Re: [RFC PATCH bpf-next v4 2/2] net: Add additional bit to support clockid_t timestamp type

[Martin KaFai Lau <martin.lau@linux.dev>]

On 4/19/24 6:13 PM, Abhishek Chauhan (ABC) wrote:
> 
> 
> On 4/18/2024 5:30 PM, Abhishek Chauhan (ABC) wrote:
>>
>>
>> On 4/18/2024 2:57 PM, Martin KaFai Lau wrote:
>>> On 4/18/24 1:10 PM, Abhishek Chauhan (ABC) wrote:
>>>>>>    #ifdef CONFIG_NET_XGRESS
>>>>>>        __u8            tc_at_ingress:1;    /* See TC_AT_INGRESS_MASK */
>>>>>>        __u8            tc_skip_classify:1;
>>>>>> @@ -1096,10 +1100,12 @@ struct sk_buff {
>>>>>>     */
>>>>>>    #ifdef __BIG_ENDIAN_BITFIELD
>>>>>>    #define SKB_MONO_DELIVERY_TIME_MASK    (1 << 7)
>>>>>> -#define TC_AT_INGRESS_MASK        (1 << 6)
>>>>>> +#define SKB_TAI_DELIVERY_TIME_MASK    (1 << 6)
>>>>>
>>>>> SKB_TSTAMP_TYPE_BIT2_MASK?
>>>
>>> nit. Shorten it to just SKB_TSTAMP_TYPE_MASK?
>>>
>> Okay i will do the same. Noted!
>>> #ifdef __BIG_ENDIAN_BITFIELD
>>> #define SKB_TSTAMP_TYPE_MASK    (3 << 6)
>>> #define SKB_TSTAMP_TYPE_RSH    (6)    /* more on this later */
>>> #else
>>> #define SKB_TSTAMP_TYPE_MASK    (3)
>>> #endif
>>>
>>>>>
>>>> I was thinking to keep it as TAI because it will confuse developers. I hope thats okay.
>>>
>>> I think it is not very useful to distinguish each bit since it is an enum value now. It becomes more like the "pkt_type:3" and its PKT_TYPE_MAX.
>>> I see what you are saying.
>>>>>> +#define TC_AT_INGRESS_MASK        (1 << 5)
>>>>>>    #else
>>>>>>    #define SKB_MONO_DELIVERY_TIME_MASK    (1 << 0)
>>>>>> -#define TC_AT_INGRESS_MASK        (1 << 1)
>>>>>> +#define SKB_TAI_DELIVERY_TIME_MASK    (1 << 1)
>>>>>> +#define TC_AT_INGRESS_MASK        (1 << 2)
>>>>>>    #endif
>>>>>>    #define SKB_BF_MONO_TC_OFFSET        offsetof(struct sk_buff, __mono_tc_offset)
>>>>>>    @@ -4206,6 +4212,11 @@ static inline void skb_set_delivery_time(struct sk_buff *skb, ktime_t kt,
>>>>>>        case CLOCK_MONOTONIC:
>>>>>>            skb->tstamp_type = SKB_CLOCK_MONO;
>>>>>>            break;
>>>>>> +    case CLOCK_TAI:
>>>>>> +        skb->tstamp_type = SKB_CLOCK_TAI;
>>>>>> +        break;
>>>>>> +    default:
>>>>>> +        WARN_ONCE(true, "clockid %d not supported", tstamp_type);
>>>>>
>>>>> and set to 0 and default tstamp_type?
>>>>> Actually thinking about it. I feel if its unsupported just fall back to default is the correct thing. I will take care of this.
>>>>>>        }
>>>>>>    }
>>>>>
>>>>>>    >
>>>>>    @@ -9372,10 +9378,16 @@ static struct bpf_insn *bpf_convert_tstamp_type_read(const struct bpf_insn *si,
>>>>>>        *insn++ = BPF_LDX_MEM(BPF_B, tmp_reg, skb_reg,
>>>>>>                      SKB_BF_MONO_TC_OFFSET);
>>>>>>        *insn++ = BPF_JMP32_IMM(BPF_JSET, tmp_reg,
>>>>>> -                SKB_MONO_DELIVERY_TIME_MASK, 2);
>>>>>> +                SKB_MONO_DELIVERY_TIME_MASK | SKB_TAI_DELIVERY_TIME_MASK, 2);
>>>>>> +    *insn++ = BPF_JMP32_IMM(BPF_JSET, tmp_reg,
>>>>>> +                SKB_MONO_DELIVERY_TIME_MASK, 3);
>>>>>> +    *insn++ = BPF_JMP32_IMM(BPF_JSET, tmp_reg,
>>>>>> +                SKB_TAI_DELIVERY_TIME_MASK, 4);
>>>>>>        *insn++ = BPF_MOV32_IMM(value_reg, BPF_SKB_TSTAMP_UNSPEC);
>>>>>>        *insn++ = BPF_JMP_A(1);
>>>>>>        *insn++ = BPF_MOV32_IMM(value_reg, BPF_SKB_TSTAMP_DELIVERY_MONO);
>>>>>> +    *insn++ = BPF_JMP_A(1);
>>>>>> +    *insn++ = BPF_MOV32_IMM(value_reg, BPF_SKB_TSTAMP_DELIVERY_TAI);
>>>
>>> With SKB_TSTAMP_TYPE_MASK defined like above, this could be simplified like this (untested):
>>>
>> Let me think this through and raise it as part of the next rfc patch.
>>> static struct bpf_insn *bpf_convert_tstamp_type_read(const struct bpf_insn *si,
>>>                                                       struct bpf_insn *insn)
>>> {
>>>      __u8 value_reg = si->dst_reg;
>>>      __u8 skb_reg = si->src_reg;
>>>
>>>      BUILD_BUG_ON(__SKB_CLOCK_MAX != BPF_SKB_TSTAMP_DELIVERY_TAI);
>>>      *insn++ = BPF_LDX_MEM(BPF_B, value_reg, skb_reg, SKB_BF_MONO_TC_OFFSET);
>>>      *insn++ = BPF_ALU32_IMM(BPF_AND, value_reg, SKB_TSTAMP_TYPE_MASK);
>>> #ifdef __BIG_ENDIAN_BITFIELD
>>>      *insn++ = BPF_ALU32_IMM(BPF_RSH, value_reg, SKB_TSTAMP_TYPE_RSH);
>>> #else
>>>      BUILD_BUG_ON(!(SKB_TSTAMP_TYPE_MASK & 0x1));
>>> #endif
>>>
>>>      return insn;
>>> }
>>>
>>>>>>          return insn;
>>>>>>    }
>>>>>> @@ -9418,10 +9430,26 @@ static struct bpf_insn *bpf_convert_tstamp_read(const struct bpf_prog *prog,
>>>>>>            __u8 tmp_reg = BPF_REG_AX;
>>>>>>              *insn++ = BPF_LDX_MEM(BPF_B, tmp_reg, skb_reg, SKB_BF_MONO_TC_OFFSET);
>>>>>> +        /*check if all three bits are set*/
>>>>>>            *insn++ = BPF_ALU32_IMM(BPF_AND, tmp_reg,
>>>>>> -                    TC_AT_INGRESS_MASK | SKB_MONO_DELIVERY_TIME_MASK);
>>>>>> -        *insn++ = BPF_JMP32_IMM(BPF_JNE, tmp_reg,
>>>>>> -                    TC_AT_INGRESS_MASK | SKB_MONO_DELIVERY_TIME_MASK, 2);
>>>>>> +                    TC_AT_INGRESS_MASK | SKB_MONO_DELIVERY_TIME_MASK |
>>>>>> +                    SKB_TAI_DELIVERY_TIME_MASK);
>>>>>> +        /*if all 3 bits are set jump 3 instructions and clear the register */
>>>>>> +        *insn++ = BPF_JMP32_IMM(BPF_JEQ, tmp_reg,
>>>>>> +                    TC_AT_INGRESS_MASK | SKB_MONO_DELIVERY_TIME_MASK |
>>>>>> +                    SKB_TAI_DELIVERY_TIME_MASK, 4);
>>>>>> +        /*Now check Mono is set with ingress mask if so clear */
>>>>>> +        *insn++ = BPF_JMP32_IMM(BPF_JEQ, tmp_reg,
>>>>>> +                    TC_AT_INGRESS_MASK | SKB_MONO_DELIVERY_TIME_MASK, 3);
>>>>>> +        /*Now Check tai is set with ingress mask if so clear */
>>>>>> +        *insn++ = BPF_JMP32_IMM(BPF_JEQ, tmp_reg,
>>>>>> +                    TC_AT_INGRESS_MASK | SKB_TAI_DELIVERY_TIME_MASK, 2);
>>>>>> +        /*Now Check tai and mono are set if so clear */
>>>>>> +        *insn++ = BPF_JMP32_IMM(BPF_JEQ, tmp_reg,
>>>>>> +                    SKB_MONO_DELIVERY_TIME_MASK |
>>>>>> +                    SKB_TAI_DELIVERY_TIME_MASK, 1);
>>>
>>> Same as the bpf_convert_tstamp_type_read, this could be simplified with SKB_TSTAMP_TYPE_MASK.
>>>
> Willem and Martin,
> When do we clear the tstamp and make it 0 in bpf_convert_tstamp_read? meaning which configuration?

When the bpf prog does not check the skb->tstamp_type. It is
the "if (!prog->tstamp_type_access)" in bpf_convert_tstamp_read().

If bpf prog does not check the skb->tstamp_type and it is at ingress,
bpf prog expects recv tstamp (ie. real clock), so it needs to clear
out the tstamp (i.e read as 0 tstamp).

> I see previously(current upstream code) if mono_delivery is set and tc_ingress_mask is set
> upstream code used to set the tstamp as 0.
> 
> Which means with addition of tai mask the new implementation should take care of following cases(correct me if i am wrong)
> 1. ( tai mask set + ingress mask set ) = Clear tstamp
> 2. ( mono mask set + ingress mask set ) = Clear tstamp
> 3. ( mono mask set + tai mask set + ingress mask set ) = Clear tstamp
> 4. ( No mask set ) = Clear tstamp
> 5. ( Tai mask set + mono mask set ) = Clear tstamp

No need to check the individual mono and tai bit here. Check the
tstamp_type as a whole. Like in pseudo C:

if (skb->tc_at_ingress && skb->tstamp_type)
	value_reg = 0;

untested code for tstamp_read() and tstamp_write():

static struct bpf_insn *bpf_convert_tstamp_read(const struct bpf_prog *prog,
                                                 const struct bpf_insn *si,
                                                 struct bpf_insn *insn)
{
	__u8 value_reg = si->dst_reg;
	__u8 skb_reg = si->src_reg;

#ifdef CONFIG_NET_XGRESS
	/* If the tstamp_type is read,
	 * the bpf prog is aware the tstamp could have delivery time.
	 * Thus, read skb->tstamp as is if tstamp_type_access is true.
	 */
	if (!prog->tstamp_type_access) {
		/* AX is needed because src_reg and dst_reg could be the same */
		__u8 tmp_reg = BPF_REG_AX;

		*insn++ = BPF_LDX_MEM(BPF_B, tmp_reg, skb_reg, SKB_BF_MONO_TC_OFFSET);
		*insn++ = BPF_JMP32_IMM(BPF_JSET, tmp_reg, TC_AT_INGRESS_MASK, 1);
		/* goto <read> */
		BPF_JMP_A(4);
		*insn++ = BPF_JMP32_IMM(BPF_JSET, tmp_reg, SKB_TSTAMP_TYPE_MASK, 1);
		/* goto <read> */
		BPF_JMP_A(2);
		/* skb->tc_at_ingress && skb->tstamp_type,
		 * read 0 as the (rcv) timestamp.
		 */
		*insn++ = BPF_MOV64_IMM(value_reg, 0);
		*insn++ = BPF_JMP_A(1);
	}
#endif

	/* <read>: value_reg = skb->tstamp */
	*insn++ = BPF_LDX_MEM(BPF_DW, value_reg, skb_reg,
			      offsetof(struct sk_buff, tstamp));
	return insn;
}

static struct bpf_insn *bpf_convert_tstamp_write(const struct bpf_prog *prog,
                                                  const struct bpf_insn *si,
	                                         struct bpf_insn *insn)
{
	__u8 value_reg = si->src_reg;
	__u8 skb_reg = si->dst_reg;

#ifdef CONFIG_NET_XGRESS
	/* If the tstamp_type is read,
	 * the bpf prog is aware the tstamp could have delivery time.
	 * Thus, write skb->tstamp as is if tstamp_type_access is true.
	 * Otherwise, writing at ingress will have to clear the
	 * mono_delivery_time (skb->tstamp_type:1)bit also.
	 */
         if (!prog->tstamp_type_access) {
		__u8 tmp_reg = BPF_REG_AX;

		*insn++ = BPF_LDX_MEM(BPF_B, tmp_reg, skb_reg, SKB_BF_MONO_TC_OFFSET);
		/* Writing __sk_buff->tstamp as ingress, goto <clear> */
		*insn++ = BPF_JMP32_IMM(BPF_JSET, tmp_reg, TC_AT_INGRESS_MASK, 1);
		/* goto <store> */
		*insn++ = BPF_JMP_A(2);
		/* <clear>: skb->tstamp_type */
		*insn++ = BPF_ALU32_IMM(BPF_AND, tmp_reg, ~SKB_TSTAMP_TYPE_MASK);
		*insn++ = BPF_STX_MEM(BPF_B, skb_reg, tmp_reg, SKB_BF_MONO_TC_OFFSET);
	}
#endif

	/* <store>: skb->tstamp = tstamp */
	*insn++ = BPF_RAW_INSN(BPF_CLASS(si->code) | BPF_DW | BPF_MEM,
			       skb_reg, value_reg, offsetof(struct sk_buff, tstamp), si->imm);
         return insn;
}

> 
> This leaves us with only two values which can be support which is 0x1 and 0x2
> 
> This means the tstamp_type should be either 0x1(mono) and tstamp_type 0x2 (tai) to set the value_reg with tstamp
> Is my understanding correct ?
> 
> Do you think the below simplified version looks okay ?
> 
> static struct bpf_insn *bpf_convert_tstamp_read(const struct bpf_prog *prog,
> 						const struct bpf_insn *si,
> 						struct bpf_insn *insn)
> {
> 	__u8 value_reg = si->dst_reg;
> 	__u8 skb_reg = si->src_reg;
> 
> BUILD_BUG_ON(__SKB_CLOCK_MAX != BPF_SKB_TSTAMP_DELIVERY_TAI);
> #ifdef CONFIG_NET_XGRESS
> 	/* If the tstamp_type is read,
> 	 * the bpf prog is aware the tstamp could have delivery time.
> 	 * Thus, read skb->tstamp as is if tstamp_type_access is true.
> 	 */
> 	if (!prog->tstamp_type_access) {
> 		/* AX is needed because src_reg and dst_reg could be the same */
> 		__u8 tmp_reg = BPF_REG_AX;
> 
> 		*insn++ = BPF_LDX_MEM(BPF_B, tmp_reg, skb_reg, SKB_BF_MONO_TC_OFFSET);
> 		/* check if all three bits are set*/
> 		*insn++ = BPF_ALU32_IMM(BPF_AND, tmp_reg,
> 					TC_AT_INGRESS_MASK | SKB_TSTAMP_TYPE_MASK);
> 
> 		/* If the value of tmp_reg is 7,6,5,4,3,0 which means invalid
> 		 * configuration set the tstamp to 0, value 0x1 and 0x2
> 		 * is correct configuration
> 		 */
> #ifdef __BIG_ENDIAN_BITFIELD
> 		*insn++ = BPF_JMP32_IMM(BPF_JEQ, tmp_reg, 0x1 << SKB_TSTAMP_TYPE_RSH, 3);
> 		*insn++ = BPF_JMP32_IMM(BPF_JEQ, tmp_reg, 0x2 << SKB_TSTAMP_TYPE_RSH, 2);
> #endif
> 		*insn++ = BPF_JMP32_IMM(BPF_JEQ, tmp_reg, 0x1, 3);
> 		*insn++ = BPF_JMP32_IMM(BPF_JEQ, tmp_reg, 0x2, 2);
> #endif
> 		/* skb->tc_at_ingress && skb->tstamp_type:2,
> 		 * read 0 as the (rcv) timestamp.
> 		 */
> 		*insn++ = BPF_MOV64_IMM(value_reg, 0);
> 		*insn++ = BPF_JMP_A(1);
> 	}
> #endif
> 
> 	*insn++ = BPF_LDX_MEM(BPF_DW, value_reg, skb_reg,
> 			      offsetof(struct sk_buff, tstamp));
> 	return insn;
> }
> 
>