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 volatile bool timerOverflowed
= false;
68 BaseKvmCPU::BaseKvmCPU(BaseKvmCPUParams
*params
)
72 dataPort(name() + ".dcache_port", this),
73 instPort(name() + ".icache_port", this),
74 threadContextDirty(true),
76 vcpuID(vm
.allocVCPUID()), vcpuFD(-1), vcpuMMapSize(0),
77 _kvmRun(NULL
), mmioRing(NULL
),
78 pageSize(sysconf(_SC_PAGE_SIZE
)),
81 perfControlledByTimer(params
->usePerfOverflow
),
82 hostFreq(params
->hostFreq
),
83 hostFactor(params
->hostFactor
),
88 panic("KVM: Failed to determine host page size (%i)\n",
91 thread
= new SimpleThread(this, 0, params
->system
,
92 params
->itb
, params
->dtb
, params
->isa
[0]);
93 thread
->setStatus(ThreadContext::Halted
);
95 threadContexts
.push_back(tc
);
99 if (params
->usePerfOverflow
)
100 runTimer
.reset(new PerfKvmTimer(hwCycles
,
105 runTimer
.reset(new PosixKvmTimer(KVM_TIMER_SIGNAL
, CLOCK_MONOTONIC
,
110 BaseKvmCPU::~BaseKvmCPU()
113 munmap(_kvmRun
, vcpuMMapSize
);
123 fatal("KVM: Multithreading not supported");
125 tc
->initMemProxies(tc
);
127 // initialize CPU, including PC
128 if (FullSystem
&& !switchedOut())
129 TheISA::initCPU(tc
, tc
->contextId());
131 mmio_req
.setThreadContext(tc
->contextId(), 0);
135 BaseKvmCPU::startup()
137 const BaseKvmCPUParams
* const p(
138 dynamic_cast<const BaseKvmCPUParams
*>(params()));
144 assert(vcpuFD
== -1);
146 // Tell the VM that a CPU is about to start.
149 // We can't initialize KVM CPUs in BaseKvmCPU::init() since we are
150 // not guaranteed that the parent KVM VM has initialized at that
151 // point. Initialize virtual CPUs here instead.
152 vcpuFD
= vm
.createVCPU(vcpuID
);
154 // Setup signal handlers. This has to be done after the vCPU is
155 // created since it manipulates the vCPU signal mask.
156 setupSignalHandler();
158 // Map the KVM run structure */
159 vcpuMMapSize
= kvm
.getVCPUMMapSize();
160 _kvmRun
= (struct kvm_run
*)mmap(0, vcpuMMapSize
,
161 PROT_READ
| PROT_WRITE
, MAP_SHARED
,
163 if (_kvmRun
== MAP_FAILED
)
164 panic("KVM: Failed to map run data structure\n");
166 // Setup a pointer to the MMIO ring buffer if coalesced MMIO is
167 // available. The offset into the KVM's communication page is
168 // provided by the coalesced MMIO capability.
169 int mmioOffset(kvm
.capCoalescedMMIO());
170 if (!p
->useCoalescedMMIO
) {
171 inform("KVM: Coalesced MMIO disabled by config.\n");
172 } else if (mmioOffset
) {
173 inform("KVM: Coalesced IO available\n");
174 mmioRing
= (struct kvm_coalesced_mmio_ring
*)(
175 (char *)_kvmRun
+ (mmioOffset
* pageSize
));
177 inform("KVM: Coalesced not supported by host OS\n");
184 BaseKvmCPU::regStats()
186 using namespace Stats
;
191 .name(name() + ".committedInsts")
192 .desc("Number of instructions committed")
196 .name(name() + ".numVMExits")
197 .desc("total number of KVM exits")
201 .name(name() + ".numVMHalfEntries")
202 .desc("number of KVM entries to finalize pending operations")
206 .name(name() + ".numExitSignal")
207 .desc("exits due to signal delivery")
211 .name(name() + ".numMMIO")
212 .desc("number of VM exits due to memory mapped IO")
216 .name(name() + ".numCoalescedMMIO")
217 .desc("number of coalesced memory mapped IO requests")
221 .name(name() + ".numIO")
222 .desc("number of VM exits due to legacy IO")
226 .name(name() + ".numHalt")
227 .desc("number of VM exits due to wait for interrupt instructions")
231 .name(name() + ".numInterrupts")
232 .desc("number of interrupts delivered")
236 .name(name() + ".numHypercalls")
237 .desc("number of hypercalls")
242 BaseKvmCPU::serializeThread(std::ostream
&os
, ThreadID tid
)
244 if (DTRACE(Checkpoint
)) {
245 DPRINTF(Checkpoint
, "KVM: Serializing thread %i:\n", tid
);
250 assert(_status
== Idle
);
251 thread
->serialize(os
);
255 BaseKvmCPU::unserializeThread(Checkpoint
*cp
, const std::string
§ion
,
258 DPRINTF(Checkpoint
, "KVM: Unserialize thread %i:\n", tid
);
261 assert(_status
== Idle
);
262 thread
->unserialize(cp
, section
);
263 threadContextDirty
= true;
267 BaseKvmCPU::drain(DrainManager
*dm
)
272 DPRINTF(Drain
, "BaseKvmCPU::drain\n");
275 // The base KVM code is normally ready when it is in the
276 // Running state, but the architecture specific code might be
277 // of a different opinion. This may happen when the CPU been
278 // notified of an event that hasn't been accepted by the vCPU
280 if (!archIsDrained()) {
285 // The state of the CPU is consistent, so we don't need to do
286 // anything special to drain it. We simply de-schedule the
287 // tick event and enter the Idle state to prevent nasty things
288 // like MMIOs from happening.
289 if (tickEvent
.scheduled())
290 deschedule(tickEvent
);
295 // Idle, no need to drain
296 assert(!tickEvent
.scheduled());
298 // Sync the thread context here since we'll need it when we
299 // switch CPUs or checkpoint the CPU.
304 case RunningServiceCompletion
:
305 // The CPU has just requested a service that was handled in
306 // the RunningService state, but the results have still not
307 // been reported to the CPU. Now, we /could/ probably just
308 // update the register state ourselves instead of letting KVM
309 // handle it, but that would be tricky. Instead, we enter KVM
310 // and let it do its stuff.
313 DPRINTF(Drain
, "KVM CPU is waiting for service completion, "
314 "requesting drain.\n");
318 // We need to drain since the CPU is waiting for service (e.g., MMIOs)
321 DPRINTF(Drain
, "KVM CPU is waiting for service, requesting drain.\n");
325 panic("KVM: Unhandled CPU state in drain()\n");
331 BaseKvmCPU::drainResume()
333 assert(!tickEvent
.scheduled());
335 // We might have been switched out. In that case, we don't need to
340 DPRINTF(Kvm
, "drainResume\n");
343 // The tick event is de-scheduled as a part of the draining
344 // process. Re-schedule it if the thread context is active.
345 if (tc
->status() == ThreadContext::Active
) {
346 schedule(tickEvent
, nextCycle());
354 BaseKvmCPU::switchOut()
356 DPRINTF(Kvm
, "switchOut\n");
358 BaseCPU::switchOut();
360 // We should have drained prior to executing a switchOut, which
361 // means that the tick event shouldn't be scheduled and the CPU is
363 assert(!tickEvent
.scheduled());
364 assert(_status
== Idle
);
368 BaseKvmCPU::takeOverFrom(BaseCPU
*cpu
)
370 DPRINTF(Kvm
, "takeOverFrom\n");
372 BaseCPU::takeOverFrom(cpu
);
374 // We should have drained prior to executing a switchOut, which
375 // means that the tick event shouldn't be scheduled and the CPU is
377 assert(!tickEvent
.scheduled());
378 assert(_status
== Idle
);
379 assert(threadContexts
.size() == 1);
381 // Force an update of the KVM state here instead of flagging the
382 // TC as dirty. This is not ideal from a performance point of
383 // view, but it makes debugging easier as it allows meaningful KVM
384 // state to be dumped before and after a takeover.
386 threadContextDirty
= false;
390 BaseKvmCPU::verifyMemoryMode() const
392 if (!(system
->isAtomicMode() && system
->bypassCaches())) {
393 fatal("The KVM-based CPUs requires the memory system to be in the "
394 "'atomic_noncaching' mode.\n");
401 DPRINTF(Kvm
, "wakeup()\n");
403 if (thread
->status() != ThreadContext::Suspended
)
410 BaseKvmCPU::activateContext(ThreadID thread_num
, Cycles delay
)
412 DPRINTF(Kvm
, "ActivateContext %d (%d cycles)\n", thread_num
, delay
);
414 assert(thread_num
== 0);
417 assert(_status
== Idle
);
418 assert(!tickEvent
.scheduled());
420 numCycles
+= ticksToCycles(thread
->lastActivate
- thread
->lastSuspend
);
422 schedule(tickEvent
, clockEdge(delay
));
428 BaseKvmCPU::suspendContext(ThreadID thread_num
)
430 DPRINTF(Kvm
, "SuspendContext %d\n", thread_num
);
432 assert(thread_num
== 0);
438 assert(_status
== Running
);
440 // The tick event may no be scheduled if the quest has requested
441 // the monitor to wait for interrupts. The normal CPU models can
442 // get their tick events descheduled by quiesce instructions, but
443 // that can't happen here.
444 if (tickEvent
.scheduled())
445 deschedule(tickEvent
);
451 BaseKvmCPU::deallocateContext(ThreadID thread_num
)
453 // for now, these are equivalent
454 suspendContext(thread_num
);
458 BaseKvmCPU::haltContext(ThreadID thread_num
)
460 // for now, these are equivalent
461 suspendContext(thread_num
);
465 BaseKvmCPU::getContext(int tn
)
474 BaseKvmCPU::totalInsts() const
480 BaseKvmCPU::totalOps() const
482 hack_once("Pretending totalOps is equivalent to totalInsts()\n");
489 inform("State dumping not implemented.");
496 assert(_status
!= Idle
);
500 // handleKvmExit() will determine the next state of the CPU
501 delay
= handleKvmExit();
507 case RunningServiceCompletion
:
509 Tick
ticksToExecute(mainEventQueue
.nextTick() - curTick());
511 // We might need to update the KVM state.
514 // Setup any pending instruction count breakpoints using
518 DPRINTF(KvmRun
, "Entering KVM...\n");
520 // Force an immediate exit from KVM after completing
521 // pending operations. The architecture-specific code
522 // takes care to run until it is in a state where it can
523 // safely be drained.
524 delay
= kvmRunDrain();
526 delay
= kvmRun(ticksToExecute
);
529 // Entering into KVM implies that we'll have to reload the thread
530 // context from KVM if we want to access it. Flag the KVM state as
531 // dirty with respect to the cached thread context.
532 kvmStateDirty
= true;
534 // Enter into the RunningService state unless the
535 // simulation was stopped by a timer.
536 if (_kvmRun
->exit_reason
!= KVM_EXIT_INTR
) {
537 _status
= RunningService
;
548 panic("BaseKvmCPU entered tick() in an illegal state (%i)\n",
552 // Schedule a new tick if we are still running
554 schedule(tickEvent
, clockEdge(ticksToCycles(delay
)));
558 BaseKvmCPU::kvmRunDrain()
560 // By default, the only thing we need to drain is a pending IO
561 // operation which assumes that we are in the
562 // RunningServiceCompletion state.
563 assert(_status
== RunningServiceCompletion
);
565 // Deliver the data from the pending IO operation and immediately
571 BaseKvmCPU::getHostCycles() const
573 return hwCycles
.read();
577 BaseKvmCPU::kvmRun(Tick ticks
)
580 DPRINTF(KvmRun
, "KVM: Executing for %i ticks\n", ticks
);
581 timerOverflowed
= false;
584 // Settings ticks == 0 is a special case which causes an entry
585 // into KVM that finishes pending operations (e.g., IO) and
586 // then immediately exits.
587 DPRINTF(KvmRun
, "KVM: Delivering IO without full guest entry\n");
591 // This signal is always masked while we are executing in gem5
592 // and gets unmasked temporarily as soon as we enter into
593 // KVM. See setSignalMask() and setupSignalHandler().
594 raise(KVM_TIMER_SIGNAL
);
596 // Enter into KVM. KVM will check for signals after completing
597 // pending operations (IO). Since the KVM_TIMER_SIGNAL is
598 // pending, this forces an immediate exit into gem5 again. We
599 // don't bother to setup timers since this shouldn't actually
600 // execute any code in the guest.
603 // We always execute at least one cycle to prevent the
604 // BaseKvmCPU::tick() to be rescheduled on the same tick
606 ticksExecuted
= clockPeriod();
608 if (ticks
< runTimer
->resolution()) {
609 DPRINTF(KvmRun
, "KVM: Adjusting tick count (%i -> %i)\n",
610 ticks
, runTimer
->resolution());
611 ticks
= runTimer
->resolution();
614 // Get hardware statistics after synchronizing contexts. The KVM
615 // state update might affect guest cycle counters.
616 uint64_t baseCycles(getHostCycles());
617 uint64_t baseInstrs(hwInstructions
.read());
619 // Arm the run timer and start the cycle timer if it isn't
620 // controlled by the overflow timer. Starting/stopping the cycle
621 // timer automatically starts the other perf timers as they are in
622 // the same counter group.
623 runTimer
->arm(ticks
);
624 if (!perfControlledByTimer
)
630 if (!perfControlledByTimer
)
633 // The timer signal may have been delivered after we exited
634 // from KVM. It will be pending in that case since it is
635 // masked when we aren't executing in KVM. Discard it to make
636 // sure we don't deliver it immediately next time we try to
638 discardPendingSignal(KVM_TIMER_SIGNAL
);
639 discardPendingSignal(KVM_INST_SIGNAL
);
641 const uint64_t hostCyclesExecuted(getHostCycles() - baseCycles
);
642 const uint64_t simCyclesExecuted(hostCyclesExecuted
* hostFactor
);
643 const uint64_t instsExecuted(hwInstructions
.read() - baseInstrs
);
644 ticksExecuted
= runTimer
->ticksFromHostCycles(hostCyclesExecuted
);
646 if (ticksExecuted
< ticks
&&
648 _kvmRun
->exit_reason
== KVM_EXIT_INTR
) {
649 // TODO: We should probably do something clever here...
650 warn("KVM: Early timer event, requested %i ticks but got %i ticks.\n",
651 ticks
, ticksExecuted
);
654 /* Update statistics */
655 numCycles
+= simCyclesExecuted
;;
656 numInsts
+= instsExecuted
;
657 ctrInsts
+= instsExecuted
;
658 system
->totalNumInsts
+= instsExecuted
;
661 "KVM: Executed %i instructions in %i cycles "
662 "(%i ticks, sim cycles: %i).\n",
663 instsExecuted
, hostCyclesExecuted
, ticksExecuted
, simCyclesExecuted
);
668 return ticksExecuted
+ flushCoalescedMMIO();
672 BaseKvmCPU::kvmNonMaskableInterrupt()
675 if (ioctl(KVM_NMI
) == -1)
676 panic("KVM: Failed to deliver NMI to virtual CPU\n");
680 BaseKvmCPU::kvmInterrupt(const struct kvm_interrupt
&interrupt
)
683 if (ioctl(KVM_INTERRUPT
, (void *)&interrupt
) == -1)
684 panic("KVM: Failed to deliver interrupt to virtual CPU\n");
688 BaseKvmCPU::getRegisters(struct kvm_regs
®s
) const
690 if (ioctl(KVM_GET_REGS
, ®s
) == -1)
691 panic("KVM: Failed to get guest registers\n");
695 BaseKvmCPU::setRegisters(const struct kvm_regs
®s
)
697 if (ioctl(KVM_SET_REGS
, (void *)®s
) == -1)
698 panic("KVM: Failed to set guest registers\n");
702 BaseKvmCPU::getSpecialRegisters(struct kvm_sregs
®s
) const
704 if (ioctl(KVM_GET_SREGS
, ®s
) == -1)
705 panic("KVM: Failed to get guest special registers\n");
709 BaseKvmCPU::setSpecialRegisters(const struct kvm_sregs
®s
)
711 if (ioctl(KVM_SET_SREGS
, (void *)®s
) == -1)
712 panic("KVM: Failed to set guest special registers\n");
716 BaseKvmCPU::getFPUState(struct kvm_fpu
&state
) const
718 if (ioctl(KVM_GET_FPU
, &state
) == -1)
719 panic("KVM: Failed to get guest FPU state\n");
723 BaseKvmCPU::setFPUState(const struct kvm_fpu
&state
)
725 if (ioctl(KVM_SET_FPU
, (void *)&state
) == -1)
726 panic("KVM: Failed to set guest FPU state\n");
731 BaseKvmCPU::setOneReg(uint64_t id
, const void *addr
)
733 #ifdef KVM_SET_ONE_REG
734 struct kvm_one_reg reg
;
736 reg
.addr
= (uint64_t)addr
;
738 if (ioctl(KVM_SET_ONE_REG
, ®
) == -1) {
739 panic("KVM: Failed to set register (0x%x) value (errno: %i)\n",
743 panic("KVM_SET_ONE_REG is unsupported on this platform.\n");
748 BaseKvmCPU::getOneReg(uint64_t id
, void *addr
) const
750 #ifdef KVM_GET_ONE_REG
751 struct kvm_one_reg reg
;
753 reg
.addr
= (uint64_t)addr
;
755 if (ioctl(KVM_GET_ONE_REG
, ®
) == -1) {
756 panic("KVM: Failed to get register (0x%x) value (errno: %i)\n",
760 panic("KVM_GET_ONE_REG is unsupported on this platform.\n");
765 BaseKvmCPU::getAndFormatOneReg(uint64_t id
) const
767 #ifdef KVM_GET_ONE_REG
768 std::ostringstream ss
;
770 ss
.setf(std::ios::hex
, std::ios::basefield
);
771 ss
.setf(std::ios::showbase
);
772 #define HANDLE_INTTYPE(len) \
773 case KVM_REG_SIZE_U ## len: { \
774 uint ## len ## _t value; \
775 getOneReg(id, &value); \
779 #define HANDLE_ARRAY(len) \
780 case KVM_REG_SIZE_U ## len: { \
781 uint8_t value[len / 8]; \
782 getOneReg(id, value); \
783 ss << "[" << value[0]; \
784 for (int i = 1; i < len / 8; ++i) \
785 ss << ", " << value[i]; \
789 switch (id
& KVM_REG_SIZE_MASK
) {
802 #undef HANDLE_INTTYPE
807 panic("KVM_GET_ONE_REG is unsupported on this platform.\n");
812 BaseKvmCPU::syncThreadContext()
817 assert(!threadContextDirty
);
819 updateThreadContext();
820 kvmStateDirty
= false;
824 BaseKvmCPU::syncKvmState()
826 if (!threadContextDirty
)
829 assert(!kvmStateDirty
);
832 threadContextDirty
= false;
836 BaseKvmCPU::handleKvmExit()
838 DPRINTF(KvmRun
, "handleKvmExit (exit_reason: %i)\n", _kvmRun
->exit_reason
);
839 assert(_status
== RunningService
);
841 // Switch into the running state by default. Individual handlers
842 // can override this.
844 switch (_kvmRun
->exit_reason
) {
845 case KVM_EXIT_UNKNOWN
:
846 return handleKvmExitUnknown();
848 case KVM_EXIT_EXCEPTION
:
849 return handleKvmExitException();
852 _status
= RunningServiceCompletion
;
854 return handleKvmExitIO();
856 case KVM_EXIT_HYPERCALL
:
858 return handleKvmExitHypercall();
861 /* The guest has halted and is waiting for interrupts */
862 DPRINTF(Kvm
, "handleKvmExitHalt\n");
865 // Suspend the thread until the next interrupt arrives
868 // This is actually ignored since the thread is suspended.
872 _status
= RunningServiceCompletion
;
873 /* Service memory mapped IO requests */
874 DPRINTF(KvmIO
, "KVM: Handling MMIO (w: %u, addr: 0x%x, len: %u)\n",
875 _kvmRun
->mmio
.is_write
,
876 _kvmRun
->mmio
.phys_addr
, _kvmRun
->mmio
.len
);
879 return doMMIOAccess(_kvmRun
->mmio
.phys_addr
, _kvmRun
->mmio
.data
,
880 _kvmRun
->mmio
.len
, _kvmRun
->mmio
.is_write
);
882 case KVM_EXIT_IRQ_WINDOW_OPEN
:
883 return handleKvmExitIRQWindowOpen();
885 case KVM_EXIT_FAIL_ENTRY
:
886 return handleKvmExitFailEntry();
889 /* KVM was interrupted by a signal, restart it in the next
893 case KVM_EXIT_INTERNAL_ERROR
:
894 panic("KVM: Internal error (suberror: %u)\n",
895 _kvmRun
->internal
.suberror
);
899 panic("KVM: Unexpected exit (exit_reason: %u)\n", _kvmRun
->exit_reason
);
904 BaseKvmCPU::handleKvmExitIO()
906 panic("KVM: Unhandled guest IO (dir: %i, size: %i, port: 0x%x, count: %i)\n",
907 _kvmRun
->io
.direction
, _kvmRun
->io
.size
,
908 _kvmRun
->io
.port
, _kvmRun
->io
.count
);
912 BaseKvmCPU::handleKvmExitHypercall()
914 panic("KVM: Unhandled hypercall\n");
918 BaseKvmCPU::handleKvmExitIRQWindowOpen()
920 warn("KVM: Unhandled IRQ window.\n");
926 BaseKvmCPU::handleKvmExitUnknown()
929 panic("KVM: Unknown error when starting vCPU (hw reason: 0x%llx)\n",
930 _kvmRun
->hw
.hardware_exit_reason
);
934 BaseKvmCPU::handleKvmExitException()
937 panic("KVM: Got exception when starting vCPU "
938 "(exception: %u, error_code: %u)\n",
939 _kvmRun
->ex
.exception
, _kvmRun
->ex
.error_code
);
943 BaseKvmCPU::handleKvmExitFailEntry()
946 panic("KVM: Failed to enter virtualized mode (hw reason: 0x%llx)\n",
947 _kvmRun
->fail_entry
.hardware_entry_failure_reason
);
951 BaseKvmCPU::doMMIOAccess(Addr paddr
, void *data
, int size
, bool write
)
953 ThreadContext
*tc(thread
->getTC());
956 mmio_req
.setPhys(paddr
, size
, Request::UNCACHEABLE
, dataMasterId());
957 // Some architectures do need to massage physical addresses a bit
958 // before they are inserted into the memory system. This enables
959 // APIC accesses on x86 and m5ops where supported through a MMIO
961 BaseTLB::Mode
tlb_mode(write
? BaseTLB::Write
: BaseTLB::Read
);
962 Fault
fault(tc
->getDTBPtr()->finalizePhysical(&mmio_req
, tc
, tlb_mode
));
963 if (fault
!= NoFault
)
964 warn("Finalization of MMIO address failed: %s\n", fault
->name());
967 const MemCmd
cmd(write
? MemCmd::WriteReq
: MemCmd::ReadReq
);
968 Packet
pkt(&mmio_req
, cmd
);
969 pkt
.dataStatic(data
);
971 if (mmio_req
.isMmappedIpr()) {
972 const Cycles
ipr_delay(write
?
973 TheISA::handleIprWrite(tc
, &pkt
) :
974 TheISA::handleIprRead(tc
, &pkt
));
975 return clockEdge(ipr_delay
);
977 return dataPort
.sendAtomic(&pkt
);
982 BaseKvmCPU::setSignalMask(const sigset_t
*mask
)
984 std::unique_ptr
<struct kvm_signal_mask
> kvm_mask
;
987 kvm_mask
.reset((struct kvm_signal_mask
*)operator new(
988 sizeof(struct kvm_signal_mask
) + sizeof(*mask
)));
989 // The kernel and the user-space headers have different ideas
990 // about the size of sigset_t. This seems like a massive hack,
991 // but is actually what qemu does.
992 assert(sizeof(*mask
) >= 8);
994 memcpy(kvm_mask
->sigset
, mask
, kvm_mask
->len
);
997 if (ioctl(KVM_SET_SIGNAL_MASK
, (void *)kvm_mask
.get()) == -1)
998 panic("KVM: Failed to set vCPU signal mask (errno: %i)\n",
1003 BaseKvmCPU::ioctl(int request
, long p1
) const
1006 panic("KVM: CPU ioctl called before initialization\n");
1008 return ::ioctl(vcpuFD
, request
, p1
);
1012 BaseKvmCPU::flushCoalescedMMIO()
1017 DPRINTF(KvmIO
, "KVM: Flushing the coalesced MMIO ring buffer\n");
1019 // TODO: We might need to do synchronization when we start to
1020 // support multiple CPUs
1022 while (mmioRing
->first
!= mmioRing
->last
) {
1023 struct kvm_coalesced_mmio
&ent(
1024 mmioRing
->coalesced_mmio
[mmioRing
->first
]);
1026 DPRINTF(KvmIO
, "KVM: Handling coalesced MMIO (addr: 0x%x, len: %u)\n",
1027 ent
.phys_addr
, ent
.len
);
1030 ticks
+= doMMIOAccess(ent
.phys_addr
, ent
.data
, ent
.len
, true);
1032 mmioRing
->first
= (mmioRing
->first
+ 1) % KVM_COALESCED_MMIO_MAX
;
1039 * Cycle timer overflow when running in KVM. Forces the KVM syscall to
1040 * exit with EINTR and allows us to run the event queue.
1043 onTimerOverflow(int signo
, siginfo_t
*si
, void *data
)
1045 timerOverflowed
= true;
1049 * Instruction counter overflow when running in KVM. Forces the KVM
1050 * syscall to exit with EINTR and allows us to handle instruction
1054 onInstEvent(int signo
, siginfo_t
*si
, void *data
)
1059 BaseKvmCPU::setupSignalHandler()
1061 struct sigaction sa
;
1063 memset(&sa
, 0, sizeof(sa
));
1064 sa
.sa_sigaction
= onTimerOverflow
;
1065 sa
.sa_flags
= SA_SIGINFO
| SA_RESTART
;
1066 if (sigaction(KVM_TIMER_SIGNAL
, &sa
, NULL
) == -1)
1067 panic("KVM: Failed to setup vCPU timer signal handler\n");
1069 memset(&sa
, 0, sizeof(sa
));
1070 sa
.sa_sigaction
= onInstEvent
;
1071 sa
.sa_flags
= SA_SIGINFO
| SA_RESTART
;
1072 if (sigaction(KVM_INST_SIGNAL
, &sa
, NULL
) == -1)
1073 panic("KVM: Failed to setup vCPU instruction signal handler\n");
1076 if (sigprocmask(SIG_BLOCK
, NULL
, &sigset
) == -1)
1077 panic("KVM: Failed get signal mask\n");
1079 // Request KVM to setup the same signal mask as we're currently
1080 // running with. We'll sometimes need to mask the KVM_TIMER_SIGNAL
1081 // to cause immediate exits from KVM after servicing IO
1082 // requests. See kvmRun().
1083 setSignalMask(&sigset
);
1085 // Mask our control signals so they aren't delivered unless we're
1086 // actually executing inside KVM.
1087 sigaddset(&sigset
, KVM_TIMER_SIGNAL
);
1088 sigaddset(&sigset
, KVM_INST_SIGNAL
);
1089 if (sigprocmask(SIG_SETMASK
, &sigset
, NULL
) == -1)
1090 panic("KVM: Failed mask the KVM control signals\n");
1094 BaseKvmCPU::discardPendingSignal(int signum
) const
1096 int discardedSignal
;
1098 // Setting the timeout to zero causes sigtimedwait to return
1100 struct timespec timeout
;
1102 timeout
.tv_nsec
= 0;
1105 sigemptyset(&sigset
);
1106 sigaddset(&sigset
, signum
);
1109 discardedSignal
= sigtimedwait(&sigset
, NULL
, &timeout
);
1110 } while (discardedSignal
== -1 && errno
== EINTR
);
1112 if (discardedSignal
== signum
)
1114 else if (discardedSignal
== -1 && errno
== EAGAIN
)
1117 panic("Unexpected return value from sigtimedwait: %i (errno: %i)\n",
1118 discardedSignal
, errno
);
1122 BaseKvmCPU::setupCounters()
1124 DPRINTF(Kvm
, "Attaching cycle counter...\n");
1125 PerfKvmCounterConfig
cfgCycles(PERF_TYPE_HARDWARE
,
1126 PERF_COUNT_HW_CPU_CYCLES
);
1127 cfgCycles
.disabled(true)
1130 if (perfControlledByTimer
) {
1131 // We need to configure the cycles counter to send overflows
1132 // since we are going to use it to trigger timer signals that
1133 // trap back into m5 from KVM. In practice, this means that we
1134 // need to set some non-zero sample period that gets
1135 // overridden when the timer is armed.
1136 cfgCycles
.wakeupEvents(1)
1140 hwCycles
.attach(cfgCycles
,
1141 0); // TID (0 => currentThread)
1147 BaseKvmCPU::tryDrain()
1152 if (!archIsDrained()) {
1153 DPRINTF(Drain
, "tryDrain: Architecture code is not ready.\n");
1157 if (_status
== Idle
|| _status
== Running
) {
1159 "tryDrain: CPU transitioned into the Idle state, drain done\n");
1160 drainManager
->signalDrainDone();
1161 drainManager
= NULL
;
1164 DPRINTF(Drain
, "tryDrain: CPU not ready.\n");
1170 BaseKvmCPU::ioctlRun()
1172 if (ioctl(KVM_RUN
) == -1) {
1174 panic("KVM: Failed to start virtual CPU (errno: %i)\n",
1180 BaseKvmCPU::setupInstStop()
1183 if (comInstEventQueue
[0]->empty()) {
1184 setupInstCounter(0);
1186 const uint64_t next(comInstEventQueue
[0]->nextTick());
1188 assert(next
> ctrInsts
);
1189 setupInstCounter(next
- ctrInsts
);
1194 BaseKvmCPU::setupInstCounter(uint64_t period
)
1196 // No need to do anything if we aren't attaching for the first
1197 // time or the period isn't changing.
1198 if (period
== activeInstPeriod
&& hwInstructions
.attached())
1201 PerfKvmCounterConfig
cfgInstructions(PERF_TYPE_HARDWARE
,
1202 PERF_COUNT_HW_INSTRUCTIONS
);
1205 // Setup a sampling counter if that has been requested.
1206 cfgInstructions
.wakeupEvents(1)
1207 .samplePeriod(period
);
1210 // We need to detach and re-attach the counter to reliably change
1211 // sampling settings. See PerfKvmCounter::period() for details.
1212 if (hwInstructions
.attached())
1213 hwInstructions
.detach();
1214 assert(hwCycles
.attached());
1215 hwInstructions
.attach(cfgInstructions
,
1216 0, // TID (0 => currentThread)
1220 hwInstructions
.enableSignals(KVM_INST_SIGNAL
);
1222 activeInstPeriod
= period
;