Linux kernel: BPF verifier divide-by-zero on empty flexible array struct

In the Linux kernel, the following vulnerability has been resolved:

bpf: Fix divide-by-zero in btf_struct_walk()

When an access goes past the struct and the last member is a flexible array, btf_struct_walk() folds the offset back into a single element with (off - moff) % t->size, but never checks that the element type has a size.

BTF takes an empty struct, so this in program BTF

/* event could be empty */
struct event {

#ifdef HAVE_TIMESTAMP __u64 ts; #endif };

struct batch {
	int nr;
	struct event events[];
};

divides by zero at prog load time. Getting there needs a PTR_TO_BTF_ID that is not MEM_ALLOC, e.g. a plain read of a local kptr stashed in a map from a sleepable program.

Oops: divide error: 0000 [#1] SMP KASAN PTI RIP: 0010:btf_struct_walk+0x53f/0x1570 Call Trace: <TASK> btf_struct_access+0x42a/0xcd0 check_ptr_to_btf_access+0x4dc/0x1160 check_mem_access+0x3a45/0x8740 check_load_mem+0x36a/0xd10 do_check_common+0x3ef0/0xb210 bpf_check+0x6d3b/0x8580 bpf_prog_load+0xf7c/0x2720 __sys_bpf+0xa83/0x3690 __x64_sys_bpf+0xc7/0x150 x64_sys_call+0x1f3f/0x27e0 do_syscall_64+0xe5/0x610 entry_SYSCALL_64_after_hwframe+0x76/0x7e </TASK>

Reject a zero-sized element type. The fixed array path in the same function already bails out on the same thing:

btf_struct_walk()
...
	/* skip empty array */
	if (moff == mtrue_end)
		continue;

	msize /= total_nelems;