Date:   Sat, 27 Jan 2018 07:35:20 +0000
From:   Lihao Liang <lihao.liang@...il.com>
To:     Paul McKenney <paulmck@...ux.vnet.ibm.com>,
        Peter Zijlstra <peterz@...radead.org>
Cc:     "Guohanjun (Hanjun Guo)" <guohanjun@...wei.com>, heng.z@...wei.com,
        hb.chen@...wei.com, linux-kernel@...r.kernel.org
Subject: Re: [PATCH RFC 01/16] prcu: Add PRCU implementation

On Thu, Jan 25, 2018 at 6:16 AM, Paul E. McKenney
<paulmck@...ux.vnet.ibm.com> wrote:
> On Tue, Jan 23, 2018 at 03:59:26PM +0800, lianglihao@...wei.com wrote:
>> From: Heng Zhang <heng.z@...wei.com>
>>
>> This RCU implementation (PRCU) is based on a fast consensus protocol
>> published in the following paper:
>>
>> Fast Consensus Using Bounded Staleness for Scalable Read-mostly Synchronization.
>> Haibo Chen, Heng Zhang, Ran Liu, Binyu Zang, and Haibing Guan.
>> IEEE Transactions on Parallel and Distributed Systems (TPDS), 2016.
>> https://dl.acm.org/citation.cfm?id=3024114.3024143
>>
>> Signed-off-by: Heng Zhang <heng.z@...wei.com>
>> Signed-off-by: Lihao Liang <lianglihao@...wei.com>
>
> A few comments and questions interspersed.
>
>                                                         Thanx, Paul
>
>> ---
>>  include/linux/prcu.h |  37 +++++++++++++++
>>  kernel/rcu/Makefile  |   2 +-
>>  kernel/rcu/prcu.c    | 125 +++++++++++++++++++++++++++++++++++++++++++++++++++
>>  kernel/sched/core.c  |   2 +
>>  4 files changed, 165 insertions(+), 1 deletion(-)
>>  create mode 100644 include/linux/prcu.h
>>  create mode 100644 kernel/rcu/prcu.c
>>
>> diff --git a/include/linux/prcu.h b/include/linux/prcu.h
>> new file mode 100644
>> index 00000000..653b4633
>> --- /dev/null
>> +++ b/include/linux/prcu.h
>> @@ -0,0 +1,37 @@
>> +#ifndef __LINUX_PRCU_H
>> +#define __LINUX_PRCU_H
>> +
>> +#include <linux/atomic.h>
>> +#include <linux/mutex.h>
>> +#include <linux/wait.h>
>> +
>> +#define CONFIG_PRCU
>> +
>> +struct prcu_local_struct {
>> +     unsigned int locked;
>> +     unsigned int online;
>> +     unsigned long long version;
>> +};
>> +
>> +struct prcu_struct {
>> +     atomic64_t global_version;
>> +     atomic_t active_ctr;
>> +     struct mutex mtx;
>> +     wait_queue_head_t wait_q;
>> +};
>> +
>> +#ifdef CONFIG_PRCU
>> +void prcu_read_lock(void);
>> +void prcu_read_unlock(void);
>> +void synchronize_prcu(void);
>> +void prcu_note_context_switch(void);
>> +
>> +#else /* #ifdef CONFIG_PRCU */
>> +
>> +#define prcu_read_lock() do {} while (0)
>> +#define prcu_read_unlock() do {} while (0)
>> +#define synchronize_prcu() do {} while (0)
>> +#define prcu_note_context_switch() do {} while (0)
>
> If CONFIG_PRCU=n and some code is built that uses PRCU, shouldn't you
> get a build error rather than an error-free but inoperative PRCU?
>

Very good point, thank you!

> Of course, Peter's question about purpose of the patch set applies
> here as well.
>

The main motivation of this patch set is the comparison results of
rcuperf between PRCU and Tree RCU in which PRCU outperformed Tree RCU
by a large margin.

As indicated in your reply of the email in this patch series

[PATCH RFC 00/16] A new RCU implementation based on a fast consensus protocol

this may be a bug on either expedited RCU grace-period latency or on
rcuperf's measurements.

Many thanks,
Lihao.

>> +
>> +#endif /* #ifdef CONFIG_PRCU */
>> +#endif /* __LINUX_PRCU_H */
>> diff --git a/kernel/rcu/Makefile b/kernel/rcu/Makefile
>> index 23803c7d..8791419c 100644
>> --- a/kernel/rcu/Makefile
>> +++ b/kernel/rcu/Makefile
>> @@ -2,7 +2,7 @@
>>  # and is generally not a function of system call inputs.
>>  KCOV_INSTRUMENT := n
>>
>> -obj-y += update.o sync.o
>> +obj-y += update.o sync.o prcu.o
>>  obj-$(CONFIG_CLASSIC_SRCU) += srcu.o
>>  obj-$(CONFIG_TREE_SRCU) += srcutree.o
>>  obj-$(CONFIG_TINY_SRCU) += srcutiny.o
>> diff --git a/kernel/rcu/prcu.c b/kernel/rcu/prcu.c
>> new file mode 100644
>> index 00000000..a00b9420
>> --- /dev/null
>> +++ b/kernel/rcu/prcu.c
>> @@ -0,0 +1,125 @@
>> +#include <linux/smp.h>
>> +#include <linux/prcu.h>
>> +#include <linux/percpu.h>
>> +#include <linux/compiler.h>
>> +#include <linux/sched.h>
>> +
>> +#include <asm/barrier.h>
>> +
>> +DEFINE_PER_CPU_SHARED_ALIGNED(struct prcu_local_struct, prcu_local);
>> +
>> +struct prcu_struct global_prcu = {
>> +     .global_version = ATOMIC64_INIT(0),
>> +     .active_ctr = ATOMIC_INIT(0),
>> +     .mtx = __MUTEX_INITIALIZER(global_prcu.mtx),
>> +     .wait_q = __WAIT_QUEUE_HEAD_INITIALIZER(global_prcu.wait_q)
>> +};
>> +struct prcu_struct *prcu = &global_prcu;
>> +
>> +static inline void prcu_report(struct prcu_local_struct *local)
>> +{
>> +     unsigned long long global_version;
>> +     unsigned long long local_version;
>> +
>> +     global_version = atomic64_read(&prcu->global_version);
>> +     local_version = local->version;
>> +     if (global_version > local_version)
>> +             cmpxchg(&local->version, local_version, global_version);
>> +}
>> +
>> +void prcu_read_lock(void)
>> +{
>> +     struct prcu_local_struct *local;
>> +
>> +     local = get_cpu_ptr(&prcu_local);
>> +     if (!local->online) {
>> +             WRITE_ONCE(local->online, 1);
>> +             smp_mb();
>> +     }
>> +
>> +     local->locked++;
>> +     put_cpu_ptr(&prcu_local);
>> +}
>> +EXPORT_SYMBOL(prcu_read_lock);
>> +
>> +void prcu_read_unlock(void)
>> +{
>> +     int locked;
>> +     struct prcu_local_struct *local;
>> +
>> +     barrier();
>> +     local = get_cpu_ptr(&prcu_local);
>> +     locked = local->locked;
>> +     if (locked) {
>> +             local->locked--;
>> +             if (locked == 1)
>> +                     prcu_report(local);
>
> Is ordering important here?  It looks to me that the compiler could
> rearrange some of the accesses within prcu_report() with the local->locked
> decrement.  There appears to be some potential for load and store tearing,
> though perhaps you have verified that your compiler avoids this on
> the architecture that you are using.
>
>> +             put_cpu_ptr(&prcu_local);
>> +     } else {
>
> Hmmm...  We get here if the RCU read-side critical section was preempted.
> If none of them are preempted, ->active_ctr remains zero.
>
>> +             put_cpu_ptr(&prcu_local);
>> +             if (!atomic_dec_return(&prcu->active_ctr))
>> +                     wake_up(&prcu->wait_q);
>> +     }
>> +}
>> +EXPORT_SYMBOL(prcu_read_unlock);
>> +
>> +static void prcu_handler(void *info)
>> +{
>> +     struct prcu_local_struct *local;
>> +
>> +     local = this_cpu_ptr(&prcu_local);
>> +     if (!local->locked)
>> +             WRITE_ONCE(local->version, atomic64_read(&prcu->global_version));
>> +}
>> +
>> +void synchronize_prcu(void)
>> +{
>> +     int cpu;
>> +     cpumask_t cpus;
>> +     unsigned long long version;
>> +     struct prcu_local_struct *local;
>> +
>> +     version = atomic64_add_return(1, &prcu->global_version);
>> +     mutex_lock(&prcu->mtx);
>> +
>> +     local = get_cpu_ptr(&prcu_local);
>> +     local->version = version;
>> +     put_cpu_ptr(&prcu_local);
>> +
>> +     cpumask_clear(&cpus);
>> +     for_each_possible_cpu(cpu) {
>> +             local = per_cpu_ptr(&prcu_local, cpu);
>> +             if (!READ_ONCE(local->online))
>> +                     continue;
>> +             if (READ_ONCE(local->version) < version) {
>
> On 32-bit systems, given that ->version is long long, you might see
> load tearing.  And on some 32-bit systems, the cmpxchg() in prcu_hander()
> might not build.
>
> Or is the idea that only prcu_handler() updates ->version?  But in that
> case, you wouldn't need the READ_ONCE() above.  What am I missing here?
>
>> +                     smp_call_function_single(cpu, prcu_handler, NULL, 0);
>> +                     cpumask_set_cpu(cpu, &cpus);
>> +             }
>> +     }
>> +
>> +     for_each_cpu(cpu, &cpus) {
>> +             local = per_cpu_ptr(&prcu_local, cpu);
>> +             while (READ_ONCE(local->version) < version)
>
> This ->version read can also tear on some 32-bit systems, and this
> one most definitely can race with the prcu_handler() above.  Does the
> algorithm operate correctly in that case?  (It doesn't look that way
> to me, but I might be missing something.) Or are 32-bit systems excluded?
>
>> +                     cpu_relax();
>> +     }
>
> I might be missing something, but I believe we need a memory barrier
> here on non-TSO systems.  Without that, couldn't we miss a preemption?
>
>> +
>> +     if (atomic_read(&prcu->active_ctr))
>> +             wait_event(prcu->wait_q, !atomic_read(&prcu->active_ctr));
>> +
>> +     mutex_unlock(&prcu->mtx);
>> +}
>> +EXPORT_SYMBOL(synchronize_prcu);
>> +
>> +void prcu_note_context_switch(void)
>> +{
>> +     struct prcu_local_struct *local;
>> +
>> +     local = get_cpu_ptr(&prcu_local);
>> +     if (local->locked) {
>> +             atomic_add(local->locked, &prcu->active_ctr);
>> +             local->locked = 0;
>> +     }
>> +     local->online = 0;
>> +     prcu_report(local);
>> +     put_cpu_ptr(&prcu_local);
>> +}
>> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
>> index 326d4f88..a308581b 100644
>> --- a/kernel/sched/core.c
>> +++ b/kernel/sched/core.c
>> @@ -15,6 +15,7 @@
>>  #include <linux/init_task.h>
>>  #include <linux/context_tracking.h>
>>  #include <linux/rcupdate_wait.h>
>> +#include <linux/prcu.h>
>>
>>  #include <linux/blkdev.h>
>>  #include <linux/kprobes.h>
>> @@ -3383,6 +3384,7 @@ static void __sched notrace __schedule(bool preempt)
>>
>>       local_irq_disable();
>>       rcu_note_context_switch(preempt);
>> +     prcu_note_context_switch();
>>
>>       /*
>>        * Make sure that signal_pending_state()->signal_pending() below
>> --
>> 2.14.1.729.g59c0ea183
>>
>

Powered by blists - more mailing lists