PM-runtime: supplier looses track of consumer during probe

[Tushar Nimkar <quic_tnimkar@quicinc.com>]

Hi linux-pm/linux-scsi,

We have included fix [1] but continuing to observe supplier loosing 
track of consumer.

Below is trace snippet with additional logging added.
Here consumer is 0:0:0:0 and supplier is 0:0:0:49488. In Last three 
lines consumer resume is completed but supplier is put down.

    kworker/u16:0-7     0.880014: rpm_idle:             0:0:0:0 flags-4 
cnt-0  dep-0  auto-1 p-0 irq-0 child-0
    kworker/u16:0-7     0.880017: bprint: 
pm_runtime_mark_last_busy.46700: :#205 dev_name:0:0:0:0 
ktime_get_mono_fast_ns():852365364
    kworker/u16:0-7     0.880019: rpm_suspend:          0:0:0:0 flags-8 
cnt-0  dep-0  auto-1 p-0 irq-0 child-0
    kworker/u16:0-7     0.880022: bprint: pm_runtime_put_noidle.44083: 
pm_runtime_put_noidle: #112 dev_name:0:0:0:49488 dev usage_count:5 
decremented usage count
    kworker/u16:0-7     0.880023: bprint: pm_runtime_put_noidle.44083: 
pm_runtime_put_noidle: #112 dev_name:0:0:0:49488 dev usage_count:4 
decremented usage count
    kworker/u16:2-142   0.880024: rpm_resume:           0:0:0:0 flags-4 
cnt-1  dep-0  auto-1 p-0 irq-0 child-0
    kworker/u16:0-7     0.880025: bprint: __rpm_put_suppliers: 
__rpm_put_suppliers: #348 consumer:0:0:0:0 supplier:0:0:0:49488 
usage_count:4
    kworker/u16:0-7     0.880061: rpm_idle:             0:0:0:49488 
flags-1 cnt-4  dep-0  auto-1 p-0 irq-0 child-0
    kworker/u16:0-7     0.880062: rpm_return_int: 
rpm_idle+0x16c:0:0:0:49488 ret=-11
    kworker/u16:2-142   0.880062: bprint: __pm_runtime_resume: 
__pm_runtime_resume: #1147 dev_name:0:0:0:49488 dev usage_count:5 
incremented usage count
    kworker/u16:2-142   0.880063: rpm_resume:           0:0:0:49488 
flags-4 cnt-5  dep-0  auto-1 p-0 irq-0 child-0
    kworker/u16:2-142   0.880063: rpm_return_int: 
rpm_resume+0x690:0:0:0:49488 ret=1
    kworker/u16:0-7     0.880063: rpm_return_int: 
rpm_suspend+0x68:0:0:0:0 ret=0
    kworker/u16:2-142   0.880063: bprint: rpm_get_suppliers: 
rpm_get_suppliers: #300 consumer:0:0:0:0 supplier:0:0:0:49488 usage_count:5
    kworker/u16:0-7     0.880065: bprint: pm_runtime_put_noidle.44083: 
pm_runtime_put_noidle: #112 dev_name:0:0:0:49488 dev usage_count:4 
decremented usage count
    kworker/u16:2-142   0.880065: bprint: 
pm_runtime_mark_last_busy.44088: :#205 dev_name:0:0:0:0 
ktime_get_mono_fast_ns():852413749
    kworker/u16:2-142   0.880065: rpm_idle:             0:0:0:0 flags-1 
cnt-1  dep-0  auto-1 p-0 irq-0 child-0
    kworker/u16:2-142   0.880065: rpm_return_int: rpm_idle+0x16c:0:0:0:0 
ret=-11
    kworker/u16:0-7     0.880066: bprint: __rpm_put_suppliers: 
__rpm_put_suppliers: #348 consumer:0:0:0:0 supplier:0:0:0:49488 
usage_count:4
    kworker/u16:0-7     0.880067: rpm_return_int: rpm_idle+0x16c:0:0:0:0 
ret=-16
    kworker/u16:2-142   0.880067: rpm_return_int: 
rpm_resume+0x690:0:0:0:0 ret=0

Upon looking into this further the race looks to be in below two 
processes running in parallel and process-1 is putting down supplier at 
[C] because process-2 is setting runtime_status as resuming at [D].

Also as per runtime PM documentation
In order to use autosuspend, subsystems or drivers must call
pm_runtime_use_autosuspend(), and thereafter they should use the various 
`*_autosuspend()` helper functions...

It was also observed that *_autosuspend() API at point [A] was invoked 
without first invoking pm_runtime_use_autosuspend() which return 
expiration as zero at point [B] and proceeds ahead for immediate runtime 
suspend of device which seems lead to this race condition.

Process -1
ufshcd_async_scan context (process 1)
scsi_autopm_put_device() //0:0:0:0
pm_runtime_put_sync()
__pm_runtime_idle()
rpm_idle() -- RPM_GET_PUT(4)
     __rpm_callback
         scsi_runtime_idle()
             pm_runtime_mark_last_busy()
             pm_runtime_autosuspend()  --[A]
                 rpm_suspend() -- RPM_AUTO(8)
                     pm_runtime_autosuspend_expiration() 
use_autosuspend    is false return 0   --- [B]
                         __update_runtime_status to RPM_SUSPENDING
                     __rpm_callback()
                         __rpm_put_suppliers(dev, false)
                     __update_runtime_status to RPM_SUSPENDED
                 rpm_suspend_suppliers()
                     rpm_idle() for supplier -- RPM_ASYNC(1) return 
(-EAGAIN) [ Other consumer active for supplier]
                 rpm_suspend() – END with return=0
         scsi_runtime_idle() END return (-EBUSY) always.
      /* Do that if resume fails too.*/
     (dev->power.runtime_status == RPM_RESUMING && retval)))  return -EBUSY
         __rpm_put_suppliers(dev, false)  -- [C]
rpm_idle() END return -EBUSY

Process -2
sd_probe context (Process 2)
scsi_autopm_get_device() //0:0:0:0
__pm_runtime_resume(RPM_GET_PUT)
rpm_resume() -- RPM_GET_PUT(4)
     __update_runtime_status to RPM_RESUMING --[D]
     __rpm_callback()
         rpm_get_suppliers()
             __pm_runtime_resume() - RPM_GET_PUT(4) – supplier
                 rpm_resume() for supplier.
     __update_runtime_status to RPM_ACTIVE
     pm_runtime_mark_last_busy ()
rpm_resume() END return 0

Can you please provide your suggestions on addressing above race condition?

This is also reported at [2].

[1]: https://lore.kernel.org/lkml/4748074.GXAFRqVoOG@kreacher/T/
[2]: https://lkml.org/lkml/2022/10/12/259

Thanks,
Tushar Nimkar