diff mbox series

tracing/user_events: align uaddr on unsigned long alignment

Message ID 20230914131102.179100-1-cleger@rivosinc.com (mailing list archive)
State Superseded
Headers show
Series tracing/user_events: align uaddr on unsigned long alignment | expand

Commit Message

Clément Léger Sept. 14, 2023, 1:11 p.m. UTC
enabler->uaddr can be aligned on 32 or 64 bits. If aligned on 32 bits,
this will result in a misaligned access on 64 bits architectures since
set_bit()/clear_bit() are expecting an unsigned long (aligned) pointer.
On architecture that do not support misaligned access, this will crash
the kernel. Align uaddr on unsigned long size to avoid such behavior.
This bug was found while running kselftests on RISC-V.

Fixes: 7235759084a4 ("tracing/user_events: Use remote writes for event enablement")
Signed-off-by: Clément Léger <cleger@rivosinc.com>
---
 kernel/trace/trace_events_user.c | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

Comments

Beau Belgrave Sept. 14, 2023, 4:42 p.m. UTC | #1
On Thu, Sep 14, 2023 at 03:11:02PM +0200, Clément Léger wrote:
> enabler->uaddr can be aligned on 32 or 64 bits. If aligned on 32 bits,
> this will result in a misaligned access on 64 bits architectures since
> set_bit()/clear_bit() are expecting an unsigned long (aligned) pointer.
> On architecture that do not support misaligned access, this will crash
> the kernel. Align uaddr on unsigned long size to avoid such behavior.
> This bug was found while running kselftests on RISC-V.
> 
> Fixes: 7235759084a4 ("tracing/user_events: Use remote writes for event enablement")
> Signed-off-by: Clément Léger <cleger@rivosinc.com>

Thanks for fixing! I have a few comments on this.

I unfortunately do not have RISC-V hardware to validate this on.

> ---
>  kernel/trace/trace_events_user.c | 12 +++++++++---
>  1 file changed, 9 insertions(+), 3 deletions(-)
> 
> diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c
> index 6f046650e527..580c0fe4b23e 100644
> --- a/kernel/trace/trace_events_user.c
> +++ b/kernel/trace/trace_events_user.c
> @@ -479,7 +479,7 @@ static int user_event_enabler_write(struct user_event_mm *mm,
>  				    bool fixup_fault, int *attempt)
>  {
>  	unsigned long uaddr = enabler->addr;
> -	unsigned long *ptr;
> +	unsigned long *ptr, bit_offset;
>  	struct page *page;
>  	void *kaddr;
>  	int ret;
> @@ -511,13 +511,19 @@ static int user_event_enabler_write(struct user_event_mm *mm,
>  	}
>  
>  	kaddr = kmap_local_page(page);
> +
> +	bit_offset = uaddr & (sizeof(unsigned long) - 1);
> +	if (bit_offset) {
> +		bit_offset *= 8;

I think for future readers of this code it would be more clear to use
BITS_PER_BYTE instead of the hardcoded 8. Given we always align on a
"natural" boundary, I believe the bit_offset will always be 32 bits.

A comment here might help clarify why we do this as well in case folks
don't see the change description.

> +		uaddr &= ~(sizeof(unsigned long) - 1);

Shouldn't this uaddr change be done before calling pin_user_pages_remote()
to ensure things cannot go bad? (I don't think they can, but it looks a
little odd).

Thanks,
-Beau

> +	}
>  	ptr = kaddr + (uaddr & ~PAGE_MASK);
>  
>  	/* Update bit atomically, user tracers must be atomic as well */
>  	if (enabler->event && enabler->event->status)
> -		set_bit(ENABLE_BIT(enabler), ptr);
> +		set_bit(ENABLE_BIT(enabler) + bit_offset, ptr);
>  	else
> -		clear_bit(ENABLE_BIT(enabler), ptr);
> +		clear_bit(ENABLE_BIT(enabler) + bit_offset, ptr);
>  
>  	kunmap_local(kaddr);
>  	unpin_user_pages_dirty_lock(&page, 1, true);
> -- 
> 2.40.1
Steven Rostedt Sept. 14, 2023, 5:17 p.m. UTC | #2
On Thu, 14 Sep 2023 15:11:02 +0200
Clément Léger <cleger@rivosinc.com> wrote:

> @@ -511,13 +511,19 @@ static int user_event_enabler_write(struct user_event_mm *mm,
>  	}
>  
>  	kaddr = kmap_local_page(page);
> +
> +	bit_offset = uaddr & (sizeof(unsigned long) - 1);
> +	if (bit_offset) {
> +		bit_offset *= 8;
> +		uaddr &= ~(sizeof(unsigned long) - 1);
> +	}
>  	ptr = kaddr + (uaddr & ~PAGE_MASK);
>  
>  	/* Update bit atomically, user tracers must be atomic as well */
>  	if (enabler->event && enabler->event->status)
> -		set_bit(ENABLE_BIT(enabler), ptr);
> +		set_bit(ENABLE_BIT(enabler) + bit_offset, ptr);
>  	else
> -		clear_bit(ENABLE_BIT(enabler), ptr);
> +		clear_bit(ENABLE_BIT(enabler) + bit_offset, ptr);
>  
>  	kunmap_local(kaddr);
>  	unpin_user_pages_dirty_lock(&page, 1, true);

Does this work for big endian machines too?

Little endian layout:

 uaddr = 0xbeef0004
 enabler = 1;

 memory at 0xbeef0000:  00 00 00 00 02 00 00 00
                                    ^
                                    addr: 0xbeef0004

				(enabler is set )

	bitoffset = uaddr & (sizeof(unsigned long) - 1); bitoffset = 4
	bit_offset *= 8;				 bitoffset = 32
	uaddr &= ~(sizeof(unsigned long) - 1);		 uaddr = 0xbeef0000

	ptr = kaddr + (uaddr & ~PAGE_MASK);

	clear_bit(1 + 32, ptr);

 memory at 0xbeef0000:  00 00 00 00 02 00 00 00
                                     ^
				bit 33 of 0xbeef0000

Now lets look at big endian layout:

 uaddr = 0xbeef0004
 enabler = 1;

 memory at 0xbeef0000:  00 00 00 00 00 00 00 02
                                    ^
                                    addr: 0xbeef0004

				(enabler is set )

	bitoffset = uaddr & (sizeof(unsigned long) - 1); bitoffset = 4
	bit_offset *= 8;				 bitoffset = 32
	uaddr &= ~(sizeof(unsigned long) - 1);		 uaddr = 0xbeef0000

	ptr = kaddr + (uaddr & ~PAGE_MASK);

	clear_bit(1 + 32, ptr);

 memory at 0xbeef0000:  00 00 00 00 00 00 00 02
                                  ^
				bit 33 of 0xbeef0000

I don't think that's what you expected!

-- Steve
Steven Rostedt Sept. 14, 2023, 5:29 p.m. UTC | #3
On Thu, 14 Sep 2023 13:17:00 -0400
Steven Rostedt <rostedt@goodmis.org> wrote:

> Now lets look at big endian layout:
> 
>  uaddr = 0xbeef0004
>  enabler = 1;
> 
>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
>                                     ^
>                                     addr: 0xbeef0004
> 
> 				(enabler is set )
> 
> 	bitoffset = uaddr & (sizeof(unsigned long) - 1); bitoffset = 4
> 	bit_offset *= 8;				 bitoffset = 32
> 	uaddr &= ~(sizeof(unsigned long) - 1);		 uaddr = 0xbeef0000
> 
> 	ptr = kaddr + (uaddr & ~PAGE_MASK);
> 
> 	clear_bit(1 + 32, ptr);
> 
>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
>                                   ^
> 				bit 33 of 0xbeef0000
> 
> I don't think that's what you expected!

I believe the above can be fixed with:

	bit_offset = uaddr & (sizeof(unsigned long) - 1);
	if (bit_offset) {
#ifdef CONFIG_CPU_BIG_ENDIAN
		bit_offest = 0;
#else
		bit_offset *= BITS_PER_BYTE;
#endif
		uaddr &= ~(sizeof(unsigned long) - 1);
	}

-- Steve
Clément Léger Sept. 14, 2023, 5:32 p.m. UTC | #4
On 14/09/2023 19:29, Steven Rostedt wrote:
> On Thu, 14 Sep 2023 13:17:00 -0400
> Steven Rostedt <rostedt@goodmis.org> wrote:
> 
>> Now lets look at big endian layout:
>>
>>  uaddr = 0xbeef0004
>>  enabler = 1;
>>
>>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
>>                                     ^
>>                                     addr: 0xbeef0004
>>
>> 				(enabler is set )
>>
>> 	bitoffset = uaddr & (sizeof(unsigned long) - 1); bitoffset = 4
>> 	bit_offset *= 8;				 bitoffset = 32
>> 	uaddr &= ~(sizeof(unsigned long) - 1);		 uaddr = 0xbeef0000
>>
>> 	ptr = kaddr + (uaddr & ~PAGE_MASK);
>>
>> 	clear_bit(1 + 32, ptr);
>>
>>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
>>                                   ^
>> 				bit 33 of 0xbeef0000
>>
>> I don't think that's what you expected!
> 
> I believe the above can be fixed with:
> 
> 	bit_offset = uaddr & (sizeof(unsigned long) - 1);
> 	if (bit_offset) {
> #ifdef CONFIG_CPU_BIG_ENDIAN
> 		bit_offest = 0;
> #else
> 		bit_offset *= BITS_PER_BYTE;
> #endif
> 		uaddr &= ~(sizeof(unsigned long) - 1);
> 	}

Hi Steven,

Nice catch ! I don't have big endian at end but I'll check that.

Thanks,

> 
> -- Steve
Clément Léger Sept. 14, 2023, 5:42 p.m. UTC | #5
On 14/09/2023 18:42, Beau Belgrave wrote:
> On Thu, Sep 14, 2023 at 03:11:02PM +0200, Clément Léger wrote:
>> enabler->uaddr can be aligned on 32 or 64 bits. If aligned on 32 bits,
>> this will result in a misaligned access on 64 bits architectures since
>> set_bit()/clear_bit() are expecting an unsigned long (aligned) pointer.
>> On architecture that do not support misaligned access, this will crash
>> the kernel. Align uaddr on unsigned long size to avoid such behavior.
>> This bug was found while running kselftests on RISC-V.
>>
>> Fixes: 7235759084a4 ("tracing/user_events: Use remote writes for event enablement")
>> Signed-off-by: Clément Léger <cleger@rivosinc.com>
> 
> Thanks for fixing! I have a few comments on this.
> 
> I unfortunately do not have RISC-V hardware to validate this on.
> 
>> ---
>>  kernel/trace/trace_events_user.c | 12 +++++++++---
>>  1 file changed, 9 insertions(+), 3 deletions(-)
>>
>> diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c
>> index 6f046650e527..580c0fe4b23e 100644
>> --- a/kernel/trace/trace_events_user.c
>> +++ b/kernel/trace/trace_events_user.c
>> @@ -479,7 +479,7 @@ static int user_event_enabler_write(struct user_event_mm *mm,
>>  				    bool fixup_fault, int *attempt)
>>  {
>>  	unsigned long uaddr = enabler->addr;
>> -	unsigned long *ptr;
>> +	unsigned long *ptr, bit_offset;
>>  	struct page *page;
>>  	void *kaddr;
>>  	int ret;
>> @@ -511,13 +511,19 @@ static int user_event_enabler_write(struct user_event_mm *mm,
>>  	}
>>  
>>  	kaddr = kmap_local_page(page);
>> +
>> +	bit_offset = uaddr & (sizeof(unsigned long) - 1);
>> +	if (bit_offset) {
>> +		bit_offset *= 8;
> 
> I think for future readers of this code it would be more clear to use
> BITS_PER_BYTE instead of the hardcoded 8. Given we always align on a
> "natural" boundary, I believe the bit_offset will always be 32 bits.
> 
> A comment here might help clarify why we do this as well in case folks
> don't see the change description.

Hi Beau,

Yes sure, I'll add a comment and use the define as well.

> 
>> +		uaddr &= ~(sizeof(unsigned long) - 1);
> 
> Shouldn't this uaddr change be done before calling pin_user_pages_remote()
> to ensure things cannot go bad? (I don't think they can, but it looks a
> little odd).

Indeed, I don't think that will cause any problem since pin_user_pages
will return a page aligned address anyway and that aligning uaddr will
not yield any page crossing. But I'll check to be sure and move that
before the call if needed.

Clément

> 
> Thanks,
> -Beau
> 
>> +	}
>>  	ptr = kaddr + (uaddr & ~PAGE_MASK);
>>  
>>  	/* Update bit atomically, user tracers must be atomic as well */
>>  	if (enabler->event && enabler->event->status)
>> -		set_bit(ENABLE_BIT(enabler), ptr);
>> +		set_bit(ENABLE_BIT(enabler) + bit_offset, ptr);
>>  	else
>> -		clear_bit(ENABLE_BIT(enabler), ptr);
>> +		clear_bit(ENABLE_BIT(enabler) + bit_offset, ptr);
>>  
>>  	kunmap_local(kaddr);
>>  	unpin_user_pages_dirty_lock(&page, 1, true);
>> -- 
>> 2.40.1
Masami Hiramatsu (Google) Sept. 15, 2023, 2:54 a.m. UTC | #6
On Thu, 14 Sep 2023 15:11:02 +0200
Clément Léger <cleger@rivosinc.com> wrote:

> enabler->uaddr can be aligned on 32 or 64 bits. If aligned on 32 bits,
> this will result in a misaligned access on 64 bits architectures since
> set_bit()/clear_bit() are expecting an unsigned long (aligned) pointer.
> On architecture that do not support misaligned access, this will crash
> the kernel. Align uaddr on unsigned long size to avoid such behavior.
> This bug was found while running kselftests on RISC-V.
> 
> Fixes: 7235759084a4 ("tracing/user_events: Use remote writes for event enablement")
> Signed-off-by: Clément Léger <cleger@rivosinc.com>
> ---
>  kernel/trace/trace_events_user.c | 12 +++++++++---
>  1 file changed, 9 insertions(+), 3 deletions(-)
> 
> diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c
> index 6f046650e527..580c0fe4b23e 100644
> --- a/kernel/trace/trace_events_user.c
> +++ b/kernel/trace/trace_events_user.c
> @@ -479,7 +479,7 @@ static int user_event_enabler_write(struct user_event_mm *mm,
>  				    bool fixup_fault, int *attempt)
>  {
>  	unsigned long uaddr = enabler->addr;
> -	unsigned long *ptr;
> +	unsigned long *ptr, bit_offset;
>  	struct page *page;
>  	void *kaddr;
>  	int ret;
> @@ -511,13 +511,19 @@ static int user_event_enabler_write(struct user_event_mm *mm,
>  	}
>  
>  	kaddr = kmap_local_page(page);
> +
> +	bit_offset = uaddr & (sizeof(unsigned long) - 1);
> +	if (bit_offset) {
> +		bit_offset *= 8;
> +		uaddr &= ~(sizeof(unsigned long) - 1);
> +	}
>  	ptr = kaddr + (uaddr & ~PAGE_MASK);
>  
>  	/* Update bit atomically, user tracers must be atomic as well */
>  	if (enabler->event && enabler->event->status)
> -		set_bit(ENABLE_BIT(enabler), ptr);
> +		set_bit(ENABLE_BIT(enabler) + bit_offset, ptr);
>  	else
> -		clear_bit(ENABLE_BIT(enabler), ptr);
> +		clear_bit(ENABLE_BIT(enabler) + bit_offset, ptr);

What we need are generic set_bit_aligned() and clear_bit_aligned(), which align the ptr
by unsigned long. (I think it should be done in set_bit/clear_bit, for architecture
which requires aligned access...)

#define LONG_ALIGN_DIFF(p)	(p) & (sizeof(long) -1)
#define LONG_ALINGNED(p)		(p) & ~(sizeof(long) - 1)

static inline void set_bit_aligned(int bit, unsigned long *ptr)
{
	int offs = LONG_ALIGN_DIFF(ptr) * 8;

#ifdef __BIGENDIAN
	if (bit >= offs) {
		set_bit(bit - offs, LONG_ALIGNED(ptr));
	} else {
		set_bit(bit + BITS_PER_LONG - offs, LONG_ALIGNED(ptr) + 1);
	}
#else
	if (bit < BITS_PER_LONG - offs) {
		set_bit(bit + offs, LONG_ALIGNED(ptr));
	} else {
		set_bit(bit - BITS_PER_LONG + offs, LONG_ALIGNED(ptr) + 1);
	}
#endif
}

And use it.

Thank you,

>  
>  	kunmap_local(kaddr);
>  	unpin_user_pages_dirty_lock(&page, 1, true);
> -- 
> 2.40.1
>
Clément Léger Sept. 17, 2023, 2:10 p.m. UTC | #7
On 15/09/2023 04:54, Masami Hiramatsu (Google) wrote:
> On Thu, 14 Sep 2023 15:11:02 +0200
> Clément Léger <cleger@rivosinc.com> wrote:
> 
>> enabler->uaddr can be aligned on 32 or 64 bits. If aligned on 32 bits,
>> this will result in a misaligned access on 64 bits architectures since
>> set_bit()/clear_bit() are expecting an unsigned long (aligned) pointer.
>> On architecture that do not support misaligned access, this will crash
>> the kernel. Align uaddr on unsigned long size to avoid such behavior.
>> This bug was found while running kselftests on RISC-V.
>>
>> Fixes: 7235759084a4 ("tracing/user_events: Use remote writes for event enablement")
>> Signed-off-by: Clément Léger <cleger@rivosinc.com>
>> ---
>>  kernel/trace/trace_events_user.c | 12 +++++++++---
>>  1 file changed, 9 insertions(+), 3 deletions(-)
>>
>> diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c
>> index 6f046650e527..580c0fe4b23e 100644
>> --- a/kernel/trace/trace_events_user.c
>> +++ b/kernel/trace/trace_events_user.c
>> @@ -479,7 +479,7 @@ static int user_event_enabler_write(struct user_event_mm *mm,
>>  				    bool fixup_fault, int *attempt)
>>  {
>>  	unsigned long uaddr = enabler->addr;
>> -	unsigned long *ptr;
>> +	unsigned long *ptr, bit_offset;
>>  	struct page *page;
>>  	void *kaddr;
>>  	int ret;
>> @@ -511,13 +511,19 @@ static int user_event_enabler_write(struct user_event_mm *mm,
>>  	}
>>  
>>  	kaddr = kmap_local_page(page);
>> +
>> +	bit_offset = uaddr & (sizeof(unsigned long) - 1);
>> +	if (bit_offset) {
>> +		bit_offset *= 8;
>> +		uaddr &= ~(sizeof(unsigned long) - 1);
>> +	}
>>  	ptr = kaddr + (uaddr & ~PAGE_MASK);
>>  
>>  	/* Update bit atomically, user tracers must be atomic as well */
>>  	if (enabler->event && enabler->event->status)
>> -		set_bit(ENABLE_BIT(enabler), ptr);
>> +		set_bit(ENABLE_BIT(enabler) + bit_offset, ptr);
>>  	else
>> -		clear_bit(ENABLE_BIT(enabler), ptr);
>> +		clear_bit(ENABLE_BIT(enabler) + bit_offset, ptr);
> 
> What we need are generic set_bit_aligned() and clear_bit_aligned(), which align the ptr
> by unsigned long. (I think it should be done in set_bit/clear_bit, for architecture
> which requires aligned access...)
> 
> #define LONG_ALIGN_DIFF(p)	(p) & (sizeof(long) -1)
> #define LONG_ALINGNED(p)		(p) & ~(sizeof(long) - 1)
> 
> static inline void set_bit_aligned(int bit, unsigned long *ptr)
> {
> 	int offs = LONG_ALIGN_DIFF(ptr) * 8;
> 
> #ifdef __BIGENDIAN
> 	if (bit >= offs) {
> 		set_bit(bit - offs, LONG_ALIGNED(ptr));
> 	} else {
> 		set_bit(bit + BITS_PER_LONG - offs, LONG_ALIGNED(ptr) + 1);
> 	}
> #else
> 	if (bit < BITS_PER_LONG - offs) {
> 		set_bit(bit + offs, LONG_ALIGNED(ptr));
> 	} else {
> 		set_bit(bit - BITS_PER_LONG + offs, LONG_ALIGNED(ptr) + 1);
> 	}
> #endif
> }

Hi Masami,

Indeed, that is a more elegant version, thanks for the snippet.

Clément

> 
> And use it.
> 
> Thank you,
> 
>>  
>>  	kunmap_local(kaddr);
>>  	unpin_user_pages_dirty_lock(&page, 1, true);
>> -- 
>> 2.40.1
>>
> 
>
David Laight Sept. 17, 2023, 9:09 p.m. UTC | #8
From: Clément Léger
> Sent: 14 September 2023 14:11
> 
> enabler->uaddr can be aligned on 32 or 64 bits. If aligned on 32 bits,
> this will result in a misaligned access on 64 bits architectures since
> set_bit()/clear_bit() are expecting an unsigned long (aligned) pointer.
> On architecture that do not support misaligned access, this will crash
> the kernel. Align uaddr on unsigned long size to avoid such behavior.
> This bug was found while running kselftests on RISC-V.

You don't want to do it on x86-64 either.
A locked accesses that crosses a cache line boundary is horrid.
So horrid that recent cpu's can be made to fault.

I'd also doubt that other cpu that can do misaligned transfers
can even do locked ones.

For x86 (and LE) the long[] bitmap can be treated as char[]
avoiding all the problems.

Perhaps there ought to be bit a bit-array based on char[]
(not long[]) that would be endianness independent and
use byte-sized atomics.
(IIRC that is still an issue on sparc32...)

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)
Clément Léger Sept. 18, 2023, 8:37 a.m. UTC | #9
On 17/09/2023 23:09, David Laight wrote:
> From: Clément Léger
>> Sent: 14 September 2023 14:11
>>
>> enabler->uaddr can be aligned on 32 or 64 bits. If aligned on 32 bits,
>> this will result in a misaligned access on 64 bits architectures since
>> set_bit()/clear_bit() are expecting an unsigned long (aligned) pointer.
>> On architecture that do not support misaligned access, this will crash
>> the kernel. Align uaddr on unsigned long size to avoid such behavior.
>> This bug was found while running kselftests on RISC-V.
> 
> You don't want to do it on x86-64 either.
> A locked accesses that crosses a cache line boundary is horrid.
> So horrid that recent cpu's can be made to fault.

Hi David,

Thanks for the additional information.

> 
> I'd also doubt that other cpu that can do misaligned transfers
> can even do locked ones.
> 
> For x86 (and LE) the long[] bitmap can be treated as char[]
> avoiding all the problems.
> 
> Perhaps there ought to be bit a bit-array based on char[]
> (not long[]) that would be endianness independent and
> use byte-sized atomics.

That would work for a few architectures but I don't think all of them
have byte "grain" atomics. So I guess Masami solution (long aligned
set/clear_bit()) remains the best out there.

Clément

> (IIRC that is still an issue on sparc32...)
> 
> 	David
> 
> -
> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> Registration No: 1397386 (Wales)
Clément Léger Sept. 19, 2023, 12:59 p.m. UTC | #10
On 14/09/2023 19:29, Steven Rostedt wrote:
> On Thu, 14 Sep 2023 13:17:00 -0400
> Steven Rostedt <rostedt@goodmis.org> wrote:
> 
>> Now lets look at big endian layout:
>>
>>  uaddr = 0xbeef0004
>>  enabler = 1;
>>
>>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
>>                                     ^
>>                                     addr: 0xbeef0004
>>
>> 				(enabler is set )
>>
>> 	bitoffset = uaddr & (sizeof(unsigned long) - 1); bitoffset = 4
>> 	bit_offset *= 8;				 bitoffset = 32
>> 	uaddr &= ~(sizeof(unsigned long) - 1);		 uaddr = 0xbeef0000
>>
>> 	ptr = kaddr + (uaddr & ~PAGE_MASK);
>>
>> 	clear_bit(1 + 32, ptr);
>>
>>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
>>                                   ^
>> 				bit 33 of 0xbeef0000
>>
>> I don't think that's what you expected!
> 
> I believe the above can be fixed with:
> 
> 	bit_offset = uaddr & (sizeof(unsigned long) - 1);
> 	if (bit_offset) {
> #ifdef CONFIG_CPU_BIG_ENDIAN
> 		bit_offest = 0;
> #else
> 		bit_offset *= BITS_PER_BYTE;
> #endif
> 		uaddr &= ~(sizeof(unsigned long) - 1);
> 	}
> 
> -- Steve


Actually, after looking more in depth at that, it seems like there are
actually 2 problems that can happen.

First one is atomic access misalignment due to enable_size == 4 and
uaddr not being aligned on a (long) boundary on 64 bits architecture.
This can generate misaligned exceptions on various architectures. This
can be fixed in a more general way according to Masami snippet.

Second one that I can see is on 64 bits, big endian architectures with
enable_size == 4. In that case, the bit provided by the userspace won't
be correctly set since this code kind of assume that the atomic are done
on 32bits value. Since the kernel assume long sized atomic operation, on
big endian 64 bits architecture, the updated bit will actually be in the
next 32 bits word.

Can someone confirm my understanding ?

Clément
Beau Belgrave Sept. 22, 2023, 7:22 p.m. UTC | #11
On Tue, Sep 19, 2023 at 02:59:12PM +0200, Clément Léger wrote:
> 
> 
> On 14/09/2023 19:29, Steven Rostedt wrote:
> > On Thu, 14 Sep 2023 13:17:00 -0400
> > Steven Rostedt <rostedt@goodmis.org> wrote:
> > 
> >> Now lets look at big endian layout:
> >>
> >>  uaddr = 0xbeef0004
> >>  enabler = 1;
> >>
> >>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
> >>                                     ^
> >>                                     addr: 0xbeef0004
> >>
> >> 				(enabler is set )
> >>
> >> 	bitoffset = uaddr & (sizeof(unsigned long) - 1); bitoffset = 4
> >> 	bit_offset *= 8;				 bitoffset = 32
> >> 	uaddr &= ~(sizeof(unsigned long) - 1);		 uaddr = 0xbeef0000
> >>
> >> 	ptr = kaddr + (uaddr & ~PAGE_MASK);
> >>
> >> 	clear_bit(1 + 32, ptr);
> >>
> >>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
> >>                                   ^
> >> 				bit 33 of 0xbeef0000
> >>
> >> I don't think that's what you expected!
> > 
> > I believe the above can be fixed with:
> > 
> > 	bit_offset = uaddr & (sizeof(unsigned long) - 1);
> > 	if (bit_offset) {
> > #ifdef CONFIG_CPU_BIG_ENDIAN
> > 		bit_offest = 0;
> > #else
> > 		bit_offset *= BITS_PER_BYTE;
> > #endif
> > 		uaddr &= ~(sizeof(unsigned long) - 1);
> > 	}
> > 
> > -- Steve
> 
> 
> Actually, after looking more in depth at that, it seems like there are
> actually 2 problems that can happen.
> 
> First one is atomic access misalignment due to enable_size == 4 and
> uaddr not being aligned on a (long) boundary on 64 bits architecture.
> This can generate misaligned exceptions on various architectures. This
> can be fixed in a more general way according to Masami snippet.
> 
> Second one that I can see is on 64 bits, big endian architectures with
> enable_size == 4. In that case, the bit provided by the userspace won't
> be correctly set since this code kind of assume that the atomic are done
> on 32bits value. Since the kernel assume long sized atomic operation, on
> big endian 64 bits architecture, the updated bit will actually be in the
> next 32 bits word.
> 
> Can someone confirm my understanding ?
> 

I have a ppc 64bit BE VM I've been validating this on. If we do the
shifting within user_events (vs a generic set_bit_aligned approach)
64bit BE does not need additional bit manipulation. However, if we were
to blindly pass the address and bit as is to set_bit_aligned() it
assumes the bit number is for a long, not a 32 bit word. So for that
approach we would need to offset the bit in the unaligned case.

Here's a patch I have that I've validated on ppc64 BE, aarch64 LE, and
x86_64 LE. I personally feel more comfortable with this approach than
the generic set_bit_aligned() one.

Thanks,
-Beau

diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c
index e3f2b8d72e01..ae854374d0b7 100644
--- a/kernel/trace/trace_events_user.c
+++ b/kernel/trace/trace_events_user.c
@@ -162,6 +162,23 @@ struct user_event_validator {
 	int			flags;
 };
 
+static inline void align_addr_bit(unsigned long *addr, int *bit)
+{
+	if (IS_ALIGNED(*addr, sizeof(long)))
+		return;
+
+	*addr = ALIGN_DOWN(*addr, sizeof(long));
+
+	/*
+	 * We only support 32 and 64 bit values. The only time we need
+	 * to align is a 32 bit value on a 64 bit kernel, which on LE
+	 * is always 32 bits, and on BE requires no change.
+	 */
+#ifdef __LITTLE_ENDIAN
+	*bit += 32;
+#endif
+}
+
 typedef void (*user_event_func_t) (struct user_event *user, struct iov_iter *i,
 				   void *tpdata, bool *faulted);
 
@@ -481,6 +498,7 @@ static int user_event_enabler_write(struct user_event_mm *mm,
 	unsigned long *ptr;
 	struct page *page;
 	void *kaddr;
+	int bit = ENABLE_BIT(enabler);
 	int ret;
 
 	lockdep_assert_held(&event_mutex);
@@ -496,6 +514,8 @@ static int user_event_enabler_write(struct user_event_mm *mm,
 		     test_bit(ENABLE_VAL_FREEING_BIT, ENABLE_BITOPS(enabler))))
 		return -EBUSY;
 
+	align_addr_bit(&uaddr, &bit);
+
 	ret = pin_user_pages_remote(mm->mm, uaddr, 1, FOLL_WRITE | FOLL_NOFAULT,
 				    &page, NULL);
 
@@ -514,9 +534,9 @@ static int user_event_enabler_write(struct user_event_mm *mm,
 
 	/* Update bit atomically, user tracers must be atomic as well */
 	if (enabler->event && enabler->event->status)
-		set_bit(ENABLE_BIT(enabler), ptr);
+		set_bit(bit, ptr);
 	else
-		clear_bit(ENABLE_BIT(enabler), ptr);
+		clear_bit(bit, ptr);
 
 	kunmap_local(kaddr);
 	unpin_user_pages_dirty_lock(&page, 1, true);
Beau Belgrave Sept. 22, 2023, 8 p.m. UTC | #12
On Tue, Sep 19, 2023 at 02:59:12PM +0200, Clément Léger wrote:
> 
> 
> On 14/09/2023 19:29, Steven Rostedt wrote:
> > On Thu, 14 Sep 2023 13:17:00 -0400
> > Steven Rostedt <rostedt@goodmis.org> wrote:
> > 
> >> Now lets look at big endian layout:
> >>
> >>  uaddr = 0xbeef0004
> >>  enabler = 1;
> >>
> >>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
> >>                                     ^
> >>                                     addr: 0xbeef0004
> >>
> >> 				(enabler is set )
> >>
> >> 	bitoffset = uaddr & (sizeof(unsigned long) - 1); bitoffset = 4
> >> 	bit_offset *= 8;				 bitoffset = 32
> >> 	uaddr &= ~(sizeof(unsigned long) - 1);		 uaddr = 0xbeef0000
> >>
> >> 	ptr = kaddr + (uaddr & ~PAGE_MASK);
> >>
> >> 	clear_bit(1 + 32, ptr);
> >>
> >>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
> >>                                   ^
> >> 				bit 33 of 0xbeef0000
> >>
> >> I don't think that's what you expected!
> > 
> > I believe the above can be fixed with:
> > 
> > 	bit_offset = uaddr & (sizeof(unsigned long) - 1);
> > 	if (bit_offset) {
> > #ifdef CONFIG_CPU_BIG_ENDIAN
> > 		bit_offest = 0;
> > #else
> > 		bit_offset *= BITS_PER_BYTE;
> > #endif
> > 		uaddr &= ~(sizeof(unsigned long) - 1);
> > 	}
> > 
> > -- Steve
> 
> 
> Actually, after looking more in depth at that, it seems like there are
> actually 2 problems that can happen.
> 
> First one is atomic access misalignment due to enable_size == 4 and
> uaddr not being aligned on a (long) boundary on 64 bits architecture.
> This can generate misaligned exceptions on various architectures. This
> can be fixed in a more general way according to Masami snippet.
> 
> Second one that I can see is on 64 bits, big endian architectures with
> enable_size == 4. In that case, the bit provided by the userspace won't
> be correctly set since this code kind of assume that the atomic are done
> on 32bits value. Since the kernel assume long sized atomic operation, on
> big endian 64 bits architecture, the updated bit will actually be in the
> next 32 bits word.
> 
> Can someone confirm my understanding ?
> 

Actually, I take back some of what I said [1]. If a 32-bit on a 64-bit
kernel comes in on BE, and is aligned, we do need to offset the bits as
well (just verified on my ppc64 BE VM).

You should be able to use that patch as a base though and add a flag to
struct user_event_enabler when this case occurs. Then in the
align_addr_bit() adjust the bits as well upon aligned cases.

Thanks,
-Beau

1. https://lore.kernel.org/linux-trace-kernel/20230922192231.GA1828-beaub@linux.microsoft.com/

> Clément
Clément Léger Sept. 25, 2023, 7:53 a.m. UTC | #13
On 22/09/2023 21:22, Beau Belgrave wrote:
> On Tue, Sep 19, 2023 at 02:59:12PM +0200, Clément Léger wrote:
>>
>>
>> On 14/09/2023 19:29, Steven Rostedt wrote:
>>> On Thu, 14 Sep 2023 13:17:00 -0400
>>> Steven Rostedt <rostedt@goodmis.org> wrote:
>>>
>>>> Now lets look at big endian layout:
>>>>
>>>>  uaddr = 0xbeef0004
>>>>  enabler = 1;
>>>>
>>>>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
>>>>                                     ^
>>>>                                     addr: 0xbeef0004
>>>>
>>>> 				(enabler is set )
>>>>
>>>> 	bitoffset = uaddr & (sizeof(unsigned long) - 1); bitoffset = 4
>>>> 	bit_offset *= 8;				 bitoffset = 32
>>>> 	uaddr &= ~(sizeof(unsigned long) - 1);		 uaddr = 0xbeef0000
>>>>
>>>> 	ptr = kaddr + (uaddr & ~PAGE_MASK);
>>>>
>>>> 	clear_bit(1 + 32, ptr);
>>>>
>>>>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
>>>>                                   ^
>>>> 				bit 33 of 0xbeef0000
>>>>
>>>> I don't think that's what you expected!
>>>
>>> I believe the above can be fixed with:
>>>
>>> 	bit_offset = uaddr & (sizeof(unsigned long) - 1);
>>> 	if (bit_offset) {
>>> #ifdef CONFIG_CPU_BIG_ENDIAN
>>> 		bit_offest = 0;
>>> #else
>>> 		bit_offset *= BITS_PER_BYTE;
>>> #endif
>>> 		uaddr &= ~(sizeof(unsigned long) - 1);
>>> 	}
>>>
>>> -- Steve
>>
>>
>> Actually, after looking more in depth at that, it seems like there are
>> actually 2 problems that can happen.
>>
>> First one is atomic access misalignment due to enable_size == 4 and
>> uaddr not being aligned on a (long) boundary on 64 bits architecture.
>> This can generate misaligned exceptions on various architectures. This
>> can be fixed in a more general way according to Masami snippet.
>>
>> Second one that I can see is on 64 bits, big endian architectures with
>> enable_size == 4. In that case, the bit provided by the userspace won't
>> be correctly set since this code kind of assume that the atomic are done
>> on 32bits value. Since the kernel assume long sized atomic operation, on
>> big endian 64 bits architecture, the updated bit will actually be in the
>> next 32 bits word.
>>
>> Can someone confirm my understanding ?
>>
> 
> I have a ppc 64bit BE VM I've been validating this on. If we do the
> shifting within user_events (vs a generic set_bit_aligned approach)
> 64bit BE does not need additional bit manipulation. However, if we were
> to blindly pass the address and bit as is to set_bit_aligned() it
> assumes the bit number is for a long, not a 32 bit word. So for that
> approach we would need to offset the bit in the unaligned case.
> 
> Here's a patch I have that I've validated on ppc64 BE, aarch64 LE, and
> x86_64 LE. I personally feel more comfortable with this approach than
> the generic set_bit_aligned() one.
> 
> Thanks,
> -Beau
> 
> diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c
> index e3f2b8d72e01..ae854374d0b7 100644
> --- a/kernel/trace/trace_events_user.c
> +++ b/kernel/trace/trace_events_user.c
> @@ -162,6 +162,23 @@ struct user_event_validator {
>  	int			flags;
>  };
>  
> +static inline void align_addr_bit(unsigned long *addr, int *bit)
> +{
> +	if (IS_ALIGNED(*addr, sizeof(long)))
> +		return;
> +
> +	*addr = ALIGN_DOWN(*addr, sizeof(long));
> +
> +	/*
> +	 * We only support 32 and 64 bit values. The only time we need
> +	 * to align is a 32 bit value on a 64 bit kernel, which on LE
> +	 * is always 32 bits, and on BE requires no change.
> +	 */
> +#ifdef __LITTLE_ENDIAN
> +	*bit += 32;
> +#endif

Hi Beau, except the specific alignment that is basically what I ended up
with for the BE 64bit case (ie just bit += 32). Regarding the generic
alignment, depends on what the maintainers wishes (generic of user_event
specific). I also feel like this shoulmd be handle specifically for
user_events which uses set_bit in some non standard way. Any suggestion ?

Thanks,

Clément

> +}
> +
>  typedef void (*user_event_func_t) (struct user_event *user, struct iov_iter *i,
>  				   void *tpdata, bool *faulted);
>  
> @@ -481,6 +498,7 @@ static int user_event_enabler_write(struct user_event_mm *mm,
>  	unsigned long *ptr;
>  	struct page *page;
>  	void *kaddr;
> +	int bit = ENABLE_BIT(enabler);
>  	int ret;
>  
>  	lockdep_assert_held(&event_mutex);
> @@ -496,6 +514,8 @@ static int user_event_enabler_write(struct user_event_mm *mm,
>  		     test_bit(ENABLE_VAL_FREEING_BIT, ENABLE_BITOPS(enabler))))
>  		return -EBUSY;
>  
> +	align_addr_bit(&uaddr, &bit);
> +
>  	ret = pin_user_pages_remote(mm->mm, uaddr, 1, FOLL_WRITE | FOLL_NOFAULT,
>  				    &page, NULL);
>  
> @@ -514,9 +534,9 @@ static int user_event_enabler_write(struct user_event_mm *mm,
>  
>  	/* Update bit atomically, user tracers must be atomic as well */
>  	if (enabler->event && enabler->event->status)
> -		set_bit(ENABLE_BIT(enabler), ptr);
> +		set_bit(bit, ptr);
>  	else
> -		clear_bit(ENABLE_BIT(enabler), ptr);
> +		clear_bit(bit, ptr);
>  
>  	kunmap_local(kaddr);
>  	unpin_user_pages_dirty_lock(&page, 1, true);
Clément Léger Sept. 25, 2023, 8:10 a.m. UTC | #14
On 22/09/2023 22:00, Beau Belgrave wrote:
> On Tue, Sep 19, 2023 at 02:59:12PM +0200, Clément Léger wrote:
>>
>>
>> On 14/09/2023 19:29, Steven Rostedt wrote:
>>> On Thu, 14 Sep 2023 13:17:00 -0400
>>> Steven Rostedt <rostedt@goodmis.org> wrote:
>>>
>>>> Now lets look at big endian layout:
>>>>
>>>>  uaddr = 0xbeef0004
>>>>  enabler = 1;
>>>>
>>>>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
>>>>                                     ^
>>>>                                     addr: 0xbeef0004
>>>>
>>>> 				(enabler is set )
>>>>
>>>> 	bitoffset = uaddr & (sizeof(unsigned long) - 1); bitoffset = 4
>>>> 	bit_offset *= 8;				 bitoffset = 32
>>>> 	uaddr &= ~(sizeof(unsigned long) - 1);		 uaddr = 0xbeef0000
>>>>
>>>> 	ptr = kaddr + (uaddr & ~PAGE_MASK);
>>>>
>>>> 	clear_bit(1 + 32, ptr);
>>>>
>>>>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
>>>>                                   ^
>>>> 				bit 33 of 0xbeef0000
>>>>
>>>> I don't think that's what you expected!
>>>
>>> I believe the above can be fixed with:
>>>
>>> 	bit_offset = uaddr & (sizeof(unsigned long) - 1);
>>> 	if (bit_offset) {
>>> #ifdef CONFIG_CPU_BIG_ENDIAN
>>> 		bit_offest = 0;
>>> #else
>>> 		bit_offset *= BITS_PER_BYTE;
>>> #endif
>>> 		uaddr &= ~(sizeof(unsigned long) - 1);
>>> 	}
>>>
>>> -- Steve
>>
>>
>> Actually, after looking more in depth at that, it seems like there are
>> actually 2 problems that can happen.
>>
>> First one is atomic access misalignment due to enable_size == 4 and
>> uaddr not being aligned on a (long) boundary on 64 bits architecture.
>> This can generate misaligned exceptions on various architectures. This
>> can be fixed in a more general way according to Masami snippet.
>>
>> Second one that I can see is on 64 bits, big endian architectures with
>> enable_size == 4. In that case, the bit provided by the userspace won't
>> be correctly set since this code kind of assume that the atomic are done
>> on 32bits value. Since the kernel assume long sized atomic operation, on
>> big endian 64 bits architecture, the updated bit will actually be in the
>> next 32 bits word.
>>
>> Can someone confirm my understanding ?
>>
> 
> Actually, I take back some of what I said [1]. If a 32-bit on a 64-bit
> kernel comes in on BE, and is aligned, we do need to offset the bits as
> well (just verified on my ppc64 BE VM).

Yes, that is what I meant in my previous comment. I'll resend my series
which handles that properly (and which includes generic
set_bit_unaligned()).

Thanks,

Clément

> 
> You should be able to use that patch as a base though and add a flag to
> struct user_event_enabler when this case occurs. Then in the
> align_addr_bit() adjust the bits as well upon aligned cases.
> 
> Thanks,
> -Beau
> 
> 1. https://lore.kernel.org/linux-trace-kernel/20230922192231.GA1828-beaub@linux.microsoft.com/
> 
>> Clément
Beau Belgrave Sept. 25, 2023, 4:04 p.m. UTC | #15
On Mon, Sep 25, 2023 at 09:53:16AM +0200, Clément Léger wrote:
> 
> 
> On 22/09/2023 21:22, Beau Belgrave wrote:
> > On Tue, Sep 19, 2023 at 02:59:12PM +0200, Clément Léger wrote:
> >>
> >>
> >> On 14/09/2023 19:29, Steven Rostedt wrote:
> >>> On Thu, 14 Sep 2023 13:17:00 -0400
> >>> Steven Rostedt <rostedt@goodmis.org> wrote:
> >>>
> >>>> Now lets look at big endian layout:
> >>>>
> >>>>  uaddr = 0xbeef0004
> >>>>  enabler = 1;
> >>>>
> >>>>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
> >>>>                                     ^
> >>>>                                     addr: 0xbeef0004
> >>>>
> >>>> 				(enabler is set )
> >>>>
> >>>> 	bitoffset = uaddr & (sizeof(unsigned long) - 1); bitoffset = 4
> >>>> 	bit_offset *= 8;				 bitoffset = 32
> >>>> 	uaddr &= ~(sizeof(unsigned long) - 1);		 uaddr = 0xbeef0000
> >>>>
> >>>> 	ptr = kaddr + (uaddr & ~PAGE_MASK);
> >>>>
> >>>> 	clear_bit(1 + 32, ptr);
> >>>>
> >>>>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
> >>>>                                   ^
> >>>> 				bit 33 of 0xbeef0000
> >>>>
> >>>> I don't think that's what you expected!
> >>>
> >>> I believe the above can be fixed with:
> >>>
> >>> 	bit_offset = uaddr & (sizeof(unsigned long) - 1);
> >>> 	if (bit_offset) {
> >>> #ifdef CONFIG_CPU_BIG_ENDIAN
> >>> 		bit_offest = 0;
> >>> #else
> >>> 		bit_offset *= BITS_PER_BYTE;
> >>> #endif
> >>> 		uaddr &= ~(sizeof(unsigned long) - 1);
> >>> 	}
> >>>
> >>> -- Steve
> >>
> >>
> >> Actually, after looking more in depth at that, it seems like there are
> >> actually 2 problems that can happen.
> >>
> >> First one is atomic access misalignment due to enable_size == 4 and
> >> uaddr not being aligned on a (long) boundary on 64 bits architecture.
> >> This can generate misaligned exceptions on various architectures. This
> >> can be fixed in a more general way according to Masami snippet.
> >>
> >> Second one that I can see is on 64 bits, big endian architectures with
> >> enable_size == 4. In that case, the bit provided by the userspace won't
> >> be correctly set since this code kind of assume that the atomic are done
> >> on 32bits value. Since the kernel assume long sized atomic operation, on
> >> big endian 64 bits architecture, the updated bit will actually be in the
> >> next 32 bits word.
> >>
> >> Can someone confirm my understanding ?
> >>
> > 
> > I have a ppc 64bit BE VM I've been validating this on. If we do the
> > shifting within user_events (vs a generic set_bit_aligned approach)
> > 64bit BE does not need additional bit manipulation. However, if we were
> > to blindly pass the address and bit as is to set_bit_aligned() it
> > assumes the bit number is for a long, not a 32 bit word. So for that
> > approach we would need to offset the bit in the unaligned case.
> > 
> > Here's a patch I have that I've validated on ppc64 BE, aarch64 LE, and
> > x86_64 LE. I personally feel more comfortable with this approach than
> > the generic set_bit_aligned() one.
> > 
> > Thanks,
> > -Beau
> > 
> > diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c
> > index e3f2b8d72e01..ae854374d0b7 100644
> > --- a/kernel/trace/trace_events_user.c
> > +++ b/kernel/trace/trace_events_user.c
> > @@ -162,6 +162,23 @@ struct user_event_validator {
> >  	int			flags;
> >  };
> >  
> > +static inline void align_addr_bit(unsigned long *addr, int *bit)
> > +{
> > +	if (IS_ALIGNED(*addr, sizeof(long)))
> > +		return;
> > +
> > +	*addr = ALIGN_DOWN(*addr, sizeof(long));
> > +
> > +	/*
> > +	 * We only support 32 and 64 bit values. The only time we need
> > +	 * to align is a 32 bit value on a 64 bit kernel, which on LE
> > +	 * is always 32 bits, and on BE requires no change.
> > +	 */
> > +#ifdef __LITTLE_ENDIAN
> > +	*bit += 32;
> > +#endif
> 
> Hi Beau, except the specific alignment that is basically what I ended up
> with for the BE 64bit case (ie just bit += 32). Regarding the generic
> alignment, depends on what the maintainers wishes (generic of user_event
> specific). I also feel like this shoulmd be handle specifically for
> user_events which uses set_bit in some non standard way. Any suggestion ?
> 

Looking at this deeper, user_events needs to track some of this
alignment requirements within each enabler. This is needed because when
enablers are disabled/removed they do not have the actual size. The size
is needed to know if we need to update the bits, etc. (IE: If originally
a 32-bit value was used and it's aligned and it's on BE, it needs the
bits shifted.)

My final version that I played around with looked like this for just the
alignment requirements:
+static inline void align_addr_bit(unsigned long *addr, int *bit,
+                                 unsigned long *flags)
+{
+       if (IS_ALIGNED(*addr, sizeof(long))) {
+#ifdef __BIG_ENDIAN
+               if (test_bit(ENABLE_VAL_32_BIT, flags))
+                       *bit += 32;
+#endif
+               return;
+       }
+
+       *addr = ALIGN_DOWN(*addr, sizeof(long));
+
+       /*
+        * We only support 32 and 64 bit values. The only time we need
+        * to align is a 32 bit value on a 64 bit kernel, which on LE
+        * is always 32 bits, and on BE requires no change.
+        */
+#ifdef __LITTLE_ENDIAN
+       *bit += 32;
+#endif
+}

Below is the full patch, which is currently appearing to pass everything
on BE (Please note, the abi_test in selftests needs fixes that I plan to
push out. I'd like to get Steven to take those changes along with yours
together from tracing to ensure we can test BE with these changes
properly).

As you'll see, it requires a bit more work than just a generic unaligned
solution due to the bits having to move for 32-bit on BE and the
tracking requirement on when to do so during delete/clear.

Thanks,
-Beau

diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c
index 6f046650e527..b05db15eaea6 100644
--- a/kernel/trace/trace_events_user.c
+++ b/kernel/trace/trace_events_user.c
@@ -127,8 +127,11 @@ struct user_event_enabler {
 /* Bit 7 is for freeing status of enablement */
 #define ENABLE_VAL_FREEING_BIT 7
 
+/* Bit 8 is for marking 32-bit on 64-bit */
+#define ENABLE_VAL_32_BIT 8
+
 /* Only duplicate the bit value */
-#define ENABLE_VAL_DUP_MASK ENABLE_VAL_BIT_MASK
+#define ENABLE_VAL_DUP_MASK (ENABLE_VAL_BIT_MASK | 1 << ENABLE_VAL_32_BIT)
 
 #define ENABLE_BITOPS(e) (&(e)->values)
 
@@ -174,6 +177,29 @@ struct user_event_validator {
 	int			flags;
 };
 
+static inline void align_addr_bit(unsigned long *addr, int *bit,
+				  unsigned long *flags)
+{
+	if (IS_ALIGNED(*addr, sizeof(long))) {
+#ifdef __BIG_ENDIAN
+		if (test_bit(ENABLE_VAL_32_BIT, flags))
+			*bit += 32;
+#endif
+		return;
+	}
+
+	*addr = ALIGN_DOWN(*addr, sizeof(long));
+
+	/*
+	 * We only support 32 and 64 bit values. The only time we need
+	 * to align is a 32 bit value on a 64 bit kernel, which on LE
+	 * is always 32 bits, and on BE requires no change.
+	 */
+#ifdef __LITTLE_ENDIAN
+	*bit += 32;
+#endif
+}
+
 typedef void (*user_event_func_t) (struct user_event *user, struct iov_iter *i,
 				   void *tpdata, bool *faulted);
 
@@ -482,6 +508,7 @@ static int user_event_enabler_write(struct user_event_mm *mm,
 	unsigned long *ptr;
 	struct page *page;
 	void *kaddr;
+	int bit = ENABLE_BIT(enabler);
 	int ret;
 
 	lockdep_assert_held(&event_mutex);
@@ -497,6 +524,8 @@ static int user_event_enabler_write(struct user_event_mm *mm,
 		     test_bit(ENABLE_VAL_FREEING_BIT, ENABLE_BITOPS(enabler))))
 		return -EBUSY;
 
+	align_addr_bit(&uaddr, &bit, ENABLE_BITOPS(enabler));
+
 	ret = pin_user_pages_remote(mm->mm, uaddr, 1, FOLL_WRITE | FOLL_NOFAULT,
 				    &page, NULL);
 
@@ -515,9 +544,9 @@ static int user_event_enabler_write(struct user_event_mm *mm,
 
 	/* Update bit atomically, user tracers must be atomic as well */
 	if (enabler->event && enabler->event->status)
-		set_bit(ENABLE_BIT(enabler), ptr);
+		set_bit(bit, ptr);
 	else
-		clear_bit(ENABLE_BIT(enabler), ptr);
+		clear_bit(bit, ptr);
 
 	kunmap_local(kaddr);
 	unpin_user_pages_dirty_lock(&page, 1, true);
@@ -849,6 +878,12 @@ static struct user_event_enabler
 	enabler->event = user;
 	enabler->addr = uaddr;
 	enabler->values = reg->enable_bit;
+
+#if BITS_PER_LONG >= 64
+	if (reg->enable_size == 4)
+		set_bit(ENABLE_VAL_32_BIT, ENABLE_BITOPS(enabler));
+#endif
+
 retry:
 	/* Prevents state changes from racing with new enablers */
 	mutex_lock(&event_mutex);
@@ -2377,7 +2412,8 @@ static long user_unreg_get(struct user_unreg __user *ureg,
 }
 
 static int user_event_mm_clear_bit(struct user_event_mm *user_mm,
-				   unsigned long uaddr, unsigned char bit)
+				   unsigned long uaddr, unsigned char bit,
+				   unsigned long flags)
 {
 	struct user_event_enabler enabler;
 	int result;
@@ -2385,7 +2421,7 @@ static int user_event_mm_clear_bit(struct user_event_mm *user_mm,
 
 	memset(&enabler, 0, sizeof(enabler));
 	enabler.addr = uaddr;
-	enabler.values = bit;
+	enabler.values = bit | flags;
 retry:
 	/* Prevents state changes from racing with new enablers */
 	mutex_lock(&event_mutex);
@@ -2415,6 +2451,7 @@ static long user_events_ioctl_unreg(unsigned long uarg)
 	struct user_event_mm *mm = current->user_event_mm;
 	struct user_event_enabler *enabler, *next;
 	struct user_unreg reg;
+	unsigned long flags;
 	long ret;
 
 	ret = user_unreg_get(ureg, &reg);
@@ -2425,6 +2462,7 @@ static long user_events_ioctl_unreg(unsigned long uarg)
 	if (!mm)
 		return -ENOENT;
 
+	flags = 0;
 	ret = -ENOENT;
 
 	/*
@@ -2441,6 +2479,10 @@ static long user_events_ioctl_unreg(unsigned long uarg)
 		    ENABLE_BIT(enabler) == reg.disable_bit) {
 			set_bit(ENABLE_VAL_FREEING_BIT, ENABLE_BITOPS(enabler));
 
+			/* We must keep compat flags for the clear */
+			if (test_bit(ENABLE_VAL_32_BIT, ENABLE_BITOPS(enabler)))
+				flags |= 1 << ENABLE_VAL_32_BIT;
+
 			if (!test_bit(ENABLE_VAL_FAULTING_BIT, ENABLE_BITOPS(enabler)))
 				user_event_enabler_destroy(enabler, true);
 
@@ -2454,7 +2496,7 @@ static long user_events_ioctl_unreg(unsigned long uarg)
 	/* Ensure bit is now cleared for user, regardless of event status */
 	if (!ret)
 		ret = user_event_mm_clear_bit(mm, reg.disable_addr,
-					      reg.disable_bit);
+					      reg.disable_bit, flags);
 
 	return ret;
 }
Clément Léger Sept. 25, 2023, 6:04 p.m. UTC | #16
On 25/09/2023 18:04, Beau Belgrave wrote:
> On Mon, Sep 25, 2023 at 09:53:16AM +0200, Clément Léger wrote:
>>
>>
>> On 22/09/2023 21:22, Beau Belgrave wrote:
>>> On Tue, Sep 19, 2023 at 02:59:12PM +0200, Clément Léger wrote:
>>>>
>>>>
>>>> On 14/09/2023 19:29, Steven Rostedt wrote:
>>>>> On Thu, 14 Sep 2023 13:17:00 -0400
>>>>> Steven Rostedt <rostedt@goodmis.org> wrote:
>>>>>
>>>>>> Now lets look at big endian layout:
>>>>>>
>>>>>>  uaddr = 0xbeef0004
>>>>>>  enabler = 1;
>>>>>>
>>>>>>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
>>>>>>                                     ^
>>>>>>                                     addr: 0xbeef0004
>>>>>>
>>>>>> 				(enabler is set )
>>>>>>
>>>>>> 	bitoffset = uaddr & (sizeof(unsigned long) - 1); bitoffset = 4
>>>>>> 	bit_offset *= 8;				 bitoffset = 32
>>>>>> 	uaddr &= ~(sizeof(unsigned long) - 1);		 uaddr = 0xbeef0000
>>>>>>
>>>>>> 	ptr = kaddr + (uaddr & ~PAGE_MASK);
>>>>>>
>>>>>> 	clear_bit(1 + 32, ptr);
>>>>>>
>>>>>>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
>>>>>>                                   ^
>>>>>> 				bit 33 of 0xbeef0000
>>>>>>
>>>>>> I don't think that's what you expected!
>>>>>
>>>>> I believe the above can be fixed with:
>>>>>
>>>>> 	bit_offset = uaddr & (sizeof(unsigned long) - 1);
>>>>> 	if (bit_offset) {
>>>>> #ifdef CONFIG_CPU_BIG_ENDIAN
>>>>> 		bit_offest = 0;
>>>>> #else
>>>>> 		bit_offset *= BITS_PER_BYTE;
>>>>> #endif
>>>>> 		uaddr &= ~(sizeof(unsigned long) - 1);
>>>>> 	}
>>>>>
>>>>> -- Steve
>>>>
>>>>
>>>> Actually, after looking more in depth at that, it seems like there are
>>>> actually 2 problems that can happen.
>>>>
>>>> First one is atomic access misalignment due to enable_size == 4 and
>>>> uaddr not being aligned on a (long) boundary on 64 bits architecture.
>>>> This can generate misaligned exceptions on various architectures. This
>>>> can be fixed in a more general way according to Masami snippet.
>>>>
>>>> Second one that I can see is on 64 bits, big endian architectures with
>>>> enable_size == 4. In that case, the bit provided by the userspace won't
>>>> be correctly set since this code kind of assume that the atomic are done
>>>> on 32bits value. Since the kernel assume long sized atomic operation, on
>>>> big endian 64 bits architecture, the updated bit will actually be in the
>>>> next 32 bits word.
>>>>
>>>> Can someone confirm my understanding ?
>>>>
>>>
>>> I have a ppc 64bit BE VM I've been validating this on. If we do the
>>> shifting within user_events (vs a generic set_bit_aligned approach)
>>> 64bit BE does not need additional bit manipulation. However, if we were
>>> to blindly pass the address and bit as is to set_bit_aligned() it
>>> assumes the bit number is for a long, not a 32 bit word. So for that
>>> approach we would need to offset the bit in the unaligned case.
>>>
>>> Here's a patch I have that I've validated on ppc64 BE, aarch64 LE, and
>>> x86_64 LE. I personally feel more comfortable with this approach than
>>> the generic set_bit_aligned() one.
>>>
>>> Thanks,
>>> -Beau
>>>
>>> diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c
>>> index e3f2b8d72e01..ae854374d0b7 100644
>>> --- a/kernel/trace/trace_events_user.c
>>> +++ b/kernel/trace/trace_events_user.c
>>> @@ -162,6 +162,23 @@ struct user_event_validator {
>>>  	int			flags;
>>>  };
>>>  
>>> +static inline void align_addr_bit(unsigned long *addr, int *bit)
>>> +{
>>> +	if (IS_ALIGNED(*addr, sizeof(long)))
>>> +		return;
>>> +
>>> +	*addr = ALIGN_DOWN(*addr, sizeof(long));
>>> +
>>> +	/*
>>> +	 * We only support 32 and 64 bit values. The only time we need
>>> +	 * to align is a 32 bit value on a 64 bit kernel, which on LE
>>> +	 * is always 32 bits, and on BE requires no change.
>>> +	 */
>>> +#ifdef __LITTLE_ENDIAN
>>> +	*bit += 32;
>>> +#endif
>>
>> Hi Beau, except the specific alignment that is basically what I ended up
>> with for the BE 64bit case (ie just bit += 32). Regarding the generic
>> alignment, depends on what the maintainers wishes (generic of user_event
>> specific). I also feel like this shoulmd be handle specifically for
>> user_events which uses set_bit in some non standard way. Any suggestion ?
>>
> 
> Looking at this deeper, user_events needs to track some of this
> alignment requirements within each enabler. This is needed because when
> enablers are disabled/removed they do not have the actual size. The size
> is needed to know if we need to update the bits, etc. (IE: If originally
> a 32-bit value was used and it's aligned and it's on BE, it needs the
> bits shifted.)
> 
> My final version that I played around with looked like this for just the
> alignment requirements:
> +static inline void align_addr_bit(unsigned long *addr, int *bit,
> +                                 unsigned long *flags)
> +{
> +       if (IS_ALIGNED(*addr, sizeof(long))) {
> +#ifdef __BIG_ENDIAN
> +               if (test_bit(ENABLE_VAL_32_BIT, flags))
> +                       *bit += 32;
> +#endif
> +               return;
> +       }
> +
> +       *addr = ALIGN_DOWN(*addr, sizeof(long));
> +
> +       /*
> +        * We only support 32 and 64 bit values. The only time we need
> +        * to align is a 32 bit value on a 64 bit kernel, which on LE
> +        * is always 32 bits, and on BE requires no change.
> +        */
> +#ifdef __LITTLE_ENDIAN
> +       *bit += 32;
> +#endif
> +}
> 
> Below is the full patch, which is currently appearing to pass everything
> on BE (Please note, the abi_test in selftests needs fixes that I plan to
> push out. I'd like to get Steven to take those changes along with yours
> together from tracing to ensure we can test BE with these changes
> properly).
> 
> As you'll see, it requires a bit more work than just a generic unaligned
> solution due to the bits having to move for 32-bit on BE and the
> tracking requirement on when to do so during delete/clear.

I actually had the same kind of handling (using a size field rather than
a flag though). However, the generic set_bit/clear bit requires a
different handling of the 32 bits on BE 64 bits which does not sounds
quite right, ie unconditionally add 32 bits to fixup the offset which is
 implicit in case you are not aligned with your code (ie shifting the
address actually give the correct bit on BE 64 bits).

Since you already wrote the code, I think we can proceed with your
version as I actually thinks its more clearer to understand.

Thanks,

Clément

> 
> Thanks,
> -Beau
> 
> diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c
> index 6f046650e527..b05db15eaea6 100644
> --- a/kernel/trace/trace_events_user.c
> +++ b/kernel/trace/trace_events_user.c
> @@ -127,8 +127,11 @@ struct user_event_enabler {
>  /* Bit 7 is for freeing status of enablement */
>  #define ENABLE_VAL_FREEING_BIT 7
>  
> +/* Bit 8 is for marking 32-bit on 64-bit */
> +#define ENABLE_VAL_32_BIT 8
> +
>  /* Only duplicate the bit value */
> -#define ENABLE_VAL_DUP_MASK ENABLE_VAL_BIT_MASK
> +#define ENABLE_VAL_DUP_MASK (ENABLE_VAL_BIT_MASK | 1 << ENABLE_VAL_32_BIT)
>  
>  #define ENABLE_BITOPS(e) (&(e)->values)
>  
> @@ -174,6 +177,29 @@ struct user_event_validator {
>  	int			flags;
>  };
>  
> +static inline void align_addr_bit(unsigned long *addr, int *bit,
> +				  unsigned long *flags)
> +{
> +	if (IS_ALIGNED(*addr, sizeof(long))) {
> +#ifdef __BIG_ENDIAN
> +		if (test_bit(ENABLE_VAL_32_BIT, flags))
> +			*bit += 32;
> +#endif
> +		return;
> +	}
> +
> +	*addr = ALIGN_DOWN(*addr, sizeof(long));
> +
> +	/*
> +	 * We only support 32 and 64 bit values. The only time we need
> +	 * to align is a 32 bit value on a 64 bit kernel, which on LE
> +	 * is always 32 bits, and on BE requires no change.
> +	 */
> +#ifdef __LITTLE_ENDIAN
> +	*bit += 32;
> +#endif
> +}
> +
>  typedef void (*user_event_func_t) (struct user_event *user, struct iov_iter *i,
>  				   void *tpdata, bool *faulted);
>  
> @@ -482,6 +508,7 @@ static int user_event_enabler_write(struct user_event_mm *mm,
>  	unsigned long *ptr;
>  	struct page *page;
>  	void *kaddr;
> +	int bit = ENABLE_BIT(enabler);
>  	int ret;
>  
>  	lockdep_assert_held(&event_mutex);
> @@ -497,6 +524,8 @@ static int user_event_enabler_write(struct user_event_mm *mm,
>  		     test_bit(ENABLE_VAL_FREEING_BIT, ENABLE_BITOPS(enabler))))
>  		return -EBUSY;
>  
> +	align_addr_bit(&uaddr, &bit, ENABLE_BITOPS(enabler));
> +
>  	ret = pin_user_pages_remote(mm->mm, uaddr, 1, FOLL_WRITE | FOLL_NOFAULT,
>  				    &page, NULL);
>  
> @@ -515,9 +544,9 @@ static int user_event_enabler_write(struct user_event_mm *mm,
>  
>  	/* Update bit atomically, user tracers must be atomic as well */
>  	if (enabler->event && enabler->event->status)
> -		set_bit(ENABLE_BIT(enabler), ptr);
> +		set_bit(bit, ptr);
>  	else
> -		clear_bit(ENABLE_BIT(enabler), ptr);
> +		clear_bit(bit, ptr);
>  
>  	kunmap_local(kaddr);
>  	unpin_user_pages_dirty_lock(&page, 1, true);
> @@ -849,6 +878,12 @@ static struct user_event_enabler
>  	enabler->event = user;
>  	enabler->addr = uaddr;
>  	enabler->values = reg->enable_bit;
> +
> +#if BITS_PER_LONG >= 64
> +	if (reg->enable_size == 4)
> +		set_bit(ENABLE_VAL_32_BIT, ENABLE_BITOPS(enabler));
> +#endif
> +
>  retry:
>  	/* Prevents state changes from racing with new enablers */
>  	mutex_lock(&event_mutex);
> @@ -2377,7 +2412,8 @@ static long user_unreg_get(struct user_unreg __user *ureg,
>  }
>  
>  static int user_event_mm_clear_bit(struct user_event_mm *user_mm,
> -				   unsigned long uaddr, unsigned char bit)
> +				   unsigned long uaddr, unsigned char bit,
> +				   unsigned long flags)
>  {
>  	struct user_event_enabler enabler;
>  	int result;
> @@ -2385,7 +2421,7 @@ static int user_event_mm_clear_bit(struct user_event_mm *user_mm,
>  
>  	memset(&enabler, 0, sizeof(enabler));
>  	enabler.addr = uaddr;
> -	enabler.values = bit;
> +	enabler.values = bit | flags;
>  retry:
>  	/* Prevents state changes from racing with new enablers */
>  	mutex_lock(&event_mutex);
> @@ -2415,6 +2451,7 @@ static long user_events_ioctl_unreg(unsigned long uarg)
>  	struct user_event_mm *mm = current->user_event_mm;
>  	struct user_event_enabler *enabler, *next;
>  	struct user_unreg reg;
> +	unsigned long flags;
>  	long ret;
>  
>  	ret = user_unreg_get(ureg, &reg);
> @@ -2425,6 +2462,7 @@ static long user_events_ioctl_unreg(unsigned long uarg)
>  	if (!mm)
>  		return -ENOENT;
>  
> +	flags = 0;
>  	ret = -ENOENT;
>  
>  	/*
> @@ -2441,6 +2479,10 @@ static long user_events_ioctl_unreg(unsigned long uarg)
>  		    ENABLE_BIT(enabler) == reg.disable_bit) {
>  			set_bit(ENABLE_VAL_FREEING_BIT, ENABLE_BITOPS(enabler));
>  
> +			/* We must keep compat flags for the clear */
> +			if (test_bit(ENABLE_VAL_32_BIT, ENABLE_BITOPS(enabler)))
> +				flags |= 1 << ENABLE_VAL_32_BIT;
> +
>  			if (!test_bit(ENABLE_VAL_FAULTING_BIT, ENABLE_BITOPS(enabler)))
>  				user_event_enabler_destroy(enabler, true);
>  
> @@ -2454,7 +2496,7 @@ static long user_events_ioctl_unreg(unsigned long uarg)
>  	/* Ensure bit is now cleared for user, regardless of event status */
>  	if (!ret)
>  		ret = user_event_mm_clear_bit(mm, reg.disable_addr,
> -					      reg.disable_bit);
> +					      reg.disable_bit, flags);
>  
>  	return ret;
>  }
Clément Léger Sept. 25, 2023, 6:22 p.m. UTC | #17
On 25/09/2023 20:04, Clément Léger wrote:
> 
> 
> On 25/09/2023 18:04, Beau Belgrave wrote:
>> On Mon, Sep 25, 2023 at 09:53:16AM +0200, Clément Léger wrote:
>>>
>>>
>>> On 22/09/2023 21:22, Beau Belgrave wrote:
>>>> On Tue, Sep 19, 2023 at 02:59:12PM +0200, Clément Léger wrote:
>>>>>
>>>>>
>>>>> On 14/09/2023 19:29, Steven Rostedt wrote:
>>>>>> On Thu, 14 Sep 2023 13:17:00 -0400
>>>>>> Steven Rostedt <rostedt@goodmis.org> wrote:
>>>>>>
>>>>>>> Now lets look at big endian layout:
>>>>>>>
>>>>>>>  uaddr = 0xbeef0004
>>>>>>>  enabler = 1;
>>>>>>>
>>>>>>>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
>>>>>>>                                     ^
>>>>>>>                                     addr: 0xbeef0004
>>>>>>>
>>>>>>> 				(enabler is set )
>>>>>>>
>>>>>>> 	bitoffset = uaddr & (sizeof(unsigned long) - 1); bitoffset = 4
>>>>>>> 	bit_offset *= 8;				 bitoffset = 32
>>>>>>> 	uaddr &= ~(sizeof(unsigned long) - 1);		 uaddr = 0xbeef0000
>>>>>>>
>>>>>>> 	ptr = kaddr + (uaddr & ~PAGE_MASK);
>>>>>>>
>>>>>>> 	clear_bit(1 + 32, ptr);
>>>>>>>
>>>>>>>  memory at 0xbeef0000:  00 00 00 00 00 00 00 02
>>>>>>>                                   ^
>>>>>>> 				bit 33 of 0xbeef0000
>>>>>>>
>>>>>>> I don't think that's what you expected!
>>>>>>
>>>>>> I believe the above can be fixed with:
>>>>>>
>>>>>> 	bit_offset = uaddr & (sizeof(unsigned long) - 1);
>>>>>> 	if (bit_offset) {
>>>>>> #ifdef CONFIG_CPU_BIG_ENDIAN
>>>>>> 		bit_offest = 0;
>>>>>> #else
>>>>>> 		bit_offset *= BITS_PER_BYTE;
>>>>>> #endif
>>>>>> 		uaddr &= ~(sizeof(unsigned long) - 1);
>>>>>> 	}
>>>>>>
>>>>>> -- Steve
>>>>>
>>>>>
>>>>> Actually, after looking more in depth at that, it seems like there are
>>>>> actually 2 problems that can happen.
>>>>>
>>>>> First one is atomic access misalignment due to enable_size == 4 and
>>>>> uaddr not being aligned on a (long) boundary on 64 bits architecture.
>>>>> This can generate misaligned exceptions on various architectures. This
>>>>> can be fixed in a more general way according to Masami snippet.
>>>>>
>>>>> Second one that I can see is on 64 bits, big endian architectures with
>>>>> enable_size == 4. In that case, the bit provided by the userspace won't
>>>>> be correctly set since this code kind of assume that the atomic are done
>>>>> on 32bits value. Since the kernel assume long sized atomic operation, on
>>>>> big endian 64 bits architecture, the updated bit will actually be in the
>>>>> next 32 bits word.
>>>>>
>>>>> Can someone confirm my understanding ?
>>>>>
>>>>
>>>> I have a ppc 64bit BE VM I've been validating this on. If we do the
>>>> shifting within user_events (vs a generic set_bit_aligned approach)
>>>> 64bit BE does not need additional bit manipulation. However, if we were
>>>> to blindly pass the address and bit as is to set_bit_aligned() it
>>>> assumes the bit number is for a long, not a 32 bit word. So for that
>>>> approach we would need to offset the bit in the unaligned case.
>>>>
>>>> Here's a patch I have that I've validated on ppc64 BE, aarch64 LE, and
>>>> x86_64 LE. I personally feel more comfortable with this approach than
>>>> the generic set_bit_aligned() one.
>>>>
>>>> Thanks,
>>>> -Beau
>>>>
>>>> diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c
>>>> index e3f2b8d72e01..ae854374d0b7 100644
>>>> --- a/kernel/trace/trace_events_user.c
>>>> +++ b/kernel/trace/trace_events_user.c
>>>> @@ -162,6 +162,23 @@ struct user_event_validator {
>>>>  	int			flags;
>>>>  };
>>>>  
>>>> +static inline void align_addr_bit(unsigned long *addr, int *bit)
>>>> +{
>>>> +	if (IS_ALIGNED(*addr, sizeof(long)))
>>>> +		return;
>>>> +
>>>> +	*addr = ALIGN_DOWN(*addr, sizeof(long));
>>>> +
>>>> +	/*
>>>> +	 * We only support 32 and 64 bit values. The only time we need
>>>> +	 * to align is a 32 bit value on a 64 bit kernel, which on LE
>>>> +	 * is always 32 bits, and on BE requires no change.
>>>> +	 */
>>>> +#ifdef __LITTLE_ENDIAN
>>>> +	*bit += 32;
>>>> +#endif
>>>
>>> Hi Beau, except the specific alignment that is basically what I ended up
>>> with for the BE 64bit case (ie just bit += 32). Regarding the generic
>>> alignment, depends on what the maintainers wishes (generic of user_event
>>> specific). I also feel like this shoulmd be handle specifically for
>>> user_events which uses set_bit in some non standard way. Any suggestion ?
>>>
>>
>> Looking at this deeper, user_events needs to track some of this
>> alignment requirements within each enabler. This is needed because when
>> enablers are disabled/removed they do not have the actual size. The size
>> is needed to know if we need to update the bits, etc. (IE: If originally
>> a 32-bit value was used and it's aligned and it's on BE, it needs the
>> bits shifted.)
>>
>> My final version that I played around with looked like this for just the
>> alignment requirements:
>> +static inline void align_addr_bit(unsigned long *addr, int *bit,
>> +                                 unsigned long *flags)
>> +{
>> +       if (IS_ALIGNED(*addr, sizeof(long))) {
>> +#ifdef __BIG_ENDIAN
>> +               if (test_bit(ENABLE_VAL_32_BIT, flags))
>> +                       *bit += 32;
>> +#endif
>> +               return;
>> +       }
>> +
>> +       *addr = ALIGN_DOWN(*addr, sizeof(long));
>> +
>> +       /*
>> +        * We only support 32 and 64 bit values. The only time we need
>> +        * to align is a 32 bit value on a 64 bit kernel, which on LE
>> +        * is always 32 bits, and on BE requires no change.
>> +        */
>> +#ifdef __LITTLE_ENDIAN
>> +       *bit += 32;
>> +#endif
>> +}
>>
>> Below is the full patch, which is currently appearing to pass everything
>> on BE (Please note, the abi_test in selftests needs fixes that I plan to
>> push out. I'd like to get Steven to take those changes along with yours
>> together from tracing to ensure we can test BE with these changes
>> properly).
>>
>> As you'll see, it requires a bit more work than just a generic unaligned
>> solution due to the bits having to move for 32-bit on BE and the
>> tracking requirement on when to do so during delete/clear.
> 
> I actually had the same kind of handling (using a size field rather than
> a flag though). However, the generic set_bit/clear bit requires a
> different handling of the 32 bits on BE 64 bits which does not sounds
> quite right, ie unconditionally add 32 bits to fixup the offset which is
>  implicit in case you are not aligned with your code (ie shifting the
> address actually give the correct bit on BE 64 bits).
> 
> Since you already wrote the code, I think we can proceed with your
> version as I actually thinks its more clearer to understand.

FWIW, here is my series, as I said, the generic set/clear bit makes it
less clear in the BE 64 bits case:


diff --git a/include/linux/bitops.h b/include/linux/bitops.h
index 2ba557e067fe..b247c24695b9 100644
--- a/include/linux/bitops.h
+++ b/include/linux/bitops.h
@@ -290,6 +290,65 @@ static __always_inline void __assign_bit(long nr,
volatile unsigned long *addr,
 		__clear_bit(nr, addr);
 }

+#define LONG_ALIGN_DIFF(p)	((unsigned long)(p)) & (sizeof(long) -1)
+#define LONG_ALIGNED_PTR(p) \
+		(unsigned long *)(((unsigned long)(p)) & ~(sizeof(long) - 1))
+
+static inline unsigned long *__fix_unaligned(int *bit, void *ptr)
+{
+	int offs = LONG_ALIGN_DIFF(ptr) * BITS_PER_BYTE;
+	unsigned long *uptr;
+
+	if (offs == 0)
+		return uptr;
+
+	uptr = LONG_ALIGNED_PTR(ptr);
+
+#ifdef __BIG_ENDIAN
+	if (*bit >= offs)
+		*bit -= offs;
+	else
+		*bit += BITS_PER_LONG + offs;
+#else
+	*bit += offs;
+#endif
+
+	return uptr;
+}
+
+/**
+ * set_bit_unaligned - Atomically set a bit in memory starting from an
unaligned
+ *		       pointer
+ * @nr: the bit to set
+ * @addr: the address to start counting from
+ *
+ * This is a relaxed atomic operation (no implied memory barriers).
+ *
+ * Note that @nr may be almost arbitrarily large; this function is not
+ * restricted to acting on a single-word quantity.
+ */
+static inline void set_bit_unaligned(int bit, void *ptr)
+{
+	unsigned long *uptr = __fix_unaligned(&bit, ptr);
+
+	set_bit(bit, uptr);
+}
+
+/**
+ * clear_bit_unaligned - Clears a bit in memory starting from an unaligned
+ *			 pointer
+ * @nr: Bit to clear
+ * @addr: Address to start counting from
+ *
+ * This is a relaxed atomic operation (no implied memory barriers).
+ */
+static inline void clear_bit_unaligned(int bit, unsigned long *ptr)
+{
+	unsigned long *uptr = __fix_unaligned(&bit, ptr);
+
+	clear_bit(bit, uptr);
+}
+
 /**
  * __ptr_set_bit - Set bit in a pointer's value
  * @nr: the bit to set
diff --git a/kernel/trace/trace_events_user.c
b/kernel/trace/trace_events_user.c
index 6f046650e527..641ab6dedb30 100644
--- a/kernel/trace/trace_events_user.c
+++ b/kernel/trace/trace_events_user.c
@@ -113,6 +113,7 @@ struct user_event_enabler {
 	struct list_head	mm_enablers_link;
 	struct user_event	*event;
 	unsigned long		addr;
+	u8			size;

 	/* Track enable bit, flags, etc. Aligned for bitops. */
 	unsigned long		values;
@@ -481,9 +482,15 @@ static int user_event_enabler_write(struct
user_event_mm *mm,
 	unsigned long uaddr = enabler->addr;
 	unsigned long *ptr;
 	struct page *page;
+	int bit = ENABLER_BIT(enabler);
 	void *kaddr;
 	int ret;

+#if defined(CONFIG_CPU_BIG_ENDIAN) && defined(CONFIG_64BIT)
+	if (enabler->size == 4)
+		bit += 32;
+#endif
+
 	lockdep_assert_held(&event_mutex);
 	mmap_assert_locked(mm->mm);

@@ -515,9 +522,9 @@ static int user_event_enabler_write(struct
user_event_mm *mm,

 	/* Update bit atomically, user tracers must be atomic as well */
 	if (enabler->event && enabler->event->status)
-		set_bit(ENABLE_BIT(enabler), ptr);
+		set_bit_unaligned(bit, ptr);
 	else
-		clear_bit(ENABLE_BIT(enabler), ptr);
+		clear_bit_unaligned(bit, ptr);

 	kunmap_local(kaddr);
 	unpin_user_pages_dirty_lock(&page, 1, true);
@@ -849,6 +856,7 @@ static struct user_event_enabler
 	enabler->event = user;
 	enabler->addr = uaddr;
 	enabler->values = reg->enable_bit;
+	enabler->size = reg->enable_size;
 retry:
 	/* Prevents state changes from racing with new enablers */
 	mutex_lock(&event_mutex);
@@ -2377,7 +2385,8 @@ static long user_unreg_get(struct user_unreg
__user *ureg,
 }

 static int user_event_mm_clear_bit(struct user_event_mm *user_mm,
-				   unsigned long uaddr, unsigned char bit)
+				   unsigned long uaddr, unsigned char bit,
+				   u8 size)
 {
 	struct user_event_enabler enabler;
 	int result;
@@ -2386,6 +2395,7 @@ static int user_event_mm_clear_bit(struct
user_event_mm *user_mm,
 	memset(&enabler, 0, sizeof(enabler));
 	enabler.addr = uaddr;
 	enabler.values = bit;
+	enabler.size = size;
 retry:
 	/* Prevents state changes from racing with new enablers */
 	mutex_lock(&event_mutex);
@@ -2416,6 +2426,7 @@ static long user_events_ioctl_unreg(unsigned long
uarg)
 	struct user_event_enabler *enabler, *next;
 	struct user_unreg reg;
 	long ret;
+	u8 size;

 	ret = user_unreg_get(ureg, &reg);

@@ -2441,6 +2452,8 @@ static long user_events_ioctl_unreg(unsigned long
uarg)
 		    ENABLE_BIT(enabler) == reg.disable_bit) {
 			set_bit(ENABLE_VAL_FREEING_BIT, ENABLE_BITOPS(enabler));

+			size = enabler->size;
+
 			if (!test_bit(ENABLE_VAL_FAULTING_BIT, ENABLE_BITOPS(enabler)))
 				user_event_enabler_destroy(enabler, true);

@@ -2454,7 +2467,7 @@ static long user_events_ioctl_unreg(unsigned long
uarg)
 	/* Ensure bit is now cleared for user, regardless of event status */
 	if (!ret)
 		ret = user_event_mm_clear_bit(mm, reg.disable_addr,
-					      reg.disable_bit);
+					      reg.disable_bit, size);

 	return ret;
 }

--

Clément

> 
> Thanks,
> 
> Clément
> 
>>
>> Thanks,
>> -Beau
>>
>> diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c
>> index 6f046650e527..b05db15eaea6 100644
>> --- a/kernel/trace/trace_events_user.c
>> +++ b/kernel/trace/trace_events_user.c
>> @@ -127,8 +127,11 @@ struct user_event_enabler {
>>  /* Bit 7 is for freeing status of enablement */
>>  #define ENABLE_VAL_FREEING_BIT 7
>>  
>> +/* Bit 8 is for marking 32-bit on 64-bit */
>> +#define ENABLE_VAL_32_BIT 8
>> +
>>  /* Only duplicate the bit value */
>> -#define ENABLE_VAL_DUP_MASK ENABLE_VAL_BIT_MASK
>> +#define ENABLE_VAL_DUP_MASK (ENABLE_VAL_BIT_MASK | 1 << ENABLE_VAL_32_BIT)
>>  
>>  #define ENABLE_BITOPS(e) (&(e)->values)
>>  
>> @@ -174,6 +177,29 @@ struct user_event_validator {
>>  	int			flags;
>>  };
>>  
>> +static inline void align_addr_bit(unsigned long *addr, int *bit,
>> +				  unsigned long *flags)
>> +{
>> +	if (IS_ALIGNED(*addr, sizeof(long))) {
>> +#ifdef __BIG_ENDIAN
>> +		if (test_bit(ENABLE_VAL_32_BIT, flags))
>> +			*bit += 32;
>> +#endif
>> +		return;
>> +	}
>> +
>> +	*addr = ALIGN_DOWN(*addr, sizeof(long));
>> +
>> +	/*
>> +	 * We only support 32 and 64 bit values. The only time we need
>> +	 * to align is a 32 bit value on a 64 bit kernel, which on LE
>> +	 * is always 32 bits, and on BE requires no change.
>> +	 */
>> +#ifdef __LITTLE_ENDIAN
>> +	*bit += 32;
>> +#endif
>> +}
>> +
>>  typedef void (*user_event_func_t) (struct user_event *user, struct iov_iter *i,
>>  				   void *tpdata, bool *faulted);
>>  
>> @@ -482,6 +508,7 @@ static int user_event_enabler_write(struct user_event_mm *mm,
>>  	unsigned long *ptr;
>>  	struct page *page;
>>  	void *kaddr;
>> +	int bit = ENABLE_BIT(enabler);
>>  	int ret;
>>  
>>  	lockdep_assert_held(&event_mutex);
>> @@ -497,6 +524,8 @@ static int user_event_enabler_write(struct user_event_mm *mm,
>>  		     test_bit(ENABLE_VAL_FREEING_BIT, ENABLE_BITOPS(enabler))))
>>  		return -EBUSY;
>>  
>> +	align_addr_bit(&uaddr, &bit, ENABLE_BITOPS(enabler));
>> +
>>  	ret = pin_user_pages_remote(mm->mm, uaddr, 1, FOLL_WRITE | FOLL_NOFAULT,
>>  				    &page, NULL);
>>  
>> @@ -515,9 +544,9 @@ static int user_event_enabler_write(struct user_event_mm *mm,
>>  
>>  	/* Update bit atomically, user tracers must be atomic as well */
>>  	if (enabler->event && enabler->event->status)
>> -		set_bit(ENABLE_BIT(enabler), ptr);
>> +		set_bit(bit, ptr);
>>  	else
>> -		clear_bit(ENABLE_BIT(enabler), ptr);
>> +		clear_bit(bit, ptr);
>>  
>>  	kunmap_local(kaddr);
>>  	unpin_user_pages_dirty_lock(&page, 1, true);
>> @@ -849,6 +878,12 @@ static struct user_event_enabler
>>  	enabler->event = user;
>>  	enabler->addr = uaddr;
>>  	enabler->values = reg->enable_bit;
>> +
>> +#if BITS_PER_LONG >= 64
>> +	if (reg->enable_size == 4)
>> +		set_bit(ENABLE_VAL_32_BIT, ENABLE_BITOPS(enabler));
>> +#endif
>> +
>>  retry:
>>  	/* Prevents state changes from racing with new enablers */
>>  	mutex_lock(&event_mutex);
>> @@ -2377,7 +2412,8 @@ static long user_unreg_get(struct user_unreg __user *ureg,
>>  }
>>  
>>  static int user_event_mm_clear_bit(struct user_event_mm *user_mm,
>> -				   unsigned long uaddr, unsigned char bit)
>> +				   unsigned long uaddr, unsigned char bit,
>> +				   unsigned long flags)
>>  {
>>  	struct user_event_enabler enabler;
>>  	int result;
>> @@ -2385,7 +2421,7 @@ static int user_event_mm_clear_bit(struct user_event_mm *user_mm,
>>  
>>  	memset(&enabler, 0, sizeof(enabler));
>>  	enabler.addr = uaddr;
>> -	enabler.values = bit;
>> +	enabler.values = bit | flags;
>>  retry:
>>  	/* Prevents state changes from racing with new enablers */
>>  	mutex_lock(&event_mutex);
>> @@ -2415,6 +2451,7 @@ static long user_events_ioctl_unreg(unsigned long uarg)
>>  	struct user_event_mm *mm = current->user_event_mm;
>>  	struct user_event_enabler *enabler, *next;
>>  	struct user_unreg reg;
>> +	unsigned long flags;
>>  	long ret;
>>  
>>  	ret = user_unreg_get(ureg, &reg);
>> @@ -2425,6 +2462,7 @@ static long user_events_ioctl_unreg(unsigned long uarg)
>>  	if (!mm)
>>  		return -ENOENT;
>>  
>> +	flags = 0;
>>  	ret = -ENOENT;
>>  
>>  	/*
>> @@ -2441,6 +2479,10 @@ static long user_events_ioctl_unreg(unsigned long uarg)
>>  		    ENABLE_BIT(enabler) == reg.disable_bit) {
>>  			set_bit(ENABLE_VAL_FREEING_BIT, ENABLE_BITOPS(enabler));
>>  
>> +			/* We must keep compat flags for the clear */
>> +			if (test_bit(ENABLE_VAL_32_BIT, ENABLE_BITOPS(enabler)))
>> +				flags |= 1 << ENABLE_VAL_32_BIT;
>> +
>>  			if (!test_bit(ENABLE_VAL_FAULTING_BIT, ENABLE_BITOPS(enabler)))
>>  				user_event_enabler_destroy(enabler, true);
>>  
>> @@ -2454,7 +2496,7 @@ static long user_events_ioctl_unreg(unsigned long uarg)
>>  	/* Ensure bit is now cleared for user, regardless of event status */
>>  	if (!ret)
>>  		ret = user_event_mm_clear_bit(mm, reg.disable_addr,
>> -					      reg.disable_bit);
>> +					      reg.disable_bit, flags);
>>  
>>  	return ret;
>>  }
diff mbox series

Patch

diff --git a/kernel/trace/trace_events_user.c b/kernel/trace/trace_events_user.c
index 6f046650e527..580c0fe4b23e 100644
--- a/kernel/trace/trace_events_user.c
+++ b/kernel/trace/trace_events_user.c
@@ -479,7 +479,7 @@  static int user_event_enabler_write(struct user_event_mm *mm,
 				    bool fixup_fault, int *attempt)
 {
 	unsigned long uaddr = enabler->addr;
-	unsigned long *ptr;
+	unsigned long *ptr, bit_offset;
 	struct page *page;
 	void *kaddr;
 	int ret;
@@ -511,13 +511,19 @@  static int user_event_enabler_write(struct user_event_mm *mm,
 	}
 
 	kaddr = kmap_local_page(page);
+
+	bit_offset = uaddr & (sizeof(unsigned long) - 1);
+	if (bit_offset) {
+		bit_offset *= 8;
+		uaddr &= ~(sizeof(unsigned long) - 1);
+	}
 	ptr = kaddr + (uaddr & ~PAGE_MASK);
 
 	/* Update bit atomically, user tracers must be atomic as well */
 	if (enabler->event && enabler->event->status)
-		set_bit(ENABLE_BIT(enabler), ptr);
+		set_bit(ENABLE_BIT(enabler) + bit_offset, ptr);
 	else
-		clear_bit(ENABLE_BIT(enabler), ptr);
+		clear_bit(ENABLE_BIT(enabler) + bit_offset, ptr);
 
 	kunmap_local(kaddr);
 	unpin_user_pages_dirty_lock(&page, 1, true);