Re: [PATCH] files: Use rcu lock to get the file structures for better performance

[Matthew Wilcox <willy@infradead.org>]

On Thu, May 21, 2020 at 08:38:35PM +0800, Muchun Song wrote:
> +++ b/fs/proc/fd.c
> @@ -34,19 +34,27 @@ static int seq_show(struct seq_file *m, void *v)
>  	if (files) {
>  		unsigned int fd = proc_fd(m->private);
>  
> -		spin_lock(&files->file_lock);
> +		rcu_read_lock();
> +again:
>  		file = fcheck_files(files, fd);
>  		if (file) {
> -			struct fdtable *fdt = files_fdtable(files);
> +			struct fdtable *fdt;
> +
> +			if (!get_file_rcu(file)) {
> +				/*
> +				 * we loop to catch the new file (or NULL
> +				 * pointer).
> +				 */
> +				goto again;
> +			}
>  
> +			fdt = files_fdtable(files);

This is unusual, and may not be safe.

fcheck_files() loads files->fdt.  Then it loads file from fdt->fd[].
Now you're loading files->fdt again here, and it could have been changed
by another thread expanding the fd table.

You have to write a changelog which convinces me you've thought about
this race and that it's safe.  Because I don't think you even realise
it's a possibility at this point.

> @@ -160,14 +168,23 @@ static int proc_fd_link(struct dentry *dentry, struct path *path)
>  		unsigned int fd = proc_fd(d_inode(dentry));
>  		struct file *fd_file;
>  
> -		spin_lock(&files->file_lock);
> +		rcu_read_lock();
> +again:
>  		fd_file = fcheck_files(files, fd);
>  		if (fd_file) {
> +			if (!get_file_rcu(fd_file)) {
> +				/*
> +				 * we loop to catch the new file
> +				 * (or NULL pointer).
> +				 */
> +				goto again;
> +			}
>  			*path = fd_file->f_path;
>  			path_get(&fd_file->f_path);
> +			fput(fd_file);
>  			ret = 0;
>  		}
> -		spin_unlock(&files->file_lock);
> +		rcu_read_unlock();

Why is it an improvement to increment/decrement the refcount on the
struct file here, rather than take/release the spinlock?