12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031 |
- Please note that the "What is RCU?" LWN series is an excellent place
- to start learning about RCU:
- 1. What is RCU, Fundamentally? http://lwn.net/Articles/262464/
- 2. What is RCU? Part 2: Usage http://lwn.net/Articles/263130/
- 3. RCU part 3: the RCU API http://lwn.net/Articles/264090/
- 4. The RCU API, 2010 Edition http://lwn.net/Articles/418853/
- 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 once you understand it,
- getting there can sometimes be a challenge. Part of the problem is that
- most of the past descriptions of RCU have been written with the mistaken
- assumption that there is "one true way" to describe RCU. Instead,
- the experience has been that different people must take different paths
- to arrive at an understanding of RCU. 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. FULL LIST OF RCU APIs
- 8. 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. Section 7 serves as an index to the docbook API
- documentation, and Section 8 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:
- a. Remove pointers to a data structure, so that subsequent
- readers cannot gain a reference to it.
- b. Wait for all previous readers to complete their RCU read-side
- critical sections.
- c. 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:
- a. rcu_read_lock()
- b. rcu_read_unlock()
- c. synchronize_rcu() / call_rcu()
- d. rcu_assign_pointer()
- e. 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);
- 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.
- rcu_read_unlock()
- void rcu_read_unlock(void);
- Used by a reader to inform the reclaimer that the reader is
- exiting 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);
- 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 a 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
- checklist.txt for some approaches to limiting the update rate.
- rcu_assign_pointer()
- typeof(p) 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.
- (Compiler experts will no doubt disagree.)
- The updater uses this function 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 function returns
- the new value, and also executes any memory-barrier instructions
- required for a given CPU architecture.
- Perhaps just as important, it 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 rcu_dereference() to fetch an RCU-protected
- pointer, which returns a value that may then be safely
- dereferenced. Note that rcu_deference() 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 nothing, not even a compiler directive.
- 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.
- 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().
- 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 time 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.
- There are no fewer than three RCU mechanisms in the Linux kernel; the
- diagram above shows the first one, which is by far the most commonly used.
- The rcu_dereference() and rcu_assign_pointer() primitives are used for
- all three mechanisms, but different defer and protect primitives are
- used as follows:
- Defer Protect
- a. synchronize_rcu() rcu_read_lock() / rcu_read_unlock()
- call_rcu() rcu_dereference()
- b. synchronize_rcu_bh() rcu_read_lock_bh() / rcu_read_unlock_bh()
- call_rcu_bh() rcu_dereference_bh()
- c. synchronize_sched() rcu_read_lock_sched() / rcu_read_unlock_sched()
- call_rcu_sched() preempt_disable() / preempt_enable()
- local_irq_save() / local_irq_restore()
- hardirq enter / hardirq exit
- NMI enter / NMI exit
- rcu_dereference_sched()
- These three mechanisms are used as follows:
- a. RCU applied to normal data structures.
- b. RCU applied to networking data structures that may be subjected
- to remote denial-of-service attacks.
- c. 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.
- 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 listRCU.txt, arrayRCU.txt, and NMI-RCU.txt.
- 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:
- o Use rcu_read_lock() and rcu_read_unlock() to guard RCU
- read-side critical sections.
- o Within an RCU read-side critical section, use rcu_dereference()
- to dereference RCU-protected pointers.
- o Use some solid scheme (such as locks or semaphores) to
- keep concurrent updates from interfering with each other.
- o 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.
- o 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 checklist.txt for additional rules to follow when using RCU.
- And again, more-typical uses of RCU may be found in listRCU.txt,
- arrayRCU.txt, and NMI-RCU.txt.
- 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,
- void (*func)(struct rcu_head *head));
- 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():
- o 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);
- Again, see checklist.txt 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/rcupdate.c for a
- production-quality implementation, and see:
- http://www.rdrop.com/users/paulmck/RCU
- 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.
- 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);
- write_unlock(&rcu_gp_mutex);
- }
- [You can ignore rcu_assign_pointer() and rcu_dereference() without
- missing much. But here they are anyway. And whatever you do, don't
- forget about them when submitting patches making use of RCU!]
- #define rcu_assign_pointer(p, v) ({ \
- smp_wmb(); \
- (p) = (v); \
- })
- #define rcu_dereference(p) ({ \
- typeof(p) _________p1 = p; \
- smp_read_barrier_depends(); \
- (_________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 immediately 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.
- 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?
- 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_PREEMPT
- 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-.
- Quick Quiz #3: If it is illegal to block in an RCU read-side
- critical section, what the heck do you do in
- PREEMPT_RT, where normal spinlocks can block???
- 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.
- @@ -13,15 +14,15 @@
- struct list_head *lp;
- struct el *p;
- - read_lock();
- - 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();
- + rcu_read_unlock();
- return 1;
- }
- }
- - read_unlock();
- + 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 spinlock_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(); 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(); 10 rcu_read_unlock();
- 11 return 1; 11 return 1;
- 12 } 12 }
- 13 } 13 }
- 14 read_unlock(); 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. 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_first_entry_rcu
- list_next_rcu
- list_for_each_entry_rcu
- list_for_each_entry_continue_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_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_del_rcu
- hlist_del_init_rcu
- hlist_replace_rcu
- list_splice_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_bh rcu_barrier_bh
- rcu_read_unlock_bh synchronize_rcu_bh
- rcu_dereference_bh synchronize_rcu_bh_expedited
- rcu_dereference_bh_check
- rcu_dereference_bh_protected
- rcu_read_lock_bh_held
- sched: Critical sections Grace period Barrier
- rcu_read_lock_sched synchronize_sched rcu_barrier_sched
- rcu_read_unlock_sched call_rcu_sched
- [preempt_disable] synchronize_sched_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
- SRCU: Critical sections Grace period Barrier
- srcu_read_lock synchronize_srcu srcu_barrier
- srcu_read_unlock call_srcu
- srcu_dereference synchronize_srcu_expedited
- srcu_dereference_check
- srcu_read_lock_held
- SRCU: Initialization/cleanup
- init_srcu_struct
- cleanup_srcu_struct
- All: lockdep-checked RCU-protected pointer access
- rcu_access_pointer
- rcu_dereference_raw
- RCU_LOCKDEP_WARN
- rcu_sleep_check
- RCU_NONIDLE
- 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:
- a. Will readers need to block? If so, you need SRCU.
- b. What about the -rt patchset? If readers would need to block
- in an non-rt kernel, you need SRCU. If readers would block
- in a -rt kernel, but not in a non-rt kernel, SRCU is not
- necessary.
- c. 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 is the only choice that will work for you.
- d. 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 need RCU-bh.
- e. Is your workload too update-intensive for normal use of
- RCU, but inappropriate for other synchronization mechanisms?
- If so, consider SLAB_DESTROY_BY_RCU. But please be careful!
- f. Do you need read-side critical sections that are respected
- even though they are in the middle of the idle loop, during
- user-mode execution, or on an offlined CPU? If so, SRCU is the
- only choice that will work for you.
- g. Otherwise, use RCU.
- Of course, this all assumes that you have determined that RCU is in fact
- the right tool for your job.
- 8. 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().
- 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_PREEMPT
- 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. ;-)
- Quick Quiz #3: If it is illegal to block in an RCU read-side
- critical section, what the heck do you do in
- PREEMPT_RT, where normal spinlocks can block???
- Answer: Just as 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 it
- 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???
- 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.
|