x86/mm/kmmio: Have mmiotracer play nice with lockdep

The mmiotrace tracer is "special". The purpose is to help reverse engineer binary drivers by removing the memory allocated by the driver and when the driver goes to access it, a fault occurs, the mmiotracer will record what the driver was doing and then do the work on its behalf by single stepping through the process.

But to achieve this ability, it must do some special things. One is it needs to grab a lock while in the breakpoint handler. This is considered an NMI state, and then lockdep warns that the lock is being held in both an NMI state (really a breakpoint handler) and also in normal context.

As the breakpoint/NMI state only happens when the driver is accessing memory, there's no concern of a race condition against the setup and tear-down of mmiotracer.

To make lockdep and mmiotrace work together, convert the locks used in the breakpoint handler into arch_spin_lock().

It also takes the rcu_read_lock() at the fault, and releases it in the single step breakpoint. This makes lockdep sad. Luckily, it also disables preemption when taking the rcu_read_lock() and enables it when releasing the rcu_read_lock(). As now the sched RCU variant is basically the same as the "normal" RCU variant, there's no reason to take the rcu_read_lock(), so just remove it.

Steven Rostedt (2):
x86/mm/kmmio: Switch to arch_spin_lock()
x86/mm/kmmio: Remove rcu_read_lock()

----

 arch/x86/mm/kmmio.c | 34 ++++++++++++++++++++++------------
 1 file changed, 22 insertions(+), 12 deletions(-)

From: Steven Rostedt <[email protected]>

The mmiotrace tracer is "special". The purpose is to help reverse engineer binary drivers by removing the memory allocated by the driver and when the driver goes to access it, a fault occurs, the mmiotracer will record what the driver was doing and then do the work on its behalf by single stepping through the process.

But to achieve this ability, it must do some special things. One is it needs to grab a lock while in the breakpoint handler. This is considered an NMI state, and then lockdep warns that the lock is being held in both an NMI state (really a breakpoint handler) and also in normal context.

As the breakpoint/NMI state only happens when the driver is accessing memory, there's no concern of a race condition against the setup and tear-down of mmiotracer.

To make lockdep and mmiotrace work together, convert the locks used in the breakpoint handler into arch_spin_lock().

---
 arch/x86/mm/kmmio.c | 31 ++++++++++++++++++++++---------
 1 file changed, 22 insertions(+), 9 deletions(-)

index d3efbc5b3449..edb486450158 100644
@@ -62,7 +62,13 @@ struct kmmio_context {
 	int active;
 };
 
-static DEFINE_SPINLOCK(kmmio_lock);
+/*
+ * The kmmio_lock is taken in int3 context, which is treated as NMI context.
+ * This causes lockdep to complain about it bein in both NMI and normal
+ * context. Hide it from lockdep, as it should not have any other locks
+ * taken under it, and this is only enabled for debugging mmio anyway.
+ */
+static arch_spinlock_t kmmio_lock = __ARCH_SPIN_LOCK_UNLOCKED;
 
 /* Protected by kmmio_lock */
 unsigned int kmmio_count;

@@ -346,10 +352,10 @@ static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
 		ctx->probe->post_handler(ctx->probe, condition, regs);
 
 	/* Prevent racing against release_kmmio_fault_page(). */
-	spin_lock(&kmmio_lock);
+	arch_spin_lock(&kmmio_lock);
 	if (ctx->fpage->count)
 		arm_kmmio_fault_page(ctx->fpage);
-	spin_unlock(&kmmio_lock);
+	arch_spin_unlock(&kmmio_lock);
 
 	regs->flags &= ~X86_EFLAGS_TF;
 	regs->flags |= ctx->saved_flags;

@@ -440,7 +446,8 @@ int register_kmmio_probe(struct kmmio_probe *p)
 	unsigned int l;
 	pte_t *pte;
 
-	spin_lock_irqsave(&kmmio_lock, flags);
+	local_irq_save(flags);
+	arch_spin_lock(&kmmio_lock);
 	if (get_kmmio_probe(addr)) {
 		ret = -EEXIST;
 		goto out;

@@ -460,7 +467,9 @@ int register_kmmio_probe(struct kmmio_probe *p)
 		size += page_level_size(l);
 	}
 out:
-	spin_unlock_irqrestore(&kmmio_lock, flags);
+	arch_spin_unlock(&kmmio_lock);
+	local_irq_restore(flags);
+
 	/*
 	 * XXX: What should I do here?
 	 * Here was a call to global_flush_tlb(), but it does not exist

@@ -494,7 +503,8 @@ static void remove_kmmio_fault_pages(struct rcu_head *head)
 	struct kmmio_fault_page **prevp = &dr->release_list;
 	unsigned long flags;
 
-	spin_lock_irqsave(&kmmio_lock, flags);
+	local_irq_save(flags);
+	arch_spin_lock(&kmmio_lock);
 	while (f) {
 		if (!f->count) {
 			list_del_rcu(&f->list);

@@ -506,7 +516,8 @@ static void remove_kmmio_fault_pages(struct rcu_head *head)
 		}
 		f = *prevp;
 	}
-	spin_unlock_irqrestore(&kmmio_lock, flags);
+	arch_spin_unlock(&kmmio_lock);
+	local_irq_restore(flags);
 
 	/* This is the real RCU destroy call. */
 	call_rcu(&dr->rcu, rcu_free_kmmio_fault_pages);

@@ -540,14 +551,16 @@ void unregister_kmmio_probe(struct kmmio_probe *p)
 	if (!pte)
 		return;
 
-	spin_lock_irqsave(&kmmio_lock, flags);
+	local_irq_save(flags);
+	arch_spin_lock(&kmmio_lock);
 	while (size < size_lim) {
 		release_kmmio_fault_page(addr + size, &release_list);
 		size += page_level_size(l);
 	}
 	list_del_rcu(&p->list);
 	kmmio_count--;
-	spin_unlock_irqrestore(&kmmio_lock, flags);
+	arch_spin_unlock(&kmmio_lock);
+	local_irq_restore(flags);
 
 	if (!release_list)
 		return;

2.35.1

From: Steven Rostedt <[email protected]>

The mmiotrace tracer is "special". The purpose is to help reverse engineer binary drivers by removing the memory allocated by the driver and when the driver goes to access it, a fault occurs, the mmiotracer will record what the driver was doing and then do the work on its behalf by single stepping through the process.

But to achieve this ability, it must do some special things. One is it take the rcu_read_lock() when the fault occurs, and then release it in the breakpoint that in the single stepping. This makes lockdep unhappy, as it changes the state of RCU from within an exception that is not contained in that exception, and we get a nasty splat from lockdep.

As it also disables preemption everywhere rcu_read_lock() is taken, and enables preemption everywhere rcu_read_unlock(), and does not enable preemption in between, it is the same as synchronize_rcu_sched(). But as the RCU sched variant has the same grace period as normal RCU, there's no reason to take the rcu_read_lock(). Simply remove it.

---
 arch/x86/mm/kmmio.c | 3 ---
 1 file changed, 3 deletions(-)

index edb486450158..e15e3aaaf94c 100644
@@ -254,7 +254,6 @@ int kmmio_handler(struct pt_regs *regs, unsigned long addr)
 	 * again.
 	 */
 	preempt_disable();
-	rcu_read_lock();
 
 	faultpage = get_kmmio_fault_page(page_base);
 	if (!faultpage) {

@@ -323,7 +322,6 @@ int kmmio_handler(struct pt_regs *regs, unsigned long addr)
 	return 1; /* fault handled */
 
 no_kmmio:
-	rcu_read_unlock();
 	preempt_enable_no_resched();
 	return ret;
 }

@@ -363,7 +361,6 @@ static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs)
 	/* These were acquired in kmmio_handler(). */
 	ctx->active--;
 	BUG_ON(ctx->active);
-	rcu_read_unlock();
 	preempt_enable_no_resched();
 
 	/*

2.35.1

Might be worth adding a comment saying that others are using this preempt_disable() to block an RCU grace period, but that is up to you guys. I will let you and your future selves be the judges.

Good point. I'll add a comment in v2.

Thanks!

-- Steve

Actually, rcu_read_lock_sched_notrace() may work instead. Let me test it.

-- Steve