[PATCH] arch : arm : add a criteria for pfn_valid

[PATCH] arch : arm : add a criteria for pfn_valid

[Zhaoyang Huang]

From: Zhaoyang Huang <zhaoyang.huang@unisoc.com>

pfn_valid can be wrong while the MSB of physical address be trimed as pfn
larger than the max_pfn.

Signed-off-by: Zhaoyang Huang <huangzhaoyang@gmail.com>
---
 arch/arm/mm/init.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/arch/arm/mm/init.c b/arch/arm/mm/init.c
index c2daabb..9c4d938 100644
--- a/arch/arm/mm/init.c
+++ b/arch/arm/mm/init.c
@@ -177,7 +177,8 @@ static void __init zone_sizes_init(unsigned long min, unsigned long max_low,
 #ifdef CONFIG_HAVE_ARCH_PFN_VALID
 int pfn_valid(unsigned long pfn)
 {
-	return memblock_is_map_memory(__pfn_to_phys(pfn));
+	return (pfn > max_pfn) ?
+		false : memblock_is_map_memory(__pfn_to_phys(pfn));
 }
 EXPORT_SYMBOL(pfn_valid);
 #endif
-- 
1.9.1

Re: [PATCH] arch : arm : add a criteria for pfn_valid

[Matthew Wilcox]

On Sat, Aug 17, 2019 at 11:00:13AM +0800, Zhaoyang Huang wrote:
>  #ifdef CONFIG_HAVE_ARCH_PFN_VALID
>  int pfn_valid(unsigned long pfn)
>  {
> -	return memblock_is_map_memory(__pfn_to_phys(pfn));
> +	return (pfn > max_pfn) ?
> +		false : memblock_is_map_memory(__pfn_to_phys(pfn));
>  }

This is a really awkward way to use the ternary operator.  It's easier to
read if you just:

+	if (pfn > max_pfn)
+		return 0;
 	return memblock_is_map_memory(__pfn_to_phys(pfn));

(if you really wanted to be clever ... er, obscure, you'd've written:

	return (pfn <= max_pfn) && memblock_is_map_memory(__pfn_to_phys(pfn));

... but don't do that)

Also, why is this diverged between arm and arm64?

Re: [PATCH] arch : arm : add a criteria for pfn_valid

[Mike Rapoport]

On Sat, Aug 17, 2019 at 11:00:13AM +0800, Zhaoyang Huang wrote:
> From: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
> 
> pfn_valid can be wrong while the MSB of physical address be trimed as pfn
> larger than the max_pfn.

How the overflow of __pfn_to_phys() is related to max_pfn?
Where is the guarantee that __pfn_to_phys(max_pfn) won't overflow?
 
> Signed-off-by: Zhaoyang Huang <huangzhaoyang@gmail.com>
> ---
>  arch/arm/mm/init.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/arm/mm/init.c b/arch/arm/mm/init.c
> index c2daabb..9c4d938 100644
> --- a/arch/arm/mm/init.c
> +++ b/arch/arm/mm/init.c
> @@ -177,7 +177,8 @@ static void __init zone_sizes_init(unsigned long min, unsigned long max_low,
>  #ifdef CONFIG_HAVE_ARCH_PFN_VALID
>  int pfn_valid(unsigned long pfn)
>  {
> -	return memblock_is_map_memory(__pfn_to_phys(pfn));
> +	return (pfn > max_pfn) ?
> +		false : memblock_is_map_memory(__pfn_to_phys(pfn));
>  }
>  EXPORT_SYMBOL(pfn_valid);
>  #endif
> -- 
> 1.9.1
> 

-- 
Sincerely yours,
Mike.

Re: [PATCH] arch : arm : add a criteria for pfn_valid

[Zhaoyang Huang]

On Sat, Aug 17, 2019 at 5:00 PM Mike Rapoport <rppt@linux.ibm.com> wrote:
>
> On Sat, Aug 17, 2019 at 11:00:13AM +0800, Zhaoyang Huang wrote:
> > From: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
> >
> > pfn_valid can be wrong while the MSB of physical address be trimed as pfn
> > larger than the max_pfn.
>
> How the overflow of __pfn_to_phys() is related to max_pfn?
> Where is the guarantee that __pfn_to_phys(max_pfn) won't overflow?
eg, the invalid pfn value as 0x1bffc0 will pass pfn_valid if there is
a memory block while the max_pfn is 0xbffc0.
In ARM64, bellowing condition check will help to
>
> > Signed-off-by: Zhaoyang Huang <huangzhaoyang@gmail.com>
> > ---
> >  arch/arm/mm/init.c | 3 ++-
> >  1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/arch/arm/mm/init.c b/arch/arm/mm/init.c
> > index c2daabb..9c4d938 100644
> > --- a/arch/arm/mm/init.c
> > +++ b/arch/arm/mm/init.c
> > @@ -177,7 +177,8 @@ static void __init zone_sizes_init(unsigned long min, unsigned long max_low,
> >  #ifdef CONFIG_HAVE_ARCH_PFN_VALID
> >  int pfn_valid(unsigned long pfn)
> >  {
> > -     return memblock_is_map_memory(__pfn_to_phys(pfn));
> > +     return (pfn > max_pfn) ?
> > +             false : memblock_is_map_memory(__pfn_to_phys(pfn));
> >  }
> >  EXPORT_SYMBOL(pfn_valid);
> >  #endif
> > --
> > 1.9.1
> >
>
> --
> Sincerely yours,
> Mike.
>

Re: [PATCH] arch : arm : add a criteria for pfn_valid

[Zhaoyang Huang]

On Sat, Aug 17, 2019 at 5:00 PM Mike Rapoport <rppt@linux.ibm.com> wrote:
>
> On Sat, Aug 17, 2019 at 11:00:13AM +0800, Zhaoyang Huang wrote:
> > From: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
> >
> > pfn_valid can be wrong while the MSB of physical address be trimed as pfn
> > larger than the max_pfn.
>
> How the overflow of __pfn_to_phys() is related to max_pfn?
> Where is the guarantee that __pfn_to_phys(max_pfn) won't overflow?
eg, the invalid pfn value as 0x1bffc0 will pass pfn_valid if there is
a memory block while the max_pfn is 0xbffc0.
In ARM64, bellowing condition check will help to
>
> > Signed-off-by: Zhaoyang Huang <huangzhaoyang@gmail.com>
> > ---
> >  arch/arm/mm/init.c | 3 ++-
> >  1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/arch/arm/mm/init.c b/arch/arm/mm/init.c
> > index c2daabb..9c4d938 100644
> > --- a/arch/arm/mm/init.c
> > +++ b/arch/arm/mm/init.c
> > @@ -177,7 +177,8 @@ static void __init zone_sizes_init(unsigned long min, unsigned long max_low,
> >  #ifdef CONFIG_HAVE_ARCH_PFN_VALID
> >  int pfn_valid(unsigned long pfn)
> >  {
> > -     return memblock_is_map_memory(__pfn_to_phys(pfn));
> > +     return (pfn > max_pfn) ?
> > +             false : memblock_is_map_memory(__pfn_to_phys(pfn));
> >  }
> >  EXPORT_SYMBOL(pfn_valid);
> >  #endif
> > --
> > 1.9.1
> >
>
> --
> Sincerely yours,
> Mike.
>

Re: [PATCH] arch : arm : add a criteria for pfn_valid

[Russell King - ARM Linux admin]

On Sat, Aug 17, 2019 at 11:00:13AM +0800, Zhaoyang Huang wrote:
> From: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
> 
> pfn_valid can be wrong while the MSB of physical address be trimed as pfn
> larger than the max_pfn.

What scenario are you addressing here?  At a guess, you're addressing
the non-LPAE case with PFNs that correspond with >= 4GiB of memory?

> 
> Signed-off-by: Zhaoyang Huang <huangzhaoyang@gmail.com>
> ---
>  arch/arm/mm/init.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/arm/mm/init.c b/arch/arm/mm/init.c
> index c2daabb..9c4d938 100644
> --- a/arch/arm/mm/init.c
> +++ b/arch/arm/mm/init.c
> @@ -177,7 +177,8 @@ static void __init zone_sizes_init(unsigned long min, unsigned long max_low,
>  #ifdef CONFIG_HAVE_ARCH_PFN_VALID
>  int pfn_valid(unsigned long pfn)
>  {
> -	return memblock_is_map_memory(__pfn_to_phys(pfn));
> +	return (pfn > max_pfn) ?
> +		false : memblock_is_map_memory(__pfn_to_phys(pfn));
>  }
>  EXPORT_SYMBOL(pfn_valid);
>  #endif
> -- 
> 1.9.1
> 
> 

-- 
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTC broadband for 0.8mile line in suburbia: sync at 12.1Mbps down 622kbps up
According to speedtest.net: 11.9Mbps down 500kbps up

Re: [PATCH] arch : arm : add a criteria for pfn_valid

[Zhaoyang Huang]

On Sun, Aug 18, 2019 at 2:32 AM Russell King - ARM Linux admin
<linux@armlinux.org.uk> wrote:
>
> On Sat, Aug 17, 2019 at 11:00:13AM +0800, Zhaoyang Huang wrote:
> > From: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
> >
> > pfn_valid can be wrong while the MSB of physical address be trimed as pfn
> > larger than the max_pfn.
>
> What scenario are you addressing here?  At a guess, you're addressing
> the non-LPAE case with PFNs that correspond with >= 4GiB of memory?
Please find bellowing for the callstack caused by this defect. The
original reason is a invalid PFN passed from userspace which will
introduce a invalid page within stable_page_flags and then kernel
panic.

[46886.723249] c7 [<c031ff98>] (stable_page_flags) from [<c03203f8>]
(kpageflags_read+0x90/0x11c)
[46886.723256] c7  r9:c101ce04 r8:c2d0bf70 r7:c2d0bf70 r6:1fbb10fb
r5:a8686f08 r4:a8686f08
[46886.723264] c7 [<c0320368>] (kpageflags_read) from [<c0312030>]
(proc_reg_read+0x80/0x94)
[46886.723270] c7  r10:000000b4 r9:00000008 r8:c2d0bf70 r7:00000000
r6:00000001 r5:ed8e7240
[46886.723272] c7  r4:00000000
[46886.723280] c7 [<c0311fb0>] (proc_reg_read) from [<c02a6e6c>]
(__vfs_read+0x48/0x150)
[46886.723284] c7  r7:c2d0bf70 r6:c0f09208 r5:c0a4f940 r4:c40326c0
[46886.723290] c7 [<c02a6e24>] (__vfs_read) from [<c02a7018>]
(vfs_read+0xa4/0x158)
[46886.723296] c7  r9:a8686f08 r8:00000008 r7:c2d0bf70 r6:a8686f08
r5:c40326c0 r4:00000008
[46886.723301] c7 [<c02a6f74>] (vfs_read) from [<c02a778c>]
(SyS_pread64+0x80/0xb8)
[46886.723306] c7  r8:00000008 r7:c0f09208 r6:c40326c0 r5:c40326c0 r4:fdd887d8
[46886.723315] c7 [<c02a770c>] (SyS_pread64) from [<c0108620>]
(ret_fast_syscall+0x0/0x28)

>
> >
> > Signed-off-by: Zhaoyang Huang <huangzhaoyang@gmail.com>
> > ---
> >  arch/arm/mm/init.c | 3 ++-
> >  1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/arch/arm/mm/init.c b/arch/arm/mm/init.c
> > index c2daabb..9c4d938 100644
> > --- a/arch/arm/mm/init.c
> > +++ b/arch/arm/mm/init.c
> > @@ -177,7 +177,8 @@ static void __init zone_sizes_init(unsigned long min, unsigned long max_low,
> >  #ifdef CONFIG_HAVE_ARCH_PFN_VALID
> >  int pfn_valid(unsigned long pfn)
> >  {
> > -     return memblock_is_map_memory(__pfn_to_phys(pfn));
> > +     return (pfn > max_pfn) ?
> > +             false : memblock_is_map_memory(__pfn_to_phys(pfn));
> >  }
> >  EXPORT_SYMBOL(pfn_valid);
> >  #endif
> > --
> > 1.9.1
> >
> >
>
> --
> RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
> FTTC broadband for 0.8mile line in suburbia: sync at 12.1Mbps down 622kbps up
> According to speedtest.net: 11.9Mbps down 500kbps up

Re: [PATCH] arch : arm : add a criteria for pfn_valid

[Zhaoyang Huang]

On Sun, Aug 18, 2019 at 2:32 AM Russell King - ARM Linux admin
<linux@armlinux.org.uk> wrote:
>
> On Sat, Aug 17, 2019 at 11:00:13AM +0800, Zhaoyang Huang wrote:
> > From: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
> >
> > pfn_valid can be wrong while the MSB of physical address be trimed as pfn
> > larger than the max_pfn.
>
> What scenario are you addressing here?  At a guess, you're addressing
> the non-LPAE case with PFNs that correspond with >= 4GiB of memory?
Please find bellowing for the callstack caused by this defect. The
original reason is a invalid PFN passed from userspace which will
introduce a invalid page within stable_page_flags and then kernel
panic.

[46886.723249] c7 [<c031ff98>] (stable_page_flags) from [<c03203f8>]
(kpageflags_read+0x90/0x11c)
[46886.723256] c7  r9:c101ce04 r8:c2d0bf70 r7:c2d0bf70 r6:1fbb10fb
r5:a8686f08 r4:a8686f08
[46886.723264] c7 [<c0320368>] (kpageflags_read) from [<c0312030>]
(proc_reg_read+0x80/0x94)
[46886.723270] c7  r10:000000b4 r9:00000008 r8:c2d0bf70 r7:00000000
r6:00000001 r5:ed8e7240
[46886.723272] c7  r4:00000000
[46886.723280] c7 [<c0311fb0>] (proc_reg_read) from [<c02a6e6c>]
(__vfs_read+0x48/0x150)
[46886.723284] c7  r7:c2d0bf70 r6:c0f09208 r5:c0a4f940 r4:c40326c0
[46886.723290] c7 [<c02a6e24>] (__vfs_read) from [<c02a7018>]
(vfs_read+0xa4/0x158)
[46886.723296] c7  r9:a8686f08 r8:00000008 r7:c2d0bf70 r6:a8686f08
r5:c40326c0 r4:00000008
[46886.723301] c7 [<c02a6f74>] (vfs_read) from [<c02a778c>]
(SyS_pread64+0x80/0xb8)
[46886.723306] c7  r8:00000008 r7:c0f09208 r6:c40326c0 r5:c40326c0 r4:fdd887d8
[46886.723315] c7 [<c02a770c>] (SyS_pread64) from [<c0108620>]
(ret_fast_syscall+0x0/0x28)

>
> >
> > Signed-off-by: Zhaoyang Huang <huangzhaoyang@gmail.com>
> > ---
> >  arch/arm/mm/init.c | 3 ++-
> >  1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/arch/arm/mm/init.c b/arch/arm/mm/init.c
> > index c2daabb..9c4d938 100644
> > --- a/arch/arm/mm/init.c
> > +++ b/arch/arm/mm/init.c
> > @@ -177,7 +177,8 @@ static void __init zone_sizes_init(unsigned long min, unsigned long max_low,
> >  #ifdef CONFIG_HAVE_ARCH_PFN_VALID
> >  int pfn_valid(unsigned long pfn)
> >  {
> > -     return memblock_is_map_memory(__pfn_to_phys(pfn));
> > +     return (pfn > max_pfn) ?
> > +             false : memblock_is_map_memory(__pfn_to_phys(pfn));
> >  }
> >  EXPORT_SYMBOL(pfn_valid);
> >  #endif
> > --
> > 1.9.1
> >
> >
>
> --
> RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
> FTTC broadband for 0.8mile line in suburbia: sync at 12.1Mbps down 622kbps up
> According to speedtest.net: 11.9Mbps down 500kbps up

Re: [PATCH] arch : arm : add a criteria for pfn_valid

[Mike Rapoport]

On Sun, Aug 18, 2019 at 03:46:51PM +0800, Zhaoyang Huang wrote:
> On Sun, Aug 18, 2019 at 2:32 AM Russell King - ARM Linux admin
> <linux@armlinux.org.uk> wrote:
> >
> > On Sat, Aug 17, 2019 at 11:00:13AM +0800, Zhaoyang Huang wrote:
> > > From: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
> > >
> > > pfn_valid can be wrong while the MSB of physical address be trimed as pfn
> > > larger than the max_pfn.
> >
> > What scenario are you addressing here?  At a guess, you're addressing
> > the non-LPAE case with PFNs that correspond with >= 4GiB of memory?
> Please find bellowing for the callstack caused by this defect. The
> original reason is a invalid PFN passed from userspace which will
> introduce a invalid page within stable_page_flags and then kernel
> panic.

Yeah, arm64 hit this issue a while ago and it was fixed with commit
5ad356eabc47 ("arm64: mm: check for upper PAGE_SHIFT bits in pfn_valid()").

IMHO, the check 

	if ((addr >> PAGE_SHIFT) != pfn)

is more robust than comparing pfn to max_pfn.

 
> [46886.723249] c7 [<c031ff98>] (stable_page_flags) from [<c03203f8>]
> (kpageflags_read+0x90/0x11c)
> [46886.723256] c7  r9:c101ce04 r8:c2d0bf70 r7:c2d0bf70 r6:1fbb10fb
> r5:a8686f08 r4:a8686f08
> [46886.723264] c7 [<c0320368>] (kpageflags_read) from [<c0312030>]
> (proc_reg_read+0x80/0x94)
> [46886.723270] c7  r10:000000b4 r9:00000008 r8:c2d0bf70 r7:00000000
> r6:00000001 r5:ed8e7240
> [46886.723272] c7  r4:00000000
> [46886.723280] c7 [<c0311fb0>] (proc_reg_read) from [<c02a6e6c>]
> (__vfs_read+0x48/0x150)
> [46886.723284] c7  r7:c2d0bf70 r6:c0f09208 r5:c0a4f940 r4:c40326c0
> [46886.723290] c7 [<c02a6e24>] (__vfs_read) from [<c02a7018>]
> (vfs_read+0xa4/0x158)
> [46886.723296] c7  r9:a8686f08 r8:00000008 r7:c2d0bf70 r6:a8686f08
> r5:c40326c0 r4:00000008
> [46886.723301] c7 [<c02a6f74>] (vfs_read) from [<c02a778c>]
> (SyS_pread64+0x80/0xb8)
> [46886.723306] c7  r8:00000008 r7:c0f09208 r6:c40326c0 r5:c40326c0 r4:fdd887d8
> [46886.723315] c7 [<c02a770c>] (SyS_pread64) from [<c0108620>]
> (ret_fast_syscall+0x0/0x28)
> 
> >
> > >
> > > Signed-off-by: Zhaoyang Huang <huangzhaoyang@gmail.com>
> > > ---
> > >  arch/arm/mm/init.c | 3 ++-
> > >  1 file changed, 2 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/arch/arm/mm/init.c b/arch/arm/mm/init.c
> > > index c2daabb..9c4d938 100644
> > > --- a/arch/arm/mm/init.c
> > > +++ b/arch/arm/mm/init.c
> > > @@ -177,7 +177,8 @@ static void __init zone_sizes_init(unsigned long min, unsigned long max_low,
> > >  #ifdef CONFIG_HAVE_ARCH_PFN_VALID
> > >  int pfn_valid(unsigned long pfn)
> > >  {
> > > -     return memblock_is_map_memory(__pfn_to_phys(pfn));
> > > +     return (pfn > max_pfn) ?
> > > +             false : memblock_is_map_memory(__pfn_to_phys(pfn));
> > >  }
> > >  EXPORT_SYMBOL(pfn_valid);
> > >  #endif
> > > --
> > > 1.9.1
> > >
> > >
> >
> > --
> > RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
> > FTTC broadband for 0.8mile line in suburbia: sync at 12.1Mbps down 622kbps up
> > According to speedtest.net: 11.9Mbps down 500kbps up
> 

-- 
Sincerely yours,
Mike.

Re: [PATCH] arch : arm : add a criteria for pfn_valid

[Russell King - ARM Linux admin]

On Sun, Aug 18, 2019 at 11:20:35AM +0300, Mike Rapoport wrote:
> On Sun, Aug 18, 2019 at 03:46:51PM +0800, Zhaoyang Huang wrote:
> > On Sun, Aug 18, 2019 at 2:32 AM Russell King - ARM Linux admin
> > <linux@armlinux.org.uk> wrote:
> > >
> > > On Sat, Aug 17, 2019 at 11:00:13AM +0800, Zhaoyang Huang wrote:
> > > > From: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
> > > >
> > > > pfn_valid can be wrong while the MSB of physical address be trimed as pfn
> > > > larger than the max_pfn.
> > >
> > > What scenario are you addressing here?  At a guess, you're addressing
> > > the non-LPAE case with PFNs that correspond with >= 4GiB of memory?
> > Please find bellowing for the callstack caused by this defect. The
> > original reason is a invalid PFN passed from userspace which will
> > introduce a invalid page within stable_page_flags and then kernel
> > panic.

Thanks.

> Yeah, arm64 hit this issue a while ago and it was fixed with commit
> 5ad356eabc47 ("arm64: mm: check for upper PAGE_SHIFT bits in pfn_valid()").
> 
> IMHO, the check 
> 
> 	if ((addr >> PAGE_SHIFT) != pfn)
> 
> is more robust than comparing pfn to max_pfn.

Yep, I'd prefer to see:

	phys_addr_t addr = __pfn_to_phys(pfn);

	if (__pfn_to_phys(addr) != pfn)
		return 0;

	return memblock_is_map_memory(addr);

-- 
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTC broadband for 0.8mile line in suburbia: sync at 12.1Mbps down 622kbps up
According to speedtest.net: 11.9Mbps down 500kbps up