Date:   Tue, 24 Jan 2023 23:36:45 -0600
From:   David Vernet <void@...ifault.com>
To:     Alexei Starovoitov <alexei.starovoitov@...il.com>
Cc:     bpf@...r.kernel.org, ast@...nel.org, daniel@...earbox.net,
        andrii@...nel.org, martin.lau@...ux.dev, song@...nel.org,
        yhs@...a.com, john.fastabend@...il.com, kpsingh@...nel.org,
        sdf@...gle.com, haoluo@...gle.com, jolsa@...nel.org,
        linux-kernel@...r.kernel.org, kernel-team@...a.com, tj@...nel.org,
        memxor@...il.com
Subject: Re: [PATCH bpf-next v2 4/9] bpf: Enable cpumasks to be queried and
 used as kptrs

On Tue, Jan 24, 2023 at 08:36:02PM -0800, Alexei Starovoitov wrote:
> On Fri, Jan 20, 2023 at 01:25:18PM -0600, David Vernet wrote:
> > +
> > +/**
> > + * struct bpf_cpumask - refcounted BPF cpumask wrapper structure
> > + * @cpumask:	The actual cpumask embedded in the struct.
> > + * @usage:	Object reference counter. When the refcount goes to 0, the
> > + *		memory is released back to the BPF allocator, which provides
> > + *		RCU safety.
> > + *
> > + * Note that we explicitly embed a cpumask_t rather than a cpumask_var_t.  This
> > + * is done to avoid confusing the verifier due to the typedef of cpumask_var_t
> > + * changing depending on whether CONFIG_CPUMASK_OFFSTACK is defined or not. See
> > + * the details in <linux/cpumask.h>. The consequence is that this structure is
> > + * likely a bit larger than it needs to be when CONFIG_CPUMASK_OFFSTACK is
> > + * defined due to embedding the whole NR_CPUS-size bitmap, but the extra memory
> > + * overhead is minimal. For the more typical case of CONFIG_CPUMASK_OFFSTACK
> > + * not being defined, the structure is the same size regardless.
> > + */
> > +struct bpf_cpumask {
> > +	cpumask_t cpumask;
> > +	refcount_t usage;
> > +};
> > +
> > +static struct bpf_mem_alloc bpf_cpumask_ma;
> > +
> > +static bool cpu_valid(u32 cpu)
> > +{
> > +	return cpu < nr_cpu_ids;
> > +}
> > +
> > +__diag_push();
> > +__diag_ignore_all("-Wmissing-prototypes",
> > +		  "Global kfuncs as their definitions will be in BTF");
> > +
> > +struct bpf_cpumask *bpf_cpumask_create(void)
> > +{
> > +	struct bpf_cpumask *cpumask;
> > +
> > +	cpumask = bpf_mem_alloc(&bpf_cpumask_ma, sizeof(*cpumask));
> > +	if (!cpumask)
> > +		return NULL;
> > +
> > +	memset(cpumask, 0, sizeof(*cpumask));
> > +	refcount_set(&cpumask->usage, 1);
> > +
> > +	return cpumask;
> > +}
> 
> Applied patches 1 and 2. Patch 3 doesn't apply anymore. Pls rebase.

Ack, will rebase for v3.

> I'm fine with existing bpf_cpumask proposal, but can we do better?
> This is so close to be a bitmap template.

Agreed that they're close, but I'm not a fan of the UX taxes for what we
get out of it. More below.

> Can we generalize it as
> struct bpf_bitmap {
> 	refcount_t refcnt;
>         int num_bits;
>         u64 bits[];
> };
> 
> struct bpf_bitmap *bpf_bitmap_create(int bits)
> {
>   bitmap = bpf_mem_alloc(&bpf_cpumask_ma, sizeof(*bitmap) + BITS_TO_LONGS(bits) * sizeof(u64));
>   bitmap->num_bits = bits;
> }

+1 that having bitmap kfuncs would be nice to expose, and should be
pretty easy to add. Happy to do so in a follow-on patch set.

> 
> and special case few custom kfuncs in the verifier that allow
> type cast from bpf_bitmap with to 'struct cpumask *' ? Like
> struct cpumask *bpf_bitmap_cast_to_cpumask(struct bpf_bitmap *bitmap)
> {
>   if (bitmap->num_bits == nr_cpu_ids)
>      return bitmap->bits;
>   return NULL;
> }
> BTF_ID_FLAGS(func, bpf_bitmap_cast_to_cpumask, KF_TRUSTED_ARGS | KF_RET_NULL)

This I'm not a huge fan of though. It seems like we're removing a useful
abstraction and adding a UX tax just to avoid defining and exporting an
additional small set of kfuncs for allocating, and acquire/releasing a
struct bpf_cpumask. That logic is very minimal, just around 100 lines of
code including doxygen comments.

It's kind of unfortunate that cpumask is so close to bitmap, but that's
nothing new -- <linux/cpumask.h> in the kernel is little more than a
thin wrapper around a bitmap that simply provides some more ergonomic
APIs, along with some magic that makes it safe to access cpumask_var_t
on the stack regardless of NR_CPUS. The latter doesn't apply to BPF, but
the former does.

> The UX will be a bit worse, since bpf prog would need to do !=NULL check
> but with future bpf_assert() we may get rid of !=NULL check.
> 
> We can keep direct cpumask accessors as kfuncs:
> 
> u32 bpf_cpumask_first(const struct cpumask *cpumask);
> u32 bpf_cpumask_first_zero(const struct cpumask *cpumask);
> 
> and add bpf_find_first_bit() and the rest of bit manipulations.

Worth noting as well is that I think struct bpf_bitmap is going to be
treated somewhat differently than struct bpf_cpumask and struct cpumask.
There is no type-safety for bitmaps in the kernel. They're just
represented as unsigned long *, so I don't we'll be able to allow
programs to pass bitmaps allocated elsewhere in the kernel to read-only
bitmap kfuncs like we do for struct cpumask *, as the verifier will just
interpret them as pointers to statically sized scalars.

> Since all of the bpf_cpumask do run-time cpu_valid() check we're not
> sacrificing performance.
> 
> Feels more generic with wider applicability at the expense of little bit worse UX.
> I haven't thought about acq/rel consequences.

The TL;DR from me is that I agree that having bitmap kfuncs is a great
idea, but I don't see the need to tie the two at the hip at the cost of
a worse UX. I'd prefer to push the extra complexity into the BPF backend
in favor of a simpler programming front-end for users.

Thoughts?

Powered by blists - more mailing lists