Re: [PATCH net v2 2/2] net/smc: correct the reason code in smc_listen_find_device when fallback

From: Guangguan Wang
Date: Wed Oct 18 2023 - 04:35:47 EST




On 2023/10/18 15:01, Dust Li wrote:
> On Tue, Oct 17, 2023 at 08:42:34PM +0800, Guangguan Wang wrote:
>> The ini->rc is used to store the last error happened when finding usable
>> ism or rdma device in smc_listen_find_device, and is set by calling smc_
>> find_device_store_rc. Once the ini->rc is assigned to an none-zero value,
>> the value can not be overwritten anymore. So the ini-rc should be set to
>> the error reason only when an error actually occurs.
>>
>> When finding ISM/RDMA devices, device not found is not a real error, as
>> not all machine have ISM/RDMA devices. Failures after device found, when
>> initializing device or when initializing connection, is real errors, and
>> should be store in ini->rc.
>>
>> SMC_CLC_DECL_DIFFPREFIX also is not a real error, as for SMC-RV2, it is
>> not require same prefix.
>>
>> Signed-off-by: Guangguan Wang <guangguan.wang@xxxxxxxxxxxxxxxxx>
>> ---
>> net/smc/af_smc.c | 12 +++---------
>> 1 file changed, 3 insertions(+), 9 deletions(-)
>>
>> diff --git a/net/smc/af_smc.c b/net/smc/af_smc.c
>> index b3a67a168495..21e9c6ec4d01 100644
>> --- a/net/smc/af_smc.c
>> +++ b/net/smc/af_smc.c
>> @@ -2163,10 +2163,8 @@ static void smc_find_ism_v2_device_serv(struct smc_sock *new_smc,
>> }
>> mutex_unlock(&smcd_dev_list.mutex);
>>
>> - if (!ini->ism_dev[0]) {
>> - smc_find_device_store_rc(SMC_CLC_DECL_NOSMCD2DEV, ini);
>> + if (!ini->ism_dev[0])
>
> Hi Guangguan,
>
> Generally, I think this is right. Fallback in one kind of device should
> not be the final fallback reason.
>
> But what if we have more than one device and failed more than once, for
> example:
> Let's say we have an ISM device, a RDMA device. First we looked the ISM device
> and failed during the initialization, we got a fallback reason A. Then we
> looked for the RDMA device, and failed again, with another reason B.
> Finally, we fallback to TCP. What should fallback reason be ?

IMO, the order of finding devices is defined by preference. ISM v2, ISM v1, RDMA v2, RDMA v1, the former the prefer.
I think it should return the most preferred device's failure reason if found and failed during the initialization.
So, here should return the first reason(reason A).

>
> OTOH, SMC_CLC_DECL_NOSMCD2DEV is only used here. Removing it would mean
> that we would never see SMC_CLC_DECL_NOSMCD2DEV in the fallback reason,
> which makes it meaningless.
>

Is SMC_CLC_DECL_NOSMCD2DEV really necessary? There is no reason names SMC_CLC_DECL_NOSMCR2DEV.

Thanks,
Guangguan Wang

> Best regards,
> Dust
>
>> goto not_found;
>> - }
>>
>> smc_ism_get_system_eid(&eid);
>> if (!smc_clc_match_eid(ini->negotiated_eid, smc_v2_ext,
>> @@ -2216,9 +2214,9 @@ static void smc_find_ism_v1_device_serv(struct smc_sock *new_smc,
>> rc = smc_listen_ism_init(new_smc, ini);
>> if (!rc)
>> return; /* V1 ISM device found */
>> + smc_find_device_store_rc(rc, ini);
>>
>> not_found:
>> - smc_find_device_store_rc(rc, ini);
>> ini->smcd_version &= ~SMC_V1;
>> ini->ism_dev[0] = NULL;
>> ini->is_smcd = false;
>> @@ -2267,10 +2265,8 @@ static void smc_find_rdma_v2_device_serv(struct smc_sock *new_smc,
>> ini->smcrv2.saddr = new_smc->clcsock->sk->sk_rcv_saddr;
>> ini->smcrv2.daddr = smc_ib_gid_to_ipv4(smc_v2_ext->roce);
>> rc = smc_find_rdma_device(new_smc, ini);
>> - if (rc) {
>> - smc_find_device_store_rc(rc, ini);
>> + if (rc)
>> goto not_found;
>> - }
>> if (!ini->smcrv2.uses_gateway)
>> memcpy(ini->smcrv2.nexthop_mac, pclc->lcl.mac, ETH_ALEN);
>>
>> @@ -2331,8 +2327,6 @@ static int smc_listen_find_device(struct smc_sock *new_smc,
>>
>> /* check for matching IP prefix and subnet length (V1) */
>> prfx_rc = smc_listen_prfx_check(new_smc, pclc);
>> - if (prfx_rc)
>> - smc_find_device_store_rc(prfx_rc, ini);
>>
>> /* get vlan id from IP device */
>> if (smc_vlan_by_tcpsk(new_smc->clcsock, ini))
>> --
>> 2.24.3 (Apple Git-128)