Implement FlushProcessWriteBuffers in PAL #3952

sergiy-k · 2015-02-11T05:05:26Z

Currently we have an empty stub in coreclr/src/pal/src/thread/process.cpp.

This function is called by GC as part of synchronization process between GC and managed threads. We need to provide proper implementation for Linux and Mac and add a test under coreclr/src/pal/tests/palsuite.

kangaroo · 2015-02-15T02:09:40Z

I'm not really sure what to do here. AFAICT there really isn't an analogue on Linux or MAC. We probably need to drop this function, and audit the call sites to see what they really need (note, I haven't audited the callsites yet)

Is this just a matter of appropriate fencing?
Does it actually need cache invalidation to ensure coherency across clusters?
Is this called due to weak memory ordering on ARM?
something else?

jkotas · 2015-02-15T06:01:20Z

The crucial callsite is in https://github.com/dotnet/coreclr/blob/master/src/vm/threadsuspend.cpp. The resolution for this issue may dependent on design for #3947.

This API does not need to invalidate cache. It just needs to flush pending writes (and invalidate speculative reads) on all cores. It is needed for correctness on all common platforms, including x86/x64.

Conceptually, this API makes every core in the system execute memory barrier. Windows kernel implements it by sending Inter-processor interrupt (http://en.wikipedia.org/wiki/Inter-processor_interrupt) to all cores in the system.

janvorli · 2015-05-15T21:09:27Z

I have spent some time thinking about how to implement the FlushProcessWriteBuffers functionality on Linux, OSX and FreeBSD.
I have also considered a possibility to remove the need for this function on non-Windows systems by adding some fences, but it turns out there are other places where fences would cause performance degradation and it would also be quite error prone due to the complexity of the codes and possibly many places that would have to be touched. See this comment in gc.cpp right before a FlushProcessWriteBuffers call :

// We need to make sure that other threads executing checked write barriers
// will see the g_card_table update before g_lowest/highest_address updates.
// Otherwise, the checked write barrier may AV accessing the old card table
// with address that it does not cover. Write barriers access card table
// without memory barriers for performance reasons, so we need to flush
// the store buffers here.

Here are possible solutions that I've figured. I currently tend to the second solution.

1. Helper threads

Have a set of helper threads, one per each CPU on the machine, each affinitized to one of the CPUs. All the threads would be waiting on an event. The FlushProcessWriteBuffers would set a global counter to the number of CPUs - 1 (to exclude itself) and signal the event.The helper threads would wake up on the event, decrement using interlocked operation the global counter (which both flushes the buffers and ensures atomicity) and go to wait on the event again.
The thread running the FlushProcessWriteBuffers would wait until the global counter reaches zero, indicating that all the threads have ran.
It seems there is an issue on OSX though. Its affinity APIs don't allow specifying direct affinity to specific CPU. It only allows specifying affinity groups and if there are more threads then the number of processors, I'm not sure if the OS will honor the even spreading, which is required for the functionality to work.

2. Signals

Use signals on Linux/FreeBSD and similar functionality on OSX. The FlushProcessWriteBuffers would first call the PROCProcessLock to ensure the list of PAL threads won't change under our hands. Then we would iterate over the list of threads and consider only runnable / running threads since the others don't have anything to flush. In fact, we really care only about running threads, but on Linux, there is no way to distinguish between those two states. It seems that on OSX and FreeBSD, we can distinguish between them, which will make it simpler.
First, we zero a global counter. Then as we scan the list of threads, we would send a realtime signal to the first thread we find running on each CPU. Before sending the signal, we would atomically increment the global counter.
On Linux, it gets a bit more complicated due to the fact that we cannot distinguish between running / runnable threads. So we set the affinity of the thread before we send it a signal to the CPU it was running on and reset the affinity after all signals are processed or right at the end of the signal handler.
The signal handler would just atomically decrement the global counter and return. The FlushProcessWriteBuffers would finish iterating over the threads and then wait until the global counter reaches zero, indicating that all the signals were processed.
Finally, we would release the process lock.

You may wonder if it is ok to rely on reading the thread's state and its CPU when they can change anytime, even during the thread enumeration. The reason why it is ok is that we care about flushing buffers that were on each CPU at the time the FlushProcessWriteBuffers was called. So we have the following possible cases:

A thread was running on a processor, but before we checked its state, it got into waiting state. This is ok, since the fact that it went into the waiting state means its buffers were flushed.
A thread was waiting, but before we checked its state, it started to run. We will send a signal unnecessarily here.
A thread was waiting, but before we checked its state, it became runnable. On Linux, we will send a signal unnecessarily here. On OSX / FreeBSD, we won't.
A thread was runnable, but before we checked its state, it started to run. That's the same as 2.
A thread was running, but before we checked its state, it became runnable. That's the same as 1. on OSX / FreeBSD and same as 2. on Linux.
So there are few cases when we send a signal unnecessarily, but the overhead should be really low since it happens only when thread state changes.

@kangaroo I have one unknown here though. I wasn't able to find a way to get a CPU a thread is running on yet for OSX. Do you know how we can get it? If there is no way, then it would still work, but we would send more signals in case when a thread 1 was running on CPU1 and while we were iterating the thread list, another thread started to run on CPU1 instead.

jkotas · 2015-05-16T07:32:41Z

Before FlushProcessWriteBuffers was added as Windows API, we used to simulate it by flipping protection on a page.

We are not the only ones solving this problem. There is a long discussion about it in the context of liburcu library: http://www.bing.com/search?q=liburcu%2Bsys_membarrier. This discussion suggests that patch to introduce sys_membarrier syscall to Linux kernel is going to be accepted soon. @compudj - thanks a lot for making it happen!

I think we should simulate the FlushProcessWriteBuffers using the page protection flip for now - it is fragile, but it has the best performance characteristics among the various options and it should be reliable on all current system. And switch over to sys_membarrier() once it becomes available in the Linux kernel, and hopefully other OSes will eventually follow the suit too.

janvorli · 2015-05-27T20:41:51Z

Implemented in dotnet/coreclr#1057 using the page protection changes.

Here is an implementation of a new system call, sys_membarrier(), which executes a memory barrier on all threads running on the system. It is implemented by calling synchronize_sched(). It can be used to distribute the cost of user-space memory barriers asymmetrically by transforming pairs of memory barriers into pairs consisting of sys_membarrier() and a compiler barrier. For synchronization primitives that distinguish between read-side and write-side (e.g. userspace RCU [1], rwlocks), the read-side can be accelerated significantly by moving the bulk of the memory barrier overhead to the write-side. The existing applications of which I am aware that would be improved by this system call are as follows: * Through Userspace RCU library (http://urcu.so) - DNS server (Knot DNS) https://www.knot-dns.cz/ - Network sniffer (http://netsniff-ng.org/) - Distributed object storage (https://sheepdog.github.io/sheepdog/) - User-space tracing (http://lttng.org) - Network storage system (https://www.gluster.org/) - Virtual routers (https://events.linuxfoundation.org/sites/events/files/slides/DPDK_RCU_0MQ.pdf) - Financial software (https://lkml.org/lkml/2015/3/23/189) Those projects use RCU in userspace to increase read-side speed and scalability compared to locking. Especially in the case of RCU used by libraries, sys_membarrier can speed up the read-side by moving the bulk of the memory barrier cost to synchronize_rcu(). * Direct users of sys_membarrier - core dotnet garbage collector (https://github.com/dotnet/coreclr/issues/198) Microsoft core dotnet GC developers are planning to use the mprotect() side-effect of issuing memory barriers through IPIs as a way to implement Windows FlushProcessWriteBuffers() on Linux. They are referring to sys_membarrier in their github thread, specifically stating that sys_membarrier() is what they are looking for. This implementation is based on kernel v4.1-rc2. 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 are ordering memory accesses with respect to smp_mb() present in Thread B, we can change each smp_mb() within Thread A into calls to sys_membarrier() and each smp_mb() within Thread B into compiler barriers "barrier()". Before the change, we had, for each smp_mb() pairs: Thread A Thread B previous mem accesses previous mem accesses smp_mb() smp_mb() following mem accesses following 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() by synchronize_sched(). * Benchmarks On Intel Xeon E5405 (8 cores) (one thread is calling sys_membarrier, the other 7 threads are busy looping) 1000 non-expedited sys_membarrier calls in 33s = 33 milliseconds/call. * User-space user of this system call: Userspace RCU library 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 invocation 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. Results in liburcu: Operations in 10s, 6 readers, 2 writers: memory barriers in reader: 1701557485 reads, 2202847 writes signal-based scheme: 9830061167 reads, 6700 writes sys_membarrier: 9952759104 reads, 425 writes sys_membarrier (dyn. check): 7970328887 reads, 425 writes The dynamic sys_membarrier availability check adds some overhead to the read-side compared to the signal-based scheme, but besides that, sys_membarrier slightly outperforms the signal-based scheme. However, this non-expedited sys_membarrier implementation has a much slower grace period than signal and memory barrier schemes. Besides diminishing the number of wake-ups, one major advantage of the membarrier system call over the signal-based scheme is that it does not need to reserve a signal. This plays much more nicely with libraries, and with processes injected into for tracing purposes, for which we cannot expect that signals will be unused by the application. An expedited version of this system call can be added later on to speed up the grace period. Its implementation will likely depend on reading the cpu_curr()->mm without holding each CPU's rq lock. This patch adds the system call to x86 and to asm-generic. [1] http://urcu.so Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org> CC: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> CC: Steven Rostedt <rostedt@goodmis.org> CC: Nicholas Miell <nmiell@comcast.net> CC: Linus Torvalds <torvalds@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com> CC: Alan Cox <gnomes@lxorguk.ukuu.org.uk> CC: Lai Jiangshan <laijs@cn.fujitsu.com> CC: Stephen Hemminger <stephen@networkplumber.org> CC: Andrew Morton <akpm@linux-foundation.org> CC: Thomas Gleixner <tglx@linutronix.de> CC: Peter Zijlstra <peterz@infradead.org> CC: David Howells <dhowells@redhat.com> CC: Pranith Kumar <bobby.prani@gmail.com> CC: Michael Kerrisk <mtk.manpages@gmail.com> CC: linux-api@vger.kernel.org --- membarrier(2) man page: --------------- snip ------------------- MEMBARRIER(2) Linux Programmer's Manual MEMBARRIER(2) NAME membarrier - issue memory barriers on a set of threads SYNOPSIS #include <linux/membarrier.h> int membarrier(int cmd, int flags); DESCRIPTION The cmd argument is one of the following: MEMBARRIER_CMD_QUERY Query the set of supported commands. It returns a bitmask of supported commands. MEMBARRIER_CMD_SHARED Execute a memory barrier on all threads running on the system. Upon return from system call, the caller thread is ensured that all running threads have passed through a state where all memory accesses to user-space addresses match program order between entry to and return from the system call (non-running threads are de facto in such a state). This covers threads from all pro‐ cesses running on the system. This command returns 0. The flags argument needs to be 0. For future extensions. All memory accesses performed in program order from each targeted thread is guaranteed to be ordered with respect to sys_membarrier(). If we use the semantic "barrier()" to represent a compiler barrier forcing memory accesses to be performed in program order across the barrier, and smp_mb() to represent explicit memory barriers forcing full memory ordering across the barrier, we have the following ordering table for each pair of barrier(), sys_membarrier() and smp_mb(): The pair ordering is detailed as (O: ordered, X: not ordered): barrier() smp_mb() sys_membarrier() barrier() X X O smp_mb() X O O sys_membarrier() O O O RETURN VALUE On success, these system calls return zero. On error, -1 is returned, and errno is set appropriately. For a given command, with flags argument set to 0, this system call is guaranteed to always return the same value until reboot. ERRORS ENOSYS System call is not implemented. EINVAL Invalid arguments. Linux 2015-04-15 MEMBARRIER(2) --------------- snip ------------------- Changes since v18: - Add unlikely() check to flags, - Describe current users in changelog. Changes since v17: - Update commit message. Changes since v16: - Update documentation. - Add man page to changelog. - Build sys_membarrier on !CONFIG_SMP. It allows userspace applications to not care about the number of processors on the system. Based on recommendations from Stephen Hemminger and Steven Rostedt. - Check that flags argument is 0, update documentation to require it. Changes since v15: - Add flags argument in addition to cmd. - Update documentation. Changes since v14: - Take care of Thomas Gleixner's comments. Changes since v13: - Move to kernel/membarrier.c. - Remove MEMBARRIER_PRIVATE flag. - Add MAINTAINERS file entry. Changes since v12: - Remove _FLAG suffix from uapi flags. - Add Expert menuconfig option CONFIG_MEMBARRIER (default=y). - Remove EXPEDITED mode. Only implement non-expedited for now, until reading the cpu_curr()->mm can be done without holding the CPU's rq lock. Changes since v11: - 5 years have passed. - Rebase on v3.19 kernel. - Add futex-alike PRIVATE vs SHARED semantic: private for per-process barriers, non-private for memory mappings shared between processes. - Simplify user API. - Code refactoring. Changes since v10: - Apply Randy's comments. - Rebase on 2.6.34-rc4 -tip. Changes since v9: - Clean up #ifdef CONFIG_SMP. Changes since v8: - Go back to rq spin locks taken by sys_membarrier() rather than adding memory barriers to the scheduler. It implies a potential RoS (reduction of service) if sys_membarrier() is executed in a busy-loop by a user, but nothing more than what is already possible with other existing system calls, but saves memory barriers in the scheduler fast path. - re-add the memory barrier comments to x86 switch_mm() as an example to other architectures. - Update documentation of the memory barriers in sys_membarrier and switch_mm(). - Append execution scenarios to the changelog showing the purpose of each memory barrier. Changes since v7: - Move spinlock-mb and scheduler related changes to separate patches. - Add support for sys_membarrier on x86_32. - Only x86 32/64 system calls are reserved in this patch. It is planned to incrementally reserve syscall IDs on other architectures as these are tested. Changes 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. Changes 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. Changes 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. Changes 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. Changes 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. Changes 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.

Here is an implementation of a new system call, sys_membarrier(), which executes a memory barrier on all threads running on the system. It is implemented by calling synchronize_sched(). It can be used to distribute the cost of user-space memory barriers asymmetrically by transforming pairs of memory barriers into pairs consisting of sys_membarrier() and a compiler barrier. For synchronization primitives that distinguish between read-side and write-side (e.g. userspace RCU [1], rwlocks), the read-side can be accelerated significantly by moving the bulk of the memory barrier overhead to the write-side. The existing applications of which I am aware that would be improved by this system call are as follows: * Through Userspace RCU library (http://urcu.so) - DNS server (Knot DNS) https://www.knot-dns.cz/ - Network sniffer (http://netsniff-ng.org/) - Distributed object storage (https://sheepdog.github.io/sheepdog/) - User-space tracing (http://lttng.org) - Network storage system (https://www.gluster.org/) - Virtual routers (https://events.linuxfoundation.org/sites/events/files/slides/DPDK_RCU_0MQ.pdf) - Financial software (https://lkml.org/lkml/2015/3/23/189) Those projects use RCU in userspace to increase read-side speed and scalability compared to locking. Especially in the case of RCU used by libraries, sys_membarrier can speed up the read-side by moving the bulk of the memory barrier cost to synchronize_rcu(). * Direct users of sys_membarrier - core dotnet garbage collector (https://github.com/dotnet/coreclr/issues/198) Microsoft core dotnet GC developers are planning to use the mprotect() side-effect of issuing memory barriers through IPIs as a way to implement Windows FlushProcessWriteBuffers() on Linux. They are referring to sys_membarrier in their github thread, specifically stating that sys_membarrier() is what they are looking for. This implementation is based on kernel v4.1-rc8. 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 are ordering memory accesses with respect to smp_mb() present in Thread B, we can change each smp_mb() within Thread A into calls to sys_membarrier() and each smp_mb() within Thread B into compiler barriers "barrier()". Before the change, we had, for each smp_mb() pairs: Thread A Thread B previous mem accesses previous mem accesses smp_mb() smp_mb() following mem accesses following 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() by synchronize_sched(). * Benchmarks On Intel Xeon E5405 (8 cores) (one thread is calling sys_membarrier, the other 7 threads are busy looping) 1000 non-expedited sys_membarrier calls in 33s = 33 milliseconds/call. * User-space user of this system call: Userspace RCU library 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 invocation 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. Results in liburcu: Operations in 10s, 6 readers, 2 writers: memory barriers in reader: 1701557485 reads, 2202847 writes signal-based scheme: 9830061167 reads, 6700 writes sys_membarrier: 9952759104 reads, 425 writes sys_membarrier (dyn. check): 7970328887 reads, 425 writes The dynamic sys_membarrier availability check adds some overhead to the read-side compared to the signal-based scheme, but besides that, sys_membarrier slightly outperforms the signal-based scheme. However, this non-expedited sys_membarrier implementation has a much slower grace period than signal and memory barrier schemes. Besides diminishing the number of wake-ups, one major advantage of the membarrier system call over the signal-based scheme is that it does not need to reserve a signal. This plays much more nicely with libraries, and with processes injected into for tracing purposes, for which we cannot expect that signals will be unused by the application. An expedited version of this system call can be added later on to speed up the grace period. Its implementation will likely depend on reading the cpu_curr()->mm without holding each CPU's rq lock. This patch adds the system call to x86 and to asm-generic. [1] http://urcu.so Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org> CC: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> CC: Steven Rostedt <rostedt@goodmis.org> CC: Nicholas Miell <nmiell@comcast.net> CC: Linus Torvalds <torvalds@linux-foundation.org> CC: Ingo Molnar <mingo@redhat.com> CC: Alan Cox <gnomes@lxorguk.ukuu.org.uk> CC: Lai Jiangshan <laijs@cn.fujitsu.com> CC: Stephen Hemminger <stephen@networkplumber.org> CC: Andrew Morton <akpm@linux-foundation.org> CC: Thomas Gleixner <tglx@linutronix.de> CC: Peter Zijlstra <peterz@infradead.org> CC: David Howells <dhowells@redhat.com> CC: Pranith Kumar <bobby.prani@gmail.com> CC: Michael Kerrisk <mtk.manpages@gmail.com> CC: linux-api@vger.kernel.org --- membarrier(2) man page: --------------- snip ------------------- MEMBARRIER(2) Linux Programmer's Manual MEMBARRIER(2) NAME membarrier - issue memory barriers on a set of threads SYNOPSIS #include <linux/membarrier.h> int membarrier(int cmd, int flags); DESCRIPTION The cmd argument is one of the following: MEMBARRIER_CMD_QUERY Query the set of supported commands. It returns a bitmask of supported commands. MEMBARRIER_CMD_SHARED Execute a memory barrier on all threads running on the system. Upon return from system call, the caller thread is ensured that all running threads have passed through a state where all memory accesses to user-space addresses match program order between entry to and return from the system call (non-running threads are de facto in such a state). This covers threads from all pro‐ cesses running on the system. This command returns 0. The flags argument needs to be 0. For future extensions. All memory accesses performed in program order from each targeted thread is guaranteed to be ordered with respect to sys_membarrier(). If we use the semantic "barrier()" to represent a compiler barrier forcing memory accesses to be performed in program order across the barrier, and smp_mb() to represent explicit memory barriers forcing full memory ordering across the barrier, we have the following ordering table for each pair of barrier(), sys_membarrier() and smp_mb(): The pair ordering is detailed as (O: ordered, X: not ordered): barrier() smp_mb() sys_membarrier() barrier() X X O smp_mb() X O O sys_membarrier() O O O RETURN VALUE On success, these system calls return zero. On error, -1 is returned, and errno is set appropriately. For a given command, with flags argument set to 0, this system call is guaranteed to always return the same value until reboot. ERRORS ENOSYS System call is not implemented. EINVAL Invalid arguments. Linux 2015-04-15 MEMBARRIER(2) --------------- snip ------------------- Changes since v18: - Add unlikely() check to flags, - Describe current users in changelog. Changes since v17: - Update commit message. Changes since v16: - Update documentation. - Add man page to changelog. - Build sys_membarrier on !CONFIG_SMP. It allows userspace applications to not care about the number of processors on the system. Based on recommendations from Stephen Hemminger and Steven Rostedt. - Check that flags argument is 0, update documentation to require it. Changes since v15: - Add flags argument in addition to cmd. - Update documentation. Changes since v14: - Take care of Thomas Gleixner's comments. Changes since v13: - Move to kernel/membarrier.c. - Remove MEMBARRIER_PRIVATE flag. - Add MAINTAINERS file entry. Changes since v12: - Remove _FLAG suffix from uapi flags. - Add Expert menuconfig option CONFIG_MEMBARRIER (default=y). - Remove EXPEDITED mode. Only implement non-expedited for now, until reading the cpu_curr()->mm can be done without holding the CPU's rq lock. Changes since v11: - 5 years have passed. - Rebase on v3.19 kernel. - Add futex-alike PRIVATE vs SHARED semantic: private for per-process barriers, non-private for memory mappings shared between processes. - Simplify user API. - Code refactoring. Changes since v10: - Apply Randy's comments. - Rebase on 2.6.34-rc4 -tip. Changes since v9: - Clean up #ifdef CONFIG_SMP. Changes since v8: - Go back to rq spin locks taken by sys_membarrier() rather than adding memory barriers to the scheduler. It implies a potential RoS (reduction of service) if sys_membarrier() is executed in a busy-loop by a user, but nothing more than what is already possible with other existing system calls, but saves memory barriers in the scheduler fast path. - re-add the memory barrier comments to x86 switch_mm() as an example to other architectures. - Update documentation of the memory barriers in sys_membarrier and switch_mm(). - Append execution scenarios to the changelog showing the purpose of each memory barrier. Changes since v7: - Move spinlock-mb and scheduler related changes to separate patches. - Add support for sys_membarrier on x86_32. - Only x86 32/64 system calls are reserved in this patch. It is planned to incrementally reserve syscall IDs on other architectures as these are tested. Changes 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. Changes 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. Changes 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. Changes 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. Changes 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. Changes 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.

Here is an implementation of a new system call, sys_membarrier(), which executes a memory barrier on all threads running on the system. It is implemented by calling synchronize_sched(). It can be used to distribute the cost of user-space memory barriers asymmetrically by transforming pairs of memory barriers into pairs consisting of sys_membarrier() and a compiler barrier. For synchronization primitives that distinguish between read-side and write-side (e.g. userspace RCU [1], rwlocks), the read-side can be accelerated significantly by moving the bulk of the memory barrier overhead to the write-side. The existing applications of which I am aware that would be improved by this system call are as follows: * Through Userspace RCU library (http://urcu.so) - DNS server (Knot DNS) https://www.knot-dns.cz/ - Network sniffer (http://netsniff-ng.org/) - Distributed object storage (https://sheepdog.github.io/sheepdog/) - User-space tracing (http://lttng.org) - Network storage system (https://www.gluster.org/) - Virtual routers (https://events.linuxfoundation.org/sites/events/files/slides/DPDK_RCU_0MQ.pdf) - Financial software (https://lkml.org/lkml/2015/3/23/189) Those projects use RCU in userspace to increase read-side speed and scalability compared to locking. Especially in the case of RCU used by libraries, sys_membarrier can speed up the read-side by moving the bulk of the memory barrier cost to synchronize_rcu(). * Direct users of sys_membarrier - core dotnet garbage collector (https://github.com/dotnet/coreclr/issues/198) Microsoft core dotnet GC developers are planning to use the mprotect() side-effect of issuing memory barriers through IPIs as a way to implement Windows FlushProcessWriteBuffers() on Linux. They are referring to sys_membarrier in their github thread, specifically stating that sys_membarrier() is what they are looking for. 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 are ordering memory accesses with respect to smp_mb() present in Thread B, we can change each smp_mb() within Thread A into calls to sys_membarrier() and each smp_mb() within Thread B into compiler barriers "barrier()". Before the change, we had, for each smp_mb() pairs: Thread A Thread B previous mem accesses previous mem accesses smp_mb() smp_mb() following mem accesses following 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() by synchronize_sched(). * Benchmarks On Intel Xeon E5405 (8 cores) (one thread is calling sys_membarrier, the other 7 threads are busy looping) 1000 non-expedited sys_membarrier calls in 33s =3D 33 milliseconds/call. * User-space user of this system call: Userspace RCU library 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 invocation 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. Results in liburcu: Operations in 10s, 6 readers, 2 writers: memory barriers in reader: 1701557485 reads, 2202847 writes signal-based scheme: 9830061167 reads, 6700 writes sys_membarrier: 9952759104 reads, 425 writes sys_membarrier (dyn. check): 7970328887 reads, 425 writes The dynamic sys_membarrier availability check adds some overhead to the read-side compared to the signal-based scheme, but besides that, sys_membarrier slightly outperforms the signal-based scheme. However, this non-expedited sys_membarrier implementation has a much slower grace period than signal and memory barrier schemes. Besides diminishing the number of wake-ups, one major advantage of the membarrier system call over the signal-based scheme is that it does not need to reserve a signal. This plays much more nicely with libraries, and with processes injected into for tracing purposes, for which we cannot expect that signals will be unused by the application. An expedited version of this system call can be added later on to speed up the grace period. Its implementation will likely depend on reading the cpu_curr()->mm without holding each CPU's rq lock. This patch adds the system call to x86 and to asm-generic. [1] http://urcu.so membarrier(2) man page: MEMBARRIER(2) Linux Programmer's Manual MEMBARRIER(2) NAME membarrier - issue memory barriers on a set of threads SYNOPSIS #include <linux/membarrier.h> int membarrier(int cmd, int flags); DESCRIPTION The cmd argument is one of the following: MEMBARRIER_CMD_QUERY Query the set of supported commands. It returns a bitmask of supported commands. MEMBARRIER_CMD_SHARED Execute a memory barrier on all threads running on the system. Upon return from system call, the caller thread is ensured that all running threads have passed through a state where all memory accesses to user-space addresses match program order between entry to and return from the system call (non-running threads are de facto in such a state). This covers threads from all pro=E2=80=90 cesses running on the system. This command returns 0. The flags argument needs to be 0. For future extensions. All memory accesses performed in program order from each targeted thread is guaranteed to be ordered with respect to sys_membarrier(). If we use the semantic "barrier()" to represent a compiler barrier forcing memory accesses to be performed in program order across the barrier, and smp_mb() to represent explicit memory barriers forcing full memory ordering across the barrier, we have the following ordering table for each pair of barrier(), sys_membarrier() and smp_mb(): The pair ordering is detailed as (O: ordered, X: not ordered): barrier() smp_mb() sys_membarrier() barrier() X X O smp_mb() X O O sys_membarrier() O O O RETURN VALUE On success, these system calls return zero. On error, -1 is returned, and errno is set appropriately. For a given command, with flags argument set to 0, this system call is guaranteed to always return the same value until reboot. ERRORS ENOSYS System call is not implemented. EINVAL Invalid arguments. Linux 2015-04-15 MEMBARRIER(2) Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Nicholas Miell <nmiell@comcast.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Ingo Molnar <mingo@redhat.com> Cc: Alan Cox <gnomes@lxorguk.ukuu.org.uk> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Stephen Hemminger <stephen@networkplumber.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: David Howells <dhowells@redhat.com> Cc: Pranith Kumar <bobby.prani@gmail.com> Cc: Michael Kerrisk <mtk.manpages@gmail.com> Cc: Shuah Khan <shuahkh@osg.samsung.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

Here is an implementation of a new system call, sys_membarrier(), which executes a memory barrier on all threads running on the system. It is implemented by calling synchronize_sched(). It can be used to distribute the cost of user-space memory barriers asymmetrically by transforming pairs of memory barriers into pairs consisting of sys_membarrier() and a compiler barrier. For synchronization primitives that distinguish between read-side and write-side (e.g. userspace RCU [1], rwlocks), the read-side can be accelerated significantly by moving the bulk of the memory barrier overhead to the write-side. The existing applications of which I am aware that would be improved by this system call are as follows: * Through Userspace RCU library (http://urcu.so) - DNS server (Knot DNS) https://www.knot-dns.cz/ - Network sniffer (http://netsniff-ng.org/) - Distributed object storage (https://sheepdog.github.io/sheepdog/) - User-space tracing (http://lttng.org) - Network storage system (https://www.gluster.org/) - Virtual routers (https://events.linuxfoundation.org/sites/events/files/slides/DPDK_RCU_0MQ.pdf) - Financial software (https://lkml.org/lkml/2015/3/23/189) Those projects use RCU in userspace to increase read-side speed and scalability compared to locking. Especially in the case of RCU used by libraries, sys_membarrier can speed up the read-side by moving the bulk of the memory barrier cost to synchronize_rcu(). * Direct users of sys_membarrier - core dotnet garbage collector (https://github.com/dotnet/coreclr/issues/198) Microsoft core dotnet GC developers are planning to use the mprotect() side-effect of issuing memory barriers through IPIs as a way to implement Windows FlushProcessWriteBuffers() on Linux. They are referring to sys_membarrier in their github thread, specifically stating that sys_membarrier() is what they are looking for. 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 are ordering memory accesses with respect to smp_mb() present in Thread B, we can change each smp_mb() within Thread A into calls to sys_membarrier() and each smp_mb() within Thread B into compiler barriers "barrier()". Before the change, we had, for each smp_mb() pairs: Thread A Thread B previous mem accesses previous mem accesses smp_mb() smp_mb() following mem accesses following 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() by synchronize_sched(). * Benchmarks On Intel Xeon E5405 (8 cores) (one thread is calling sys_membarrier, the other 7 threads are busy looping) 1000 non-expedited sys_membarrier calls in 33s =3D 33 milliseconds/call. * User-space user of this system call: Userspace RCU library 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 invocation 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. Results in liburcu: Operations in 10s, 6 readers, 2 writers: memory barriers in reader: 1701557485 reads, 2202847 writes signal-based scheme: 9830061167 reads, 6700 writes sys_membarrier: 9952759104 reads, 425 writes sys_membarrier (dyn. check): 7970328887 reads, 425 writes The dynamic sys_membarrier availability check adds some overhead to the read-side compared to the signal-based scheme, but besides that, sys_membarrier slightly outperforms the signal-based scheme. However, this non-expedited sys_membarrier implementation has a much slower grace period than signal and memory barrier schemes. Besides diminishing the number of wake-ups, one major advantage of the membarrier system call over the signal-based scheme is that it does not need to reserve a signal. This plays much more nicely with libraries, and with processes injected into for tracing purposes, for which we cannot expect that signals will be unused by the application. An expedited version of this system call can be added later on to speed up the grace period. Its implementation will likely depend on reading the cpu_curr()->mm without holding each CPU's rq lock. This patch adds the system call to x86 and to asm-generic. [1] http://urcu.so membarrier(2) man page: MEMBARRIER(2) Linux Programmer's Manual MEMBARRIER(2) NAME membarrier - issue memory barriers on a set of threads SYNOPSIS #include <linux/membarrier.h> int membarrier(int cmd, int flags); DESCRIPTION The cmd argument is one of the following: MEMBARRIER_CMD_QUERY Query the set of supported commands. It returns a bitmask of supported commands. MEMBARRIER_CMD_SHARED Execute a memory barrier on all threads running on the system. Upon return from system call, the caller thread is ensured that all running threads have passed through a state where all memory accesses to user-space addresses match program order between entry to and return from the system call (non-running threads are de facto in such a state). This covers threads from all pro=E2=80=90 cesses running on the system. This command returns 0. The flags argument needs to be 0. For future extensions. All memory accesses performed in program order from each targeted thread is guaranteed to be ordered with respect to sys_membarrier(). If we use the semantic "barrier()" to represent a compiler barrier forcing memory accesses to be performed in program order across the barrier, and smp_mb() to represent explicit memory barriers forcing full memory ordering across the barrier, we have the following ordering table for each pair of barrier(), sys_membarrier() and smp_mb(): The pair ordering is detailed as (O: ordered, X: not ordered): barrier() smp_mb() sys_membarrier() barrier() X X O smp_mb() X O O sys_membarrier() O O O RETURN VALUE On success, these system calls return zero. On error, -1 is returned, and errno is set appropriately. For a given command, with flags argument set to 0, this system call is guaranteed to always return the same value until reboot. ERRORS ENOSYS System call is not implemented. EINVAL Invalid arguments. Linux 2015-04-15 MEMBARRIER(2) Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Nicholas Miell <nmiell@comcast.net> Cc: Ingo Molnar <mingo@redhat.com> Cc: Alan Cox <gnomes@lxorguk.ukuu.org.uk> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Stephen Hemminger <stephen@networkplumber.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: David Howells <dhowells@redhat.com> Cc: Pranith Kumar <bobby.prani@gmail.com> Cc: Michael Kerrisk <mtk.manpages@gmail.com> Cc: Shuah Khan <shuahkh@osg.samsung.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

Here is an implementation of a new system call, sys_membarrier(), which executes a memory barrier on all threads running on the system. It is implemented by calling synchronize_sched(). It can be used to distribute the cost of user-space memory barriers asymmetrically by transforming pairs of memory barriers into pairs consisting of sys_membarrier() and a compiler barrier. For synchronization primitives that distinguish between read-side and write-side (e.g. userspace RCU [1], rwlocks), the read-side can be accelerated significantly by moving the bulk of the memory barrier overhead to the write-side. The existing applications of which I am aware that would be improved by this system call are as follows: * Through Userspace RCU library (http://urcu.so) - DNS server (Knot DNS) https://www.knot-dns.cz/ - Network sniffer (http://netsniff-ng.org/) - Distributed object storage (https://sheepdog.github.io/sheepdog/) - User-space tracing (http://lttng.org) - Network storage system (https://www.gluster.org/) - Virtual routers (https://events.linuxfoundation.org/sites/events/files/slides/DPDK_RCU_0MQ.pdf) - Financial software (https://lkml.org/lkml/2015/3/23/189) Those projects use RCU in userspace to increase read-side speed and scalability compared to locking. Especially in the case of RCU used by libraries, sys_membarrier can speed up the read-side by moving the bulk of the memory barrier cost to synchronize_rcu(). * Direct users of sys_membarrier - core dotnet garbage collector (https://github.com/dotnet/coreclr/issues/198) Microsoft core dotnet GC developers are planning to use the mprotect() side-effect of issuing memory barriers through IPIs as a way to implement Windows FlushProcessWriteBuffers() on Linux. They are referring to sys_membarrier in their github thread, specifically stating that sys_membarrier() is what they are looking for. 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 are ordering memory accesses with respect to smp_mb() present in Thread B, we can change each smp_mb() within Thread A into calls to sys_membarrier() and each smp_mb() within Thread B into compiler barriers "barrier()". Before the change, we had, for each smp_mb() pairs: Thread A Thread B previous mem accesses previous mem accesses smp_mb() smp_mb() following mem accesses following 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() by synchronize_sched(). * Benchmarks On Intel Xeon E5405 (8 cores) (one thread is calling sys_membarrier, the other 7 threads are busy looping) 1000 non-expedited sys_membarrier calls in 33s =3D 33 milliseconds/call. * User-space user of this system call: Userspace RCU library 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 invocation 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. Results in liburcu: Operations in 10s, 6 readers, 2 writers: memory barriers in reader: 1701557485 reads, 2202847 writes signal-based scheme: 9830061167 reads, 6700 writes sys_membarrier: 9952759104 reads, 425 writes sys_membarrier (dyn. check): 7970328887 reads, 425 writes The dynamic sys_membarrier availability check adds some overhead to the read-side compared to the signal-based scheme, but besides that, sys_membarrier slightly outperforms the signal-based scheme. However, this non-expedited sys_membarrier implementation has a much slower grace period than signal and memory barrier schemes. Besides diminishing the number of wake-ups, one major advantage of the membarrier system call over the signal-based scheme is that it does not need to reserve a signal. This plays much more nicely with libraries, and with processes injected into for tracing purposes, for which we cannot expect that signals will be unused by the application. An expedited version of this system call can be added later on to speed up the grace period. Its implementation will likely depend on reading the cpu_curr()->mm without holding each CPU's rq lock. This patch adds the system call to x86 and to asm-generic. [1] http://urcu.so membarrier(2) man page: MEMBARRIER(2) Linux Programmer's Manual MEMBARRIER(2) NAME membarrier - issue memory barriers on a set of threads SYNOPSIS #include <linux/membarrier.h> int membarrier(int cmd, int flags); DESCRIPTION The cmd argument is one of the following: MEMBARRIER_CMD_QUERY Query the set of supported commands. It returns a bitmask of supported commands. MEMBARRIER_CMD_SHARED Execute a memory barrier on all threads running on the system. Upon return from system call, the caller thread is ensured that all running threads have passed through a state where all memory accesses to user-space addresses match program order between entry to and return from the system call (non-running threads are de facto in such a state). This covers threads from all pro=E2=80=90 cesses running on the system. This command returns 0. The flags argument needs to be 0. For future extensions. All memory accesses performed in program order from each targeted thread is guaranteed to be ordered with respect to sys_membarrier(). If we use the semantic "barrier()" to represent a compiler barrier forcing memory accesses to be performed in program order across the barrier, and smp_mb() to represent explicit memory barriers forcing full memory ordering across the barrier, we have the following ordering table for each pair of barrier(), sys_membarrier() and smp_mb(): The pair ordering is detailed as (O: ordered, X: not ordered): barrier() smp_mb() sys_membarrier() barrier() X X O smp_mb() X O O sys_membarrier() O O O RETURN VALUE On success, these system calls return zero. On error, -1 is returned, and errno is set appropriately. For a given command, with flags argument set to 0, this system call is guaranteed to always return the same value until reboot. ERRORS ENOSYS System call is not implemented. EINVAL Invalid arguments. Linux 2015-04-15 MEMBARRIER(2) Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Nicholas Miell <nmiell@comcast.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Ingo Molnar <mingo@redhat.com> Cc: Alan Cox <gnomes@lxorguk.ukuu.org.uk> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Stephen Hemminger <stephen@networkplumber.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: David Howells <dhowells@redhat.com> Cc: Pranith Kumar <bobby.prani@gmail.com> Cc: Michael Kerrisk <mtk.manpages@gmail.com> Cc: Shuah Khan <shuahkh@osg.samsung.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

compudj · 2015-09-16T14:38:58Z

FYI, sys_membarrier has been merged into Linux 4.3-rc1 (currently x86 32/64 support, other architectures should be wired up soon).

jkotas · 2015-09-16T16:00:35Z

@compudj Perfect - I have opened issue for CoreCLR to take advantage of it.

Here is an implementation of a new system call, sys_membarrier(), which executes a memory barrier on all threads running on the system. It is implemented by calling synchronize_sched(). It can be used to distribute the cost of user-space memory barriers asymmetrically by transforming pairs of memory barriers into pairs consisting of sys_membarrier() and a compiler barrier. For synchronization primitives that distinguish between read-side and write-side (e.g. userspace RCU [1], rwlocks), the read-side can be accelerated significantly by moving the bulk of the memory barrier overhead to the write-side. The existing applications of which I am aware that would be improved by this system call are as follows: * Through Userspace RCU library (http://urcu.so) - DNS server (Knot DNS) https://www.knot-dns.cz/ - Network sniffer (http://netsniff-ng.org/) - Distributed object storage (https://sheepdog.github.io/sheepdog/) - User-space tracing (http://lttng.org) - Network storage system (https://www.gluster.org/) - Virtual routers (https://events.linuxfoundation.org/sites/events/files/slides/DPDK_RCU_0MQ.pdf) - Financial software (https://lkml.org/lkml/2015/3/23/189) Those projects use RCU in userspace to increase read-side speed and scalability compared to locking. Especially in the case of RCU used by libraries, sys_membarrier can speed up the read-side by moving the bulk of the memory barrier cost to synchronize_rcu(). * Direct users of sys_membarrier - core dotnet garbage collector (https://github.com/dotnet/coreclr/issues/198) Microsoft core dotnet GC developers are planning to use the mprotect() side-effect of issuing memory barriers through IPIs as a way to implement Windows FlushProcessWriteBuffers() on Linux. They are referring to sys_membarrier in their github thread, specifically stating that sys_membarrier() is what they are looking for. 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 are ordering memory accesses with respect to smp_mb() present in Thread B, we can change each smp_mb() within Thread A into calls to sys_membarrier() and each smp_mb() within Thread B into compiler barriers "barrier()". Before the change, we had, for each smp_mb() pairs: Thread A Thread B previous mem accesses previous mem accesses smp_mb() smp_mb() following mem accesses following 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() by synchronize_sched(). * Benchmarks On Intel Xeon E5405 (8 cores) (one thread is calling sys_membarrier, the other 7 threads are busy looping) 1000 non-expedited sys_membarrier calls in 33s =3D 33 milliseconds/call. * User-space user of this system call: Userspace RCU library 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 invocation 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. Results in liburcu: Operations in 10s, 6 readers, 2 writers: memory barriers in reader: 1701557485 reads, 2202847 writes signal-based scheme: 9830061167 reads, 6700 writes sys_membarrier: 9952759104 reads, 425 writes sys_membarrier (dyn. check): 7970328887 reads, 425 writes The dynamic sys_membarrier availability check adds some overhead to the read-side compared to the signal-based scheme, but besides that, sys_membarrier slightly outperforms the signal-based scheme. However, this non-expedited sys_membarrier implementation has a much slower grace period than signal and memory barrier schemes. Besides diminishing the number of wake-ups, one major advantage of the membarrier system call over the signal-based scheme is that it does not need to reserve a signal. This plays much more nicely with libraries, and with processes injected into for tracing purposes, for which we cannot expect that signals will be unused by the application. An expedited version of this system call can be added later on to speed up the grace period. Its implementation will likely depend on reading the cpu_curr()->mm without holding each CPU's rq lock. This patch adds the system call to x86 and to asm-generic. [1] http://urcu.so membarrier(2) man page: MEMBARRIER(2) Linux Programmer's Manual MEMBARRIER(2) NAME membarrier - issue memory barriers on a set of threads SYNOPSIS #include <linux/membarrier.h> int membarrier(int cmd, int flags); DESCRIPTION The cmd argument is one of the following: MEMBARRIER_CMD_QUERY Query the set of supported commands. It returns a bitmask of supported commands. MEMBARRIER_CMD_SHARED Execute a memory barrier on all threads running on the system. Upon return from system call, the caller thread is ensured that all running threads have passed through a state where all memory accesses to user-space addresses match program order between entry to and return from the system call (non-running threads are de facto in such a state). This covers threads from all pro=E2=80=90 cesses running on the system. This command returns 0. The flags argument needs to be 0. For future extensions. All memory accesses performed in program order from each targeted thread is guaranteed to be ordered with respect to sys_membarrier(). If we use the semantic "barrier()" to represent a compiler barrier forcing memory accesses to be performed in program order across the barrier, and smp_mb() to represent explicit memory barriers forcing full memory ordering across the barrier, we have the following ordering table for each pair of barrier(), sys_membarrier() and smp_mb(): The pair ordering is detailed as (O: ordered, X: not ordered): barrier() smp_mb() sys_membarrier() barrier() X X O smp_mb() X O O sys_membarrier() O O O RETURN VALUE On success, these system calls return zero. On error, -1 is returned, and errno is set appropriately. For a given command, with flags argument set to 0, this system call is guaranteed to always return the same value until reboot. ERRORS ENOSYS System call is not implemented. EINVAL Invalid arguments. Linux 2015-04-15 MEMBARRIER(2) Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Nicholas Miell <nmiell@comcast.net> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Ingo Molnar <mingo@redhat.com> Cc: Alan Cox <gnomes@lxorguk.ukuu.org.uk> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Stephen Hemminger <stephen@networkplumber.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: David Howells <dhowells@redhat.com> Cc: Pranith Kumar <bobby.prani@gmail.com> Cc: Michael Kerrisk <mtk.manpages@gmail.com> Cc: Shuah Khan <shuahkh@osg.samsung.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

Here is an implementation of a new system call, sys_membarrier(), which executes a memory barrier on all threads running on the system. It is implemented by calling synchronize_sched(). It can be used to distribute the cost of user-space memory barriers asymmetrically by transforming pairs of memory barriers into pairs consisting of sys_membarrier() and a compiler barrier. For synchronization primitives that distinguish between read-side and write-side (e.g. userspace RCU [1], rwlocks), the read-side can be accelerated significantly by moving the bulk of the memory barrier overhead to the write-side. The existing applications of which I am aware that would be improved by this system call are as follows: * Through Userspace RCU library (http://urcu.so) - DNS server (Knot DNS) https://www.knot-dns.cz/ - Network sniffer (http://netsniff-ng.org/) - Distributed object storage (https://sheepdog.github.io/sheepdog/) - User-space tracing (http://lttng.org) - Network storage system (https://www.gluster.org/) - Virtual routers (https://events.linuxfoundation.org/sites/events/files/slides/DPDK_RCU_0MQ.pdf) - Financial software (https://lkml.org/lkml/2015/3/23/189) Those projects use RCU in userspace to increase read-side speed and scalability compared to locking. Especially in the case of RCU used by libraries, sys_membarrier can speed up the read-side by moving the bulk of the memory barrier cost to synchronize_rcu(). * Direct users of sys_membarrier - core dotnet garbage collector (https://github.com/dotnet/coreclr/issues/198) Microsoft core dotnet GC developers are planning to use the mprotect() side-effect of issuing memory barriers through IPIs as a way to implement Windows FlushProcessWriteBuffers() on Linux. They are referring to sys_membarrier in their github thread, specifically stating that sys_membarrier() is what they are looking for. 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 are ordering memory accesses with respect to smp_mb() present in Thread B, we can change each smp_mb() within Thread A into calls to sys_membarrier() and each smp_mb() within Thread B into compiler barriers "barrier()". Before the change, we had, for each smp_mb() pairs: Thread A Thread B previous mem accesses previous mem accesses smp_mb() smp_mb() following mem accesses following 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() by synchronize_sched(). * Benchmarks On Intel Xeon E5405 (8 cores) (one thread is calling sys_membarrier, the other 7 threads are busy looping) 1000 non-expedited sys_membarrier calls in 33s =3D 33 milliseconds/call. * User-space user of this system call: Userspace RCU library 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 invocation 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. Results in liburcu: Operations in 10s, 6 readers, 2 writers: memory barriers in reader: 1701557485 reads, 2202847 writes signal-based scheme: 9830061167 reads, 6700 writes sys_membarrier: 9952759104 reads, 425 writes sys_membarrier (dyn. check): 7970328887 reads, 425 writes The dynamic sys_membarrier availability check adds some overhead to the read-side compared to the signal-based scheme, but besides that, sys_membarrier slightly outperforms the signal-based scheme. However, this non-expedited sys_membarrier implementation has a much slower grace period than signal and memory barrier schemes. Besides diminishing the number of wake-ups, one major advantage of the membarrier system call over the signal-based scheme is that it does not need to reserve a signal. This plays much more nicely with libraries, and with processes injected into for tracing purposes, for which we cannot expect that signals will be unused by the application. An expedited version of this system call can be added later on to speed up the grace period. Its implementation will likely depend on reading the cpu_curr()->mm without holding each CPU's rq lock. This patch adds the system call to x86 and to asm-generic. [1] http://urcu.so membarrier(2) man page: MEMBARRIER(2) Linux Programmer's Manual MEMBARRIER(2) NAME membarrier - issue memory barriers on a set of threads SYNOPSIS #include <linux/membarrier.h> int membarrier(int cmd, int flags); DESCRIPTION The cmd argument is one of the following: MEMBARRIER_CMD_QUERY Query the set of supported commands. It returns a bitmask of supported commands. MEMBARRIER_CMD_SHARED Execute a memory barrier on all threads running on the system. Upon return from system call, the caller thread is ensured that all running threads have passed through a state where all memory accesses to user-space addresses match program order between entry to and return from the system call (non-running threads are de facto in such a state). This covers threads from all pro=E2=80=90 cesses running on the system. This command returns 0. The flags argument needs to be 0. For future extensions. All memory accesses performed in program order from each targeted thread is guaranteed to be ordered with respect to sys_membarrier(). If we use the semantic "barrier()" to represent a compiler barrier forcing memory accesses to be performed in program order across the barrier, and smp_mb() to represent explicit memory barriers forcing full memory ordering across the barrier, we have the following ordering table for each pair of barrier(), sys_membarrier() and smp_mb(): The pair ordering is detailed as (O: ordered, X: not ordered): barrier() smp_mb() sys_membarrier() barrier() X X O smp_mb() X O O sys_membarrier() O O O RETURN VALUE On success, these system calls return zero. On error, -1 is returned, and errno is set appropriately. For a given command, with flags argument set to 0, this system call is guaranteed to always return the same value until reboot. ERRORS ENOSYS System call is not implemented. EINVAL Invalid arguments. Linux 2015-04-15 MEMBARRIER(2) Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Nicholas Miell <nmiell@comcast.net> Cc: Ingo Molnar <mingo@redhat.com> Cc: Alan Cox <gnomes@lxorguk.ukuu.org.uk> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Stephen Hemminger <stephen@networkplumber.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: David Howells <dhowells@redhat.com> Cc: Pranith Kumar <bobby.prani@gmail.com> Cc: Michael Kerrisk <mtk.manpages@gmail.com> Cc: Shuah Khan <shuahkh@osg.samsung.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: mostafa-z <mostafazarghami@gmail.com>

Here is an implementation of a new system call, sys_membarrier(), which executes a memory barrier on all threads running on the system. It is implemented by calling synchronize_sched(). It can be used to distribute the cost of user-space memory barriers asymmetrically by transforming pairs of memory barriers into pairs consisting of sys_membarrier() and a compiler barrier. For synchronization primitives that distinguish between read-side and write-side (e.g. userspace RCU [1], rwlocks), the read-side can be accelerated significantly by moving the bulk of the memory barrier overhead to the write-side. The existing applications of which I am aware that would be improved by this system call are as follows: * Through Userspace RCU library (http://urcu.so) - DNS server (Knot DNS) https://www.knot-dns.cz/ - Network sniffer (http://netsniff-ng.org/) - Distributed object storage (https://sheepdog.github.io/sheepdog/) - User-space tracing (http://lttng.org) - Network storage system (https://www.gluster.org/) - Virtual routers (https://events.linuxfoundation.org/sites/events/files/slides/DPDK_RCU_0MQ.pdf) - Financial software (https://lkml.org/lkml/2015/3/23/189) Those projects use RCU in userspace to increase read-side speed and scalability compared to locking. Especially in the case of RCU used by libraries, sys_membarrier can speed up the read-side by moving the bulk of the memory barrier cost to synchronize_rcu(). * Direct users of sys_membarrier - core dotnet garbage collector (https://github.com/dotnet/coreclr/issues/198) Microsoft core dotnet GC developers are planning to use the mprotect() side-effect of issuing memory barriers through IPIs as a way to implement Windows FlushProcessWriteBuffers() on Linux. They are referring to sys_membarrier in their github thread, specifically stating that sys_membarrier() is what they are looking for. 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 are ordering memory accesses with respect to smp_mb() present in Thread B, we can change each smp_mb() within Thread A into calls to sys_membarrier() and each smp_mb() within Thread B into compiler barriers "barrier()". Before the change, we had, for each smp_mb() pairs: Thread A Thread B previous mem accesses previous mem accesses smp_mb() smp_mb() following mem accesses following 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() by synchronize_sched(). * Benchmarks On Intel Xeon E5405 (8 cores) (one thread is calling sys_membarrier, the other 7 threads are busy looping) 1000 non-expedited sys_membarrier calls in 33s =3D 33 milliseconds/call. * User-space user of this system call: Userspace RCU library 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 invocation 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. Results in liburcu: Operations in 10s, 6 readers, 2 writers: memory barriers in reader: 1701557485 reads, 2202847 writes signal-based scheme: 9830061167 reads, 6700 writes sys_membarrier: 9952759104 reads, 425 writes sys_membarrier (dyn. check): 7970328887 reads, 425 writes The dynamic sys_membarrier availability check adds some overhead to the read-side compared to the signal-based scheme, but besides that, sys_membarrier slightly outperforms the signal-based scheme. However, this non-expedited sys_membarrier implementation has a much slower grace period than signal and memory barrier schemes. Besides diminishing the number of wake-ups, one major advantage of the membarrier system call over the signal-based scheme is that it does not need to reserve a signal. This plays much more nicely with libraries, and with processes injected into for tracing purposes, for which we cannot expect that signals will be unused by the application. An expedited version of this system call can be added later on to speed up the grace period. Its implementation will likely depend on reading the cpu_curr()->mm without holding each CPU's rq lock. This patch adds the system call to x86 and to asm-generic. [1] http://urcu.so membarrier(2) man page: MEMBARRIER(2) Linux Programmer's Manual MEMBARRIER(2) NAME membarrier - issue memory barriers on a set of threads SYNOPSIS #include <linux/membarrier.h> int membarrier(int cmd, int flags); DESCRIPTION The cmd argument is one of the following: MEMBARRIER_CMD_QUERY Query the set of supported commands. It returns a bitmask of supported commands. MEMBARRIER_CMD_SHARED Execute a memory barrier on all threads running on the system. Upon return from system call, the caller thread is ensured that all running threads have passed through a state where all memory accesses to user-space addresses match program order between entry to and return from the system call (non-running threads are de facto in such a state). This covers threads from all pro=E2=80=90 cesses running on the system. This command returns 0. The flags argument needs to be 0. For future extensions. All memory accesses performed in program order from each targeted thread is guaranteed to be ordered with respect to sys_membarrier(). If we use the semantic "barrier()" to represent a compiler barrier forcing memory accesses to be performed in program order across the barrier, and smp_mb() to represent explicit memory barriers forcing full memory ordering across the barrier, we have the following ordering table for each pair of barrier(), sys_membarrier() and smp_mb(): The pair ordering is detailed as (O: ordered, X: not ordered): barrier() smp_mb() sys_membarrier() barrier() X X O smp_mb() X O O sys_membarrier() O O O RETURN VALUE On success, these system calls return zero. On error, -1 is returned, and errno is set appropriately. For a given command, with flags argument set to 0, this system call is guaranteed to always return the same value until reboot. ERRORS ENOSYS System call is not implemented. EINVAL Invalid arguments. Linux 2015-04-15 MEMBARRIER(2) Change-Id: If02eda5e20e3de58074175f56ae0b3b118397c69 Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Nicholas Miell <nmiell@comcast.net> Cc: Ingo Molnar <mingo@redhat.com> Cc: Alan Cox <gnomes@lxorguk.ukuu.org.uk> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Stephen Hemminger <stephen@networkplumber.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: David Howells <dhowells@redhat.com> Cc: Pranith Kumar <bobby.prani@gmail.com> Cc: Michael Kerrisk <mtk.manpages@gmail.com> Cc: Shuah Khan <shuahkh@osg.samsung.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

Here is an implementation of a new system call, sys_membarrier(), which executes a memory barrier on all threads running on the system. It is implemented by calling synchronize_sched(). It can be used to distribute the cost of user-space memory barriers asymmetrically by transforming pairs of memory barriers into pairs consisting of sys_membarrier() and a compiler barrier. For synchronization primitives that distinguish between read-side and write-side (e.g. userspace RCU [1], rwlocks), the read-side can be accelerated significantly by moving the bulk of the memory barrier overhead to the write-side. The existing applications of which I am aware that would be improved by this system call are as follows: * Through Userspace RCU library (http://urcu.so) - DNS server (Knot DNS) https://www.knot-dns.cz/ - Network sniffer (http://netsniff-ng.org/) - Distributed object storage (https://sheepdog.github.io/sheepdog/) - User-space tracing (http://lttng.org) - Network storage system (https://www.gluster.org/) - Virtual routers (https://events.linuxfoundation.org/sites/events/files/slides/DPDK_RCU_0MQ.pdf) - Financial software (https://lkml.org/lkml/2015/3/23/189) Those projects use RCU in userspace to increase read-side speed and scalability compared to locking. Especially in the case of RCU used by libraries, sys_membarrier can speed up the read-side by moving the bulk of the memory barrier cost to synchronize_rcu(). * Direct users of sys_membarrier - core dotnet garbage collector (https://github.com/dotnet/coreclr/issues/198) Microsoft core dotnet GC developers are planning to use the mprotect() side-effect of issuing memory barriers through IPIs as a way to implement Windows FlushProcessWriteBuffers() on Linux. They are referring to sys_membarrier in their github thread, specifically stating that sys_membarrier() is what they are looking for. 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 are ordering memory accesses with respect to smp_mb() present in Thread B, we can change each smp_mb() within Thread A into calls to sys_membarrier() and each smp_mb() within Thread B into compiler barriers "barrier()". Before the change, we had, for each smp_mb() pairs: Thread A Thread B previous mem accesses previous mem accesses smp_mb() smp_mb() following mem accesses following 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() by synchronize_sched(). * Benchmarks On Intel Xeon E5405 (8 cores) (one thread is calling sys_membarrier, the other 7 threads are busy looping) 1000 non-expedited sys_membarrier calls in 33s =3D 33 milliseconds/call. * User-space user of this system call: Userspace RCU library 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 invocation 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. Results in liburcu: Operations in 10s, 6 readers, 2 writers: memory barriers in reader: 1701557485 reads, 2202847 writes signal-based scheme: 9830061167 reads, 6700 writes sys_membarrier: 9952759104 reads, 425 writes sys_membarrier (dyn. check): 7970328887 reads, 425 writes The dynamic sys_membarrier availability check adds some overhead to the read-side compared to the signal-based scheme, but besides that, sys_membarrier slightly outperforms the signal-based scheme. However, this non-expedited sys_membarrier implementation has a much slower grace period than signal and memory barrier schemes. Besides diminishing the number of wake-ups, one major advantage of the membarrier system call over the signal-based scheme is that it does not need to reserve a signal. This plays much more nicely with libraries, and with processes injected into for tracing purposes, for which we cannot expect that signals will be unused by the application. An expedited version of this system call can be added later on to speed up the grace period. Its implementation will likely depend on reading the cpu_curr()->mm without holding each CPU's rq lock. This patch adds the system call to x86 and to asm-generic. [1] http://urcu.so membarrier(2) man page: MEMBARRIER(2) Linux Programmer's Manual MEMBARRIER(2) NAME membarrier - issue memory barriers on a set of threads SYNOPSIS #include <linux/membarrier.h> int membarrier(int cmd, int flags); DESCRIPTION The cmd argument is one of the following: MEMBARRIER_CMD_QUERY Query the set of supported commands. It returns a bitmask of supported commands. MEMBARRIER_CMD_SHARED Execute a memory barrier on all threads running on the system. Upon return from system call, the caller thread is ensured that all running threads have passed through a state where all memory accesses to user-space addresses match program order between entry to and return from the system call (non-running threads are de facto in such a state). This covers threads from all pro=E2=80=90 cesses running on the system. This command returns 0. The flags argument needs to be 0. For future extensions. All memory accesses performed in program order from each targeted thread is guaranteed to be ordered with respect to sys_membarrier(). If we use the semantic "barrier()" to represent a compiler barrier forcing memory accesses to be performed in program order across the barrier, and smp_mb() to represent explicit memory barriers forcing full memory ordering across the barrier, we have the following ordering table for each pair of barrier(), sys_membarrier() and smp_mb(): The pair ordering is detailed as (O: ordered, X: not ordered): barrier() smp_mb() sys_membarrier() barrier() X X O smp_mb() X O O sys_membarrier() O O O RETURN VALUE On success, these system calls return zero. On error, -1 is returned, and errno is set appropriately. For a given command, with flags argument set to 0, this system call is guaranteed to always return the same value until reboot. ERRORS ENOSYS System call is not implemented. EINVAL Invalid arguments. Linux 2015-04-15 MEMBARRIER(2) Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Nicholas Miell <nmiell@comcast.net> Cc: Ingo Molnar <mingo@redhat.com> Cc: Alan Cox <gnomes@lxorguk.ukuu.org.uk> Cc: Lai Jiangshan <laijs@cn.fujitsu.com> Cc: Stephen Hemminger <stephen@networkplumber.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: David Howells <dhowells@redhat.com> Cc: Pranith Kumar <bobby.prani@gmail.com> Cc: Michael Kerrisk <mtk.manpages@gmail.com> Cc: Shuah Khan <shuahkh@osg.samsung.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

xiangzhai · 2020-04-27T01:43:24Z

:mips-interest

janvorli self-assigned this May 27, 2015

janvorli closed this as completed May 27, 2015

msftgits transferred this issue from dotnet/coreclr Jan 30, 2020

jkotas mentioned this issue Jan 30, 2020

Use sys_membarrier to implement FlushProcessWriteBuffers on Linux 4.3+ #4501

Closed

ghost locked as resolved and limited conversation to collaborators Jan 7, 2021

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Implement FlushProcessWriteBuffers in PAL #3952

Implement FlushProcessWriteBuffers in PAL #3952

sergiy-k commented Feb 11, 2015

kangaroo commented Feb 15, 2015

jkotas commented Feb 15, 2015

janvorli commented May 15, 2015

jkotas commented May 16, 2015

janvorli commented May 27, 2015

compudj commented Sep 16, 2015

jkotas commented Sep 16, 2015

xiangzhai commented Apr 27, 2020

Implement FlushProcessWriteBuffers in PAL #3952

Implement FlushProcessWriteBuffers in PAL #3952

Comments

sergiy-k commented Feb 11, 2015

kangaroo commented Feb 15, 2015

jkotas commented Feb 15, 2015

janvorli commented May 15, 2015

1. Helper threads

2. Signals

jkotas commented May 16, 2015

janvorli commented May 27, 2015

compudj commented Sep 16, 2015

jkotas commented Sep 16, 2015

xiangzhai commented Apr 27, 2020