On Wed, Feb 28, 2024 at 6:16 PM Yafang Shao <laoar.shao@xxxxxxxxx> wrote:
>
> On Wed, Feb 28, 2024 at 2:04 PM Andrii Nakryiko
> <andrii.nakryiko@xxxxxxxxx> wrote:
> >
> > On Tue, Feb 27, 2024 at 6:25 PM Yafang Shao <laoar.shao@xxxxxxxxx> wrote:
> > >
> > > On Wed, Feb 28, 2024 at 9:24 AM Andrii Nakryiko
> > > <andrii.nakryiko@xxxxxxxxx> wrote:
> > > >
> > > > On Sun, Feb 25, 2024 at 2:07 AM Yafang Shao <laoar.shao@xxxxxxxxx> wrote:
> > > > >
> > > > > Add three new kfuncs for the bits iterator:
> > > > > - bpf_iter_bits_new
> > > > > Initialize a new bits iterator for a given memory area. Due to the
> > > > > limitation of bpf memalloc, the max number of bits that can be iterated
> > > > > over is limited to (4096 * 8).
> > > > > - bpf_iter_bits_next
> > > > > Get the next bit in a bpf_iter_bits
> > > > > - bpf_iter_bits_destroy
> > > > > Destroy a bpf_iter_bits
> > > > >
> > > > > The bits iterator facilitates the iteration of the bits of a memory area,
> > > > > such as cpumask. It can be used in any context and on any address.
> > > > >
> > > > > Signed-off-by: Yafang Shao <laoar.shao@xxxxxxxxx>
> > > > > ---
> > > > > kernel/bpf/helpers.c | 100 +++++++++++++++++++++++++++++++++++++++++++
> > > > > 1 file changed, 100 insertions(+)
> > > > >
> > > > > diff --git a/kernel/bpf/helpers.c b/kernel/bpf/helpers.c
> > > > > index 93edf730d288..052f63891834 100644
> > > > > --- a/kernel/bpf/helpers.c
> > > > > +++ b/kernel/bpf/helpers.c
> > > > > @@ -2542,6 +2542,103 @@ __bpf_kfunc void bpf_throw(u64 cookie)
> > > > > WARN(1, "A call to BPF exception callback should never return\n");
> > > > > }
> > > > >
> > > > > +struct bpf_iter_bits {
> > > > > + __u64 __opaque[2];
> > > > > +} __aligned(8);
> > > > > +
> > > > > +struct bpf_iter_bits_kern {
> > > > > + unsigned long *bits;
> > > > > + u32 nr_bits;
> > > > > + int bit;
> > > > > +} __aligned(8);
> > > > > +
> > > > > +/**
> > > > > + * bpf_iter_bits_new() - Initialize a new bits iterator for a given memory area
> > > > > + * @it: The new bpf_iter_bits to be created
> > > > > + * @unsafe_ptr__ign: A ponter pointing to a memory area to be iterated over
> > > > > + * @nr_bits: The number of bits to be iterated over. Due to the limitation of
> > > > > + * memalloc, it can't greater than (4096 * 8).
> > > > > + *
> > > > > + * This function initializes a new bpf_iter_bits structure for iterating over
> > > > > + * a memory area which is specified by the @unsafe_ptr__ign and @nr_bits. It
> > > > > + * copy the data of the memory area to the newly created bpf_iter_bits @it for
> > > > > + * subsequent iteration operations.
> > > > > + *
> > > > > + * On success, 0 is returned. On failure, ERR is returned.
> > > > > + */
> > > > > +__bpf_kfunc int
> > > > > +bpf_iter_bits_new(struct bpf_iter_bits *it, const void *unsafe_ptr__ign, u32 nr_bits)
> > > > > +{
> > > > > + struct bpf_iter_bits_kern *kit = (void *)it;
> > > > > + u32 size = BITS_TO_BYTES(nr_bits);
> > > > > + int err;
> > > > > +
> > > > > + BUILD_BUG_ON(sizeof(struct bpf_iter_bits_kern) != sizeof(struct bpf_iter_bits));
> > > > > + BUILD_BUG_ON(__alignof__(struct bpf_iter_bits_kern) !=
> > > > > + __alignof__(struct bpf_iter_bits));
> > > > > +
> > > > > + if (!unsafe_ptr__ign || !nr_bits) {
> > > > > + kit->bits = NULL;
> > > > > + return -EINVAL;
> > > > > + }
> > > > > +
> > > > > + kit->bits = bpf_mem_alloc(&bpf_global_ma, size);
> > > > > + if (!kit->bits)
> > > > > + return -ENOMEM;
> > > >
> > > > it's probably going to be a pretty common case to do bits iteration
> > > > for nr_bits<=64, right?
> > >
> > > It's highly unlikely.
> > > Consider the CPU count as an example; There are 256 CPUs on our AMD
> > > EPYC servers.
> >
> > Also consider u64-based bit masks (like struct backtrack_state in
> > verifier code, which has u32 reg_mask and u64 stack_mask). This
> > iterator is a generic bits iterator, there are tons of cases of
> > u64/u32 masks in practice.
>
> Should we optimize it as follows?
>
> if (nr_bits <= 64) {
> // do the optimization
> } else {
> // fallback to memalloc
> }
>
Yep, that's what I'm proposing
> >
> > >
> > > > So as an optimization, instead of doing
> > > > bpf_mem_alloc() for this case, you can just copy up to 8 bytes and
> > > > store it in a union of `unsigned long *bits` and `unsigned long
> > > > bits_copy`. As a performance optimization (and to reduce dependency on
> > > > memory allocation). WDYT?
> > > >
> > >
> > > --
> > > Regards
> > > Yafang
>
>
>
> --
> Regards
> Yafang
