Merge branch 'fs-file-descriptor-optimization'

Merge file descriptor allocation speedup.

Eric Dumazet has a test-case for a fairly common network deamon load
pattern: openign and closing a lot of sockets that each have very little
work done on them.  It turns out that in that case, the cost of just
finding the correct file descriptor number can be a dominating factor.

We've long had a trivial optimization for allocating file descriptors
sequentially, but that optimization ends up being not very effective
when other file descriptors are being closed concurrently, and the fd
patterns are not some simple FIFO pattern.  In such cases we ended up
spending a lot of time just scanning the bitmap of open file descriptors
in order to find the next file descriptor number to open.

This trivial patch-series mitigates that by simply introducing a
second-level bitmap of which words in the first bitmap are already fully
allocated.  That cuts down the cost of scanning by an order of magnitude
in some pathological (but realistic) cases.

The second patch is an even more trivial patch to avoid unnecessarily
dirtying the cacheline for the close-on-exec bit array that normally
ends up being all empty.

* fs-file-descriptor-optimization:
  vfs: conditionally clear close-on-exec flag
  vfs: Fix pathological performance case for __alloc_fd()

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];
 };