f3f 100644
--- a/drivers/gpu/drm/drm_prime.c
+++ b/drivers/gpu/drm/drm_prime.c
@@ -713,6 +713,15 @@ int drm_gem_prime_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)
struct file *fil;
int ret;
+ if (obj->funcs && obj->funcs->mmap) {
+ ret = obj->funcs->mmap(obj, vma);
+ if (ret)
+ return ret;
+ vma->vm_private_data = obj;
+ drm_gem_object_get(obj);
+ return 0;
+ }
+
priv = kzalloc(sizeof(*priv), GFP_KERNEL);
fil = kzalloc(sizeof(*fil), GFP_KERNEL);
if (!priv || !fil) {
--
2.18.1
On Fri, Sep 6, 2019 at 2:13 PM Gerd Hoffmann <[email protected]> wrote:
>
> Hi,
>
> > I think if we do an mmap callback, it should replace all the mmap handling
> > (except the drm_gem_object_get) that drm_gem_mmap_obj does. So maybe
> > something like the below:
>
> [ snip ]
>
> > Since I remember quite a few discussions where the default vma flag
> > wrangling we're doing is seriously getting in the way of things too.
>
> Yep, makes sense.
>
> > I think even better would be if this new ->mmap hook could also be used
> > directly by the dma-buf mmap code, without having to jump through hoops
> > creating a fake file and fake vma offset and everything. I think with that
> > we'd have a really solid case to add this ->mmap hook.
>
> Looks easy on a quick glance. So something like the patch below (quick
> sketch, not tested yet other than compiling it)?
Yup, looks good, and if it works I'm happy to smash r-b and a-b tags over this.
One thought below.
> cheers,
> Gerd
>
> From c13b30d776fb593a03473fa3bc93baf470cba728 Mon Sep 17 00:00:00 2001
> From: Gerd Hoffmann <[email protected]>
> Date: Wed, 19 Jun 2019 14:26:51 +0200
> Subject: [PATCH] drm: add mmap() to drm_gem_object_funcs
>
> drm_gem_object_funcs->vm_ops alone can't handle everything which needs
> to be done for mmap(), tweaking vm_flags for example. So add a new
> mmap() callback to drm_gem_object_funcs where this code can go to.
>
> Note that the vm_ops field is not used in case the mmap callback is
> presnt, it is expected that the callback sets vma->vm_ops instead.
>
> drm_gem_mmap_obj() will use the new callback for object specific mmap
> setup. With this in place the need for driver-speific fops->mmap
> callbacks goes away, drm_gem_mmap can be hooked instead.
>
> drm_gem_prime_mmap() will use the new callback too to just mmap gem
> objects directly instead of jumping though loops to make
> drm_gem_object_lookup() and fops->mmap work.
>
> Signed-off-by: Gerd Hoffmann <[email protected]>
> ---
> include/drm/drm_gem.h | 14 ++++++++++++++
> drivers/gpu/drm/drm_gem.c | 26 +++++++++++++++++---------
> drivers/gpu/drm/drm_prime.c | 9 +++++++++
> 3 files changed, 40 insertions(+), 9 deletions(-)
>
> diff --git a/include/drm/drm_gem.h b/include/drm/drm_gem.h
> index 6aaba14f5972..e71f75a2ab57 100644
> --- a/include/drm/drm_gem.h
> +++ b/include/drm/drm_gem.h
> @@ -150,6 +150,20 @@ struct drm_gem_object_funcs {
> */
> void (*vunmap)(struct drm_gem_object *obj, void *vaddr);
>
> + /**
> + * @mmap:
> + *
> + * Handle mmap() of the gem object, setup vma accordingly.
> + *
> + * This callback is optional.
> + *
> + * The callback is used by by both drm_gem_mmap_obj() and
> + * drm_gem_prime_mmap(). When @mmap is present @vm_ops is not
> + * used, the @mmap callback must set vma->vm_ops instead.
> + *
> + */
> + int (*mmap)(struct drm_gem_object *obj, struct vm_area_struct *vma);
> +
> /**
> * @vm_ops:
> *
> diff --git a/drivers/gpu/drm/drm_gem.c b/drivers/gpu/drm/drm_gem.c
> index 6854f5867d51..aabde003ac4a 100644
> --- a/drivers/gpu/drm/drm_gem.c
> +++ b/drivers/gpu/drm/drm_gem.c
> @@ -1099,22 +1099,30 @@ int drm_gem_mmap_obj(struct drm_gem_object *obj, unsigned long obj_size,
> struct vm_area_struct *vma)
> {
> struct drm_device *dev = obj->dev;
> + int ret;
>
> /* Check for valid size. */
> if (obj_size < vma->vm_end - vma->vm_start)
> return -EINVAL;
>
> - if (obj->funcs && obj->funcs->vm_ops)
> - vma->vm_ops = obj->funcs->vm_ops;
> - else if (dev->driver->gem_vm_ops)
> - vma->vm_ops = dev->driver->gem_vm_ops;
> - else
> - return -EINVAL;
> + if (obj->funcs && obj->funcs->mmap) {
> + ret = obj->funcs->mmap(obj, vma);
> + if (ret)
> + return ret;
> + } else {
> + if (obj->funcs && obj->funcs->vm_ops)
> + vma->vm_ops = obj->funcs->vm_ops;
> + else if (dev->driver->gem_vm_ops)
> + vma->vm_ops = dev->driver->gem_vm_ops;
> + else
> + return -EINVAL;
> +
> + vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
> + vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
> + vma->vm_page_prot = pgprot_decrypted(vma->vm_page_prot);
> + }
Totally unrelated discussion around HMM lead me to a bit a chase, and
realizing that we probably want a
WARN_ON(!(vma->vm_flags & VM_SPECIAL));
here, to make sure drivers set at least one of the "this is a special
vma, don't try to treat it like pagecache/anon memory". Just to be on
the safe side. Maybe we also want to keep the entire vma->vm_flags
assignment in the common code, but at least the WARN_ON would be a
good check I think. Maybe also check for VM_DONTEXPAND while at it
(that would also break things badly if it's not set). VM_DONTDUMP I
think is leftover from when drm drivers exposed register mmaps to
userspace.
Anyway, just some detail questions, I think this looks real good.
Thanks, Daniel
> - vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
> vma->vm_private_data = obj;
> - vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
> - vma->vm_page_prot = pgprot_decrypted(vma->vm_page_prot);
>
> /* Take a ref for this mapping of the object, so that the fault
> * handler can dereference the mmap offset's pointer to the object.
> diff --git a/drivers/gpu/drm/drm_prime.c b/drivers/gpu/drm/drm_prime.c
> index 0a2316e0e812..0814211b0f3f 100644
> --- a/drivers/gpu/drm/drm_prime.c
> +++ b/drivers/gpu/drm/drm_prime.c
> @@ -713,6 +713,15 @@ int drm_gem_prime_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)
> struct file *fil;
> int ret;
>
> + if (obj->funcs && obj->funcs->mmap) {
> + ret = obj->funcs->mmap(obj, vma);
> + if (ret)
> + return ret;
> + vma->vm_private_data = obj;
> + drm_gem_object_get(obj);
> + return 0;
> + }
> +
> priv = kzalloc(sizeof(*priv), GFP_KERNEL);
> fil = kzalloc(sizeof(*fil), GFP_KERNEL);
> if (!priv || !fil) {
> --
> 2.18.1
>
> _______________________________________________
> dri-devel mailing list
> [email protected]
> https://lists.freedesktop.org/mailman/listinfo/dri-devel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
Hi,
> > + vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
> > + vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
> > + vma->vm_page_prot = pgprot_decrypted(vma->vm_page_prot);
> > + }
>
> Totally unrelated discussion around HMM lead me to a bit a chase, and
> realizing that we probably want a
>
> WARN_ON(!(vma->vm_flags & VM_SPECIAL));
>
> here, to make sure drivers set at least one of the "this is a special
> vma, don't try to treat it like pagecache/anon memory". Just to be on
> the safe side. Maybe we also want to keep the entire vma->vm_flags
> assignment in the common code, but at least the WARN_ON would be a
> good check I think. Maybe also check for VM_DONTEXPAND while at it
Hmm. VM_SPECIAL is this:
#define VM_SPECIAL (VM_IO | VM_DONTEXPAND | VM_PFNMAP | VM_MIXEDMAP)
Requiring VM_DONTEXPAND makes sense for sure. Dunno about the other
ones. For drm_gem_vram_helper VM_IO + VM_PFNMAP are needed.
But we also have drm_gem_shmem_helper which backs objects with normal
pages. In fact drm_gem_shmem_mmap does this:
/* VM_PFNMAP was set by drm_gem_mmap() */
vma->vm_flags &= ~VM_PFNMAP;
vma->vm_flags |= VM_MIXEDMAP;
include/linux/mm.h isn't very helpful in explaining how VM_MIXEDMAP
should be used, only saying can be both "struct page" and pfnmap, so I'm
not sure why VM_MIXEDMAP is set here, it should always be "struct page"
for shmem, no?
cheers,
Gerd