[2/2] libbpf: BPF programs dynamic loading and attaching

BPF programs designated as dynamically loaded can be loaded and
attached independently after the initial bpf_object loading and
attaching.

These programs can also be reloaded and reattached multiple times,
enabling more flexible management of a resident BPF program set.

A key motivation for this feature is to reduce load times for
utilities that include hundreds of BPF programs. When the selection
of a resident BPF program set cannot be determined at the time of
bpf_object loading and attaching, all BPF programs would otherwise
need to be marked as autoload, leading to unnecessary overhead.
This patch addresses that inefficiency.

Signed-off-by: Slava Imameev <slava.imameev@crowdstrike.com>
---
 tools/lib/bpf/libbpf.c                        | 144 +++++++++++++++--
 tools/lib/bpf/libbpf.h                        |   5 +-
 tools/lib/bpf/libbpf.map                      |   2 +
 .../selftests/bpf/prog_tests/dynamicload.c    | 145 ++++++++++++++++++
 .../selftests/bpf/prog_tests/load_type.c      |  61 ++++++++
 .../selftests/bpf/progs/test_dynamicload.c    |  31 ++++
 .../selftests/bpf/progs/test_load_type.c      |   8 +
 7 files changed, 385 insertions(+), 11 deletions(-)
 create mode 100644 tools/testing/selftests/bpf/prog_tests/dynamicload.c
 create mode 100644 tools/testing/selftests/bpf/progs/test_dynamicload.c

On Tue, Jan 28, 2025 at 3:08 PM Martin Kelly
<martin.kelly@crowdstrike.com> wrote:
>
> On Fri, 2025-01-24 at 10:31 -0800, Andrii Nakryiko wrote:
> > > On Wed, Jan 22, 2025 at 1:53 PM Slava Imameev
> > > <slava.imameev@crowdstrike.com> wrote:
> > > > >
> > > > > BPF programs designated as dynamically loaded can be loaded and
> > > > > attached independently after the initial bpf_object loading and
> > > > > attaching.
> > > > >
> > > > > These programs can also be reloaded and reattached multiple
> > > > > times,
> > > > > enabling more flexible management of a resident BPF program
> > > > > set.
> > > > >
> > > > > A key motivation for this feature is to reduce load times for
> > > > > utilities that include hundreds of BPF programs. When the
> > > > > selection
> > > > > of a resident BPF program set cannot be determined at the time
> > > > > of
> > > > > bpf_object loading and attaching, all BPF programs would
> > > > > otherwise
> > > > > need to be marked as autoload, leading to unnecessary overhead.
> > > > > This patch addresses that inefficiency.
> > >
> > > Can you elaborate on why it's impossible to determine which BPF
> > > programs should be loaded before BPF object load step?
> > >
>
> The main use case for this patch is large applications that need to
> dynamically load/unload BPF programs. Our specific use case is a
> continuously-running security application with a dynamically-
> reconfigurable feature set. As part of that reconfiguration, BPF
> programs may get loaded/unloaded on-the-fly. Restarting the entire
> application during reconfiguration is undesirable, as critical state
> data can be lost and loading hundreds of BPF programs is time-
> consuming.

Thanks for the details (and sorry for late reply, been traveling lately)!

I don't want to complicate bpf_object internals with this third
autoload state, as this creates tons of non-obvious gotchas that have
to constantly be checked whenever any new feature is added to
bpf_object (or even during refactorings). It does seem like you have a
pretty complicated use case, and so maybe some of the alternatives
would be just fine for you.

I see two ways forward for you. Either you can break apart your BPF
object of ~100 BPF programs into more independent BPF objects (seeing
that programs can be independently loaded/unloaded depending on
configuration, seems like you do have a bunch of logic independence,
right?). I assume shared BPF maps are the biggest reason to keep all
those programs together in one BPF object. To share BPF maps between
multiple BPF objects libbpf provides two complementary interfaces:

  - bpf_map__reuse_fd() for manual control
  - BPF map pinning (could be declarative or manual)

This way you can ensure that all BPF objects would use the same BPF
map, where necessary.

Alternatively, we can look at this problem as needing libbpf to only
prepare BPF program code (doing all the relocations and stuff like
that), but then application actually taking care of loading/unloading
BPF program with bpf_prog_load() outside of bpf_object abstraction.
I've had an almost ready patches splitting bpf_object__load() into two
steps: bpf_object__prepare() and bpf_object__load() after that.
"prepare" step would create BPF maps, load BTF information, perform
necessary relocations and arrive at final state of BPF program code
(which you can get with bpf_program__insns() API), but stopping just
short of actually doing bpf_prog_load() step.

This seems like it would solve your problem as well. You'd use libbpf
to do all the low-level ELF processing and relocation, but then take
over managing BPF program lifetime. Loading/unloading as you see fit,
including in parallel.

Is this something that would work for you?

>
> The above points apply more generically to *any* application that
> requires dynamic loading/unloading. Reconfiguration can be done via a
> restart, but that has drawbacks:
> (a) Losing any non-persistent application state on restart. In our
> case, this creates a lapse in security that could be exploited by
> adversaries.
> (b) In applications with many programs, load+attach can take a long
> time. We measured load+attach of ~100 BPF programs taking ~10 seconds
> when done with current libbpf serially. Dynamically loading only the
> programs needed avoids wasting memory and CPU cycles.
> (c) The application itself might take a long time to restart, separate
> from the BPF load/attach time. By loading dynamically, the BPF programs
> can take effect much sooner and avoid wasted restart cycles.
>
> This patch set also permits loading BPF programs in parallel if the
> application wishes. We tested parallel loading with 200+ BPF programs
> and found the load time dropped from 18 seconds to 5 seconds when done
> in parallel on a 6.8 kernel.

bpf_object is intentionally single-threaded, so I don't think we'll be
supporting parallel BPF program loading in the paradigm of bpf_object
(but see the bpf_object__prepare() proposal). Even from API standpoint
this is problematic with logging and log buffers basically assuming
single-threaded execution of BPF program loading.

All that could be changed or worked around, but your use case is not
really a typical case, so I'm a bit hesitant at this point.

>
> In the future, this approach could also allow maps to not be
> autoloaded, further saving on memory if no program needs the underlying
> map.
>
> In summary, we believe dynamic loading of BPF programs is an important
> capability that will improve the performance of CrowdStrike's security
> applications as well as being useful to other applications that want to
> avoid restarts.
> >

On Wed, 2025-02-05 at 14:33 -0800, Andrii Nakryiko wrote:
> > > > 
> > > > I see two ways forward for you. Either you can break apart your
> > > > BPF
> > > > object of ~100 BPF programs into more independent BPF objects >
> > > > > (seeing
> > > > that programs can be independently loaded/unloaded depending on
> > > > configuration, seems like you do have a bunch of logic > >
> > > > independence,
> > > > right?). I assume shared BPF maps are the biggest reason to
> > > > keep > > all
> > > > those programs together in one BPF object. To share BPF maps >
> > > > > between
> > > > multiple BPF objects libbpf provides two complementary
> > > > interfaces:
> > > > 
> > > >   - bpf_map__reuse_fd() for manual control
> > > >   - BPF map pinning (could be declarative or manual)
> > > > 
> > > > This way you can ensure that all BPF objects would use the same
> > > > BPF
> > > > map, where necessary.
> > > > 

I think this approach *could* work but could easily become complex for
us because we'd need to track all the dependencies between programs and
maps, and anything missed could lead to difficult refcount bugs.

Further, splitting into objects incurs some performance and memory cost
because bpf_object__load_vmlinux_btf will be called for each object,
and there's currently no way to share BTF data across the objects.
Having a single BPF object avoids this issue. Potentially, libbpf could
cache some BTF data to make lessen the impact.

> > > > Alternatively, we can look at this problem as needing libbpf to
> > > > > > only
> > > > prepare BPF program code (doing all the relocations and stuff
> > > > like
> > > > that), but then application actually taking care of > >
> > > > loading/unloading
> > > > BPF program with bpf_prog_load() outside of bpf_object
> > > > abstraction.
> > > > I've had an almost ready patches splitting bpf_object__load()
> > > > into > > > two
> > > > steps: bpf_object__prepare() and bpf_object__load() after that.
> > > > "prepare" step would create BPF maps, load BTF information,
> > > > perform
> > > > necessary relocations and arrive at final state of BPF program
> > > > code
> > > > (which you can get with bpf_program__insns() API), but stopping
> > > > > > just
> > > > short of actually doing bpf_prog_load() step.
> > > > 
> > > > This seems like it would solve your problem as well. You'd use
> > > > > > libbpf
> > > > to do all the low-level ELF processing and relocation, but then
> > > > > > take
> > > > over managing BPF program lifetime. Loading/unloading as you
> > > > see > > fit,
> > > > including in parallel.
> > > > 
> > > > Is this something that would work for you?
> > > > 

I think this API could work, though I think we would need a few other
modifications as well in order to correctly handle program/map
dependencies and account for relocations. At a high level, I think we'd
need something that includes:

1) A way to associate each BPF program with all the maps it will use
(association of struct bpf_program * --> list of struct bpf_map * in
some form). This is so that we can load/unload associated maps when we
load/unload a program.

2) An API to create a BPF map, in case a new map needs to be loaded
after initial startup.

3) An API to allow unloading a map while keeping map->fd reserved. This
is important because the fd value is used by BPF program instructions,
so without something like this, we'd have to redo the relocation
process for any other BPF programs that access this map (and thus
reload those programs too). This API could be implemented by dup'ing a
placeholder fd.

Alternatively, if libbpf could automatically refcount maps across
multiple BPF objects to load/unload them on demand, then all of the
above work could happen behind the scenes. This would be similar to the
other approach you mentioned, but with libbpf doing the refcounting
heavy lifting instead of leaving that to each application, thus more
robust and elegant. This would mean changing libbpf to (a) synchronize
access to some map functions and (b) allowing struct bpf_map * to be
shared across BPF objects. Perhaps a concept of a "collection of BPF
objects" might allow for this.

> > > > > > > > 
> > > > > > > > This patch set also permits loading BPF programs in
> > > > > > > > parallel if > > > > the
> > > > > > > > application wishes. We tested parallel loading with
> > > > > > > > 200+ BPF > > > > > > programs
> > > > > > > > and found the load time dropped from 18 seconds to 5
> > > > > > > > seconds > > > > when > > done
> > > > > > > > in parallel on a 6.8 kernel.
> > > > 
> > > > bpf_object is intentionally single-threaded, so I don't think
> > > > we'll > > > be
> > > > supporting parallel BPF program loading in the paradigm of > >
> > > > bpf_object
> > > > (but see the bpf_object__prepare() proposal). Even from API > >
> > > > > standpoint
> > > > this is problematic with logging and log buffers basically
> > > > assuming
> > > > single-threaded execution of BPF program loading.
> > > > 
> > > > All that could be changed or worked around, but your use case
> > > > is > > not
> > > > really a typical case, so I'm a bit hesitant at this point.
> > > > 
> > > > > > > > 

I can understand where you're coming from if no one else has mentioned
a use case like this. We can do parallel loading by splitting our
programs into BPF objects, but unless the objects are split very
evenly, this results in less optimal load time. For example, if 100
programs are split into 2 objects and one object has 80 programs while
the other has 20, then the one with 80 programs creates a bottleneck.
>

On Wed, Feb 12, 2025 at 2:31 PM Martin Kelly
<martin.kelly@crowdstrike.com> wrote:
>
> On Mon, 2025-02-10 at 16:06 -0800, Andrii Nakryiko wrote:
> > > Tracking associated maps for a program is not necessary. As long as
> > > the last BPF program using the BPF map is unloaded, the kernel will
> > > automatically free not-anymore-referenced BPF map. Note that
> > > bpf_object itself will keep FDs for BPF maps, so you'd need to make
> > > sure to do bpf_object__close() to release those references.
> > >
> > > But if you are going to ask to re-create BPF maps next time BPF
> > > program is loaded... Well, I'll say you are asking for a bit too >
> > > much,
> > > tbh. If you want to be *that* sophisticated, it shouldn't be too
> > > hard
> > > for you to get all this information from BPF program's
> > > instructions.
> > >
>
> We really are that sophisticated (see below for more details). We could
> scan program instructions, but we'd then tie our logic to BPF
> implementation details and duplicate logic already present in libbpf
> (https://elixir.bootlin.com/linux/v6.13.2/source/tools/lib/bpf/libbpf.c#L6087
> ). Obviously this *can* be done but it's not at all ideal from an
> application perspective.
>

I agree it's not ideal, but it's also not some complicated and
bound-to-be-changed logic. What you point out in libbpf source code is
a bit different thing, reality is much simpler. Only so-called ldimm64
instruction (BPF_LD | BPF_IMM | BPF_DW opcode) can be referencing map
FD, so analysing this is borderline trivial. And this is part of BPF
ISA, so not going to change.

We need to double check, but I think libbpf doesn't use FD_ARRAY
approach, unless you are using light skeleton, so if you don't you
don't even have to worry about FD_ARRAY thing.

>
> > > > >
> > > bpf_object is the unit of coherence in libbpf, so I don't see us
> > > refcounting maps between bpf_objects. Kernel is doing refcounting
> > > based on FDs, so see if you can use that.
> > >
>
> I can understand that. That said, I think if there's no logic across
> objects, and bpf_object access is not thread-safe, it puts us into a
> tough situation:
> - Complex refcounting, code scanning, etc to keep consistency when
> manipulating maps used by multiple programs.
> - Parallel loading not being well-balanced, if we split programs across
> objects.
>
> We could alternatively write our own custom loader, but then we’d have
> to duplicate much of the useful logic that libbpf already implements:
> skeleton generation, map/program association, embedding programs into
> ELFs, loading logic and kernel probing, etc. We’d like some way to
> handle dynamic/parallel loading without having to replicate all the
> advantages libbpf grants us.
>

Yeah, I can understand that as well, but bpf_object's single-threaded
design and the fact that bpf_object__load is kind of the final step
where programs are loaded (or not) is pretty backed in. I don't see
bpf_object becoming multi-threaded. The dynamic program
loading/unloading/loading again is something that I can't yet justify,
tbh.

So the best I can propose you is to use libbpf's skeleton and
bpf_object concept for, effectively, ELF handling, relocations, all
the preparations up to loading BPF programs. And after that you can
take over loading and handling program lifetime outside of bpf_object.

Dynamic map creation after bpf_object__load() I think is completely
outside of the scope and you'll have to solve this problem for
yourself. I would point out, though, that internally libbpf already
switched to sort-of pre-creating stable FDs for maps before they are
actually created in the kernel. So it's conceivable that we can have
more granularity in bpf_object preparation. I.e., first step would be
to parse ELF and handle relocations, prepare everything. After that we
can have a step to create maps, and then another one to create
programs. Usually people would do all that, but you can stop right
before maps creation or before program creation, whatever fits your
use case better.

The key is that program instructions will be final and won't need
adjustments regardless of maps actually being created or not. FDs, as
I mentioned, are stable regardless.

So, not ideal for your (very complicated) use case, but you still
avoid dealing with all the ELF and relocation stuff (which is the
annoying and rather complicated part, and I can see no one wanting to
reimplement that). Map and program creation is relatively
straightforward matters compared to that.

> > >
> > >
> > > Is 100 just a nicely looking rather large number, or do you really
> > > have 100 different BPF programs? Why so many and are they really
> > > all
> > > unique?
> > >
> > > Asking because if it's just a way to attach BPF program doing more
> > > or
> > > less uniform set of actions for different hooks, then perhaps there
> > > are better ways to do this without having to duplicating BPF
> > > programs
> > > so much (like BPF cookie, multi-kprobes, etc, etc)
>
> 100 is not an arbitrary number; we have that and higher (~200 is a good
> current estimate, and that grows as new product features are added).
> The programs are really doing different things. We also have to support
> a wide range of kernels, handling cases like: "on this kernel range,
> trampolines aren't supported, so use kretprobes with a context map for
> function args instead of fexit, but on newer kernels just use an fexit
> hook."

Yes, this is typical, and bpf_program__set_autoload() and
bpf_map__set_autocreate() are meant to handle that. It's the program
loading after bpf_object load is what is not supported.

>
> The use case here is that our security monitoring agent leverages eBPF
> as its foundational technology to gather telemetry from the kernel. As
> part of that, we hook many different kernel subsystems (process,
> memory, filesystem, network, etc), tying them together and tracking
> with maps. So we legitimately have a very large number of programs all
> doing different work. For products of this scale, it increases security
> and performance to load this set of programs and their maps in an
> optimized, parallel fashion and subsequently change the loaded set of
> programs and maps dynamically without disturbing the rest of the
> application.

Yes, makes sense. You'll need to decide for yourself if it's actually
more meaningful to split those 200 programs into independent
bpf_objects by features, and be rigorous about sharing state (maps)
through bpf_map__reuse_fd(), which would allow to parallelize loading
within confines of existing libbpf APIs. Or you can be a bit more
low-level with program loading outside of bpf_object API, as I
described above.

This reply was resent as the previous email had a missing In-Reply-To
in the header.

> > > On Mon, 2025-02-10 at 16:06 -0800, Andrii Nakryiko wrote:
> > > > Tracking associated maps for a program is not necessary. As long as
> > > > the last BPF program using the BPF map is unloaded, the kernel will
> > > > automatically free not-anymore-referenced BPF map. Note that
> > > > bpf_object itself will keep FDs for BPF maps, so you'd need to make
> > > > sure to do bpf_object__close() to release those references.
> > > >
> > > > But if you are going to ask to re-create BPF maps next time BPF
> > > > program is loaded... Well, I'll say you are asking for a bit too >
> > > > much,
> > > > tbh. If you want to be *that* sophisticated, it shouldn't be too
> > > > hard
> > > > for you to get all this information from BPF program's
> > > > instructions.
> > > >
> >
> > We really are that sophisticated (see below for more details). We could
> > scan program instructions, but we'd then tie our logic to BPF
> > implementation details and duplicate logic already present in libbpf
> > implementation details and duplicate logic already present in libbpf
> > (https://elixir.bootlin.com/linux/v6.13.2/source/tools/lib/bpf/libbpf.c#L=
6087
> > ). Obviously this *can* be done but it's not at all ideal from an
> > application perspective.
> >
>
>
> I agree it's not ideal, but it's also not some complicated and
> bound-to-be-changed logic. What you point out in libbpf source code is
> a bit different thing, reality is much simpler. Only so-called ldimm64
> instruction (BPF_LD | BPF_IMM | BPF_DW opcode) can be referencing map
> FD, so analysing this is borderline trivial. And this is part of BPF
> ISA, so not going to change.


Our approach is to associate an array of maps as a property with each
BPF program, this property is initialised at the relocation stage.
So, we do not need to parse BPF program instructions. Instead, we rely on
recorded relocations. I think this is a more robust and clean solution with
advantage of all code in the same place and being at the higher level of
abstraction with a relocation table.

The mainline libbpf keeps array of maps for a bpf_object, we extended
this by adding an array of maps associated with each bpf_program.

For example, a code excerpt, from our development branch, which associates
a map with bpf_program at relocation phase:

     insn[0].src_reg = BPF_PSEUDO_MAP_FD;
     insn[0].imm = map->fd;
     err = bpf_program__add_map(prog, map);


> > > > > >
> > > > bpf_object is the unit of coherence in libbpf, so I don't see us
> > > > refcounting maps between bpf_objects. Kernel is doing refcounting
> > > > based on FDs, so see if you can use that.
> > > >
> >
> > I can understand that. That said, I think if there's no logic across
> > objects, and bpf_object access is not thread-safe, it puts us into a
> > tough situation:
> > - Complex refcounting, code scanning, etc to keep consistency when
> > manipulating maps used by multiple programs.
> > - Parallel loading not being well-balanced, if we split programs across
> > objects.
> >
> > We could alternatively write our own custom loader, but then we’d have
> > to duplicate much of the useful logic that libbpf already implements:
> > skeleton generation, map/program association, embedding programs into
> > ELFs, loading logic and kernel probing, etc. We’d like some way to
> > handle dynamic/parallel loading without having to replicate all the
> > advantages libbpf grants us.
> >
>
>
> Yeah, I can understand that as well, but bpf_object's single-threaded
> design and the fact that bpf_object__load is kind of the final step
> where programs are loaded (or not) is pretty backed in. I don't see
> bpf_object becoming multi-threaded.


We understood this, but the current bpf_object design allowed us to use
it in a multithreaded environment with minor modification for bpf_program
load.

We understand that the design choice of libbpf being single threaded
is unlikely to be reconsidered.


> > > > > >
> > > > bpf_object is the unit of coherence in libbpf, so I don't see us
> > > > refcounting maps between bpf_objects. Kernel is doing refcounting
> > > > based on FDs, so see if you can use that.
> > > >
> >
> > I can understand that. That said, I think if there's no logic across
> > objects, and bpf_object access is not thread-safe, it puts us into a
> > tough situation:
> > - Complex refcounting, code scanning, etc to keep consistency when
> > manipulating maps used by multiple programs.
> > - Parallel loading not being well-balanced, if we split programs across
> > objects.
> >
> > We could alternatively write our own custom loader, but then we’d have
> > to duplicate much of the useful logic that libbpf already implements:
> > skeleton generation, map/program association, embedding programs into
> > ELFs, loading logic and kernel probing, etc. We’d like some way to
> > handle dynamic/parallel loading without having to replicate all the
> > advantages libbpf grants us.
> >
>
>
> Yeah, I can understand that as well, but bpf_object's single-threaded
> design and the fact that bpf_object__load is kind of the final step
> where programs are loaded (or not) is pretty backed in. I don't see
> bpf_object becoming multi-threaded. The dynamic program
> loading/unloading/loading again is something that I can't yet justify,
> tbh.
>
>
> So the best I can propose you is to use libbpf's skeleton and
> bpf_object concept for, effectively, ELF handling, relocations, all
> the preparations up to loading BPF programs. And after that you can
> take over loading and handling program lifetime outside of bpf_object.
>
>
> Dynamic map creation after bpf_object__load() I think is completely
> outside of the scope and you'll have to solve this problem for
> yourself. I would point out, though, that internally libbpf already
> switched to sort-of pre-creating stable FDs for maps before they are
> actually created in the kernel. So it's conceivable that we can have
> more granularity in bpf_object preparation. I.e., first step would be
> to parse ELF and handle relocations, prepare everything. After that we
> can have a step to create maps, and then another one to create
> programs. Usually people would do all that, but you can stop right
> before maps creation or before program creation, whatever fits your
> use case better.
>
>
> The key is that program instructions will be final and won't need
> adjustments regardless of maps actually being created or not. FDs, as
> I mentioned, are stable regardless.


We used this in our design, so we did not need to scan BPF program
instructions to fix map's fds referenced by instructions from a dynamically
loaded bpf_program with dynamically created maps.


> >
> > The use case here is that our security monitoring agent leverages eBPF
> > as its foundational technology to gather telemetry from the kernel. As
> > part of that, we hook many different kernel subsystems (process,
> > memory, filesystem, network, etc), tying them together and tracking
> > with maps. So we legitimately have a very large number of programs all
> > doing different work. For products of this scale, it increases security
> > and performance to load this set of programs and their maps in an
> > optimized, parallel fashion and subsequently change the loaded set of
> > programs and maps dynamically without disturbing the rest of the
> > application.
>
>
> Yes, makes sense. You'll need to decide for yourself if it's actually
> more meaningful to split those 200 programs into independent
> bpf_objects by features, and be rigorous about sharing state (maps)
> through bpf_map__reuse_fd(), which would allow to parallelize loading
> within confines of existing libbpf APIs. Or you can be a bit more
> low-level with program loading outside of bpf_object API, as I
> described above.

Yes, this can be one of the ways to share bpf maps across multiple
bpf_objects and use existing libbpf for parallel bps programs loading,
if we want to keep a full libbpf compatibility, but at a cost of
complicating design, as we need to convert a single bpf_object model
to multiple bpf_objects with a new layer that manages these bpf_objects.

In our case, as a bpf_program can map to multiple features, which can be
modified independently, and to achieve an even load balancing across
multiple threads, it would be probably one bpf_program for a bpf_object.