diff mbox series

[2/2] KVM: selftests: add a memslot-related performance benchmark

Message ID ec67677c725d9f52d4abef3cfef07eac266c15cd.1612139762.git.maciej.szmigiero@oracle.com (mailing list archive)
State New, archived
Headers show
Series [1/2] KVM: selftests: Keep track of memslots more efficiently | expand

Commit Message

Maciej S. Szmigiero Feb. 1, 2021, 8:10 a.m. UTC
From: "Maciej S. Szmigiero" <maciej.szmigiero@oracle.com>

This benchmark contains the following tests:
* Map test, where the host unmaps guest memory while the guest writes to
it (maps it).

The test is designed in a way to make the unmap operation on the host
take a negligible amount of time in comparison with the mapping
operation in the guest.

The test area is actually split in two: the first half is being mapped
by the guest while the second half in being unmapped by the host.
Then a guest <-> host sync happens and the areas are reversed.

* Unmap test which is broadly similar to the above map test, but it is
designed in an opposite way: to make the mapping operation in the guest
take a negligible amount of time in comparison with the unmap operation
on the host.
This test is available in two variants: with per-page unmap operation
or a chunked one (using 2 MiB chunk size).

* Move active area test which involves moving the last (highest gfn)
memslot a bit back and forth on the host while the guest is
concurrently writing around the area being moved (including over the
moved memslot).

* Move inactive area test which is similar to the previous move active
area test, but now guest writes all happen outside of the area being
moved.

* Read / write test in which the guest writes to the beginning of each
page of the test area while the host writes to the middle of each such
page.
Then each side checks the values the other side has written.
This particular test is not expected to give different results depending
on particular memslots implementation, it is meant as a rough sanity
check and to provide insight on the spread of test results expected.

Each test performs its operation in a loop until a test period ends
(this is 5 seconds by default, but it is configurable).
Then the total count of loops done is divided by the actual elapsed
time to give the test result.

The tests have a configurable memslot cap with the "-s" test option, by
default the system maximum is used.
Each test is repeated a particular number of times (by default 20
times), the best result achieved is printed.

The test memory area is divided equally between memslots, the reminder
is added to the last memslot.
The test area size does not depend on the number of memslots in use.

The tests also measure the time that it took to add all these memslots.
The best result from the tests that use the whole test area is printed
after all the requested tests are done.

In general, these tests are designed to use as much memory as possible
(within reason) while still doing 100+ loops even on high memslot counts
with the default test length.
Increasing the test runtime makes it increasingly more likely that some
event will happen on the system during the test run, which might lower
the test result.

Signed-off-by: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
---
 tools/testing/selftests/kvm/.gitignore        |    1 +
 tools/testing/selftests/kvm/Makefile          |    1 +
 .../testing/selftests/kvm/memslot_perf_test.c | 1039 +++++++++++++++++
 3 files changed, 1041 insertions(+)
 create mode 100644 tools/testing/selftests/kvm/memslot_perf_test.c

Comments

Andrew Jones Feb. 8, 2021, 11:09 a.m. UTC | #1
On Mon, Feb 01, 2021 at 09:10:57AM +0100, Maciej S. Szmigiero wrote:
[...]
> diff --git a/tools/testing/selftests/kvm/.gitignore b/tools/testing/selftests/kvm/.gitignore
> index ce8f4ad39684..059a655053ca 100644
> --- a/tools/testing/selftests/kvm/.gitignore
> +++ b/tools/testing/selftests/kvm/.gitignore
> @@ -31,3 +31,4 @@
>  /kvm_create_max_vcpus
>  /set_memory_region_test
>  /steal_time
> +/memslot_perf_test
> diff --git a/tools/testing/selftests/kvm/Makefile b/tools/testing/selftests/kvm/Makefile
> index e7c6237d7383..2abc9e182c30 100644
> --- a/tools/testing/selftests/kvm/Makefile
> +++ b/tools/testing/selftests/kvm/Makefile
> @@ -65,6 +65,7 @@ TEST_GEN_PROGS_x86_64 += dirty_log_perf_test
>  TEST_GEN_PROGS_x86_64 += kvm_create_max_vcpus
>  TEST_GEN_PROGS_x86_64 += set_memory_region_test
>  TEST_GEN_PROGS_x86_64 += steal_time
> +TEST_GEN_PROGS_x86_64 += memslot_perf_test
> 
>  TEST_GEN_PROGS_aarch64 += aarch64/get-reg-list
>  TEST_GEN_PROGS_aarch64 += aarch64/get-reg-list-sve

We've been trying keep the lists in .gitignore and Makefile in alphabetic
order. It's not really important, but seems like we should keep it now
that we've got it. Well, except I see dirty_log_perf_test and
dirty_log_test are out of order already...

[...]
> +static bool prepare_vm(struct vm_data *data, int nslots, uint64_t *maxslots,
> +		       void *guest_code, uint64_t mempages,
> +		       struct timespec *slot_runtime)
> +{
> +	uint32_t max_mem_slots;
> +	uint64_t rempages;
> +	uint64_t guest_addr;
> +	uint32_t slot;
> +	struct timespec tstart;
> +	struct sync_area *sync;
> +
> +	max_mem_slots = kvm_check_cap(KVM_CAP_NR_MEMSLOTS);
> +	TEST_ASSERT(max_mem_slots > 1,
> +		    "KVM_CAP_NR_MEMSLOTS should be greater than 1");
> +	TEST_ASSERT(nslots > 1 || nslots == -1,
> +		    "Slot count cap should be greater than 1");
> +	if (nslots != -1)
> +		max_mem_slots = min(max_mem_slots, (uint32_t)nslots);
> +	pr_info_v("Allowed number of memory slots: %"PRIu32"\n", max_mem_slots);
> +
> +	TEST_ASSERT(mempages > 1,
> +		    "Can't test without any memory");
> +
> +	data->npages = mempages;
> +	data->nslots = max_mem_slots - 1;
> +	data->pages_per_slot = mempages / data->nslots;
> +	if (!data->pages_per_slot) {
> +		*maxslots = mempages + 1;
> +		return false;
> +	}
> +
> +	rempages = mempages % data->nslots;
> +	data->hva_slots = malloc(sizeof(*data->hva_slots) * data->nslots);
> +	TEST_ASSERT(data->hva_slots, "malloc() fail");
> +
> +	data->vm = vm_create_default(VCPU_ID, mempages, guest_code);
> +
> +	vcpu_set_cpuid(data->vm, VCPU_ID, kvm_get_supported_cpuid());

This vcpu_set_cpuid() call, which causes problems for non-x86 builds,
is now embedded in vm_create_default() and therefore redundant here.


Otherwise this looks good to me. I'll try to find some time to test
it on an AArch64 machine configured to use 4k pages on the host.

Reviewed-by: Andrew Jones <drjones@redhat.com>

Thanks,
drew
diff mbox series

Patch

diff --git a/tools/testing/selftests/kvm/.gitignore b/tools/testing/selftests/kvm/.gitignore
index ce8f4ad39684..059a655053ca 100644
--- a/tools/testing/selftests/kvm/.gitignore
+++ b/tools/testing/selftests/kvm/.gitignore
@@ -31,3 +31,4 @@ 
 /kvm_create_max_vcpus
 /set_memory_region_test
 /steal_time
+/memslot_perf_test
diff --git a/tools/testing/selftests/kvm/Makefile b/tools/testing/selftests/kvm/Makefile
index e7c6237d7383..2abc9e182c30 100644
--- a/tools/testing/selftests/kvm/Makefile
+++ b/tools/testing/selftests/kvm/Makefile
@@ -65,6 +65,7 @@  TEST_GEN_PROGS_x86_64 += dirty_log_perf_test
 TEST_GEN_PROGS_x86_64 += kvm_create_max_vcpus
 TEST_GEN_PROGS_x86_64 += set_memory_region_test
 TEST_GEN_PROGS_x86_64 += steal_time
+TEST_GEN_PROGS_x86_64 += memslot_perf_test
 
 TEST_GEN_PROGS_aarch64 += aarch64/get-reg-list
 TEST_GEN_PROGS_aarch64 += aarch64/get-reg-list-sve
diff --git a/tools/testing/selftests/kvm/memslot_perf_test.c b/tools/testing/selftests/kvm/memslot_perf_test.c
new file mode 100644
index 000000000000..1a632094d2eb
--- /dev/null
+++ b/tools/testing/selftests/kvm/memslot_perf_test.c
@@ -0,0 +1,1039 @@ 
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * A memslot-related performance benchmark.
+ *
+ * Copyright (C) 2021 Oracle and/or its affiliates.
+ *
+ * Basic guest setup / host vCPU thread code lifted from set_memory_region_test.
+ */
+#include <pthread.h>
+#include <sched.h>
+#include <semaphore.h>
+#include <stdatomic.h>
+#include <stdbool.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/mman.h>
+#include <time.h>
+#include <unistd.h>
+
+#include <linux/compiler.h>
+
+#include <test_util.h>
+#include <kvm_util.h>
+#include <processor.h>
+
+#define VCPU_ID 0
+
+#define MEM_SIZE		((512U << 20) + 4096)
+#define MEM_SIZE_PAGES		(MEM_SIZE / 4096)
+#define MEM_GPA		0x10000000UL
+#define MEM_AUX_GPA		MEM_GPA
+#define MEM_SYNC_GPA		MEM_AUX_GPA
+#define MEM_TEST_GPA		(MEM_AUX_GPA + 4096)
+#define MEM_TEST_SIZE		(MEM_SIZE - 4096)
+static_assert(MEM_SIZE % 4096 == 0, "invalid mem size");
+static_assert(MEM_TEST_SIZE % 4096 == 0, "invalid mem test size");
+
+/*
+ * 32 MiB is max size that gets well over 100 iterations on 509 slots.
+ * Considering that each slot needs to have at least one page up to
+ * 8194 slots in use can then be tested (although with slightly
+ * limited resolution).
+ */
+#define MEM_SIZE_MAP		((32U << 20) + 4096)
+#define MEM_SIZE_MAP_PAGES	(MEM_SIZE_MAP / 4096)
+#define MEM_TEST_MAP_SIZE	(MEM_SIZE_MAP - 4096)
+#define MEM_TEST_MAP_SIZE_PAGES (MEM_TEST_MAP_SIZE / 4096)
+static_assert(MEM_SIZE_MAP % 4096 == 0, "invalid map test region size");
+static_assert(MEM_TEST_MAP_SIZE % 4096 == 0, "invalid map test region size");
+static_assert(MEM_TEST_MAP_SIZE_PAGES % 2 == 0, "invalid map test region size");
+static_assert(MEM_TEST_MAP_SIZE_PAGES > 2, "invalid map test region size");
+
+/*
+ * 128 MiB is min size that fills 32k slots with at least one page in each
+ * while at the same time gets 100+ iterations in such test
+ */
+#define MEM_TEST_UNMAP_SIZE		(128U << 20)
+#define MEM_TEST_UNMAP_SIZE_PAGES	(MEM_TEST_UNMAP_SIZE / 4096)
+/* 2 MiB chunk size like a typical huge page */
+#define MEM_TEST_UNMAP_CHUNK_PAGES	(2U << (20 - 12))
+static_assert(MEM_TEST_UNMAP_SIZE <= MEM_TEST_SIZE,
+	      "invalid unmap test region size");
+static_assert(MEM_TEST_UNMAP_SIZE % 4096 == 0,
+	      "invalid unmap test region size");
+static_assert(MEM_TEST_UNMAP_SIZE_PAGES %
+	      (2 * MEM_TEST_UNMAP_CHUNK_PAGES) == 0,
+	      "invalid unmap test region size");
+
+/*
+ * For the move active test the middle of the test area is placed on
+ * a memslot boundary: half lies in the memslot being moved, half in
+ * other memslot(s).
+ *
+ * When running this test with 32k memslots (32764, really) each memslot
+ * contains 4 pages.
+ * The last one additionally contains the remaining 21 pages of memory,
+ * for the total size of 25 pages.
+ * Hence, the maximum size here is 50 pages.
+ */
+#define MEM_TEST_MOVE_SIZE_PAGES	(50)
+#define MEM_TEST_MOVE_SIZE		(MEM_TEST_MOVE_SIZE_PAGES * 4096)
+#define MEM_TEST_MOVE_GPA_DEST		(MEM_GPA + MEM_SIZE)
+static_assert(MEM_TEST_MOVE_SIZE <= MEM_TEST_SIZE,
+	      "invalid move test region size");
+
+#define MEM_TEST_VAL_1 0x1122334455667788
+#define MEM_TEST_VAL_2 0x99AABBCCDDEEFF00
+
+struct vm_data {
+	struct kvm_vm *vm;
+	pthread_t vcpu_thread;
+	uint32_t nslots;
+	uint64_t npages;
+	uint64_t pages_per_slot;
+	void **hva_slots;
+	bool mmio_ok;
+	uint64_t mmio_gpa_min;
+	uint64_t mmio_gpa_max;
+};
+
+struct sync_area {
+	atomic_bool start_flag;
+	atomic_bool exit_flag;
+	atomic_bool sync_flag;
+	void *move_area_ptr;
+};
+
+/*
+ * Technically, we need also for the atomic bool to be address-free, which
+ * is recommended, but not strictly required, by C11 for lockless
+ * implementations.
+ * However, in practice both GCC and Clang fulfill this requirement on
+ * all KVM-supported platforms.
+ */
+static_assert(ATOMIC_BOOL_LOCK_FREE == 2, "atomic bool is not lockless");
+
+static sem_t vcpu_ready;
+
+static bool map_unmap_verify;
+
+static bool verbose;
+#define pr_info_v(...)				\
+	do {					\
+		if (verbose)			\
+			pr_info(__VA_ARGS__);	\
+	} while (0)
+
+static void *vcpu_worker(void *data)
+{
+	struct vm_data *vm = data;
+	struct kvm_run *run;
+	struct ucall uc;
+	uint64_t cmd;
+
+	run = vcpu_state(vm->vm, VCPU_ID);
+	while (1) {
+		vcpu_run(vm->vm, VCPU_ID);
+
+		if (run->exit_reason == KVM_EXIT_IO) {
+			cmd = get_ucall(vm->vm, VCPU_ID, &uc);
+			if (cmd != UCALL_SYNC)
+				break;
+
+			sem_post(&vcpu_ready);
+			continue;
+		}
+
+		if (run->exit_reason != KVM_EXIT_MMIO)
+			break;
+
+		TEST_ASSERT(vm->mmio_ok, "Unexpected mmio exit");
+		TEST_ASSERT(run->mmio.is_write, "Unexpected mmio read");
+		TEST_ASSERT(run->mmio.len == 8,
+			    "Unexpected exit mmio size = %u", run->mmio.len);
+		TEST_ASSERT(run->mmio.phys_addr >= vm->mmio_gpa_min &&
+			    run->mmio.phys_addr <= vm->mmio_gpa_max,
+			    "Unexpected exit mmio address = 0x%llx",
+			    run->mmio.phys_addr);
+	}
+
+	if (run->exit_reason == KVM_EXIT_IO && cmd == UCALL_ABORT)
+		TEST_FAIL("%s at %s:%ld, val = %lu", (const char *)uc.args[0],
+			  __FILE__, uc.args[1], uc.args[2]);
+
+	return NULL;
+}
+
+static void wait_for_vcpu(void)
+{
+	struct timespec ts;
+
+	TEST_ASSERT(!clock_gettime(CLOCK_REALTIME, &ts),
+		    "clock_gettime() failed: %d\n", errno);
+
+	ts.tv_sec += 2;
+	TEST_ASSERT(!sem_timedwait(&vcpu_ready, &ts),
+		    "sem_timedwait() failed: %d\n", errno);
+}
+
+static void *vm_gpa2hva(struct vm_data *data, uint64_t gpa, uint64_t *rempages)
+{
+	uint64_t gpage, pgoffs;
+	uint32_t slot, slotoffs;
+	void *base;
+
+	TEST_ASSERT(gpa >= MEM_GPA, "Too low gpa to translate");
+	TEST_ASSERT(gpa < MEM_GPA + data->npages * 4096,
+		    "Too high gpa to translate");
+	gpa -= MEM_GPA;
+
+	gpage = gpa / 4096;
+	pgoffs = gpa % 4096;
+	slot = min(gpage / data->pages_per_slot, (uint64_t)data->nslots - 1);
+	slotoffs = gpage - (slot * data->pages_per_slot);
+
+	if (rempages) {
+		uint64_t slotpages;
+
+		if (slot == data->nslots - 1)
+			slotpages = data->npages - slot * data->pages_per_slot;
+		else
+			slotpages = data->pages_per_slot;
+
+		TEST_ASSERT(!pgoffs,
+			    "Asking for remaining pages in slot but gpa not page aligned");
+		*rempages = slotpages - slotoffs;
+	}
+
+	base = data->hva_slots[slot];
+	return (uint8_t *)base + slotoffs * 4096 + pgoffs;
+}
+
+static uint64_t vm_slot2gpa(struct vm_data *data, uint32_t slot)
+{
+	TEST_ASSERT(slot < data->nslots, "Too high slot number");
+
+	return MEM_GPA + slot * data->pages_per_slot * 4096;
+}
+
+static struct vm_data *alloc_vm(void)
+{
+	struct vm_data *data;
+
+	data = malloc(sizeof(*data));
+	TEST_ASSERT(data, "malloc(vmdata) failed");
+
+	data->vm = NULL;
+	data->hva_slots = NULL;
+
+	return data;
+}
+
+static bool prepare_vm(struct vm_data *data, int nslots, uint64_t *maxslots,
+		       void *guest_code, uint64_t mempages,
+		       struct timespec *slot_runtime)
+{
+	uint32_t max_mem_slots;
+	uint64_t rempages;
+	uint64_t guest_addr;
+	uint32_t slot;
+	struct timespec tstart;
+	struct sync_area *sync;
+
+	max_mem_slots = kvm_check_cap(KVM_CAP_NR_MEMSLOTS);
+	TEST_ASSERT(max_mem_slots > 1,
+		    "KVM_CAP_NR_MEMSLOTS should be greater than 1");
+	TEST_ASSERT(nslots > 1 || nslots == -1,
+		    "Slot count cap should be greater than 1");
+	if (nslots != -1)
+		max_mem_slots = min(max_mem_slots, (uint32_t)nslots);
+	pr_info_v("Allowed number of memory slots: %"PRIu32"\n", max_mem_slots);
+
+	TEST_ASSERT(mempages > 1,
+		    "Can't test without any memory");
+
+	data->npages = mempages;
+	data->nslots = max_mem_slots - 1;
+	data->pages_per_slot = mempages / data->nslots;
+	if (!data->pages_per_slot) {
+		*maxslots = mempages + 1;
+		return false;
+	}
+
+	rempages = mempages % data->nslots;
+	data->hva_slots = malloc(sizeof(*data->hva_slots) * data->nslots);
+	TEST_ASSERT(data->hva_slots, "malloc() fail");
+
+	data->vm = vm_create_default(VCPU_ID, mempages, guest_code);
+
+	vcpu_set_cpuid(data->vm, VCPU_ID, kvm_get_supported_cpuid());
+
+	pr_info_v("Adding slots 1..%i, each slot with %"PRIu64" pages + %"PRIu64" extra pages last\n",
+		max_mem_slots - 1, data->pages_per_slot, rempages);
+
+	clock_gettime(CLOCK_MONOTONIC, &tstart);
+	for (slot = 1, guest_addr = MEM_GPA; slot < max_mem_slots; slot++) {
+		uint64_t npages;
+
+		npages = data->pages_per_slot;
+		if (slot == max_mem_slots - 1)
+			npages += rempages;
+
+		vm_userspace_mem_region_add(data->vm, VM_MEM_SRC_ANONYMOUS,
+					    guest_addr, slot, npages,
+					    0);
+		guest_addr += npages * 4096;
+	}
+	*slot_runtime = timespec_diff_now(tstart);
+
+	for (slot = 0, guest_addr = MEM_GPA; slot < max_mem_slots - 1; slot++) {
+		uint64_t npages;
+		uint64_t gpa;
+
+		npages = data->pages_per_slot;
+		if (slot == max_mem_slots - 2)
+			npages += rempages;
+
+		gpa = vm_phy_pages_alloc(data->vm, npages, guest_addr,
+					 slot + 1);
+		TEST_ASSERT(gpa == guest_addr,
+			    "vm_phy_pages_alloc() failed\n");
+
+		data->hva_slots[slot] = addr_gpa2hva(data->vm, guest_addr);
+		memset(data->hva_slots[slot], 0, npages * 4096);
+
+		guest_addr += npages * 4096;
+	}
+
+	virt_map(data->vm, MEM_GPA, MEM_GPA, mempages, 0);
+
+	sync = (typeof(sync))vm_gpa2hva(data, MEM_SYNC_GPA, NULL);
+	atomic_init(&sync->start_flag, false);
+	atomic_init(&sync->exit_flag, false);
+	atomic_init(&sync->sync_flag, false);
+
+	data->mmio_ok = false;
+
+	return true;
+}
+
+static void launch_vm(struct vm_data *data)
+{
+	pr_info_v("Launching the test VM\n");
+
+	pthread_create(&data->vcpu_thread, NULL, vcpu_worker, data);
+
+	/* Ensure the guest thread is spun up. */
+	wait_for_vcpu();
+}
+
+static void free_vm(struct vm_data *data)
+{
+	kvm_vm_free(data->vm);
+	free(data->hva_slots);
+	free(data);
+}
+
+static void wait_guest_exit(struct vm_data *data)
+{
+	pthread_join(data->vcpu_thread, NULL);
+}
+
+static void let_guest_run(struct sync_area *sync)
+{
+	atomic_store_explicit(&sync->start_flag, true, memory_order_release);
+}
+
+static void guest_spin_until_start(void)
+{
+	struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
+
+	while (!atomic_load_explicit(&sync->start_flag, memory_order_acquire))
+		;
+}
+
+static void make_guest_exit(struct sync_area *sync)
+{
+	atomic_store_explicit(&sync->exit_flag, true, memory_order_release);
+}
+
+static bool _guest_should_exit(void)
+{
+	struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
+
+	return atomic_load_explicit(&sync->exit_flag, memory_order_acquire);
+}
+
+#define guest_should_exit() unlikely(_guest_should_exit())
+
+/*
+ * noinline so we can easily see how much time the host spends waiting
+ * for the guest.
+ * For the same reason use alarm() instead of polling clock_gettime()
+ * to implement a wait timeout.
+ */
+static noinline void host_perform_sync(struct sync_area *sync)
+{
+	alarm(2);
+
+	atomic_store_explicit(&sync->sync_flag, true, memory_order_release);
+	while (atomic_load_explicit(&sync->sync_flag, memory_order_acquire))
+		;
+
+	alarm(0);
+}
+
+static bool guest_perform_sync(void)
+{
+	struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
+	bool expected;
+
+	do {
+		if (guest_should_exit())
+			return false;
+
+		expected = true;
+	} while (!atomic_compare_exchange_weak_explicit(&sync->sync_flag,
+							&expected, false,
+							memory_order_acq_rel,
+							memory_order_relaxed));
+
+	return true;
+}
+
+static void guest_code_test_memslot_move(void)
+{
+	struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
+	uintptr_t base = (typeof(base))READ_ONCE(sync->move_area_ptr);
+
+	GUEST_SYNC(0);
+
+	guest_spin_until_start();
+
+	while (!guest_should_exit()) {
+		uintptr_t ptr;
+
+		for (ptr = base; ptr < base + MEM_TEST_MOVE_SIZE;
+		     ptr += 4096)
+			*(uint64_t *)ptr = MEM_TEST_VAL_1;
+
+		/*
+		 * No host sync here since the MMIO exits are so expensive
+		 * that the host would spend most of its time waiting for
+		 * the guest and so instead of measuring memslot move
+		 * performance we would measure the performance and
+		 * likelihood of MMIO exits
+		 */
+	}
+
+	GUEST_DONE();
+}
+
+static void guest_code_test_memslot_map(void)
+{
+	struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
+
+	GUEST_SYNC(0);
+
+	guest_spin_until_start();
+
+	while (1) {
+		uintptr_t ptr;
+
+		for (ptr = MEM_TEST_GPA;
+		     ptr < MEM_TEST_GPA + MEM_TEST_MAP_SIZE / 2; ptr += 4096)
+			*(uint64_t *)ptr = MEM_TEST_VAL_1;
+
+		if (!guest_perform_sync())
+			break;
+
+		for (ptr = MEM_TEST_GPA + MEM_TEST_MAP_SIZE / 2;
+		     ptr < MEM_TEST_GPA + MEM_TEST_MAP_SIZE; ptr += 4096)
+			*(uint64_t *)ptr = MEM_TEST_VAL_2;
+
+		if (!guest_perform_sync())
+			break;
+	}
+
+	GUEST_DONE();
+}
+
+static void guest_code_test_memslot_unmap(void)
+{
+	struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
+
+	GUEST_SYNC(0);
+
+	guest_spin_until_start();
+
+	while (1) {
+		uintptr_t ptr = MEM_TEST_GPA;
+
+		/*
+		 * We can afford to access (map) just a small number of pages
+		 * per host sync as otherwise the host will spend
+		 * a significant amount of its time waiting for the guest
+		 * (instead of doing unmap operations), so this will
+		 * effectively turn this test into a map performance test.
+		 *
+		 * Just access a single page to be on the safe side.
+		 */
+		*(uint64_t *)ptr = MEM_TEST_VAL_1;
+
+		if (!guest_perform_sync())
+			break;
+
+		ptr += MEM_TEST_UNMAP_SIZE / 2;
+		*(uint64_t *)ptr = MEM_TEST_VAL_2;
+
+		if (!guest_perform_sync())
+			break;
+	}
+
+	GUEST_DONE();
+}
+
+static void guest_code_test_memslot_rw(void)
+{
+	GUEST_SYNC(0);
+
+	guest_spin_until_start();
+
+	while (1) {
+		uintptr_t ptr;
+
+		for (ptr = MEM_TEST_GPA;
+		     ptr < MEM_TEST_GPA + MEM_TEST_SIZE; ptr += 4096)
+			*(uint64_t *)ptr = MEM_TEST_VAL_1;
+
+		if (!guest_perform_sync())
+			break;
+
+		for (ptr = MEM_TEST_GPA + 4096 / 2;
+		     ptr < MEM_TEST_GPA + MEM_TEST_SIZE; ptr += 4096) {
+			uint64_t val = *(uint64_t *)ptr;
+
+			GUEST_ASSERT_1(val == MEM_TEST_VAL_2, val);
+			*(uint64_t *)ptr = 0;
+		}
+
+		if (!guest_perform_sync())
+			break;
+	}
+
+	GUEST_DONE();
+}
+
+static bool test_memslot_move_prepare(struct vm_data *data,
+				      struct sync_area *sync,
+				      uint64_t *maxslots, bool isactive)
+{
+	uint64_t movesrcgpa, movetestgpa;
+
+	movesrcgpa = vm_slot2gpa(data, data->nslots - 1);
+
+	if (isactive) {
+		uint64_t lastpages;
+
+		vm_gpa2hva(data, movesrcgpa, &lastpages);
+		if (lastpages < MEM_TEST_MOVE_SIZE_PAGES / 2) {
+			*maxslots = 0;
+			return false;
+		}
+	}
+
+	movetestgpa = movesrcgpa - (MEM_TEST_MOVE_SIZE / (isactive ? 2 : 1));
+	sync->move_area_ptr = (void *)movetestgpa;
+
+	if (isactive) {
+		data->mmio_ok = true;
+		data->mmio_gpa_min = movesrcgpa;
+		data->mmio_gpa_max = movesrcgpa + MEM_TEST_MOVE_SIZE / 2 - 1;
+	}
+
+	return true;
+}
+
+static bool test_memslot_move_prepare_active(struct vm_data *data,
+					     struct sync_area *sync,
+					     uint64_t *maxslots)
+{
+	return test_memslot_move_prepare(data, sync, maxslots, true);
+}
+
+static bool test_memslot_move_prepare_inactive(struct vm_data *data,
+					       struct sync_area *sync,
+					       uint64_t *maxslots)
+{
+	return test_memslot_move_prepare(data, sync, maxslots, false);
+}
+
+static void test_memslot_move_loop(struct vm_data *data, struct sync_area *sync)
+{
+	uint64_t movesrcgpa;
+
+	movesrcgpa = vm_slot2gpa(data, data->nslots - 1);
+	vm_mem_region_move(data->vm, data->nslots - 1 + 1,
+			   MEM_TEST_MOVE_GPA_DEST);
+	vm_mem_region_move(data->vm, data->nslots - 1 + 1, movesrcgpa);
+}
+
+static void test_memslot_do_unmap(struct vm_data *data,
+				  uint64_t offsp, uint64_t count)
+{
+	uint64_t gpa, ctr;
+
+	for (gpa = MEM_TEST_GPA + offsp * 4096, ctr = 0; ctr < count; ) {
+		uint64_t npages;
+		void *hva;
+		int ret;
+
+		hva = vm_gpa2hva(data, gpa, &npages);
+		TEST_ASSERT(npages, "Empty memory slot at gptr 0x%"PRIx64, gpa);
+		npages = min(npages, count - ctr);
+		ret = madvise(hva, npages * 4096, MADV_DONTNEED);
+		TEST_ASSERT(!ret,
+			    "madvise(%p, MADV_DONTNEED) on VM memory should not fail for gptr 0x%"PRIx64,
+			    hva, gpa);
+		ctr += npages;
+		gpa += npages * 4096;
+	}
+	TEST_ASSERT(ctr == count,
+		    "madvise(MADV_DONTNEED) should exactly cover all of the requested area");
+}
+
+static void test_memslot_map_unmap_check(struct vm_data *data,
+					 uint64_t offsp, uint64_t valexp)
+{
+	uint64_t gpa;
+	uint64_t *val;
+
+	if (!map_unmap_verify)
+		return;
+
+	gpa = MEM_TEST_GPA + offsp * 4096;
+	val = (typeof(val))vm_gpa2hva(data, gpa, NULL);
+	TEST_ASSERT(*val == valexp,
+		    "Guest written values should read back correctly before unmap (%"PRIu64" vs %"PRIu64" @ %"PRIx64")",
+		    *val, valexp, gpa);
+	*val = 0;
+}
+
+static void test_memslot_map_loop(struct vm_data *data, struct sync_area *sync)
+{
+	/*
+	 * Unmap the second half of the test area while guest writes to (maps)
+	 * the first half.
+	 */
+	test_memslot_do_unmap(data, MEM_TEST_MAP_SIZE_PAGES / 2,
+			      MEM_TEST_MAP_SIZE_PAGES / 2);
+
+	/*
+	 * Wait for the guest to finish writing the first half of the test
+	 * area, verify the written value on the first and the last page of
+	 * this area and then unmap it.
+	 * Meanwhile, the guest is writing to (mapping) the second half of
+	 * the test area.
+	 */
+	host_perform_sync(sync);
+	test_memslot_map_unmap_check(data, 0, MEM_TEST_VAL_1);
+	test_memslot_map_unmap_check(data,
+				     MEM_TEST_MAP_SIZE_PAGES / 2 - 1,
+				     MEM_TEST_VAL_1);
+	test_memslot_do_unmap(data, 0, MEM_TEST_MAP_SIZE_PAGES / 2);
+
+
+	/*
+	 * Wait for the guest to finish writing the second half of the test
+	 * area and verify the written value on the first and the last page
+	 * of this area.
+	 * The area will be unmapped at the beginning of the next loop
+	 * iteration.
+	 * Meanwhile, the guest is writing to (mapping) the first half of
+	 * the test area.
+	 */
+	host_perform_sync(sync);
+	test_memslot_map_unmap_check(data, MEM_TEST_MAP_SIZE_PAGES / 2,
+				     MEM_TEST_VAL_2);
+	test_memslot_map_unmap_check(data, MEM_TEST_MAP_SIZE_PAGES - 1,
+				     MEM_TEST_VAL_2);
+}
+
+static void test_memslot_unmap_loop_common(struct vm_data *data,
+					   struct sync_area *sync,
+					   uint64_t chunk)
+{
+	uint64_t ctr;
+
+	/*
+	 * Wait for the guest to finish mapping page(s) in the first half
+	 * of the test area, verify the written value and then perform unmap
+	 * of this area.
+	 * Meanwhile, the guest is writing to (mapping) page(s) in the second
+	 * half of the test area.
+	 */
+	host_perform_sync(sync);
+	test_memslot_map_unmap_check(data, 0, MEM_TEST_VAL_1);
+	for (ctr = 0; ctr < MEM_TEST_UNMAP_SIZE_PAGES / 2; ctr += chunk)
+		test_memslot_do_unmap(data, ctr, chunk);
+
+	/* Likewise, but for the opposite host / guest areas */
+	host_perform_sync(sync);
+	test_memslot_map_unmap_check(data, MEM_TEST_UNMAP_SIZE_PAGES / 2,
+				     MEM_TEST_VAL_2);
+	for (ctr = MEM_TEST_UNMAP_SIZE_PAGES / 2;
+	     ctr < MEM_TEST_UNMAP_SIZE_PAGES; ctr += chunk)
+		test_memslot_do_unmap(data, ctr, chunk);
+}
+
+static void test_memslot_unmap_loop(struct vm_data *data,
+				    struct sync_area *sync)
+{
+	test_memslot_unmap_loop_common(data, sync, 1);
+}
+
+static void test_memslot_unmap_loop_chunked(struct vm_data *data,
+					    struct sync_area *sync)
+{
+	test_memslot_unmap_loop_common(data, sync, MEM_TEST_UNMAP_CHUNK_PAGES);
+}
+
+static void test_memslot_rw_loop(struct vm_data *data, struct sync_area *sync)
+{
+	uint64_t gptr;
+
+	for (gptr = MEM_TEST_GPA + 4096 / 2;
+	     gptr < MEM_TEST_GPA + MEM_TEST_SIZE; gptr += 4096)
+		*(uint64_t *)vm_gpa2hva(data, gptr, NULL) = MEM_TEST_VAL_2;
+
+	host_perform_sync(sync);
+
+	for (gptr = MEM_TEST_GPA;
+	     gptr < MEM_TEST_GPA + MEM_TEST_SIZE; gptr += 4096) {
+		uint64_t *vptr = (typeof(vptr))vm_gpa2hva(data, gptr, NULL);
+		uint64_t val = *vptr;
+
+		TEST_ASSERT(val == MEM_TEST_VAL_1,
+			    "Guest written values should read back correctly (is %"PRIu64" @ %"PRIx64")",
+			    val, gptr);
+		*vptr = 0;
+	}
+
+	host_perform_sync(sync);
+}
+
+struct test_data {
+	const char *name;
+	uint64_t mem_size;
+	void (*guest_code)(void);
+	bool (*prepare)(struct vm_data *data, struct sync_area *sync,
+			uint64_t *maxslots);
+	void (*loop)(struct vm_data *data, struct sync_area *sync);
+};
+
+static bool test_execute(int nslots, uint64_t *maxslots,
+			 unsigned int maxtime,
+			 const struct test_data *tdata,
+			 uint64_t *nloops,
+			 struct timespec *slot_runtime,
+			 struct timespec *guest_runtime)
+{
+	uint64_t mem_size = tdata->mem_size ? : MEM_SIZE_PAGES;
+	struct vm_data *data;
+	struct sync_area *sync;
+	struct timespec tstart;
+	bool ret = true;
+
+	data = alloc_vm();
+	if (!prepare_vm(data, nslots, maxslots, tdata->guest_code,
+			mem_size, slot_runtime)) {
+		ret = false;
+		goto exit_free;
+	}
+
+	sync = (typeof(sync))vm_gpa2hva(data, MEM_SYNC_GPA, NULL);
+
+	if (tdata->prepare &&
+	    !tdata->prepare(data, sync, maxslots)) {
+		ret = false;
+		goto exit_free;
+	}
+
+	launch_vm(data);
+
+	clock_gettime(CLOCK_MONOTONIC, &tstart);
+	let_guest_run(sync);
+
+	while (1) {
+		*guest_runtime = timespec_diff_now(tstart);
+		if (guest_runtime->tv_sec >= maxtime)
+			break;
+
+		tdata->loop(data, sync);
+
+		(*nloops)++;
+	}
+
+	make_guest_exit(sync);
+	wait_guest_exit(data);
+
+exit_free:
+	free_vm(data);
+
+	return ret;
+}
+
+static const struct test_data tests[] = {
+	{
+		.name = "map",
+		.mem_size = MEM_SIZE_MAP_PAGES,
+		.guest_code = guest_code_test_memslot_map,
+		.loop = test_memslot_map_loop,
+	},
+	{
+		.name = "unmap",
+		.mem_size = MEM_TEST_UNMAP_SIZE_PAGES + 1,
+		.guest_code = guest_code_test_memslot_unmap,
+		.loop = test_memslot_unmap_loop,
+	},
+	{
+		.name = "unmap chunked",
+		.mem_size = MEM_TEST_UNMAP_SIZE_PAGES + 1,
+		.guest_code = guest_code_test_memslot_unmap,
+		.loop = test_memslot_unmap_loop_chunked,
+	},
+	{
+		.name = "move active area",
+		.guest_code = guest_code_test_memslot_move,
+		.prepare = test_memslot_move_prepare_active,
+		.loop = test_memslot_move_loop,
+	},
+	{
+		.name = "move inactive area",
+		.guest_code = guest_code_test_memslot_move,
+		.prepare = test_memslot_move_prepare_inactive,
+		.loop = test_memslot_move_loop,
+	},
+	{
+		.name = "RW",
+		.guest_code = guest_code_test_memslot_rw,
+		.loop = test_memslot_rw_loop
+	},
+};
+
+#define NTESTS ARRAY_SIZE(tests)
+
+struct test_args {
+	int tfirst;
+	int tlast;
+	int nslots;
+	int seconds;
+	int runs;
+};
+
+static void help(char *name, struct test_args *targs)
+{
+	int ctr;
+
+	pr_info("usage: %s [-h] [-v] [-d] [-s slots] [-f first_test] [-e last_test] [-l test_length] [-r run_count]\n",
+		name);
+	pr_info(" -h: print this help screen.\n");
+	pr_info(" -v: enable verbose mode (not for benchmarking).\n");
+	pr_info(" -d: enable extra debug checks.\n");
+	pr_info(" -s: specify memslot count cap (-1 means no cap; currently: %i)\n",
+		targs->nslots);
+	pr_info(" -f: specify the first test to run (currently: %i; max %zu)\n",
+		targs->tfirst, NTESTS - 1);
+	pr_info(" -e: specify the last test to run (currently: %i; max %zu)\n",
+		targs->tlast, NTESTS - 1);
+	pr_info(" -l: specify the test length in seconds (currently: %i)\n",
+		targs->seconds);
+	pr_info(" -r: specify the number of runs per test (currently: %i)\n",
+		targs->runs);
+
+	pr_info("\nAvailable tests:\n");
+	for (ctr = 0; ctr < NTESTS; ctr++)
+		pr_info("%d: %s\n", ctr, tests[ctr].name);
+}
+
+static bool parse_args(int argc, char *argv[],
+		       struct test_args *targs)
+{
+	int opt;
+
+	while ((opt = getopt(argc, argv, "hvds:f:e:l:r:")) != -1) {
+		switch (opt) {
+		case 'h':
+		default:
+			help(argv[0], targs);
+			return false;
+		case 'v':
+			verbose = true;
+			break;
+		case 'd':
+			map_unmap_verify = true;
+			break;
+		case 's':
+			targs->nslots = atoi(optarg);
+			if (targs->nslots <= 0 && targs->nslots != -1) {
+				pr_info("Slot count cap has to be positive or -1 for no cap\n");
+				return false;
+			}
+			break;
+		case 'f':
+			targs->tfirst = atoi(optarg);
+			if (targs->tfirst < 0) {
+				pr_info("First test to run has to be non-negative\n");
+				return false;
+			}
+			break;
+		case 'e':
+			targs->tlast = atoi(optarg);
+			if (targs->tlast < 0 || targs->tlast >= NTESTS) {
+				pr_info("Last test to run has to be non-negative and less than %zu\n",
+					NTESTS);
+				return false;
+			}
+			break;
+		case 'l':
+			targs->seconds = atoi(optarg);
+			if (targs->seconds < 0) {
+				pr_info("Test length in seconds has to be non-negative\n");
+				return false;
+			}
+			break;
+		case 'r':
+			targs->runs = atoi(optarg);
+			if (targs->runs <= 0) {
+				pr_info("Runs per test has to be positive\n");
+				return false;
+			}
+			break;
+		}
+	}
+
+	if (optind < argc) {
+		help(argv[0], targs);
+		return false;
+	}
+
+	if (targs->tfirst > targs->tlast) {
+		pr_info("First test to run cannot be greater than the last test to run\n");
+		return false;
+	}
+
+	return true;
+}
+
+struct test_result {
+	struct timespec slot_runtime, guest_runtime, iter_runtime;
+	int64_t slottimens, runtimens;
+	uint64_t nloops;
+};
+
+static bool test_loop(const struct test_data *data,
+		      const struct test_args *targs,
+		      struct test_result *rbestslottime,
+		      struct test_result *rbestruntime)
+{
+	uint64_t maxslots;
+	struct test_result result;
+
+	result.nloops = 0;
+	if (!test_execute(targs->nslots, &maxslots, targs->seconds, data,
+			  &result.nloops,
+			  &result.slot_runtime, &result.guest_runtime)) {
+		if (maxslots)
+			pr_info("Memslot count too high for this test, decrease the cap (max is %"PRIu64")\n",
+				maxslots);
+		else
+			pr_info("Memslot count may be too high for this test, try adjusting the cap\n");
+
+		return false;
+	}
+
+	pr_info("Test took %ld.%.9lds for slot setup + %ld.%.9lds all iterations\n",
+		result.slot_runtime.tv_sec, result.slot_runtime.tv_nsec,
+		result.guest_runtime.tv_sec, result.guest_runtime.tv_nsec);
+	if (!result.nloops) {
+		pr_info("No full loops done - too short test time or system too loaded?\n");
+		return true;
+	}
+
+	result.iter_runtime = timespec_div(result.guest_runtime,
+					   result.nloops);
+	pr_info("Done %"PRIu64" iterations, avg %ld.%.9lds each\n",
+		result.nloops,
+		result.iter_runtime.tv_sec,
+		result.iter_runtime.tv_nsec);
+	result.slottimens = timespec_to_ns(result.slot_runtime);
+	result.runtimens = timespec_to_ns(result.iter_runtime);
+
+	/*
+	 * Only rank the slot setup time for tests using the whole test memory
+	 * area so they are comparable
+	 */
+	if (!data->mem_size &&
+	    (!rbestslottime->slottimens ||
+	     result.slottimens < rbestslottime->slottimens))
+		*rbestslottime = result;
+	if (!rbestruntime->runtimens ||
+	    result.runtimens < rbestruntime->runtimens)
+		*rbestruntime = result;
+
+	return true;
+}
+
+int main(int argc, char *argv[])
+{
+	struct test_args targs = {
+		.tfirst = 0,
+		.tlast = NTESTS - 1,
+		.nslots = -1,
+		.seconds = 5,
+		.runs = 20,
+	};
+	struct test_result rbestslottime;
+	int tctr;
+
+	/* Tell stdout not to buffer its content */
+	setbuf(stdout, NULL);
+
+	if (!parse_args(argc, argv, &targs))
+		return -1;
+
+	rbestslottime.slottimens = 0;
+	for (tctr = targs.tfirst; tctr <= targs.tlast; tctr++) {
+		const struct test_data *data = &tests[tctr];
+		unsigned int runctr;
+		struct test_result rbestruntime;
+
+		if (tctr > targs.tfirst)
+			pr_info("\n");
+
+		pr_info("Testing %s performance with %i runs, %d seconds each\n",
+			data->name, targs.runs, targs.seconds);
+
+		rbestruntime.runtimens = 0;
+		for (runctr = 0; runctr < targs.runs; runctr++)
+			if (!test_loop(data, &targs,
+				       &rbestslottime, &rbestruntime))
+				break;
+
+		if (rbestruntime.runtimens)
+			pr_info("Best runtime result was %ld.%.9lds per iteration (with %"PRIu64" iterations)\n",
+				rbestruntime.iter_runtime.tv_sec,
+				rbestruntime.iter_runtime.tv_nsec,
+				rbestruntime.nloops);
+	}
+
+	if (rbestslottime.slottimens)
+		pr_info("Best slot setup time for the whole test area was %ld.%.9lds\n",
+			rbestslottime.slot_runtime.tv_sec,
+			rbestslottime.slot_runtime.tv_nsec);
+
+	return 0;
+}