[Intel-gfx] [PATCH 3/3] drm/i915: mark GEM objects as dirty when written by the CPU

[Daniel Vetter]

On Tue, Dec 08, 2015 at 06:24:40PM +0000, Dave Gordon wrote:
> On 08/12/15 17:03, Chris Wilson wrote:
> >On Tue, Dec 08, 2015 at 04:51:18PM +0000, Dave Gordon wrote:
> >>This patch covers a couple more places where a GEM object is (or may be)
> >>modified by means of CPU writes, and should therefore be marked dirty to
> >>ensure that the changes are not lost in the evenof of the object is
> >>evicted under memory pressure.
> >>
> >>It may be possible to optimise these paths later, by marking only
> >>specific pages of the object as dirty (for objects backed by shmfs
> >>pages); but for now let's ensure correctness by dirtying the whole
> >>object.
> >>
> >>Signed-off-by: Dave Gordon <david.s.gordon at intel.com>
> >>Cc: Chris Wilson <chris at chris-wilson.co.uk>
> >>---
> >>  drivers/gpu/drm/i915/i915_gem.c        | 4 +++-
> >>  drivers/gpu/drm/i915/i915_gem_dmabuf.c | 2 ++
> >>  2 files changed, 5 insertions(+), 1 deletion(-)
> >>
> >>diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> >>index 12f68f4..36b9539 100644
> >>--- a/drivers/gpu/drm/i915/i915_gem.c
> >>+++ b/drivers/gpu/drm/i915/i915_gem.c
> >>@@ -937,7 +937,6 @@ i915_gem_shmem_pwrite(struct drm_device *dev,
> >>  	i915_gem_object_pin_pages(obj);
> >>
> >>  	offset = args->offset;
> >>-	obj->dirty = 1;
> >>
> >>  	for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents,
> >>  			 offset >> PAGE_SHIFT) {
> >>@@ -1074,6 +1073,9 @@ i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
> >>  		goto out;
> >>  	}
> >>
> >>+	/* Object backing store will be out of date hereafter */
> >>+	obj->dirty = 1;
> >
> >Possibly. I'd rather just have shmem_pwrite be consistent and use
> >set_page_dirty. It is baked into the code that it doesn't access every
> >page.
> 
> It wasn't the shmem path that was the problem; this line was previously
> inside i915_gem_shmem_pwrite() above. But i915_gem_phys_pwrite() was missing
> the corresponding line, so it was simpler to move marking the object dirty
> up to the top level of the ioctl for now, especially as
> i915_gem_gtt_pwrite_fast() might or might not have marked the object in the
> case where it returns early.
> 
> We could at some time in the future devolve object marking to a
> class-specific vfunc, at which point this line would disappear again; but
> we'd have to implement it in each class, or at least the ones that users can
> call pwrite on (shmem, phys, and eventually stolen?). Of those, shmem can do
> per-page dirtying, but phys can stolen can't (stolen doesn't even have
> "struct page" entries available).
> 
> Which is why it's simpler to just mark the whole object here and let
> put_pages() deal with it later (if ever -- if the object is never actually
> swapped out then marking the object incurs LESS overhead than marking all
> the pages).
> 
> >>  	trace_i915_gem_object_pwrite(obj, args->offset, args->size);
> >>
> >>  	ret = -EFAULT;
> >>diff --git a/drivers/gpu/drm/i915/i915_gem_dmabuf.c b/drivers/gpu/drm/i915/i915_gem_dmabuf.c
> >>index e9c2bfd..49a74c6 100644
> >>--- a/drivers/gpu/drm/i915/i915_gem_dmabuf.c
> >>+++ b/drivers/gpu/drm/i915/i915_gem_dmabuf.c
> >>@@ -208,6 +208,8 @@ static int i915_gem_begin_cpu_access(struct dma_buf *dma_buf, size_t start, size
> >>  		return ret;
> >>
> >>  	ret = i915_gem_object_set_to_cpu_domain(obj, write);
> >>+	if (write)
> >>+		obj->dirty = 1;
> >
> >No. The accessor here should already be using set_page_dirty.
> >-Chris
> 
> What function would that be? I can't find any calls to set_page_dirty() in
> this source file. OTOH, does a dmabuf object have shmfs backing store
> anyway?

I think there's indeed a bug here, and setting obj->dirty is the right
defensive solution. For mmap access from userspace (once that happens)
we'd be covered by the set_page_dirty shmem would do.

But for kernel-internal users (dma-buf vmap, used e.g. by udl) there
indeed seems to be no one setting obj->dirty. And that code definitely
needs to be somewhere in i915.

I guess we could make a case whether we should set obj->dirty in vmap
instead - since we don't support the dma-buf kmap stuff there's no problem
there. But indeed this should be set somewhere.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch