| Review Checklist for RCU Patches |
| |
| |
| This document contains a checklist for producing and reviewing patches |
| that make use of RCU. Violating any of the rules listed below will |
| result in the same sorts of problems that leaving out a locking primitive |
| would cause. This list is based on experiences reviewing such patches |
| over a rather long period of time, but improvements are always welcome! |
| |
| 0. Is RCU being applied to a read-mostly situation? If the data |
| structure is updated more than about 10% of the time, then you |
| should strongly consider some other approach, unless detailed |
| performance measurements show that RCU is nonetheless the right |
| tool for the job. Yes, RCU does reduce read-side overhead by |
| increasing write-side overhead, which is exactly why normal uses |
| of RCU will do much more reading than updating. |
| |
| Another exception is where performance is not an issue, and RCU |
| provides a simpler implementation. An example of this situation |
| is the dynamic NMI code in the Linux 2.6 kernel, at least on |
| architectures where NMIs are rare. |
| |
| Yet another exception is where the low real-time latency of RCU's |
| read-side primitives is critically important. |
| |
| 1. Does the update code have proper mutual exclusion? |
| |
| RCU does allow -readers- to run (almost) naked, but -writers- must |
| still use some sort of mutual exclusion, such as: |
| |
| a. locking, |
| b. atomic operations, or |
| c. restricting updates to a single task. |
| |
| If you choose #b, be prepared to describe how you have handled |
| memory barriers on weakly ordered machines (pretty much all of |
| them -- even x86 allows later loads to be reordered to precede |
| earlier stores), and be prepared to explain why this added |
| complexity is worthwhile. If you choose #c, be prepared to |
| explain how this single task does not become a major bottleneck on |
| big multiprocessor machines (for example, if the task is updating |
| information relating to itself that other tasks can read, there |
| by definition can be no bottleneck). |
| |
| 2. Do the RCU read-side critical sections make proper use of |
| rcu_read_lock() and friends? These primitives are needed |
| to prevent grace periods from ending prematurely, which |
| could result in data being unceremoniously freed out from |
| under your read-side code, which can greatly increase the |
| actuarial risk of your kernel. |
| |
| As a rough rule of thumb, any dereference of an RCU-protected |
| pointer must be covered by rcu_read_lock(), rcu_read_lock_bh(), |
| rcu_read_lock_sched(), or by the appropriate update-side lock. |
| Disabling of preemption can serve as rcu_read_lock_sched(), but |
| is less readable. |
| |
| 3. Does the update code tolerate concurrent accesses? |
| |
| The whole point of RCU is to permit readers to run without |
| any locks or atomic operations. This means that readers will |
| be running while updates are in progress. There are a number |
| of ways to handle this concurrency, depending on the situation: |
| |
| a. Use the RCU variants of the list and hlist update |
| primitives to add, remove, and replace elements on |
| an RCU-protected list. Alternatively, use the other |
| RCU-protected data structures that have been added to |
| the Linux kernel. |
| |
| This is almost always the best approach. |
| |
| b. Proceed as in (a) above, but also maintain per-element |
| locks (that are acquired by both readers and writers) |
| that guard per-element state. Of course, fields that |
| the readers refrain from accessing can be guarded by |
| some other lock acquired only by updaters, if desired. |
| |
| This works quite well, also. |
| |
| c. Make updates appear atomic to readers. For example, |
| pointer updates to properly aligned fields will |
| appear atomic, as will individual atomic primitives. |
| Sequences of perations performed under a lock will -not- |
| appear to be atomic to RCU readers, nor will sequences |
| of multiple atomic primitives. |
| |
| This can work, but is starting to get a bit tricky. |
| |
| d. Carefully order the updates and the reads so that |
| readers see valid data at all phases of the update. |
| This is often more difficult than it sounds, especially |
| given modern CPUs' tendency to reorder memory references. |
| One must usually liberally sprinkle memory barriers |
| (smp_wmb(), smp_rmb(), smp_mb()) through the code, |
| making it difficult to understand and to test. |
| |
| It is usually better to group the changing data into |
| a separate structure, so that the change may be made |
| to appear atomic by updating a pointer to reference |
| a new structure containing updated values. |
| |
| 4. Weakly ordered CPUs pose special challenges. Almost all CPUs |
| are weakly ordered -- even x86 CPUs allow later loads to be |
| reordered to precede earlier stores. RCU code must take all of |
| the following measures to prevent memory-corruption problems: |
| |
| a. Readers must maintain proper ordering of their memory |
| accesses. The rcu_dereference() primitive ensures that |
| the CPU picks up the pointer before it picks up the data |
| that the pointer points to. This really is necessary |
| on Alpha CPUs. If you don't believe me, see: |
| |
| http://www.openvms.compaq.com/wizard/wiz_2637.html |
| |
| The rcu_dereference() primitive is also an excellent |
| documentation aid, letting the person reading the code |
| know exactly which pointers are protected by RCU. |
| Please note that compilers can also reorder code, and |
| they are becoming increasingly aggressive about doing |
| just that. The rcu_dereference() primitive therefore |
| also prevents destructive compiler optimizations. |
| |
| The rcu_dereference() primitive is used by the |
| various "_rcu()" list-traversal primitives, such |
| as the list_for_each_entry_rcu(). Note that it is |
| perfectly legal (if redundant) for update-side code to |
| use rcu_dereference() and the "_rcu()" list-traversal |
| primitives. This is particularly useful in code that |
| is common to readers and updaters. However, lockdep |
| will complain if you access rcu_dereference() outside |
| of an RCU read-side critical section. See lockdep.txt |
| to learn what to do about this. |
| |
| Of course, neither rcu_dereference() nor the "_rcu()" |
| list-traversal primitives can substitute for a good |
| concurrency design coordinating among multiple updaters. |
| |
| b. If the list macros are being used, the list_add_tail_rcu() |
| and list_add_rcu() primitives must be used in order |
| to prevent weakly ordered machines from misordering |
| structure initialization and pointer planting. |
| Similarly, if the hlist macros are being used, the |
| hlist_add_head_rcu() primitive is required. |
| |
| c. If the list macros are being used, the list_del_rcu() |
| primitive must be used to keep list_del()'s pointer |
| poisoning from inflicting toxic effects on concurrent |
| readers. Similarly, if the hlist macros are being used, |
| the hlist_del_rcu() primitive is required. |
| |
| The list_replace_rcu() and hlist_replace_rcu() primitives |
| may be used to replace an old structure with a new one |
| in their respective types of RCU-protected lists. |
| |
| d. Rules similar to (4b) and (4c) apply to the "hlist_nulls" |
| type of RCU-protected linked lists. |
| |
| e. Updates must ensure that initialization of a given |
| structure happens before pointers to that structure are |
| publicized. Use the rcu_assign_pointer() primitive |
| when publicizing a pointer to a structure that can |
| be traversed by an RCU read-side critical section. |
| |
| 5. If call_rcu(), or a related primitive such as call_rcu_bh() or |
| call_rcu_sched(), is used, the callback function must be |
| written to be called from softirq context. In particular, |
| it cannot block. |
| |
| 6. Since synchronize_rcu() can block, it cannot be called from |
| any sort of irq context. The same rule applies for |
| synchronize_rcu_bh(), synchronize_sched(), synchronize_srcu(), |
| synchronize_rcu_expedited(), synchronize_rcu_bh_expedited(), |
| synchronize_sched_expedite(), and synchronize_srcu_expedited(). |
| |
| The expedited forms of these primitives have the same semantics |
| as the non-expedited forms, but expediting is both expensive |
| and unfriendly to real-time workloads. Use of the expedited |
| primitives should be restricted to rare configuration-change |
| operations that would not normally be undertaken while a real-time |
| workload is running. |
| |
| 7. If the updater uses call_rcu() or synchronize_rcu(), then the |
| corresponding readers must use rcu_read_lock() and |
| rcu_read_unlock(). If the updater uses call_rcu_bh() or |
| synchronize_rcu_bh(), then the corresponding readers must |
| use rcu_read_lock_bh() and rcu_read_unlock_bh(). If the |
| updater uses call_rcu_sched() or synchronize_sched(), then |
| the corresponding readers must disable preemption, possibly |
| by calling rcu_read_lock_sched() and rcu_read_unlock_sched(). |
| If the updater uses synchronize_srcu(), the the corresponding |
| readers must use srcu_read_lock() and srcu_read_unlock(), |
| and with the same srcu_struct. The rules for the expedited |
| primitives are the same as for their non-expedited counterparts. |
| Mixing things up will result in confusion and broken kernels. |
| |
| One exception to this rule: rcu_read_lock() and rcu_read_unlock() |
| may be substituted for rcu_read_lock_bh() and rcu_read_unlock_bh() |
| in cases where local bottom halves are already known to be |
| disabled, for example, in irq or softirq context. Commenting |
| such cases is a must, of course! And the jury is still out on |
| whether the increased speed is worth it. |
| |
| 8. Although synchronize_rcu() is slower than is call_rcu(), it |
| usually results in simpler code. So, unless update performance |
| is critically important or the updaters cannot block, |
| synchronize_rcu() should be used in preference to call_rcu(). |
| |
| An especially important property of the synchronize_rcu() |
| primitive is that it automatically self-limits: if grace periods |
| are delayed for whatever reason, then the synchronize_rcu() |
| primitive will correspondingly delay updates. In contrast, |
| code using call_rcu() should explicitly limit update rate in |
| cases where grace periods are delayed, as failing to do so can |
| result in excessive realtime latencies or even OOM conditions. |
| |
| Ways of gaining this self-limiting property when using call_rcu() |
| include: |
| |
| a. Keeping a count of the number of data-structure elements |
| used by the RCU-protected data structure, including |
| those waiting for a grace period to elapse. Enforce a |
| limit on this number, stalling updates as needed to allow |
| previously deferred frees to complete. Alternatively, |
| limit only the number awaiting deferred free rather than |
| the total number of elements. |
| |
| One way to stall the updates is to acquire the update-side |
| mutex. (Don't try this with a spinlock -- other CPUs |
| spinning on the lock could prevent the grace period |
| from ever ending.) Another way to stall the updates |
| is for the updates to use a wrapper function around |
| the memory allocator, so that this wrapper function |
| simulates OOM when there is too much memory awaiting an |
| RCU grace period. There are of course many other |
| variations on this theme. |
| |
| b. Limiting update rate. For example, if updates occur only |
| once per hour, then no explicit rate limiting is required, |
| unless your system is already badly broken. The dcache |
| subsystem takes this approach -- updates are guarded |
| by a global lock, limiting their rate. |
| |
| c. Trusted update -- if updates can only be done manually by |
| superuser or some other trusted user, then it might not |
| be necessary to automatically limit them. The theory |
| here is that superuser already has lots of ways to crash |
| the machine. |
| |
| d. Use call_rcu_bh() rather than call_rcu(), in order to take |
| advantage of call_rcu_bh()'s faster grace periods. |
| |
| e. Periodically invoke synchronize_rcu(), permitting a limited |
| number of updates per grace period. |
| |
| The same cautions apply to call_rcu_bh() and call_rcu_sched(). |
| |
| 9. All RCU list-traversal primitives, which include |
| rcu_dereference(), list_for_each_entry_rcu(), |
| list_for_each_continue_rcu(), and list_for_each_safe_rcu(), |
| must be either within an RCU read-side critical section or |
| must be protected by appropriate update-side locks. RCU |
| read-side critical sections are delimited by rcu_read_lock() |
| and rcu_read_unlock(), or by similar primitives such as |
| rcu_read_lock_bh() and rcu_read_unlock_bh(), in which case |
| the matching rcu_dereference() primitive must be used in order |
| to keep lockdep happy, in this case, rcu_dereference_bh(). |
| |
| The reason that it is permissible to use RCU list-traversal |
| primitives when the update-side lock is held is that doing so |
| can be quite helpful in reducing code bloat when common code is |
| shared between readers and updaters. Additional primitives |
| are provided for this case, as discussed in lockdep.txt. |
| |
| 10. Conversely, if you are in an RCU read-side critical section, |
| and you don't hold the appropriate update-side lock, you -must- |
| use the "_rcu()" variants of the list macros. Failing to do so |
| will break Alpha, cause aggressive compilers to generate bad code, |
| and confuse people trying to read your code. |
| |
| 11. Note that synchronize_rcu() -only- guarantees to wait until |
| all currently executing rcu_read_lock()-protected RCU read-side |
| critical sections complete. It does -not- necessarily guarantee |
| that all currently running interrupts, NMIs, preempt_disable() |
| code, or idle loops will complete. Therefore, if you do not have |
| rcu_read_lock()-protected read-side critical sections, do -not- |
| use synchronize_rcu(). |
| |
| Similarly, disabling preemption is not an acceptable substitute |
| for rcu_read_lock(). Code that attempts to use preemption |
| disabling where it should be using rcu_read_lock() will break |
| in real-time kernel builds. |
| |
| If you want to wait for interrupt handlers, NMI handlers, and |
| code under the influence of preempt_disable(), you instead |
| need to use synchronize_irq() or synchronize_sched(). |
| |
| 12. Any lock acquired by an RCU callback must be acquired elsewhere |
| with softirq disabled, e.g., via spin_lock_irqsave(), |
| spin_lock_bh(), etc. Failing to disable irq on a given |
| acquisition of that lock will result in deadlock as soon as |
| the RCU softirq handler happens to run your RCU callback while |
| interrupting that acquisition's critical section. |
| |
| 13. RCU callbacks can be and are executed in parallel. In many cases, |
| the callback code simply wrappers around kfree(), so that this |
| is not an issue (or, more accurately, to the extent that it is |
| an issue, the memory-allocator locking handles it). However, |
| if the callbacks do manipulate a shared data structure, they |
| must use whatever locking or other synchronization is required |
| to safely access and/or modify that data structure. |
| |
| RCU callbacks are -usually- executed on the same CPU that executed |
| the corresponding call_rcu(), call_rcu_bh(), or call_rcu_sched(), |
| but are by -no- means guaranteed to be. For example, if a given |
| CPU goes offline while having an RCU callback pending, then that |
| RCU callback will execute on some surviving CPU. (If this was |
| not the case, a self-spawning RCU callback would prevent the |
| victim CPU from ever going offline.) |
| |
| 14. SRCU (srcu_read_lock(), srcu_read_unlock(), srcu_dereference(), |
| synchronize_srcu(), and synchronize_srcu_expedited()) may only |
| be invoked from process context. Unlike other forms of RCU, it |
| -is- permissible to block in an SRCU read-side critical section |
| (demarked by srcu_read_lock() and srcu_read_unlock()), hence the |
| "SRCU": "sleepable RCU". Please note that if you don't need |
| to sleep in read-side critical sections, you should be using |
| RCU rather than SRCU, because RCU is almost always faster and |
| easier to use than is SRCU. |
| |
| Also unlike other forms of RCU, explicit initialization |
| and cleanup is required via init_srcu_struct() and |
| cleanup_srcu_struct(). These are passed a "struct srcu_struct" |
| that defines the scope of a given SRCU domain. Once initialized, |
| the srcu_struct is passed to srcu_read_lock(), srcu_read_unlock() |
| synchronize_srcu(), and synchronize_srcu_expedited(). A given |
| synchronize_srcu() waits only for SRCU read-side critical |
| sections governed by srcu_read_lock() and srcu_read_unlock() |
| calls that have been passed the same srcu_struct. This property |
| is what makes sleeping read-side critical sections tolerable -- |
| a given subsystem delays only its own updates, not those of other |
| subsystems using SRCU. Therefore, SRCU is less prone to OOM the |
| system than RCU would be if RCU's read-side critical sections |
| were permitted to sleep. |
| |
| The ability to sleep in read-side critical sections does not |
| come for free. First, corresponding srcu_read_lock() and |
| srcu_read_unlock() calls must be passed the same srcu_struct. |
| Second, grace-period-detection overhead is amortized only |
| over those updates sharing a given srcu_struct, rather than |
| being globally amortized as they are for other forms of RCU. |
| Therefore, SRCU should be used in preference to rw_semaphore |
| only in extremely read-intensive situations, or in situations |
| requiring SRCU's read-side deadlock immunity or low read-side |
| realtime latency. |
| |
| Note that, rcu_assign_pointer() relates to SRCU just as they do |
| to other forms of RCU. |
| |
| 15. The whole point of call_rcu(), synchronize_rcu(), and friends |
| is to wait until all pre-existing readers have finished before |
| carrying out some otherwise-destructive operation. It is |
| therefore critically important to -first- remove any path |
| that readers can follow that could be affected by the |
| destructive operation, and -only- -then- invoke call_rcu(), |
| synchronize_rcu(), or friends. |
| |
| Because these primitives only wait for pre-existing readers, it |
| is the caller's responsibility to guarantee that any subsequent |
| readers will execute safely. |
| |
| 16. The various RCU read-side primitives do -not- necessarily contain |
| memory barriers. You should therefore plan for the CPU |
| and the compiler to freely reorder code into and out of RCU |
| read-side critical sections. It is the responsibility of the |
| RCU update-side primitives to deal with this. |
| |
| 17. Use CONFIG_PROVE_RCU, CONFIG_DEBUG_OBJECTS_RCU_HEAD, and |
| the __rcu sparse checks to validate your RCU code. These |
| can help find problems as follows: |
| |
| CONFIG_PROVE_RCU: check that accesses to RCU-protected data |
| structures are carried out under the proper RCU |
| read-side critical section, while holding the right |
| combination of locks, or whatever other conditions |
| are appropriate. |
| |
| CONFIG_DEBUG_OBJECTS_RCU_HEAD: check that you don't pass the |
| same object to call_rcu() (or friends) before an RCU |
| grace period has elapsed since the last time that you |
| passed that same object to call_rcu() (or friends). |
| |
| __rcu sparse checks: tag the pointer to the RCU-protected data |
| structure with __rcu, and sparse will warn you if you |
| access that pointer without the services of one of the |
| variants of rcu_dereference(). |
| |
| These debugging aids can help you find problems that are |
| otherwise extremely difficult to spot. |