diff mbox

[02/20] PCI: fix pci_remap_iospace() remap attribute

Message ID 20170227151436.18698-3-lorenzo.pieralisi@arm.com (mailing list archive)
State New, archived
Delegated to: Bjorn Helgaas
Headers show

Commit Message

Lorenzo Pieralisi Feb. 27, 2017, 3:14 p.m. UTC
According to the PCI local bus specifications (Revision 3.0, 3.2.5),
I/O Address space transactions are non-posted. On architectures where
I/O space is implemented through a chunk of memory mapped space mapped
to PCI address space (ie IA64/ARM/ARM64) the memory mapping for the
region backing I/O Address Space transactions determines the I/O
transactions attributes (before the transactions actually reaches the
PCI bus where it is handled according to the PCI specifications).

Current pci_remap_iospace() interface, that is used to map the PCI I/O
Address Space into virtual address space, use pgprot_device() as memory
attribute for the virtual address mapping, that in some architectures
(ie ARM64) provides non-cacheable but write bufferable mappings (ie
posted writes), which clash with the non-posted write behaviour for I/O
Address Space mandated by the PCI specifications.

Update the prot ioremap_page_range() parameter in pci_remap_iospace()
to pgprot_noncached to ensure that the virtual mapping backing
I/O Address Space guarantee non-posted write transactions issued
when addressing I/O Address Space through the MMIO mapping.

Signed-off-by: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Will Deacon <will.deacon@arm.com>
Cc: Bjorn Helgaas <bhelgaas@google.com>
Cc: Russell King <linux@armlinux.org.uk>
Cc: Catalin Marinas <catalin.marinas@arm.com>
---
 drivers/pci/pci.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Comments

Bjorn Helgaas March 16, 2017, 9:48 p.m. UTC | #1
[+cc Luis]

On Mon, Feb 27, 2017 at 03:14:13PM +0000, Lorenzo Pieralisi wrote:
> According to the PCI local bus specifications (Revision 3.0, 3.2.5),
> I/O Address space transactions are non-posted. On architectures where
> I/O space is implemented through a chunk of memory mapped space mapped
> to PCI address space (ie IA64/ARM/ARM64) the memory mapping for the
> region backing I/O Address Space transactions determines the I/O
> transactions attributes (before the transactions actually reaches the
> PCI bus where it is handled according to the PCI specifications).
> 
> Current pci_remap_iospace() interface, that is used to map the PCI I/O
> Address Space into virtual address space, use pgprot_device() as memory
> attribute for the virtual address mapping, that in some architectures
> (ie ARM64) provides non-cacheable but write bufferable mappings (ie
> posted writes), which clash with the non-posted write behaviour for I/O
> Address Space mandated by the PCI specifications.
> 
> Update the prot ioremap_page_range() parameter in pci_remap_iospace()
> to pgprot_noncached to ensure that the virtual mapping backing
> I/O Address Space guarantee non-posted write transactions issued
> when addressing I/O Address Space through the MMIO mapping.
> 
> Signed-off-by: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>
> Cc: Arnd Bergmann <arnd@arndb.de>
> Cc: Will Deacon <will.deacon@arm.com>
> Cc: Bjorn Helgaas <bhelgaas@google.com>
> Cc: Russell King <linux@armlinux.org.uk>
> Cc: Catalin Marinas <catalin.marinas@arm.com>
> ---
>  drivers/pci/pci.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
> index bd98674..bfb3c6e 100644
> --- a/drivers/pci/pci.c
> +++ b/drivers/pci/pci.c
> @@ -3375,7 +3375,7 @@ int pci_remap_iospace(const struct resource *res, phys_addr_t phys_addr)
>  		return -EINVAL;
>  
>  	return ioremap_page_range(vaddr, vaddr + resource_size(res), phys_addr,
> -				  pgprot_device(PAGE_KERNEL));
> +				  pgprot_noncached(PAGE_KERNEL));

pgprot_device() is equivalent to pgprot_noncached() on all arches
except ARM64, and I trust you're doing the right thing on ARM64, so
I'm fine with this from a PCI perspective.

I do find this puzzling because I naively expected pgprot_noncached()
to match up with ioremap_nocache(), and apparently it doesn't.

For example, ARM64 ioremap_nocache() uses PROT_DEVICE_nGnRE, which
doesn't match the MT_DEVICE_nGnRnE in pgprot_noncached().

The point of these patches is to use non-posted mappings.  Apparently
you can do that with pgprot_noncached() here, but ioremap_nocache()
isn't enough for the config space mappings?

I suppose that's a consequence of the pgprot_noncached() vs
ioremap_nocache() mismatch, but this is all extremely confusing.

>  #else
>  	/* this architecture does not have memory mapped I/O space,
>  	   so this function should never be called */
> -- 
> 2.10.0
> 
> 
> _______________________________________________
> linux-arm-kernel mailing list
> linux-arm-kernel@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Luis Chamberlain March 17, 2017, 12:33 a.m. UTC | #2
On Thu, Mar 16, 2017 at 04:48:44PM -0500, Bjorn Helgaas wrote:
> [+cc Luis]
> 
> On Mon, Feb 27, 2017 at 03:14:13PM +0000, Lorenzo Pieralisi wrote:
> > According to the PCI local bus specifications (Revision 3.0, 3.2.5),
> > I/O Address space transactions are non-posted. On architectures where
> > I/O space is implemented through a chunk of memory mapped space mapped
> > to PCI address space (ie IA64/ARM/ARM64) the memory mapping for the
> > region backing I/O Address Space transactions determines the I/O
> > transactions attributes (before the transactions actually reaches the
> > PCI bus where it is handled according to the PCI specifications).
> > 
> > Current pci_remap_iospace() interface, that is used to map the PCI I/O
> > Address Space into virtual address space, use pgprot_device() as memory
> > attribute for the virtual address mapping, that in some architectures
> > (ie ARM64) provides non-cacheable but write bufferable mappings (ie
> > posted writes), 

<sarcasm>
Gee wiz, I am glad this is so well documented.
</sarcasm>

> > which clash with the non-posted write behaviour for I/O
> > Address Space mandated by the PCI specifications.
> > 
> > Update the prot ioremap_page_range() parameter in pci_remap_iospace()
> > to pgprot_noncached to ensure that the virtual mapping backing
> > I/O Address Space guarantee non-posted write transactions issued
> > when addressing I/O Address Space through the MMIO mapping.

How did we end up with pgprot_device() then in the first place Liviu Dudau [0] ?
I ask for two reasons:

a) should we then use a Fixes tag for this patch ?
b) it does not seem clear what the semantics for pgprot_device() or even
   pgprot_noncached(). Can you add some ?

8b921acfeffdb ("PCI: Add pci_remap_iospace() to map bus I/O resources")

Also this patch claims archs can override this call alone, as its __weak.
So is the right thing to do to change pci_remap_iospace() to pgprot_noncached()
or is it for archs to add their own pci_remap_iospace()? If so why ? Without
proper semantics defined for these helpers this is all fuzzy.

> > Signed-off-by: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>
> > Cc: Arnd Bergmann <arnd@arndb.de>
> > Cc: Will Deacon <will.deacon@arm.com>
> > Cc: Bjorn Helgaas <bhelgaas@google.com>
> > Cc: Russell King <linux@armlinux.org.uk>
> > Cc: Catalin Marinas <catalin.marinas@arm.com>
> > ---
> >  drivers/pci/pci.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
> > index bd98674..bfb3c6e 100644
> > --- a/drivers/pci/pci.c
> > +++ b/drivers/pci/pci.c
> > @@ -3375,7 +3375,7 @@ int pci_remap_iospace(const struct resource *res, phys_addr_t phys_addr)
> >  		return -EINVAL;
> >  
> >  	return ioremap_page_range(vaddr, vaddr + resource_size(res), phys_addr,
> > -				  pgprot_device(PAGE_KERNEL));
> > +				  pgprot_noncached(PAGE_KERNEL));
> 
> pgprot_device() is equivalent to pgprot_noncached() on all arches
> except ARM64, and I trust you're doing the right thing on ARM64, so
> I'm fine with this from a PCI perspective.
> 
> I do find this puzzling because I naively expected pgprot_noncached()
> to match up with ioremap_nocache(), and apparently it doesn't.
> 
> For example, ARM64 ioremap_nocache() uses PROT_DEVICE_nGnRE, which
> doesn't match the MT_DEVICE_nGnRnE in pgprot_noncached().
> 
> The point of these patches is to use non-posted mappings.  Apparently
> you can do that with pgprot_noncached() here, but ioremap_nocache()
> isn't enough for the config space mappings?

This is for iospace it seems, so the other patch I think was for
config space.

  Luis

> I suppose that's a consequence of the pgprot_noncached() vs
> ioremap_nocache() mismatch, but this is all extremely confusing.
> 
> >  #else
> >  	/* this architecture does not have memory mapped I/O space,
> >  	   so this function should never be called */
> > -- 
> > 2.10.0
> > 
> > 
> > _______________________________________________
> > linux-arm-kernel mailing list
> > linux-arm-kernel@lists.infradead.org
> > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
>
Liviu Dudau March 17, 2017, 10:43 a.m. UTC | #3
On Fri, Mar 17, 2017 at 01:33:21AM +0100, Luis R. Rodriguez wrote:
> On Thu, Mar 16, 2017 at 04:48:44PM -0500, Bjorn Helgaas wrote:
> > [+cc Luis]
> > 
> > On Mon, Feb 27, 2017 at 03:14:13PM +0000, Lorenzo Pieralisi wrote:
> > > According to the PCI local bus specifications (Revision 3.0, 3.2.5),
> > > I/O Address space transactions are non-posted. On architectures where
> > > I/O space is implemented through a chunk of memory mapped space mapped
> > > to PCI address space (ie IA64/ARM/ARM64) the memory mapping for the
> > > region backing I/O Address Space transactions determines the I/O
> > > transactions attributes (before the transactions actually reaches the
> > > PCI bus where it is handled according to the PCI specifications).
> > > 
> > > Current pci_remap_iospace() interface, that is used to map the PCI I/O
> > > Address Space into virtual address space, use pgprot_device() as memory
> > > attribute for the virtual address mapping, that in some architectures
> > > (ie ARM64) provides non-cacheable but write bufferable mappings (ie
> > > posted writes), 
> 
> <sarcasm>
> Gee wiz, I am glad this is so well documented.
> </sarcasm>
> 
> > > which clash with the non-posted write behaviour for I/O
> > > Address Space mandated by the PCI specifications.
> > > 
> > > Update the prot ioremap_page_range() parameter in pci_remap_iospace()
> > > to pgprot_noncached to ensure that the virtual mapping backing
> > > I/O Address Space guarantee non-posted write transactions issued
> > > when addressing I/O Address Space through the MMIO mapping.
> 
> How did we end up with pgprot_device() then in the first place Liviu Dudau [0] ?
> I ask for two reasons:

[replying using personal email as the corporate email system is taking its sweet time
to deliver the email to my inbox]

I've asked the people with the right knowledge about the correct API to use (Hi Catalin!),
and during the review it did not throw any red flags. I guess, given Bjorn's comment,
that everyone assumed AArch64 is the same as all other architectures and pgprot_device
is synonymous to pgprot_noncached.

> 
> a) should we then use a Fixes tag for this patch ?

I'm not aware of issues being reported, but Lorenzo might have more info on this.

> b) it does not seem clear what the semantics for pgprot_device() or even
>    pgprot_noncached(). Can you add some ?
> 
> 8b921acfeffdb ("PCI: Add pci_remap_iospace() to map bus I/O resources")
> 
> Also this patch claims archs can override this call alone, as its __weak.
> So is the right thing to do to change pci_remap_iospace() to pgprot_noncached()
> or is it for archs to add their own pci_remap_iospace()? If so why ? Without
> proper semantics defined for these helpers this is all fuzzy.

That was the initial intention, to let arches / platforms overwrite the whole
pci_remap_iospace(). I guess the reality is that no one needs to overwrite it except
for the AArch64 quirk, so probably easier to remove the __weak and fix the attributes for arm64.

Best regards,
Liviu

> 
> > > Signed-off-by: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>
> > > Cc: Arnd Bergmann <arnd@arndb.de>
> > > Cc: Will Deacon <will.deacon@arm.com>
> > > Cc: Bjorn Helgaas <bhelgaas@google.com>
> > > Cc: Russell King <linux@armlinux.org.uk>
> > > Cc: Catalin Marinas <catalin.marinas@arm.com>
> > > ---
> > >  drivers/pci/pci.c | 2 +-
> > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > 
> > > diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
> > > index bd98674..bfb3c6e 100644
> > > --- a/drivers/pci/pci.c
> > > +++ b/drivers/pci/pci.c
> > > @@ -3375,7 +3375,7 @@ int pci_remap_iospace(const struct resource *res, phys_addr_t phys_addr)
> > >  		return -EINVAL;
> > >  
> > >  	return ioremap_page_range(vaddr, vaddr + resource_size(res), phys_addr,
> > > -				  pgprot_device(PAGE_KERNEL));
> > > +				  pgprot_noncached(PAGE_KERNEL));
> > 
> > pgprot_device() is equivalent to pgprot_noncached() on all arches
> > except ARM64, and I trust you're doing the right thing on ARM64, so
> > I'm fine with this from a PCI perspective.
> > 
> > I do find this puzzling because I naively expected pgprot_noncached()
> > to match up with ioremap_nocache(), and apparently it doesn't.
> > 
> > For example, ARM64 ioremap_nocache() uses PROT_DEVICE_nGnRE, which
> > doesn't match the MT_DEVICE_nGnRnE in pgprot_noncached().
> > 
> > The point of these patches is to use non-posted mappings.  Apparently
> > you can do that with pgprot_noncached() here, but ioremap_nocache()
> > isn't enough for the config space mappings?
> 
> This is for iospace it seems, so the other patch I think was for
> config space.
> 
>   Luis
> 
> > I suppose that's a consequence of the pgprot_noncached() vs
> > ioremap_nocache() mismatch, but this is all extremely confusing.
> > 
> > >  #else
> > >  	/* this architecture does not have memory mapped I/O space,
> > >  	   so this function should never be called */
> > > -- 
> > > 2.10.0
> > > 
> > > 
> > > _______________________________________________
> > > linux-arm-kernel mailing list
> > > linux-arm-kernel@lists.infradead.org
> > > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
> > 
> 
> -- 
> Luis Rodriguez, SUSE LINUX GmbH
> Maxfeldstrasse 5; D-90409 Nuernberg
Luis Chamberlain March 17, 2017, 4:26 p.m. UTC | #4
On Fri, Mar 17, 2017 at 10:43:39AM +0000, Liviu Dudau wrote:
> On Fri, Mar 17, 2017 at 01:33:21AM +0100, Luis R. Rodriguez wrote:
> > a) should we then use a Fixes tag for this patch ?
> 
> I'm not aware of issues being reported, but Lorenzo might have more info on this.

Lorenzo ? If not what exactly made you discover this ? If it is a fix, and only 
ARM64 is implicated, seems like a worthy change to consider for stable for the
sake of stable ARM64 kernels. But, that would leave the PCI config space without
a simple 1 liner fix too -- so maybe its not worth it. Distributions wanting
to support ARM64 however would like to carry these changes, so some annotations
such as Fixes should help.

> > b) it does not seem clear what the semantics for pgprot_device() or even
> >    pgprot_noncached(). Can you add some ?
> > 
> > 8b921acfeffdb ("PCI: Add pci_remap_iospace() to map bus I/O resources")
> > 
> > Also this patch claims archs can override this call alone, as its __weak.
> > So is the right thing to do to change pci_remap_iospace() to pgprot_noncached()
> > or is it for archs to add their own pci_remap_iospace()? If so why ? Without
> > proper semantics defined for these helpers this is all fuzzy.
> 
> That was the initial intention, to let arches / platforms overwrite the whole
> pci_remap_iospace(). I guess the reality is that no one needs to overwrite it except
> for the AArch64 quirk, so probably easier to remove the __weak and fix the attributes for arm64.

Sounds much more reasonable to me.

  Luis
Bjorn Helgaas March 20, 2017, 4:06 p.m. UTC | #5
On Fri, Mar 17, 2017 at 01:33:21AM +0100, Luis R. Rodriguez wrote:
> On Thu, Mar 16, 2017 at 04:48:44PM -0500, Bjorn Helgaas wrote:
> > [+cc Luis]
> > 
> > On Mon, Feb 27, 2017 at 03:14:13PM +0000, Lorenzo Pieralisi wrote:
> > > According to the PCI local bus specifications (Revision 3.0, 3.2.5),
> > > I/O Address space transactions are non-posted. On architectures where
> > > I/O space is implemented through a chunk of memory mapped space mapped
> > > to PCI address space (ie IA64/ARM/ARM64) the memory mapping for the
> > > region backing I/O Address Space transactions determines the I/O
> > > transactions attributes (before the transactions actually reaches the
> > > PCI bus where it is handled according to the PCI specifications).
> > > 
> > > Current pci_remap_iospace() interface, that is used to map the PCI I/O
> > > Address Space into virtual address space, use pgprot_device() as memory
> > > attribute for the virtual address mapping, that in some architectures
> > > (ie ARM64) provides non-cacheable but write bufferable mappings (ie
> > > posted writes), 
> 
> <sarcasm>
> Gee wiz, I am glad this is so well documented.
> </sarcasm>

I'm not sure this is actionable feedback.  The two paragraphs above
seem clear and useful to me.  How would you like to see them improved?

> > > ...
> > > @@ -3375,7 +3375,7 @@ int pci_remap_iospace(const struct resource *res, phys_addr_t phys_addr)
> > >  		return -EINVAL;
> > >  
> > >  	return ioremap_page_range(vaddr, vaddr + resource_size(res), phys_addr,
> > > -				  pgprot_device(PAGE_KERNEL));
> > > +				  pgprot_noncached(PAGE_KERNEL));
> > 
> > ...
> > I do find this puzzling because I naively expected pgprot_noncached()
> > to match up with ioremap_nocache(), and apparently it doesn't.
> > 
> > For example, ARM64 ioremap_nocache() uses PROT_DEVICE_nGnRE, which
> > doesn't match the MT_DEVICE_nGnRnE in pgprot_noncached().
> > 
> > The point of these patches is to use non-posted mappings.  Apparently
> > you can do that with pgprot_noncached() here, but ioremap_nocache()
> > isn't enough for the config space mappings?
> 
> This is for iospace it seems, so the other patch I think was for
> config space.

I understand that 02/20 is for PCI I/O port space and 03/20 is for PCI
config space.  I'm confused because I thought we wanted the same
non-posted mapping for both, but looks like they're different.

Patch 02/20 uses ioremap_page_range(..., pgprot_noncached(PAGE_KERNEL))
to map PCI I/O port space:

  #define pgprot_noncached(prot) \
          __pgprot_modify(prot, PTE_ATTRINDX_MASK, PTE_ATTRINDX(MT_DEVICE_nGnRnE) | PTE_PXN | PTE_UXN)

Patch 03/20 uses ioremap_nocache() to map PCI config space:

  #define ioremap_nocache(addr, size)     __ioremap((addr), (size), __pgprot(PROT_DEVICE_nGnRE))

So the I/O port mapping uses MT_DEVICE_nGnRnE, while the config space
mapping uses PROT_DEVICE_nGnRE, which looks different.

Bjorn
Lorenzo Pieralisi March 20, 2017, 4:19 p.m. UTC | #6
On Fri, Mar 17, 2017 at 05:26:18PM +0100, Luis R. Rodriguez wrote:
> On Fri, Mar 17, 2017 at 10:43:39AM +0000, Liviu Dudau wrote:
> > On Fri, Mar 17, 2017 at 01:33:21AM +0100, Luis R. Rodriguez wrote:
> > > a) should we then use a Fixes tag for this patch ?
> > 
> > I'm not aware of issues being reported, but Lorenzo might have more info on this.
> 
> Lorenzo ? If not what exactly made you discover this ? If it is a fix, and only 
> ARM64 is implicated, seems like a worthy change to consider for stable for the
> sake of stable ARM64 kernels. But, that would leave the PCI config space without
> a simple 1 liner fix too -- so maybe its not worth it. Distributions wanting
> to support ARM64 however would like to carry these changes, so some annotations
> such as Fixes should help.

It started with this thread:

http://lists.infradead.org/pipermail/linux-arm-kernel/2017-January/477353.html

this series is not fixing any current issue I am aware of (but I am not
keen on leaving code as-is either) hence adding a Fixes: tag is problematic.

I would leave stable kernels alone for the time being.

Lorenzo

> > > b) it does not seem clear what the semantics for pgprot_device() or even
> > >    pgprot_noncached(). Can you add some ?
> > > 
> > > 8b921acfeffdb ("PCI: Add pci_remap_iospace() to map bus I/O resources")
> > > 
> > > Also this patch claims archs can override this call alone, as its __weak.
> > > So is the right thing to do to change pci_remap_iospace() to pgprot_noncached()
> > > or is it for archs to add their own pci_remap_iospace()? If so why ? Without
> > > proper semantics defined for these helpers this is all fuzzy.
> > 
> > That was the initial intention, to let arches / platforms overwrite the whole
> > pci_remap_iospace(). I guess the reality is that no one needs to overwrite it except
> > for the AArch64 quirk, so probably easier to remove the __weak and fix the attributes for arm64.
> 
> Sounds much more reasonable to me.
> 
>   Luis
Lorenzo Pieralisi March 20, 2017, 4:26 p.m. UTC | #7
On Mon, Mar 20, 2017 at 11:06:36AM -0500, Bjorn Helgaas wrote:

[...]

> > > > @@ -3375,7 +3375,7 @@ int pci_remap_iospace(const struct resource *res, phys_addr_t phys_addr)
> > > >  		return -EINVAL;
> > > >  
> > > >  	return ioremap_page_range(vaddr, vaddr + resource_size(res), phys_addr,
> > > > -				  pgprot_device(PAGE_KERNEL));
> > > > +				  pgprot_noncached(PAGE_KERNEL));
> > > 
> > > ...
> > > I do find this puzzling because I naively expected pgprot_noncached()
> > > to match up with ioremap_nocache(), and apparently it doesn't.
> > > 
> > > For example, ARM64 ioremap_nocache() uses PROT_DEVICE_nGnRE, which
> > > doesn't match the MT_DEVICE_nGnRnE in pgprot_noncached().
> > > 
> > > The point of these patches is to use non-posted mappings.  Apparently
> > > you can do that with pgprot_noncached() here, but ioremap_nocache()
> > > isn't enough for the config space mappings?
> > 
> > This is for iospace it seems, so the other patch I think was for
> > config space.
> 
> I understand that 02/20 is for PCI I/O port space and 03/20 is for PCI
> config space.  I'm confused because I thought we wanted the same
> non-posted mapping for both, but looks like they're different.
> 
> Patch 02/20 uses ioremap_page_range(..., pgprot_noncached(PAGE_KERNEL))
> to map PCI I/O port space:
> 
>   #define pgprot_noncached(prot) \
>           __pgprot_modify(prot, PTE_ATTRINDX_MASK, PTE_ATTRINDX(MT_DEVICE_nGnRnE) | PTE_PXN | PTE_UXN)
> 
> Patch 03/20 uses ioremap_nocache() to map PCI config space:
> 
>   #define ioremap_nocache(addr, size)     __ioremap((addr), (size), __pgprot(PROT_DEVICE_nGnRE))
> 
> So the I/O port mapping uses MT_DEVICE_nGnRnE, while the config space
> mapping uses PROT_DEVICE_nGnRE, which looks different.

On ARM64 (PATCH 4) and ARM (PATCH 5) we override pci_remap_cfgspace()
with implementations that provide non-posted writes bus attributes,
PATCH 3 is just there to provide a "safe" (well, I need input on that)
fall-back.

Thanks,
Lorenzo
Bjorn Helgaas March 20, 2017, 4:38 p.m. UTC | #8
On Mon, Mar 20, 2017 at 04:26:27PM +0000, Lorenzo Pieralisi wrote:
> On Mon, Mar 20, 2017 at 11:06:36AM -0500, Bjorn Helgaas wrote:
> 
> [...]
> 
> > > > > @@ -3375,7 +3375,7 @@ int pci_remap_iospace(const struct resource *res, phys_addr_t phys_addr)
> > > > >  		return -EINVAL;
> > > > >  
> > > > >  	return ioremap_page_range(vaddr, vaddr + resource_size(res), phys_addr,
> > > > > -				  pgprot_device(PAGE_KERNEL));
> > > > > +				  pgprot_noncached(PAGE_KERNEL));
> > > > 
> > > > ...
> > > > I do find this puzzling because I naively expected pgprot_noncached()
> > > > to match up with ioremap_nocache(), and apparently it doesn't.
> > > > 
> > > > For example, ARM64 ioremap_nocache() uses PROT_DEVICE_nGnRE, which
> > > > doesn't match the MT_DEVICE_nGnRnE in pgprot_noncached().
> > > > 
> > > > The point of these patches is to use non-posted mappings.  Apparently
> > > > you can do that with pgprot_noncached() here, but ioremap_nocache()
> > > > isn't enough for the config space mappings?
> > > 
> > > This is for iospace it seems, so the other patch I think was for
> > > config space.
> > 
> > I understand that 02/20 is for PCI I/O port space and 03/20 is for PCI
> > config space.  I'm confused because I thought we wanted the same
> > non-posted mapping for both, but looks like they're different.
> > 
> > Patch 02/20 uses ioremap_page_range(..., pgprot_noncached(PAGE_KERNEL))
> > to map PCI I/O port space:
> > 
> >   #define pgprot_noncached(prot) \
> >           __pgprot_modify(prot, PTE_ATTRINDX_MASK, PTE_ATTRINDX(MT_DEVICE_nGnRnE) | PTE_PXN | PTE_UXN)
> > 
> > Patch 03/20 uses ioremap_nocache() to map PCI config space:
> > 
> >   #define ioremap_nocache(addr, size)     __ioremap((addr), (size), __pgprot(PROT_DEVICE_nGnRE))
> > 
> > So the I/O port mapping uses MT_DEVICE_nGnRnE, while the config space
> > mapping uses PROT_DEVICE_nGnRE, which looks different.
> 
> On ARM64 (PATCH 4) and ARM (PATCH 5) we override pci_remap_cfgspace()
> with implementations that provide non-posted writes bus attributes,
> PATCH 3 is just there to provide a "safe" (well, I need input on that)
> fall-back.

Ah, thanks, that makes sense to me finally!  After patch 4,
pci_remap_cfgspace() uses PROT_DEVICE_nGnRnE, which does look like the
MT_DEVICE_nGnRnE used in the ioremap_page_range() path.

I wonder if we can get rid of pgprot_device() altogether?  Patch 2
removed the use in pci_remap_iospace(), and patch 18 removes the use
in tegra_pcie_bus_alloc().

Bjorn
diff mbox

Patch

diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
index bd98674..bfb3c6e 100644
--- a/drivers/pci/pci.c
+++ b/drivers/pci/pci.c
@@ -3375,7 +3375,7 @@  int pci_remap_iospace(const struct resource *res, phys_addr_t phys_addr)
 		return -EINVAL;
 
 	return ioremap_page_range(vaddr, vaddr + resource_size(res), phys_addr,
-				  pgprot_device(PAGE_KERNEL));
+				  pgprot_noncached(PAGE_KERNEL));
 #else
 	/* this architecture does not have memory mapped I/O space,
 	   so this function should never be called */