1 // SPDX-License-Identifier: GPL-2.0
3 * A memslot-related performance benchmark.
5 * Copyright (C) 2021 Oracle and/or its affiliates.
7 * Basic guest setup / host vCPU thread code lifted from set_memory_region_test.
11 #include <semaphore.h>
12 #include <stdatomic.h>
22 #include <linux/compiler.h>
23 #include <linux/sizes.h>
25 #include <test_util.h>
27 #include <processor.h>
29 #define MEM_EXTRA_SIZE SZ_64K
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)
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).
44 #define MEM_SIZE_MAP (SZ_32M + MEM_EXTRA_SIZE)
45 #define MEM_TEST_MAP_SIZE (MEM_SIZE_MAP - MEM_EXTRA_SIZE)
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
51 * 2 MiB chunk size like a typical huge page
53 #define MEM_TEST_UNMAP_SIZE SZ_128M
54 #define MEM_TEST_UNMAP_CHUNK_SIZE SZ_2M
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
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().
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
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");
80 #define MEM_TEST_VAL_1 0x1122334455667788
81 #define MEM_TEST_VAL_2 0x99AABBCCDDEEFF00
85 struct kvm_vcpu *vcpu;
86 pthread_t vcpu_thread;
89 uint64_t pages_per_slot;
92 uint64_t mmio_gpa_min;
93 uint64_t mmio_gpa_max;
97 uint32_t guest_page_size;
98 atomic_bool start_flag;
99 atomic_bool exit_flag;
100 atomic_bool sync_flag;
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
108 * However, in practice both GCC and Clang fulfill this requirement on
109 * all KVM-supported platforms.
111 static_assert(ATOMIC_BOOL_LOCK_FREE == 2, "atomic bool is not lockless");
113 static sem_t vcpu_ready;
115 static bool map_unmap_verify;
118 #define pr_info_v(...) \
121 pr_info(__VA_ARGS__); \
124 static void check_mmio_access(struct vm_data *data, struct kvm_run *run)
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);
136 static void *vcpu_worker(void *__data)
138 struct vm_data *data = __data;
139 struct kvm_vcpu *vcpu = data->vcpu;
140 struct kvm_run *run = vcpu->run;
146 switch (get_ucall(vcpu, &uc)) {
148 TEST_ASSERT(uc.args[1] == 0,
149 "Unexpected sync ucall, got %lx",
151 sem_post(&vcpu_ready);
154 if (run->exit_reason == KVM_EXIT_MMIO)
155 check_mmio_access(data, run);
160 REPORT_GUEST_ASSERT(uc);
165 TEST_FAIL("Unknown ucall %lu", uc.cmd);
173 static void wait_for_vcpu(void)
177 TEST_ASSERT(!clock_gettime(CLOCK_REALTIME, &ts),
178 "clock_gettime() failed: %d\n", errno);
181 TEST_ASSERT(!sem_timedwait(&vcpu_ready, &ts),
182 "sem_timedwait() failed: %d\n", errno);
185 static void *vm_gpa2hva(struct vm_data *data, uint64_t gpa, uint64_t *rempages)
187 uint64_t gpage, pgoffs;
188 uint32_t slot, slotoffs;
190 uint32_t guest_page_size = data->vm->page_size;
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");
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);
205 if (slot == data->nslots - 1)
206 slotpages = data->npages - slot * data->pages_per_slot;
208 slotpages = data->pages_per_slot;
211 "Asking for remaining pages in slot but gpa not page aligned");
212 *rempages = slotpages - slotoffs;
215 base = data->hva_slots[slot];
216 return (uint8_t *)base + slotoffs * guest_page_size + pgoffs;
219 static uint64_t vm_slot2gpa(struct vm_data *data, uint32_t slot)
221 uint32_t guest_page_size = data->vm->page_size;
223 TEST_ASSERT(slot < data->nslots, "Too high slot number");
225 return MEM_GPA + slot * data->pages_per_slot * guest_page_size;
228 static struct vm_data *alloc_vm(void)
230 struct vm_data *data;
232 data = malloc(sizeof(*data));
233 TEST_ASSERT(data, "malloc(vmdata) failed");
237 data->hva_slots = NULL;
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)
248 if ((pages_per_slot * guest_page_size) % host_page_size)
251 if ((rempages * guest_page_size) % host_page_size)
258 static uint64_t get_max_slots(struct vm_data *data, uint32_t host_page_size)
260 uint32_t guest_page_size = data->vm->page_size;
261 uint64_t mempages, pages_per_slot, rempages;
264 mempages = data->npages;
265 slots = data->nslots;
266 while (--slots > 1) {
267 pages_per_slot = mempages / slots;
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 */
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)
284 uint64_t mempages, rempages;
286 uint32_t slot, host_page_size, guest_page_size;
287 struct timespec tstart;
288 struct sync_area *sync;
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;
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");
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);
308 data->hva_slots = malloc(sizeof(*data->hva_slots) * data->nslots);
309 TEST_ASSERT(data->hva_slots, "malloc() fail");
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);
314 clock_gettime(CLOCK_MONOTONIC, &tstart);
315 for (slot = 1, guest_addr = MEM_GPA; slot <= data->nslots; slot++) {
318 npages = data->pages_per_slot;
319 if (slot == data->nslots)
322 vm_userspace_mem_region_add(data->vm, VM_MEM_SRC_ANONYMOUS,
323 guest_addr, slot, npages,
325 guest_addr += npages * guest_page_size;
327 *slot_runtime = timespec_elapsed(tstart);
329 for (slot = 1, guest_addr = MEM_GPA; slot <= data->nslots; slot++) {
333 npages = data->pages_per_slot;
334 if (slot == data->nslots)
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");
341 data->hva_slots[slot - 1] = addr_gpa2hva(data->vm, guest_addr);
342 memset(data->hva_slots[slot - 1], 0, npages * guest_page_size);
344 guest_addr += npages * guest_page_size;
347 virt_map(data->vm, MEM_GPA, MEM_GPA, data->npages);
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);
355 data->mmio_ok = false;
360 static void launch_vm(struct vm_data *data)
362 pr_info_v("Launching the test VM\n");
364 pthread_create(&data->vcpu_thread, NULL, vcpu_worker, data);
366 /* Ensure the guest thread is spun up. */
370 static void free_vm(struct vm_data *data)
372 kvm_vm_free(data->vm);
373 free(data->hva_slots);
377 static void wait_guest_exit(struct vm_data *data)
379 pthread_join(data->vcpu_thread, NULL);
382 static void let_guest_run(struct sync_area *sync)
384 atomic_store_explicit(&sync->start_flag, true, memory_order_release);
387 static void guest_spin_until_start(void)
389 struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
391 while (!atomic_load_explicit(&sync->start_flag, memory_order_acquire))
395 static void make_guest_exit(struct sync_area *sync)
397 atomic_store_explicit(&sync->exit_flag, true, memory_order_release);
400 static bool _guest_should_exit(void)
402 struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
404 return atomic_load_explicit(&sync->exit_flag, memory_order_acquire);
407 #define guest_should_exit() unlikely(_guest_should_exit())
410 * noinline so we can easily see how much time the host spends waiting
412 * For the same reason use alarm() instead of polling clock_gettime()
413 * to implement a wait timeout.
415 static noinline void host_perform_sync(struct sync_area *sync)
419 atomic_store_explicit(&sync->sync_flag, true, memory_order_release);
420 while (atomic_load_explicit(&sync->sync_flag, memory_order_acquire))
426 static bool guest_perform_sync(void)
428 struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
432 if (guest_should_exit())
436 } while (!atomic_compare_exchange_weak_explicit(&sync->sync_flag,
438 memory_order_acq_rel,
439 memory_order_relaxed));
444 static void guest_code_test_memslot_move(void)
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);
452 guest_spin_until_start();
454 while (!guest_should_exit()) {
457 for (ptr = base; ptr < base + MEM_TEST_MOVE_SIZE;
459 *(uint64_t *)ptr = MEM_TEST_VAL_1;
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
473 static void guest_code_test_memslot_map(void)
475 struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
476 uint32_t page_size = (typeof(page_size))READ_ONCE(sync->guest_page_size);
480 guest_spin_until_start();
485 for (ptr = MEM_TEST_GPA;
486 ptr < MEM_TEST_GPA + MEM_TEST_MAP_SIZE / 2;
488 *(uint64_t *)ptr = MEM_TEST_VAL_1;
490 if (!guest_perform_sync())
493 for (ptr = MEM_TEST_GPA + MEM_TEST_MAP_SIZE / 2;
494 ptr < MEM_TEST_GPA + MEM_TEST_MAP_SIZE;
496 *(uint64_t *)ptr = MEM_TEST_VAL_2;
498 if (!guest_perform_sync())
505 static void guest_code_test_memslot_unmap(void)
507 struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
511 guest_spin_until_start();
514 uintptr_t ptr = MEM_TEST_GPA;
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.
523 * Just access a single page to be on the safe side.
525 *(uint64_t *)ptr = MEM_TEST_VAL_1;
527 if (!guest_perform_sync())
530 ptr += MEM_TEST_UNMAP_SIZE / 2;
531 *(uint64_t *)ptr = MEM_TEST_VAL_2;
533 if (!guest_perform_sync())
540 static void guest_code_test_memslot_rw(void)
542 struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
543 uint32_t page_size = (typeof(page_size))READ_ONCE(sync->guest_page_size);
547 guest_spin_until_start();
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;
556 if (!guest_perform_sync())
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;
563 GUEST_ASSERT_EQ(val, MEM_TEST_VAL_2);
564 *(uint64_t *)ptr = 0;
567 if (!guest_perform_sync())
574 static bool test_memslot_move_prepare(struct vm_data *data,
575 struct sync_area *sync,
576 uint64_t *maxslots, bool isactive)
578 uint32_t guest_page_size = data->vm->page_size;
579 uint64_t movesrcgpa, movetestgpa;
581 movesrcgpa = vm_slot2gpa(data, data->nslots - 1);
586 vm_gpa2hva(data, movesrcgpa, &lastpages);
587 if (lastpages * guest_page_size < MEM_TEST_MOVE_SIZE / 2) {
593 movetestgpa = movesrcgpa - (MEM_TEST_MOVE_SIZE / (isactive ? 2 : 1));
594 sync->move_area_ptr = (void *)movetestgpa;
597 data->mmio_ok = true;
598 data->mmio_gpa_min = movesrcgpa;
599 data->mmio_gpa_max = movesrcgpa + MEM_TEST_MOVE_SIZE / 2 - 1;
605 static bool test_memslot_move_prepare_active(struct vm_data *data,
606 struct sync_area *sync,
609 return test_memslot_move_prepare(data, sync, maxslots, true);
612 static bool test_memslot_move_prepare_inactive(struct vm_data *data,
613 struct sync_area *sync,
616 return test_memslot_move_prepare(data, sync, maxslots, false);
619 static void test_memslot_move_loop(struct vm_data *data, struct sync_area *sync)
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);
629 static void test_memslot_do_unmap(struct vm_data *data,
630 uint64_t offsp, uint64_t count)
633 uint32_t guest_page_size = data->vm->page_size;
635 for (gpa = MEM_TEST_GPA + offsp * guest_page_size, ctr = 0; ctr < count; ) {
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);
645 "madvise(%p, MADV_DONTNEED) on VM memory should not fail for gptr 0x%"PRIx64,
648 gpa += npages * guest_page_size;
650 TEST_ASSERT(ctr == count,
651 "madvise(MADV_DONTNEED) should exactly cover all of the requested area");
654 static void test_memslot_map_unmap_check(struct vm_data *data,
655 uint64_t offsp, uint64_t valexp)
659 uint32_t guest_page_size = data->vm->page_size;
661 if (!map_unmap_verify)
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")",
672 static void test_memslot_map_loop(struct vm_data *data, struct sync_area *sync)
674 uint32_t guest_page_size = data->vm->page_size;
675 uint64_t guest_pages = MEM_TEST_MAP_SIZE / guest_page_size;
678 * Unmap the second half of the test area while guest writes to (maps)
681 test_memslot_do_unmap(data, guest_pages / 2, guest_pages / 2);
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
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);
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
700 * The area will be unmapped at the beginning of the next loop
702 * Meanwhile, the guest is writing to (mapping) the first half of
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);
710 static void test_memslot_unmap_loop_common(struct vm_data *data,
711 struct sync_area *sync,
714 uint32_t guest_page_size = data->vm->page_size;
715 uint64_t guest_pages = MEM_TEST_UNMAP_SIZE / guest_page_size;
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
722 * Meanwhile, the guest is writing to (mapping) page(s) in the second
723 * half of the test area.
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);
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);
737 static void test_memslot_unmap_loop(struct vm_data *data,
738 struct sync_area *sync)
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;
745 test_memslot_unmap_loop_common(data, sync, guest_chunk_pages);
748 static void test_memslot_unmap_loop_chunked(struct vm_data *data,
749 struct sync_area *sync)
751 uint32_t guest_page_size = data->vm->page_size;
752 uint64_t guest_chunk_pages = MEM_TEST_UNMAP_CHUNK_SIZE / guest_page_size;
754 test_memslot_unmap_loop_common(data, sync, guest_chunk_pages);
757 static void test_memslot_rw_loop(struct vm_data *data, struct sync_area *sync)
760 uint32_t guest_page_size = data->vm->page_size;
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;
766 host_perform_sync(sync);
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;
773 TEST_ASSERT(val == MEM_TEST_VAL_1,
774 "Guest written values should read back correctly (is %"PRIu64" @ %"PRIx64")",
779 host_perform_sync(sync);
785 void (*guest_code)(void);
786 bool (*prepare)(struct vm_data *data, struct sync_area *sync,
788 void (*loop)(struct vm_data *data, struct sync_area *sync);
791 static bool test_execute(int nslots, uint64_t *maxslots,
792 unsigned int maxtime,
793 const struct test_data *tdata,
795 struct timespec *slot_runtime,
796 struct timespec *guest_runtime)
798 uint64_t mem_size = tdata->mem_size ? : MEM_SIZE;
799 struct vm_data *data;
800 struct sync_area *sync;
801 struct timespec tstart;
805 if (!prepare_vm(data, nslots, maxslots, tdata->guest_code,
806 mem_size, slot_runtime)) {
811 sync = (typeof(sync))vm_gpa2hva(data, MEM_SYNC_GPA, NULL);
812 if (tdata->prepare &&
813 !tdata->prepare(data, sync, maxslots)) {
820 clock_gettime(CLOCK_MONOTONIC, &tstart);
824 *guest_runtime = timespec_elapsed(tstart);
825 if (guest_runtime->tv_sec >= maxtime)
828 tdata->loop(data, sync);
833 make_guest_exit(sync);
834 wait_guest_exit(data);
842 static const struct test_data tests[] = {
845 .mem_size = MEM_SIZE_MAP,
846 .guest_code = guest_code_test_memslot_map,
847 .loop = test_memslot_map_loop,
851 .mem_size = MEM_TEST_UNMAP_SIZE + MEM_EXTRA_SIZE,
852 .guest_code = guest_code_test_memslot_unmap,
853 .loop = test_memslot_unmap_loop,
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,
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,
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,
875 .guest_code = guest_code_test_memslot_rw,
876 .loop = test_memslot_rw_loop
880 #define NTESTS ARRAY_SIZE(tests)
890 static void help(char *name, struct test_args *targs)
894 pr_info("usage: %s [-h] [-v] [-d] [-s slots] [-f first_test] [-e last_test] [-l test_length] [-r run_count]\n",
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",
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",
907 pr_info(" -r: specify the number of runs per test (currently: %i)\n",
910 pr_info("\nAvailable tests:\n");
911 for (ctr = 0; ctr < NTESTS; ctr++)
912 pr_info("%d: %s\n", ctr, tests[ctr].name);
915 static bool check_memory_sizes(void)
917 uint32_t host_page_size = getpagesize();
918 uint32_t guest_page_size = vm_guest_mode_params[VM_MODE_DEFAULT].page_size;
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);
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");
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");
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");
951 static bool parse_args(int argc, char *argv[],
952 struct test_args *targs)
954 uint32_t max_mem_slots;
957 while ((opt = getopt(argc, argv, "hvds:f:e:l:r:")) != -1) {
961 help(argv[0], targs);
967 map_unmap_verify = true;
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");
977 targs->tfirst = atoi_non_negative("First test", optarg);
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",
988 targs->seconds = atoi_non_negative("Test length", optarg);
991 targs->runs = atoi_positive("Runs per test", optarg);
997 help(argv[0], targs);
1001 if (targs->tfirst > targs->tlast) {
1002 pr_info("First test to run cannot be greater than the last test to run\n");
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");
1012 /* Memory slot 0 is reserved */
1013 if (targs->nslots == -1)
1014 targs->nslots = max_mem_slots - 1;
1016 targs->nslots = min_t(int, targs->nslots, max_mem_slots) - 1;
1018 pr_info_v("Allowed Number of memory slots: %"PRIu32"\n",
1024 struct test_result {
1025 struct timespec slot_runtime, guest_runtime, iter_runtime;
1026 int64_t slottimens, runtimens;
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)
1036 struct test_result result;
1039 if (!test_execute(targs->nslots, &maxslots, targs->seconds, data,
1041 &result.slot_runtime, &result.guest_runtime)) {
1043 pr_info("Memslot count too high for this test, decrease the cap (max is %"PRIu64")\n",
1046 pr_info("Memslot count may be too high for this test, try adjusting the cap\n");
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");
1059 result.iter_runtime = timespec_div(result.guest_runtime,
1061 pr_info("Done %"PRIu64" iterations, avg %ld.%.9lds each\n",
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);
1069 * Only rank the slot setup time for tests using the whole test memory
1070 * area so they are comparable
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;
1083 int main(int argc, char *argv[])
1085 struct test_args targs = {
1087 .tlast = NTESTS - 1,
1092 struct test_result rbestslottime;
1095 if (!check_memory_sizes())
1098 if (!parse_args(argc, argv, &targs))
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;
1107 if (tctr > targs.tfirst)
1110 pr_info("Testing %s performance with %i runs, %d seconds each\n",
1111 data->name, targs.runs, targs.seconds);
1113 rbestruntime.runtimens = 0;
1114 for (runctr = 0; runctr < targs.runs; runctr++)
1115 if (!test_loop(data, &targs,
1116 &rbestslottime, &rbestruntime))
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);
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);