bpf: Avoid RCU context warning when unpinning htab with internal structs

This small patchset is about avoid RCU context warning when unpinning htab with internal structs (timer, workqueue, or task_work).

v3:
- fix nit (Yonghong Song)
- add Acked-by: Yonghong Song <[email protected]>

v2:
- rename bpf_free_inode() to bpf_destroy_inode() (Andrii)
https://lore.kernel.org/all/[email protected]/

v1:
https://lore.kernel.org/all/[email protected]/

---
KaFai Wan (2):
bpf: Avoid RCU context warning when unpinning htab with internal
structs
selftests/bpf: Add test for unpinning htab with internal timer struct

 kernel/bpf/inode.c                            |  4 +--
 .../selftests/bpf/prog_tests/pinning_htab.c   | 36 +++++++++++++++++++
 .../selftests/bpf/progs/test_pinning_htab.c   | 25 +++++++++++++
 3 files changed, 63 insertions(+), 2 deletions(-)
 create mode 100644 tools/testing/selftests/bpf/prog_tests/pinning_htab.c
 create mode 100644 tools/testing/selftests/bpf/progs/test_pinning_htab.c

2.43.0

When unpinning a BPF hash table (htab or htab_lru) that contains internal structures (timer, workqueue, or task_work) in its values, a BUG warning is triggered:

 BUG: sleeping function called from invalid context at kernel/bpf/hashtab.c:244
 in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 14, name: ksoftirqd/0
 ...

The issue arises from the interaction between BPF object unpinning and RCU callback mechanisms:
1. BPF object unpinning uses ->free_inode() which schedules cleanup via call_rcu(), deferring the actual freeing to an RCU callback that executes within the RCU_SOFTIRQ context.
2. During cleanup of hash tables containing internal structures, htab_map_free_internal_structs() is invoked, which includes cond_resched() or cond_resched_rcu() calls to yield the CPU during potentially long operations.

However, cond_resched() or cond_resched_rcu() cannot be safely called from atomic RCU softirq context, leading to the BUG warning when attempting to reschedule.

Fix this by changing from ->free_inode() to ->destroy_inode() and rename bpf_free_inode() to bpf_destroy_inode() for BPF objects (prog, map, link). This allows direct inode freeing without RCU callback scheduling, avoiding the invalid context warning.

---
 kernel/bpf/inode.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

index f90bdcc0a047..81780bcf8d25 100644
@@ -775,7 +775,7 @@ static int bpf_show_options(struct seq_file *m, struct dentry *root)
 	return 0;
 }
 
-static void bpf_free_inode(struct inode *inode)
+static void bpf_destroy_inode(struct inode *inode)
 {
 	enum bpf_type type;
 

@@ -790,7 +790,7 @@ const struct super_operations bpf_super_ops = {
 	.statfs		= simple_statfs,
 	.drop_inode	= inode_just_drop,
 	.show_options	= bpf_show_options,
-	.free_inode	= bpf_free_inode,
+	.destroy_inode	= bpf_destroy_inode,
 };
 
 enum {

2.43.0

Add test to verify that unpinning hash tables containing internal timer structures does not trigger context warnings.

Each subtest (timer_prealloc and timer_no_prealloc) can trigger the context warning when unpinning, but the warning cannot be triggered twice within a short time interval (a HZ), which is expected behavior.

---
 .../selftests/bpf/prog_tests/pinning_htab.c   | 36 +++++++++++++++++++
 .../selftests/bpf/progs/test_pinning_htab.c   | 25 +++++++++++++
 2 files changed, 61 insertions(+)
 create mode 100644 tools/testing/selftests/bpf/prog_tests/pinning_htab.c
 create mode 100644 tools/testing/selftests/bpf/progs/test_pinning_htab.c

new file mode 100644
index 000000000000..16bd74be3dbe
@@ -0,0 +1,36 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#include <test_progs.h>
+#include "test_pinning_htab.skel.h"
+
+static void unpin_map(const char *map_name, const char *pin_path)
+{
+	struct test_pinning_htab *skel;
+	struct bpf_map *map;
+	int err;
+
+	skel = test_pinning_htab__open_and_load();
+	if (!ASSERT_OK_PTR(skel, "skel open_and_load"))
+		return;
+
+	map = bpf_object__find_map_by_name(skel->obj, map_name);
+	if (!ASSERT_OK_PTR(map, "bpf_object__find_map_by_name"))
+		goto out;
+
+	err = bpf_map__pin(map, pin_path);
+	if (!ASSERT_OK(err, "bpf_map__pin"))
+		goto out;
+
+	err = bpf_map__unpin(map, pin_path);
+	ASSERT_OK(err, "bpf_map__unpin");
+out:
+	test_pinning_htab__destroy(skel);
+}
+
+void test_pinning_htab(void)
+{
+	if (test__start_subtest("timer_prealloc"))
+		unpin_map("timer_prealloc", "/sys/fs/bpf/timer_prealloc");
+	if (test__start_subtest("timer_no_prealloc"))
+		unpin_map("timer_no_prealloc", "/sys/fs/bpf/timer_no_prealloc");
+}

new file mode 100644
index 000000000000..ae227930c73c
@@ -0,0 +1,25 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#include "vmlinux.h"
+#include <bpf/bpf_helpers.h>
+
+char _license[] SEC("license") = "GPL";
+
+struct timer_val {
+	struct bpf_timer timer;
+};
+
+struct {
+	__uint(type, BPF_MAP_TYPE_HASH);
+	__type(key, __u32);
+	__type(value, struct timer_val);
+	__uint(max_entries, 1);
+} timer_prealloc SEC(".maps");
+
+struct {
+	__uint(type, BPF_MAP_TYPE_HASH);
+	__type(key, __u32);
+	__type(value, struct timer_val);
+	__uint(max_entries, 1);
+	__uint(map_flags, BPF_F_NO_PREALLOC);
+} timer_no_prealloc SEC(".maps");

2.43.0

It was difficult to decipher the above paragraph.
I think you're referring to throttling logic in __might_resched().

I kept it as-is while applying, but in the future either skip such low level implementation detail or expand it.
Here you could have said that the error is printed by __might_resched() which prints it once a HZ.

OK, I understand. Thanks for your feedback and guidance.

Thanks,
KaFai