Re: [PATCH] coresight: cti: write regsiters directly in cti_enable_hw()

From: Tingwei Zhang
Date: Thu Aug 27 2020 - 23:18:46 EST


On Fri, Aug 28, 2020 at 02:12:53AM +0800, Mathieu Poirier wrote:
> Hi Tingwei,
>
> On Tue, Aug 18, 2020 at 07:10:57PM +0800, Tingwei Zhang wrote:
> > Deadlock as below is triggered by one CPU holds drvdata->spinlock
> > and calls cti_enable_hw(). Smp_call_function_single() is called
> > in cti_enable_hw() and tries to let another CPU write CTI registers.
> > That CPU is trying to get drvdata->spinlock in cti_cpu_pm_notify()
> > and doesn't response to IPI from smp_call_function_single().
> >
> > [ 988.335937] CPU: 6 PID: 10258 Comm: sh Tainted: G W L
> > 5.8.0-rc6-mainline-16783-gc38daa79b26b-dirty #1
> > [ 988.346364] Hardware name: Thundercomm Dragonboard 845c (DT)
> > [ 988.352073] pstate: 20400005 (nzCv daif +PAN -UAO BTYPE=--)
> > [ 988.357689] pc : smp_call_function_single+0x158/0x1b8
> > [ 988.362782] lr : smp_call_function_single+0x124/0x1b8
> > ...
> > [ 988.451638] Call trace:
> > [ 988.454119] smp_call_function_single+0x158/0x1b8
> > [ 988.458866] cti_enable+0xb4/0xf8 [coresight_cti]
> > [ 988.463618] coresight_control_assoc_ectdev+0x6c/0x128 [coresight]
> > [ 988.469855] coresight_enable+0x1f0/0x364 [coresight]
> > [ 988.474957] enable_source_store+0x5c/0x9c [coresight]
> > [ 988.480140] dev_attr_store+0x14/0x28
> > [ 988.483839] sysfs_kf_write+0x38/0x4c
> > [ 988.487532] kernfs_fop_write+0x1c0/0x2b0
> > [ 988.491585] vfs_write+0xfc/0x300
> > [ 988.494931] ksys_write+0x78/0xe0
> > [ 988.498283] __arm64_sys_write+0x18/0x20
> > [ 988.502240] el0_svc_common+0x98/0x160
> > [ 988.506024] do_el0_svc+0x78/0x80
> > [ 988.509377] el0_sync_handler+0xd4/0x270
> > [ 988.513337] el0_sync+0x164/0x180
> >
>
> Was this the full log or you did cut some of it?
>

I cut some CPU registers' value since it's too long and not relevant.
The Call trace is full.

> > This change write CTI registers directly in cti_enable_hw().
> > Config->hw_powered has been checked to be true with spinlock holded.
> > CTI is powered and can be programmed until spinlock is released.
> >
>
> From your explanation above it seems that cti_enable_hw() was called from,
> say
> CPUy, to enable the CTI associated to CPUx. CTIx's drvdata->spinlock was
> taken
> and smp_call_function_single() called right after. That woke up CPUx and
> cti_cpu_pm_notify() was executed on CPUx in interrupt context, trying to
> take
> CTIx's drvdata->spinlock. That hung CPUx and the kernel got angry. Is my
> assessment correct?
>

Most of them is correct. The only difference is CPUx is power on when
cti_enable_hw() is called. Otherwise it will goto cti_state_unchanged:
and won't call cti_enable_hw_smp_call(). cti_cpu_pm_notify() is called
when CPUx tries to suspend instead of resume.

> If so I don't think the fix suggested in this patch will work. The same
> condition will happen whenever cti_enable_hw() is called on a CPU to
> enable a
> CTI that belongs to another CPU and that cti_cpu_pm_notify() is called on
> latter
> CPU at the same time.
>

I'm not sure I understand this correctly. Let me clarify it a little bit.
It's a deadlock since cti_enable_hw() holds the spinlock and calls
cti_enable_hw_smp_call() from CPUx to enable CTI associated to CPUy. It
waits for cti_enable_hw_smp_call() to return. IPI is sent to CPUy while
CPUy is in cti_cpu_pm_notify() and waits for spinlock. In this patch,
I remove cti_enable_hw_smp_call() and write CTI CPU directly on CPUx.
It won't wait for CPUy and release spinlock after program registers of
CTI. After cti_enable_hw() releases spinlock, cti_cpu_pm_notify() will
continue to run. Since spinlock is held and config->hw_powered is true,
we don't need to worry about CPUy power down when we program CTI on CPUx.

> I think a better solution is to grab the lock in cti_enable_hw() and check
> the
> value of ->ctidev.cpu. If not a global CPU, i.e >= 0, then release the
> lock and
> call smp_call_function_single(). In cti_enable_hw_smp_call() take the
> lock
> again and move forward from there.
>

After cti_enable_hw() releases the lock, it's possible that CPU is offline
by user, cti_enable_hw_smp_call() will fail in this case.



> I have applied the other two patches in this set so no need to send them
> again.
>
Thanks,
Tingwei
> Thanks,
> Mathieu
>
> > Fixes: 6a0953ce7de9 ("coresight: cti: Add CPU idle pm notifer to CTI
> devices")
> > Signed-off-by: Tingwei Zhang <tingwei@xxxxxxxxxxxxxx>
> > ---
> > drivers/hwtracing/coresight/coresight-cti.c | 17 +----------------
> > 1 file changed, 1 insertion(+), 16 deletions(-)
> >
> > diff --git a/drivers/hwtracing/coresight/coresight-cti.c
> b/drivers/hwtracing/coresight/coresight-cti.c
> > index 3ccc703dc940..869569eb8c7f 100644
> > --- a/drivers/hwtracing/coresight/coresight-cti.c
> > +++ b/drivers/hwtracing/coresight/coresight-cti.c
> > @@ -86,13 +86,6 @@ void cti_write_all_hw_regs(struct cti_drvdata
> *drvdata)
> > CS_LOCK(drvdata->base);
> > }
> >
> > -static void cti_enable_hw_smp_call(void *info)
> > -{
> > - struct cti_drvdata *drvdata = info;
> > -
> > - cti_write_all_hw_regs(drvdata);
> > -}
> > -
> > /* write regs to hardware and enable */
> > static int cti_enable_hw(struct cti_drvdata *drvdata)
> > {
> > @@ -112,15 +105,7 @@ static int cti_enable_hw(struct cti_drvdata
> *drvdata)
> > if (rc)
> > goto cti_err_not_enabled;
> >
> > - if (drvdata->ctidev.cpu >= 0) {
> > - rc = smp_call_function_single(drvdata->ctidev.cpu,
> > - cti_enable_hw_smp_call,
> > - drvdata, 1);
> > - if (rc)
> > - goto cti_err_not_enabled;
> > - } else {
> > - cti_write_all_hw_regs(drvdata);
> > - }
> > + cti_write_all_hw_regs(drvdata);
> >
> > config->hw_enabled = true;
> > atomic_inc(&drvdata->config.enable_req_count);
> > --
> > The Qualcomm Innovation Center, Inc. is a member of the Code Aurora
> Forum,
> > a Linux Foundation Collaborative Project
> >
>
> _______________________________________________
> linux-arm-kernel mailing list
> linux-arm-kernel@xxxxxxxxxxxxxxxxxxx
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel