char random: fix boot id uniqueness race (v2)

The proc file /proc/sys/kernel/random/boot_id can be read concurrently by user-space processes. If two (or more) user-space processes concurrently read boot_id when sysctl_bootid is not yet assigned, a race can occur making boot_id differ between the reads. Because the whole point of the boot id is to be unique across a kernel execution, fix this by protecting this operation with a mutex, and introduce a boot_id_generated flag, along with appropriate memory barriers, to let the fast-path know if the boot ID has been generated without having to hold the mutex.

I propose this approach rather than setting it up within an initcall(), because letting execution randomness add to entropy before populating the boot id seems to be a wanted property. Also, populating it lazily rather than at boot time only makes the performance hit be taken when boot_id is being read.

Q: Why are these memory barriers required ? Aren't the mutexes already
dealing with ordering ?

The need for memory barriers is a consequence of letting the fast-path run without holding this mutex.

Here is the race dealt with by the smp_rmb()/smp_wmb(). I'm showing the result of reversed write order here:

CPU A CPU B

Load boot_id_generated
(test -> false)
mutex_lock(&boot_id_mutex)
(implied memory barrier
with acquire semantic)
Load boot_id_generated again
(test -> false)
boot_id_generated = 1
(both the compiler and
CPU are free to reorder
the boot_id_generated
store before uuid stores)
Load boot_id_generated
(test -> true)
Load uuid content
(races with generate_random_uuid:
result either 0 or corrupted)
Return corrupted uuid.
generate_random_uuid(uuid)
mutex_unlock(&boot_id_mutex)

I prefer not requiring the fast-path to take a mutex, because this would transform a read-mostly operation into an operation that requires cache-line exchanges (the mutex). However, if we want the fast-path to be mutex-free, we need to enforce order with memory barriers: smp_rmb on the read-side, smp_wmb on the update-side. Failure to do so leads to the race shown above, where a corrupted boot_id can be returned.

* Changelog since v1:
- boot_id_mutex is now declared within the proc_do_uuid scope.
- added explanation for memory barriers.

Signed-off-by Mathieu Desnoyers <[email protected]>

CC "Theodore Ts'o" <[email protected]>

CC Matt Mackall <[email protected]>

CC Greg Kroah-Hartman <[email protected]>

---
 drivers/char/random.c |   20 +++++++++++++++++---
 1 file changed, 17 insertions(+), 3 deletions(-)

Index: linux-2.6-lttng/drivers/char/random.c
===================================================================

drivers/char/random.c +10 -3

@@ -1244,16 +1244,30 @@ static char sysctl_bootid[16];
 static int proc_do_uuid(ctl_table *table, int write,
 			void __user *buffer, size_t *lenp, loff_t *ppos)
 {
+	static int boot_id_generated;
+	static DEFINE_MUTEX(boot_id_mutex);
 	ctl_table fake_table;
 	unsigned char buf[64], tmp_uuid[16], *uuid;
 
 	uuid = table->data;
 	if (!uuid) {
 		uuid = tmp_uuid;
-		uuid[8] = 0;
-	}
-	if (uuid[8] == 0)
 		generate_random_uuid(uuid);
+	} else {
+		if (unlikely(!ACCESS_ONCE(boot_id_generated))) {
+			mutex_lock(&boot_id_mutex);
+			if (!boot_id_generated) {
+				generate_random_uuid(uuid);
+				/* Store uuid before boot_id_generated. */
+				smp_wmb();
+				boot_id_generated = 1;

Quoted 106 lines · Mathieu Desnoyers

The proc file /proc/sys/kernel/random/boot_id can be read concurrently by user-space processes. If two (or more) user-space processes concurrently read boot_id when sysctl_bootid is not yet assigned, a race can occur making boot_id differ between the reads. Because the whole point of the boot id is to be unique across a kernel execution, fix this by protecting this operation with a mutex, and introduce a boot_id_generated flag, along with appropriate memory barriers, to let the fast-path know if the boot ID has been generated without having to hold the mutex.

I propose this approach rather than setting it up within an initcall(), because letting execution randomness add to entropy before populating the boot id seems to be a wanted property. Also, populating it lazily rather than at boot time only makes the performance hit be taken when boot_id is being read.

Q: Why are these memory barriers required ? Aren't the mutexes already
dealing with ordering ?

The need for memory barriers is a consequence of letting the fast-path run without holding this mutex.

Here is the race dealt with by the smp_rmb()/smp_wmb(). I'm showing the result of reversed write order here:

CPU A CPU B

Load boot_id_generated
(test -> false)
mutex_lock(&boot_id_mutex)
(implied memory barrier
with acquire semantic)
Load boot_id_generated again
(test -> false)
boot_id_generated = 1
(both the compiler and
CPU are free to reorder
the boot_id_generated
store before uuid stores)
Load boot_id_generated
(test -> true)
Load uuid content
(races with generate_random_uuid:
result either 0 or corrupted)
Return corrupted uuid.
generate_random_uuid(uuid)
mutex_unlock(&boot_id_mutex)

I prefer not requiring the fast-path to take a mutex, because this would transform a read-mostly operation into an operation that requires cache-line exchanges (the mutex). However, if we want the fast-path to be mutex-free, we need to enforce order with memory barriers: smp_rmb on the read-side, smp_wmb on the update-side. Failure to do so leads to the race shown above, where a corrupted boot_id can be returned.

* Changelog since v1:
- boot_id_mutex is now declared within the proc_do_uuid scope.
- added explanation for memory barriers.

Signed-off-by Mathieu Desnoyers <[email protected]>

CC "Theodore Ts'o" <[email protected]>

CC Matt Mackall <[email protected]>

CC Greg Kroah-Hartman <[email protected]>

---
 drivers/char/random.c |   20 +++++++++++++++++---
 1 file changed, 17 insertions(+), 3 deletions(-)

Index: linux-2.6-lttng/drivers/char/random.c
===================================================================

drivers/char/random.c +17 -3

@@ -1244,16 +1244,30 @@ static char sysctl_bootid[16];
 static int proc_do_uuid(ctl_table *table, int write,
 			void __user *buffer, size_t *lenp, loff_t *ppos)
 {
+	static int boot_id_generated;
+	static DEFINE_MUTEX(boot_id_mutex);
 	ctl_table fake_table;
 	unsigned char buf[64], tmp_uuid[16], *uuid;
 
 	uuid = table->data;
 	if (!uuid) {
 		uuid = tmp_uuid;
-		uuid[8] = 0;
-	}
-	if (uuid[8] == 0)
 		generate_random_uuid(uuid);
+	} else {
+		if (unlikely(!ACCESS_ONCE(boot_id_generated))) {
+			mutex_lock(&boot_id_mutex);
+			if (!boot_id_generated) {
+				generate_random_uuid(uuid);
+				/* Store uuid before boot_id_generated. */
+				smp_wmb();
+				boot_id_generated = 1;
+			}
+			mutex_unlock(&boot_id_mutex);
+		} else {
+			/* Load boot_id_generated before uuid */
+			smp_rmb();
+		}
+	}
 

drivers/char/random.c static int proc_do_uuid(ctl_table *table, int write, · line 1273

Quoted 3 lines

drivers/char/random.c +0 -0

 	sprintf(buf, "%pU", uuid);
 

This seems overly complex to me.

I doubt this is performance critical path ?

What about a basic patch like :

drivers/char/random.c +0 -1

index 54ca8b2..af6040d 100644
@@ -1260,11 +1260,15 @@ static int proc_do_uuid(ctl_table *table, int write,
 	uuid = table->data;
 	if (!uuid) {
 		uuid = tmp_uuid;
-		uuid[8] = 0;

drivers/char/random.c +1 -0

+	}
 	sprintf(buf, "%pU", uuid);
 
 	fake_table.data = buf;

That would make sense, as long as we're OK about turning a read-mostly operation into a fully serialized operation that requires to exchange the lock between processor cache-lines. But as you point out, it should be fairly unfrequently used.

Any particular reason to use a spin lock rather than a mutex ? I did put a mutex in my implementation assuming that it would be a little more RT-friendly.