Skip to contentSkip to conversation

bpf: Drop unprotected find_vpid() in favour of find_get_pid()

PATCH Superseded Lee Jones 3 participants 7 messages Jul 21, 2022 – Jul 22, 2022

Subsystems (2)

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Lee Jones

author
#0 lore raw

PATCH 1/1 bpf: Drop unprotected find_vpid() in favour of find_get_pid()

The documentation for find_pid() clearly states:

"Must be called with the tasklist_lock or rcu_read_lock() held."

Presently we do neither.

In an ideal world we would wrap the in-lined call to find_vpid() along with get_pid_task() in the suggested rcu_read_lock() and have done. However, looking at get_pid_task()'s internals, it already does that independently, so this would lead to deadlock.

Instead, we'll use find_get_pid() which searches for the vpid, then takes a reference to it preventing early free, all within the safety of rcu_read_lock(). Once we have our reference we can safely make use of it up until the point it is put.

Cc Alexei Starovoitov <[email protected]>
Cc Daniel Borkmann <[email protected]>
Cc John Fastabend <[email protected]>
Cc Andrii Nakryiko <[email protected]>
Cc Martin KaFai Lau <[email protected]>
Cc Song Liu <[email protected]>
Cc Yonghong Song <[email protected]>
Cc KP Singh <[email protected]>
Cc Stanislav Fomichev <[email protected]>
Cc Hao Luo <[email protected]>
Cc Jiri Olsa <[email protected]>
Fixes 41bdc4b40ed6f ("bpf: introduce bpf subcommand BPF_TASK_FD_QUERY")
---
 kernel/bpf/syscall.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
kernel/bpf/syscall.c +4 -1
index 83c7136c5788d..c20cff30581c4 100644
@@ -4385,6 +4385,7 @@ static int bpf_task_fd_query(const union bpf_attr *attr,
 	const struct perf_event *event;
 	struct task_struct *task;
 	struct file *file;
+	struct pid *ppid;
 	int err;
 
 	if (CHECK_ATTR(BPF_TASK_FD_QUERY))
@@ -4396,7 +4397,9 @@ static int bpf_task_fd_query(const union bpf_attr *attr,
 	if (attr->task_fd_query.flags != 0)
 		return -EINVAL;
 
-	task = get_pid_task(find_vpid(pid), PIDTYPE_PID);
+	ppid = find_get_pid(pid);
+	task = get_pid_task(ppid, PIDTYPE_PID);
+	put_pid(ppid);
 	if (!task)
 		return -ENOENT;
 
Signature · 1 line
2.37.0.170.g444d1eabd0-goog
#1 lore raw
Quoted 10 lines · Lee Jones

The documentation for find_pid() clearly states:

"Must be called with the tasklist_lock or rcu_read_lock() held."

Presently we do neither.

In an ideal world we would wrap the in-lined call to find_vpid() along with get_pid_task() in the suggested rcu_read_lock() and have done. However, looking at get_pid_task()'s internals, it already does that independently, so this would lead to deadlock.

hm, we can have nested rcu_read_lock calls, right?

jirka

Quoted 50 lines

Instead, we'll use find_get_pid() which searches for the vpid, then takes a reference to it preventing early free, all within the safety of rcu_read_lock(). Once we have our reference we can safely make use of it up until the point it is put.

Cc Alexei Starovoitov <[email protected]>
Cc Daniel Borkmann <[email protected]>
Cc John Fastabend <[email protected]>
Cc Andrii Nakryiko <[email protected]>
Cc Martin KaFai Lau <[email protected]>
Cc Song Liu <[email protected]>
Cc Yonghong Song <[email protected]>
Cc KP Singh <[email protected]>
Cc Stanislav Fomichev <[email protected]>
Cc Hao Luo <[email protected]>
Cc Jiri Olsa <[email protected]>
Fixes 41bdc4b40ed6f ("bpf: introduce bpf subcommand BPF_TASK_FD_QUERY")
---
 kernel/bpf/syscall.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
kernel/bpf/syscall.c +4 -1
index 83c7136c5788d..c20cff30581c4 100644
@@ -4385,6 +4385,7 @@ static int bpf_task_fd_query(const union bpf_attr *attr,
 	const struct perf_event *event;
 	struct task_struct *task;
 	struct file *file;
+	struct pid *ppid;
 	int err;
 
 	if (CHECK_ATTR(BPF_TASK_FD_QUERY))
@@ -4396,7 +4397,9 @@ static int bpf_task_fd_query(const union bpf_attr *attr,
 	if (attr->task_fd_query.flags != 0)
 		return -EINVAL;
 
-	task = get_pid_task(find_vpid(pid), PIDTYPE_PID);
+	ppid = find_get_pid(pid);
+	task = get_pid_task(ppid, PIDTYPE_PID);
+	put_pid(ppid);
 	if (!task)
 		return -ENOENT;
 
Signature · 1 line
2.37.0.170.g444d1eabd0-goog
author
#2 lore raw
Quoted 5 lines · Jiri Olsa
Quoted 4 lines · Lee Jones

The documentation for find_pid() clearly states:

"Must be called with the tasklist_lock or rcu_read_lock() held."

Quoted 8 lines
Quoted 6 lines

Presently we do neither.

In an ideal world we would wrap the in-lined call to find_vpid() along with get_pid_task() in the suggested rcu_read_lock() and have done. However, looking at get_pid_task()'s internals, it already does that independently, so this would lead to deadlock.

hm, we can have nested rcu_read_lock calls, right?

I assumed not, but that might be an oversight on my part.

Would that be your preference?

Quoted 46 lines
Quoted 46 lines

Instead, we'll use find_get_pid() which searches for the vpid, then takes a reference to it preventing early free, all within the safety of rcu_read_lock(). Once we have our reference we can safely make use of it up until the point it is put.

Cc Alexei Starovoitov <[email protected]>
Cc Daniel Borkmann <[email protected]>
Cc John Fastabend <[email protected]>
Cc Andrii Nakryiko <[email protected]>
Cc Martin KaFai Lau <[email protected]>
Cc Song Liu <[email protected]>
Cc Yonghong Song <[email protected]>
Cc KP Singh <[email protected]>
Cc Stanislav Fomichev <[email protected]>
Cc Hao Luo <[email protected]>
Cc Jiri Olsa <[email protected]>
Fixes 41bdc4b40ed6f ("bpf: introduce bpf subcommand BPF_TASK_FD_QUERY")
---
 kernel/bpf/syscall.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
kernel/bpf/syscall.c +4 -1
index 83c7136c5788d..c20cff30581c4 100644
@@ -4385,6 +4385,7 @@ static int bpf_task_fd_query(const union bpf_attr *attr,
 	const struct perf_event *event;
 	struct task_struct *task;
 	struct file *file;
+	struct pid *ppid;
 	int err;
 
 	if (CHECK_ATTR(BPF_TASK_FD_QUERY))
@@ -4396,7 +4397,9 @@ static int bpf_task_fd_query(const union bpf_attr *attr,
 	if (attr->task_fd_query.flags != 0)
 		return -EINVAL;
 
-	task = get_pid_task(find_vpid(pid), PIDTYPE_PID);
+	ppid = find_get_pid(pid);
+	task = get_pid_task(ppid, PIDTYPE_PID);
+	put_pid(ppid);
 	if (!task)
 		return -ENOENT;
 
Signature · 1 line
Lee Jones [李琼斯]
#3 lore raw
Quoted 4 lines · Lee Jones
Quoted 2 lines · Jiri Olsa
Quoted 1 line · Lee Jones

The documentation for find_pid() clearly states:

typo find_vpid

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

"Must be called with the tasklist_lock or rcu_read_lock() held."

Presently we do neither.

just curious, did you see crash related to this or you just spot that

Quoted 11 lines
Quoted 7 lines
Quoted 5 lines

In an ideal world we would wrap the in-lined call to find_vpid() along with get_pid_task() in the suggested rcu_read_lock() and have done. However, looking at get_pid_task()'s internals, it already does that independently, so this would lead to deadlock.

hm, we can have nested rcu_read_lock calls, right?

I assumed not, but that might be an oversight on my part.

Would that be your preference?

seems simpler than calling get/put for ppid

jirka

Quoted 50 lines
Quoted 47 lines
Quoted 47 lines

Instead, we'll use find_get_pid() which searches for the vpid, then takes a reference to it preventing early free, all within the safety of rcu_read_lock(). Once we have our reference we can safely make use of it up until the point it is put.

Cc Alexei Starovoitov <[email protected]>
Cc Daniel Borkmann <[email protected]>
Cc John Fastabend <[email protected]>
Cc Andrii Nakryiko <[email protected]>
Cc Martin KaFai Lau <[email protected]>
Cc Song Liu <[email protected]>
Cc Yonghong Song <[email protected]>
Cc KP Singh <[email protected]>
Cc Stanislav Fomichev <[email protected]>
Cc Hao Luo <[email protected]>
Cc Jiri Olsa <[email protected]>
Fixes 41bdc4b40ed6f ("bpf: introduce bpf subcommand BPF_TASK_FD_QUERY")
---
 kernel/bpf/syscall.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
kernel/bpf/syscall.c +4 -1
index 83c7136c5788d..c20cff30581c4 100644
@@ -4385,6 +4385,7 @@ static int bpf_task_fd_query(const union bpf_attr *attr,
 	const struct perf_event *event;
 	struct task_struct *task;
 	struct file *file;
+	struct pid *ppid;
 	int err;
 
 	if (CHECK_ATTR(BPF_TASK_FD_QUERY))
@@ -4396,7 +4397,9 @@ static int bpf_task_fd_query(const union bpf_attr *attr,
 	if (attr->task_fd_query.flags != 0)
 		return -EINVAL;
 
-	task = get_pid_task(find_vpid(pid), PIDTYPE_PID);
+	ppid = find_get_pid(pid);
+	task = get_pid_task(ppid, PIDTYPE_PID);
+	put_pid(ppid);
 	if (!task)
 		return -ENOENT;
 
Signature · 1 line
Lee Jones [李琼斯]
#4 lore raw
Quoted 6 lines · Jiri Olsa
Quoted 4 lines · Lee Jones
Quoted 2 lines · Jiri Olsa
Quoted 1 line · Lee Jones

The documentation for find_pid() clearly states:

Quoted 18 lines

typo find_vpid

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

"Must be called with the tasklist_lock or rcu_read_lock() held."

Presently we do neither.

just curious, did you see crash related to this or you just spot that

Quoted 9 lines
Quoted 7 lines
Quoted 5 lines

In an ideal world we would wrap the in-lined call to find_vpid() along with get_pid_task() in the suggested rcu_read_lock() and have done. However, looking at get_pid_task()'s internals, it already does that independently, so this would lead to deadlock.

hm, we can have nested rcu_read_lock calls, right?

I assumed not, but that might be an oversight on my part.

From kernel documentation, nested rcu_read_lock is allowed.
https://www.kernel.org/doc/Documentation/RCU/Design/Requirements/Requirements.html

RCU's grace-period guarantee allows updaters to wait for the completion of all pre-existing RCU read-side critical sections. An RCU read-side critical section begins with the marker rcu_read_lock() and ends with the marker rcu_read_unlock(). These markers may be nested, and RCU treats a nested set as one big RCU read-side critical section. Production-quality implementations of rcu_read_lock() and rcu_read_unlock() are extremely lightweight, and in fact have exactly zero overhead in Linux kernels built for production use with CONFIG_PREEMPT=n.

Quoted 4 lines
Quoted 2 lines

Would that be your preference?

seems simpler than calling get/put for ppid

The current implementation seems okay since we can hide rcu_read_lock() inside find_get_pid(). We can also avoid nested rcu_read_lock(), which is although allowed but not pretty.

Quoted 53 lines

jirka

Quoted 50 lines
Quoted 47 lines
Quoted 47 lines

Instead, we'll use find_get_pid() which searches for the vpid, then takes a reference to it preventing early free, all within the safety of rcu_read_lock(). Once we have our reference we can safely make use of it up until the point it is put.

Cc Alexei Starovoitov <[email protected]>
Cc Daniel Borkmann <[email protected]>
Cc John Fastabend <[email protected]>
Cc Andrii Nakryiko <[email protected]>
Cc Martin KaFai Lau <[email protected]>
Cc Song Liu <[email protected]>
Cc Yonghong Song <[email protected]>
Cc KP Singh <[email protected]>
Cc Stanislav Fomichev <[email protected]>
Cc Hao Luo <[email protected]>
Cc Jiri Olsa <[email protected]>
Fixes 41bdc4b40ed6f ("bpf: introduce bpf subcommand BPF_TASK_FD_QUERY")
---
  kernel/bpf/syscall.c | 5 ++++-
  1 file changed, 4 insertions(+), 1 deletion(-)
kernel/bpf/syscall.c +4 -1
index 83c7136c5788d..c20cff30581c4 100644
@@ -4385,6 +4385,7 @@ static int bpf_task_fd_query(const union bpf_attr *attr,
  	const struct perf_event *event;
  	struct task_struct *task;
  	struct file *file;
+	struct pid *ppid;
  	int err;
 
  	if (CHECK_ATTR(BPF_TASK_FD_QUERY))
@@ -4396,7 +4397,9 @@ static int bpf_task_fd_query(const union bpf_attr *attr,
  	if (attr->task_fd_query.flags != 0)
  		return -EINVAL;
 
-	task = get_pid_task(find_vpid(pid), PIDTYPE_PID);
+	ppid = find_get_pid(pid);
+	task = get_pid_task(ppid, PIDTYPE_PID);
+	put_pid(ppid);
  	if (!task)
  		return -ENOENT;
 
Signature · 1 line
Lee Jones [李琼斯]
author
#5 lore raw
Quoted 41 lines · Yonghong Song
Quoted 6 lines · Jiri Olsa
Quoted 4 lines · Lee Jones
Quoted 2 lines · Jiri Olsa
Quoted 1 line · Lee Jones

The documentation for find_pid() clearly states:

Quoted 18 lines

typo find_vpid

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

"Must be called with the tasklist_lock or rcu_read_lock() held."

Presently we do neither.

just curious, did you see crash related to this or you just spot that

Quoted 9 lines
Quoted 7 lines
Quoted 5 lines

In an ideal world we would wrap the in-lined call to find_vpid() along with get_pid_task() in the suggested rcu_read_lock() and have done. However, looking at get_pid_task()'s internals, it already does that independently, so this would lead to deadlock.

hm, we can have nested rcu_read_lock calls, right?

I assumed not, but that might be an oversight on my part.

From kernel documentation, nested rcu_read_lock is allowed.
https://www.kernel.org/doc/Documentation/RCU/Design/Requirements/Requirements.html

RCU's grace-period guarantee allows updaters to wait for the completion of all pre-existing RCU read-side critical sections. An RCU read-side critical section begins with the marker rcu_read_lock() and ends with the marker rcu_read_unlock(). These markers may be nested, and RCU treats a nested set as one big RCU read-side critical section. Production-quality implementations of rcu_read_lock() and rcu_read_unlock() are extremely lightweight, and in fact have exactly zero overhead in Linux kernels built for production use with CONFIG_PREEMPT=n.

Quoted 1 line
Quoted 1 line
Quoted 8 lines
Quoted 3 lines
Quoted 1 line

Would that be your preference?

seems simpler than calling get/put for ppid

The current implementation seems okay since we can hide rcu_read_lock() inside find_get_pid(). We can also avoid nested rcu_read_lock(), which is although allowed but not pretty.

Right, this was my thinking.

Happy to go with whatever you guys decide though.

Make the call and I'll rework, or not.

Signature · 1 line
Lee Jones [李琼斯]
#6 lore raw
Quoted 52 lines · Lee Jones
Quoted 41 lines · Yonghong Song
Quoted 6 lines · Jiri Olsa
Quoted 4 lines · Lee Jones
Quoted 2 lines · Jiri Olsa
Quoted 1 line · Lee Jones

The documentation for find_pid() clearly states:

Quoted 18 lines

typo find_vpid

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

"Must be called with the tasklist_lock or rcu_read_lock() held."

Presently we do neither.

just curious, did you see crash related to this or you just spot that

Quoted 9 lines
Quoted 7 lines
Quoted 5 lines

In an ideal world we would wrap the in-lined call to find_vpid() along with get_pid_task() in the suggested rcu_read_lock() and have done. However, looking at get_pid_task()'s internals, it already does that independently, so this would lead to deadlock.

hm, we can have nested rcu_read_lock calls, right?

I assumed not, but that might be an oversight on my part.

From kernel documentation, nested rcu_read_lock is allowed.
https://www.kernel.org/doc/Documentation/RCU/Design/Requirements/Requirements.html

RCU's grace-period guarantee allows updaters to wait for the completion of all pre-existing RCU read-side critical sections. An RCU read-side critical section begins with the marker rcu_read_lock() and ends with the marker rcu_read_unlock(). These markers may be nested, and RCU treats a nested set as one big RCU read-side critical section. Production-quality implementations of rcu_read_lock() and rcu_read_unlock() are extremely lightweight, and in fact have exactly zero overhead in Linux kernels built for production use with CONFIG_PREEMPT=n.

Quoted 1 line
Quoted 1 line
Quoted 8 lines
Quoted 3 lines
Quoted 1 line

Would that be your preference?

seems simpler than calling get/put for ppid

The current implementation seems okay since we can hide rcu_read_lock() inside find_get_pid(). We can also avoid nested rcu_read_lock(), which is although allowed but not pretty.

Quoted 5 lines

Right, this was my thinking.

Happy to go with whatever you guys decide though.

Make the call and I'll rework, or not.

ok, I can live with the current version ;-) could you please resend with fixed changelog?

thanks,
jirka