introduce sys_membarrier(): process-wide memory barrier (v7)

Here is an implementation of a new system call, sys_membarrier(), which executes a memory barrier on all threads of the current process.

It aims at greatly simplifying and enhancing the current signal-based liburcu userspace RCU synchronize_rcu() implementation. (found at http://lttng.org/urcu)

Changelog since v1:

- Only perform the IPI in CONFIG_SMP.
- Only perform the IPI if the process has more than one thread.
- Only send IPIs to CPUs involved with threads belonging to our process.
- Adaptative IPI scheme (single vs many IPI with threshold).
- Issue smp_mb() at the beginning and end of the system call.

Changelog since v2:
- simply send-to-many to the mm_cpumask. It contains the list of processors we have to IPI to (which use the mm), and this mask is updated atomically.

Changelog since v3a:
- Confirm that each CPU indeed runs the current task's ->mm before sending an IPI. Ensures that we do not disturb RT tasks in the presence of lazy TLB shootdown.
- Document memory barriers needed in switch_mm().
- Surround helper functions with #ifdef CONFIG_SMP.

Changelog since v4:
- Add "int expedited" parameter, use synchronize_sched() in the non-expedited case. Thanks to Lai Jiangshan for making us consider seriously using synchronize_sched() to provide the low-overhead membarrier scheme.
- Check num_online_cpus() == 1, quickly return without doing nothing.

Changelog since v5:
- Plan ahead for extensibility by introducing mandatory/optional masks to the "flags" system call parameter. Past experience with accept4(), signalfd4(), eventfd2(), epoll_create1(), dup3(), pipe2(), and inotify_init1() indicates that this is the kind of thing we want to plan for. Return -EINVAL if the mandatory flags received are unknown.
- Create include/linux/membarrier.h to define these flags.
- Add MEMBARRIER_QUERY optional flag.

Changelog since v6:
- Remove some unlikely() not so unlikely.
- Add the proper scheduler memory barriers needed to only use the RCU read lock in sys_membarrier rather than take each runqueue spinlock:
- Move memory barriers from per-architecture switch_mm() to schedule() and finish_lock_switch(), where they clearly document that all data protected by the rq lock is guaranteed to have memory barriers issued between the scheduler update and the task execution. Replacing the spin lock acquire/release barriers with these memory barriers imply either no overhead (x86 spinlock atomic instruction already implies a full mb) or some hopefully small overhead caused by the upgrade of the spinlock acquire/release barriers to more heavyweight smp_mb().
- The "generic" version of spinlock-mb.h declares both a mapping to standard spinlocks and full memory barriers. Each architecture can specialize this header following their own need and declare CONFIG_HAVE_SPINLOCK_MB to use their own spinlock-mb.h.
- Note: benchmarks of scheduler overhead with specialized spinlock-mb.h implementations on a wide range of architecture would be welcome.

Both the signal-based and the sys_membarrier userspace RCU schemes permit us to remove the memory barrier from the userspace RCU rcu_read_lock() and rcu_read_unlock() primitives, thus significantly accelerating them. These memory barriers are replaced by compiler barriers on the read-side, and all matching memory barriers on the write-side are turned into an invokation of a memory barrier on all active threads in the process. By letting the kernel perform this synchronization rather than dumbly sending a signal to every process threads (as we currently do), we diminish the number of unnecessary wake ups and only issue the memory barriers on active threads. Non-running threads do not need to execute such barrier anyway, because these are implied by the scheduler context switches.

To explain the benefit of this scheme, let's introduce two example threads:

Thread A (non-frequent, e.g. executing liburcu synchronize_rcu()) Thread B (frequent, e.g. executing liburcu rcu_read_lock()/rcu_read_unlock())

In a scheme where all smp_mb() in thread A synchronize_rcu() are ordering memory accesses with respect to smp_mb() present in rcu_read_lock/unlock(), we can change all smp_mb() from synchronize_rcu() into calls to sys_membarrier() and all smp_mb() from rcu_read_lock/unlock() into compiler barriers "barrier()".

Before the change, we had, for each smp_mb() pairs:

Thread A                    Thread B
prev mem accesses           prev mem accesses
smp_mb()                    smp_mb()
follow mem accesses         follow mem accesses

After the change, these pairs become:

Thread A                    Thread B
prev mem accesses           prev mem accesses
sys_membarrier()            barrier()
follow mem accesses         follow mem accesses

As we can see, there are two possible scenarios: either Thread B memory accesses do not happen concurrently with Thread A accesses (1), or they do (2).

1) Non-concurrent Thread A vs Thread B accesses:

Thread A Thread B
prev mem accesses
sys_membarrier()
follow mem accesses
prev mem accesses
barrier()
follow mem accesses

In this case, thread B accesses will be weakly ordered. This is OK, because at that point, thread A is not particularly interested in ordering them with respect to its own accesses.

2) Concurrent Thread A vs Thread B accesses

Thread A                    Thread B
prev mem accesses           prev mem accesses
sys_membarrier()            barrier()
follow mem accesses         follow mem accesses

In this case, thread B accesses, which are ensured to be in program order thanks to the compiler barrier, will be "upgraded" to full smp_mb() thanks to the IPIs executing memory barriers on each active system threads. Each non-running process threads are intrinsically serialized by the scheduler.

For my Intel Xeon E5405
(one thread is doing the sys_membarrier, the others are busy looping)

* expedited

10,000,000 sys_membarrier calls:

T=1: 0m20.173s
T=2: 0m20.506s
T=3: 0m22.632s
T=4: 0m24.759s
T=5: 0m26.633s
T=6: 0m29.654s
T=7: 0m30.669s

For a 2-3 microseconds/call.

* non-expedited

1000 sys_membarrier calls:

T=1-7: 0m16.002s

For a 16 milliseconds/call. (~5000-8000 times slower than expedited)

The expected top pattern for the expedited scheme, when using 1 CPU for a thread doing sys_membarrier() in a loop and other 7 threads busy-waiting in user-space on a variable shows that the thread doing sys_membarrier is doing mostly system calls, and other threads are mostly running in user-space. Note that IPI handlers are not taken into account in the cpu time sampling.

Cpu0  :100.0%us,  0.0%sy,  0.0%ni,  0.0%id,  0.0%wa,  0.0%hi,  0.0%si,  0.0%st
Cpu1  : 99.7%us,  0.0%sy,  0.0%ni,  0.0%id,  0.0%wa,  0.3%hi,  0.0%si,  0.0%st
Cpu2  : 99.3%us,  0.0%sy,  0.0%ni,  0.0%id,  0.0%wa,  0.7%hi,  0.0%si,  0.0%st
Cpu3  :100.0%us,  0.0%sy,  0.0%ni,  0.0%id,  0.0%wa,  0.0%hi,  0.0%si,  0.0%st
Cpu4  :100.0%us,  0.0%sy,  0.0%ni,  0.0%id,  0.0%wa,  0.0%hi,  0.0%si,  0.0%st
Cpu5  : 96.0%us,  1.3%sy,  0.0%ni,  0.0%id,  0.0%wa,  0.0%hi,  2.6%si,  0.0%st
Cpu6  :  1.3%us, 98.7%sy,  0.0%ni,  0.0%id,  0.0%wa,  0.0%hi,  0.0%si,  0.0%st
Cpu7  : 96.1%us,  3.3%sy,  0.0%ni,  0.0%id,  0.0%wa,  0.3%hi,  0.3%si,  0.0%st

Results in liburcu:

Operations in 10s, 6 readers, 2 writers:

(what we previously had)
memory barriers in reader: 973494744 reads, 892368 writes
signal-based scheme: 6289946025 reads, 1251 writes

(what we have now, with dynamic sys_membarrier check, expedited scheme) memory barriers in reader: 907693804 reads, 817793 writes
sys_membarrier scheme: 4316818891 reads, 503790 writes

(dynamic sys_membarrier check, non-expedited scheme) memory barriers in reader: 907693804 reads, 817793 writes
sys_membarrier scheme: 8698725501 reads, 313 writes

So the dynamic sys_membarrier availability check adds some overhead to the read-side, but besides that, with the expedited scheme, we can see that we are close to the read-side performance of the signal-based scheme and also close (5/8) to the performance of the memory-barrier write-side. We have a write-side speedup of 400:1 over the signal-based scheme by using the sys_membarrier system call. This allows a 4.5:1 read-side speedup over the memory barrier scheme.

The non-expedited scheme adds indeed a much lower overhead on the read-side both because we do not send IPIs and because we perform less updates, which in turn generates less cache-line exchanges. The write-side latency becomes even higher than with the signal-based scheme. The advantage of the non-expedited sys_membarrier() scheme over signal-based scheme is that it does not require to wake up all the process threads.

The system call number is only assigned for x86_64 in this RFC patch.

---
 arch/x86/Kconfig                   |    1 
 arch/x86/include/asm/spinlock-mb.h |   22 ++++
 arch/x86/include/asm/unistd_64.h   |    2 
 include/asm-generic/spinlock-mb.h  |   21 ++++
 include/linux/Kbuild               |    1 
 include/linux/membarrier.h         |   25 +++++
 include/linux/spinlock.h           |    6 +
 init/Kconfig                       |    3 
 kernel/sched.c                     |  173 +++++++++++++++++++++++++++++++++++--
 9 files changed, 249 insertions(+), 5 deletions(-)

Index: linux-2.6-lttng/arch/x86/include/asm/unistd_64.h
===================================================================
@@ -661,6 +661,8 @@ __SYSCALL(__NR_pwritev, sys_pwritev)
 __SYSCALL(__NR_rt_tgsigqueueinfo, sys_rt_tgsigqueueinfo)
 #define __NR_perf_event_open			298
 __SYSCALL(__NR_perf_event_open, sys_perf_event_open)
+#define __NR_membarrier				299
+__SYSCALL(__NR_membarrier, sys_membarrier)
 
 #ifndef __NO_STUBS
 #define __ARCH_WANT_OLD_READDIR

Index: linux-2.6-lttng/include/linux/membarrier.h
===================================================================
@@ -0,0 +1,25 @@
+#ifndef _LINUX_MEMBARRIER_H
+#define _LINUX_MEMBARRIER_H
+
+/* First argument to membarrier syscall */
+
+/*
+ * Mandatory flags to the membarrier system call that the kernel must
+ * understand are in the low 16 bits.
+ */
+#define MEMBARRIER_MANDATORY_MASK	0x0000FFFF	/* Mandatory flags */
+
+/*
+ * Optional hints that the kernel can ignore are in the high 16 bits.
+ */
+#define MEMBARRIER_OPTIONAL_MASK	0xFFFF0000	/* Optional hints */
+
+/* Expedited: adds some overhead, fast execution (few microseconds) */
+#define MEMBARRIER_EXPEDITED		(1 << 0)
+/* Delayed: Low overhead, but slow execution (few milliseconds) */
+#define MEMBARRIER_DELAYED		(1 << 1)
+
+/* Query flag support, without performing synchronization */
+#define MEMBARRIER_QUERY		(1 << 16)
+
+#endif

Index: linux-2.6-lttng/include/linux/Kbuild
===================================================================
@@ -110,6 +110,7 @@ header-y += magic.h
 header-y += major.h
 header-y += map_to_7segment.h
 header-y += matroxfb.h
+header-y += membarrier.h
 header-y += meye.h
 header-y += minix_fs.h
 header-y += mmtimer.h

Index: linux-2.6-lttng/include/asm-generic/spinlock-mb.h
===================================================================
@@ -0,0 +1,21 @@
+#ifndef ASM_GENERIC_SPINLOCK_MB_H
+#define ASM_GENERIC_SPINLOCK_MB_H
+
+/*
+ * Generic spinlock-mb mappings. Use standard spinlocks with acquire/release
+ * semantics, and define the associated memory barriers as full memory barriers.
+ */
+
+#define spin_lock__no_acquire			spin_lock
+#define spin_unlock__no_release			spin_unlock
+
+#define spin_lock_irq__no_acquire		spin_lock_irq
+#define spin_unlock_irq__no_release		spin_unlock_irq
+
+#define smp_acquire__after_spin_lock()		do { } while (0)
+#define smp_release__before_spin_unlock()	do { } while (0)
+
+#define smp_mb__after_spin_lock()		smp_mb()
+#define smp_mb__before_spin_unlock()		smp_mb()
+
+#endif /* ASM_GENERIC_SPINLOCK_MB_H */

Index: linux-2.6-lttng/include/linux/spinlock.h
===================================================================
@@ -346,4 +346,10 @@ extern int _atomic_dec_and_lock(atomic_t
 # include <linux/spinlock_api_up.h>
 #endif
 
+#ifdef CONFIG_HAVE_SPINLOCK_MB
+# include <asm/spinlock-mb.h>
+#else
+# include <asm-generic/spinlock-mb.h>
+#endif
+
 #endif /* __LINUX_SPINLOCK_H */

Index: linux-2.6-lttng/arch/x86/include/asm/spinlock-mb.h
===================================================================
@@ -0,0 +1,22 @@
+#ifndef ASM_X86_SPINLOCK_MB_H
+#define ASM_X86_SPINLOCK_MB_H
+
+/*
+ * X86 spinlock-mb mappings. Use standard spinlocks with acquire/release
+ * semantics. Associated memory barriers are defined as no-ops, because the
+ * spinlock LOCK-prefixed atomic operations imply a full memory barrier.
+ */
+
+#define spin_lock__no_acquire			spin_lock
+#define spin_unlock__no_release			spin_unlock
+
+#define spin_lock_irq__no_acquire		spin_lock_irq
+#define spin_unlock_irq__no_release		spin_unlock_irq
+
+#define smp_acquire__after_spin_lock()		do { } while (0)
+#define smp_release__before_spin_unlock()	do { } while (0)
+
+#define smp_mb__after_spin_lock()		do { } while (0)
+#define smp_mb__before_spin_unlock()		do { } while (0)
+
+#endif /* ASM_X86_SPINLOCK_MB_H */

Index: linux-2.6-lttng/arch/x86/Kconfig
===================================================================
@@ -55,6 +55,7 @@ config X86
 	select HAVE_KERNEL_BZIP2
 	select HAVE_KERNEL_LZMA
 	select HAVE_ARCH_KMEMCHECK
+	select HAVE_SPINLOCK_MB
 
 config OUTPUT_FORMAT
 	string

Index: linux-2.6-lttng/init/Kconfig
===================================================================
@@ -310,6 +310,9 @@ config AUDIT_TREE
 	depends on AUDITSYSCALL
 	select INOTIFY
 
+config HAVE_SPINLOCK_MB
+	def_bool n
+
 menu "RCU Subsystem"
 
 choice

Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68