Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/percpu
[safe/jmp/linux-2.6] / drivers / dma / dmaengine.c
index 92438e9..87399ca 100644 (file)
@@ -284,7 +284,7 @@ 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)
 {
@@ -326,14 +326,7 @@ arch_initcall(dma_channel_table_init);
  */
 struct dma_chan *dma_find_channel(enum dma_transaction_type tx_type)
 {
-       struct dma_chan *chan;
-       int cpu;
-
-       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);
 
@@ -608,6 +601,50 @@ 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;
@@ -644,8 +681,12 @@ 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) &&
@@ -661,6 +702,12 @@ int dma_async_device_register(struct dma_device *device)
        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;
@@ -779,6 +826,7 @@ 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);
@@ -803,12 +851,14 @@ 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);
@@ -819,10 +869,10 @@ 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;
 }
@@ -849,12 +899,12 @@ 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);
@@ -865,10 +915,10 @@ 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;
 }
@@ -897,13 +947,13 @@ 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);
@@ -914,10 +964,10 @@ 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;
 }
@@ -928,55 +978,29 @@ void dma_async_tx_descriptor_init(struct dma_async_tx_descriptor *tx,
 {
        tx->chan = chan;
        spin_lock_init(&tx->lock);
-       INIT_LIST_HEAD(&tx->tx_list);
 }
 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);