Merge branch 'remove-qdisc-running-counter'

Sebastian Andrzej Siewior says:

====================
Try to simplify the gnet_stats and remove qdisc->running sequence counter.

The first few patches is a follow up to
    https://lore.kernel.org/all/20211007175000.2334713-1-bigeasy@linutronix.de/

The remaining patches (#5+) remove the seqcount_t (Qdisc::running) from
the Qdisc. The statistics (Qdisc::bstats and Qdisc::cpu_bstats) use
u64_stats_t and the "running state" is now represented by a bit in
Qdisc::state.

By removing the seqcount_t from Qdisc and decoupling the bstats
statistics from the seqcount_t it is possible to query the statistics
even if the Qdisc is running instead of waiting until it is idle again.

The try-lock like usage of the seqcount_t in qdisc_run_begin() is
problematic on PREEMPT_RT. Inside the qdisc_run_begin/end() qdisc->running
sequence counter write sections, at sch_direct_xmit(), the seqcount write
serialization lock is released then re-acquired. This is fine for !RT, because
the writer is in a BH disabled region and there is a no in-IRQ reader. For RT
though, BH sections are preemptible. The earlier introduced seqcount_LOCKNAME_t
mechanism, which for RT the reader acquires then relesaes the write
serailization lock to avoid infinite spinning if it preempts a seqcount write
section, cannot work: the qdisc->running write serialization lock is already
intermittingly released inside the seqcount write section.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>

2 files changed, 10 insertions, 4 deletions

diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index 173984414f38..f9cd6fea213f 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h
@@ -1916,7 +1916,6 @@ enum netdev_ml_priv_type {
  *	@sfp_bus:	attached &struct sfp_bus structure.
  *
  *	@qdisc_tx_busylock: lockdep class annotating Qdisc->busylock spinlock
- *	@qdisc_running_key: lockdep class annotating Qdisc->running seqcount
  *
  *	@proto_down:	protocol port state information can be sent to the
  *			switch driver and used to set the phys state of the
@@ -2250,7 +2249,6 @@ struct net_device {
 	struct phy_device	*phydev;
 	struct sfp_bus		*sfp_bus;
 	struct lock_class_key	*qdisc_tx_busylock;
-	struct lock_class_key	*qdisc_running_key;
 	bool			proto_down;
 	unsigned		wol_enabled:1;
 	unsigned		threaded:1;
@@ -2360,13 +2358,11 @@ static inline void netdev_for_each_tx_queue(struct net_device *dev,
 #define netdev_lockdep_set_classes(dev)				\
 {								\
 	static struct lock_class_key qdisc_tx_busylock_key;	\
-	static struct lock_class_key qdisc_running_key;		\
 	static struct lock_class_key qdisc_xmit_lock_key;	\
 	static struct lock_class_key dev_addr_list_lock_key;	\
 	unsigned int i;						\
 								\
 	(dev)->qdisc_tx_busylock = &qdisc_tx_busylock_key;	\
-	(dev)->qdisc_running_key = &qdisc_running_key;		\
 	lockdep_set_class(&(dev)->addr_list_lock,		\
 			  &dev_addr_list_lock_key);		\
 	for (i = 0; i < (dev)->num_tx_queues; i++)		\
diff --git a/include/linux/u64_stats_sync.h b/include/linux/u64_stats_sync.h
index e81856c0ba13..e8ec116c916b 100644
--- a/include/linux/u64_stats_sync.h
+++ b/include/linux/u64_stats_sync.h
@@ -83,6 +83,11 @@ static inline u64 u64_stats_read(const u64_stats_t *p)
 	return local64_read(&p->v);
 }
 
+static inline void u64_stats_set(u64_stats_t *p, u64 val)
+{
+	local64_set(&p->v, val);
+}
+
 static inline void u64_stats_add(u64_stats_t *p, unsigned long val)
 {
 	local64_add(val, &p->v);
@@ -104,6 +109,11 @@ static inline u64 u64_stats_read(const u64_stats_t *p)
 	return p->v;
 }
 
+static inline void u64_stats_set(u64_stats_t *p, u64 val)
+{
+	p->v = val;
+}
+
 static inline void u64_stats_add(u64_stats_t *p, unsigned long val)
 {
 	p->v += val;