Loading...
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 | // SPDX-License-Identifier: GPL-2.0-only /* * Copyright 2017, Gustavo Romero, IBM Corp. * * Check if thread endianness is flipped inadvertently to BE on trap * caught in TM whilst MSR.FP and MSR.VEC are zero (i.e. just after * load_fp and load_vec overflowed). * * The issue can be checked on LE machines simply by zeroing load_fp * and load_vec and then causing a trap in TM. Since the endianness * changes to BE on return from the signal handler, 'nop' is * thread as an illegal instruction in following sequence: * tbegin. * beq 1f * trap * tend. * 1: nop * * However, although the issue is also present on BE machines, it's a * bit trickier to check it on BE machines because MSR.LE bit is set * to zero which determines a BE endianness that is the native * endianness on BE machines, so nothing notably critical happens, * i.e. no illegal instruction is observed immediately after returning * from the signal handler (as it happens on LE machines). Thus to test * it on BE machines LE endianness is forced after a first trap and then * the endianness is verified on subsequent traps to determine if the * endianness "flipped back" to the native endianness (BE). */ #define _GNU_SOURCE #include <error.h> #include <stdio.h> #include <stdlib.h> #include <unistd.h> #include <htmintrin.h> #include <inttypes.h> #include <pthread.h> #include <sched.h> #include <signal.h> #include <stdbool.h> #include "tm.h" #include "utils.h" #define pr_error(error_code, format, ...) \ error_at_line(1, error_code, __FILE__, __LINE__, format, ##__VA_ARGS__) #define MSR_LE 1UL #define LE 1UL pthread_t t0_ping; pthread_t t1_pong; int exit_from_pong; int trap_event; int le; bool success; void trap_signal_handler(int signo, siginfo_t *si, void *uc) { ucontext_t *ucp = uc; uint64_t thread_endianness; /* Get thread endianness: extract bit LE from MSR */ thread_endianness = MSR_LE & ucp->uc_mcontext.gp_regs[PT_MSR]; /* * Little-Endian Machine */ if (le) { /* First trap event */ if (trap_event == 0) { /* Do nothing. Since it is returning from this trap * event that endianness is flipped by the bug, so just * let the process return from the signal handler and * check on the second trap event if endianness is * flipped or not. */ } /* Second trap event */ else if (trap_event == 1) { /* * Since trap was caught in TM on first trap event, if * endianness was still LE (not flipped inadvertently) * after returning from the signal handler instruction * (1) is executed (basically a 'nop'), as it's located * at address of tbegin. +4 (rollback addr). As (1) on * LE endianness does in effect nothing, instruction (2) * is then executed again as 'trap', generating a second * trap event (note that in that case 'trap' is caught * not in transacional mode). On te other hand, if after * the return from the signal handler the endianness in- * advertently flipped, instruction (1) is tread as a * branch instruction, i.e. b .+8, hence instruction (3) * and (4) are executed (tbegin.; trap;) and we get sim- * ilaly on the trap signal handler, but now in TM mode. * Either way, it's now possible to check the MSR LE bit * once in the trap handler to verify if endianness was * flipped or not after the return from the second trap * event. If endianness is flipped, the bug is present. * Finally, getting a trap in TM mode or not is just * worth noting because it affects the math to determine * the offset added to the NIP on return: the NIP for a * trap caught in TM is the rollback address, i.e. the * next instruction after 'tbegin.', whilst the NIP for * a trap caught in non-transactional mode is the very * same address of the 'trap' instruction that generated * the trap event. */ if (thread_endianness == LE) { /* Go to 'success', i.e. instruction (6) */ ucp->uc_mcontext.gp_regs[PT_NIP] += 16; } else { /* * Thread endianness is BE, so it flipped * inadvertently. Thus we flip back to LE and * set NIP to go to 'failure', instruction (5). */ ucp->uc_mcontext.gp_regs[PT_MSR] |= 1UL; ucp->uc_mcontext.gp_regs[PT_NIP] += 4; } } } /* * Big-Endian Machine */ else { /* First trap event */ if (trap_event == 0) { /* * Force thread endianness to be LE. Instructions (1), * (3), and (4) will be executed, generating a second * trap in TM mode. */ ucp->uc_mcontext.gp_regs[PT_MSR] |= 1UL; } /* Second trap event */ else if (trap_event == 1) { /* * Do nothing. If bug is present on return from this * second trap event endianness will flip back "automat- * ically" to BE, otherwise thread endianness will * continue to be LE, just as it was set above. */ } /* A third trap event */ else { /* * Once here it means that after returning from the sec- * ond trap event instruction (4) (trap) was executed * as LE, generating a third trap event. In that case * endianness is still LE as set on return from the * first trap event, hence no bug. Otherwise, bug * flipped back to BE on return from the second trap * event and instruction (4) was executed as 'tdi' (so * basically a 'nop') and branch to 'failure' in * instruction (5) was taken to indicate failure and we * never get here. */ /* * Flip back to BE and go to instruction (6), i.e. go to * 'success'. */ ucp->uc_mcontext.gp_regs[PT_MSR] &= ~1UL; ucp->uc_mcontext.gp_regs[PT_NIP] += 8; } } trap_event++; } void usr1_signal_handler(int signo, siginfo_t *si, void *not_used) { /* Got a USR1 signal from ping(), so just tell pong() to exit */ exit_from_pong = 1; } void *ping(void *not_used) { uint64_t i; trap_event = 0; /* * Wait an amount of context switches so load_fp and load_vec overflows * and MSR_[FP|VEC|V] is 0. */ for (i = 0; i < 1024*1024*512; i++) ; asm goto( /* * [NA] means "Native Endianness", i.e. it tells how a * instruction is executed on machine's native endianness (in * other words, native endianness matches kernel endianness). * [OP] means "Opposite Endianness", i.e. on a BE machine, it * tells how a instruction is executed as a LE instruction; con- * versely, on a LE machine, it tells how a instruction is * executed as a BE instruction. When [NA] is omitted, it means * that the native interpretation of a given instruction is not * relevant for the test. Likewise when [OP] is omitted. */ " tbegin. ;" /* (0) tbegin. [NA] */ " tdi 0, 0, 0x48;" /* (1) nop [NA]; b (3) [OP] */ " trap ;" /* (2) trap [NA] */ ".long 0x1D05007C;" /* (3) tbegin. [OP] */ ".long 0x0800E07F;" /* (4) trap [OP]; nop [NA] */ " b %l[failure] ;" /* (5) b [NA]; MSR.LE flipped (bug) */ " b %l[success] ;" /* (6) b [NA]; MSR.LE did not flip (ok)*/ : : : : failure, success); failure: success = false; goto exit_from_ping; success: success = true; exit_from_ping: /* Tell pong() to exit before leaving */ pthread_kill(t1_pong, SIGUSR1); return NULL; } void *pong(void *not_used) { while (!exit_from_pong) /* * Induce context switches on ping() thread * until ping() finishes its job and signs * to exit from this loop. */ sched_yield(); return NULL; } int tm_trap_test(void) { uint16_t k = 1; int cpu, rc; pthread_attr_t attr; cpu_set_t cpuset; struct sigaction trap_sa; SKIP_IF(!have_htm()); SKIP_IF(htm_is_synthetic()); trap_sa.sa_flags = SA_SIGINFO; trap_sa.sa_sigaction = trap_signal_handler; sigaction(SIGTRAP, &trap_sa, NULL); struct sigaction usr1_sa; usr1_sa.sa_flags = SA_SIGINFO; usr1_sa.sa_sigaction = usr1_signal_handler; sigaction(SIGUSR1, &usr1_sa, NULL); cpu = pick_online_cpu(); FAIL_IF(cpu < 0); // Set only one CPU in the mask. Both threads will be bound to that CPU. CPU_ZERO(&cpuset); CPU_SET(cpu, &cpuset); /* Init pthread attribute */ rc = pthread_attr_init(&attr); if (rc) pr_error(rc, "pthread_attr_init()"); /* * Bind thread ping() and pong() both to CPU 0 so they ping-pong and * speed up context switches on ping() thread, speeding up the load_fp * and load_vec overflow. */ rc = pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpuset); if (rc) pr_error(rc, "pthread_attr_setaffinity()"); /* Figure out the machine endianness */ le = (int) *(uint8_t *)&k; printf("%s machine detected. Checking if endianness flips %s", le ? "Little-Endian" : "Big-Endian", "inadvertently on trap in TM... "); rc = fflush(0); if (rc) pr_error(rc, "fflush()"); /* Launch ping() */ rc = pthread_create(&t0_ping, &attr, ping, NULL); if (rc) pr_error(rc, "pthread_create()"); exit_from_pong = 0; /* Launch pong() */ rc = pthread_create(&t1_pong, &attr, pong, NULL); if (rc) pr_error(rc, "pthread_create()"); rc = pthread_join(t0_ping, NULL); if (rc) pr_error(rc, "pthread_join()"); rc = pthread_join(t1_pong, NULL); if (rc) pr_error(rc, "pthread_join()"); if (success) { printf("no.\n"); /* no, endianness did not flip inadvertently */ return EXIT_SUCCESS; } printf("yes!\n"); /* yes, endianness did flip inadvertently */ return EXIT_FAILURE; } int main(int argc, char **argv) { return test_harness(tm_trap_test, "tm_trap_test"); } |