Message-ID: <20150401235211.GV5622@wotan.suse.de>
Date:	Thu, 2 Apr 2015 01:52:11 +0200
From:	"Luis R. Rodriguez" <mcgrof@...e.com>
To:	Ville Syrjälä <syrjala@....fi>,
	Andy Lutomirski <luto@...capital.net>,
	Bjorn Helgaas <bhelgaas@...gle.com>,
	"Luis R. Rodriguez" <mcgrof@...not-panic.com>,
	Ingo Molnar <mingo@...hat.com>,
	Thomas Gleixner <tglx@...utronix.de>,
	"H. Peter Anvin" <hpa@...or.com>, Juergen Gross <jgross@...e.com>,
	Jan Beulich <JBeulich@...e.com>, Borislav Petkov <bp@...e.de>,
	Suresh Siddha <suresh.b.siddha@...el.com>,
	venkatesh.pallipadi@...el.com, Dave Airlie <airlied@...hat.com>,
	"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
	Linux Fbdev development list <linux-fbdev@...r.kernel.org>,
	X86 ML <x86@...nel.org>,
	"xen-devel@...ts.xenproject.org" <xen-devel@...ts.xenproject.org>,
	Ingo Molnar <mingo@...e.hu>,
	Linus Torvalds <torvalds@...ux-foundation.org>,
	Daniel Vetter <daniel.vetter@...ll.ch>,
	Antonino Daplas <adaplas@...il.com>,
	Jean-Christophe Plagniol-Villard <plagnioj@...osoft.com>,
	Tomi Valkeinen <tomi.valkeinen@...com>
Subject: Re: [PATCH v1 09/47] vidoe: fbdev: atyfb: remove and fix MTRR MMIO
 "hole" work around

On Sat, Mar 28, 2015 at 02:23:34PM +0200, Ville Syrjälä wrote:
> On Sat, Mar 28, 2015 at 01:28:18AM +0100, Luis R. Rodriguez wrote:
> > On Fri, Mar 27, 2015 at 03:02:10PM -0700, Andy Lutomirski wrote:
> > > On Fri, Mar 27, 2015 at 2:56 PM, Ville Syrjälä <syrjala@....fi> wrote:
> > > > On Fri, Mar 27, 2015 at 08:57:59PM +0100, Luis R. Rodriguez wrote:
> > > >> On Fri, Mar 27, 2015 at 12:43:55PM -0700, Andy Lutomirski wrote:
> > > >> > On Fri, Mar 27, 2015 at 12:38 PM, Luis R. Rodriguez <mcgrof@...e.com> wrote:
> > > >> > > On Sat, Mar 21, 2015 at 11:15:14AM +0200, Ville Syrjälä wrote:
> > > >> > >> On Fri, Mar 20, 2015 at 04:17:59PM -0700, Luis R. Rodriguez wrote:
> > > >> > >> > diff --git a/drivers/video/fbdev/aty/atyfb_base.c b/drivers/video/fbdev/aty/atyfb_base.c
> > > >> > >> > index 8025624..8875e56 100644
> > > >> > >> > --- a/drivers/video/fbdev/aty/atyfb_base.c
> > > >> > >> > +++ b/drivers/video/fbdev/aty/atyfb_base.c
> > > >> > >> > @@ -2630,21 +2630,10 @@ static int aty_init(struct fb_info *info)
> > > >> > >> >
> > > >> > >> >  #ifdef CONFIG_MTRR
> > > >> > >> >     par->mtrr_aper = -1;
> > > >> > >> > -   par->mtrr_reg = -1;
> > > >> > >> >     if (!nomtrr) {
> > > >> > >> > -           /* Cover the whole resource. */
> > > >> > >> > -           par->mtrr_aper = mtrr_add(par->res_start, par->res_size,
> > > >> > >> > +           par->mtrr_aper = mtrr_add(info->fix.smem_start,
> > > >> > >> > +                                     info->fix.smem_len,
> > > >> > >> >                                       MTRR_TYPE_WRCOMB, 1);
> > > >> > >>
> > > >> > >> MTRRs need power of two size, so how is this supposed to work?
> > > >> > >
> > > >> > > As per mtrr_add_page() [0] the base and size are just supposed to be in units
> > > >> > > of 4 KiB, although the practice is to use powers of 2 in *some* drivers this
> > > >> > > is not standardized and by no means recorded as a requirement. Obviously
> > > >> > > powers of 2 will work too and you'd end up neatly aligned as well. mtrr_add()
> > > >> > > will use mtrr_check() to verify the the same requirement. Furthermore,
> > > >> > > as per my commit log message:
> > > >> >
> > > >> > Whatever the code may or may not do, the x86 architecture uses
> > > >> > power-of-two MTRR sizes.  So I'm confused.
> > > >>
> > > >> There should be no confusion, I simply did not know that *was* the
> > > >> requirement for x86, if that is the case we should add a check for that
> > > >> and perhaps generalize a helper that does the power of two helper changes,
> > > >> the cleanest I found was the vesafb driver solution.
> > > >>
> > > >> Thoughts?
> > > >
> > > > The vesafb solution is bad since you'll only end up covering only
> > > > the first 4MB of the framebuffer instead of the almost 8MB you want.
> > > > Which in practice will mean throwing away half the VRAM since you really
> > > > don't want the massive performance hit from accessing it as UC. And that
> > > > would mean giving up decent display resolutions as well :(
> > > >
> > > > And the other option of trying to cover the remainder with multiple ever
> > > > smaller MTRRs doesn't work either since you'll run out of MTRRs very
> > > > quickly.
> > > >
> > > > This is precisely why I used the hole method in atyfb in the first
> > > > place.
> > > >
> > > > I don't really like the idea of any new mtrr code not supporting that
> > > > use case, especially as these things tend to be present in older machines
> > > > where PAT isn't an option.
> > > 
> > > According to the Intel SDM, volume 3, section 11.5.2.1, table 11-6,
> > > non-PAT CPUs that have a WC MTRR, PCD = 1, and PWT = 1 (aka UC) have
> > > an effective memory type of UC.

This is true but non-PAT systems that use just ioremap() will default to
_PAGE_CACHE_MODE_UC_MINUS, not _PAGE_CACHE_MODE_UC, and _PAGE_CACHE_MODE_UC_MINUS
on Linux has PCD = 1, PWT = 0. The list comes from:

uint16_t __cachemode2pte_tbl[_PAGE_CACHE_MODE_NUM] = {                          
        [_PAGE_CACHE_MODE_WB      ]     = 0         | 0        ,                
        [_PAGE_CACHE_MODE_WC      ]     = _PAGE_PWT | 0        ,                
        [_PAGE_CACHE_MODE_UC_MINUS]     = 0         | _PAGE_PCD,                
        [_PAGE_CACHE_MODE_UC      ]     = _PAGE_PWT | _PAGE_PCD,                
        [_PAGE_CACHE_MODE_WT      ]     = 0         | _PAGE_PCD,                
        [_PAGE_CACHE_MODE_WP      ]     = 0         | _PAGE_PCD,                
};                                                                              

This can better be read here:

 PAT
 |PCD
 ||PWT
 |||
 000 WB          _PAGE_CACHE_MODE_WB
 001 WC          _PAGE_CACHE_MODE_WC
 010 UC-         _PAGE_CACHE_MODE_UC_MINUS
 011 UC          _PAGE_CACHE_MODE_UC

On x86 ioremap() defaults to ioremap_nocache() and right now that uses
_PAGE_CACHE_MODE_UC_MINUS not _PAGE_CACHE_MODE_UC. We have two cases
to consider for non-PAT systems then:

a) Right now as ioremap() and ioremap_nocache() default to _PAGE_CACHE_MODE_UC_MINUS
   on x86. In this case using a WC MTRR seems to use PWT=0, PCD=1, and
   table table 11-6 on non-PAT systems seems to place this situation as
   "implementation defined" and not encouraged.

a) when commit de33c442e "x86 PAT: fix performance drop for glx, use
   UC minus for ioremap(), ioremap_nocache() and pci_mmap_page_range()"
   gets reverted and we use _PAGE_CACHE_MODE_UC by default. In this
   case on x86 for both ioremap() and ioremap_nocache() as they will
   both default to _PAGE_CACHE_MODE_UC we'll end up as you note with
   an effective memory type of UC.

If I've understood this correctly then neither of these situations are good and
its just by chance that on some systems situation a) has lead to proper WC.

On a PAT system we have a bit different combinatorial results (based on Table
11-7):

a) Right now ioremap() and ioremap_nocache() defaulting to
    _PAGE_CACHE_MODE_UC_MINUS yields + MTRR WC = WC

b) When commit de33c442e gets reverted _PAGE_CACHE_MODE_UC + MTRR WC = UC

So to be clear right now atyfb should work fine on PAT systems
with de33c442e in place, once reverted as-is right now we'd end
up with UC effective memory type.

For both PAT and non-PAT systems when commit de33c442e gets reverted
we'd end up with UC as the effective memory type for atyfb. Right
now it shoud work on PAT systems and by chance its suspected to work
on non-PAT systems. We want to phase MTRR though, specially to avoid
all this insane combinatorial nightmware.

> > > Hence my suggestion to add
> > > ioremap_x86_uc and/or set_memory_x86_uc to punch a UC hole in an
> > > otherwise WC MTRR-covered region.

To be clear I think you mean then that ioremap_x86_uc() would help us avoid the
jumps between combinatorial issues with MTRR on PAT / non-PAT systems before
and after commit de33c442e gets reverted. So for instance if we had on the
atyfb driver:

ioremap_x86_uc(PCI BAR)
ioremap_wc(framebuffer)
arch_phys_add_wc(PCI BAR)

On non-PAT systems on the MMIO region with PWT=1, PCD=1 we'd end up with UC.
Sadly though since _PAGE_CACHE_WC on non-PAT has PWT=1, PCD=0, the WC
MTRR that follows would mean we'd end up with another grey area (but
similar to before as technically an effectivethe memory type of WC).

On PAT systems the above would not use MTRRs but we'd be counting on
overlapping memory types -- its not clear if aliasing here is a problem.

Also Intel SDM, volume 3, section "11.11.4 Range Size and Alignment Requirement"
describes that: "the minimum range size is 4 KiB, the base address must be on
a 4 KiB boundary. For ranges greater than 4 KiB each range must be of length
2^n and its base address must be alinged on a 2^n boundary where n is a value
equal or greatar then 12. The base-address alignment value cannot be less
than its length. For example, an 8-KiB range cannot be aligned on a
4-KiB boundary. It must be aligned on at least an 8-KiB boundary"

So to answer my own question: indeed, our framebuffer base address must be
aligned on a 2^n boundary, the size also has to be a power of 2. MTRR supports
fixed range sizes and variable range sizes, in case of the MMIO that does
not need to abide by the power of 2 rule as a fixed range size of 4 KiB
could be used although upon review ouf our own implemetnation its unclear if
that is what is used for 4 KiB sized MTRRs.

Hence my arch_phys_add_wc(PCI BAR) as above.

> > OK I think I get it now.
> > 
> > And I take it this would hopefully only be used for non-PAT systems?

Since we likely could care to use ioremap_x86_uc() on PAT systems as well we
could make the effective for both PAT and non-PAT obviously then.  Later when
we get ioremap() to default to strong UC we could drop ioremap_x86_uc() as we'd
only need it as transitory until then -- that is unless we want perhaps a strong
UC ioremap primitive which is always following strong UC when available regardless
of these default transitions.

The big issue I see here is simply the combinatorial issues, so I do think
its best to annotate these corner cases well and avoid them.

> > Would there be a use case for PAT systems? I wonder if we can wrap
> > this under some APIs to make it clean and hide this dirty thing
> > behind the scenes, it seems a fragile and error prone and my hope
> > would be that we won't need more specialization in this area for
> > PAT systems.
> 
> One potential complication is kernel vs. userspace mmap. MTRR applies to
> the physical address, but PAT applies to the virtual address, so with
> the WC MTRR you get WC for userspace "for free" as well.

What is the performance impact of having the conversion being done by the
kernel? Has anyone done measurements? If significant can't the subsystem mmap()
cache the phys address for PAT? Shouldn't the TLB take care of those considerations
for us? If this is generally desirable shouldn't we just generalize the cache
for devices for O(1) access through a generic API?

Can the difference, other than a possible performance hit, implicate a userspace
visible change?

If the performance / userspace effect is neglibable then I'd expect the gains
from cleaner code / APIs to outweight the cons. After all the goal is to
streamline PAT when possible here.

> Also the
> userspace mmaps request will have the length of smem_len (at most), so
> it won't be the nice power of two in that case.

Does that length change implicate a userspace visible change?

> Also on PAT systems w/o a BIOS provided WC MTRR, the fbdev mmap seems
> to be total crap at the moment. IIRC I have a patch to fix things a bit...

Isn't that becuase of the lack of the ioremap_wc()'s? You seem to be
alternatively doing this with pgprot_writecombine(), more on this strategy
below though.

> From 4e6d70d223f35953c8a11a58cf3376a8a001fa4f Mon Sep 17 00:00:00 2001
> From: Ville Syrjala <syrjala@....fi>
> Date: Fri, 15 Apr 2011 04:02:43 +0300
> Subject: [PATCH] fb: writecombine fb
> 
> ---
>  drivers/video/fbdev/core/fbmem.c | 12 +++++++++++-
>  1 file changed, 11 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/video/fbdev/core/fbmem.c b/drivers/video/fbdev/core/fbmem.c
> index 0705d88..ecbde0e 100644
> --- a/drivers/video/fbdev/core/fbmem.c
> +++ b/drivers/video/fbdev/core/fbmem.c
> @@ -1396,6 +1396,7 @@ fb_mmap(struct file *file, struct vm_area_struct * vma)
>  	unsigned long mmio_pgoff;
>  	unsigned long start;
>  	u32 len;
> +	bool mmio = false;
>  
>  	if (!info)
>  		return -ENODEV;
> @@ -1426,11 +1427,20 @@ fb_mmap(struct file *file, struct vm_area_struct * vma)
>  		vma->vm_pgoff -= mmio_pgoff;
>  		start = info->fix.mmio_start;
>  		len = info->fix.mmio_len;
> +		mmio = true;
>  	}
>  	mutex_unlock(&info->mm_lock);
>  
> +	if (!mmio) {
> +		vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
> +		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
> +
> +		if (!vm_iomap_memory(vma, start, len))
> +			return 0;
> +	}
> +
>  	vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
> -	fb_pgprotect(file, vma, start);
> +	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
>  
>  	return vm_iomap_memory(vma, start, len);
>  }
> 
> Perhaps it's time I tried to send that upstream properly :P

Lets assume drivers all have ioremap_wc() on the framebuffer, would the
the above not be needed then (disregarding the corner cases such as atyfb)?

If your goal is to generalize a place to make framebuffer WC instead of doing
that at mmap() why not do it at register_framebuffer() time and do it
only once? I suspect all this might also be easier to do and generalize
after this series.
 
So as we can see from this series there are tons of drivers that can safely
be moved to use ioremap_wc() already, provided there are no regressions with
the simple ioremap_wc() / arch_phys_wc_add() switch. There are only a few corner
cases to address after that. Addressing both of these *first* would simplify
the code and gramatically make it a bit more consistent while trying to avoid a
generalized regression. I believe a generalized solution is definitely in order
but we also should first address the corner cases.

So how about we:

a) convert all drivers over that are safe to convert to ioremap_wc() /
   arch_phys_add_wc()
b) address all corner cases and try to avoid further combinatorial
   issues
c) after a while push for reverting de33c442e
d) generalize a solution / for framebuffer

Ideally as I mentioned in the other thread with Bjorn we could even
have the WC be done further below for us but it was very unclear
if we could accomplish this due the definition of the PCI flags,
the way we'd use it and the way they could be integrated on hardware
by manufacturers. I think generalizing things under the frambuffer
code would be good intermediate step but I think we need to phase
this in in light of the corner cases, combinatorial issues with
PAT / non-PAT and eventual reverting goals of commit de33c442e
in order to generalize strong UC.

 Luis
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists