On Wed, 4 Oct 2023 09:54:31 -0700
Jakub Kicinski <[email protected]> wrote:
> On Wed, 4 Oct 2023 12:35:24 -0400 Steven Rostedt wrote:
> > > Potentially naive question - the trace point holds enum skb_drop_reason.
> > > The user space can get the names from BTF. Can we not teach user space
> > > to generically look up names of enums in BTF?
> >
> > That puts a hard requirement to include BTF in builds where it was not
> > needed before. I really do not want to build with BTF just to get access to
> > these symbols. And since this is used by the embedded world, and BTF is
> > extremely bloated, the short answer is "No".
>
> Dunno. BTF is there most of the time. It could make the life of
> majority of the users far more pleasant.
BTF isn't there for a lot of developers working in embedded who use this
code. Most my users that I deal with have minimal environments, so BTF is a
showstopper.
>
> I hope we can at least agree that the current methods of generating
> the string arrays at C level are... aesthetically displeasing.
I don't know, I kinda like it ;-)
-- Steve
On 04/10/2023 18:29, Steven Rostedt wrote:
> On Wed, 4 Oct 2023 09:54:31 -0700
> Jakub Kicinski <[email protected]> wrote:
>
>> On Wed, 4 Oct 2023 12:35:24 -0400 Steven Rostedt wrote:
>>>> Potentially naive question - the trace point holds enum skb_drop_reason.
>>>> The user space can get the names from BTF. Can we not teach user space
>>>> to generically look up names of enums in BTF?
>>>
>>> That puts a hard requirement to include BTF in builds where it was not
>>> needed before. I really do not want to build with BTF just to get access to
>>> these symbols. And since this is used by the embedded world, and BTF is
>>> extremely bloated, the short answer is "No".
>>
>> Dunno. BTF is there most of the time. It could make the life of
>> majority of the users far more pleasant.
>
> BTF isn't there for a lot of developers working in embedded who use this
> code. Most my users that I deal with have minimal environments, so BTF is a
> showstopper.
One thing we've heard from some embedded folks [1] is that having
kernel BTF loadable as a separate module (rather than embedded in
vmlinux) would help, as there are size limits on vmlinux that they can
workaround by having modules on a different partition. We're hoping
to get that working soon. I was wondering if you see other issues around
BTF adoption for embedded systems that we could put on the to-do list?
Not necessarily for this particular use-case (since there are
complications with trace data as you describe), but just trying to make
sure we can remove barriers to BTF adoption where possible.
Thanks!
Alan
[1]
https://lore.kernel.org/bpf/CAHBbfcUkr6fTm2X9GNsFNqV75fTG=aBQXFx_8Ayk+4hk7heB-g@mail.gmail.com/
>
>>
>> I hope we can at least agree that the current methods of generating
>> the string arrays at C level are... aesthetically displeasing.
>
> I don't know, I kinda like it ;-)
>
> -- Steve
>
On Wed, 4 Oct 2023 22:35:07 +0100
Alan Maguire <[email protected]> wrote:
> One thing we've heard from some embedded folks [1] is that having
> kernel BTF loadable as a separate module (rather than embedded in
> vmlinux) would help, as there are size limits on vmlinux that they can
> workaround by having modules on a different partition. We're hoping
> to get that working soon. I was wondering if you see other issues around
> BTF adoption for embedded systems that we could put on the to-do list?
> Not necessarily for this particular use-case (since there are
> complications with trace data as you describe), but just trying to make
> sure we can remove barriers to BTF adoption where possible.
I wonder how easy is it to create subsets of BTF. For one thing, in the
future we want to be able to trace the arguments of all functions. That is,
tracing all functions at the same time (function tracer) and getting the
arguments within the trace.
This would only require information about functions and their arguments,
which would be very useful. Is BTF easy to break apart? That is, just
generate the information needed for function arguments?
Note, pretty much all functions do not pass structures by values, and this
would not need to know the contents of a pointer to a structure. This would
mean that structure layout information is not needed.
-- Steve
On 04/10/2023 22:43, Steven Rostedt wrote:
> On Wed, 4 Oct 2023 22:35:07 +0100
> Alan Maguire <[email protected]> wrote:
>
>> One thing we've heard from some embedded folks [1] is that having
>> kernel BTF loadable as a separate module (rather than embedded in
>> vmlinux) would help, as there are size limits on vmlinux that they can
>> workaround by having modules on a different partition. We're hoping
>> to get that working soon. I was wondering if you see other issues around
>> BTF adoption for embedded systems that we could put on the to-do list?
>> Not necessarily for this particular use-case (since there are
>> complications with trace data as you describe), but just trying to make
>> sure we can remove barriers to BTF adoption where possible.
>
> I wonder how easy is it to create subsets of BTF. For one thing, in the
> future we want to be able to trace the arguments of all functions. That is,
> tracing all functions at the same time (function tracer) and getting the
> arguments within the trace.
>
> This would only require information about functions and their arguments,
> which would be very useful. Is BTF easy to break apart? That is, just
> generate the information needed for function arguments?
>
There has been a fair bit of effort around this from the userspace side;
the BTF gen efforts were focused around applications carrying the
minimum BTF for their needs, so just the structures needed by the
particular BPF programs rather than the full set of vmlinux structures
for example [1].
Parsing BTF in-kernel to pull out the BTF functions (BTF_KIND_FUNC),
their prototypes (BTF_KIND_FUNC_PROTO) and all associated parameters
would be pretty straightforward I think, especially if you don't need
the structures that are passed via pointers. So if you're starting with
the full BTF, creating a subset for use in tracing would be reasonably
straightforward. My personal preference would always be to have the
full BTF where possible, but if that wasn't feasible on some systems
we'd need to add some options to pahole/libbpf to support such trimming
during the DWARF->BTF translation process.
Alan
[1] https://lore.kernel.org/bpf/[email protected]/
> Note, pretty much all functions do not pass structures by values, and this
> would not need to know the contents of a pointer to a structure. This would
> mean that structure layout information is not needed.
>
> -- Steve
>