This commit was manufactured by cvs2svn to create branch 'vserver'.
[linux-2.6.git] / drivers / dma / dmaengine.c
diff --git a/drivers/dma/dmaengine.c b/drivers/dma/dmaengine.c
new file mode 100644 (file)
index 0000000..1527804
--- /dev/null
@@ -0,0 +1,412 @@
+/*
+ * Copyright(c) 2004 - 2006 Intel Corporation. All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License as published by the Free
+ * Software Foundation; either version 2 of the License, or (at your option)
+ * any later version.
+ *
+ * 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.
+ *
+ * The full GNU General Public License is included in this distribution in the
+ * file called COPYING.
+ */
+
+/*
+ * This code implements the DMA subsystem. It provides a HW-neutral interface
+ * for other kernel code to use asynchronous memory copy capabilities,
+ * if present, and allows different HW DMA drivers to register as providing
+ * this capability.
+ *
+ * Due to the fact we are accelerating what is already a relatively fast
+ * operation, the code goes to great lengths to avoid additional overhead,
+ * such as locking.
+ *
+ * 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.
+ *
+ * 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 has a channels list, it's only modified under the client->lock
+ * and in an RCU callback, so it's safe to read under rcu_read_lock().
+ *
+ * Each device has a kref, which is initialized to 1 when the device is
+ * registered. A kref_put is done for each class_device registered.  When the
+ * class_device is released, the coresponding 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 coresponding kref_put
+ * happens. The device's release function does a completion, so
+ * unregister_device does a remove event, class_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 single reference is set when on an
+ * ADDED event, and removed with a REMOVE event.  Net DMA client takes an
+ * extra reference per outstanding transaction.  The relase function does a
+ * kref_put on the device. -ChrisL
+ */
+
+#include <linux/init.h>
+#include <linux/module.h>
+#include <linux/device.h>
+#include <linux/dmaengine.h>
+#include <linux/hardirq.h>
+#include <linux/spinlock.h>
+#include <linux/percpu.h>
+#include <linux/rcupdate.h>
+#include <linux/mutex.h>
+
+static DEFINE_MUTEX(dma_list_mutex);
+static LIST_HEAD(dma_device_list);
+static LIST_HEAD(dma_client_list);
+
+/* --- sysfs implementation --- */
+
+static ssize_t show_memcpy_count(struct class_device *cd, char *buf)
+{
+       struct dma_chan *chan = container_of(cd, struct dma_chan, class_dev);
+       unsigned long count = 0;
+       int i;
+
+       for_each_possible_cpu(i)
+               count += per_cpu_ptr(chan->local, i)->memcpy_count;
+
+       return sprintf(buf, "%lu\n", count);
+}
+
+static ssize_t show_bytes_transferred(struct class_device *cd, char *buf)
+{
+       struct dma_chan *chan = container_of(cd, struct dma_chan, class_dev);
+       unsigned long count = 0;
+       int i;
+
+       for_each_possible_cpu(i)
+               count += per_cpu_ptr(chan->local, i)->bytes_transferred;
+
+       return sprintf(buf, "%lu\n", count);
+}
+
+static ssize_t show_in_use(struct class_device *cd, char *buf)
+{
+       struct dma_chan *chan = container_of(cd, struct dma_chan, class_dev);
+
+       return sprintf(buf, "%d\n", (chan->client ? 1 : 0));
+}
+
+static struct class_device_attribute dma_class_attrs[] = {
+       __ATTR(memcpy_count, S_IRUGO, show_memcpy_count, NULL),
+       __ATTR(bytes_transferred, S_IRUGO, show_bytes_transferred, NULL),
+       __ATTR(in_use, S_IRUGO, show_in_use, NULL),
+       __ATTR_NULL
+};
+
+static void dma_async_device_cleanup(struct kref *kref);
+
+static void dma_class_dev_release(struct class_device *cd)
+{
+       struct dma_chan *chan = container_of(cd, struct dma_chan, class_dev);
+       kref_put(&chan->device->refcount, dma_async_device_cleanup);
+}
+
+static struct class dma_devclass = {
+       .name            = "dma",
+       .class_dev_attrs = dma_class_attrs,
+       .release = dma_class_dev_release,
+};
+
+/* --- client and device registration --- */
+
+/**
+ * dma_client_chan_alloc - try to allocate a channel to a client
+ * @client: &dma_client
+ *
+ * Called with dma_list_mutex held.
+ */
+static struct dma_chan *dma_client_chan_alloc(struct dma_client *client)
+{
+       struct dma_device *device;
+       struct dma_chan *chan;
+       unsigned long flags;
+       int desc;       /* allocated descriptor count */
+
+       /* Find a channel, any DMA engine will do */
+       list_for_each_entry(device, &dma_device_list, global_node) {
+               list_for_each_entry(chan, &device->channels, device_node) {
+                       if (chan->client)
+                               continue;
+
+                       desc = chan->device->device_alloc_chan_resources(chan);
+                       if (desc >= 0) {
+                               kref_get(&device->refcount);
+                               kref_init(&chan->refcount);
+                               chan->slow_ref = 0;
+                               INIT_RCU_HEAD(&chan->rcu);
+                               chan->client = client;
+                               spin_lock_irqsave(&client->lock, flags);
+                               list_add_tail_rcu(&chan->client_node,
+                                                 &client->channels);
+                               spin_unlock_irqrestore(&client->lock, flags);
+                               return chan;
+                       }
+               }
+       }
+
+       return NULL;
+}
+
+/**
+ * dma_chan_cleanup - release a DMA channel's resources
+ * @kref: kernel reference structure that contains the DMA channel device
+ */
+void dma_chan_cleanup(struct kref *kref)
+{
+       struct dma_chan *chan = container_of(kref, struct dma_chan, refcount);
+       chan->device->device_free_chan_resources(chan);
+       chan->client = NULL;
+       kref_put(&chan->device->refcount, dma_async_device_cleanup);
+}
+
+static void dma_chan_free_rcu(struct rcu_head *rcu)
+{
+       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);
+}
+
+static void dma_client_chan_free(struct dma_chan *chan)
+{
+       atomic_add(0x7FFFFFFF, &chan->refcount.refcount);
+       chan->slow_ref = 1;
+       call_rcu(&chan->rcu, dma_chan_free_rcu);
+}
+
+/**
+ * dma_chans_rebalance - reallocate channels to clients
+ *
+ * When the number of DMA channel in the system changes,
+ * channels need to be rebalanced among clients.
+ */
+static void dma_chans_rebalance(void)
+{
+       struct dma_client *client;
+       struct dma_chan *chan;
+       unsigned long flags;
+
+       mutex_lock(&dma_list_mutex);
+
+       list_for_each_entry(client, &dma_client_list, global_node) {
+               while (client->chans_desired > client->chan_count) {
+                       chan = dma_client_chan_alloc(client);
+                       if (!chan)
+                               break;
+                       client->chan_count++;
+                       client->event_callback(client,
+                                              chan,
+                                              DMA_RESOURCE_ADDED);
+               }
+               while (client->chans_desired < client->chan_count) {
+                       spin_lock_irqsave(&client->lock, flags);
+                       chan = list_entry(client->channels.next,
+                                         struct dma_chan,
+                                         client_node);
+                       list_del_rcu(&chan->client_node);
+                       spin_unlock_irqrestore(&client->lock, flags);
+                       client->chan_count--;
+                       client->event_callback(client,
+                                              chan,
+                                              DMA_RESOURCE_REMOVED);
+                       dma_client_chan_free(chan);
+               }
+       }
+
+       mutex_unlock(&dma_list_mutex);
+}
+
+/**
+ * dma_async_client_register - allocate and register a &dma_client
+ * @event_callback: callback for notification of channel addition/removal
+ */
+struct dma_client *dma_async_client_register(dma_event_callback event_callback)
+{
+       struct dma_client *client;
+
+       client = kzalloc(sizeof(*client), GFP_KERNEL);
+       if (!client)
+               return NULL;
+
+       INIT_LIST_HEAD(&client->channels);
+       spin_lock_init(&client->lock);
+       client->chans_desired = 0;
+       client->chan_count = 0;
+       client->event_callback = event_callback;
+
+       mutex_lock(&dma_list_mutex);
+       list_add_tail(&client->global_node, &dma_client_list);
+       mutex_unlock(&dma_list_mutex);
+
+       return client;
+}
+
+/**
+ * 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
+ */
+void dma_async_client_unregister(struct dma_client *client)
+{
+       struct dma_chan *chan;
+
+       if (!client)
+               return;
+
+       rcu_read_lock();
+       list_for_each_entry_rcu(chan, &client->channels, client_node)
+               dma_client_chan_free(chan);
+       rcu_read_unlock();
+
+       mutex_lock(&dma_list_mutex);
+       list_del(&client->global_node);
+       mutex_unlock(&dma_list_mutex);
+
+       kfree(client);
+       dma_chans_rebalance();
+}
+
+/**
+ * dma_async_client_chan_request - request DMA channels
+ * @client: &dma_client
+ * @number: count of DMA channels requested
+ *
+ * Clients call dma_async_client_chan_request() to specify how many
+ * DMA channels they need, 0 to free all currently allocated.
+ * The resulting allocations/frees are indicated to the client via the
+ * event callback.
+ */
+void dma_async_client_chan_request(struct dma_client *client,
+                       unsigned int number)
+{
+       client->chans_desired = number;
+       dma_chans_rebalance();
+}
+
+/**
+ * dma_async_device_register - registers DMA devices found
+ * @device: &dma_device
+ */
+int dma_async_device_register(struct dma_device *device)
+{
+       static int id;
+       int chancnt = 0;
+       struct dma_chan* chan;
+
+       if (!device)
+               return -ENODEV;
+
+       init_completion(&device->done);
+       kref_init(&device->refcount);
+       device->dev_id = id++;
+
+       /* represent channels in sysfs. Probably want devs too */
+       list_for_each_entry(chan, &device->channels, device_node) {
+               chan->local = alloc_percpu(typeof(*chan->local));
+               if (chan->local == NULL)
+                       continue;
+
+               chan->chan_id = chancnt++;
+               chan->class_dev.class = &dma_devclass;
+               chan->class_dev.dev = NULL;
+               snprintf(chan->class_dev.class_id, BUS_ID_SIZE, "dma%dchan%d",
+                        device->dev_id, chan->chan_id);
+
+               kref_get(&device->refcount);
+               class_device_register(&chan->class_dev);
+       }
+
+       mutex_lock(&dma_list_mutex);
+       list_add_tail(&device->global_node, &dma_device_list);
+       mutex_unlock(&dma_list_mutex);
+
+       dma_chans_rebalance();
+
+       return 0;
+}
+
+/**
+ * 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
+ * @device: &dma_device
+ */
+void dma_async_device_unregister(struct dma_device *device)
+{
+       struct dma_chan *chan;
+       unsigned long flags;
+
+       mutex_lock(&dma_list_mutex);
+       list_del(&device->global_node);
+       mutex_unlock(&dma_list_mutex);
+
+       list_for_each_entry(chan, &device->channels, device_node) {
+               if (chan->client) {
+                       spin_lock_irqsave(&chan->client->lock, flags);
+                       list_del(&chan->client_node);
+                       chan->client->chan_count--;
+                       spin_unlock_irqrestore(&chan->client->lock, flags);
+                       chan->client->event_callback(chan->client,
+                                                    chan,
+                                                    DMA_RESOURCE_REMOVED);
+                       dma_client_chan_free(chan);
+               }
+               class_device_unregister(&chan->class_dev);
+       }
+       dma_chans_rebalance();
+
+       kref_put(&device->refcount, dma_async_device_cleanup);
+       wait_for_completion(&device->done);
+}
+
+static int __init dma_bus_init(void)
+{
+       mutex_init(&dma_list_mutex);
+       return class_register(&dma_devclass);
+}
+
+subsys_initcall(dma_bus_init);
+
+EXPORT_SYMBOL(dma_async_client_register);
+EXPORT_SYMBOL(dma_async_client_unregister);
+EXPORT_SYMBOL(dma_async_client_chan_request);
+EXPORT_SYMBOL(dma_async_memcpy_buf_to_buf);
+EXPORT_SYMBOL(dma_async_memcpy_buf_to_pg);
+EXPORT_SYMBOL(dma_async_memcpy_pg_to_pg);
+EXPORT_SYMBOL(dma_async_memcpy_complete);
+EXPORT_SYMBOL(dma_async_memcpy_issue_pending);
+EXPORT_SYMBOL(dma_async_device_register);
+EXPORT_SYMBOL(dma_async_device_unregister);
+EXPORT_SYMBOL(dma_chan_cleanup);