diff options
author | Linus Torvalds <torvalds@linux-foundation.org> | 2014-03-31 11:05:24 -0700 |
---|---|---|
committer | Linus Torvalds <torvalds@linux-foundation.org> | 2014-03-31 11:05:24 -0700 |
commit | b3fd4ea9df2d5c39cd6ce08faf965ed669eb3b56 (patch) | |
tree | de45e2aff3106f4fa413ea3f08a7f8c969c0479b /Documentation | |
parent | 462bf234a82ae1ae9d7628f59bc81022591e1348 (diff) | |
parent | 7de700e6806cafa30c70bc84478431a11197a5ea (diff) |
Merge branch 'core-rcu-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip
Pull RCU updates from Ingo Molnar:
"Main changes:
- Torture-test changes, including refactoring of rcutorture and
introduction of a vestigial locktorture.
- Real-time latency fixes.
- Documentation updates.
- Miscellaneous fixes"
* 'core-rcu-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (77 commits)
rcu: Provide grace-period piggybacking API
rcu: Ensure kernel/rcu/rcu.h can be sourced/used stand-alone
rcu: Fix sparse warning for rcu_expedited from kernel/ksysfs.c
notifier: Substitute rcu_access_pointer() for rcu_dereference_raw()
Documentation/memory-barriers.txt: Clarify release/acquire ordering
rcutorture: Save kvm.sh output to log
rcutorture: Add a lock_busted to test the test
rcutorture: Place kvm-test-1-run.sh output into res directory
rcutorture: Rename TREE_RCU-Kconfig.txt
locktorture: Add kvm-recheck.sh plug-in for locktorture
rcutorture: Gracefully handle NULL cleanup hooks
locktorture: Add vestigial locktorture configuration
rcutorture: Introduce "rcu" directory level underneath configs
rcutorture: Rename kvm-test-1-rcu.sh
rcutorture: Remove RCU dependencies from ver_functions.sh API
rcutorture: Create CFcommon file for common Kconfig parameters
rcutorture: Create config files for scripted test-the-test testing
rcutorture: Add an rcu_busted to test the test
locktorture: Add a lock-torture kernel module
rcutorture: Abstract kvm-recheck.sh
...
Diffstat (limited to 'Documentation')
-rw-r--r-- | Documentation/RCU/RTFP.txt | 149 | ||||
-rw-r--r-- | Documentation/RCU/checklist.txt | 18 | ||||
-rw-r--r-- | Documentation/kernel-per-CPU-kthreads.txt | 13 | ||||
-rw-r--r-- | Documentation/memory-barriers.txt | 137 |
4 files changed, 249 insertions, 68 deletions
diff --git a/Documentation/RCU/RTFP.txt b/Documentation/RCU/RTFP.txt index 273e654d7d08..2f0fcb2112d2 100644 --- a/Documentation/RCU/RTFP.txt +++ b/Documentation/RCU/RTFP.txt @@ -31,6 +31,14 @@ has lapsed, so this approach may be used in non-GPL software, if desired. (In contrast, implementation of RCU is permitted only in software licensed under either GPL or LGPL. Sorry!!!) +In 1987, Rashid et al. described lazy TLB-flush [RichardRashid87a]. +At first glance, this has nothing to do with RCU, but nevertheless +this paper helped inspire the update-side batching used in the later +RCU implementation in DYNIX/ptx. In 1988, Barbara Liskov published +a description of Argus that noted that use of out-of-date values can +be tolerated in some situations. Thus, this paper provides some early +theoretical justification for use of stale data. + In 1990, Pugh [Pugh90] noted that explicitly tracking which threads were reading a given data structure permitted deferred free to operate in the presence of non-terminating threads. However, this explicit @@ -41,11 +49,11 @@ providing a fine-grained locking design, however, it would be interesting to see how much of the performance advantage reported in 1990 remains today. -At about this same time, Adams [Adams91] described ``chaotic relaxation'', -where the normal barriers between successive iterations of convergent -numerical algorithms are relaxed, so that iteration $n$ might use -data from iteration $n-1$ or even $n-2$. This introduces error, -which typically slows convergence and thus increases the number of +At about this same time, Andrews [Andrews91textbook] described ``chaotic +relaxation'', where the normal barriers between successive iterations +of convergent numerical algorithms are relaxed, so that iteration $n$ +might use data from iteration $n-1$ or even $n-2$. This introduces +error, which typically slows convergence and thus increases the number of iterations required. However, this increase is sometimes more than made up for by a reduction in the number of expensive barrier operations, which are otherwise required to synchronize the threads at the end @@ -55,7 +63,8 @@ is thus inapplicable to most data structures in operating-system kernels. In 1992, Henry (now Alexia) Massalin completed a dissertation advising parallel programmers to defer processing when feasible to simplify -synchronization. RCU makes extremely heavy use of this advice. +synchronization [HMassalinPhD]. RCU makes extremely heavy use of +this advice. In 1993, Jacobson [Jacobson93] verbally described what is perhaps the simplest deferred-free technique: simply waiting a fixed amount of time @@ -90,27 +99,29 @@ mechanism, which is quite similar to RCU [Gamsa99]. These operating systems made pervasive use of RCU in place of "existence locks", which greatly simplifies locking hierarchies and helps avoid deadlocks. -2001 saw the first RCU presentation involving Linux [McKenney01a] -at OLS. The resulting abundance of RCU patches was presented the -following year [McKenney02a], and use of RCU in dcache was first -described that same year [Linder02a]. +The year 2000 saw an email exchange that would likely have +led to yet another independent invention of something like RCU +[RustyRussell2000a,RustyRussell2000b]. Instead, 2001 saw the first +RCU presentation involving Linux [McKenney01a] at OLS. The resulting +abundance of RCU patches was presented the following year [McKenney02a], +and use of RCU in dcache was first described that same year [Linder02a]. Also in 2002, Michael [Michael02b,Michael02a] presented "hazard-pointer" techniques that defer the destruction of data structures to simplify non-blocking synchronization (wait-free synchronization, lock-free synchronization, and obstruction-free synchronization are all examples of -non-blocking synchronization). In particular, this technique eliminates -locking, reduces contention, reduces memory latency for readers, and -parallelizes pipeline stalls and memory latency for writers. However, -these techniques still impose significant read-side overhead in the -form of memory barriers. Researchers at Sun worked along similar lines -in the same timeframe [HerlihyLM02]. These techniques can be thought -of as inside-out reference counts, where the count is represented by the -number of hazard pointers referencing a given data structure rather than -the more conventional counter field within the data structure itself. -The key advantage of inside-out reference counts is that they can be -stored in immortal variables, thus allowing races between access and -deletion to be avoided. +non-blocking synchronization). The corresponding journal article appeared +in 2004 [MagedMichael04a]. This technique eliminates locking, reduces +contention, reduces memory latency for readers, and parallelizes pipeline +stalls and memory latency for writers. However, these techniques still +impose significant read-side overhead in the form of memory barriers. +Researchers at Sun worked along similar lines in the same timeframe +[HerlihyLM02]. These techniques can be thought of as inside-out reference +counts, where the count is represented by the number of hazard pointers +referencing a given data structure rather than the more conventional +counter field within the data structure itself. The key advantage +of inside-out reference counts is that they can be stored in immortal +variables, thus allowing races between access and deletion to be avoided. By the same token, RCU can be thought of as a "bulk reference count", where some form of reference counter covers all reference by a given CPU @@ -123,8 +134,10 @@ can be thought of in other terms as well. In 2003, the K42 group described how RCU could be used to create hot-pluggable implementations of operating-system functions [Appavoo03a]. -Later that year saw a paper describing an RCU implementation of System -V IPC [Arcangeli03], and an introduction to RCU in Linux Journal +Later that year saw a paper describing an RCU implementation +of System V IPC [Arcangeli03] (following up on a suggestion by +Hugh Dickins [Dickins02a] and an implementation by Mingming Cao +[MingmingCao2002IPCRCU]), and an introduction to RCU in Linux Journal [McKenney03a]. 2004 has seen a Linux-Journal article on use of RCU in dcache @@ -383,6 +396,21 @@ for Programming Languages and Operating Systems}" } } +@phdthesis{HMassalinPhD +,author="H. Massalin" +,title="Synthesis: An Efficient Implementation of Fundamental Operating +System Services" +,school="Columbia University" +,address="New York, NY" +,year="1992" +,annotation={ + Mondo optimizing compiler. + Wait-free stuff. + Good advice: defer work to avoid synchronization. See page 90 + (PDF page 106), Section 5.4, fourth bullet point. +} +} + @unpublished{Jacobson93 ,author="Van Jacobson" ,title="Avoid Read-Side Locking Via Delayed Free" @@ -671,6 +699,20 @@ Orran Krieger and Rusty Russell and Dipankar Sarma and Maneesh Soni" [Viewed October 18, 2004]" } +@conference{Michael02b +,author="Maged M. Michael" +,title="High Performance Dynamic Lock-Free Hash Tables and List-Based Sets" +,Year="2002" +,Month="August" +,booktitle="{Proceedings of the 14\textsuperscript{th} Annual ACM +Symposium on Parallel +Algorithms and Architecture}" +,pages="73-82" +,annotation={ +Like the title says... +} +} + @Conference{Linder02a ,Author="Hanna Linder and Dipankar Sarma and Maneesh Soni" ,Title="Scalability of the Directory Entry Cache" @@ -727,6 +769,24 @@ Andrea Arcangeli and Andi Kleen and Orran Krieger and Rusty Russell" } } +@conference{Michael02a +,author="Maged M. Michael" +,title="Safe Memory Reclamation for Dynamic Lock-Free Objects Using Atomic +Reads and Writes" +,Year="2002" +,Month="August" +,booktitle="{Proceedings of the 21\textsuperscript{st} Annual ACM +Symposium on Principles of Distributed Computing}" +,pages="21-30" +,annotation={ + Each thread keeps an array of pointers to items that it is + currently referencing. Sort of an inside-out garbage collection + mechanism, but one that requires the accessing code to explicitly + state its needs. Also requires read-side memory barriers on + most architectures. +} +} + @unpublished{Dickins02a ,author="Hugh Dickins" ,title="Use RCU for System-V IPC" @@ -735,6 +795,17 @@ Andrea Arcangeli and Andi Kleen and Orran Krieger and Rusty Russell" ,note="private communication" } +@InProceedings{HerlihyLM02 +,author={Maurice Herlihy and Victor Luchangco and Mark Moir} +,title="The Repeat Offender Problem: A Mechanism for Supporting Dynamic-Sized, +Lock-Free Data Structures" +,booktitle={Proceedings of 16\textsuperscript{th} International +Symposium on Distributed Computing} +,year=2002 +,month="October" +,pages="339-353" +} + @unpublished{Sarma02b ,Author="Dipankar Sarma" ,Title="Some dcache\_rcu benchmark numbers" @@ -749,6 +820,19 @@ Andrea Arcangeli and Andi Kleen and Orran Krieger and Rusty Russell" } } +@unpublished{MingmingCao2002IPCRCU +,Author="Mingming Cao" +,Title="[PATCH]updated ipc lock patch" +,month="October" +,year="2002" +,note="Available: +\url{https://lkml.org/lkml/2002/10/24/262} +[Viewed February 15, 2014]" +,annotation={ + Mingming Cao's patch to introduce RCU to SysV IPC. +} +} + @unpublished{LinusTorvalds2003a ,Author="Linus Torvalds" ,Title="Re: {[PATCH]} small fixes in brlock.h" @@ -982,6 +1066,23 @@ Realtime Applications" } } +@article{MagedMichael04a +,author="Maged M. Michael" +,title="Hazard Pointers: Safe Memory Reclamation for Lock-Free Objects" +,Year="2004" +,Month="June" +,journal="IEEE Transactions on Parallel and Distributed Systems" +,volume="15" +,number="6" +,pages="491-504" +,url="Available: +\url{http://www.research.ibm.com/people/m/michael/ieeetpds-2004.pdf} +[Viewed March 1, 2005]" +,annotation={ + New canonical hazard-pointer citation. +} +} + @phdthesis{PaulEdwardMcKenneyPhD ,author="Paul E. McKenney" ,title="Exploiting Deferred Destruction: diff --git a/Documentation/RCU/checklist.txt b/Documentation/RCU/checklist.txt index 91266193b8f4..9d10d1db16a5 100644 --- a/Documentation/RCU/checklist.txt +++ b/Documentation/RCU/checklist.txt @@ -256,10 +256,10 @@ over a rather long period of time, but improvements are always welcome! 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. + once per hour, then no explicit rate limiting is + required, unless your system is already badly broken. + Older versions of the dcache subsystem take this approach, + guarding updates with 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 @@ -268,7 +268,8 @@ over a rather long period of time, but improvements are always welcome! the machine. d. Use call_rcu_bh() rather than call_rcu(), in order to take - advantage of call_rcu_bh()'s faster grace periods. + advantage of call_rcu_bh()'s faster grace periods. (This + is only a partial solution, though.) e. Periodically invoke synchronize_rcu(), permitting a limited number of updates per grace period. @@ -276,6 +277,13 @@ over a rather long period of time, but improvements are always welcome! The same cautions apply to call_rcu_bh(), call_rcu_sched(), call_srcu(), and kfree_rcu(). + Note that although these primitives do take action to avoid memory + exhaustion when any given CPU has too many callbacks, a determined + user could still exhaust memory. This is especially the case + if a system with a large number of CPUs has been configured to + offload all of its RCU callbacks onto a single CPU, or if the + system has relatively little free memory. + 9. All RCU list-traversal primitives, which include rcu_dereference(), list_for_each_entry_rcu(), and list_for_each_safe_rcu(), must be either within an RCU read-side diff --git a/Documentation/kernel-per-CPU-kthreads.txt b/Documentation/kernel-per-CPU-kthreads.txt index 827104fb9364..f3cd299fcc41 100644 --- a/Documentation/kernel-per-CPU-kthreads.txt +++ b/Documentation/kernel-per-CPU-kthreads.txt @@ -162,7 +162,18 @@ Purpose: Execute workqueue requests To reduce its OS jitter, do any of the following: 1. Run your workload at a real-time priority, which will allow preempting the kworker daemons. -2. Do any of the following needed to avoid jitter that your +2. A given workqueue can be made visible in the sysfs filesystem + by passing the WQ_SYSFS to that workqueue's alloc_workqueue(). + Such a workqueue can be confined to a given subset of the + CPUs using the /sys/devices/virtual/workqueue/*/cpumask sysfs + files. The set of WQ_SYSFS workqueues can be displayed using + "ls sys/devices/virtual/workqueue". That said, the workqueues + maintainer would like to caution people against indiscriminately + sprinkling WQ_SYSFS across all the workqueues. The reason for + caution is that it is easy to add WQ_SYSFS, but because sysfs is + part of the formal user/kernel API, it can be nearly impossible + to remove it, even if its addition was a mistake. +3. Do any of the following needed to avoid jitter that your application cannot tolerate: a. Build your kernel with CONFIG_SLUB=y rather than CONFIG_SLAB=y, thus avoiding the slab allocator's periodic diff --git a/Documentation/memory-barriers.txt b/Documentation/memory-barriers.txt index 102dc19c4119..11c1d2049662 100644 --- a/Documentation/memory-barriers.txt +++ b/Documentation/memory-barriers.txt @@ -608,26 +608,30 @@ as follows: b = p; /* BUG: Compiler can reorder!!! */ do_something(); -The solution is again ACCESS_ONCE(), which preserves the ordering between -the load from variable 'a' and the store to variable 'b': +The solution is again ACCESS_ONCE() and barrier(), which preserves the +ordering between the load from variable 'a' and the store to variable 'b': q = ACCESS_ONCE(a); if (q) { + barrier(); ACCESS_ONCE(b) = p; do_something(); } else { + barrier(); ACCESS_ONCE(b) = p; do_something_else(); } -You could also use barrier() to prevent the compiler from moving -the stores to variable 'b', but barrier() would not prevent the -compiler from proving to itself that a==1 always, so ACCESS_ONCE() -is also needed. +The initial ACCESS_ONCE() is required to prevent the compiler from +proving the value of 'a', and the pair of barrier() invocations are +required to prevent the compiler from pulling the two identical stores +to 'b' out from the legs of the "if" statement. It is important to note that control dependencies absolutely require a a conditional. For example, the following "optimized" version of -the above example breaks ordering: +the above example breaks ordering, which is why the barrier() invocations +are absolutely required if you have identical stores in both legs of +the "if" statement: q = ACCESS_ONCE(a); ACCESS_ONCE(b) = p; /* BUG: No ordering vs. load from a!!! */ @@ -643,9 +647,11 @@ It is of course legal for the prior load to be part of the conditional, for example, as follows: if (ACCESS_ONCE(a) > 0) { + barrier(); ACCESS_ONCE(b) = q / 2; do_something(); } else { + barrier(); ACCESS_ONCE(b) = q / 3; do_something_else(); } @@ -659,9 +665,11 @@ the needed conditional. For example: q = ACCESS_ONCE(a); if (q % MAX) { + barrier(); ACCESS_ONCE(b) = p; do_something(); } else { + barrier(); ACCESS_ONCE(b) = p; do_something_else(); } @@ -723,8 +731,13 @@ In summary: use smb_rmb(), smp_wmb(), or, in the case of prior stores and later loads, smp_mb(). + (*) If both legs of the "if" statement begin with identical stores + to the same variable, a barrier() statement is required at the + beginning of each leg of the "if" statement. + (*) Control dependencies require at least one run-time conditional - between the prior load and the subsequent store. If the compiler + between the prior load and the subsequent store, and this + conditional must involve the prior load. If the compiler is able to optimize the conditional away, it will have also optimized away the ordering. Careful use of ACCESS_ONCE() can help to preserve the needed conditional. @@ -1249,6 +1262,23 @@ The ACCESS_ONCE() function can prevent any number of optimizations that, while perfectly safe in single-threaded code, can be fatal in concurrent code. Here are some examples of these sorts of optimizations: + (*) The compiler is within its rights to reorder loads and stores + to the same variable, and in some cases, the CPU is within its + rights to reorder loads to the same variable. This means that + the following code: + + a[0] = x; + a[1] = x; + + Might result in an older value of x stored in a[1] than in a[0]. + Prevent both the compiler and the CPU from doing this as follows: + + a[0] = ACCESS_ONCE(x); + a[1] = ACCESS_ONCE(x); + + In short, ACCESS_ONCE() provides cache coherence for accesses from + multiple CPUs to a single variable. + (*) The compiler is within its rights to merge successive loads from the same variable. Such merging can cause the compiler to "optimize" the following code: @@ -1644,12 +1674,12 @@ for each construct. These operations all imply certain barriers: Memory operations issued after the ACQUIRE will be completed after the ACQUIRE operation has completed. - Memory operations issued before the ACQUIRE may be completed after the - ACQUIRE operation has completed. An smp_mb__before_spinlock(), combined - with a following ACQUIRE, orders prior loads against subsequent stores and - stores and prior stores against subsequent stores. Note that this is - weaker than smp_mb()! The smp_mb__before_spinlock() primitive is free on - many architectures. + Memory operations issued before the ACQUIRE may be completed after + the ACQUIRE operation has completed. An smp_mb__before_spinlock(), + combined with a following ACQUIRE, orders prior loads against + subsequent loads and stores and also orders prior stores against + subsequent stores. Note that this is weaker than smp_mb()! The + smp_mb__before_spinlock() primitive is free on many architectures. (2) RELEASE operation implication: @@ -1694,24 +1724,21 @@ may occur as: ACQUIRE M, STORE *B, STORE *A, RELEASE M -This same reordering can of course occur if the lock's ACQUIRE and RELEASE are -to the same lock variable, but only from the perspective of another CPU not -holding that lock. - -In short, a RELEASE followed by an ACQUIRE may -not- be assumed to be a full -memory barrier because it is possible for a preceding RELEASE to pass a -later ACQUIRE from the viewpoint of the CPU, but not from the viewpoint -of the compiler. Note that deadlocks cannot be introduced by this -interchange because if such a deadlock threatened, the RELEASE would -simply complete. - -If it is necessary for a RELEASE-ACQUIRE pair to produce a full barrier, the -ACQUIRE can be followed by an smp_mb__after_unlock_lock() invocation. This -will produce a full barrier if either (a) the RELEASE and the ACQUIRE are -executed by the same CPU or task, or (b) the RELEASE and ACQUIRE act on the -same variable. The smp_mb__after_unlock_lock() primitive is free on many -architectures. Without smp_mb__after_unlock_lock(), the critical sections -corresponding to the RELEASE and the ACQUIRE can cross: +When the ACQUIRE and RELEASE are a lock acquisition and release, +respectively, this same reordering can occur if the lock's ACQUIRE and +RELEASE are to the same lock variable, but only from the perspective of +another CPU not holding that lock. In short, a ACQUIRE followed by an +RELEASE may -not- be assumed to be a full memory barrier. + +Similarly, the reverse case of a RELEASE followed by an ACQUIRE does not +imply a full memory barrier. If it is necessary for a RELEASE-ACQUIRE +pair to produce a full barrier, the ACQUIRE can be followed by an +smp_mb__after_unlock_lock() invocation. This will produce a full barrier +if either (a) the RELEASE and the ACQUIRE are executed by the same +CPU or task, or (b) the RELEASE and ACQUIRE act on the same variable. +The smp_mb__after_unlock_lock() primitive is free on many architectures. +Without smp_mb__after_unlock_lock(), the CPU's execution of the critical +sections corresponding to the RELEASE and the ACQUIRE can cross, so that: *A = a; RELEASE M @@ -1722,7 +1749,36 @@ could occur as: ACQUIRE N, STORE *B, STORE *A, RELEASE M -With smp_mb__after_unlock_lock(), they cannot, so that: +It might appear that this reordering could introduce a deadlock. +However, this cannot happen because if such a deadlock threatened, +the RELEASE would simply complete, thereby avoiding the deadlock. + + Why does this work? + + One key point is that we are only talking about the CPU doing + the reordering, not the compiler. If the compiler (or, for + that matter, the developer) switched the operations, deadlock + -could- occur. + + But suppose the CPU reordered the operations. In this case, + the unlock precedes the lock in the assembly code. The CPU + simply elected to try executing the later lock operation first. + If there is a deadlock, this lock operation will simply spin (or + try to sleep, but more on that later). The CPU will eventually + execute the unlock operation (which preceded the lock operation + in the assembly code), which will unravel the potential deadlock, + allowing the lock operation to succeed. + + But what if the lock is a sleeplock? In that case, the code will + try to enter the scheduler, where it will eventually encounter + a memory barrier, which will force the earlier unlock operation + to complete, again unraveling the deadlock. There might be + a sleep-unlock race, but the locking primitive needs to resolve + such races properly in any case. + +With smp_mb__after_unlock_lock(), the two critical sections cannot overlap. +For example, with the following code, the store to *A will always be +seen by other CPUs before the store to *B: *A = a; RELEASE M @@ -1730,13 +1786,18 @@ With smp_mb__after_unlock_lock(), they cannot, so that: smp_mb__after_unlock_lock(); *B = b; -will always occur as either of the following: +The operations will always occur in one of the following orders: - STORE *A, RELEASE, ACQUIRE, STORE *B - STORE *A, ACQUIRE, RELEASE, STORE *B + STORE *A, RELEASE, ACQUIRE, smp_mb__after_unlock_lock(), STORE *B + STORE *A, ACQUIRE, RELEASE, smp_mb__after_unlock_lock(), STORE *B + ACQUIRE, STORE *A, RELEASE, smp_mb__after_unlock_lock(), STORE *B If the RELEASE and ACQUIRE were instead both operating on the same lock -variable, only the first of these two alternatives can occur. +variable, only the first of these alternatives can occur. In addition, +the more strongly ordered systems may rule out some of the above orders. +But in any case, as noted earlier, the smp_mb__after_unlock_lock() +ensures that the store to *A will always be seen as happening before +the store to *B. Locks and semaphores may not provide any guarantee of ordering on UP compiled systems, and so cannot be counted on in such a situation to actually achieve @@ -2757,7 +2818,7 @@ in that order, but, without intervention, the sequence may have almost any combination of elements combined or discarded, provided the program's view of the world remains consistent. Note that ACCESS_ONCE() is -not- optional in the above example, as there are architectures where a given CPU might -interchange successive loads to the same location. On such architectures, +reorder successive loads to the same location. On such architectures, ACCESS_ONCE() does whatever is necessary to prevent this, for example, on Itanium the volatile casts used by ACCESS_ONCE() cause GCC to emit the special ld.acq and st.rel instructions that prevent such reordering. |