On Sun, Feb 09, 2025 at 07:44:18AM -0800, Eric Biggers wrote:
> On Sun, Feb 09, 2025 at 11:12:55AM +0200, Leon Romanovsky wrote:
> > On Thu, Feb 06, 2025 at 07:57:50PM -0800, Eric Biggers wrote:
> > > On Thu, Feb 06, 2025 at 07:36:43PM -0800, Eric Biggers wrote:
> > > > +int irdma_ieq_check_mpacrc(const void *addr, u32 len, u32 val)
> > > > {
> > > > - u32 crc = 0;
> > > > -
> > > > - crypto_shash_digest(desc, addr, len, (u8 *)&crc);
> > > > - if (crc != val)
> > > > + if (~crc32c(~0, addr, len) != val)
> > > > return -EINVAL;
> > > >
> > > > return 0;
> > > > }
> > >
> > > Sorry, I just realized this isn't actually equivalent on big endian CPUs, since
> > > the byte array produced by crypto_shash_digest() used little endian byte order,
> > > whereas crc32c() just returns a CPU endian value.
> > >
> > > And of course this broken subsystem uses u32 for the little endian values
> > > instead of __le32 like the result of the kernel.
> > >
> > > Not sure it's worth my time to continue to try to fix this subsystem properly.
> >
> > There is no need to be such dramatic. You are not fixing anything by
> > switch to new APIs
>
> Exactly. That's because I dropped the patches that actually did fix real
> endianness bugs, because of the pointless pushback I received -- see
> https://lore.kernel.org/linux-rdma/20250127223840.67280-1-ebiggers@xxxxxxxxxx/T/#u
Anyway, I already sent v3 of this patch that keeps the cpu_to_le32() to maintain
the exact same behavior as the old code, so please consider that if you are
interested. Note that I had to add '(__force u32)' to be compatible with this
driver's incorrect types, but that was effectively already there before, just
hidden by writing bytes into a u32.
- Eric
