Re: [PATCH v2 7/7] scsi: libsas: fix issue of swapping two sas disks

From: John Garry
Date: Fri Feb 01 2019 - 04:27:51 EST

On 01/02/2019 02:04, Jason Yan wrote:

On 2019/2/1 0:34, John Garry wrote:
On 31/01/2019 02:55, Jason Yan wrote:

On 2019/1/31 1:53, John Garry wrote:
On 30/01/2019 08:24, Jason Yan wrote:
The work flow of revalidation now is scanning expander phy by the
sequence of the phy and check if the phy have changed. This will leads
to an issue of swapping two sas disks on one expander.

Assume we have two sas disks, connected with expander phy10 and phy11:

phy10: 5000cca04eb1001d port-0:0:10
phy11: 5000cca04eb043ad port-0:0:11

Swap these two disks, and imaging the following scenario:

revalidation 1:

What does "revalidation 1" actually mean?

'revalidation 1' means one entry in sas_discover_domain().

-->phy10: 0 --> delete phy10 domain device
-->phy11: 5000cca04eb043ad (no change)

so is disk 11 still inserted at this stage?

Maybe, but that's what we read from the hardware.

revalidation done

revalidation 2:

is port-0:0:10 deleted now?

Yes. But we don't care about it.

-->step 1, check phy10:
-->phy10: 5000cca04eb043ad --> add to wide port(port-0:0:11)
address is still 5000cca04eb043ad now)

We do not want this to happen and it seems to be the crux of the problem.

As an alternate to your solution, how about check if the PHY is an end
device. If so, it should not form/join a wideport; that is, apart from
dual-port disks, which I am not sure about - I think each port still has
a unique WWN, so should be ok.

If the PHY do not join a wideport, then it have to form a wideport of
it's own. I'm not sure if we can have two ports with the same address
and do not break anything?

I'm not sure, but port-0:0:11 should be deleted from step 2, just after this step, below.


So this should not happen. How are you physcially swapping them such
that phy11 address is still 5000cca04eb043ad? I don't see how this
be true at revalidation 1.

This issue is because we always process the PHYs from 0 to max phy
number. And please be aware of the real physcial address of the PHY and
the address stored in the memory is not always the same.
Actually when you checking phy10, phy11 physcial address is not
5000cca04eb043ad. But the address stored in domain device is still
5000cca04eb043ad. We have not get a chance to to read it because we are
processing phy10 now, right?

I see.

It's very easy to reproduce. I suggest you to do it yourself and look at
the logs.

I can't physically access the backpane, and this is not the sort of
thing which is easy to fake by hacking the driver.

And the log which you provided internally does not have much - if any -
libsas logs to help me understand it.

-->step 2, check phy11:
-->phy11: 0 --> phy11 address is 0 now, but it's part of wide
port(port-0:0:11), the domain device will not be deleted.
revalidation done

revalidation 3:
-->phy10, 5000cca04eb043ad (no change)
-->phy11: 5000cca04eb1001d --> try to add port-0:0:11 but failed,
port-0:0:11 already exist, trigger a warning as follows
revalidation done

[14790.189699] sysfs: cannot create duplicate filename

[14790.201081] CPU: 25 PID: 5031 Comm: kworker/u192:3 Not tainted
4.16.0-rc1-191134-g138f084-dirty #228
[14790.210199] Hardware name: Huawei D06/D06, BIOS Hisilicon D06 EC
Nemo 2.0 RC0 - B303 05/16/2018
[14790.219323] Workqueue: 0000:74:02.0_disco_q sas_revalidate_domain
[14790.225404] Call trace:
[14790.227842] dump_backtrace+0x0/0x18c
[14790.231492] show_stack+0x14/0x1c
[14790.234798] dump_stack+0x88/0xac
[14790.238101] sysfs_warn_dup+0x64/0x7c
[14790.241751] sysfs_create_dir_ns+0x90/0xa0
[14790.245835] kobject_add_internal+0xa0/0x284
[14790.250092] kobject_add+0xb8/0x11c
[14790.253570] device_add+0xe8/0x598
[14790.256960] sas_port_add+0x24/0x50
[14790.260436] sas_ex_discover_devices+0xb10/0xc30
[14790.265040] sas_ex_revalidate_domain+0x1d8/0x518
[14790.269731] sas_revalidate_domain+0x12c/0x154
[14790.274163] process_one_work+0x128/0x2b0
[14790.278160] worker_thread+0x14c/0x408
[14790.281897] kthread+0xfc/0x128
[14790.285026] ret_from_fork+0x10/0x18
[14790.288598] ------------[ cut here ]------------

At last, the disk 5000cca04eb1001d is lost.