On 2024/1/6 01:47, Alexei Starovoitov wrote:
> On Fri, Jan 5, 2024 at 2:34 AM Leon Hwang <hffilwlqm@xxxxxxxxx> wrote:
>>
>>
>>
>> On 5/1/24 12:15, Alexei Starovoitov wrote:
>>> On Thu, Jan 4, 2024 at 6:23 AM Leon Hwang <hffilwlqm@xxxxxxxxx> wrote:
>>>>
>>>>
>>>> diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
>>>> index fe30b9ebb8de4..67fa337fc2e0c 100644
>>>> --- a/arch/x86/net/bpf_jit_comp.c
>>>> +++ b/arch/x86/net/bpf_jit_comp.c
>>>> @@ -259,7 +259,7 @@ struct jit_context {
>>>> /* Number of bytes emit_patch() needs to generate instructions */
>>>> #define X86_PATCH_SIZE 5
>>>> /* Number of bytes that will be skipped on tailcall */
>>>> -#define X86_TAIL_CALL_OFFSET (11 + ENDBR_INSN_SIZE)
>>>> +#define X86_TAIL_CALL_OFFSET (22 + ENDBR_INSN_SIZE)
>>>>
>>>> static void push_r12(u8 **pprog)
>>>> {
>>>> @@ -406,14 +406,21 @@ static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf,
>>>> */
>>>> emit_nops(&prog, X86_PATCH_SIZE);
>>>> if (!ebpf_from_cbpf) {
>>>> - if (tail_call_reachable && !is_subprog)
>>>> + if (tail_call_reachable && !is_subprog) {
>>>> /* When it's the entry of the whole tailcall context,
>>>> * zeroing rax means initialising tail_call_cnt.
>>>> */
>>>> - EMIT2(0x31, 0xC0); /* xor eax, eax */
>>>> - else
>>>> - /* Keep the same instruction layout. */
>>>> - EMIT2(0x66, 0x90); /* nop2 */
>>>> + EMIT2(0x31, 0xC0); /* xor eax, eax */
>>>> + EMIT1(0x50); /* push rax */
>>>> + /* Make rax as ptr that points to tail_call_cnt. */
>>>> + EMIT3(0x48, 0x89, 0xE0); /* mov rax, rsp */
>>>> + EMIT1_off32(0xE8, 2); /* call main prog */
>>>> + EMIT1(0x59); /* pop rcx, get rid of tail_call_cnt */
>>>> + EMIT1(0xC3); /* ret */
>>>> + } else {
>>>> + /* Keep the same instruction size. */
>>>> + emit_nops(&prog, 13);
>>>> + }
>>>
>>> I'm afraid the extra call breaks stack unwinding and many other things.
>>> The proper frame needs to be setup (push rbp; etc)
>>> and 'leave' + emit_return() is used.
>>> Plain 'ret' is not ok.
>>> x86_call_depth_emit_accounting() needs to be used too.
>>> That will make X86_TAIL_CALL_OFFSET adjustment very complicated.
>>> Also the fix doesn't address the stack size issue.
>>> We shouldn't allow all the extra frames at run-time.
>>>
>>> The tail_cnt_ptr approach is interesting but too heavy,
>>> since arm64, s390 and other JITs would need to repeat it with equally
>>> complicated calculations in TAIL_CALL_OFFSET.
>>>
>>> The fix should really be thought through for all JITs. Not just x86.
>>>
>>> I'm thinking whether we should do the following instead:
>>>
>>> diff --git a/kernel/bpf/arraymap.c b/kernel/bpf/arraymap.c
>>> index 0bdbbbeab155..0b45571559be 100644
>>> --- a/kernel/bpf/arraymap.c
>>> +++ b/kernel/bpf/arraymap.c
>>> @@ -910,7 +910,7 @@ static void *prog_fd_array_get_ptr(struct bpf_map *map,
>>> if (IS_ERR(prog))
>>> return prog;
>>>
>>> - if (!bpf_prog_map_compatible(map, prog)) {
>>> + if (!bpf_prog_map_compatible(map, prog) || prog->aux->func_cnt) {
>>> bpf_prog_put(prog);
>>> return ERR_PTR(-EINVAL);
>>> }
>>>
>>> This will stop stack growth, but it will break a few existing tests.
>>> I feel it's a price worth paying.
>>
>> I don't think this can avoid this issue completely.
>>
>> For example:
>>
>> #include "vmlinux.h"
>>
>> #include "bpf_helpers.h"
>>
>> struct {
>> __uint(type, BPF_MAP_TYPE_PROG_ARRAY);
>> __uint(max_entries, 1);
>> __uint(key_size, sizeof(__u32));
>> __uint(value_size, sizeof(__u32));
>> } prog_array SEC(".maps");
>>
>>
>> static __noinline int
>> subprog(struct __sk_buff *skb)
>> {
>> volatile int retval = 0;
>>
>> bpf_tail_call(skb, &prog_array, 0);
>>
>> return retval;
>> }
>>
>> SEC("tc")
>> int entry(struct __sk_buff *skb)
>> {
>> const int N = 10000;
>>
>> for (int i = 0; i < N; i++)
>> subprog(skb);
>>
>> return 0;
>> }
>>
>> char _license[] SEC("license") = "GPL";
>>
>> Then, objdump its asm:
>>
>> Disassembly of section .text:
>>
>> 0000000000000000 <subprog>:
>> ; {
>> 0: b7 02 00 00 00 00 00 00 r2 = 0x0
>> ; volatile int retval = 0;
>> 1: 63 2a fc ff 00 00 00 00 *(u32 *)(r10 - 0x4) = r2
>> ; bpf_tail_call(skb, &prog_array, 0);
>> 2: 18 02 00 00 00 00 00 00 00 00 00 00 00 00 00 00 r2 = 0x0 ll
>> 4: b7 03 00 00 00 00 00 00 r3 = 0x0
>> 5: 85 00 00 00 0c 00 00 00 call 0xc
>> ; return retval;
>> 6: 61 a1 fc ff 00 00 00 00 r1 = *(u32 *)(r10 - 0x4)
>> 7: 95 00 00 00 00 00 00 00 exit
>>
>> Disassembly of section tc:
>>
>> 0000000000000000 <entry>:
>> ; {
>> 0: bf 16 00 00 00 00 00 00 r6 = r1
>> 1: b7 07 00 00 10 27 00 00 r7 = 0x2710
>>
>> 0000000000000010 <LBB0_1>:
>> ; subprog(skb);
>> 2: bf 61 00 00 00 00 00 00 r1 = r6
>> 3: 85 10 00 00 ff ff ff ff call -0x1
>> ; for (int i = 0; i < N; i++)
>> 4: 07 07 00 00 ff ff ff ff r7 += -0x1
>> 5: bf 71 00 00 00 00 00 00 r1 = r7
>> 6: 67 01 00 00 20 00 00 00 r1 <<= 0x20
>> 7: 77 01 00 00 20 00 00 00 r1 >>= 0x20
>> 8: 15 01 01 00 00 00 00 00 if r1 == 0x0 goto +0x1 <LBB0_2>
>> 9: 05 00 f8 ff 00 00 00 00 goto -0x8 <LBB0_1>
>>
>> 0000000000000050 <LBB0_2>:
>> ; return 0;
>> 10: b7 00 00 00 00 00 00 00 r0 = 0x0
>> 11: 95 00 00 00 00 00 00 00 exit
>>
>> As a result, the bpf prog in prog_array can be tailcalled for N times,
>> even though there's no subprog in the bpf prog in prog_array.
>
> You mean that total execution time is N*N ?
No, it's N. There's N tailcalls in subprog() to be called in entry().
> and tailcall is a way to increase loop count?
Yes, this is a way. And MAX_TAIL_CALL_CNT limit does not work for this case.
> We allow BPF_MAX_LOOPS = 8 * 1024 * 1024 in bpf_loop,
> so many calls to subprog(skb); is not an issue
> as long as they don't stall cpu and don't increase stack size.
What if there are BPF_MAX_LOOPS subprog(skb) and there are BPF_MAX_LOOPS
loops in the tail-callee bpf prog?
Thanks,
Leon
