Re: [PATCH net-next v4] net: Re-use and set mono_delivery_time bit for userspace tstamp packets

From: Martin KaFai Lau
Date: Wed Mar 20 2024 - 16:31:06 EST


On 3/19/24 11:22 PM, Abhishek Chauhan (ABC) wrote:


On 3/18/2024 12:02 PM, Abhishek Chauhan (ABC) wrote:


On 3/14/2024 3:29 PM, Abhishek Chauhan (ABC) wrote:


On 3/14/2024 2:48 PM, Martin KaFai Lau wrote:
On 3/14/24 1:53 PM, Abhishek Chauhan (ABC) wrote:
The bpf_convert_tstamp_{read,write} and the helper bpf_skb_set_tstamp need to be
changed to handle the new "user_delivery_time" bit anyway, e.g.
bpf_skb_set_tstamp(BPF_SKB_TSTAMP_DELIVERY_MONO) needs to clear the
"user_delivery_time" bit.

I think the "struct inet_frag_queue" also needs a new "user_delivery_time"
field. "mono_delivery_time" is already in there.

[ ... ]


Martin, Do we really need to add user_delivery_time as part of inet_frag_queue struct? I was wondering why is this required since we are using tstamp_type:2 to
distinguish between timestamp anyway .

Let me know what you think ?

I would think the first step is to revert this patch. I don't think much of the current patch can be reused.

1. I will raise one patch to introduce rename mono_delivery_time to
tstamp_type

Right, I expect something like this:

struct sk_buff {
        /* ... */
-            __u8                    mono_delivery_time:1;
+        __u8            tstamp_type:1;
        /* ... */
};


Okay ,This should be straight-forward.

2. I will introduce setting of userspace timestamp type as the second bit
whem transmit_time is set.

I expect the second patch should be introducing the enum first

enum skb_tstamp_type {
    SKB_TSTAMP_TYPE_RX_REAL = 0, /* A RX (receive) time in real */
    SKB_TSTAMP_TYPE_TX_MONO = 1, /* A TX (delivery) time in mono */
};

and start doing "skb->tstamp_type = SKB_TSTAMP_TYPE_TX_MONO;" instead of
"skb->tstamp_type = 1;"

and the same for "skb->tstamp_type = SKB_TSTAMP_TYPE_RX_REAL;" instead of
"skb->tstamp_type = 0;"


This one I am not sure but probably need to change the skb_set_delivery_time() function signature also:

static inline void skb_set_delivery_time(struct sk_buff *skb, ktime_t kt,
-                                        bool mono)
+                     enum skb_tstamp_type tstamp_type)

This should be straight-forward as well

The third patch is to change tstamp_type from 1 bit to 2 bits and add SKB_TSTAMP_TYPE_TX_USER.

struct sk_buff {
        /* ... */
-        __u8            tstamp_type:1;
+        __u8            tstamp_type:2;
        /* ... */
};

enum skb_tstamp_type {
    SKB_TSTAMP_TYPE_RX_REAL = 0,    /* A RX (receive) time in real */
    SKB_TSTAMP_TYPE_TX_MONO = 1,    /* A TX (delivery) time in mono */
+    SKB_TSTAMP_TYPE_TX_USER = 2,    /* A TX (delivery) time and its clock
                     * is in skb->sk->sk_clockid.
                     */
};

This will shift a bit out of the byte where tstamp_type lives. It should be the "inner_protocol_type" bit by my hand count. Please check if it is directly used in bpf instruction (filter.c). As far as I look, it is not, so should be fine. Some details about bpf instruction accessible skb bit field here: https://lore.kernel.org/all/20230321014115.997841-1-kuba@xxxxxxxxxx/
This is where i would need thorough reviews from you and Willem as my area of expertise is limited to part of network stack and BPF is not one of them.
But i have plan on this and i know how to do it.

Expect patches to be arriving to your inboxes next week, as we have a long weekend in Qualcomm
Fingers crossed :)



3. This will be a first step to make the design scalable.
4. Tomorrow if we have more timestamp to support, upstream community has to do is
update the enum and increase the bitfield from 2=>3 and so on.

I need help from Martin to test the patch which renames the mono_delivery_time
to tstamp_type (Which i feel should be straight forward as the value of the bit is 1)

The bpf change is not a no-op rename of mono_delivery_time. It needs to take care of the new bit added to the tstamp_type. Please see the previous email (and I also left it in the beginning of this email).

Thus, you need to compile the selftests/bpf/ and run it to verify the changes when handling the new bit. The Documentation/bpf/bpf_devel_QA.rst has the howto details. You probably only need the newer llvm (newer gcc should work also as bpf CI has been using it) and the newer pahole. I can definitely help if there is issue in running the test_progs in selftests/bpf or you have question on making the changes in filter.c. To run the test: "./test_progs -t tc_redirect/tc_redirect_dtime"


Martin,
I was able to compile test_progs and execute the above command mentioned by you . Does the output look okay for you ?

[ 3076.040766] IPv6: ADDRCONF(NETDEV_CHANGE): veth_src_fwd: link becomes ready
[ 3076.040809] IPv6: ADDRCONF(NETDEV_CHANGE): veth_src: link becomes ready
[ 3076.072844] IPv6: ADDRCONF(NETDEV_CHANGE): veth_dst: link becomes ready
[ 3076.072880] IPv6: ADDRCONF(NETDEV_CHANGE): veth_dst_fwd: link becomes ready
#214/5 tc_redirect/tc_redirect_dtime:OK
#214 tc_redirect:OK

lgtm. thanks.