DMA: PL330: Add dma api driver
[safe/jmp/linux-2.6] / drivers / dma / dmaengine.c
index 6df144a..9d31d5e 100644 (file)
@@ -58,6 +58,7 @@
 #include <linux/jiffies.h>
 #include <linux/rculist.h>
 #include <linux/idr.h>
+#include <linux/slab.h>
 
 static DEFINE_MUTEX(dma_list_mutex);
 static LIST_HEAD(dma_device_list);
@@ -284,7 +285,7 @@ struct dma_chan_tbl_ent {
 /**
  * channel_table - percpu lookup table for memory-to-memory offload providers
  */
-static struct dma_chan_tbl_ent *channel_table[DMA_TX_TYPE_END];
+static struct dma_chan_tbl_ent __percpu *channel_table[DMA_TX_TYPE_END];
 
 static int __init dma_channel_table_init(void)
 {
@@ -326,17 +327,7 @@ arch_initcall(dma_channel_table_init);
  */
 struct dma_chan *dma_find_channel(enum dma_transaction_type tx_type)
 {
-       struct dma_chan *chan;
-       int cpu;
-
-       WARN_ONCE(dmaengine_ref_count == 0,
-                 "client called %s without a reference", __func__);
-
-       cpu = get_cpu();
-       chan = per_cpu_ptr(channel_table[tx_type], cpu)->chan;
-       put_cpu();
-
-       return chan;
+       return this_cpu_read(channel_table[tx_type]->chan);
 }
 EXPORT_SYMBOL(dma_find_channel);
 
@@ -348,9 +339,6 @@ void dma_issue_pending_all(void)
        struct dma_device *device;
        struct dma_chan *chan;
 
-       WARN_ONCE(dmaengine_ref_count == 0,
-                 "client called %s without a reference", __func__);
-
        rcu_read_lock();
        list_for_each_entry_rcu(device, &dma_device_list, global_node) {
                if (dma_has_cap(DMA_PRIVATE, device->cap_mask))
@@ -513,6 +501,7 @@ struct dma_chan *__dma_request_channel(dma_cap_mask_t *mask, dma_filter_fn fn, v
                         * published in the general-purpose allocator
                         */
                        dma_cap_set(DMA_PRIVATE, device->cap_mask);
+                       device->privatecnt++;
                        err = dma_chan_get(chan);
 
                        if (err == -ENODEV) {
@@ -524,6 +513,8 @@ struct dma_chan *__dma_request_channel(dma_cap_mask_t *mask, dma_filter_fn fn, v
                                       dma_chan_name(chan), err);
                        else
                                break;
+                       if (--device->privatecnt == 0)
+                               dma_cap_clear(DMA_PRIVATE, device->cap_mask);
                        chan = NULL;
                }
        }
@@ -542,6 +533,9 @@ void dma_release_channel(struct dma_chan *chan)
        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_GPL(dma_release_channel);
@@ -606,6 +600,68 @@ void dmaengine_put(void)
 }
 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
  * @device: &dma_device
@@ -624,8 +680,12 @@ 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) &&
@@ -633,38 +693,42 @@ int dma_async_device_register(struct dma_device *device)
        BUG_ON(dma_has_cap(DMA_SLAVE, device->cap_mask) &&
                !device->device_prep_slave_sg);
        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);
 
+       /* note: this only matters in the
+        * CONFIG_ASYNC_TX_DISABLE_CHANNEL_SWITCH=y case
+        */
+       if (device_has_all_tx_types(device))
+               dma_cap_set(DMA_ASYNC_TX, device->cap_mask);
+
        idr_ref = kmalloc(sizeof(*idr_ref), GFP_KERNEL);
        if (!idr_ref)
                return -ENOMEM;
-       atomic_set(idr_ref, 0);
- 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)
+       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);
-                       continue;
+                       chan->local = NULL;
+                       goto err_out;
                }
 
                chan->chan_id = chancnt++;
@@ -681,6 +745,8 @@ int dma_async_device_register(struct dma_device *device)
                if (rc) {
                        free_percpu(chan->local);
                        chan->local = NULL;
+                       kfree(chan->dev);
+                       atomic_dec(idr_ref);
                        goto err_out;
                }
                chan->client_count = 0;
@@ -705,12 +771,23 @@ int dma_async_device_register(struct dma_device *device)
                        }
                }
        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);
 
        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;
@@ -748,6 +825,7 @@ void dma_async_device_unregister(struct dma_device *device)
                chan->dev->chan = NULL;
                mutex_unlock(&dma_list_mutex);
                device_unregister(&chan->dev->device);
+               free_percpu(chan->local);
        }
 }
 EXPORT_SYMBOL(dma_async_device_unregister);
@@ -772,12 +850,14 @@ 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);
@@ -788,10 +868,10 @@ 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;
 }
@@ -818,12 +898,12 @@ 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);
@@ -834,10 +914,10 @@ 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;
 }
@@ -866,13 +946,13 @@ 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);
@@ -883,10 +963,10 @@ 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;
 }
@@ -896,55 +976,32 @@ void dma_async_tx_descriptor_init(struct dma_async_tx_descriptor *tx,
        struct dma_chan *chan)
 {
        tx->chan = chan;
+       #ifndef CONFIG_ASYNC_TX_DISABLE_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
- *
- * This routine assumes that tx was obtained from a call to async_memcpy,
- * async_xor, async_memset, etc which ensures that tx is "in-flight" (prepped
- * and submitted).  Walking the parent chain is only meant to cover for DMA
- * drivers that do not implement the DMA_INTERRUPT capability and may race with
- * the driver's descriptor cleanup routine.
  */
 enum dma_status
 dma_wait_for_async_tx(struct dma_async_tx_descriptor *tx)
 {
-       enum dma_status status;
-       struct dma_async_tx_descriptor *iter;
-       struct dma_async_tx_descriptor *parent;
+       unsigned long dma_sync_wait_timeout = jiffies + msecs_to_jiffies(5000);
 
        if (!tx)
                return DMA_SUCCESS;
 
-       WARN_ONCE(tx->parent, "%s: speculatively walking dependency chain for"
-                 " %s\n", __func__, dma_chan_name(tx->chan));
-
-       /* poll through the dependency chain, return when tx is complete */
-       do {
-               iter = tx;
-
-               /* find the root of the unsubmitted dependency chain */
-               do {
-                       parent = iter->parent;
-                       if (!parent)
-                               break;
-                       else
-                               iter = parent;
-               } while (parent);
-
-               /* there is a small window for ->parent == NULL and
-                * ->cookie == -EBUSY
-                */
-               while (iter->cookie == -EBUSY)
-                       cpu_relax();
-
-               status = dma_sync_wait(iter->chan, iter->cookie);
-       } while (status == DMA_IN_PROGRESS || (iter != tx));
-
-       return status;
+       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);
 
@@ -954,7 +1011,7 @@ EXPORT_SYMBOL_GPL(dma_wait_for_async_tx);
  */
 void dma_run_dependencies(struct dma_async_tx_descriptor *tx)
 {
-       struct dma_async_tx_descriptor *dep = tx->next;
+       struct dma_async_tx_descriptor *dep = txd_next(tx);
        struct dma_async_tx_descriptor *dep_next;
        struct dma_chan *chan;
 
@@ -962,7 +1019,7 @@ void dma_run_dependencies(struct dma_async_tx_descriptor *tx)
                return;
 
        /* we'll submit tx->next now, so clear the link */
-       tx->next = NULL;
+       txd_clear_next(tx);
        chan = dep->chan;
 
        /* keep submitting up until a channel switch is detected
@@ -970,14 +1027,14 @@ void dma_run_dependencies(struct dma_async_tx_descriptor *tx)
         * processing the interrupt from async_tx_channel_switch
         */
        for (; dep; dep = dep_next) {
-               spin_lock_bh(&dep->lock);
-               dep->parent = NULL;
-               dep_next = dep->next;
+               txd_lock(dep);
+               txd_clear_parent(dep);
+               dep_next = txd_next(dep);
                if (dep_next && dep_next->chan == chan)
-                       dep->next = NULL; /* ->next will be submitted */
+                       txd_clear_next(dep); /* ->next will be submitted */
                else
                        dep_next = NULL; /* submit current dep and terminate */
-               spin_unlock_bh(&dep->lock);
+               txd_unlock(dep);
 
                dep->tx_submit(dep);
        }