Re: [PATCH bpf-next 1/2] bpf: Add a bpf_kallsyms_lookup helper

From: Alexei Starovoitov
Date: Fri Dec 18 2020 - 15:37:45 EST


On Fri, Dec 18, 2020 at 10:53:57AM -0800, Andrii Nakryiko wrote:
> On Thu, Dec 17, 2020 at 7:20 PM Alexei Starovoitov
> <alexei.starovoitov@xxxxxxxxx> wrote:
> >
> > On Thu, Dec 17, 2020 at 09:26:09AM -0800, Yonghong Song wrote:
> > >
> > >
> > > On 12/17/20 7:31 AM, Florent Revest wrote:
> > > > On Mon, Dec 14, 2020 at 7:47 AM Yonghong Song <yhs@xxxxxx> wrote:
> > > > > On 12/11/20 6:40 AM, Florent Revest wrote:
> > > > > > On Wed, Dec 2, 2020 at 10:18 PM Alexei Starovoitov
> > > > > > <alexei.starovoitov@xxxxxxxxx> wrote:
> > > > > > > I still think that adopting printk/vsnprintf for this instead of
> > > > > > > reinventing the wheel
> > > > > > > is more flexible and easier to maintain long term.
> > > > > > > Almost the same layout can be done with vsnprintf
> > > > > > > with exception of \0 char.
> > > > > > > More meaningful names, etc.
> > > > > > > See Documentation/core-api/printk-formats.rst
> > > > > >
> > > > > > I agree this would be nice. I finally got a bit of time to experiment
> > > > > > with this and I noticed a few things:
> > > > > >
> > > > > > First of all, because helpers only have 5 arguments, if we use two for
> > > > > > the output buffer and its size and two for the format string and its
> > > > > > size, we are only left with one argument for a modifier. This is still
> > > > > > enough for our usecase (where we'd only use "%ps" for example) but it
> > > > > > does not strictly-speaking allow for the same layout that Andrii
> > > > > > proposed.
> > > > >
> > > > > See helper bpf_seq_printf. It packs all arguments for format string and
> > > > > puts them into an array. bpf_seq_printf will unpack them as it parsed
> > > > > through the format string. So it should be doable to have more than
> > > > > "%ps" in format string.
> > > >
> > > > This could be a nice trick, thank you for the suggestion Yonghong :)
> > > >
> > > > My understanding is that this would also require two extra args (one
> > > > for the array of arguments and one for the size of this array) so it
> > > > would still not fit the 5 arguments limit I described in my previous
> > > > email.
> > > > eg: this would not be possible:
> > > > long bpf_snprintf(const char *out, u32 out_size,
> > > > const char *fmt, u32 fmt_size,
> > > > const void *data, u32 data_len)
> > >
> > > Right. bpf allows only up to 5 parameters.
> > > >
> > > > Would you then suggest that we also put the format string and its
> > > > length in the first and second cells of this array and have something
> > > > along the line of:
> > > > long bpf_snprintf(const char *out, u32 out_size,
> > > > const void *args, u32 args_len) ?
> > > > This seems like a fairly opaque signature to me and harder to verify.
> > >
> > > One way is to define an explicit type for args, something like
> > > struct bpf_fmt_str_data {
> > > char *fmt;
> > > u64 fmt_len;
> > > u64 data[];
> > > };
> >
> > that feels a bit convoluted.
> >
> > The reason I feel unease with the helper as was originally proposed
> > and with Andrii's proposal is all the extra strlen and strcpy that
> > needs to be done. In the helper we have to call kallsyms_lookup()
> > which is ok interface for what it was desinged to do,
> > but it's awkward to use to construct new string ("%s [%s]", sym, modname)
> > or to send two strings into a ring buffer.
> > Andrii's zero separator idea will simplify bpf prog, but user space
> > would need to do strlen anyway if it needs to pretty print.
> > If we take pain on converting addr to sym+modname let's figure out
> > how to make it easy for the bpf prog to do and easy for user space to consume.
> > That's why I proposed snprintf.
>
> I have nothing against snprintf support for symbols. But
> bpf_ksym_resolve() solves only a partially overlapping problem, so
> deserves to be added in addition to snprintf support. With snprintf,
> it will be hard to avoid two lookups of the same symbol to print "%s
> [%s]" form, so there is a performance loss, which is probably bigger
> than a simple search for a zero-byte.

I suspect we're not on the same page in terms of what printf can do.
See Documentation/core-api/printk-formats.rst and lib/vsprintf.c:symbol_string()
It's exactly one lookup in sprintf implementation.
bpf_snprintf(buf, "%ps", addr) would be equivalent to
{
ksym_resolve(sym, modname, addr, SYM | MOD);
printf("%s [%s]", sym, modname);
}

> But bpf_ksym_resolve() can be
> used flexibly. You can either do two separate bpf_ksym_resolve() calls
> to get symbol name (and its length) and symbol's module (and its
> length), if you need to process it programmatically in BPF program. Or
> you can bundle it together and let user-space process it. User-space
> will need to copy data anyways because it can't stay in
> perfbuf/ringbuf for long. So scanning for zero delimiters will be
> negligible, it will just bring data into cache. All I'm saying is that
> ksym_resolve() gives flexibility which snprintf can't provide.

Well, with snprintf there will be no way to print mod symbol
without modname, but imo it's a good thing.
What is the use case for getting mod symbol without modname?

> Additionally, with ksym_resolve() being able to return base address,
> it's now possible to do a bunch of new stuff, from in-BPF
> symbolization to additional things like correlating memory accesses or
> function calls, etc.

Getting adjusted base address could be useful some day, but why now? What for?

> bits), my point is that ksym_resolve() is more powerful than
> snprintf(): the latter can be used pretty much only for
> pretty-printing.

Potentially yes. I think the stated goal was pretty printing.

>
> >
> > As far as 6 arg issue:
> > long bpf_snprintf(const char *out, u32 out_size,
> > const char *fmt, u32 fmt_size,
> > const void *data, u32 data_len);
> > Yeah. It won't work as-is, but fmt_size is unnecessary nowadays.
> > The verifier understands read-only data.
> > Hence the helper can be:
> > long bpf_snprintf(const char *out, u32 out_size,
>
> With the power of BTF, we can also put these two correlated values
> into a single struct and pass a pointer to it. It will take only one
> parameter for one memory region. Alternative is the "fat pointer"
> approach that Go and Rust use, but it's less flexible overall.

I think it will be less flexible when output size is fixed by the type info.
With explicit size the bpf_snprintf() can print directly into ringbuffer.
Multiple bpf_snprintf() will be able to fill it one by one reducing
space available at every step.
bpf_snprintf() would need to return the number of bytes, of course.
Just like probe_read_str.