Skip to contentSkip to conversation

Fix cgroup attach flags being assigned to effective progs

PATCH In review Pu Lehui 2 participants 6 messages Aug 20, 2022 – Sep 8, 2022

Subsystems (2)
lore
Share & tools
lore

Pu Lehui

author
#0 lore raw

cover Fix cgroup attach flags being assigned to effective progs

Attach flags is only valid for attached progs of this layer cgroup, but not for effective progs.

Pu Lehui (2):
bpf, cgroup: Fix attach flags being assigned to effective progs
bpftool: Fix cgroup attach flags being assigned to effective progs

 kernel/bpf/cgroup.c        | 5 ++++-
 tools/bpf/bpftool/cgroup.c | 9 +++------
 2 files changed, 7 insertions(+), 7 deletions(-)
Signature · 1 line
2.25.1
author
#1 lore raw

PATCH 1/2 bpf, cgroup: Fix attach flags being assigned to effective progs

Attach flags is only valid for attached progs of this layer cgroup, but not for effective progs. We know that the attached progs is at the beginning of the effective progs array, so we can just populate the elements in front of the prog_attach_flags array.

---
 kernel/bpf/cgroup.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
kernel/bpf/cgroup.c +4 -1
index 59b7eb60d5b4..9adf72e99907 100644
@@ -1091,11 +1091,14 @@ static int __cgroup_bpf_query(struct cgroup *cgrp, const union bpf_attr *attr,
 		}
 
 		if (prog_attach_flags) {
+			int progs_cnt = prog_list_length(&cgrp->bpf.progs[atype]);
 			flags = cgrp->bpf.flags[atype];
 
-			for (i = 0; i < cnt; i++)
+			/* attach flags only for attached progs, but not effective progs */
+			for (i = 0; i < progs_cnt; i++)
 				if (copy_to_user(prog_attach_flags + i, &flags, sizeof(flags)))
 					return -EFAULT;
+
 			prog_attach_flags += cnt;
 		}
 
Signature · 1 line
2.25.1
#2 lore raw

RE: [PATCH bpf-next 1/2] bpf, cgroup: Fix attach flags being assigned to effective progs

Quoted 6 lines · Pu Lehui

Attach flags is only valid for attached progs of this layer cgroup, but not for effective progs. We know that the attached progs is at the beginning of the effective progs array, so we can just populate the elements in front of the prog_attach_flags array.

Trying to parse above, could you add a bit more detail on why this is problem so readers don't need to track it down.

kernel/bpf/cgroup.c static int __cgroup_bpf_query(struct cgroup *cgrp, const union bpf_attr *attr, · line 1092

Quoted 11 lines
---
 kernel/bpf/cgroup.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
kernel/bpf/cgroup.c +0 -0
index 59b7eb60d5b4..9adf72e99907 100644
@@ -1091,11 +1091,14 @@ static int __cgroup_bpf_query(struct cgroup *cgrp, const union bpf_attr *attr,
 		}
 

Because we are looking at it let me try to understand. There are two paths that set cnt relative bits here,

  if (attr->query.query_flags & BPF_F_QUERY_EFFECTIVE) {
      ...     
      cnt = min_t(int, bpf_prog_array_length(effective), total_cnt);                                       
      ...     
  } else {
     ...
     progs = &cgrp->bpf.progs[atype];
     cnt = min_t(int, prog_list_length(progs), total_cnt);
     ...
  }

And the docs claim

* **BPF_F_QUERY_EFFECTIVE**
* Only return information regarding programs which are
* currently effective at the specified *target_fd*.

so in the EFFECTIVE case should we be reporting flags at all if the commit message says "attach flags is only valid for attached progs of this layer cgroup, but not for effective progs."

And then in the else branch the change is what you have in the diff anyways correct?

kernel/bpf/cgroup.c static int __cgroup_bpf_query(struct cgroup *cgrp, const union bpf_attr *attr, · line 1096

Quoted 5 lines
kernel/bpf/cgroup.c +1 -1
 		if (prog_attach_flags) {
+			int progs_cnt = prog_list_length(&cgrp->bpf.progs[atype]);
 			flags = cgrp->bpf.flags[atype];
 
-			for (i = 0; i < cnt; i++)

Do we need to min with total_cnt here so we don't walk off a short user list?

Quoted 11 lines
kernel/bpf/cgroup.c +3 -0
+			/* attach flags only for attached progs, but not effective progs */
+			for (i = 0; i < progs_cnt; i++)
 				if (copy_to_user(prog_attach_flags + i, &flags, sizeof(flags)))
 					return -EFAULT;
+
 			prog_attach_flags += cnt;
 		}
 
Signature · 1 line
2.25.1
author
#3 lore raw

Re: [PATCH bpf-next 1/2] bpf, cgroup: Fix attach flags being assigned to effective progs

Quoted 15 lines · John Fastabend
Quoted 6 lines · Pu Lehui

Attach flags is only valid for attached progs of this layer cgroup, but not for effective progs. We know that the attached progs is at the beginning of the effective progs array, so we can just populate the elements in front of the prog_attach_flags array.

Trying to parse above, could you add a bit more detail on why this is problem so readers don't need to track it down.

Quoted 4 lines
---
  kernel/bpf/cgroup.c | 5 ++++-
  1 file changed, 4 insertions(+), 1 deletion(-)
Quoted 11 lines

kernel/bpf/cgroup.c static int __cgroup_bpf_query(struct cgroup *cgrp, const union bpf_attr *attr, · line 1092

Quoted 7 lines
kernel/bpf/cgroup.c +0 -0
index 59b7eb60d5b4..9adf72e99907 100644
@@ -1091,11 +1091,14 @@ static int __cgroup_bpf_query(struct cgroup *cgrp, const union bpf_attr *attr,
  		}
 

Because we are looking at it let me try to understand. There are two paths that set cnt relative bits here,

Hi John,

Thanks for your review. The reason is that:
For the following cgroup tree:
root

    |
   cg1
    |

cg2

I attached prog1 and prog2 to root cgroup with MULTI attach type, attached prog3 to cg1 with OVERRIDE attach type, and then used bpftool to show this cgroup tree with effective query flag:

$ bpftool cgroup tree /sys/fs/cgroup effective
CgroupPath
ID       AttachType      AttachFlags     Name
/sys/fs/cgroup
1        sysctl          multi           sysctl_tcp_mem
2        sysctl          multi           sysctl_tcp_mem
/sys/fs/cgroup/cg1
3        sysctl          override        sysctl_tcp_mem
1        sysctl          override        sysctl_tcp_mem <- wrong
2        sysctl          override        sysctl_tcp_mem <- wrong
/sys/fs/cgroup/cg1/cg2
3        sysctl                          sysctl_tcp_mem
1        sysctl                          sysctl_tcp_mem
2        sysctl                          sysctl_tcp_mem

For cg1, obviously, the attach flags of prog1 and prog2 can not be OVERRIDE, and the attach flags of prog1 and prog2 is meaningless for cg1. We only need to care the attach flags of prog which attached to cg1, other progs attach flags should be omit.

Quoted 23 lines
   if (attr->query.query_flags & BPF_F_QUERY_EFFECTIVE) {
       ...
       cnt = min_t(int, bpf_prog_array_length(effective), total_cnt);
       ...
   } else {
      ...
      progs = &cgrp->bpf.progs[atype];
      cnt = min_t(int, prog_list_length(progs), total_cnt);
      ...
   }

And the docs claim

* **BPF_F_QUERY_EFFECTIVE**
* Only return information regarding programs which are
* currently effective at the specified *target_fd*.

so in the EFFECTIVE case should we be reporting flags at all if the commit message says "attach flags is only valid for attached progs of this layer cgroup, but not for effective progs."

And then in the else branch the change is what you have in the diff anyways correct?

Quoted 8 lines

kernel/bpf/cgroup.c static int __cgroup_bpf_query(struct cgroup *cgrp, const union bpf_attr *attr, · line 1096

Quoted 5 lines
kernel/bpf/cgroup.c +1 -1
  		if (prog_attach_flags) {
+			int progs_cnt = prog_list_length(&cgrp->bpf.progs[atype]);
  			flags = cgrp->bpf.flags[atype];
 
-			for (i = 0; i < cnt; i++)

Do we need to min with total_cnt here so we don't walk off a short user list?

We should limit it, will fix it in v2. For query without effective flag, progs_cnt will equal to cnt, and for effective flag situation, progs_cnt only calculate prog count which attached to this cgroup layer.

Quoted 13 lines
Quoted 11 lines
kernel/bpf/cgroup.c +3 -0
+			/* attach flags only for attached progs, but not effective progs */
+			for (i = 0; i < progs_cnt; i++)
  				if (copy_to_user(prog_attach_flags + i, &flags, sizeof(flags)))
  					return -EFAULT;
+
  			prog_attach_flags += cnt;
  		}
 
Signature · 1 line
2.25.1

.

author
#4 lore raw

Re: [PATCH bpf-next 1/2] bpf, cgroup: Fix attach flags being assigned to effective progs

Quoted 40 lines · Pu Lehui
Quoted 15 lines · John Fastabend
Quoted 6 lines · Pu Lehui

Attach flags is only valid for attached progs of this layer cgroup, but not for effective progs. We know that the attached progs is at the beginning of the effective progs array, so we can just populate the elements in front of the prog_attach_flags array.

Trying to parse above, could you add a bit more detail on why this is problem so readers don't need to track it down.

Quoted 4 lines

---
  kernel/bpf/cgroup.c | 5 ++++-

  1 file changed, 4 insertions(+), 1 deletion(-)
Quoted 11 lines

kernel/bpf/cgroup.c static int __cgroup_bpf_query(struct cgroup · line 1092

Quoted 7 lines
kernel/bpf/cgroup.c +0 -0
index 59b7eb60d5b4..9adf72e99907 100644
@@ -1091,11 +1091,14 @@ static int __cgroup_bpf_query(struct cgroup 
 *cgrp, const union bpf_attr *attr,
           }

Because we are looking at it let me try to understand. There are two paths that set cnt relative bits here,

Hi John,

Thanks for your review. The reason is that:
For the following cgroup tree:
   root
    |
   cg1
    |
   cg2

Quoted 1 line

I attached prog1 and prog2 to root cgroup with MULTI attach type,

sorry, typo, attach type -> attach flag

Quoted 67 lines

attached prog3 to cg1 with OVERRIDE attach type, and then used bpftool to show this cgroup tree with effective query flag:

$ bpftool cgroup tree /sys/fs/cgroup effective
CgroupPath
ID       AttachType      AttachFlags     Name
/sys/fs/cgroup
1        sysctl          multi           sysctl_tcp_mem
2        sysctl          multi           sysctl_tcp_mem
/sys/fs/cgroup/cg1
3        sysctl          override        sysctl_tcp_mem
1        sysctl          override        sysctl_tcp_mem <- wrong
2        sysctl          override        sysctl_tcp_mem <- wrong
/sys/fs/cgroup/cg1/cg2
3        sysctl                          sysctl_tcp_mem
1        sysctl                          sysctl_tcp_mem
2        sysctl                          sysctl_tcp_mem

For cg1, obviously, the attach flags of prog1 and prog2 can not be OVERRIDE, and the attach flags of prog1 and prog2 is meaningless for cg1. We only need to care the attach flags of prog which attached to cg1, other progs attach flags should be omit.

Quoted 21 lines

   if (attr->query.query_flags & BPF_F_QUERY_EFFECTIVE) {
       ...
       cnt = min_t(int, bpf_prog_array_length(effective), total_cnt);
       ...
   } else {
      ...
      progs = &cgrp->bpf.progs[atype];
      cnt = min_t(int, prog_list_length(progs), total_cnt);
      ...
   }

And the docs claim

  *              **BPF_F_QUERY_EFFECTIVE**
  *                      Only return information regarding programs which are *                      currently effective at the specified *target_fd*.

so in the EFFECTIVE case should we be reporting flags at all if the commit message says "attach flags is only valid for attached progs of this layer cgroup, but not for effective progs."

Quoted 9 lines

And then in the else branch the change is what you have in the diff anyways correct?

kernel/bpf/cgroup.c static int __cgroup_bpf_query(struct cgroup *cgrp, const union bpf_attr *attr, · line 1096

Quoted 4 lines
kernel/bpf/cgroup.c +1 -1
           if (prog_attach_flags) {
+            int progs_cnt = prog_list_length(&cgrp->bpf.progs[atype]);
               flags = cgrp->bpf.flags[atype];
-            for (i = 0; i < cnt; i++)

Do we need to min with total_cnt here so we don't walk off a short user list?

We should limit it, will fix it in v2. For query without effective flag, progs_cnt will equal to cnt, and for effective flag situation, progs_cnt only calculate prog count which attached to this cgroup layer.

Quoted 11 lines
Quoted 11 lines
kernel/bpf/cgroup.c +1 -0
+            /* attach flags only for attached progs, but not 

effective progs */
+            for (i = 0; i < progs_cnt; i++)
                  if (copy_to_user(prog_attach_flags + i, &flags, sizeof(flags))) return -EFAULT;
+
              prog_attach_flags += cnt;
          }

Signature · 1 line
2.25.1
Quoted 2 lines

.

.

author
#5 lore raw

PATCH 2/2 bpftool: Fix cgroup attach flags being assigned to effective progs

In reply to Pu Lehui

When root-cgroup attach multi progs and sub-cgroup attach a override prog, bpftool will display incorrectly for the attach flags of the sub-cgroup’s effective progs:

$ bpftool c t /sys/fs/cgroup effective
CgroupPath
ID       AttachType      AttachFlags     Name
/sys/fs/cgroup
6        sysctl          multi           sysctl_tcp_mem
13       sysctl          multi           sysctl_tcp_mem
/sys/fs/cgroup/cg1
    20       sysctl          override        sysctl_tcp_mem
    6        sysctl          override        sysctl_tcp_mem
    13       sysctl          override        sysctl_tcp_mem

Attach flags is only valid for attached progs of this layer cgroup, but not for effective progs. Since prog_attach_flags array is already bypass the effective progs, so we can just use it.

---
 tools/bpf/bpftool/cgroup.c | 9 +++------
 1 file changed, 3 insertions(+), 6 deletions(-)
tools/bpf/bpftool/cgroup.c +3 -6
index cced668fb2a3..fa3eef0ff860 100644
@@ -219,11 +219,7 @@ static int show_attached_bpf_progs(int cgroup_fd, enum bpf_attach_type type,
 		return 0;
 
 	for (iter = 0; iter < p.prog_cnt; iter++) {
-		__u32 attach_flags;
-
-		attach_flags = prog_attach_flags[iter] ?: p.attach_flags;
-
-		switch (attach_flags) {
+		switch (prog_attach_flags[iter]) {
 		case BPF_F_ALLOW_MULTI:
 			attach_flags_str = "multi";
 			break;
@@ -234,7 +230,8 @@ static int show_attached_bpf_progs(int cgroup_fd, enum bpf_attach_type type,
 			attach_flags_str = "";
 			break;
 		default:
-			snprintf(buf, sizeof(buf), "unknown(%x)", attach_flags);
+			snprintf(buf, sizeof(buf), "unknown(%x)",
+				 prog_attach_flags[iter]);
 			attach_flags_str = buf;
 		}
 
Signature · 1 line
2.25.1