Re: [RFC PATCH 1/2] tracing/user_events: Use remote writes for event enablement

[Masami Hiramatsu (Google) <mhiramat@kernel.org>]

On Mon, 31 Oct 2022 09:46:03 -0700
Beau Belgrave <beaub@linux.microsoft.com> wrote:

> On Mon, Oct 31, 2022 at 11:47:03PM +0900, Masami Hiramatsu wrote:
> > Hi,
> > 
> > I have some comments.
> > 
> > On Thu, 27 Oct 2022 15:40:10 -0700
> > Beau Belgrave <beaub@linux.microsoft.com> wrote:
> > 
> > [...]
> > > @@ -1570,11 +1610,12 @@ static long user_reg_get(struct user_reg __user *ureg, struct user_reg *kreg)
> > >   * Registers a user_event on behalf of a user process.
> > >   */
> > >  static long user_events_ioctl_reg(struct user_event_file_info *info,
> > > -				  unsigned long uarg)
> > > +				  struct file *file, unsigned long uarg)
> > >  {
> > >  	struct user_reg __user *ureg = (struct user_reg __user *)uarg;
> > >  	struct user_reg reg;
> > >  	struct user_event *user;
> > > +	struct user_event_enabler *enabler;
> > >  	char *name;
> > >  	long ret;
> > >  
> > > @@ -1607,8 +1648,12 @@ static long user_events_ioctl_reg(struct user_event_file_info *info,
> > >  	if (ret < 0)
> > >  		return ret;
> > >  
> > > +	enabler = user_event_enabler_create(file, &reg, user);
> > > +
> > > +	if (!enabler)
> > 
> > Shouldn't we free the user_event if needed here?
> > (I found the similar memory leak pattern in the above failure case
> >  of the user_events_ref_add().)
> > 
> 
> user_events are shared across the entire group. They cannot be cleaned
> up until all references are gone. This is true both in this case and the
> in the user_events_ref_add() case.
> 
> The pattern is to register events in the group's hashtable, then add
> them to the local file ref array that is RCU protected. If the file ref
> cannot be allocated, etc. the refcount on user is decremented. If we
> cannot create an enabler, the refcount is still held until file release.

OK, when the ioctl returns, there should be 3 cases;

- Return success, a new(existing) user_event added.
- Return error, no new user_event added.
- Return error, a new user_event added but enabler is not initialized.

And in the last case, the new user_event will be released when user
closes the file. Could you comment it here?

> 
> If the event has already been added to the local file ref array, it is
> returned to prevent another reference.

I'm not sure this point. Did you mean returning an error to prevent
registering the same event again?


> 
> > > +		return -ENOMEM;
> > > +
> > >  	put_user((u32)ret, &ureg->write_index);
> > > -	put_user(user->index, &ureg->status_bit);
> > >  
> > >  	return 0;
> > >  }
> > [...]
> > > @@ -1849,7 +1863,6 @@ static int user_status_open(struct inode *node, struct file *file)
> > >  
> > >  static const struct file_operations user_status_fops = {
> > >  	.open = user_status_open,
> > > -	.mmap = user_status_mmap,
> > 
> > So, if this drops the mmap operation, can we drop the writable flag from
> > the status tracefs file?
> > 
> 
> Good catch, yes I'll remove this.

Thanks!

> 
> > static int create_user_tracefs(void)
> > {
> > [...]
> >         /* mmap with MAP_SHARED requires writable fd */
> >         emmap = tracefs_create_file("user_events_status", TRACE_MODE_WRITE,
> >                                     NULL, NULL, &user_status_fops);
> > 
> > Thank you,
> > 
> > -- 
> > Masami Hiramatsu (Google) <mhiramat@kernel.org>
> 
> Thanks,
> -Beau


-- 
Masami Hiramatsu (Google) <mhiramat@kernel.org>