Re: Major 2.6.38 / 2.6.39 / 3.0 regression ignored?

[Kirill Smelkov <kirr@mns.spb.ru>]

On Tue, Jul 26, 2011 at 05:48:27PM +0400, Kirill Smelkov wrote:
> On Sat, Jul 23, 2011 at 12:23:36AM +0400, Kirill Smelkov wrote:
> > Keith,
> > 
> > first of all thanks for your prompt reply. Then...
> > 
> > On Fri, Jul 22, 2011 at 11:00:41AM -0700, Keith Packard wrote:
> > > On Fri, 22 Jul 2011 15:08:06 +0400, Kirill Smelkov <kirr@mns.spb.ru> wrote:
> > > 
> > > > And now after v3.0 is out, I've tested it again, and yes, like it was
> > > > broken on v3.0-rc5, it is (now even more) broken on v3.0 -- after first
> > > > bad io access the system freezes completely:
> > > 
> > > I looked at this when I first saw it (a couple of weeks ago), and I
> > > couldn't see any obvious reason this patch would cause this particular
> > > problem. I didn't want to revert the patch at that point as I feared it
> > > would cause other subtle problems. Given that you've got a work-around,
> > > it seemed best to just push this off past 3.0.
> > 
> > What kind of a workaround are you talking about? Sorry, to me it all
> > looked like "UMS is being ignored forever". Anyway, let's move on to try
> > to solve the issue.
> > 
> > 
> > > Given the failing address passed to ioread32, this seems like it's
> > > probably the call to READ_BREADCRUMB -- I915_BREADCRUMB_INDEX is 0x21,
> > > which is an offset in 32-bit units within the hardware status page. If
> > > the status_page.page_addr value was zero, then the computed address
> > > would end up being 0x84.
> > > 
> > > And, it looks like status_page.page_addr *will* end up being zero as a
> > > result of the patch in question. The patch resets the entire ring
> > > structure contents back to the initial values, which includes smashing
> > > the status_page structure to zero, clearing the value of
> > > status_page.page_addr set in i915_init_phys_hws.
> > > 
> > > Here's an untested patch which moves the initialization of
> > > status_page.page_addr into intel_render_ring_init_dri. I note that
> > > intel_init_render_ring_buffer *already* has the setting of the
> > > status_page.page_addr value, and so I've removed the setting of
> > > status_page.page_addr from i915_init_phys_hws.
> > > 
> > > I suspect we could remove the memset from intel_init_render_ring_buffer;
> > > it seems entirely superfluous given the memset in i915_init_phys_hws.
> > > 
> > > From 159ba1dd207fc52590ce8a3afd83f40bd2cedf46 Mon Sep 17 00:00:00 2001
> > > From: Keith Packard <keithp@keithp.com>
> > > Date: Fri, 22 Jul 2011 10:44:39 -0700
> > > Subject: [PATCH] drm/i915: Initialize RCS ring status page address in
> > >  intel_render_ring_init_dri
> > > 
> > > Physically-addressed hardware status pages are initialized early in
> > > the driver load process by i915_init_phys_hws. For UMS environments,
> > > the ring structure is not initialized until the X server starts. At
> > > that point, the entire ring structure is re-initialized with all new
> > > values. Any values set in the ring structure (including
> > > ring->status_page.page_addr) will be lost when the ring is
> > > re-initialized.
> > > 
> > > This patch moves the initialization of the status_page.page_addr value
> > > to intel_render_ring_init_dri.
> > > 
> > > Signed-off-by: Keith Packard <keithp@keithp.com>
> > > ---
> > >  drivers/gpu/drm/i915/i915_dma.c         |    6 ++----
> > >  drivers/gpu/drm/i915/intel_ringbuffer.c |    3 +++
> > >  2 files changed, 5 insertions(+), 4 deletions(-)
> > > 
> > > diff --git a/drivers/gpu/drm/i915/i915_dma.c b/drivers/gpu/drm/i915/i915_dma.c
> > > index 1271282..8a3942c 100644
> > > --- a/drivers/gpu/drm/i915/i915_dma.c
> > > +++ b/drivers/gpu/drm/i915/i915_dma.c
> > > @@ -61,7 +61,6 @@ static void i915_write_hws_pga(struct drm_device *dev)
> > >  static int i915_init_phys_hws(struct drm_device *dev)
> > >  {
> > >  	drm_i915_private_t *dev_priv = dev->dev_private;
> > > -	struct intel_ring_buffer *ring = LP_RING(dev_priv);
> > >  
> > >  	/* Program Hardware Status Page */
> > >  	dev_priv->status_page_dmah =
> > > @@ -71,10 +70,9 @@ static int i915_init_phys_hws(struct drm_device *dev)
> > >  		DRM_ERROR("Can not allocate hardware status page\n");
> > >  		return -ENOMEM;
> > >  	}
> > > -	ring->status_page.page_addr =
> > > -		(void __force __iomem *)dev_priv->status_page_dmah->vaddr;
> > >  
> > > -	memset_io(ring->status_page.page_addr, 0, PAGE_SIZE);
> > > +	memset_io((void __force __iomem *)dev_priv->status_page_dmah->vaddr,
> > > +		  0, PAGE_SIZE);
> > >  
> > >  	i915_write_hws_pga(dev);
> > >  
> > > diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c
> > > index e961568..47b9b27 100644
> > > --- a/drivers/gpu/drm/i915/intel_ringbuffer.c
> > > +++ b/drivers/gpu/drm/i915/intel_ringbuffer.c
> > > @@ -1321,6 +1321,9 @@ int intel_render_ring_init_dri(struct drm_device *dev, u64 start, u32 size)
> > >  		ring->get_seqno = pc_render_get_seqno;
> > >  	}
> > >  
> > > +	if (!I915_NEED_GFX_HWS(dev))
> > > +		ring->status_page.page_addr = dev_priv->status_page_dmah->vaddr;
> > > +
> > >  	ring->dev = dev;
> > >  	INIT_LIST_HEAD(&ring->active_list);
> > >  	INIT_LIST_HEAD(&ring->request_list);
> > 
> > I can't tell whether this is correct, because intel gfx driver is
> > unknown to me, but from the first glance your description sounds reasonable.
> > 
> > I'm out of office till ~ next week's tuesday, and on return I'll try
> > to test it on the hardware in question.
> 
> Keith, thanks again for the patch. As promised I've tested it on the
> hardware in question and yes, bad_access is gone and X seems to work,
> so thank you, but...
> 
> 
> I see there are more such bugs in introduced-in-guilty-patch
> intel_render_ring_init_dri(). For example ring->irq_queue is
> left uninitialized and also ring->irq_lock etc...
>
>
> I'm X newbie, so if here is something stupid X-wise, please don't
> beat me too hard, but to me the gist of the problem is the original
> patch, where Chris does
> 
> ( git show e8616b6ced6137085e6657cc63bc2fe3900b8616 )
> > diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c
> > index 03e3370..51fbc5e 100644
> > --- a/drivers/gpu/drm/i915/intel_ringbuffer.c
> > +++ b/drivers/gpu/drm/i915/intel_ringbuffer.c
> > @@ -1291,6 +1291,48 @@ int intel_init_render_ring_buffer(struct drm_device *dev)
> >         return intel_init_ring_buffer(dev, ring);
> >  }
> >  
> > +int intel_render_ring_init_dri(struct drm_device *dev, u64 start, u32 size)
> > +{
> > +       drm_i915_private_t *dev_priv = dev->dev_private;
> > +       struct intel_ring_buffer *ring = &dev_priv->ring[RCS];
> > +
> > +       *ring = render_ring;
>           ^^^^^^^^^^^^^^^^^^^
>           here resets
> 
> > +       if (INTEL_INFO(dev)->gen >= 6) {
> > +               ring->add_request = gen6_add_request;
> > +               ring->irq_get = gen6_render_ring_get_irq;
> > +               ring->irq_put = gen6_render_ring_put_irq;
> > +       } else if (IS_GEN5(dev)) {
> > +               ring->add_request = pc_render_add_request;
> > +               ring->get_seqno = pc_render_get_seqno;
> > +       }
> 
> and then the rest of the `ring` is initialized seemingly copy-pasted
> from intel_init_ring_buffer():
> 
> > +       ring->dev = dev;
> > +       INIT_LIST_HEAD(&ring->active_list);
> > +       INIT_LIST_HEAD(&ring->request_list);
> > +       INIT_LIST_HEAD(&ring->gpu_write_list);
> > +
> > +       ring->size = size;
> > +       ring->effective_size = ring->size;
> > +       if (IS_I830(ring->dev))
> > +               ring->effective_size -= 128;
> > +
> > +       ring->map.offset = start;
> > +       ring->map.size = size;
> > +       ring->map.type = 0;
> > +       ring->map.flags = 0;
> > +       ring->map.mtrr = 0;
> ...
> 
> where both 3 chunks go almost exactly from intel_init_ring_buffer(), and
> ring->effective_size tweak even stripped original comment:
> 
> # original version from intel_init_ring_buffer():
>         /* Workaround an erratum on the i830 which causes a hang if
>          * the TAIL pointer points to within the last 2 cachelines
>          * of the buffer.
>          */
>         ring->effective_size = ring->size;
>         if (IS_I830(ring->dev))
>                 ring->effective_size -= 128;
> 
> ...
> 
> 
> The line marked "here resets" resets all the fields, and maybe it's not a good
> idea to re-initialize them all afterwards (missing some as this thread show),
> or at least if it is really needed, share initialization code between
> intel_render_ring_init_dri() and intel_init_ring_buffer() ?
> 
> >From the outside it looks like the offending patch was done as a quick
> fix in a hurry (lots of copy-paste), and maybe it would be better to
> re-do it properly...

Silence... ?

I read UMS is still ignored, because e.g. that uninitialized
ring->irq_lock which I've wrote about above is for sure used e.g. in
gen6_render_ring_get_irq() added to ring vtable in
intel_render_ring_init_dri().

And also is copy-pasting, instead of properly structuring things, ok?


Why not revert what caused trouble and introduced other subtle bugs, and
redo things properly in the first place?