Fix the csum_and_copy_..._user() idiocy
We need a way for csum_and_copy_{from,to}_user() to report faults.
The approach taken back in 2020 (avoid 0 as return value by starting
summing from ~0U, use 0 to report faults) had been broken; it does
yield the right value modulo 2^16-1, but the case when data is
entirely zero-filled is not handled right. It almost works, since
for most of the codepaths we have a non-zero value added in
and there 0 is not different from anything divisible by 0xffff.
However, there are cases (ICMPv4 replies, for example) where we
are not guaranteed that.
In other words, we really need to have those primitives return 0
on filled-with-zeroes input. So let's make them return a 64bit
value instead; we can do that cheaply (all supported architectures
do that via a couple of registers) and we can use that to report
faults without disturbing the 32bit csum.
New type: __wsum_fault. 64bit, returned by csum_and_copy_..._user().
Primitives:
* CSUM_FAULT representing the fault
* to_wsum_fault() folding __wsum value into that
* from_wsum_fault() extracting __wsum value
* wsum_fault_check() checking if it's a fault value
Representation depends upon the target.
CSUM_FAULT: ~0ULL
to_wsum_fault(v32): (u64)v32 for 64bit and 32bit l-e,
(u64)v32 << 32 for 32bit b-e.
Rationale: relationship between the calling conventions for returning 64bit
and those for returning 32bit values. On 64bit architectures the same
register is used; on 32bit l-e the lower half of the value goes in the
same register that is used for returning 32bit values and the upper half
goes into additional register. On 32bit b-e the opposite happens -
upper 32 bits go into the register used for returning 32bit values and
the lower 32 bits get stuffed into additional register.
So with this choice of representation we need minimal changes on the
asm side (zero an extra register in 32bit case, nothing in 64bit case),
and from_wsum_fault() is as cheap as it gets.
Sum calculation is back to "start from 0".
X-paperbag: brown
Fixes: c693cc4676a0 "saner calling conventions for csum_and_copy_..._user()"
Fucked-up-by: Al Viro <viro@zeniv.linux.org.uk>
Reported-by: gus Gusenleitner Klaus <gus@keba.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
30 files changed
