On Tue, Feb 1, 2022 at 6:36 PM Hou Tao <hotforest@xxxxxxxxx> wrote:
>
> Hi,
>
> > >
> > > Hi Andrii,
> > >
> > > > >
> > > > > 4096 is OK for x86-64, but for other archs with greater than 4KB
> > > > > page size (e.g. 64KB under arm64), test_verifier for test case
> > > > > "check valid spill/fill, ptr to mem" will fail, so just use
> > > > > getpagesize() to initialize the ring buffer size. Do this for
> > > > > test_progs as well.
> > > > >
> > > [...]
> > >
> > > > > diff --git a/tools/testing/selftests/bpf/progs/ima.c b/tools/testing/selftests/bpf/progs/ima.c
> > > > > index 96060ff4ffc6..e192a9f16aea 100644
> > > > > --- a/tools/testing/selftests/bpf/progs/ima.c
> > > > > +++ b/tools/testing/selftests/bpf/progs/ima.c
> > > > > @@ -13,7 +13,6 @@ u32 monitored_pid = 0;
> > > > >
> > > > > struct {
> > > > > __uint(type, BPF_MAP_TYPE_RINGBUF);
> > > > > - __uint(max_entries, 1 << 12);
> > > >
> > > > Should we just bump it to 64/128/256KB instead? It's quite annoying to
> > > > do a split open and then load just due to this...
> > > >
> > > Agreed.
> > >
> > > > I'm also wondering if we should either teach kernel to round up to
> > > > closes power-of-2 of page_size internally, or teach libbpf to do this
> > > > for RINGBUF maps. Thoughts?
> > > >
> > > It seems that max_entries doesn't need to be page-aligned. For example
> > > if max_entries is 4096 and page size is 65536, we can allocate a
> > > 65536-sized page and set rb->mask 4095 and it will work. The only
> > > downside is 60KB memory is waster, but it is the implementation
> > > details and can be improved if subpage mapping can be supported.
> > >
> > > So how about removing the page-aligned restraint in kernel ?
> > >
> >
> > No, if you read BPF ringbuf code carefully you'll see that we map the
> > entire ringbuf data twice in the memory (see [0] for lame ASCII
> > diagram), so that records that are wrapped at the end of the ringbuf
> > and go back to the start are still accessible as a linear array. It's
> > a very important guarantee, so it has to be page size multiple. But
> > auto-increasing it to the closest power-of-2 of page size seems like a
> > pretty low-impact change. Hard to imagine breaking anything except
> > some carefully crafted tests for ENOSPC behavior.
> >
>
> Yes, i know the double map trick. What i tried to say is that:
> (1) remove the page-aligned restrain for max_entries
> (2) still allocate page-aligned memory for ringbuf
>
> instead of rounding max_entries up to closest power-of-2 page size
> directly, so max_entries from userspace is unchanged and double map trick
> still works.
I don't see how. Knowing the correct and exact size of the ringbuf
data area is mandatory for correctly consuming ringbuf data from
user-space. But if I'm missing something, feel free to give it a try
and see if it actually works.
>
> > [0] https://github.com/torvalds/linux/blob/master/kernel/bpf/ringbuf.c#L73-L89
>
> > > Regards,
> > > Tao
>
