src/dev/dma_device.cc - public/gem5 - Git at Google

 /*
  * Copyright (c) 2012, 2015, 2017, 2019 ARM Limited
  * All rights reserved.
  *
  * The license below extends only to copyright in the software and shall
  * not be construed as granting a license to any other intellectual
  * property including but not limited to intellectual property relating
  * to a hardware implementation of the functionality of the software
  * licensed hereunder.  You may use the software subject to the license
  * terms below provided that you ensure that this notice is replicated
  * unmodified and in its entirety in all distributions of the software,
  * modified or unmodified, in source code or in binary form.
  *
  * Copyright (c) 2006 The Regents of The University of Michigan
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
  * modification, are permitted provided that the following conditions are
  * met: redistributions of source code must retain the above copyright
  * notice, this list of conditions and the following disclaimer;
  * redistributions in binary form must reproduce the above copyright
  * notice, this list of conditions and the following disclaimer in the
  * documentation and/or other materials provided with the distribution;
  * neither the name of the copyright holders nor the names of its
  * contributors may be used to endorse or promote products derived from
  * this software without specific prior written permission.
  *
  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  */

 #include "dev/dma_device.hh"

 #include <algorithm>
 #include <cassert>
 #include <cstring>
 #include <utility>

 #include "base/logging.hh"
 #include "base/trace.hh"
 #include "debug/DMA.hh"
 #include "debug/Drain.hh"
 #include "sim/clocked_object.hh"
 #include "sim/system.hh"

 namespace gem5
 {

 DmaPort::DmaPort(ClockedObject *dev, System *s,
                  uint32_t sid, uint32_t ssid)
     : RequestPort(dev->name() + ".dma"),
       device(dev), sys(s), requestorId(s->getRequestorId(dev)),
       sendEvent([this]{ sendDma(); }, dev->name()),
       defaultSid(sid), defaultSSid(ssid), cacheLineSize(s->cacheLineSize())
 { }

 void
 DmaPort::handleRespPacket(PacketPtr pkt, Tick delay)
 {
     // Should always see a response with a sender state.
     assert(pkt->isResponse());

     // Get the DMA sender state.
     auto *state = dynamic_cast<DmaReqState*>(pkt->senderState);
     assert(state);

     handleResp(state, pkt->getAddr(), pkt->req->getSize(), delay);

     delete pkt;
 }

 void
 DmaPort::handleResp(DmaReqState *state, Addr addr, Addr size, Tick delay)
 {
     DPRINTF(DMA, "Received response %s for addr: %#x size: %d nb: %d,"  \
             " tot: %d sched %d\n",
             MemCmd(state->cmd).toString(), addr, size,
             state->numBytes, state->totBytes,
             state->completionEvent ?
             state->completionEvent->scheduled() : 0);

     // Update the number of bytes received based on the request rather
     // than the packet as the latter could be rounded up to line sizes.
     state->numBytes += size;
     assert(state->totBytes >= state->numBytes);

     // If we have reached the total number of bytes for this DMA request,
     // then signal the completion and delete the sate.
     if (state->totBytes == state->numBytes) {
         assert(pendingCount != 0);
         pendingCount--;
         if (state->completionEvent) {
             delay += state->delay;
             device->schedule(state->completionEvent, curTick() + delay);
         }
         delete state;
     }

     // We might be drained at this point, if so signal the drain event.
     if (pendingCount == 0)
         signalDrainDone();
 }

 PacketPtr
 DmaPort::DmaReqState::createPacket()
 {
     RequestPtr req = std::make_shared<Request>(
             gen.addr(), gen.size(), flags, id);
     req->setStreamId(sid);
     req->setSubstreamId(ssid);
     req->taskId(context_switch_task_id::DMA);

     PacketPtr pkt = new Packet(req, cmd);

     if (data)
         pkt->dataStatic(data + gen.complete());

     pkt->senderState = this;
     return pkt;
 }

 bool
 DmaPort::recvTimingResp(PacketPtr pkt)
 {
     // We shouldn't ever get a cacheable block in Modified state.
     assert(pkt->req->isUncacheable() ||
            !(pkt->cacheResponding() && !pkt->hasSharers()));

     handleRespPacket(pkt);

     return true;
 }

 DmaDevice::DmaDevice(const Params &p)
     : PioDevice(p), dmaPort(this, sys, p.sid, p.ssid)
 { }

 void
 DmaDevice::init()
 {
     panic_if(!dmaPort.isConnected(),
              "DMA port of %s not connected to anything!", name());
     PioDevice::init();
 }

 DrainState
 DmaPort::drain()
 {
     if (pendingCount == 0) {
         return DrainState::Drained;
     } else {
         DPRINTF(Drain, "DmaPort not drained\n");
         return DrainState::Draining;
     }
 }

 void
 DmaPort::recvReqRetry()
 {
     assert(transmitList.size());
     trySendTimingReq();
 }

 void
 DmaPort::dmaAction(Packet::Command cmd, Addr addr, int size, Event *event,
                    uint8_t *data, uint32_t sid, uint32_t ssid, Tick delay,
                    Request::Flags flag)
 {
     DPRINTF(DMA, "Starting DMA for addr: %#x size: %d sched: %d\n", addr, size,
             event ? event->scheduled() : -1);

     // One DMA request sender state for every action, that is then
     // split into many requests and packets based on the block size,
     // i.e. cache line size.
     transmitList.push_back(
             new DmaReqState(cmd, addr, cacheLineSize, size,
                 data, flag, requestorId, sid, ssid, event, delay));
     pendingCount++;

     // In zero time, also initiate the sending of the packets for the request
     // we have just created. For atomic this involves actually completing all
     // the requests.
     sendDma();
 }

 void
 DmaPort::dmaAction(Packet::Command cmd, Addr addr, int size, Event *event,
                    uint8_t *data, Tick delay, Request::Flags flag)
 {
     dmaAction(cmd, addr, size, event, data,
               defaultSid, defaultSSid, delay, flag);
 }

 void
 DmaPort::trySendTimingReq()
 {
     // Send the next packet for the first DMA request on the transmit list,
     // and schedule the following send if it is successful
     DmaReqState *state = transmitList.front();

     PacketPtr pkt = inRetry ? inRetry : state->createPacket();
     inRetry = nullptr;

     DPRINTF(DMA, "Trying to send %s addr %#x\n", pkt->cmdString(),
             pkt->getAddr());

     // Check if this was the last packet now, since hypothetically the packet
     // response may come immediately, and state may be deleted.
     bool last = state->gen.last();
     if (!sendTimingReq(pkt))
         inRetry = pkt;
     if (!inRetry) {
         // If that was the last packet from this request, pop it from the list.
         if (last)
             transmitList.pop_front();
         else
             state->gen.next();
         DPRINTF(DMA, "-- Done\n");
         // If there is more to do, then do so.
         if (!transmitList.empty()) {
             // This should ultimately wait for as many cycles as the device
             // needs to send the packet, but currently the port does not have
             // any known width so simply wait a single cycle.
             device->schedule(sendEvent, device->clockEdge(Cycles(1)));
         }
     } else {
         DPRINTF(DMA, "-- Failed, waiting for retry\n");
     }

     DPRINTF(DMA, "TransmitList: %d, inRetry: %d\n",
             transmitList.size(), inRetry ? 1 : 0);
 }

 bool
 DmaPort::sendAtomicReq(DmaReqState *state)
 {
     PacketPtr pkt = state->createPacket();
     DPRINTF(DMA, "Sending  DMA for addr: %#x size: %d\n",
             state->gen.addr(), state->gen.size());
     Tick lat = sendAtomic(pkt);

     // Check if we're done, since handleResp may delete state.
     bool done = !state->gen.next();
     handleRespPacket(pkt, lat);
     return done;
 }

 bool
 DmaPort::sendAtomicBdReq(DmaReqState *state)
 {
     bool done = false;

     auto bd_it = memBackdoors.contains(state->gen.addr());
     if (bd_it == memBackdoors.end()) {
         // We don't have a backdoor for this address, so use a packet.

         PacketPtr pkt = state->createPacket();
         DPRINTF(DMA, "Sending DMA for addr: %#x size: %d\n",
                 state->gen.addr(), state->gen.size());

         MemBackdoorPtr bd = nullptr;
         Tick lat = sendAtomicBackdoor(pkt, bd);

         // If we got a backdoor, record it.
         if (bd && memBackdoors.insert(bd->range(), bd) != memBackdoors.end()) {
             // Invalidation callback which finds this backdoor and removes it.
             auto callback = [this](const MemBackdoor &backdoor) {
                 for (auto it = memBackdoors.begin();
                         it != memBackdoors.end(); it++) {
                     if (it->second == &backdoor) {
                         memBackdoors.erase(it);
                         return;
                     }
                 }
                 panic("Got invalidation for unknown memory backdoor.");
             };
             bd->addInvalidationCallback(callback);
         }

         // Check if we're done now, since handleResp may delete state.
         done = !state->gen.next();
         handleRespPacket(pkt, lat);
     } else {
         // We have a backdoor that can at least partially satisfy this request.
         DPRINTF(DMA, "Handling DMA for addr: %#x size %d through backdoor\n",
                 state->gen.addr(), state->gen.size());

         const auto *bd = bd_it->second;
         // Offset of this access into the backdoor.
         const Addr offset = state->gen.addr() - bd->range().start();
         // How many bytes we still need.
         const Addr remaining = state->totBytes - state->gen.complete();
         // How many bytes this backdoor can provide, starting from offset.
         const Addr available = bd->range().size() - offset;

         // How many bytes we're going to handle through this backdoor.
         const Addr handled = std::min(remaining, available);

         // If there's a buffer for data, read/write it.
         if (state->data) {
             uint8_t *bd_data = bd->ptr() + offset;
             uint8_t *state_data = state->data + state->gen.complete();
             if (MemCmd(state->cmd).isRead())
                 memcpy(state_data, bd_data, handled);
             else
                 memcpy(bd_data, state_data, handled);
         }

         // Advance the chunk generator past this region of memory.
         state->gen.setNext(state->gen.addr() + handled);

         // Check if we're done now, since handleResp may delete state.
         done = !state->gen.next();
         handleResp(state, state->gen.addr(), handled);
     }

     return done;
 }

 void
 DmaPort::sendDma()
 {
     // Some kind of selection between access methods. More work is going to
     // have to be done to make switching actually work.
     assert(transmitList.size());

     if (sys->isTimingMode()) {
         // If we are either waiting for a retry or are still waiting after
         // sending the last packet, then do not proceed.
         if (inRetry || sendEvent.scheduled()) {
             DPRINTF(DMA, "Can't send immediately, waiting to send\n");
             return;
         }

         trySendTimingReq();
     } else if (sys->isAtomicMode()) {
         const bool bypass = sys->bypassCaches();

         // Send everything there is to send in zero time.
         while (!transmitList.empty()) {
             DmaReqState *state = transmitList.front();
             transmitList.pop_front();

             bool done = state->gen.done();
             while (!done)
                 done = bypass ? sendAtomicBdReq(state) : sendAtomicReq(state);
         }
     } else {
         panic("Unknown memory mode.");
     }
 }

 Port &
 DmaDevice::getPort(const std::string &if_name, PortID idx)
 {
     if (if_name == "dma") {
         return dmaPort;
     }
     return PioDevice::getPort(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), cacheLineSize(port.sys->cacheLineSize()),
       buffer(size)
 {
     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)
 {
     panic_if(!tryGet(dst, len), "Buffer underrun in DmaReadFifo::get()");
 }

 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());

     if (port.sys->bypassCaches())
         resumeFillBypass();
     else
         resumeFillTiming();

     if (!old_eob && atEndOfBlock())
         onEndOfBlock();
 }

 void
 DmaReadFifo::resumeFillBypass()
 {
     const size_t fifo_space = buffer.capacity() - buffer.size();
     if (fifo_space >= cacheLineSize || buffer.capacity() < cacheLineSize) {
         const size_t block_remaining = endAddr - nextAddr;
         const size_t xfer_size = std::min(fifo_space, block_remaining);
         std::vector<uint8_t> tmp_buffer(xfer_size);

         assert(pendingRequests.empty());
         DPRINTF(DMA, "Direct bypass startAddr=%#x xfer_size=%#x " \
                 "fifo_space=%#x block_remaining=%#x\n",
                 nextAddr, xfer_size, fifo_space, block_remaining);

         port.dmaAction(MemCmd::ReadReq, nextAddr, xfer_size, nullptr,
                 tmp_buffer.data(), 0, reqFlags);

         buffer.write(tmp_buffer.begin(), xfer_size);
         nextAddr += xfer_size;
     }
 }

 void
 DmaReadFifo::resumeFillTiming()
 {
     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));
     }
 }

 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), _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();
 }

 } // namespace gem5
	/*
	* Copyright (c) 2012, 2015, 2017, 2019 ARM Limited
	* All rights reserved.
	*
	* The license below extends only to copyright in the software and shall
	* not be construed as granting a license to any other intellectual
	* property including but not limited to intellectual property relating
	* to a hardware implementation of the functionality of the software
	* licensed hereunder. You may use the software subject to the license
	* terms below provided that you ensure that this notice is replicated
	* unmodified and in its entirety in all distributions of the software,
	* modified or unmodified, in source code or in binary form.
	*
	* Copyright (c) 2006 The Regents of The University of Michigan
	* All rights reserved.
	*
	* Redistribution and use in source and binary forms, with or without
	* modification, are permitted provided that the following conditions are
	* met: redistributions of source code must retain the above copyright
	* notice, this list of conditions and the following disclaimer;
	* redistributions in binary form must reproduce the above copyright
	* notice, this list of conditions and the following disclaimer in the
	* documentation and/or other materials provided with the distribution;
	* neither the name of the copyright holders nor the names of its
	* contributors may be used to endorse or promote products derived from
	* this software without specific prior written permission.
	*
	* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
	* "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
	* LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
	* A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
	* OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
	* SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
	* LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
	* DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
	* THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
	* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
	* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
	*/

	#include "dev/dma_device.hh"

	#include <algorithm>
	#include <cassert>
	#include <cstring>
	#include <utility>

	#include "base/logging.hh"
	#include "base/trace.hh"
	#include "debug/DMA.hh"
	#include "debug/Drain.hh"
	#include "sim/clocked_object.hh"
	#include "sim/system.hh"

	namespace gem5
	{

	DmaPort::DmaPort(ClockedObject dev, System s,
	uint32_t sid, uint32_t ssid)
	: RequestPort(dev->name() + ".dma"),
	device(dev), sys(s), requestorId(s->getRequestorId(dev)),
	sendEvent([this]{ sendDma(); }, dev->name()),
	defaultSid(sid), defaultSSid(ssid), cacheLineSize(s->cacheLineSize())
	{ }

	void
	DmaPort::handleRespPacket(PacketPtr pkt, Tick delay)
	{
	// Should always see a response with a sender state.
	assert(pkt->isResponse());

	// Get the DMA sender state.
	auto state = dynamic_cast<DmaReqState>(pkt->senderState);
	assert(state);

	handleResp(state, pkt->getAddr(), pkt->req->getSize(), delay);

	delete pkt;
	}

	void
	DmaPort::handleResp(DmaReqState *state, Addr addr, Addr size, Tick delay)
	{
	DPRINTF(DMA, "Received response %s for addr: %#x size: %d nb: %d," \
	" tot: %d sched %d\n",
	MemCmd(state->cmd).toString(), addr, size,
	state->numBytes, state->totBytes,
	state->completionEvent ?
	state->completionEvent->scheduled() : 0);

	// Update the number of bytes received based on the request rather
	// than the packet as the latter could be rounded up to line sizes.
	state->numBytes += size;
	assert(state->totBytes >= state->numBytes);

	// If we have reached the total number of bytes for this DMA request,
	// then signal the completion and delete the sate.
	if (state->totBytes == state->numBytes) {
	assert(pendingCount != 0);
	pendingCount--;
	if (state->completionEvent) {
	delay += state->delay;
	device->schedule(state->completionEvent, curTick() + delay);
	}
	delete state;
	}

	// We might be drained at this point, if so signal the drain event.
	if (pendingCount == 0)
	signalDrainDone();
	}

	PacketPtr
	DmaPort::DmaReqState::createPacket()
	{
	RequestPtr req = std::make_shared<Request>(
	gen.addr(), gen.size(), flags, id);
	req->setStreamId(sid);
	req->setSubstreamId(ssid);
	req->taskId(context_switch_task_id::DMA);

	PacketPtr pkt = new Packet(req, cmd);

	if (data)
	pkt->dataStatic(data + gen.complete());

	pkt->senderState = this;
	return pkt;
	}

	bool
	DmaPort::recvTimingResp(PacketPtr pkt)
	{
	// We shouldn't ever get a cacheable block in Modified state.
	assert(pkt->req->isUncacheable() \|\|
	!(pkt->cacheResponding() && !pkt->hasSharers()));

	handleRespPacket(pkt);

	return true;
	}

	DmaDevice::DmaDevice(const Params &p)
	: PioDevice(p), dmaPort(this, sys, p.sid, p.ssid)
	{ }

	void
	DmaDevice::init()
	{
	panic_if(!dmaPort.isConnected(),
	"DMA port of %s not connected to anything!", name());
	PioDevice::init();
	}

	DrainState
	DmaPort::drain()
	{
	if (pendingCount == 0) {
	return DrainState::Drained;
	} else {
	DPRINTF(Drain, "DmaPort not drained\n");
	return DrainState::Draining;
	}
	}

	void
	DmaPort::recvReqRetry()
	{
	assert(transmitList.size());
	trySendTimingReq();
	}

	void
	DmaPort::dmaAction(Packet::Command cmd, Addr addr, int size, Event *event,
	uint8_t *data, uint32_t sid, uint32_t ssid, Tick delay,
	Request::Flags flag)
	{
	DPRINTF(DMA, "Starting DMA for addr: %#x size: %d sched: %d\n", addr, size,
	event ? event->scheduled() : -1);

	// One DMA request sender state for every action, that is then
	// split into many requests and packets based on the block size,
	// i.e. cache line size.
	transmitList.push_back(
	new DmaReqState(cmd, addr, cacheLineSize, size,
	data, flag, requestorId, sid, ssid, event, delay));
	pendingCount++;

	// In zero time, also initiate the sending of the packets for the request
	// we have just created. For atomic this involves actually completing all
	// the requests.
	sendDma();
	}

	void
	DmaPort::dmaAction(Packet::Command cmd, Addr addr, int size, Event *event,
	uint8_t *data, Tick delay, Request::Flags flag)
	{
	dmaAction(cmd, addr, size, event, data,
	defaultSid, defaultSSid, delay, flag);
	}

	void
	DmaPort::trySendTimingReq()
	{
	// Send the next packet for the first DMA request on the transmit list,
	// and schedule the following send if it is successful
	DmaReqState *state = transmitList.front();

	PacketPtr pkt = inRetry ? inRetry : state->createPacket();
	inRetry = nullptr;

	DPRINTF(DMA, "Trying to send %s addr %#x\n", pkt->cmdString(),
	pkt->getAddr());

	// Check if this was the last packet now, since hypothetically the packet
	// response may come immediately, and state may be deleted.
	bool last = state->gen.last();
	if (!sendTimingReq(pkt))
	inRetry = pkt;
	if (!inRetry) {
	// If that was the last packet from this request, pop it from the list.
	if (last)
	transmitList.pop_front();
	else
	state->gen.next();
	DPRINTF(DMA, "-- Done\n");
	// If there is more to do, then do so.
	if (!transmitList.empty()) {
	// This should ultimately wait for as many cycles as the device
	// needs to send the packet, but currently the port does not have
	// any known width so simply wait a single cycle.
	device->schedule(sendEvent, device->clockEdge(Cycles(1)));
	}
	} else {
	DPRINTF(DMA, "-- Failed, waiting for retry\n");
	}

	DPRINTF(DMA, "TransmitList: %d, inRetry: %d\n",
	transmitList.size(), inRetry ? 1 : 0);
	}

	bool
	DmaPort::sendAtomicReq(DmaReqState *state)
	{
	PacketPtr pkt = state->createPacket();
	DPRINTF(DMA, "Sending DMA for addr: %#x size: %d\n",
	state->gen.addr(), state->gen.size());
	Tick lat = sendAtomic(pkt);

	// Check if we're done, since handleResp may delete state.
	bool done = !state->gen.next();
	handleRespPacket(pkt, lat);
	return done;
	}

	bool
	DmaPort::sendAtomicBdReq(DmaReqState *state)
	{
	bool done = false;

	auto bd_it = memBackdoors.contains(state->gen.addr());
	if (bd_it == memBackdoors.end()) {
	// We don't have a backdoor for this address, so use a packet.

	PacketPtr pkt = state->createPacket();
	DPRINTF(DMA, "Sending DMA for addr: %#x size: %d\n",
	state->gen.addr(), state->gen.size());

	MemBackdoorPtr bd = nullptr;
	Tick lat = sendAtomicBackdoor(pkt, bd);

	// If we got a backdoor, record it.
	if (bd && memBackdoors.insert(bd->range(), bd) != memBackdoors.end()) {
	// Invalidation callback which finds this backdoor and removes it.
	auto callback = [this](const MemBackdoor &backdoor) {
	for (auto it = memBackdoors.begin();
	it != memBackdoors.end(); it++) {
	if (it->second == &backdoor) {
	memBackdoors.erase(it);
	return;
	}
	}
	panic("Got invalidation for unknown memory backdoor.");
	};
	bd->addInvalidationCallback(callback);
	}

	// Check if we're done now, since handleResp may delete state.
	done = !state->gen.next();
	handleRespPacket(pkt, lat);
	} else {
	// We have a backdoor that can at least partially satisfy this request.
	DPRINTF(DMA, "Handling DMA for addr: %#x size %d through backdoor\n",
	state->gen.addr(), state->gen.size());

	const auto *bd = bd_it->second;
	// Offset of this access into the backdoor.
	const Addr offset = state->gen.addr() - bd->range().start();
	// How many bytes we still need.
	const Addr remaining = state->totBytes - state->gen.complete();
	// How many bytes this backdoor can provide, starting from offset.
	const Addr available = bd->range().size() - offset;

	// How many bytes we're going to handle through this backdoor.
	const Addr handled = std::min(remaining, available);

	// If there's a buffer for data, read/write it.
	if (state->data) {
	uint8_t *bd_data = bd->ptr() + offset;
	uint8_t *state_data = state->data + state->gen.complete();
	if (MemCmd(state->cmd).isRead())
	memcpy(state_data, bd_data, handled);
	else
	memcpy(bd_data, state_data, handled);
	}

	// Advance the chunk generator past this region of memory.
	state->gen.setNext(state->gen.addr() + handled);

	// Check if we're done now, since handleResp may delete state.
	done = !state->gen.next();
	handleResp(state, state->gen.addr(), handled);
	}

	return done;
	}

	void
	DmaPort::sendDma()
	{
	// Some kind of selection between access methods. More work is going to
	// have to be done to make switching actually work.
	assert(transmitList.size());

	if (sys->isTimingMode()) {
	// If we are either waiting for a retry or are still waiting after
	// sending the last packet, then do not proceed.
	if (inRetry \|\| sendEvent.scheduled()) {
	DPRINTF(DMA, "Can't send immediately, waiting to send\n");
	return;
	}

	trySendTimingReq();
	} else if (sys->isAtomicMode()) {
	const bool bypass = sys->bypassCaches();

	// Send everything there is to send in zero time.
	while (!transmitList.empty()) {
	DmaReqState *state = transmitList.front();
	transmitList.pop_front();

	bool done = state->gen.done();
	while (!done)
	done = bypass ? sendAtomicBdReq(state) : sendAtomicReq(state);
	}
	} else {
	panic("Unknown memory mode.");
	}
	}

	Port &
	DmaDevice::getPort(const std::string &if_name, PortID idx)
	{
	if (if_name == "dma") {
	return dmaPort;
	}
	return PioDevice::getPort(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), cacheLineSize(port.sys->cacheLineSize()),
	buffer(size)
	{
	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)
	{
	panic_if(!tryGet(dst, len), "Buffer underrun in DmaReadFifo::get()");
	}

	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());

	if (port.sys->bypassCaches())
	resumeFillBypass();
	else
	resumeFillTiming();

	if (!old_eob && atEndOfBlock())
	onEndOfBlock();
	}

	void
	DmaReadFifo::resumeFillBypass()
	{
	const size_t fifo_space = buffer.capacity() - buffer.size();
	if (fifo_space >= cacheLineSize \|\| buffer.capacity() < cacheLineSize) {
	const size_t block_remaining = endAddr - nextAddr;
	const size_t xfer_size = std::min(fifo_space, block_remaining);
	std::vector<uint8_t> tmp_buffer(xfer_size);

	assert(pendingRequests.empty());
	DPRINTF(DMA, "Direct bypass startAddr=%#x xfer_size=%#x " \
	"fifo_space=%#x block_remaining=%#x\n",
	nextAddr, xfer_size, fifo_space, block_remaining);

	port.dmaAction(MemCmd::ReadReq, nextAddr, xfer_size, nullptr,
	tmp_buffer.data(), 0, reqFlags);

	buffer.write(tmp_buffer.begin(), xfer_size);
	nextAddr += xfer_size;
	}
	}

	void
	DmaReadFifo::resumeFillTiming()
	{
	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));
	}
	}

	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), _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();
	}

	} // namespace gem5