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

From: Mathieu Poirier
Date: Thu Aug 27 2020 - 14:12:59 EST


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?

> 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?

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 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.

I have applied the other two patches in this set so no need to send them again.

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
>