Merge branch 'support-associating-bpf-programs-with-struct_ops'

Amery Hung says:

====================
Support associating BPF programs with struct_ops

Hi,

This patchset adds a new BPF command BPF_PROG_ASSOC_STRUCT_OPS to
the bpf() syscall to allow associating a BPF program with a struct_ops.
The command is introduced to address a emerging need from struct_ops
users. As the number of subsystems adopting struct_ops grows, more
users are building their struct_ops-based solution with some help from
other BPF programs. For example, scx_layer uses a syscall program as
a user space trigger to refresh layers [0]. It also uses tracing program
to infer whether a task is using GPU and needs to be prioritized [1]. In
these use cases, when there are multiple struct_ops instances, the
struct_ops kfuncs called from different BPF programs, whether struct_ops
or not needs to be able to refer to a specific one, which currently is
not possible.

The new BPF command will allow users to explicitly associate a BPF
program with a struct_ops map. The libbpf wrapper can be called after
loading programs and before attaching programs and struct_ops.

Internally, it will set prog->aux->st_ops_assoc to the struct_ops
map. struct_ops kfuncs can then get the associated struct_ops struct
by calling bpf_prog_get_assoc_struct_ops() with prog->aux, which can
be acquired from a "__prog" argument. The value of the special
argument will be fixed up by the verifier during verification.

The command conceptually associates the implementation of BPF programs
with struct_ops map, not the attachment. A program associated with the
map will take a refcount of it so that st_ops_assoc always points to a
valid struct_ops struct. struct_ops implementers can use the helper,
bpf_prog_get_assoc_struct_ops to get the pointer. The returned
struct_ops if not NULL is guaranteed to be valid and initialized.
However, it is not guaranteed that the struct_ops is attached. The
struct_ops implementer still need to take steps to track and check the
state of the struct_ops in kdata, if the use case demand the struct_ops
to be attached.

We can also consider support associating struct_ops link with BPF
programs, which on one hand make struct_ops implementer's job easier,
but might complicate libbpf workflow and does not apply to legacy
struct_ops attachment.

[0] https://github.com/sched-ext/scx/blob/main/scheds/rust/scx_layered/src/bpf/main.bpf.c#L557
[1] https://github.com/sched-ext/scx/blob/main/scheds/rust/scx_layered/src/bpf/main.bpf.c#L754
---
v7 -> v8
   - Fix libbpf return (Andrii)
   - Follow kfunc _impl suffic naming convention in selftest (Alexei)
   Link: https://lore.kernel.org/bpf/20251121231352.4032020-1-ameryhung@gmail.com/

v6 -> v7
   - Drop the guarantee that bpf_prog_get_assoc_struct_ops() will always return
     an initialized struct_ops (Martin)
   - Minor misc. changes in selftests
   Link: https://lore.kernel.org/bpf/20251114221741.317631-1-ameryhung@gmail.com/

v5 -> v6
   - Drop refcnt bumping for async callbacks and add RCU annotation (Martin)
   - Fix libbpf bug and update comments (Andrii)
   - Fix refcount bug in bpf_prog_assoc_struct_ops() (AI)
   Link: https://lore.kernel.org/bpf/20251104172652.1746988-1-ameryhung@gmail.com/

v4 -> v5
   - Simplify the API for getting associated struct_ops and dont't
     expose struct_ops map lifecycle management (Andrii, Alexei)
   Link: https://lore.kernel.org/bpf/20251024212914.1474337-1-ameryhung@gmail.com/

v3 -> v4
   - Fix potential dangling pointer in timer callback. Protect
     st_ops_assoc with RCU. The get helper now needs to be paired with
     bpf_struct_ops_put()
   - The command should only increase refcount once for a program
     (Andrii)
   - Test a struct_ops program reused in two struct_ops maps
   - Test getting associated struct_ops in timer callback
   Link: https://lore.kernel.org/bpf/20251017215627.722338-1-ameryhung@gmail.com/

v2 -> v3
   - Change the type of st_ops_assoc from void* (i.e., kdata) to bpf_map
     (Andrii)
   - Fix a bug that clears BPF_PTR_POISON when a struct_ops map is freed
     (Andrii)
   - Return NULL if the map is not fully initialized (Martin)
   - Move struct_ops map refcount inc/dec into internal helpers (Martin)
   - Add libbpf API, bpf_program__assoc_struct_ops (Andrii)
   Link: https://lore.kernel.org/bpf/20251016204503.3203690-1-ameryhung@gmail.com/

v1 -> v2
   - Poison st_ops_assoc when reusing the program in more than one
     struct_ops maps and add a helper to access the pointer (Andrii)
   - Minor style and naming changes (Andrii)
   Link: https://lore.kernel.org/bpf/20251010174953.2884682-1-ameryhung@gmail.com/

---
====================

Link: https://patch.msgid.link/20251203233748.668365-1-ameryhung@gmail.com
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>

2 files changed, 33 insertions, 0 deletions

diff --git a/include/linux/bpf.h b/include/linux/bpf.h
index 6498be4c44f8..28d8d6b7bb1e 100644
--- a/include/linux/bpf.h
+++ b/include/linux/bpf.h
@@ -1739,6 +1739,8 @@ struct bpf_prog_aux {
 		struct rcu_head	rcu;
 	};
 	struct bpf_stream stream[2];
+	struct mutex st_ops_assoc_mutex;
+	struct bpf_map __rcu *st_ops_assoc;
 };
 
 struct bpf_prog {
@@ -2041,6 +2043,9 @@ static inline void bpf_module_put(const void *data, struct module *owner)
 		module_put(owner);
 }
 int bpf_struct_ops_link_create(union bpf_attr *attr);
+int bpf_prog_assoc_struct_ops(struct bpf_prog *prog, struct bpf_map *map);
+void bpf_prog_disassoc_struct_ops(struct bpf_prog *prog);
+void *bpf_prog_get_assoc_struct_ops(const struct bpf_prog_aux *aux);
 u32 bpf_struct_ops_id(const void *kdata);
 
 #ifdef CONFIG_NET
@@ -2088,6 +2093,17 @@ static inline int bpf_struct_ops_link_create(union bpf_attr *attr)
 {
 	return -EOPNOTSUPP;
 }
+static inline int bpf_prog_assoc_struct_ops(struct bpf_prog *prog, struct bpf_map *map)
+{
+	return -EOPNOTSUPP;
+}
+static inline void bpf_prog_disassoc_struct_ops(struct bpf_prog *prog)
+{
+}
+static inline void *bpf_prog_get_assoc_struct_ops(const struct bpf_prog_aux *aux)
+{
+	return NULL;
+}
 static inline void bpf_map_struct_ops_info_fill(struct bpf_map_info *info, struct bpf_map *map)
 {
 }
diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
index f8d8513eda27..84ced3ed2d21 100644
--- a/include/uapi/linux/bpf.h
+++ b/include/uapi/linux/bpf.h
@@ -918,6 +918,16 @@ union bpf_iter_link_info {
  *		Number of bytes read from the stream on success, or -1 if an
  *		error occurred (in which case, *errno* is set appropriately).
  *
+ * BPF_PROG_ASSOC_STRUCT_OPS
+ * 	Description
+ * 		Associate a BPF program with a struct_ops map. The struct_ops
+ * 		map is identified by *map_fd* and the BPF program is
+ * 		identified by *prog_fd*.
+ *
+ * 	Return
+ * 		0 on success or -1 if an error occurred (in which case,
+ * 		*errno* is set appropriately).
+ *
  * NOTES
  *	eBPF objects (maps and programs) can be shared between processes.
  *
@@ -974,6 +984,7 @@ enum bpf_cmd {
 	BPF_PROG_BIND_MAP,
 	BPF_TOKEN_CREATE,
 	BPF_PROG_STREAM_READ_BY_FD,
+	BPF_PROG_ASSOC_STRUCT_OPS,
 	__MAX_BPF_CMD,
 };
 
@@ -1894,6 +1905,12 @@ union bpf_attr {
 		__u32		prog_fd;
 	} prog_stream_read;
 
+	struct {
+		__u32		map_fd;
+		__u32		prog_fd;
+		__u32		flags;
+	} prog_assoc_struct_ops;
+
 } __attribute__((aligned(8)));
 
 /* The description below is an attempt at providing documentation to eBPF