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;
117 static bool disable_slot_zap_quirk;
121 #define pr_info_v(...) \
124 pr_info(__VA_ARGS__); \
127 static void check_mmio_access(struct vm_data *data, struct kvm_run *run)
129 TEST_ASSERT(data->mmio_ok, "Unexpected mmio exit");
130 TEST_ASSERT(run->mmio.is_write, "Unexpected mmio read");
131 TEST_ASSERT(run->mmio.len == 8,
132 "Unexpected exit mmio size = %u", run->mmio.len);
133 TEST_ASSERT(run->mmio.phys_addr >= data->mmio_gpa_min &&
134 run->mmio.phys_addr <= data->mmio_gpa_max,
135 "Unexpected exit mmio address = 0x%llx",
136 run->mmio.phys_addr);
139 static void *vcpu_worker(void *__data)
141 struct vm_data *data = __data;
142 struct kvm_vcpu *vcpu = data->vcpu;
143 struct kvm_run *run = vcpu->run;
149 switch (get_ucall(vcpu, &uc)) {
151 TEST_ASSERT(uc.args[1] == 0,
152 "Unexpected sync ucall, got %lx",
154 sem_post(&vcpu_ready);
157 if (run->exit_reason == KVM_EXIT_MMIO)
158 check_mmio_access(data, run);
163 REPORT_GUEST_ASSERT(uc);
168 TEST_FAIL("Unknown ucall %lu", uc.cmd);
176 static void wait_for_vcpu(void)
180 TEST_ASSERT(!clock_gettime(CLOCK_REALTIME, &ts),
181 "clock_gettime() failed: %d", errno);
184 TEST_ASSERT(!sem_timedwait(&vcpu_ready, &ts),
185 "sem_timedwait() failed: %d", errno);
188 static void *vm_gpa2hva(struct vm_data *data, uint64_t gpa, uint64_t *rempages)
190 uint64_t gpage, pgoffs;
191 uint32_t slot, slotoffs;
193 uint32_t guest_page_size = data->vm->page_size;
195 TEST_ASSERT(gpa >= MEM_GPA, "Too low gpa to translate");
196 TEST_ASSERT(gpa < MEM_GPA + data->npages * guest_page_size,
197 "Too high gpa to translate");
200 gpage = gpa / guest_page_size;
201 pgoffs = gpa % guest_page_size;
202 slot = min(gpage / data->pages_per_slot, (uint64_t)data->nslots - 1);
203 slotoffs = gpage - (slot * data->pages_per_slot);
208 if (slot == data->nslots - 1)
209 slotpages = data->npages - slot * data->pages_per_slot;
211 slotpages = data->pages_per_slot;
214 "Asking for remaining pages in slot but gpa not page aligned");
215 *rempages = slotpages - slotoffs;
218 base = data->hva_slots[slot];
219 return (uint8_t *)base + slotoffs * guest_page_size + pgoffs;
222 static uint64_t vm_slot2gpa(struct vm_data *data, uint32_t slot)
224 uint32_t guest_page_size = data->vm->page_size;
226 TEST_ASSERT(slot < data->nslots, "Too high slot number");
228 return MEM_GPA + slot * data->pages_per_slot * guest_page_size;
231 static struct vm_data *alloc_vm(void)
233 struct vm_data *data;
235 data = malloc(sizeof(*data));
236 TEST_ASSERT(data, "malloc(vmdata) failed");
240 data->hva_slots = NULL;
245 static bool check_slot_pages(uint32_t host_page_size, uint32_t guest_page_size,
246 uint64_t pages_per_slot, uint64_t rempages)
251 if ((pages_per_slot * guest_page_size) % host_page_size)
254 if ((rempages * guest_page_size) % host_page_size)
261 static uint64_t get_max_slots(struct vm_data *data, uint32_t host_page_size)
263 uint32_t guest_page_size = data->vm->page_size;
264 uint64_t mempages, pages_per_slot, rempages;
267 mempages = data->npages;
268 slots = data->nslots;
269 while (--slots > 1) {
270 pages_per_slot = mempages / slots;
274 rempages = mempages % pages_per_slot;
275 if (check_slot_pages(host_page_size, guest_page_size,
276 pages_per_slot, rempages))
277 return slots + 1; /* slot 0 is reserved */
283 static bool prepare_vm(struct vm_data *data, int nslots, uint64_t *maxslots,
284 void *guest_code, uint64_t mem_size,
285 struct timespec *slot_runtime)
287 uint64_t mempages, rempages;
289 uint32_t slot, host_page_size, guest_page_size;
290 struct timespec tstart;
291 struct sync_area *sync;
293 host_page_size = getpagesize();
294 guest_page_size = vm_guest_mode_params[VM_MODE_DEFAULT].page_size;
295 mempages = mem_size / guest_page_size;
297 data->vm = __vm_create_with_one_vcpu(&data->vcpu, mempages, guest_code);
298 TEST_ASSERT(data->vm->page_size == guest_page_size, "Invalid VM page size");
300 data->npages = mempages;
301 TEST_ASSERT(data->npages > 1, "Can't test without any memory");
302 data->nslots = nslots;
303 data->pages_per_slot = data->npages / data->nslots;
304 rempages = data->npages % data->nslots;
305 if (!check_slot_pages(host_page_size, guest_page_size,
306 data->pages_per_slot, rempages)) {
307 *maxslots = get_max_slots(data, host_page_size);
311 data->hva_slots = malloc(sizeof(*data->hva_slots) * data->nslots);
312 TEST_ASSERT(data->hva_slots, "malloc() fail");
314 pr_info_v("Adding slots 1..%i, each slot with %"PRIu64" pages + %"PRIu64" extra pages last\n",
315 data->nslots, data->pages_per_slot, rempages);
317 clock_gettime(CLOCK_MONOTONIC, &tstart);
318 for (slot = 1, guest_addr = MEM_GPA; slot <= data->nslots; slot++) {
321 npages = data->pages_per_slot;
322 if (slot == data->nslots)
325 vm_userspace_mem_region_add(data->vm, VM_MEM_SRC_ANONYMOUS,
326 guest_addr, slot, npages,
328 guest_addr += npages * guest_page_size;
330 *slot_runtime = timespec_elapsed(tstart);
332 for (slot = 1, guest_addr = MEM_GPA; slot <= data->nslots; slot++) {
336 npages = data->pages_per_slot;
337 if (slot == data->nslots)
340 gpa = vm_phy_pages_alloc(data->vm, npages, guest_addr, slot);
341 TEST_ASSERT(gpa == guest_addr,
342 "vm_phy_pages_alloc() failed");
344 data->hva_slots[slot - 1] = addr_gpa2hva(data->vm, guest_addr);
345 memset(data->hva_slots[slot - 1], 0, npages * guest_page_size);
347 guest_addr += npages * guest_page_size;
350 virt_map(data->vm, MEM_GPA, MEM_GPA, data->npages);
352 sync = (typeof(sync))vm_gpa2hva(data, MEM_SYNC_GPA, NULL);
353 sync->guest_page_size = data->vm->page_size;
354 atomic_init(&sync->start_flag, false);
355 atomic_init(&sync->exit_flag, false);
356 atomic_init(&sync->sync_flag, false);
358 data->mmio_ok = false;
363 static void launch_vm(struct vm_data *data)
365 pr_info_v("Launching the test VM\n");
367 pthread_create(&data->vcpu_thread, NULL, vcpu_worker, data);
369 /* Ensure the guest thread is spun up. */
373 static void free_vm(struct vm_data *data)
375 kvm_vm_free(data->vm);
376 free(data->hva_slots);
380 static void wait_guest_exit(struct vm_data *data)
382 pthread_join(data->vcpu_thread, NULL);
385 static void let_guest_run(struct sync_area *sync)
387 atomic_store_explicit(&sync->start_flag, true, memory_order_release);
390 static void guest_spin_until_start(void)
392 struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
394 while (!atomic_load_explicit(&sync->start_flag, memory_order_acquire))
398 static void make_guest_exit(struct sync_area *sync)
400 atomic_store_explicit(&sync->exit_flag, true, memory_order_release);
403 static bool _guest_should_exit(void)
405 struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
407 return atomic_load_explicit(&sync->exit_flag, memory_order_acquire);
410 #define guest_should_exit() unlikely(_guest_should_exit())
413 * noinline so we can easily see how much time the host spends waiting
415 * For the same reason use alarm() instead of polling clock_gettime()
416 * to implement a wait timeout.
418 static noinline void host_perform_sync(struct sync_area *sync)
422 atomic_store_explicit(&sync->sync_flag, true, memory_order_release);
423 while (atomic_load_explicit(&sync->sync_flag, memory_order_acquire))
429 static bool guest_perform_sync(void)
431 struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
435 if (guest_should_exit())
439 } while (!atomic_compare_exchange_weak_explicit(&sync->sync_flag,
441 memory_order_acq_rel,
442 memory_order_relaxed));
447 static void guest_code_test_memslot_move(void)
449 struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
450 uint32_t page_size = (typeof(page_size))READ_ONCE(sync->guest_page_size);
451 uintptr_t base = (typeof(base))READ_ONCE(sync->move_area_ptr);
455 guest_spin_until_start();
457 while (!guest_should_exit()) {
460 for (ptr = base; ptr < base + MEM_TEST_MOVE_SIZE;
462 *(uint64_t *)ptr = MEM_TEST_VAL_1;
465 * No host sync here since the MMIO exits are so expensive
466 * that the host would spend most of its time waiting for
467 * the guest and so instead of measuring memslot move
468 * performance we would measure the performance and
469 * likelihood of MMIO exits
476 static void guest_code_test_memslot_map(void)
478 struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
479 uint32_t page_size = (typeof(page_size))READ_ONCE(sync->guest_page_size);
483 guest_spin_until_start();
488 for (ptr = MEM_TEST_GPA;
489 ptr < MEM_TEST_GPA + MEM_TEST_MAP_SIZE / 2;
491 *(uint64_t *)ptr = MEM_TEST_VAL_1;
493 if (!guest_perform_sync())
496 for (ptr = MEM_TEST_GPA + MEM_TEST_MAP_SIZE / 2;
497 ptr < MEM_TEST_GPA + MEM_TEST_MAP_SIZE;
499 *(uint64_t *)ptr = MEM_TEST_VAL_2;
501 if (!guest_perform_sync())
508 static void guest_code_test_memslot_unmap(void)
510 struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
514 guest_spin_until_start();
517 uintptr_t ptr = MEM_TEST_GPA;
520 * We can afford to access (map) just a small number of pages
521 * per host sync as otherwise the host will spend
522 * a significant amount of its time waiting for the guest
523 * (instead of doing unmap operations), so this will
524 * effectively turn this test into a map performance test.
526 * Just access a single page to be on the safe side.
528 *(uint64_t *)ptr = MEM_TEST_VAL_1;
530 if (!guest_perform_sync())
533 ptr += MEM_TEST_UNMAP_SIZE / 2;
534 *(uint64_t *)ptr = MEM_TEST_VAL_2;
536 if (!guest_perform_sync())
543 static void guest_code_test_memslot_rw(void)
545 struct sync_area *sync = (typeof(sync))MEM_SYNC_GPA;
546 uint32_t page_size = (typeof(page_size))READ_ONCE(sync->guest_page_size);
550 guest_spin_until_start();
555 for (ptr = MEM_TEST_GPA;
556 ptr < MEM_TEST_GPA + MEM_TEST_SIZE; ptr += page_size)
557 *(uint64_t *)ptr = MEM_TEST_VAL_1;
559 if (!guest_perform_sync())
562 for (ptr = MEM_TEST_GPA + page_size / 2;
563 ptr < MEM_TEST_GPA + MEM_TEST_SIZE; ptr += page_size) {
564 uint64_t val = *(uint64_t *)ptr;
566 GUEST_ASSERT_EQ(val, MEM_TEST_VAL_2);
567 *(uint64_t *)ptr = 0;
570 if (!guest_perform_sync())
577 static bool test_memslot_move_prepare(struct vm_data *data,
578 struct sync_area *sync,
579 uint64_t *maxslots, bool isactive)
581 uint32_t guest_page_size = data->vm->page_size;
582 uint64_t movesrcgpa, movetestgpa;
585 if (disable_slot_zap_quirk)
586 vm_enable_cap(data->vm, KVM_CAP_DISABLE_QUIRKS2, KVM_X86_QUIRK_SLOT_ZAP_ALL);
589 movesrcgpa = vm_slot2gpa(data, data->nslots - 1);
594 vm_gpa2hva(data, movesrcgpa, &lastpages);
595 if (lastpages * guest_page_size < MEM_TEST_MOVE_SIZE / 2) {
601 movetestgpa = movesrcgpa - (MEM_TEST_MOVE_SIZE / (isactive ? 2 : 1));
602 sync->move_area_ptr = (void *)movetestgpa;
605 data->mmio_ok = true;
606 data->mmio_gpa_min = movesrcgpa;
607 data->mmio_gpa_max = movesrcgpa + MEM_TEST_MOVE_SIZE / 2 - 1;
613 static bool test_memslot_move_prepare_active(struct vm_data *data,
614 struct sync_area *sync,
617 return test_memslot_move_prepare(data, sync, maxslots, true);
620 static bool test_memslot_move_prepare_inactive(struct vm_data *data,
621 struct sync_area *sync,
624 return test_memslot_move_prepare(data, sync, maxslots, false);
627 static void test_memslot_move_loop(struct vm_data *data, struct sync_area *sync)
631 movesrcgpa = vm_slot2gpa(data, data->nslots - 1);
632 vm_mem_region_move(data->vm, data->nslots - 1 + 1,
633 MEM_TEST_MOVE_GPA_DEST);
634 vm_mem_region_move(data->vm, data->nslots - 1 + 1, movesrcgpa);
637 static void test_memslot_do_unmap(struct vm_data *data,
638 uint64_t offsp, uint64_t count)
641 uint32_t guest_page_size = data->vm->page_size;
643 for (gpa = MEM_TEST_GPA + offsp * guest_page_size, ctr = 0; ctr < count; ) {
648 hva = vm_gpa2hva(data, gpa, &npages);
649 TEST_ASSERT(npages, "Empty memory slot at gptr 0x%"PRIx64, gpa);
650 npages = min(npages, count - ctr);
651 ret = madvise(hva, npages * guest_page_size, MADV_DONTNEED);
653 "madvise(%p, MADV_DONTNEED) on VM memory should not fail for gptr 0x%"PRIx64,
656 gpa += npages * guest_page_size;
658 TEST_ASSERT(ctr == count,
659 "madvise(MADV_DONTNEED) should exactly cover all of the requested area");
662 static void test_memslot_map_unmap_check(struct vm_data *data,
663 uint64_t offsp, uint64_t valexp)
667 uint32_t guest_page_size = data->vm->page_size;
669 if (!map_unmap_verify)
672 gpa = MEM_TEST_GPA + offsp * guest_page_size;
673 val = (typeof(val))vm_gpa2hva(data, gpa, NULL);
674 TEST_ASSERT(*val == valexp,
675 "Guest written values should read back correctly before unmap (%"PRIu64" vs %"PRIu64" @ %"PRIx64")",
680 static void test_memslot_map_loop(struct vm_data *data, struct sync_area *sync)
682 uint32_t guest_page_size = data->vm->page_size;
683 uint64_t guest_pages = MEM_TEST_MAP_SIZE / guest_page_size;
686 * Unmap the second half of the test area while guest writes to (maps)
689 test_memslot_do_unmap(data, guest_pages / 2, guest_pages / 2);
692 * Wait for the guest to finish writing the first half of the test
693 * area, verify the written value on the first and the last page of
694 * this area and then unmap it.
695 * Meanwhile, the guest is writing to (mapping) the second half of
698 host_perform_sync(sync);
699 test_memslot_map_unmap_check(data, 0, MEM_TEST_VAL_1);
700 test_memslot_map_unmap_check(data, guest_pages / 2 - 1, MEM_TEST_VAL_1);
701 test_memslot_do_unmap(data, 0, guest_pages / 2);
705 * Wait for the guest to finish writing the second half of the test
706 * area and verify the written value on the first and the last page
708 * The area will be unmapped at the beginning of the next loop
710 * Meanwhile, the guest is writing to (mapping) the first half of
713 host_perform_sync(sync);
714 test_memslot_map_unmap_check(data, guest_pages / 2, MEM_TEST_VAL_2);
715 test_memslot_map_unmap_check(data, guest_pages - 1, MEM_TEST_VAL_2);
718 static void test_memslot_unmap_loop_common(struct vm_data *data,
719 struct sync_area *sync,
722 uint32_t guest_page_size = data->vm->page_size;
723 uint64_t guest_pages = MEM_TEST_UNMAP_SIZE / guest_page_size;
727 * Wait for the guest to finish mapping page(s) in the first half
728 * of the test area, verify the written value and then perform unmap
730 * Meanwhile, the guest is writing to (mapping) page(s) in the second
731 * half of the test area.
733 host_perform_sync(sync);
734 test_memslot_map_unmap_check(data, 0, MEM_TEST_VAL_1);
735 for (ctr = 0; ctr < guest_pages / 2; ctr += chunk)
736 test_memslot_do_unmap(data, ctr, chunk);
738 /* Likewise, but for the opposite host / guest areas */
739 host_perform_sync(sync);
740 test_memslot_map_unmap_check(data, guest_pages / 2, MEM_TEST_VAL_2);
741 for (ctr = guest_pages / 2; ctr < guest_pages; ctr += chunk)
742 test_memslot_do_unmap(data, ctr, chunk);
745 static void test_memslot_unmap_loop(struct vm_data *data,
746 struct sync_area *sync)
748 uint32_t host_page_size = getpagesize();
749 uint32_t guest_page_size = data->vm->page_size;
750 uint64_t guest_chunk_pages = guest_page_size >= host_page_size ?
751 1 : host_page_size / guest_page_size;
753 test_memslot_unmap_loop_common(data, sync, guest_chunk_pages);
756 static void test_memslot_unmap_loop_chunked(struct vm_data *data,
757 struct sync_area *sync)
759 uint32_t guest_page_size = data->vm->page_size;
760 uint64_t guest_chunk_pages = MEM_TEST_UNMAP_CHUNK_SIZE / guest_page_size;
762 test_memslot_unmap_loop_common(data, sync, guest_chunk_pages);
765 static void test_memslot_rw_loop(struct vm_data *data, struct sync_area *sync)
768 uint32_t guest_page_size = data->vm->page_size;
770 for (gptr = MEM_TEST_GPA + guest_page_size / 2;
771 gptr < MEM_TEST_GPA + MEM_TEST_SIZE; gptr += guest_page_size)
772 *(uint64_t *)vm_gpa2hva(data, gptr, NULL) = MEM_TEST_VAL_2;
774 host_perform_sync(sync);
776 for (gptr = MEM_TEST_GPA;
777 gptr < MEM_TEST_GPA + MEM_TEST_SIZE; gptr += guest_page_size) {
778 uint64_t *vptr = (typeof(vptr))vm_gpa2hva(data, gptr, NULL);
779 uint64_t val = *vptr;
781 TEST_ASSERT(val == MEM_TEST_VAL_1,
782 "Guest written values should read back correctly (is %"PRIu64" @ %"PRIx64")",
787 host_perform_sync(sync);
793 void (*guest_code)(void);
794 bool (*prepare)(struct vm_data *data, struct sync_area *sync,
796 void (*loop)(struct vm_data *data, struct sync_area *sync);
799 static bool test_execute(int nslots, uint64_t *maxslots,
800 unsigned int maxtime,
801 const struct test_data *tdata,
803 struct timespec *slot_runtime,
804 struct timespec *guest_runtime)
806 uint64_t mem_size = tdata->mem_size ? : MEM_SIZE;
807 struct vm_data *data;
808 struct sync_area *sync;
809 struct timespec tstart;
813 if (!prepare_vm(data, nslots, maxslots, tdata->guest_code,
814 mem_size, slot_runtime)) {
819 sync = (typeof(sync))vm_gpa2hva(data, MEM_SYNC_GPA, NULL);
820 if (tdata->prepare &&
821 !tdata->prepare(data, sync, maxslots)) {
828 clock_gettime(CLOCK_MONOTONIC, &tstart);
832 *guest_runtime = timespec_elapsed(tstart);
833 if (guest_runtime->tv_sec >= maxtime)
836 tdata->loop(data, sync);
841 make_guest_exit(sync);
842 wait_guest_exit(data);
850 static const struct test_data tests[] = {
853 .mem_size = MEM_SIZE_MAP,
854 .guest_code = guest_code_test_memslot_map,
855 .loop = test_memslot_map_loop,
859 .mem_size = MEM_TEST_UNMAP_SIZE + MEM_EXTRA_SIZE,
860 .guest_code = guest_code_test_memslot_unmap,
861 .loop = test_memslot_unmap_loop,
864 .name = "unmap chunked",
865 .mem_size = MEM_TEST_UNMAP_SIZE + MEM_EXTRA_SIZE,
866 .guest_code = guest_code_test_memslot_unmap,
867 .loop = test_memslot_unmap_loop_chunked,
870 .name = "move active area",
871 .guest_code = guest_code_test_memslot_move,
872 .prepare = test_memslot_move_prepare_active,
873 .loop = test_memslot_move_loop,
876 .name = "move inactive area",
877 .guest_code = guest_code_test_memslot_move,
878 .prepare = test_memslot_move_prepare_inactive,
879 .loop = test_memslot_move_loop,
883 .guest_code = guest_code_test_memslot_rw,
884 .loop = test_memslot_rw_loop
888 #define NTESTS ARRAY_SIZE(tests)
898 static void help(char *name, struct test_args *targs)
902 pr_info("usage: %s [-h] [-v] [-d] [-s slots] [-f first_test] [-e last_test] [-l test_length] [-r run_count]\n",
904 pr_info(" -h: print this help screen.\n");
905 pr_info(" -v: enable verbose mode (not for benchmarking).\n");
906 pr_info(" -d: enable extra debug checks.\n");
907 pr_info(" -q: Disable memslot zap quirk during memslot move.\n");
908 pr_info(" -s: specify memslot count cap (-1 means no cap; currently: %i)\n",
910 pr_info(" -f: specify the first test to run (currently: %i; max %zu)\n",
911 targs->tfirst, NTESTS - 1);
912 pr_info(" -e: specify the last test to run (currently: %i; max %zu)\n",
913 targs->tlast, NTESTS - 1);
914 pr_info(" -l: specify the test length in seconds (currently: %i)\n",
916 pr_info(" -r: specify the number of runs per test (currently: %i)\n",
919 pr_info("\nAvailable tests:\n");
920 for (ctr = 0; ctr < NTESTS; ctr++)
921 pr_info("%d: %s\n", ctr, tests[ctr].name);
924 static bool check_memory_sizes(void)
926 uint32_t host_page_size = getpagesize();
927 uint32_t guest_page_size = vm_guest_mode_params[VM_MODE_DEFAULT].page_size;
929 if (host_page_size > SZ_64K || guest_page_size > SZ_64K) {
930 pr_info("Unsupported page size on host (0x%x) or guest (0x%x)\n",
931 host_page_size, guest_page_size);
935 if (MEM_SIZE % guest_page_size ||
936 MEM_TEST_SIZE % guest_page_size) {
937 pr_info("invalid MEM_SIZE or MEM_TEST_SIZE\n");
941 if (MEM_SIZE_MAP % guest_page_size ||
942 MEM_TEST_MAP_SIZE % guest_page_size ||
943 (MEM_TEST_MAP_SIZE / guest_page_size) <= 2 ||
944 (MEM_TEST_MAP_SIZE / guest_page_size) % 2) {
945 pr_info("invalid MEM_SIZE_MAP or MEM_TEST_MAP_SIZE\n");
949 if (MEM_TEST_UNMAP_SIZE > MEM_TEST_SIZE ||
950 MEM_TEST_UNMAP_SIZE % guest_page_size ||
951 (MEM_TEST_UNMAP_SIZE / guest_page_size) %
952 (2 * MEM_TEST_UNMAP_CHUNK_SIZE / guest_page_size)) {
953 pr_info("invalid MEM_TEST_UNMAP_SIZE or MEM_TEST_UNMAP_CHUNK_SIZE\n");
960 static bool parse_args(int argc, char *argv[],
961 struct test_args *targs)
963 uint32_t max_mem_slots;
966 while ((opt = getopt(argc, argv, "hvdqs:f:e:l:r:")) != -1) {
970 help(argv[0], targs);
976 map_unmap_verify = true;
980 disable_slot_zap_quirk = true;
981 TEST_REQUIRE(kvm_check_cap(KVM_CAP_DISABLE_QUIRKS2) &
982 KVM_X86_QUIRK_SLOT_ZAP_ALL);
986 targs->nslots = atoi_paranoid(optarg);
987 if (targs->nslots <= 1 && targs->nslots != -1) {
988 pr_info("Slot count cap must be larger than 1 or -1 for no cap\n");
993 targs->tfirst = atoi_non_negative("First test", optarg);
996 targs->tlast = atoi_non_negative("Last test", optarg);
997 if (targs->tlast >= NTESTS) {
998 pr_info("Last test to run has to be non-negative and less than %zu\n",
1004 targs->seconds = atoi_non_negative("Test length", optarg);
1007 targs->runs = atoi_positive("Runs per test", optarg);
1012 if (optind < argc) {
1013 help(argv[0], targs);
1017 if (targs->tfirst > targs->tlast) {
1018 pr_info("First test to run cannot be greater than the last test to run\n");
1022 max_mem_slots = kvm_check_cap(KVM_CAP_NR_MEMSLOTS);
1023 if (max_mem_slots <= 1) {
1024 pr_info("KVM_CAP_NR_MEMSLOTS should be greater than 1\n");
1028 /* Memory slot 0 is reserved */
1029 if (targs->nslots == -1)
1030 targs->nslots = max_mem_slots - 1;
1032 targs->nslots = min_t(int, targs->nslots, max_mem_slots) - 1;
1034 pr_info_v("Allowed Number of memory slots: %"PRIu32"\n",
1040 struct test_result {
1041 struct timespec slot_runtime, guest_runtime, iter_runtime;
1042 int64_t slottimens, runtimens;
1046 static bool test_loop(const struct test_data *data,
1047 const struct test_args *targs,
1048 struct test_result *rbestslottime,
1049 struct test_result *rbestruntime)
1052 struct test_result result = {};
1054 if (!test_execute(targs->nslots, &maxslots, targs->seconds, data,
1056 &result.slot_runtime, &result.guest_runtime)) {
1058 pr_info("Memslot count too high for this test, decrease the cap (max is %"PRIu64")\n",
1061 pr_info("Memslot count may be too high for this test, try adjusting the cap\n");
1066 pr_info("Test took %ld.%.9lds for slot setup + %ld.%.9lds all iterations\n",
1067 result.slot_runtime.tv_sec, result.slot_runtime.tv_nsec,
1068 result.guest_runtime.tv_sec, result.guest_runtime.tv_nsec);
1069 if (!result.nloops) {
1070 pr_info("No full loops done - too short test time or system too loaded?\n");
1074 result.iter_runtime = timespec_div(result.guest_runtime,
1076 pr_info("Done %"PRIu64" iterations, avg %ld.%.9lds each\n",
1078 result.iter_runtime.tv_sec,
1079 result.iter_runtime.tv_nsec);
1080 result.slottimens = timespec_to_ns(result.slot_runtime);
1081 result.runtimens = timespec_to_ns(result.iter_runtime);
1084 * Only rank the slot setup time for tests using the whole test memory
1085 * area so they are comparable
1087 if (!data->mem_size &&
1088 (!rbestslottime->slottimens ||
1089 result.slottimens < rbestslottime->slottimens))
1090 *rbestslottime = result;
1091 if (!rbestruntime->runtimens ||
1092 result.runtimens < rbestruntime->runtimens)
1093 *rbestruntime = result;
1098 int main(int argc, char *argv[])
1100 struct test_args targs = {
1102 .tlast = NTESTS - 1,
1107 struct test_result rbestslottime = {};
1110 if (!check_memory_sizes())
1113 if (!parse_args(argc, argv, &targs))
1116 for (tctr = targs.tfirst; tctr <= targs.tlast; tctr++) {
1117 const struct test_data *data = &tests[tctr];
1118 unsigned int runctr;
1119 struct test_result rbestruntime = {};
1121 if (tctr > targs.tfirst)
1124 pr_info("Testing %s performance with %i runs, %d seconds each\n",
1125 data->name, targs.runs, targs.seconds);
1127 for (runctr = 0; runctr < targs.runs; runctr++)
1128 if (!test_loop(data, &targs,
1129 &rbestslottime, &rbestruntime))
1132 if (rbestruntime.runtimens)
1133 pr_info("Best runtime result was %ld.%.9lds per iteration (with %"PRIu64" iterations)\n",
1134 rbestruntime.iter_runtime.tv_sec,
1135 rbestruntime.iter_runtime.tv_nsec,
1136 rbestruntime.nloops);
1139 if (rbestslottime.slottimens)
1140 pr_info("Best slot setup time for the whole test area was %ld.%.9lds\n",
1141 rbestslottime.slot_runtime.tv_sec,
1142 rbestslottime.slot_runtime.tv_nsec);