1 // SPDX-License-Identifier: GPL-2.0
3 * access_tracking_perf_test
5 * Copyright (C) 2021, Google, Inc.
7 * This test measures the performance effects of KVM's access tracking.
8 * Access tracking is driven by the MMU notifiers test_young, clear_young, and
9 * clear_flush_young. These notifiers do not have a direct userspace API,
10 * however the clear_young notifier can be triggered by marking a pages as idle
11 * in /sys/kernel/mm/page_idle/bitmap. This test leverages that mechanism to
12 * enable access tracking on guest memory.
14 * To measure performance this test runs a VM with a configurable number of
15 * vCPUs that each touch every page in disjoint regions of memory. Performance
16 * is measured in the time it takes all vCPUs to finish touching their
19 * Note that a deterministic correctness test of access tracking is not possible
20 * by using page_idle as it exists today. This is for a few reasons:
22 * 1. page_idle only issues clear_young notifiers, which lack a TLB flush. This
23 * means subsequent guest accesses are not guaranteed to see page table
24 * updates made by KVM until some time in the future.
26 * 2. page_idle only operates on LRU pages. Newly allocated pages are not
27 * immediately allocated to LRU lists. Instead they are held in a "pagevec",
28 * which is drained to LRU lists some time in the future. There is no
29 * userspace API to force this drain to occur.
31 * These limitations are worked around in this test by using a large enough
32 * region of memory for each vCPU such that the number of translations cached in
33 * the TLB and the number of pages held in pagevecs are a small fraction of the
34 * overall workload. And if either of those conditions are not true this test
35 * will fail rather than silently passing.
41 #include <sys/types.h>
45 #include "test_util.h"
46 #include "perf_test_util.h"
47 #include "guest_modes.h"
49 /* Global variable used to synchronize all of the vCPU threads. */
50 static int iteration = -1;
52 /* Defines what vCPU threads should do during a given iteration. */
54 /* Run the vCPU to access all its memory. */
55 ITERATION_ACCESS_MEMORY,
56 /* Mark the vCPU's memory idle in page_idle. */
60 /* Set to true when vCPU threads should exit. */
63 /* The iteration that was last completed by each vCPU. */
64 static int vcpu_last_completed_iteration[KVM_MAX_VCPUS];
66 /* Whether to overlap the regions of memory vCPUs access. */
67 static bool overlap_memory_access;
70 /* The backing source for the region of memory. */
71 enum vm_mem_backing_src_type backing_src;
73 /* The amount of memory to allocate for each vCPU. */
74 uint64_t vcpu_memory_bytes;
76 /* The number of vCPUs to create in the VM. */
80 static uint64_t pread_uint64(int fd, const char *filename, uint64_t index)
83 off_t offset = index * sizeof(value);
85 TEST_ASSERT(pread(fd, &value, sizeof(value), offset) == sizeof(value),
86 "pread from %s offset 0x%" PRIx64 " failed!",
93 #define PAGEMAP_PRESENT (1ULL << 63)
94 #define PAGEMAP_PFN_MASK ((1ULL << 55) - 1)
96 static uint64_t lookup_pfn(int pagemap_fd, struct kvm_vm *vm, uint64_t gva)
98 uint64_t hva = (uint64_t) addr_gva2hva(vm, gva);
102 entry = pread_uint64(pagemap_fd, "pagemap", hva / getpagesize());
103 if (!(entry & PAGEMAP_PRESENT))
106 pfn = entry & PAGEMAP_PFN_MASK;
108 print_skip("Looking up PFNs requires CAP_SYS_ADMIN");
115 static bool is_page_idle(int page_idle_fd, uint64_t pfn)
117 uint64_t bits = pread_uint64(page_idle_fd, "page_idle", pfn / 64);
119 return !!((bits >> (pfn % 64)) & 1);
122 static void mark_page_idle(int page_idle_fd, uint64_t pfn)
124 uint64_t bits = 1ULL << (pfn % 64);
126 TEST_ASSERT(pwrite(page_idle_fd, &bits, 8, 8 * (pfn / 64)) == 8,
127 "Set page_idle bits for PFN 0x%" PRIx64, pfn);
130 static void mark_vcpu_memory_idle(struct kvm_vm *vm, int vcpu_id)
132 uint64_t base_gva = perf_test_args.vcpu_args[vcpu_id].gva;
133 uint64_t pages = perf_test_args.vcpu_args[vcpu_id].pages;
135 uint64_t still_idle = 0;
140 /* If vCPUs are using an overlapping region, let vCPU 0 mark it idle. */
141 if (overlap_memory_access && vcpu_id)
144 page_idle_fd = open("/sys/kernel/mm/page_idle/bitmap", O_RDWR);
145 TEST_ASSERT(page_idle_fd > 0, "Failed to open page_idle.");
147 pagemap_fd = open("/proc/self/pagemap", O_RDONLY);
148 TEST_ASSERT(pagemap_fd > 0, "Failed to open pagemap.");
150 for (page = 0; page < pages; page++) {
151 uint64_t gva = base_gva + page * perf_test_args.guest_page_size;
152 uint64_t pfn = lookup_pfn(pagemap_fd, vm, gva);
159 if (is_page_idle(page_idle_fd, pfn)) {
164 mark_page_idle(page_idle_fd, pfn);
168 * Assumption: Less than 1% of pages are going to be swapped out from
169 * under us during this test.
171 TEST_ASSERT(no_pfn < pages / 100,
172 "vCPU %d: No PFN for %" PRIu64 " out of %" PRIu64 " pages.",
173 vcpu_id, no_pfn, pages);
176 * Test that at least 90% of memory has been marked idle (the rest might
177 * not be marked idle because the pages have not yet made it to an LRU
178 * list or the translations are still cached in the TLB). 90% is
179 * arbitrary; high enough that we ensure most memory access went through
180 * access tracking but low enough as to not make the test too brittle
181 * over time and across architectures.
183 TEST_ASSERT(still_idle < pages / 10,
184 "vCPU%d: Too many pages still idle (%"PRIu64 " out of %"
186 vcpu_id, still_idle, pages);
192 static void assert_ucall(struct kvm_vm *vm, uint32_t vcpu_id,
193 uint64_t expected_ucall)
196 uint64_t actual_ucall = get_ucall(vm, vcpu_id, &uc);
198 TEST_ASSERT(expected_ucall == actual_ucall,
199 "Guest exited unexpectedly (expected ucall %" PRIu64
200 ", got %" PRIu64 ")",
201 expected_ucall, actual_ucall);
204 static bool spin_wait_for_next_iteration(int *current_iteration)
206 int last_iteration = *current_iteration;
212 *current_iteration = READ_ONCE(iteration);
213 } while (last_iteration == *current_iteration);
218 static void *vcpu_thread_main(void *arg)
220 struct perf_test_vcpu_args *vcpu_args = arg;
221 struct kvm_vm *vm = perf_test_args.vm;
222 int vcpu_id = vcpu_args->vcpu_id;
223 int current_iteration = -1;
225 while (spin_wait_for_next_iteration(¤t_iteration)) {
226 switch (READ_ONCE(iteration_work)) {
227 case ITERATION_ACCESS_MEMORY:
228 vcpu_run(vm, vcpu_id);
229 assert_ucall(vm, vcpu_id, UCALL_SYNC);
231 case ITERATION_MARK_IDLE:
232 mark_vcpu_memory_idle(vm, vcpu_id);
236 vcpu_last_completed_iteration[vcpu_id] = current_iteration;
242 static void spin_wait_for_vcpu(int vcpu_id, int target_iteration)
244 while (READ_ONCE(vcpu_last_completed_iteration[vcpu_id]) !=
250 /* The type of memory accesses to perform in the VM. */
256 static void run_iteration(struct kvm_vm *vm, int vcpus, const char *description)
258 struct timespec ts_start;
259 struct timespec ts_elapsed;
263 /* Kick off the vCPUs by incrementing iteration. */
264 next_iteration = ++iteration;
266 clock_gettime(CLOCK_MONOTONIC, &ts_start);
268 /* Wait for all vCPUs to finish the iteration. */
269 for (vcpu_id = 0; vcpu_id < vcpus; vcpu_id++)
270 spin_wait_for_vcpu(vcpu_id, next_iteration);
272 ts_elapsed = timespec_elapsed(ts_start);
273 pr_info("%-30s: %ld.%09lds\n",
274 description, ts_elapsed.tv_sec, ts_elapsed.tv_nsec);
277 static void access_memory(struct kvm_vm *vm, int vcpus, enum access_type access,
278 const char *description)
280 perf_test_args.wr_fract = (access == ACCESS_READ) ? INT_MAX : 1;
281 sync_global_to_guest(vm, perf_test_args);
282 iteration_work = ITERATION_ACCESS_MEMORY;
283 run_iteration(vm, vcpus, description);
286 static void mark_memory_idle(struct kvm_vm *vm, int vcpus)
289 * Even though this parallelizes the work across vCPUs, this is still a
290 * very slow operation because page_idle forces the test to mark one pfn
291 * at a time and the clear_young notifier serializes on the KVM MMU
294 pr_debug("Marking VM memory idle (slow)...\n");
295 iteration_work = ITERATION_MARK_IDLE;
296 run_iteration(vm, vcpus, "Mark memory idle");
299 static pthread_t *create_vcpu_threads(int vcpus)
301 pthread_t *vcpu_threads;
304 vcpu_threads = malloc(vcpus * sizeof(vcpu_threads[0]));
305 TEST_ASSERT(vcpu_threads, "Failed to allocate vcpu_threads.");
307 for (i = 0; i < vcpus; i++) {
308 vcpu_last_completed_iteration[i] = iteration;
309 pthread_create(&vcpu_threads[i], NULL, vcpu_thread_main,
310 &perf_test_args.vcpu_args[i]);
316 static void terminate_vcpu_threads(pthread_t *vcpu_threads, int vcpus)
320 /* Set done to signal the vCPU threads to exit */
323 for (i = 0; i < vcpus; i++)
324 pthread_join(vcpu_threads[i], NULL);
327 static void run_test(enum vm_guest_mode mode, void *arg)
329 struct test_params *params = arg;
331 pthread_t *vcpu_threads;
332 int vcpus = params->vcpus;
334 vm = perf_test_create_vm(mode, vcpus, params->vcpu_memory_bytes, 1,
335 params->backing_src);
337 perf_test_setup_vcpus(vm, vcpus, params->vcpu_memory_bytes,
338 !overlap_memory_access);
340 vcpu_threads = create_vcpu_threads(vcpus);
343 access_memory(vm, vcpus, ACCESS_WRITE, "Populating memory");
345 /* As a control, read and write to the populated memory first. */
346 access_memory(vm, vcpus, ACCESS_WRITE, "Writing to populated memory");
347 access_memory(vm, vcpus, ACCESS_READ, "Reading from populated memory");
349 /* Repeat on memory that has been marked as idle. */
350 mark_memory_idle(vm, vcpus);
351 access_memory(vm, vcpus, ACCESS_WRITE, "Writing to idle memory");
352 mark_memory_idle(vm, vcpus);
353 access_memory(vm, vcpus, ACCESS_READ, "Reading from idle memory");
355 terminate_vcpu_threads(vcpu_threads, vcpus);
357 perf_test_destroy_vm(vm);
360 static void help(char *name)
363 printf("usage: %s [-h] [-m mode] [-b vcpu_bytes] [-v vcpus] [-o] [-s mem_type]\n",
366 printf(" -h: Display this help message.");
368 printf(" -b: specify the size of the memory region which should be\n"
369 " dirtied by each vCPU. e.g. 10M or 3G.\n"
371 printf(" -v: specify the number of vCPUs to run.\n");
372 printf(" -o: Overlap guest memory accesses instead of partitioning\n"
373 " them into a separate region of memory for each vCPU.\n");
374 printf(" -s: specify the type of memory that should be used to\n"
375 " back the guest data region.\n\n");
381 int main(int argc, char *argv[])
383 struct test_params params = {
384 .backing_src = VM_MEM_SRC_ANONYMOUS,
385 .vcpu_memory_bytes = DEFAULT_PER_VCPU_MEM_SIZE,
391 guest_modes_append_default();
393 while ((opt = getopt(argc, argv, "hm:b:v:os:")) != -1) {
396 guest_modes_cmdline(optarg);
399 params.vcpu_memory_bytes = parse_size(optarg);
402 params.vcpus = atoi(optarg);
405 overlap_memory_access = true;
408 params.backing_src = parse_backing_src_type(optarg);
417 page_idle_fd = open("/sys/kernel/mm/page_idle/bitmap", O_RDWR);
418 if (page_idle_fd < 0) {
419 print_skip("CONFIG_IDLE_PAGE_TRACKING is not enabled");
424 for_each_guest_mode(run_test, ¶ms);