Re: [PATCH 1/3] coresight: ultrasoc-smb: fix sleep while close preempt in enable_smb

[Jonathan Cameron]

On Thu, 12 Oct 2023 17:47:04 +0800
Junhao He <hejunhao3@xxxxxxxxxx> wrote:

> When we to enable the SMB by perf, the perf sched will call perf_ctx_lock()
> to close system preempt in event_function_call(). But SMB::enable_smb() use
> mutex to lock the critical section, which may sleep.
> 
>  BUG: sleeping function called from invalid context at kernel/locking/mutex.c:580
>  in_atomic(): 1, irqs_disabled(): 1, non_block: 0, pid: 153023, name: perf
>  preempt_count: 2, expected: 0
>  RCU nest depth: 0, expected: 0
>  INFO: lockdep is turned off.
>  irq event stamp: 0
>  hardirqs last  enabled at (0): [<0000000000000000>] 0x0
>  hardirqs last disabled at (0): [<ffffa2983f5c5f40>] copy_process+0xae8/0x2b48
>  softirqs last  enabled at (0): [<ffffa2983f5c5f40>] copy_process+0xae8/0x2b48
>  softirqs last disabled at (0): [<0000000000000000>] 0x0
>  CPU: 2 PID: 153023 Comm: perf Kdump: loaded Tainted: G   W  O   6.5.0-rc4+ #1
> 
>  Call trace:
>  ...
>   __mutex_lock+0xbc/0xa70
>   mutex_lock_nested+0x34/0x48
>   smb_update_buffer+0x58/0x360 [ultrasoc_smb]
>   etm_event_stop+0x204/0x2d8 [coresight]
>   etm_event_del+0x1c/0x30 [coresight]
>   event_sched_out+0x17c/0x3b8
>   group_sched_out.part.0+0x5c/0x208
>   __perf_event_disable+0x15c/0x210
>   event_function+0xe0/0x230
>   remote_function+0xb4/0xe8
>   generic_exec_single+0x160/0x268
>   smp_call_function_single+0x20c/0x2a0
>   event_function_call+0x20c/0x220
>   _perf_event_disable+0x5c/0x90
>   perf_event_for_each_child+0x58/0xc0
>   _perf_ioctl+0x34c/0x1250
>   perf_ioctl+0x64/0x98
> ...
> 
> Use spinlock replace mutex to control driver data access to one at a
> time. But the function copy_to_user() may sleep so spinlock do not to
> lock it.

I'd like to see a comment on why we no longer need to lock over the copy_to_user
rather than simply that we can't.

> 
> Fixes: 06f5c2926aaa ("drivers/coresight: Add UltraSoc System Memory Buffer driver")
> Signed-off-by: Junhao He <hejunhao3@xxxxxxxxxx>

A follow up patch could change a lot of this to use the new cleanup.h (don't want that
in the fix though as will make back porting trickier.).
That should let you do
guard(spin_lock)(&drvdata->spinlock);
and then use direct returns instead of goto complexity.



Jonathan

> @@ -132,10 +132,8 @@ static ssize_t smb_read(struct file *file, char __user *data, size_t len,
>  	if (!len)
>  		return 0;
>  
> -	mutex_lock(&drvdata->mutex);
> -
>  	if (!sdb->data_size)
> -		goto out;
> +		return 0;
>  
>  	to_copy = min(sdb->data_size, len);
>  
> @@ -145,20 +143,18 @@ static ssize_t smb_read(struct file *file, char __user *data, size_t len,
>  
>  	if (copy_to_user(data, sdb->buf_base + sdb->buf_rdptr, to_copy)) {
>  		dev_dbg(dev, "Failed to copy data to user\n");
> -		to_copy = -EFAULT;
> -		goto out;
> +		return -EFAULT;
>  	}
>  
> +	spin_lock(&drvdata->spinlock);
>  	*ppos += to_copy;
> -

Unrelated white space change that shouldn't be here.

>  	smb_update_read_ptr(drvdata, to_copy);
>  
> -	dev_dbg(dev, "%zu bytes copied\n", to_copy);
> -out:
>  	if (!sdb->data_size)
>  		smb_reset_buffer(drvdata);
> -	mutex_unlock(&drvdata->mutex);
> +	spin_unlock(&drvdata->spinlock);
>  
> +	dev_dbg(dev, "%zu bytes copied\n", to_copy);
>  	return to_copy;
>  }