diff mbox series

scsi/sg: don't grab scsi host module reference

Message ID 20230621160111.1433521-1-yukuai1@huaweicloud.com
State New
Headers show
Series scsi/sg: don't grab scsi host module reference | expand

Commit Message

Yu Kuai June 21, 2023, 4:01 p.m. UTC
From: Yu Kuai <yukuai3@huawei.com>

In order to prevent request_queue to be freed before cleaning up
blktrace debugfs entries, commit db59133e9279 ("scsi: sg: fix blktrace
debugfs entries leakage") use scsi_device_get(), however,
scsi_device_get() will also grab scsi module reference and scsi module
can't be removed.

It's reported that blktests can't unload scsi_debug after block/001:

blktests (master) # ./check block
block/001 (stress device hotplugging) [failed]
     +++ /root/blktests/results/nodev/block/001.out.bad 2023-06-19
      Running block/001
      Stressing sd
     +modprobe: FATAL: Module scsi_debug is in use.

Fix this problem by grabbing request_queue reference directly, so that
scsi host module can still be unloaded while request_queue will be
pinged by sg device.

Reported-by: Chaitanya Kulkarni <chaitanyak@nvidia.com>
Link: https://lore.kernel.org/all/1760da91-876d-fc9c-ab51-999a6f66ad50@nvidia.com/
Fixes: db59133e9279 ("scsi: sg: fix blktrace debugfs entries leakage")
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
---
 drivers/scsi/sg.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

Comments

Marc Hartmayer July 4, 2023, 5:04 p.m. UTC | #1
On Thu, Jun 22, 2023 at 12:01 AM +0800, Yu Kuai <yukuai1@huaweicloud.com> wrote:
> From: Yu Kuai <yukuai3@huawei.com>
>
> In order to prevent request_queue to be freed before cleaning up
> blktrace debugfs entries, commit db59133e9279 ("scsi: sg: fix blktrace
> debugfs entries leakage") use scsi_device_get(), however,
> scsi_device_get() will also grab scsi module reference and scsi module
> can't be removed.
>
> It's reported that blktests can't unload scsi_debug after block/001:
>
> blktests (master) # ./check block
> block/001 (stress device hotplugging) [failed]
>      +++ /root/blktests/results/nodev/block/001.out.bad 2023-06-19
>       Running block/001
>       Stressing sd
>      +modprobe: FATAL: Module scsi_debug is in use.
>
> Fix this problem by grabbing request_queue reference directly, so that
> scsi host module can still be unloaded while request_queue will be
> pinged by sg device.
>
> Reported-by: Chaitanya Kulkarni <chaitanyak@nvidia.com>
> Link: https://lore.kernel.org/all/1760da91-876d-fc9c-ab51-999a6f66ad50@nvidia.com/
> Fixes: db59133e9279 ("scsi: sg: fix blktrace debugfs entries leakage")
> Signed-off-by: Yu Kuai <yukuai3@huawei.com>
> ---
>  drivers/scsi/sg.c | 6 +++---
>  1 file changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/scsi/sg.c b/drivers/scsi/sg.c
> index 2433eeef042a..dcb73787c29d 100644
> --- a/drivers/scsi/sg.c
> +++ b/drivers/scsi/sg.c
> @@ -1497,7 +1497,7 @@ sg_add_device(struct device *cl_dev)
>  	int error;
>  	unsigned long iflags;
>  
> -	error = scsi_device_get(scsidp);
> +	error = blk_get_queue(scsidp->request_queue);
>  	if (error)
>  		return error;
>  
> @@ -1558,7 +1558,7 @@ sg_add_device(struct device *cl_dev)
>  out:
>  	if (cdev)
>  		cdev_del(cdev);
> -	scsi_device_put(scsidp);
> +	blk_put_queue(scsidp->request_queue);
>  	return error;
>  }
>  
> @@ -1575,7 +1575,7 @@ sg_device_destroy(struct kref *kref)
>  	 */
>  
>  	blk_trace_remove(q);
> -	scsi_device_put(sdp->device);
> +	blk_put_queue(q);
>  
>  	write_lock_irqsave(&sg_index_lock, flags);
>  	idr_remove(&sg_index_idr, sdp->index);
> -- 
> 2.39.2

Hi,

This change (bisected) triggers a regression in our KVM on s390x CI. The
symptom is that a “scsi_debug device” does not bind to the scsi_generic
driver. On s390x you can reproduce the problem as follows (I have not
tested on x86):

With this patch applied:

$ sudo modprobe scsi_debug
$ # Get the 'scsi_host,channel,target_number,LUN' tuple for the scsi_debug device
$ lsscsi |grep scsi_debug |awk '{ print $1 }'
[0:0:0:0]
$ sudo stat /sys/bus/scsi/devices/0:0:0:0/scsi_generic
stat: cannot statx '/sys/bus/scsi/devices/0:0:0:0/scsi_generic': No such file or directory


Patch reverted:

$ sudo modprobe scsi_debug
$ lsscsi |grep scsi_debug |awk '{ print $1 }'
[0:0:0:0]
$ sudo stat /sys/bus/scsi/devices/0:0:0:0/scsi_generic
  File: /sys/bus/scsi/devices/0:0:0:0/scsi_generic
  Size: 0         	Blocks: 0          IO Block: 4096   directory
Device: 0,20	Inode: 12155       Links: 3
…

Any ideas?

 Marc
Benjamin Block July 4, 2023, 6:51 p.m. UTC | #2
On Tue, Jul 04, 2023 at 07:04:00PM +0200, Marc Hartmayer wrote:
> On Thu, Jun 22, 2023 at 12:01 AM +0800, Yu Kuai <yukuai1@huaweicloud.com> wrote:
> > From: Yu Kuai <yukuai3@huawei.com>
> > diff --git a/drivers/scsi/sg.c b/drivers/scsi/sg.c
> > index 2433eeef042a..dcb73787c29d 100644
> > --- a/drivers/scsi/sg.c
> > +++ b/drivers/scsi/sg.c
> > @@ -1497,7 +1497,7 @@ sg_add_device(struct device *cl_dev)
> >  	int error;
> >  	unsigned long iflags;
> >  
> > -	error = scsi_device_get(scsidp);
> > +	error = blk_get_queue(scsidp->request_queue);
> >  	if (error)
> >  		return error;
> >  

Might be interesting as well. Marc showed me a `dmesg` snipped earlier
from when the bind fails:

  [   15.441817] scsi host2: scsi_eh_2: sleeping
  [   15.441899] scsi_debug:sdebug_driver_probe: scsi_debug: trim poll_queues to 0. poll_q/nr_hw = (0/1)
  [   15.441907] scsi host2: scsi_debug: version 0191 [20210520]
                   dev_size_mb=8, opts=0x0, submit_queues=1, statistics=0
  [   15.442078] scsi host2: scsi_scan_host_selected: <4294967295:4294967295:18446744073709551615>
  [   15.442267] scsi 2:0:0:0: scsi scan: INQUIRY pass 1 length 36
  [   15.442286] scsi 2:0:0:0: scsi scan: INQUIRY successful with code 0x0
  [   15.442296] scsi 2:0:0:0: scsi scan: INQUIRY pass 2 length 96
  [   15.442308] scsi 2:0:0:0: scsi scan: INQUIRY successful with code 0x0
  [   15.442317] scsi 2:0:0:0: Direct-Access     Linux    scsi_debug       0191 PQ: 0 ANSI: 7
  [   15.442554] scsi 2:0:0:0: Power-on or device reset occurred
  [   15.442560] scsi 2:0:0:0: tag#50 Done: SUCCESS Result: hostbyte=DID_OK driverbyte=DRIVER_OK cmd_age=0s
  [   15.442565] scsi 2:0:0:0: tag#50 CDB: Report supported operation codes a3 0c 01 88 00 00 00 00 00 14 00 00
  [   15.442569] scsi 2:0:0:0: tag#50 Sense Key : Unit Attention [current]
  [   15.442573] scsi 2:0:0:0: tag#50 Add. Sense: Power on occurred

The bind should happend around here somewhere I think.

  [   15.472680] sd 2:0:0:0: scsi scan: Sending REPORT LUNS to (try 0)
  [   15.472703] sd 2:0:0:0: scsi scan: REPORT LUNS successful (try 0) result 0x0
  [   15.472706] sd 2:0:0:0: scsi scan: REPORT LUN scan
  [   15.472709] sd 2:0:0:0: scsi scan: device exists on 2:0:0:0
  [   15.492874] sd 2:0:0:0: [sdi] 16384 512-byte logical blocks: (8.39 MB/8.00 MiB)
  [   15.502853] sd 2:0:0:0: [sdi] Write Protect is off
  [   15.502856] sd 2:0:0:0: [sdi] Mode Sense: 73 00 10 08
  [   15.522819] sd 2:0:0:0: [sdi] Write cache: enabled, read cache: enabled, supports DPO and FUA
  [   15.552773] sd 2:0:0:0: [sdi] Preferred minimum I/O size 512 bytes
  [   15.552776] sd 2:0:0:0: [sdi] Optimal transfer size 524288 bytes
  [   15.575373] sd 2:0:0:0: [sdi] tag#62 Done: SUCCESS Result: hostbyte=DID_OK driverbyte=DRIVER_OK cmd_age=0s
  [   15.575377] sd 2:0:0:0: [sdi] tag#62 CDB: Inquiry 12 01 b9 00 04 00
  [   15.575380] sd 2:0:0:0: [sdi] tag#62 Sense Key : Illegal Request [current]
  [   15.575383] sd 2:0:0:0: [sdi] tag#62 Add. Sense: Invalid field in cdb
  [   15.645749] sd 2:0:0:0: [sdi] Attached SCSI disk

But we don't even see the `sg_alloc: dev=...` message that is logged in
`sg_alloc()`. And between the change above and the call to `sg_alloc()`,
there is only the character device allocation; and if that failed, it
would print an error. So either the bind is never even tried, or the new
`blk_get_queue()` fails to get a reference.
    Which is odd, since the only way that would happen is, if the queue
was marked dying; but we see that the stack is using it for LUN probing
in `sd`.

> This change (bisected) triggers a regression in our KVM on s390x CI. The
> symptom is that a “scsi_debug device” does not bind to the scsi_generic
> driver. On s390x you can reproduce the problem as follows (I have not
> tested on x86):
> 
> With this patch applied:
> 
> $ sudo modprobe scsi_debug

One more thing maybe worth mentioning: in the kernel configuration we
use in the CI we have `sg` built-in. I guess most have it built as
module.

> $ # Get the 'scsi_host,channel,target_number,LUN' tuple for the scsi_debug device
> $ lsscsi |grep scsi_debug |awk '{ print $1 }'
> [0:0:0:0]
> $ sudo stat /sys/bus/scsi/devices/0:0:0:0/scsi_generic
> stat: cannot statx '/sys/bus/scsi/devices/0:0:0:0/scsi_generic': No such file or directory
> 
> 
> Patch reverted:
> 
> $ sudo modprobe scsi_debug
> $ lsscsi |grep scsi_debug |awk '{ print $1 }'
> [0:0:0:0]
> $ sudo stat /sys/bus/scsi/devices/0:0:0:0/scsi_generic
>   File: /sys/bus/scsi/devices/0:0:0:0/scsi_generic
>   Size: 0         	Blocks: 0          IO Block: 4096   directory
> Device: 0,20	Inode: 12155       Links: 3
> …

That's all I got from looking at it earlier, so far.
Yu Kuai July 5, 2023, 1:43 a.m. UTC | #3
Hi,

在 2023/07/05 1:04, Marc Hartmayer 写道:
> On Thu, Jun 22, 2023 at 12:01 AM +0800, Yu Kuai <yukuai1@huaweicloud.com> wrote:
>> From: Yu Kuai <yukuai3@huawei.com>
>>
>> In order to prevent request_queue to be freed before cleaning up
>> blktrace debugfs entries, commit db59133e9279 ("scsi: sg: fix blktrace
>> debugfs entries leakage") use scsi_device_get(), however,
>> scsi_device_get() will also grab scsi module reference and scsi module
>> can't be removed.
>>
>> It's reported that blktests can't unload scsi_debug after block/001:
>>
>> blktests (master) # ./check block
>> block/001 (stress device hotplugging) [failed]
>>       +++ /root/blktests/results/nodev/block/001.out.bad 2023-06-19
>>        Running block/001
>>        Stressing sd
>>       +modprobe: FATAL: Module scsi_debug is in use.
>>
>> Fix this problem by grabbing request_queue reference directly, so that
>> scsi host module can still be unloaded while request_queue will be
>> pinged by sg device.
>>
>> Reported-by: Chaitanya Kulkarni <chaitanyak@nvidia.com>
>> Link: https://lore.kernel.org/all/1760da91-876d-fc9c-ab51-999a6f66ad50@nvidia.com/
>> Fixes: db59133e9279 ("scsi: sg: fix blktrace debugfs entries leakage")
>> Signed-off-by: Yu Kuai <yukuai3@huawei.com>
>> ---
>>   drivers/scsi/sg.c | 6 +++---
>>   1 file changed, 3 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/scsi/sg.c b/drivers/scsi/sg.c
>> index 2433eeef042a..dcb73787c29d 100644
>> --- a/drivers/scsi/sg.c
>> +++ b/drivers/scsi/sg.c
>> @@ -1497,7 +1497,7 @@ sg_add_device(struct device *cl_dev)
>>   	int error;
>>   	unsigned long iflags;
>>   
>> -	error = scsi_device_get(scsidp);
>> +	error = blk_get_queue(scsidp->request_queue);
>>   	if (error)
>>   		return error;
>>   
>> @@ -1558,7 +1558,7 @@ sg_add_device(struct device *cl_dev)
>>   out:
>>   	if (cdev)
>>   		cdev_del(cdev);
>> -	scsi_device_put(scsidp);
>> +	blk_put_queue(scsidp->request_queue);
>>   	return error;
>>   }
>>   
>> @@ -1575,7 +1575,7 @@ sg_device_destroy(struct kref *kref)
>>   	 */
>>   
>>   	blk_trace_remove(q);
>> -	scsi_device_put(sdp->device);
>> +	blk_put_queue(q);
>>   
>>   	write_lock_irqsave(&sg_index_lock, flags);
>>   	idr_remove(&sg_index_idr, sdp->index);
>> -- 
>> 2.39.2
> 
> Hi,
> 
> This change (bisected) triggers a regression in our KVM on s390x CI. The
> symptom is that a “scsi_debug device” does not bind to the scsi_generic
> driver. On s390x you can reproduce the problem as follows (I have not
> tested on x86):
> 
> With this patch applied:
> 
> $ sudo modprobe scsi_debug
> $ # Get the 'scsi_host,channel,target_number,LUN' tuple for the scsi_debug device
> $ lsscsi |grep scsi_debug |awk '{ print $1 }'
> [0:0:0:0]
> $ sudo stat /sys/bus/scsi/devices/0:0:0:0/scsi_generic
> stat: cannot statx '/sys/bus/scsi/devices/0:0:0:0/scsi_generic': No such file or directory
> 
> 
> Patch reverted:
> 

I didn't figure out the root cause, howver, have you tried to reviert
this patch as well?

db59133e9279 ("scsi: sg: fix blktrace debugfs entries leakage"

Thanks,
Kuai
> $ sudo modprobe scsi_debug
> $ lsscsi |grep scsi_debug |awk '{ print $1 }'
> [0:0:0:0]
> $ sudo stat /sys/bus/scsi/devices/0:0:0:0/scsi_generic
>    File: /sys/bus/scsi/devices/0:0:0:0/scsi_generic
>    Size: 0         	Blocks: 0          IO Block: 4096   directory
> Device: 0,20	Inode: 12155       Links: 3
> …
> 
> Any ideas?
> 
>   Marc
> .
>
Yu Kuai July 5, 2023, 2:16 a.m. UTC | #4
Hi,

在 2023/07/05 2:51, Benjamin Block 写道:
> On Tue, Jul 04, 2023 at 07:04:00PM +0200, Marc Hartmayer wrote:
>> On Thu, Jun 22, 2023 at 12:01 AM +0800, Yu Kuai <yukuai1@huaweicloud.com> wrote:
>>> From: Yu Kuai <yukuai3@huawei.com>
>>> diff --git a/drivers/scsi/sg.c b/drivers/scsi/sg.c
>>> index 2433eeef042a..dcb73787c29d 100644
>>> --- a/drivers/scsi/sg.c
>>> +++ b/drivers/scsi/sg.c
>>> @@ -1497,7 +1497,7 @@ sg_add_device(struct device *cl_dev)
>>>   	int error;
>>>   	unsigned long iflags;
>>>   
>>> -	error = scsi_device_get(scsidp);
>>> +	error = blk_get_queue(scsidp->request_queue);
>>>   	if (error)
>>>   		return error;
>>>   
> 
> Might be interesting as well. Marc showed me a `dmesg` snipped earlier
> from when the bind fails:
> 
>    [   15.441817] scsi host2: scsi_eh_2: sleeping
>    [   15.441899] scsi_debug:sdebug_driver_probe: scsi_debug: trim poll_queues to 0. poll_q/nr_hw = (0/1)
>    [   15.441907] scsi host2: scsi_debug: version 0191 [20210520]
>                     dev_size_mb=8, opts=0x0, submit_queues=1, statistics=0
>    [   15.442078] scsi host2: scsi_scan_host_selected: <4294967295:4294967295:18446744073709551615>
>    [   15.442267] scsi 2:0:0:0: scsi scan: INQUIRY pass 1 length 36
>    [   15.442286] scsi 2:0:0:0: scsi scan: INQUIRY successful with code 0x0
>    [   15.442296] scsi 2:0:0:0: scsi scan: INQUIRY pass 2 length 96
>    [   15.442308] scsi 2:0:0:0: scsi scan: INQUIRY successful with code 0x0
>    [   15.442317] scsi 2:0:0:0: Direct-Access     Linux    scsi_debug       0191 PQ: 0 ANSI: 7
>    [   15.442554] scsi 2:0:0:0: Power-on or device reset occurred
>    [   15.442560] scsi 2:0:0:0: tag#50 Done: SUCCESS Result: hostbyte=DID_OK driverbyte=DRIVER_OK cmd_age=0s
>    [   15.442565] scsi 2:0:0:0: tag#50 CDB: Report supported operation codes a3 0c 01 88 00 00 00 00 00 14 00 00
>    [   15.442569] scsi 2:0:0:0: tag#50 Sense Key : Unit Attention [current]
>    [   15.442573] scsi 2:0:0:0: tag#50 Add. Sense: Power on occurred
> 
> The bind should happend around here somewhere I think.
> 
>    [   15.472680] sd 2:0:0:0: scsi scan: Sending REPORT LUNS to (try 0)
>    [   15.472703] sd 2:0:0:0: scsi scan: REPORT LUNS successful (try 0) result 0x0
>    [   15.472706] sd 2:0:0:0: scsi scan: REPORT LUN scan
>    [   15.472709] sd 2:0:0:0: scsi scan: device exists on 2:0:0:0
>    [   15.492874] sd 2:0:0:0: [sdi] 16384 512-byte logical blocks: (8.39 MB/8.00 MiB)
>    [   15.502853] sd 2:0:0:0: [sdi] Write Protect is off
>    [   15.502856] sd 2:0:0:0: [sdi] Mode Sense: 73 00 10 08
>    [   15.522819] sd 2:0:0:0: [sdi] Write cache: enabled, read cache: enabled, supports DPO and FUA
>    [   15.552773] sd 2:0:0:0: [sdi] Preferred minimum I/O size 512 bytes
>    [   15.552776] sd 2:0:0:0: [sdi] Optimal transfer size 524288 bytes
>    [   15.575373] sd 2:0:0:0: [sdi] tag#62 Done: SUCCESS Result: hostbyte=DID_OK driverbyte=DRIVER_OK cmd_age=0s
>    [   15.575377] sd 2:0:0:0: [sdi] tag#62 CDB: Inquiry 12 01 b9 00 04 00
>    [   15.575380] sd 2:0:0:0: [sdi] tag#62 Sense Key : Illegal Request [current]
>    [   15.575383] sd 2:0:0:0: [sdi] tag#62 Add. Sense: Invalid field in cdb
>    [   15.645749] sd 2:0:0:0: [sdi] Attached SCSI disk
> 
> But we don't even see the `sg_alloc: dev=...` message that is logged in
> `sg_alloc()`. And between the change above and the call to `sg_alloc()`,
> there is only the character device allocation; and if that failed, it
> would print an error. So either the bind is never even tried, or the new
> `blk_get_queue()` fails to get a reference.
>      Which is odd, since the only way that would happen is, if the queue
> was marked dying; but we see that the stack is using it for LUN probing
> in `sd`.

Yes, if scsi_device_get() works fine, but blk_get_queue() has problems,
it seems to me that sg_add_device() can be called with scsi_device queue
mark dying? This is odd, but I'm not sure if it's the case.

Thanks,
Kuai
> 
>> This change (bisected) triggers a regression in our KVM on s390x CI. The
>> symptom is that a “scsi_debug device” does not bind to the scsi_generic
>> driver. On s390x you can reproduce the problem as follows (I have not
>> tested on x86):
>>
>> With this patch applied:
>>
>> $ sudo modprobe scsi_debug
> 
> One more thing maybe worth mentioning: in the kernel configuration we
> use in the CI we have `sg` built-in. I guess most have it built as
> module.
> 
>> $ # Get the 'scsi_host,channel,target_number,LUN' tuple for the scsi_debug device
>> $ lsscsi |grep scsi_debug |awk '{ print $1 }'
>> [0:0:0:0]
>> $ sudo stat /sys/bus/scsi/devices/0:0:0:0/scsi_generic
>> stat: cannot statx '/sys/bus/scsi/devices/0:0:0:0/scsi_generic': No such file or directory
>>
>>
>> Patch reverted:
>>
>> $ sudo modprobe scsi_debug
>> $ lsscsi |grep scsi_debug |awk '{ print $1 }'
>> [0:0:0:0]
>> $ sudo stat /sys/bus/scsi/devices/0:0:0:0/scsi_generic
>>    File: /sys/bus/scsi/devices/0:0:0:0/scsi_generic
>>    Size: 0         	Blocks: 0          IO Block: 4096   directory
>> Device: 0,20	Inode: 12155       Links: 3
>> …
> 
> That's all I got from looking at it earlier, so far.
>
Yu Kuai July 5, 2023, 2:26 a.m. UTC | #5
Hi,

在 2023/07/05 10:16, Yu Kuai 写道:
> Hi,
> 
> 在 2023/07/05 2:51, Benjamin Block 写道:
>> On Tue, Jul 04, 2023 at 07:04:00PM +0200, Marc Hartmayer wrote:
>>> On Thu, Jun 22, 2023 at 12:01 AM +0800, Yu Kuai 
>>> <yukuai1@huaweicloud.com> wrote:
>>>> From: Yu Kuai <yukuai3@huawei.com>
>>>> diff --git a/drivers/scsi/sg.c b/drivers/scsi/sg.c
>>>> index 2433eeef042a..dcb73787c29d 100644
>>>> --- a/drivers/scsi/sg.c
>>>> +++ b/drivers/scsi/sg.c
>>>> @@ -1497,7 +1497,7 @@ sg_add_device(struct device *cl_dev)
>>>>       int error;
>>>>       unsigned long iflags;
>>>> -    error = scsi_device_get(scsidp);
>>>> +    error = blk_get_queue(scsidp->request_queue);
>>>>       if (error)
>>>>           return error;
>>
>> Might be interesting as well. Marc showed me a `dmesg` snipped earlier
>> from when the bind fails:
>>
>>    [   15.441817] scsi host2: scsi_eh_2: sleeping
>>    [   15.441899] scsi_debug:sdebug_driver_probe: scsi_debug: trim 
>> poll_queues to 0. poll_q/nr_hw = (0/1)
>>    [   15.441907] scsi host2: scsi_debug: version 0191 [20210520]
>>                     dev_size_mb=8, opts=0x0, submit_queues=1, 
>> statistics=0
>>    [   15.442078] scsi host2: scsi_scan_host_selected: 
>> <4294967295:4294967295:18446744073709551615>
>>    [   15.442267] scsi 2:0:0:0: scsi scan: INQUIRY pass 1 length 36
>>    [   15.442286] scsi 2:0:0:0: scsi scan: INQUIRY successful with 
>> code 0x0
>>    [   15.442296] scsi 2:0:0:0: scsi scan: INQUIRY pass 2 length 96
>>    [   15.442308] scsi 2:0:0:0: scsi scan: INQUIRY successful with 
>> code 0x0
>>    [   15.442317] scsi 2:0:0:0: Direct-Access     Linux    
>> scsi_debug       0191 PQ: 0 ANSI: 7
>>    [   15.442554] scsi 2:0:0:0: Power-on or device reset occurred
>>    [   15.442560] scsi 2:0:0:0: tag#50 Done: SUCCESS Result: 
>> hostbyte=DID_OK driverbyte=DRIVER_OK cmd_age=0s
>>    [   15.442565] scsi 2:0:0:0: tag#50 CDB: Report supported operation 
>> codes a3 0c 01 88 00 00 00 00 00 14 00 00
>>    [   15.442569] scsi 2:0:0:0: tag#50 Sense Key : Unit Attention 
>> [current]
>>    [   15.442573] scsi 2:0:0:0: tag#50 Add. Sense: Power on occurred
>>
>> The bind should happend around here somewhere I think.
>>
>>    [   15.472680] sd 2:0:0:0: scsi scan: Sending REPORT LUNS to (try 0)
>>    [   15.472703] sd 2:0:0:0: scsi scan: REPORT LUNS successful (try 
>> 0) result 0x0
>>    [   15.472706] sd 2:0:0:0: scsi scan: REPORT LUN scan
>>    [   15.472709] sd 2:0:0:0: scsi scan: device exists on 2:0:0:0
>>    [   15.492874] sd 2:0:0:0: [sdi] 16384 512-byte logical blocks: 
>> (8.39 MB/8.00 MiB)
>>    [   15.502853] sd 2:0:0:0: [sdi] Write Protect is off
>>    [   15.502856] sd 2:0:0:0: [sdi] Mode Sense: 73 00 10 08
>>    [   15.522819] sd 2:0:0:0: [sdi] Write cache: enabled, read cache: 
>> enabled, supports DPO and FUA
>>    [   15.552773] sd 2:0:0:0: [sdi] Preferred minimum I/O size 512 bytes
>>    [   15.552776] sd 2:0:0:0: [sdi] Optimal transfer size 524288 bytes
>>    [   15.575373] sd 2:0:0:0: [sdi] tag#62 Done: SUCCESS Result: 
>> hostbyte=DID_OK driverbyte=DRIVER_OK cmd_age=0s
>>    [   15.575377] sd 2:0:0:0: [sdi] tag#62 CDB: Inquiry 12 01 b9 00 04 00
>>    [   15.575380] sd 2:0:0:0: [sdi] tag#62 Sense Key : Illegal Request 
>> [current]
>>    [   15.575383] sd 2:0:0:0: [sdi] tag#62 Add. Sense: Invalid field 
>> in cdb
>>    [   15.645749] sd 2:0:0:0: [sdi] Attached SCSI disk
>>
>> But we don't even see the `sg_alloc: dev=...` message that is logged in
>> `sg_alloc()`. And between the change above and the call to `sg_alloc()`,
>> there is only the character device allocation; and if that failed, it
>> would print an error. So either the bind is never even tried, or the new
>> `blk_get_queue()` fails to get a reference.
>>      Which is odd, since the only way that would happen is, if the queue
>> was marked dying; but we see that the stack is using it for LUN probing
>> in `sd`.
> 
> Yes, if scsi_device_get() works fine, but blk_get_queue() has problems,
> it seems to me that sg_add_device() can be called with scsi_device queue
> mark dying? This is odd, but I'm not sure if it's the case.

Sorry that I totally messed how blk_get_queue() is called, it returns
true on success and false on failure, not 0 on success and errno on
failure.

Sorry for all the troble.

Kuai
> 
> Thanks,
> Kuai
>>
>>> This change (bisected) triggers a regression in our KVM on s390x CI. The
>>> symptom is that a “scsi_debug device” does not bind to the scsi_generic
>>> driver. On s390x you can reproduce the problem as follows (I have not
>>> tested on x86):
>>>
>>> With this patch applied:
>>>
>>> $ sudo modprobe scsi_debug
>>
>> One more thing maybe worth mentioning: in the kernel configuration we
>> use in the CI we have `sg` built-in. I guess most have it built as
>> module.
>>
>>> $ # Get the 'scsi_host,channel,target_number,LUN' tuple for the 
>>> scsi_debug device
>>> $ lsscsi |grep scsi_debug |awk '{ print $1 }'
>>> [0:0:0:0]
>>> $ sudo stat /sys/bus/scsi/devices/0:0:0:0/scsi_generic
>>> stat: cannot statx '/sys/bus/scsi/devices/0:0:0:0/scsi_generic': No 
>>> such file or directory
>>>
>>>
>>> Patch reverted:
>>>
>>> $ sudo modprobe scsi_debug
>>> $ lsscsi |grep scsi_debug |awk '{ print $1 }'
>>> [0:0:0:0]
>>> $ sudo stat /sys/bus/scsi/devices/0:0:0:0/scsi_generic
>>>    File: /sys/bus/scsi/devices/0:0:0:0/scsi_generic
>>>    Size: 0             Blocks: 0          IO Block: 4096   directory
>>> Device: 0,20    Inode: 12155       Links: 3
>>> …
>>
>> That's all I got from looking at it earlier, so far.
>>
> 
> .
>
Yu Kuai July 5, 2023, 2:28 a.m. UTC | #6
Hi,

在 2023/07/05 9:43, Yu Kuai 写道:
> Hi,
> 
> 在 2023/07/05 1:04, Marc Hartmayer 写道:
>> On Thu, Jun 22, 2023 at 12:01 AM +0800, Yu Kuai 
>> <yukuai1@huaweicloud.com> wrote:
>>> From: Yu Kuai <yukuai3@huawei.com>
>>>
>>> In order to prevent request_queue to be freed before cleaning up
>>> blktrace debugfs entries, commit db59133e9279 ("scsi: sg: fix blktrace
>>> debugfs entries leakage") use scsi_device_get(), however,
>>> scsi_device_get() will also grab scsi module reference and scsi module
>>> can't be removed.
>>>
>>> It's reported that blktests can't unload scsi_debug after block/001:
>>>
>>> blktests (master) # ./check block
>>> block/001 (stress device hotplugging) [failed]
>>>       +++ /root/blktests/results/nodev/block/001.out.bad 2023-06-19
>>>        Running block/001
>>>        Stressing sd
>>>       +modprobe: FATAL: Module scsi_debug is in use.
>>>
>>> Fix this problem by grabbing request_queue reference directly, so that
>>> scsi host module can still be unloaded while request_queue will be
>>> pinged by sg device.
>>>
>>> Reported-by: Chaitanya Kulkarni <chaitanyak@nvidia.com>
>>> Link: 
>>> https://lore.kernel.org/all/1760da91-876d-fc9c-ab51-999a6f66ad50@nvidia.com/ 
>>>
>>> Fixes: db59133e9279 ("scsi: sg: fix blktrace debugfs entries leakage")
>>> Signed-off-by: Yu Kuai <yukuai3@huawei.com>
>>> ---
>>>   drivers/scsi/sg.c | 6 +++---
>>>   1 file changed, 3 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/drivers/scsi/sg.c b/drivers/scsi/sg.c
>>> index 2433eeef042a..dcb73787c29d 100644
>>> --- a/drivers/scsi/sg.c
>>> +++ b/drivers/scsi/sg.c
>>> @@ -1497,7 +1497,7 @@ sg_add_device(struct device *cl_dev)
>>>       int error;
>>>       unsigned long iflags;
>>> -    error = scsi_device_get(scsidp);
>>> +    error = blk_get_queue(scsidp->request_queue);
>>>       if (error)
>>>           return error;
>>> @@ -1558,7 +1558,7 @@ sg_add_device(struct device *cl_dev)
>>>   out:
>>>       if (cdev)
>>>           cdev_del(cdev);
>>> -    scsi_device_put(scsidp);
>>> +    blk_put_queue(scsidp->request_queue);
>>>       return error;
>>>   }
>>> @@ -1575,7 +1575,7 @@ sg_device_destroy(struct kref *kref)
>>>        */
>>>       blk_trace_remove(q);
>>> -    scsi_device_put(sdp->device);
>>> +    blk_put_queue(q);
>>>       write_lock_irqsave(&sg_index_lock, flags);
>>>       idr_remove(&sg_index_idr, sdp->index);
>>> -- 
>>> 2.39.2
>>
>> Hi,
>>
>> This change (bisected) triggers a regression in our KVM on s390x CI. The
>> symptom is that a “scsi_debug device” does not bind to the scsi_generic
>> driver. On s390x you can reproduce the problem as follows (I have not
>> tested on x86):
>>
>> With this patch applied:
>>
>> $ sudo modprobe scsi_debug
>> $ # Get the 'scsi_host,channel,target_number,LUN' tuple for the 
>> scsi_debug device
>> $ lsscsi |grep scsi_debug |awk '{ print $1 }'
>> [0:0:0:0]
>> $ sudo stat /sys/bus/scsi/devices/0:0:0:0/scsi_generic
>> stat: cannot statx '/sys/bus/scsi/devices/0:0:0:0/scsi_generic': No 
>> such file or directory
>>
>>
>> Patch reverted:
>>
> 
> I didn't figure out the root cause, howver, have you tried to reviert
> this patch as well?
> 
> db59133e9279 ("scsi: sg: fix blktrace debugfs entries leakage"

Never mind this, root cause is that the checking of return value of
blk_get_queue() is wrong.

This shoud be fixed by following patch:

diff --git a/drivers/scsi/sg.c b/drivers/scsi/sg.c
index 89fa046c7158..0d8afffd1683 100644
--- a/drivers/scsi/sg.c
+++ b/drivers/scsi/sg.c
@@ -1497,9 +1497,10 @@ sg_add_device(struct device *cl_dev)
         int error;
         unsigned long iflags;

-       error = blk_get_queue(scsidp->request_queue);
-       if (error)
-               return error;
+       if (!blk_get_queue(scsidp->request_queue)) {
+               pr_warn("%s: get scsi_device queue failed\n", __func__);
+               return -ENODEV;
+       }

         error = -ENOMEM;
         cdev = cdev_alloc();

> 
> Thanks,
> Kuai
>> $ sudo modprobe scsi_debug
>> $ lsscsi |grep scsi_debug |awk '{ print $1 }'
>> [0:0:0:0]
>> $ sudo stat /sys/bus/scsi/devices/0:0:0:0/scsi_generic
>>    File: /sys/bus/scsi/devices/0:0:0:0/scsi_generic
>>    Size: 0             Blocks: 0          IO Block: 4096   directory
>> Device: 0,20    Inode: 12155       Links: 3
>> …
>>
>> Any ideas?
>>
>>   Marc
>> .
>>
> 
> .
>
Marc Hartmayer July 5, 2023, 6:13 a.m. UTC | #7
On Wed, Jul 05, 2023 at 10:28 AM +0800, Yu Kuai <yukuai1@huaweicloud.com> wrote:
> Hi,
>
> 在 2023/07/05 9:43, Yu Kuai 写道:
>> Hi,
>> 
>> 在 2023/07/05 1:04, Marc Hartmayer 写道:
>>> On Thu, Jun 22, 2023 at 12:01 AM +0800, Yu Kuai 
>>> <yukuai1@huaweicloud.com> wrote:
>>>> From: Yu Kuai <yukuai3@huawei.com>
>>>>
>>>> In order to prevent request_queue to be freed before cleaning up
>>>> blktrace debugfs entries, commit db59133e9279 ("scsi: sg: fix blktrace
>>>> debugfs entries leakage") use scsi_device_get(), however,
>>>> scsi_device_get() will also grab scsi module reference and scsi module
>>>> can't be removed.
>>>>
>>>> It's reported that blktests can't unload scsi_debug after block/001:
>>>>
>>>> blktests (master) # ./check block
>>>> block/001 (stress device hotplugging) [failed]
>>>>       +++ /root/blktests/results/nodev/block/001.out.bad 2023-06-19
>>>>        Running block/001
>>>>        Stressing sd
>>>>       +modprobe: FATAL: Module scsi_debug is in use.
>>>>
>>>> Fix this problem by grabbing request_queue reference directly, so that
>>>> scsi host module can still be unloaded while request_queue will be
>>>> pinged by sg device.
>>>>
>>>> Reported-by: Chaitanya Kulkarni <chaitanyak@nvidia.com>
>>>> Link: 
>>>> https://lore.kernel.org/all/1760da91-876d-fc9c-ab51-999a6f66ad50@nvidia.com/ 
>>>>
>>>> Fixes: db59133e9279 ("scsi: sg: fix blktrace debugfs entries leakage")
>>>> Signed-off-by: Yu Kuai <yukuai3@huawei.com>
>>>> ---
>>>>   drivers/scsi/sg.c | 6 +++---
>>>>   1 file changed, 3 insertions(+), 3 deletions(-)
>>>>
>>>> diff --git a/drivers/scsi/sg.c b/drivers/scsi/sg.c
>>>> index 2433eeef042a..dcb73787c29d 100644
>>>> --- a/drivers/scsi/sg.c
>>>> +++ b/drivers/scsi/sg.c
>>>> @@ -1497,7 +1497,7 @@ sg_add_device(struct device *cl_dev)
>>>>       int error;
>>>>       unsigned long iflags;
>>>> -    error = scsi_device_get(scsidp);
>>>> +    error = blk_get_queue(scsidp->request_queue);
>>>>       if (error)
>>>>           return error;
>>>> @@ -1558,7 +1558,7 @@ sg_add_device(struct device *cl_dev)
>>>>   out:
>>>>       if (cdev)
>>>>           cdev_del(cdev);
>>>> -    scsi_device_put(scsidp);
>>>> +    blk_put_queue(scsidp->request_queue);
>>>>       return error;
>>>>   }
>>>> @@ -1575,7 +1575,7 @@ sg_device_destroy(struct kref *kref)
>>>>        */
>>>>       blk_trace_remove(q);
>>>> -    scsi_device_put(sdp->device);
>>>> +    blk_put_queue(q);
>>>>       write_lock_irqsave(&sg_index_lock, flags);
>>>>       idr_remove(&sg_index_idr, sdp->index);
>>>> -- 
>>>> 2.39.2
>>>
>>> Hi,
>>>
>>> This change (bisected) triggers a regression in our KVM on s390x CI. The
>>> symptom is that a “scsi_debug device” does not bind to the scsi_generic
>>> driver. On s390x you can reproduce the problem as follows (I have not
>>> tested on x86):
>>>
>>> With this patch applied:
>>>
>>> $ sudo modprobe scsi_debug
>>> $ # Get the 'scsi_host,channel,target_number,LUN' tuple for the 
>>> scsi_debug device
>>> $ lsscsi |grep scsi_debug |awk '{ print $1 }'
>>> [0:0:0:0]
>>> $ sudo stat /sys/bus/scsi/devices/0:0:0:0/scsi_generic
>>> stat: cannot statx '/sys/bus/scsi/devices/0:0:0:0/scsi_generic': No 
>>> such file or directory
>>>
>>>
>>> Patch reverted:
>>>
>> 
>> I didn't figure out the root cause, howver, have you tried to reviert
>> this patch as well?
>> 
>> db59133e9279 ("scsi: sg: fix blktrace debugfs entries leakage"
>
> Never mind this, root cause is that the checking of return value of
> blk_get_queue() is wrong.
>
> This shoud be fixed by following patch:
>
> diff --git a/drivers/scsi/sg.c b/drivers/scsi/sg.c
> index 89fa046c7158..0d8afffd1683 100644
> --- a/drivers/scsi/sg.c
> +++ b/drivers/scsi/sg.c
> @@ -1497,9 +1497,10 @@ sg_add_device(struct device *cl_dev)
>          int error;
>          unsigned long iflags;
>
> -       error = blk_get_queue(scsidp->request_queue);
> -       if (error)
> -               return error;
> +       if (!blk_get_queue(scsidp->request_queue)) {
> +               pr_warn("%s: get scsi_device queue failed\n", __func__);
> +               return -ENODEV;
> +       }

Hi Kuai,

I just tried your fix and it works - thanks.

Tested-by: Marc Hartmayer <mhartmay@linux.ibm.com>

 Marc

[…snip]
diff mbox series

Patch

diff --git a/drivers/scsi/sg.c b/drivers/scsi/sg.c
index 2433eeef042a..dcb73787c29d 100644
--- a/drivers/scsi/sg.c
+++ b/drivers/scsi/sg.c
@@ -1497,7 +1497,7 @@  sg_add_device(struct device *cl_dev)
 	int error;
 	unsigned long iflags;
 
-	error = scsi_device_get(scsidp);
+	error = blk_get_queue(scsidp->request_queue);
 	if (error)
 		return error;
 
@@ -1558,7 +1558,7 @@  sg_add_device(struct device *cl_dev)
 out:
 	if (cdev)
 		cdev_del(cdev);
-	scsi_device_put(scsidp);
+	blk_put_queue(scsidp->request_queue);
 	return error;
 }
 
@@ -1575,7 +1575,7 @@  sg_device_destroy(struct kref *kref)
 	 */
 
 	blk_trace_remove(q);
-	scsi_device_put(sdp->device);
+	blk_put_queue(q);
 
 	write_lock_irqsave(&sg_index_lock, flags);
 	idr_remove(&sg_index_idr, sdp->index);