[PATCH] Optimize select/poll by putting small data sets on the stack

Optimize select and poll by a using stack space for small fd sets

This brings back an old optimization from Linux 2.0.  Using the stack is
faster than kmalloc.  On a Intel P4 system it speeds up a select of a
single pty fd by about 13% (~4000 cycles -> ~3500)

It also saves memory because a daemon hanging in select or poll will
usually save one or two less pages.  This can add up - e.g.  if you have 10
daemons blocking in poll/select you save 40KB of memory.

I did a patch for this long ago, but it was never applied.  This version is
a reimplementation of the old patch that tries to be less intrusive.  I
only did the minimal changes needed for the stack allocation.

The cut off point before external memory is allocated is currently at
832bytes.  The system calls always allocate this much memory on the stack.

These 832 bytes are divided into 256 bytes frontend data (for the select
bitmaps of the pollfds) and the rest of the space for the wait queues used
by the low level drivers.  There are some extreme cases where this won't
work out for select and it falls back to allocating memory too early -
especially with very sparse large select bitmaps - but the majority of
processes who only have a small number of file descriptors should be ok.
[TBD: 832/256 might not be the best split for select or poll]

I suspect more optimizations might be possible, but they would be more
complicated.  One way would be to cache the select/poll context over
multiple system calls because typically the input values should be similar.
 Problem is when to flush the file descriptors out though.

Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

1 files changed, 17 insertions, 0 deletions

diff --git a/include/linux/poll.h b/include/linux/poll.h
index 8e8f6098508a..51e1b56741fb 100644
--- a/include/linux/poll.h
+++ b/include/linux/poll.h
@@ -11,6 +11,15 @@
 #include <linux/mm.h>
 #include <asm/uaccess.h>
 
+/* ~832 bytes of stack space used max in sys_select/sys_poll before allocating
+   additional memory. */
+#define MAX_STACK_ALLOC 832
+#define FRONTEND_STACK_ALLOC	256
+#define SELECT_STACK_ALLOC	FRONTEND_STACK_ALLOC
+#define POLL_STACK_ALLOC	FRONTEND_STACK_ALLOC
+#define WQUEUES_STACK_ALLOC	(MAX_STACK_ALLOC - FRONTEND_STACK_ALLOC)
+#define N_INLINE_POLL_ENTRIES	(WQUEUES_STACK_ALLOC / sizeof(struct poll_table_entry))
+
 struct poll_table_struct;
 
 /* 
@@ -33,6 +42,12 @@ static inline void init_poll_funcptr(poll_table *pt, poll_queue_proc qproc)
 	pt->qproc = qproc;
 }
 
+struct poll_table_entry {
+	struct file * filp;
+	wait_queue_t wait;
+	wait_queue_head_t * wait_address;
+};
+
 /*
  * Structures and helpers for sys_poll/sys_poll
  */
@@ -40,6 +55,8 @@ struct poll_wqueues {
 	poll_table pt;
 	struct poll_table_page * table;
 	int error;
+	int inline_index;
+	struct poll_table_entry inline_entries[N_INLINE_POLL_ENTRIES];
 };
 
 extern void poll_initwait(struct poll_wqueues *pwq);