[RFC] bpf: change bpf syacall to use u64 temp variables

Subsystems: bpf [core], bpf [general] (safe dynamic programs and tools), the rest

3 messages, 3 authors, 2015-10-19 · open the first message on its own page

[RFC] bpf: change bpf syacall to use u64 temp variables

From: yalin wang <hidden>
Date: 2015-10-19 07:10:58

This patch change map_lookup_elem() and map_update_elem() function
to use u64 temp variable if the key_size or value_size is less than
u64, we don't need use kmalloc() for these small variables.

Signed-off-by: yalin wang <redacted>
---
 kernel/bpf/syscall.c | 30 ++++++++++++++++++++----------
 1 file changed, 20 insertions(+), 10 deletions(-)

diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
index f640e5f..c82d7bf 100644
--- a/kernel/bpf/syscall.c
+++ b/kernel/bpf/syscall.c

@@ -189,7 +189,8 @@ static int map_lookup_elem(union bpf_attr *attr)
 	void __user *uvalue = u64_to_ptr(attr->value);
 	int ufd = attr->map_fd;
 	struct bpf_map *map;
-	void *key, *value, *ptr;
+	u64 key_buf, value_buf;
+	void *key = &key_buf, *value = &value_buf, *ptr;
 	struct fd f;
 	int err;

@@ -202,7 +203,8 @@ static int map_lookup_elem(union bpf_attr *attr)
 		return PTR_ERR(map);
 
 	err = -ENOMEM;
-	key = kmalloc(map->key_size, GFP_USER);
+	if (map->key_size > sizeof(u64))
+		key = kmalloc(map->key_size, GFP_USER);
 	if (!key)
 		goto err_put;

@@ -211,7 +213,8 @@ static int map_lookup_elem(union bpf_attr *attr)
 		goto free_key;
 
 	err = -ENOMEM;
-	value = kmalloc(map->value_size, GFP_USER);
+	if (map->value_size > sizeof(u64))
+		value = kmalloc(map->value_size, GFP_USER);
 	if (!value)
 		goto free_key;

@@ -232,9 +235,11 @@ static int map_lookup_elem(union bpf_attr *attr)
 	err = 0;
 
 free_value:
-	kfree(value);
+	if (value != &value_buf)
+		kfree(value);
 free_key:
-	kfree(key);
+	if (key != &key_buf)
+		kfree(key);
 err_put:
 	fdput(f);
 	return err;

@@ -248,7 +253,8 @@ static int map_update_elem(union bpf_attr *attr)
 	void __user *uvalue = u64_to_ptr(attr->value);
 	int ufd = attr->map_fd;
 	struct bpf_map *map;
-	void *key, *value;
+	u64 key_buf, value_buf;
+	void *key = &key_buf, *value = &value_buf;
 	struct fd f;
 	int err;

@@ -261,7 +267,8 @@ static int map_update_elem(union bpf_attr *attr)
 		return PTR_ERR(map);
 
 	err = -ENOMEM;
-	key = kmalloc(map->key_size, GFP_USER);
+	if (map->key_size > sizeof(u64))
+		key = kmalloc(map->key_size, GFP_USER);
 	if (!key)
 		goto err_put;

@@ -270,7 +277,8 @@ static int map_update_elem(union bpf_attr *attr)
 		goto free_key;
 
 	err = -ENOMEM;
-	value = kmalloc(map->value_size, GFP_USER);
+	if (map->value_size > sizeof(u64))
+		value = kmalloc(map->value_size, GFP_USER);
 	if (!value)
 		goto free_key;

@@ -286,9 +294,11 @@ static int map_update_elem(union bpf_attr *attr)
 	rcu_read_unlock();
 
 free_value:
-	kfree(value);
+	if (value != &value_buf)
+		kfree(value);
 free_key:
-	kfree(key);
+	if (key != &key_buf)
+		kfree(key);
 err_put:
 	fdput(f);
 	return err;

-- 
1.9.1

Re: [RFC] bpf: change bpf syacall to use u64 temp variables

From: Daniel Borkmann <daniel@iogearbox.net>
Date: 2015-10-19 08:55:28

On 10/19/2015 09:10 AM, yalin wang wrote:

This patch change map_lookup_elem() and map_update_elem() function
to use u64 temp variable if the key_size or value_size is less than
u64, we don't need use kmalloc() for these small variables.

Signed-off-by: yalin wang <redacted>

 From an application PoV that has to make the bpf(2) syscall, how
much do we actually gain from this? I'm curious, did you perform
some benchmarks that show a noticeable difference?

Re: [RFC] bpf: change bpf syacall to use u64 temp variables

From: Alexei Starovoitov <hidden>
Date: 2015-10-19 19:08:39

On Mon, Oct 19, 2015 at 03:10:46PM +0800, yalin wang wrote:

quoted hunk

This patch change map_lookup_elem() and map_update_elem() function
to use u64 temp variable if the key_size or value_size is less than
u64, we don't need use kmalloc() for these small variables.

Signed-off-by: yalin wang <redacted>
---
 kernel/bpf/syscall.c | 30 ++++++++++++++++++++----------
 1 file changed, 20 insertions(+), 10 deletions(-)

diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
index f640e5f..c82d7bf 100644
--- a/kernel/bpf/syscall.c
+++ b/kernel/bpf/syscall.c

@@ -189,7 +189,8 @@ static int map_lookup_elem(union bpf_attr *attr)
 	void __user *uvalue = u64_to_ptr(attr->value);
 	int ufd = attr->map_fd;
 	struct bpf_map *map;
-	void *key, *value, *ptr;
+	u64 key_buf, value_buf;
+	void *key = &key_buf, *value = &value_buf, *ptr;
 	struct fd f;
 	int err;

@@ -202,7 +203,8 @@ static int map_lookup_elem(union bpf_attr *attr)
 		return PTR_ERR(map);
 
 	err = -ENOMEM;
-	key = kmalloc(map->key_size, GFP_USER);
+	if (map->key_size > sizeof(u64))
+		key = kmalloc(map->key_size, GFP_USER);

I think it's a good optimization for common case.
Performance numbers would be good to prove the point.
Thanks

`h`	back out one level
`j`	next message in thread
`k`	previous message in thread
`l`	drill in
`Esc`	close help / fold thread tree
`?`	toggle this help