slub: add ability to list alloc / free callers per slab
Christoph Lameter [Sun, 6 May 2007 21:49:45 +0000 (14:49 -0700)]
This patch enables listing the callers who allocated or freed objects in a
cache.

For example to list the allocators for kmalloc-128 do

cat /sys/slab/kmalloc-128/alloc_calls
      7 sn_io_slot_fixup+0x40/0x700
      7 sn_io_slot_fixup+0x80/0x700
      9 sn_bus_fixup+0xe0/0x380
      6 param_sysfs_setup+0xf0/0x280
    276 percpu_populate+0xf0/0x1a0
     19 __register_chrdev_region+0x30/0x360
      8 expand_files+0x2e0/0x6e0
      1 sys_epoll_create+0x60/0x200
      1 __mounts_open+0x140/0x2c0
     65 kmem_alloc+0x110/0x280
      3 alloc_disk_node+0xe0/0x200
     33 as_get_io_context+0x90/0x280
     74 kobject_kset_add_dir+0x40/0x140
     12 pci_create_bus+0x2a0/0x5c0
      1 acpi_ev_create_gpe_block+0x120/0x9e0
     41 con_insert_unipair+0x100/0x1c0
      1 uart_open+0x1c0/0xba0
      1 dma_pool_create+0xe0/0x340
      2 neigh_table_init_no_netlink+0x260/0x4c0
      6 neigh_parms_alloc+0x30/0x200
      1 netlink_kernel_create+0x130/0x320
      5 fz_hash_alloc+0x50/0xe0
      2 sn_common_hubdev_init+0xd0/0x6e0
     28 kernel_param_sysfs_setup+0x30/0x180
     72 process_zones+0x70/0x2e0

cat /sys/slab/kmalloc-128/free_calls
    558 <not-available>
      3 sn_io_slot_fixup+0x600/0x700
     84 free_fdtable_rcu+0x120/0x260
      2 seq_release+0x40/0x60
      6 kmem_free+0x70/0xc0
     24 free_as_io_context+0x20/0x200
      1 acpi_get_object_info+0x3a0/0x3e0
      1 acpi_add_single_object+0xcf0/0x1e40
      2 con_release_unimap+0x80/0x140
      1 free+0x20/0x40

SLAB_STORE_USER must be enabled for a slab cache by either booting with
"slab_debug" or enabling user tracking specifically for the slab of interest.

Signed-off-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/slub.c

index 4251917..a623196 100644 (file)
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -97,9 +97,6 @@
  *
  * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
  *
- * - Support DEBUG_SLAB_LEAK. Trouble is we do not know where the full
- *   slabs are in SLUB.
- *
  * - SLAB_DEBUG_INITIAL is not supported but I have never seen a use of
  *   it.
  *
@@ -2659,6 +2656,169 @@ static unsigned long validate_slab_cache(struct kmem_cache *s)
        return count;
 }
 
+/*
+ * Generate lists of locations where slabcache objects are allocated
+ * and freed.
+ */
+
+struct location {
+       unsigned long count;
+       void *addr;
+};
+
+struct loc_track {
+       unsigned long max;
+       unsigned long count;
+       struct location *loc;
+};
+
+static void free_loc_track(struct loc_track *t)
+{
+       if (t->max)
+               free_pages((unsigned long)t->loc,
+                       get_order(sizeof(struct location) * t->max));
+}
+
+static int alloc_loc_track(struct loc_track *t, unsigned long max)
+{
+       struct location *l;
+       int order;
+
+       if (!max)
+               max = PAGE_SIZE / sizeof(struct location);
+
+       order = get_order(sizeof(struct location) * max);
+
+       l = (void *)__get_free_pages(GFP_KERNEL, order);
+
+       if (!l)
+               return 0;
+
+       if (t->count) {
+               memcpy(l, t->loc, sizeof(struct location) * t->count);
+               free_loc_track(t);
+       }
+       t->max = max;
+       t->loc = l;
+       return 1;
+}
+
+static int add_location(struct loc_track *t, struct kmem_cache *s,
+                                               void *addr)
+{
+       long start, end, pos;
+       struct location *l;
+       void *caddr;
+
+       start = -1;
+       end = t->count;
+
+       for ( ; ; ) {
+               pos = start + (end - start + 1) / 2;
+
+               /*
+                * There is nothing at "end". If we end up there
+                * we need to add something to before end.
+                */
+               if (pos == end)
+                       break;
+
+               caddr = t->loc[pos].addr;
+               if (addr == caddr) {
+                       t->loc[pos].count++;
+                       return 1;
+               }
+
+               if (addr < caddr)
+                       end = pos;
+               else
+                       start = pos;
+       }
+
+       /*
+        * Not found. Insert new tracking element
+        */
+       if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max))
+               return 0;
+
+       l = t->loc + pos;
+       if (pos < t->count)
+               memmove(l + 1, l,
+                       (t->count - pos) * sizeof(struct location));
+       t->count++;
+       l->count = 1;
+       l->addr = addr;
+       return 1;
+}
+
+static void process_slab(struct loc_track *t, struct kmem_cache *s,
+               struct page *page, enum track_item alloc)
+{
+       void *addr = page_address(page);
+       unsigned long map[BITS_TO_LONGS(s->objects)];
+       void *p;
+
+       bitmap_zero(map, s->objects);
+       for (p = page->freelist; p; p = get_freepointer(s, p))
+               set_bit((p - addr) / s->size, map);
+
+       for (p = addr; p < addr + s->objects * s->size; p += s->size)
+               if (!test_bit((p - addr) / s->size, map)) {
+                       void *addr = get_track(s, p, alloc)->addr;
+
+                       add_location(t, s, addr);
+               }
+}
+
+static int list_locations(struct kmem_cache *s, char *buf,
+                                       enum track_item alloc)
+{
+       int n = 0;
+       unsigned long i;
+       struct loc_track t;
+       int node;
+
+       t.count = 0;
+       t.max = 0;
+
+       /* Push back cpu slabs */
+       flush_all(s);
+
+       for_each_online_node(node) {
+               struct kmem_cache_node *n = get_node(s, node);
+               unsigned long flags;
+               struct page *page;
+
+               if (!atomic_read(&n->nr_slabs))
+                       continue;
+
+               spin_lock_irqsave(&n->list_lock, flags);
+               list_for_each_entry(page, &n->partial, lru)
+                       process_slab(&t, s, page, alloc);
+               list_for_each_entry(page, &n->full, lru)
+                       process_slab(&t, s, page, alloc);
+               spin_unlock_irqrestore(&n->list_lock, flags);
+       }
+
+       for (i = 0; i < t.count; i++) {
+               void *addr = t.loc[i].addr;
+
+               if (n > PAGE_SIZE - 100)
+                       break;
+               n += sprintf(buf + n, "%7ld ", t.loc[i].count);
+               if (addr)
+                       n += sprint_symbol(buf + n, (unsigned long)t.loc[i].addr);
+               else
+                       n += sprintf(buf + n, "<not-available>");
+               n += sprintf(buf + n, "\n");
+       }
+
+       free_loc_track(&t);
+       if (!t.count)
+               n += sprintf(buf, "No data\n");
+       return n;
+}
+
 static unsigned long count_partial(struct kmem_cache_node *n)
 {
        unsigned long flags;
@@ -3009,6 +3169,22 @@ static ssize_t validate_store(struct kmem_cache *s,
 }
 SLAB_ATTR(validate);
 
+static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
+{
+       if (!(s->flags & SLAB_STORE_USER))
+               return -ENOSYS;
+       return list_locations(s, buf, TRACK_ALLOC);
+}
+SLAB_ATTR_RO(alloc_calls);
+
+static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
+{
+       if (!(s->flags & SLAB_STORE_USER))
+               return -ENOSYS;
+       return list_locations(s, buf, TRACK_FREE);
+}
+SLAB_ATTR_RO(free_calls);
+
 #ifdef CONFIG_NUMA
 static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf)
 {
@@ -3049,6 +3225,8 @@ static struct attribute * slab_attrs[] = {
        &poison_attr.attr,
        &store_user_attr.attr,
        &validate_attr.attr,
+       &alloc_calls_attr.attr,
+       &free_calls_attr.attr,
 #ifdef CONFIG_ZONE_DMA
        &cache_dma_attr.attr,
 #endif