Re: [PATCH v8 net-next 1/3] filter: add Extended BPF interpreter and converter

[Alexei Starovoitov]

On Mon, Mar 10, 2014 at 7:20 PM, H. Peter Anvin <hpa@xxxxxxxxx> wrote:
> Bool vs bitfield are orthogonal, at least under gcc, unless I'm completely out to sea.  It is probably not a good idea to create a bitfield when it doesn't buy you anything, lest you generate rmw instructions when byte stores would do.
>

access to upper fields of sk_filter is not in critical path, therefore
my preference
is to save space.
In this case:
struct sk_filter {
        atomic_t                refcnt;
        unsigned int            len;
        unsigned int            len_ext;
        bool or unsigned int:1            jited;
        and the rest are pointers.

so either u32:1 or bool approach adds 4 bytes.
The reason I used bitfield is that I was expecting someone to complain
about overall size
increase and I can optimize it later into:
        unsigned int            len:15;
        unsigned int            len_ext:15;
        unsigned int            jited:1;
without touching arch/*/net/*jit files, since filter len fits.

Only lower part is cache sensitive, since 'insns/insns_ext' need to come close
enough to 'bpf_func', so that interpreter (called via SK_RUN_FILTER macro)
starts executing insns out of cache.

Thanks
Alexei

> On March 10, 2014 7:02:18 PM PDT, Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>>On Mon, Mar 10, 2014 at 6:51 PM, David Miller <davem@xxxxxxxxxxxxx>
>>wrote:
>>> From: Alexei Starovoitov <ast@xxxxxxxxxxxx>
>>> Date: Sun,  9 Mar 2014 23:04:02 -0700
>>>
>>>> +     unsigned int            jited:1;
>>>
>>> The C language has a proper type for boolean states, please therefore
>>> use 'bool', true, and false.
>>
>>No, the C standard actually has no such thing.
>>
>>In a structure, a bitfield is actually better than bool, because it
>>takes only one bit. A "bool" takes at least a byte.
>>
>>Now, in this case it may not be an issue (looks like there are no
>>other uses that can use the better packing, so bit/byte/word is all
>>the same), but I really really want to make it clear that it is not at
>>all true that "bool" is somehow better than a single-bit bitfield. The
>>bitfield can pack *much* better, and I would actually say that it's
>>generally a *better* idea to use a bitfield, because you can much more
>>easily expand on it later by adding other bitfields.
>>
>>There are very few actual real advantages to "bool". The magic casting
>>behavior is arguably an advantage (the implicit cast in assigning to a
>>bitfield truncates to the low bits, the implicit cast on assignment to
>>"bool" does a test against zero), but is also quite arguably a
>>possible source of confusion and can cause problems down the line when
>>converting from bool to a bitfield (for the afore-mentioned packing
>>reasons).
>>
>>So please don't sell "bool" as some kind of panacea. It has at least
>>as many problems as it has advantages.
>>
>>I would generally suggest that people only use "bool" for function
>>return types, and absolutely nothing else. Seriously.
>>
>>              Linus
>
> --
> Sent from my mobile phone.  Please pardon brevity and lack of formatting.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/