Hi Mike, On Tue, Jul 28, 2020 at 06:08:37PM +0800, Mike Leach wrote:
Hi Tingwei,
On Tue, 28 Jul 2020 at 02:07, Tingwei Zhang tingweiz@codeaurora.org wrote:
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
> 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?
Sorry - I may not have been clear here - I do not want to make having a CTI mandatory for enabling the trace session - this is an optional component.
We can't prevent user to unload CTI module when there's no active trace session.
Agreed - the same applies to all the other components.
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.
If the CTI is present at the time the trace path is created, then it must be enabled, and a failure to enable the CTI fails the entire path enable - this is the operation that is implemented by coresight_control_assoc_ectdev() at this time. If a CTI is not present at the time the trace path is created, then as you have said previously, the ect_dev pointer will be NULL and the check and enable of CTI will be skipped by coresight_control_assoc_ectdev() and have no effect on the trace path.
For the module version, if CTI is present and enabled then the module must be locked so it cannot be removed while the trace is active - this is the only change that is required. So adding get / put module in coresight_control_assoc_ectdev() will lock the module in for the duration that trace is active.
I've tested the code below to ensure that the cti module is correctly locked while trace is active, but can be removed as required when trace is inactive.
Thanks a lot for coming up with the patch and testing it, Mike.
Regards
Mike
static int coresight_control_assoc_ectdev(struct coresight_device *csdev, bool enable) { int ect_ret = 0; struct coresight_device *ect_csdev = csdev->ect_dev; struct module *mod;
if (!ect_csdev) return 0; if ((!ect_ops(ect_csdev)->enable) || (!ect_ops(ect_csdev)->disable)) return 0; mod = ect_csdev->dev.parent->driver->owner; if (enable) { if (try_module_get(mod)) { ect_ret = ect_ops(ect_csdev)->enable(ect_csdev); if (ect_ret) module_put(mod); } else ect_ret = -ENODEV; } else { module_put(mod); ect_ret = ect_ops(ect_csdev)->disable(ect_csdev); }
If CTI module is removed before enablement of ETM and installed after ETM enablement and before ETM disablement. We have a unbalanced module reference number here. I made one test to prove this with above code.
We may need move module_put() after disable(). Return proper error from disable() to indicate CTI was not enabled before so we can skip module_put().
Tests: db845c:/sys/bus/coresight/devices # echo 1 > tmc_etr0/enable_sink db845c:/sys/bus/coresight/devices # echo 1 > etm1/enable_source db845c:/sys/bus/coresight/devices # rmmod coresight_cti rmmod: failed to unload coresight_cti: Try again 1|db845c:/sys/bus/coresight/devices # echo 0 > etm1/enable_source db845c:/sys/bus/coresight/devices # rmmod coresight_cti db845c:/sys/bus/coresight/devices # echo 1 > etm1/enable_source db845c:/sys/bus/coresight/devices # insmod /data/modules/coresight-cti.ko db845c:/sys/bus/coresight/devices # echo 1 > etm2/enable_source db845c:/sys/bus/coresight/devices # echo 0 > etm1/enable_source db845c:/sys/bus/coresight/devices # rmmod coresight_cti db845c:/sys/bus/coresight/devices #
Above module removal should fail since etm2 is still active now.
/* output warning if ECT enable is preventing trace operation */ if (ect_ret) dev_info(&csdev->dev, "Associated ECT device (%s) %s failed\n", dev_name(&ect_csdev->dev), enable ? "enable" : "disable"); return ect_ret;}
Tests:-
root@linaro-developer:/home/linaro/scripts# ./trace-sysfs-start.bash Tracing via sysfs enabling etf wait before enable etm enabling etm 0. enabling etm 1. enabling etm 2. enabling etm 3. waiting for trace root@linaro-developer:/home/linaro/scripts# rmmod ../cs-mods/coresight-etm4x.ko rmmod: ERROR: Module coresight_etm4x is in use root@linaro-developer:/home/linaro/scripts# rmmod ../cs-mods/coresight-cti.ko rmmod: ERROR: Module coresight_cti is in use root@linaro-developer:/home/linaro/scripts# ./trace-sysfs-end.bash ending trace disabling etm 0. disabling etm 1. disabling etm 2. disabling etm 3. disabling etf 16+0 records in 16+0 records out 8192 bytes (8.2 kB) copied, 0.000834116 s, 9.8 MB/s root@linaro-developer:/home/linaro/scripts# rmmod ../cs-mods/coresight-cti.ko root@linaro-developer:/home/linaro/scripts# rmmod ../cs-mods/coresight-etm4x.ko root@linaro-developer:/home/linaro/scripts# insmod ../cs-mods/coresight-etm4x.ko root@linaro-developer:/home/linaro/scripts# ./trace-sysfs-start.bash Tracing via sysfs enabling etf wait before enable etm enabling etm 0. enabling etm 1. enabling etm 2. enabling etm 3. waiting for trace root@linaro-developer:/home/linaro/scripts# rmmod ../cs-mods/coresight-etm4x.ko rmmod: ERROR: Module coresight_etm4x is in use root@linaro-developer:/home/linaro/scripts# rmmod ../cs-mods/coresight-cti.ko rmmod: ERROR: Module coresight_cti is not currently loaded root@linaro-developer:/home/linaro/scripts# ./trace-sysfs-end.bash ending trace disabling etm 0. disabling etm 1. disabling etm 2. disabling etm 3. disabling etf 16+0 records in 16+0 records out 8192 bytes (8.2 kB) copied, 0.000819117 s, 10.0 MB/s
}
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
-- Mike Leach Principal Engineer, ARM Ltd. Manchester Design Centre. UK