Re: [PATCH v8 2/2] lib: checksum: Use aligned accesses for ip_fast_csum and csum_ipv6_magic tests

[Guenter Roeck <linux@roeck-us.net>]

Hi Charlie,

On 2/14/24 17:30, Charlie Jenkins wrote:
> On Wed, Feb 14, 2024 at 03:03:07PM -0800, Guenter Roeck wrote:
>> On 2/14/24 13:41, Charlie Jenkins wrote:
>>> The test cases for ip_fast_csum and csum_ipv6_magic were failing on a
>>> variety of architectures that are big endian or do not support
>>> misalgined accesses. Both of these test cases are changed to support big
>>> and little endian architectures.
>>>
>>> The test for ip_fast_csum is changed to align the data along (14 +
>>> NET_IP_ALIGN) bytes which is the alignment of an IP header. The test for
>>> csum_ipv6_magic aligns the data using a struct. An extra padding field
>>> is added to the struct to ensure that the size of the struct is the same
>>> on all architectures (44 bytes).
>>>
>>> The test for csum_ipv6_magic somewhat arbitrarily aligned saddr and
>>> daddr. This would fail on parisc64 due to the following code snippet in
>>> arch/parisc/include/asm/checksum.h:
>>>
>>> add		%4, %0, %0\n"
>>> ldd,ma		8(%1), %6\n"
>>> ldd,ma		8(%2), %7\n"
>>> add,dc		%5, %0, %0\n"
>>>
>>> The second add is expecting carry flags from the first add. Normally,
>>> a double word load (ldd) does not modify the carry flags. However,
>>> because saddr and daddr may be misaligned, ldd triggers a misalignment
>>> trap that gets handled in arch/parisc/kernel/unaligned.c. This causes
>>> many additional instructions to be executed between the two adds. This
>>> can be easily solved by adding the carry into %0 before executing the
>>> ldd.
>>>
>>
>> I really think this is a bug either in the trap handler or in the hppa64
>> qemu emulation. Only unaligned ldd instructions affect (actually,
>> unconditionally set) the carry flag. That doesn't happen with unaligned
>> ldw instructions. It would be worthwhile tracking this down since there are
>> lots of unaligned data accesses (8-byte accesses on 4-byte aligned addresses)
>> when running the kernel in 64-bit mode. On the other side, I guess this
>> is a different problem. Not sure though if that should even be mentioned
>> here since that makes it sound as if it would be expected that such
>> accesses impact the carry flag.
> 
> I wasn't confident it was a bug somewhere so that's why I sent this patch.
> 
> However, I have just found the section of the processor manual [1] I was
> looking for (Section Privileged Software-Accessible Registers subsection
> Processor Status Word (PSW)):
> 
> "Processor state is encoded in a 64-bit register called the Processor
> Status Word (PSW). When an interruption occurs, the current value of the
> PSW is saved in the Interruption Processor Status Word (IPSW) and
> usually all defined PSW bits are set to 0.
> 
> "The PSW is set to the contents of the IPSW by the RETURN FROM
> INTERRUPTION instruction. The interruption handler may restore the
> original PSW, modify selected bits, or may change the PSW to an entirely
> new value."
> 
> Stored in the PSW register are the "Carry/borrow bits". This confirms
> that the carry/borrow bits should be restored. The save is supposed to
> automatically happen upon an interrupt and restored by the RETURN FROM
> INTERRUPTION, thus this is a QEMU bug and not a Linux bug (please
> correct me if I am wrong).
> 

I know that much (I looked into the manual as well), I just really don't
know if this is a Linux bug or a QEMU bug, and I have not been able to
nail it down. I think someone with access to hardware will need to confirm.

Specifically: Yes, the carry/borrow bits should be restored. Question is
if the Linux kernel's interrupt handler doesn't restore the carry bits
or if the problem is on the qemu side.

> This v8 was not needed after-all it seems. It would be best to stick
> with the v7.
> 
I tend to agree; after all, v7 exposes the problem, making it easier to
determine if the problem can be reproduced on real hardware.

FWIW,I wrote some test code which exposes the problem. It is quite
easy to show that carry is always set after executing ldd on an unaligned
address. That is also why I know for sure that the problem is not
seen with ldw on unaligned addresses.

Thanks,
Guenter