Re: [PATCH 3/3] fs/file.c: move sanity_check from alloc_fd() to put_unused_fd()

[Mateusz Guzik]

On Sat, Jun 15, 2024 at 06:41:45AM +0200, Mateusz Guzik wrote:
> On Fri, Jun 14, 2024 at 12:34:16PM -0400, Yu Ma wrote:
> > alloc_fd() has a sanity check inside to make sure the FILE object mapping to the
> 
> Total nitpick: FILE is the libc thing, I would refer to it as 'struct
> file'. See below for the actual point.
> 
> > Combined with patch 1 and 2 in series, pts/blogbench-1.1.0 read improved by
> > 32%, write improved by 15% on Intel ICX 160 cores configuration with v6.8-rc6.
> > 
> > Reviewed-by: Tim Chen <tim.c.chen@xxxxxxxxxxxxxxx>
> > Signed-off-by: Yu Ma <yu.ma@xxxxxxxxx>
> > ---
> >  fs/file.c | 14 ++++++--------
> >  1 file changed, 6 insertions(+), 8 deletions(-)
> > 
> > diff --git a/fs/file.c b/fs/file.c
> > index a0e94a178c0b..59d62909e2e3 100644
> > --- a/fs/file.c
> > +++ b/fs/file.c
> > @@ -548,13 +548,6 @@ static int alloc_fd(unsigned start, unsigned end, unsigned flags)
> >  	else
> >  		__clear_close_on_exec(fd, fdt);
> >  	error = fd;
> > -#if 1
> > -	/* Sanity check */
> > -	if (rcu_access_pointer(fdt->fd[fd]) != NULL) {
> > -		printk(KERN_WARNING "alloc_fd: slot %d not NULL!\n", fd);
> > -		rcu_assign_pointer(fdt->fd[fd], NULL);
> > -	}
> > -#endif
> >  
> 
> I was going to ask when was the last time anyone seen this fire and
> suggest getting rid of it if enough time(tm) passed. Turns out it does
> show up sometimes, latest result I found is 2017 vintage:
> https://groups.google.com/g/syzkaller-bugs/c/jfQ7upCDf9s/m/RQjhDrZ7AQAJ
> 
> So you are moving this to another locked area, but one which does not
> execute in the benchmark?
> 
> Patch 2/3 states 28% read and 14% write increase, this commit message
> claims it goes up to 32% and 15% respectively -- pretty big. I presume
> this has to do with bouncing a line containing the fd.
> 
> I would argue moving this check elsewhere is about as good as removing
> it altogether, but that's for the vfs overlords to decide.
> 
> All that aside, looking at disasm of alloc_fd it is pretty clear there
> is time to save, for example:
> 
> 	if (unlikely(nr >= fdt->max_fds)) {
> 		if (fd >= end) {
> 			error = -EMFILE;
> 			goto out;
> 		}
> 		error = expand_files(fd, fd);
> 		if (error < 0)
> 			goto out;
> 		if (error)
> 			goto repeat;
> 	}
> 

Now that I wrote it I noticed the fd < end check has to be performed
regardless of max_fds -- someone could have changed rlimit to a lower
value after using a higher fd. But the main point stands: the call to
expand_files and associated error handling don't have to be there.

> This elides 2 branches and a func call in the common case. Completely
> untested, maybe has some brainfarts, feel free to take without credit
> and further massage the routine.
> 
> Moreover my disasm shows that even looking for a bit results in
> a func call(!) to _find_next_zero_bit -- someone(tm) should probably
> massage it into another inline.
> 
> After the above massaging is done and if it turns out the check has to
> stay, you can plausibly damage-control it with prefetch -- issue it
> immediately after finding the fd number, before any other work.
> 
> All that said, by the above I'm confident there is still *some*
> performance left on the table despite the lock.
> 
> >  out:
> >  	spin_unlock(&files->file_lock);
> > @@ -572,7 +565,7 @@ int get_unused_fd_flags(unsigned flags)
> >  }
> >  EXPORT_SYMBOL(get_unused_fd_flags);
> >  
> > -static void __put_unused_fd(struct files_struct *files, unsigned int fd)
> > +static inline void __put_unused_fd(struct files_struct *files, unsigned int fd)
> >  {
> >  	struct fdtable *fdt = files_fdtable(files);
> >  	__clear_open_fd(fd, fdt);
> > @@ -583,7 +576,12 @@ static void __put_unused_fd(struct files_struct *files, unsigned int fd)
> >  void put_unused_fd(unsigned int fd)
> >  {
> >  	struct files_struct *files = current->files;
> > +	struct fdtable *fdt = files_fdtable(files);
> >  	spin_lock(&files->file_lock);
> > +	if (unlikely(rcu_access_pointer(fdt->fd[fd]))) {
> > +		printk(KERN_WARNING "put_unused_fd: slot %d not NULL!\n", fd);
> > +		rcu_assign_pointer(fdt->fd[fd], NULL);
> > +	}
> >  	__put_unused_fd(files, fd);
> >  	spin_unlock(&files->file_lock);
> >  }