aboutsummaryrefslogtreecommitdiffstats
path: root/kernel
diff options
context:
space:
mode:
authorOleg Nesterov <oleg@redhat.com>2019-04-25 18:50:55 +0200
committerPaul E. McKenney <paulmck@linux.ibm.com>2019-05-28 09:05:23 -0700
commit89da3b94bb97417ca2c5b0ce3a28643819030247 (patch)
tree3a494bdabb013141bbb60f7380b5a0f64d23872b /kernel
parent3f2947b78151ec938dc06aea4ba0e11e56becdff (diff)
downloadtip-89da3b94bb97417ca2c5b0ce3a28643819030247.tar.gz
tip-89da3b94bb97417ca2c5b0ce3a28643819030247.tar.xz
tip-89da3b94bb97417ca2c5b0ce3a28643819030247.zip
rcu/sync: Simplify the state machine
With this patch rcu_sync has a single state variable and the transition rules become really simple: GP_IDLE - owned by the first rcu_sync_enter() which moves it to GP_ENTER - owned by rcu-callback which moves it to GP_PASSED - owned by the last rcu_sync_exit() which moves it to GP_EXIT - and this is the only "nontrivial" state. rcu-callback moves it back to GP_IDLE unless another enter() comes before a GP pass. If rcu-callback is invoked before the next rcu_sync_exit() it must see gp_count incremented by that enter() and set GP_PASSED. Otherwise, if the next rcu_sync_exit() wins the race, it will move it to GP_REPLAY - owned by rcu-callback which moves it to GP_EXIT Signed-off-by: Oleg Nesterov <oleg@redhat.com> [ paulmck: While here, apply READ_ONCE() and WRITE_ONCE() to ->gp_state. ] [ paulmck: Tweaks to make htmldocs happy. (Reported by kbuild test robot.) ] Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
Diffstat (limited to 'kernel')
-rw-r--r--kernel/rcu/sync.c165
1 files changed, 95 insertions, 70 deletions
diff --git a/kernel/rcu/sync.c b/kernel/rcu/sync.c
index ee427e138dad..d4558ab7a07d 100644
--- a/kernel/rcu/sync.c
+++ b/kernel/rcu/sync.c
@@ -10,15 +10,13 @@
#include <linux/rcu_sync.h>
#include <linux/sched.h>
-enum { GP_IDLE = 0, GP_PENDING, GP_PASSED };
-enum { CB_IDLE = 0, CB_PENDING, CB_REPLAY };
+enum { GP_IDLE = 0, GP_ENTER, GP_PASSED, GP_EXIT, GP_REPLAY };
#define rss_lock gp_wait.lock
/**
* rcu_sync_init() - Initialize an rcu_sync structure
* @rsp: Pointer to rcu_sync structure to be initialized
- * @type: Flavor of RCU with which to synchronize rcu_sync structure
*/
void rcu_sync_init(struct rcu_sync *rsp)
{
@@ -41,56 +39,26 @@ void rcu_sync_enter_start(struct rcu_sync *rsp)
rsp->gp_state = GP_PASSED;
}
-/**
- * rcu_sync_enter() - Force readers onto slowpath
- * @rsp: Pointer to rcu_sync structure to use for synchronization
- *
- * This function is used by updaters who need readers to make use of
- * a slowpath during the update. After this function returns, all
- * subsequent calls to rcu_sync_is_idle() will return false, which
- * tells readers to stay off their fastpaths. A later call to
- * rcu_sync_exit() re-enables reader slowpaths.
- *
- * When called in isolation, rcu_sync_enter() must wait for a grace
- * period, however, closely spaced calls to rcu_sync_enter() can
- * optimize away the grace-period wait via a state machine implemented
- * by rcu_sync_enter(), rcu_sync_exit(), and rcu_sync_func().
- */
-void rcu_sync_enter(struct rcu_sync *rsp)
-{
- bool need_wait, need_sync;
- spin_lock_irq(&rsp->rss_lock);
- need_wait = rsp->gp_count++;
- need_sync = rsp->gp_state == GP_IDLE;
- if (need_sync)
- rsp->gp_state = GP_PENDING;
- spin_unlock_irq(&rsp->rss_lock);
+static void rcu_sync_func(struct rcu_head *rhp);
- WARN_ON_ONCE(need_wait && need_sync);
- if (need_sync) {
- synchronize_rcu();
- rsp->gp_state = GP_PASSED;
- wake_up_all(&rsp->gp_wait);
- } else if (need_wait) {
- wait_event(rsp->gp_wait, rsp->gp_state == GP_PASSED);
- } else {
- /*
- * Possible when there's a pending CB from a rcu_sync_exit().
- * Nobody has yet been allowed the 'fast' path and thus we can
- * avoid doing any sync(). The callback will get 'dropped'.
- */
- WARN_ON_ONCE(rsp->gp_state != GP_PASSED);
- }
+static void rcu_sync_call(struct rcu_sync *rsp)
+{
+ call_rcu(&rsp->cb_head, rcu_sync_func);
}
/**
* rcu_sync_func() - Callback function managing reader access to fastpath
* @rhp: Pointer to rcu_head in rcu_sync structure to use for synchronization
*
- * This function is passed to one of the call_rcu() functions by
+ * This function is passed to call_rcu() function by rcu_sync_enter() and
* rcu_sync_exit(), so that it is invoked after a grace period following the
- * that invocation of rcu_sync_exit(). It takes action based on events that
+ * that invocation of enter/exit.
+ *
+ * If it is called by rcu_sync_enter() it signals that all the readers were
+ * switched onto slow path.
+ *
+ * If it is called by rcu_sync_exit() it takes action based on events that
* have taken place in the meantime, so that closely spaced rcu_sync_enter()
* and rcu_sync_exit() pairs need not wait for a grace period.
*
@@ -107,35 +75,88 @@ static void rcu_sync_func(struct rcu_head *rhp)
struct rcu_sync *rsp = container_of(rhp, struct rcu_sync, cb_head);
unsigned long flags;
- WARN_ON_ONCE(rsp->gp_state != GP_PASSED);
- WARN_ON_ONCE(rsp->cb_state == CB_IDLE);
+ WARN_ON_ONCE(READ_ONCE(rsp->gp_state) == GP_IDLE);
+ WARN_ON_ONCE(READ_ONCE(rsp->gp_state) == GP_PASSED);
spin_lock_irqsave(&rsp->rss_lock, flags);
if (rsp->gp_count) {
/*
- * A new rcu_sync_begin() has happened; drop the callback.
+ * We're at least a GP after the GP_IDLE->GP_ENTER transition.
*/
- rsp->cb_state = CB_IDLE;
- } else if (rsp->cb_state == CB_REPLAY) {
+ WRITE_ONCE(rsp->gp_state, GP_PASSED);
+ wake_up_locked(&rsp->gp_wait);
+ } else if (rsp->gp_state == GP_REPLAY) {
/*
- * A new rcu_sync_exit() has happened; requeue the callback
- * to catch a later GP.
+ * A new rcu_sync_exit() has happened; requeue the callback to
+ * catch a later GP.
*/
- rsp->cb_state = CB_PENDING;
- call_rcu(&rsp->cb_head, rcu_sync_func);
+ WRITE_ONCE(rsp->gp_state, GP_EXIT);
+ rcu_sync_call(rsp);
} else {
/*
- * We're at least a GP after rcu_sync_exit(); eveybody will now
- * have observed the write side critical section. Let 'em rip!.
+ * We're at least a GP after the last rcu_sync_exit(); eveybody
+ * will now have observed the write side critical section.
+ * Let 'em rip!.
*/
- rsp->cb_state = CB_IDLE;
- rsp->gp_state = GP_IDLE;
+ WRITE_ONCE(rsp->gp_state, GP_IDLE);
}
spin_unlock_irqrestore(&rsp->rss_lock, flags);
}
/**
- * rcu_sync_exit() - Allow readers back onto fast patch after grace period
+ * rcu_sync_enter() - Force readers onto slowpath
+ * @rsp: Pointer to rcu_sync structure to use for synchronization
+ *
+ * This function is used by updaters who need readers to make use of
+ * a slowpath during the update. After this function returns, all
+ * subsequent calls to rcu_sync_is_idle() will return false, which
+ * tells readers to stay off their fastpaths. A later call to
+ * rcu_sync_exit() re-enables reader slowpaths.
+ *
+ * When called in isolation, rcu_sync_enter() must wait for a grace
+ * period, however, closely spaced calls to rcu_sync_enter() can
+ * optimize away the grace-period wait via a state machine implemented
+ * by rcu_sync_enter(), rcu_sync_exit(), and rcu_sync_func().
+ */
+void rcu_sync_enter(struct rcu_sync *rsp)
+{
+ int gp_state;
+
+ spin_lock_irq(&rsp->rss_lock);
+ gp_state = rsp->gp_state;
+ if (gp_state == GP_IDLE) {
+ WRITE_ONCE(rsp->gp_state, GP_ENTER);
+ WARN_ON_ONCE(rsp->gp_count);
+ /*
+ * Note that we could simply do rcu_sync_call(rsp) here and
+ * avoid the "if (gp_state == GP_IDLE)" block below.
+ *
+ * However, synchronize_rcu() can be faster if rcu_expedited
+ * or rcu_blocking_is_gp() is true.
+ *
+ * Another reason is that we can't wait for rcu callback if
+ * we are called at early boot time but this shouldn't happen.
+ */
+ }
+ rsp->gp_count++;
+ spin_unlock_irq(&rsp->rss_lock);
+
+ if (gp_state == GP_IDLE) {
+ /*
+ * See the comment above, this simply does the "synchronous"
+ * call_rcu(rcu_sync_func) which does GP_ENTER -> GP_PASSED.
+ */
+ synchronize_rcu();
+ rcu_sync_func(&rsp->cb_head);
+ /* Not really needed, wait_event() would see GP_PASSED. */
+ return;
+ }
+
+ wait_event(rsp->gp_wait, READ_ONCE(rsp->gp_state) >= GP_PASSED);
+}
+
+/**
+ * rcu_sync_exit() - Allow readers back onto fast path after grace period
* @rsp: Pointer to rcu_sync structure to use for synchronization
*
* This function is used by updaters who have completed, and can therefore
@@ -146,13 +167,16 @@ static void rcu_sync_func(struct rcu_head *rhp)
*/
void rcu_sync_exit(struct rcu_sync *rsp)
{
+ WARN_ON_ONCE(READ_ONCE(rsp->gp_state) == GP_IDLE);
+ WARN_ON_ONCE(READ_ONCE(rsp->gp_count) == 0);
+
spin_lock_irq(&rsp->rss_lock);
if (!--rsp->gp_count) {
- if (rsp->cb_state == CB_IDLE) {
- rsp->cb_state = CB_PENDING;
- call_rcu(&rsp->cb_head, rcu_sync_func);
- } else if (rsp->cb_state == CB_PENDING) {
- rsp->cb_state = CB_REPLAY;
+ if (rsp->gp_state == GP_PASSED) {
+ WRITE_ONCE(rsp->gp_state, GP_EXIT);
+ rcu_sync_call(rsp);
+ } else if (rsp->gp_state == GP_EXIT) {
+ WRITE_ONCE(rsp->gp_state, GP_REPLAY);
}
}
spin_unlock_irq(&rsp->rss_lock);
@@ -164,18 +188,19 @@ void rcu_sync_exit(struct rcu_sync *rsp)
*/
void rcu_sync_dtor(struct rcu_sync *rsp)
{
- int cb_state;
+ int gp_state;
- WARN_ON_ONCE(rsp->gp_count);
+ WARN_ON_ONCE(READ_ONCE(rsp->gp_count));
+ WARN_ON_ONCE(READ_ONCE(rsp->gp_state) == GP_PASSED);
spin_lock_irq(&rsp->rss_lock);
- if (rsp->cb_state == CB_REPLAY)
- rsp->cb_state = CB_PENDING;
- cb_state = rsp->cb_state;
+ if (rsp->gp_state == GP_REPLAY)
+ WRITE_ONCE(rsp->gp_state, GP_EXIT);
+ gp_state = rsp->gp_state;
spin_unlock_irq(&rsp->rss_lock);
- if (cb_state != CB_IDLE) {
+ if (gp_state != GP_IDLE) {
rcu_barrier();
- WARN_ON_ONCE(rsp->cb_state != CB_IDLE);
+ WARN_ON_ONCE(rsp->gp_state != GP_IDLE);
}
}