Fix cpusets update_cpumask

Cause writes to cpuset "cpus" file to update cpus_allowed for member tasks:

- collect batches of tasks under tasklist_lock and then call
set_cpus_allowed() on them outside the lock (since this can sleep).

- add a simple generic priority heap type to allow efficient collection
of batches of tasks to be processed without duplicating or missing any
tasks in subsequent batches.

- make "cpus" file update a no-op if the mask hasn't changed

- fix race between update_cpumask() and sched_setaffinity() by making
sched_setaffinity() post-check that it's not running on any cpus outside
cpuset_cpus_allowed().

[[e-mail]: coding-style fixes]
Signed-off-by: Paul Menage <[e-mail]>
Cc: Paul Jackson <[e-mail]>
Cc: David Rientjes <[e-mail]>
Cc: Nick Piggin <[e-mail]>
Cc: Peter Zijlstra <[e-mail]>
Cc: Balbir Singh <[e-mail]>
Cc: Cedric Le Goater <[e-mail]>
Cc: "Eric W. Biederman" <[e-mail]>
Cc: Serge Hallyn <[e-mail]>
Signed-off-by: Andrew Morton <[e-mail]>
Signed-off-by: Linus Torvalds <[e-mail]>

include/linux/prio_heap.h

···
+#ifndef _LINUX_PRIO_HEAP_H
+#define _LINUX_PRIO_HEAP_H
+
+/*
+ * Simple insertion-only static-sized priority heap containing
+ * pointers, based on CLR, chapter 7
+ */
+
+#include <linux/gfp.h>
+
+/**
+ * struct ptr_heap - simple static-sized priority heap
+ * @ptrs - pointer to data area
+ * @max - max number of elements that can be stored in @ptrs
+ * @size - current number of valid elements in @ptrs (in the range 0..@size-1
+ * @gt: comparison operator, which should implement "greater than"
+ */
+struct ptr_heap {
+	void **ptrs;
+	int max;
+	int size;
+	int (*gt)(void *, void *);
+};
+
+/**
+ * heap_init - initialize an empty heap with a given memory size
+ * @heap: the heap structure to be initialized
+ * @size: amount of memory to use in bytes
+ * @gfp_mask: mask to pass to kmalloc()
+ * @gt: comparison operator, which should implement "greater than"
+ */
+extern int heap_init(struct ptr_heap *heap, size_t size, gfp_t gfp_mask,
+		     int (*gt)(void *, void *));
+
+/**
+ * heap_free - release a heap's storage
+ * @heap: the heap structure whose data should be released
+ */
+void heap_free(struct ptr_heap *heap);
+
+/**
+ * heap_insert - insert a value into the heap and return any overflowed value
+ * @heap: the heap to be operated on
+ * @p: the pointer to be inserted
+ *
+ * Attempts to insert the given value into the priority heap. If the
+ * heap is full prior to the insertion, then the resulting heap will
+ * consist of the smallest @max elements of the original heap and the
+ * new element; the greatest element will be removed from the heap and
+ * returned. Note that the returned element will be the new element
+ * (i.e. no change to the heap) if the new element is greater than all
+ * elements currently in the heap.
+ */
+extern void *heap_insert(struct ptr_heap *heap, void *p);
+
+
+
+#endif /* _LINUX_PRIO_HEAP_H */

kernel/cpuset.c

···
 #include <linux/mount.h>
 #include <linux/namei.h>
 #include <linux/pagemap.h>
+#include <linux/prio_heap.h>
 #include <linux/proc_fs.h>
 #include <linux/rcupdate.h>
 #include <linux/sched.h>
···
 	/* Don't kfree(doms) -- partition_sched_domains() does that. */
 }
 
+static inline int started_after_time(struct task_struct *t1,
+				     struct timespec *time,
+				     struct task_struct *t2)
+{
+	int start_diff = timespec_compare(&t1->start_time, time);
+	if (start_diff > 0) {
+		return 1;
+	} else if (start_diff < 0) {
+		return 0;
+	} else {
+		/*
+		 * Arbitrarily, if two processes started at the same
+		 * time, we'll say that the lower pointer value
+		 * started first. Note that t2 may have exited by now
+		 * so this may not be a valid pointer any longer, but
+		 * that's fine - it still serves to distinguish
+		 * between two tasks started (effectively)
+		 * simultaneously.
+		 */
+		return t1 > t2;
+	}
+}
+
+static inline int started_after(void *p1, void *p2)
+{
+	struct task_struct *t1 = p1;
+	struct task_struct *t2 = p2;
+	return started_after_time(t1, &t2->start_time, t2);
+}
+
 /*
  * Call with manage_mutex held.  May take callback_mutex during call.
  */
···
 static int update_cpumask(struct cpuset *cs, char *buf)
 {
 	struct cpuset trialcs;
-	int retval;
-	int cpus_changed, is_load_balanced;
+	int retval, i;
+	int is_load_balanced;
+	struct cgroup_iter it;
+	struct cgroup *cgrp = cs->css.cgroup;
+	struct task_struct *p, *dropped;
+	/* Never dereference latest_task, since it's not refcounted */
+	struct task_struct *latest_task = NULL;
+	struct ptr_heap heap;
+	struct timespec latest_time = { 0, 0 };
 
 	/* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
 	if (cs == &top_cpuset)
···
 	if (retval < 0)
 		return retval;
 
-	cpus_changed = !cpus_equal(cs->cpus_allowed, trialcs.cpus_allowed);
+	/* Nothing to do if the cpus didn't change */
+	if (cpus_equal(cs->cpus_allowed, trialcs.cpus_allowed))
+		return 0;
+	retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, &started_after);
+	if (retval)
+		return retval;
+
 	is_load_balanced = is_sched_load_balance(&trialcs);
 
 	mutex_lock(&callback_mutex);
 	cs->cpus_allowed = trialcs.cpus_allowed;
 	mutex_unlock(&callback_mutex);
 
-	if (cpus_changed && is_load_balanced)
+ again:
+	/*
+	 * Scan tasks in the cpuset, and update the cpumasks of any
+	 * that need an update. Since we can't call set_cpus_allowed()
+	 * while holding tasklist_lock, gather tasks to be processed
+	 * in a heap structure. If the statically-sized heap fills up,
+	 * overflow tasks that started later, and in future iterations
+	 * only consider tasks that started after the latest task in
+	 * the previous pass. This guarantees forward progress and
+	 * that we don't miss any tasks
+	 */
+	heap.size = 0;
+	cgroup_iter_start(cgrp, &it);
+	while ((p = cgroup_iter_next(cgrp, &it))) {
+		/* Only affect tasks that don't have the right cpus_allowed */
+		if (cpus_equal(p->cpus_allowed, cs->cpus_allowed))
+			continue;
+		/*
+		 * Only process tasks that started after the last task
+		 * we processed
+		 */
+		if (!started_after_time(p, &latest_time, latest_task))
+			continue;
+		dropped = heap_insert(&heap, p);
+		if (dropped == NULL) {
+			get_task_struct(p);
+		} else if (dropped != p) {
+			get_task_struct(p);
+			put_task_struct(dropped);
+		}
+	}
+	cgroup_iter_end(cgrp, &it);
+	if (heap.size) {
+		for (i = 0; i < heap.size; i++) {
+			struct task_struct *p = heap.ptrs[i];
+			if (i == 0) {
+				latest_time = p->start_time;
+				latest_task = p;
+			}
+			set_cpus_allowed(p, cs->cpus_allowed);
+			put_task_struct(p);
+		}
+		/*
+		 * If we had to process any tasks at all, scan again
+		 * in case some of them were in the middle of forking
+		 * children that didn't notice the new cpumask
+		 * restriction.  Not the most efficient way to do it,
+		 * but it avoids having to take callback_mutex in the
+		 * fork path
+		 */
+		goto again;
+	}
+	heap_free(&heap);
+	if (is_load_balanced)
 		rebuild_sched_domains();
 
 	return 0;

kernel/sched.c

···
 
 	cpus_allowed = cpuset_cpus_allowed(p);
 	cpus_and(new_mask, new_mask, cpus_allowed);
+ again:
 	retval = set_cpus_allowed(p, new_mask);
 
+	if (!retval) {
+		cpus_allowed = cpuset_cpus_allowed(p);
+		if (!cpus_subset(new_mask, cpus_allowed)) {
+			/*
+			 * We must have raced with a concurrent cpuset
+			 * update. Just reset the cpus_allowed to the
+			 * cpuset's cpus_allowed
+			 */
+			new_mask = cpus_allowed;
+			goto again;
+		}
+	}
 out_unlock:
 	put_task_struct(p);
 	mutex_unlock(&sched_hotcpu_mutex);

lib/Makefile

···
 	 rbtree.o radix-tree.o dump_stack.o \
 	 idr.o int_sqrt.o bitmap.o extable.o prio_tree.o \
 	 sha1.o irq_regs.o reciprocal_div.o argv_split.o \
-	 proportions.o
+	 proportions.o prio_heap.o
 
 lib-$(CONFIG_MMU) += ioremap.o
 lib-$(CONFIG_SMP) += cpumask.o

lib/prio_heap.c

···
+/*
+ * Simple insertion-only static-sized priority heap containing
+ * pointers, based on CLR, chapter 7
+ */
+
+#include <linux/slab.h>
+#include <linux/prio_heap.h>
+
+int heap_init(struct ptr_heap *heap, size_t size, gfp_t gfp_mask,
+	      int (*gt)(void *, void *))
+{
+	heap->ptrs = kmalloc(size, gfp_mask);
+	if (!heap->ptrs)
+		return -ENOMEM;
+	heap->size = 0;
+	heap->max = size / sizeof(void *);
+	heap->gt = gt;
+	return 0;
+}
+
+void heap_free(struct ptr_heap *heap)
+{
+	kfree(heap->ptrs);
+}
+
+void *heap_insert(struct ptr_heap *heap, void *p)
+{
+	void *res;
+	void **ptrs = heap->ptrs;
+	int pos;
+
+	if (heap->size < heap->max) {
+		/* Heap insertion */
+		int pos = heap->size++;
+		while (pos > 0 && heap->gt(p, ptrs[(pos-1)/2])) {
+			ptrs[pos] = ptrs[(pos-1)/2];
+			pos = (pos-1)/2;
+		}
+		ptrs[pos] = p;
+		return NULL;
+	}
+
+	/* The heap is full, so something will have to be dropped */
+
+	/* If the new pointer is greater than the current max, drop it */
+	if (heap->gt(p, ptrs[0]))
+		return p;
+
+	/* Replace the current max and heapify */
+	res = ptrs[0];
+	ptrs[0] = p;
+	pos = 0;
+
+	while (1) {
+		int left = 2 * pos + 1;
+		int right = 2 * pos + 2;
+		int largest = pos;
+		if (left < heap->size && heap->gt(ptrs[left], p))
+			largest = left;
+		if (right < heap->size && heap->gt(ptrs[right], ptrs[largest]))
+			largest = right;
+		if (largest == pos)
+			break;
+		/* Push p down the heap one level and bump one up */
+		ptrs[pos] = ptrs[largest];
+		ptrs[largest] = p;
+		pos = largest;
+	}
+	return res;
+}