vfs: Fix pathological performance case for __alloc_fd()

Al Viro points out that:
> >     * [Linux-specific aside] our __alloc_fd() can degrade quite badly
> > with some use patterns.  The cacheline pingpong in the bitmap is probably
> > inevitable, unless we accept considerably heavier memory footprint,
> > but we also have a case when alloc_fd() takes O(n) and it's _not_ hard
> > to trigger - close(3);open(...); will have the next open() after that
> > scanning the entire in-use bitmap.

And Eric Dumazet has a somewhat realistic multithreaded microbenchmark
that opens and closes a lot of sockets with minimal work per socket.

This patch largely fixes it.  We keep a 2nd-level bitmap of the open
file bitmaps, showing which words are already full.  So then we can
traverse that second-level bitmap to efficiently skip already allocated
file descriptors.

On his benchmark, this improves performance by up to an order of
magnitude, by avoiding the excessive open file bitmap scanning.

Tested-and-acked-by: Eric Dumazet <edumazet@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 2 insertions, 0 deletions

diff --git a/include/linux/fdtable.h b/include/linux/fdtable.h
index 674e3e226465..5295535b60c6 100644
--- a/include/linux/fdtable.h
+++ b/include/linux/fdtable.h
@@ -26,6 +26,7 @@ struct fdtable {
 	struct file __rcu **fd;      /* current fd array */
 	unsigned long *close_on_exec;
 	unsigned long *open_fds;
+	unsigned long *full_fds_bits;
 	struct rcu_head rcu;
 };
 
@@ -59,6 +60,7 @@ struct files_struct {
 	int next_fd;
 	unsigned long close_on_exec_init[1];
 	unsigned long open_fds_init[1];
+	unsigned long full_fds_bits_init[1];
 	struct file __rcu * fd_array[NR_OPEN_DEFAULT];
 };