Re: [PATCH V2 7/7] x86,rcu: use percpu rcu_preempt_depth

Lai Jiangshan <laijs@xxxxxxxxxxxxxxxxx> · Wed, 20 Nov 2019 10:47:53 +0800

On 2019/11/20 5:14 上午, Paul E. McKenney wrote:
On Tue, Nov 19, 2019 at 09:59:03AM +0800, Lai Jiangshan wrote:


On 2019/11/18 10:59 下午, Paul E. McKenney wrote:
On Mon, Nov 18, 2019 at 10:02:50AM +0800, Lai Jiangshan wrote:


On 2019/11/16 11:48 下午, Paul E. McKenney wrote:
On Sat, Nov 02, 2019 at 12:45:59PM +0000, Lai Jiangshan wrote:
Convert x86 to use a per-cpu rcu_preempt_depth. The reason for doing so
is that accessing per-cpu variables is a lot cheaper than accessing
task_struct or thread_info variables.

We need to save/restore the actual rcu_preempt_depth when switch.
We also place the per-cpu rcu_preempt_depth close to __preempt_count
and current_task variable.

Using the idea of per-cpu __preempt_count.

No function call when using rcu_read_[un]lock().
Single instruction for rcu_read_lock().
2 instructions for fast path of rcu_read_unlock().

CC: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Signed-off-by: Lai Jiangshan <laijs@xxxxxxxxxxxxxxxxx>

Wouldn't putting RCU's nesting-depth counter in task info be just as fast,
just as nice for #include/inlining, and a lot less per-architecture code?

Or am I missing some issue with the task-info approach?

struct thread_info itself is per-architecture definition.
All the arches would have to be touched if RCU's nesting-depth counter
is put int struct thread_info. Though the inlining functions can be
defined in include/asm-generic/ so that it serves for all arches
and X86 can have its own implementation in arch/x86/include/asm/.

True enough.

But doesn't the per-CPU code require per-architecture changes to copy
to and from the per-CPU variable?  If that code simpler and smaller than
the thread_info access code, I will be -very- surprised.

The per-CPU code is not simpler. And my code touch X86 ONLY so that
it requires an additional CONFIG and some more "#if" in rcu code
which adds little more complicity.

As far as I understand, the rcu_read_lock_nesting can only be possible
put in per CPU in X86, not possible in other ARCHs.

Putting the rcu_read_lock_nesting in struct thread_info and putting
the inlining functions in include/asm-generic/ are OK for me. It will
be a good-shaped framework and reduce function calls, and still allow
x86 has its own implementation.

In addition, even for x86, the struct thread_info approach would avoid
adding more instructions to the context-switch path.

For all arches, these is no additional instructions to the
context-switch path for the struct thread_info approach.

The X86 per-cpu approach adds only a 2loads and 2 stores in the
context-switch path.

But there are tens of read_read_[un]lock calls between context-switchs
if not hundreds nor thousands. The read_read_[un]lock is more frequent
than context-switch. Every read_read_[un]lock pair needs about 20
instructions now, reducing the pair to 3 instructions in the percpu
approach is a win in my opinion.


The framework will only be more appealing when the x86 percpu
rcu_read_lock_nesting is implemented. This series implements
the x86 percpu rcu_read_lock_nesting first and avoid touch too much
files. The framework can be implemented later

If I remember correctly, you mentioned earlier that the struct
thread_info approach was just as few instructions for rcu_read_lock()
and rcu_read_unlock() as is the per-CPU approach.  If that is the
case, and if the struct thread_info approach avoids adding code to the
context-switch path, why would the x86 maintainers be at all happy with
the per-CPU approach?  What am I missing here?


The struct thread_info approach will save only function calls compared
with the current struct task_struct approach. The struct thread_info
approach and the current struct task_struct approach are the same
excepct that the later one has to do function calls to avoid c-header
files including hell.

These is no additional instructions to the context-switch path for the
struct thread_info approach. But I'm not sure whether each Arch
maintainers be at all happy with the struct thread_info approach
since the benefit is not so large as x86.

And maybe gcc LTO can reduce the function calls too, so that the
struct thread_info approach and the current struct task_struct
approach will be really exactly the same. But I know nothing about
LTO.

The struct thread_info approach will still need 12(=5+7) instructions
for read_read_[un]lock pair whereas the per-CPU approach needs only 3
instructions for every pair (fast path).

In x86, on one hand, this patchset reduces the number of instructions
needed for the read_read_[un]lock pair markedly (12 basic instructions
+ 2 functions calls => 3 instructions).

On the other hand, it can be compared with preempt_disable/enable()
pair, which had also been changed from the struct thread_info approach
by Peter Zijlstra <peterz@xxxxxxxxxxxxx> at c2daa3bed53a in 2013.
The importance of reducing the overhead of the preempt_disable/enable()
pair and the read_read_[un]lock pair is the same to me. Especially,
preempt_disable/enable() is discouraged except really needed, while
the usage of read_read_[un]lock is increasing.

Thanks
Lai


							Thanx, Paul

Thanks
Lai



							Thanx, Paul

---
    arch/x86/Kconfig                         |  2 +
    arch/x86/include/asm/rcu_preempt_depth.h | 87 ++++++++++++++++++++++++
    arch/x86/kernel/cpu/common.c             |  7 ++
    arch/x86/kernel/process_32.c             |  2 +
    arch/x86/kernel/process_64.c             |  2 +
    include/linux/rcupdate.h                 | 24 +++++++
    init/init_task.c                         |  2 +-
    kernel/fork.c                            |  2 +-
    kernel/rcu/Kconfig                       |  3 +
    kernel/rcu/tree_exp.h                    |  2 +
    kernel/rcu/tree_plugin.h                 | 37 +++++++---
    11 files changed, 158 insertions(+), 12 deletions(-)
    create mode 100644 arch/x86/include/asm/rcu_preempt_depth.h

diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
index d6e1faa28c58..af9fedc0fdc4 100644
--- a/arch/x86/Kconfig
+++ b/arch/x86/Kconfig
@@ -18,6 +18,7 @@ config X86_32
    	select MODULES_USE_ELF_REL
    	select OLD_SIGACTION
    	select GENERIC_VDSO_32
+	select ARCH_HAVE_RCU_PREEMPT_DEEPTH
    config X86_64
    	def_bool y
@@ -31,6 +32,7 @@ config X86_64
    	select NEED_DMA_MAP_STATE
    	select SWIOTLB
    	select ARCH_HAS_SYSCALL_WRAPPER
+	select ARCH_HAVE_RCU_PREEMPT_DEEPTH
    config FORCE_DYNAMIC_FTRACE
    	def_bool y
diff --git a/arch/x86/include/asm/rcu_preempt_depth.h b/arch/x86/include/asm/rcu_preempt_depth.h
new file mode 100644
index 000000000000..88010ad59c20
--- /dev/null
+++ b/arch/x86/include/asm/rcu_preempt_depth.h
@@ -0,0 +1,87 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef __ASM_RCU_PREEMPT_DEPTH_H
+#define __ASM_RCU_PREEMPT_DEPTH_H
+
+#include <asm/rmwcc.h>
+#include <asm/percpu.h>
+
+#ifdef CONFIG_PREEMPT_RCU
+DECLARE_PER_CPU(int, __rcu_preempt_depth);
+
+/*
+ * We use the RCU_NEED_SPECIAL bit as an inverted need_special
+ * such that a decrement hitting 0 means we can and should do
+ * rcu_read_unlock_special().
+ */
+#define RCU_NEED_SPECIAL	0x80000000
+
+#define INIT_RCU_PREEMPT_DEPTH	(RCU_NEED_SPECIAL)
+
+/* We mask the RCU_NEED_SPECIAL bit so that it return real depth */
+static __always_inline int rcu_preempt_depth(void)
+{
+	return raw_cpu_read_4(__rcu_preempt_depth) & ~RCU_NEED_SPECIAL;
+}
+
+static __always_inline void rcu_preempt_depth_set(int pc)
+{
+	int old, new;
+
+	do {
+		old = raw_cpu_read_4(__rcu_preempt_depth);
+		new = (old & RCU_NEED_SPECIAL) |
+			(pc & ~RCU_NEED_SPECIAL);
+	} while (raw_cpu_cmpxchg_4(__rcu_preempt_depth, old, new) != old);
+}
+
+/*
+ * We fold the RCU_NEED_SPECIAL bit into the rcu_preempt_depth such that
+ * rcu_read_unlock() can decrement and test for needing to do special
+ * with a single instruction.
+ *
+ * We invert the actual bit, so that when the decrement hits 0 we know
+ * both it just exited the outmost rcu_read_lock() critical section and
+ * we need to do specail (the bit is cleared) if it doesn't need to be
+ * deferred.
+ */
+
+static inline void set_rcu_preempt_need_special(void)
+{
+	raw_cpu_and_4(__rcu_preempt_depth, ~RCU_NEED_SPECIAL);
+}
+
+/*
+ * irq needs to be disabled for clearing any bits of ->rcu_read_unlock_special
+ * and calling this function. Otherwise it may clear the work done
+ * by set_rcu_preempt_need_special() in interrupt.
+ */
+static inline void clear_rcu_preempt_need_special(void)
+{
+	raw_cpu_or_4(__rcu_preempt_depth, RCU_NEED_SPECIAL);
+}
+
+static __always_inline void rcu_preempt_depth_inc(void)
+{
+	raw_cpu_add_4(__rcu_preempt_depth, 1);
+}
+
+static __always_inline bool rcu_preempt_depth_dec_and_test(void)
+{
+	return GEN_UNARY_RMWcc("decl", __rcu_preempt_depth, e, __percpu_arg([var]));
+}
+
+/* must be macros to avoid header recursion hell */
+#define save_restore_rcu_preempt_depth(prev_p, next_p) do {				\
+		prev_p->rcu_read_lock_nesting = this_cpu_read(__rcu_preempt_depth);	\
+		this_cpu_write(__rcu_preempt_depth, next_p->rcu_read_lock_nesting);	\
+	} while (0)
+
+#define DEFINE_PERCPU_RCU_PREEMP_DEPTH						\
+	DEFINE_PER_CPU(int, __rcu_preempt_depth) = INIT_RCU_PREEMPT_DEPTH;	\
+	EXPORT_PER_CPU_SYMBOL(__rcu_preempt_depth)
+#else /* #ifdef CONFIG_PREEMPT_RCU */
+#define save_restore_rcu_preempt_depth(prev_p, next_p) do {} while (0)
+#define DEFINE_PERCPU_RCU_PREEMP_DEPTH	/* empty */
+#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
+
+#endif /* __ASM_RCU_PREEMPT_DEPTH_H */
diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c
index 9ae7d1bcd4f4..0151737e196c 100644
--- a/arch/x86/kernel/cpu/common.c
+++ b/arch/x86/kernel/cpu/common.c
@@ -46,6 +46,7 @@
    #include <asm/asm.h>
    #include <asm/bugs.h>
    #include <asm/cpu.h>
+#include <asm/rcu_preempt_depth.h>
    #include <asm/mce.h>
    #include <asm/msr.h>
    #include <asm/pat.h>
@@ -1633,6 +1634,9 @@ DEFINE_PER_CPU(unsigned int, irq_count) __visible = -1;
    DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
    EXPORT_PER_CPU_SYMBOL(__preempt_count);
+/* close to __preempt_count */
+DEFINE_PERCPU_RCU_PREEMP_DEPTH;
+
    /* May not be marked __init: used by software suspend */
    void syscall_init(void)
    {
@@ -1690,6 +1694,9 @@ EXPORT_PER_CPU_SYMBOL(current_task);
    DEFINE_PER_CPU(int, __preempt_count) = INIT_PREEMPT_COUNT;
    EXPORT_PER_CPU_SYMBOL(__preempt_count);
+/* close to __preempt_count */
+DEFINE_PERCPU_RCU_PREEMP_DEPTH;
+
    /*
     * On x86_32, vm86 modifies tss.sp0, so sp0 isn't a reliable way to find
     * the top of the kernel stack.  Use an extra percpu variable to track the
diff --git a/arch/x86/kernel/process_32.c b/arch/x86/kernel/process_32.c
index b8ceec4974fe..ab1f20353663 100644
--- a/arch/x86/kernel/process_32.c
+++ b/arch/x86/kernel/process_32.c
@@ -51,6 +51,7 @@
    #include <asm/cpu.h>
    #include <asm/syscalls.h>
    #include <asm/debugreg.h>
+#include <asm/rcu_preempt_depth.h>
    #include <asm/switch_to.h>
    #include <asm/vm86.h>
    #include <asm/resctrl_sched.h>
@@ -290,6 +291,7 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
    	if (prev->gs | next->gs)
    		lazy_load_gs(next->gs);
+	save_restore_rcu_preempt_depth(prev_p, next_p);
    	this_cpu_write(current_task, next_p);
    	switch_fpu_finish(next_fpu);
diff --git a/arch/x86/kernel/process_64.c b/arch/x86/kernel/process_64.c
index af64519b2695..2e1c6e829d30 100644
--- a/arch/x86/kernel/process_64.c
+++ b/arch/x86/kernel/process_64.c
@@ -50,6 +50,7 @@
    #include <asm/ia32.h>
    #include <asm/syscalls.h>
    #include <asm/debugreg.h>
+#include <asm/rcu_preempt_depth.h>
    #include <asm/switch_to.h>
    #include <asm/xen/hypervisor.h>
    #include <asm/vdso.h>
@@ -559,6 +560,7 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
    	x86_fsgsbase_load(prev, next);
+	save_restore_rcu_preempt_depth(prev_p, next_p);
    	/*
    	 * Switch the PDA and FPU contexts.
    	 */
diff --git a/include/linux/rcupdate.h b/include/linux/rcupdate.h
index a35daab95d14..0d2abf08b694 100644
--- a/include/linux/rcupdate.h
+++ b/include/linux/rcupdate.h
@@ -41,6 +41,29 @@ void synchronize_rcu(void);
    #ifdef CONFIG_PREEMPT_RCU
+#ifdef CONFIG_ARCH_HAVE_RCU_PREEMPT_DEEPTH
+#include <asm/rcu_preempt_depth.h>
+
+#ifndef CONFIG_PROVE_LOCKING
+extern void rcu_read_unlock_special(void);
+
+static inline void __rcu_read_lock(void)
+{
+	rcu_preempt_depth_inc();
+}
+
+static inline void __rcu_read_unlock(void)
+{
+	if (unlikely(rcu_preempt_depth_dec_and_test()))
+		rcu_read_unlock_special();
+}
+#else
+void __rcu_read_lock(void);
+void __rcu_read_unlock(void);
+#endif
+
+#else /* #ifdef CONFIG_ARCH_HAVE_RCU_PREEMPT_DEEPTH */
+#define INIT_RCU_PREEMPT_DEPTH (0)
    void __rcu_read_lock(void);
    void __rcu_read_unlock(void);
@@ -51,6 +74,7 @@ void __rcu_read_unlock(void);
     * types of kernel builds, the rcu_read_lock() nesting depth is unknowable.
     */
    #define rcu_preempt_depth() (current->rcu_read_lock_nesting)
+#endif /* #else #ifdef CONFIG_ARCH_HAVE_RCU_PREEMPT_DEEPTH */
    #else /* #ifdef CONFIG_PREEMPT_RCU */
diff --git a/init/init_task.c b/init/init_task.c
index 9e5cbe5eab7b..0a91e38fba37 100644
--- a/init/init_task.c
+++ b/init/init_task.c
@@ -130,7 +130,7 @@ struct task_struct init_task
    	.perf_event_list = LIST_HEAD_INIT(init_task.perf_event_list),
    #endif
    #ifdef CONFIG_PREEMPT_RCU
-	.rcu_read_lock_nesting = 0,
+	.rcu_read_lock_nesting = INIT_RCU_PREEMPT_DEPTH,
    	.rcu_read_unlock_special.s = 0,
    	.rcu_node_entry = LIST_HEAD_INIT(init_task.rcu_node_entry),
    	.rcu_blocked_node = NULL,
diff --git a/kernel/fork.c b/kernel/fork.c
index f9572f416126..7368d4ccb857 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -1665,7 +1665,7 @@ init_task_pid(struct task_struct *task, enum pid_type type, struct pid *pid)
    static inline void rcu_copy_process(struct task_struct *p)
    {
    #ifdef CONFIG_PREEMPT_RCU
-	p->rcu_read_lock_nesting = 0;
+	p->rcu_read_lock_nesting = INIT_RCU_PREEMPT_DEPTH;
    	p->rcu_read_unlock_special.s = 0;
    	p->rcu_blocked_node = NULL;
    	INIT_LIST_HEAD(&p->rcu_node_entry);
diff --git a/kernel/rcu/Kconfig b/kernel/rcu/Kconfig
index 1cc940fef17c..d2ecca49a1a4 100644
--- a/kernel/rcu/Kconfig
+++ b/kernel/rcu/Kconfig
@@ -14,6 +14,9 @@ config TREE_RCU
    	  thousands of CPUs.  It also scales down nicely to
    	  smaller systems.
+config ARCH_HAVE_RCU_PREEMPT_DEEPTH
+	def_bool n
+
    config PREEMPT_RCU
    	bool
    	default y if PREEMPTION
diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h
index dcb2124203cf..f919881832d4 100644
--- a/kernel/rcu/tree_exp.h
+++ b/kernel/rcu/tree_exp.h
@@ -588,6 +588,7 @@ static void wait_rcu_exp_gp(struct work_struct *wp)
    }
    #ifdef CONFIG_PREEMPT_RCU
+static inline void set_rcu_preempt_need_special(void);
    /*
     * Remote handler for smp_call_function_single().  If there is an
@@ -637,6 +638,7 @@ static void rcu_exp_handler(void *unused)
    		if (rnp->expmask & rdp->grpmask) {
    			rdp->exp_deferred_qs = true;
    			t->rcu_read_unlock_special.b.exp_hint = true;
+			set_rcu_preempt_need_special();
    		}
    		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
    		return;
diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
index e16c3867d2ff..e6774a7ab16b 100644
--- a/kernel/rcu/tree_plugin.h
+++ b/kernel/rcu/tree_plugin.h
@@ -82,7 +82,7 @@ static void __init rcu_bootup_announce_oddness(void)
    #ifdef CONFIG_PREEMPT_RCU
    static void rcu_report_exp_rnp(struct rcu_node *rnp, bool wake);
-static void rcu_read_unlock_special(struct task_struct *t);
+void rcu_read_unlock_special(void);
    /*
     * Tell them what RCU they are running.
@@ -298,6 +298,7 @@ void rcu_note_context_switch(bool preempt)
    		t->rcu_read_unlock_special.b.need_qs = false;
    		t->rcu_read_unlock_special.b.blocked = true;
    		t->rcu_blocked_node = rnp;
+		set_rcu_preempt_need_special();
    		/*
    		 * Verify the CPU's sanity, trace the preemption, and
@@ -345,6 +346,7 @@ static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
    /* Bias and limit values for ->rcu_read_lock_nesting. */
    #define RCU_NEST_PMAX (INT_MAX / 2)
+#ifndef CONFIG_ARCH_HAVE_RCU_PREEMPT_DEEPTH
    static inline void rcu_preempt_depth_inc(void)
    {
    	current->rcu_read_lock_nesting++;
@@ -352,7 +354,12 @@ static inline void rcu_preempt_depth_inc(void)
    static inline bool rcu_preempt_depth_dec_and_test(void)
    {
-	return --current->rcu_read_lock_nesting == 0;
+	if (--current->rcu_read_lock_nesting == 0) {
+		/* check speical after dec ->rcu_read_lock_nesting */
+		barrier();
+		return READ_ONCE(current->rcu_read_unlock_special.s);
+	}
+	return 0;
    }
    static inline void rcu_preempt_depth_set(int val)
@@ -360,6 +367,12 @@ static inline void rcu_preempt_depth_set(int val)
    	current->rcu_read_lock_nesting = val;
    }
+static inline void clear_rcu_preempt_need_special(void) {}
+static inline void set_rcu_preempt_need_special(void) {}
+
+#endif /* #ifndef CONFIG_ARCH_HAVE_RCU_PREEMPT_DEEPTH */
+
+#if !defined(CONFIG_ARCH_HAVE_RCU_PREEMPT_DEEPTH) || defined (CONFIG_PROVE_LOCKING)
    /*
     * Preemptible RCU implementation for rcu_read_lock().
     * Just increment ->rcu_read_lock_nesting, shared state will be updated
@@ -383,18 +396,16 @@ EXPORT_SYMBOL_GPL(__rcu_read_lock);
     */
    void __rcu_read_unlock(void)
    {
-	struct task_struct *t = current;
-
    	if (rcu_preempt_depth_dec_and_test()) {
-		barrier();  /* critical section before exit code. */
-		if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
-			rcu_read_unlock_special(t);
+		rcu_read_unlock_special();
    	}
    	if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
-		WARN_ON_ONCE(rcu_preempt_depth() < 0);
+		WARN_ON_ONCE(rcu_preempt_depth() < 0 ||
+			     rcu_preempt_depth() > RCU_NEST_PMAX);
    	}
    }
    EXPORT_SYMBOL_GPL(__rcu_read_unlock);
+#endif /* #if !defined(CONFIG_ARCH_HAVE_RCU_PREEMPT_DEEPTH) || defined (CONFIG_PROVE_LOCKING) */
    /*
     * Advance a ->blkd_tasks-list pointer to the next entry, instead
@@ -449,6 +460,7 @@ rcu_preempt_deferred_qs_irqrestore(struct task_struct *t, unsigned long flags)
    		return;
    	}
    	t->rcu_read_unlock_special.s = 0;
+	clear_rcu_preempt_need_special();
    	if (special.b.need_qs)
    		rcu_qs();
@@ -579,8 +591,9 @@ static void rcu_preempt_deferred_qs_handler(struct irq_work *iwp)
     * notify RCU core processing or task having blocked during the RCU
     * read-side critical section.
     */
-static void rcu_read_unlock_special(struct task_struct *t)
+void rcu_read_unlock_special(void)
    {
+	struct task_struct *t = current;
    	unsigned long flags;
    	bool preempt_bh_were_disabled =
    			!!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
@@ -631,6 +644,7 @@ static void rcu_read_unlock_special(struct task_struct *t)
    	}
    	rcu_preempt_deferred_qs_irqrestore(t, flags);
    }
+EXPORT_SYMBOL_GPL(rcu_read_unlock_special);
    /*
     * Check that the list of blocked tasks for the newly completed grace
@@ -694,8 +708,10 @@ static void rcu_flavor_sched_clock_irq(int user)
    	    __this_cpu_read(rcu_data.core_needs_qs) &&
    	    __this_cpu_read(rcu_data.cpu_no_qs.b.norm) &&
    	    !t->rcu_read_unlock_special.b.need_qs &&
-	    time_after(jiffies, rcu_state.gp_start + HZ))
+	    time_after(jiffies, rcu_state.gp_start + HZ)) {
    		t->rcu_read_unlock_special.b.need_qs = true;
+		set_rcu_preempt_need_special();
+	}
    }
    /*
@@ -714,6 +730,7 @@ void exit_rcu(void)
    		rcu_preempt_depth_set(1);
    		barrier();
    		WRITE_ONCE(t->rcu_read_unlock_special.b.blocked, true);
+		set_rcu_preempt_need_special();
    	} else if (unlikely(rcu_preempt_depth())) {
    		rcu_preempt_depth_set(1);
    	} else {
--
2.20.1







