sched/numa: Use {cpu, pid} to create task groups for shared faults
authorPeter Zijlstra <peterz@infradead.org>
Mon, 7 Oct 2013 10:29:21 +0000 (11:29 +0100)
committerIngo Molnar <mingo@kernel.org>
Wed, 9 Oct 2013 12:47:47 +0000 (14:47 +0200)
While parallel applications tend to align their data on the cache
boundary, they tend not to align on the page or THP boundary.
Consequently tasks that partition their data can still "false-share"
pages presenting a problem for optimal NUMA placement.

This patch uses NUMA hinting faults to chain tasks together into
numa_groups. As well as storing the NID a task was running on when
accessing a page a truncated representation of the faulting PID is
stored. If subsequent faults are from different PIDs it is reasonable
to assume that those two tasks share a page and are candidates for
being grouped together. Note that this patch makes no scheduling
decisions based on the grouping information.

Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Link: http://lkml.kernel.org/r/1381141781-10992-44-git-send-email-mgorman@suse.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
include/linux/mm.h
include/linux/sched.h
kernel/sched/core.c
kernel/sched/fair.c
kernel/sched/sched.h
mm/memory.c

index ce464cd4777eea900ec92f218675785cacd50d1d..81443d557a2e179a7d3768c63abc2519d9d1ed19 100644 (file)
@@ -691,6 +691,12 @@ static inline bool cpupid_cpu_unset(int cpupid)
        return cpupid_to_cpu(cpupid) == (-1 & LAST__CPU_MASK);
 }
 
+static inline bool __cpupid_match_pid(pid_t task_pid, int cpupid)
+{
+       return (task_pid & LAST__PID_MASK) == cpupid_to_pid(cpupid);
+}
+
+#define cpupid_match_pid(task, cpupid) __cpupid_match_pid(task->pid, cpupid)
 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
 static inline int page_cpupid_xchg_last(struct page *page, int cpupid)
 {
@@ -760,6 +766,11 @@ static inline bool cpupid_pid_unset(int cpupid)
 static inline void page_cpupid_reset_last(struct page *page)
 {
 }
+
+static inline bool cpupid_match_pid(struct task_struct *task, int cpupid)
+{
+       return false;
+}
 #endif /* CONFIG_NUMA_BALANCING */
 
 static inline struct zone *page_zone(const struct page *page)
index b6619792bb13348001094f55287713bf3025be45..f587ded5c148defea335e6406fa3545807a327f3 100644 (file)
@@ -1347,6 +1347,9 @@ struct task_struct {
        u64 node_stamp;                 /* migration stamp  */
        struct callback_head numa_work;
 
+       struct list_head numa_entry;
+       struct numa_group *numa_group;
+
        /*
         * Exponential decaying average of faults on a per-node basis.
         * Scheduling placement decisions are made based on the these counts.
index 1fe59da280e31501cb141a9eae2febee43f4e2a8..51092d5cc64c75980ea724e23c9cdeeb86acbc56 100644 (file)
@@ -1733,6 +1733,9 @@ static void __sched_fork(struct task_struct *p)
        p->numa_work.next = &p->numa_work;
        p->numa_faults = NULL;
        p->numa_faults_buffer = NULL;
+
+       INIT_LIST_HEAD(&p->numa_entry);
+       p->numa_group = NULL;
 #endif /* CONFIG_NUMA_BALANCING */
 }
 
index dbe0f628efa3c03963665cf32b86d07f9ecf4fd4..85565053a6ed2e626e532b59bbd678cd2dea91ed 100644 (file)
@@ -888,6 +888,17 @@ static unsigned int task_scan_max(struct task_struct *p)
  */
 unsigned int sysctl_numa_balancing_settle_count __read_mostly = 4;
 
+struct numa_group {
+       atomic_t refcount;
+
+       spinlock_t lock; /* nr_tasks, tasks */
+       int nr_tasks;
+       struct list_head task_list;
+
+       struct rcu_head rcu;
+       atomic_long_t faults[0];
+};
+
 static inline int task_faults_idx(int nid, int priv)
 {
        return 2 * nid + priv;
@@ -1182,7 +1193,10 @@ static void task_numa_placement(struct task_struct *p)
                int priv, i;
 
                for (priv = 0; priv < 2; priv++) {
+                       long diff;
+
                        i = task_faults_idx(nid, priv);
+                       diff = -p->numa_faults[i];
 
                        /* Decay existing window, copy faults since last scan */
                        p->numa_faults[i] >>= 1;
@@ -1190,6 +1204,11 @@ static void task_numa_placement(struct task_struct *p)
                        p->numa_faults_buffer[i] = 0;
 
                        faults += p->numa_faults[i];
+                       diff += p->numa_faults[i];
+                       if (p->numa_group) {
+                               /* safe because we can only change our own group */
+                               atomic_long_add(diff, &p->numa_group->faults[i]);
+                       }
                }
 
                if (faults > max_faults) {
@@ -1207,6 +1226,131 @@ static void task_numa_placement(struct task_struct *p)
        }
 }
 
+static inline int get_numa_group(struct numa_group *grp)
+{
+       return atomic_inc_not_zero(&grp->refcount);
+}
+
+static inline void put_numa_group(struct numa_group *grp)
+{
+       if (atomic_dec_and_test(&grp->refcount))
+               kfree_rcu(grp, rcu);
+}
+
+static void double_lock(spinlock_t *l1, spinlock_t *l2)
+{
+       if (l1 > l2)
+               swap(l1, l2);
+
+       spin_lock(l1);
+       spin_lock_nested(l2, SINGLE_DEPTH_NESTING);
+}
+
+static void task_numa_group(struct task_struct *p, int cpupid)
+{
+       struct numa_group *grp, *my_grp;
+       struct task_struct *tsk;
+       bool join = false;
+       int cpu = cpupid_to_cpu(cpupid);
+       int i;
+
+       if (unlikely(!p->numa_group)) {
+               unsigned int size = sizeof(struct numa_group) +
+                                   2*nr_node_ids*sizeof(atomic_long_t);
+
+               grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
+               if (!grp)
+                       return;
+
+               atomic_set(&grp->refcount, 1);
+               spin_lock_init(&grp->lock);
+               INIT_LIST_HEAD(&grp->task_list);
+
+               for (i = 0; i < 2*nr_node_ids; i++)
+                       atomic_long_set(&grp->faults[i], p->numa_faults[i]);
+
+               list_add(&p->numa_entry, &grp->task_list);
+               grp->nr_tasks++;
+               rcu_assign_pointer(p->numa_group, grp);
+       }
+
+       rcu_read_lock();
+       tsk = ACCESS_ONCE(cpu_rq(cpu)->curr);
+
+       if (!cpupid_match_pid(tsk, cpupid))
+               goto unlock;
+
+       grp = rcu_dereference(tsk->numa_group);
+       if (!grp)
+               goto unlock;
+
+       my_grp = p->numa_group;
+       if (grp == my_grp)
+               goto unlock;
+
+       /*
+        * Only join the other group if its bigger; if we're the bigger group,
+        * the other task will join us.
+        */
+       if (my_grp->nr_tasks > grp->nr_tasks)
+               goto unlock;
+
+       /*
+        * Tie-break on the grp address.
+        */
+       if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
+               goto unlock;
+
+       if (!get_numa_group(grp))
+               goto unlock;
+
+       join = true;
+
+unlock:
+       rcu_read_unlock();
+
+       if (!join)
+               return;
+
+       for (i = 0; i < 2*nr_node_ids; i++) {
+               atomic_long_sub(p->numa_faults[i], &my_grp->faults[i]);
+               atomic_long_add(p->numa_faults[i], &grp->faults[i]);
+       }
+
+       double_lock(&my_grp->lock, &grp->lock);
+
+       list_move(&p->numa_entry, &grp->task_list);
+       my_grp->nr_tasks--;
+       grp->nr_tasks++;
+
+       spin_unlock(&my_grp->lock);
+       spin_unlock(&grp->lock);
+
+       rcu_assign_pointer(p->numa_group, grp);
+
+       put_numa_group(my_grp);
+}
+
+void task_numa_free(struct task_struct *p)
+{
+       struct numa_group *grp = p->numa_group;
+       int i;
+
+       if (grp) {
+               for (i = 0; i < 2*nr_node_ids; i++)
+                       atomic_long_sub(p->numa_faults[i], &grp->faults[i]);
+
+               spin_lock(&grp->lock);
+               list_del(&p->numa_entry);
+               grp->nr_tasks--;
+               spin_unlock(&grp->lock);
+               rcu_assign_pointer(p->numa_group, NULL);
+               put_numa_group(grp);
+       }
+
+       kfree(p->numa_faults);
+}
+
 /*
  * Got a PROT_NONE fault for a page on @node.
  */
@@ -1222,15 +1366,6 @@ void task_numa_fault(int last_cpupid, int node, int pages, bool migrated)
        if (!p->mm)
                return;
 
-       /*
-        * First accesses are treated as private, otherwise consider accesses
-        * to be private if the accessing pid has not changed
-        */
-       if (!cpupid_pid_unset(last_cpupid))
-               priv = ((p->pid & LAST__PID_MASK) == cpupid_to_pid(last_cpupid));
-       else
-               priv = 1;
-
        /* Allocate buffer to track faults on a per-node basis */
        if (unlikely(!p->numa_faults)) {
                int size = sizeof(*p->numa_faults) * 2 * nr_node_ids;
@@ -1244,6 +1379,18 @@ void task_numa_fault(int last_cpupid, int node, int pages, bool migrated)
                p->numa_faults_buffer = p->numa_faults + (2 * nr_node_ids);
        }
 
+       /*
+        * First accesses are treated as private, otherwise consider accesses
+        * to be private if the accessing pid has not changed
+        */
+       if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
+               priv = 1;
+       } else {
+               priv = cpupid_match_pid(p, last_cpupid);
+               if (!priv)
+                       task_numa_group(p, last_cpupid);
+       }
+
        /*
         * If pages are properly placed (did not migrate) then scan slower.
         * This is reset periodically in case of phase changes
index 691e96964dcccf1abfd9cbb1c4066442f5676d56..8037b10a256f2d1619fc3cfe3861197c8b52e128 100644 (file)
@@ -559,10 +559,7 @@ static inline u64 rq_clock_task(struct rq *rq)
 #ifdef CONFIG_NUMA_BALANCING
 extern int migrate_task_to(struct task_struct *p, int cpu);
 extern int migrate_swap(struct task_struct *, struct task_struct *);
-static inline void task_numa_free(struct task_struct *p)
-{
-       kfree(p->numa_faults);
-}
+extern void task_numa_free(struct task_struct *p);
 #else /* CONFIG_NUMA_BALANCING */
 static inline void task_numa_free(struct task_struct *p)
 {
index 5162e6d0d6522eca38de500299cc676c1fbfeb2e..c57efa25cdbb235fa05d3a93222d5d72c9b0c387 100644 (file)
@@ -2719,6 +2719,14 @@ static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
                get_page(dirty_page);
 
 reuse:
+               /*
+                * Clear the pages cpupid information as the existing
+                * information potentially belongs to a now completely
+                * unrelated process.
+                */
+               if (old_page)
+                       page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1);
+
                flush_cache_page(vma, address, pte_pfn(orig_pte));
                entry = pte_mkyoung(orig_pte);
                entry = maybe_mkwrite(pte_mkdirty(entry), vma);