Re: [PATCH v3] scsi: libsas: Fix exp-attached end device cannot be scanned in again after probe failed

[John Garry <john.g.garry@oracle.com>]

On 18/06/2024 14:10, yangxingui wrote:
>>>>
>>>>> We found that it is judged as broadcast flutter when the 
>>>>> exp-attached end
>>>>> device reconnects after probe failed, as follows:
>>>>>
>>>>> [78779.654026] sas: broadcast received: 0
>>>>> [78779.654037] sas: REVALIDATING DOMAIN on port 0, pid:10
>>>>> [78779.654680] sas: ex 500e004aaaaaaa1f phy05 change count has changed
>>>>> [78779.662977] sas: ex 500e004aaaaaaa1f phy05 originated 
>>>>> BROADCAST(CHANGE)
>>>>> [78779.662986] sas: ex 500e004aaaaaaa1f phy05 new device attached
>>>>> [78779.663079] sas: ex 500e004aaaaaaa1f phy05:U:8 attached: 
>>>>> 500e004aaaaaaa05 (stp)
>>>>> [78779.693542] hisi_sas_v3_hw 0000:b4:02.0: dev[16:5] found
>>>>> [78779.701155] sas: done REVALIDATING DOMAIN on port 0, pid:10, res 
>>>>> 0x0
>>>>> [78779.707864] sas: Enter sas_scsi_recover_host busy: 0 failed: 0
>>>>> ...
>>>>> [78835.161307] sas: --- Exit sas_scsi_recover_host: busy: 0 failed: 
>>>>> 0 tries: 1
>>>>> [78835.171344] sas: sas_probe_sata: for exp-attached device 
>>>>> 500e004aaaaaaa05 returned -19
>>>>> [78835.180879] hisi_sas_v3_hw 0000:b4:02.0: dev[16:5] is gone
>>>>> [78835.187487] sas: broadcast received: 0
>>>>> [78835.187504] sas: REVALIDATING DOMAIN on port 0, pid:10
>>>>> [78835.188263] sas: ex 500e004aaaaaaa1f phy05 change count has changed
>>>>> [78835.195870] sas: ex 500e004aaaaaaa1f phy05 originated 
>>>>> BROADCAST(CHANGE)
>>>>> [78835.195875] sas: ex 500e004aaaaaaa1f rediscovering phy05
>>>>> [78835.196022] sas: ex 500e004aaaaaaa1f phy05:U:A attached: 
>>>>> 500e004aaaaaaa05 (stp)
>>>>> [78835.196026] sas: ex 500e004aaaaaaa1f phy05 broadcast flutter
>>>>> [78835.197615] sas: done REVALIDATING DOMAIN on port 0, pid:10, res 
>>>>> 0x0
>>>>>
>>>>> The cause of the problem is that the related ex_phy's 
>>>>> attached_sas_addr was
>>>>> not cleared after the end device probe failed. In order to solve 
>>>>> the above
>>>>> problem, a function sas_ex_unregister_end_dev() is defined to clear 
>>>>> the
>>>>> ex_phy information and unregister the end device after the 
>>>>> exp-attached end
>>>>> device probe failed.
>>>>
>>>> Can you just manually clear the ex_phy's attached_sas_addr at the 
>>>> appropiate point (along with calling sas_unregister_dev())? It seems 
>>>> that we are using heavy-handed approach in calling 
>>>> sas_unregister_devs_sas_addr(), which does the clearing and much more.
>>>
>>> I just tried it and it worked. If we only clear ex_phy's 
>>> attached_sas_addr, there is no need to call sas_destruct_ports(). We 
>>> are currently using sas_unregister_devs_sas_addr() which will add the 
>>> port to sas_port_del_list, so we need to call sas_destruct_ports() 
>>> separately to delete the port.
>>>
>>> Should we also delete the port after the devices probe failed?
>>
>> I'm not sure. Please check it.
>>
>> sas_fail_probe() would still call sas_unregister_dev(), as required.
>>
>> And you said that the sas_fail_probe() probe call would be 
>> asynchronous to sas_revalidate_domainin(). I actually expected you to 
>> have the new call to sas_destruct_ports() at the top of 
>> sas_revalidate_domainin(), like v2, but it is in sas_probe_devices().
>>
>> Anyway, please check whether you require this additional call to 
>> delete the port.
>>
> Sorry, there was something wrong with the previous process description.
> the correct is:
> 
> 1. REVALIDATING DOMAIN
> 2. new device attached, create port,etc.
> 4. done REVALIDATING DOMAIN
> 5. @out, handle parent->port->sas_port_del_list
> 6. sas_probe_devices()
> 7. if device probe failed in step 6 and call 
> sas_unregister_devs_sas_addr(), then add phy->port->list to 
> parent->port->sas_port_del_list // port won't delete
> 
> 8. next, REVALIDATING DOMAIN
> 9. new device attached
> 10. new port create failed, as port already exits.
> 
> 
> So, v3 delete port at then end of sas_probe_devices(). And if we don't 
> use sas_unregister_devs_sas_addr() follow your suggestion then we don't 
> need to call sas_destruct_ports().

I am finding it hard to follow you now.

Can you show the complete change which you think that we now require to 
fix this issue?