| Message ID | 11eb10d95109abc104beb94c0409d49644ea8517.1458140105.git.robin.murphy@arm.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Wed, Mar 16, 2016 at 02:57:49PM +0000, Robin Murphy wrote: > In the absence of an fb_mmap callback, the fbdev code falls back to a > naive implementation which relies upon the DMA address being the same > as the physical address, and the buffer being physically contiguous > from there. Whilst this often holds for standard CMA allocations via > the platform's regular DMA ops, if the allocation is provided by an > IOMMU then such assumptions can fall apart spectacularly. > > To resolve this, reroute the fb_mmap call to the appropriate DMA API > implementation, as per the other cma_helper calls. > > Signed-off-by: Robin Murphy <robin.murphy@arm.com> > --- > > Hi dri-devel, > > This is an empirical fix for something I tickled via the newly-added > ARM HDLCD driver on a Juno platform - I have no idea whatsoever about > how "proper" it is in terms of the DRM infrastructure, so feel free to > treat this as a bug report rather than an actual patch if appropriate ;) I think the best case would be if we could have a generic fbdev helper that remaps to dumb mmap support. But that's a bit tricky to pull of: 1. from fb_info we can get at the fbdev drm_framebuffer. 2. from a drm_framebuffer we can get at the underlying backing storage object using fb->funcs->get_handle. 3. With that handle we could go into the dumb mmap support (using als the vma) and create the mmap. Except that ->get_handle needs a file_priv, and that just exist for the fbdev emulation kms client. I guess we could fix that by creating a minimal fake drm file_priv for the fbdev emulation (and treat it more like any other kms client), but I think that's way too much work when this simple patch here gets the job done. Long story short, assuming this doesn't break other cma-helper using drivers: Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch> > > drivers/gpu/drm/drm_fb_cma_helper.c | 8 ++++++++ > 1 file changed, 8 insertions(+) > > diff --git a/drivers/gpu/drm/drm_fb_cma_helper.c b/drivers/gpu/drm/drm_fb_cma_helper.c > index bb88e3d..117d7ea 100644 > --- a/drivers/gpu/drm/drm_fb_cma_helper.c > +++ b/drivers/gpu/drm/drm_fb_cma_helper.c > @@ -23,6 +23,7 @@ > #include <drm/drm_crtc_helper.h> > #include <drm/drm_gem_cma_helper.h> > #include <drm/drm_fb_cma_helper.h> > +#include <linux/dma-mapping.h> > #include <linux/module.h> > > struct drm_fb_cma { > @@ -221,6 +222,12 @@ int drm_fb_cma_debugfs_show(struct seq_file *m, void *arg) > EXPORT_SYMBOL_GPL(drm_fb_cma_debugfs_show); > #endif > > +static int drm_fb_cma_mmap(struct fb_info *info, struct vm_area_struct *vma) > +{ > + return dma_mmap_writecombine(info->device, vma, info->screen_base, > + info->fix.smem_start, info->fix.smem_len); > +} > + > static struct fb_ops drm_fbdev_cma_ops = { > .owner = THIS_MODULE, > .fb_fillrect = drm_fb_helper_sys_fillrect, > @@ -231,6 +238,7 @@ static struct fb_ops drm_fbdev_cma_ops = { > .fb_blank = drm_fb_helper_blank, > .fb_pan_display = drm_fb_helper_pan_display, > .fb_setcmap = drm_fb_helper_setcmap, > + .fb_mmap = drm_fb_cma_mmap, > }; > > static int drm_fbdev_cma_create(struct drm_fb_helper *helper, > -- > 2.7.3.dirty > > _______________________________________________ > dri-devel mailing list > dri-devel@lists.freedesktop.org > https://lists.freedesktop.org/mailman/listinfo/dri-devel
On Wed, Mar 16, 2016 at 04:28:25PM +0100, Daniel Vetter wrote: > On Wed, Mar 16, 2016 at 02:57:49PM +0000, Robin Murphy wrote: > > In the absence of an fb_mmap callback, the fbdev code falls back to a > > naive implementation which relies upon the DMA address being the same > > as the physical address, and the buffer being physically contiguous > > from there. Whilst this often holds for standard CMA allocations via > > the platform's regular DMA ops, if the allocation is provided by an > > IOMMU then such assumptions can fall apart spectacularly. > > > > To resolve this, reroute the fb_mmap call to the appropriate DMA API > > implementation, as per the other cma_helper calls. > > > > Signed-off-by: Robin Murphy <robin.murphy@arm.com> > > --- > > > > Hi dri-devel, > > > > This is an empirical fix for something I tickled via the newly-added > > ARM HDLCD driver on a Juno platform - I have no idea whatsoever about > > how "proper" it is in terms of the DRM infrastructure, so feel free to > > treat this as a bug report rather than an actual patch if appropriate ;) > > I think the best case would be if we could have a generic fbdev helper > that remaps to dumb mmap support. But that's a bit tricky to pull of: > 1. from fb_info we can get at the fbdev drm_framebuffer. > 2. from a drm_framebuffer we can get at the underlying backing storage > object using fb->funcs->get_handle. > 3. With that handle we could go into the dumb mmap support (using als the > vma) and create the mmap. > > Except that ->get_handle needs a file_priv, and that just exist for the > fbdev emulation kms client. I guess we could fix that by creating a > minimal fake drm file_priv for the fbdev emulation (and treat it more like > any other kms client), but I think that's way too much work when this > simple patch here gets the job done. I think first, a different question needs to be answered: include/uapi/linux/fb.h: struct fb_fix_screeninfo { char id[16]; /* identification string eg "TT Builtin" */ unsigned long smem_start; /* Start of frame buffer mem */ /* (physical address) */ Should a DMA address be exposed through smem_start, rather than a physical address as the long-standing documentation quoted above has stated? Is it, in fact, a driver bug to store something that isn't a physical address there?
On 16/03/16 19:14, Russell King - ARM Linux wrote: > On Wed, Mar 16, 2016 at 04:28:25PM +0100, Daniel Vetter wrote: >> On Wed, Mar 16, 2016 at 02:57:49PM +0000, Robin Murphy wrote: >>> In the absence of an fb_mmap callback, the fbdev code falls back to a >>> naive implementation which relies upon the DMA address being the same >>> as the physical address, and the buffer being physically contiguous >>> from there. Whilst this often holds for standard CMA allocations via >>> the platform's regular DMA ops, if the allocation is provided by an >>> IOMMU then such assumptions can fall apart spectacularly. >>> >>> To resolve this, reroute the fb_mmap call to the appropriate DMA API >>> implementation, as per the other cma_helper calls. >>> >>> Signed-off-by: Robin Murphy <robin.murphy@arm.com> >>> --- >>> >>> Hi dri-devel, >>> >>> This is an empirical fix for something I tickled via the newly-added >>> ARM HDLCD driver on a Juno platform - I have no idea whatsoever about >>> how "proper" it is in terms of the DRM infrastructure, so feel free to >>> treat this as a bug report rather than an actual patch if appropriate ;) >> >> I think the best case would be if we could have a generic fbdev helper >> that remaps to dumb mmap support. But that's a bit tricky to pull of: >> 1. from fb_info we can get at the fbdev drm_framebuffer. >> 2. from a drm_framebuffer we can get at the underlying backing storage >> object using fb->funcs->get_handle. >> 3. With that handle we could go into the dumb mmap support (using als the >> vma) and create the mmap. >> >> Except that ->get_handle needs a file_priv, and that just exist for the >> fbdev emulation kms client. I guess we could fix that by creating a >> minimal fake drm file_priv for the fbdev emulation (and treat it more like >> any other kms client), but I think that's way too much work when this >> simple patch here gets the job done. > > I think first, a different question needs to be answered: > > include/uapi/linux/fb.h: > > struct fb_fix_screeninfo { > char id[16]; /* identification string eg "TT Builtin" */ > unsigned long smem_start; /* Start of frame buffer mem */ > /* (physical address) */ > > Should a DMA address be exposed through smem_start, rather than a > physical address as the long-standing documentation quoted above > has stated? > > Is it, in fact, a driver bug to store something that isn't a physical > address there? We could also go into whether it's right to store even a physical address in something which isn't necessarily big enough... After the time I spent in the bowels of the fbdev code figuring out this crash, I fear that if we dig too deep we may awaken something in the darkness ;) Robin.
diff --git a/drivers/gpu/drm/drm_fb_cma_helper.c b/drivers/gpu/drm/drm_fb_cma_helper.c index bb88e3d..117d7ea 100644 --- a/drivers/gpu/drm/drm_fb_cma_helper.c +++ b/drivers/gpu/drm/drm_fb_cma_helper.c @@ -23,6 +23,7 @@ #include <drm/drm_crtc_helper.h> #include <drm/drm_gem_cma_helper.h> #include <drm/drm_fb_cma_helper.h> +#include <linux/dma-mapping.h> #include <linux/module.h> struct drm_fb_cma { @@ -221,6 +222,12 @@ int drm_fb_cma_debugfs_show(struct seq_file *m, void *arg) EXPORT_SYMBOL_GPL(drm_fb_cma_debugfs_show); #endif +static int drm_fb_cma_mmap(struct fb_info *info, struct vm_area_struct *vma) +{ + return dma_mmap_writecombine(info->device, vma, info->screen_base, + info->fix.smem_start, info->fix.smem_len); +} + static struct fb_ops drm_fbdev_cma_ops = { .owner = THIS_MODULE, .fb_fillrect = drm_fb_helper_sys_fillrect, @@ -231,6 +238,7 @@ static struct fb_ops drm_fbdev_cma_ops = { .fb_blank = drm_fb_helper_blank, .fb_pan_display = drm_fb_helper_pan_display, .fb_setcmap = drm_fb_helper_setcmap, + .fb_mmap = drm_fb_cma_mmap, }; static int drm_fbdev_cma_create(struct drm_fb_helper *helper,
In the absence of an fb_mmap callback, the fbdev code falls back to a naive implementation which relies upon the DMA address being the same as the physical address, and the buffer being physically contiguous from there. Whilst this often holds for standard CMA allocations via the platform's regular DMA ops, if the allocation is provided by an IOMMU then such assumptions can fall apart spectacularly. To resolve this, reroute the fb_mmap call to the appropriate DMA API implementation, as per the other cma_helper calls. Signed-off-by: Robin Murphy <robin.murphy@arm.com> --- Hi dri-devel, This is an empirical fix for something I tickled via the newly-added ARM HDLCD driver on a Juno platform - I have no idea whatsoever about how "proper" it is in terms of the DRM infrastructure, so feel free to treat this as a bug report rather than an actual patch if appropriate ;) drivers/gpu/drm/drm_fb_cma_helper.c | 8 ++++++++ 1 file changed, 8 insertions(+)