What is RCU? -- “Read, Copy, Update”

Please note that the “What is RCU?” LWN series is an excellent place to start learning about RCU:

1. What is RCU, Fundamentally? https://lwn.net/Articles/262464/
2. What is RCU? Part 2: Usage https://lwn.net/Articles/263130/
3. RCU part 3: the RCU API https://lwn.net/Articles/264090/
4. The RCU API, 2010 Edition https://lwn.net/Articles/418853/
5. The RCU API, 2014 Edition https://lwn.net/Articles/609904/
6. The RCU API, 2019 Edition https://lwn.net/Articles/777036/

For those preferring video:

What is RCU?

RCU is a synchronization mechanism that was added to the Linux kernel during the 2.5 development effort that is optimized for read-mostly situations. Although RCU is actually quite simple, making effective use of it requires you to think differently about your code. Another part of the problem is the mistaken assumption that there is “one true way” to describe and to use RCU. Instead, the experience has been that different people must take different paths to arrive at an understanding of RCU, depending on their experiences and use cases. This document provides several different paths, as follows:

1. RCU OVERVIEW

2. WHAT IS RCU’S CORE API?

3. WHAT ARE SOME EXAMPLE USES OF CORE RCU API?

4. WHAT IF MY UPDATING THREAD CANNOT BLOCK?

5. WHAT ARE SOME SIMPLE IMPLEMENTATIONS OF RCU?

6. ANALOGY WITH READER-WRITER LOCKING

7. ANALOGY WITH REFERENCE COUNTING

8. FULL LIST OF RCU APIs

9. ANSWERS TO QUICK QUIZZES

People who prefer starting with a conceptual overview should focus on Section 1, though most readers will profit by reading this section at some point. People who prefer to start with an API that they can then experiment with should focus on Section 2. People who prefer to start with example uses should focus on Sections 3 and 4. People who need to understand the RCU implementation should focus on Section 5, then dive into the kernel source code. People who reason best by analogy should focus on Section 6 and 7. Section 8 serves as an index to the docbook API documentation, and Section 9 is the traditional answer key.

So, start with the section that makes the most sense to you and your preferred method of learning. If you need to know everything about everything, feel free to read the whole thing -- but if you are really that type of person, you have perused the source code and will therefore never need this document anyway. ;-)

1. RCU OVERVIEW

The basic idea behind RCU is to split updates into “removal” and “reclamation” phases. The removal phase removes references to data items within a data structure (possibly by replacing them with references to new versions of these data items), and can run concurrently with readers. The reason that it is safe to run the removal phase concurrently with readers is the semantics of modern CPUs guarantee that readers will see either the old or the new version of the data structure rather than a partially updated reference. The reclamation phase does the work of reclaiming (e.g., freeing) the data items removed from the data structure during the removal phase. Because reclaiming data items can disrupt any readers concurrently referencing those data items, the reclamation phase must not start until readers no longer hold references to those data items.

Splitting the update into removal and reclamation phases permits the updater to perform the removal phase immediately, and to defer the reclamation phase until all readers active during the removal phase have completed, either by blocking until they finish or by registering a callback that is invoked after they finish. Only readers that are active during the removal phase need be considered, because any reader starting after the removal phase will be unable to gain a reference to the removed data items, and therefore cannot be disrupted by the reclamation phase.

So the typical RCU update sequence goes something like the following:

  1. Remove pointers to a data structure, so that subsequent readers cannot gain a reference to it.

  2. Wait for all previous readers to complete their RCU read-side critical sections.

  3. At this point, there cannot be any readers who hold references to the data structure, so it now may safely be reclaimed (e.g., kfree()d).

Step (b) above is the key idea underlying RCU’s deferred destruction. The ability to wait until all readers are done allows RCU readers to use much lighter-weight synchronization, in some cases, absolutely no synchronization at all. In contrast, in more conventional lock-based schemes, readers must use heavy-weight synchronization in order to prevent an updater from deleting the data structure out from under them. This is because lock-based updaters typically update data items in place, and must therefore exclude readers. In contrast, RCU-based updaters typically take advantage of the fact that writes to single aligned pointers are atomic on modern CPUs, allowing atomic insertion, removal, and replacement of data items in a linked structure without disrupting readers. Concurrent RCU readers can then continue accessing the old versions, and can dispense with the atomic operations, memory barriers, and communications cache misses that are so expensive on present-day SMP computer systems, even in absence of lock contention.

In the three-step procedure shown above, the updater is performing both the removal and the reclamation step, but it is often helpful for an entirely different thread to do the reclamation, as is in fact the case in the Linux kernel’s directory-entry cache (dcache). Even if the same thread performs both the update step (step (a) above) and the reclamation step (step (c) above), it is often helpful to think of them separately. For example, RCU readers and updaters need not communicate at all, but RCU provides implicit low-overhead communication between readers and reclaimers, namely, in step (b) above.

So how the heck can a reclaimer tell when a reader is done, given that readers are not doing any sort of synchronization operations??? Read on to learn about how RCU’s API makes this easy.

2. WHAT IS RCU’S CORE API?

The core RCU API is quite small:

  1. rcu_read_lock()

  2. rcu_read_unlock()

  3. synchronize_rcu() / call_rcu()

  4. rcu_assign_pointer()

  5. rcu_dereference()

There are many other members of the RCU API, but the rest can be expressed in terms of these five, though most implementations instead express synchronize_rcu() in terms of the call_rcu() callback API.

The five core RCU APIs are described below, the other 18 will be enumerated later. See the kernel docbook documentation for more info, or look directly at the function header comments.

rcu_read_lock()

void rcu_read_lock(void);

This temporal primitive is used by a reader to inform the reclaimer that the reader is entering an RCU read-side critical section. It is illegal to block while in an RCU read-side critical section, though kernels built with CONFIG_PREEMPT_RCU can preempt RCU read-side critical sections. Any RCU-protected data structure accessed during an RCU read-side critical section is guaranteed to remain unreclaimed for the full duration of that critical section. Reference counts may be used in conjunction with RCU to maintain longer-term references to data structures.

Note that anything that disables bottom halves, preemption, or interrupts also enters an RCU read-side critical section. Acquiring a spinlock also enters an RCU read-side critical sections, even for spinlocks that do not disable preemption, as is the case in kernels built with CONFIG_PREEMPT_RT=y. Sleeplocks do not enter RCU read-side critical sections.

rcu_read_unlock()

void rcu_read_unlock(void);

This temporal primitives is used by a reader to inform the reclaimer that the reader is exiting an RCU read-side critical section. Anything that enables bottom halves, preemption, or interrupts also exits an RCU read-side critical section. Releasing a spinlock also exits an RCU read-side critical section.

Note that RCU read-side critical sections may be nested and/or overlapping.

synchronize_rcu()

void synchronize_rcu(void);

This temporal primitive marks the end of updater code and the beginning of reclaimer code. It does this by blocking until all pre-existing RCU read-side critical sections on all CPUs have completed. Note that synchronize_rcu() will not necessarily wait for any subsequent RCU read-side critical sections to complete. For example, consider the following sequence of events:

        CPU 0                  CPU 1                 CPU 2
    ----------------- ------------------------- ---------------
1.  rcu_read_lock()
2.                    enters synchronize_rcu()
3.                                               rcu_read_lock()
4.  rcu_read_unlock()
5.                     exits synchronize_rcu()
6.                                              rcu_read_unlock()

To reiterate, synchronize_rcu() waits only for ongoing RCU read-side critical sections to complete, not necessarily for any that begin after synchronize_rcu() is invoked.

Of course, synchronize_rcu() does not necessarily return immediately after the last pre-existing RCU read-side critical section completes. For one thing, there might well be scheduling delays. For another thing, many RCU implementations process requests in batches in order to improve efficiencies, which can further delay synchronize_rcu().

Since synchronize_rcu() is the API that must figure out when readers are done, its implementation is key to RCU. For RCU to be useful in all but the most read-intensive situations, synchronize_rcu()’s overhead must also be quite small.

The call_rcu() API is an asynchronous callback form of synchronize_rcu(), and is described in more detail in a later section. Instead of blocking, it registers a function and argument which are invoked after all ongoing RCU read-side critical sections have completed. This callback variant is particularly useful in situations where it is illegal to block or where update-side performance is critically important.

However, the call_rcu() API should not be used lightly, as use of the synchronize_rcu() API generally results in simpler code. In addition, the synchronize_rcu() API has the nice property of automatically limiting update rate should grace periods be delayed. This property results in system resilience in face of denial-of-service attacks. Code using call_rcu() should limit update rate in order to gain this same sort of resilience. See Review Checklist for RCU Patches for some approaches to limiting the update rate.

rcu_assign_pointer()

void rcu_assign_pointer(p, typeof(p) v);

Yes, rcu_assign_pointer() is implemented as a macro, though it would be cool to be able to declare a function in this manner. (And there has been some discussion of adding overloaded functions to the C language, so who knows?)

The updater uses this spatial macro to assign a new value to an RCU-protected pointer, in order to safely communicate the change in value from the updater to the reader. This is a spatial (as opposed to temporal) macro. It does not evaluate to an rvalue, but it does provide any compiler directives and memory-barrier instructions required for a given compile or CPU architecture. Its ordering properties are that of a store-release operation, that is, any prior loads and stores required to initialize the structure are ordered before the store that publishes the pointer to that structure.

Perhaps just as important, rcu_assign_pointer() serves to document (1) which pointers are protected by RCU and (2) the point at which a given structure becomes accessible to other CPUs. That said, rcu_assign_pointer() is most frequently used indirectly, via the _rcu list-manipulation primitives such as list_add_rcu().

rcu_dereference()

typeof(p) rcu_dereference(p);

Like rcu_assign_pointer(), rcu_dereference() must be implemented as a macro.

The reader uses the spatial rcu_dereference() macro to fetch an RCU-protected pointer, which returns a value that may then be safely dereferenced. Note that rcu_dereference() does not actually dereference the pointer, instead, it protects the pointer for later dereferencing. It also executes any needed memory-barrier instructions for a given CPU architecture. Currently, only Alpha needs memory barriers within rcu_dereference() -- on other CPUs, it compiles to a volatile load. However, no mainstream C compilers respect address dependencies, so rcu_dereference() uses volatile casts, which, in combination with the coding guidelines listed in PROPER CARE AND FEEDING OF RETURN VALUES FROM rcu_dereference(), prevent current compilers from breaking these dependencies.

Common coding practice uses rcu_dereference() to copy an RCU-protected pointer to a local variable, then dereferences this local variable, for example as follows:

p = rcu_dereference(head.next);
return p->data;

However, in this case, one could just as easily combine these into one statement:

return rcu_dereference(head.next)->data;

If you are going to be fetching multiple fields from the RCU-protected structure, using the local variable is of course preferred. Repeated rcu_dereference() calls look ugly, do not guarantee that the same pointer will be returned if an update happened while in the critical section, and incur unnecessary overhead on Alpha CPUs.

Note that the value returned by rcu_dereference() is valid only within the enclosing RCU read-side critical section [1]. For example, the following is not legal:

rcu_read_lock();
p = rcu_dereference(head.next);
rcu_read_unlock();
x = p->address; /* BUG!!! */
rcu_read_lock();
y = p->data;    /* BUG!!! */
rcu_read_unlock();

Holding a reference from one RCU read-side critical section to another is just as illegal as holding a reference from one lock-based critical section to another! Similarly, using a reference outside of the critical section in which it was acquired is just as illegal as doing so with normal locking.

As with rcu_assign_pointer(), an important function of rcu_dereference() is to document which pointers are protected by RCU, in particular, flagging a pointer that is subject to changing at any time, including immediately after the rcu_dereference(). And, again like rcu_assign_pointer(), rcu_dereference() is typically used indirectly, via the _rcu list-manipulation primitives, such as list_for_each_entry_rcu() [2].

The following diagram shows how each API communicates among the reader, updater, and reclaimer.

rcu_assign_pointer()
                        +--------+
+---------------------->| reader |---------+
|                       +--------+         |
|                           |              |
|                           |              | Protect:
|                           |              | rcu_read_lock()
|                           |              | rcu_read_unlock()
|        rcu_dereference()  |              |
+---------+                 |              |
| updater |<----------------+              |
+---------+                                V
|                                    +-----------+
+----------------------------------->| reclaimer |
                                     +-----------+
  Defer:
  synchronize_rcu() & call_rcu()

The RCU infrastructure observes the temporal sequence of rcu_read_lock(), rcu_read_unlock(), synchronize_rcu(), and call_rcu() invocations in order to determine when (1) synchronize_rcu() invocations may return to their callers and (2) call_rcu() callbacks may be invoked. Efficient implementations of the RCU infrastructure make heavy use of batching in order to amortize their overhead over many uses of the corresponding APIs. The rcu_assign_pointer() and rcu_dereference() invocations communicate spatial changes via stores to and loads from the RCU-protected pointer in question.

There are at least three flavors of RCU usage in the Linux kernel. The diagram above shows the most common one. On the updater side, the rcu_assign_pointer(), synchronize_rcu() and call_rcu() primitives used are the same for all three flavors. However for protection (on the reader side), the primitives used vary depending on the flavor:

  1. rcu_read_lock() / rcu_read_unlock() rcu_dereference()

  2. rcu_read_lock_bh() / rcu_read_unlock_bh() local_bh_disable() / local_bh_enable() rcu_dereference_bh()

  3. rcu_read_lock_sched() / rcu_read_unlock_sched() preempt_disable() / preempt_enable() local_irq_save() / local_irq_restore() hardirq enter / hardirq exit NMI enter / NMI exit rcu_dereference_sched()

These three flavors are used as follows:

  1. RCU applied to normal data structures.

  2. RCU applied to networking data structures that may be subjected to remote denial-of-service attacks.

  3. RCU applied to scheduler and interrupt/NMI-handler tasks.

Again, most uses will be of (a). The (b) and (c) cases are important for specialized uses, but are relatively uncommon. The SRCU, RCU-Tasks, RCU-Tasks-Rude, and RCU-Tasks-Trace have similar relationships among their assorted primitives.

3. WHAT ARE SOME EXAMPLE USES OF CORE RCU API?

This section shows a simple use of the core RCU API to protect a global pointer to a dynamically allocated structure. More-typical uses of RCU may be found in Using RCU to Protect Read-Mostly Linked Lists and Using RCU to Protect Dynamic NMI Handlers.

struct foo {
        int a;
        char b;
        long c;
};
DEFINE_SPINLOCK(foo_mutex);

struct foo __rcu *gbl_foo;

/*
 * Create a new struct foo that is the same as the one currently
 * pointed to by gbl_foo, except that field "a" is replaced
 * with "new_a".  Points gbl_foo to the new structure, and
 * frees up the old structure after a grace period.
 *
 * Uses rcu_assign_pointer() to ensure that concurrent readers
 * see the initialized version of the new structure.
 *
 * Uses synchronize_rcu() to ensure that any readers that might
 * have references to the old structure complete before freeing
 * the old structure.
 */
void foo_update_a(int new_a)
{
        struct foo *new_fp;
        struct foo *old_fp;

        new_fp = kmalloc(sizeof(*new_fp), GFP_KERNEL);
        spin_lock(&foo_mutex);
        old_fp = rcu_dereference_protected(gbl_foo, lockdep_is_held(&foo_mutex));
        *new_fp = *old_fp;
        new_fp->a = new_a;
        rcu_assign_pointer(gbl_foo, new_fp);
        spin_unlock(&foo_mutex);
        synchronize_rcu();
        kfree(old_fp);
}

/*
 * Return the value of field "a" of the current gbl_foo
 * structure.  Use rcu_read_lock() and rcu_read_unlock()
 * to ensure that the structure does not get deleted out
 * from under us, and use rcu_dereference() to ensure that
 * we see the initialized version of the structure (important
 * for DEC Alpha and for people reading the code).
 */
int foo_get_a(void)
{
        int retval;

        rcu_read_lock();
        retval = rcu_dereference(gbl_foo)->a;
        rcu_read_unlock();
        return retval;
}

So, to sum up:

  • Use rcu_read_lock() and rcu_read_unlock() to guard RCU read-side critical sections.

  • Within an RCU read-side critical section, use rcu_dereference() to dereference RCU-protected pointers.

  • Use some solid design (such as locks or semaphores) to keep concurrent updates from interfering with each other.

  • Use rcu_assign_pointer() to update an RCU-protected pointer. This primitive protects concurrent readers from the updater, not concurrent updates from each other! You therefore still need to use locking (or something similar) to keep concurrent rcu_assign_pointer() primitives from interfering with each other.

  • Use synchronize_rcu() after removing a data element from an RCU-protected data structure, but before reclaiming/freeing the data element, in order to wait for the completion of all RCU read-side critical sections that might be referencing that data item.

See Review Checklist for RCU Patches for additional rules to follow when using RCU. And again, more-typical uses of RCU may be found in Using RCU to Protect Read-Mostly Linked Lists and Using RCU to Protect Dynamic NMI Handlers.

4. WHAT IF MY UPDATING THREAD CANNOT BLOCK?

In the example above, foo_update_a() blocks until a grace period elapses. This is quite simple, but in some cases one cannot afford to wait so long -- there might be other high-priority work to be done.

In such cases, one uses call_rcu() rather than synchronize_rcu(). The call_rcu() API is as follows:

void call_rcu(struct rcu_head *head, rcu_callback_t func);

This function invokes func(head) after a grace period has elapsed. This invocation might happen from either softirq or process context, so the function is not permitted to block. The foo struct needs to have an rcu_head structure added, perhaps as follows:

struct foo {
        int a;
        char b;
        long c;
        struct rcu_head rcu;
};

The foo_update_a() function might then be written as follows:

/*
 * Create a new struct foo that is the same as the one currently
 * pointed to by gbl_foo, except that field "a" is replaced
 * with "new_a".  Points gbl_foo to the new structure, and
 * frees up the old structure after a grace period.
 *
 * Uses rcu_assign_pointer() to ensure that concurrent readers
 * see the initialized version of the new structure.
 *
 * Uses call_rcu() to ensure that any readers that might have
 * references to the old structure complete before freeing the
 * old structure.
 */
void foo_update_a(int new_a)
{
        struct foo *new_fp;
        struct foo *old_fp;

        new_fp = kmalloc(sizeof(*new_fp), GFP_KERNEL);
        spin_lock(&foo_mutex);
        old_fp = rcu_dereference_protected(gbl_foo, lockdep_is_held(&foo_mutex));
        *new_fp = *old_fp;
        new_fp->a = new_a;
        rcu_assign_pointer(gbl_foo, new_fp);
        spin_unlock(&foo_mutex);
        call_rcu(&old_fp->rcu, foo_reclaim);
}

The foo_reclaim() function might appear as follows:

void foo_reclaim(struct rcu_head *rp)
{
        struct foo *fp = container_of(rp, struct foo, rcu);

        foo_cleanup(fp->a);

        kfree(fp);
}

The container_of() primitive is a macro that, given a pointer into a struct, the type of the struct, and the pointed-to field within the struct, returns a pointer to the beginning of the struct.

The use of call_rcu() permits the caller of foo_update_a() to immediately regain control, without needing to worry further about the old version of the newly updated element. It also clearly shows the RCU distinction between updater, namely foo_update_a(), and reclaimer, namely foo_reclaim().

The summary of advice is the same as for the previous section, except that we are now using call_rcu() rather than synchronize_rcu():

  • Use call_rcu() after removing a data element from an RCU-protected data structure in order to register a callback function that will be invoked after the completion of all RCU read-side critical sections that might be referencing that data item.

If the callback for call_rcu() is not doing anything more than calling kfree() on the structure, you can use kfree_rcu() instead of call_rcu() to avoid having to write your own callback:

kfree_rcu(old_fp, rcu);

If the occasional sleep is permitted, the single-argument form may be used, omitting the rcu_head structure from struct foo.

kfree_rcu_mightsleep(old_fp);

This variant almost never blocks, but might do so by invoking synchronize_rcu() in response to memory-allocation failure.

Again, see Review Checklist for RCU Patches for additional rules governing the use of RCU.

5. WHAT ARE SOME SIMPLE IMPLEMENTATIONS OF RCU?

One of the nice things about RCU is that it has extremely simple “toy” implementations that are a good first step towards understanding the production-quality implementations in the Linux kernel. This section presents two such “toy” implementations of RCU, one that is implemented in terms of familiar locking primitives, and another that more closely resembles “classic” RCU. Both are way too simple for real-world use, lacking both functionality and performance. However, they are useful in getting a feel for how RCU works. See kernel/rcu/update.c for a production-quality implementation, and see:

for papers describing the Linux kernel RCU implementation. The OLS’01 and OLS’02 papers are a good introduction, and the dissertation provides more details on the current implementation as of early 2004.

5A. “TOY” IMPLEMENTATION #1: LOCKING

This section presents a “toy” RCU implementation that is based on familiar locking primitives. Its overhead makes it a non-starter for real-life use, as does its lack of scalability. It is also unsuitable for realtime use, since it allows scheduling latency to “bleed” from one read-side critical section to another. It also assumes recursive reader-writer locks: If you try this with non-recursive locks, and you allow nested rcu_read_lock() calls, you can deadlock.

However, it is probably the easiest implementation to relate to, so is a good starting point.

It is extremely simple:

static DEFINE_RWLOCK(rcu_gp_mutex);

void rcu_read_lock(void)
{
        read_lock(&rcu_gp_mutex);
}

void rcu_read_unlock(void)
{
        read_unlock(&rcu_gp_mutex);
}

void synchronize_rcu(void)
{
        write_lock(&rcu_gp_mutex);
        smp_mb__after_spinlock();
        write_unlock(&rcu_gp_mutex);
}

[You can ignore rcu_assign_pointer() and rcu_dereference() without missing much. But here are simplified versions anyway. And whatever you do, don’t forget about them when submitting patches making use of RCU!]:

#define rcu_assign_pointer(p, v) \
({ \
        smp_store_release(&(p), (v)); \
})

#define rcu_dereference(p) \
({ \
        typeof(p) _________p1 = READ_ONCE(p); \
        (_________p1); \
})

The rcu_read_lock() and rcu_read_unlock() primitive read-acquire and release a global reader-writer lock. The synchronize_rcu() primitive write-acquires this same lock, then releases it. This means that once synchronize_rcu() exits, all RCU read-side critical sections that were in progress before synchronize_rcu() was called are guaranteed to have completed -- there is no way that synchronize_rcu() would have been able to write-acquire the lock otherwise. The smp_mb__after_spinlock() promotes synchronize_rcu() to a full memory barrier in compliance with the “Memory-Barrier Guarantees” listed in:

It is possible to nest rcu_read_lock(), since reader-writer locks may be recursively acquired. Note also that rcu_read_lock() is immune from deadlock (an important property of RCU). The reason for this is that the only thing that can block rcu_read_lock() is a synchronize_rcu(). But synchronize_rcu() does not acquire any locks while holding rcu_gp_mutex, so there can be no deadlock cycle.

Quick Quiz #1:

Why is this argument naive? How could a deadlock occur when using this algorithm in a real-world Linux kernel? How could this deadlock be avoided?

Answers to Quick Quiz

5B. “TOY” EXAMPLE #2: CLASSIC RCU

This section presents a “toy” RCU implementation that is based on “classic RCU”. It is also short on performance (but only for updates) and on features such as hotplug CPU and the ability to run in CONFIG_PREEMPTION kernels. The definitions of rcu_dereference() and rcu_assign_pointer() are the same as those shown in the preceding section, so they are omitted.

void rcu_read_lock(void) { }

void rcu_read_unlock(void) { }

void synchronize_rcu(void)
{
        int cpu;

        for_each_possible_cpu(cpu)
                run_on(cpu);
}

Note that rcu_read_lock() and rcu_read_unlock() do absolutely nothing. This is the great strength of classic RCU in a non-preemptive kernel: read-side overhead is precisely zero, at least on non-Alpha CPUs. And there is absolutely no way that rcu_read_lock() can possibly participate in a deadlock cycle!

The implementation of synchronize_rcu() simply schedules itself on each CPU in turn. The run_on() primitive can be implemented straightforwardly in terms of the sched_setaffinity() primitive. Of course, a somewhat less “toy” implementation would restore the affinity upon completion rather than just leaving all tasks running on the last CPU, but when I said “toy”, I meant toy!

So how the heck is this supposed to work???

Remember that it is illegal to block while in an RCU read-side critical section. Therefore, if a given CPU executes a context switch, we know that it must have completed all preceding RCU read-side critical sections. Once all CPUs have executed a context switch, then all preceding RCU read-side critical sections will have completed.

So, suppose that we remove a data item from its structure and then invoke synchronize_rcu(). Once synchronize_rcu() returns, we are guaranteed that there are no RCU read-side critical sections holding a reference to that data item, so we can safely reclaim it.

Quick Quiz #2:

Give an example where Classic RCU’s read-side overhead is negative.

Answers to Quick Quiz

Quick Quiz #3:

If it is illegal to block in an RCU read-side critical section, what the heck do you do in CONFIG_PREEMPT_RT, where normal spinlocks can block???

Answers to Quick Quiz

6. ANALOGY WITH READER-WRITER LOCKING

Although RCU can be used in many different ways, a very common use of RCU is analogous to reader-writer locking. The following unified diff shows how closely related RCU and reader-writer locking can be.

@@ -5,5 +5,5 @@ struct el {
        int data;
        /* Other data fields */
 };
-rwlock_t listmutex;
+spinlock_t listmutex;
 struct el head;

@@ -13,15 +14,15 @@
        struct list_head *lp;
        struct el *p;

-       read_lock(&listmutex);
-       list_for_each_entry(p, head, lp) {
+       rcu_read_lock();
+       list_for_each_entry_rcu(p, head, lp) {
                if (p->key == key) {
                        *result = p->data;
-                       read_unlock(&listmutex);
+                       rcu_read_unlock();
                        return 1;
                }
        }
-       read_unlock(&listmutex);
+       rcu_read_unlock();
        return 0;
 }

@@ -29,15 +30,16 @@
 {
        struct el *p;

-       write_lock(&listmutex);
+       spin_lock(&listmutex);
        list_for_each_entry(p, head, lp) {
                if (p->key == key) {
-                       list_del(&p->list);
-                       write_unlock(&listmutex);
+                       list_del_rcu(&p->list);
+                       spin_unlock(&listmutex);
+                       synchronize_rcu();
                        kfree(p);
                        return 1;
                }
        }
-       write_unlock(&listmutex);
+       spin_unlock(&listmutex);
        return 0;
 }

Or, for those who prefer a side-by-side listing:

1 struct el {                          1 struct el {
2   struct list_head list;             2   struct list_head list;
3   long key;                          3   long key;
4   spinlock_t mutex;                  4   spinlock_t mutex;
5   int data;                          5   int data;
6   /* Other data fields */            6   /* Other data fields */
7 };                                   7 };
8 rwlock_t listmutex;                  8 spinlock_t listmutex;
9 struct el head;                      9 struct el head;
 1 int search(long key, int *result)    1 int search(long key, int *result)
 2 {                                    2 {
 3   struct list_head *lp;              3   struct list_head *lp;
 4   struct el *p;                      4   struct el *p;
 5                                      5
 6   read_lock(&listmutex);             6   rcu_read_lock();
 7   list_for_each_entry(p, head, lp) { 7   list_for_each_entry_rcu(p, head, lp) {
 8     if (p->key == key) {             8     if (p->key == key) {
 9       *result = p->data;             9       *result = p->data;
10       read_unlock(&listmutex);      10       rcu_read_unlock();
11       return 1;                     11       return 1;
12     }                               12     }
13   }                                 13   }
14   read_unlock(&listmutex);          14   rcu_read_unlock();
15   return 0;                         15   return 0;
16 }                                   16 }
 1 int delete(long key)                 1 int delete(long key)
 2 {                                    2 {
 3   struct el *p;                      3   struct el *p;
 4                                      4
 5   write_lock(&listmutex);            5   spin_lock(&listmutex);
 6   list_for_each_entry(p, head, lp) { 6   list_for_each_entry(p, head, lp) {
 7     if (p->key == key) {             7     if (p->key == key) {
 8       list_del(&p->list);            8       list_del_rcu(&p->list);
 9       write_unlock(&listmutex);      9       spin_unlock(&listmutex);
                                       10       synchronize_rcu();
10       kfree(p);                     11       kfree(p);
11       return 1;                     12       return 1;
12     }                               13     }
13   }                                 14   }
14   write_unlock(&listmutex);         15   spin_unlock(&listmutex);
15   return 0;                         16   return 0;
16 }                                   17 }

Either way, the differences are quite small. Read-side locking moves to rcu_read_lock() and rcu_read_unlock, update-side locking moves from a reader-writer lock to a simple spinlock, and a synchronize_rcu() precedes the kfree().

However, there is one potential catch: the read-side and update-side critical sections can now run concurrently. In many cases, this will not be a problem, but it is necessary to check carefully regardless. For example, if multiple independent list updates must be seen as a single atomic update, converting to RCU will require special care.

Also, the presence of synchronize_rcu() means that the RCU version of delete() can now block. If this is a problem, there is a callback-based mechanism that never blocks, namely call_rcu() or kfree_rcu(), that can be used in place of synchronize_rcu().

7. ANALOGY WITH REFERENCE COUNTING

The reader-writer analogy (illustrated by the previous section) is not always the best way to think about using RCU. Another helpful analogy considers RCU an effective reference count on everything which is protected by RCU.

A reference count typically does not prevent the referenced object’s values from changing, but does prevent changes to type -- particularly the gross change of type that happens when that object’s memory is freed and re-allocated for some other purpose. Once a type-safe reference to the object is obtained, some other mechanism is needed to ensure consistent access to the data in the object. This could involve taking a spinlock, but with RCU the typical approach is to perform reads with SMP-aware operations such as smp_load_acquire(), to perform updates with atomic read-modify-write operations, and to provide the necessary ordering. RCU provides a number of support functions that embed the required operations and ordering, such as the list_for_each_entry_rcu() macro used in the previous section.

A more focused view of the reference counting behavior is that, between rcu_read_lock() and rcu_read_unlock(), any reference taken with rcu_dereference() on a pointer marked as __rcu can be treated as though a reference-count on that object has been temporarily increased. This prevents the object from changing type. Exactly what this means will depend on normal expectations of objects of that type, but it typically includes that spinlocks can still be safely locked, normal reference counters can be safely manipulated, and __rcu pointers can be safely dereferenced.

Some operations that one might expect to see on an object for which an RCU reference is held include:

  • Copying out data that is guaranteed to be stable by the object’s type.

  • Using kref_get_unless_zero() or similar to get a longer-term reference. This may fail of course.

  • Acquiring a spinlock in the object, and checking if the object still is the expected object and if so, manipulating it freely.

The understanding that RCU provides a reference that only prevents a change of type is particularly visible with objects allocated from a slab cache marked SLAB_TYPESAFE_BY_RCU. RCU operations may yield a reference to an object from such a cache that has been concurrently freed and the memory reallocated to a completely different object, though of the same type. In this case RCU doesn’t even protect the identity of the object from changing, only its type. So the object found may not be the one expected, but it will be one where it is safe to take a reference (and then potentially acquiring a spinlock), allowing subsequent code to check whether the identity matches expectations. It is tempting to simply acquire the spinlock without first taking the reference, but unfortunately any spinlock in a SLAB_TYPESAFE_BY_RCU object must be initialized after each and every call to kmem_cache_alloc(), which renders reference-free spinlock acquisition completely unsafe. Therefore, when using SLAB_TYPESAFE_BY_RCU, make proper use of a reference counter. (Those willing to initialize their locks in a kmem_cache constructor may also use locking, including cache-friendly sequence locking.)

With traditional reference counting -- such as that implemented by the kref library in Linux -- there is typically code that runs when the last reference to an object is dropped. With kref, this is the function passed to kref_put(). When RCU is being used, such finalization code must not be run until all __rcu pointers referencing the object have been updated, and then a grace period has passed. Every remaining globally visible pointer to the object must be considered to be a potential counted reference, and the finalization code is typically run using call_rcu() only after all those pointers have been changed.

To see how to choose between these two analogies -- of RCU as a reader-writer lock and RCU as a reference counting system -- it is useful to reflect on the scale of the thing being protected. The reader-writer lock analogy looks at larger multi-part objects such as a linked list and shows how RCU can facilitate concurrency while elements are added to, and removed from, the list. The reference-count analogy looks at the individual objects and looks at how they can be accessed safely within whatever whole they are a part of.

8. FULL LIST OF RCU APIs

The RCU APIs are documented in docbook-format header comments in the Linux-kernel source code, but it helps to have a full list of the APIs, since there does not appear to be a way to categorize them in docbook. Here is the list, by category.

RCU list traversal:

list_entry_rcu
list_entry_lockless
list_first_entry_rcu
list_next_rcu
list_for_each_entry_rcu
list_for_each_entry_continue_rcu
list_for_each_entry_from_rcu
list_first_or_null_rcu
list_next_or_null_rcu
hlist_first_rcu
hlist_next_rcu
hlist_pprev_rcu
hlist_for_each_entry_rcu
hlist_for_each_entry_rcu_bh
hlist_for_each_entry_from_rcu
hlist_for_each_entry_continue_rcu
hlist_for_each_entry_continue_rcu_bh
hlist_nulls_first_rcu
hlist_nulls_for_each_entry_rcu
hlist_bl_first_rcu
hlist_bl_for_each_entry_rcu

RCU pointer/list update:

rcu_assign_pointer
list_add_rcu
list_add_tail_rcu
list_del_rcu
list_replace_rcu
hlist_add_behind_rcu
hlist_add_before_rcu
hlist_add_head_rcu
hlist_add_tail_rcu
hlist_del_rcu
hlist_del_init_rcu
hlist_replace_rcu
list_splice_init_rcu
list_splice_tail_init_rcu
hlist_nulls_del_init_rcu
hlist_nulls_del_rcu
hlist_nulls_add_head_rcu
hlist_bl_add_head_rcu
hlist_bl_del_init_rcu
hlist_bl_del_rcu
hlist_bl_set_first_rcu

RCU:

Critical sections       Grace period            Barrier

rcu_read_lock           synchronize_net         rcu_barrier
rcu_read_unlock         synchronize_rcu
rcu_dereference         synchronize_rcu_expedited
rcu_read_lock_held      call_rcu
rcu_dereference_check   kfree_rcu
rcu_dereference_protected

bh:

Critical sections       Grace period            Barrier

rcu_read_lock_bh        call_rcu                rcu_barrier
rcu_read_unlock_bh      synchronize_rcu
[local_bh_disable]      synchronize_rcu_expedited
[and friends]
rcu_dereference_bh
rcu_dereference_bh_check
rcu_dereference_bh_protected
rcu_read_lock_bh_held

sched:

Critical sections       Grace period            Barrier

rcu_read_lock_sched     call_rcu                rcu_barrier
rcu_read_unlock_sched   synchronize_rcu
[preempt_disable]       synchronize_rcu_expedited
[and friends]
rcu_read_lock_sched_notrace
rcu_read_unlock_sched_notrace
rcu_dereference_sched
rcu_dereference_sched_check
rcu_dereference_sched_protected
rcu_read_lock_sched_held

RCU-Tasks:

Critical sections       Grace period            Barrier

N/A                     call_rcu_tasks          rcu_barrier_tasks
                        synchronize_rcu_tasks

RCU-Tasks-Rude:

Critical sections       Grace period            Barrier

N/A                     call_rcu_tasks_rude     rcu_barrier_tasks_rude
                        synchronize_rcu_tasks_rude

RCU-Tasks-Trace:

Critical sections       Grace period            Barrier

rcu_read_lock_trace     call_rcu_tasks_trace    rcu_barrier_tasks_trace
rcu_read_unlock_trace   synchronize_rcu_tasks_trace

SRCU:

Critical sections       Grace period            Barrier

srcu_read_lock          call_srcu               srcu_barrier
srcu_read_unlock        synchronize_srcu
srcu_dereference        synchronize_srcu_expedited
srcu_dereference_check
srcu_read_lock_held

SRCU: Initialization/cleanup:

DEFINE_SRCU
DEFINE_STATIC_SRCU
init_srcu_struct
cleanup_srcu_struct

All: lockdep-checked RCU utility APIs:

RCU_LOCKDEP_WARN
rcu_sleep_check

All: Unchecked RCU-protected pointer access:

rcu_dereference_raw

All: Unchecked RCU-protected pointer access with dereferencing prohibited:

rcu_access_pointer

See the comment headers in the source code (or the docbook generated from them) for more information.

However, given that there are no fewer than four families of RCU APIs in the Linux kernel, how do you choose which one to use? The following list can be helpful:

  1. Will readers need to block? If so, you need SRCU.

  2. Will readers need to block and are you doing tracing, for example, ftrace or BPF? If so, you need RCU-tasks, RCU-tasks-rude, and/or RCU-tasks-trace.

  3. What about the -rt patchset? If readers would need to block in an non-rt kernel, you need SRCU. If readers would block when acquiring spinlocks in a -rt kernel, but not in a non-rt kernel, SRCU is not necessary. (The -rt patchset turns spinlocks into sleeplocks, hence this distinction.)

  4. Do you need to treat NMI handlers, hardirq handlers, and code segments with preemption disabled (whether via preempt_disable(), local_irq_save(), local_bh_disable(), or some other mechanism) as if they were explicit RCU readers? If so, RCU-sched readers are the only choice that will work for you, but since about v4.20 you use can use the vanilla RCU update primitives.

  5. Do you need RCU grace periods to complete even in the face of softirq monopolization of one or more of the CPUs? For example, is your code subject to network-based denial-of-service attacks? If so, you should disable softirq across your readers, for example, by using rcu_read_lock_bh(). Since about v4.20 you use can use the vanilla RCU update primitives.

  6. Is your workload too update-intensive for normal use of RCU, but inappropriate for other synchronization mechanisms? If so, consider SLAB_TYPESAFE_BY_RCU (which was originally named SLAB_DESTROY_BY_RCU). But please be careful!

  7. Do you need read-side critical sections that are respected even on CPUs that are deep in the idle loop, during entry to or exit from user-mode execution, or on an offlined CPU? If so, SRCU and RCU Tasks Trace are the only choices that will work for you, with SRCU being strongly preferred in almost all cases.

  8. Otherwise, use RCU.

Of course, this all assumes that you have determined that RCU is in fact the right tool for your job.

9. ANSWERS TO QUICK QUIZZES

Quick Quiz #1:

Why is this argument naive? How could a deadlock occur when using this algorithm in a real-world Linux kernel? [Referring to the lock-based “toy” RCU algorithm.]

Answer:

Consider the following sequence of events:

  1. CPU 0 acquires some unrelated lock, call it “problematic_lock”, disabling irq via spin_lock_irqsave().

  2. CPU 1 enters synchronize_rcu(), write-acquiring rcu_gp_mutex.

  3. CPU 0 enters rcu_read_lock(), but must wait because CPU 1 holds rcu_gp_mutex.

  4. CPU 1 is interrupted, and the irq handler attempts to acquire problematic_lock.

The system is now deadlocked.

One way to avoid this deadlock is to use an approach like that of CONFIG_PREEMPT_RT, where all normal spinlocks become blocking locks, and all irq handlers execute in the context of special tasks. In this case, in step 4 above, the irq handler would block, allowing CPU 1 to release rcu_gp_mutex, avoiding the deadlock.

Even in the absence of deadlock, this RCU implementation allows latency to “bleed” from readers to other readers through synchronize_rcu(). To see this, consider task A in an RCU read-side critical section (thus read-holding rcu_gp_mutex), task B blocked attempting to write-acquire rcu_gp_mutex, and task C blocked in rcu_read_lock() attempting to read_acquire rcu_gp_mutex. Task A’s RCU read-side latency is holding up task C, albeit indirectly via task B.

Realtime RCU implementations therefore use a counter-based approach where tasks in RCU read-side critical sections cannot be blocked by tasks executing synchronize_rcu().

Back to Quick Quiz #1

Quick Quiz #2:

Give an example where Classic RCU’s read-side overhead is negative.

Answer:

Imagine a single-CPU system with a non-CONFIG_PREEMPTION kernel where a routing table is used by process-context code, but can be updated by irq-context code (for example, by an “ICMP REDIRECT” packet). The usual way of handling this would be to have the process-context code disable interrupts while searching the routing table. Use of RCU allows such interrupt-disabling to be dispensed with. Thus, without RCU, you pay the cost of disabling interrupts, and with RCU you don’t.

One can argue that the overhead of RCU in this case is negative with respect to the single-CPU interrupt-disabling approach. Others might argue that the overhead of RCU is merely zero, and that replacing the positive overhead of the interrupt-disabling scheme with the zero-overhead RCU scheme does not constitute negative overhead.

In real life, of course, things are more complex. But even the theoretical possibility of negative overhead for a synchronization primitive is a bit unexpected. ;-)

Back to Quick Quiz #2

Quick Quiz #3:

If it is illegal to block in an RCU read-side critical section, what the heck do you do in CONFIG_PREEMPT_RT, where normal spinlocks can block???

Answer:

Just as CONFIG_PREEMPT_RT permits preemption of spinlock critical sections, it permits preemption of RCU read-side critical sections. It also permits spinlocks blocking while in RCU read-side critical sections.

Why the apparent inconsistency? Because it is possible to use priority boosting to keep the RCU grace periods short if need be (for example, if running short of memory). In contrast, if blocking waiting for (say) network reception, there is no way to know what should be boosted. Especially given that the process we need to boost might well be a human being who just went out for a pizza or something. And although a computer-operated cattle prod might arouse serious interest, it might also provoke serious objections. Besides, how does the computer know what pizza parlor the human being went to???

Back to Quick Quiz #3

ACKNOWLEDGEMENTS

My thanks to the people who helped make this human-readable, including Jon Walpole, Josh Triplett, Serge Hallyn, Suzanne Wood, and Alan Stern.

For more information, see http://www.rdrop.com/users/paulmck/RCU.