tools/testing/selftests/kvm/x86_64/nx_huge_pages_test.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * tools/testing/selftests/kvm/nx_huge_page_test.c
   4  *
   5  * Usage: to be run via nx_huge_page_test.sh, which does the necessary
   6  * environment setup and teardown
   7  *
   8  * Copyright (C) 2022, Google LLC.
   9  */
  10
  11 #define _GNU_SOURCE
  12
  13 #include <fcntl.h>
  14 #include <stdint.h>
  15 #include <time.h>
  16
  17 #include <test_util.h>
  18 #include "kvm_util.h"
  19 #include "processor.h"
  20
  21 #define HPAGE_SLOT              10
  22 #define HPAGE_GPA               (4UL << 30) /* 4G prevents collision w/ slot 0 */
  23 #define HPAGE_GVA               HPAGE_GPA /* GVA is arbitrary, so use GPA. */
  24 #define PAGES_PER_2MB_HUGE_PAGE 512
  25 #define HPAGE_SLOT_NPAGES       (3 * PAGES_PER_2MB_HUGE_PAGE)
  26
  27 /*
  28  * Passed by nx_huge_pages_test.sh to provide an easy warning if this test is
  29  * being run without it.
  30  */
  31 #define MAGIC_TOKEN 887563923
  32
  33 /*
  34  * x86 opcode for the return instruction. Used to call into, and then
  35  * immediately return from, memory backed with hugepages.
  36  */
  37 #define RETURN_OPCODE 0xC3
  38
  39 /* Call the specified memory address. */
  40 static void guest_do_CALL(uint64_t target)
  41 {
  42         ((void (*)(void)) target)();
  43 }
  44
  45 /*
  46  * Exit the VM after each memory access so that the userspace component of the
  47  * test can make assertions about the pages backing the VM.
  48  *
  49  * See the below for an explanation of how each access should affect the
  50  * backing mappings.
  51  */
  52 void guest_code(void)
  53 {
  54         uint64_t hpage_1 = HPAGE_GVA;
  55         uint64_t hpage_2 = hpage_1 + (PAGE_SIZE * 512);
  56         uint64_t hpage_3 = hpage_2 + (PAGE_SIZE * 512);
  57
  58         READ_ONCE(*(uint64_t *)hpage_1);
  59         GUEST_SYNC(1);
  60
  61         READ_ONCE(*(uint64_t *)hpage_2);
  62         GUEST_SYNC(2);
  63
  64         guest_do_CALL(hpage_1);
  65         GUEST_SYNC(3);
  66
  67         guest_do_CALL(hpage_3);
  68         GUEST_SYNC(4);
  69
  70         READ_ONCE(*(uint64_t *)hpage_1);
  71         GUEST_SYNC(5);
  72
  73         READ_ONCE(*(uint64_t *)hpage_3);
  74         GUEST_SYNC(6);
  75 }
  76
  77 static void check_2m_page_count(struct kvm_vm *vm, int expected_pages_2m)
  78 {
  79         int actual_pages_2m;
  80
  81         actual_pages_2m = vm_get_stat(vm, "pages_2m");
  82
  83         TEST_ASSERT(actual_pages_2m == expected_pages_2m,
  84                     "Unexpected 2m page count. Expected %d, got %d",
  85                     expected_pages_2m, actual_pages_2m);
  86 }
  87
  88 static void check_split_count(struct kvm_vm *vm, int expected_splits)
  89 {
  90         int actual_splits;
  91
  92         actual_splits = vm_get_stat(vm, "nx_lpage_splits");
  93
  94         TEST_ASSERT(actual_splits == expected_splits,
  95                     "Unexpected NX huge page split count. Expected %d, got %d",
  96                     expected_splits, actual_splits);
  97 }
  98
  99 static void wait_for_reclaim(int reclaim_period_ms)
 100 {
 101         long reclaim_wait_ms;
 102         struct timespec ts;
 103
 104         reclaim_wait_ms = reclaim_period_ms * 5;
 105         ts.tv_sec = reclaim_wait_ms / 1000;
 106         ts.tv_nsec = (reclaim_wait_ms - (ts.tv_sec * 1000)) * 1000000;
 107         nanosleep(&ts, NULL);
 108 }
 109
 110 void run_test(int reclaim_period_ms, bool disable_nx_huge_pages,
 111               bool reboot_permissions)
 112 {
 113         struct kvm_vcpu *vcpu;
 114         struct kvm_vm *vm;
 115         uint64_t nr_bytes;
 116         void *hva;
 117         int r;
 118
 119         vm = vm_create(1);
 120
 121         if (disable_nx_huge_pages) {
 122                 r = __vm_disable_nx_huge_pages(vm);
 123                 if (reboot_permissions) {
 124                         TEST_ASSERT(!r, "Disabling NX huge pages should succeed if process has reboot permissions");
 125                 } else {
 126                         TEST_ASSERT(r == -1 && errno == EPERM,
 127                                     "This process should not have permission to disable NX huge pages");
 128                         return;
 129                 }
 130         }
 131
 132         vcpu = vm_vcpu_add(vm, 0, guest_code);
 133
 134         vm_userspace_mem_region_add(vm, VM_MEM_SRC_ANONYMOUS_HUGETLB,
 135                                     HPAGE_GPA, HPAGE_SLOT,
 136                                     HPAGE_SLOT_NPAGES, 0);
 137
 138         nr_bytes = HPAGE_SLOT_NPAGES * vm->page_size;
 139
 140         /*
 141          * Ensure that KVM can map HPAGE_SLOT with huge pages by mapping the
 142          * region into the guest with 2MiB pages whenever TDP is disabled (i.e.
 143          * whenever KVM is shadowing the guest page tables).
 144          *
 145          * When TDP is enabled, KVM should be able to map HPAGE_SLOT with huge
 146          * pages irrespective of the guest page size, so map with 4KiB pages
 147          * to test that that is the case.
 148          */
 149         if (kvm_is_tdp_enabled())
 150                 virt_map_level(vm, HPAGE_GVA, HPAGE_GPA, nr_bytes, PG_LEVEL_4K);
 151         else
 152                 virt_map_level(vm, HPAGE_GVA, HPAGE_GPA, nr_bytes, PG_LEVEL_2M);
 153
 154         hva = addr_gpa2hva(vm, HPAGE_GPA);
 155         memset(hva, RETURN_OPCODE, nr_bytes);
 156
 157         check_2m_page_count(vm, 0);
 158         check_split_count(vm, 0);
 159
 160         /*
 161          * The guest code will first read from the first hugepage, resulting
 162          * in a huge page mapping being created.
 163          */
 164         vcpu_run(vcpu);
 165         check_2m_page_count(vm, 1);
 166         check_split_count(vm, 0);
 167
 168         /*
 169          * Then the guest code will read from the second hugepage, resulting
 170          * in another huge page mapping being created.
 171          */
 172         vcpu_run(vcpu);
 173         check_2m_page_count(vm, 2);
 174         check_split_count(vm, 0);
 175
 176         /*
 177          * Next, the guest will execute from the first huge page, causing it
 178          * to be remapped at 4k.
 179          *
 180          * If NX huge pages are disabled, this should have no effect.
 181          */
 182         vcpu_run(vcpu);
 183         check_2m_page_count(vm, disable_nx_huge_pages ? 2 : 1);
 184         check_split_count(vm, disable_nx_huge_pages ? 0 : 1);
 185
 186         /*
 187          * Executing from the third huge page (previously unaccessed) will
 188          * cause part to be mapped at 4k.
 189          *
 190          * If NX huge pages are disabled, it should be mapped at 2M.
 191          */
 192         vcpu_run(vcpu);
 193         check_2m_page_count(vm, disable_nx_huge_pages ? 3 : 1);
 194         check_split_count(vm, disable_nx_huge_pages ? 0 : 2);
 195
 196         /* Reading from the first huge page again should have no effect. */
 197         vcpu_run(vcpu);
 198         check_2m_page_count(vm, disable_nx_huge_pages ? 3 : 1);
 199         check_split_count(vm, disable_nx_huge_pages ? 0 : 2);
 200
 201         /* Give recovery thread time to run. */
 202         wait_for_reclaim(reclaim_period_ms);
 203
 204         /*
 205          * Now that the reclaimer has run, all the split pages should be gone.
 206          *
 207          * If NX huge pages are disabled, the relaimer will not run, so
 208          * nothing should change from here on.
 209          */
 210         check_2m_page_count(vm, disable_nx_huge_pages ? 3 : 1);
 211         check_split_count(vm, 0);
 212
 213         /*
 214          * The 4k mapping on hpage 3 should have been removed, so check that
 215          * reading from it causes a huge page mapping to be installed.
 216          */
 217         vcpu_run(vcpu);
 218         check_2m_page_count(vm, disable_nx_huge_pages ? 3 : 2);
 219         check_split_count(vm, 0);
 220
 221         kvm_vm_free(vm);
 222 }
 223
 224 static void help(char *name)
 225 {
 226         puts("");
 227         printf("usage: %s [-h] [-p period_ms] [-t token]\n", name);
 228         puts("");
 229         printf(" -p: The NX reclaim period in milliseconds.\n");
 230         printf(" -t: The magic token to indicate environment setup is done.\n");
 231         printf(" -r: The test has reboot permissions and can disable NX huge pages.\n");
 232         puts("");
 233         exit(0);
 234 }
 235
 236 int main(int argc, char **argv)
 237 {
 238         int reclaim_period_ms = 0, token = 0, opt;
 239         bool reboot_permissions = false;
 240
 241         while ((opt = getopt(argc, argv, "hp:t:r")) != -1) {
 242                 switch (opt) {
 243                 case 'p':
 244                         reclaim_period_ms = atoi_positive("Reclaim period", optarg);
 245                         break;
 246                 case 't':
 247                         token = atoi_paranoid(optarg);
 248                         break;
 249                 case 'r':
 250                         reboot_permissions = true;
 251                         break;
 252                 case 'h':
 253                 default:
 254                         help(argv[0]);
 255                         break;
 256                 }
 257         }
 258
 259         TEST_REQUIRE(kvm_has_cap(KVM_CAP_VM_DISABLE_NX_HUGE_PAGES));
 260
 261         __TEST_REQUIRE(token == MAGIC_TOKEN,
 262                        "This test must be run with the magic token %d.\n"
 263                        "This is done by nx_huge_pages_test.sh, which\n"
 264                        "also handles environment setup for the test.");
 265
 266         run_test(reclaim_period_ms, false, reboot_permissions);
 267         run_test(reclaim_period_ms, true, reboot_permissions);
 268
 269         return 0;
 270 }
 271