Merge branch 'mfd/wm8994' of git://git.kernel.org/pub/scm/linux/kernel/git/broonie...
[linux-2.6.git] / drivers / dma / dmaengine.c
index 6579965..b48967b 100644 (file)
  *
  * LOCKING:
  *
- * The subsystem keeps two global lists, dma_device_list and dma_client_list.
- * Both of these are protected by a mutex, dma_list_mutex.
+ * The subsystem keeps a global list of dma_device structs it is protected by a
+ * mutex, dma_list_mutex.
+ *
+ * A subsystem can get access to a channel by calling dmaengine_get() followed
+ * by dma_find_channel(), or if it has need for an exclusive channel it can call
+ * dma_request_channel().  Once a channel is allocated a reference is taken
+ * against its corresponding driver to disable removal.
  *
  * Each device has a channels list, which runs unlocked but is never modified
  * once the device is registered, it's just setup by the driver.
  *
- * Each client is responsible for keeping track of the channels it uses.  See
- * the definition of dma_event_callback in dmaengine.h.
- *
- * Each device has a kref, which is initialized to 1 when the device is
- * registered. A kref_get is done for each device registered.  When the
- * device is released, the corresponding kref_put is done in the release
- * method. Every time one of the device's channels is allocated to a client,
- * a kref_get occurs.  When the channel is freed, the corresponding kref_put
- * happens. The device's release function does a completion, so
- * unregister_device does a remove event, device_unregister, a kref_put
- * for the first reference, then waits on the completion for all other
- * references to finish.
- *
- * Each channel has an open-coded implementation of Rusty Russell's "bigref,"
- * with a kref and a per_cpu local_t.  A dma_chan_get is called when a client
- * signals that it wants to use a channel, and dma_chan_put is called when
- * a channel is removed or a client using it is unregistered.  A client can
- * take extra references per outstanding transaction, as is the case with
- * the NET DMA client.  The release function does a kref_put on the device.
- *     -ChrisL, DanW
+ * See Documentation/dmaengine.txt for more details
  */
 
+#include <linux/dma-mapping.h>
 #include <linux/init.h>
 #include <linux/module.h>
 #include <linux/mm.h>
 #include <linux/rcupdate.h>
 #include <linux/mutex.h>
 #include <linux/jiffies.h>
+#include <linux/rculist.h>
+#include <linux/idr.h>
+#include <linux/slab.h>
 
 static DEFINE_MUTEX(dma_list_mutex);
+static DEFINE_IDR(dma_idr);
 static LIST_HEAD(dma_device_list);
-static LIST_HEAD(dma_client_list);
+static long dmaengine_ref_count;
 
 /* --- sysfs implementation --- */
 
+/**
+ * dev_to_dma_chan - convert a device pointer to the its sysfs container object
+ * @dev - device node
+ *
+ * Must be called under dma_list_mutex
+ */
+static struct dma_chan *dev_to_dma_chan(struct device *dev)
+{
+       struct dma_chan_dev *chan_dev;
+
+       chan_dev = container_of(dev, typeof(*chan_dev), device);
+       return chan_dev->chan;
+}
+
 static ssize_t show_memcpy_count(struct device *dev, struct device_attribute *attr, char *buf)
 {
-       struct dma_chan *chan = to_dma_chan(dev);
+       struct dma_chan *chan;
        unsigned long count = 0;
        int i;
+       int err;
 
-       for_each_possible_cpu(i)
-               count += per_cpu_ptr(chan->local, i)->memcpy_count;
+       mutex_lock(&dma_list_mutex);
+       chan = dev_to_dma_chan(dev);
+       if (chan) {
+               for_each_possible_cpu(i)
+                       count += per_cpu_ptr(chan->local, i)->memcpy_count;
+               err = sprintf(buf, "%lu\n", count);
+       } else
+               err = -ENODEV;
+       mutex_unlock(&dma_list_mutex);
 
-       return sprintf(buf, "%lu\n", count);
+       return err;
 }
 
 static ssize_t show_bytes_transferred(struct device *dev, struct device_attribute *attr,
                                      char *buf)
 {
-       struct dma_chan *chan = to_dma_chan(dev);
+       struct dma_chan *chan;
        unsigned long count = 0;
        int i;
+       int err;
 
-       for_each_possible_cpu(i)
-               count += per_cpu_ptr(chan->local, i)->bytes_transferred;
+       mutex_lock(&dma_list_mutex);
+       chan = dev_to_dma_chan(dev);
+       if (chan) {
+               for_each_possible_cpu(i)
+                       count += per_cpu_ptr(chan->local, i)->bytes_transferred;
+               err = sprintf(buf, "%lu\n", count);
+       } else
+               err = -ENODEV;
+       mutex_unlock(&dma_list_mutex);
 
-       return sprintf(buf, "%lu\n", count);
+       return err;
 }
 
 static ssize_t show_in_use(struct device *dev, struct device_attribute *attr, char *buf)
 {
-       struct dma_chan *chan = to_dma_chan(dev);
-       int in_use = 0;
-
-       if (unlikely(chan->slow_ref) &&
-               atomic_read(&chan->refcount.refcount) > 1)
-               in_use = 1;
-       else {
-               if (local_read(&(per_cpu_ptr(chan->local,
-                       get_cpu())->refcount)) > 0)
-                       in_use = 1;
-               put_cpu();
-       }
+       struct dma_chan *chan;
+       int err;
 
-       return sprintf(buf, "%d\n", in_use);
+       mutex_lock(&dma_list_mutex);
+       chan = dev_to_dma_chan(dev);
+       if (chan)
+               err = sprintf(buf, "%d\n", chan->client_count);
+       else
+               err = -ENODEV;
+       mutex_unlock(&dma_list_mutex);
+
+       return err;
 }
 
 static struct device_attribute dma_attrs[] = {
@@ -127,76 +146,110 @@ static struct device_attribute dma_attrs[] = {
        __ATTR_NULL
 };
 
-static void dma_async_device_cleanup(struct kref *kref);
-
-static void dma_dev_release(struct device *dev)
+static void chan_dev_release(struct device *dev)
 {
-       struct dma_chan *chan = to_dma_chan(dev);
-       kref_put(&chan->device->refcount, dma_async_device_cleanup);
+       struct dma_chan_dev *chan_dev;
+
+       chan_dev = container_of(dev, typeof(*chan_dev), device);
+       if (atomic_dec_and_test(chan_dev->idr_ref)) {
+               mutex_lock(&dma_list_mutex);
+               idr_remove(&dma_idr, chan_dev->dev_id);
+               mutex_unlock(&dma_list_mutex);
+               kfree(chan_dev->idr_ref);
+       }
+       kfree(chan_dev);
 }
 
 static struct class dma_devclass = {
        .name           = "dma",
        .dev_attrs      = dma_attrs,
-       .dev_release    = dma_dev_release,
+       .dev_release    = chan_dev_release,
 };
 
 /* --- client and device registration --- */
 
-#define dma_chan_satisfies_mask(chan, mask) \
-       __dma_chan_satisfies_mask((chan), &(mask))
+#define dma_device_satisfies_mask(device, mask) \
+       __dma_device_satisfies_mask((device), &(mask))
 static int
-__dma_chan_satisfies_mask(struct dma_chan *chan, dma_cap_mask_t *want)
+__dma_device_satisfies_mask(struct dma_device *device, dma_cap_mask_t *want)
 {
        dma_cap_mask_t has;
 
-       bitmap_and(has.bits, want->bits, chan->device->cap_mask.bits,
+       bitmap_and(has.bits, want->bits, device->cap_mask.bits,
                DMA_TX_TYPE_END);
        return bitmap_equal(want->bits, has.bits, DMA_TX_TYPE_END);
 }
 
+static struct module *dma_chan_to_owner(struct dma_chan *chan)
+{
+       return chan->device->dev->driver->owner;
+}
+
 /**
- * dma_client_chan_alloc - try to allocate channels to a client
- * @client: &dma_client
+ * balance_ref_count - catch up the channel reference count
+ * @chan - channel to balance ->client_count versus dmaengine_ref_count
  *
- * Called with dma_list_mutex held.
+ * balance_ref_count must be called under dma_list_mutex
  */
-static void dma_client_chan_alloc(struct dma_client *client)
+static void balance_ref_count(struct dma_chan *chan)
 {
-       struct dma_device *device;
-       struct dma_chan *chan;
-       int desc;       /* allocated descriptor count */
-       enum dma_state_client ack;
+       struct module *owner = dma_chan_to_owner(chan);
 
-       /* Find a channel */
-       list_for_each_entry(device, &dma_device_list, global_node) {
-               /* Does the client require a specific DMA controller? */
-               if (client->slave && client->slave->dma_dev
-                               && client->slave->dma_dev != device->dev)
-                       continue;
+       while (chan->client_count < dmaengine_ref_count) {
+               __module_get(owner);
+               chan->client_count++;
+       }
+}
 
-               list_for_each_entry(chan, &device->channels, device_node) {
-                       if (!dma_chan_satisfies_mask(chan, client->cap_mask))
-                               continue;
+/**
+ * dma_chan_get - try to grab a dma channel's parent driver module
+ * @chan - channel to grab
+ *
+ * Must be called under dma_list_mutex
+ */
+static int dma_chan_get(struct dma_chan *chan)
+{
+       int err = -ENODEV;
+       struct module *owner = dma_chan_to_owner(chan);
+
+       if (chan->client_count) {
+               __module_get(owner);
+               err = 0;
+       } else if (try_module_get(owner))
+               err = 0;
+
+       if (err == 0)
+               chan->client_count++;
+
+       /* allocate upon first client reference */
+       if (chan->client_count == 1 && err == 0) {
+               int desc_cnt = chan->device->device_alloc_chan_resources(chan);
+
+               if (desc_cnt < 0) {
+                       err = desc_cnt;
+                       chan->client_count = 0;
+                       module_put(owner);
+               } else if (!dma_has_cap(DMA_PRIVATE, chan->device->cap_mask))
+                       balance_ref_count(chan);
+       }
 
-                       desc = chan->device->device_alloc_chan_resources(
-                                       chan, client);
-                       if (desc >= 0) {
-                               ack = client->event_callback(client,
-                                               chan,
-                                               DMA_RESOURCE_AVAILABLE);
+       return err;
+}
 
-                               /* we are done once this client rejects
-                                * an available resource
-                                */
-                               if (ack == DMA_ACK) {
-                                       dma_chan_get(chan);
-                                       chan->client_count++;
-                               } else if (ack == DMA_NAK)
-                                       return;
-                       }
-               }
-       }
+/**
+ * dma_chan_put - drop a reference to a dma channel's parent driver module
+ * @chan - channel to release
+ *
+ * Must be called under dma_list_mutex
+ */
+static void dma_chan_put(struct dma_chan *chan)
+{
+       if (!chan->client_count)
+               return; /* this channel failed alloc_chan_resources */
+       chan->client_count--;
+       module_put(dma_chan_to_owner(chan));
+       if (chan->client_count == 0)
+               chan->device->device_free_chan_resources(chan);
 }
 
 enum dma_status dma_sync_wait(struct dma_chan *chan, dma_cookie_t cookie)
@@ -218,138 +271,397 @@ enum dma_status dma_sync_wait(struct dma_chan *chan, dma_cookie_t cookie)
 EXPORT_SYMBOL(dma_sync_wait);
 
 /**
- * dma_chan_cleanup - release a DMA channel's resources
- * @kref: kernel reference structure that contains the DMA channel device
+ * dma_cap_mask_all - enable iteration over all operation types
+ */
+static dma_cap_mask_t dma_cap_mask_all;
+
+/**
+ * dma_chan_tbl_ent - tracks channel allocations per core/operation
+ * @chan - associated channel for this entry
  */
-void dma_chan_cleanup(struct kref *kref)
+struct dma_chan_tbl_ent {
+       struct dma_chan *chan;
+};
+
+/**
+ * channel_table - percpu lookup table for memory-to-memory offload providers
+ */
+static struct dma_chan_tbl_ent __percpu *channel_table[DMA_TX_TYPE_END];
+
+static int __init dma_channel_table_init(void)
 {
-       struct dma_chan *chan = container_of(kref, struct dma_chan, refcount);
-       chan->device->device_free_chan_resources(chan);
-       kref_put(&chan->device->refcount, dma_async_device_cleanup);
+       enum dma_transaction_type cap;
+       int err = 0;
+
+       bitmap_fill(dma_cap_mask_all.bits, DMA_TX_TYPE_END);
+
+       /* 'interrupt', 'private', and 'slave' are channel capabilities,
+        * but are not associated with an operation so they do not need
+        * an entry in the channel_table
+        */
+       clear_bit(DMA_INTERRUPT, dma_cap_mask_all.bits);
+       clear_bit(DMA_PRIVATE, dma_cap_mask_all.bits);
+       clear_bit(DMA_SLAVE, dma_cap_mask_all.bits);
+
+       for_each_dma_cap_mask(cap, dma_cap_mask_all) {
+               channel_table[cap] = alloc_percpu(struct dma_chan_tbl_ent);
+               if (!channel_table[cap]) {
+                       err = -ENOMEM;
+                       break;
+               }
+       }
+
+       if (err) {
+               pr_err("dmaengine: initialization failure\n");
+               for_each_dma_cap_mask(cap, dma_cap_mask_all)
+                       if (channel_table[cap])
+                               free_percpu(channel_table[cap]);
+       }
+
+       return err;
 }
-EXPORT_SYMBOL(dma_chan_cleanup);
+arch_initcall(dma_channel_table_init);
 
-static void dma_chan_free_rcu(struct rcu_head *rcu)
+/**
+ * dma_find_channel - find a channel to carry out the operation
+ * @tx_type: transaction type
+ */
+struct dma_chan *dma_find_channel(enum dma_transaction_type tx_type)
 {
-       struct dma_chan *chan = container_of(rcu, struct dma_chan, rcu);
-       int bias = 0x7FFFFFFF;
-       int i;
-       for_each_possible_cpu(i)
-               bias -= local_read(&per_cpu_ptr(chan->local, i)->refcount);
-       atomic_sub(bias, &chan->refcount.refcount);
-       kref_put(&chan->refcount, dma_chan_cleanup);
+       return this_cpu_read(channel_table[tx_type]->chan);
 }
+EXPORT_SYMBOL(dma_find_channel);
 
-static void dma_chan_release(struct dma_chan *chan)
+/**
+ * dma_issue_pending_all - flush all pending operations across all channels
+ */
+void dma_issue_pending_all(void)
 {
-       atomic_add(0x7FFFFFFF, &chan->refcount.refcount);
-       chan->slow_ref = 1;
-       call_rcu(&chan->rcu, dma_chan_free_rcu);
+       struct dma_device *device;
+       struct dma_chan *chan;
+
+       rcu_read_lock();
+       list_for_each_entry_rcu(device, &dma_device_list, global_node) {
+               if (dma_has_cap(DMA_PRIVATE, device->cap_mask))
+                       continue;
+               list_for_each_entry(chan, &device->channels, device_node)
+                       if (chan->client_count)
+                               device->device_issue_pending(chan);
+       }
+       rcu_read_unlock();
 }
+EXPORT_SYMBOL(dma_issue_pending_all);
 
 /**
- * dma_chans_notify_available - broadcast available channels to the clients
+ * nth_chan - returns the nth channel of the given capability
+ * @cap: capability to match
+ * @n: nth channel desired
+ *
+ * Defaults to returning the channel with the desired capability and the
+ * lowest reference count when 'n' cannot be satisfied.  Must be called
+ * under dma_list_mutex.
  */
-static void dma_clients_notify_available(void)
+static struct dma_chan *nth_chan(enum dma_transaction_type cap, int n)
 {
-       struct dma_client *client;
+       struct dma_device *device;
+       struct dma_chan *chan;
+       struct dma_chan *ret = NULL;
+       struct dma_chan *min = NULL;
 
-       mutex_lock(&dma_list_mutex);
+       list_for_each_entry(device, &dma_device_list, global_node) {
+               if (!dma_has_cap(cap, device->cap_mask) ||
+                   dma_has_cap(DMA_PRIVATE, device->cap_mask))
+                       continue;
+               list_for_each_entry(chan, &device->channels, device_node) {
+                       if (!chan->client_count)
+                               continue;
+                       if (!min)
+                               min = chan;
+                       else if (chan->table_count < min->table_count)
+                               min = chan;
+
+                       if (n-- == 0) {
+                               ret = chan;
+                               break; /* done */
+                       }
+               }
+               if (ret)
+                       break; /* done */
+       }
 
-       list_for_each_entry(client, &dma_client_list, global_node)
-               dma_client_chan_alloc(client);
+       if (!ret)
+               ret = min;
 
-       mutex_unlock(&dma_list_mutex);
+       if (ret)
+               ret->table_count++;
+
+       return ret;
 }
 
 /**
- * dma_chans_notify_available - tell the clients that a channel is going away
- * @chan: channel on its way out
+ * dma_channel_rebalance - redistribute the available channels
+ *
+ * Optimize for cpu isolation (each cpu gets a dedicated channel for an
+ * operation type) in the SMP case,  and operation isolation (avoid
+ * multi-tasking channels) in the non-SMP case.  Must be called under
+ * dma_list_mutex.
  */
-static void dma_clients_notify_removed(struct dma_chan *chan)
+static void dma_channel_rebalance(void)
 {
-       struct dma_client *client;
-       enum dma_state_client ack;
+       struct dma_chan *chan;
+       struct dma_device *device;
+       int cpu;
+       int cap;
+       int n;
 
-       mutex_lock(&dma_list_mutex);
+       /* undo the last distribution */
+       for_each_dma_cap_mask(cap, dma_cap_mask_all)
+               for_each_possible_cpu(cpu)
+                       per_cpu_ptr(channel_table[cap], cpu)->chan = NULL;
 
-       list_for_each_entry(client, &dma_client_list, global_node) {
-               ack = client->event_callback(client, chan,
-                               DMA_RESOURCE_REMOVED);
+       list_for_each_entry(device, &dma_device_list, global_node) {
+               if (dma_has_cap(DMA_PRIVATE, device->cap_mask))
+                       continue;
+               list_for_each_entry(chan, &device->channels, device_node)
+                       chan->table_count = 0;
+       }
 
-               /* client was holding resources for this channel so
-                * free it
-                */
-               if (ack == DMA_ACK) {
-                       dma_chan_put(chan);
-                       chan->client_count--;
+       /* don't populate the channel_table if no clients are available */
+       if (!dmaengine_ref_count)
+               return;
+
+       /* redistribute available channels */
+       n = 0;
+       for_each_dma_cap_mask(cap, dma_cap_mask_all)
+               for_each_online_cpu(cpu) {
+                       if (num_possible_cpus() > 1)
+                               chan = nth_chan(cap, n++);
+                       else
+                               chan = nth_chan(cap, -1);
+
+                       per_cpu_ptr(channel_table[cap], cpu)->chan = chan;
                }
+}
+
+static struct dma_chan *private_candidate(dma_cap_mask_t *mask, struct dma_device *dev,
+                                         dma_filter_fn fn, void *fn_param)
+{
+       struct dma_chan *chan;
+
+       if (!__dma_device_satisfies_mask(dev, mask)) {
+               pr_debug("%s: wrong capabilities\n", __func__);
+               return NULL;
        }
+       /* devices with multiple channels need special handling as we need to
+        * ensure that all channels are either private or public.
+        */
+       if (dev->chancnt > 1 && !dma_has_cap(DMA_PRIVATE, dev->cap_mask))
+               list_for_each_entry(chan, &dev->channels, device_node) {
+                       /* some channels are already publicly allocated */
+                       if (chan->client_count)
+                               return NULL;
+               }
 
-       mutex_unlock(&dma_list_mutex);
+       list_for_each_entry(chan, &dev->channels, device_node) {
+               if (chan->client_count) {
+                       pr_debug("%s: %s busy\n",
+                                __func__, dma_chan_name(chan));
+                       continue;
+               }
+               if (fn && !fn(chan, fn_param)) {
+                       pr_debug("%s: %s filter said false\n",
+                                __func__, dma_chan_name(chan));
+                       continue;
+               }
+               return chan;
+       }
+
+       return NULL;
 }
 
 /**
- * dma_async_client_register - register a &dma_client
- * @client: ptr to a client structure with valid 'event_callback' and 'cap_mask'
+ * dma_request_channel - try to allocate an exclusive channel
+ * @mask: capabilities that the channel must satisfy
+ * @fn: optional callback to disposition available channels
+ * @fn_param: opaque parameter to pass to dma_filter_fn
  */
-void dma_async_client_register(struct dma_client *client)
+struct dma_chan *__dma_request_channel(dma_cap_mask_t *mask, dma_filter_fn fn, void *fn_param)
 {
-       /* validate client data */
-       BUG_ON(dma_has_cap(DMA_SLAVE, client->cap_mask) &&
-               !client->slave);
+       struct dma_device *device, *_d;
+       struct dma_chan *chan = NULL;
+       int err;
+
+       /* Find a channel */
+       mutex_lock(&dma_list_mutex);
+       list_for_each_entry_safe(device, _d, &dma_device_list, global_node) {
+               chan = private_candidate(mask, device, fn, fn_param);
+               if (chan) {
+                       /* Found a suitable channel, try to grab, prep, and
+                        * return it.  We first set DMA_PRIVATE to disable
+                        * balance_ref_count as this channel will not be
+                        * published in the general-purpose allocator
+                        */
+                       dma_cap_set(DMA_PRIVATE, device->cap_mask);
+                       device->privatecnt++;
+                       err = dma_chan_get(chan);
+
+                       if (err == -ENODEV) {
+                               pr_debug("%s: %s module removed\n", __func__,
+                                        dma_chan_name(chan));
+                               list_del_rcu(&device->global_node);
+                       } else if (err)
+                               pr_debug("dmaengine: failed to get %s: (%d)\n",
+                                        dma_chan_name(chan), err);
+                       else
+                               break;
+                       if (--device->privatecnt == 0)
+                               dma_cap_clear(DMA_PRIVATE, device->cap_mask);
+                       chan = NULL;
+               }
+       }
+       mutex_unlock(&dma_list_mutex);
 
+       pr_debug("%s: %s (%s)\n", __func__, chan ? "success" : "fail",
+                chan ? dma_chan_name(chan) : NULL);
+
+       return chan;
+}
+EXPORT_SYMBOL_GPL(__dma_request_channel);
+
+void dma_release_channel(struct dma_chan *chan)
+{
        mutex_lock(&dma_list_mutex);
-       list_add_tail(&client->global_node, &dma_client_list);
+       WARN_ONCE(chan->client_count != 1,
+                 "chan reference count %d != 1\n", chan->client_count);
+       dma_chan_put(chan);
+       /* drop PRIVATE cap enabled by __dma_request_channel() */
+       if (--chan->device->privatecnt == 0)
+               dma_cap_clear(DMA_PRIVATE, chan->device->cap_mask);
        mutex_unlock(&dma_list_mutex);
 }
-EXPORT_SYMBOL(dma_async_client_register);
+EXPORT_SYMBOL_GPL(dma_release_channel);
 
 /**
- * dma_async_client_unregister - unregister a client and free the &dma_client
- * @client: &dma_client to free
- *
- * Force frees any allocated DMA channels, frees the &dma_client memory
+ * dmaengine_get - register interest in dma_channels
  */
-void dma_async_client_unregister(struct dma_client *client)
+void dmaengine_get(void)
 {
-       struct dma_device *device;
+       struct dma_device *device, *_d;
        struct dma_chan *chan;
-       enum dma_state_client ack;
-
-       if (!client)
-               return;
+       int err;
 
        mutex_lock(&dma_list_mutex);
-       /* free all channels the client is holding */
-       list_for_each_entry(device, &dma_device_list, global_node)
-               list_for_each_entry(chan, &device->channels, device_node) {
-                       ack = client->event_callback(client, chan,
-                               DMA_RESOURCE_REMOVED);
+       dmaengine_ref_count++;
 
-                       if (ack == DMA_ACK) {
-                               dma_chan_put(chan);
-                               chan->client_count--;
-                       }
+       /* try to grab channels */
+       list_for_each_entry_safe(device, _d, &dma_device_list, global_node) {
+               if (dma_has_cap(DMA_PRIVATE, device->cap_mask))
+                       continue;
+               list_for_each_entry(chan, &device->channels, device_node) {
+                       err = dma_chan_get(chan);
+                       if (err == -ENODEV) {
+                               /* module removed before we could use it */
+                               list_del_rcu(&device->global_node);
+                               break;
+                       } else if (err)
+                               pr_err("dmaengine: failed to get %s: (%d)\n",
+                                      dma_chan_name(chan), err);
                }
+       }
 
-       list_del(&client->global_node);
+       /* if this is the first reference and there were channels
+        * waiting we need to rebalance to get those channels
+        * incorporated into the channel table
+        */
+       if (dmaengine_ref_count == 1)
+               dma_channel_rebalance();
        mutex_unlock(&dma_list_mutex);
 }
-EXPORT_SYMBOL(dma_async_client_unregister);
+EXPORT_SYMBOL(dmaengine_get);
 
 /**
- * dma_async_client_chan_request - send all available channels to the
- * client that satisfy the capability mask
- * @client - requester
+ * dmaengine_put - let dma drivers be removed when ref_count == 0
  */
-void dma_async_client_chan_request(struct dma_client *client)
+void dmaengine_put(void)
 {
+       struct dma_device *device;
+       struct dma_chan *chan;
+
        mutex_lock(&dma_list_mutex);
-       dma_client_chan_alloc(client);
+       dmaengine_ref_count--;
+       BUG_ON(dmaengine_ref_count < 0);
+       /* drop channel references */
+       list_for_each_entry(device, &dma_device_list, global_node) {
+               if (dma_has_cap(DMA_PRIVATE, device->cap_mask))
+                       continue;
+               list_for_each_entry(chan, &device->channels, device_node)
+                       dma_chan_put(chan);
+       }
        mutex_unlock(&dma_list_mutex);
 }
-EXPORT_SYMBOL(dma_async_client_chan_request);
+EXPORT_SYMBOL(dmaengine_put);
+
+static bool device_has_all_tx_types(struct dma_device *device)
+{
+       /* A device that satisfies this test has channels that will never cause
+        * an async_tx channel switch event as all possible operation types can
+        * be handled.
+        */
+       #ifdef CONFIG_ASYNC_TX_DMA
+       if (!dma_has_cap(DMA_INTERRUPT, device->cap_mask))
+               return false;
+       #endif
+
+       #if defined(CONFIG_ASYNC_MEMCPY) || defined(CONFIG_ASYNC_MEMCPY_MODULE)
+       if (!dma_has_cap(DMA_MEMCPY, device->cap_mask))
+               return false;
+       #endif
+
+       #if defined(CONFIG_ASYNC_MEMSET) || defined(CONFIG_ASYNC_MEMSET_MODULE)
+       if (!dma_has_cap(DMA_MEMSET, device->cap_mask))
+               return false;
+       #endif
+
+       #if defined(CONFIG_ASYNC_XOR) || defined(CONFIG_ASYNC_XOR_MODULE)
+       if (!dma_has_cap(DMA_XOR, device->cap_mask))
+               return false;
+
+       #ifndef CONFIG_ASYNC_TX_DISABLE_XOR_VAL_DMA
+       if (!dma_has_cap(DMA_XOR_VAL, device->cap_mask))
+               return false;
+       #endif
+       #endif
+
+       #if defined(CONFIG_ASYNC_PQ) || defined(CONFIG_ASYNC_PQ_MODULE)
+       if (!dma_has_cap(DMA_PQ, device->cap_mask))
+               return false;
+
+       #ifndef CONFIG_ASYNC_TX_DISABLE_PQ_VAL_DMA
+       if (!dma_has_cap(DMA_PQ_VAL, device->cap_mask))
+               return false;
+       #endif
+       #endif
+
+       return true;
+}
+
+static int get_dma_id(struct dma_device *device)
+{
+       int rc;
+
+ idr_retry:
+       if (!idr_pre_get(&dma_idr, GFP_KERNEL))
+               return -ENOMEM;
+       mutex_lock(&dma_list_mutex);
+       rc = idr_get_new(&dma_idr, NULL, &device->dev_id);
+       mutex_unlock(&dma_list_mutex);
+       if (rc == -EAGAIN)
+               goto idr_retry;
+       else if (rc != 0)
+               return rc;
+
+       return 0;
+}
 
 /**
  * dma_async_device_register - registers DMA devices found
@@ -357,9 +669,9 @@ EXPORT_SYMBOL(dma_async_client_chan_request);
  */
 int dma_async_device_register(struct dma_device *device)
 {
-       static int id;
        int chancnt = 0, rc;
        struct dma_chan* chan;
+       atomic_t *idr_ref;
 
        if (!device)
                return -ENODEV;
@@ -369,74 +681,125 @@ int dma_async_device_register(struct dma_device *device)
                !device->device_prep_dma_memcpy);
        BUG_ON(dma_has_cap(DMA_XOR, device->cap_mask) &&
                !device->device_prep_dma_xor);
-       BUG_ON(dma_has_cap(DMA_ZERO_SUM, device->cap_mask) &&
-               !device->device_prep_dma_zero_sum);
+       BUG_ON(dma_has_cap(DMA_XOR_VAL, device->cap_mask) &&
+               !device->device_prep_dma_xor_val);
+       BUG_ON(dma_has_cap(DMA_PQ, device->cap_mask) &&
+               !device->device_prep_dma_pq);
+       BUG_ON(dma_has_cap(DMA_PQ_VAL, device->cap_mask) &&
+               !device->device_prep_dma_pq_val);
        BUG_ON(dma_has_cap(DMA_MEMSET, device->cap_mask) &&
                !device->device_prep_dma_memset);
        BUG_ON(dma_has_cap(DMA_INTERRUPT, device->cap_mask) &&
                !device->device_prep_dma_interrupt);
+       BUG_ON(dma_has_cap(DMA_SG, device->cap_mask) &&
+               !device->device_prep_dma_sg);
        BUG_ON(dma_has_cap(DMA_SLAVE, device->cap_mask) &&
                !device->device_prep_slave_sg);
+       BUG_ON(dma_has_cap(DMA_CYCLIC, device->cap_mask) &&
+               !device->device_prep_dma_cyclic);
        BUG_ON(dma_has_cap(DMA_SLAVE, device->cap_mask) &&
-               !device->device_terminate_all);
+               !device->device_control);
 
        BUG_ON(!device->device_alloc_chan_resources);
        BUG_ON(!device->device_free_chan_resources);
-       BUG_ON(!device->device_is_tx_complete);
+       BUG_ON(!device->device_tx_status);
        BUG_ON(!device->device_issue_pending);
        BUG_ON(!device->dev);
 
-       init_completion(&device->done);
-       kref_init(&device->refcount);
+       /* note: this only matters in the
+        * CONFIG_ASYNC_TX_ENABLE_CHANNEL_SWITCH=n case
+        */
+       if (device_has_all_tx_types(device))
+               dma_cap_set(DMA_ASYNC_TX, device->cap_mask);
 
-       mutex_lock(&dma_list_mutex);
-       device->dev_id = id++;
-       mutex_unlock(&dma_list_mutex);
+       idr_ref = kmalloc(sizeof(*idr_ref), GFP_KERNEL);
+       if (!idr_ref)
+               return -ENOMEM;
+       rc = get_dma_id(device);
+       if (rc != 0) {
+               kfree(idr_ref);
+               return rc;
+       }
+
+       atomic_set(idr_ref, 0);
 
        /* represent channels in sysfs. Probably want devs too */
        list_for_each_entry(chan, &device->channels, device_node) {
+               rc = -ENOMEM;
                chan->local = alloc_percpu(typeof(*chan->local));
                if (chan->local == NULL)
-                       continue;
+                       goto err_out;
+               chan->dev = kzalloc(sizeof(*chan->dev), GFP_KERNEL);
+               if (chan->dev == NULL) {
+                       free_percpu(chan->local);
+                       chan->local = NULL;
+                       goto err_out;
+               }
 
                chan->chan_id = chancnt++;
-               chan->dev.class = &dma_devclass;
-               chan->dev.parent = device->dev;
-               dev_set_name(&chan->dev, "dma%dchan%d",
+               chan->dev->device.class = &dma_devclass;
+               chan->dev->device.parent = device->dev;
+               chan->dev->chan = chan;
+               chan->dev->idr_ref = idr_ref;
+               chan->dev->dev_id = device->dev_id;
+               atomic_inc(idr_ref);
+               dev_set_name(&chan->dev->device, "dma%dchan%d",
                             device->dev_id, chan->chan_id);
 
-               rc = device_register(&chan->dev);
+               rc = device_register(&chan->dev->device);
                if (rc) {
-                       chancnt--;
                        free_percpu(chan->local);
                        chan->local = NULL;
+                       kfree(chan->dev);
+                       atomic_dec(idr_ref);
                        goto err_out;
                }
-
-               /* One for the channel, one of the class device */
-               kref_get(&device->refcount);
-               kref_get(&device->refcount);
-               kref_init(&chan->refcount);
                chan->client_count = 0;
-               chan->slow_ref = 0;
-               INIT_RCU_HEAD(&chan->rcu);
        }
+       device->chancnt = chancnt;
 
        mutex_lock(&dma_list_mutex);
-       list_add_tail(&device->global_node, &dma_device_list);
+       /* take references on public channels */
+       if (dmaengine_ref_count && !dma_has_cap(DMA_PRIVATE, device->cap_mask))
+               list_for_each_entry(chan, &device->channels, device_node) {
+                       /* if clients are already waiting for channels we need
+                        * to take references on their behalf
+                        */
+                       if (dma_chan_get(chan) == -ENODEV) {
+                               /* note we can only get here for the first
+                                * channel as the remaining channels are
+                                * guaranteed to get a reference
+                                */
+                               rc = -ENODEV;
+                               mutex_unlock(&dma_list_mutex);
+                               goto err_out;
+                       }
+               }
+       list_add_tail_rcu(&device->global_node, &dma_device_list);
+       if (dma_has_cap(DMA_PRIVATE, device->cap_mask))
+               device->privatecnt++;   /* Always private */
+       dma_channel_rebalance();
        mutex_unlock(&dma_list_mutex);
 
-       dma_clients_notify_available();
-
        return 0;
 
 err_out:
+       /* if we never registered a channel just release the idr */
+       if (atomic_read(idr_ref) == 0) {
+               mutex_lock(&dma_list_mutex);
+               idr_remove(&dma_idr, device->dev_id);
+               mutex_unlock(&dma_list_mutex);
+               kfree(idr_ref);
+               return rc;
+       }
+
        list_for_each_entry(chan, &device->channels, device_node) {
                if (chan->local == NULL)
                        continue;
-               kref_put(&device->refcount, dma_async_device_cleanup);
-               device_unregister(&chan->dev);
-               chancnt--;
+               mutex_lock(&dma_list_mutex);
+               chan->dev->chan = NULL;
+               mutex_unlock(&dma_list_mutex);
+               device_unregister(&chan->dev->device);
                free_percpu(chan->local);
        }
        return rc;
@@ -444,37 +807,31 @@ err_out:
 EXPORT_SYMBOL(dma_async_device_register);
 
 /**
- * dma_async_device_cleanup - function called when all references are released
- * @kref: kernel reference object
- */
-static void dma_async_device_cleanup(struct kref *kref)
-{
-       struct dma_device *device;
-
-       device = container_of(kref, struct dma_device, refcount);
-       complete(&device->done);
-}
-
-/**
- * dma_async_device_unregister - unregisters DMA devices
+ * dma_async_device_unregister - unregister a DMA device
  * @device: &dma_device
+ *
+ * This routine is called by dma driver exit routines, dmaengine holds module
+ * references to prevent it being called while channels are in use.
  */
 void dma_async_device_unregister(struct dma_device *device)
 {
        struct dma_chan *chan;
 
        mutex_lock(&dma_list_mutex);
-       list_del(&device->global_node);
+       list_del_rcu(&device->global_node);
+       dma_channel_rebalance();
        mutex_unlock(&dma_list_mutex);
 
        list_for_each_entry(chan, &device->channels, device_node) {
-               dma_clients_notify_removed(chan);
-               device_unregister(&chan->dev);
-               dma_chan_release(chan);
+               WARN_ONCE(chan->client_count,
+                         "%s called while %d clients hold a reference\n",
+                         __func__, chan->client_count);
+               mutex_lock(&dma_list_mutex);
+               chan->dev->chan = NULL;
+               mutex_unlock(&dma_list_mutex);
+               device_unregister(&chan->dev->device);
+               free_percpu(chan->local);
        }
-
-       kref_put(&device->refcount, dma_async_device_cleanup);
-       wait_for_completion(&device->done);
 }
 EXPORT_SYMBOL(dma_async_device_unregister);
 
@@ -498,12 +855,14 @@ dma_async_memcpy_buf_to_buf(struct dma_chan *chan, void *dest,
        struct dma_async_tx_descriptor *tx;
        dma_addr_t dma_dest, dma_src;
        dma_cookie_t cookie;
-       int cpu;
+       unsigned long flags;
 
        dma_src = dma_map_single(dev->dev, src, len, DMA_TO_DEVICE);
        dma_dest = dma_map_single(dev->dev, dest, len, DMA_FROM_DEVICE);
-       tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len,
-                                        DMA_CTRL_ACK);
+       flags = DMA_CTRL_ACK |
+               DMA_COMPL_SRC_UNMAP_SINGLE |
+               DMA_COMPL_DEST_UNMAP_SINGLE;
+       tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len, flags);
 
        if (!tx) {
                dma_unmap_single(dev->dev, dma_src, len, DMA_TO_DEVICE);
@@ -514,10 +873,10 @@ dma_async_memcpy_buf_to_buf(struct dma_chan *chan, void *dest,
        tx->callback = NULL;
        cookie = tx->tx_submit(tx);
 
-       cpu = get_cpu();
-       per_cpu_ptr(chan->local, cpu)->bytes_transferred += len;
-       per_cpu_ptr(chan->local, cpu)->memcpy_count++;
-       put_cpu();
+       preempt_disable();
+       __this_cpu_add(chan->local->bytes_transferred, len);
+       __this_cpu_inc(chan->local->memcpy_count);
+       preempt_enable();
 
        return cookie;
 }
@@ -544,12 +903,12 @@ dma_async_memcpy_buf_to_pg(struct dma_chan *chan, struct page *page,
        struct dma_async_tx_descriptor *tx;
        dma_addr_t dma_dest, dma_src;
        dma_cookie_t cookie;
-       int cpu;
+       unsigned long flags;
 
        dma_src = dma_map_single(dev->dev, kdata, len, DMA_TO_DEVICE);
        dma_dest = dma_map_page(dev->dev, page, offset, len, DMA_FROM_DEVICE);
-       tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len,
-                                        DMA_CTRL_ACK);
+       flags = DMA_CTRL_ACK | DMA_COMPL_SRC_UNMAP_SINGLE;
+       tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len, flags);
 
        if (!tx) {
                dma_unmap_single(dev->dev, dma_src, len, DMA_TO_DEVICE);
@@ -560,10 +919,10 @@ dma_async_memcpy_buf_to_pg(struct dma_chan *chan, struct page *page,
        tx->callback = NULL;
        cookie = tx->tx_submit(tx);
 
-       cpu = get_cpu();
-       per_cpu_ptr(chan->local, cpu)->bytes_transferred += len;
-       per_cpu_ptr(chan->local, cpu)->memcpy_count++;
-       put_cpu();
+       preempt_disable();
+       __this_cpu_add(chan->local->bytes_transferred, len);
+       __this_cpu_inc(chan->local->memcpy_count);
+       preempt_enable();
 
        return cookie;
 }
@@ -592,13 +951,13 @@ dma_async_memcpy_pg_to_pg(struct dma_chan *chan, struct page *dest_pg,
        struct dma_async_tx_descriptor *tx;
        dma_addr_t dma_dest, dma_src;
        dma_cookie_t cookie;
-       int cpu;
+       unsigned long flags;
 
        dma_src = dma_map_page(dev->dev, src_pg, src_off, len, DMA_TO_DEVICE);
        dma_dest = dma_map_page(dev->dev, dest_pg, dest_off, len,
                                DMA_FROM_DEVICE);
-       tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len,
-                                        DMA_CTRL_ACK);
+       flags = DMA_CTRL_ACK;
+       tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len, flags);
 
        if (!tx) {
                dma_unmap_page(dev->dev, dma_src, len, DMA_TO_DEVICE);
@@ -609,10 +968,10 @@ dma_async_memcpy_pg_to_pg(struct dma_chan *chan, struct page *dest_pg,
        tx->callback = NULL;
        cookie = tx->tx_submit(tx);
 
-       cpu = get_cpu();
-       per_cpu_ptr(chan->local, cpu)->bytes_transferred += len;
-       per_cpu_ptr(chan->local, cpu)->memcpy_count++;
-       put_cpu();
+       preempt_disable();
+       __this_cpu_add(chan->local->bytes_transferred, len);
+       __this_cpu_inc(chan->local->memcpy_count);
+       preempt_enable();
 
        return cookie;
 }
@@ -622,14 +981,77 @@ void dma_async_tx_descriptor_init(struct dma_async_tx_descriptor *tx,
        struct dma_chan *chan)
 {
        tx->chan = chan;
+       #ifdef CONFIG_ASYNC_TX_ENABLE_CHANNEL_SWITCH
        spin_lock_init(&tx->lock);
+       #endif
 }
 EXPORT_SYMBOL(dma_async_tx_descriptor_init);
 
+/* dma_wait_for_async_tx - spin wait for a transaction to complete
+ * @tx: in-flight transaction to wait on
+ */
+enum dma_status
+dma_wait_for_async_tx(struct dma_async_tx_descriptor *tx)
+{
+       unsigned long dma_sync_wait_timeout = jiffies + msecs_to_jiffies(5000);
+
+       if (!tx)
+               return DMA_SUCCESS;
+
+       while (tx->cookie == -EBUSY) {
+               if (time_after_eq(jiffies, dma_sync_wait_timeout)) {
+                       pr_err("%s timeout waiting for descriptor submission\n",
+                               __func__);
+                       return DMA_ERROR;
+               }
+               cpu_relax();
+       }
+       return dma_sync_wait(tx->chan, tx->cookie);
+}
+EXPORT_SYMBOL_GPL(dma_wait_for_async_tx);
+
+/* dma_run_dependencies - helper routine for dma drivers to process
+ *     (start) dependent operations on their target channel
+ * @tx: transaction with dependencies
+ */
+void dma_run_dependencies(struct dma_async_tx_descriptor *tx)
+{
+       struct dma_async_tx_descriptor *dep = txd_next(tx);
+       struct dma_async_tx_descriptor *dep_next;
+       struct dma_chan *chan;
+
+       if (!dep)
+               return;
+
+       /* we'll submit tx->next now, so clear the link */
+       txd_clear_next(tx);
+       chan = dep->chan;
+
+       /* keep submitting up until a channel switch is detected
+        * in that case we will be called again as a result of
+        * processing the interrupt from async_tx_channel_switch
+        */
+       for (; dep; dep = dep_next) {
+               txd_lock(dep);
+               txd_clear_parent(dep);
+               dep_next = txd_next(dep);
+               if (dep_next && dep_next->chan == chan)
+                       txd_clear_next(dep); /* ->next will be submitted */
+               else
+                       dep_next = NULL; /* submit current dep and terminate */
+               txd_unlock(dep);
+
+               dep->tx_submit(dep);
+       }
+
+       chan->device->device_issue_pending(chan);
+}
+EXPORT_SYMBOL_GPL(dma_run_dependencies);
+
 static int __init dma_bus_init(void)
 {
-       mutex_init(&dma_list_mutex);
        return class_register(&dma_devclass);
 }
-subsys_initcall(dma_bus_init);
+arch_initcall(dma_bus_init);
+