1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright Intel Corporation, 2023 4 * 5 * Author: Chao Peng <chao.p.peng@linux.intel.com> 6 */ 7 #include <stdlib.h> 8 #include <string.h> 9 #include <unistd.h> 10 #include <errno.h> 11 #include <stdio.h> 12 #include <fcntl.h> 13 14 #include <linux/bitmap.h> 15 #include <linux/falloc.h> 16 #include <linux/sizes.h> 17 #include <sys/mman.h> 18 #include <sys/types.h> 19 #include <sys/stat.h> 20 21 #include "kvm_util.h" 22 #include "numaif.h" 23 #include "test_util.h" 24 #include "ucall_common.h" 25 26 static size_t page_size; 27 28 static void test_file_read_write(int fd, size_t total_size) 29 { 30 char buf[64]; 31 32 TEST_ASSERT(read(fd, buf, sizeof(buf)) < 0, 33 "read on a guest_mem fd should fail"); 34 TEST_ASSERT(write(fd, buf, sizeof(buf)) < 0, 35 "write on a guest_mem fd should fail"); 36 TEST_ASSERT(pread(fd, buf, sizeof(buf), 0) < 0, 37 "pread on a guest_mem fd should fail"); 38 TEST_ASSERT(pwrite(fd, buf, sizeof(buf), 0) < 0, 39 "pwrite on a guest_mem fd should fail"); 40 } 41 42 static void test_mmap_cow(int fd, size_t size) 43 { 44 void *mem; 45 46 mem = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0); 47 TEST_ASSERT(mem == MAP_FAILED, "Copy-on-write not allowed by guest_memfd."); 48 } 49 50 static void test_mmap_supported(int fd, size_t total_size) 51 { 52 const char val = 0xaa; 53 char *mem; 54 size_t i; 55 int ret; 56 57 mem = kvm_mmap(total_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd); 58 59 memset(mem, val, total_size); 60 for (i = 0; i < total_size; i++) 61 TEST_ASSERT_EQ(READ_ONCE(mem[i]), val); 62 63 ret = fallocate(fd, FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE, 0, 64 page_size); 65 TEST_ASSERT(!ret, "fallocate the first page should succeed."); 66 67 for (i = 0; i < page_size; i++) 68 TEST_ASSERT_EQ(READ_ONCE(mem[i]), 0x00); 69 for (; i < total_size; i++) 70 TEST_ASSERT_EQ(READ_ONCE(mem[i]), val); 71 72 memset(mem, val, page_size); 73 for (i = 0; i < total_size; i++) 74 TEST_ASSERT_EQ(READ_ONCE(mem[i]), val); 75 76 kvm_munmap(mem, total_size); 77 } 78 79 static void test_mbind(int fd, size_t total_size) 80 { 81 const unsigned long nodemask_0 = 1; /* nid: 0 */ 82 unsigned long nodemask = 0; 83 unsigned long maxnode = BITS_PER_TYPE(nodemask); 84 int policy; 85 char *mem; 86 int ret; 87 88 if (!is_multi_numa_node_system()) 89 return; 90 91 mem = kvm_mmap(total_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd); 92 93 /* Test MPOL_INTERLEAVE policy */ 94 kvm_mbind(mem, page_size * 2, MPOL_INTERLEAVE, &nodemask_0, maxnode, 0); 95 kvm_get_mempolicy(&policy, &nodemask, maxnode, mem, MPOL_F_ADDR); 96 TEST_ASSERT(policy == MPOL_INTERLEAVE && nodemask == nodemask_0, 97 "Wanted MPOL_INTERLEAVE (%u) and nodemask 0x%lx, got %u and 0x%lx", 98 MPOL_INTERLEAVE, nodemask_0, policy, nodemask); 99 100 /* Test basic MPOL_BIND policy */ 101 kvm_mbind(mem + page_size * 2, page_size * 2, MPOL_BIND, &nodemask_0, maxnode, 0); 102 kvm_get_mempolicy(&policy, &nodemask, maxnode, mem + page_size * 2, MPOL_F_ADDR); 103 TEST_ASSERT(policy == MPOL_BIND && nodemask == nodemask_0, 104 "Wanted MPOL_BIND (%u) and nodemask 0x%lx, got %u and 0x%lx", 105 MPOL_BIND, nodemask_0, policy, nodemask); 106 107 /* Test MPOL_DEFAULT policy */ 108 kvm_mbind(mem, total_size, MPOL_DEFAULT, NULL, 0, 0); 109 kvm_get_mempolicy(&policy, &nodemask, maxnode, mem, MPOL_F_ADDR); 110 TEST_ASSERT(policy == MPOL_DEFAULT && !nodemask, 111 "Wanted MPOL_DEFAULT (%u) and nodemask 0x0, got %u and 0x%lx", 112 MPOL_DEFAULT, policy, nodemask); 113 114 /* Test with invalid policy */ 115 ret = mbind(mem, page_size, 999, &nodemask_0, maxnode, 0); 116 TEST_ASSERT(ret == -1 && errno == EINVAL, 117 "mbind with invalid policy should fail with EINVAL"); 118 119 kvm_munmap(mem, total_size); 120 } 121 122 static void test_numa_allocation(int fd, size_t total_size) 123 { 124 unsigned long node0_mask = 1; /* Node 0 */ 125 unsigned long node1_mask = 2; /* Node 1 */ 126 unsigned long maxnode = 8; 127 void *pages[4]; 128 int status[4]; 129 char *mem; 130 int i; 131 132 if (!is_multi_numa_node_system()) 133 return; 134 135 mem = kvm_mmap(total_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd); 136 137 for (i = 0; i < 4; i++) 138 pages[i] = (char *)mem + page_size * i; 139 140 /* Set NUMA policy after allocation */ 141 memset(mem, 0xaa, page_size); 142 kvm_mbind(pages[0], page_size, MPOL_BIND, &node0_mask, maxnode, 0); 143 kvm_fallocate(fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE, 0, page_size); 144 145 /* Set NUMA policy before allocation */ 146 kvm_mbind(pages[0], page_size * 2, MPOL_BIND, &node1_mask, maxnode, 0); 147 kvm_mbind(pages[2], page_size * 2, MPOL_BIND, &node0_mask, maxnode, 0); 148 memset(mem, 0xaa, total_size); 149 150 /* Validate if pages are allocated on specified NUMA nodes */ 151 kvm_move_pages(0, 4, pages, NULL, status, 0); 152 TEST_ASSERT(status[0] == 1, "Expected page 0 on node 1, got it on node %d", status[0]); 153 TEST_ASSERT(status[1] == 1, "Expected page 1 on node 1, got it on node %d", status[1]); 154 TEST_ASSERT(status[2] == 0, "Expected page 2 on node 0, got it on node %d", status[2]); 155 TEST_ASSERT(status[3] == 0, "Expected page 3 on node 0, got it on node %d", status[3]); 156 157 /* Punch hole for all pages */ 158 kvm_fallocate(fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE, 0, total_size); 159 160 /* Change NUMA policy nodes and reallocate */ 161 kvm_mbind(pages[0], page_size * 2, MPOL_BIND, &node0_mask, maxnode, 0); 162 kvm_mbind(pages[2], page_size * 2, MPOL_BIND, &node1_mask, maxnode, 0); 163 memset(mem, 0xaa, total_size); 164 165 kvm_move_pages(0, 4, pages, NULL, status, 0); 166 TEST_ASSERT(status[0] == 0, "Expected page 0 on node 0, got it on node %d", status[0]); 167 TEST_ASSERT(status[1] == 0, "Expected page 1 on node 0, got it on node %d", status[1]); 168 TEST_ASSERT(status[2] == 1, "Expected page 2 on node 1, got it on node %d", status[2]); 169 TEST_ASSERT(status[3] == 1, "Expected page 3 on node 1, got it on node %d", status[3]); 170 171 kvm_munmap(mem, total_size); 172 } 173 174 static void test_collapse(int fd, uint64_t flags) 175 { 176 const size_t pmd_size = get_trans_hugepagesz(); 177 void *reserved_addr; 178 void *aligned_addr; 179 char *mem; 180 off_t i; 181 182 /* 183 * To even reach the point where the guest_memfd folios will 184 * get collapsed, both the userspace address and the offset 185 * within the guest_memfd have to be aligned to pmd_size. 186 * 187 * To achieve that alignment, reserve virtual address space 188 * with regular mmap, then use MAP_FIXED to allocate memory 189 * from a pmd_size-aligned offset (0) at a known, available 190 * virtual address. 191 */ 192 reserved_addr = kvm_mmap(pmd_size * 2, PROT_NONE, 193 MAP_PRIVATE | MAP_ANONYMOUS, -1); 194 aligned_addr = align_ptr_up(reserved_addr, pmd_size); 195 196 mem = mmap(aligned_addr, pmd_size, PROT_READ | PROT_WRITE, 197 MAP_FIXED | MAP_SHARED, fd, 0); 198 TEST_ASSERT(IS_ALIGNED((u64)mem, pmd_size), 199 "Userspace address must be aligned to PMD size."); 200 201 /* 202 * Use reads to populate page table to avoid setting dirty 203 * flag on page. 204 */ 205 for (i = 0; i < pmd_size; i += getpagesize()) 206 READ_ONCE(mem[i]); 207 208 /* 209 * Advising the use of huge pages in guest_memfd should be 210 * fine... 211 */ 212 kvm_madvise(mem, pmd_size, MADV_HUGEPAGE); 213 214 /* 215 * ... but collapsing folios must not be supported to avoid 216 * mapping beyond shared ranges into host userspace page 217 * tables. 218 */ 219 TEST_ASSERT_EQ(madvise(mem, pmd_size, MADV_COLLAPSE), -1); 220 TEST_ASSERT_EQ(errno, EINVAL); 221 222 /* 223 * Removing from host page tables and re-faulting should be 224 * fine; should not end up faulting in a collapsed/huge folio. 225 */ 226 kvm_madvise(mem, pmd_size, MADV_DONTNEED); 227 READ_ONCE(mem[0]); 228 229 kvm_munmap(reserved_addr, pmd_size * 2); 230 } 231 232 static void test_fault_sigbus(int fd, size_t accessible_size, size_t map_size) 233 { 234 const char val = 0xaa; 235 char *mem; 236 size_t i; 237 238 mem = kvm_mmap(map_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd); 239 240 TEST_EXPECT_SIGBUS(memset(mem, val, map_size)); 241 TEST_EXPECT_SIGBUS((void)READ_ONCE(mem[accessible_size])); 242 243 for (i = 0; i < accessible_size; i++) 244 TEST_ASSERT_EQ(READ_ONCE(mem[i]), val); 245 246 kvm_munmap(mem, map_size); 247 } 248 249 static void test_fault_overflow(int fd, size_t total_size) 250 { 251 test_fault_sigbus(fd, total_size, total_size * 4); 252 } 253 254 static void test_fault_private(int fd, size_t total_size) 255 { 256 test_fault_sigbus(fd, 0, total_size); 257 } 258 259 static void test_mmap_not_supported(int fd, size_t total_size) 260 { 261 char *mem; 262 263 mem = mmap(NULL, page_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0); 264 TEST_ASSERT_EQ(mem, MAP_FAILED); 265 266 mem = mmap(NULL, total_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0); 267 TEST_ASSERT_EQ(mem, MAP_FAILED); 268 } 269 270 static void test_file_size(int fd, size_t total_size) 271 { 272 struct stat sb; 273 int ret; 274 275 ret = fstat(fd, &sb); 276 TEST_ASSERT(!ret, "fstat should succeed"); 277 TEST_ASSERT_EQ(sb.st_size, total_size); 278 TEST_ASSERT_EQ(sb.st_blksize, page_size); 279 } 280 281 static void test_fallocate(int fd, size_t total_size) 282 { 283 int ret; 284 285 ret = fallocate(fd, FALLOC_FL_KEEP_SIZE, 0, total_size); 286 TEST_ASSERT(!ret, "fallocate with aligned offset and size should succeed"); 287 288 ret = fallocate(fd, FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE, 289 page_size - 1, page_size); 290 TEST_ASSERT(ret, "fallocate with unaligned offset should fail"); 291 292 ret = fallocate(fd, FALLOC_FL_KEEP_SIZE, total_size, page_size); 293 TEST_ASSERT(ret, "fallocate beginning at total_size should fail"); 294 295 ret = fallocate(fd, FALLOC_FL_KEEP_SIZE, total_size + page_size, page_size); 296 TEST_ASSERT(ret, "fallocate beginning after total_size should fail"); 297 298 ret = fallocate(fd, FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE, 299 total_size, page_size); 300 TEST_ASSERT(!ret, "fallocate(PUNCH_HOLE) at total_size should succeed"); 301 302 ret = fallocate(fd, FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE, 303 total_size + page_size, page_size); 304 TEST_ASSERT(!ret, "fallocate(PUNCH_HOLE) after total_size should succeed"); 305 306 ret = fallocate(fd, FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE, 307 page_size, page_size - 1); 308 TEST_ASSERT(ret, "fallocate with unaligned size should fail"); 309 310 ret = fallocate(fd, FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE, 311 page_size, page_size); 312 TEST_ASSERT(!ret, "fallocate(PUNCH_HOLE) with aligned offset and size should succeed"); 313 314 ret = fallocate(fd, FALLOC_FL_KEEP_SIZE, page_size, page_size); 315 TEST_ASSERT(!ret, "fallocate to restore punched hole should succeed"); 316 } 317 318 static void test_invalid_punch_hole(int fd, size_t total_size) 319 { 320 struct { 321 off_t offset; 322 off_t len; 323 } testcases[] = { 324 {0, 1}, 325 {0, page_size - 1}, 326 {0, page_size + 1}, 327 328 {1, 1}, 329 {1, page_size - 1}, 330 {1, page_size}, 331 {1, page_size + 1}, 332 333 {page_size, 1}, 334 {page_size, page_size - 1}, 335 {page_size, page_size + 1}, 336 }; 337 int ret, i; 338 339 for (i = 0; i < ARRAY_SIZE(testcases); i++) { 340 ret = fallocate(fd, FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE, 341 testcases[i].offset, testcases[i].len); 342 TEST_ASSERT(ret == -1 && errno == EINVAL, 343 "PUNCH_HOLE with !PAGE_SIZE offset (%lx) and/or length (%lx) should fail", 344 testcases[i].offset, testcases[i].len); 345 } 346 } 347 348 static void test_create_guest_memfd_invalid_sizes(struct kvm_vm *vm, 349 uint64_t guest_memfd_flags) 350 { 351 size_t size; 352 int fd; 353 354 for (size = 1; size < page_size; size++) { 355 fd = __vm_create_guest_memfd(vm, size, guest_memfd_flags); 356 TEST_ASSERT(fd < 0 && errno == EINVAL, 357 "guest_memfd() with non-page-aligned page size '0x%lx' should fail with EINVAL", 358 size); 359 } 360 } 361 362 static void test_create_guest_memfd_multiple(struct kvm_vm *vm) 363 { 364 int fd1, fd2, ret; 365 struct stat st1, st2; 366 367 fd1 = __vm_create_guest_memfd(vm, page_size, 0); 368 TEST_ASSERT(fd1 != -1, "memfd creation should succeed"); 369 370 ret = fstat(fd1, &st1); 371 TEST_ASSERT(ret != -1, "memfd fstat should succeed"); 372 TEST_ASSERT(st1.st_size == page_size, "memfd st_size should match requested size"); 373 374 fd2 = __vm_create_guest_memfd(vm, page_size * 2, 0); 375 TEST_ASSERT(fd2 != -1, "memfd creation should succeed"); 376 377 ret = fstat(fd2, &st2); 378 TEST_ASSERT(ret != -1, "memfd fstat should succeed"); 379 TEST_ASSERT(st2.st_size == page_size * 2, "second memfd st_size should match requested size"); 380 381 ret = fstat(fd1, &st1); 382 TEST_ASSERT(ret != -1, "memfd fstat should succeed"); 383 TEST_ASSERT(st1.st_size == page_size, "first memfd st_size should still match requested size"); 384 TEST_ASSERT(st1.st_ino != st2.st_ino, "different memfd should have different inode numbers"); 385 386 close(fd2); 387 close(fd1); 388 } 389 390 static void test_guest_memfd_flags(struct kvm_vm *vm) 391 { 392 uint64_t valid_flags = vm_check_cap(vm, KVM_CAP_GUEST_MEMFD_FLAGS); 393 uint64_t flag; 394 int fd; 395 396 for (flag = BIT(0); flag; flag <<= 1) { 397 fd = __vm_create_guest_memfd(vm, page_size, flag); 398 if (flag & valid_flags) { 399 TEST_ASSERT(fd >= 0, 400 "guest_memfd() with flag '0x%lx' should succeed", 401 flag); 402 close(fd); 403 } else { 404 TEST_ASSERT(fd < 0 && errno == EINVAL, 405 "guest_memfd() with flag '0x%lx' should fail with EINVAL", 406 flag); 407 } 408 } 409 } 410 411 #define __gmem_test(__test, __vm, __flags, __gmem_size) \ 412 do { \ 413 int fd = vm_create_guest_memfd(__vm, __gmem_size, __flags); \ 414 \ 415 test_##__test(fd, __gmem_size); \ 416 close(fd); \ 417 } while (0) 418 419 #define gmem_test(__test, __vm, __flags) \ 420 __gmem_test(__test, __vm, __flags, page_size * 4) 421 422 static void __test_guest_memfd(struct kvm_vm *vm, uint64_t flags) 423 { 424 test_create_guest_memfd_multiple(vm); 425 test_create_guest_memfd_invalid_sizes(vm, flags); 426 427 gmem_test(file_read_write, vm, flags); 428 429 if (flags & GUEST_MEMFD_FLAG_MMAP) { 430 if (flags & GUEST_MEMFD_FLAG_INIT_SHARED) { 431 size_t pmd_size = get_trans_hugepagesz(); 432 433 gmem_test(mmap_supported, vm, flags); 434 gmem_test(fault_overflow, vm, flags); 435 gmem_test(numa_allocation, vm, flags); 436 __gmem_test(collapse, vm, flags, pmd_size); 437 } else { 438 gmem_test(fault_private, vm, flags); 439 } 440 441 gmem_test(mmap_cow, vm, flags); 442 gmem_test(mbind, vm, flags); 443 } else { 444 gmem_test(mmap_not_supported, vm, flags); 445 } 446 447 gmem_test(file_size, vm, flags); 448 gmem_test(fallocate, vm, flags); 449 gmem_test(invalid_punch_hole, vm, flags); 450 } 451 452 static void test_guest_memfd(unsigned long vm_type) 453 { 454 struct kvm_vm *vm = vm_create_barebones_type(vm_type); 455 uint64_t flags; 456 457 test_guest_memfd_flags(vm); 458 459 __test_guest_memfd(vm, 0); 460 461 flags = vm_check_cap(vm, KVM_CAP_GUEST_MEMFD_FLAGS); 462 if (flags & GUEST_MEMFD_FLAG_MMAP) 463 __test_guest_memfd(vm, GUEST_MEMFD_FLAG_MMAP); 464 465 /* MMAP should always be supported if INIT_SHARED is supported. */ 466 if (flags & GUEST_MEMFD_FLAG_INIT_SHARED) 467 __test_guest_memfd(vm, GUEST_MEMFD_FLAG_MMAP | 468 GUEST_MEMFD_FLAG_INIT_SHARED); 469 470 kvm_vm_free(vm); 471 } 472 473 static void guest_code(uint8_t *mem, uint64_t size) 474 { 475 size_t i; 476 477 for (i = 0; i < size; i++) 478 __GUEST_ASSERT(mem[i] == 0xaa, 479 "Guest expected 0xaa at offset %lu, got 0x%x", i, mem[i]); 480 481 memset(mem, 0xff, size); 482 GUEST_DONE(); 483 } 484 485 static void test_guest_memfd_guest(void) 486 { 487 /* 488 * Skip the first 4gb and slot0. slot0 maps <1gb and is used to back 489 * the guest's code, stack, and page tables, and low memory contains 490 * the PCI hole and other MMIO regions that need to be avoided. 491 */ 492 const uint64_t gpa = SZ_4G; 493 const int slot = 1; 494 495 struct kvm_vcpu *vcpu; 496 struct kvm_vm *vm; 497 uint8_t *mem; 498 size_t size; 499 int fd, i; 500 501 if (!kvm_check_cap(KVM_CAP_GUEST_MEMFD_FLAGS)) 502 return; 503 504 vm = __vm_create_shape_with_one_vcpu(VM_SHAPE_DEFAULT, &vcpu, 1, guest_code); 505 506 TEST_ASSERT(vm_check_cap(vm, KVM_CAP_GUEST_MEMFD_FLAGS) & GUEST_MEMFD_FLAG_MMAP, 507 "Default VM type should support MMAP, supported flags = 0x%x", 508 vm_check_cap(vm, KVM_CAP_GUEST_MEMFD_FLAGS)); 509 TEST_ASSERT(vm_check_cap(vm, KVM_CAP_GUEST_MEMFD_FLAGS) & GUEST_MEMFD_FLAG_INIT_SHARED, 510 "Default VM type should support INIT_SHARED, supported flags = 0x%x", 511 vm_check_cap(vm, KVM_CAP_GUEST_MEMFD_FLAGS)); 512 513 size = vm->page_size; 514 fd = vm_create_guest_memfd(vm, size, GUEST_MEMFD_FLAG_MMAP | 515 GUEST_MEMFD_FLAG_INIT_SHARED); 516 vm_set_user_memory_region2(vm, slot, KVM_MEM_GUEST_MEMFD, gpa, size, NULL, fd, 0); 517 518 mem = kvm_mmap(size, PROT_READ | PROT_WRITE, MAP_SHARED, fd); 519 memset(mem, 0xaa, size); 520 kvm_munmap(mem, size); 521 522 virt_pg_map(vm, gpa, gpa); 523 vcpu_args_set(vcpu, 2, gpa, size); 524 vcpu_run(vcpu); 525 526 TEST_ASSERT_EQ(get_ucall(vcpu, NULL), UCALL_DONE); 527 528 mem = kvm_mmap(size, PROT_READ | PROT_WRITE, MAP_SHARED, fd); 529 for (i = 0; i < size; i++) 530 TEST_ASSERT_EQ(mem[i], 0xff); 531 532 close(fd); 533 kvm_vm_free(vm); 534 } 535 536 int main(int argc, char *argv[]) 537 { 538 unsigned long vm_types, vm_type; 539 540 TEST_REQUIRE(kvm_has_cap(KVM_CAP_GUEST_MEMFD)); 541 542 page_size = getpagesize(); 543 544 /* 545 * Not all architectures support KVM_CAP_VM_TYPES. However, those that 546 * support guest_memfd have that support for the default VM type. 547 */ 548 vm_types = kvm_check_cap(KVM_CAP_VM_TYPES); 549 if (!vm_types) 550 vm_types = BIT(VM_TYPE_DEFAULT); 551 552 for_each_set_bit(vm_type, &vm_types, BITS_PER_TYPE(vm_types)) 553 test_guest_memfd(vm_type); 554 555 test_guest_memfd_guest(); 556 } 557