ebpf can allow sub-word load results to be used as 64-bit pointers
When I modify the bpf_cubic_init() function from
https://github.com/aroodgar/bpf-tcp-congestion-control-algorithm
to read:
SEC("struct_ops/bpf_cubic_init")
void BPF_PROG(bpf_cubic_init, struct sock *sk)
{
asm volatile("r2 = *(u16*)(r1 + 0)"); // verifier should demand u64
asm volatile("*(u32 *)(r2 +1504) = 0"); // 1280 in some configs
}
the verifier accepts it, but the second line crashes when it runs during a TCP connect() because the "*(u16*)" in the load from context yields only the low bits of the pointer.
Linux ubuntu66 6.12.0-11677-g2ba9f676d0a2 #10 SMP Sat Nov 30 11:28:09 EST 2024 x86_64 x86_64 x86_64 GNU/Linux
BUG: unable to handle page fault for address: 0000000000001020 #PF: supervisor write access in kernel mode #PF: error_code(0x0002) - not-present page PGD 0 P4D 0 Oops: Oops: 0002 [#1] SMP PTI CPU: 6 UID: 0 PID: 1546 Comm: a.out Not tainted 6.12.0-11677-g2ba9f676d0a2 #10 Hardware name: FreeBSD BHYVE/BHYVE, BIOS 14.0 10/17/2021 RIP: 0010:bpf_prog_0e20ff5294b59096_bpf_cubic_init+0x19/0x25 Code: 00 cc cc cc cc cc cc cc cc cc cc cc cc cc cc cc cc f3 0f 1e fa 0f 1f 44 00 00 0f 1f 00 55 48 89 e5 f3 0f 1e fa 48 0f b7 77 00 <c7> 86 e0 05 00 00 00 00 00 00 c9 c3 cc cc cc cc cc cc cc cc cc cc RSP: 0018:ffffc9000076bc58 EFLAGS: 00010202 RAX: 0000000000000001 RBX: 0000000000000001 RCX: 0000000000000000 RDX: 0000000000000000 RSI: 0000000000000a40 RDI: ffffc9000076bc88 RBP: ffffc9000076bc58 R08: ffffc9000076bc40 R09: ffffc9000076bc20 R10: ffffffff842f3200 R11: ffffffff827659c0 R12: 0000000000000004 R13: ffff888105493098 R14: 0000000000000000 R15: 00000000ffffff8d FS: 00007fa5348d1740(0000) GS:ffff88842fb80000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000001020 CR3: 000000010bbfa003 CR4: 00000000003706f0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: <TASK> ? __die+0x1e/0x60 ? page_fault_oops+0x157/0x450 ? eth_header+0x25/0xb0 ? exc_page_fault+0x66/0x140 ? asm_exc_page_fault+0x26/0x30 ? bpf_prog_0e20ff5294b59096_bpf_cubic_init+0x19/0x25 ? __bpf_prog_enter+0x14/0x60 bpf__tcp_congestion_ops_init+0x47/0xa3 tcp_init_congestion_control+0x2a/0xe0 tcp_init_transfer+0x2b2/0x2d0 tcp_finish_connect+0x82/0x130 tcp_rcv_state_process+0x352/0xf20 tcp_v4_do_rcv+0xca/0x240 __release_sock+0xc6/0xd0 release_sock+0x2a/0x90 __inet_stream_connect+0x208/0x3c0 ? __pfx_woken_wake_function+0x10/0x10 inet_stream_connect+0x35/0x50 __sys_connect+0x93/0xb0 ? ksys_write+0x67/0xe0 __x64_sys_connect+0x13/0x20 ? ksys_write+0x67/0xe0 __x64_sys_connect+0x13/0x20 do_syscall_64+0x3f/0xd0 entry_SYSCALL_64_after_hwframe+0x76/0x7e
Robert Morris
[email protected]
Kumar Kartikeya Dwivedi
maintainerQuoted 14 lines · On Sat, 30 Nov 2024 at 18:10, <[email protected]> wrote:
When I modify the
bpf_cubic_init()function fromhttps://github.com/aroodgar/bpf-tcp-congestion-control-algorithm
to read:
SEC("struct_ops/bpf_cubic_init")
voidBPF_PROG(bpf_cubic_init, struct sock *sk)
{
asmvolatile("r2 = *(u16*)(r1 + 0)"); // verifier should demand u64
asmvolatile("*(u32 *)(r2 +1504) = 0"); // 1280 in some configs
}
Quoted 3 lines
the verifier accepts it, but the second line crashes when it runs during a TCP
connect()because the "*(u16*)" in the load from context yields only the low bits of the pointer.
Thanks for the bug report.
It is a missing check on "size" of the access in btf_ctx_access in kernel/bpf/btf.c.
It is probably not seen in practice as desugaring of such arguments is done by libbpf macros etc.
This won't just be limited to struct_ops, but many other program types (tracing, etc.) which use this code.
I have prepared a fix here: https://github.com/kkdwivedi/linux/commits/sops
The relevant diff is:
kernel/bpf/btf.c +0 -0
@@ -6458,6 +6458,11 @@ bool btf_ctx_access(int off, int size, enum
bpf_access_type type,
tname, off);
return false;
}+ if (size != sizeof(u64)) {
+ bpf_log(log, "func '%s' size %d is not multiple of 8\n",
+ tname, size);
+ return false;
+ }
arg = get_ctx_arg_idx(btf, t, off);
args = (const struct btf_param *)(t + 1);
/* if (t == NULL) Fall back to default BPF prog withThis is not the right 100% fix though, as fields may be u32 scalars etc. so size needs to be checked against the member's type. Right now this will reject some valid programs. But you get the idea. The size of access needs to be checked.
Will test out in our CI before posting to the list.
May I ask how you happened to stumble upon this?
Quoted 57 lines
Linux ubuntu66 6.12.0-11677-g2ba9f676d0a2 #10 SMP Sat Nov 30 11:28:09 EST 2024
x86_64x86_64x86_64GNU/LinuxBUG: unable to handle page fault for address: 0000000000001020 #PF: supervisor write access in kernel mode #PF: error_code(0x0002) - not-present page PGD 0 P4D 0 Oops: Oops: 0002 [#1] SMP PTI CPU: 6 UID: 0 PID: 1546 Comm: a.out Not tainted 6.12.0-11677-g2ba9f676d0a2 #10 Hardware name: FreeBSD BHYVE/BHYVE, BIOS 14.0 10/17/2021 RIP: 0010:bpf_prog_0e20ff5294b59096_bpf_cubic_init+0x19/0x25 Code: 00 cc cc cc cc cc cc cc cc cc cc cc cc cc cc cc cc f3 0f 1e fa 0f 1f 44 00 00 0f 1f 00 55 48 89 e5 f3 0f 1e fa 48 0f b7 77 00 <c7> 86 e0 05 00 00 00 00 00 00 c9 c3 cc cc cc cc cc cc cc cc cc cc RSP: 0018:ffffc9000076bc58 EFLAGS: 00010202 RAX: 0000000000000001 RBX: 0000000000000001 RCX: 0000000000000000 RDX: 0000000000000000 RSI: 0000000000000a40 RDI: ffffc9000076bc88 RBP: ffffc9000076bc58 R08: ffffc9000076bc40 R09: ffffc9000076bc20 R10: ffffffff842f3200 R11: ffffffff827659c0 R12: 0000000000000004 R13: ffff888105493098 R14: 0000000000000000 R15: 00000000ffffff8d FS: 00007fa5348d1740(0000) GS:ffff88842fb80000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000000001020 CR3: 000000010bbfa003 CR4: 00000000003706f0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: <TASK> ? __die+0x1e/0x60 ? page_fault_oops+0x157/0x450 ? eth_header+0x25/0xb0 ? exc_page_fault+0x66/0x140 ? asm_exc_page_fault+0x26/0x30 ? bpf_prog_0e20ff5294b59096_bpf_cubic_init+0x19/0x25 ? __bpf_prog_enter+0x14/0x60 bpf__tcp_congestion_ops_init+0x47/0xa3 tcp_init_congestion_control+0x2a/0xe0 tcp_init_transfer+0x2b2/0x2d0 tcp_finish_connect+0x82/0x130 tcp_rcv_state_process+0x352/0xf20 tcp_v4_do_rcv+0xca/0x240 __release_sock+0xc6/0xd0 release_sock+0x2a/0x90 __inet_stream_connect+0x208/0x3c0 ? __pfx_woken_wake_function+0x10/0x10 inet_stream_connect+0x35/0x50 __sys_connect+0x93/0xb0 ? ksys_write+0x67/0xe0 __x64_sys_connect+0x13/0x20 ? ksys_write+0x67/0xe0 __x64_sys_connect+0x13/0x20 do_syscall_64+0x3f/0xd0 entry_SYSCALL_64_after_hwframe+0x76/0x7eRobert Morris
[email protected]