Fix cpusets update_cpumask
Paul Menage [Fri, 19 Oct 2007 06:40:22 +0000 (23:40 -0700)]
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().

[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Paul Menage <menage@google.com>
Cc: Paul Jackson <pj@sgi.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Balbir Singh <balbir@in.ibm.com>
Cc: Cedric Le Goater <clg@fr.ibm.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Serge Hallyn <serue@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

include/linux/prio_heap.h [new file with mode: 0644]
kernel/cpuset.c
kernel/sched.c
lib/Makefile
lib/prio_heap.c [new file with mode: 0644]

diff --git a/include/linux/prio_heap.h b/include/linux/prio_heap.h
new file mode 100644 (file)
index 0000000..0809435
--- /dev/null
@@ -0,0 +1,58 @@
+#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 */
index 64ad59c..fa31cb9 100644 (file)
@@ -38,6 +38,7 @@
 #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>
@@ -701,6 +702,36 @@ done:
        /* 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.
  */
@@ -708,8 +739,15 @@ done:
 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)
@@ -736,14 +774,73 @@ static int update_cpumask(struct cpuset *cs, char *buf)
        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;
index 39d6354..72a809a 100644 (file)
@@ -4471,8 +4471,21 @@ long sched_setaffinity(pid_t pid, cpumask_t new_mask)
 
        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);
index c5f215d..3a0983b 100644 (file)
@@ -6,7 +6,7 @@ lib-y := ctype.o string.o vsprintf.o cmdline.o \
         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
diff --git a/lib/prio_heap.c b/lib/prio_heap.c
new file mode 100644 (file)
index 0000000..471944a
--- /dev/null
@@ -0,0 +1,70 @@
+/*
+ * 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;
+}