2021-09-01 18:49:48

by Luo Jiaxing

[permalink] [raw]
Subject: Some concurrent actions cause __device_links_no_driver to report the warning calltrace.

Hi, rafael


I found one issue about device link, and want to ask for your help.


During the kernel test recently, we find that some concurrent actions
generate a calltrace, as shown in the following:

<4>[  606.102307] WARNING: CPU: 0 PID: 7 at drivers/base/core.c:1339
__device_links_no_driver+0x138/0x170
<4>[  606.284685] Call trace:
<4>[  606.287122]  __device_links_no_driver+0x138/0x170
<4>[  606.291804]  device_links_driver_cleanup+0xb0/0xfc
<4>[  606.296575]  __device_release_driver+0x148/0x1d8
<4>[  606.301173]  device_release_driver+0x38/0x50
<4>[  606.305423]  bus_remove_device+0x130/0x140
<4>[  606.309502]  device_del+0x174/0x430
<4>[  606.312975]  __scsi_remove_device+0x114/0x14c
<4>[  606.317313]  scsi_remove_target+0x1bc/0x240
<4>[  606.321469]  sas_rphy_remove+0x90/0x94
<4>[  606.325202]  sas_rphy_delete+0x44/0x5c
<4>[  606.328935]  sas_destruct_devices+0x64/0xa0 [libsas]
<4>[  606.333883]  sas_revalidate_domain+0xf8/0x1d0 [libsas]
<4>[  606.339002]  process_one_work+0x1dc/0x48c
<4>[  606.342994]  worker_thread+0x15c/0x464
<4>[  606.346726]  kthread+0x168/0x16c
<4>[  606.349940]  ret_from_fork+0x10/0x18
<4>[  606.353502] ---[ end trace cceb4f5db8bdcd25 ]---


The test method is to rmmod device driver and perform hard reset on the
hard disk at the same time.


We know device_links_unbind_consumers() is called during rmmod device
driver to release all consumers under the device in sequence.

As we are storage controller driver, so it look as follows:

supplier: storage controller

consumer: sda->sdb->sdc...

As the device_links_unbind_consumers () releases the consumer device in
serial mode. If a concurrent action is performed to hard reset a hard
disk, as the following software call stack show :

scsi_remove_target->device_del->bus_remove_device->device_release_driver->__device_links_no_driver().

The hardreset process also calls __device_links_no_driver.

Assume that device_links_unbind_consumers () is releasing sda and sdb is
queuing, but scsi_remove_target() calls __device_links_no_driver() to
release sdb in advance.Then a warning calltrace is generated.

We got some further analysis, it shows that sdb's link->status is now
DL_STATE_ACTIVE(sda's sdb's link->status is modified to
DL_STATE_SUPPLIER_UNBIND by device_links_unbind_consumers).

The if() in the following code will be false and pass through.

if (link->status != DL_STATE_CONSUMER_PROBE &&
    link->status != DL_STATE_ACTIVE)
    continue;

Since link->supplier->links.status has been set to DL_DEV_UNBINDING,
next code enters the else branch.

if (link->supplier->links.status == DL_DEV_DRIVER_BOUND) {
        WRITE_ONCE(link->status, DL_STATE_AVAILABLE);
} else {
        WARN_ON(!(link->flags & DL_FLAG_SYNC_STATE_ONLY));
        WRITE_ONCE(link->status, DL_STATE_DORMANT);
}


Because link->flags is set to DL_FLAG_MANAGED, calltrace is generated
based on WARN_ON.


In conclusion, we know that the call trace is generated because
link->supplier->links.status and link->status are not modified
synchronously.

After link->supplier->links.status is changed to DL_DEV_UNBINDING, the
value of link->status is changed to DL_STATE_SUPPLIER_UNBIND in sequence.

During this time difference, if a concurrent kernel thread invokes
__device_links_no_driver, warning calltrace will occurs.


I wonder if there is any way to solve this warning call trace?


Thanks

Jiaxing