Re: [PATCH v5 3/3] scsi: libsas: Fix the failure of adding phy with zero-address to port

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

On 05/12/2023 13:22, yangxingui wrote:
> 
> On 2023/12/5 2:05, John Garry wrote:
>> On 04/12/2023 12:29, Xingui Yang wrote:
>>> When the expander device which attached many SATA disks is connected to
>>> the host, first disable and then enable the local phy. The following 
>>> BUG()
>>> will be triggered with a small probability:
>>>
>>> [562240.051046] sas: phy19 part of wide port with phy16
>>
>> Please use code from latest kernel. This again seems to be the old 
>> comment format.
> Ok.
>>
>>> [562240.051197] sas: ex 500e004aaaaaaa1f phy19:U:0 attached: 
>>> 0000000000000000 (no device)
>>
>> The log at 562240.051046 tells that phy19 formed a wideport with 
>> phy16, but then here we see that phy19 has attached SAS address 0. How 
>> did we form a wideport with a phy with sas address 0? Sorry if I asked 
>> this before, but I looked through the thread and it is not clear.
> Ok, the early address of phy19 is not 0, and forms a wide port with 
> phy16. But now phy19 has been unregistered and the sas address of phy19 
> is set to 0.

ok, so the old logs are simply misleading: "sas: phy19 part of wide port 
with phy16" implies that we have joined phy19 to a wideport with phy16.

Indeed, my change to that vague print is more than 4.5 years old now - 
see commit a5b38d3159.

Sorry to say, but that does not fill me full of confidence that the 
changes in this series are suitable for a mainline kernel. Please don't 
do that. Test against the very recent mainline kernel.

> 
>>
>>> [562240.051203] sas: done REVALIDATING DOMAIN on port 0, pid:435909, 
>>> res 0x0
>>> <...>
>>> [562240.062536] sas: ex 500e004aaaaaaa1f phy0 new device attached
>>> [562240.062616] sas: ex 500e004aaaaaaa1f phy00:U:5 attached: 
>>> 0000000000000000 (stp)
>>> [562240.062680]  port-7:7:0: trying to add phy phy-7:7:19 fails: it's 
>>> already part of another port
>>> [562240.085064] ------------[ cut here ]------------
>>> [562240.096612] kernel BUG at drivers/scsi/scsi_transport_sas.c:1083!
>>> [562240.109611] Internal error: Oops - BUG: 0 [#1] SMP
>>> [562240.343518] Process kworker/u256:3 (pid: 435909, stack limit = 
>>> 0x0000000003bcbebf)
>>> [562240.421714] Workqueue: 0000:b4:02.0_disco_q sas_revalidate_domain 
>>> [libsas]
>>> [562240.437173] pstate: 40c00009 (nZcv daif +PAN +UAO)
>>> [562240.450478] pc : sas_port_add_phy+0x13c/0x168 [scsi_transport_sas]
>>> [562240.465283] lr : sas_port_add_phy+0x13c/0x168 [scsi_transport_sas]
>>> [562240.479751] sp : ffff0000300cfa70
>>> [562240.674822] Call trace:
>>> [562240.682709]  sas_port_add_phy+0x13c/0x168 [scsi_transport_sas]
>>> [562240.694013]  sas_ex_get_linkrate.isra.5+0xcc/0x128 [libsas]
>>> [562240.704957]  sas_ex_discover_end_dev+0xfc/0x538 [libsas]
>>> [562240.715508]  sas_ex_discover_dev+0x3cc/0x4b8 [libsas]
>>> [562240.725634]  sas_ex_discover_devices+0x9c/0x1a8 [libsas]
>>> [562240.735855]  sas_ex_revalidate_domain+0x2f0/0x450 [libsas]
>>> [562240.746123]  sas_revalidate_domain+0x158/0x160 [libsas]
>>> [562240.756014]  process_one_work+0x1b4/0x448
>>> [562240.764548]  worker_thread+0x54/0x468
>>> [562240.772562]  kthread+0x134/0x138
>>> [562240.779989]  ret_from_fork+0x10/0x18
>>>
>>> What causes this problem:
>>> 1. For phy19, when the phy is attached and added to the parent wide 
>>> port,
>>> the path is:
>>> sas_rediscover()
>>>      ->sas_discover_new()
>>>          ->sas_ex_discover_devices()
>>>              ->sas_ex_discover_dev()
>>>                  -> sas_add_parent_port()
>>>
>>> ex_phy->port was not set and when it is removed from parent wide port 
>>> the
>>> path is:
>>> sas_rediscover()
>>>      ->sas_unregister_devs_sas_addr()
>>
>>
>> Sorry, but that is not a callpath. Maybe you condensed it. Please 
>> expand it.
> Ok.
>>
>>>
>>> Then the sas address of phy19 becomes 0, and since ex_phy->port is NULL,
>>> phy19 was not removed from the parent wide port's phy_list.
>>>
>>> 2. For phy0, it is connected to a new sata device and the path is:
>>> sas_rediscover()
>>>      ->sas_discover_new()->sas_ex_phy_discover()
>>>                              ->sas_ex_phy_discover_helper()
>>>                                  ->sas_set_ex_phy()
>>>                          ->sas_ex_discover_devices()
>>>                              ->sas_ex_discover_dev()
>>>                                  ->sas_ex_discover_end_dev()
>>>                                      ->sas_port_alloc() // Create 
>>> port-7:7:0
>>>                                      ->sas_ex_get_linkrate()
>>>                                          ->sas_port_add_phy()
>>>
>>> The type of the newly connected device is stp, but the linkrate is 5 
>>> which
>>> less than 1.5G, then the sas address is set to 0 in sas_set_ex_phy().
>>
>> I don't understand why we do anything when in this state. linkrate == 
>> 5 means phy reset in progress. Can we just bail out until the SATA phy 
>> is in a decent shape? I assume that when the SATA phy is in "up" state 
>> that we get a broadcast event and can re-evaluate.
> You are saying that we use a method similar to SAS_SATA_SPINUP_HOLD?

Maybe. Can we simply re-use SAS_SATA_SPINUP_HOLD handling? Is it suitable?

>>
>>> Subsequently, a new port port-7:7:0 was created and tried to add 
>>> phy19 with
>>> the same zero-address to this new port. However, phy19 still belongs to
>>> another port, then a BUG() was triggered in sas_ex_get_linkrate().
>>>
>>> Fix the problem as follows:
>>> 1. Use sas_port_add_ex_phy() instead of sas_port_add_phy() when 
>>> ex_phy is
>>> added to the parent port.
>>
>> this seems ok
>>
>>>
>>> 2. Set ex_dev->parent_port to NULL when the number of phy on the port
>>> becomes 0.
>>>
>>> 3. When phy->attached_dev_type != NO_DEVICE, do not set the zero address
>>> for phy->attached_sas_addr.
>>>
>>> Fixes: 2908d778ab3e ("[SCSI] aic94xx: new driver")
>>> Fixes: 7d1d86518118 ("[SCSI] libsas: fix false positive 'device 
>>> attached' conditions")
>>> Signed-off-by: Xingui Yang <yangxingui@huawei.com>
>>> ---
>>>   drivers/scsi/libsas/sas_expander.c | 10 ++++++----
>>>   1 file changed, 6 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/drivers/scsi/libsas/sas_expander.c 
>>> b/drivers/scsi/libsas/sas_expander.c
>>> index 7aa968b85e1e..9152152d5e10 100644
>>> --- a/drivers/scsi/libsas/sas_expander.c
>>> +++ b/drivers/scsi/libsas/sas_expander.c
>>> @@ -45,7 +45,7 @@ static void sas_add_parent_port(struct 
>>> domain_device *dev, int phy_id)
>>>           BUG_ON(sas_port_add(ex->parent_port));
>>>           sas_port_mark_backlink(ex->parent_port);
>>>       }
>>> -    sas_port_add_phy(ex->parent_port, ex_phy->phy);
>>> +    sas_port_add_ex_phy(ex->parent_port, ex_phy);
>>>   }
>>>   /* ---------- SMP task management ---------- */
>>> @@ -261,8 +261,7 @@ static void sas_set_ex_phy(struct domain_device 
>>> *dev, int phy_id,
>>>       /* help some expanders that fail to zero sas_address in the 'no
>>>        * device' case
>>>        */
>>
>> Please pay attention to this comment. It seems that some expanders 
>> require us to explicitly zero the SAS address.
> Yes, we have reviewed this point, and its modification is for some 
> expanders to report that the sas address isn't zero in the "no device" 
> case. The current modification does not affect its original problem fix, 
> we just removed its linkrate judgment.

ok

>>
>>> -    if (phy->attached_dev_type == SAS_PHY_UNUSED ||
>>> -        phy->linkrate < SAS_LINK_RATE_1_5_GBPS)
>>> +    if (phy->attached_dev_type == SAS_PHY_UNUSED)
>>>           memset(phy->attached_sas_addr, 0, SAS_ADDR_SIZE);
>>>       else
>>>           memcpy(phy->attached_sas_addr, dr->attached_sas_addr, 
>>> SAS_ADDR_SIZE);
>>> @@ -1864,9 +1863,12 @@ static void 
>>> sas_unregister_devs_sas_addr(struct domain_device *parent,
>>>       if (phy->port) {
>>>           sas_port_delete_phy(phy->port, phy->phy);
>>>           sas_device_set_phy(found, phy->port);
>>> -        if (phy->port->num_phys == 0)
>>> +        if (phy->port->num_phys == 0) {
>>>               list_add_tail(&phy->port->del_list,
>>>                   &parent->port->sas_port_del_list);
>>> +            if (ex_dev->parent_port == phy->port)
>>> +                ex_dev->parent_port = NULL;
>>
>> This does not feel like the right place to do this. So the port which 
>> we queue to free is the ex_dev->parent_port, right?
> Yes, we found that if ex_dev->parent_port is not set to NULL, after the 
> port is released, if there is a new ex_phy connection, use-after-free 

Do you mean really a memory use-after-free, like which KASAN would report?


> problems will occur. And the current branch is to determine whether the 
> number of phys on the port is 0. I think it is more appropriate to set 
> parent_port. Do you have any better suggestions?

Let me check again...

>>
>> BTW, do you know why it's called ex_dev->parent_port and not 
>> ex_dev->port? I find the name parent_port confusing...
> It is the port connected to the upper-level device, so named  parent_port.

But isn't this just the sas_port for the expander device attachment to 
the host and more closely associated to the expander itself?

Thanks,
John