On 5/28/19 7:05 PM, Paul Burton wrote:

The virt_addr_valid() function is meant to return true iff virt_to_page() will return a valid struct page reference. This is true iff the address provided is found within the unmapped address range between PAGE_OFFSET & MAP_BASE, but we don't currently check for that condition. Instead we simply mask the address to obtain what will be a physical address if the virtual address is indeed in the desired range, shift it to form a PFN & then call pfn_valid(). This can incorrectly return true if called with a virtual address which, after masking, happens to form a physical address corresponding to a valid PFN.

For example we may vmalloc an address in the kernel mapped region starting a MAP_BASE & obtain the virtual address:

addr = 0xc000000000002000

When masked by virt_to_phys(), which uses __pa() & in turn CPHYSADDR(), we obtain the following (bogus) physical address:

addr = 0x2000

In a common system with PHYS_OFFSET=0 this will correspond to a valid struct page which should really be accessed by virtual address PAGE_OFFSET+0x2000, causing virt_addr_valid() to incorrectly return 1 indicating that the original address corresponds to a struct page.

This is equivalent to the ARM64 change made in commit ca219452c6b8 ("arm64: Correctly bounds check virt_addr_valid").

This fixes fallout when hardened usercopy is enabled caused by the related commit 517e1fbeb65f ("mm/usercopy: Drop extra is_vmalloc_or_module() check") which removed a check for the vmalloc range that was present from the introduction of the hardened usercopy feature.

Signed-off-by: Paul Burton paul.burton@mips.com References: ca219452c6b8 ("arm64: Correctly bounds check virt_addr_valid") References: 517e1fbeb65f ("mm/usercopy: Drop extra is_vmalloc_or_module() check") Reported-by: Julien Cristau jcristau@debian.org Tested-by: YunQiang Su ysu@wavecomp.com

Reviewed-by: Philippe Mathieu-Daudé f4bug@amsat.org

URL: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=929366 Cc: stable@vger.kernel.org # v4.12+

---

arch/mips/mm/mmap.c | 5 +++++ 1 file changed, 5 insertions(+)

diff --git a/arch/mips/mm/mmap.c b/arch/mips/mm/mmap.c index 2f616ebeb7e0..7755a1fad05a 100644 --- a/arch/mips/mm/mmap.c +++ b/arch/mips/mm/mmap.c @@ -203,6 +203,11 @@ unsigned long arch_randomize_brk(struct mm_struct *mm) int __virt_addr_valid(const volatile void *kaddr) {

• unsigned long vaddr = (unsigned long)vaddr;
• if ((vaddr < PAGE_OFFSET) || (vaddr >= MAP_BASE))

• return 0;
  

• return pfn_valid(PFN_DOWN(virt_to_phys(kaddr)));

} EXPORT_SYMBOL_GPL(__virt_addr_valid);

On 5/28/19 7:05 PM, Paul Burton wrote:

The virt_addr_valid() function is meant to return true iff virt_to_page() will return a valid struct page reference. This is true iff the address provided is found within the unmapped address range between PAGE_OFFSET & MAP_BASE, but we don't currently check for that condition. Instead we simply mask the address to obtain what will be a physical address if the virtual address is indeed in the desired range, shift it to form a PFN & then call pfn_valid(). This can incorrectly return true if called with a virtual address which, after masking, happens to form a physical address corresponding to a valid PFN.

For example we may vmalloc an address in the kernel mapped region starting a MAP_BASE & obtain the virtual address:

addr = 0xc000000000002000

When masked by virt_to_phys(), which uses __pa() & in turn CPHYSADDR(), we obtain the following (bogus) physical address:

addr = 0x2000

In a common system with PHYS_OFFSET=0 this will correspond to a valid struct page which should really be accessed by virtual address PAGE_OFFSET+0x2000, causing virt_addr_valid() to incorrectly return 1 indicating that the original address corresponds to a struct page.

This is equivalent to the ARM64 change made in commit ca219452c6b8 ("arm64: Correctly bounds check virt_addr_valid").

This fixes fallout when hardened usercopy is enabled caused by the related commit 517e1fbeb65f ("mm/usercopy: Drop extra is_vmalloc_or_module() check") which removed a check for the vmalloc range that was present from the introduction of the hardened usercopy feature.

Signed-off-by: Paul Burton paul.burton@mips.com References: ca219452c6b8 ("arm64: Correctly bounds check virt_addr_valid") References: 517e1fbeb65f ("mm/usercopy: Drop extra is_vmalloc_or_module() check") Reported-by: Julien Cristau jcristau@debian.org Tested-by: YunQiang Su ysu@wavecomp.com URL: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=929366 Cc: stable@vger.kernel.org # v4.12+

---

arch/mips/mm/mmap.c | 5 +++++ 1 file changed, 5 insertions(+)

diff --git a/arch/mips/mm/mmap.c b/arch/mips/mm/mmap.c index 2f616ebeb7e0..7755a1fad05a 100644 --- a/arch/mips/mm/mmap.c +++ b/arch/mips/mm/mmap.c @@ -203,6 +203,11 @@ unsigned long arch_randomize_brk(struct mm_struct *mm) int __virt_addr_valid(const volatile void *kaddr) {

• unsigned long vaddr = (unsigned long)vaddr;
• if ((vaddr < PAGE_OFFSET) || (vaddr >= MAP_BASE))

• return 0;
  

• return pfn_valid(PFN_DOWN(virt_to_phys(kaddr)));

} EXPORT_SYMBOL_GPL(__virt_addr_valid);

Hi Paul,

Someone complained that this compiled to a constant "return 0" for him: https://bugs.openwrt.org/index.php?do=details&task_id=2305#comment6554

I just checked this on a unmodified 5.2-rc4 with the xway_defconfig and I get this:

0001915c <__virt_addr_valid>: 1915c: 03e00008 jr ra 19160: 00001025 move v0,zero

Is this intended?

Hauke