Copyright IBM Corporation, 2015
Author: Paul E. McKenney
The initial version of this document appeared in the LWN articles here, here, and here.
Read-copy update (RCU) is a synchronization mechanism that is often used as a replacement for reader-writer locking. RCU is unusual in that updaters do not block readers, which means that RCU's read-side primitives can be exceedingly fast and scalable. In addition, updaters can make useful forward progress concurrently with readers. However, all this concurrency between RCU readers and updaters does raise the question of exactly what RCU readers are doing, which in turn raises the question of exactly what RCU's requirements are.
This document therefore summarizes RCU's requirements, and can be thought of as an informal, high-level specification for RCU. It is important to understand that RCU's specification is primarily empirical in nature; in fact, I learned about many of these requirements the hard way. This situation might cause some consternation, however, not only has this learning process been a lot of fun, but it has also been a great privilege to work with so many people willing to apply technologies in interesting new ways.
All that aside, here are the categories of currently known RCU requirements:
This is followed by a summary, which is in turn followed by the inevitable answers to the quick quizzes.
RCU's fundamental requirements are the closest thing RCU has to hard mathematical requirements. These are:
RCU's grace-period guarantee is unusual in being premeditated: Jack Slingwine and I had this guarantee firmly in mind when we started work on RCU (then called “rclock”) in the early 1990s. That said, the past two decades of experience with RCU have produced a much more detailed understanding of this guarantee.
RCU's grace-period guarantee allows updaters to wait for the completion of all pre-existing RCU read-side critical sections. An RCU read-side critical section begins with the marker rcu_read_lock() and ends with the marker rcu_read_unlock(). These markers may be nested, and RCU treats a nested set as one big RCU read-side critical section. Production-quality implementations of rcu_read_lock() and rcu_read_unlock() are extremely lightweight, and in fact have exactly zero overhead in Linux kernels built for production use with CONFIG_PREEMPT=n.
This guarantee allows ordering to be enforced with extremely low overhead to readers, for example:
1 int x, y;
2
3 void thread0(void)
4 {
5 rcu_read_lock();
6 r1 = READ_ONCE(x);
7 r2 = READ_ONCE(y);
8 rcu_read_unlock();
9 }
10
11 void thread1(void)
12 {
13 WRITE_ONCE(x, 1);
14 synchronize_rcu();
15 WRITE_ONCE(y, 1);
16 }
Because the synchronize_rcu() on line 14 waits for all pre-existing readers, any instance of thread0() that loads a value of zero from x must complete before thread1() stores to y, so that instance must also load a value of zero from y. Similarly, any instance of thread0() that loads a value of one from y must have started after the synchronize_rcu() started, and must therefore also load a value of one from x. Therefore, the outcome:
cannot happen.(r1 == 0 && r2 == 1)
@@QQ@@ Wait a minute! You said that updaters can make useful forward progress concurrently with readers, but pre-existing readers will block synchronize_rcu()!!! Just who are you trying to fool???
@@QQA@@ First, if updaters do not wish to be blocked by readers, they can use call_rcu() or kfree_rcu(), which will be discussed later. Second, even when using synchronize_rcu(), the other update-side code does run concurrently with readers, whether pre-existing or not.
@@QQE@@
This scenario resembles one of the first uses of RCU in DYNIX/ptx, which managed a distributed lock manager's transition into a state suitable for handling recovery from node failure, more or less as follows:
1 #define STATE_NORMAL 0
2 #define STATE_WANT_RECOVERY 1
3 #define STATE_RECOVERING 2
4 #define STATE_WANT_NORMAL 3
5
6 int state = STATE_NORMAL;
7
8 void do_something_dlm(void)
9 {
10 int state_snap;
11
12 rcu_read_lock();
13 state_snap = READ_ONCE(state);
14 if (state_snap == STATE_NORMAL)
15 do_something();
16 else
17 do_something_carefully();
18 rcu_read_unlock();
19 }
20
21 void start_recovery(void)
22 {
23 WRITE_ONCE(state, STATE_WANT_RECOVERY);
24 synchronize_rcu();
25 WRITE_ONCE(state, STATE_RECOVERING);
26 recovery();
27 WRITE_ONCE(state, STATE_WANT_NORMAL);
28 synchronize_rcu();
29 WRITE_ONCE(state, STATE_NORMAL);
30 }
The RCU read-side critical section in do_something_dlm() works with the synchronize_rcu() in start_recovery() to guarantee that do_something() never runs concurrently with recovery(), but with little or no synchronization overhead in do_something_dlm().
@@QQ@@ Why is the synchronize_rcu() on line 28 needed?
@@QQA@@ Without that extra grace period, memory reordering could result in do_something_dlm() executing do_something() concurrently with the last bits of recovery().
@@QQE@@
In order to avoid fatal problems such as deadlocks, an RCU read-side critical section must not contain calls to synchronize_rcu(). Similarly, an RCU read-side critical section must not contain anything that waits, directly or indirectly, on completion of an invocation of synchronize_rcu().
Although RCU's grace-period guarantee is useful in and of itself, with quite a few use cases, it would be good to be able to use RCU to coordinate read-side access to linked data structures. For this, the grace-period guarantee is not sufficient, as can be seen in function add_gp_buggy() below. We will look at the reader's code later, but in the meantime, just think of the reader as locklessly picking up the gp pointer, and, if the value loaded is non-NULL, locklessly accessing the ->a and ->b fields.
1 bool add_gp_buggy(int a, int b)
2 {
3 p = kmalloc(sizeof(*p), GFP_KERNEL);
4 if (!p)
5 return -ENOMEM;
6 spin_lock(&gp_lock);
7 if (rcu_access_pointer(gp)) {
8 spin_unlock(&gp_lock);
9 return false;
10 }
11 p->a = a;
12 p->b = a;
13 gp = p; /* ORDERING BUG */
14 spin_unlock(&gp_lock);
15 return true;
16 }
The problem is that both the compiler and weakly ordered CPUs are within their rights to reorder this code as follows:
1 bool add_gp_buggy_optimized(int a, int b)
2 {
3 p = kmalloc(sizeof(*p), GFP_KERNEL);
4 if (!p)
5 return -ENOMEM;
6 spin_lock(&gp_lock);
7 if (rcu_access_pointer(gp)) {
8 spin_unlock(&gp_lock);
9 return false;
10 }
11 gp = p; /* ORDERING BUG */
12 p->a = a;
13 p->b = a;
14 spin_unlock(&gp_lock);
15 return true;
16 }
If an RCU reader fetches gp just after add_gp_buggy_optimized executes line 11, it will see garbage in the ->a and ->b fields. And this is but one of many ways in which compiler and hardware optimizations could cause trouble. Therefore, we clearly need some way to prevent the compiler and the CPU from reordering in this manner, which brings us to the publish-subscribe guarantee discussed in the next section.
RCU's publish-subscribe guarantee allows data to be inserted into a linked data structure without disrupting RCU readers. The updater uses rcu_assign_pointer() to insert the new data, and readers use rcu_dereference() to access data, whether new or old. The following shows an example of insertion:
1 bool add_gp(int a, int b)
2 {
3 p = kmalloc(sizeof(*p), GFP_KERNEL);
4 if (!p)
5 return -ENOMEM;
6 spin_lock(&gp_lock);
7 if (rcu_access_pointer(gp)) {
8 spin_unlock(&gp_lock);
9 return false;
10 }
11 p->a = a;
12 p->b = a;
13 rcu_assign_pointer(gp, p);
14 spin_unlock(&gp_lock);
15 return true;
16 }
The rcu_assign_pointer() on line 13 is conceptually equivalent to a simple assignment statement, but also guarantees that its assignment will happen after the two assignments in lines 11 and 12, similar to the C11 memory_order_release store operation. It also prevents any number of “interesting” compiler optimizations, for example, the use of gp as a scratch location immediately preceding the assignment.
@@QQ@@ But rcu_assign_pointer() does nothing to prevent the two assignments to p->a and p->b from being reordered. Can't that also cause problems?
@@QQA@@ No, it cannot. The readers cannot see either of these two fields until the assignment to gp, by which time both fields are fully initialized. So reordering the assignments to p->a and p->b cannot possibly cause any problems.
@@QQE@@
It is tempting to assume that the reader need not do anything special to control its accesses to the RCU-protected data, as shown in do_something_gp_buggy() below:
1 bool do_something_gp_buggy(void)
2 {
3 rcu_read_lock();
4 p = gp; /* OPTIMIZATIONS GALORE!!! */
5 if (p) {
6 do_something(p->a, p->b);
7 rcu_read_unlock();
8 return true;
9 }
10 rcu_read_unlock();
11 return false;
12 }
However, this temptation must be resisted because there are a surprisingly large number of ways that the compiler (to say nothing of DEC Alpha CPUs) can trip this code up. For but one example, if the compiler were short of registers, it might choose to refetch from gp rather than keeping a separate copy in p as follows:
1 bool do_something_gp_buggy_optimized(void)
2 {
3 rcu_read_lock();
4 if (gp) { /* OPTIMIZATIONS GALORE!!! */
5 do_something(gp->a, gp->b);
6 rcu_read_unlock();
7 return true;
8 }
9 rcu_read_unlock();
10 return false;
11 }
If this function ran concurrently with a series of updates that replaced the current structure with a new one, the fetches of gp->a and gp->b might well come from two different structures, which could cause serious confusion. To prevent this (and much else besides), do_something_gp() uses rcu_dereference() to fetch from gp:
1 bool do_something_gp(void)
2 {
3 rcu_read_lock();
4 p = rcu_dereference(gp);
5 if (p) {
6 do_something(p->a, p->b);
7 rcu_read_unlock();
8 return true;
9 }
10 rcu_read_unlock();
11 return false;
12 }
The rcu_dereference() uses volatile casts and (for DEC Alpha) memory barriers in the Linux kernel. Should a high-quality implementation of C11 memory_order_consume [PDF] ever appear, then rcu_dereference() could be implemented as a memory_order_consume load. Regardless of the exact implementation, a pointer fetched by rcu_dereference() may not be used outside of the outermost RCU read-side critical section containing that rcu_dereference(), unless protection of the corresponding data element has been passed from RCU to some other synchronization mechanism, most commonly locking or reference counting.
In short, updaters use rcu_assign_pointer() and readers use rcu_dereference(), and these two RCU API elements work together to ensure that readers have a consistent view of newly added data elements.
Of course, it is also necessary to remove elements from RCU-protected data structures, for example, using the following process:
1 bool remove_gp_synchronous(void)
2 {
3 struct foo *p;
4
5 spin_lock(&gp_lock);
6 p = rcu_access_pointer(gp);
7 if (!p) {
8 spin_unlock(&gp_lock);
9 return false;
10 }
11 rcu_assign_pointer(gp, NULL);
12 spin_unlock(&gp_lock);
13 synchronize_rcu();
14 kfree(p);
15 return true;
16 }
This function is straightforward, with line 13 waiting for a grace period before line 14 frees the old data element. This waiting ensures that readers will reach line 7 of do_something_gp() before the data element referenced by p is freed. The rcu_access_pointer() on line 6 is similar to rcu_dereference(), except that:
@@QQ@@ Without the rcu_dereference() or the rcu_access_pointer(), what destructive optimizations might the compiler make use of?
@@QQA@@ Let's start with what happens to do_something_gp() if it fails to use rcu_dereference(). It could reuse a value formerly fetched from this same pointer. It could also fetch the pointer from gp in a byte-at-a-time manner, resulting in load tearing, in turn resulting a bytewise mash-up of two distince pointer values. It might even use value-speculation optimizations, where it makes a wrong guess, but by the time it gets around to checking the value, an update has changed the pointer to match the wrong guess. Too bad about any dereferences that returned pre-initialization garbage in the meantime!
For remove_gp_synchronous(), as long as all modifications to gp are carried out while holding gp_lock, the above optimizations are harmless. However, with CONFIG_SPARSE_RCU_POINTER=y, sparse will complain if you define gp with __rcu and then access it without using either rcu_access_pointer() or rcu_dereference().
@@QQE@@
In short, RCU's publish-subscribe guarantee is provided by the combination of rcu_assign_pointer() and rcu_dereference(). This guarantee allows data elements to be safely added to RCU-protected linked data structures without disrupting RCU readers. This guarantee can be used in combination with the grace-period guarantee to also allow data elements to be removed from RCU-protected linked data structures, again without disrupting RCU readers.
This guarantee was only partially premeditated. DYNIX/ptx used an explicit memory barrier for publication, but had nothing resembling rcu_dereference() for subscription, nor did it have anything resembling the smp_read_barrier_depends() that was later subsumed into rcu_dereference(). The need for these operations made itself known quite suddenly at a late-1990s meeting with the DEC Alpha architects, back in the days when DEC was still a free-standing company. It took the Alpha architects a good hour to convince me that any sort of barrier would ever be needed, and it then took me a good two hours to convince them that their documentation did not make this point clear. More recent work with the C and C++ standards committees have provided much education on tricks and traps from the compiler. In short, compilers were much less tricky in the early 1990s, but in 2015, don't even think about omitting rcu_dereference()!
The previous section's simple linked-data-structure scenario clearly demonstrates the need for RCU's stringent memory-ordering guarantees on systems with more than one CPU:
@@QQ@@ Given that multiple CPUs can start RCU read-side critical sections at any time without any ordering whatsoever, how can RCU possibly tell whether or not a given RCU read-side critical section starts before a given instance of synchronize_rcu()?
@@QQA@@ If RCU cannot tell whether or not a given RCU read-side critical section starts before a given instance of synchronize_rcu(), then it must assume that the RCU read-side critical section started first. In other words, a given instance of synchronize_rcu() can avoid waiting on a given RCU read-side critical section only if it can prove that synchronize_rcu() started first.
@@QQE@@
@@QQ@@ The first and second guarantees require unbelievably strict ordering! Are all these memory barriers really required?
@@QQA@@ Yes, they really are required. To see why the first guarantee is required, consider the following sequence of events:
Therefore, there absolutely must be a full memory barrier between the end of the RCU read-side critical section and the end of the grace period.
The sequence of events demonstrating the necessity of the second rule is roughly similar:
And similarly, without a memory barrier between the beginning of the grace period and the beginning of the RCU read-side critical section, CPU 1 might end up accessing the freelist.
The “as if” rule of course applies, so that any implementation that acts as if the appropriate memory barriers were in place is a correct implementation. That said, it is much easier to fool yourself into believing that you have adhered to the as-if rule than it is to actually adhere to it!
@@QQE@@
Note that these memory-barrier requirements do not replace the fundamental RCU requirement that a grace period wait for all pre-existing readers. On the contrary, the memory barriers called out in this section must operate in such a way as to enforce this fundamental requirement. Of course, different implementations enforce this requirement in different ways, but enforce it they must.
The common-case RCU primitives are unconditional. They are invoked, they do their job, and they return, with no possibility of error, and no need to retry. This is a key RCU design philosophy.
However, this philosophy is pragmatic rather than pigheaded. If someone comes up with a good justification for a particular conditional RCU primitive, it might well be implemented and added. After all, this guarantee was reverse-engineered, not premeditated. The unconditional nature of the RCU primitives was initially an accident of implementation, and later experience with synchronization primitives with conditional primitives caused me to elevate this accident to a guarantee. Therefore, the justification for adding a conditional primitive to RCU would need to be based on detailed and compelling use cases.
As far as RCU is concerned, it is always possible to carry out an update within an RCU read-side critical section. For example, that RCU read-side critical section might search for a given data element, and then might acquire the update-side spinlock in order to update that element, all while remaining in that RCU read-side critical section. Of course, it is necessary to exit the RCU read-side critical section before invoking synchronize_rcu(), however, this inconvenience can be avoided through use of the call_rcu() and kfree_rcu() API members described later in this document.
@@QQ@@ But how does the upgrade-to-write operation exclude other readers?
@@QQA@@ It doesn't, just like normal RCU updates, which also do not exclude RCU readers.
@@QQE@@
This guarantee allows lookup code to be shared between read-side and update-side code, and was premeditated, appearing in the earliest DYNIX/ptx RCU documentation.
RCU provides extremely lightweight readers, and its read-side guarantees, though quite useful, are correspondingly lightweight. It is therefore all too easy to assume that RCU is guaranteeing more than it really is. Of course, the list of things that RCU does not guarantee is infinitely long, however, the following sections list a few non-guarantees that have caused confusion. Except where otherwise noted, these non-guarantees were premeditated.
Reader-side markers such as rcu_read_lock() and rcu_read_unlock() provide absolutely no ordering guarantees except through their interaction with the grace-period APIs such as synchronize_rcu(). To see this, consider the following pair of threads:
1 void thread0(void)
2 {
3 rcu_read_lock();
4 WRITE_ONCE(x, 1);
5 rcu_read_unlock();
6 rcu_read_lock();
7 WRITE_ONCE(y, 1);
8 rcu_read_unlock();
9 }
10
11 void thread1(void)
12 {
13 rcu_read_lock();
14 r1 = READ_ONCE(y);
15 rcu_read_unlock();
16 rcu_read_lock();
17 r2 = READ_ONCE(x);
18 rcu_read_unlock();
19 }
After thread0() and thread1() execute concurrently, it is quite possible to have
(that is, y appears to have been assigned before x), which would not be possible if rcu_read_lock() and rcu_read_unlock() had much in the way of ordering properties. But they do not, so the CPU is within its rights to do significant reordering. This is by design: Any significant ordering constraints would slow down these fast-path APIs.(r1 == 1 && r2 == 0)
@@QQ@@ Can't the compiler also reorder this code?
@@QQA@@ No, the volatile casts in READ_ONCE() and WRITE_ONCE() prevent the compiler from reordering in this particular case.
@@QQE@@
Neither rcu_read_lock() nor rcu_read_unlock() exclude updates. All they do is to prevent grace periods from ending. The following example illustrates this:
1 void thread0(void)
2 {
3 rcu_read_lock();
4 r1 = READ_ONCE(y);
5 if (r1) {
6 do_something_with_nonzero_x();
7 r2 = READ_ONCE(x);
8 WARN_ON(!r2); /* BUG!!! */
9 }
10 rcu_read_unlock();
11 }
12
13 void thread1(void)
14 {
15 spin_lock(&my_lock);
16 WRITE_ONCE(x, 1);
17 WRITE_ONCE(y, 1);
18 spin_unlock(&my_lock);
19 }
If the thread0() function's rcu_read_lock() excluded the thread1() function's update, the WARN_ON() could never fire. But the fact is that rcu_read_lock() does not exclude much of anything aside from subsequent grace periods, of which thread1() has none, so the WARN_ON() can and does fire.
It might be tempting to assume that after synchronize_rcu() completes, there are no readers executing. This temptation must be avoided because new readers can start immediately after synchronize_rcu() starts, and synchronize_rcu() is under no obligation to wait for these new readers.
@@QQ@@ Suppose that synchronize_rcu() did wait until all readers had completed. Would the updater be able to rely on this?
@@QQA@@ No. Even if synchronize_rcu() were to wait until all readers had completed, a new reader might start immediately after synchronize_rcu() completed. Therefore, the code following synchronize_rcu() cannot rely on there being no readers in any case.
@@QQE@@
It is tempting to assume that if any part of one RCU read-side critical section precedes a given grace period, and if any part of another RCU read-side critical section follows that same grace period, then all of the first RCU read-side critical section must precede all of the second. However, this just isn't the case: A single grace period does not partition the set of RCU read-side critical sections. An example of this situation can be illustrated as follows, where x, y, and z are initially all zero:
1 void thread0(void)
2 {
3 rcu_read_lock();
4 WRITE_ONCE(a, 1);
5 WRITE_ONCE(b, 1);
6 rcu_read_unlock();
7 }
8
9 void thread1(void)
10 {
11 r1 = READ_ONCE(a);
12 synchronize_rcu();
13 WRITE_ONCE(c, 1);
14 }
15
16 void thread2(void)
17 {
18 rcu_read_lock();
19 r2 = READ_ONCE(b);
20 r3 = READ_ONCE(c);
21 rcu_read_unlock();
22 }
It turns out that the outcome:
is entirely possible. The following figure show how this can happen, with each circled QS indicating the point at which RCU recorded a quiescent state for each thread, that is, a state in which RCU knows that the thread cannot be in the midst of an RCU read-side critical section that started before the current grace period:(r1 == 1 && r2 == 0 && r3 == 1)
If it is necessary to partition RCU read-side critical sections in this manner, it is necessary to use two grace periods, where the first grace period is known to end before the second grace period starts:
1 void thread0(void)
2 {
3 rcu_read_lock();
4 WRITE_ONCE(a, 1);
5 WRITE_ONCE(b, 1);
6 rcu_read_unlock();
7 }
8
9 void thread1(void)
10 {
11 r1 = READ_ONCE(a);
12 synchronize_rcu();
13 WRITE_ONCE(c, 1);
14 }
15
16 void thread2(void)
17 {
18 r2 = READ_ONCE(c);
19 synchronize_rcu();
20 WRITE_ONCE(d, 1);
21 }
22
23 void thread3(void)
24 {
25 rcu_read_lock();
26 r3 = READ_ONCE(b);
27 r4 = READ_ONCE(d);
28 rcu_read_unlock();
29 }
Here, if (r1 == 1), then thread0()'s write to b must happen before the end of thread1()'s grace period. If in addition (r4 == 1), then thread3()'s read from b must happen after the beginning of thread2()'s grace period. If it is also the case that (r2 == 1), then the end of thread1()'s grace period must precede the beginning of thread2()'s grace period. This mean that the two RCU read-side critical sections cannot overlap, guaranteeing that (r3 == 1). As a result, the outcome:
cannot happen.(r1 == 1 && r2 == 1 && r3 == 0 && r4 == 1)
This non-requirement was also non-premeditated, but became apparent when studying RCU's interaction with memory ordering.
It is also tempting to assume that if an RCU read-side critical section happens between a pair of grace periods, then those grace periods cannot overlap. However, this temptation leads nowhere good, as can be illustrated by the following, with all variables initially zero:
1 void thread0(void)
2 {
3 rcu_read_lock();
4 WRITE_ONCE(a, 1);
5 WRITE_ONCE(b, 1);
6 rcu_read_unlock();
7 }
8
9 void thread1(void)
10 {
11 r1 = READ_ONCE(a);
12 synchronize_rcu();
13 WRITE_ONCE(c, 1);
14 }
15
16 void thread2(void)
17 {
18 rcu_read_lock();
19 WRITE_ONCE(d, 1);
20 r2 = READ_ONCE(c);
21 rcu_read_unlock();
22 }
23
24 void thread3(void)
25 {
26 r3 = READ_ONCE(d);
27 synchronize_rcu();
28 WRITE_ONCE(e, 1);
29 }
30
31 void thread4(void)
32 {
33 rcu_read_lock();
34 r4 = READ_ONCE(b);
35 r5 = READ_ONCE(e);
36 rcu_read_unlock();
37 }
In this case, the outcome:
is entirely possible, as illustrated below:(r1 == 1 && r2 == 1 && r3 == 1 && r4 == 0 && r5 == 1)
Again, an RCU read-side critical section can overlap almost all of a given grace period, just so long as it does not overlap the entire grace period. As a result, an RCU read-side critical section cannot partition a pair of RCU grace periods.
@@QQ@@ How long a sequence of grace periods, each separated by an RCU read-side critical section, would be required to partition the RCU read-side critical sections at the beginning and end of the chain?
@@QQA@@ In theory, an infinite number. In practice, an unknown number that is sensitive to both implementation details and timing considerations. Therefore, even in practice, RCU users must abide by the theoretical rather than the practical answer.
@@QQE@@
There was a time when disabling preemption on any given CPU would block subsequent grace periods. However, this was an accident of implementation and is not a requirement. And in the current Linux-kernel implementation, disabling preemption on a given CPU in fact does not block grace periods, as Oleg Nesterov demonstrated.
If you need a preempt-disable region to block grace periods, you need to add rcu_read_lock() and rcu_read_unlock(), for example as follows:
1 preempt_disable(); 2 rcu_read_lock(); 3 do_something(); 4 rcu_read_unlock(); 5 preempt_enable(); 6 7 /* Spinlocks implicitly disable preemption. */ 8 spin_lock(&mylock); 9 rcu_read_lock(); 10 do_something(); 11 rcu_read_unlock(); 12 spin_unlock(&mylock);
In theory, you could enter the RCU read-side critical section first, but it is more efficient to keep the entire RCU read-side critical section contained in the preempt-disable region as shown above. Of course, RCU read-side critical sections that extend outside of preempt-disable regions will work correctly, but such critical sections can be preempted, which forces rcu_read_unlock() to do more work. And no, this is not an invitation to enclose all of your RCU read-side critical sections within preempt-disable regions, because doing so would degrade real-time response.
This non-requirement appeared with preemptible RCU. If you need a grace period that waits on non-preemptible code regions, use RCU-sched.
These parallelism facts of life are by no means specific to RCU, but the RCU implementation must abide by them. They therefore bear repeating:
This last parallelism fact of life means that RCU must pay special attention to the preceding facts of life. The idea that Linux might scale to systems with thousands of CPUs would have been met with some skepticism in the 1990s, but these requirements would have otherwise have been unsurprising, even in the early 1990s.
These sections list quality-of-implementation requirements. Although an RCU implementation that ignores these requirements could still be used, it would likely be subject to limitations that would make it inappropriate for industrial-strength production use. Classes of quality-of-implementation requirements are as follows:
These classes is covered in the following sections.
RCU is and always has been intended primarily for read-mostly situations, as illustrated by the following figure. This means that RCU's read-side primitives are optimized, often at the expense of its update-side primitives.
This focus on read-mostly situations means that RCU must interoperate with other synchronization primitives. For example, the add_gp() and remove_gp_synchronous() examples discussed earlier use RCU to protect readers and locking to coordinate updaters. However, the need extends much farther, requiring that a variety of synchronization primitives be legal within RCU read-side critical sections, including spinlocks, sequence locks, atomic operations, reference counters, and memory barriers.
@@QQ@@ What about sleeping locks?
@@QQA@@ These are forbidden within Linux-kernel RCU read-side critical sections because it is not legal to place a quiescent state (in this case, voluntary context switch) within an RCU read-side critical section. However, sleeping locks may be used within userspace RCU read-side critical sections, and also within Linux-kernel sleepable RCU (SRCU) read-side critical sections. In addition, the -rt patchset turns spinlocks into a sleeping locks so that the corresponding critical sections can be preempted, which also means that these sleeplockified spinlocks (but not other sleeping locks!) may be acquire within -rt-Linux-kernel RCU read-side critical sections.
Note that it is legal for a normal RCU read-side critical section to conditionally acquire a sleeping locks (as in mutex_trylock()), but only as long as it does not loop indefinitely attempting to conditionally acquire that sleeping locks. The key point is that things like mutex_trylock() either return with the mutex held, or return an error indication if the mutex was not immediately available. Either way, mutex_trylock() returns immediately without sleeping.
@@QQE@@
It often comes as a surprise that many algorithms do not require a consistent view of data, but many can function in that mode, with network routing being the poster child. Internet routing algorithms take significant time to propagate updates, so that by the time an update arrives at a given system, that system has been sending network traffic the wrong way for a considerable length of time. Having a few threads continue to send traffic the wrong way for a few more milliseconds is clearly not a problem: In the worst case, TCP retransmissions will eventually get the data where it needs to go. In general, when tracking the state of the universe outside of the computer, some level of inconsistency must be tolerated due to speed-of-light delays if nothing else.
Furthermore, uncertainty about external state is inherent in many cases. For example, a pair of veternarians might use heartbeat to determine whether or not a given cat was alive. But how long should they wait after the last heartbeat to decide that the cat is in fact dead? Waiting less than 400 milliseconds makes no sense because this would mean that a relaxed cat would be considered to cycle between death and life more than 100 times per minute. Moreover, just as with human beings, a cat's heart might stop for some period of time, so the exact wait period is a judgment call. One of our pair of veternarians might wait 30 seconds before pronouncing the cat dead, while the other might insist on waiting a full minute. The two veternarians would then disagree on the state of the cat during the final 30 seconds of the minute following the last heartbeat, as fancifully illustrated below:
Interestingly enough, this same situation applies to hardware. When push comes to shove, how do we tell whether or not some external server has failed? We send messages to it periodically, and declare it failed if we don't receive a response within a given period of time. Policy decisions can usually tolerate short periods of inconsistency. The policy was decided some time ago, and is only now being put into effect, so a few milliseconds of delay is normally inconsequential.
However, there are algorithms that absolutely must see consistent data. For example, the translation between a user-level SystemV semaphore ID to the corresponding in-kernel data structure is protected by RCU, but it is absolutely forbidden to update a semaphore that has just been removed. In the Linux kernel, this need for consistency is accommodated by acquiring spinlocks located in the in-kernel data structure from within the RCU read-side critical section, and this is indicated by the green box in the figure above. Many other techniques may be used, and are in fact used within the Linux kernel.
In short, RCU is not required to maintain consistency, and other mechanisms may be used in concert with RCU when consistency is required. RCU's specialization allows it to do its job extremely well, and its ability to interoperate with other synchronization mechanisms allows the right mix of synchronization tools to be used for a given job.
Energy efficiency is a critical component of performance today, and Linux-kernel RCU implementations must therefore avoid unnecessarily awakening idle CPUs. I cannot claim that this requirement was premeditated. In fact, I learned of it during a telephone conversation in which I was given “frank and open” feedback on the importance of energy efficiency in battery-powered systems and on specific energy-efficiency shortcomings of the Linux-kernel RCU implementation. In my experience, the battery-powered embedded community will consider any unnecessary wakeups to be extremely unfriendly acts. So much so that mere Linux-kernel-mailing-list posts are insufficient to vent their ire.
Memory consumption is not particularly important for in most situations, and has become decreasingly so as memory sizes have expanded and memory costs have plummeted. However, as I learned from Matt Mackall's bloatwatch efforts, memory footprint is critically important on single-CPU systems with non-preemptible (CONFIG_PREEMPT=n) kernels, and thus tiny RCU was born. Josh Triplett has since taken over the small-memory banner with his Linux kernel tinification project, which resulted in SRCU becoming optional for those kernels not needing it.
The remaining performance requirements are, for the most part, unsurprising. For example, in keeping with RCU's read-side specialization, rcu_dereference() should have negligible overhead (for example, suppression of a few minor compiler optimizations). Similarly, in non-preemptible environments, rcu_read_lock() and rcu_read_unlock() should have exactly zero overhead.
In preemptible environments, in the case where the RCU read-side critical section was not preempted (as will be the case for the highest-priority real-time process), rcu_read_lock() and rcu_read_unlock() should have minimal overhead. In particular, they should not contain atomic read-modify-write operations, memory-barrier instructions, preemption disabling, interrupt disabling, or backwards branches. However, in the case where the RCU read-side critical section was preempted, rcu_read_unlock() may acquire spinlocks and disable interrupts. This is why it is better to nest an RCU read-side critical section within a preempt-disable region than vice versa, at least in cases where that critical section is short enough to avoid unduly degrading real-time latencies.
The synchronize_rcu() grace-period-wait primitive is optimized for throughput. It may therefore incur several milliseconds of latency in addition to the duration of the longest RCU read-side critical section. On the other hand, multiple concurrent invocations of synchronize_rcu() are required to use batching optimizations so that they can be satisfied by a single underlying grace-period-wait operation. For example, in the Linux kernel, it is not unusual for a single grace-period-wait operation to serve more than 1,000 separate invocations of synchronize_rcu(), thus amortizing the per-invocation overhead down to nearly zero. However, the grace-period optimization is also required to avoid measurable degradation of real-time scheduling and interrupt latencies.
In some cases, the multi-millisecond synchronize_rcu() latencies are unacceptable. In these cases, synchronize_rcu_expedited() may be used instead, reducing the grace-period latency down to a few tens of microseconds on small systems, at least in cases where the RCU read-side critical sections are short. There are currently no special latency requirements for synchronize_rcu_expedited() on large systems, but, consistent with the empirical nature of the RCU specification, that is subject to change. However, there most definitely are scalability requirements: A storm of synchronize_rcu_expedited() invocations on 4096 CPUs should at least make reasonable forward progress. In return for its shorter latencies, synchronize_rcu_expedited() is permitted to impose modest degradation of real-time latency on non-idle online CPUs. That said, it will likely be necessary to take further steps to reduce this degradation, hopefully to roughly that of a scheduling-clock interrupt.
There are a number of situations where even synchronize_rcu_expedited()'s reduced grace-period latency is unacceptable. In these situations, the asynchronous call_rcu() can be used in place of synchronize_rcu() as follows:
1 struct foo {
2 int a;
3 int b;
4 struct rcu_head rh;
5 };
6
7 static void remove_gp_cb(struct rcu_head *rhp)
8 {
9 struct foo *p = container_of(rhp, struct foo, rh);
10
11 kfree(p);
12 }
13
14 bool remove_gp_asynchronous(void)
15 {
16 struct foo *p;
17
18 spin_lock(&gp_lock);
19 p = rcu_dereference(gp);
20 if (!p) {
21 spin_unlock(&gp_lock);
22 return false;
23 }
24 rcu_assign_pointer(gp, NULL);
25 call_rcu(&p->rh, remove_gp_cb);
26 spin_unlock(&gp_lock);
27 return true;
28 }
A definition of struct foo is finally needed, and appears on lines 1-5. The function remove_gp_cb() is passed to call_rcu() on line 25, and will be invoked after the end of a subsequent grace period. This gets the same effect as remove_gp_synchronous(), but without forcing the updater to wait for a grace period to elapse. The call_rcu() function may be used in a number of situations where neither synchronize_rcu() nor synchronize_rcu_expedited() would be legal, including within preempt-disable code, local_bh_disable() code, interrupt-disable code, and interrupt handlers. However, even call_rcu() is illegal within NMI handlers. The callback function (remove_gp_cb() in this case) will be executed within softirq (software interrupt) environment within the Linux kernel, either within a real softirq handler or under the protection of local_bh_disable(). In both the Linux kernel and in userspace, it is bad practice to write an RCU callback function that takes too long. Long-running operations should be relegated to separate threads or (in the Linux kernel) workqueues.
@@QQ@@ Why does line 19 use rcu_access_pointer()? After all, call_rcu() on line 25 stores into the structure, which would interact badly with concurrent insertions. Doesn't this mean that rcu_dereference() is required?
@@QQA@@ Presumably the ->gp_lock acquired on line 18 excludes any changes, including any insertions that rcu_dereference() would protect against. Therefore, any insertions will be delayed until after ->gp_lock is released on line 25, which in turn means that rcu_access_pointer() suffices.
@@QQE@@
However, all that remove_gp_cb() is doing is invoking kfree() on the data element. This is a common idiom, and is supported by kfree_rcu(), which allows “fire and forget” operation as shown below:
1 struct foo {
2 int a;
3 int b;
4 struct rcu_head rh;
5 };
6
7 bool remove_gp_faf(void)
8 {
9 struct foo *p;
10
11 spin_lock(&gp_lock);
12 p = rcu_dereference(gp);
13 if (!p) {
14 spin_unlock(&gp_lock);
15 return false;
16 }
17 rcu_assign_pointer(gp, NULL);
18 kfree_rcu(p, rh);
19 spin_unlock(&gp_lock);
20 return true;
21 }
Note that remove_gp_faf() simply invokes kfree_rcu() and proceeds, without any need to pay any further attention to the subsequent grace period and kfree(). It is permissible to invoke kfree_rcu() from the same environments as for call_rcu(). Interestingly enough, DYNIX/ptx had the equivalents of call_rcu() and kfree_rcu(), but not synchronize_rcu(). This was due to the fact that RCU was not heavily used within DYNIX/ptx, so the very few places that needed something like synchronize_rcu() simply open-coded it.
@@QQ@@ Earlier it was claimed that call_rcu() and kfree_rcu() allowed updaters to avoid being blocked by readers. But how can that be correct, given that the invocation of the callback and the freeing of the memory (respectively) must still wait for a grace period to elapse?
@@QQA@@ We could define things this way, but keep in mind that this sort of definition would say that updates in garbage-collected languages cannot complete until the next time the garbage collector runs, which does not seem at all reasonable. The key point is that in most cases, an updater using either call_rcu() or kfree_rcu() can proceed to the next update as soon as it has invoked call_rcu() or kfree_rcu(), without having to wait for a subsequent grace period.
@@QQE@@
But what if the updater must wait for the completion of code to be executed after the end of the grace period, but has other tasks that can be carried out in the meantime? The polling-style get_state_synchronize_rcu() and cond_synchronize_rcu() functions may be used for this purpose, as shown below:
1 bool remove_gp_poll(void)
2 {
3 struct foo *p;
4 unsigned long s;
5
6 spin_lock(&gp_lock);
7 p = rcu_access_pointer(gp);
8 if (!p) {
9 spin_unlock(&gp_lock);
10 return false;
11 }
12 rcu_assign_pointer(gp, NULL);
13 spin_unlock(&gp_lock);
14 s = get_state_synchronize_rcu();
15 do_something_while_waiting();
16 cond_synchronize_rcu(s);
17 kfree(p);
18 return true;
19 }
On line 14, get_state_synchronize_rcu() obtains a “cookie” from RCU, then line 15 carries out other tasks, and finally, line 16 returns immediately if a grace period has elapsed in the meantime, but otherwise waits as required. The need for get_state_synchronize_rcu and cond_synchronize_rcu() has appeared quite recently, so it is too early to tell whether they will stand the test of time.
RCU thus provides a range of tools to allow updaters to strike the required tradeoff between latency, flexibility and CPU overhead.
Composability has received much attention in recent years, perhaps in part due to the collision of multicore hardware with object-oriented techniques designed in single-threaded environments for single-threaded use. And in theory, RCU read-side critical sections may be composed, and in fact may be nested arbitrarily deeply. In practice, as with all real-world implementations of composable constructs, there are limitations.
Implementations of RCU for which rcu_read_lock() and rcu_read_unlock() generate no code, such as Linux-kernel RCU when CONFIG_PREEMPT=n, can be nested arbitrarily deeply. After all, there is no overhead. Except that if all these instances of rcu_read_lock() and rcu_read_unlock() are visible to the compiler, compilation will eventually fail due to exhausting memory, mass storage, or user patience, whichever comes first. If the nesting is not visible to the compiler, as is the case with mutually recursive functions each in its own translation unit, stack overflow will result. If the nesting takes the form of loops, either the control variable will overflow or (in the Linux kernel) you will get an RCU CPU stall warning. Nevertheless, this class of RCU implementations is one of the most composable constructs in existence.
RCU implementations that explicitly track nesting depth are limited by the nesting-depth counter. For example, the Linux kernel's preemptible RCU limits nesting to INT_MAX. This should suffice for almost all practical purposes. That said, a consecutive pair of RCU read-side critical sections between which there is an operation that waits for a grace period cannot be enclosed in another RCU read-side critical section. This is because it is not legal to wait for a grace period within an RCU read-side critical section: To do so would result either in deadlock or in RCU implicitly splitting the enclosing RCU read-side critical section, neither of which is conducive to a long-lived and prosperous kernel.
It is worth noting that RCU is not alone in limiting composability. For example, many transactional-memory implementations prohibit composing a pair of transactions separated by an irrevocable operation (for example, a network receive operation). For another example, lock-based critical sections can be composed surprisingly freely, but only if deadlock is avoided.
In short, although RCU read-side critical sections are highly composable, care is required in some situations, just as is the case for any other composable synchronization mechanism.
A given RCU workload might have an endless and intense stream of RCU read-side critical sections, perhaps even so intense that there was never a point in time during which there was not at least one RCU read-side critical section in flight. RCU cannot allow this situation to block grace periods: As long as all the RCU read-side critical sections are finite, grace periods must also be finite.
That said, preemptible RCU implementations could potentially result in RCU read-side critical sections being preempted for long durations, which has the effect of creating a long-duration RCU read-side critical section. This situation can arise only in heavily loaded systems, but systems using real-time priorities are of course more vulnerable. Therefore, RCU priority boosting is provided to help deal with this case. That said, the exact requirements on RCU priority boosting will likely evolve as more experience accumulates.
Other workloads might have very high update rates. Although one can argue that such workloads should instead use something other than RCU, the fact remains that RCU must handle such workloads gracefully. This requirement is another factor driving batching of grace periods, but it is also the driving force behind the checks for large numbers of queued RCU callbacks in the call_rcu() code path. Finally, high update rates should not delay RCU read-side critical sections, although some read-side delays can occur when using synchronize_rcu_expedited(), courtesy of this function's use of try_stop_cpus(). (In the future, synchronize_rcu_expedited() will be converted to use lighter-weight inter-processor interrupts (IPIs), but this will still disturb readers, though to a much smaller degree.)
Although all three of these corner cases were understood in the early 1990s, a simple user-level test consisting of close(open(path)) in a tight loop in the early 2000s suddenly provided a much deeper appreciation of the high-update-rate corner case. This test also motivated addition of some RCU code to react to high update rates, for example, if a given CPU finds itself with more than 10,000 RCU callbacks queued, it will cause RCU to take evasive action by more aggressively starting grace periods and more aggressively forcing completion of grace-period processing. This evasive action causes the grace period to complete more quickly, but at the cost of restricting RCU's batching optimizations, thus increasing the CPU overhead incurred by that grace period.
Between Murphy's Law and “To err is human”, it is necessary to guard against mishaps and misuse:
Code shared between readers and updaters can use rcu_dereference_check(), which also takes a lockdep expression, and emits a lockdep splat if neither rcu_read_lock() nor the indicated protection is in place. In addition, rcu_dereference_raw() is used in those (hopefully rare) cases where the required protection cannot be easily described. Finally, rcu_read_lock_held() is provided to allow a function to verify that it has been invoked within an RCU read-side critical section. I was made aware of this set of requirements shortly after Thomas Gleixner audited a number of RCU uses.
Some extreme workloads might intentionally delay RCU grace periods, and systems running those workloads can be booted with rcupdate.rcu_cpu_stall_suppress to suppress the splats. This kernel parameter may also be set via sysfs. Furthermore, RCU CPU stall warnings are counter-productive during sysrq dumps and during panics. RCU therefore supplies the rcu_sysrq_start() and rcu_sysrq_end() API members to be called before and after long sysrq dumps. RCU also supplies the rcu_panic() notifier that is automatically invoked at the beginning of a panic to suppress further RCU CPU stall warnings.
This requirement made itself known in the early 1990s, pretty much the first time that it was necessary to debug a CPU stall. That said, the initial implementation in DYNIX/ptx was quite generic in comparison with that of Linux.
This not a hard-and-fast list: RCU's diagnostic capabilities will continue to be guided by the number and type of usage bugs found in real-world RCU usage.
The Linux kernel provides an interesting environment for all kinds of software, including RCU. Some of the relevant points of interest are as follows:
This list is probably incomplete, but it does give a feel for the most notable Linux-kernel complications. Each of the following sections covers one of the above topics.
RCU's goal is automatic configuration, so that almost nobody needs to worry about RCU's Kconfig options. And for almost all users, RCU does in fact work well “out of the box.”
However, there are specialized use cases that are handled by kernel boot parameters and Kconfig options. Unfortunately, the Kconfig system will explicitly ask users about new Kconfig options, which requires almost all of them be hidden behind a CONFIG_RCU_EXPERT Kconfig option.
This all should be quite obvious, but the fact remains that Linus Torvalds recently had to remind me of this requirement.
In many cases, kernel obtains information about the system from the firmware, and sometimes things are lost in translation. Or the translation is accurate, but the original message is bogus.
For example, some systems' firmware overreports the number of CPUs, sometimes by a large factor. If RCU naively believed the firmware, as it used to do, it would create too many per-CPU kthreads. Although the resulting system will still run correctly, the extra kthreads needlessly consume memory and can cause confusion when they show up in ps listings.
RCU must therefore wait for a given CPU to actually come online before it can allow itself to believe that the CPU actually exists. The resulting “ghost CPUs” (which are never going to come online) cause a number of interesting complications.
The Linux kernel's boot sequence is an interesting process, and RCU is used early, even before rcu_init() is invoked. In fact, a number of RCU's primitives can be used as soon as the initial task's task_struct is available and the boot CPU's per-CPU variables are set up. The read-side primitives (rcu_read_lock(), rcu_read_unlock(), rcu_dereference(), and rcu_access_pointer()) will operate normally very early on, as will rcu_assign_pointer().
Although call_rcu() may be invoked at any time during boot, callbacks are not guaranteed to be invoked until after the scheduler is fully up and running. This delay in callback invocation is due to the fact that RCU does not invoke callbacks until it is fully initialized, and this full initialization cannot occur until after the scheduler has initialized itself to the point where RCU can spawn and run its kthreads. In theory, it would be possible to invoke callbacks earlier, however, this is not a panacea because there would be severe restrictions on what operations those callbacks could invoke.
Perhaps surprisingly, synchronize_rcu(), synchronize_rcu_bh() (discussed below), and synchronize_sched() will all operate normally during very early boot, the reason being that there is only one CPU and preemption is disabled. This means that the call synchronize_rcu() (or friends) itself is a quiescent state and thus a grace period, so the early-boot implementation can be a no-op.
Both synchronize_rcu_bh() and synchronize_sched() continue to operate normally through the remainder of boot, courtesy of the fact that preemption is disabled across their RCU read-side critical sections and also courtesy of the fact that there is still only one CPU. However, once the scheduler starts initializing, preemption is enabled. There is still only a single CPU, but the fact that preemption is enabled means that the no-op implementation of synchronize_rcu() no longer works in CONFIG_PREEMPT=y kernels. Therefore, as soon as the scheduler starts initializing, the early-boot fastpath is disabled. This means that synchronize_rcu() switches to its runtime mode of operation where it posts callbacks, which in turn means that any call to synchronize_rcu() will block until the corresponding callback is invoked. Unfortunately, the callback cannot be invoked until RCU's runtime grace-period machinery is up and running, which cannot happen until the scheduler has initialized itself sufficiently to allow RCU's kthreads to be spawned. Therefore, invoking synchronize_rcu() during scheduler initialization can result in deadlock.
@@QQ@@ So what happens with synchronize_rcu() during scheduler initialization for CONFIG_PREEMPT=n kernels?
@@QQA@@ In CONFIG_PREEMPT=n kernel, synchronize_rcu() maps directly to synchronize_sched(). Therefore, synchronize_rcu() works normally throughout boot in CONFIG_PREEMPT=n kernels. However, your code must also work in CONFIG_PREEMPT=y kernels, so it is still necessary to avoid invoking synchronize_rcu() during scheduler initialization.
@@QQE@@
I learned of these boot-time requirements as a result of a series of system hangs.
The Linux kernel has interrupts, and RCU read-side critical sections are legal within interrupt handlers and within interrupt-disabled regions of code, as are invocations of call_rcu().
Some Linux-kernel architectures can enter an interrupt handler from non-idle process context, and then just never leave it, instead stealthily transitioning back to process context. This trick is sometimes used to invoke system calls from inside the kernel. These “half-interrupts” mean that RCU has to be very careful about how it counts interrupt nesting levels. I learned of this requirement the hard way during a rewrite of RCU's dyntick-idle code.
The Linux kernel has non-maskable interrupts (NMIs), and RCU read-side critical sections are legal within NMI handlers. Thankfully, RCU update-side primitives, including call_rcu(), are prohibited within NMI handlers.
The name notwithstanding, some Linux-kernel architectures can have nested NMIs, which RCU must handle correctly. Andy Lutomirski surprised me with this requirement; he also kindly surprised me with an algorithm that meets this requirement.
The Linux kernel has loadable modules, and these modules can also be unloaded. After a given module has been unloaded, any attempt to call one of its functions results in a segmentation fault. The module-unload functions must therefore cancel any delayed calls to loadable-module functions, for example, any outstanding mod_timer() must be dealt with via del_timer_sync() or similar.
Unfortunately, there is no way to cancel an RCU callback; once you invoke call_rcu(), the callback function is going to eventually be invoked, unless the system goes down first. Because it is normally considered socially irresponsible to crash the system in response to a module unload request, we need some other way to deal with in-flight RCU callbacks.
RCU therefore provides rcu_barrier(), which waits until all in-flight RCU callbacks have been invoked. If a module uses call_rcu(), its exit function should therefore prevent any future invocation of call_rcu(), then invoke rcu_barrier(). In theory, the underlying module-unload code could invoke rcu_barrier() unconditionally, but in practice this would incur unacceptable latencies.
Nikita Danilov noted this requirement for an analogous filesystem-unmount situation, and Dipankar Sarma incorporated rcu_barrier() into RCU. The need for rcu_barrier() for module unloading became apparent later.
The Linux kernel supports CPU hotplug, which means that CPUs can come and go. It is of course illegal to use any RCU API member from an offline CPU. This requirement was present from day one in DYNIX/ptx, but on the other hand, the Linux kernel's CPU-hotplug implementation is “interesting.”
The Linux-kernel CPU-hotplug implementation has notifiers that are used to allow the various kernel subsystems (including RCU) to respond appropriately to a given CPU-hotplug operation. Most RCU operations may be invoked from CPU-hotplug notifiers, including even normal synchronous grace-period operations such as synchronize_rcu(). However, expedited grace-period operations such as synchronize_rcu_expedited() are not supported, due to the fact that current implementations block CPU-hotplug operations, which could result in deadlock.
In addition, all-callback-wait operations such as rcu_barrier() are also not supported, due to the fact that there are phases of CPU-hotplug operations where the outgoing CPU's callbacks will not be invoked until after the CPU-hotplug operation ends, which could also result in deadlock.
RCU depends on the scheduler, and the scheduler uses RCU to protect some of its data structures. This means the scheduler is forbidden from acquiring the runqueue locks and the priority-inheritance locks in the middle of an outermost RCU read-side critical section unless either (1) it releases them before exiting that same RCU read-side critical section, or (2) interrupts are disabled across that entire RCU read-side critical section. This same prohibition also applies (recursively!) to any lock that is acquired while holding any lock to which this prohibition applies. Adhering to this rule prevents preemptible RCU from invoking rcu_read_unlock_special() while either runqueue or priority-inheritance locks are held, thus avoiding deadlock.
Prior to v4.4, it was only necessary to disable preemption across RCU read-side critical sections that acquired scheduler locks. In v4.4, expedited grace periods started using IPIs, and these IPIs could force a rcu_read_unlock() to take the slowpath. Therefore, this expedited-grace-period change required disabling of interrupts, not just preemption.
For RCU's part, the preemptible-RCU rcu_read_unlock() implementation must be written carefully to avoid similar deadlocks. In particular, rcu_read_unlock() must tolerate an interrupt where the interrupt handler invokes both rcu_read_lock() and rcu_read_unlock(). This possibility requires rcu_read_unlock() to use negative nesting levels to avoid destructive recursion via interrupt handler's use of RCU.
This pair of mutual scheduler-RCU requirements came as a complete surprise.
As noted above, RCU makes use of kthreads, and it is necessary to avoid excessive CPU-time accumulation by these kthreads. This requirement was no surprise, but RCU's violation of it when running context-switch-heavy workloads when built with CONFIG_NO_HZ_FULL=y did come as a surprise [PDF]. RCU has made good progress towards meeting this requirement, even for context-switch-have CONFIG_NO_HZ_FULL=y workloads, but there is room for further improvement.
It is possible to use tracing on RCU code, but tracing itself uses RCU. For this reason, rcu_dereference_raw_notrace() is provided for use by tracing, which avoids the destructive recursion that could otherwise ensue. This API is also used by virtualization in some architectures, where RCU readers execute in environments in which tracing cannot be used. The tracing folks both located the requirement and provided the needed fix, so this surprise requirement was relatively painless.
Interrupting idle CPUs is considered socially unacceptable, especially by people with battery-powered embedded systems. RCU therefore conserves energy by detecting which CPUs are idle, including tracking CPUs that have been interrupted from idle. This is a large part of the energy-efficiency requirement, so I learned of this via an irate phone call.
Because RCU avoids interrupting idle CPUs, it is illegal to execute an RCU read-side critical section on an idle CPU. (Kernels built with CONFIG_PROVE_RCU=y will splat if you try it.) The RCU_NONIDLE() macro and _rcuidle event tracing is provided to work around this restriction. In addition, rcu_is_watching() may be used to test whether or not it is currently legal to run RCU read-side critical sections on this CPU. I learned of the need for diagnostics on the one hand and RCU_NONIDLE() on the other while inspecting idle-loop code. Steven Rostedt supplied _rcuidle event tracing, which is used quite heavily in the idle loop.
It is similarly socially unacceptable to interrupt an nohz_full CPU running in userspace. RCU must therefore track nohz_full userspace execution. And in CONFIG_NO_HZ_FULL_SYSIDLE=y kernels, RCU must separately track idle CPUs on the one hand and CPUs that are either idle or executing in userspace on the other. In both cases, RCU must be able to sample state at two points in time, and be able to determine whether or not some other CPU spent any time idle and/or executing in userspace.
These energy-efficiency requirements have proven quite difficult to understand and to meet, for example, there have been more than five clean-sheet rewrites of RCU's energy-efficiency code, the last of which was finally able to demonstrate real energy savings running on real hardware [PDF]. As noted earlier, I learned of many of these requirements via angry phone calls: Flaming me on the Linux-kernel mailing list was apparently not sufficient to fully vent their ire at RCU's energy-efficiency bugs!
Although small-memory non-realtime systems can simply use Tiny RCU, code size is only one aspect of memory efficiency. Another aspect is the size of the rcu_head structure used by call_rcu() and kfree_rcu(). Although this structure contains nothing more than a pair of pointers, it does appear in many RCU-protected data structures, including some that are size critical. The page structure is a case in point, as evidenced by the many occurrences of the union keyword within that structure.
This need for memory efficiency is one reason that RCU uses hand-crafted singly linked lists to track the rcu_head structures that are waiting for a grace period to elapse. It is also the reason why rcu_head structures do not contain debug information, such as fields tracking the file and line of the call_rcu() or kfree_rcu() that posted them. Although this information might appear in debug-only kernel builds at some point, in the meantime, the ->func field will often provide the needed debug information.
However, in some cases, the need for memory efficiency leads to even more extreme measures. Returning to the page structure, the rcu_head field shares storage with a great many other structures that are used at various points in the corresponding page's lifetime. In order to correctly resolve certain race conditions, the Linux kernel's memory-management subsystem needs a particular bit to remain zero during all phases of grace-period processing, and that bit happens to map to the bottom bit of the rcu_head structure's ->next field. RCU makes this guarantee as long as call_rcu() is used to post the callback, as opposed to kfree_rcu() or some future “lazy” variant of call_rcu() that might one day be created for energy-efficiency purposes.
Expanding on the earlier discussion, RCU is used heavily by hot code paths in performance-critical portions of the Linux kernel's networking, security, virtualization, and scheduling code paths. RCU must therefore use efficient implementations, especially in its read-side primitives. To that end, it would be good if preemptible RCU's implementation of rcu_read_lock() could be inlined, however, doing this requires resolving #include issues with the task_struct structure.
The Linux kernel supports hardware configurations with up to 4096 CPUs, which means that RCU must be extremely scalable. Algorithms that involve frequent acquisitions of global locks or frequent atomic operations on global variables simply cannot be tolerated within the RCU implementation. RCU therefore makes heavy use of a combining tree based on the rcu_node structure. RCU is required to tolerate all CPUs continuously invoking any combination of RCU's runtime primitives with minimal per-operation overhead. In fact, in many cases, increasing load must decrease the per-operation overhead, witness the batching optimizations for synchronize_rcu(), call_rcu(), synchronize_rcu_expedited(), and rcu_barrier(). As a general rule, RCU must cheerfully accept whatever the rest of the Linux kernel decides to throw at it.
The Linux kernel is used for real-time workloads, especially in conjunction with the -rt patchset. The real-time-latency response requirements are such that the traditional approach of disabling preemption across RCU read-side critical sections is inappropriate. Kernels built with CONFIG_PREEMPT=y therefore use an RCU implementation that allows RCU read-side critical sections to be preempted. This requirement made its presence known after users made it clear that an earlier real-time patch did not meet their needs, in conjunction with some RCU issues encountered by a very early version of the -rt patchset.
In addition, RCU must make do with a sub-100-microsecond real-time latency budget. In fact, on smaller systems with the -rt patchset, the Linux kernel provides sub-20-microsecond real-time latencies for the whole kernel, including RCU. RCU's scalability and latency must therefore be sufficient for these sorts of configurations. To my surprise, the sub-100-microsecond real-time latency budget applies to even the largest systems [PDF], up to and including systems with 4096 CPUs. This real-time requirement motivated the grace-period kthread, which also simplified handling of a number of race conditions.
Finally, RCU's status as a synchronization primitive means that any RCU failure can result in arbitrary memory corruption that can be extremely difficult to debug. This means that RCU must be extremely reliable, which in practice also means that RCU must have an aggressive stress-test suite. This stress-test suite is called rcutorture.
Although the need for rcutorture was no surprise, the current immense popularity of the Linux kernel is posing interesting—and perhaps unprecedented—validation challenges. To see this, keep in mind that there are well over one billion instances of the Linux kernel running today, given Android smartphones, Linux-powered televisions, and servers. This number can be expected to increase sharply with the advent of the celebrated Internet of Things.
Suppose that RCU contains a race condition that manifests on average once per million years of runtime. This bug will be occurring about three times per day across the installed base. RCU could simply hide behind hardware error rates, given that no one should really expect their smartphone to last for a million years. However, anyone taking too much comfort from this thought should consider the fact that in most jurisdictions, a successful multi-year test of a given mechanism, which might include a Linux kernel, suffices for a number of types of safety-critical certifications. In fact, rumor has it that the Linux kernel is already being used in production for safety-critical applications. I don't know about you, but I would feel quite bad if a bug in RCU killed someone. Which might explain my recent focus on validation and verification.
One of the more surprising things about RCU is that there are now no fewer than five flavors, or API families. In addition, the primary flavor that has been the sole focus up to this point has two different implementations, non-preemptible and preemptible. The other four flavors are listed below, with requirements for each described in a separate section.
The softirq-disable (AKA “bottom-half”, hence the “_bh” abbreviations) flavor of RCU, or RCU-bh, was developed by Dipankar Sarma to provide a flavor of RCU that could withstand the network-based denial-of-service attacks researched by Robert Olsson. These attacks placed so much networking load on the system that some of the CPUs never exited softirq execution, which in turn prevented those CPUs from ever executing a context switch, which, in the RCU implementation of that time, prevented grace periods from ever ending. The result was an out-of-memory condition and a system hang.
The solution was the creation of RCU-bh, which does local_bh_disable() across its read-side critical sections, and which uses the transition from one type of softirq processing to another as a quiescent state in addition to context switch, idle, user mode, and offline. This means that RCU-bh grace periods can complete even when some of the CPUs execute in softirq indefinitely, thus allowing algorithms based on RCU-bh to withstand network-based denial-of-service attacks.
Because rcu_read_lock_bh() and rcu_read_unlock_bh() disable and re-enable softirq handlers, any attempt to start a softirq handlers during the RCU-bh read-side critical section will be deferred. In this case, rcu_read_unlock_bh() will invoke softirq processing, which can take considerable time. One can of course argue that this softirq overhead should be associated with the code following the RCU-bh read-side critical section rather than rcu_read_unlock_bh(), but the fact is that most profiling tools cannot be expected to make this sort of fine distinction. For example, suppose that a three-millisecond-long RCU-bh read-side critical section executes during a time of heavy networking load. There will very likely be an attempt to invoke at least one softirq handler during that three milliseconds, but any such invocation will be delayed until the time of the rcu_read_unlock_bh(). This can of course make it appear at first glance as if rcu_read_unlock_bh() was executing very slowly.
The RCU-bh API includes rcu_read_lock_bh(), rcu_read_unlock_bh(), rcu_dereference_bh(), rcu_dereference_bh_check(), synchronize_rcu_bh(), synchronize_rcu_bh_expedited(), call_rcu_bh(), rcu_barrier_bh(), and rcu_read_lock_bh_held().
Before preemptible RCU, waiting for an RCU grace period had the side effect of also waiting for all pre-existing interrupt and NMI handlers. However, there are legitimate preemptible-RCU implementations that do not have this property, given that any point in the code outside of an RCU read-side critical section can be a quiescent state. Therefore, RCU-sched was created, which follows “classic” RCU in that an RCU-sched grace period waits for for pre-existing interrupt and NMI handlers. In kernels built with CONFIG_PREEMPT=n, the RCU and RCU-sched APIs have identical implementations, while kernels built with CONFIG_PREEMPT=y provide a separate implementation for each.
Note well that in CONFIG_PREEMPT=y kernels, rcu_read_lock_sched() and rcu_read_unlock_sched() disable and re-enable preemption, respectively. This means that if there was a preemption attempt during the RCU-sched read-side critical section, rcu_read_unlock_sched() will enter the scheduler, with all the latency and overhead entailed. Just as with rcu_read_unlock_bh(), this can make it look as if rcu_read_unlock_sched() was executing very slowly. However, the highest-priority task won't be preempted, so that task will enjoy low-overhead rcu_read_unlock_sched() invocations.
The RCU-sched API includes rcu_read_lock_sched(), rcu_read_unlock_sched(), rcu_read_lock_sched_notrace(), rcu_read_unlock_sched_notrace(), rcu_dereference_sched(), rcu_dereference_sched_check(), synchronize_sched(), synchronize_rcu_sched_expedited(), call_rcu_sched(), rcu_barrier_sched(), and rcu_read_lock_sched_held(). However, anything that disables preemption also marks an RCU-sched read-side critical section, including preempt_disable() and preempt_enable(), local_irq_save() and local_irq_restore(), and so on.
For well over a decade, someone saying “I need to block within an RCU read-side critical section” was a reliable indication that this someone did not understand RCU. After all, if you are always blocking in an RCU read-side critical section, you can probably afford to use a higher-overhead synchronization mechanism. However, that changed with the advent of the Linux kernel's notifiers, whose RCU read-side critical sections almost never sleep, but sometimes need to. This resulted in the introduction of sleepable RCU, or SRCU.
SRCU allows different domains to be defined, with each such domain defined by an instance of an srcu_struct structure. A pointer to this structure must be passed in to each SRCU function, for example, synchronize_srcu(&ss), where ss is the srcu_struct structure. The key benefit of these domains is that a slow SRCU reader in one domain does not delay an SRCU grace period in some other domain. That said, one consequence of these domains is that read-side code must pass a “cookie” from srcu_read_lock() to srcu_read_unlock(), for example, as follows:
1 int idx; 2 3 idx = srcu_read_lock(&ss); 4 do_something(); 5 srcu_read_unlock(&ss, idx);
As noted above, it is legal to block within SRCU read-side critical sections, however, with great power comes great responsibility. If you block forever in one of a given domain's SRCU read-side critical sections, then that domain's grace periods will also be blocked forever. Of course, one good way to block forever is to deadlock, which can happen if any operation in a given domain's SRCU read-side critical section can block waiting, either directly or indirectly, for that domain's grace period to elapse. For example, this results in a self-deadlock:
1 int idx; 2 3 idx = srcu_read_lock(&ss); 4 do_something(); 5 synchronize_srcu(&ss); 6 srcu_read_unlock(&ss, idx);
However, if line 5 acquired a mutex that was held across a synchronize_srcu() for domain ss, deadlock would still be possible. Furthermore, if line 5 acquired a mutex that was held across a synchronize_srcu() for some other domain ss1, and if an ss1-domain SRCU read-side critical section acquired another mutex that was held across as ss-domain synchronize_srcu(), deadlock would again be possible. Such a deadlock cycle could extend across an arbitrarily large number of different SRCU domains. Again, with great power comes great responsibility.
Unlike the other RCU flavors, SRCU read-side critical sections can run on idle and even offline CPUs. This ability requires that srcu_read_lock() and srcu_read_unlock() contain memory barriers, which means that SRCU readers will run a bit slower than would RCU readers. It also motivates the smp_mb__after_srcu_read_unlock() API, which, in combination with srcu_read_unlock(), guarantees a full memory barrier.
The SRCU API includes srcu_read_lock(), srcu_read_unlock(), srcu_dereference(), srcu_dereference_check(), synchronize_srcu(), synchronize_srcu_expedited(), call_srcu(), srcu_barrier(), and srcu_read_lock_held(). It also includes DEFINE_SRCU(), DEFINE_STATIC_SRCU(), and init_srcu_struct() APIs for defining and initializing srcu_struct structures.
Some forms of tracing use “tramopolines” to handle the binary rewriting required to install different types of probes. It would be good to be able to free old trampolines, which sounds like a job for some form of RCU. However, because it is necessary to be able to install a trace anywhere in the code, it is not possible to use read-side markers such as rcu_read_lock() and rcu_read_unlock(). In addition, it does not work to have these markers in the trampoline itself, because there would need to be instructions following rcu_read_unlock(). Although synchronize_rcu() would guarantee that execution reached the rcu_read_unlock(), it would not be able to guarantee that execution had completely left the trampoline.
The solution, in the form of Tasks RCU, is to have implicit read-side critical sections that are delimited by voluntary context switches, that is, calls to schedule(), cond_resched_rcu_qs(), and synchronize_rcu_tasks(). In addition, transitions to and from userspace execution also delimit tasks-RCU read-side critical sections.
The tasks-RCU API is quite compact, consisting only of call_rcu_tasks(), synchronize_rcu_tasks(), and rcu_barrier_tasks().
One of the tricks that RCU uses to attain update-side scalability is to increase grace-period latency with increasing numbers of CPUs. If this becomes a serious problem, it will be necessary to rework the grace-period state machine so as to avoid the need for the additional latency.
Expedited grace periods scan the CPUs, so their latency and overhead increases with increasing numbers of CPUs. If this becomes a serious problem on large systems, it will be necessary to do some redesign to avoid this scalability problem.
RCU disables CPU hotplug in a few places, perhaps most notably in the expedited grace-period and rcu_barrier() operations. If there is a strong reason to use expedited grace periods in CPU-hotplug notifiers, it will be necessary to avoid disabling CPU hotplug. This would introduce some complexity, so there had better be a very good reason.
The tradeoff between grace-period latency on the one hand and interruptions of other CPUs on the other hand may need to be re-examined. The desire is of course for zero grace-period latency as well as zero interprocessor interrupts undertaken during an expedited grace period operation. While this ideal is unlikely to be achievable, it is quite possible that further improvements can be made.
The multiprocessor implementations of RCU use a combining tree that groups CPUs so as to reduce lock contention and increase cache locality. However, this combining tree does not spread its memory across NUMA nodes nor does it align the CPU groups with hardware features such as sockets or cores. Such spreading and alignment is currently believed to be unnecessary because the hotpath read-side primitives do not access the combining tree, nor does call_rcu() in the common case. If you believe that your architecture needs such spreading and alignment, then your architecture should also benefit from the rcutree.rcu_fanout_leaf boot parameter, which can be set to the number of CPUs in a socket, NUMA node, or whatever. If the number of CPUs is too large, use a fraction of the number of CPUs. If the number of CPUs is a large prime number, well, that certainly is an “interesting” architectural choice! More flexible arrangements might be considered, but only if rcutree.rcu_fanout_leaf has proven inadequate, and only if the inadequacy has been demonstrated by a carefully run and realistic system-level workload.
Please note that arrangements that require RCU to remap CPU numbers will require extremely good demonstration of need and full exploration of alternatives.
There is an embarrassingly large number of flavors of RCU, and this number has been increasing over time. Perhaps it will be possible to combine some at some future date.
RCU's various kthreads are reasonably recent additions. It is quite likely that adjustments will be required to more gracefully handle extreme loads. It might also be necessary to be able to relate CPU utilization by RCU's kthreads and softirq handlers to the code that instigated this CPU utilization. For example, RCU callback overhead might be charged back to the originating call_rcu() instance, though probably not in production kernels.
This document has presented more than two decade's worth of RCU requirements. Given that the requirements keep changing, this will not be the last word on this subject, but at least it serves to get an important subset of the requirements set forth.
@@QQAL@@