On Mon, May 08, 2023 at 06:58:23PM +0100, Jeremy Sowden wrote: > On 2023-05-03, at 22:54:11 +0200, Pablo Neira Ayuso wrote: > > On Tue, Apr 25, 2023 at 08:51:43PM +0100, Jeremy Sowden wrote: > > > On 2023-04-12, at 13:06:02 +0200, Pablo Neira Ayuso wrote: > > > > I mean, would it be possible to add a NFT_BITWISE_BOOL variant that > > > > takes _SREG2 via select_ops? > > > > > > In an earlier version, instead of adding new boolean ops, I added > > > support for passing the mask and xor arguments in registers: > > > > > > https://lore.kernel.org/netfilter-devel/20200224124931.512416-1-jeremy@xxxxxxxxxx/ > > > > > > Doing the same thing with one extra register is straightforward for AND > > > and XOR: > > > > > > AND(x, y) = (x & y) ^ 0 > > > XOR(x, y) = (x & 1) ^ y > > > > > > since we can pass y in _SREG2 and 0 in _XOR for AND, and 1 in _MASK and > > > y in _SREG2 for XOR. For OR: > > > > > > OR(x, y) = (x & ~y) ^ y > > > > > > it's a bit more complicated. Instead of getting both the mask and xor > > > arguments from user space, we need to do something like passing y in > > > _SREG2 alone, and then constructing the bitwise negation in the kernel. > > > > > > Obviously, this means that the kernel is no longer completely agnostic > > > about the sorts of mask-and-xor expressions user space may send. > > > > > > Since that is the case, we could go further and just perform the > > > original ope- rations. Thus if we get an boolean op with an _SREG2 > > > argument: > > > > > > * if there is an _XOR of 0, compute: > > > > > > _SREG & _SREG2 > > > > > > * if there is a _MASK of 1, compute: > > > > > > _SREG ^ _SREG2 > > > > > > * if there are no _MASK or _XOR arguments, compute: > > > > > > _SREG | _SREG2 > > > > OK, if my understanding is correct, these are the two options: > > > > 1) Infer from arguments the type of operation. > > 2) Have explicit NFT_BITWISE_{AND,OR,XOR} operations. > > > > If so, I think it is better to stick to your original patch, where > > explicit bitwise operations NFT_BITWISE_{_AND,_OR,_XOR} are added > > (which is what you proposed last time IIRC). > > > > Thanks for explaining. > > No problem. I'll get rebasing. Please, small batch, not larger than 10 patches if possible.
