Re: [PATCH v4 06/20] coresight: add try_get_module() in coresight_grab_device()

[Tingwei Zhang <tingweiz@codeaurora.org>]

Hi Mike,

On Tue, Jul 28, 2020 at 01:04:21AM +0800, Mike Leach wrote:
> Hi Tingwei,
> 
> On Sun, 26 Jul 2020 at 02:32, Tingwei Zhang <tingweiz@codeaurora.org>
> wrote:
> >
> > Hi Mike,
> > On Sat, Jul 25, 2020 at 09:51:31PM +0800, Mike Leach wrote:
> > > Hi Tingwei,
> > >
> > > On Sat, 25 Jul 2020 at 11:05, Tingwei Zhang <tingweiz@codeaurora.org>
> > > wrote:
> > > >
> > > > Hi Mike,
> > > >
> > > >  Thu, Jul 23, 2020 at 06:55:47PM +0800, Mike Leach wrote:
> > > > > Hi Tingwei,
> > > > >
> > > > > On Thu, 23 Jul 2020 at 05:28, Tingwei Zhang
> <tingwei@codeaurora.org>
> > > > > wrote:
> > > > > >
> > > > > > When coresight device is in an active session, driver module of
> > > > > > that device should not be removed. Use try_get_module() in
> > > > > > coresight_grab_device() to prevent module to be unloaded.
> > > > > >
> > > > > > Signed-off-by: Tingwei Zhang <tingwei@codeaurora.org>
> > > > > > ---
> > > > > >  drivers/hwtracing/coresight/coresight.c | 27
> > > +++++++++++++++++++++----
> > > > > >  1 file changed, 23 insertions(+), 4 deletions(-)
> > > > > >
> > > > > > diff --git a/drivers/hwtracing/coresight/coresight.c
> > > > > b/drivers/hwtracing/coresight/coresight.c
> > > > > > index b7151c5f81b1..17bc76ea86ae 100644
> > > > > > --- a/drivers/hwtracing/coresight/coresight.c
> > > > > > +++ b/drivers/hwtracing/coresight/coresight.c
> > > > > > @@ -640,7 +640,7 @@ struct coresight_device
> > > > > *coresight_get_sink_by_id(u32 id)
> > > > > >   * don't appear on the trace path, they should be handled along
> > > with
> > > > > the
> > > > > >   * the master device.
> > > > > >   */
> > > > > > -static void coresight_grab_device(struct coresight_device
> *csdev)
> > > > > > +static int coresight_grab_device(struct coresight_device
> *csdev)
> > > > > >  {
> > > > > >         int i;
> > > > > >
> > > > > > @@ -648,10 +648,25 @@ static void coresight_grab_device(struct
> > > > > coresight_device *csdev)
> > > > > >                 struct coresight_device *child;
> > > > > >
> > > > > >                 child  = csdev->pdata->conns[i].child_dev;
> > > > > > -               if (child && child->type ==
> > > CORESIGHT_DEV_TYPE_HELPER)
> > > > > > +               if (child && child->type ==
> > > CORESIGHT_DEV_TYPE_HELPER) {
> > > > > > +                       if
> > > > > (!try_module_get(child->dev.parent->driver->owner))
> > > > > > +                               goto err;
> > > > > >                         pm_runtime_get_sync(child->dev.parent);
> > > > > > +               }
> > > > > >         }
> > > > > > +       if (!try_module_get(csdev->dev.parent->driver->owner))
> > > > > > +               goto err;
> > > > > >         pm_runtime_get_sync(csdev->dev.parent);
> > > > > > +       return 0;
> > > > > > +err:
> > > > > > +       for (i--; i >= 0; i--) {
> > > > > > +               struct coresight_device *child;
> > > > > > +
> > > > > > +               child  = csdev->pdata->conns[i].child_dev;
> > > > > > +               if (child && child->type ==
> > > CORESIGHT_DEV_TYPE_HELPER)
> > > > > > +                       pm_runtime_put(child->dev.parent);
> > > > > > +       }
> > > > > > +       return -ENODEV;
> > > > > >  }
> > > > > >
> > > > > >  /*
> > > > > > @@ -663,12 +678,15 @@ static void coresight_drop_device(struct
> > > > > coresight_device *csdev)
> > > > > >         int i;
> > > > > >
> > > > > >         pm_runtime_put(csdev->dev.parent);
> > > > > > +       module_put(csdev->dev.parent->driver->owner);
> > > > > >         for (i = 0; i < csdev->pdata->nr_outport; i++) {
> > > > > >                 struct coresight_device *child;
> > > > > >
> > > > > >                 child  = csdev->pdata->conns[i].child_dev;
> > > > > > -               if (child && child->type ==
> > > CORESIGHT_DEV_TYPE_HELPER)
> > > > > > +               if (child && child->type ==
> > > CORESIGHT_DEV_TYPE_HELPER) {
> > > > > >                         pm_runtime_put(child->dev.parent);
> > > > > > +
> > > module_put(child->dev.parent->driver->owner);
> > > > > > +               }
> > > > > >         }
> > > > > >  }
> > > > > >
> > > > > > @@ -721,7 +739,8 @@ static int _coresight_build_path(struct
> > > > > coresight_device *csdev,
> > > > > >         if (!node)
> > > > > >                 return -ENOMEM;
> > > > > >
> > > > > > -       coresight_grab_device(csdev);
> > > > > > +       if (coresight_grab_device(csdev))
> > > > > > +               return -ENODEV;
> > > > > >         node->csdev = csdev;
> > > > > >         list_add(&node->link, path);
> > > > > >
> > > > > > --
> > > > > > The Qualcomm Innovation Center, Inc. is a member of the Code
> Aurora
> > > > > Forum,
> > > > > > a Linux Foundation Collaborative Project
> > > > > >
> > > > >
> > > > > The CTI devices are associated with other coresight components
> using
> > > > > csdev->ect_dev; These are not handled in the main linear trace
> path as
> > > > > these have a star topology interlinking all components.
> > > > > However, when a component has an associated ect_dev then is
> enabled at
> > > > > the same time as the other component is enabled.
> > > > >
> > > > > So a module_get/put will be needed for this association to prevent
> the
> > > > > CTI being removed while a trace session is active.
> > > > > I suggest in coresight.c : coresight_control_assoc_ectdev()
> > > > >
> > > >
> > > > In module unload process, devices of that module will be removed. In
> > > > this case, cti_remove() is called on all cti device when unloading
> > > > cti module. All cti connections is cleaned at that time.
> > > > csdev->ect_dev is set to NULL. coresight_control_assoc_ectdev()
> > > > will return since since ect_csdev is NULL.
> > > >
> > > > I think we are safe without holding module reference since cti
> > > > driver has done a pretty good job on clean up.
> > > >
> > >
> > > The issue here is not about clean up. We need to keep all the
> > > programmed coresight modules loaded and operational while a coresight
> > > trace session is in progress. The CTI can be used to control the
> > > generation of trace, trigger trace related events etc.
> > > If the module is removed while a session is in progress then this
> > > programming is lost and the trace data collected may not be what is
> > > expected.
> > >
> >
> > Got your point now. In my opinion, CTI is kinda different to other
> > coresight components since it's optional.
> >
> > For other components, they have to been there when path is built or
> > the enablement fails. Use can either get a successfully return which
> > means it's good and all devices/modules are there until path is
> > released, or a failed return which means trace session can't be
> > setup.
> >
> 
> The module get/put do more than ensure that the system can be enabled.
> They ensure that no components in the coresight path can be unloaded
> until all paths in use are released. This ensures that clients do not
> get the trace system pulled out from underneath - corrupting trace
> data, and possibly affecting the client program.
>

I agree.
 
> > In CTI case, there's no garrantee that CTI related to the trace
> > session is there even the enablement is successful.
> 
> And there is no guarantee that it is not programmed to perform a key
> task in controlling the generation of trace.
> 
> The entire programmed coresight system must be protected when in use.
> That includes CTI devices. Otherwise the integrity of the captured
> trace could be compromised, and indeed there may be no trace captured
> if this is relying on CTI triggers.
> 

That's correct.  If CTI is not there, the trace may be wrong or no
trace as all.

> >For example,
> > One cti is required for ETM tracing session. If cti module is
> > unloaded before enable that trace session. The enablement will
> > still be successfully done since there's no ect_dev connected
> > at all.
> >
> > My point is holding cti module reference in enable path won't
> > fix all the problem. Altenatively, user should be aware that
> > unload cti module leads to all cti functioniol failure.
> > With current design, the behavior is consistant on cti. User
> > can unload cti module whenever he wants but it will break
> > the function.
> 
> Precisely - and this is what should never be allowed - why can the CTI
> be used to break the trace system but not other components?
> 

We can't prevent user to unload CTI module when there's no active
trace session. If we don't allow CTI to be not available, we need to
check whether CTI is there in build_path routine or coresight enable
routine. We need return failure if CTI is not there when user enables
the trace session. Currently, coresight framework check associate CTI
device, but if CTI deivce is not probed due to probe failure or
module unload, coresight framework assume there's no assocaite CTI
device and just continue the enablement.  It won't return failure.

If we have plan to change CTI check to be mandatory, make the coresight
enablement fail if CTI is not there like other trace components.  I'm
perfect fine to add module_get/module_put for CTI.

> Moving forwards we will see users who simply use pre-programmed
> settings from client programs such as perf. It may be that some users
> do not have the credentials to add and remove modules - this could be
> an admin function where larger systems with multiple sessions are in
> use.
> 
> 
> Regards
> 
> Mike
> 
> 
> >
> > Thanks, Tingwei
> >
> > > Regards
> > >
> > > Mike
> > >
> > > > Let me know if you still think we need hold the module reference.
> > > >
> > > > Thanks, Tingwei
> > > > > Regards
> > > > >
> > > > > Mike
> > > > >
> > > > >
> > > > > --
> > > > > Mike Leach
> > > > > Principal Engineer, ARM Ltd.
> > > > > Manchester Design Centre. UK
> > > > >
> > > > > _______________________________________________
> > > > > linux-arm-kernel mailing list
> > > > > linux-arm-kernel@lists.infradead.org
> > > > > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
> > >
> > >
> > >
> > > --
> > > Mike Leach
> > > Principal Engineer, ARM Ltd.
> > > Manchester Design Centre. UK
> > >
> > > _______________________________________________
> > > linux-arm-kernel mailing list
> > > linux-arm-kernel@lists.infradead.org
> > > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
> 
> 
> 
> -- 
> Mike Leach
> Principal Engineer, ARM Ltd.
> Manchester Design Centre. UK
> 
> _______________________________________________
> linux-arm-kernel mailing list
> linux-arm-kernel@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel