2 * Copyright (c) 2012 ARM Limited
5 * The license below extends only to copyright in the software and shall
6 * not be construed as granting a license to any other intellectual
7 * property including but not limited to intellectual property relating
8 * to a hardware implementation of the functionality of the software
9 * licensed hereunder. You may use the software subject to the license
10 * terms below provided that you ensure that this notice is replicated
11 * unmodified and in its entirety in all distributions of the software,
12 * modified or unmodified, in source code or in binary form.
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions are
16 * met: redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer;
18 * redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution;
21 * neither the name of the copyright holders nor the names of its
22 * contributors may be used to endorse or promote products derived from
23 * this software without specific prior written permission.
25 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
26 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
27 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
28 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
29 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
30 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
31 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
32 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
33 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
34 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
35 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
37 * Authors: Andreas Sandberg
40 #include <linux/kvm.h>
41 #include <sys/ioctl.h>
49 #include "arch/mmapped_ipr.hh"
50 #include "arch/utility.hh"
51 #include "cpu/kvm/base.hh"
52 #include "debug/Checkpoint.hh"
53 #include "debug/Drain.hh"
54 #include "debug/Kvm.hh"
55 #include "debug/KvmIO.hh"
56 #include "debug/KvmRun.hh"
57 #include "params/BaseKvmCPU.hh"
58 #include "sim/process.hh"
59 #include "sim/system.hh"
63 /* Used by some KVM macros */
64 #define PAGE_SIZE pageSize
66 BaseKvmCPU::BaseKvmCPU(BaseKvmCPUParams
*params
)
70 dataPort(name() + ".dcache_port", this),
71 instPort(name() + ".icache_port", this),
72 threadContextDirty(true),
74 vcpuID(vm
.allocVCPUID()), vcpuFD(-1), vcpuMMapSize(0),
75 _kvmRun(NULL
), mmioRing(NULL
),
76 pageSize(sysconf(_SC_PAGE_SIZE
)),
79 perfControlledByTimer(params
->usePerfOverflow
),
80 hostFactor(params
->hostFactor
),
85 panic("KVM: Failed to determine host page size (%i)\n",
89 thread
= new SimpleThread(this, 0, params
->system
, params
->itb
, params
->dtb
,
92 thread
= new SimpleThread(this, /* thread_num */ 0, params
->system
,
93 params
->workload
[0], params
->itb
,
94 params
->dtb
, params
->isa
[0]);
96 thread
->setStatus(ThreadContext::Halted
);
98 threadContexts
.push_back(tc
);
101 BaseKvmCPU::~BaseKvmCPU()
104 munmap(_kvmRun
, vcpuMMapSize
);
114 fatal("KVM: Multithreading not supported");
116 tc
->initMemProxies(tc
);
118 // initialize CPU, including PC
119 if (FullSystem
&& !switchedOut())
120 TheISA::initCPU(tc
, tc
->contextId());
124 BaseKvmCPU::startup()
126 const BaseKvmCPUParams
* const p(
127 dynamic_cast<const BaseKvmCPUParams
*>(params()));
133 assert(vcpuFD
== -1);
135 // Tell the VM that a CPU is about to start.
138 // We can't initialize KVM CPUs in BaseKvmCPU::init() since we are
139 // not guaranteed that the parent KVM VM has initialized at that
140 // point. Initialize virtual CPUs here instead.
141 vcpuFD
= vm
.createVCPU(vcpuID
);
143 // Map the KVM run structure */
144 vcpuMMapSize
= kvm
.getVCPUMMapSize();
145 _kvmRun
= (struct kvm_run
*)mmap(0, vcpuMMapSize
,
146 PROT_READ
| PROT_WRITE
, MAP_SHARED
,
148 if (_kvmRun
== MAP_FAILED
)
149 panic("KVM: Failed to map run data structure\n");
151 // Setup a pointer to the MMIO ring buffer if coalesced MMIO is
152 // available. The offset into the KVM's communication page is
153 // provided by the coalesced MMIO capability.
154 int mmioOffset(kvm
.capCoalescedMMIO());
155 if (!p
->useCoalescedMMIO
) {
156 inform("KVM: Coalesced MMIO disabled by config.\n");
157 } else if (mmioOffset
) {
158 inform("KVM: Coalesced IO available\n");
159 mmioRing
= (struct kvm_coalesced_mmio_ring
*)(
160 (char *)_kvmRun
+ (mmioOffset
* pageSize
));
162 inform("KVM: Coalesced not supported by host OS\n");
168 new EventWrapper
<BaseKvmCPU
,
169 &BaseKvmCPU::startupThread
>(this, true));
170 schedule(startupEvent
, curTick());
174 BaseKvmCPU::startupThread()
176 // Do thread-specific initialization. We need to setup signal
177 // delivery for counters and timers from within the thread that
178 // will execute the event queue to ensure that signals are
179 // delivered to the right threads.
180 const BaseKvmCPUParams
* const p(
181 dynamic_cast<const BaseKvmCPUParams
*>(params()));
183 vcpuThread
= pthread_self();
185 // Setup signal handlers. This has to be done after the vCPU is
186 // created since it manipulates the vCPU signal mask.
187 setupSignalHandler();
191 if (p
->usePerfOverflow
)
192 runTimer
.reset(new PerfKvmTimer(hwCycles
,
197 runTimer
.reset(new PosixKvmTimer(KVM_KICK_SIGNAL
, CLOCK_MONOTONIC
,
204 BaseKvmCPU::regStats()
206 using namespace Stats
;
211 .name(name() + ".committedInsts")
212 .desc("Number of instructions committed")
216 .name(name() + ".numVMExits")
217 .desc("total number of KVM exits")
221 .name(name() + ".numVMHalfEntries")
222 .desc("number of KVM entries to finalize pending operations")
226 .name(name() + ".numExitSignal")
227 .desc("exits due to signal delivery")
231 .name(name() + ".numMMIO")
232 .desc("number of VM exits due to memory mapped IO")
236 .name(name() + ".numCoalescedMMIO")
237 .desc("number of coalesced memory mapped IO requests")
241 .name(name() + ".numIO")
242 .desc("number of VM exits due to legacy IO")
246 .name(name() + ".numHalt")
247 .desc("number of VM exits due to wait for interrupt instructions")
251 .name(name() + ".numInterrupts")
252 .desc("number of interrupts delivered")
256 .name(name() + ".numHypercalls")
257 .desc("number of hypercalls")
262 BaseKvmCPU::serializeThread(std::ostream
&os
, ThreadID tid
)
264 if (DTRACE(Checkpoint
)) {
265 DPRINTF(Checkpoint
, "KVM: Serializing thread %i:\n", tid
);
270 assert(_status
== Idle
);
271 thread
->serialize(os
);
275 BaseKvmCPU::unserializeThread(Checkpoint
*cp
, const std::string
§ion
,
278 DPRINTF(Checkpoint
, "KVM: Unserialize thread %i:\n", tid
);
281 assert(_status
== Idle
);
282 thread
->unserialize(cp
, section
);
283 threadContextDirty
= true;
287 BaseKvmCPU::drain(DrainManager
*dm
)
292 DPRINTF(Drain
, "BaseKvmCPU::drain\n");
295 // The base KVM code is normally ready when it is in the
296 // Running state, but the architecture specific code might be
297 // of a different opinion. This may happen when the CPU been
298 // notified of an event that hasn't been accepted by the vCPU
300 if (!archIsDrained()) {
305 // The state of the CPU is consistent, so we don't need to do
306 // anything special to drain it. We simply de-schedule the
307 // tick event and enter the Idle state to prevent nasty things
308 // like MMIOs from happening.
309 if (tickEvent
.scheduled())
310 deschedule(tickEvent
);
315 // Idle, no need to drain
316 assert(!tickEvent
.scheduled());
318 // Sync the thread context here since we'll need it when we
319 // switch CPUs or checkpoint the CPU.
324 case RunningServiceCompletion
:
325 // The CPU has just requested a service that was handled in
326 // the RunningService state, but the results have still not
327 // been reported to the CPU. Now, we /could/ probably just
328 // update the register state ourselves instead of letting KVM
329 // handle it, but that would be tricky. Instead, we enter KVM
330 // and let it do its stuff.
333 DPRINTF(Drain
, "KVM CPU is waiting for service completion, "
334 "requesting drain.\n");
338 // We need to drain since the CPU is waiting for service (e.g., MMIOs)
341 DPRINTF(Drain
, "KVM CPU is waiting for service, requesting drain.\n");
345 panic("KVM: Unhandled CPU state in drain()\n");
351 BaseKvmCPU::drainResume()
353 assert(!tickEvent
.scheduled());
355 // We might have been switched out. In that case, we don't need to
360 DPRINTF(Kvm
, "drainResume\n");
363 // The tick event is de-scheduled as a part of the draining
364 // process. Re-schedule it if the thread context is active.
365 if (tc
->status() == ThreadContext::Active
) {
366 schedule(tickEvent
, nextCycle());
374 BaseKvmCPU::switchOut()
376 DPRINTF(Kvm
, "switchOut\n");
378 BaseCPU::switchOut();
380 // We should have drained prior to executing a switchOut, which
381 // means that the tick event shouldn't be scheduled and the CPU is
383 assert(!tickEvent
.scheduled());
384 assert(_status
== Idle
);
388 BaseKvmCPU::takeOverFrom(BaseCPU
*cpu
)
390 DPRINTF(Kvm
, "takeOverFrom\n");
392 BaseCPU::takeOverFrom(cpu
);
394 // We should have drained prior to executing a switchOut, which
395 // means that the tick event shouldn't be scheduled and the CPU is
397 assert(!tickEvent
.scheduled());
398 assert(_status
== Idle
);
399 assert(threadContexts
.size() == 1);
401 // Force an update of the KVM state here instead of flagging the
402 // TC as dirty. This is not ideal from a performance point of
403 // view, but it makes debugging easier as it allows meaningful KVM
404 // state to be dumped before and after a takeover.
406 threadContextDirty
= false;
410 BaseKvmCPU::verifyMemoryMode() const
412 if (!(system
->isAtomicMode() && system
->bypassCaches())) {
413 fatal("The KVM-based CPUs requires the memory system to be in the "
414 "'atomic_noncaching' mode.\n");
421 DPRINTF(Kvm
, "wakeup()\n");
422 // This method might have been called from another
423 // context. Migrate to this SimObject's event queue when
424 // delivering the wakeup signal.
425 EventQueue::ScopedMigration
migrate(eventQueue());
427 // Kick the vCPU to get it to come out of KVM.
430 if (thread
->status() != ThreadContext::Suspended
)
437 BaseKvmCPU::activateContext(ThreadID thread_num
)
439 DPRINTF(Kvm
, "ActivateContext %d\n", thread_num
);
441 assert(thread_num
== 0);
444 assert(_status
== Idle
);
445 assert(!tickEvent
.scheduled());
447 numCycles
+= ticksToCycles(thread
->lastActivate
- thread
->lastSuspend
);
449 schedule(tickEvent
, clockEdge(Cycles(0)));
455 BaseKvmCPU::suspendContext(ThreadID thread_num
)
457 DPRINTF(Kvm
, "SuspendContext %d\n", thread_num
);
459 assert(thread_num
== 0);
465 assert(_status
== Running
|| _status
== RunningServiceCompletion
);
467 // The tick event may no be scheduled if the quest has requested
468 // the monitor to wait for interrupts. The normal CPU models can
469 // get their tick events descheduled by quiesce instructions, but
470 // that can't happen here.
471 if (tickEvent
.scheduled())
472 deschedule(tickEvent
);
478 BaseKvmCPU::deallocateContext(ThreadID thread_num
)
480 // for now, these are equivalent
481 suspendContext(thread_num
);
485 BaseKvmCPU::haltContext(ThreadID thread_num
)
487 // for now, these are equivalent
488 suspendContext(thread_num
);
492 BaseKvmCPU::getContext(int tn
)
501 BaseKvmCPU::totalInsts() const
507 BaseKvmCPU::totalOps() const
509 hack_once("Pretending totalOps is equivalent to totalInsts()\n");
516 inform("State dumping not implemented.");
523 assert(_status
!= Idle
);
527 // handleKvmExit() will determine the next state of the CPU
528 delay
= handleKvmExit();
534 case RunningServiceCompletion
:
536 EventQueue
*q
= curEventQueue();
537 Tick
ticksToExecute(q
->nextTick() - curTick());
539 // We might need to update the KVM state.
542 // Setup any pending instruction count breakpoints using
546 DPRINTF(KvmRun
, "Entering KVM...\n");
548 // Force an immediate exit from KVM after completing
549 // pending operations. The architecture-specific code
550 // takes care to run until it is in a state where it can
551 // safely be drained.
552 delay
= kvmRunDrain();
554 delay
= kvmRun(ticksToExecute
);
557 // The CPU might have been suspended before entering into
558 // KVM. Assume that the CPU was suspended /before/ entering
559 // into KVM and skip the exit handling.
563 // Entering into KVM implies that we'll have to reload the thread
564 // context from KVM if we want to access it. Flag the KVM state as
565 // dirty with respect to the cached thread context.
566 kvmStateDirty
= true;
568 // Enter into the RunningService state unless the
569 // simulation was stopped by a timer.
570 if (_kvmRun
->exit_reason
!= KVM_EXIT_INTR
) {
571 _status
= RunningService
;
577 // Service any pending instruction events. The vCPU should
578 // have exited in time for the event using the instruction
579 // counter configured by setupInstStop().
580 comInstEventQueue
[0]->serviceEvents(ctrInsts
);
581 system
->instEventQueue
.serviceEvents(system
->totalNumInsts
);
588 panic("BaseKvmCPU entered tick() in an illegal state (%i)\n",
592 // Schedule a new tick if we are still running
594 schedule(tickEvent
, clockEdge(ticksToCycles(delay
)));
598 BaseKvmCPU::kvmRunDrain()
600 // By default, the only thing we need to drain is a pending IO
601 // operation which assumes that we are in the
602 // RunningServiceCompletion state.
603 assert(_status
== RunningServiceCompletion
);
605 // Deliver the data from the pending IO operation and immediately
611 BaseKvmCPU::getHostCycles() const
613 return hwCycles
.read();
617 BaseKvmCPU::kvmRun(Tick ticks
)
620 DPRINTF(KvmRun
, "KVM: Executing for %i ticks\n", ticks
);
623 // Settings ticks == 0 is a special case which causes an entry
624 // into KVM that finishes pending operations (e.g., IO) and
625 // then immediately exits.
626 DPRINTF(KvmRun
, "KVM: Delivering IO without full guest entry\n");
630 // Send a KVM_KICK_SIGNAL to the vCPU thread (i.e., this
631 // thread). The KVM control signal is masked while executing
632 // in gem5 and gets unmasked temporarily as when entering
633 // KVM. See setSignalMask() and setupSignalHandler().
636 // Start the vCPU. KVM will check for signals after completing
637 // pending operations (IO). Since the KVM_KICK_SIGNAL is
638 // pending, this forces an immediate exit to gem5 again. We
639 // don't bother to setup timers since this shouldn't actually
640 // execute any code (other than completing half-executed IO
641 // instructions) in the guest.
644 // We always execute at least one cycle to prevent the
645 // BaseKvmCPU::tick() to be rescheduled on the same tick
647 ticksExecuted
= clockPeriod();
649 // This method is executed as a result of a tick event. That
650 // means that the event queue will be locked when entering the
651 // method. We temporarily unlock the event queue to allow
652 // other threads to steal control of this thread to inject
653 // interrupts. They will typically lock the queue and then
654 // force an exit from KVM by kicking the vCPU.
655 EventQueue::ScopedRelease
release(curEventQueue());
657 if (ticks
< runTimer
->resolution()) {
658 DPRINTF(KvmRun
, "KVM: Adjusting tick count (%i -> %i)\n",
659 ticks
, runTimer
->resolution());
660 ticks
= runTimer
->resolution();
663 // Get hardware statistics after synchronizing contexts. The KVM
664 // state update might affect guest cycle counters.
665 uint64_t baseCycles(getHostCycles());
666 uint64_t baseInstrs(hwInstructions
.read());
668 // Arm the run timer and start the cycle timer if it isn't
669 // controlled by the overflow timer. Starting/stopping the cycle
670 // timer automatically starts the other perf timers as they are in
671 // the same counter group.
672 runTimer
->arm(ticks
);
673 if (!perfControlledByTimer
)
679 if (!perfControlledByTimer
)
682 // The control signal may have been delivered after we exited
683 // from KVM. It will be pending in that case since it is
684 // masked when we aren't executing in KVM. Discard it to make
685 // sure we don't deliver it immediately next time we try to
687 discardPendingSignal(KVM_KICK_SIGNAL
);
689 const uint64_t hostCyclesExecuted(getHostCycles() - baseCycles
);
690 const uint64_t simCyclesExecuted(hostCyclesExecuted
* hostFactor
);
691 const uint64_t instsExecuted(hwInstructions
.read() - baseInstrs
);
692 ticksExecuted
= runTimer
->ticksFromHostCycles(hostCyclesExecuted
);
694 /* Update statistics */
695 numCycles
+= simCyclesExecuted
;;
696 numInsts
+= instsExecuted
;
697 ctrInsts
+= instsExecuted
;
698 system
->totalNumInsts
+= instsExecuted
;
701 "KVM: Executed %i instructions in %i cycles "
702 "(%i ticks, sim cycles: %i).\n",
703 instsExecuted
, hostCyclesExecuted
, ticksExecuted
, simCyclesExecuted
);
708 return ticksExecuted
+ flushCoalescedMMIO();
712 BaseKvmCPU::kvmNonMaskableInterrupt()
715 if (ioctl(KVM_NMI
) == -1)
716 panic("KVM: Failed to deliver NMI to virtual CPU\n");
720 BaseKvmCPU::kvmInterrupt(const struct kvm_interrupt
&interrupt
)
723 if (ioctl(KVM_INTERRUPT
, (void *)&interrupt
) == -1)
724 panic("KVM: Failed to deliver interrupt to virtual CPU\n");
728 BaseKvmCPU::getRegisters(struct kvm_regs
®s
) const
730 if (ioctl(KVM_GET_REGS
, ®s
) == -1)
731 panic("KVM: Failed to get guest registers\n");
735 BaseKvmCPU::setRegisters(const struct kvm_regs
®s
)
737 if (ioctl(KVM_SET_REGS
, (void *)®s
) == -1)
738 panic("KVM: Failed to set guest registers\n");
742 BaseKvmCPU::getSpecialRegisters(struct kvm_sregs
®s
) const
744 if (ioctl(KVM_GET_SREGS
, ®s
) == -1)
745 panic("KVM: Failed to get guest special registers\n");
749 BaseKvmCPU::setSpecialRegisters(const struct kvm_sregs
®s
)
751 if (ioctl(KVM_SET_SREGS
, (void *)®s
) == -1)
752 panic("KVM: Failed to set guest special registers\n");
756 BaseKvmCPU::getFPUState(struct kvm_fpu
&state
) const
758 if (ioctl(KVM_GET_FPU
, &state
) == -1)
759 panic("KVM: Failed to get guest FPU state\n");
763 BaseKvmCPU::setFPUState(const struct kvm_fpu
&state
)
765 if (ioctl(KVM_SET_FPU
, (void *)&state
) == -1)
766 panic("KVM: Failed to set guest FPU state\n");
771 BaseKvmCPU::setOneReg(uint64_t id
, const void *addr
)
773 #ifdef KVM_SET_ONE_REG
774 struct kvm_one_reg reg
;
776 reg
.addr
= (uint64_t)addr
;
778 if (ioctl(KVM_SET_ONE_REG
, ®
) == -1) {
779 panic("KVM: Failed to set register (0x%x) value (errno: %i)\n",
783 panic("KVM_SET_ONE_REG is unsupported on this platform.\n");
788 BaseKvmCPU::getOneReg(uint64_t id
, void *addr
) const
790 #ifdef KVM_GET_ONE_REG
791 struct kvm_one_reg reg
;
793 reg
.addr
= (uint64_t)addr
;
795 if (ioctl(KVM_GET_ONE_REG
, ®
) == -1) {
796 panic("KVM: Failed to get register (0x%x) value (errno: %i)\n",
800 panic("KVM_GET_ONE_REG is unsupported on this platform.\n");
805 BaseKvmCPU::getAndFormatOneReg(uint64_t id
) const
807 #ifdef KVM_GET_ONE_REG
808 std::ostringstream ss
;
810 ss
.setf(std::ios::hex
, std::ios::basefield
);
811 ss
.setf(std::ios::showbase
);
812 #define HANDLE_INTTYPE(len) \
813 case KVM_REG_SIZE_U ## len: { \
814 uint ## len ## _t value; \
815 getOneReg(id, &value); \
819 #define HANDLE_ARRAY(len) \
820 case KVM_REG_SIZE_U ## len: { \
821 uint8_t value[len / 8]; \
822 getOneReg(id, value); \
823 ss << "[" << value[0]; \
824 for (int i = 1; i < len / 8; ++i) \
825 ss << ", " << value[i]; \
829 switch (id
& KVM_REG_SIZE_MASK
) {
842 #undef HANDLE_INTTYPE
847 panic("KVM_GET_ONE_REG is unsupported on this platform.\n");
852 BaseKvmCPU::syncThreadContext()
857 assert(!threadContextDirty
);
859 updateThreadContext();
860 kvmStateDirty
= false;
864 BaseKvmCPU::syncKvmState()
866 if (!threadContextDirty
)
869 assert(!kvmStateDirty
);
872 threadContextDirty
= false;
876 BaseKvmCPU::handleKvmExit()
878 DPRINTF(KvmRun
, "handleKvmExit (exit_reason: %i)\n", _kvmRun
->exit_reason
);
879 assert(_status
== RunningService
);
881 // Switch into the running state by default. Individual handlers
882 // can override this.
884 switch (_kvmRun
->exit_reason
) {
885 case KVM_EXIT_UNKNOWN
:
886 return handleKvmExitUnknown();
888 case KVM_EXIT_EXCEPTION
:
889 return handleKvmExitException();
892 _status
= RunningServiceCompletion
;
894 return handleKvmExitIO();
896 case KVM_EXIT_HYPERCALL
:
898 return handleKvmExitHypercall();
901 /* The guest has halted and is waiting for interrupts */
902 DPRINTF(Kvm
, "handleKvmExitHalt\n");
905 // Suspend the thread until the next interrupt arrives
908 // This is actually ignored since the thread is suspended.
912 _status
= RunningServiceCompletion
;
913 /* Service memory mapped IO requests */
914 DPRINTF(KvmIO
, "KVM: Handling MMIO (w: %u, addr: 0x%x, len: %u)\n",
915 _kvmRun
->mmio
.is_write
,
916 _kvmRun
->mmio
.phys_addr
, _kvmRun
->mmio
.len
);
919 return doMMIOAccess(_kvmRun
->mmio
.phys_addr
, _kvmRun
->mmio
.data
,
920 _kvmRun
->mmio
.len
, _kvmRun
->mmio
.is_write
);
922 case KVM_EXIT_IRQ_WINDOW_OPEN
:
923 return handleKvmExitIRQWindowOpen();
925 case KVM_EXIT_FAIL_ENTRY
:
926 return handleKvmExitFailEntry();
929 /* KVM was interrupted by a signal, restart it in the next
933 case KVM_EXIT_INTERNAL_ERROR
:
934 panic("KVM: Internal error (suberror: %u)\n",
935 _kvmRun
->internal
.suberror
);
939 panic("KVM: Unexpected exit (exit_reason: %u)\n", _kvmRun
->exit_reason
);
944 BaseKvmCPU::handleKvmExitIO()
946 panic("KVM: Unhandled guest IO (dir: %i, size: %i, port: 0x%x, count: %i)\n",
947 _kvmRun
->io
.direction
, _kvmRun
->io
.size
,
948 _kvmRun
->io
.port
, _kvmRun
->io
.count
);
952 BaseKvmCPU::handleKvmExitHypercall()
954 panic("KVM: Unhandled hypercall\n");
958 BaseKvmCPU::handleKvmExitIRQWindowOpen()
960 warn("KVM: Unhandled IRQ window.\n");
966 BaseKvmCPU::handleKvmExitUnknown()
969 panic("KVM: Unknown error when starting vCPU (hw reason: 0x%llx)\n",
970 _kvmRun
->hw
.hardware_exit_reason
);
974 BaseKvmCPU::handleKvmExitException()
977 panic("KVM: Got exception when starting vCPU "
978 "(exception: %u, error_code: %u)\n",
979 _kvmRun
->ex
.exception
, _kvmRun
->ex
.error_code
);
983 BaseKvmCPU::handleKvmExitFailEntry()
986 panic("KVM: Failed to enter virtualized mode (hw reason: 0x%llx)\n",
987 _kvmRun
->fail_entry
.hardware_entry_failure_reason
);
991 BaseKvmCPU::doMMIOAccess(Addr paddr
, void *data
, int size
, bool write
)
993 ThreadContext
*tc(thread
->getTC());
996 Request
mmio_req(paddr
, size
, Request::UNCACHEABLE
, dataMasterId());
997 mmio_req
.setThreadContext(tc
->contextId(), 0);
998 // Some architectures do need to massage physical addresses a bit
999 // before they are inserted into the memory system. This enables
1000 // APIC accesses on x86 and m5ops where supported through a MMIO
1002 BaseTLB::Mode
tlb_mode(write
? BaseTLB::Write
: BaseTLB::Read
);
1003 Fault
fault(tc
->getDTBPtr()->finalizePhysical(&mmio_req
, tc
, tlb_mode
));
1004 if (fault
!= NoFault
)
1005 warn("Finalization of MMIO address failed: %s\n", fault
->name());
1008 const MemCmd
cmd(write
? MemCmd::WriteReq
: MemCmd::ReadReq
);
1009 Packet
pkt(&mmio_req
, cmd
);
1010 pkt
.dataStatic(data
);
1012 if (mmio_req
.isMmappedIpr()) {
1013 // We currently assume that there is no need to migrate to a
1014 // different event queue when doing IPRs. Currently, IPRs are
1015 // only used for m5ops, so it should be a valid assumption.
1016 const Cycles
ipr_delay(write
?
1017 TheISA::handleIprWrite(tc
, &pkt
) :
1018 TheISA::handleIprRead(tc
, &pkt
));
1019 threadContextDirty
= true;
1020 return clockPeriod() * ipr_delay
;
1022 // Temporarily lock and migrate to the event queue of the
1023 // VM. This queue is assumed to "own" all devices we need to
1024 // access if running in multi-core mode.
1025 EventQueue::ScopedMigration
migrate(vm
.eventQueue());
1027 return dataPort
.sendAtomic(&pkt
);
1032 BaseKvmCPU::setSignalMask(const sigset_t
*mask
)
1034 std::unique_ptr
<struct kvm_signal_mask
> kvm_mask
;
1037 kvm_mask
.reset((struct kvm_signal_mask
*)operator new(
1038 sizeof(struct kvm_signal_mask
) + sizeof(*mask
)));
1039 // The kernel and the user-space headers have different ideas
1040 // about the size of sigset_t. This seems like a massive hack,
1041 // but is actually what qemu does.
1042 assert(sizeof(*mask
) >= 8);
1044 memcpy(kvm_mask
->sigset
, mask
, kvm_mask
->len
);
1047 if (ioctl(KVM_SET_SIGNAL_MASK
, (void *)kvm_mask
.get()) == -1)
1048 panic("KVM: Failed to set vCPU signal mask (errno: %i)\n",
1053 BaseKvmCPU::ioctl(int request
, long p1
) const
1056 panic("KVM: CPU ioctl called before initialization\n");
1058 return ::ioctl(vcpuFD
, request
, p1
);
1062 BaseKvmCPU::flushCoalescedMMIO()
1067 DPRINTF(KvmIO
, "KVM: Flushing the coalesced MMIO ring buffer\n");
1069 // TODO: We might need to do synchronization when we start to
1070 // support multiple CPUs
1072 while (mmioRing
->first
!= mmioRing
->last
) {
1073 struct kvm_coalesced_mmio
&ent(
1074 mmioRing
->coalesced_mmio
[mmioRing
->first
]);
1076 DPRINTF(KvmIO
, "KVM: Handling coalesced MMIO (addr: 0x%x, len: %u)\n",
1077 ent
.phys_addr
, ent
.len
);
1080 ticks
+= doMMIOAccess(ent
.phys_addr
, ent
.data
, ent
.len
, true);
1082 mmioRing
->first
= (mmioRing
->first
+ 1) % KVM_COALESCED_MMIO_MAX
;
1089 * Dummy handler for KVM kick signals.
1091 * @note This function is usually not called since the kernel doesn't
1092 * seem to deliver signals when the signal is only unmasked when
1093 * running in KVM. This doesn't matter though since we are only
1094 * interested in getting KVM to exit, which happens as expected. See
1095 * setupSignalHandler() and kvmRun() for details about KVM signal
1099 onKickSignal(int signo
, siginfo_t
*si
, void *data
)
1104 BaseKvmCPU::setupSignalHandler()
1106 struct sigaction sa
;
1108 memset(&sa
, 0, sizeof(sa
));
1109 sa
.sa_sigaction
= onKickSignal
;
1110 sa
.sa_flags
= SA_SIGINFO
| SA_RESTART
;
1111 if (sigaction(KVM_KICK_SIGNAL
, &sa
, NULL
) == -1)
1112 panic("KVM: Failed to setup vCPU timer signal handler\n");
1115 if (pthread_sigmask(SIG_BLOCK
, NULL
, &sigset
) == -1)
1116 panic("KVM: Failed get signal mask\n");
1118 // Request KVM to setup the same signal mask as we're currently
1119 // running with except for the KVM control signal. We'll sometimes
1120 // need to raise the KVM_KICK_SIGNAL to cause immediate exits from
1121 // KVM after servicing IO requests. See kvmRun().
1122 sigdelset(&sigset
, KVM_KICK_SIGNAL
);
1123 setSignalMask(&sigset
);
1125 // Mask our control signals so they aren't delivered unless we're
1126 // actually executing inside KVM.
1127 sigaddset(&sigset
, KVM_KICK_SIGNAL
);
1128 if (pthread_sigmask(SIG_SETMASK
, &sigset
, NULL
) == -1)
1129 panic("KVM: Failed mask the KVM control signals\n");
1133 BaseKvmCPU::discardPendingSignal(int signum
) const
1135 int discardedSignal
;
1137 // Setting the timeout to zero causes sigtimedwait to return
1139 struct timespec timeout
;
1141 timeout
.tv_nsec
= 0;
1144 sigemptyset(&sigset
);
1145 sigaddset(&sigset
, signum
);
1148 discardedSignal
= sigtimedwait(&sigset
, NULL
, &timeout
);
1149 } while (discardedSignal
== -1 && errno
== EINTR
);
1151 if (discardedSignal
== signum
)
1153 else if (discardedSignal
== -1 && errno
== EAGAIN
)
1156 panic("Unexpected return value from sigtimedwait: %i (errno: %i)\n",
1157 discardedSignal
, errno
);
1161 BaseKvmCPU::setupCounters()
1163 DPRINTF(Kvm
, "Attaching cycle counter...\n");
1164 PerfKvmCounterConfig
cfgCycles(PERF_TYPE_HARDWARE
,
1165 PERF_COUNT_HW_CPU_CYCLES
);
1166 cfgCycles
.disabled(true)
1169 // Try to exclude the host. We set both exclude_hv and
1170 // exclude_host since different architectures use slightly
1171 // different APIs in the kernel.
1172 cfgCycles
.exclude_hv(true)
1173 .exclude_host(true);
1175 if (perfControlledByTimer
) {
1176 // We need to configure the cycles counter to send overflows
1177 // since we are going to use it to trigger timer signals that
1178 // trap back into m5 from KVM. In practice, this means that we
1179 // need to set some non-zero sample period that gets
1180 // overridden when the timer is armed.
1181 cfgCycles
.wakeupEvents(1)
1185 hwCycles
.attach(cfgCycles
,
1186 0); // TID (0 => currentThread)
1192 BaseKvmCPU::tryDrain()
1197 if (!archIsDrained()) {
1198 DPRINTF(Drain
, "tryDrain: Architecture code is not ready.\n");
1202 if (_status
== Idle
|| _status
== Running
) {
1204 "tryDrain: CPU transitioned into the Idle state, drain done\n");
1205 drainManager
->signalDrainDone();
1206 drainManager
= NULL
;
1209 DPRINTF(Drain
, "tryDrain: CPU not ready.\n");
1215 BaseKvmCPU::ioctlRun()
1217 if (ioctl(KVM_RUN
) == -1) {
1219 panic("KVM: Failed to start virtual CPU (errno: %i)\n",
1225 BaseKvmCPU::setupInstStop()
1227 if (comInstEventQueue
[0]->empty()) {
1228 setupInstCounter(0);
1230 const uint64_t next(comInstEventQueue
[0]->nextTick());
1232 assert(next
> ctrInsts
);
1233 setupInstCounter(next
- ctrInsts
);
1238 BaseKvmCPU::setupInstCounter(uint64_t period
)
1240 // No need to do anything if we aren't attaching for the first
1241 // time or the period isn't changing.
1242 if (period
== activeInstPeriod
&& hwInstructions
.attached())
1245 PerfKvmCounterConfig
cfgInstructions(PERF_TYPE_HARDWARE
,
1246 PERF_COUNT_HW_INSTRUCTIONS
);
1248 // Try to exclude the host. We set both exclude_hv and
1249 // exclude_host since different architectures use slightly
1250 // different APIs in the kernel.
1251 cfgInstructions
.exclude_hv(true)
1252 .exclude_host(true);
1255 // Setup a sampling counter if that has been requested.
1256 cfgInstructions
.wakeupEvents(1)
1257 .samplePeriod(period
);
1260 // We need to detach and re-attach the counter to reliably change
1261 // sampling settings. See PerfKvmCounter::period() for details.
1262 if (hwInstructions
.attached())
1263 hwInstructions
.detach();
1264 assert(hwCycles
.attached());
1265 hwInstructions
.attach(cfgInstructions
,
1266 0, // TID (0 => currentThread)
1270 hwInstructions
.enableSignals(KVM_KICK_SIGNAL
);
1272 activeInstPeriod
= period
;