]> pilppa.com Git - linux-2.6-omap-h63xx.git/commitdiff
dmaengine: introduce dma_request_channel and private channels
authorDan Williams <dan.j.williams@intel.com>
Tue, 6 Jan 2009 18:38:15 +0000 (11:38 -0700)
committerDan Williams <dan.j.williams@intel.com>
Tue, 6 Jan 2009 18:38:15 +0000 (11:38 -0700)
This interface is primarily for device-to-memory clients which need to
search for dma channels with platform-specific characteristics.  The
prototype is:

struct dma_chan *dma_request_channel(dma_cap_mask_t mask,
                                     dma_filter_fn filter_fn,
                                     void *filter_param);

When the optional 'filter_fn' parameter is set to NULL
dma_request_channel simply returns the first channel that satisfies the
capability mask.  Otherwise, when the mask parameter is insufficient for
specifying the necessary channel, the filter_fn routine can be used to
disposition the available channels in the system. The filter_fn routine
is called once for each free channel in the system.  Upon seeing a
suitable channel filter_fn returns DMA_ACK which flags that channel to
be the return value from dma_request_channel.  A channel allocated via
this interface is exclusive to the caller, until dma_release_channel()
is called.

To ensure that all channels are not consumed by the general-purpose
allocator the DMA_PRIVATE capability is provided to exclude a dma_device
from general-purpose (memory-to-memory) consideration.

Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Maciej Sosnowski <maciej.sosnowski@intel.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
drivers/dma/dmaengine.c
include/linux/dmaengine.h

index 418eca28d472baf28909c6b4f30110dcef8af61e..7a0594f24a3f688ec16fcf79c04c8cb306c5f142 100644 (file)
@@ -134,14 +134,14 @@ static struct class dma_devclass = {
 
 /* --- 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);
 }
@@ -195,7 +195,7 @@ static int dma_chan_get(struct dma_chan *chan)
                        err = desc_cnt;
                        chan->client_count = 0;
                        module_put(owner);
-               } else
+               } else if (!dma_has_cap(DMA_PRIVATE, chan->device->cap_mask))
                        balance_ref_count(chan);
        }
 
@@ -232,14 +232,16 @@ static void dma_client_chan_alloc(struct dma_client *client)
 
        /* Find a channel */
        list_for_each_entry(device, &dma_device_list, global_node) {
+               if (dma_has_cap(DMA_PRIVATE, device->cap_mask))
+                       continue;
                /* Does the client require a specific DMA controller? */
                if (client->slave && client->slave->dma_dev
                                && client->slave->dma_dev != device->dev)
                        continue;
+               if (!dma_device_satisfies_mask(device, client->cap_mask))
+                       continue;
 
                list_for_each_entry(chan, &device->channels, device_node) {
-                       if (!dma_chan_satisfies_mask(chan, client->cap_mask))
-                               continue;
                        if (!chan->client_count)
                                continue;
                        ack = client->event_callback(client, chan,
@@ -320,11 +322,12 @@ static int __init dma_channel_table_init(void)
 
        bitmap_fill(dma_cap_mask_all.bits, DMA_TX_TYPE_END);
 
-       /* 'interrupt' and 'slave' are channel capabilities, but are not
-        * associated with an operation so they do not need an entry in the
-        * channel_table
+       /* '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) {
@@ -378,10 +381,13 @@ void dma_issue_pending_all(void)
                  "client called %s without a reference", __func__);
 
        rcu_read_lock();
-       list_for_each_entry_rcu(device, &dma_device_list, global_node)
+       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);
@@ -403,7 +409,8 @@ static struct dma_chan *nth_chan(enum dma_transaction_type cap, int n)
        struct dma_chan *min = NULL;
 
        list_for_each_entry(device, &dma_device_list, global_node) {
-               if (!dma_has_cap(cap, device->cap_mask))
+               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)
@@ -452,9 +459,12 @@ static void dma_channel_rebalance(void)
                for_each_possible_cpu(cpu)
                        per_cpu_ptr(channel_table[cap], cpu)->chan = NULL;
 
-       list_for_each_entry(device, &dma_device_list, global_node)
+       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;
+       }
 
        /* don't populate the channel_table if no clients are available */
        if (!dmaengine_ref_count)
@@ -473,6 +483,111 @@ static void dma_channel_rebalance(void)
                }
 }
 
+static struct dma_chan *private_candidate(dma_cap_mask_t *mask, struct dma_device *dev)
+{
+       struct dma_chan *chan;
+       struct dma_chan *ret = NULL;
+
+       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;
+               }
+
+       list_for_each_entry(chan, &dev->channels, device_node) {
+               if (chan->client_count) {
+                       pr_debug("%s: %s busy\n",
+                                __func__, dev_name(&chan->dev));
+                       continue;
+               }
+               ret = chan;
+               break;
+       }
+
+       return ret;
+}
+
+/**
+ * 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
+ */
+struct dma_chan *__dma_request_channel(dma_cap_mask_t *mask, dma_filter_fn fn, void *fn_param)
+{
+       struct dma_device *device, *_d;
+       struct dma_chan *chan = NULL;
+       enum dma_state_client ack;
+       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);
+               if (!chan)
+                       continue;
+
+               if (fn)
+                       ack = fn(chan, fn_param);
+               else
+                       ack = DMA_ACK;
+
+               if (ack == DMA_ACK) {
+                       /* 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);
+                       err = dma_chan_get(chan);
+
+                       if (err == -ENODEV) {
+                               pr_debug("%s: %s module removed\n", __func__,
+                                        dev_name(&chan->dev));
+                               list_del_rcu(&device->global_node);
+                       } else if (err)
+                               pr_err("dmaengine: failed to get %s: (%d)\n",
+                                      dev_name(&chan->dev), err);
+                       else
+                               break;
+               } else if (ack == DMA_DUP) {
+                       pr_debug("%s: %s filter said DMA_DUP\n",
+                                __func__, dev_name(&chan->dev));
+               } else if (ack == DMA_NAK) {
+                       pr_debug("%s: %s filter said DMA_NAK\n",
+                                __func__, dev_name(&chan->dev));
+                       break;
+               } else
+                       WARN_ONCE(1, "filter_fn: unknown response?\n");
+               chan = NULL;
+       }
+       mutex_unlock(&dma_list_mutex);
+
+       pr_debug("%s: %s (%s)\n", __func__, chan ? "success" : "fail",
+                chan ? dev_name(&chan->dev) : NULL);
+
+       return chan;
+}
+EXPORT_SYMBOL_GPL(__dma_request_channel);
+
+void dma_release_channel(struct dma_chan *chan)
+{
+       mutex_lock(&dma_list_mutex);
+       WARN_ONCE(chan->client_count != 1,
+                 "chan reference count %d != 1\n", chan->client_count);
+       dma_chan_put(chan);
+       mutex_unlock(&dma_list_mutex);
+}
+EXPORT_SYMBOL_GPL(dma_release_channel);
+
 /**
  * dma_chans_notify_available - broadcast available channels to the clients
  */
@@ -506,7 +621,9 @@ void dma_async_client_register(struct dma_client *client)
        dmaengine_ref_count++;
 
        /* try to grab channels */
-       list_for_each_entry_safe(device, _d, &dma_device_list, global_node)
+       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) {
@@ -517,6 +634,7 @@ void dma_async_client_register(struct dma_client *client)
                                pr_err("dmaengine: failed to get %s: (%d)\n",
                                       dev_name(&chan->dev), err);
                }
+       }
 
        /* if this is the first reference and there were channels
         * waiting we need to rebalance to get those channels
@@ -547,9 +665,12 @@ void dma_async_client_unregister(struct dma_client *client)
        dmaengine_ref_count--;
        BUG_ON(dmaengine_ref_count < 0);
        /* drop channel references */
-       list_for_each_entry(device, &dma_device_list, global_node)
+       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);
+       }
 
        list_del(&client->global_node);
        mutex_unlock(&dma_list_mutex);
@@ -639,9 +760,11 @@ int dma_async_device_register(struct dma_device *device)
                chan->slow_ref = 0;
                INIT_RCU_HEAD(&chan->rcu);
        }
+       device->chancnt = chancnt;
 
        mutex_lock(&dma_list_mutex);
-       if (dmaengine_ref_count)
+       /* 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
index 57a43adfc39ee35de82a9ecae8df54de6e5061bd..fe40bc020af638de72b5692e9d66dcad5212d8d5 100644 (file)
@@ -89,6 +89,7 @@ enum dma_transaction_type {
        DMA_MEMSET,
        DMA_MEMCPY_CRC32C,
        DMA_INTERRUPT,
+       DMA_PRIVATE,
        DMA_SLAVE,
 };
 
@@ -223,6 +224,18 @@ struct dma_client;
 typedef enum dma_state_client (*dma_event_callback) (struct dma_client *client,
                struct dma_chan *chan, enum dma_state state);
 
+/**
+ * typedef dma_filter_fn - callback filter for dma_request_channel
+ * @chan: channel to be reviewed
+ * @filter_param: opaque parameter passed through dma_request_channel
+ *
+ * When this optional parameter is specified in a call to dma_request_channel a
+ * suitable channel is passed to this routine for further dispositioning before
+ * being returned.  Where 'suitable' indicates a non-busy channel that
+ * satisfies the given capability mask.
+ */
+typedef enum dma_state_client (*dma_filter_fn)(struct dma_chan *chan, void *filter_param);
+
 /**
  * struct dma_client - info on the entity making use of DMA services
  * @event_callback: func ptr to call when something happens
@@ -472,6 +485,9 @@ void dma_async_device_unregister(struct dma_device *device);
 void dma_run_dependencies(struct dma_async_tx_descriptor *tx);
 struct dma_chan *dma_find_channel(enum dma_transaction_type tx_type);
 void dma_issue_pending_all(void);
+#define dma_request_channel(mask, x, y) __dma_request_channel(&(mask), x, y)
+struct dma_chan *__dma_request_channel(dma_cap_mask_t *mask, dma_filter_fn fn, void *fn_param);
+void dma_release_channel(struct dma_chan *chan);
 
 /* --- Helper iov-locking functions --- */