| Message ID | 1546860080-13027-1-git-send-email-miles.chen@mediatek.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | arm64: trap illegal translations in __virt_to_phys() | expand |
On Mon, Jan 07, 2019 at 07:21:20PM +0800, Miles Chen wrote: > Current __virt_to_phys() only print warning messages for non-linear > addresses. It's hard to catch all warnings by those messages. Why? Are you seeing a large number of warnings somewhere? > So add a VIRTUAL_BUG_ON() to trap all non-linear and non-symbol > addresses (e.g., stack addresses) > > Tested by pass stack addresses and symbol addresses to __pa(). Result: > stack addresses: kernel BUG() Either: * Stacks are vmap'd, and __is_lm_address(stack_addr) is false. We'll produce a WARNING() today (and return a junk physical address). * Stacks are linear mapped, and cannot be distinguished from other linear mapped addresses. The physical address will be valid. ... so I don't understand why you need to change this. > symbol addresses: kernel warning message That should already be the case today, since the kernel image is mapped separately from the linear map, so __is_lm_address(symbol_addr) should be false. > > Maybe we should trap all non-linear address translations in the future. > > Signed-off-by: Miles Chen <miles.chen@mediatek.com> > --- > arch/arm64/mm/physaddr.c | 5 +++++ > 1 file changed, 5 insertions(+) > > diff --git a/arch/arm64/mm/physaddr.c b/arch/arm64/mm/physaddr.c > index 67a9ba9eaa96..f6b935dad19c 100644 > --- a/arch/arm64/mm/physaddr.c > +++ b/arch/arm64/mm/physaddr.c > @@ -14,6 +14,11 @@ phys_addr_t __virt_to_phys(unsigned long x) > (void *)x, > (void *)x); > > + /* trap all non-linear and non-symbol addresses */ > + VIRTUAL_BUG_ON(!__is_lm_address(x) && > + (x < (unsigned long)KERNEL_START || > + x > (unsigned long)KERNEL_END)); The KERNEL_START and KERNEL_END definitions refer to the kernel image, not the linear map, so it doesn't make any sense to permit those here. It is *not* valid to call __virt_to_phys() with a symbol address. We only support those in __virt_to_phys_nodebug() so that broken code has a chance of stumbling on. If you want the kernel to die immediately when it hits a warning here, please set panic_on_warn. Thanks, Mark.
On Mon, 2019-01-07 at 15:00 +0000, Mark Rutland wrote: > On Mon, Jan 07, 2019 at 07:21:20PM +0800, Miles Chen wrote: > > Current __virt_to_phys() only print warning messages for non-linear > > addresses. It's hard to catch all warnings by those messages. > > Why? Are you seeing a large number of warnings somewhere? Official kernel works fine. I saw some cases in our internal branch and we're fixing them. > > > So add a VIRTUAL_BUG_ON() to trap all non-linear and non-symbol > > addresses (e.g., stack addresses) > > > > Tested by pass stack addresses and symbol addresses to __pa(). Result: > > stack addresses: kernel BUG() > > Either: > > * Stacks are vmap'd, and __is_lm_address(stack_addr) is false. We'll > produce a WARNING() today (and return a junk physical address). > > * Stacks are linear mapped, and cannot be distinguished from other > linear mapped addresses. The physical address will be valid. > > ... so I don't understand why you need to change this. For the first case: for vmap'd stack, __pa() returns a junk physical address and it might be easier to debug this incorrect address translation by a BUG() call instead of monitoring the warning log. Like __phys_addr_symbol() does. __phys_addr_symbol() uses VIRTUAL_BUG_ON() for non symbol addresses. > > > symbol addresses: kernel warning message > > That should already be the case today, since the kernel image is mapped > separately from the linear map, so __is_lm_address(symbol_addr) should > be false. > > > > > Maybe we should trap all non-linear address translations in the future. > > > > Signed-off-by: Miles Chen <miles.chen@mediatek.com> > > --- > > arch/arm64/mm/physaddr.c | 5 +++++ > > 1 file changed, 5 insertions(+) > > > > diff --git a/arch/arm64/mm/physaddr.c b/arch/arm64/mm/physaddr.c > > index 67a9ba9eaa96..f6b935dad19c 100644 > > --- a/arch/arm64/mm/physaddr.c > > +++ b/arch/arm64/mm/physaddr.c > > @@ -14,6 +14,11 @@ phys_addr_t __virt_to_phys(unsigned long x) > > (void *)x, > > (void *)x); > > > > + /* trap all non-linear and non-symbol addresses */ > > + VIRTUAL_BUG_ON(!__is_lm_address(x) && > > + (x < (unsigned long)KERNEL_START || > > + x > (unsigned long)KERNEL_END)); > > The KERNEL_START and KERNEL_END definitions refer to the kernel image, > not the linear map, so it doesn't make any sense to permit those here. > > It is *not* valid to call __virt_to_phys() with a symbol address. We > only support those in __virt_to_phys_nodebug() so that broken code has a > chance of stumbling on. > > If you want the kernel to die immediately when it hits a warning here, > please set panic_on_warn. > > Thanks, > Mark.
On Tue, Jan 08, 2019 at 11:24:43AM +0800, Miles Chen wrote: > On Mon, 2019-01-07 at 15:00 +0000, Mark Rutland wrote: > > On Mon, Jan 07, 2019 at 07:21:20PM +0800, Miles Chen wrote: > > > Current __virt_to_phys() only print warning messages for non-linear > > > addresses. It's hard to catch all warnings by those messages. > > > > Why? Are you seeing a large number of warnings somewhere? > > Official kernel works fine. I saw some cases in our internal branch and > we're fixing them. > > > > > > So add a VIRTUAL_BUG_ON() to trap all non-linear and non-symbol > > > addresses (e.g., stack addresses) > > > > > > Tested by pass stack addresses and symbol addresses to __pa(). Result: > > > stack addresses: kernel BUG() > > > > Either: > > > > * Stacks are vmap'd, and __is_lm_address(stack_addr) is false. We'll > > produce a WARNING() today (and return a junk physical address). > > > > * Stacks are linear mapped, and cannot be distinguished from other > > linear mapped addresses. The physical address will be valid. > > > > ... so I don't understand why you need to change this. > > For the first case: for vmap'd stack, __pa() returns a junk > physical address and it might be easier to debug this incorrect address > translation by a BUG() call instead of monitoring the warning log. I think that's an argument for upgrading the existing WARN() to a BUG(), rather than adding a separate VIRTUAL_BUG_ON(). However, there are cases where the junk physical address is not used to perform an access, and the WARN() is more helpful. You can set panic_on_warn to get an immediate panic() when the WARN() fires. Is there some reason that approach doesn't work for you? Thanks, Mark.
On Tue, 2019-01-08 at 12:14 +0000, Mark Rutland wrote: > On Tue, Jan 08, 2019 at 11:24:43AM +0800, Miles Chen wrote: > > On Mon, 2019-01-07 at 15:00 +0000, Mark Rutland wrote: > > > On Mon, Jan 07, 2019 at 07:21:20PM +0800, Miles Chen wrote: > > > > Current __virt_to_phys() only print warning messages for non-linear > > > > addresses. It's hard to catch all warnings by those messages. > > > > > > Why? Are you seeing a large number of warnings somewhere? > > > > Official kernel works fine. I saw some cases in our internal branch and > > we're fixing them. > > > > > > > > > So add a VIRTUAL_BUG_ON() to trap all non-linear and non-symbol > > > > addresses (e.g., stack addresses) > > > > > > > > Tested by pass stack addresses and symbol addresses to __pa(). Result: > > > > stack addresses: kernel BUG() > > > > > > Either: > > > > > > * Stacks are vmap'd, and __is_lm_address(stack_addr) is false. We'll > > > produce a WARNING() today (and return a junk physical address). > > > > > > * Stacks are linear mapped, and cannot be distinguished from other > > > linear mapped addresses. The physical address will be valid. > > > > > > ... so I don't understand why you need to change this. > > > > For the first case: for vmap'd stack, __pa() returns a junk > > physical address and it might be easier to debug this incorrect address > > translation by a BUG() call instead of monitoring the warning log. > > I think that's an argument for upgrading the existing WARN() to a BUG(), > rather than adding a separate VIRTUAL_BUG_ON(). > > However, there are cases where the junk physical address is not used to > perform an access, and the WARN() is more helpful. got it. > > You can set panic_on_warn to get an immediate panic() when the WARN() > fires. Is there some reason that approach doesn't work for you? panic_on_warn works fine. thanks for your comment. cheers, Miles > > Thanks, > Mark.
diff --git a/arch/arm64/mm/physaddr.c b/arch/arm64/mm/physaddr.c index 67a9ba9eaa96..f6b935dad19c 100644 --- a/arch/arm64/mm/physaddr.c +++ b/arch/arm64/mm/physaddr.c @@ -14,6 +14,11 @@ phys_addr_t __virt_to_phys(unsigned long x) (void *)x, (void *)x); + /* trap all non-linear and non-symbol addresses */ + VIRTUAL_BUG_ON(!__is_lm_address(x) && + (x < (unsigned long)KERNEL_START || + x > (unsigned long)KERNEL_END)); + return __virt_to_phys_nodebug(x); } EXPORT_SYMBOL(__virt_to_phys);
Current __virt_to_phys() only print warning messages for non-linear addresses. It's hard to catch all warnings by those messages. So add a VIRTUAL_BUG_ON() to trap all non-linear and non-symbol addresses (e.g., stack addresses) Tested by pass stack addresses and symbol addresses to __pa(). Result: stack addresses: kernel BUG() symbol addresses: kernel warning message Maybe we should trap all non-linear address translations in the future. Signed-off-by: Miles Chen <miles.chen@mediatek.com> --- arch/arm64/mm/physaddr.c | 5 +++++ 1 file changed, 5 insertions(+)