Skip to contentSkip to conversation

bpf: Fix perf bpf event and audit prog id logging

RFC In review Jiri Olsa 4 participants 5 messages Nov 15, 2022 – Nov 16, 2022

Subsystems (3)
lore
Share & tools
lore

Jiri Olsa

author reviewer
#0 lore raw

RFC 1/1 bpf: Fix perf bpf event and audit prog id logging

hi,
perf_event_bpf_event and bpf_audit_prog calls currently report zero program id for unload path.

It's because of the [1] change moved those audit calls into work queue and they are executed after the id is zeroed in bpf_prog_free_id.

I originally made a change that added 'id_audit' field to struct bpf_prog, which would be initialized as id, untouched and used in audit callbacks.

Then I realized we might actually not need to zero prog->aux->id in bpf_prog_free_id. It seems to be called just once on release paths. Tests seems ok with that.

thoughts?

thanks,
jirka

[1] d809e134be7a ("bpf: Prepare bpf_prog_put() to be called from irq context.")

---
 kernel/bpf/syscall.c | 1 -
 1 file changed, 1 deletion(-)
kernel/bpf/syscall.c +0 -1
index fdbae52f463f..426529355c29 100644
@@ -1991,7 +1991,6 @@ void bpf_prog_free_id(struct bpf_prog *prog, bool do_idr_lock)
 		__acquire(&prog_idr_lock);
 
 	idr_remove(&prog_idr, prog->aux->id);
-	prog->aux->id = 0;
 
 	if (do_idr_lock)
 		spin_unlock_irqrestore(&prog_idr_lock, flags);
Signature · 1 line
2.38.1
maintainer
#1 lore raw
Quoted 31 lines · Jiri Olsa

hi,
perf_event_bpf_event and bpf_audit_prog calls currently report zero program id for unload path.

It's because of the [1] change moved those audit calls into work queue and they are executed after the id is zeroed in bpf_prog_free_id.

I originally made a change that added 'id_audit' field to struct bpf_prog, which would be initialized as id, untouched and used in audit callbacks.

Then I realized we might actually not need to zero prog->aux->id in bpf_prog_free_id. It seems to be called just once on release paths. Tests seems ok with that.

thoughts?

thanks,
jirka

[1] d809e134be7a ("bpf: Prepare bpf_prog_put() to be called from irq context.")

---
  kernel/bpf/syscall.c | 1 -
  1 file changed, 1 deletion(-)
kernel/bpf/syscall.c +0 -0
index fdbae52f463f..426529355c29 100644

kernel/bpf/syscall.c void bpf_prog_free_id(struct bpf_prog *prog, bool do_idr_lock) · line 1994

Quoted 4 lines
kernel/bpf/syscall.c +0 -1
@@ -1991,7 +1991,6 @@ void bpf_prog_free_id(struct bpf_prog *prog, bool do_idr_lock)
  		__acquire(&prog_idr_lock);
 
  	idr_remove(&prog_idr, prog->aux->id);
-	prog->aux->id = 0;

This would trigger a race when offloaded progs are used, see also ad8ad79f4f60 ("bpf:
offload: free program id when device disappears"). __bpf_prog_offload_destroy() calls it, and my read is that later bpf_prog_free_id() then hits the early !prog->aux->id return path. Is there a reason for irq context to not defer the bpf_prog_free_id()?

Quoted 3 lines
kernel/bpf/syscall.c +0 -0
  	if (do_idr_lock)
  		spin_unlock_irqrestore(&prog_idr_lock, flags);
 
#2 lore raw
Quoted 27 lines · Daniel Borkmann
Quoted 26 lines · Jiri Olsa

hi,
perf_event_bpf_event and bpf_audit_prog calls currently report zero program id for unload path.

It's because of the [1] change moved those audit calls into work queue and they are executed after the id is zeroed in bpf_prog_free_id.

I originally made a change that added 'id_audit' field to struct bpf_prog, which would be initialized as id, untouched and used in audit callbacks.

Then I realized we might actually not need to zero prog->aux->id in bpf_prog_free_id. It seems to be called just once on release paths. Tests seems ok with that.

thoughts?

thanks,
jirka

[1] d809e134be7a ("bpf: Prepare bpf_prog_put() to be called from irq context.")

---
  kernel/bpf/syscall.c | 1 -
  1 file changed, 1 deletion(-)
Quoted 13 lines

kernel/bpf/syscall.c void bpf_prog_free_id(struct bpf_prog *prog, bool do_idr_lock) · line 1993

Quoted 8 lines
kernel/bpf/syscall.c +0 -1
index fdbae52f463f..426529355c29 100644
@@ -1991,7 +1991,6 @@ void bpf_prog_free_id(struct bpf_prog *prog, bool do_idr_lock)
  		__acquire(&prog_idr_lock);
  	idr_remove(&prog_idr, prog->aux->id);
-	prog->aux->id = 0;

This would trigger a race when offloaded progs are used, see also ad8ad79f4f60 ("bpf:
offload: free program id when device disappears"). __bpf_prog_offload_destroy() calls it, and my read is that later bpf_prog_free_id() then hits the early !prog->aux->id return path. Is there a reason for irq context to not defer the bpf_prog_free_id()?

there's comment saying:
/* bpf_prog_free_id() must be called first */

it was added together with the BPF_(PROG|MAP)_GET_NEXT_ID api:
34ad5580f8f9 bpf: Add BPF_(PROG|MAP)_GET_NEXT_ID command

Martin, any idea?

while looking on this I noticed we can remove the do_idr_lock argument (patch below) because it's always true and I think we need to upgrade all the prog_idr_lock locks to spin_lock_irqsave because bpf_prog_put could be called from irq context because of d809e134be7a

thanks,
jirka

---

include/linux/bpf.h +1 -1
index 49f9d2bec401..50c71ce698d0 100644
@@ -1761,7 +1761,7 @@ void bpf_prog_inc(struct bpf_prog *prog);
 struct bpf_prog * __must_check bpf_prog_inc_not_zero(struct bpf_prog *prog);
 void bpf_prog_put(struct bpf_prog *prog);
 
-void bpf_prog_free_id(struct bpf_prog *prog, bool do_idr_lock);
+void bpf_prog_free_id(struct bpf_prog *prog);
 void bpf_map_free_id(struct bpf_map *map, bool do_idr_lock);
 
 struct btf_field *btf_record_find(const struct btf_record *rec,
kernel/bpf/offload.c +1 -1
index 13e4efc971e6..327dab644200 100644
@@ -217,7 +217,7 @@ static void __bpf_prog_offload_destroy(struct bpf_prog *prog)
 		offload->offdev->ops->destroy(prog);
 
 	/* Make sure BPF_PROG_GET_NEXT_ID can't find this dead program */
-	bpf_prog_free_id(prog, true);
+	bpf_prog_free_id(prog);
 
 	list_del_init(&offload->offloads);
 	kfree(offload);
kernel/bpf/syscall.c +4 -10
index fdbae52f463f..9b929d8ab06d 100644
@@ -1973,7 +1973,7 @@ static int bpf_prog_alloc_id(struct bpf_prog *prog)
 	return id > 0 ? 0 : id;
 }
 
-void bpf_prog_free_id(struct bpf_prog *prog, bool do_idr_lock)
+void bpf_prog_free_id(struct bpf_prog *prog)
 {
 	unsigned long flags;
 
@@ -1985,18 +1985,12 @@ void bpf_prog_free_id(struct bpf_prog *prog, bool do_idr_lock)
 	if (!prog->aux->id)
 		return;
 
-	if (do_idr_lock)
-		spin_lock_irqsave(&prog_idr_lock, flags);
-	else
-		__acquire(&prog_idr_lock);
+	spin_lock_irqsave(&prog_idr_lock, flags);
 
 	idr_remove(&prog_idr, prog->aux->id);
 	prog->aux->id = 0;
 
-	if (do_idr_lock)
-		spin_unlock_irqrestore(&prog_idr_lock, flags);
-	else
-		__release(&prog_idr_lock);
+	spin_unlock_irqrestore(&prog_idr_lock, flags);
 }
 
 static void __bpf_prog_put_rcu(struct rcu_head *rcu)
@@ -2048,7 +2042,7 @@ static void __bpf_prog_put(struct bpf_prog *prog, bool do_idr_lock)
 
 	if (atomic64_dec_and_test(&aux->refcnt)) {
 		/* bpf_prog_free_id() must be called first */
-		bpf_prog_free_id(prog, do_idr_lock);
+		bpf_prog_free_id(prog);
 
 		if (in_irq() || irqs_disabled()) {
 			INIT_WORK(&aux->work, bpf_prog_put_deferred);
#3 lore raw
Quoted 49 lines · Jiri Olsa
Quoted 27 lines · Daniel Borkmann
Quoted 26 lines · Jiri Olsa

hi,
perf_event_bpf_event and bpf_audit_prog calls currently report zero program id for unload path.

It's because of the [1] change moved those audit calls into work queue and they are executed after the id is zeroed in bpf_prog_free_id.

I originally made a change that added 'id_audit' field to struct bpf_prog, which would be initialized as id, untouched and used in audit callbacks.

Then I realized we might actually not need to zero prog->aux->id in bpf_prog_free_id. It seems to be called just once on release paths. Tests seems ok with that.

thoughts?

thanks,
jirka

[1] d809e134be7a ("bpf: Prepare bpf_prog_put() to be called from irq context.")

---
  kernel/bpf/syscall.c | 1 -
  1 file changed, 1 deletion(-)
Quoted 13 lines

kernel/bpf/syscall.c void bpf_prog_free_id(struct bpf_prog *prog, bool do_idr_lock) · line 1993

Quoted 8 lines
kernel/bpf/syscall.c +0 -1
index fdbae52f463f..426529355c29 100644
@@ -1991,7 +1991,6 @@ void bpf_prog_free_id(struct bpf_prog *prog, bool do_idr_lock)
            __acquire(&prog_idr_lock);
    idr_remove(&prog_idr, prog->aux->id);
-   prog->aux->id = 0;

This would trigger a race when offloaded progs are used, see also ad8ad79f4f60 ("bpf:
offload: free program id when device disappears"). __bpf_prog_offload_destroy() calls it, and my read is that later bpf_prog_free_id() then hits the early !prog->aux->id return path. Is there a reason for irq context to not defer the bpf_prog_free_id()?

there's comment saying:
/* bpf_prog_free_id() must be called first */

it was added together with the BPF_(PROG|MAP)_GET_NEXT_ID api:
34ad5580f8f9 bpf: Add BPF_(PROG|MAP)_GET_NEXT_ID command

Quoted 6 lines

Martin, any idea?

while looking on this I noticed we can remove the do_idr_lock argument (patch below) because it's always true and I think we need to upgrade all the prog_idr_lock locks to spin_lock_irqsave because bpf_prog_put could be called from irq context because of d809e134be7a

before we dive into rabbit hole let's understand
the severity of
"perf_event_bpf_event and bpf_audit_prog calls currently report zero
program id for unload path"

and why we cannot leave it as-is.

#4 lore raw
Quoted 57 lines · Alexei Starovoitov
Quoted 49 lines · Jiri Olsa
Quoted 27 lines · Daniel Borkmann
Quoted 26 lines · Jiri Olsa

hi,
perf_event_bpf_event and bpf_audit_prog calls currently report zero program id for unload path.

It's because of the [1] change moved those audit calls into work queue and they are executed after the id is zeroed in bpf_prog_free_id.

I originally made a change that added 'id_audit' field to struct bpf_prog, which would be initialized as id, untouched and used in audit callbacks.

Then I realized we might actually not need to zero prog->aux->id in bpf_prog_free_id. It seems to be called just once on release paths. Tests seems ok with that.

thoughts?

thanks,
jirka

[1] d809e134be7a ("bpf: Prepare bpf_prog_put() to be called from irq context.")

---
  kernel/bpf/syscall.c | 1 -
  1 file changed, 1 deletion(-)
Quoted 13 lines

kernel/bpf/syscall.c void bpf_prog_free_id(struct bpf_prog *prog, bool do_idr_lock) · line 1993

Quoted 8 lines
kernel/bpf/syscall.c +0 -1
index fdbae52f463f..426529355c29 100644
@@ -1991,7 +1991,6 @@ void bpf_prog_free_id(struct bpf_prog *prog, bool do_idr_lock)
            __acquire(&prog_idr_lock);
    idr_remove(&prog_idr, prog->aux->id);
-   prog->aux->id = 0;

This would trigger a race when offloaded progs are used, see also ad8ad79f4f60 ("bpf:
offload: free program id when device disappears"). __bpf_prog_offload_destroy() calls it, and my read is that later bpf_prog_free_id() then hits the early !prog->aux->id return path. Is there a reason for irq context to not defer the bpf_prog_free_id()?

there's comment saying:
/* bpf_prog_free_id() must be called first */

it was added together with the BPF_(PROG|MAP)_GET_NEXT_ID api:
34ad5580f8f9 bpf: Add BPF_(PROG|MAP)_GET_NEXT_ID command

Quoted 6 lines

Martin, any idea?

while looking on this I noticed we can remove the do_idr_lock argument (patch below) because it's always true and I think we need to upgrade all the prog_idr_lock locks to spin_lock_irqsave because bpf_prog_put could be called from irq context because of d809e134be7a

Quoted 6 lines

before we dive into rabbit hole let's understand
the severity of
"perf_event_bpf_event and bpf_audit_prog calls currently report zero
program id for unload path"

and why we cannot leave it as-is.

I found this because I wanted use perf bpf-event to monitor bpf programs loads/unloads.. and I need 'id' on the unload event

perf_event_bpf_event is irq safe so quick fix for me would be to move it back:

        if (atomic64_dec_and_test(&aux->refcnt)) {
+               perf_event_bpf_event(prog, PERF_BPF_EVENT_PROG_UNLOAD, 0);
                /* bpf_prog_free_id() must be called first */
                bpf_prog_free_id(prog, do_idr_lock);

FWIW I also used fix below for a while for testing

jirka

---

include/linux/bpf.h +1 -0
index d88e0741b381..6b752d4815e6 100644
@@ -1163,6 +1163,7 @@ struct bpf_prog_aux {
 	u32 max_tp_access;
 	u32 stack_depth;
 	u32 id;
+	u32 id_audit;
 	u32 func_cnt; /* used by non-func prog as the number of func progs */
 	u32 func_idx; /* 0 for non-func prog, the index in func array for func prog */
 	u32 attach_btf_id; /* in-kernel BTF type id to attach to */
kernel/bpf/syscall.c +2 -2
index d6081e8336c6..1f4fdf0aaac6 100644
@@ -1930,7 +1930,7 @@ static void bpf_audit_prog(const struct bpf_prog *prog, unsigned int op)
 	if (unlikely(!ab))
 		return;
 	audit_log_format(ab, "prog-id=%u op=%s",
-			 prog->aux->id, bpf_audit_str[op]);
+			 prog->aux->id_audit, bpf_audit_str[op]);
 	audit_log_end(ab);
 }
 
@@ -1942,7 +1942,7 @@ static int bpf_prog_alloc_id(struct bpf_prog *prog)
 	spin_lock_bh(&prog_idr_lock);
 	id = idr_alloc_cyclic(&prog_idr, prog, 1, INT_MAX, GFP_ATOMIC);
 	if (id > 0)
-		prog->aux->id = id;
+		prog->aux->id = prog->aux->id_audit = id;
 	spin_unlock_bh(&prog_idr_lock);
 	idr_preload_end();
 
kernel/events/core.c +1 -1
index 0f0bfcf5adce..32edb3a4360e 100644
@@ -9065,7 +9065,7 @@ void perf_event_bpf_event(struct bpf_prog *prog,
 			},
 			.type = type,
 			.flags = flags,
-			.id = prog->aux->id,
+			.id = prog->aux->id_audit,
 		},
 	};