random: add a spinlock_t to struct batched_entropy
authorSebastian Andrzej Siewior <bigeasy@linutronix.de>
Sat, 20 Apr 2019 04:09:51 +0000 (00:09 -0400)
committerTheodore Ts'o <tytso@mit.edu>
Sat, 20 Apr 2019 04:09:56 +0000 (00:09 -0400)
The per-CPU variable batched_entropy_uXX is protected by get_cpu_var().
This is just a preempt_disable() which ensures that the variable is only
from the local CPU. It does not protect against users on the same CPU
from another context. It is possible that a preemptible context reads
slot 0 and then an interrupt occurs and the same value is read again.

The above scenario is confirmed by lockdep if we add a spinlock:
| ================================
| WARNING: inconsistent lock state
| 5.1.0-rc3+ #42 Not tainted
| --------------------------------
| inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W} usage.
| ksoftirqd/9/56 [HC0[0]:SC1[1]:HE0:SE0] takes:
| (____ptrval____) (batched_entropy_u32.lock){+.?.}, at: get_random_u32+0x3e/0xe0
| {SOFTIRQ-ON-W} state was registered at:
|   _raw_spin_lock+0x2a/0x40
|   get_random_u32+0x3e/0xe0
|   new_slab+0x15c/0x7b0
|   ___slab_alloc+0x492/0x620
|   __slab_alloc.isra.73+0x53/0xa0
|   kmem_cache_alloc_node+0xaf/0x2a0
|   copy_process.part.41+0x1e1/0x2370
|   _do_fork+0xdb/0x6d0
|   kernel_thread+0x20/0x30
|   kthreadd+0x1ba/0x220
|   ret_from_fork+0x3a/0x50

| other info that might help us debug this:
|  Possible unsafe locking scenario:
|
|        CPU0
|        ----
|   lock(batched_entropy_u32.lock);
|   <Interrupt>
|     lock(batched_entropy_u32.lock);
|
|  *** DEADLOCK ***
|
| stack backtrace:
| Call Trace:

|  kmem_cache_alloc_trace+0x20e/0x270
|  ipmi_alloc_recv_msg+0x16/0x40

|  __do_softirq+0xec/0x48d
|  run_ksoftirqd+0x37/0x60
|  smpboot_thread_fn+0x191/0x290
|  kthread+0xfe/0x130
|  ret_from_fork+0x3a/0x50

Add a spinlock_t to the batched_entropy data structure and acquire the
lock while accessing it. Acquire the lock with disabled interrupts
because this function may be used from interrupt context.

Remove the batched_entropy_reset_lock lock. Now that we have a lock for
the data scructure, we can access it from a remote CPU.

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
drivers/char/random.c

index 587df86..a42b3d7 100644 (file)
@@ -2282,8 +2282,8 @@ struct batched_entropy {
                u32 entropy_u32[CHACHA_BLOCK_SIZE / sizeof(u32)];
        };
        unsigned int position;
+       spinlock_t batch_lock;
 };
-static rwlock_t batched_entropy_reset_lock = __RW_LOCK_UNLOCKED(batched_entropy_reset_lock);
 
 /*
  * Get a random word for internal kernel use only. The quality of the random
@@ -2293,12 +2293,14 @@ static rwlock_t batched_entropy_reset_lock = __RW_LOCK_UNLOCKED(batched_entropy_
  * wait_for_random_bytes() should be called and return 0 at least once
  * at any point prior.
  */
-static DEFINE_PER_CPU(struct batched_entropy, batched_entropy_u64);
+static DEFINE_PER_CPU(struct batched_entropy, batched_entropy_u64) = {
+       .batch_lock     = __SPIN_LOCK_UNLOCKED(batched_entropy_u64.lock),
+};
+
 u64 get_random_u64(void)
 {
        u64 ret;
-       bool use_lock;
-       unsigned long flags = 0;
+       unsigned long flags;
        struct batched_entropy *batch;
        static void *previous;
 
@@ -2313,28 +2315,25 @@ u64 get_random_u64(void)
 
        warn_unseeded_randomness(&previous);
 
-       use_lock = READ_ONCE(crng_init) < 2;
-       batch = &get_cpu_var(batched_entropy_u64);
-       if (use_lock)
-               read_lock_irqsave(&batched_entropy_reset_lock, flags);
+       batch = raw_cpu_ptr(&batched_entropy_u64);
+       spin_lock_irqsave(&batch->batch_lock, flags);
        if (batch->position % ARRAY_SIZE(batch->entropy_u64) == 0) {
                extract_crng((u8 *)batch->entropy_u64);
                batch->position = 0;
        }
        ret = batch->entropy_u64[batch->position++];
-       if (use_lock)
-               read_unlock_irqrestore(&batched_entropy_reset_lock, flags);
-       put_cpu_var(batched_entropy_u64);
+       spin_unlock_irqrestore(&batch->batch_lock, flags);
        return ret;
 }
 EXPORT_SYMBOL(get_random_u64);
 
-static DEFINE_PER_CPU(struct batched_entropy, batched_entropy_u32);
+static DEFINE_PER_CPU(struct batched_entropy, batched_entropy_u32) = {
+       .batch_lock     = __SPIN_LOCK_UNLOCKED(batched_entropy_u32.lock),
+};
 u32 get_random_u32(void)
 {
        u32 ret;
-       bool use_lock;
-       unsigned long flags = 0;
+       unsigned long flags;
        struct batched_entropy *batch;
        static void *previous;
 
@@ -2343,18 +2342,14 @@ u32 get_random_u32(void)
 
        warn_unseeded_randomness(&previous);
 
-       use_lock = READ_ONCE(crng_init) < 2;
-       batch = &get_cpu_var(batched_entropy_u32);
-       if (use_lock)
-               read_lock_irqsave(&batched_entropy_reset_lock, flags);
+       batch = raw_cpu_ptr(&batched_entropy_u32);
+       spin_lock_irqsave(&batch->batch_lock, flags);
        if (batch->position % ARRAY_SIZE(batch->entropy_u32) == 0) {
                extract_crng((u8 *)batch->entropy_u32);
                batch->position = 0;
        }
        ret = batch->entropy_u32[batch->position++];
-       if (use_lock)
-               read_unlock_irqrestore(&batched_entropy_reset_lock, flags);
-       put_cpu_var(batched_entropy_u32);
+       spin_unlock_irqrestore(&batch->batch_lock, flags);
        return ret;
 }
 EXPORT_SYMBOL(get_random_u32);
@@ -2368,12 +2363,19 @@ static void invalidate_batched_entropy(void)
        int cpu;
        unsigned long flags;
 
-       write_lock_irqsave(&batched_entropy_reset_lock, flags);
        for_each_possible_cpu (cpu) {
-               per_cpu_ptr(&batched_entropy_u32, cpu)->position = 0;
-               per_cpu_ptr(&batched_entropy_u64, cpu)->position = 0;
+               struct batched_entropy *batched_entropy;
+
+               batched_entropy = per_cpu_ptr(&batched_entropy_u32, cpu);
+               spin_lock_irqsave(&batched_entropy->batch_lock, flags);
+               batched_entropy->position = 0;
+               spin_unlock(&batched_entropy->batch_lock);
+
+               batched_entropy = per_cpu_ptr(&batched_entropy_u64, cpu);
+               spin_lock(&batched_entropy->batch_lock);
+               batched_entropy->position = 0;
+               spin_unlock_irqrestore(&batched_entropy->batch_lock, flags);
        }
-       write_unlock_irqrestore(&batched_entropy_reset_lock, flags);
 }
 
 /**