3 * All rights reserved
4 *
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.
13 *
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.
24 *
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.
36 *
37 * Authors: Andreas Sandberg
38 */
39
40#include <linux/kvm.h>
41#include <sys/ioctl.h>
42#include <sys/mman.h>
43#include <unistd.h>
44
45#include <cerrno>
46#include <csignal>
47#include <ostream>
48
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"
60
61#include <signal.h>
62
63/* Used by some KVM macros */
64#define PAGE_SIZE pageSize
65
66BaseKvmCPU::BaseKvmCPU(BaseKvmCPUParams *params)
67 : BaseCPU(params),
68 vm(*params->kvmVM),
69 _status(Idle),
70 dataPort(name() + ".dcache_port", this),
71 instPort(name() + ".icache_port", this),
72 threadContextDirty(true),
73 kvmStateDirty(false),
74 vcpuID(vm.allocVCPUID()), vcpuFD(-1), vcpuMMapSize(0),
75 _kvmRun(NULL), mmioRing(NULL),
76 pageSize(sysconf(_SC_PAGE_SIZE)),
77 tickEvent(*this),
78 activeInstPeriod(0),
79 perfControlledByTimer(params->usePerfOverflow),
80 hostFactor(params->hostFactor),
81 drainManager(NULL),
82 ctrInsts(0)
83{
84 if (pageSize == -1)
85 panic("KVM: Failed to determine host page size (%i)\n",
86 errno);
87
88 if (FullSystem)
89 thread = new SimpleThread(this, 0, params->system, params->itb, params->dtb,
90 params->isa[0]);
91 else
92 thread = new SimpleThread(this, /* thread_num */ 0, params->system,
93 params->workload[0], params->itb,
94 params->dtb, params->isa[0]);
95
96 thread->setStatus(ThreadContext::Halted);
97 tc = thread->getTC();
98 threadContexts.push_back(tc);
99}
100
101BaseKvmCPU::~BaseKvmCPU()
102{
103 if (_kvmRun)
104 munmap(_kvmRun, vcpuMMapSize);
105 close(vcpuFD);
106}
107
108void
109BaseKvmCPU::init()
110{
111 BaseCPU::init();
112
113 if (numThreads != 1)
114 fatal("KVM: Multithreading not supported");
115
116 tc->initMemProxies(tc);
117
118 // initialize CPU, including PC
119 if (FullSystem && !switchedOut())
120 TheISA::initCPU(tc, tc->contextId());
121}
122
123void
124BaseKvmCPU::startup()
125{
126 const BaseKvmCPUParams * const p(
127 dynamic_cast<const BaseKvmCPUParams *>(params()));
128
129 Kvm &kvm(vm.kvm);
130
131 BaseCPU::startup();
132
133 assert(vcpuFD == -1);
134
135 // Tell the VM that a CPU is about to start.
136 vm.cpuStartup();
137
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);
142
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,
147 vcpuFD, 0);
148 if (_kvmRun == MAP_FAILED)
149 panic("KVM: Failed to map run data structure\n");
150
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));
161 } else {
162 inform("KVM: Coalesced not supported by host OS\n");
163 }
164
165 thread->startup();
166
167 Event *startupEvent(
168 new EventWrapper<BaseKvmCPU,
169 &BaseKvmCPU::startupThread>(this, true));
170 schedule(startupEvent, curTick());
171}
172
173void
174BaseKvmCPU::startupThread()
175{
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()));
182
183 vcpuThread = pthread_self();
184
185 // Setup signal handlers. This has to be done after the vCPU is
186 // created since it manipulates the vCPU signal mask.
187 setupSignalHandler();
188
189 setupCounters();
190
191 if (p->usePerfOverflow)
192 runTimer.reset(new PerfKvmTimer(hwCycles,
193 KVM_KICK_SIGNAL,
194 p->hostFactor,
195 p->hostFreq));
196 else
197 runTimer.reset(new PosixKvmTimer(KVM_KICK_SIGNAL, CLOCK_MONOTONIC,
198 p->hostFactor,
199 p->hostFreq));
200
201}
202
203void
204BaseKvmCPU::regStats()
205{
206 using namespace Stats;
207
208 BaseCPU::regStats();
209
210 numInsts
211 .name(name() + ".committedInsts")
212 .desc("Number of instructions committed")
213 ;
214
215 numVMExits
216 .name(name() + ".numVMExits")
217 .desc("total number of KVM exits")
218 ;
219
220 numVMHalfEntries
221 .name(name() + ".numVMHalfEntries")
222 .desc("number of KVM entries to finalize pending operations")
223 ;
224
225 numExitSignal
226 .name(name() + ".numExitSignal")
227 .desc("exits due to signal delivery")
228 ;
229
230 numMMIO
231 .name(name() + ".numMMIO")
232 .desc("number of VM exits due to memory mapped IO")
233 ;
234
235 numCoalescedMMIO
236 .name(name() + ".numCoalescedMMIO")
237 .desc("number of coalesced memory mapped IO requests")
238 ;
239
240 numIO
241 .name(name() + ".numIO")
242 .desc("number of VM exits due to legacy IO")
243 ;
244
245 numHalt
246 .name(name() + ".numHalt")
247 .desc("number of VM exits due to wait for interrupt instructions")
248 ;
249
250 numInterrupts
251 .name(name() + ".numInterrupts")
252 .desc("number of interrupts delivered")
253 ;
254
255 numHypercalls
256 .name(name() + ".numHypercalls")
257 .desc("number of hypercalls")
258 ;
259}
260
261void
262BaseKvmCPU::serializeThread(std::ostream &os, ThreadID tid)
263{
264 if (DTRACE(Checkpoint)) {
265 DPRINTF(Checkpoint, "KVM: Serializing thread %i:\n", tid);
266 dump();
267 }
268
269 assert(tid == 0);
270 assert(_status == Idle);
271 thread->serialize(os);
272}
273
274void
275BaseKvmCPU::unserializeThread(Checkpoint *cp, const std::string §ion,
276 ThreadID tid)
277{
278 DPRINTF(Checkpoint, "KVM: Unserialize thread %i:\n", tid);
279
280 assert(tid == 0);
281 assert(_status == Idle);
282 thread->unserialize(cp, section);
283 threadContextDirty = true;
284}
285
286unsigned int
287BaseKvmCPU::drain(DrainManager *dm)
288{
289 if (switchedOut())
290 return 0;
291
292 DPRINTF(Drain, "BaseKvmCPU::drain\n");
293 switch (_status) {
294 case Running:
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
299 // yet.
300 if (!archIsDrained()) {
301 drainManager = dm;
302 return 1;
303 }
304
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);
311 _status = Idle;
312
313 /** FALLTHROUGH */
314 case Idle:
315 // Idle, no need to drain
316 assert(!tickEvent.scheduled());
317
318 // Sync the thread context here since we'll need it when we
319 // switch CPUs or checkpoint the CPU.
320 syncThreadContext();
321
322 return 0;
323
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.
331 drainManager = dm;
332
333 DPRINTF(Drain, "KVM CPU is waiting for service completion, "
334 "requesting drain.\n");
335 return 1;
336
337 case RunningService:
338 // We need to drain since the CPU is waiting for service (e.g., MMIOs)
339 drainManager = dm;
340
341 DPRINTF(Drain, "KVM CPU is waiting for service, requesting drain.\n");
342 return 1;
343
344 default:
345 panic("KVM: Unhandled CPU state in drain()\n");
346 return 0;
347 }
348}
349
350void
351BaseKvmCPU::drainResume()
352{
353 assert(!tickEvent.scheduled());
354
355 // We might have been switched out. In that case, we don't need to
356 // do anything.
357 if (switchedOut())
358 return;
359
360 DPRINTF(Kvm, "drainResume\n");
361 verifyMemoryMode();
362
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());
367 _status = Running;
368 } else {
369 _status = Idle;
370 }
371}
372
373void
374BaseKvmCPU::switchOut()
375{
376 DPRINTF(Kvm, "switchOut\n");
377
378 BaseCPU::switchOut();
379
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
382 // idle.
383 assert(!tickEvent.scheduled());
384 assert(_status == Idle);
385}
386
387void
388BaseKvmCPU::takeOverFrom(BaseCPU *cpu)
389{
390 DPRINTF(Kvm, "takeOverFrom\n");
391
392 BaseCPU::takeOverFrom(cpu);
393
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
396 // idle.
397 assert(!tickEvent.scheduled());
398 assert(_status == Idle);
399 assert(threadContexts.size() == 1);
400
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.
405 updateKvmState();
406 threadContextDirty = false;
407}
408
409void
410BaseKvmCPU::verifyMemoryMode() const
411{
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");
415 }
416}
417
418void
419BaseKvmCPU::wakeup()
420{
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());
426
427 // Kick the vCPU to get it to come out of KVM.
428 kick();
429
430 if (thread->status() != ThreadContext::Suspended)
431 return;
432
433 thread->activate();
434}
435
436void
437BaseKvmCPU::activateContext(ThreadID thread_num)
438{
439 DPRINTF(Kvm, "ActivateContext %d\n", thread_num);
440
441 assert(thread_num == 0);
442 assert(thread);
443
444 assert(_status == Idle);
445 assert(!tickEvent.scheduled());
446
447 numCycles += ticksToCycles(thread->lastActivate - thread->lastSuspend);
448
449 schedule(tickEvent, clockEdge(Cycles(0)));
450 _status = Running;
451}
452
453
454void
455BaseKvmCPU::suspendContext(ThreadID thread_num)
456{
457 DPRINTF(Kvm, "SuspendContext %d\n", thread_num);
458
459 assert(thread_num == 0);
460 assert(thread);
461
462 if (_status == Idle)
463 return;
464
465 assert(_status == Running || _status == RunningServiceCompletion);
466
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);
473
474 _status = Idle;
475}
476
477void
478BaseKvmCPU::deallocateContext(ThreadID thread_num)
479{
480 // for now, these are equivalent
481 suspendContext(thread_num);
482}
483
484void
485BaseKvmCPU::haltContext(ThreadID thread_num)
486{
487 // for now, these are equivalent
488 suspendContext(thread_num);
489}
490
491ThreadContext *
492BaseKvmCPU::getContext(int tn)
493{
494 assert(tn == 0);
495 syncThreadContext();
496 return tc;
497}
498
499
500Counter
501BaseKvmCPU::totalInsts() const
502{
503 return ctrInsts;
504}
505
506Counter
507BaseKvmCPU::totalOps() const
508{
509 hack_once("Pretending totalOps is equivalent to totalInsts()\n");
510 return ctrInsts;
511}
512
513void
514BaseKvmCPU::dump()
515{
516 inform("State dumping not implemented.");
517}
518
519void
520BaseKvmCPU::tick()
521{
522 Tick delay(0);
523 assert(_status != Idle);
524
525 switch (_status) {
526 case RunningService:
527 // handleKvmExit() will determine the next state of the CPU
528 delay = handleKvmExit();
529
530 if (tryDrain())
531 _status = Idle;
532 break;
533
534 case RunningServiceCompletion:
535 case Running: {
| 3 * All rights reserved
4 *
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.
13 *
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.
24 *
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.
36 *
37 * Authors: Andreas Sandberg
38 */
39
40#include <linux/kvm.h>
41#include <sys/ioctl.h>
42#include <sys/mman.h>
43#include <unistd.h>
44
45#include <cerrno>
46#include <csignal>
47#include <ostream>
48
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"
60
61#include <signal.h>
62
63/* Used by some KVM macros */
64#define PAGE_SIZE pageSize
65
66BaseKvmCPU::BaseKvmCPU(BaseKvmCPUParams *params)
67 : BaseCPU(params),
68 vm(*params->kvmVM),
69 _status(Idle),
70 dataPort(name() + ".dcache_port", this),
71 instPort(name() + ".icache_port", this),
72 threadContextDirty(true),
73 kvmStateDirty(false),
74 vcpuID(vm.allocVCPUID()), vcpuFD(-1), vcpuMMapSize(0),
75 _kvmRun(NULL), mmioRing(NULL),
76 pageSize(sysconf(_SC_PAGE_SIZE)),
77 tickEvent(*this),
78 activeInstPeriod(0),
79 perfControlledByTimer(params->usePerfOverflow),
80 hostFactor(params->hostFactor),
81 drainManager(NULL),
82 ctrInsts(0)
83{
84 if (pageSize == -1)
85 panic("KVM: Failed to determine host page size (%i)\n",
86 errno);
87
88 if (FullSystem)
89 thread = new SimpleThread(this, 0, params->system, params->itb, params->dtb,
90 params->isa[0]);
91 else
92 thread = new SimpleThread(this, /* thread_num */ 0, params->system,
93 params->workload[0], params->itb,
94 params->dtb, params->isa[0]);
95
96 thread->setStatus(ThreadContext::Halted);
97 tc = thread->getTC();
98 threadContexts.push_back(tc);
99}
100
101BaseKvmCPU::~BaseKvmCPU()
102{
103 if (_kvmRun)
104 munmap(_kvmRun, vcpuMMapSize);
105 close(vcpuFD);
106}
107
108void
109BaseKvmCPU::init()
110{
111 BaseCPU::init();
112
113 if (numThreads != 1)
114 fatal("KVM: Multithreading not supported");
115
116 tc->initMemProxies(tc);
117
118 // initialize CPU, including PC
119 if (FullSystem && !switchedOut())
120 TheISA::initCPU(tc, tc->contextId());
121}
122
123void
124BaseKvmCPU::startup()
125{
126 const BaseKvmCPUParams * const p(
127 dynamic_cast<const BaseKvmCPUParams *>(params()));
128
129 Kvm &kvm(vm.kvm);
130
131 BaseCPU::startup();
132
133 assert(vcpuFD == -1);
134
135 // Tell the VM that a CPU is about to start.
136 vm.cpuStartup();
137
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);
142
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,
147 vcpuFD, 0);
148 if (_kvmRun == MAP_FAILED)
149 panic("KVM: Failed to map run data structure\n");
150
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));
161 } else {
162 inform("KVM: Coalesced not supported by host OS\n");
163 }
164
165 thread->startup();
166
167 Event *startupEvent(
168 new EventWrapper<BaseKvmCPU,
169 &BaseKvmCPU::startupThread>(this, true));
170 schedule(startupEvent, curTick());
171}
172
173void
174BaseKvmCPU::startupThread()
175{
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()));
182
183 vcpuThread = pthread_self();
184
185 // Setup signal handlers. This has to be done after the vCPU is
186 // created since it manipulates the vCPU signal mask.
187 setupSignalHandler();
188
189 setupCounters();
190
191 if (p->usePerfOverflow)
192 runTimer.reset(new PerfKvmTimer(hwCycles,
193 KVM_KICK_SIGNAL,
194 p->hostFactor,
195 p->hostFreq));
196 else
197 runTimer.reset(new PosixKvmTimer(KVM_KICK_SIGNAL, CLOCK_MONOTONIC,
198 p->hostFactor,
199 p->hostFreq));
200
201}
202
203void
204BaseKvmCPU::regStats()
205{
206 using namespace Stats;
207
208 BaseCPU::regStats();
209
210 numInsts
211 .name(name() + ".committedInsts")
212 .desc("Number of instructions committed")
213 ;
214
215 numVMExits
216 .name(name() + ".numVMExits")
217 .desc("total number of KVM exits")
218 ;
219
220 numVMHalfEntries
221 .name(name() + ".numVMHalfEntries")
222 .desc("number of KVM entries to finalize pending operations")
223 ;
224
225 numExitSignal
226 .name(name() + ".numExitSignal")
227 .desc("exits due to signal delivery")
228 ;
229
230 numMMIO
231 .name(name() + ".numMMIO")
232 .desc("number of VM exits due to memory mapped IO")
233 ;
234
235 numCoalescedMMIO
236 .name(name() + ".numCoalescedMMIO")
237 .desc("number of coalesced memory mapped IO requests")
238 ;
239
240 numIO
241 .name(name() + ".numIO")
242 .desc("number of VM exits due to legacy IO")
243 ;
244
245 numHalt
246 .name(name() + ".numHalt")
247 .desc("number of VM exits due to wait for interrupt instructions")
248 ;
249
250 numInterrupts
251 .name(name() + ".numInterrupts")
252 .desc("number of interrupts delivered")
253 ;
254
255 numHypercalls
256 .name(name() + ".numHypercalls")
257 .desc("number of hypercalls")
258 ;
259}
260
261void
262BaseKvmCPU::serializeThread(std::ostream &os, ThreadID tid)
263{
264 if (DTRACE(Checkpoint)) {
265 DPRINTF(Checkpoint, "KVM: Serializing thread %i:\n", tid);
266 dump();
267 }
268
269 assert(tid == 0);
270 assert(_status == Idle);
271 thread->serialize(os);
272}
273
274void
275BaseKvmCPU::unserializeThread(Checkpoint *cp, const std::string §ion,
276 ThreadID tid)
277{
278 DPRINTF(Checkpoint, "KVM: Unserialize thread %i:\n", tid);
279
280 assert(tid == 0);
281 assert(_status == Idle);
282 thread->unserialize(cp, section);
283 threadContextDirty = true;
284}
285
286unsigned int
287BaseKvmCPU::drain(DrainManager *dm)
288{
289 if (switchedOut())
290 return 0;
291
292 DPRINTF(Drain, "BaseKvmCPU::drain\n");
293 switch (_status) {
294 case Running:
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
299 // yet.
300 if (!archIsDrained()) {
301 drainManager = dm;
302 return 1;
303 }
304
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);
311 _status = Idle;
312
313 /** FALLTHROUGH */
314 case Idle:
315 // Idle, no need to drain
316 assert(!tickEvent.scheduled());
317
318 // Sync the thread context here since we'll need it when we
319 // switch CPUs or checkpoint the CPU.
320 syncThreadContext();
321
322 return 0;
323
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.
331 drainManager = dm;
332
333 DPRINTF(Drain, "KVM CPU is waiting for service completion, "
334 "requesting drain.\n");
335 return 1;
336
337 case RunningService:
338 // We need to drain since the CPU is waiting for service (e.g., MMIOs)
339 drainManager = dm;
340
341 DPRINTF(Drain, "KVM CPU is waiting for service, requesting drain.\n");
342 return 1;
343
344 default:
345 panic("KVM: Unhandled CPU state in drain()\n");
346 return 0;
347 }
348}
349
350void
351BaseKvmCPU::drainResume()
352{
353 assert(!tickEvent.scheduled());
354
355 // We might have been switched out. In that case, we don't need to
356 // do anything.
357 if (switchedOut())
358 return;
359
360 DPRINTF(Kvm, "drainResume\n");
361 verifyMemoryMode();
362
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());
367 _status = Running;
368 } else {
369 _status = Idle;
370 }
371}
372
373void
374BaseKvmCPU::switchOut()
375{
376 DPRINTF(Kvm, "switchOut\n");
377
378 BaseCPU::switchOut();
379
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
382 // idle.
383 assert(!tickEvent.scheduled());
384 assert(_status == Idle);
385}
386
387void
388BaseKvmCPU::takeOverFrom(BaseCPU *cpu)
389{
390 DPRINTF(Kvm, "takeOverFrom\n");
391
392 BaseCPU::takeOverFrom(cpu);
393
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
396 // idle.
397 assert(!tickEvent.scheduled());
398 assert(_status == Idle);
399 assert(threadContexts.size() == 1);
400
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.
405 updateKvmState();
406 threadContextDirty = false;
407}
408
409void
410BaseKvmCPU::verifyMemoryMode() const
411{
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");
415 }
416}
417
418void
419BaseKvmCPU::wakeup()
420{
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());
426
427 // Kick the vCPU to get it to come out of KVM.
428 kick();
429
430 if (thread->status() != ThreadContext::Suspended)
431 return;
432
433 thread->activate();
434}
435
436void
437BaseKvmCPU::activateContext(ThreadID thread_num)
438{
439 DPRINTF(Kvm, "ActivateContext %d\n", thread_num);
440
441 assert(thread_num == 0);
442 assert(thread);
443
444 assert(_status == Idle);
445 assert(!tickEvent.scheduled());
446
447 numCycles += ticksToCycles(thread->lastActivate - thread->lastSuspend);
448
449 schedule(tickEvent, clockEdge(Cycles(0)));
450 _status = Running;
451}
452
453
454void
455BaseKvmCPU::suspendContext(ThreadID thread_num)
456{
457 DPRINTF(Kvm, "SuspendContext %d\n", thread_num);
458
459 assert(thread_num == 0);
460 assert(thread);
461
462 if (_status == Idle)
463 return;
464
465 assert(_status == Running || _status == RunningServiceCompletion);
466
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);
473
474 _status = Idle;
475}
476
477void
478BaseKvmCPU::deallocateContext(ThreadID thread_num)
479{
480 // for now, these are equivalent
481 suspendContext(thread_num);
482}
483
484void
485BaseKvmCPU::haltContext(ThreadID thread_num)
486{
487 // for now, these are equivalent
488 suspendContext(thread_num);
489}
490
491ThreadContext *
492BaseKvmCPU::getContext(int tn)
493{
494 assert(tn == 0);
495 syncThreadContext();
496 return tc;
497}
498
499
500Counter
501BaseKvmCPU::totalInsts() const
502{
503 return ctrInsts;
504}
505
506Counter
507BaseKvmCPU::totalOps() const
508{
509 hack_once("Pretending totalOps is equivalent to totalInsts()\n");
510 return ctrInsts;
511}
512
513void
514BaseKvmCPU::dump()
515{
516 inform("State dumping not implemented.");
517}
518
519void
520BaseKvmCPU::tick()
521{
522 Tick delay(0);
523 assert(_status != Idle);
524
525 switch (_status) {
526 case RunningService:
527 // handleKvmExit() will determine the next state of the CPU
528 delay = handleKvmExit();
529
530 if (tryDrain())
531 _status = Idle;
532 break;
533
534 case RunningServiceCompletion:
535 case Running: {
|
545
546 DPRINTF(KvmRun, "Entering KVM...\n");
547 if (drainManager) {
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();
553 } else {
554 delay = kvmRun(ticksToExecute);
555 }
556
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.
560 if (_status == Idle)
561 break;
562
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;
567
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;
572 } else {
573 ++numExitSignal;
574 _status = Running;
575 }
576
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);
582
583 if (tryDrain())
584 _status = Idle;
585 } break;
586
587 default:
588 panic("BaseKvmCPU entered tick() in an illegal state (%i)\n",
589 _status);
590 }
591
592 // Schedule a new tick if we are still running
593 if (_status != Idle)
594 schedule(tickEvent, clockEdge(ticksToCycles(delay)));
595}
596
597Tick
598BaseKvmCPU::kvmRunDrain()
599{
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);
604
605 // Deliver the data from the pending IO operation and immediately
606 // exit.
607 return kvmRun(0);
608}
609
610uint64_t
611BaseKvmCPU::getHostCycles() const
612{
613 return hwCycles.read();
614}
615
616Tick
617BaseKvmCPU::kvmRun(Tick ticks)
618{
619 Tick ticksExecuted;
620 DPRINTF(KvmRun, "KVM: Executing for %i ticks\n", ticks);
621
622 if (ticks == 0) {
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");
627
628 ++numVMHalfEntries;
629
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().
634 kick();
635
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.
642 ioctlRun();
643
644 // We always execute at least one cycle to prevent the
645 // BaseKvmCPU::tick() to be rescheduled on the same tick
646 // twice.
647 ticksExecuted = clockPeriod();
648 } else {
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());
656
657 if (ticks < runTimer->resolution()) {
658 DPRINTF(KvmRun, "KVM: Adjusting tick count (%i -> %i)\n",
659 ticks, runTimer->resolution());
660 ticks = runTimer->resolution();
661 }
662
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());
667
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)
674 hwCycles.start();
675
676 ioctlRun();
677
678 runTimer->disarm();
679 if (!perfControlledByTimer)
680 hwCycles.stop();
681
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
686 // enter into KVM.
687 discardPendingSignal(KVM_KICK_SIGNAL);
688
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);
693
694 /* Update statistics */
695 numCycles += simCyclesExecuted;;
696 numInsts += instsExecuted;
697 ctrInsts += instsExecuted;
698 system->totalNumInsts += instsExecuted;
699
700 DPRINTF(KvmRun,
701 "KVM: Executed %i instructions in %i cycles "
702 "(%i ticks, sim cycles: %i).\n",
703 instsExecuted, hostCyclesExecuted, ticksExecuted, simCyclesExecuted);
704 }
705
706 ++numVMExits;
707
708 return ticksExecuted + flushCoalescedMMIO();
709}
710
711void
712BaseKvmCPU::kvmNonMaskableInterrupt()
713{
714 ++numInterrupts;
715 if (ioctl(KVM_NMI) == -1)
716 panic("KVM: Failed to deliver NMI to virtual CPU\n");
717}
718
719void
720BaseKvmCPU::kvmInterrupt(const struct kvm_interrupt &interrupt)
721{
722 ++numInterrupts;
723 if (ioctl(KVM_INTERRUPT, (void *)&interrupt) == -1)
724 panic("KVM: Failed to deliver interrupt to virtual CPU\n");
725}
726
727void
728BaseKvmCPU::getRegisters(struct kvm_regs ®s) const
729{
730 if (ioctl(KVM_GET_REGS, ®s) == -1)
731 panic("KVM: Failed to get guest registers\n");
732}
733
734void
735BaseKvmCPU::setRegisters(const struct kvm_regs ®s)
736{
737 if (ioctl(KVM_SET_REGS, (void *)®s) == -1)
738 panic("KVM: Failed to set guest registers\n");
739}
740
741void
742BaseKvmCPU::getSpecialRegisters(struct kvm_sregs ®s) const
743{
744 if (ioctl(KVM_GET_SREGS, ®s) == -1)
745 panic("KVM: Failed to get guest special registers\n");
746}
747
748void
749BaseKvmCPU::setSpecialRegisters(const struct kvm_sregs ®s)
750{
751 if (ioctl(KVM_SET_SREGS, (void *)®s) == -1)
752 panic("KVM: Failed to set guest special registers\n");
753}
754
755void
756BaseKvmCPU::getFPUState(struct kvm_fpu &state) const
757{
758 if (ioctl(KVM_GET_FPU, &state) == -1)
759 panic("KVM: Failed to get guest FPU state\n");
760}
761
762void
763BaseKvmCPU::setFPUState(const struct kvm_fpu &state)
764{
765 if (ioctl(KVM_SET_FPU, (void *)&state) == -1)
766 panic("KVM: Failed to set guest FPU state\n");
767}
768
769
770void
771BaseKvmCPU::setOneReg(uint64_t id, const void *addr)
772{
773#ifdef KVM_SET_ONE_REG
774 struct kvm_one_reg reg;
775 reg.id = id;
776 reg.addr = (uint64_t)addr;
777
778 if (ioctl(KVM_SET_ONE_REG, ®) == -1) {
779 panic("KVM: Failed to set register (0x%x) value (errno: %i)\n",
780 id, errno);
781 }
782#else
783 panic("KVM_SET_ONE_REG is unsupported on this platform.\n");
784#endif
785}
786
787void
788BaseKvmCPU::getOneReg(uint64_t id, void *addr) const
789{
790#ifdef KVM_GET_ONE_REG
791 struct kvm_one_reg reg;
792 reg.id = id;
793 reg.addr = (uint64_t)addr;
794
795 if (ioctl(KVM_GET_ONE_REG, ®) == -1) {
796 panic("KVM: Failed to get register (0x%x) value (errno: %i)\n",
797 id, errno);
798 }
799#else
800 panic("KVM_GET_ONE_REG is unsupported on this platform.\n");
801#endif
802}
803
804std::string
805BaseKvmCPU::getAndFormatOneReg(uint64_t id) const
806{
807#ifdef KVM_GET_ONE_REG
808 std::ostringstream ss;
809
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); \
816 ss << value; \
817 } break
818
819#define HANDLE_ARRAY(len) \
820 case KVM_REG_SIZE_U ## len: { \
821 uint8_t value[len / 8]; \
822 getOneReg(id, value); \
823 ccprintf(ss, "[0x%x", value[0]); \
824 for (int i = 1; i < len / 8; ++i) \
825 ccprintf(ss, ", 0x%x", value[i]); \
826 ccprintf(ss, "]"); \
827 } break
828
829 switch (id & KVM_REG_SIZE_MASK) {
830 HANDLE_INTTYPE(8);
831 HANDLE_INTTYPE(16);
832 HANDLE_INTTYPE(32);
833 HANDLE_INTTYPE(64);
834 HANDLE_ARRAY(128);
835 HANDLE_ARRAY(256);
836 HANDLE_ARRAY(512);
837 HANDLE_ARRAY(1024);
838 default:
839 ss << "??";
840 }
841
842#undef HANDLE_INTTYPE
843#undef HANDLE_ARRAY
844
845 return ss.str();
846#else
847 panic("KVM_GET_ONE_REG is unsupported on this platform.\n");
848#endif
849}
850
851void
852BaseKvmCPU::syncThreadContext()
853{
854 if (!kvmStateDirty)
855 return;
856
857 assert(!threadContextDirty);
858
859 updateThreadContext();
860 kvmStateDirty = false;
861}
862
863void
864BaseKvmCPU::syncKvmState()
865{
866 if (!threadContextDirty)
867 return;
868
869 assert(!kvmStateDirty);
870
871 updateKvmState();
872 threadContextDirty = false;
873}
874
875Tick
876BaseKvmCPU::handleKvmExit()
877{
878 DPRINTF(KvmRun, "handleKvmExit (exit_reason: %i)\n", _kvmRun->exit_reason);
879 assert(_status == RunningService);
880
881 // Switch into the running state by default. Individual handlers
882 // can override this.
883 _status = Running;
884 switch (_kvmRun->exit_reason) {
885 case KVM_EXIT_UNKNOWN:
886 return handleKvmExitUnknown();
887
888 case KVM_EXIT_EXCEPTION:
889 return handleKvmExitException();
890
891 case KVM_EXIT_IO:
892 _status = RunningServiceCompletion;
893 ++numIO;
894 return handleKvmExitIO();
895
896 case KVM_EXIT_HYPERCALL:
897 ++numHypercalls;
898 return handleKvmExitHypercall();
899
900 case KVM_EXIT_HLT:
901 /* The guest has halted and is waiting for interrupts */
902 DPRINTF(Kvm, "handleKvmExitHalt\n");
903 ++numHalt;
904
905 // Suspend the thread until the next interrupt arrives
906 thread->suspend();
907
908 // This is actually ignored since the thread is suspended.
909 return 0;
910
911 case KVM_EXIT_MMIO:
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);
917
918 ++numMMIO;
919 return doMMIOAccess(_kvmRun->mmio.phys_addr, _kvmRun->mmio.data,
920 _kvmRun->mmio.len, _kvmRun->mmio.is_write);
921
922 case KVM_EXIT_IRQ_WINDOW_OPEN:
923 return handleKvmExitIRQWindowOpen();
924
925 case KVM_EXIT_FAIL_ENTRY:
926 return handleKvmExitFailEntry();
927
928 case KVM_EXIT_INTR:
929 /* KVM was interrupted by a signal, restart it in the next
930 * tick. */
931 return 0;
932
933 case KVM_EXIT_INTERNAL_ERROR:
934 panic("KVM: Internal error (suberror: %u)\n",
935 _kvmRun->internal.suberror);
936
937 default:
938 dump();
939 panic("KVM: Unexpected exit (exit_reason: %u)\n", _kvmRun->exit_reason);
940 }
941}
942
943Tick
944BaseKvmCPU::handleKvmExitIO()
945{
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);
949}
950
951Tick
952BaseKvmCPU::handleKvmExitHypercall()
953{
954 panic("KVM: Unhandled hypercall\n");
955}
956
957Tick
958BaseKvmCPU::handleKvmExitIRQWindowOpen()
959{
960 warn("KVM: Unhandled IRQ window.\n");
961 return 0;
962}
963
964
965Tick
966BaseKvmCPU::handleKvmExitUnknown()
967{
968 dump();
969 panic("KVM: Unknown error when starting vCPU (hw reason: 0x%llx)\n",
970 _kvmRun->hw.hardware_exit_reason);
971}
972
973Tick
974BaseKvmCPU::handleKvmExitException()
975{
976 dump();
977 panic("KVM: Got exception when starting vCPU "
978 "(exception: %u, error_code: %u)\n",
979 _kvmRun->ex.exception, _kvmRun->ex.error_code);
980}
981
982Tick
983BaseKvmCPU::handleKvmExitFailEntry()
984{
985 dump();
986 panic("KVM: Failed to enter virtualized mode (hw reason: 0x%llx)\n",
987 _kvmRun->fail_entry.hardware_entry_failure_reason);
988}
989
990Tick
991BaseKvmCPU::doMMIOAccess(Addr paddr, void *data, int size, bool write)
992{
993 ThreadContext *tc(thread->getTC());
994 syncThreadContext();
995
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
1001 // interface.
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());
1006
1007
1008 const MemCmd cmd(write ? MemCmd::WriteReq : MemCmd::ReadReq);
1009 Packet pkt(&mmio_req, cmd);
1010 pkt.dataStatic(data);
1011
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;
1021 } else {
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());
1026
1027 return dataPort.sendAtomic(&pkt);
1028 }
1029}
1030
1031void
1032BaseKvmCPU::setSignalMask(const sigset_t *mask)
1033{
1034 std::unique_ptr<struct kvm_signal_mask> kvm_mask;
1035
1036 if (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);
1043 kvm_mask->len = 8;
1044 memcpy(kvm_mask->sigset, mask, kvm_mask->len);
1045 }
1046
1047 if (ioctl(KVM_SET_SIGNAL_MASK, (void *)kvm_mask.get()) == -1)
1048 panic("KVM: Failed to set vCPU signal mask (errno: %i)\n",
1049 errno);
1050}
1051
1052int
1053BaseKvmCPU::ioctl(int request, long p1) const
1054{
1055 if (vcpuFD == -1)
1056 panic("KVM: CPU ioctl called before initialization\n");
1057
1058 return ::ioctl(vcpuFD, request, p1);
1059}
1060
1061Tick
1062BaseKvmCPU::flushCoalescedMMIO()
1063{
1064 if (!mmioRing)
1065 return 0;
1066
1067 DPRINTF(KvmIO, "KVM: Flushing the coalesced MMIO ring buffer\n");
1068
1069 // TODO: We might need to do synchronization when we start to
1070 // support multiple CPUs
1071 Tick ticks(0);
1072 while (mmioRing->first != mmioRing->last) {
1073 struct kvm_coalesced_mmio &ent(
1074 mmioRing->coalesced_mmio[mmioRing->first]);
1075
1076 DPRINTF(KvmIO, "KVM: Handling coalesced MMIO (addr: 0x%x, len: %u)\n",
1077 ent.phys_addr, ent.len);
1078
1079 ++numCoalescedMMIO;
1080 ticks += doMMIOAccess(ent.phys_addr, ent.data, ent.len, true);
1081
1082 mmioRing->first = (mmioRing->first + 1) % KVM_COALESCED_MMIO_MAX;
1083 }
1084
1085 return ticks;
1086}
1087
1088/**
1089 * Dummy handler for KVM kick signals.
1090 *
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
1096 * handling.
1097 */
1098static void
1099onKickSignal(int signo, siginfo_t *si, void *data)
1100{
1101}
1102
1103void
1104BaseKvmCPU::setupSignalHandler()
1105{
1106 struct sigaction sa;
1107
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");
1113
1114 sigset_t sigset;
1115 if (pthread_sigmask(SIG_BLOCK, NULL, &sigset) == -1)
1116 panic("KVM: Failed get signal mask\n");
1117
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);
1124
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");
1130}
1131
1132bool
1133BaseKvmCPU::discardPendingSignal(int signum) const
1134{
1135 int discardedSignal;
1136
1137 // Setting the timeout to zero causes sigtimedwait to return
1138 // immediately.
1139 struct timespec timeout;
1140 timeout.tv_sec = 0;
1141 timeout.tv_nsec = 0;
1142
1143 sigset_t sigset;
1144 sigemptyset(&sigset);
1145 sigaddset(&sigset, signum);
1146
1147 do {
1148 discardedSignal = sigtimedwait(&sigset, NULL, &timeout);
1149 } while (discardedSignal == -1 && errno == EINTR);
1150
1151 if (discardedSignal == signum)
1152 return true;
1153 else if (discardedSignal == -1 && errno == EAGAIN)
1154 return false;
1155 else
1156 panic("Unexpected return value from sigtimedwait: %i (errno: %i)\n",
1157 discardedSignal, errno);
1158}
1159
1160void
1161BaseKvmCPU::setupCounters()
1162{
1163 DPRINTF(Kvm, "Attaching cycle counter...\n");
1164 PerfKvmCounterConfig cfgCycles(PERF_TYPE_HARDWARE,
1165 PERF_COUNT_HW_CPU_CYCLES);
1166 cfgCycles.disabled(true)
1167 .pinned(true);
1168
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);
1174
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)
1182 .samplePeriod(42);
1183 }
1184
1185 hwCycles.attach(cfgCycles,
1186 0); // TID (0 => currentThread)
1187
1188 setupInstCounter();
1189}
1190
1191bool
1192BaseKvmCPU::tryDrain()
1193{
1194 if (!drainManager)
1195 return false;
1196
1197 if (!archIsDrained()) {
1198 DPRINTF(Drain, "tryDrain: Architecture code is not ready.\n");
1199 return false;
1200 }
1201
1202 if (_status == Idle || _status == Running) {
1203 DPRINTF(Drain,
1204 "tryDrain: CPU transitioned into the Idle state, drain done\n");
1205 drainManager->signalDrainDone();
1206 drainManager = NULL;
1207 return true;
1208 } else {
1209 DPRINTF(Drain, "tryDrain: CPU not ready.\n");
1210 return false;
1211 }
1212}
1213
1214void
1215BaseKvmCPU::ioctlRun()
1216{
1217 if (ioctl(KVM_RUN) == -1) {
1218 if (errno != EINTR)
1219 panic("KVM: Failed to start virtual CPU (errno: %i)\n",
1220 errno);
1221 }
1222}
1223
1224void
1225BaseKvmCPU::setupInstStop()
1226{
1227 if (comInstEventQueue[0]->empty()) {
1228 setupInstCounter(0);
1229 } else {
1230 const uint64_t next(comInstEventQueue[0]->nextTick());
1231
1232 assert(next > ctrInsts);
1233 setupInstCounter(next - ctrInsts);
1234 }
1235}
1236
1237void
1238BaseKvmCPU::setupInstCounter(uint64_t period)
1239{
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())
1243 return;
1244
1245 PerfKvmCounterConfig cfgInstructions(PERF_TYPE_HARDWARE,
1246 PERF_COUNT_HW_INSTRUCTIONS);
1247
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);
1253
1254 if (period) {
1255 // Setup a sampling counter if that has been requested.
1256 cfgInstructions.wakeupEvents(1)
1257 .samplePeriod(period);
1258 }
1259
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)
1267 hwCycles);
1268
1269 if (period)
1270 hwInstructions.enableSignals(KVM_KICK_SIGNAL);
1271
1272 activeInstPeriod = period;
1273}
| 553
554 DPRINTF(KvmRun, "Entering KVM...\n");
555 if (drainManager) {
556 // Force an immediate exit from KVM after completing
557 // pending operations. The architecture-specific code
558 // takes care to run until it is in a state where it can
559 // safely be drained.
560 delay = kvmRunDrain();
561 } else {
562 delay = kvmRun(ticksToExecute);
563 }
564
565 // The CPU might have been suspended before entering into
566 // KVM. Assume that the CPU was suspended /before/ entering
567 // into KVM and skip the exit handling.
568 if (_status == Idle)
569 break;
570
571 // Entering into KVM implies that we'll have to reload the thread
572 // context from KVM if we want to access it. Flag the KVM state as
573 // dirty with respect to the cached thread context.
574 kvmStateDirty = true;
575
576 // Enter into the RunningService state unless the
577 // simulation was stopped by a timer.
578 if (_kvmRun->exit_reason != KVM_EXIT_INTR) {
579 _status = RunningService;
580 } else {
581 ++numExitSignal;
582 _status = Running;
583 }
584
585 // Service any pending instruction events. The vCPU should
586 // have exited in time for the event using the instruction
587 // counter configured by setupInstStop().
588 comInstEventQueue[0]->serviceEvents(ctrInsts);
589 system->instEventQueue.serviceEvents(system->totalNumInsts);
590
591 if (tryDrain())
592 _status = Idle;
593 } break;
594
595 default:
596 panic("BaseKvmCPU entered tick() in an illegal state (%i)\n",
597 _status);
598 }
599
600 // Schedule a new tick if we are still running
601 if (_status != Idle)
602 schedule(tickEvent, clockEdge(ticksToCycles(delay)));
603}
604
605Tick
606BaseKvmCPU::kvmRunDrain()
607{
608 // By default, the only thing we need to drain is a pending IO
609 // operation which assumes that we are in the
610 // RunningServiceCompletion state.
611 assert(_status == RunningServiceCompletion);
612
613 // Deliver the data from the pending IO operation and immediately
614 // exit.
615 return kvmRun(0);
616}
617
618uint64_t
619BaseKvmCPU::getHostCycles() const
620{
621 return hwCycles.read();
622}
623
624Tick
625BaseKvmCPU::kvmRun(Tick ticks)
626{
627 Tick ticksExecuted;
628 DPRINTF(KvmRun, "KVM: Executing for %i ticks\n", ticks);
629
630 if (ticks == 0) {
631 // Settings ticks == 0 is a special case which causes an entry
632 // into KVM that finishes pending operations (e.g., IO) and
633 // then immediately exits.
634 DPRINTF(KvmRun, "KVM: Delivering IO without full guest entry\n");
635
636 ++numVMHalfEntries;
637
638 // Send a KVM_KICK_SIGNAL to the vCPU thread (i.e., this
639 // thread). The KVM control signal is masked while executing
640 // in gem5 and gets unmasked temporarily as when entering
641 // KVM. See setSignalMask() and setupSignalHandler().
642 kick();
643
644 // Start the vCPU. KVM will check for signals after completing
645 // pending operations (IO). Since the KVM_KICK_SIGNAL is
646 // pending, this forces an immediate exit to gem5 again. We
647 // don't bother to setup timers since this shouldn't actually
648 // execute any code (other than completing half-executed IO
649 // instructions) in the guest.
650 ioctlRun();
651
652 // We always execute at least one cycle to prevent the
653 // BaseKvmCPU::tick() to be rescheduled on the same tick
654 // twice.
655 ticksExecuted = clockPeriod();
656 } else {
657 // This method is executed as a result of a tick event. That
658 // means that the event queue will be locked when entering the
659 // method. We temporarily unlock the event queue to allow
660 // other threads to steal control of this thread to inject
661 // interrupts. They will typically lock the queue and then
662 // force an exit from KVM by kicking the vCPU.
663 EventQueue::ScopedRelease release(curEventQueue());
664
665 if (ticks < runTimer->resolution()) {
666 DPRINTF(KvmRun, "KVM: Adjusting tick count (%i -> %i)\n",
667 ticks, runTimer->resolution());
668 ticks = runTimer->resolution();
669 }
670
671 // Get hardware statistics after synchronizing contexts. The KVM
672 // state update might affect guest cycle counters.
673 uint64_t baseCycles(getHostCycles());
674 uint64_t baseInstrs(hwInstructions.read());
675
676 // Arm the run timer and start the cycle timer if it isn't
677 // controlled by the overflow timer. Starting/stopping the cycle
678 // timer automatically starts the other perf timers as they are in
679 // the same counter group.
680 runTimer->arm(ticks);
681 if (!perfControlledByTimer)
682 hwCycles.start();
683
684 ioctlRun();
685
686 runTimer->disarm();
687 if (!perfControlledByTimer)
688 hwCycles.stop();
689
690 // The control signal may have been delivered after we exited
691 // from KVM. It will be pending in that case since it is
692 // masked when we aren't executing in KVM. Discard it to make
693 // sure we don't deliver it immediately next time we try to
694 // enter into KVM.
695 discardPendingSignal(KVM_KICK_SIGNAL);
696
697 const uint64_t hostCyclesExecuted(getHostCycles() - baseCycles);
698 const uint64_t simCyclesExecuted(hostCyclesExecuted * hostFactor);
699 const uint64_t instsExecuted(hwInstructions.read() - baseInstrs);
700 ticksExecuted = runTimer->ticksFromHostCycles(hostCyclesExecuted);
701
702 /* Update statistics */
703 numCycles += simCyclesExecuted;;
704 numInsts += instsExecuted;
705 ctrInsts += instsExecuted;
706 system->totalNumInsts += instsExecuted;
707
708 DPRINTF(KvmRun,
709 "KVM: Executed %i instructions in %i cycles "
710 "(%i ticks, sim cycles: %i).\n",
711 instsExecuted, hostCyclesExecuted, ticksExecuted, simCyclesExecuted);
712 }
713
714 ++numVMExits;
715
716 return ticksExecuted + flushCoalescedMMIO();
717}
718
719void
720BaseKvmCPU::kvmNonMaskableInterrupt()
721{
722 ++numInterrupts;
723 if (ioctl(KVM_NMI) == -1)
724 panic("KVM: Failed to deliver NMI to virtual CPU\n");
725}
726
727void
728BaseKvmCPU::kvmInterrupt(const struct kvm_interrupt &interrupt)
729{
730 ++numInterrupts;
731 if (ioctl(KVM_INTERRUPT, (void *)&interrupt) == -1)
732 panic("KVM: Failed to deliver interrupt to virtual CPU\n");
733}
734
735void
736BaseKvmCPU::getRegisters(struct kvm_regs ®s) const
737{
738 if (ioctl(KVM_GET_REGS, ®s) == -1)
739 panic("KVM: Failed to get guest registers\n");
740}
741
742void
743BaseKvmCPU::setRegisters(const struct kvm_regs ®s)
744{
745 if (ioctl(KVM_SET_REGS, (void *)®s) == -1)
746 panic("KVM: Failed to set guest registers\n");
747}
748
749void
750BaseKvmCPU::getSpecialRegisters(struct kvm_sregs ®s) const
751{
752 if (ioctl(KVM_GET_SREGS, ®s) == -1)
753 panic("KVM: Failed to get guest special registers\n");
754}
755
756void
757BaseKvmCPU::setSpecialRegisters(const struct kvm_sregs ®s)
758{
759 if (ioctl(KVM_SET_SREGS, (void *)®s) == -1)
760 panic("KVM: Failed to set guest special registers\n");
761}
762
763void
764BaseKvmCPU::getFPUState(struct kvm_fpu &state) const
765{
766 if (ioctl(KVM_GET_FPU, &state) == -1)
767 panic("KVM: Failed to get guest FPU state\n");
768}
769
770void
771BaseKvmCPU::setFPUState(const struct kvm_fpu &state)
772{
773 if (ioctl(KVM_SET_FPU, (void *)&state) == -1)
774 panic("KVM: Failed to set guest FPU state\n");
775}
776
777
778void
779BaseKvmCPU::setOneReg(uint64_t id, const void *addr)
780{
781#ifdef KVM_SET_ONE_REG
782 struct kvm_one_reg reg;
783 reg.id = id;
784 reg.addr = (uint64_t)addr;
785
786 if (ioctl(KVM_SET_ONE_REG, ®) == -1) {
787 panic("KVM: Failed to set register (0x%x) value (errno: %i)\n",
788 id, errno);
789 }
790#else
791 panic("KVM_SET_ONE_REG is unsupported on this platform.\n");
792#endif
793}
794
795void
796BaseKvmCPU::getOneReg(uint64_t id, void *addr) const
797{
798#ifdef KVM_GET_ONE_REG
799 struct kvm_one_reg reg;
800 reg.id = id;
801 reg.addr = (uint64_t)addr;
802
803 if (ioctl(KVM_GET_ONE_REG, ®) == -1) {
804 panic("KVM: Failed to get register (0x%x) value (errno: %i)\n",
805 id, errno);
806 }
807#else
808 panic("KVM_GET_ONE_REG is unsupported on this platform.\n");
809#endif
810}
811
812std::string
813BaseKvmCPU::getAndFormatOneReg(uint64_t id) const
814{
815#ifdef KVM_GET_ONE_REG
816 std::ostringstream ss;
817
818 ss.setf(std::ios::hex, std::ios::basefield);
819 ss.setf(std::ios::showbase);
820#define HANDLE_INTTYPE(len) \
821 case KVM_REG_SIZE_U ## len: { \
822 uint ## len ## _t value; \
823 getOneReg(id, &value); \
824 ss << value; \
825 } break
826
827#define HANDLE_ARRAY(len) \
828 case KVM_REG_SIZE_U ## len: { \
829 uint8_t value[len / 8]; \
830 getOneReg(id, value); \
831 ccprintf(ss, "[0x%x", value[0]); \
832 for (int i = 1; i < len / 8; ++i) \
833 ccprintf(ss, ", 0x%x", value[i]); \
834 ccprintf(ss, "]"); \
835 } break
836
837 switch (id & KVM_REG_SIZE_MASK) {
838 HANDLE_INTTYPE(8);
839 HANDLE_INTTYPE(16);
840 HANDLE_INTTYPE(32);
841 HANDLE_INTTYPE(64);
842 HANDLE_ARRAY(128);
843 HANDLE_ARRAY(256);
844 HANDLE_ARRAY(512);
845 HANDLE_ARRAY(1024);
846 default:
847 ss << "??";
848 }
849
850#undef HANDLE_INTTYPE
851#undef HANDLE_ARRAY
852
853 return ss.str();
854#else
855 panic("KVM_GET_ONE_REG is unsupported on this platform.\n");
856#endif
857}
858
859void
860BaseKvmCPU::syncThreadContext()
861{
862 if (!kvmStateDirty)
863 return;
864
865 assert(!threadContextDirty);
866
867 updateThreadContext();
868 kvmStateDirty = false;
869}
870
871void
872BaseKvmCPU::syncKvmState()
873{
874 if (!threadContextDirty)
875 return;
876
877 assert(!kvmStateDirty);
878
879 updateKvmState();
880 threadContextDirty = false;
881}
882
883Tick
884BaseKvmCPU::handleKvmExit()
885{
886 DPRINTF(KvmRun, "handleKvmExit (exit_reason: %i)\n", _kvmRun->exit_reason);
887 assert(_status == RunningService);
888
889 // Switch into the running state by default. Individual handlers
890 // can override this.
891 _status = Running;
892 switch (_kvmRun->exit_reason) {
893 case KVM_EXIT_UNKNOWN:
894 return handleKvmExitUnknown();
895
896 case KVM_EXIT_EXCEPTION:
897 return handleKvmExitException();
898
899 case KVM_EXIT_IO:
900 _status = RunningServiceCompletion;
901 ++numIO;
902 return handleKvmExitIO();
903
904 case KVM_EXIT_HYPERCALL:
905 ++numHypercalls;
906 return handleKvmExitHypercall();
907
908 case KVM_EXIT_HLT:
909 /* The guest has halted and is waiting for interrupts */
910 DPRINTF(Kvm, "handleKvmExitHalt\n");
911 ++numHalt;
912
913 // Suspend the thread until the next interrupt arrives
914 thread->suspend();
915
916 // This is actually ignored since the thread is suspended.
917 return 0;
918
919 case KVM_EXIT_MMIO:
920 _status = RunningServiceCompletion;
921 /* Service memory mapped IO requests */
922 DPRINTF(KvmIO, "KVM: Handling MMIO (w: %u, addr: 0x%x, len: %u)\n",
923 _kvmRun->mmio.is_write,
924 _kvmRun->mmio.phys_addr, _kvmRun->mmio.len);
925
926 ++numMMIO;
927 return doMMIOAccess(_kvmRun->mmio.phys_addr, _kvmRun->mmio.data,
928 _kvmRun->mmio.len, _kvmRun->mmio.is_write);
929
930 case KVM_EXIT_IRQ_WINDOW_OPEN:
931 return handleKvmExitIRQWindowOpen();
932
933 case KVM_EXIT_FAIL_ENTRY:
934 return handleKvmExitFailEntry();
935
936 case KVM_EXIT_INTR:
937 /* KVM was interrupted by a signal, restart it in the next
938 * tick. */
939 return 0;
940
941 case KVM_EXIT_INTERNAL_ERROR:
942 panic("KVM: Internal error (suberror: %u)\n",
943 _kvmRun->internal.suberror);
944
945 default:
946 dump();
947 panic("KVM: Unexpected exit (exit_reason: %u)\n", _kvmRun->exit_reason);
948 }
949}
950
951Tick
952BaseKvmCPU::handleKvmExitIO()
953{
954 panic("KVM: Unhandled guest IO (dir: %i, size: %i, port: 0x%x, count: %i)\n",
955 _kvmRun->io.direction, _kvmRun->io.size,
956 _kvmRun->io.port, _kvmRun->io.count);
957}
958
959Tick
960BaseKvmCPU::handleKvmExitHypercall()
961{
962 panic("KVM: Unhandled hypercall\n");
963}
964
965Tick
966BaseKvmCPU::handleKvmExitIRQWindowOpen()
967{
968 warn("KVM: Unhandled IRQ window.\n");
969 return 0;
970}
971
972
973Tick
974BaseKvmCPU::handleKvmExitUnknown()
975{
976 dump();
977 panic("KVM: Unknown error when starting vCPU (hw reason: 0x%llx)\n",
978 _kvmRun->hw.hardware_exit_reason);
979}
980
981Tick
982BaseKvmCPU::handleKvmExitException()
983{
984 dump();
985 panic("KVM: Got exception when starting vCPU "
986 "(exception: %u, error_code: %u)\n",
987 _kvmRun->ex.exception, _kvmRun->ex.error_code);
988}
989
990Tick
991BaseKvmCPU::handleKvmExitFailEntry()
992{
993 dump();
994 panic("KVM: Failed to enter virtualized mode (hw reason: 0x%llx)\n",
995 _kvmRun->fail_entry.hardware_entry_failure_reason);
996}
997
998Tick
999BaseKvmCPU::doMMIOAccess(Addr paddr, void *data, int size, bool write)
1000{
1001 ThreadContext *tc(thread->getTC());
1002 syncThreadContext();
1003
1004 Request mmio_req(paddr, size, Request::UNCACHEABLE, dataMasterId());
1005 mmio_req.setThreadContext(tc->contextId(), 0);
1006 // Some architectures do need to massage physical addresses a bit
1007 // before they are inserted into the memory system. This enables
1008 // APIC accesses on x86 and m5ops where supported through a MMIO
1009 // interface.
1010 BaseTLB::Mode tlb_mode(write ? BaseTLB::Write : BaseTLB::Read);
1011 Fault fault(tc->getDTBPtr()->finalizePhysical(&mmio_req, tc, tlb_mode));
1012 if (fault != NoFault)
1013 warn("Finalization of MMIO address failed: %s\n", fault->name());
1014
1015
1016 const MemCmd cmd(write ? MemCmd::WriteReq : MemCmd::ReadReq);
1017 Packet pkt(&mmio_req, cmd);
1018 pkt.dataStatic(data);
1019
1020 if (mmio_req.isMmappedIpr()) {
1021 // We currently assume that there is no need to migrate to a
1022 // different event queue when doing IPRs. Currently, IPRs are
1023 // only used for m5ops, so it should be a valid assumption.
1024 const Cycles ipr_delay(write ?
1025 TheISA::handleIprWrite(tc, &pkt) :
1026 TheISA::handleIprRead(tc, &pkt));
1027 threadContextDirty = true;
1028 return clockPeriod() * ipr_delay;
1029 } else {
1030 // Temporarily lock and migrate to the event queue of the
1031 // VM. This queue is assumed to "own" all devices we need to
1032 // access if running in multi-core mode.
1033 EventQueue::ScopedMigration migrate(vm.eventQueue());
1034
1035 return dataPort.sendAtomic(&pkt);
1036 }
1037}
1038
1039void
1040BaseKvmCPU::setSignalMask(const sigset_t *mask)
1041{
1042 std::unique_ptr<struct kvm_signal_mask> kvm_mask;
1043
1044 if (mask) {
1045 kvm_mask.reset((struct kvm_signal_mask *)operator new(
1046 sizeof(struct kvm_signal_mask) + sizeof(*mask)));
1047 // The kernel and the user-space headers have different ideas
1048 // about the size of sigset_t. This seems like a massive hack,
1049 // but is actually what qemu does.
1050 assert(sizeof(*mask) >= 8);
1051 kvm_mask->len = 8;
1052 memcpy(kvm_mask->sigset, mask, kvm_mask->len);
1053 }
1054
1055 if (ioctl(KVM_SET_SIGNAL_MASK, (void *)kvm_mask.get()) == -1)
1056 panic("KVM: Failed to set vCPU signal mask (errno: %i)\n",
1057 errno);
1058}
1059
1060int
1061BaseKvmCPU::ioctl(int request, long p1) const
1062{
1063 if (vcpuFD == -1)
1064 panic("KVM: CPU ioctl called before initialization\n");
1065
1066 return ::ioctl(vcpuFD, request, p1);
1067}
1068
1069Tick
1070BaseKvmCPU::flushCoalescedMMIO()
1071{
1072 if (!mmioRing)
1073 return 0;
1074
1075 DPRINTF(KvmIO, "KVM: Flushing the coalesced MMIO ring buffer\n");
1076
1077 // TODO: We might need to do synchronization when we start to
1078 // support multiple CPUs
1079 Tick ticks(0);
1080 while (mmioRing->first != mmioRing->last) {
1081 struct kvm_coalesced_mmio &ent(
1082 mmioRing->coalesced_mmio[mmioRing->first]);
1083
1084 DPRINTF(KvmIO, "KVM: Handling coalesced MMIO (addr: 0x%x, len: %u)\n",
1085 ent.phys_addr, ent.len);
1086
1087 ++numCoalescedMMIO;
1088 ticks += doMMIOAccess(ent.phys_addr, ent.data, ent.len, true);
1089
1090 mmioRing->first = (mmioRing->first + 1) % KVM_COALESCED_MMIO_MAX;
1091 }
1092
1093 return ticks;
1094}
1095
1096/**
1097 * Dummy handler for KVM kick signals.
1098 *
1099 * @note This function is usually not called since the kernel doesn't
1100 * seem to deliver signals when the signal is only unmasked when
1101 * running in KVM. This doesn't matter though since we are only
1102 * interested in getting KVM to exit, which happens as expected. See
1103 * setupSignalHandler() and kvmRun() for details about KVM signal
1104 * handling.
1105 */
1106static void
1107onKickSignal(int signo, siginfo_t *si, void *data)
1108{
1109}
1110
1111void
1112BaseKvmCPU::setupSignalHandler()
1113{
1114 struct sigaction sa;
1115
1116 memset(&sa, 0, sizeof(sa));
1117 sa.sa_sigaction = onKickSignal;
1118 sa.sa_flags = SA_SIGINFO | SA_RESTART;
1119 if (sigaction(KVM_KICK_SIGNAL, &sa, NULL) == -1)
1120 panic("KVM: Failed to setup vCPU timer signal handler\n");
1121
1122 sigset_t sigset;
1123 if (pthread_sigmask(SIG_BLOCK, NULL, &sigset) == -1)
1124 panic("KVM: Failed get signal mask\n");
1125
1126 // Request KVM to setup the same signal mask as we're currently
1127 // running with except for the KVM control signal. We'll sometimes
1128 // need to raise the KVM_KICK_SIGNAL to cause immediate exits from
1129 // KVM after servicing IO requests. See kvmRun().
1130 sigdelset(&sigset, KVM_KICK_SIGNAL);
1131 setSignalMask(&sigset);
1132
1133 // Mask our control signals so they aren't delivered unless we're
1134 // actually executing inside KVM.
1135 sigaddset(&sigset, KVM_KICK_SIGNAL);
1136 if (pthread_sigmask(SIG_SETMASK, &sigset, NULL) == -1)
1137 panic("KVM: Failed mask the KVM control signals\n");
1138}
1139
1140bool
1141BaseKvmCPU::discardPendingSignal(int signum) const
1142{
1143 int discardedSignal;
1144
1145 // Setting the timeout to zero causes sigtimedwait to return
1146 // immediately.
1147 struct timespec timeout;
1148 timeout.tv_sec = 0;
1149 timeout.tv_nsec = 0;
1150
1151 sigset_t sigset;
1152 sigemptyset(&sigset);
1153 sigaddset(&sigset, signum);
1154
1155 do {
1156 discardedSignal = sigtimedwait(&sigset, NULL, &timeout);
1157 } while (discardedSignal == -1 && errno == EINTR);
1158
1159 if (discardedSignal == signum)
1160 return true;
1161 else if (discardedSignal == -1 && errno == EAGAIN)
1162 return false;
1163 else
1164 panic("Unexpected return value from sigtimedwait: %i (errno: %i)\n",
1165 discardedSignal, errno);
1166}
1167
1168void
1169BaseKvmCPU::setupCounters()
1170{
1171 DPRINTF(Kvm, "Attaching cycle counter...\n");
1172 PerfKvmCounterConfig cfgCycles(PERF_TYPE_HARDWARE,
1173 PERF_COUNT_HW_CPU_CYCLES);
1174 cfgCycles.disabled(true)
1175 .pinned(true);
1176
1177 // Try to exclude the host. We set both exclude_hv and
1178 // exclude_host since different architectures use slightly
1179 // different APIs in the kernel.
1180 cfgCycles.exclude_hv(true)
1181 .exclude_host(true);
1182
1183 if (perfControlledByTimer) {
1184 // We need to configure the cycles counter to send overflows
1185 // since we are going to use it to trigger timer signals that
1186 // trap back into m5 from KVM. In practice, this means that we
1187 // need to set some non-zero sample period that gets
1188 // overridden when the timer is armed.
1189 cfgCycles.wakeupEvents(1)
1190 .samplePeriod(42);
1191 }
1192
1193 hwCycles.attach(cfgCycles,
1194 0); // TID (0 => currentThread)
1195
1196 setupInstCounter();
1197}
1198
1199bool
1200BaseKvmCPU::tryDrain()
1201{
1202 if (!drainManager)
1203 return false;
1204
1205 if (!archIsDrained()) {
1206 DPRINTF(Drain, "tryDrain: Architecture code is not ready.\n");
1207 return false;
1208 }
1209
1210 if (_status == Idle || _status == Running) {
1211 DPRINTF(Drain,
1212 "tryDrain: CPU transitioned into the Idle state, drain done\n");
1213 drainManager->signalDrainDone();
1214 drainManager = NULL;
1215 return true;
1216 } else {
1217 DPRINTF(Drain, "tryDrain: CPU not ready.\n");
1218 return false;
1219 }
1220}
1221
1222void
1223BaseKvmCPU::ioctlRun()
1224{
1225 if (ioctl(KVM_RUN) == -1) {
1226 if (errno != EINTR)
1227 panic("KVM: Failed to start virtual CPU (errno: %i)\n",
1228 errno);
1229 }
1230}
1231
1232void
1233BaseKvmCPU::setupInstStop()
1234{
1235 if (comInstEventQueue[0]->empty()) {
1236 setupInstCounter(0);
1237 } else {
1238 const uint64_t next(comInstEventQueue[0]->nextTick());
1239
1240 assert(next > ctrInsts);
1241 setupInstCounter(next - ctrInsts);
1242 }
1243}
1244
1245void
1246BaseKvmCPU::setupInstCounter(uint64_t period)
1247{
1248 // No need to do anything if we aren't attaching for the first
1249 // time or the period isn't changing.
1250 if (period == activeInstPeriod && hwInstructions.attached())
1251 return;
1252
1253 PerfKvmCounterConfig cfgInstructions(PERF_TYPE_HARDWARE,
1254 PERF_COUNT_HW_INSTRUCTIONS);
1255
1256 // Try to exclude the host. We set both exclude_hv and
1257 // exclude_host since different architectures use slightly
1258 // different APIs in the kernel.
1259 cfgInstructions.exclude_hv(true)
1260 .exclude_host(true);
1261
1262 if (period) {
1263 // Setup a sampling counter if that has been requested.
1264 cfgInstructions.wakeupEvents(1)
1265 .samplePeriod(period);
1266 }
1267
1268 // We need to detach and re-attach the counter to reliably change
1269 // sampling settings. See PerfKvmCounter::period() for details.
1270 if (hwInstructions.attached())
1271 hwInstructions.detach();
1272 assert(hwCycles.attached());
1273 hwInstructions.attach(cfgInstructions,
1274 0, // TID (0 => currentThread)
1275 hwCycles);
1276
1277 if (period)
1278 hwInstructions.enableSignals(KVM_KICK_SIGNAL);
1279
1280 activeInstPeriod = period;
1281}
|