Date: Tue, 26 Dec 2023 19:52:32 +0800
From: Wen Gu <guwen@...ux.alibaba.com>
To: shaozhengchao <shaozhengchao@...wei.com>
Cc: "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
 yuehaibing <yuehaibing@...wei.com>, "Libin (Huawei)"
 <huawei.libin@...wei.com>, Dust Li <dust.li@...ux.alibaba.com>,
 tonylu_linux <tonylu@...ux.alibaba.com>, "D. Wythe"
 <alibuda@...ux.alibaba.com>
Subject: Re: [question] smc: how to enable SMC_LO feature



On 2023/12/14 11:17, shaozhengchao wrote:
> 
> 
> On 2023/12/13 20:59, Wen Gu wrote:
>> On 2023/12/13 17:00, shaozhengchao wrote:
>>>
>>>
>>> On 2023/12/5 14:45, shaozhengchao wrote:
>>>>
>>>>
>>>> On 2023/12/4 12:06, shaozhengchao wrote:
>>>>>
>>>>>
>>>>> On 2023/12/4 11:52, Wen Gu wrote:
>>>>>>
>>>>>> On 2023/12/4 11:22, shaozhengchao wrote:
>>>>>>>
>>>>>>>
>>>>>>> On 2023/11/23 14:15, shaozhengchao wrote:
>>>>>>>>
>>>>>>>>
>>>>>>>> On 2023/11/23 10:21, Wen Gu wrote:
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> On 2023/11/21 20:14, shaozhengchao wrote:
>>>>>>>>>> Hi Wen Gu:
>>>>>>>>>> Currently, I am interested in the SMC_LOOPBACK feature proposed
>>>>>>>>>> by you. Therefore, I use your patchset[1] to test the SMC_LO feature on
>>>>>>>>>> my x86_64 environment and kernel is based on linux-next, commit: 5ba73bec5e7b.
>>>>>>>>>> The test result shows that the smc_lo feature cannot be enabled. Here's
>>>>>>>>>> my analysis:
>>>>>>>>>>
>>>>>>>>>> 1. Run the following command to perform the test, and then capture
>>>>>>>>>> packets on the lo device.
>>>>>>>>>> - serv:  smc_run taskset -c <cpu> sockperf sr --tcp
>>>>>>>>>> - clnt:  smc_run taskset -c <cpu> sockperf  tp --tcp --msg-size=64000 -i 127.0.0.1 -t 30
>>>>>>>>>>
>>>>>>>>>> 2. Use Wireshark to open packets. It is found that the VCE port replies with
>>>>>>>>>> SMC-R-Deline packets.
>>>>>>>>>> [cid:image001.png@...A1CB4.F1052C30]
>>>>>>>>>>
>>>>>>>>>> 3. Rx
>>>>>>>>>> When smc_listen_work invokes smc_listen_v2_check, the VCE port returns
>>>>>>>>>> a Decline packet because eid_cnt and flag.seid in the received packet are both 0.
>>>>>>>>>>
>>>>>>>>>> 4. Tx
>>>>>>>>>> In smc_clc_send_proposal,
>>>>>>>>>> v2_ext->hdr.eid_cnt = smc_clc_eid_table.ueid_cnt;
>>>>>>>>>> v2_ext->hdr.flag.seid = smc_clc_eid_table.seid_enabled;
>>>>>>>>>>
>>>>>>>>>> When smc_clc_init, ueid_cnt=0, and in the x86_64 environment, seid_enabled is
>>>>>>>>>> always equal to 0.
>>>>>>>>>>
>>>>>>>>>> So, I must call smc_clc_ueid_add function to increase ueid count?
>>>>>>>>>> But I don't see where operations can be added, may I missed something?
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Hi Zhengchao Shao,
>>>>>>>>>
>>>>>>>>> Yes. When using SMC-D in non-s390 architecture (like x86 here), A common
>>>>>>>>> UEID should be set. It can be set by following steps:
>>>>>>>>>
>>>>>>>>> - Install smc-tools[1].
>>>>>>>>>
>>>>>>>>> - Run # smcd ueid add <ueid> in loopback test environment.
>>>>>>>>>
>>>>>>>>>    EID works as an ID to indicate the max communication space of SMC. When SEID is
>>>>>>>>>    unavailable, an UEID is required.
>>>>>>>>>
>>>>>>>> Hi Wen Gu:
>>>>>>>>      Thank you for your reply. This is very useful for me. And I will
>>>>>>>> be happy to learn from it.
>>>>>>>>
>>>>>>>> Thanks
>>>>>>>>
>>>>>>>> Zhengchao Shao
>>>>>>>>> - Then run the test.
>>>>>>>>>
>>>>>>>>> Hope this works for you :)
>>>>>>>>>
>>>>>>>>> [1] https://github.com/ibm-s390-linux/smc-tools
>>>>>>>>>
>>>>>>>>> Regards,
>>>>>>>>> Wen Gu
>>>>>>>>>
>>>>>>>>>> Could you give me some advice? Thanks very much.
>>>>>>>>>>
>>>>>>>>>> Zhengchao Shao
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> [1]link: 
>>>>>>>>>> https://patchwork.kernel.org/project/netdevbpf/cover/1695568613-125057-1-git-send-email-guwen@linux.alibaba.com/
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>> Hi Wen Gu:
>>>>>>>      I have test as following, but the performance is really
>>>>>>> degraded. Now I have no idea.
>>>>>>> 1. add ueid
>>>>>>> run: smcd ueid add 16
>>>>>>> kernel message:
>>>>>>> [ 5252.009133] NET: Registered PF_SMC protocol family
>>>>>>> [ 5252.009233] smc: adding smcd device smc_lo with pnetid
>>>>>>> 2. start server
>>>>>>> smc_run taskset -c 1 sockperf sr --tcp
>>>>>>> 3. start client
>>>>>>> smc_run taskset -c 3 sockperf tp  --tcp --msg-size=64000 -i 127.0.0.1 -t 30
>>>>>>>
>>>>>>> The test results are as follows:
>>>>>>>                TCP                  SMC-lo
>>>>>>> Bandwidth(MBps)         1890.56               1300.41(-31.22%)
>>>>>>>
>>>>>>> I didn't find a better direction when I initially positioned it. No
>>>>>>> error is recorded in the kernel log, and the smcd statistics are normal.
>>>>>>> [root@...alhost smc-tools]# smcd stats
>>>>>>> SMC-D Connections Summary
>>>>>>>    Total connections handled             2
>>>>>>>    SMC connections                       2
>>>>>>>    Handshake errors                      0
>>>>>>>    Avg requests per SMC conn       1277462.0
>>>>>>>    TCP fallback                          0
>>>>>>>
>>>>>>> RX Stats
>>>>>>>    Data transmitted (Bytes)    40907328000 (40.91G)
>>>>>>>    Total requests                  1277190
>>>>>>>    Buffer full                          45 (0.00%)
>>>>>>>              8KB    16KB    32KB    64KB   128KB   256KB   512KB >512KB
>>>>>>>    Bufs        0       0       0       2       0       0 0       0
>>>>>>>    Reqs   638.0K       0       0  639.2K       0       0 0       0
>>>>>>>
>>>>>>> TX Stats
>>>>>>>    Data transmitted (Bytes)    40907328000 (40.91G)
>>>>>>>    Total requests                  1277734
>>>>>>>    Buffer full                      638239 (49.95%)
>>>>>>>    Buffer full (remote)                  0 (0.00%)
>>>>>>>    Buffer too small                      0 (0.00%)
>>>>>>>    Buffer too small (remote)             0 (0.00%)
>>>>>>>              8KB    16KB    32KB    64KB   128KB   256KB   512KB >512KB
>>>>>>>    Bufs        0       0       0       0       0       0 0       0
>>>>>>>    Reqs        0       0       0  1.278M       0       0 0       0
>>>>>>>
>>>>>>> Extras
>>>>>>>    Special socket calls                  1
>>>>>>>
>>>>>>> I captured the perf information and found that the percentage of
>>>>>>> rep_movs_alternative and _raw_spin_unlock_irqrestore functions was high
>>>>>>> during tx and rx.
>>>>>>> 36.12%  [kernel]         [k]rep_movs_alternative
>>>>>>> 14.23%  [kernel]         [k]_raw_spin_unlock_irqrestore
>>>>>>>
>>>>>>> I've attached the flame map. Could you help analyze it? What I missed?
>>>>>>> Thanks.
>>>>>>
>>>>>> Hi Zhengchao Shao,
>>>>>>
>>>>>> Since sndbuf and RMB in SMC are pre-alloced ringbuf and won't grow dynamically
>>>>>> like TCP, it is necessary to appropriately increase the default value of smc
>>>>>> sk_sndbuf and sk_rcvbuf before testing throughput.
>>>>>>
>>>>>> Set this and try again:
>>>>>>
>>>>>> # sysctl -w net.smc.wmem=1048576
>>>>>> # sysctl -w net.smc.rmem=1048576
>>>>>>
>>>>>> (The initial value of wmem and rmem are 64K)
>>>>>>
>>>>>> Regards,
>>>>>> Wen Gu
>>>>>>
>>>>>>>
>>>>>>> Zhengchao Shao
>>>>> Hi Wen Gu:
>>>>>      It solves the issue. Thank you very much.
>>>>>
>>>>> Zhengchao Shao
>>>>>
>>>> Hi Wen Gu:
>>>>    I've tested all the performance test items in the patchset. The
>>>> performance improvement is to be expected, except for nignx.
>>>> My VM is configured with 48 cores and 32 GB memory. Therefore, run
>>>> the following command:
>>>> <smc_run> nignx
>>>> <smc_run>./wrk -t 96 -c 1000 -d 30 http://127.0.0.1:80
>>>>
>>>> The test results are as follows:
>>>>                          TCP                         SMC_lo
>>>> Requests/s           309425.42               135547.25(-56.19%)
>>>> The performance decreases by 56.19%.
>>>>
>>>> I capture packets and find that wrk can perform HTTP GET after each
>>>> connect when smc_loopback is disabled.
>>>> However, when smc_loopback is enabled, there is no HTTP GET behavior.
>>>> I wonder if there is some compatibility problem with the SMC protocol when encapsulate packet? Could you give me 
>>>> some advice?
>>>> In the attachment, I captured some of the packets.
>>>> nosmc_nginx.pcap is for SMC disabled and smc_nginx.pcap is for SMC
>>>> enabled.
>>>> Thank you very much.
>>>>
>>>> Zhengchao Shao
>>>>
>>>>
>>>>
>>> Hi Wen Gu:
>>>      When the VM is configured with 8 cores and 16 GB memory, run
>>> the following command:
>>> <smc_run> nignx
>>> <smc_run>./wrk -t 8 -c 1000 -d 30 http://127.0.0.1:80
>>> the test data is as follows:
>>>           TCP          SMC_lo
>>> Requests/s  66056.66    94526.66(43.10%)
>>>
>>> But When the VM is configured with 48 cores and 32 GB memory, run
>>> the following command:
>>> <smc_run> nignx
>>> <smc_run>./wrk -t 96 -c 1000 -d 30 http://127.0.0.1:80
>>> the test data is as follows:
>>>           TCP          SMC_lo
>>> Requests/s  309425.42     135547.25(-56.19%)
>>>
>>> It seems that in the scenario with a large number of CPU cores,
>>> performance is not optimized, but performance deteriorates. What I
>>> missed?
>>> Thank you.
>>>
>>> Zhengchao Shao
>>
>> Hi Zhengchao,
>>
>> I failed to reproduce this large regression. Could you please share some
>> information about your test environment?
>>
> Hi Wen Gu:
>> - The nginx configure.
> See the nginx.conf file in the attachment.
>> - The guest(VM) cpu topology.
> See the vm_cpuinfo file in the attachment.
>> - The host(physical machine) cpu topology.
> See the host_cpuinfo file in the attachment.
>> - The mapping relationship between vcpu of guest(VM) and physical cpu of host.
> See the cpu_map file in the attachment.
>> - The cpu usage (top) when regression happens.
>>
> See the perf_top and perf.svg file in the attachment.

Hi Zhengchao,

Thank you for the detailed information.

In the flame graph you provided, there are clearly prolonged spin-wait in
both smc_rx_recvmsg and smc_tx_sendmsg. The footprint involves the
__check_object_size() and find_vmap_area(). I think the regression relates
to the spin lock contention when CONFIG_HARDENED_USERCOPY is set and Tx / Rx
concurrently copy data between userspace and kernel vzalloced DMB.

       App1           App2
         |              ^
         |              |  userspace
      ----------------------
         |    +-----+   |  kernel
         +--->| DMB |---+
              +-----+

- smc_tx_sendmsg -> memcpy_from_msg -> copy_from_iter -> check_copy_size ->
   check_object_size -> check_heap_object -> if(vm) find_vmap_area -> try to hold spin lock vmap_area_lock

- smc_rx_recvmsg -> memcpy_to_msg -> copy_to_iter -> check_copy_size ->
   check_object_size -> check_heap_object -> if(vm) find_vmap_area -> try to hold spin lock vmap_area_lock

So I reproduced your test (thanks again for the details) and changed the DMB
creation from vzalloc to kzalloc(or alloc_page), thereby avoiding the spin lock
contention in find_vmap_area. Then the regression disappears.
(The attachments include flame graphs that use vzalloc and kzalloc respectively.)

                       SMC
-c1000 -t8      615397.66
-c1000 -t96     625627.69

diff --git a/net/smc/smc_loopback.c b/net/smc/smc_loopback.c
index 909c0699e91c..d6c9cd1a2f5b 100644
--- a/net/smc/smc_loopback.c
+++ b/net/smc/smc_loopback.c
@@ -191,7 +191,9 @@ static int smc_lo_register_dmb(struct smcd_dev *smcd, struct smcd_dmb *dmb,
         }

         dmb_node->sba_idx = sba_idx;
-       dmb_node->cpu_addr = vzalloc(dmb->dmb_len);
+       dmb_node->cpu_addr = kzalloc(dmb->dmb_len, GFP_KERNEL |
+                       __GFP_NOWARN | __GFP_NORETRY |
+                       __GFP_NOMEMALLOC);
         if (!dmb_node->cpu_addr) {
                 rc = -ENOMEM;
                 goto err_node;
@@ -260,7 +262,7 @@ static int smc_lo_unregister_dmb(struct smcd_dev *smcd, struct smcd_dmb *dmb)
         write_unlock(&ldev->dmb_ht_lock);

         clear_bit(dmb_node->sba_idx, ldev->sba_idx_mask);
-       vfree(dmb_node->cpu_addr);
+       kfree(dmb_node->cpu_addr);
         kfree(dmb_node);
         SMC_LO_STAT_DMBS_DEC(ldev);


Hope this works for you. And it needs to reconsider if the virtual alloced
memory is the right way for DMB in loopback-ism.

Best regards,
Wen Gu

>> Thank you.
>>
> 
> Thank you very much.
> 
> Zhengchao Shao
Download attachment "smc_vzalloc_t96.svg.svg" of type "image/svg+xml" (202182 bytes)

Download attachment "smc_kzalloc_t96.svg.svg" of type "image/svg+xml" (296582 bytes)

Powered by blists - more mailing lists