1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * A memslot-related performance benchmark.
4  *
5  * Copyright (C) 2021 Oracle and/or its affiliates.
6  *
7  * Basic guest setup / host vCPU thread code lifted from set_memory_region_test.
8  */
9 #include <pthread.h>
10 #include <sched.h>
11 #include <semaphore.h>
12 #include <stdatomic.h>
13 #include <stdbool.h>
14 #include <stdint.h>
15 #include <stdio.h>
16 #include <stdlib.h>
17 #include <string.h>
18 #include <sys/mman.h>
19 #include <time.h>
20 #include <unistd.h>
21 
22 #include <linux/compiler.h>
23 #include <linux/sizes.h>
24 
25 #include <test_util.h>
26 #include <kvm_util.h>
27 #include <processor.h>
28 
29 #define MEM_EXTRA_SIZE		SZ_64K
30 
31 #define MEM_SIZE		(SZ_512M + MEM_EXTRA_SIZE)
32 #define MEM_GPA			SZ_256M
33 #define MEM_AUX_GPA		MEM_GPA
34 #define MEM_SYNC_GPA		MEM_AUX_GPA
35 #define MEM_TEST_GPA		(MEM_AUX_GPA + MEM_EXTRA_SIZE)
36 #define MEM_TEST_SIZE		(MEM_SIZE - MEM_EXTRA_SIZE)
37 
38 /*
39  * 32 MiB is max size that gets well over 100 iterations on 509 slots.
40  * Considering that each slot needs to have at least one page up to
41  * 8194 slots in use can then be tested (although with slightly
42  * limited resolution).
43  */
44 #define MEM_SIZE_MAP		(SZ_32M + MEM_EXTRA_SIZE)
45 #define MEM_TEST_MAP_SIZE	(MEM_SIZE_MAP - MEM_EXTRA_SIZE)
46 
47 /*
48  * 128 MiB is min size that fills 32k slots with at least one page in each
49  * while at the same time gets 100+ iterations in such test
50  *
51  * 2 MiB chunk size like a typical huge page
52  */
53 #define MEM_TEST_UNMAP_SIZE		SZ_128M
54 #define MEM_TEST_UNMAP_CHUNK_SIZE	SZ_2M
55 
56 /*
57  * For the move active test the middle of the test area is placed on
58  * a memslot boundary: half lies in the memslot being moved, half in
59  * other memslot(s).
60  *
61  * We have different number of memory slots, excluding the reserved
62  * memory slot 0, on various architectures and configurations. The
63  * memory size in this test is calculated by picking the maximal
64  * last memory slot's memory size, with alignment to the largest
65  * supported page size (64KB). In this way, the selected memory
66  * size for this test is compatible with test_memslot_move_prepare().
67  *
68  * architecture   slots    memory-per-slot    memory-on-last-slot
69  * --------------------------------------------------------------
70  * x86-4KB        32763    16KB               160KB
71  * arm64-4KB      32766    16KB               112KB
72  * arm64-16KB     32766    16KB               112KB
73  * arm64-64KB     8192     64KB               128KB
74  */
75 #define MEM_TEST_MOVE_SIZE		(3 * SZ_64K)
76 #define MEM_TEST_MOVE_GPA_DEST		(MEM_GPA + MEM_SIZE)
77 static_assert(MEM_TEST_MOVE_SIZE <= MEM_TEST_SIZE,
78 	      "invalid move test region size");
79 
80 #define MEM_TEST_VAL_1 0x1122334455667788
81 #define MEM_TEST_VAL_2 0x99AABBCCDDEEFF00
82 
83 struct vm_data {
84 	struct kvm_vm *vm;
85 	struct kvm_vcpu *vcpu;
86 	pthread_t vcpu_thread;
87 	uint32_t nslots;
88 	uint64_t npages;
89 	uint64_t pages_per_slot;
90 	void **hva_slots;
91 	bool mmio_ok;
92 	uint64_t mmio_gpa_min;
93 	uint64_t mmio_gpa_max;
94 };
95 
96 struct sync_area {
97 	uint32_t    guest_page_size;
98 	atomic_bool start_flag;
99 	atomic_bool exit_flag;
100 	atomic_bool sync_flag;
101 	void *move_area_ptr;
102 };
103 
104 /*
105  * Technically, we need also for the atomic bool to be address-free, which
106  * is recommended, but not strictly required, by C11 for lockless
107  * implementations.
108  * However, in practice both GCC and Clang fulfill this requirement on
109  * all KVM-supported platforms.
110  */
111 static_assert(ATOMIC_BOOL_LOCK_FREE == 2, "atomic bool is not lockless");
112 
113 static sem_t vcpu_ready;
114 
115 static bool map_unmap_verify;
116 
117 static bool verbose;
118 #define pr_info_v(...)				\
119 	do {					\
120 		if (verbose)			\
121 			pr_info(__VA_ARGS__);	\
122 	} while (0)
123 
124 static void check_mmio_access(struct vm_data *data, struct kvm_run *run)
125 {
126 	TEST_ASSERT(data->mmio_ok, "Unexpected mmio exit");
127 	TEST_ASSERT(run->mmio.is_write, "Unexpected mmio read");
128 	TEST_ASSERT(run->mmio.len == 8,
129 		    "Unexpected exit mmio size = %u", run->mmio.len);
130 	TEST_ASSERT(run->mmio.phys_addr >= data->mmio_gpa_min &&
131 		    run->mmio.phys_addr <= data->mmio_gpa_max,
132 		    "Unexpected exit mmio address = 0x%llx",
133 		    run->mmio.phys_addr);
134 }
135 
136 static void *vcpu_worker(void *__data)
137 {
138 	struct vm_data *data = __data;
139 	struct kvm_vcpu *vcpu = data->vcpu;
140 	struct kvm_run *run = vcpu->run;
141 	struct ucall uc;
142 
143 	while (1) {
144 		vcpu_run(vcpu);
145 
146 		switch (get_ucall(vcpu, &uc)) {
147 		case UCALL_SYNC:
148 			TEST_ASSERT(uc.args[1] == 0,
149 				"Unexpected sync ucall, got %lx",
150 				(ulong)uc.args[1]);
151 			sem_post(&vcpu_ready);
152 			continue;
153 		case UCALL_NONE:
154 			if (run->exit_reason == KVM_EXIT_MMIO)
155 				check_mmio_access(data, run);
156 			else
157 				goto done;
158 			break;
159 		case UCALL_ABORT:
160 			REPORT_GUEST_ASSERT_1(uc, "val = %lu");
161 			break;
162 		case UCALL_DONE:
163 			goto done;
164 		default:
165 			TEST_FAIL("Unknown ucall %lu", uc.cmd);
166 		}
167 	}
168 
169 done:
170 	return NULL;
171 }
172 
173 static void wait_for_vcpu(void)
174 {
175 	struct timespec ts;
176 
177 	TEST_ASSERT(!clock_gettime(CLOCK_REALTIME, &ts),
178 		    "clock_gettime() failed: %d\n", errno);
179 
180 	ts.tv_sec += 2;
181 	TEST_ASSERT(!sem_timedwait(&vcpu_ready, &ts),
182 		    "sem_timedwait() failed: %d\n", errno);
183 }
184 
185 static void *vm_gpa2hva(struct vm_data *data, uint64_t gpa, uint64_t *rempages)
186 {
187 	uint64_t gpage, pgoffs;
188 	uint32_t slot, slotoffs;
189 	void *base;
190 	uint32_t guest_page_size = data->vm->page_size;
191 
192 	TEST_ASSERT(gpa >= MEM_GPA, "Too low gpa to translate");
193 	TEST_ASSERT(gpa < MEM_GPA + data->npages * guest_page_size,
194 		    "Too high gpa to translate");
195 	gpa -= MEM_GPA;
196 
197 	gpage = gpa / guest_page_size;
198 	pgoffs = gpa % guest_page_size;
199 	slot = min(gpage / data->pages_per_slot, (uint64_t)data->nslots - 1);
200 	slotoffs = gpage - (slot * data->pages_per_slot);
201 
202 	if (rempages) {
203 		uint64_t slotpages;
204 
205 		if (slot == data->nslots - 1)
206 			slotpages = data->npages - slot * data->pages_per_slot;
207 		else
208 			slotpages = data->pages_per_slot;
209 
210 		TEST_ASSERT(!pgoffs,
211 			    "Asking for remaining pages in slot but gpa not page aligned");
212 		*rempages = slotpages - slotoffs;
213 	}
214 
215 	base = data->hva_slots[slot];
216 	return (uint8_t *)base + slotoffs * guest_page_size + pgoffs;
217 }
218 
219 static uint64_t vm_slot2gpa(struct vm_data *data, uint32_t slot)
220 {
221 	uint32_t guest_page_size = data->vm->page_size;
222 
223 	TEST_ASSERT(slot < data->nslots, "Too high slot number");
224 
225 	return MEM_GPA + slot * data->pages_per_slot * guest_page_size;
226 }
227 
228 static struct vm_data *alloc_vm(void)
229 {
230 	struct vm_data *data;
231 
232 	data = malloc(sizeof(*data));
233 	TEST_ASSERT(data, "malloc(vmdata) failed");
234 
235 	data->vm = NULL;
236 	data->vcpu = NULL;
237 	data->hva_slots = NULL;
238 
239 	return data;
240 }
241 
242 static bool check_slot_pages(uint32_t host_page_size, uint32_t guest_page_size,
243 			     uint64_t pages_per_slot, uint64_t rempages)
244 {
245 	if (!pages_per_slot)
246 		return false;
247 
248 	if ((pages_per_slot * guest_page_size) % host_page_size)
249 		return false;
250 
251 	if ((rempages * guest_page_size) % host_page_size)
252 		return false;
253 
254 	return true;
255 }
256 
257 
258 static uint64_t get_max_slots(struct vm_data *data, uint32_t host_page_size)
259 {
260 	uint32_t guest_page_size = data->vm->page_size;
261 	uint64_t mempages, pages_per_slot, rempages;
262 	uint64_t slots;
263 
264 	mempages = data->npages;
265 	slots = data->nslots;
266 	while (--slots > 1) {
267 		pages_per_slot = mempages / slots;
268 		if (!pages_per_slot)
269 			continue;
270 
271 		rempages = mempages % pages_per_slot;
272 		if (check_slot_pages(host_page_size, guest_page_size,
273 				     pages_per_slot, rempages))
274 			return slots + 1;	/* slot 0 is reserved */
275 	}
276 
277 	return 0;
278 }
279 
280 static bool prepare_vm(struct vm_data *data, int nslots, uint64_t *maxslots,
281 		       void *guest_code, uint64_t mem_size,
282 		       struct timespec *slot_runtime)
283 {
284 	uint64_t mempages, rempages;
285 	uint64_t guest_addr;
286 	uint32_t slot, host_page_size, guest_page_size;
287 	struct timespec tstart;
288 	struct sync_area *sync;
289 
290 	host_page_size = getpagesize();
291 	guest_page_size = vm_guest_mode_params[VM_MODE_DEFAULT].page_size;
292 	mempages = mem_size / guest_page_size;
293 
294 	data->vm = __vm_create_with_one_vcpu(&data->vcpu, mempages, guest_code);
295 	TEST_ASSERT(data->vm->page_size == guest_page_size, "Invalid VM page size");
296 
297 	data->npages = mempages;
298 	TEST_ASSERT(data->npages > 1, "Can't test without any memory");
299 	data->nslots = nslots;
300 	data->pages_per_slot = data->npages / data->nslots;
301 	rempages = data->npages % data->nslots;
302 	if (!check_slot_pages(host_page_size, guest_page_size,
303 			      data->pages_per_slot, rempages)) {
304 		*maxslots = get_max_slots(data, host_page_size);
305 		return false;
306 	}
307 
308 	data->hva_slots = malloc(sizeof(*data->hva_slots) * data->nslots);
309 	TEST_ASSERT(data->hva_slots, "malloc() fail");
310 
311 	pr_info_v("Adding slots 1..%i, each slot with %"PRIu64" pages + %"PRIu64" extra pages last\n",
312 		data->nslots, data->pages_per_slot, rempages);
313 
314 	clock_gettime(CLOCK_MONOTONIC, &tstart);
315 	for (slot = 1, guest_addr = MEM_GPA; slot <= data->nslots; slot++) {
316 		uint64_t npages;
317 
318 		npages = data->pages_per_slot;
319 		if (slot == data->nslots)
320 			npages += rempages;
321 
322 		vm_userspace_mem_region_add(data->vm, VM_MEM_SRC_ANONYMOUS,
323 					    guest_addr, slot, npages,
324 					    0);
325 		guest_addr += npages * guest_page_size;
326 	}
327 	*slot_runtime = timespec_elapsed(tstart);
328 
329 	for (slot = 1, guest_addr = MEM_GPA; slot <= data->nslots; slot++) {
330 		uint64_t npages;
331 		uint64_t gpa;
332 
333 		npages = data->pages_per_slot;
334 		if (slot == data->nslots)
335 			npages += rempages;
336 
337 		gpa = vm_phy_pages_alloc(data->vm, npages, guest_addr, slot);
338 		TEST_ASSERT(gpa == guest_addr,
339 			    "vm_phy_pages_alloc() failed\n");
340 
341 		data->hva_slots[slot - 1] = addr_gpa2hva(data->vm, guest_addr);
342 		memset(data->hva_slots[slot - 1], 0, npages * guest_page_size);
343 
344 		guest_addr += npages * guest_page_size;
345 	}
346 
347 	virt_map(data->vm, MEM_GPA, MEM_GPA, data->npages);
348 
349 	sync = (typeof(sync))vm_gpa2hva(data, MEM_SYNC_GPA, NULL);
350 	sync->guest_page_size = data->vm->page_size;
351 	atomic_init(&sync->start_flag, false);
352 	atomic_init(&sync->exit_flag, false);
353 	atomic_init(&sync->sync_flag, false);
354 
355 	data->mmio_ok = false;
356 
357 	return true;
358 }
359 
360 static void launch_vm(struct vm_data *data)
361 {
362 	pr_info_v("Launching the test VM\n");
363 
364 	pthread_create(&data->vcpu_thread, NULL, vcpu_worker, data);
365 
366 	/* Ensure the guest thread is spun up. */
367 	wait_for_vcpu();
368 }
369 
370 static void free_vm(struct vm_data *data)
371 {
372 	kvm_vm_free(data->vm);
373 	free(data->hva_slots);
374 	free(data);
375 }
376 
377 static void wait_guest_exit(struct vm_data *data)
378 {
379 	pthread_join(data->vcpu_thread, NULL);
380 }
381 
382 static void let_guest_run(struct sync_area *sync)
383 {
384 	atomic_store_explicit(&sync->start_flag, true, memory_order_release);
385 }
386 
387 static void guest_spin_until_start(void)
388 {
389 	struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
390 
391 	while (!atomic_load_explicit(&sync->start_flag, memory_order_acquire))
392 		;
393 }
394 
395 static void make_guest_exit(struct sync_area *sync)
396 {
397 	atomic_store_explicit(&sync->exit_flag, true, memory_order_release);
398 }
399 
400 static bool _guest_should_exit(void)
401 {
402 	struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
403 
404 	return atomic_load_explicit(&sync->exit_flag, memory_order_acquire);
405 }
406 
407 #define guest_should_exit() unlikely(_guest_should_exit())
408 
409 /*
410  * noinline so we can easily see how much time the host spends waiting
411  * for the guest.
412  * For the same reason use alarm() instead of polling clock_gettime()
413  * to implement a wait timeout.
414  */
415 static noinline void host_perform_sync(struct sync_area *sync)
416 {
417 	alarm(2);
418 
419 	atomic_store_explicit(&sync->sync_flag, true, memory_order_release);
420 	while (atomic_load_explicit(&sync->sync_flag, memory_order_acquire))
421 		;
422 
423 	alarm(0);
424 }
425 
426 static bool guest_perform_sync(void)
427 {
428 	struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
429 	bool expected;
430 
431 	do {
432 		if (guest_should_exit())
433 			return false;
434 
435 		expected = true;
436 	} while (!atomic_compare_exchange_weak_explicit(&sync->sync_flag,
437 							&expected, false,
438 							memory_order_acq_rel,
439 							memory_order_relaxed));
440 
441 	return true;
442 }
443 
444 static void guest_code_test_memslot_move(void)
445 {
446 	struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
447 	uint32_t page_size = (typeof(page_size))READ_ONCE(sync->guest_page_size);
448 	uintptr_t base = (typeof(base))READ_ONCE(sync->move_area_ptr);
449 
450 	GUEST_SYNC(0);
451 
452 	guest_spin_until_start();
453 
454 	while (!guest_should_exit()) {
455 		uintptr_t ptr;
456 
457 		for (ptr = base; ptr < base + MEM_TEST_MOVE_SIZE;
458 		     ptr += page_size)
459 			*(uint64_t *)ptr = MEM_TEST_VAL_1;
460 
461 		/*
462 		 * No host sync here since the MMIO exits are so expensive
463 		 * that the host would spend most of its time waiting for
464 		 * the guest and so instead of measuring memslot move
465 		 * performance we would measure the performance and
466 		 * likelihood of MMIO exits
467 		 */
468 	}
469 
470 	GUEST_DONE();
471 }
472 
473 static void guest_code_test_memslot_map(void)
474 {
475 	struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
476 	uint32_t page_size = (typeof(page_size))READ_ONCE(sync->guest_page_size);
477 
478 	GUEST_SYNC(0);
479 
480 	guest_spin_until_start();
481 
482 	while (1) {
483 		uintptr_t ptr;
484 
485 		for (ptr = MEM_TEST_GPA;
486 		     ptr < MEM_TEST_GPA + MEM_TEST_MAP_SIZE / 2;
487 		     ptr += page_size)
488 			*(uint64_t *)ptr = MEM_TEST_VAL_1;
489 
490 		if (!guest_perform_sync())
491 			break;
492 
493 		for (ptr = MEM_TEST_GPA + MEM_TEST_MAP_SIZE / 2;
494 		     ptr < MEM_TEST_GPA + MEM_TEST_MAP_SIZE;
495 		     ptr += page_size)
496 			*(uint64_t *)ptr = MEM_TEST_VAL_2;
497 
498 		if (!guest_perform_sync())
499 			break;
500 	}
501 
502 	GUEST_DONE();
503 }
504 
505 static void guest_code_test_memslot_unmap(void)
506 {
507 	struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
508 
509 	GUEST_SYNC(0);
510 
511 	guest_spin_until_start();
512 
513 	while (1) {
514 		uintptr_t ptr = MEM_TEST_GPA;
515 
516 		/*
517 		 * We can afford to access (map) just a small number of pages
518 		 * per host sync as otherwise the host will spend
519 		 * a significant amount of its time waiting for the guest
520 		 * (instead of doing unmap operations), so this will
521 		 * effectively turn this test into a map performance test.
522 		 *
523 		 * Just access a single page to be on the safe side.
524 		 */
525 		*(uint64_t *)ptr = MEM_TEST_VAL_1;
526 
527 		if (!guest_perform_sync())
528 			break;
529 
530 		ptr += MEM_TEST_UNMAP_SIZE / 2;
531 		*(uint64_t *)ptr = MEM_TEST_VAL_2;
532 
533 		if (!guest_perform_sync())
534 			break;
535 	}
536 
537 	GUEST_DONE();
538 }
539 
540 static void guest_code_test_memslot_rw(void)
541 {
542 	struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
543 	uint32_t page_size = (typeof(page_size))READ_ONCE(sync->guest_page_size);
544 
545 	GUEST_SYNC(0);
546 
547 	guest_spin_until_start();
548 
549 	while (1) {
550 		uintptr_t ptr;
551 
552 		for (ptr = MEM_TEST_GPA;
553 		     ptr < MEM_TEST_GPA + MEM_TEST_SIZE; ptr += page_size)
554 			*(uint64_t *)ptr = MEM_TEST_VAL_1;
555 
556 		if (!guest_perform_sync())
557 			break;
558 
559 		for (ptr = MEM_TEST_GPA + page_size / 2;
560 		     ptr < MEM_TEST_GPA + MEM_TEST_SIZE; ptr += page_size) {
561 			uint64_t val = *(uint64_t *)ptr;
562 
563 			GUEST_ASSERT_1(val == MEM_TEST_VAL_2, val);
564 			*(uint64_t *)ptr = 0;
565 		}
566 
567 		if (!guest_perform_sync())
568 			break;
569 	}
570 
571 	GUEST_DONE();
572 }
573 
574 static bool test_memslot_move_prepare(struct vm_data *data,
575 				      struct sync_area *sync,
576 				      uint64_t *maxslots, bool isactive)
577 {
578 	uint32_t guest_page_size = data->vm->page_size;
579 	uint64_t movesrcgpa, movetestgpa;
580 
581 	movesrcgpa = vm_slot2gpa(data, data->nslots - 1);
582 
583 	if (isactive) {
584 		uint64_t lastpages;
585 
586 		vm_gpa2hva(data, movesrcgpa, &lastpages);
587 		if (lastpages * guest_page_size < MEM_TEST_MOVE_SIZE / 2) {
588 			*maxslots = 0;
589 			return false;
590 		}
591 	}
592 
593 	movetestgpa = movesrcgpa - (MEM_TEST_MOVE_SIZE / (isactive ? 2 : 1));
594 	sync->move_area_ptr = (void *)movetestgpa;
595 
596 	if (isactive) {
597 		data->mmio_ok = true;
598 		data->mmio_gpa_min = movesrcgpa;
599 		data->mmio_gpa_max = movesrcgpa + MEM_TEST_MOVE_SIZE / 2 - 1;
600 	}
601 
602 	return true;
603 }
604 
605 static bool test_memslot_move_prepare_active(struct vm_data *data,
606 					     struct sync_area *sync,
607 					     uint64_t *maxslots)
608 {
609 	return test_memslot_move_prepare(data, sync, maxslots, true);
610 }
611 
612 static bool test_memslot_move_prepare_inactive(struct vm_data *data,
613 					       struct sync_area *sync,
614 					       uint64_t *maxslots)
615 {
616 	return test_memslot_move_prepare(data, sync, maxslots, false);
617 }
618 
619 static void test_memslot_move_loop(struct vm_data *data, struct sync_area *sync)
620 {
621 	uint64_t movesrcgpa;
622 
623 	movesrcgpa = vm_slot2gpa(data, data->nslots - 1);
624 	vm_mem_region_move(data->vm, data->nslots - 1 + 1,
625 			   MEM_TEST_MOVE_GPA_DEST);
626 	vm_mem_region_move(data->vm, data->nslots - 1 + 1, movesrcgpa);
627 }
628 
629 static void test_memslot_do_unmap(struct vm_data *data,
630 				  uint64_t offsp, uint64_t count)
631 {
632 	uint64_t gpa, ctr;
633 	uint32_t guest_page_size = data->vm->page_size;
634 
635 	for (gpa = MEM_TEST_GPA + offsp * guest_page_size, ctr = 0; ctr < count; ) {
636 		uint64_t npages;
637 		void *hva;
638 		int ret;
639 
640 		hva = vm_gpa2hva(data, gpa, &npages);
641 		TEST_ASSERT(npages, "Empty memory slot at gptr 0x%"PRIx64, gpa);
642 		npages = min(npages, count - ctr);
643 		ret = madvise(hva, npages * guest_page_size, MADV_DONTNEED);
644 		TEST_ASSERT(!ret,
645 			    "madvise(%p, MADV_DONTNEED) on VM memory should not fail for gptr 0x%"PRIx64,
646 			    hva, gpa);
647 		ctr += npages;
648 		gpa += npages * guest_page_size;
649 	}
650 	TEST_ASSERT(ctr == count,
651 		    "madvise(MADV_DONTNEED) should exactly cover all of the requested area");
652 }
653 
654 static void test_memslot_map_unmap_check(struct vm_data *data,
655 					 uint64_t offsp, uint64_t valexp)
656 {
657 	uint64_t gpa;
658 	uint64_t *val;
659 	uint32_t guest_page_size = data->vm->page_size;
660 
661 	if (!map_unmap_verify)
662 		return;
663 
664 	gpa = MEM_TEST_GPA + offsp * guest_page_size;
665 	val = (typeof(val))vm_gpa2hva(data, gpa, NULL);
666 	TEST_ASSERT(*val == valexp,
667 		    "Guest written values should read back correctly before unmap (%"PRIu64" vs %"PRIu64" @ %"PRIx64")",
668 		    *val, valexp, gpa);
669 	*val = 0;
670 }
671 
672 static void test_memslot_map_loop(struct vm_data *data, struct sync_area *sync)
673 {
674 	uint32_t guest_page_size = data->vm->page_size;
675 	uint64_t guest_pages = MEM_TEST_MAP_SIZE / guest_page_size;
676 
677 	/*
678 	 * Unmap the second half of the test area while guest writes to (maps)
679 	 * the first half.
680 	 */
681 	test_memslot_do_unmap(data, guest_pages / 2, guest_pages / 2);
682 
683 	/*
684 	 * Wait for the guest to finish writing the first half of the test
685 	 * area, verify the written value on the first and the last page of
686 	 * this area and then unmap it.
687 	 * Meanwhile, the guest is writing to (mapping) the second half of
688 	 * the test area.
689 	 */
690 	host_perform_sync(sync);
691 	test_memslot_map_unmap_check(data, 0, MEM_TEST_VAL_1);
692 	test_memslot_map_unmap_check(data, guest_pages / 2 - 1, MEM_TEST_VAL_1);
693 	test_memslot_do_unmap(data, 0, guest_pages / 2);
694 
695 
696 	/*
697 	 * Wait for the guest to finish writing the second half of the test
698 	 * area and verify the written value on the first and the last page
699 	 * of this area.
700 	 * The area will be unmapped at the beginning of the next loop
701 	 * iteration.
702 	 * Meanwhile, the guest is writing to (mapping) the first half of
703 	 * the test area.
704 	 */
705 	host_perform_sync(sync);
706 	test_memslot_map_unmap_check(data, guest_pages / 2, MEM_TEST_VAL_2);
707 	test_memslot_map_unmap_check(data, guest_pages - 1, MEM_TEST_VAL_2);
708 }
709 
710 static void test_memslot_unmap_loop_common(struct vm_data *data,
711 					   struct sync_area *sync,
712 					   uint64_t chunk)
713 {
714 	uint32_t guest_page_size = data->vm->page_size;
715 	uint64_t guest_pages = MEM_TEST_UNMAP_SIZE / guest_page_size;
716 	uint64_t ctr;
717 
718 	/*
719 	 * Wait for the guest to finish mapping page(s) in the first half
720 	 * of the test area, verify the written value and then perform unmap
721 	 * of this area.
722 	 * Meanwhile, the guest is writing to (mapping) page(s) in the second
723 	 * half of the test area.
724 	 */
725 	host_perform_sync(sync);
726 	test_memslot_map_unmap_check(data, 0, MEM_TEST_VAL_1);
727 	for (ctr = 0; ctr < guest_pages / 2; ctr += chunk)
728 		test_memslot_do_unmap(data, ctr, chunk);
729 
730 	/* Likewise, but for the opposite host / guest areas */
731 	host_perform_sync(sync);
732 	test_memslot_map_unmap_check(data, guest_pages / 2, MEM_TEST_VAL_2);
733 	for (ctr = guest_pages / 2; ctr < guest_pages; ctr += chunk)
734 		test_memslot_do_unmap(data, ctr, chunk);
735 }
736 
737 static void test_memslot_unmap_loop(struct vm_data *data,
738 				    struct sync_area *sync)
739 {
740 	uint32_t host_page_size = getpagesize();
741 	uint32_t guest_page_size = data->vm->page_size;
742 	uint64_t guest_chunk_pages = guest_page_size >= host_page_size ?
743 					1 : host_page_size / guest_page_size;
744 
745 	test_memslot_unmap_loop_common(data, sync, guest_chunk_pages);
746 }
747 
748 static void test_memslot_unmap_loop_chunked(struct vm_data *data,
749 					    struct sync_area *sync)
750 {
751 	uint32_t guest_page_size = data->vm->page_size;
752 	uint64_t guest_chunk_pages = MEM_TEST_UNMAP_CHUNK_SIZE / guest_page_size;
753 
754 	test_memslot_unmap_loop_common(data, sync, guest_chunk_pages);
755 }
756 
757 static void test_memslot_rw_loop(struct vm_data *data, struct sync_area *sync)
758 {
759 	uint64_t gptr;
760 	uint32_t guest_page_size = data->vm->page_size;
761 
762 	for (gptr = MEM_TEST_GPA + guest_page_size / 2;
763 	     gptr < MEM_TEST_GPA + MEM_TEST_SIZE; gptr += guest_page_size)
764 		*(uint64_t *)vm_gpa2hva(data, gptr, NULL) = MEM_TEST_VAL_2;
765 
766 	host_perform_sync(sync);
767 
768 	for (gptr = MEM_TEST_GPA;
769 	     gptr < MEM_TEST_GPA + MEM_TEST_SIZE; gptr += guest_page_size) {
770 		uint64_t *vptr = (typeof(vptr))vm_gpa2hva(data, gptr, NULL);
771 		uint64_t val = *vptr;
772 
773 		TEST_ASSERT(val == MEM_TEST_VAL_1,
774 			    "Guest written values should read back correctly (is %"PRIu64" @ %"PRIx64")",
775 			    val, gptr);
776 		*vptr = 0;
777 	}
778 
779 	host_perform_sync(sync);
780 }
781 
782 struct test_data {
783 	const char *name;
784 	uint64_t mem_size;
785 	void (*guest_code)(void);
786 	bool (*prepare)(struct vm_data *data, struct sync_area *sync,
787 			uint64_t *maxslots);
788 	void (*loop)(struct vm_data *data, struct sync_area *sync);
789 };
790 
791 static bool test_execute(int nslots, uint64_t *maxslots,
792 			 unsigned int maxtime,
793 			 const struct test_data *tdata,
794 			 uint64_t *nloops,
795 			 struct timespec *slot_runtime,
796 			 struct timespec *guest_runtime)
797 {
798 	uint64_t mem_size = tdata->mem_size ? : MEM_SIZE;
799 	struct vm_data *data;
800 	struct sync_area *sync;
801 	struct timespec tstart;
802 	bool ret = true;
803 
804 	data = alloc_vm();
805 	if (!prepare_vm(data, nslots, maxslots, tdata->guest_code,
806 			mem_size, slot_runtime)) {
807 		ret = false;
808 		goto exit_free;
809 	}
810 
811 	sync = (typeof(sync))vm_gpa2hva(data, MEM_SYNC_GPA, NULL);
812 	if (tdata->prepare &&
813 	    !tdata->prepare(data, sync, maxslots)) {
814 		ret = false;
815 		goto exit_free;
816 	}
817 
818 	launch_vm(data);
819 
820 	clock_gettime(CLOCK_MONOTONIC, &tstart);
821 	let_guest_run(sync);
822 
823 	while (1) {
824 		*guest_runtime = timespec_elapsed(tstart);
825 		if (guest_runtime->tv_sec >= maxtime)
826 			break;
827 
828 		tdata->loop(data, sync);
829 
830 		(*nloops)++;
831 	}
832 
833 	make_guest_exit(sync);
834 	wait_guest_exit(data);
835 
836 exit_free:
837 	free_vm(data);
838 
839 	return ret;
840 }
841 
842 static const struct test_data tests[] = {
843 	{
844 		.name = "map",
845 		.mem_size = MEM_SIZE_MAP,
846 		.guest_code = guest_code_test_memslot_map,
847 		.loop = test_memslot_map_loop,
848 	},
849 	{
850 		.name = "unmap",
851 		.mem_size = MEM_TEST_UNMAP_SIZE + MEM_EXTRA_SIZE,
852 		.guest_code = guest_code_test_memslot_unmap,
853 		.loop = test_memslot_unmap_loop,
854 	},
855 	{
856 		.name = "unmap chunked",
857 		.mem_size = MEM_TEST_UNMAP_SIZE + MEM_EXTRA_SIZE,
858 		.guest_code = guest_code_test_memslot_unmap,
859 		.loop = test_memslot_unmap_loop_chunked,
860 	},
861 	{
862 		.name = "move active area",
863 		.guest_code = guest_code_test_memslot_move,
864 		.prepare = test_memslot_move_prepare_active,
865 		.loop = test_memslot_move_loop,
866 	},
867 	{
868 		.name = "move inactive area",
869 		.guest_code = guest_code_test_memslot_move,
870 		.prepare = test_memslot_move_prepare_inactive,
871 		.loop = test_memslot_move_loop,
872 	},
873 	{
874 		.name = "RW",
875 		.guest_code = guest_code_test_memslot_rw,
876 		.loop = test_memslot_rw_loop
877 	},
878 };
879 
880 #define NTESTS ARRAY_SIZE(tests)
881 
882 struct test_args {
883 	int tfirst;
884 	int tlast;
885 	int nslots;
886 	int seconds;
887 	int runs;
888 };
889 
890 static void help(char *name, struct test_args *targs)
891 {
892 	int ctr;
893 
894 	pr_info("usage: %s [-h] [-v] [-d] [-s slots] [-f first_test] [-e last_test] [-l test_length] [-r run_count]\n",
895 		name);
896 	pr_info(" -h: print this help screen.\n");
897 	pr_info(" -v: enable verbose mode (not for benchmarking).\n");
898 	pr_info(" -d: enable extra debug checks.\n");
899 	pr_info(" -s: specify memslot count cap (-1 means no cap; currently: %i)\n",
900 		targs->nslots);
901 	pr_info(" -f: specify the first test to run (currently: %i; max %zu)\n",
902 		targs->tfirst, NTESTS - 1);
903 	pr_info(" -e: specify the last test to run (currently: %i; max %zu)\n",
904 		targs->tlast, NTESTS - 1);
905 	pr_info(" -l: specify the test length in seconds (currently: %i)\n",
906 		targs->seconds);
907 	pr_info(" -r: specify the number of runs per test (currently: %i)\n",
908 		targs->runs);
909 
910 	pr_info("\nAvailable tests:\n");
911 	for (ctr = 0; ctr < NTESTS; ctr++)
912 		pr_info("%d: %s\n", ctr, tests[ctr].name);
913 }
914 
915 static bool check_memory_sizes(void)
916 {
917 	uint32_t host_page_size = getpagesize();
918 	uint32_t guest_page_size = vm_guest_mode_params[VM_MODE_DEFAULT].page_size;
919 
920 	if (host_page_size > SZ_64K || guest_page_size > SZ_64K) {
921 		pr_info("Unsupported page size on host (0x%x) or guest (0x%x)\n",
922 			host_page_size, guest_page_size);
923 		return false;
924 	}
925 
926 	if (MEM_SIZE % guest_page_size ||
927 	    MEM_TEST_SIZE % guest_page_size) {
928 		pr_info("invalid MEM_SIZE or MEM_TEST_SIZE\n");
929 		return false;
930 	}
931 
932 	if (MEM_SIZE_MAP % guest_page_size		||
933 	    MEM_TEST_MAP_SIZE % guest_page_size		||
934 	    (MEM_TEST_MAP_SIZE / guest_page_size) <= 2	||
935 	    (MEM_TEST_MAP_SIZE / guest_page_size) % 2) {
936 		pr_info("invalid MEM_SIZE_MAP or MEM_TEST_MAP_SIZE\n");
937 		return false;
938 	}
939 
940 	if (MEM_TEST_UNMAP_SIZE > MEM_TEST_SIZE		||
941 	    MEM_TEST_UNMAP_SIZE % guest_page_size	||
942 	    (MEM_TEST_UNMAP_SIZE / guest_page_size) %
943 	    (2 * MEM_TEST_UNMAP_CHUNK_SIZE / guest_page_size)) {
944 		pr_info("invalid MEM_TEST_UNMAP_SIZE or MEM_TEST_UNMAP_CHUNK_SIZE\n");
945 		return false;
946 	}
947 
948 	return true;
949 }
950 
951 static bool parse_args(int argc, char *argv[],
952 		       struct test_args *targs)
953 {
954 	uint32_t max_mem_slots;
955 	int opt;
956 
957 	while ((opt = getopt(argc, argv, "hvds:f:e:l:r:")) != -1) {
958 		switch (opt) {
959 		case 'h':
960 		default:
961 			help(argv[0], targs);
962 			return false;
963 		case 'v':
964 			verbose = true;
965 			break;
966 		case 'd':
967 			map_unmap_verify = true;
968 			break;
969 		case 's':
970 			targs->nslots = atoi_paranoid(optarg);
971 			if (targs->nslots <= 1 && targs->nslots != -1) {
972 				pr_info("Slot count cap must be larger than 1 or -1 for no cap\n");
973 				return false;
974 			}
975 			break;
976 		case 'f':
977 			targs->tfirst = atoi_non_negative("First test", optarg);
978 			break;
979 		case 'e':
980 			targs->tlast = atoi_non_negative("Last test", optarg);
981 			if (targs->tlast >= NTESTS) {
982 				pr_info("Last test to run has to be non-negative and less than %zu\n",
983 					NTESTS);
984 				return false;
985 			}
986 			break;
987 		case 'l':
988 			targs->seconds = atoi_non_negative("Test length", optarg);
989 			break;
990 		case 'r':
991 			targs->runs = atoi_positive("Runs per test", optarg);
992 			break;
993 		}
994 	}
995 
996 	if (optind < argc) {
997 		help(argv[0], targs);
998 		return false;
999 	}
1000 
1001 	if (targs->tfirst > targs->tlast) {
1002 		pr_info("First test to run cannot be greater than the last test to run\n");
1003 		return false;
1004 	}
1005 
1006 	max_mem_slots = kvm_check_cap(KVM_CAP_NR_MEMSLOTS);
1007 	if (max_mem_slots <= 1) {
1008 		pr_info("KVM_CAP_NR_MEMSLOTS should be greater than 1\n");
1009 		return false;
1010 	}
1011 
1012 	/* Memory slot 0 is reserved */
1013 	if (targs->nslots == -1)
1014 		targs->nslots = max_mem_slots - 1;
1015 	else
1016 		targs->nslots = min_t(int, targs->nslots, max_mem_slots) - 1;
1017 
1018 	pr_info_v("Allowed Number of memory slots: %"PRIu32"\n",
1019 		  targs->nslots + 1);
1020 
1021 	return true;
1022 }
1023 
1024 struct test_result {
1025 	struct timespec slot_runtime, guest_runtime, iter_runtime;
1026 	int64_t slottimens, runtimens;
1027 	uint64_t nloops;
1028 };
1029 
1030 static bool test_loop(const struct test_data *data,
1031 		      const struct test_args *targs,
1032 		      struct test_result *rbestslottime,
1033 		      struct test_result *rbestruntime)
1034 {
1035 	uint64_t maxslots;
1036 	struct test_result result;
1037 
1038 	result.nloops = 0;
1039 	if (!test_execute(targs->nslots, &maxslots, targs->seconds, data,
1040 			  &result.nloops,
1041 			  &result.slot_runtime, &result.guest_runtime)) {
1042 		if (maxslots)
1043 			pr_info("Memslot count too high for this test, decrease the cap (max is %"PRIu64")\n",
1044 				maxslots);
1045 		else
1046 			pr_info("Memslot count may be too high for this test, try adjusting the cap\n");
1047 
1048 		return false;
1049 	}
1050 
1051 	pr_info("Test took %ld.%.9lds for slot setup + %ld.%.9lds all iterations\n",
1052 		result.slot_runtime.tv_sec, result.slot_runtime.tv_nsec,
1053 		result.guest_runtime.tv_sec, result.guest_runtime.tv_nsec);
1054 	if (!result.nloops) {
1055 		pr_info("No full loops done - too short test time or system too loaded?\n");
1056 		return true;
1057 	}
1058 
1059 	result.iter_runtime = timespec_div(result.guest_runtime,
1060 					   result.nloops);
1061 	pr_info("Done %"PRIu64" iterations, avg %ld.%.9lds each\n",
1062 		result.nloops,
1063 		result.iter_runtime.tv_sec,
1064 		result.iter_runtime.tv_nsec);
1065 	result.slottimens = timespec_to_ns(result.slot_runtime);
1066 	result.runtimens = timespec_to_ns(result.iter_runtime);
1067 
1068 	/*
1069 	 * Only rank the slot setup time for tests using the whole test memory
1070 	 * area so they are comparable
1071 	 */
1072 	if (!data->mem_size &&
1073 	    (!rbestslottime->slottimens ||
1074 	     result.slottimens < rbestslottime->slottimens))
1075 		*rbestslottime = result;
1076 	if (!rbestruntime->runtimens ||
1077 	    result.runtimens < rbestruntime->runtimens)
1078 		*rbestruntime = result;
1079 
1080 	return true;
1081 }
1082 
1083 int main(int argc, char *argv[])
1084 {
1085 	struct test_args targs = {
1086 		.tfirst = 0,
1087 		.tlast = NTESTS - 1,
1088 		.nslots = -1,
1089 		.seconds = 5,
1090 		.runs = 1,
1091 	};
1092 	struct test_result rbestslottime;
1093 	int tctr;
1094 
1095 	if (!check_memory_sizes())
1096 		return -1;
1097 
1098 	if (!parse_args(argc, argv, &targs))
1099 		return -1;
1100 
1101 	rbestslottime.slottimens = 0;
1102 	for (tctr = targs.tfirst; tctr <= targs.tlast; tctr++) {
1103 		const struct test_data *data = &tests[tctr];
1104 		unsigned int runctr;
1105 		struct test_result rbestruntime;
1106 
1107 		if (tctr > targs.tfirst)
1108 			pr_info("\n");
1109 
1110 		pr_info("Testing %s performance with %i runs, %d seconds each\n",
1111 			data->name, targs.runs, targs.seconds);
1112 
1113 		rbestruntime.runtimens = 0;
1114 		for (runctr = 0; runctr < targs.runs; runctr++)
1115 			if (!test_loop(data, &targs,
1116 				       &rbestslottime, &rbestruntime))
1117 				break;
1118 
1119 		if (rbestruntime.runtimens)
1120 			pr_info("Best runtime result was %ld.%.9lds per iteration (with %"PRIu64" iterations)\n",
1121 				rbestruntime.iter_runtime.tv_sec,
1122 				rbestruntime.iter_runtime.tv_nsec,
1123 				rbestruntime.nloops);
1124 	}
1125 
1126 	if (rbestslottime.slottimens)
1127 		pr_info("Best slot setup time for the whole test area was %ld.%.9lds\n",
1128 			rbestslottime.slot_runtime.tv_sec,
1129 			rbestslottime.slot_runtime.tv_nsec);
1130 
1131 	return 0;
1132 }
1133