dev: Add a simple DMA engine that can be used by devices

Add a simple DMA engine that sits behind a FIFO. This engine can be
used by devices that need to read large amounts of data (e.g., display
controllers). Most aspects of the controller, such as FIFO size,
maximum number of in-flight accesses, and maximum request sizes can be
configured.

The DMA copies blocks of data into its FIFO. Transfers are initiated
with a call to startFill() command that takes a start address and a
size. Advanced users can create a derived class that overrides the
onEndOfBlock() callback that is triggered when the last request to a
block has been issued. At this point, the DMA engine is ready to start
fetching a new block of data, potentially from a different address
range.

The DMA engine stops issuing new requests while it is draining. Care
must be taken to ensure that devices that are fed by a DMA engine are
suspended while the system is draining to avoid buffer underruns.

2 files changed, 472 insertions, 14 deletions

diff --git a/src/dev/dma_device.cc b/src/dev/dma_device.cc
index 343f706b0..67de5ec4b 100644
--- a/src/dev/dma_device.cc
+++ b/src/dev/dma_device.cc
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2012 ARM Limited
+ * Copyright (c) 2012, 2015 ARM Limited
  * All rights reserved.
  *
  * The license below extends only to copyright in the software and shall
@@ -40,18 +40,22 @@
  * Authors: Ali Saidi
  *          Nathan Binkert
  *          Andreas Hansson
+ *          Andreas Sandberg
  */
 
+#include "dev/dma_device.hh"
+
+#include <utility>
+
 #include "base/chunk_generator.hh"
 #include "debug/DMA.hh"
 #include "debug/Drain.hh"
-#include "dev/dma_device.hh"
 #include "sim/system.hh"
 
 DmaPort::DmaPort(MemObject *dev, System *s)
-    : MasterPort(dev->name() + ".dma", dev), device(dev), sendEvent(this),
-      sys(s), masterId(s->getMasterId(dev->name())),
-      pendingCount(0), inRetry(false)
+    : MasterPort(dev->name() + ".dma", dev),
+      device(dev), sys(s), masterId(s->getMasterId(dev->name())),
+      sendEvent(this), pendingCount(0), inRetry(false)
 { }
 
 void
@@ -262,3 +266,214 @@ DmaDevice::getMasterPort(const std::string &if_name, PortID idx)
     }
     return PioDevice::getMasterPort(if_name, idx);
 }
+
+
+
+
+
+DmaReadFifo::DmaReadFifo(DmaPort &_port, size_t size,
+                         unsigned max_req_size,
+                         unsigned max_pending,
+                         Request::Flags flags)
+    : maxReqSize(max_req_size), fifoSize(size),
+      reqFlags(flags), port(_port),
+      buffer(size),
+      nextAddr(0), endAddr(0)
+{
+    freeRequests.resize(max_pending);
+    for (auto &e : freeRequests)
+        e.reset(new DmaDoneEvent(this, max_req_size));
+
+}
+
+DmaReadFifo::~DmaReadFifo()
+{
+    for (auto &p : pendingRequests) {
+        DmaDoneEvent *e(p.release());
+
+        if (e->done()) {
+            delete e;
+        } else {
+            // We can't kill in-flight DMAs, so we'll just transfer
+            // ownership to the event queue so that they get freed
+            // when they are done.
+            e->kill();
+        }
+    }
+}
+
+void
+DmaReadFifo::serialize(CheckpointOut &cp) const
+{
+    assert(pendingRequests.empty());
+
+    SERIALIZE_CONTAINER(buffer);
+    SERIALIZE_SCALAR(endAddr);
+    SERIALIZE_SCALAR(nextAddr);
+}
+
+void
+DmaReadFifo::unserialize(CheckpointIn &cp)
+{
+    UNSERIALIZE_CONTAINER(buffer);
+    UNSERIALIZE_SCALAR(endAddr);
+    UNSERIALIZE_SCALAR(nextAddr);
+}
+
+bool
+DmaReadFifo::tryGet(uint8_t *dst, size_t len)
+{
+    if (buffer.size() >= len) {
+        buffer.read(dst, len);
+        resumeFill();
+        return true;
+    } else {
+        return false;
+    }
+}
+
+void
+DmaReadFifo::get(uint8_t *dst, size_t len)
+{
+    const bool success(tryGet(dst, len));
+    panic_if(!success, "Buffer underrun in DmaReadFifo::get()\n");
+}
+
+void
+DmaReadFifo::startFill(Addr start, size_t size)
+{
+    assert(atEndOfBlock());
+
+    nextAddr = start;
+    endAddr = start + size;
+    resumeFill();
+}
+
+void
+DmaReadFifo::stopFill()
+{
+    // Prevent new DMA requests by setting the next address to the end
+    // address. Pending requests will still complete.
+    nextAddr = endAddr;
+
+    // Flag in-flight accesses as canceled. This prevents their data
+    // from being written to the FIFO.
+    for (auto &p : pendingRequests)
+        p->cancel();
+}
+
+void
+DmaReadFifo::resumeFill()
+{
+    // Don't try to fetch more data if we are draining. This ensures
+    // that the DMA engine settles down before we checkpoint it.
+    if (drainState() == DrainState::Draining)
+        return;
+
+    const bool old_eob(atEndOfBlock());
+    size_t size_pending(0);
+    for (auto &e : pendingRequests)
+        size_pending += e->requestSize();
+
+    while (!freeRequests.empty() && !atEndOfBlock()) {
+        const size_t req_size(std::min(maxReqSize, endAddr - nextAddr));
+        if (buffer.size() + size_pending + req_size > fifoSize)
+            break;
+
+        DmaDoneEventUPtr event(std::move(freeRequests.front()));
+        freeRequests.pop_front();
+        assert(event);
+
+        event->reset(req_size);
+        port.dmaAction(MemCmd::ReadReq, nextAddr, req_size, event.get(),
+                       event->data(), 0, reqFlags);
+        nextAddr += req_size;
+        size_pending += req_size;
+
+        pendingRequests.emplace_back(std::move(event));
+    }
+
+    // EOB can be set before a call to dmaDone() if in-flight accesses
+    // have been canceled.
+    if (!old_eob && atEndOfBlock())
+        onEndOfBlock();
+}
+
+void
+DmaReadFifo::dmaDone()
+{
+    const bool old_active(isActive());
+
+    handlePending();
+    resumeFill();
+
+    if (!old_active && isActive())
+        onIdle();
+}
+
+void
+DmaReadFifo::handlePending()
+{
+    while (!pendingRequests.empty() && pendingRequests.front()->done()) {
+        // Get the first finished pending request
+        DmaDoneEventUPtr event(std::move(pendingRequests.front()));
+        pendingRequests.pop_front();
+
+        if (!event->canceled())
+            buffer.write(event->data(), event->requestSize());
+
+        // Move the event to the list of free requests
+        freeRequests.emplace_back(std::move(event));
+    }
+
+    if (pendingRequests.empty())
+        signalDrainDone();
+}
+
+
+
+DrainState
+DmaReadFifo::drain()
+{
+    return pendingRequests.empty() ? DrainState::Drained : DrainState::Draining;
+}
+
+
+DmaReadFifo::DmaDoneEvent::DmaDoneEvent(DmaReadFifo *_parent,
+                                        size_t max_size)
+    : parent(_parent), _done(false), _canceled(false), _data(max_size, 0)
+{
+}
+
+void
+DmaReadFifo::DmaDoneEvent::kill()
+{
+    parent = nullptr;
+    setFlags(AutoDelete);
+}
+
+void
+DmaReadFifo::DmaDoneEvent::cancel()
+{
+    _canceled = true;
+}
+
+void
+DmaReadFifo::DmaDoneEvent::reset(size_t size)
+{
+    assert(size <= _data.size());
+    _done = false;
+    _canceled = false;
+    _requestSize = size;
+}
+
+void
+DmaReadFifo::DmaDoneEvent::process()
+{
+    if (!parent)
+        return;
+
+    assert(!_done);
+    _done = true;
+    parent->dmaDone();
+}
diff --git a/src/dev/dma_device.hh b/src/dev/dma_device.hh
index b0e35e44f..a7a03d520 100644
--- a/src/dev/dma_device.hh
+++ b/src/dev/dma_device.hh
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2012-2013 ARM Limited
+ * Copyright (c) 2012-2013, 2015 ARM Limited
  * All rights reserved.
  *
  * The license below extends only to copyright in the software and shall
@@ -39,13 +39,16 @@
  *
  * Authors: Ali Saidi
  *          Nathan Binkert
+ *          Andreas Sandberg
  */
 
 #ifndef __DEV_DMA_DEVICE_HH__
 #define __DEV_DMA_DEVICE_HH__
 
 #include <deque>
+#include <memory>
 
+#include "base/circlebuf.hh"
 #include "dev/io_device.hh"
 #include "params/DmaDevice.hh"
 #include "sim/drain.hh"
@@ -104,22 +107,24 @@ class DmaPort : public MasterPort, public Drainable
         {}
     };
 
+  public:
     /** The device that owns this port. */
-    MemObject *device;
-
-    /** Use a deque as we never do any insertion or removal in the middle */
-    std::deque<PacketPtr> transmitList;
-
-    /** Event used to schedule a future sending from the transmit list. */
-    EventWrapper<DmaPort, &DmaPort::sendDma> sendEvent;
+    MemObject *const device;
 
     /** The system that device/port are in. This is used to select which mode
      * we are currently operating in. */
-    System *sys;
+    System *const sys;
 
     /** Id for all requests */
     const MasterID masterId;
 
+  protected:
+    /** Use a deque as we never do any insertion or removal in the middle */
+    std::deque<PacketPtr> transmitList;
+
+    /** Event used to schedule a future sending from the transmit list. */
+    EventWrapper<DmaPort, &DmaPort::sendDma> sendEvent;
+
     /** Number of outstanding packets the dma port has. */
     uint32_t pendingCount;
 
@@ -179,4 +184,242 @@ class DmaDevice : public PioDevice
 
 };
 
+/**
+ * Buffered DMA engine helper class
+ *
+ * This class implements a simple DMA engine that feeds a FIFO
+ * buffer. The size of the buffer, the maximum number of pending
+ * requests and the maximum request size are all set when the engine
+ * is instantiated.
+ *
+ * An <i>asynchronous</i> transfer of a <i>block</i> of data
+ * (designated by a start address and a size) is started by calling
+ * the startFill() method. The DMA engine will aggressively try to
+ * keep the internal FIFO full. As soon as there is room in the FIFO
+ * for more data <i>and</i> there are free request slots, a new fill
+ * will be started.
+ *
+ * Data in the FIFO can be read back using the get() and tryGet()
+ * methods. Both request a block of data from the FIFO. However, get()
+ * panics if the block cannot be satisfied, while tryGet() simply
+ * returns false. The latter call makes it possible to implement
+ * custom buffer underrun handling.
+ *
+ * A simple use case would be something like this:
+ * \code{.cpp}
+ *     // Create a DMA engine with a 1KiB buffer. Issue up to 8 concurrent
+ *     // uncacheable 64 byte (maximum) requests.
+ *     DmaReadFifo *dma = new DmaReadFifo(port, 1024, 64, 8,
+ *                                        Request::UNCACHEABLE);
+ *
+ *     // Start copying 4KiB data from 0xFF000000
+ *     dma->startFill(0xFF000000, 0x1000);
+ *
+ *     // Some time later when there is data in the FIFO.
+ *     uint8_t data[8];
+ *     dma->get(data, sizeof(data))
+ * \endcode
+ *
+ *
+ * The DMA engine allows new blocks to be requested as soon as the
+ * last request for a block has been sent (i.e., there is no need to
+ * wait for pending requests to complete). This can be queried with
+ * the atEndOfBlock() method and more advanced implementations may
+ * override the onEndOfBlock() callback.
+ */
+class DmaReadFifo : public Drainable, public Serializable
+{
+  public:
+    DmaReadFifo(DmaPort &port, size_t size,
+                unsigned max_req_size,
+                unsigned max_pending,
+                Request::Flags flags = 0);
+
+    ~DmaReadFifo();
+
+  public: // Serializable
+    void serialize(CheckpointOut &cp) const M5_ATTR_OVERRIDE;
+    void unserialize(CheckpointIn &cp) M5_ATTR_OVERRIDE;
+
+  public: // Drainable
+    DrainState drain() M5_ATTR_OVERRIDE;
+
+  public: // FIFO access
+    /**
+     * @{
+     * @name FIFO access
+     */
+    /**
+     * Try to read data from the FIFO.
+     *
+     * This method reads len bytes of data from the FIFO and stores
+     * them in the memory location pointed to by dst. The method
+     * fails, and no data is written to the buffer, if the FIFO
+     * doesn't contain enough data to satisfy the request.
+     *
+     * @param dst Pointer to a destination buffer
+     * @param len Amount of data to read.
+     * @return true on success, false otherwise.
+     */
+    bool tryGet(uint8_t *dst, size_t len);
+
+    template<typename T>
+    bool tryGet(T &value) {
+        return tryGet(static_cast<T *>(&value), sizeof(T));
+    };
+
+    /**
+     * Read data from the FIFO and panic on failure.
+     *
+     * @see tryGet()
+     *
+     * @param dst Pointer to a destination buffer
+     * @param len Amount of data to read.
+     */
+    void get(uint8_t *dst, size_t len);
+
+    template<typename T>
+    T get() {
+        T value;
+        get(static_cast<uint8_t *>(&value), sizeof(T));
+        return value;
+    };
+
+    /** Get the amount of data stored in the FIFO */
+    size_t size() const { return buffer.size(); }
+    /** Flush the FIFO */
+    void flush() { buffer.flush(); }
+
+    /** @} */
+  public: // FIFO fill control
+    /**
+     * @{
+     * @name FIFO fill control
+     */
+    /**
+     * Start filling the FIFO.
+     *
+     * @warn It's considered an error to call start on an active DMA
+     * engine unless the last request from the active block has been
+     * sent (i.e., atEndOfBlock() is true).
+     *
+     * @param start Physical address to copy from.
+     * @param size Size of the block to copy.
+     */
+    void startFill(Addr start, size_t size);
+
+    /**
+     * Stop the DMA engine.
+     *
+     * Stop filling the FIFO and ignore incoming responses for pending
+     * requests. The onEndOfBlock() callback will not be called after
+     * this method has been invoked. However, once the last response
+     * has been received, the onIdle() callback will still be called.
+     */
+    void stopFill();
+
+    /**
+     * Has the DMA engine sent out the last request for the active
+     * block?
+     */
+    bool atEndOfBlock() const {
+        return nextAddr == endAddr;
+    }
+
+    /**
+     * Is the DMA engine active (i.e., are there still in-flight
+     * accesses)?
+     */
+    bool isActive() const {
+        return !(pendingRequests.empty() && atEndOfBlock());
+    }
+
+    /** @} */
+  protected: // Callbacks
+    /**
+     * @{
+     * @name Callbacks
+     */
+    /**
+     * End of block callback
+     *
+     * This callback is called <i>once</i> after the last access in a
+     * block has been sent. It is legal for a derived class to call
+     * startFill() from this method to initiate a transfer.
+     */
+    virtual void onEndOfBlock() {};
+
+    /**
+     * Last response received callback
+     *
+     * This callback is called when the DMA engine becomes idle (i.e.,
+     * there are no pending requests).
+     *
+     * It is possible for a DMA engine to reach the end of block and
+     * become idle at the same tick. In such a case, the
+     * onEndOfBlock() callback will be called first. This callback
+     * will <i>NOT</i> be called if that callback initiates a new DMA transfer.
+     */
+    virtual void onIdle() {};
+
+    /** @} */
+  private: // Configuration
+    /** Maximum request size in bytes */
+    const Addr maxReqSize;
+    /** Maximum FIFO size in bytes */
+    const size_t fifoSize;
+    /** Request flags */
+    const Request::Flags reqFlags;
+
+    DmaPort &port;
+
+  private:
+    class DmaDoneEvent : public Event
+    {
+      public:
+        DmaDoneEvent(DmaReadFifo *_parent, size_t max_size);
+
+        void kill();
+        void cancel();
+        bool canceled() const { return _canceled; }
+        void reset(size_t size);
+        void process();
+
+        bool done() const { return _done; }
+        size_t requestSize() const { return _requestSize; }
+        const uint8_t *data() const { return _data.data(); }
+        uint8_t *data() { return _data.data(); }
+
+      private:
+        DmaReadFifo *parent;
+        bool _done;
+        bool _canceled;
+        size_t _requestSize;
+        std::vector<uint8_t> _data;
+    };
+
+    typedef std::unique_ptr<DmaDoneEvent> DmaDoneEventUPtr;
+
+    /**
+     * DMA request done, handle incoming data and issue new
+     * request.
+     */
+    void dmaDone();
+
+    /** Handle pending requests that have been flagged as done. */
+    void handlePending();
+
+    /** Try to issue new DMA requests */
+    void resumeFill();
+
+  private: // Internal state
+    Fifo<uint8_t> buffer;
+
+    Addr nextAddr;
+    Addr endAddr;
+
+    std::deque<DmaDoneEventUPtr> pendingRequests;
+    std::deque<DmaDoneEventUPtr> freeRequests;
+};
+
 #endif // __DEV_DMA_DEVICE_HH__