Skip to contentSkip to conversation

bpf: support BPF_PROG_QUERY for progs attached to sockmap

PATCH zhudi (E) 2 participants 2 messages Nov 3, 2021 BPF

lore
Share & tools
lore

zhudi (E)

author
#0 lore raw
Quoted 25 lines · Di Zhu
Quoted 7 lines

Right now there is no way to query whether BPF programs are attached to a sockmap or not.

we can use the standard interface in libbpf to query, such as:
bpf_prog_query(mapFd, BPF_SK_SKB_STREAM_PARSER, 0, NULL, ...); the mapFd is the fd of sockmap.

Quoted 5 lines
---
  include/linux/bpf.h  |  9 +++++
  kernel/bpf/syscall.c |  5 +++
  net/core/sock_map.c  | 88

++++++++++++++++++++++++++++++++++++++++----

include/linux/bpf.h int bpf_prog_test_run_syscall(struct bpf_prog

Quoted 7 lines
  3 files changed, 95 insertions(+), 7 deletions(-)
include/linux/bpf.h +0 -0
index d604c8251d88..594ca91992db 100644
@@ -1961,6 +1961,9 @@ int bpf_prog_test_run_syscall(struct bpf_prog

*prog,

include/linux/bpf.h int bpf_prog_test_run_syscall(struct bpf_prog · line 1961

Quoted 1 line
include/linux/bpf.h +0 -0
  int sock_map_get_from_fd(const union bpf_attr *attr, struct bpf_prog

*prog);

include/linux/bpf.h int bpf_prog_test_run_syscall(struct bpf_prog · line 1962

Quoted 1 line
include/linux/bpf.h +0 -0
  int sock_map_prog_detach(const union bpf_attr *attr, enum
Quoted 9 lines

bpf_prog_type ptype);

include/linux/bpf.h int bpf_prog_test_run_syscall(struct bpf_prog · line 1963

Quoted 1 line
include/linux/bpf.h +0 -0
  int sock_map_update_elem_sys(struct bpf_map *map, void *key, void

*value, u64 flags);

include/linux/bpf.h int bpf_prog_test_run_syscall(struct bpf_prog · line 1965

Quoted 2 lines
include/linux/bpf.h +2 -0
+int sockmap_bpf_prog_query(const union bpf_attr *attr,
+			   union bpf_attr __user *uattr);

All previous functions are with prefix "sock_map". Why you choose a different prefix "sockmap"?

Thanks for all your suggestions, I will make changes to the inappropriate code.

Quoted 50 lines

include/linux/bpf.h static inline int

Quoted 5 lines
include/linux/bpf.h +1 -0
+
  void sock_map_unhash(struct sock *sk);
  void sock_map_close(struct sock *sk, long timeout);
  #else
@@ -2014,6 +2017,12 @@ static inline int

sock_map_update_elem_sys(struct bpf_map *map, void *key, void

kernel/bpf/syscall.c static int bpf_prog_query(const union bpf_attr

Quoted 17 lines
include/linux/bpf.h +6 -0
  {
  	return -EOPNOTSUPP;
  }
+
+static inline int sockmap_bpf_prog_query(const union bpf_attr *attr,
+					 union bpf_attr __user *uattr)
+{
+	return -EINVAL;
+}
  #endif /* CONFIG_BPF_SYSCALL */
  #endif /* CONFIG_NET && CONFIG_BPF_SYSCALL */
 
kernel/bpf/syscall.c +0 -0
index 4e50c0bfdb7d..17faeff8f85f 100644
@@ -3275,6 +3275,11 @@ static int bpf_prog_query(const union bpf_attr

*attr,

net/core/sock_map.c static struct sk_psock_progs

Quoted 16 lines
kernel/bpf/syscall.c +5 -0
  	case BPF_FLOW_DISSECTOR:
  	case BPF_SK_LOOKUP:
  		return netns_bpf_prog_query(attr, uattr);
+	case BPF_SK_SKB_STREAM_PARSER:
+	case BPF_SK_SKB_STREAM_VERDICT:
+	case BPF_SK_MSG_VERDICT:
+	case BPF_SK_SKB_VERDICT:
+		return sockmap_bpf_prog_query(attr, uattr);
  	default:
  		return -EINVAL;
  	}
net/core/sock_map.c +0 -0
index e252b8ec2b85..ca65ed0004d3 100644
@@ -1412,38 +1412,50 @@ static struct sk_psock_progs

*sock_map_progs(struct bpf_map *map)

net/core/sock_map.c static struct sk_psock_progs · line 1415

Quoted 4 lines
net/core/sock_map.c +0 -1
  	return NULL;
  }
 
-static int sock_map_prog_update(struct bpf_map *map, struct bpf_prog

*prog,

net/core/sock_map.c static struct sk_psock_progs · line 1415

Quoted 2 lines
net/core/sock_map.c +1 -1
-				struct bpf_prog *old, u32 which)
+static int sock_map_prog_lookup(struct bpf_map *map, struct bpf_prog

**pprog[],

Quoted 7 lines

Can we just change "**pprog[]" to "***pprog"? In the code, you really just pass the address of the decl "struct bpf_prog **pprog;" to the function.

net/core/sock_map.c static struct sk_psock_progs · line 1416

Quoted 1 line
net/core/sock_map.c +1 -0
+				u32 which)

Some format issue here?

Format is right, passed the checkpatch script check.

Quoted 107 lines
Quoted 29 lines
net/core/sock_map.c +3 -5
  {
  	struct sk_psock_progs *progs = sock_map_progs(map);
-	struct bpf_prog **pprog;
 
  	if (!progs)
  		return -EOPNOTSUPP;
 
  	switch (which) {
  	case BPF_SK_MSG_VERDICT:
-		pprog = &progs->msg_parser;
+		*pprog = &progs->msg_parser;
  		break;
  #if IS_ENABLED(CONFIG_BPF_STREAM_PARSER)
  	case BPF_SK_SKB_STREAM_PARSER:
-		pprog = &progs->stream_parser;
+		*pprog = &progs->stream_parser;
  		break;
  #endif
  	case BPF_SK_SKB_STREAM_VERDICT:
  		if (progs->skb_verdict)
  			return -EBUSY;
-		pprog = &progs->stream_verdict;
+		*pprog = &progs->stream_verdict;
  		break;
  	case BPF_SK_SKB_VERDICT:
  		if (progs->stream_verdict)
  			return -EBUSY;
-		pprog = &progs->skb_verdict;

net/core/sock_map.c static struct sk_psock_progs · line 1449

Quoted 10 lines
net/core/sock_map.c +5 -0
+		*pprog = &progs->skb_verdict;
  		break;
  	default:
  		return -EOPNOTSUPP;
  	}
 
+	return 0;
+}
+
+static int sock_map_prog_update(struct bpf_map *map, struct bpf_prog

*prog,

net/core/sock_map.c static struct sk_psock_progs · line 1450

Quoted 1 line
net/core/sock_map.c +1 -0
+				struct bpf_prog *old, u32 which)

Some format issue here?

net/core/sock_map.c static int sock_map_prog_update(struct

Quoted 12 lines
net/core/sock_map.c +8 -0
+{
+	struct bpf_prog **pprog;
+	int ret;
+
+	ret = sock_map_prog_lookup(map, &pprog, which);
+	if (ret)
+		return ret;
+
  	if (old)
  		return psock_replace_prog(pprog, prog, old);
 
@@ -1451,6 +1463,68 @@ static int sock_map_prog_update(struct

bpf_map *map, struct bpf_prog *prog,

net/core/sock_map.c static int sock_map_prog_update(struct · line 1467

Quoted 5 lines
net/core/sock_map.c +2 -0
  	return 0;
  }
 
+int sockmap_bpf_prog_query(const union bpf_attr *attr,
+			   union bpf_attr __user *uattr)

Format issue here?

net/core/sock_map.c static int sock_map_prog_update(struct · line 1469

Quoted 2 lines
net/core/sock_map.c +2 -0
+{
+	__u32 __user *prog_ids = u64_to_user_ptr(attr->query.prog_ids);

Typically we use u32 in the kernel code. But I know there are __u32 usage as well, esp. with __user attributes. I put a comment here just in case that somebody else has a different opinion.

net/core/sock_map.c static int sock_map_prog_update(struct · line 1471

Quoted 2 lines
net/core/sock_map.c +2 -0
+	u32 prog_cnt = 0, flags = 0;
+	u32 ufd = attr->target_fd;

You can merge the above u32 together.

net/core/sock_map.c static int sock_map_prog_update(struct · line 1477

Quoted 6 lines
net/core/sock_map.c +6 -0
+	struct bpf_prog **pprog;
+	struct bpf_prog *prog;
+	struct bpf_map *map;
+	struct fd f;
+	int ret;
+	u32 id = 0;

to maintain reverse christmas tree?

Quoted 21 lines
net/core/sock_map.c +21 -0
+
+	if (attr->query.query_flags)
+		return -EINVAL;
+
+	f = fdget(ufd);
+	map = __bpf_map_get(f);
+	if (IS_ERR(map))
+		return PTR_ERR(map);
+
+	rcu_read_lock();
+
+	ret = sock_map_prog_lookup(map, &pprog, attr->query.attach_type);
+	if (ret)
+		goto end;
+
+	prog = *pprog;
+	prog_cnt = (!prog) ? 0 : 1;
+
+	if (!attr->query.prog_cnt || !prog_ids || !prog_cnt)
+		goto end;
+
Quoted 8 lines

net/core/sock_map.c static int sock_map_prog_update(struct · line 1499

Quoted 1 line
net/core/sock_map.c +1 -0
+	prog = bpf_prog_inc_not_zero(prog);

Could you explain why we need bpf_prog_inc_not_zero here? We are inside rcu_read_lock/unlock region. We got a program from *pprog. If this program is not NULL, this program should not disappear since we are in rcu read lock region, right? Maybe I missed something, it would be good you can explain the scenario you try to pretect here.

bpf_prog_inc_not_zero() return a failure indicating that the program is being released and prog->aux->id will be set to 0.
Yes, it is just ok for accessing prog->aux->id directly.

Quoted 42 lines
Quoted 12 lines
net/core/sock_map.c +12 -0
+	if (IS_ERR(prog)) {
+		ret = PTR_ERR(prog);
+		goto end;
+	}
+	id = prog->aux->id;
+	bpf_prog_put(prog);
+
+end:
+	rcu_read_unlock();
+
+	if (copy_to_user(&uattr->query.attach_flags, &flags, sizeof(flags))) {
+		ret = -EFAULT;

net/core/sock_map.c static int sock_map_prog_update(struct · line 1521

Quoted 10 lines
net/core/sock_map.c +10 -0
+		goto err;
+	}
+	if (id != 0 && copy_to_user(prog_ids, &id, sizeof(u32))) {
+		ret = -EFAULT;
+		goto err;
+	}
+	if (copy_to_user(&uattr->query.prog_cnt, &prog_cnt, sizeof(prog_cnt))) {
+		ret = -EFAULT;
+		goto err;
+	}

You can do

	if (copy_to_user(&uattr->query.attach_flags, &flags, sizeof(flags)) ||
	    (id != 0 && copy_to_user(prog_ids, &id, sizeof(u32))) ||
	    copy_to_user(&uattr->query.prog_cnt, &prog_cnt, sizeof(prog_cnt)))
		ret = -EFAULT;

to make code a little bit concise.

Quoted 10 lines
net/core/sock_map.c +6 -0
+
+err:
+	fdput(f);
+	return ret;
+}
+
  static void sock_map_unlink(struct sock *sk, struct sk_psock_link *link)
  {
  	switch (link->map->map_type) {
#1 lore raw
Quoted 91 lines · zhudi (E)
Quoted 25 lines · Di Zhu
Quoted 7 lines

Right now there is no way to query whether BPF programs are attached to a sockmap or not.

we can use the standard interface in libbpf to query, such as:
bpf_prog_query(mapFd, BPF_SK_SKB_STREAM_PARSER, 0, NULL, ...); the mapFd is the fd of sockmap.

Quoted 5 lines
---
   include/linux/bpf.h  |  9 +++++
   kernel/bpf/syscall.c |  5 +++
   net/core/sock_map.c  | 88

++++++++++++++++++++++++++++++++++++++++----

include/linux/bpf.h int bpf_prog_test_run_syscall(struct bpf_prog

Quoted 7 lines
   3 files changed, 95 insertions(+), 7 deletions(-)
include/linux/bpf.h +0 -0
index d604c8251d88..594ca91992db 100644
@@ -1961,6 +1961,9 @@ int bpf_prog_test_run_syscall(struct bpf_prog

*prog,

include/linux/bpf.h int bpf_prog_test_run_syscall(struct bpf_prog · line 1961

Quoted 1 line
include/linux/bpf.h +0 -0
   int sock_map_get_from_fd(const union bpf_attr *attr, struct bpf_prog

*prog);

include/linux/bpf.h int bpf_prog_test_run_syscall(struct bpf_prog · line 1962

Quoted 1 line
include/linux/bpf.h +0 -0
   int sock_map_prog_detach(const union bpf_attr *attr, enum
Quoted 9 lines

bpf_prog_type ptype);

include/linux/bpf.h int bpf_prog_test_run_syscall(struct bpf_prog · line 1963

Quoted 1 line
include/linux/bpf.h +0 -0
   int sock_map_update_elem_sys(struct bpf_map *map, void *key, void

*value, u64 flags);

include/linux/bpf.h int bpf_prog_test_run_syscall(struct bpf_prog · line 1965

Quoted 2 lines
include/linux/bpf.h +2 -0
+int sockmap_bpf_prog_query(const union bpf_attr *attr,
+			   union bpf_attr __user *uattr);

All previous functions are with prefix "sock_map". Why you choose a different prefix "sockmap"?

Thanks for all your suggestions, I will make changes to the inappropriate code.

Quoted 54 lines

include/linux/bpf.h static inline int

Quoted 5 lines
include/linux/bpf.h +1 -0
+
   void sock_map_unhash(struct sock *sk);
   void sock_map_close(struct sock *sk, long timeout);
   #else
@@ -2014,6 +2017,12 @@ static inline int

sock_map_update_elem_sys(struct bpf_map *map, void *key, void

kernel/bpf/syscall.c static int bpf_prog_query(const union bpf_attr

Quoted 17 lines
include/linux/bpf.h +6 -0
   {
   	return -EOPNOTSUPP;
   }
+
+static inline int sockmap_bpf_prog_query(const union bpf_attr *attr,
+					 union bpf_attr __user *uattr)
+{
+	return -EINVAL;
+}
   #endif /* CONFIG_BPF_SYSCALL */
   #endif /* CONFIG_NET && CONFIG_BPF_SYSCALL */
 
kernel/bpf/syscall.c +0 -0
index 4e50c0bfdb7d..17faeff8f85f 100644
@@ -3275,6 +3275,11 @@ static int bpf_prog_query(const union bpf_attr

*attr,

net/core/sock_map.c static struct sk_psock_progs

Quoted 16 lines
kernel/bpf/syscall.c +5 -0
   	case BPF_FLOW_DISSECTOR:
   	case BPF_SK_LOOKUP:
   		return netns_bpf_prog_query(attr, uattr);
+	case BPF_SK_SKB_STREAM_PARSER:
+	case BPF_SK_SKB_STREAM_VERDICT:
+	case BPF_SK_MSG_VERDICT:
+	case BPF_SK_SKB_VERDICT:
+		return sockmap_bpf_prog_query(attr, uattr);
   	default:
   		return -EINVAL;
   	}
net/core/sock_map.c +0 -0
index e252b8ec2b85..ca65ed0004d3 100644
@@ -1412,38 +1412,50 @@ static struct sk_psock_progs

*sock_map_progs(struct bpf_map *map)

net/core/sock_map.c static struct sk_psock_progs · line 1415

Quoted 4 lines
net/core/sock_map.c +0 -1
   	return NULL;
   }
 
-static int sock_map_prog_update(struct bpf_map *map, struct bpf_prog

*prog,

net/core/sock_map.c static struct sk_psock_progs · line 1415

Quoted 2 lines
net/core/sock_map.c +1 -1
-				struct bpf_prog *old, u32 which)
+static int sock_map_prog_lookup(struct bpf_map *map, struct bpf_prog

**pprog[],

Can we just change "**pprog[]" to "***pprog"? In the code, you really just pass the address of the decl "struct bpf_prog **pprog;" to the function.

Quoted 6 lines
Quoted 3 lines

net/core/sock_map.c static struct sk_psock_progs · line 1416

Quoted 1 line
net/core/sock_map.c +1 -0
+				u32 which)

Some format issue here?

Format is right, passed the checkpatch script check.

Sorry about this. I guess my reply formating cheated me:

Quoted 3 lines
Quoted 3 lines

net/core/sock_map.c static struct sk_psock_progs · line 1417

Quoted 1 line
net/core/sock_map.c +1 -0
+static int sock_map_prog_lookup(struct bpf_map *map, struct bpf_prog

**pprog[],

net/core/sock_map.c static struct sk_psock_progs · line 1418

Quoted 1 line
net/core/sock_map.c +1 -0
+				u32 which)

I see a larger misalignment between "struct bpf_map *map" and "u32 which" in the reply email. But looking at original patch, there are no issues.

Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
net/core/sock_map.c +0 -1
   {
   	struct sk_psock_progs *progs = sock_map_progs(map);
-	struct bpf_prog **pprog;
 
   	if (!progs)
   		return -EOPNOTSUPP;

[...]