Skip to contentSkip to conversation

pid: Replace struct pid 1-element array with flex-array

PATCH In review Kees Cook 3 participants 3 messages Jun 30, 2023 – Jul 1, 2023 Namespaces:

lore
Share & tools
lore

Kees Cook

author
#0 lore raw

PATCH v2 1/1 pid: Replace struct pid 1-element array with flex-array

For pid namespaces, struct pid uses a dynamically sized array member, "numbers". This was implemented using the ancient 1-element fake flexible array, which has been deprecated for decades. Replace it with a C99 flexible array, refactor the array size calculations to use struct_size(), and address elements via indexes. Note that the static initializer (which defines a single element) works as-is, and requires no special handling.

Without this, CONFIG_UBSAN_BOUNDS (and potentially CONFIG_FORTIFY_SOURCE) will trigger bounds checks when entering a pid namespace:
https://lore.kernel.org/lkml/20230517-bushaltestelle-super-e223978c1ba6@brauner

For example: unshare --fork --pid --mount-proc readlink /proc/self

Cc Christian Brauner <[email protected]>
Cc Jan Kara <[email protected]>
Cc Andreas Gruenbacher <[email protected]>
Cc Daniel Verkamp <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Jeff Xu <[email protected]>
Cc Andrew Morton <[email protected]>
Cc Boqun Feng <[email protected]>
Cc Luis Chamberlain <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Reported-by [email protected]

---
v2:
- drop pointer math to array index conversions (torvalds)
- use struct_size_t now that it exists (torvalds)
- updated commit log with reproducer example
v1: https://lore.kernel.org/lkml/[email protected]/

---
 include/linux/pid.h    | 2 +-
 kernel/pid.c           | 7 +++++--
 kernel/pid_namespace.c | 2 +-
 3 files changed, 7 insertions(+), 4 deletions(-)
include/linux/pid.h +1 -1
index b75de288a8c2..653a527574c4 100644
@@ -67,7 +67,7 @@ struct pid
 	/* wait queue for pidfd notifications */
 	wait_queue_head_t wait_pidfd;
 	struct rcu_head rcu;
-	struct upid numbers[1];
+	struct upid numbers[];
 };
 
 extern struct pid init_struct_pid;
kernel/pid.c +5 -2
index f93954a0384d..6a1d23a11026 100644
@@ -656,8 +656,11 @@ void __init pid_idr_init(void)
 
 	idr_init(&init_pid_ns.idr);
 
-	init_pid_ns.pid_cachep = KMEM_CACHE(pid,
-			SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT);
+	init_pid_ns.pid_cachep = kmem_cache_create("pid",
+			struct_size_t(struct pid, numbers, 1),
+			__alignof__(struct pid),
+			SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT,
+			NULL);
 }
 
 static struct file *__pidfd_fget(struct task_struct *task, int fd)
kernel/pid_namespace.c +1 -1
index b43eee07b00c..0bf44afe04dd 100644
@@ -48,7 +48,7 @@ static struct kmem_cache *create_pid_cachep(unsigned int level)
 		return kc;
 
 	snprintf(name, sizeof(name), "pid_%u", level + 1);
-	len = sizeof(struct pid) + level * sizeof(struct upid);
+	len = struct_size_t(struct pid, numbers, level + 1);
 	mutex_lock(&pid_caches_mutex);
 	/* Name collision forces to do allocation under mutex. */
 	if (!*pkc)
Signature · 1 line
2.34.1
Reviewed-by on v2 1/1
#1 lore raw
Quoted 13 lines · Kees Cook

For pid namespaces, struct pid uses a dynamically sized array member, "numbers". This was implemented using the ancient 1-element fake flexible array, which has been deprecated for decades. Replace it with a C99 flexible array, refactor the array size calculations to use struct_size(), and address elements via indexes. Note that the static initializer (which defines a single element) works as-is, and requires no special handling.

Without this, CONFIG_UBSAN_BOUNDS (and potentially CONFIG_FORTIFY_SOURCE) will trigger bounds checks when entering a pid namespace:
https://lore.kernel.org/lkml/20230517-bushaltestelle-super-e223978c1ba6@brauner

For example: unshare --fork --pid --mount-proc readlink /proc/self

Quoted 14 lines
Cc Christian Brauner <[email protected]>
Cc Jan Kara <[email protected]>
Cc Andreas Gruenbacher <[email protected]>
Cc Daniel Verkamp <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Jeff Xu <[email protected]>
Cc Andrew Morton <[email protected]>
Cc Boqun Feng <[email protected]>
Cc Luis Chamberlain <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Reported-by [email protected]

I would vote in favor of the array over the pointer notation, but anyway... we need to land this change, and that's what matters.

Thanks!
--
Gustavo

Quoted 56 lines

---
v2:
- drop pointer math to array index conversions (torvalds)
- use struct_size_t now that it exists (torvalds)
- updated commit log with reproducer example
v1: https://lore.kernel.org/lkml/[email protected]/

---
  include/linux/pid.h    | 2 +-
  kernel/pid.c           | 7 +++++--
  kernel/pid_namespace.c | 2 +-
  3 files changed, 7 insertions(+), 4 deletions(-)
include/linux/pid.h +1 -1
index b75de288a8c2..653a527574c4 100644
@@ -67,7 +67,7 @@ struct pid
  	/* wait queue for pidfd notifications */
  	wait_queue_head_t wait_pidfd;
  	struct rcu_head rcu;
-	struct upid numbers[1];
+	struct upid numbers[];
  };
 
  extern struct pid init_struct_pid;
kernel/pid.c +5 -2
index f93954a0384d..6a1d23a11026 100644
@@ -656,8 +656,11 @@ void __init pid_idr_init(void)
 
  	idr_init(&init_pid_ns.idr);
 
-	init_pid_ns.pid_cachep = KMEM_CACHE(pid,
-			SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT);
+	init_pid_ns.pid_cachep = kmem_cache_create("pid",
+			struct_size_t(struct pid, numbers, 1),
+			__alignof__(struct pid),
+			SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT,
+			NULL);
  }
 
  static struct file *__pidfd_fget(struct task_struct *task, int fd)
kernel/pid_namespace.c +1 -1
index b43eee07b00c..0bf44afe04dd 100644
@@ -48,7 +48,7 @@ static struct kmem_cache *create_pid_cachep(unsigned int level)
  		return kc;
 
  	snprintf(name, sizeof(name), "pid_%u", level + 1);
-	len = sizeof(struct pid) + level * sizeof(struct upid);
+	len = struct_size_t(struct pid, numbers, level + 1);
  	mutex_lock(&pid_caches_mutex);
  	/* Name collision forces to do allocation under mutex. */
  	if (!*pkc)
#2 lore raw
In reply to Kees Cook
Quoted 27 lines · Kees Cook

For pid namespaces, struct pid uses a dynamically sized array member, "numbers". This was implemented using the ancient 1-element fake flexible array, which has been deprecated for decades. Replace it with a C99 flexible array, refactor the array size calculations to use struct_size(), and address elements via indexes. Note that the static initializer (which defines a single element) works as-is, and requires no special handling.

Without this, CONFIG_UBSAN_BOUNDS (and potentially CONFIG_FORTIFY_SOURCE) will trigger bounds checks when entering a pid namespace:
https://lore.kernel.org/lkml/20230517-bushaltestelle-super-e223978c1ba6@brauner

For example: unshare --fork --pid --mount-proc readlink /proc/self

Cc Christian Brauner <[email protected]>
Cc Jan Kara <[email protected]>
Cc Andreas Gruenbacher <[email protected]>
Cc Daniel Verkamp <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
Cc Jeff Xu <[email protected]>
Cc Andrew Morton <[email protected]>
Cc Boqun Feng <[email protected]>
Cc Luis Chamberlain <[email protected]>
Cc Frederic Weisbecker <[email protected]>
Reported-by [email protected]
Quoted 5 lines

---
v2:
- drop pointer math to array index conversions (torvalds)
- use struct_size_t now that it exists (torvalds)
- updated commit log with reproducer example

Thanks for that. Linus already merged the other version I gave him with his requested changes but without your struct_size_t() update. It didn't exist when you originally sent that patch afaiu and I just realized that it existed right now. In any case, I'm just going to send a trivial follow-up patch switching the two places to struct_size_t().

Thanks, Kees!