tree-wide: fix assorted typos all over the place
[safe/jmp/linux-2.6] / lib / dma-debug.c
index 20d6cdb..d9b08e0 100644 (file)
  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
  */
 
+#include <linux/scatterlist.h>
+#include <linux/dma-mapping.h>
+#include <linux/stacktrace.h>
 #include <linux/dma-debug.h>
 #include <linux/spinlock.h>
 #include <linux/debugfs.h>
+#include <linux/uaccess.h>
+#include <linux/device.h>
 #include <linux/types.h>
+#include <linux/sched.h>
+#include <linux/ctype.h>
 #include <linux/list.h>
 #include <linux/slab.h>
 
+#include <asm/sections.h>
+
 #define HASH_SIZE       1024ULL
 #define HASH_FN_SHIFT   13
 #define HASH_FN_MASK    (HASH_SIZE - 1)
@@ -35,6 +44,8 @@ enum {
        dma_debug_coherent,
 };
 
+#define DMA_DEBUG_STACKTRACE_ENTRIES 5
+
 struct dma_debug_entry {
        struct list_head list;
        struct device    *dev;
@@ -45,12 +56,16 @@ struct dma_debug_entry {
        int              direction;
        int              sg_call_ents;
        int              sg_mapped_ents;
+#ifdef CONFIG_STACKTRACE
+       struct           stack_trace stacktrace;
+       unsigned long    st_entries[DMA_DEBUG_STACKTRACE_ENTRIES];
+#endif
 };
 
 struct hash_bucket {
        struct list_head list;
        spinlock_t lock;
-} __cacheline_aligned_in_smp;
+} ____cacheline_aligned_in_smp;
 
 /* Hash list to save the allocated dma addresses */
 static struct hash_bucket dma_entry_hash[HASH_SIZE];
@@ -72,6 +87,7 @@ static u32 show_num_errors = 1;
 
 static u32 num_free_entries;
 static u32 min_free_entries;
+static u32 nr_total_entries;
 
 /* number of preallocated entries requested by kernel cmdline */
 static u32 req_entries;
@@ -84,6 +100,105 @@ static struct dentry *show_all_errors_dent  __read_mostly;
 static struct dentry *show_num_errors_dent  __read_mostly;
 static struct dentry *num_free_entries_dent __read_mostly;
 static struct dentry *min_free_entries_dent __read_mostly;
+static struct dentry *filter_dent           __read_mostly;
+
+/* per-driver filter related state */
+
+#define NAME_MAX_LEN   64
+
+static char                  current_driver_name[NAME_MAX_LEN] __read_mostly;
+static struct device_driver *current_driver                    __read_mostly;
+
+static DEFINE_RWLOCK(driver_name_lock);
+
+static const char *type2name[4] = { "single", "page",
+                                   "scather-gather", "coherent" };
+
+static const char *dir2name[4] = { "DMA_BIDIRECTIONAL", "DMA_TO_DEVICE",
+                                  "DMA_FROM_DEVICE", "DMA_NONE" };
+
+/* little merge helper - remove it after the merge window */
+#ifndef BUS_NOTIFY_UNBOUND_DRIVER
+#define BUS_NOTIFY_UNBOUND_DRIVER 0x0005
+#endif
+
+/*
+ * The access to some variables in this macro is racy. We can't use atomic_t
+ * here because all these variables are exported to debugfs. Some of them even
+ * writeable. This is also the reason why a lock won't help much. But anyway,
+ * the races are no big deal. Here is why:
+ *
+ *   error_count: the addition is racy, but the worst thing that can happen is
+ *                that we don't count some errors
+ *   show_num_errors: the subtraction is racy. Also no big deal because in
+ *                    worst case this will result in one warning more in the
+ *                    system log than the user configured. This variable is
+ *                    writeable via debugfs.
+ */
+static inline void dump_entry_trace(struct dma_debug_entry *entry)
+{
+#ifdef CONFIG_STACKTRACE
+       if (entry) {
+               pr_warning("Mapped at:\n");
+               print_stack_trace(&entry->stacktrace, 0);
+       }
+#endif
+}
+
+static bool driver_filter(struct device *dev)
+{
+       struct device_driver *drv;
+       unsigned long flags;
+       bool ret;
+
+       /* driver filter off */
+       if (likely(!current_driver_name[0]))
+               return true;
+
+       /* driver filter on and initialized */
+       if (current_driver && dev && dev->driver == current_driver)
+               return true;
+
+       /* driver filter on, but we can't filter on a NULL device... */
+       if (!dev)
+               return false;
+
+       if (current_driver || !current_driver_name[0])
+               return false;
+
+       /* driver filter on but not yet initialized */
+       drv = get_driver(dev->driver);
+       if (!drv)
+               return false;
+
+       /* lock to protect against change of current_driver_name */
+       read_lock_irqsave(&driver_name_lock, flags);
+
+       ret = false;
+       if (drv->name &&
+           strncmp(current_driver_name, drv->name, NAME_MAX_LEN - 1) == 0) {
+               current_driver = drv;
+               ret = true;
+       }
+
+       read_unlock_irqrestore(&driver_name_lock, flags);
+       put_driver(drv);
+
+       return ret;
+}
+
+#define err_printk(dev, entry, format, arg...) do {                    \
+               error_count += 1;                                       \
+               if (driver_filter(dev) &&                               \
+                   (show_all_errors || show_num_errors > 0)) {         \
+                       WARN(1, "%s %s: " format,                       \
+                            dev ? dev_driver_string(dev) : "NULL",     \
+                            dev ? dev_name(dev) : "NULL", ## arg);     \
+                       dump_entry_trace(entry);                        \
+               }                                                       \
+               if (!show_all_errors && show_num_errors > 0)            \
+                       show_num_errors -= 1;                           \
+       } while (0);
 
 /*
  * Hash related functions
@@ -131,15 +246,51 @@ static void put_hash_bucket(struct hash_bucket *bucket,
 static struct dma_debug_entry *hash_bucket_find(struct hash_bucket *bucket,
                                                struct dma_debug_entry *ref)
 {
-       struct dma_debug_entry *entry;
+       struct dma_debug_entry *entry, *ret = NULL;
+       int matches = 0, match_lvl, last_lvl = 0;
 
        list_for_each_entry(entry, &bucket->list, list) {
-               if ((entry->dev_addr == ref->dev_addr) &&
-                   (entry->dev == ref->dev))
+               if ((entry->dev_addr != ref->dev_addr) ||
+                   (entry->dev != ref->dev))
+                       continue;
+
+               /*
+                * Some drivers map the same physical address multiple
+                * times. Without a hardware IOMMU this results in the
+                * same device addresses being put into the dma-debug
+                * hash multiple times too. This can result in false
+                * positives being reported. Therefore we implement a
+                * best-fit algorithm here which returns the entry from
+                * the hash which fits best to the reference value
+                * instead of the first-fit.
+                */
+               matches += 1;
+               match_lvl = 0;
+               entry->size         == ref->size         ? ++match_lvl : 0;
+               entry->type         == ref->type         ? ++match_lvl : 0;
+               entry->direction    == ref->direction    ? ++match_lvl : 0;
+               entry->sg_call_ents == ref->sg_call_ents ? ++match_lvl : 0;
+
+               if (match_lvl == 4) {
+                       /* perfect-fit - return the result */
                        return entry;
+               } else if (match_lvl > last_lvl) {
+                       /*
+                        * We found an entry that fits better then the
+                        * previous one
+                        */
+                       last_lvl = match_lvl;
+                       ret      = entry;
+               }
        }
 
-       return NULL;
+       /*
+        * If we have multiple matches but no perfect-fit, just return
+        * NULL.
+        */
+       ret = (matches == 1) ? ret : NULL;
+
+       return ret;
 }
 
 /*
@@ -160,6 +311,36 @@ static void hash_bucket_del(struct dma_debug_entry *entry)
 }
 
 /*
+ * Dump mapping entries for debugging purposes
+ */
+void debug_dma_dump_mappings(struct device *dev)
+{
+       int idx;
+
+       for (idx = 0; idx < HASH_SIZE; idx++) {
+               struct hash_bucket *bucket = &dma_entry_hash[idx];
+               struct dma_debug_entry *entry;
+               unsigned long flags;
+
+               spin_lock_irqsave(&bucket->lock, flags);
+
+               list_for_each_entry(entry, &bucket->list, list) {
+                       if (!dev || dev == entry->dev) {
+                               dev_info(entry->dev,
+                                        "%s idx %d P=%Lx D=%Lx L=%Lx %s\n",
+                                        type2name[entry->type], idx,
+                                        (unsigned long long)entry->paddr,
+                                        entry->dev_addr, entry->size,
+                                        dir2name[entry->direction]);
+                       }
+               }
+
+               spin_unlock_irqrestore(&bucket->lock, flags);
+       }
+}
+EXPORT_SYMBOL(debug_dma_dump_mappings);
+
+/*
  * Wrapper function for adding an entry to the hash.
  * This function takes care of locking itself.
  */
@@ -173,6 +354,21 @@ static void add_dma_entry(struct dma_debug_entry *entry)
        put_hash_bucket(bucket, &flags);
 }
 
+static struct dma_debug_entry *__dma_entry_alloc(void)
+{
+       struct dma_debug_entry *entry;
+
+       entry = list_entry(free_entries.next, struct dma_debug_entry, list);
+       list_del(&entry->list);
+       memset(entry, 0, sizeof(*entry));
+
+       num_free_entries -= 1;
+       if (num_free_entries < min_free_entries)
+               min_free_entries = num_free_entries;
+
+       return entry;
+}
+
 /* struct dma_entry allocator
  *
  * The next two functions implement the allocator for
@@ -186,19 +382,19 @@ static struct dma_debug_entry *dma_entry_alloc(void)
        spin_lock_irqsave(&free_entries_lock, flags);
 
        if (list_empty(&free_entries)) {
-               printk(KERN_ERR "DMA-API: debugging out of memory "
-                               "- disabling\n");
+               pr_err("DMA-API: debugging out of memory - disabling\n");
                global_disable = true;
                goto out;
        }
 
-       entry = list_entry(free_entries.next, struct dma_debug_entry, list);
-       list_del(&entry->list);
-       memset(entry, 0, sizeof(*entry));
+       entry = __dma_entry_alloc();
 
-       num_free_entries -= 1;
-       if (num_free_entries < min_free_entries)
-               min_free_entries = num_free_entries;
+#ifdef CONFIG_STACKTRACE
+       entry->stacktrace.max_entries = DMA_DEBUG_STACKTRACE_ENTRIES;
+       entry->stacktrace.entries = entry->st_entries;
+       entry->stacktrace.skip = 2;
+       save_stack_trace(&entry->stacktrace);
+#endif
 
 out:
        spin_unlock_irqrestore(&free_entries_lock, flags);
@@ -220,6 +416,53 @@ static void dma_entry_free(struct dma_debug_entry *entry)
        spin_unlock_irqrestore(&free_entries_lock, flags);
 }
 
+int dma_debug_resize_entries(u32 num_entries)
+{
+       int i, delta, ret = 0;
+       unsigned long flags;
+       struct dma_debug_entry *entry;
+       LIST_HEAD(tmp);
+
+       spin_lock_irqsave(&free_entries_lock, flags);
+
+       if (nr_total_entries < num_entries) {
+               delta = num_entries - nr_total_entries;
+
+               spin_unlock_irqrestore(&free_entries_lock, flags);
+
+               for (i = 0; i < delta; i++) {
+                       entry = kzalloc(sizeof(*entry), GFP_KERNEL);
+                       if (!entry)
+                               break;
+
+                       list_add_tail(&entry->list, &tmp);
+               }
+
+               spin_lock_irqsave(&free_entries_lock, flags);
+
+               list_splice(&tmp, &free_entries);
+               nr_total_entries += i;
+               num_free_entries += i;
+       } else {
+               delta = nr_total_entries - num_entries;
+
+               for (i = 0; i < delta && !list_empty(&free_entries); i++) {
+                       entry = __dma_entry_alloc();
+                       kfree(entry);
+               }
+
+               nr_total_entries -= i;
+       }
+
+       if (nr_total_entries != num_entries)
+               ret = 1;
+
+       spin_unlock_irqrestore(&free_entries_lock, flags);
+
+       return ret;
+}
+EXPORT_SYMBOL(dma_debug_resize_entries);
+
 /*
  * DMA-API debugging init code
  *
@@ -244,8 +487,7 @@ static int prealloc_memory(u32 num_entries)
        num_free_entries = num_entries;
        min_free_entries = num_entries;
 
-       printk(KERN_INFO "DMA-API: preallocated %d debug entries\n",
-                       num_entries);
+       pr_info("DMA-API: preallocated %d debug entries\n", num_entries);
 
        return 0;
 
@@ -259,11 +501,102 @@ out_err:
        return -ENOMEM;
 }
 
+static ssize_t filter_read(struct file *file, char __user *user_buf,
+                          size_t count, loff_t *ppos)
+{
+       char buf[NAME_MAX_LEN + 1];
+       unsigned long flags;
+       int len;
+
+       if (!current_driver_name[0])
+               return 0;
+
+       /*
+        * We can't copy to userspace directly because current_driver_name can
+        * only be read under the driver_name_lock with irqs disabled. So
+        * create a temporary copy first.
+        */
+       read_lock_irqsave(&driver_name_lock, flags);
+       len = scnprintf(buf, NAME_MAX_LEN + 1, "%s\n", current_driver_name);
+       read_unlock_irqrestore(&driver_name_lock, flags);
+
+       return simple_read_from_buffer(user_buf, count, ppos, buf, len);
+}
+
+static ssize_t filter_write(struct file *file, const char __user *userbuf,
+                           size_t count, loff_t *ppos)
+{
+       char buf[NAME_MAX_LEN];
+       unsigned long flags;
+       size_t len;
+       int i;
+
+       /*
+        * We can't copy from userspace directly. Access to
+        * current_driver_name is protected with a write_lock with irqs
+        * disabled. Since copy_from_user can fault and may sleep we
+        * need to copy to temporary buffer first
+        */
+       len = min(count, (size_t)(NAME_MAX_LEN - 1));
+       if (copy_from_user(buf, userbuf, len))
+               return -EFAULT;
+
+       buf[len] = 0;
+
+       write_lock_irqsave(&driver_name_lock, flags);
+
+       /*
+        * Now handle the string we got from userspace very carefully.
+        * The rules are:
+        *         - only use the first token we got
+        *         - token delimiter is everything looking like a space
+        *           character (' ', '\n', '\t' ...)
+        *
+        */
+       if (!isalnum(buf[0])) {
+               /*
+                * If the first character userspace gave us is not
+                * alphanumerical then assume the filter should be
+                * switched off.
+                */
+               if (current_driver_name[0])
+                       pr_info("DMA-API: switching off dma-debug driver filter\n");
+               current_driver_name[0] = 0;
+               current_driver = NULL;
+               goto out_unlock;
+       }
+
+       /*
+        * Now parse out the first token and use it as the name for the
+        * driver to filter for.
+        */
+       for (i = 0; i < NAME_MAX_LEN; ++i) {
+               current_driver_name[i] = buf[i];
+               if (isspace(buf[i]) || buf[i] == ' ' || buf[i] == 0)
+                       break;
+       }
+       current_driver_name[i] = 0;
+       current_driver = NULL;
+
+       pr_info("DMA-API: enable driver filter for driver [%s]\n",
+               current_driver_name);
+
+out_unlock:
+       write_unlock_irqrestore(&driver_name_lock, flags);
+
+       return count;
+}
+
+const struct file_operations filter_fops = {
+       .read  = filter_read,
+       .write = filter_write,
+};
+
 static int dma_debug_fs_init(void)
 {
        dma_debug_dent = debugfs_create_dir("dma-api", NULL);
        if (!dma_debug_dent) {
-               printk(KERN_ERR "DMA-API: can not create debugfs directory\n");
+               pr_err("DMA-API: can not create debugfs directory\n");
                return -ENOMEM;
        }
 
@@ -302,6 +635,11 @@ static int dma_debug_fs_init(void)
        if (!min_free_entries_dent)
                goto out_err;
 
+       filter_dent = debugfs_create_file("driver_filter", 0644,
+                                         dma_debug_dent, NULL, &filter_fops);
+       if (!filter_dent)
+               goto out_err;
+
        return 0;
 
 out_err:
@@ -310,6 +648,65 @@ out_err:
        return -ENOMEM;
 }
 
+static int device_dma_allocations(struct device *dev)
+{
+       struct dma_debug_entry *entry;
+       unsigned long flags;
+       int count = 0, i;
+
+       local_irq_save(flags);
+
+       for (i = 0; i < HASH_SIZE; ++i) {
+               spin_lock(&dma_entry_hash[i].lock);
+               list_for_each_entry(entry, &dma_entry_hash[i].list, list) {
+                       if (entry->dev == dev)
+                               count += 1;
+               }
+               spin_unlock(&dma_entry_hash[i].lock);
+       }
+
+       local_irq_restore(flags);
+
+       return count;
+}
+
+static int dma_debug_device_change(struct notifier_block *nb,
+                                   unsigned long action, void *data)
+{
+       struct device *dev = data;
+       int count;
+
+
+       switch (action) {
+       case BUS_NOTIFY_UNBOUND_DRIVER:
+               count = device_dma_allocations(dev);
+               if (count == 0)
+                       break;
+               err_printk(dev, NULL, "DMA-API: device driver has pending "
+                               "DMA allocations while released from device "
+                               "[count=%d]\n", count);
+               break;
+       default:
+               break;
+       }
+
+       return 0;
+}
+
+void dma_debug_add_bus(struct bus_type *bus)
+{
+       struct notifier_block *nb;
+
+       nb = kzalloc(sizeof(struct notifier_block), GFP_KERNEL);
+       if (nb == NULL) {
+               pr_err("dma_debug_add_bus: out of memory\n");
+               return;
+       }
+
+       nb->notifier_call = dma_debug_device_change;
+
+       bus_register_notifier(bus, nb);
+}
 
 /*
  * Let the architectures decide how many entries should be preallocated.
@@ -323,12 +720,11 @@ void dma_debug_init(u32 num_entries)
 
        for (i = 0; i < HASH_SIZE; ++i) {
                INIT_LIST_HEAD(&dma_entry_hash[i].list);
-               dma_entry_hash[i].lock = SPIN_LOCK_UNLOCKED;
+               spin_lock_init(&dma_entry_hash[i].lock);
        }
 
        if (dma_debug_fs_init() != 0) {
-               printk(KERN_ERR "DMA-API: error creating debugfs entries "
-                               "- disabling\n");
+               pr_err("DMA-API: error creating debugfs entries - disabling\n");
                global_disable = true;
 
                return;
@@ -338,14 +734,15 @@ void dma_debug_init(u32 num_entries)
                num_entries = req_entries;
 
        if (prealloc_memory(num_entries) != 0) {
-               printk(KERN_ERR "DMA-API: debugging out of memory error "
-                               "- disabled\n");
+               pr_err("DMA-API: debugging out of memory error - disabled\n");
                global_disable = true;
 
                return;
        }
 
-       printk(KERN_INFO "DMA-API: debugging enabled by kernel config\n");
+       nr_total_entries = num_free_entries;
+
+       pr_info("DMA-API: debugging enabled by kernel config\n");
 }
 
 static __init int dma_debug_cmdline(char *str)
@@ -354,8 +751,7 @@ static __init int dma_debug_cmdline(char *str)
                return -EINVAL;
 
        if (strncmp(str, "off", 3) == 0) {
-               printk(KERN_INFO "DMA-API: debugging disabled on kernel "
-                                "command line\n");
+               pr_info("DMA-API: debugging disabled on kernel command line\n");
                global_disable = true;
        }
 
@@ -380,3 +776,524 @@ static __init int dma_debug_entries_cmdline(char *str)
 __setup("dma_debug=", dma_debug_cmdline);
 __setup("dma_debug_entries=", dma_debug_entries_cmdline);
 
+static void check_unmap(struct dma_debug_entry *ref)
+{
+       struct dma_debug_entry *entry;
+       struct hash_bucket *bucket;
+       unsigned long flags;
+
+       if (dma_mapping_error(ref->dev, ref->dev_addr)) {
+               err_printk(ref->dev, NULL, "DMA-API: device driver tries "
+                          "to free an invalid DMA memory address\n");
+               return;
+       }
+
+       bucket = get_hash_bucket(ref, &flags);
+       entry = hash_bucket_find(bucket, ref);
+
+       if (!entry) {
+               err_printk(ref->dev, NULL, "DMA-API: device driver tries "
+                          "to free DMA memory it has not allocated "
+                          "[device address=0x%016llx] [size=%llu bytes]\n",
+                          ref->dev_addr, ref->size);
+               goto out;
+       }
+
+       if (ref->size != entry->size) {
+               err_printk(ref->dev, entry, "DMA-API: device driver frees "
+                          "DMA memory with different size "
+                          "[device address=0x%016llx] [map size=%llu bytes] "
+                          "[unmap size=%llu bytes]\n",
+                          ref->dev_addr, entry->size, ref->size);
+       }
+
+       if (ref->type != entry->type) {
+               err_printk(ref->dev, entry, "DMA-API: device driver frees "
+                          "DMA memory with wrong function "
+                          "[device address=0x%016llx] [size=%llu bytes] "
+                          "[mapped as %s] [unmapped as %s]\n",
+                          ref->dev_addr, ref->size,
+                          type2name[entry->type], type2name[ref->type]);
+       } else if ((entry->type == dma_debug_coherent) &&
+                  (ref->paddr != entry->paddr)) {
+               err_printk(ref->dev, entry, "DMA-API: device driver frees "
+                          "DMA memory with different CPU address "
+                          "[device address=0x%016llx] [size=%llu bytes] "
+                          "[cpu alloc address=0x%016llx] "
+                          "[cpu free address=0x%016llx]",
+                          ref->dev_addr, ref->size,
+                          (unsigned long long)entry->paddr,
+                          (unsigned long long)ref->paddr);
+       }
+
+       if (ref->sg_call_ents && ref->type == dma_debug_sg &&
+           ref->sg_call_ents != entry->sg_call_ents) {
+               err_printk(ref->dev, entry, "DMA-API: device driver frees "
+                          "DMA sg list with different entry count "
+                          "[map count=%d] [unmap count=%d]\n",
+                          entry->sg_call_ents, ref->sg_call_ents);
+       }
+
+       /*
+        * This may be no bug in reality - but most implementations of the
+        * DMA API don't handle this properly, so check for it here
+        */
+       if (ref->direction != entry->direction) {
+               err_printk(ref->dev, entry, "DMA-API: device driver frees "
+                          "DMA memory with different direction "
+                          "[device address=0x%016llx] [size=%llu bytes] "
+                          "[mapped with %s] [unmapped with %s]\n",
+                          ref->dev_addr, ref->size,
+                          dir2name[entry->direction],
+                          dir2name[ref->direction]);
+       }
+
+       hash_bucket_del(entry);
+       dma_entry_free(entry);
+
+out:
+       put_hash_bucket(bucket, &flags);
+}
+
+static void check_for_stack(struct device *dev, void *addr)
+{
+       if (object_is_on_stack(addr))
+               err_printk(dev, NULL, "DMA-API: device driver maps memory from"
+                               "stack [addr=%p]\n", addr);
+}
+
+static inline bool overlap(void *addr, unsigned long len, void *start, void *end)
+{
+       unsigned long a1 = (unsigned long)addr;
+       unsigned long b1 = a1 + len;
+       unsigned long a2 = (unsigned long)start;
+       unsigned long b2 = (unsigned long)end;
+
+       return !(b1 <= a2 || a1 >= b2);
+}
+
+static void check_for_illegal_area(struct device *dev, void *addr, unsigned long len)
+{
+       if (overlap(addr, len, _text, _etext) ||
+           overlap(addr, len, __start_rodata, __end_rodata))
+               err_printk(dev, NULL, "DMA-API: device driver maps memory from kernel text or rodata [addr=%p] [len=%lu]\n", addr, len);
+}
+
+static void check_sync(struct device *dev,
+                      struct dma_debug_entry *ref,
+                      bool to_cpu)
+{
+       struct dma_debug_entry *entry;
+       struct hash_bucket *bucket;
+       unsigned long flags;
+
+       bucket = get_hash_bucket(ref, &flags);
+
+       entry = hash_bucket_find(bucket, ref);
+
+       if (!entry) {
+               err_printk(dev, NULL, "DMA-API: device driver tries "
+                               "to sync DMA memory it has not allocated "
+                               "[device address=0x%016llx] [size=%llu bytes]\n",
+                               (unsigned long long)ref->dev_addr, ref->size);
+               goto out;
+       }
+
+       if (ref->size > entry->size) {
+               err_printk(dev, entry, "DMA-API: device driver syncs"
+                               " DMA memory outside allocated range "
+                               "[device address=0x%016llx] "
+                               "[allocation size=%llu bytes] "
+                               "[sync offset+size=%llu]\n",
+                               entry->dev_addr, entry->size,
+                               ref->size);
+       }
+
+       if (ref->direction != entry->direction) {
+               err_printk(dev, entry, "DMA-API: device driver syncs "
+                               "DMA memory with different direction "
+                               "[device address=0x%016llx] [size=%llu bytes] "
+                               "[mapped with %s] [synced with %s]\n",
+                               (unsigned long long)ref->dev_addr, entry->size,
+                               dir2name[entry->direction],
+                               dir2name[ref->direction]);
+       }
+
+       if (entry->direction == DMA_BIDIRECTIONAL)
+               goto out;
+
+       if (to_cpu && !(entry->direction == DMA_FROM_DEVICE) &&
+                     !(ref->direction == DMA_TO_DEVICE))
+               err_printk(dev, entry, "DMA-API: device driver syncs "
+                               "device read-only DMA memory for cpu "
+                               "[device address=0x%016llx] [size=%llu bytes] "
+                               "[mapped with %s] [synced with %s]\n",
+                               (unsigned long long)ref->dev_addr, entry->size,
+                               dir2name[entry->direction],
+                               dir2name[ref->direction]);
+
+       if (!to_cpu && !(entry->direction == DMA_TO_DEVICE) &&
+                      !(ref->direction == DMA_FROM_DEVICE))
+               err_printk(dev, entry, "DMA-API: device driver syncs "
+                               "device write-only DMA memory to device "
+                               "[device address=0x%016llx] [size=%llu bytes] "
+                               "[mapped with %s] [synced with %s]\n",
+                               (unsigned long long)ref->dev_addr, entry->size,
+                               dir2name[entry->direction],
+                               dir2name[ref->direction]);
+
+out:
+       put_hash_bucket(bucket, &flags);
+
+}
+
+void debug_dma_map_page(struct device *dev, struct page *page, size_t offset,
+                       size_t size, int direction, dma_addr_t dma_addr,
+                       bool map_single)
+{
+       struct dma_debug_entry *entry;
+
+       if (unlikely(global_disable))
+               return;
+
+       if (unlikely(dma_mapping_error(dev, dma_addr)))
+               return;
+
+       entry = dma_entry_alloc();
+       if (!entry)
+               return;
+
+       entry->dev       = dev;
+       entry->type      = dma_debug_page;
+       entry->paddr     = page_to_phys(page) + offset;
+       entry->dev_addr  = dma_addr;
+       entry->size      = size;
+       entry->direction = direction;
+
+       if (map_single)
+               entry->type = dma_debug_single;
+
+       if (!PageHighMem(page)) {
+               void *addr = page_address(page) + offset;
+
+               check_for_stack(dev, addr);
+               check_for_illegal_area(dev, addr, size);
+       }
+
+       add_dma_entry(entry);
+}
+EXPORT_SYMBOL(debug_dma_map_page);
+
+void debug_dma_unmap_page(struct device *dev, dma_addr_t addr,
+                         size_t size, int direction, bool map_single)
+{
+       struct dma_debug_entry ref = {
+               .type           = dma_debug_page,
+               .dev            = dev,
+               .dev_addr       = addr,
+               .size           = size,
+               .direction      = direction,
+       };
+
+       if (unlikely(global_disable))
+               return;
+
+       if (map_single)
+               ref.type = dma_debug_single;
+
+       check_unmap(&ref);
+}
+EXPORT_SYMBOL(debug_dma_unmap_page);
+
+void debug_dma_map_sg(struct device *dev, struct scatterlist *sg,
+                     int nents, int mapped_ents, int direction)
+{
+       struct dma_debug_entry *entry;
+       struct scatterlist *s;
+       int i;
+
+       if (unlikely(global_disable))
+               return;
+
+       for_each_sg(sg, s, mapped_ents, i) {
+               entry = dma_entry_alloc();
+               if (!entry)
+                       return;
+
+               entry->type           = dma_debug_sg;
+               entry->dev            = dev;
+               entry->paddr          = sg_phys(s);
+               entry->size           = sg_dma_len(s);
+               entry->dev_addr       = sg_dma_address(s);
+               entry->direction      = direction;
+               entry->sg_call_ents   = nents;
+               entry->sg_mapped_ents = mapped_ents;
+
+               if (!PageHighMem(sg_page(s))) {
+                       check_for_stack(dev, sg_virt(s));
+                       check_for_illegal_area(dev, sg_virt(s), sg_dma_len(s));
+               }
+
+               add_dma_entry(entry);
+       }
+}
+EXPORT_SYMBOL(debug_dma_map_sg);
+
+static int get_nr_mapped_entries(struct device *dev,
+                                struct dma_debug_entry *ref)
+{
+       struct dma_debug_entry *entry;
+       struct hash_bucket *bucket;
+       unsigned long flags;
+       int mapped_ents;
+
+       bucket       = get_hash_bucket(ref, &flags);
+       entry        = hash_bucket_find(bucket, ref);
+       mapped_ents  = 0;
+
+       if (entry)
+               mapped_ents = entry->sg_mapped_ents;
+       put_hash_bucket(bucket, &flags);
+
+       return mapped_ents;
+}
+
+void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
+                       int nelems, int dir)
+{
+       struct scatterlist *s;
+       int mapped_ents = 0, i;
+
+       if (unlikely(global_disable))
+               return;
+
+       for_each_sg(sglist, s, nelems, i) {
+
+               struct dma_debug_entry ref = {
+                       .type           = dma_debug_sg,
+                       .dev            = dev,
+                       .paddr          = sg_phys(s),
+                       .dev_addr       = sg_dma_address(s),
+                       .size           = sg_dma_len(s),
+                       .direction      = dir,
+                       .sg_call_ents   = nelems,
+               };
+
+               if (mapped_ents && i >= mapped_ents)
+                       break;
+
+               if (!i)
+                       mapped_ents = get_nr_mapped_entries(dev, &ref);
+
+               check_unmap(&ref);
+       }
+}
+EXPORT_SYMBOL(debug_dma_unmap_sg);
+
+void debug_dma_alloc_coherent(struct device *dev, size_t size,
+                             dma_addr_t dma_addr, void *virt)
+{
+       struct dma_debug_entry *entry;
+
+       if (unlikely(global_disable))
+               return;
+
+       if (unlikely(virt == NULL))
+               return;
+
+       entry = dma_entry_alloc();
+       if (!entry)
+               return;
+
+       entry->type      = dma_debug_coherent;
+       entry->dev       = dev;
+       entry->paddr     = virt_to_phys(virt);
+       entry->size      = size;
+       entry->dev_addr  = dma_addr;
+       entry->direction = DMA_BIDIRECTIONAL;
+
+       add_dma_entry(entry);
+}
+EXPORT_SYMBOL(debug_dma_alloc_coherent);
+
+void debug_dma_free_coherent(struct device *dev, size_t size,
+                        void *virt, dma_addr_t addr)
+{
+       struct dma_debug_entry ref = {
+               .type           = dma_debug_coherent,
+               .dev            = dev,
+               .paddr          = virt_to_phys(virt),
+               .dev_addr       = addr,
+               .size           = size,
+               .direction      = DMA_BIDIRECTIONAL,
+       };
+
+       if (unlikely(global_disable))
+               return;
+
+       check_unmap(&ref);
+}
+EXPORT_SYMBOL(debug_dma_free_coherent);
+
+void debug_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
+                                  size_t size, int direction)
+{
+       struct dma_debug_entry ref;
+
+       if (unlikely(global_disable))
+               return;
+
+       ref.type         = dma_debug_single;
+       ref.dev          = dev;
+       ref.dev_addr     = dma_handle;
+       ref.size         = size;
+       ref.direction    = direction;
+       ref.sg_call_ents = 0;
+
+       check_sync(dev, &ref, true);
+}
+EXPORT_SYMBOL(debug_dma_sync_single_for_cpu);
+
+void debug_dma_sync_single_for_device(struct device *dev,
+                                     dma_addr_t dma_handle, size_t size,
+                                     int direction)
+{
+       struct dma_debug_entry ref;
+
+       if (unlikely(global_disable))
+               return;
+
+       ref.type         = dma_debug_single;
+       ref.dev          = dev;
+       ref.dev_addr     = dma_handle;
+       ref.size         = size;
+       ref.direction    = direction;
+       ref.sg_call_ents = 0;
+
+       check_sync(dev, &ref, false);
+}
+EXPORT_SYMBOL(debug_dma_sync_single_for_device);
+
+void debug_dma_sync_single_range_for_cpu(struct device *dev,
+                                        dma_addr_t dma_handle,
+                                        unsigned long offset, size_t size,
+                                        int direction)
+{
+       struct dma_debug_entry ref;
+
+       if (unlikely(global_disable))
+               return;
+
+       ref.type         = dma_debug_single;
+       ref.dev          = dev;
+       ref.dev_addr     = dma_handle;
+       ref.size         = offset + size;
+       ref.direction    = direction;
+       ref.sg_call_ents = 0;
+
+       check_sync(dev, &ref, true);
+}
+EXPORT_SYMBOL(debug_dma_sync_single_range_for_cpu);
+
+void debug_dma_sync_single_range_for_device(struct device *dev,
+                                           dma_addr_t dma_handle,
+                                           unsigned long offset,
+                                           size_t size, int direction)
+{
+       struct dma_debug_entry ref;
+
+       if (unlikely(global_disable))
+               return;
+
+       ref.type         = dma_debug_single;
+       ref.dev          = dev;
+       ref.dev_addr     = dma_handle;
+       ref.size         = offset + size;
+       ref.direction    = direction;
+       ref.sg_call_ents = 0;
+
+       check_sync(dev, &ref, false);
+}
+EXPORT_SYMBOL(debug_dma_sync_single_range_for_device);
+
+void debug_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
+                              int nelems, int direction)
+{
+       struct scatterlist *s;
+       int mapped_ents = 0, i;
+
+       if (unlikely(global_disable))
+               return;
+
+       for_each_sg(sg, s, nelems, i) {
+
+               struct dma_debug_entry ref = {
+                       .type           = dma_debug_sg,
+                       .dev            = dev,
+                       .paddr          = sg_phys(s),
+                       .dev_addr       = sg_dma_address(s),
+                       .size           = sg_dma_len(s),
+                       .direction      = direction,
+                       .sg_call_ents   = nelems,
+               };
+
+               if (!i)
+                       mapped_ents = get_nr_mapped_entries(dev, &ref);
+
+               if (i >= mapped_ents)
+                       break;
+
+               check_sync(dev, &ref, true);
+       }
+}
+EXPORT_SYMBOL(debug_dma_sync_sg_for_cpu);
+
+void debug_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
+                                 int nelems, int direction)
+{
+       struct scatterlist *s;
+       int mapped_ents = 0, i;
+
+       if (unlikely(global_disable))
+               return;
+
+       for_each_sg(sg, s, nelems, i) {
+
+               struct dma_debug_entry ref = {
+                       .type           = dma_debug_sg,
+                       .dev            = dev,
+                       .paddr          = sg_phys(s),
+                       .dev_addr       = sg_dma_address(s),
+                       .size           = sg_dma_len(s),
+                       .direction      = direction,
+                       .sg_call_ents   = nelems,
+               };
+               if (!i)
+                       mapped_ents = get_nr_mapped_entries(dev, &ref);
+
+               if (i >= mapped_ents)
+                       break;
+
+               check_sync(dev, &ref, false);
+       }
+}
+EXPORT_SYMBOL(debug_dma_sync_sg_for_device);
+
+static int __init dma_debug_driver_setup(char *str)
+{
+       int i;
+
+       for (i = 0; i < NAME_MAX_LEN - 1; ++i, ++str) {
+               current_driver_name[i] = *str;
+               if (*str == 0)
+                       break;
+       }
+
+       if (current_driver_name[0])
+               pr_info("DMA-API: enable driver filter for driver [%s]\n",
+                       current_driver_name);
+
+
+       return 1;
+}
+__setup("dma_debug_driver=", dma_debug_driver_setup);