Re: [PATCH v12 15/17] counter: Implement events_queue_size sysfs attribute

From: David Lechner
Date: Sat Jul 10 2021 - 13:54:23 EST


On 7/10/21 11:08 AM, David Lechner wrote:
On 7/10/21 10:43 AM, David Lechner wrote:
On 7/10/21 5:25 AM, William Breathitt Gray wrote:
and after that

[   16.564403] ================================================
[   16.570725] WARNING: lock held when returning to user space!
[   16.577044] 5.13.0-next-20210706+ #4 Not tainted
[   16.582198] ------------------------------------------------
[   16.588507] cat/331 is leaving the kernel with locks still held!
[   16.595214] 1 lock held by cat/331:
[   16.599103]  #0: ffff888102bb3630
(&counter->chrdev_lock){+.+.}-{3:3}, at: counter_chrdev_open+0x21/0x60
[counter]

Jarkko
I'm not sure how to resolve this warning. The purpose of this lock is to
limit chrdev to a single open at a time. To accomplish this I grab this
lock in counter_chrdev_open() and hold it until counter_chrdev_release()
is called. Is there a better way to accomplish this?

How about using an atomic flag, e.g test_and_set_bit()?

Another option could be to rethink it at a higher level and avoid the
need for a lock (and sysfs attribute) altogether. For example, would it
work to (re)allocate the kfifo buffer in the COUNTER_ENABLE_EVENTS_IOCTL
callback and add a parameter to that ioctl to specify the buffer size
(with units of events rather than bytes)?


Thinking about it a bit more, this probably isn't a good suggestion.
It would create issues with being able to start reading from the chrdev
before enabling events.