kmemleak: Add the base support
authorCatalin Marinas <catalin.marinas@arm.com>
Thu, 11 Jun 2009 12:22:39 +0000 (13:22 +0100)
committerCatalin Marinas <catalin.marinas@arm.com>
Thu, 11 Jun 2009 16:03:28 +0000 (17:03 +0100)
This patch adds the base support for the kernel memory leak
detector. It traces the memory allocation/freeing in a way similar to
the Boehm's conservative garbage collector, the difference being that
the unreferenced objects are not freed but only shown in
/sys/kernel/debug/kmemleak. Enabling this feature introduces an
overhead to memory allocations.

Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Cc: Ingo Molnar <mingo@elte.hu>
Acked-by: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
include/linux/kmemleak.h [new file with mode: 0644]
init/main.c
mm/kmemleak.c [new file with mode: 0644]

diff --git a/include/linux/kmemleak.h b/include/linux/kmemleak.h
new file mode 100644 (file)
index 0000000..7796aed
--- /dev/null
@@ -0,0 +1,96 @@
+/*
+ * include/linux/kmemleak.h
+ *
+ * Copyright (C) 2008 ARM Limited
+ * Written by Catalin Marinas <catalin.marinas@arm.com>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
+ */
+
+#ifndef __KMEMLEAK_H
+#define __KMEMLEAK_H
+
+#ifdef CONFIG_DEBUG_KMEMLEAK
+
+extern void kmemleak_init(void);
+extern void kmemleak_alloc(const void *ptr, size_t size, int min_count,
+                          gfp_t gfp);
+extern void kmemleak_free(const void *ptr);
+extern void kmemleak_padding(const void *ptr, unsigned long offset,
+                            size_t size);
+extern void kmemleak_not_leak(const void *ptr);
+extern void kmemleak_ignore(const void *ptr);
+extern void kmemleak_scan_area(const void *ptr, unsigned long offset,
+                              size_t length, gfp_t gfp);
+extern void kmemleak_no_scan(const void *ptr);
+
+static inline void kmemleak_alloc_recursive(const void *ptr, size_t size,
+                                           int min_count, unsigned long flags,
+                                           gfp_t gfp)
+{
+       if (!(flags & SLAB_NOLEAKTRACE))
+               kmemleak_alloc(ptr, size, min_count, gfp);
+}
+
+static inline void kmemleak_free_recursive(const void *ptr, unsigned long flags)
+{
+       if (!(flags & SLAB_NOLEAKTRACE))
+               kmemleak_free(ptr);
+}
+
+static inline void kmemleak_erase(void **ptr)
+{
+       *ptr = NULL;
+}
+
+#else
+
+static inline void kmemleak_init(void)
+{
+}
+static inline void kmemleak_alloc(const void *ptr, size_t size, int min_count,
+                                 gfp_t gfp)
+{
+}
+static inline void kmemleak_alloc_recursive(const void *ptr, size_t size,
+                                           int min_count, unsigned long flags,
+                                           gfp_t gfp)
+{
+}
+static inline void kmemleak_free(const void *ptr)
+{
+}
+static inline void kmemleak_free_recursive(const void *ptr, unsigned long flags)
+{
+}
+static inline void kmemleak_not_leak(const void *ptr)
+{
+}
+static inline void kmemleak_ignore(const void *ptr)
+{
+}
+static inline void kmemleak_scan_area(const void *ptr, unsigned long offset,
+                                     size_t length, gfp_t gfp)
+{
+}
+static inline void kmemleak_erase(void **ptr)
+{
+}
+static inline void kmemleak_no_scan(const void *ptr)
+{
+}
+
+#endif /* CONFIG_DEBUG_KMEMLEAK */
+
+#endif /* __KMEMLEAK_H */
index bb7dc57eee36ed8079e63bdf1a4c725d4c97e45f..9d759d68d7a0a6ce6d8b5c23b8660c0a83cd0645 100644 (file)
@@ -56,6 +56,7 @@
 #include <linux/debug_locks.h>
 #include <linux/debugobjects.h>
 #include <linux/lockdep.h>
+#include <linux/kmemleak.h>
 #include <linux/pid_namespace.h>
 #include <linux/device.h>
 #include <linux/kthread.h>
@@ -603,6 +604,7 @@ asmlinkage void __init start_kernel(void)
        /* init some links before init_ISA_irqs() */
        early_irq_init();
        init_IRQ();
+       prio_tree_init();
        pidhash_init();
        init_timers();
        hrtimers_init();
@@ -654,6 +656,7 @@ asmlinkage void __init start_kernel(void)
        cpu_hotplug_init();
        kmem_cache_init();
        kmemtrace_init();
+       kmemleak_init();
        debug_objects_mem_init();
        idr_init_cache();
        setup_per_cpu_pageset();
@@ -663,7 +666,6 @@ asmlinkage void __init start_kernel(void)
        calibrate_delay();
        pidmap_init();
        pgtable_cache_init();
-       prio_tree_init();
        anon_vma_init();
 #ifdef CONFIG_X86
        if (efi_enabled)
diff --git a/mm/kmemleak.c b/mm/kmemleak.c
new file mode 100644 (file)
index 0000000..58ec86c
--- /dev/null
@@ -0,0 +1,1498 @@
+/*
+ * mm/kmemleak.c
+ *
+ * Copyright (C) 2008 ARM Limited
+ * Written by Catalin Marinas <catalin.marinas@arm.com>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
+ *
+ *
+ * For more information on the algorithm and kmemleak usage, please see
+ * Documentation/kmemleak.txt.
+ *
+ * Notes on locking
+ * ----------------
+ *
+ * The following locks and mutexes are used by kmemleak:
+ *
+ * - kmemleak_lock (rwlock): protects the object_list modifications and
+ *   accesses to the object_tree_root. The object_list is the main list
+ *   holding the metadata (struct kmemleak_object) for the allocated memory
+ *   blocks. The object_tree_root is a priority search tree used to look-up
+ *   metadata based on a pointer to the corresponding memory block.  The
+ *   kmemleak_object structures are added to the object_list and
+ *   object_tree_root in the create_object() function called from the
+ *   kmemleak_alloc() callback and removed in delete_object() called from the
+ *   kmemleak_free() callback
+ * - kmemleak_object.lock (spinlock): protects a kmemleak_object. Accesses to
+ *   the metadata (e.g. count) are protected by this lock. Note that some
+ *   members of this structure may be protected by other means (atomic or
+ *   kmemleak_lock). This lock is also held when scanning the corresponding
+ *   memory block to avoid the kernel freeing it via the kmemleak_free()
+ *   callback. This is less heavyweight than holding a global lock like
+ *   kmemleak_lock during scanning
+ * - scan_mutex (mutex): ensures that only one thread may scan the memory for
+ *   unreferenced objects at a time. The gray_list contains the objects which
+ *   are already referenced or marked as false positives and need to be
+ *   scanned. This list is only modified during a scanning episode when the
+ *   scan_mutex is held. At the end of a scan, the gray_list is always empty.
+ *   Note that the kmemleak_object.use_count is incremented when an object is
+ *   added to the gray_list and therefore cannot be freed
+ * - kmemleak_mutex (mutex): prevents multiple users of the "kmemleak" debugfs
+ *   file together with modifications to the memory scanning parameters
+ *   including the scan_thread pointer
+ *
+ * The kmemleak_object structures have a use_count incremented or decremented
+ * using the get_object()/put_object() functions. When the use_count becomes
+ * 0, this count can no longer be incremented and put_object() schedules the
+ * kmemleak_object freeing via an RCU callback. All calls to the get_object()
+ * function must be protected by rcu_read_lock() to avoid accessing a freed
+ * structure.
+ */
+
+#include <linux/init.h>
+#include <linux/kernel.h>
+#include <linux/list.h>
+#include <linux/sched.h>
+#include <linux/jiffies.h>
+#include <linux/delay.h>
+#include <linux/module.h>
+#include <linux/kthread.h>
+#include <linux/prio_tree.h>
+#include <linux/gfp.h>
+#include <linux/fs.h>
+#include <linux/debugfs.h>
+#include <linux/seq_file.h>
+#include <linux/cpumask.h>
+#include <linux/spinlock.h>
+#include <linux/mutex.h>
+#include <linux/rcupdate.h>
+#include <linux/stacktrace.h>
+#include <linux/cache.h>
+#include <linux/percpu.h>
+#include <linux/hardirq.h>
+#include <linux/mmzone.h>
+#include <linux/slab.h>
+#include <linux/thread_info.h>
+#include <linux/err.h>
+#include <linux/uaccess.h>
+#include <linux/string.h>
+#include <linux/nodemask.h>
+#include <linux/mm.h>
+
+#include <asm/sections.h>
+#include <asm/processor.h>
+#include <asm/atomic.h>
+
+#include <linux/kmemleak.h>
+
+/*
+ * Kmemleak configuration and common defines.
+ */
+#define MAX_TRACE              16      /* stack trace length */
+#define REPORTS_NR             50      /* maximum number of reported leaks */
+#define MSECS_MIN_AGE          5000    /* minimum object age for reporting */
+#define MSECS_SCAN_YIELD       10      /* CPU yielding period */
+#define SECS_FIRST_SCAN                60      /* delay before the first scan */
+#define SECS_SCAN_WAIT         600     /* subsequent auto scanning delay */
+
+#define BYTES_PER_POINTER      sizeof(void *)
+
+/* scanning area inside a memory block */
+struct kmemleak_scan_area {
+       struct hlist_node node;
+       unsigned long offset;
+       size_t length;
+};
+
+/*
+ * Structure holding the metadata for each allocated memory block.
+ * Modifications to such objects should be made while holding the
+ * object->lock. Insertions or deletions from object_list, gray_list or
+ * tree_node are already protected by the corresponding locks or mutex (see
+ * the notes on locking above). These objects are reference-counted
+ * (use_count) and freed using the RCU mechanism.
+ */
+struct kmemleak_object {
+       spinlock_t lock;
+       unsigned long flags;            /* object status flags */
+       struct list_head object_list;
+       struct list_head gray_list;
+       struct prio_tree_node tree_node;
+       struct rcu_head rcu;            /* object_list lockless traversal */
+       /* object usage count; object freed when use_count == 0 */
+       atomic_t use_count;
+       unsigned long pointer;
+       size_t size;
+       /* minimum number of a pointers found before it is considered leak */
+       int min_count;
+       /* the total number of pointers found pointing to this object */
+       int count;
+       /* memory ranges to be scanned inside an object (empty for all) */
+       struct hlist_head area_list;
+       unsigned long trace[MAX_TRACE];
+       unsigned int trace_len;
+       unsigned long jiffies;          /* creation timestamp */
+       pid_t pid;                      /* pid of the current task */
+       char comm[TASK_COMM_LEN];       /* executable name */
+};
+
+/* flag representing the memory block allocation status */
+#define OBJECT_ALLOCATED       (1 << 0)
+/* flag set after the first reporting of an unreference object */
+#define OBJECT_REPORTED                (1 << 1)
+/* flag set to not scan the object */
+#define OBJECT_NO_SCAN         (1 << 2)
+
+/* the list of all allocated objects */
+static LIST_HEAD(object_list);
+/* the list of gray-colored objects (see color_gray comment below) */
+static LIST_HEAD(gray_list);
+/* prio search tree for object boundaries */
+static struct prio_tree_root object_tree_root;
+/* rw_lock protecting the access to object_list and prio_tree_root */
+static DEFINE_RWLOCK(kmemleak_lock);
+
+/* allocation caches for kmemleak internal data */
+static struct kmem_cache *object_cache;
+static struct kmem_cache *scan_area_cache;
+
+/* set if tracing memory operations is enabled */
+static atomic_t kmemleak_enabled = ATOMIC_INIT(0);
+/* set in the late_initcall if there were no errors */
+static atomic_t kmemleak_initialized = ATOMIC_INIT(0);
+/* enables or disables early logging of the memory operations */
+static atomic_t kmemleak_early_log = ATOMIC_INIT(1);
+/* set if a fata kmemleak error has occurred */
+static atomic_t kmemleak_error = ATOMIC_INIT(0);
+
+/* minimum and maximum address that may be valid pointers */
+static unsigned long min_addr = ULONG_MAX;
+static unsigned long max_addr;
+
+/* used for yielding the CPU to other tasks during scanning */
+static unsigned long next_scan_yield;
+static struct task_struct *scan_thread;
+static unsigned long jiffies_scan_yield;
+static unsigned long jiffies_min_age;
+/* delay between automatic memory scannings */
+static signed long jiffies_scan_wait;
+/* enables or disables the task stacks scanning */
+static int kmemleak_stack_scan;
+/* mutex protecting the memory scanning */
+static DEFINE_MUTEX(scan_mutex);
+/* mutex protecting the access to the /sys/kernel/debug/kmemleak file */
+static DEFINE_MUTEX(kmemleak_mutex);
+
+/* number of leaks reported (for limitation purposes) */
+static int reported_leaks;
+
+/*
+ * Early object allocation/freeing logging. Kkmemleak is initialized after the
+ * kernel allocator. However, both the kernel allocator and kmemleak may
+ * allocate memory blocks which need to be tracked. Kkmemleak defines an
+ * arbitrary buffer to hold the allocation/freeing information before it is
+ * fully initialized.
+ */
+
+/* kmemleak operation type for early logging */
+enum {
+       KMEMLEAK_ALLOC,
+       KMEMLEAK_FREE,
+       KMEMLEAK_NOT_LEAK,
+       KMEMLEAK_IGNORE,
+       KMEMLEAK_SCAN_AREA,
+       KMEMLEAK_NO_SCAN
+};
+
+/*
+ * Structure holding the information passed to kmemleak callbacks during the
+ * early logging.
+ */
+struct early_log {
+       int op_type;                    /* kmemleak operation type */
+       const void *ptr;                /* allocated/freed memory block */
+       size_t size;                    /* memory block size */
+       int min_count;                  /* minimum reference count */
+       unsigned long offset;           /* scan area offset */
+       size_t length;                  /* scan area length */
+};
+
+/* early logging buffer and current position */
+static struct early_log early_log[200];
+static int crt_early_log;
+
+static void kmemleak_disable(void);
+
+/*
+ * Print a warning and dump the stack trace.
+ */
+#define kmemleak_warn(x...)    do {    \
+       pr_warning(x);                  \
+       dump_stack();                   \
+} while (0)
+
+/*
+ * Macro invoked when a serious kmemleak condition occured and cannot be
+ * recovered from. Kkmemleak will be disabled and further allocation/freeing
+ * tracing no longer available.
+ */
+#define kmemleak_panic(x...)   do {    \
+       kmemleak_warn(x);               \
+       kmemleak_disable();             \
+} while (0)
+
+/*
+ * Object colors, encoded with count and min_count:
+ * - white - orphan object, not enough references to it (count < min_count)
+ * - gray  - not orphan, not marked as false positive (min_count == 0) or
+ *             sufficient references to it (count >= min_count)
+ * - black - ignore, it doesn't contain references (e.g. text section)
+ *             (min_count == -1). No function defined for this color.
+ * Newly created objects don't have any color assigned (object->count == -1)
+ * before the next memory scan when they become white.
+ */
+static int color_white(const struct kmemleak_object *object)
+{
+       return object->count != -1 && object->count < object->min_count;
+}
+
+static int color_gray(const struct kmemleak_object *object)
+{
+       return object->min_count != -1 && object->count >= object->min_count;
+}
+
+/*
+ * Objects are considered referenced if their color is gray and they have not
+ * been deleted.
+ */
+static int referenced_object(struct kmemleak_object *object)
+{
+       return (object->flags & OBJECT_ALLOCATED) && color_gray(object);
+}
+
+/*
+ * Objects are considered unreferenced only if their color is white, they have
+ * not be deleted and have a minimum age to avoid false positives caused by
+ * pointers temporarily stored in CPU registers.
+ */
+static int unreferenced_object(struct kmemleak_object *object)
+{
+       return (object->flags & OBJECT_ALLOCATED) && color_white(object) &&
+               time_is_before_eq_jiffies(object->jiffies + jiffies_min_age);
+}
+
+/*
+ * Printing of the (un)referenced objects information, either to the seq file
+ * or to the kernel log. The print_referenced/print_unreferenced functions
+ * must be called with the object->lock held.
+ */
+#define print_helper(seq, x...)        do {    \
+       struct seq_file *s = (seq);     \
+       if (s)                          \
+               seq_printf(s, x);       \
+       else                            \
+               pr_info(x);             \
+} while (0)
+
+static void print_referenced(struct kmemleak_object *object)
+{
+       pr_info("kmemleak: referenced object 0x%08lx (size %zu)\n",
+               object->pointer, object->size);
+}
+
+static void print_unreferenced(struct seq_file *seq,
+                              struct kmemleak_object *object)
+{
+       int i;
+
+       print_helper(seq, "kmemleak: unreferenced object 0x%08lx (size %zu):\n",
+                    object->pointer, object->size);
+       print_helper(seq, "  comm \"%s\", pid %d, jiffies %lu\n",
+                    object->comm, object->pid, object->jiffies);
+       print_helper(seq, "  backtrace:\n");
+
+       for (i = 0; i < object->trace_len; i++) {
+               void *ptr = (void *)object->trace[i];
+               print_helper(seq, "    [<%p>] %pS\n", ptr, ptr);
+       }
+}
+
+/*
+ * Print the kmemleak_object information. This function is used mainly for
+ * debugging special cases when kmemleak operations. It must be called with
+ * the object->lock held.
+ */
+static void dump_object_info(struct kmemleak_object *object)
+{
+       struct stack_trace trace;
+
+       trace.nr_entries = object->trace_len;
+       trace.entries = object->trace;
+
+       pr_notice("kmemleak: Object 0x%08lx (size %zu):\n",
+                 object->tree_node.start, object->size);
+       pr_notice("  comm \"%s\", pid %d, jiffies %lu\n",
+                 object->comm, object->pid, object->jiffies);
+       pr_notice("  min_count = %d\n", object->min_count);
+       pr_notice("  count = %d\n", object->count);
+       pr_notice("  backtrace:\n");
+       print_stack_trace(&trace, 4);
+}
+
+/*
+ * Look-up a memory block metadata (kmemleak_object) in the priority search
+ * tree based on a pointer value. If alias is 0, only values pointing to the
+ * beginning of the memory block are allowed. The kmemleak_lock must be held
+ * when calling this function.
+ */
+static struct kmemleak_object *lookup_object(unsigned long ptr, int alias)
+{
+       struct prio_tree_node *node;
+       struct prio_tree_iter iter;
+       struct kmemleak_object *object;
+
+       prio_tree_iter_init(&iter, &object_tree_root, ptr, ptr);
+       node = prio_tree_next(&iter);
+       if (node) {
+               object = prio_tree_entry(node, struct kmemleak_object,
+                                        tree_node);
+               if (!alias && object->pointer != ptr) {
+                       kmemleak_warn("kmemleak: Found object by alias");
+                       object = NULL;
+               }
+       } else
+               object = NULL;
+
+       return object;
+}
+
+/*
+ * Increment the object use_count. Return 1 if successful or 0 otherwise. Note
+ * that once an object's use_count reached 0, the RCU freeing was already
+ * registered and the object should no longer be used. This function must be
+ * called under the protection of rcu_read_lock().
+ */
+static int get_object(struct kmemleak_object *object)
+{
+       return atomic_inc_not_zero(&object->use_count);
+}
+
+/*
+ * RCU callback to free a kmemleak_object.
+ */
+static void free_object_rcu(struct rcu_head *rcu)
+{
+       struct hlist_node *elem, *tmp;
+       struct kmemleak_scan_area *area;
+       struct kmemleak_object *object =
+               container_of(rcu, struct kmemleak_object, rcu);
+
+       /*
+        * Once use_count is 0 (guaranteed by put_object), there is no other
+        * code accessing this object, hence no need for locking.
+        */
+       hlist_for_each_entry_safe(area, elem, tmp, &object->area_list, node) {
+               hlist_del(elem);
+               kmem_cache_free(scan_area_cache, area);
+       }
+       kmem_cache_free(object_cache, object);
+}
+
+/*
+ * Decrement the object use_count. Once the count is 0, free the object using
+ * an RCU callback. Since put_object() may be called via the kmemleak_free() ->
+ * delete_object() path, the delayed RCU freeing ensures that there is no
+ * recursive call to the kernel allocator. Lock-less RCU object_list traversal
+ * is also possible.
+ */
+static void put_object(struct kmemleak_object *object)
+{
+       if (!atomic_dec_and_test(&object->use_count))
+               return;
+
+       /* should only get here after delete_object was called */
+       WARN_ON(object->flags & OBJECT_ALLOCATED);
+
+       call_rcu(&object->rcu, free_object_rcu);
+}
+
+/*
+ * Look up an object in the prio search tree and increase its use_count.
+ */
+static struct kmemleak_object *find_and_get_object(unsigned long ptr, int alias)
+{
+       unsigned long flags;
+       struct kmemleak_object *object = NULL;
+
+       rcu_read_lock();
+       read_lock_irqsave(&kmemleak_lock, flags);
+       if (ptr >= min_addr && ptr < max_addr)
+               object = lookup_object(ptr, alias);
+       read_unlock_irqrestore(&kmemleak_lock, flags);
+
+       /* check whether the object is still available */
+       if (object && !get_object(object))
+               object = NULL;
+       rcu_read_unlock();
+
+       return object;
+}
+
+/*
+ * Create the metadata (struct kmemleak_object) corresponding to an allocated
+ * memory block and add it to the object_list and object_tree_root.
+ */
+static void create_object(unsigned long ptr, size_t size, int min_count,
+                         gfp_t gfp)
+{
+       unsigned long flags;
+       struct kmemleak_object *object;
+       struct prio_tree_node *node;
+       struct stack_trace trace;
+
+       object = kmem_cache_alloc(object_cache, gfp & ~GFP_SLAB_BUG_MASK);
+       if (!object) {
+               kmemleak_panic("kmemleak: Cannot allocate a kmemleak_object "
+                              "structure\n");
+               return;
+       }
+
+       INIT_LIST_HEAD(&object->object_list);
+       INIT_LIST_HEAD(&object->gray_list);
+       INIT_HLIST_HEAD(&object->area_list);
+       spin_lock_init(&object->lock);
+       atomic_set(&object->use_count, 1);
+       object->flags = OBJECT_ALLOCATED;
+       object->pointer = ptr;
+       object->size = size;
+       object->min_count = min_count;
+       object->count = -1;                     /* no color initially */
+       object->jiffies = jiffies;
+
+       /* task information */
+       if (in_irq()) {
+               object->pid = 0;
+               strncpy(object->comm, "hardirq", sizeof(object->comm));
+       } else if (in_softirq()) {
+               object->pid = 0;
+               strncpy(object->comm, "softirq", sizeof(object->comm));
+       } else {
+               object->pid = current->pid;
+               /*
+                * There is a small chance of a race with set_task_comm(),
+                * however using get_task_comm() here may cause locking
+                * dependency issues with current->alloc_lock. In the worst
+                * case, the command line is not correct.
+                */
+               strncpy(object->comm, current->comm, sizeof(object->comm));
+       }
+
+       /* kernel backtrace */
+       trace.max_entries = MAX_TRACE;
+       trace.nr_entries = 0;
+       trace.entries = object->trace;
+       trace.skip = 1;
+       save_stack_trace(&trace);
+       object->trace_len = trace.nr_entries;
+
+       INIT_PRIO_TREE_NODE(&object->tree_node);
+       object->tree_node.start = ptr;
+       object->tree_node.last = ptr + size - 1;
+
+       write_lock_irqsave(&kmemleak_lock, flags);
+       min_addr = min(min_addr, ptr);
+       max_addr = max(max_addr, ptr + size);
+       node = prio_tree_insert(&object_tree_root, &object->tree_node);
+       /*
+        * The code calling the kernel does not yet have the pointer to the
+        * memory block to be able to free it.  However, we still hold the
+        * kmemleak_lock here in case parts of the kernel started freeing
+        * random memory blocks.
+        */
+       if (node != &object->tree_node) {
+               unsigned long flags;
+
+               kmemleak_panic("kmemleak: Cannot insert 0x%lx into the object "
+                              "search tree (already existing)\n", ptr);
+               object = lookup_object(ptr, 1);
+               spin_lock_irqsave(&object->lock, flags);
+               dump_object_info(object);
+               spin_unlock_irqrestore(&object->lock, flags);
+
+               goto out;
+       }
+       list_add_tail_rcu(&object->object_list, &object_list);
+out:
+       write_unlock_irqrestore(&kmemleak_lock, flags);
+}
+
+/*
+ * Remove the metadata (struct kmemleak_object) for a memory block from the
+ * object_list and object_tree_root and decrement its use_count.
+ */
+static void delete_object(unsigned long ptr)
+{
+       unsigned long flags;
+       struct kmemleak_object *object;
+
+       write_lock_irqsave(&kmemleak_lock, flags);
+       object = lookup_object(ptr, 0);
+       if (!object) {
+               kmemleak_warn("kmemleak: Freeing unknown object at 0x%08lx\n",
+                             ptr);
+               write_unlock_irqrestore(&kmemleak_lock, flags);
+               return;
+       }
+       prio_tree_remove(&object_tree_root, &object->tree_node);
+       list_del_rcu(&object->object_list);
+       write_unlock_irqrestore(&kmemleak_lock, flags);
+
+       WARN_ON(!(object->flags & OBJECT_ALLOCATED));
+       WARN_ON(atomic_read(&object->use_count) < 1);
+
+       /*
+        * Locking here also ensures that the corresponding memory block
+        * cannot be freed when it is being scanned.
+        */
+       spin_lock_irqsave(&object->lock, flags);
+       if (object->flags & OBJECT_REPORTED)
+               print_referenced(object);
+       object->flags &= ~OBJECT_ALLOCATED;
+       spin_unlock_irqrestore(&object->lock, flags);
+       put_object(object);
+}
+
+/*
+ * Make a object permanently as gray-colored so that it can no longer be
+ * reported as a leak. This is used in general to mark a false positive.
+ */
+static void make_gray_object(unsigned long ptr)
+{
+       unsigned long flags;
+       struct kmemleak_object *object;
+
+       object = find_and_get_object(ptr, 0);
+       if (!object) {
+               kmemleak_warn("kmemleak: Graying unknown object at 0x%08lx\n",
+                             ptr);
+               return;
+       }
+
+       spin_lock_irqsave(&object->lock, flags);
+       object->min_count = 0;
+       spin_unlock_irqrestore(&object->lock, flags);
+       put_object(object);
+}
+
+/*
+ * Mark the object as black-colored so that it is ignored from scans and
+ * reporting.
+ */
+static void make_black_object(unsigned long ptr)
+{
+       unsigned long flags;
+       struct kmemleak_object *object;
+
+       object = find_and_get_object(ptr, 0);
+       if (!object) {
+               kmemleak_warn("kmemleak: Blacking unknown object at 0x%08lx\n",
+                             ptr);
+               return;
+       }
+
+       spin_lock_irqsave(&object->lock, flags);
+       object->min_count = -1;
+       spin_unlock_irqrestore(&object->lock, flags);
+       put_object(object);
+}
+
+/*
+ * Add a scanning area to the object. If at least one such area is added,
+ * kmemleak will only scan these ranges rather than the whole memory block.
+ */
+static void add_scan_area(unsigned long ptr, unsigned long offset,
+                         size_t length, gfp_t gfp)
+{
+       unsigned long flags;
+       struct kmemleak_object *object;
+       struct kmemleak_scan_area *area;
+
+       object = find_and_get_object(ptr, 0);
+       if (!object) {
+               kmemleak_warn("kmemleak: Adding scan area to unknown "
+                             "object at 0x%08lx\n", ptr);
+               return;
+       }
+
+       area = kmem_cache_alloc(scan_area_cache, gfp & ~GFP_SLAB_BUG_MASK);
+       if (!area) {
+               kmemleak_warn("kmemleak: Cannot allocate a scan area\n");
+               goto out;
+       }
+
+       spin_lock_irqsave(&object->lock, flags);
+       if (offset + length > object->size) {
+               kmemleak_warn("kmemleak: Scan area larger than object "
+                             "0x%08lx\n", ptr);
+               dump_object_info(object);
+               kmem_cache_free(scan_area_cache, area);
+               goto out_unlock;
+       }
+
+       INIT_HLIST_NODE(&area->node);
+       area->offset = offset;
+       area->length = length;
+
+       hlist_add_head(&area->node, &object->area_list);
+out_unlock:
+       spin_unlock_irqrestore(&object->lock, flags);
+out:
+       put_object(object);
+}
+
+/*
+ * Set the OBJECT_NO_SCAN flag for the object corresponding to the give
+ * pointer. Such object will not be scanned by kmemleak but references to it
+ * are searched.
+ */
+static void object_no_scan(unsigned long ptr)
+{
+       unsigned long flags;
+       struct kmemleak_object *object;
+
+       object = find_and_get_object(ptr, 0);
+       if (!object) {
+               kmemleak_warn("kmemleak: Not scanning unknown object at "
+                             "0x%08lx\n", ptr);
+               return;
+       }
+
+       spin_lock_irqsave(&object->lock, flags);
+       object->flags |= OBJECT_NO_SCAN;
+       spin_unlock_irqrestore(&object->lock, flags);
+       put_object(object);
+}
+
+/*
+ * Log an early kmemleak_* call to the early_log buffer. These calls will be
+ * processed later once kmemleak is fully initialized.
+ */
+static void log_early(int op_type, const void *ptr, size_t size,
+                     int min_count, unsigned long offset, size_t length)
+{
+       unsigned long flags;
+       struct early_log *log;
+
+       if (crt_early_log >= ARRAY_SIZE(early_log)) {
+               kmemleak_panic("kmemleak: Early log buffer exceeded\n");
+               return;
+       }
+
+       /*
+        * There is no need for locking since the kernel is still in UP mode
+        * at this stage. Disabling the IRQs is enough.
+        */
+       local_irq_save(flags);
+       log = &early_log[crt_early_log];
+       log->op_type = op_type;
+       log->ptr = ptr;
+       log->size = size;
+       log->min_count = min_count;
+       log->offset = offset;
+       log->length = length;
+       crt_early_log++;
+       local_irq_restore(flags);
+}
+
+/*
+ * Memory allocation function callback. This function is called from the
+ * kernel allocators when a new block is allocated (kmem_cache_alloc, kmalloc,
+ * vmalloc etc.).
+ */
+void kmemleak_alloc(const void *ptr, size_t size, int min_count, gfp_t gfp)
+{
+       pr_debug("%s(0x%p, %zu, %d)\n", __func__, ptr, size, min_count);
+
+       if (atomic_read(&kmemleak_enabled) && ptr && !IS_ERR(ptr))
+               create_object((unsigned long)ptr, size, min_count, gfp);
+       else if (atomic_read(&kmemleak_early_log))
+               log_early(KMEMLEAK_ALLOC, ptr, size, min_count, 0, 0);
+}
+EXPORT_SYMBOL_GPL(kmemleak_alloc);
+
+/*
+ * Memory freeing function callback. This function is called from the kernel
+ * allocators when a block is freed (kmem_cache_free, kfree, vfree etc.).
+ */
+void kmemleak_free(const void *ptr)
+{
+       pr_debug("%s(0x%p)\n", __func__, ptr);
+
+       if (atomic_read(&kmemleak_enabled) && ptr && !IS_ERR(ptr))
+               delete_object((unsigned long)ptr);
+       else if (atomic_read(&kmemleak_early_log))
+               log_early(KMEMLEAK_FREE, ptr, 0, 0, 0, 0);
+}
+EXPORT_SYMBOL_GPL(kmemleak_free);
+
+/*
+ * Mark an already allocated memory block as a false positive. This will cause
+ * the block to no longer be reported as leak and always be scanned.
+ */
+void kmemleak_not_leak(const void *ptr)
+{
+       pr_debug("%s(0x%p)\n", __func__, ptr);
+
+       if (atomic_read(&kmemleak_enabled) && ptr && !IS_ERR(ptr))
+               make_gray_object((unsigned long)ptr);
+       else if (atomic_read(&kmemleak_early_log))
+               log_early(KMEMLEAK_NOT_LEAK, ptr, 0, 0, 0, 0);
+}
+EXPORT_SYMBOL(kmemleak_not_leak);
+
+/*
+ * Ignore a memory block. This is usually done when it is known that the
+ * corresponding block is not a leak and does not contain any references to
+ * other allocated memory blocks.
+ */
+void kmemleak_ignore(const void *ptr)
+{
+       pr_debug("%s(0x%p)\n", __func__, ptr);
+
+       if (atomic_read(&kmemleak_enabled) && ptr && !IS_ERR(ptr))
+               make_black_object((unsigned long)ptr);
+       else if (atomic_read(&kmemleak_early_log))
+               log_early(KMEMLEAK_IGNORE, ptr, 0, 0, 0, 0);
+}
+EXPORT_SYMBOL(kmemleak_ignore);
+
+/*
+ * Limit the range to be scanned in an allocated memory block.
+ */
+void kmemleak_scan_area(const void *ptr, unsigned long offset, size_t length,
+                       gfp_t gfp)
+{
+       pr_debug("%s(0x%p)\n", __func__, ptr);
+
+       if (atomic_read(&kmemleak_enabled) && ptr && !IS_ERR(ptr))
+               add_scan_area((unsigned long)ptr, offset, length, gfp);
+       else if (atomic_read(&kmemleak_early_log))
+               log_early(KMEMLEAK_SCAN_AREA, ptr, 0, 0, offset, length);
+}
+EXPORT_SYMBOL(kmemleak_scan_area);
+
+/*
+ * Inform kmemleak not to scan the given memory block.
+ */
+void kmemleak_no_scan(const void *ptr)
+{
+       pr_debug("%s(0x%p)\n", __func__, ptr);
+
+       if (atomic_read(&kmemleak_enabled) && ptr && !IS_ERR(ptr))
+               object_no_scan((unsigned long)ptr);
+       else if (atomic_read(&kmemleak_early_log))
+               log_early(KMEMLEAK_NO_SCAN, ptr, 0, 0, 0, 0);
+}
+EXPORT_SYMBOL(kmemleak_no_scan);
+
+/*
+ * Yield the CPU so that other tasks get a chance to run.  The yielding is
+ * rate-limited to avoid excessive number of calls to the schedule() function
+ * during memory scanning.
+ */
+static void scan_yield(void)
+{
+       might_sleep();
+
+       if (time_is_before_eq_jiffies(next_scan_yield)) {
+               schedule();
+               next_scan_yield = jiffies + jiffies_scan_yield;
+       }
+}
+
+/*
+ * Memory scanning is a long process and it needs to be interruptable. This
+ * function checks whether such interrupt condition occured.
+ */
+static int scan_should_stop(void)
+{
+       if (!atomic_read(&kmemleak_enabled))
+               return 1;
+
+       /*
+        * This function may be called from either process or kthread context,
+        * hence the need to check for both stop conditions.
+        */
+       if (current->mm)
+               return signal_pending(current);
+       else
+               return kthread_should_stop();
+
+       return 0;
+}
+
+/*
+ * Scan a memory block (exclusive range) for valid pointers and add those
+ * found to the gray list.
+ */
+static void scan_block(void *_start, void *_end,
+                      struct kmemleak_object *scanned)
+{
+       unsigned long *ptr;
+       unsigned long *start = PTR_ALIGN(_start, BYTES_PER_POINTER);
+       unsigned long *end = _end - (BYTES_PER_POINTER - 1);
+
+       for (ptr = start; ptr < end; ptr++) {
+               unsigned long flags;
+               unsigned long pointer = *ptr;
+               struct kmemleak_object *object;
+
+               if (scan_should_stop())
+                       break;
+
+               /*
+                * When scanning a memory block with a corresponding
+                * kmemleak_object, the CPU yielding is handled in the calling
+                * code since it holds the object->lock to avoid the block
+                * freeing.
+                */
+               if (!scanned)
+                       scan_yield();
+
+               object = find_and_get_object(pointer, 1);
+               if (!object)
+                       continue;
+               if (object == scanned) {
+                       /* self referenced, ignore */
+                       put_object(object);
+                       continue;
+               }
+
+               /*
+                * Avoid the lockdep recursive warning on object->lock being
+                * previously acquired in scan_object(). These locks are
+                * enclosed by scan_mutex.
+                */
+               spin_lock_irqsave_nested(&object->lock, flags,
+                                        SINGLE_DEPTH_NESTING);
+               if (!color_white(object)) {
+                       /* non-orphan, ignored or new */
+                       spin_unlock_irqrestore(&object->lock, flags);
+                       put_object(object);
+                       continue;
+               }
+
+               /*
+                * Increase the object's reference count (number of pointers
+                * to the memory block). If this count reaches the required
+                * minimum, the object's color will become gray and it will be
+                * added to the gray_list.
+                */
+               object->count++;
+               if (color_gray(object))
+                       list_add_tail(&object->gray_list, &gray_list);
+               else
+                       put_object(object);
+               spin_unlock_irqrestore(&object->lock, flags);
+       }
+}
+
+/*
+ * Scan a memory block corresponding to a kmemleak_object. A condition is
+ * that object->use_count >= 1.
+ */
+static void scan_object(struct kmemleak_object *object)
+{
+       struct kmemleak_scan_area *area;
+       struct hlist_node *elem;
+       unsigned long flags;
+
+       /*
+        * Once the object->lock is aquired, the corresponding memory block
+        * cannot be freed (the same lock is aquired in delete_object).
+        */
+       spin_lock_irqsave(&object->lock, flags);
+       if (object->flags & OBJECT_NO_SCAN)
+               goto out;
+       if (!(object->flags & OBJECT_ALLOCATED))
+               /* already freed object */
+               goto out;
+       if (hlist_empty(&object->area_list))
+               scan_block((void *)object->pointer,
+                          (void *)(object->pointer + object->size), object);
+       else
+               hlist_for_each_entry(area, elem, &object->area_list, node)
+                       scan_block((void *)(object->pointer + area->offset),
+                                  (void *)(object->pointer + area->offset
+                                           + area->length), object);
+out:
+       spin_unlock_irqrestore(&object->lock, flags);
+}
+
+/*
+ * Scan data sections and all the referenced memory blocks allocated via the
+ * kernel's standard allocators. This function must be called with the
+ * scan_mutex held.
+ */
+static void kmemleak_scan(void)
+{
+       unsigned long flags;
+       struct kmemleak_object *object, *tmp;
+       struct task_struct *task;
+       int i;
+
+       /* prepare the kmemleak_object's */
+       rcu_read_lock();
+       list_for_each_entry_rcu(object, &object_list, object_list) {
+               spin_lock_irqsave(&object->lock, flags);
+#ifdef DEBUG
+               /*
+                * With a few exceptions there should be a maximum of
+                * 1 reference to any object at this point.
+                */
+               if (atomic_read(&object->use_count) > 1) {
+                       pr_debug("kmemleak: object->use_count = %d\n",
+                                atomic_read(&object->use_count));
+                       dump_object_info(object);
+               }
+#endif
+               /* reset the reference count (whiten the object) */
+               object->count = 0;
+               if (color_gray(object) && get_object(object))
+                       list_add_tail(&object->gray_list, &gray_list);
+
+               spin_unlock_irqrestore(&object->lock, flags);
+       }
+       rcu_read_unlock();
+
+       /* data/bss scanning */
+       scan_block(_sdata, _edata, NULL);
+       scan_block(__bss_start, __bss_stop, NULL);
+
+#ifdef CONFIG_SMP
+       /* per-cpu sections scanning */
+       for_each_possible_cpu(i)
+               scan_block(__per_cpu_start + per_cpu_offset(i),
+                          __per_cpu_end + per_cpu_offset(i), NULL);
+#endif
+
+       /*
+        * Struct page scanning for each node. The code below is not yet safe
+        * with MEMORY_HOTPLUG.
+        */
+       for_each_online_node(i) {
+               pg_data_t *pgdat = NODE_DATA(i);
+               unsigned long start_pfn = pgdat->node_start_pfn;
+               unsigned long end_pfn = start_pfn + pgdat->node_spanned_pages;
+               unsigned long pfn;
+
+               for (pfn = start_pfn; pfn < end_pfn; pfn++) {
+                       struct page *page;
+
+                       if (!pfn_valid(pfn))
+                               continue;
+                       page = pfn_to_page(pfn);
+                       /* only scan if page is in use */
+                       if (page_count(page) == 0)
+                               continue;
+                       scan_block(page, page + 1, NULL);
+               }
+       }
+
+       /*
+        * Scanning the task stacks may introduce false negatives and it is
+        * not enabled by default.
+        */
+       if (kmemleak_stack_scan) {
+               read_lock(&tasklist_lock);
+               for_each_process(task)
+                       scan_block(task_stack_page(task),
+                                  task_stack_page(task) + THREAD_SIZE, NULL);
+               read_unlock(&tasklist_lock);
+       }
+
+       /*
+        * Scan the objects already referenced from the sections scanned
+        * above. More objects will be referenced and, if there are no memory
+        * leaks, all the objects will be scanned. The list traversal is safe
+        * for both tail additions and removals from inside the loop. The
+        * kmemleak objects cannot be freed from outside the loop because their
+        * use_count was increased.
+        */
+       object = list_entry(gray_list.next, typeof(*object), gray_list);
+       while (&object->gray_list != &gray_list) {
+               scan_yield();
+
+               /* may add new objects to the list */
+               if (!scan_should_stop())
+                       scan_object(object);
+
+               tmp = list_entry(object->gray_list.next, typeof(*object),
+                                gray_list);
+
+               /* remove the object from the list and release it */
+               list_del(&object->gray_list);
+               put_object(object);
+
+               object = tmp;
+       }
+       WARN_ON(!list_empty(&gray_list));
+}
+
+/*
+ * Thread function performing automatic memory scanning. Unreferenced objects
+ * at the end of a memory scan are reported but only the first time.
+ */
+static int kmemleak_scan_thread(void *arg)
+{
+       static int first_run = 1;
+
+       pr_info("kmemleak: Automatic memory scanning thread started\n");
+
+       /*
+        * Wait before the first scan to allow the system to fully initialize.
+        */
+       if (first_run) {
+               first_run = 0;
+               ssleep(SECS_FIRST_SCAN);
+       }
+
+       while (!kthread_should_stop()) {
+               struct kmemleak_object *object;
+               signed long timeout = jiffies_scan_wait;
+
+               mutex_lock(&scan_mutex);
+
+               kmemleak_scan();
+               reported_leaks = 0;
+
+               rcu_read_lock();
+               list_for_each_entry_rcu(object, &object_list, object_list) {
+                       unsigned long flags;
+
+                       if (reported_leaks >= REPORTS_NR)
+                               break;
+                       spin_lock_irqsave(&object->lock, flags);
+                       if (!(object->flags & OBJECT_REPORTED) &&
+                           unreferenced_object(object)) {
+                               print_unreferenced(NULL, object);
+                               object->flags |= OBJECT_REPORTED;
+                               reported_leaks++;
+                       } else if ((object->flags & OBJECT_REPORTED) &&
+                                  referenced_object(object)) {
+                               print_referenced(object);
+                               object->flags &= ~OBJECT_REPORTED;
+                       }
+                       spin_unlock_irqrestore(&object->lock, flags);
+               }
+               rcu_read_unlock();
+
+               mutex_unlock(&scan_mutex);
+               /* wait before the next scan */
+               while (timeout && !kthread_should_stop())
+                       timeout = schedule_timeout_interruptible(timeout);
+       }
+
+       pr_info("kmemleak: Automatic memory scanning thread ended\n");
+
+       return 0;
+}
+
+/*
+ * Start the automatic memory scanning thread. This function must be called
+ * with the kmemleak_mutex held.
+ */
+void start_scan_thread(void)
+{
+       if (scan_thread)
+               return;
+       scan_thread = kthread_run(kmemleak_scan_thread, NULL, "kmemleak");
+       if (IS_ERR(scan_thread)) {
+               pr_warning("kmemleak: Failed to create the scan thread\n");
+               scan_thread = NULL;
+       }
+}
+
+/*
+ * Stop the automatic memory scanning thread. This function must be called
+ * with the kmemleak_mutex held.
+ */
+void stop_scan_thread(void)
+{
+       if (scan_thread) {
+               kthread_stop(scan_thread);
+               scan_thread = NULL;
+       }
+}
+
+/*
+ * Iterate over the object_list and return the first valid object at or after
+ * the required position with its use_count incremented. The function triggers
+ * a memory scanning when the pos argument points to the first position.
+ */
+static void *kmemleak_seq_start(struct seq_file *seq, loff_t *pos)
+{
+       struct kmemleak_object *object;
+       loff_t n = *pos;
+
+       if (!n) {
+               kmemleak_scan();
+               reported_leaks = 0;
+       }
+       if (reported_leaks >= REPORTS_NR)
+               return NULL;
+
+       rcu_read_lock();
+       list_for_each_entry_rcu(object, &object_list, object_list) {
+               if (n-- > 0)
+                       continue;
+               if (get_object(object))
+                       goto out;
+       }
+       object = NULL;
+out:
+       rcu_read_unlock();
+       return object;
+}
+
+/*
+ * Return the next object in the object_list. The function decrements the
+ * use_count of the previous object and increases that of the next one.
+ */
+static void *kmemleak_seq_next(struct seq_file *seq, void *v, loff_t *pos)
+{
+       struct kmemleak_object *prev_obj = v;
+       struct kmemleak_object *next_obj = NULL;
+       struct list_head *n = &prev_obj->object_list;
+
+       ++(*pos);
+       if (reported_leaks >= REPORTS_NR)
+               goto out;
+
+       rcu_read_lock();
+       list_for_each_continue_rcu(n, &object_list) {
+               next_obj = list_entry(n, struct kmemleak_object, object_list);
+               if (get_object(next_obj))
+                       break;
+       }
+       rcu_read_unlock();
+out:
+       put_object(prev_obj);
+       return next_obj;
+}
+
+/*
+ * Decrement the use_count of the last object required, if any.
+ */
+static void kmemleak_seq_stop(struct seq_file *seq, void *v)
+{
+       if (v)
+               put_object(v);
+}
+
+/*
+ * Print the information for an unreferenced object to the seq file.
+ */
+static int kmemleak_seq_show(struct seq_file *seq, void *v)
+{
+       struct kmemleak_object *object = v;
+       unsigned long flags;
+
+       spin_lock_irqsave(&object->lock, flags);
+       if (!unreferenced_object(object))
+               goto out;
+       print_unreferenced(seq, object);
+       reported_leaks++;
+out:
+       spin_unlock_irqrestore(&object->lock, flags);
+       return 0;
+}
+
+static const struct seq_operations kmemleak_seq_ops = {
+       .start = kmemleak_seq_start,
+       .next  = kmemleak_seq_next,
+       .stop  = kmemleak_seq_stop,
+       .show  = kmemleak_seq_show,
+};
+
+static int kmemleak_open(struct inode *inode, struct file *file)
+{
+       int ret = 0;
+
+       if (!atomic_read(&kmemleak_enabled))
+               return -EBUSY;
+
+       ret = mutex_lock_interruptible(&kmemleak_mutex);
+       if (ret < 0)
+               goto out;
+       if (file->f_mode & FMODE_READ) {
+               ret = mutex_lock_interruptible(&scan_mutex);
+               if (ret < 0)
+                       goto kmemleak_unlock;
+               ret = seq_open(file, &kmemleak_seq_ops);
+               if (ret < 0)
+                       goto scan_unlock;
+       }
+       return ret;
+
+scan_unlock:
+       mutex_unlock(&scan_mutex);
+kmemleak_unlock:
+       mutex_unlock(&kmemleak_mutex);
+out:
+       return ret;
+}
+
+static int kmemleak_release(struct inode *inode, struct file *file)
+{
+       int ret = 0;
+
+       if (file->f_mode & FMODE_READ) {
+               seq_release(inode, file);
+               mutex_unlock(&scan_mutex);
+       }
+       mutex_unlock(&kmemleak_mutex);
+
+       return ret;
+}
+
+/*
+ * File write operation to configure kmemleak at run-time. The following
+ * commands can be written to the /sys/kernel/debug/kmemleak file:
+ *   off       - disable kmemleak (irreversible)
+ *   stack=on  - enable the task stacks scanning
+ *   stack=off - disable the tasks stacks scanning
+ *   scan=on   - start the automatic memory scanning thread
+ *   scan=off  - stop the automatic memory scanning thread
+ *   scan=...  - set the automatic memory scanning period in seconds (0 to
+ *               disable it)
+ */
+static ssize_t kmemleak_write(struct file *file, const char __user *user_buf,
+                             size_t size, loff_t *ppos)
+{
+       char buf[64];
+       int buf_size;
+
+       if (!atomic_read(&kmemleak_enabled))
+               return -EBUSY;
+
+       buf_size = min(size, (sizeof(buf) - 1));
+       if (strncpy_from_user(buf, user_buf, buf_size) < 0)
+               return -EFAULT;
+       buf[buf_size] = 0;
+
+       if (strncmp(buf, "off", 3) == 0)
+               kmemleak_disable();
+       else if (strncmp(buf, "stack=on", 8) == 0)
+               kmemleak_stack_scan = 1;
+       else if (strncmp(buf, "stack=off", 9) == 0)
+               kmemleak_stack_scan = 0;
+       else if (strncmp(buf, "scan=on", 7) == 0)
+               start_scan_thread();
+       else if (strncmp(buf, "scan=off", 8) == 0)
+               stop_scan_thread();
+       else if (strncmp(buf, "scan=", 5) == 0) {
+               unsigned long secs;
+               int err;
+
+               err = strict_strtoul(buf + 5, 0, &secs);
+               if (err < 0)
+                       return err;
+               stop_scan_thread();
+               if (secs) {
+                       jiffies_scan_wait = msecs_to_jiffies(secs * 1000);
+                       start_scan_thread();
+               }
+       } else
+               return -EINVAL;
+
+       /* ignore the rest of the buffer, only one command at a time */
+       *ppos += size;
+       return size;
+}
+
+static const struct file_operations kmemleak_fops = {
+       .owner          = THIS_MODULE,
+       .open           = kmemleak_open,
+       .read           = seq_read,
+       .write          = kmemleak_write,
+       .llseek         = seq_lseek,
+       .release        = kmemleak_release,
+};
+
+/*
+ * Perform the freeing of the kmemleak internal objects after waiting for any
+ * current memory scan to complete.
+ */
+static int kmemleak_cleanup_thread(void *arg)
+{
+       struct kmemleak_object *object;
+
+       mutex_lock(&kmemleak_mutex);
+       stop_scan_thread();
+       mutex_unlock(&kmemleak_mutex);
+
+       mutex_lock(&scan_mutex);
+       rcu_read_lock();
+       list_for_each_entry_rcu(object, &object_list, object_list)
+               delete_object(object->pointer);
+       rcu_read_unlock();
+       mutex_unlock(&scan_mutex);
+
+       return 0;
+}
+
+/*
+ * Start the clean-up thread.
+ */
+static void kmemleak_cleanup(void)
+{
+       struct task_struct *cleanup_thread;
+
+       cleanup_thread = kthread_run(kmemleak_cleanup_thread, NULL,
+                                    "kmemleak-clean");
+       if (IS_ERR(cleanup_thread))
+               pr_warning("kmemleak: Failed to create the clean-up thread\n");
+}
+
+/*
+ * Disable kmemleak. No memory allocation/freeing will be traced once this
+ * function is called. Disabling kmemleak is an irreversible operation.
+ */
+static void kmemleak_disable(void)
+{
+       /* atomically check whether it was already invoked */
+       if (atomic_cmpxchg(&kmemleak_error, 0, 1))
+               return;
+
+       /* stop any memory operation tracing */
+       atomic_set(&kmemleak_early_log, 0);
+       atomic_set(&kmemleak_enabled, 0);
+
+       /* check whether it is too early for a kernel thread */
+       if (atomic_read(&kmemleak_initialized))
+               kmemleak_cleanup();
+
+       pr_info("Kernel memory leak detector disabled\n");
+}
+
+/*
+ * Allow boot-time kmemleak disabling (enabled by default).
+ */
+static int kmemleak_boot_config(char *str)
+{
+       if (!str)
+               return -EINVAL;
+       if (strcmp(str, "off") == 0)
+               kmemleak_disable();
+       else if (strcmp(str, "on") != 0)
+               return -EINVAL;
+       return 0;
+}
+early_param("kmemleak", kmemleak_boot_config);
+
+/*
+ * Kkmemleak initialization.
+ */
+void __init kmemleak_init(void)
+{
+       int i;
+       unsigned long flags;
+
+       jiffies_scan_yield = msecs_to_jiffies(MSECS_SCAN_YIELD);
+       jiffies_min_age = msecs_to_jiffies(MSECS_MIN_AGE);
+       jiffies_scan_wait = msecs_to_jiffies(SECS_SCAN_WAIT * 1000);
+
+       object_cache = KMEM_CACHE(kmemleak_object, SLAB_NOLEAKTRACE);
+       scan_area_cache = KMEM_CACHE(kmemleak_scan_area, SLAB_NOLEAKTRACE);
+       INIT_PRIO_TREE_ROOT(&object_tree_root);
+
+       /* the kernel is still in UP mode, so disabling the IRQs is enough */
+       local_irq_save(flags);
+       if (!atomic_read(&kmemleak_error)) {
+               atomic_set(&kmemleak_enabled, 1);
+               atomic_set(&kmemleak_early_log, 0);
+       }
+       local_irq_restore(flags);
+
+       /*
+        * This is the point where tracking allocations is safe. Automatic
+        * scanning is started during the late initcall. Add the early logged
+        * callbacks to the kmemleak infrastructure.
+        */
+       for (i = 0; i < crt_early_log; i++) {
+               struct early_log *log = &early_log[i];
+
+               switch (log->op_type) {
+               case KMEMLEAK_ALLOC:
+                       kmemleak_alloc(log->ptr, log->size, log->min_count,
+                                      GFP_KERNEL);
+                       break;
+               case KMEMLEAK_FREE:
+                       kmemleak_free(log->ptr);
+                       break;
+               case KMEMLEAK_NOT_LEAK:
+                       kmemleak_not_leak(log->ptr);
+                       break;
+               case KMEMLEAK_IGNORE:
+                       kmemleak_ignore(log->ptr);
+                       break;
+               case KMEMLEAK_SCAN_AREA:
+                       kmemleak_scan_area(log->ptr, log->offset, log->length,
+                                          GFP_KERNEL);
+                       break;
+               case KMEMLEAK_NO_SCAN:
+                       kmemleak_no_scan(log->ptr);
+                       break;
+               default:
+                       WARN_ON(1);
+               }
+       }
+}
+
+/*
+ * Late initialization function.
+ */
+static int __init kmemleak_late_init(void)
+{
+       struct dentry *dentry;
+
+       atomic_set(&kmemleak_initialized, 1);
+
+       if (atomic_read(&kmemleak_error)) {
+               /*
+                * Some error occured and kmemleak was disabled. There is a
+                * small chance that kmemleak_disable() was called immediately
+                * after setting kmemleak_initialized and we may end up with
+                * two clean-up threads but serialized by scan_mutex.
+                */
+               kmemleak_cleanup();
+               return -ENOMEM;
+       }
+
+       dentry = debugfs_create_file("kmemleak", S_IRUGO, NULL, NULL,
+                                    &kmemleak_fops);
+       if (!dentry)
+               pr_warning("kmemleak: Failed to create the debugfs kmemleak "
+                          "file\n");
+       mutex_lock(&kmemleak_mutex);
+       start_scan_thread();
+       mutex_unlock(&kmemleak_mutex);
+
+       pr_info("Kernel memory leak detector initialized\n");
+
+       return 0;
+}
+late_initcall(kmemleak_late_init);