srcu: Move work-scheduling fields from srcu_struct to srcu_usage
authorPaul E. McKenney <paulmck@kernel.org>
Sat, 18 Mar 2023 04:30:32 +0000 (21:30 -0700)
committerPaul E. McKenney <paulmck@kernel.org>
Tue, 4 Apr 2023 15:36:37 +0000 (08:36 -0700)
This commit moves the ->reschedule_jiffies, ->reschedule_count, and
->work fields from the srcu_struct structure to the srcu_usage structure
to reduce the size of the former in order to improve cache locality.

However, this means that the container_of() calls cannot get a pointer
to the srcu_struct because they are no longer in the srcu_struct.
This issue is addressed by adding a ->srcu_ssp field in the srcu_usage
structure that references the corresponding srcu_struct structure.
And given the presence of the sup pointer to the srcu_usage structure,
replace some ssp->srcu_usage-> instances with sup->.

[ paulmck Apply feedback from kernel test robot. ]

Link: https://lore.kernel.org/oe-kbuild-all/202303191400.iO5BOqka-lkp@intel.com/
Suggested-by: Christoph Hellwig <hch@lst.de>
Tested-by: Sachin Sant <sachinp@linux.ibm.com>
Tested-by: "Zhang, Qiang1" <qiang1.zhang@intel.com>
Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
include/linux/srcutree.h
kernel/rcu/srcutree.c

index d544ec1c0c8ef4efb0ddb407663daa01a16e29e4..cd0cdd8142c51e567133f1a1c46baf280f967e22 100644 (file)
@@ -84,6 +84,10 @@ struct srcu_usage {
        atomic_t srcu_barrier_cpu_cnt;          /* # CPUs not yet posting a */
                                                /*  callback for the barrier */
                                                /*  operation. */
+       unsigned long reschedule_jiffies;
+       unsigned long reschedule_count;
+       struct delayed_work work;
+       struct srcu_struct *srcu_ssp;
 };
 
 /*
@@ -92,9 +96,6 @@ struct srcu_usage {
 struct srcu_struct {
        unsigned int srcu_idx;                  /* Current rdr array element. */
        struct srcu_data __percpu *sda;         /* Per-CPU srcu_data array. */
-       unsigned long reschedule_jiffies;
-       unsigned long reschedule_count;
-       struct delayed_work work;
        struct lockdep_map dep_map;
        struct srcu_usage *srcu_sup;            /* Update-side data. */
 };
@@ -119,10 +120,10 @@ struct srcu_struct {
 {                                                                                              \
        .lock = __SPIN_LOCK_UNLOCKED(name.lock),                                                \
        .srcu_gp_seq_needed = -1UL,                                                             \
+       .work = __DELAYED_WORK_INITIALIZER(name.work, NULL, 0),                                 \
 }
 
 #define __SRCU_STRUCT_INIT_COMMON(name, usage_name)                                            \
-       .work = __DELAYED_WORK_INITIALIZER(name.work, NULL, 0),                                 \
        .srcu_sup = &usage_name,                                                                \
        __SRCU_DEP_MAP_INIT(name)
 
index 97d1fe9a160cce5eec5d09f5d714722b5910bc7d..169a6513b73979ca9bd2baac92eba9f932642a79 100644 (file)
@@ -251,7 +251,7 @@ static int init_srcu_struct_fields(struct srcu_struct *ssp, bool is_static)
        ssp->srcu_sup->srcu_barrier_seq = 0;
        mutex_init(&ssp->srcu_sup->srcu_barrier_mutex);
        atomic_set(&ssp->srcu_sup->srcu_barrier_cpu_cnt, 0);
-       INIT_DELAYED_WORK(&ssp->work, process_srcu);
+       INIT_DELAYED_WORK(&ssp->srcu_sup->work, process_srcu);
        ssp->srcu_sup->sda_is_static = is_static;
        if (!is_static)
                ssp->sda = alloc_percpu(struct srcu_data);
@@ -275,6 +275,7 @@ static int init_srcu_struct_fields(struct srcu_struct *ssp, bool is_static)
                        WRITE_ONCE(ssp->srcu_sup->srcu_size_state, SRCU_SIZE_BIG);
                }
        }
+       ssp->srcu_sup->srcu_ssp = ssp;
        smp_store_release(&ssp->srcu_sup->srcu_gp_seq_needed, 0); /* Init done. */
        return 0;
 }
@@ -647,7 +648,7 @@ void cleanup_srcu_struct(struct srcu_struct *ssp)
                return; /* Just leak it! */
        if (WARN_ON(srcu_readers_active(ssp)))
                return; /* Just leak it! */
-       flush_delayed_work(&ssp->work);
+       flush_delayed_work(&ssp->srcu_sup->work);
        for_each_possible_cpu(cpu) {
                struct srcu_data *sdp = per_cpu_ptr(ssp->sda, cpu);
 
@@ -1059,10 +1060,10 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
                // can only be executed during early boot when there is only
                // the one boot CPU running with interrupts still disabled.
                if (likely(srcu_init_done))
-                       queue_delayed_work(rcu_gp_wq, &ssp->work,
+                       queue_delayed_work(rcu_gp_wq, &ssp->srcu_sup->work,
                                           !!srcu_get_delay(ssp));
-               else if (list_empty(&ssp->work.work.entry))
-                       list_add(&ssp->work.work.entry, &srcu_boot_list);
+               else if (list_empty(&ssp->srcu_sup->work.work.entry))
+                       list_add(&ssp->srcu_sup->work.work.entry, &srcu_boot_list);
        }
        spin_unlock_irqrestore_rcu_node(ssp->srcu_sup, flags);
 }
@@ -1723,7 +1724,7 @@ static void srcu_reschedule(struct srcu_struct *ssp, unsigned long delay)
        spin_unlock_irq_rcu_node(ssp->srcu_sup);
 
        if (pushgp)
-               queue_delayed_work(rcu_gp_wq, &ssp->work, delay);
+               queue_delayed_work(rcu_gp_wq, &ssp->srcu_sup->work, delay);
 }
 
 /*
@@ -1734,22 +1735,24 @@ static void process_srcu(struct work_struct *work)
        unsigned long curdelay;
        unsigned long j;
        struct srcu_struct *ssp;
+       struct srcu_usage *sup;
 
-       ssp = container_of(work, struct srcu_struct, work.work);
+       sup = container_of(work, struct srcu_usage, work.work);
+       ssp = sup->srcu_ssp;
 
        srcu_advance_state(ssp);
        curdelay = srcu_get_delay(ssp);
        if (curdelay) {
-               WRITE_ONCE(ssp->reschedule_count, 0);
+               WRITE_ONCE(sup->reschedule_count, 0);
        } else {
                j = jiffies;
-               if (READ_ONCE(ssp->reschedule_jiffies) == j) {
-                       WRITE_ONCE(ssp->reschedule_count, READ_ONCE(ssp->reschedule_count) + 1);
-                       if (READ_ONCE(ssp->reschedule_count) > srcu_max_nodelay)
+               if (READ_ONCE(sup->reschedule_jiffies) == j) {
+                       WRITE_ONCE(sup->reschedule_count, READ_ONCE(sup->reschedule_count) + 1);
+                       if (READ_ONCE(sup->reschedule_count) > srcu_max_nodelay)
                                curdelay = 1;
                } else {
-                       WRITE_ONCE(ssp->reschedule_count, 1);
-                       WRITE_ONCE(ssp->reschedule_jiffies, j);
+                       WRITE_ONCE(sup->reschedule_count, 1);
+                       WRITE_ONCE(sup->reschedule_jiffies, j);
                }
        }
        srcu_reschedule(ssp, curdelay);
@@ -1848,7 +1851,7 @@ early_initcall(srcu_bootup_announce);
 
 void __init srcu_init(void)
 {
-       struct srcu_struct *ssp;
+       struct srcu_usage *sup;
 
        /* Decide on srcu_struct-size strategy. */
        if (SRCU_SIZING_IS(SRCU_SIZING_AUTO)) {
@@ -1868,13 +1871,13 @@ void __init srcu_init(void)
         */
        srcu_init_done = true;
        while (!list_empty(&srcu_boot_list)) {
-               ssp = list_first_entry(&srcu_boot_list, struct srcu_struct,
+               sup = list_first_entry(&srcu_boot_list, struct srcu_usage,
                                      work.work.entry);
-               list_del_init(&ssp->work.work.entry);
+               list_del_init(&sup->work.work.entry);
                if (SRCU_SIZING_IS(SRCU_SIZING_INIT) &&
-                   ssp->srcu_sup->srcu_size_state == SRCU_SIZE_SMALL)
-                       ssp->srcu_sup->srcu_size_state = SRCU_SIZE_ALLOC;
-               queue_work(rcu_gp_wq, &ssp->work.work);
+                   sup->srcu_size_state == SRCU_SIZE_SMALL)
+                       sup->srcu_size_state = SRCU_SIZE_ALLOC;
+               queue_work(rcu_gp_wq, &sup->work.work);
        }
 }