Skip to contentSkip to conversation

bpf: Fix NULL pointer dereference in bpf_sock_from_file

PATCH In review Syeda Mahnur Asif 4 participants 4 messages Aug 20 – Aug 21

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 bpf: Fix NULL pointer dereference in bpf_sock_from_file

bpf_sock_from_file should not dereference a NULL file pointer. KASAN detects a null-ptr-deref when eBPF tracing fentry/fexit programs are attached to points such as __mmap_region and file_map_prot_check kernel functions. This can result in a NULL file pointer flowing from context to the helper.

A minimal check before dereferencing can fix this.

Fixes b60da4955f53 ("bpf: Only provide bpf_sock_from_file with CONFIG_NET")

---
v2:
- Added Reviewed-by from Emil Tsalapatis
- Fixed name in Signed-off-by

 net/core/filter.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
net/core/filter.c +4 -1
index 16845987b244..7c3caae4bafc 100644
@@ -12182,7 +12182,10 @@ const struct bpf_func_proto bpf_skc_to_mptcp_sock_proto = {
 
 BPF_CALL_1(bpf_sock_from_file, struct file *, file)
 {
-	return (unsigned long)sock_from_file(file);
+	if (file)
+		return (unsigned long)sock_from_file(file);
+
+	return (unsigned long)NULL;
 }
 
 BTF_ID_LIST(bpf_sock_from_file_btf_ids)
Signature · 1 line
2.53.0
#1 lore raw
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Pre-existing issues:
- [High] Incomplete validation allows ERR_PTR to cause kernel panic in bpf_sock_from_file.
- [High] Missing NULL pointer checks in BPF helpers taking ARG_PTR_TO_BTF_ID lead to kernel panics. --

--- Patch [1]: [PATCH bpf v2] bpf: Fix NULL pointer dereference in bpf_sock_from_file --- commit a54db842fd2086671fb2b904beab51f691e8cc8b
Author: Syeda Mahnur Asif <[email protected]>

bpf: Fix NULL pointer dereference in bpf_sock_from_file

Adds a NULL check to bpf_sock_from_file to prevent kernel panics when fentry/fexit programs pass a NULL file pointer.

net/core/filter.c const struct bpf_func_proto bpf_skc_to_mptcp_sock_proto = { · line 12186

Quoted 11 lines
net/core/filter.c +2 -1
index 16845987b2449..7c3caae4bafc3 100644
@@ -12182,7 +12182,10 @@ const struct bpf_func_proto bpf_skc_to_mptcp_sock_proto = {
 
 BPF_CALL_1(bpf_sock_from_file, struct file *, file)
 {
-	return (unsigned long)sock_from_file(file);
+	if (file)
+		return (unsigned long)sock_from_file(file);

[Severity: High]
This isn't a bug introduced by this patch, but does this check adequately handle error pointers?

If an eBPF fexit program attaches to a kernel function returning a struct file pointer that fails (e.g., returning ERR_PTR(-ENOENT)), the BPF verifier treats this as a valid PTR_TO_BTF_ID.

Since an error pointer is not NULL, this condition will pass, and sock_from_file() will attempt to dereference it, leading to a crash.

Could we use IS_ERR_OR_NULL(file) instead to catch these cases as well?

Also, while not introduced by this patch, are there similar missing NULL pointer checks in other BPF helpers taking ARG_PTR_TO_BTF_ID?

When fentry/fexit programs pass context arguments, the verifier lacks BTF nullability annotations for most parameters. This means helpers can receive and blindly dereference NULL pointers.

For instance, bpf_task_pt_regs() directly dereferences the task without checking:

kernel/trace/bpf_trace.c:bpf_task_pt_regs() {
	return (unsigned long) task_pt_regs(task);
}

Similarly, __bpf_get_task_stack() calls try_get_task_stack() which will dereference the task:

kernel/bpf/stackmap.c:__bpf_get_task_stack() {
	if (!try_get_task_stack(task))
		return -EFAULT;
}

And bpf_xdp_get_buff_len() accesses the xdp buffer directly:

net/core/filter.c:bpf_xdp_get_buff_len() {
	return xdp_get_buff_len(xdp);
}

Do these helpers need similar runtime validation to prevent panics in the same execution context?

Quoted 5 lines
net/core/filter.c +2 -0
+
+	return (unsigned long)NULL;
 }
 
 BTF_ID_LIST(bpf_sock_from_file_btf_ids)
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
#2 lore raw
In reply to Syeda Mahnur Asif
Quoted 25 lines · Syeda Mahnur Asif

bpf_sock_from_file should not dereference a NULL file pointer. KASAN detects a null-ptr-deref when eBPF tracing fentry/fexit programs are attached to points such as __mmap_region and file_map_prot_check kernel functions. This can result in a NULL file pointer flowing from context to the helper.

A minimal check before dereferencing can fix this.

Fixes b60da4955f53 ("bpf: Only provide bpf_sock_from_file with CONFIG_NET")

---
v2:
- Added Reviewed-by from Emil Tsalapatis
- Fixed name in Signed-off-by

 net/core/filter.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
net/core/filter.c +0 -0
index 16845987b244..7c3caae4bafc 100644

net/core/filter.c const struct bpf_func_proto bpf_skc_to_mptcp_sock_proto = { · line 12186

Quoted 6 lines
net/core/filter.c +2 -1
@@ -12182,7 +12182,10 @@ const struct bpf_func_proto bpf_skc_to_mptcp_sock_proto = {
 
 BPF_CALL_1(bpf_sock_from_file, struct file *, file)
 {
-       return (unsigned long)sock_from_file(file);
+       if (file)
+               return (unsigned long)sock_from_file(file);

given this is a legacy ARG_PTR_TO_BTF_ID, it can be not just NULL, but also a) small value bogus "pointer" resulting from NULL + offset calculation and/or b) just random garbage casted to struct file pointer with bpf_core_cast() (maybe this hole we've closed with explicitly untrusted, not sure).

Either way, this is not a sufficient fix at least.

But I wonder if the proper fix is actually to mark this (and other similar) argument as explicitly requiring trusted PTR_TO_BTF_ID?

Thoughts?

Quoted 8 lines
net/core/filter.c +2 -1
+
+       return (unsigned long)NULL;
 }
 
 BTF_ID_LIST(bpf_sock_from_file_btf_ids)
--

2.53.0

#3 lore raw
Quoted 39 lines · Andrii Nakryiko
Quoted 25 lines · Syeda Mahnur Asif

bpf_sock_from_file should not dereference a NULL file pointer. KASAN detects a null-ptr-deref when eBPF tracing fentry/fexit programs are attached to points such as __mmap_region and file_map_prot_check kernel functions. This can result in a NULL file pointer flowing from context to the helper.

A minimal check before dereferencing can fix this.

Fixes b60da4955f53 ("bpf: Only provide bpf_sock_from_file with CONFIG_NET")

---
v2:
- Added Reviewed-by from Emil Tsalapatis
- Fixed name in Signed-off-by

 net/core/filter.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
net/core/filter.c +0 -0
index 16845987b244..7c3caae4bafc 100644

net/core/filter.c const struct bpf_func_proto bpf_skc_to_mptcp_sock_proto = { · line 12186

Quoted 6 lines
net/core/filter.c +2 -1
@@ -12182,7 +12182,10 @@ const struct bpf_func_proto bpf_skc_to_mptcp_sock_proto = {
 
 BPF_CALL_1(bpf_sock_from_file, struct file *, file)
 {
-       return (unsigned long)sock_from_file(file);
+       if (file)
+               return (unsigned long)sock_from_file(file);

given this is a legacy ARG_PTR_TO_BTF_ID, it can be not just NULL, but also a) small value bogus "pointer" resulting from NULL + offset calculation and/or b) just random garbage casted to struct file pointer with bpf_core_cast() (maybe this hole we've closed with explicitly untrusted, not sure).

Quoted 7 lines

Either way, this is not a sufficient fix at least.

But I wonder if the proper fix is actually to mark this (and other similar) argument as explicitly requiring trusted PTR_TO_BTF_ID?

Thoughts?

+1, should bite the bullet and accelerate deprecation of legacy PTR_TO_BTF_ID. We can mark specific arguments / fields in structs as trusted where there is a legitimate use case and the kernel guarantees their lifetime appropriately.

Quoted 8 lines
Quoted 8 lines
net/core/filter.c +2 -1
+
+       return (unsigned long)NULL;
 }
 
 BTF_ID_LIST(bpf_sock_from_file_btf_ids)
--

2.53.0