[PATCH v2 1/1] scsi: Fix racing between dev init and dev reset

[Miles Chen]

Hi Alice, 

>Hi Miles,
>
>On Wed, 2022-03-09 at 07:23 +0800, Miles Chen wrote:
>> Hi Alice,
>> 
>> Thanks for your patch. I have some questions.
>> 
>> > Device reset thread uses kobject_uevent_env() to get kobj.parent
>> > after scsi_evt_emit(), and it races with device init thread which
>> > calls device_add() to create kobj.parent before
>> > kobject_uevent_env().
>> > 
>> > Device reset call trace:
>> > fill_kobj_path
>> > kobject_get_path
>> > kobject_uevent_env
>> > scsi_evt_emit			<- add wait_event()
>> > scsi_evt_thread
>> > 
>> > Device init call trace:
>> > fill_kobj_path
>> > kobject_get_path
>> > kobject_uevent_env
>> > device_add				<- create kobj.parent
>> > scsi_target_add
>> > scsi_sysfs_add_sdev
>> > scsi_add_lun
>> > scsi_probe_and_add_lun
>> 
>> These are backtraces, and the things above looks like your solution,
>> not
>> the problem.
>> It's better to describe the problem first and solution later:
>> 
>> Device init call:                              Device reset call:   
>> scsi_probe_and_add_lun()                       scsi_evt_thread()
>>   scsi_add_lun()
>>     scsi_sysfs_add_sdev()                        scsi_evt_emit()
>>       scsi_target_add()                            kobject_uevent_env
>> () 
>>         device_add()                                 kobject_get_path
>> () // something wrong?
>> 	                                               fill_kobj_path()
>> 	  get_device_parent() // I see a get, not create parent kobj,
>> could you show me that point?
>> 	  kobject_uevent()
>>             kobject_uevent_env()
>>               kobject_get_path()
>>                 fill_kobj_path()
>> 
>
>Device init call:	   Device reset call:
>scsi_probe_and_add_lun()   scsi_evt_thread()
>  scsi_add_lun()	    scsi_evt_emit()
>   scsi_sysfs_add_sdev()     kobject_uevent_env() //get kobj.parent
>    scsi_target_add()	      kobject_get_path() //get wrong
>kobj.parent     device_add() // add kobj.parent  fill_kobj_path()
>      kobject_uevent_env()
>       kobject_get_path()
>        fill_kobj_path()
>Above backtrace describes the problem, device reset thread will get
>wrong kobj.parent when device init thread didn’t add kobj.parent yet.

Thanks for sharing the error log, I think it is clearer this way:

Device init call:	           Device reset call:
scsi_probe_and_add_lun()              scsi_evt_thread()
  scsi_add_lun()	                scsi_evt_emit()
   scsi_sysfs_add_sdev()                  kobject_uevent_env() //get kobj.parent
    scsi_target_add()	                    kobject_get_path()
                                              len = get_kobj_path_length () // len=1 because parent is not create yet     
    device_add() // add kobj.parent               
      kobject_uevent_env()                 
       kobject_get_path()                     path = kzalloc()
        fill_kobj_path()	              fill_kobj_path() // --length; length -= cur is a negative value
	                                        memcpy(path + length, kobject_name(parent), cur); // slab OOB!

Now I understood this slab OOB issue, please add the kasan log and the
analysis to the commit message so the maintainers can see it and help.

>
>> > 
>> > These two jobs are scheduled asynchronously, we can't guaranteed
>> > that
>> > kobj.parent will be created in device init thread before device
>> > reset
>> > thread calls kobj_get_path().
>> 
>> What happens if device reset thread calls kobj_get_path() before
>> kobj.parent's
>> creation? a crash or something?
>> 
>
>It will cause slab-out-of-bounds in kobject_get_path and kernel panic
>than crash.

Okay, please add the KASAN slab-out-of-bounds error in the commit message.

Thanks,
Miles
>
>> > 
>> > To resolve the racing issue between device init thread and device
>> > reset thread, we use wait_event() in scsi_evt_emit() to wait for
>> > device_add() to complete the creation of kobj.parent.
>> > 
>> > Signed-off-by: Alice Chao <alice.chao at mediatek.com>
>> 
>> When we send patch N+1, we have to put the change log between patch
>> v1 to vN
>> under a "---", so others can understand the changes between patch N
>> and N+1.
>> e.g.,
>> 
>> ---
>> 
>> Change since v1
>> - remove Change-Id
>> 
>> > ---
>> >  drivers/scsi/scsi_lib.c  | 1 +
>> >  drivers/scsi/scsi_scan.c | 1 +
>> >  2 files changed, 2 insertions(+)
>> > 
>> > diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
>> > index 0a70aa763a96..abf9a71ed77c 100644
>> > --- a/drivers/scsi/scsi_lib.c
>> > +++ b/drivers/scsi/scsi_lib.c
>> > @@ -2461,6 +2461,7 @@ static void scsi_evt_emit(struct scsi_device
>> > *sdev, struct scsi_event *evt)
>> >  		break;
>> >  	case SDEV_EVT_POWER_ON_RESET_OCCURRED:
>> >  		envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
>> > +		wait_event(sdev->host->host_wait, sdev-
>> > >sdev_gendev.kobj.parent != NULL);
>> >  		break;
>> >  	default:
>> >  		/* do nothing */
>> > diff --git a/drivers/scsi/scsi_scan.c b/drivers/scsi/scsi_scan.c
>> > index f4e6c68ac99e..431f229ac435 100644
>> > --- a/drivers/scsi/scsi_scan.c
>> > +++ b/drivers/scsi/scsi_scan.c
>> > @@ -1904,6 +1904,7 @@ static void do_scsi_scan_host(struct
>> > Scsi_Host *shost)
>> >  	} else {
>> >  		scsi_scan_host_selected(shost, SCAN_WILD_CARD,
>> > SCAN_WILD_CARD,
>> >  				SCAN_WILD_CARD, 0);
>> > +		wake_up(&shost->host_wait);
>> 
>> do_scsi_scan_host() is not in the call trace above, could you show
>> the relationship
>> of do_scsi_scan_host in the call flow?
>> 
>
>ufshcd_async_scan()
>scsi_scan_host()
>do_scsi_scan_host()  << here
>scsi_scan_host_selected()
> __scsi_scan_target()
>scsi_probe_and_add_lun()
>scsi_sysfs_add_sdev()
>device_add()
>kobject_uevent_env()
>kobject_get_path()
>fill_kobj_path()
>
>After we add wake_up here, we can ensure that device reset thread will
>get kobject after device init thread finishes adding parent.
>
>> thanks,
>> Miles
>> 
>> >  	}
>> >  }
>
>