1/* SPDX-License-Identifier: GPL-2.0 */
2/*
3 * syscall_numbering.c - test calling the x86-64 kernel with various
4 * valid and invalid system call numbers.
5 *
6 * Copyright (c) 2018 Andrew Lutomirski
7 */
8
9#define _GNU_SOURCE
10
11#include <stdlib.h>
12#include <stdio.h>
13#include <stdbool.h>
14#include <errno.h>
15#include <unistd.h>
16#include <string.h>
17#include <fcntl.h>
18#include <limits.h>
19#include <signal.h>
20#include <sysexits.h>
21
22#include <sys/ptrace.h>
23#include <sys/user.h>
24#include <sys/wait.h>
25#include <sys/mman.h>
26
27#include <linux/ptrace.h>
28
29/* Common system call numbers */
30#define SYS_READ 0
31#define SYS_WRITE 1
32#define SYS_GETPID 39
33/* x64-only system call numbers */
34#define X64_IOCTL 16
35#define X64_READV 19
36#define X64_WRITEV 20
37/* x32-only system call numbers (without X32_BIT) */
38#define X32_IOCTL 514
39#define X32_READV 515
40#define X32_WRITEV 516
41
42#define X32_BIT 0x40000000
43
44static int nullfd = -1; /* File descriptor for /dev/null */
45static bool with_x32; /* x32 supported on this kernel? */
46
47enum ptrace_pass {
48 PTP_NOTHING,
49 PTP_GETREGS,
50 PTP_WRITEBACK,
51 PTP_FUZZRET,
52 PTP_FUZZHIGH,
53 PTP_INTNUM,
54 PTP_DONE
55};
56
57static const char * const ptrace_pass_name[] =
58{
59 [PTP_NOTHING] = "just stop, no data read",
60 [PTP_GETREGS] = "only getregs",
61 [PTP_WRITEBACK] = "getregs, unmodified setregs",
62 [PTP_FUZZRET] = "modifying the default return",
63 [PTP_FUZZHIGH] = "clobbering the top 32 bits",
64 [PTP_INTNUM] = "sign-extending the syscall number",
65};
66
67/*
68 * Shared memory block between tracer and test
69 */
70struct shared {
71 unsigned int nerr; /* Total error count */
72 unsigned int indent; /* Message indentation level */
73 enum ptrace_pass ptrace_pass;
74 bool probing_syscall; /* In probe_syscall() */
75};
76static volatile struct shared *sh;
77
78static inline unsigned int offset(void)
79{
80 unsigned int level = sh ? sh->indent : 0;
81
82 return 8 + level * 4;
83}
84
85#define msg(lvl, fmt, ...) printf("%-*s" fmt, offset(), "[" #lvl "]", \
86 ## __VA_ARGS__)
87
88#define run(fmt, ...) msg(RUN, fmt, ## __VA_ARGS__)
89#define info(fmt, ...) msg(INFO, fmt, ## __VA_ARGS__)
90#define ok(fmt, ...) msg(OK, fmt, ## __VA_ARGS__)
91
92#define fail(fmt, ...) \
93 do { \
94 msg(FAIL, fmt, ## __VA_ARGS__); \
95 sh->nerr++; \
96 } while (0)
97
98#define crit(fmt, ...) \
99 do { \
100 sh->indent = 0; \
101 msg(FAIL, fmt, ## __VA_ARGS__); \
102 msg(SKIP, "Unable to run test\n"); \
103 exit(EX_OSERR); \
104 } while (0)
105
106/* Sentinel for ptrace-modified return value */
107#define MODIFIED_BY_PTRACE -9999
108
109/*
110 * Directly invokes the given syscall with nullfd as the first argument
111 * and the rest zero. Avoids involving glibc wrappers in case they ever
112 * end up intercepting some system calls for some reason, or modify
113 * the system call number itself.
114 */
115static long long probe_syscall(int msb, int lsb)
116{
117 register long long arg1 asm("rdi") = nullfd;
118 register long long arg2 asm("rsi") = 0;
119 register long long arg3 asm("rdx") = 0;
120 register long long arg4 asm("r10") = 0;
121 register long long arg5 asm("r8") = 0;
122 register long long arg6 asm("r9") = 0;
123 long long nr = ((long long)msb << 32) | (unsigned int)lsb;
124 long long ret;
125
126 /*
127 * We pass in an extra copy of the extended system call number
128 * in %rbx, so we can examine it from the ptrace handler without
129 * worrying about it being possibly modified. This is to test
130 * the validity of struct user regs.orig_rax a.k.a.
131 * struct pt_regs.orig_ax.
132 */
133 sh->probing_syscall = true;
134 asm volatile("syscall"
135 : "=a" (ret)
136 : "a" (nr), "b" (nr),
137 "r" (arg1), "r" (arg2), "r" (arg3),
138 "r" (arg4), "r" (arg5), "r" (arg6)
139 : "rcx", "r11", "memory", "cc");
140 sh->probing_syscall = false;
141
142 return ret;
143}
144
145static const char *syscall_str(int msb, int start, int end)
146{
147 static char buf[64];
148 const char * const type = (start & X32_BIT) ? "x32" : "x64";
149 int lsb = start;
150
151 /*
152 * Improve readability by stripping the x32 bit, but round
153 * toward zero so we don't display -1 as -1073741825.
154 */
155 if (lsb < 0)
156 lsb |= X32_BIT;
157 else
158 lsb &= ~X32_BIT;
159
160 if (start == end)
161 snprintf(buf, size: sizeof buf, fmt: "%s syscall %d:%d",
162 type, msb, lsb);
163 else
164 snprintf(buf, size: sizeof buf, fmt: "%s syscalls %d:%d..%d",
165 type, msb, lsb, lsb + (end-start));
166
167 return buf;
168}
169
170static unsigned int _check_for(int msb, int start, int end, long long expect,
171 const char *expect_str)
172{
173 unsigned int err = 0;
174
175 sh->indent++;
176 if (start != end)
177 sh->indent++;
178
179 for (int nr = start; nr <= end; nr++) {
180 long long ret = probe_syscall(msb, lsb: nr);
181
182 if (ret != expect) {
183 fail("%s returned %lld, but it should have returned %s\n",
184 syscall_str(msb, nr, nr),
185 ret, expect_str);
186 err++;
187 }
188 }
189
190 if (start != end)
191 sh->indent--;
192
193 if (err) {
194 if (start != end)
195 fail("%s had %u failure%s\n",
196 syscall_str(msb, start, end),
197 err, err == 1 ? "s" : "");
198 } else {
199 ok("%s returned %s as expected\n",
200 syscall_str(msb, start, end), expect_str);
201 }
202
203 sh->indent--;
204
205 return err;
206}
207
208#define check_for(msb,start,end,expect) \
209 _check_for(msb,start,end,expect,#expect)
210
211static bool check_zero(int msb, int nr)
212{
213 return check_for(msb, nr, nr, 0);
214}
215
216static bool check_enosys(int msb, int nr)
217{
218 return check_for(msb, nr, nr, -ENOSYS);
219}
220
221/*
222 * Anyone diagnosing a failure will want to know whether the kernel
223 * supports x32. Tell them. This can also be used to conditionalize
224 * tests based on existence or nonexistence of x32.
225 */
226static bool test_x32(void)
227{
228 long long ret;
229 pid_t mypid = getpid();
230
231 run("Checking for x32 by calling x32 getpid()\n");
232 ret = probe_syscall(msb: 0, SYS_GETPID | X32_BIT);
233
234 sh->indent++;
235 if (ret == mypid) {
236 info("x32 is supported\n");
237 with_x32 = true;
238 } else if (ret == -ENOSYS) {
239 info("x32 is not supported\n");
240 with_x32 = false;
241 } else {
242 fail("x32 getpid() returned %lld, but it should have returned either %lld or -ENOSYS\n", ret, (long long)mypid);
243 with_x32 = false;
244 }
245 sh->indent--;
246 return with_x32;
247}
248
249static void test_syscalls_common(int msb)
250{
251 enum ptrace_pass pass = sh->ptrace_pass;
252
253 run("Checking some common syscalls as 64 bit\n");
254 check_zero(msb, SYS_READ);
255 check_zero(msb, SYS_WRITE);
256
257 run("Checking some 64-bit only syscalls as 64 bit\n");
258 check_zero(msb, X64_READV);
259 check_zero(msb, X64_WRITEV);
260
261 run("Checking out of range system calls\n");
262 check_for(msb, -64, -2, -ENOSYS);
263 if (pass >= PTP_FUZZRET)
264 check_for(msb, -1, -1, MODIFIED_BY_PTRACE);
265 else
266 check_for(msb, -1, -1, -ENOSYS);
267 check_for(msb, X32_BIT-64, X32_BIT-1, -ENOSYS);
268 check_for(msb, -64-X32_BIT, -1-X32_BIT, -ENOSYS);
269 check_for(msb, INT_MAX-64, INT_MAX-1, -ENOSYS);
270}
271
272static void test_syscalls_with_x32(int msb)
273{
274 /*
275 * Syscalls 512-547 are "x32" syscalls. They are
276 * intended to be called with the x32 (0x40000000) bit
277 * set. Calling them without the x32 bit set is
278 * nonsense and should not work.
279 */
280 run("Checking x32 syscalls as 64 bit\n");
281 check_for(msb, 512, 547, -ENOSYS);
282
283 run("Checking some common syscalls as x32\n");
284 check_zero(msb, SYS_READ | X32_BIT);
285 check_zero(msb, SYS_WRITE | X32_BIT);
286
287 run("Checking some x32 syscalls as x32\n");
288 check_zero(msb, X32_READV | X32_BIT);
289 check_zero(msb, X32_WRITEV | X32_BIT);
290
291 run("Checking some 64-bit syscalls as x32\n");
292 check_enosys(msb, X64_IOCTL | X32_BIT);
293 check_enosys(msb, X64_READV | X32_BIT);
294 check_enosys(msb, X64_WRITEV | X32_BIT);
295}
296
297static void test_syscalls_without_x32(int msb)
298{
299 run("Checking for absence of x32 system calls\n");
300 check_for(msb, 0 | X32_BIT, 999 | X32_BIT, -ENOSYS);
301}
302
303static void test_syscall_numbering(void)
304{
305 static const int msbs[] = {
306 0, 1, -1, X32_BIT-1, X32_BIT, X32_BIT-1, -X32_BIT, INT_MAX,
307 INT_MIN, INT_MIN+1
308 };
309
310 sh->indent++;
311
312 /*
313 * The MSB is supposed to be ignored, so we loop over a few
314 * to test that out.
315 */
316 for (size_t i = 0; i < sizeof(msbs)/sizeof(msbs[0]); i++) {
317 int msb = msbs[i];
318 run("Checking system calls with msb = %d (0x%x)\n",
319 msb, msb);
320
321 sh->indent++;
322
323 test_syscalls_common(msb);
324 if (with_x32)
325 test_syscalls_with_x32(msb);
326 else
327 test_syscalls_without_x32(msb);
328
329 sh->indent--;
330 }
331
332 sh->indent--;
333}
334
335static void syscall_numbering_tracee(void)
336{
337 enum ptrace_pass pass;
338
339 if (ptrace(PTRACE_TRACEME, 0, 0, 0)) {
340 crit("Failed to request tracing\n");
341 return;
342 }
343 raise(SIGSTOP);
344
345 for (sh->ptrace_pass = pass = PTP_NOTHING; pass < PTP_DONE;
346 sh->ptrace_pass = ++pass) {
347 run("Running tests under ptrace: %s\n", ptrace_pass_name[pass]);
348 test_syscall_numbering();
349 }
350}
351
352static void mess_with_syscall(pid_t testpid, enum ptrace_pass pass)
353{
354 struct user_regs_struct regs;
355
356 sh->probing_syscall = false; /* Do this on entry only */
357
358 /* For these, don't even getregs */
359 if (pass == PTP_NOTHING || pass == PTP_DONE)
360 return;
361
362 ptrace(PTRACE_GETREGS, testpid, NULL, &regs);
363
364 if (regs.orig_rax != regs.rbx) {
365 fail("orig_rax %#llx doesn't match syscall number %#llx\n",
366 (unsigned long long)regs.orig_rax,
367 (unsigned long long)regs.rbx);
368 }
369
370 switch (pass) {
371 case PTP_GETREGS:
372 /* Just read, no writeback */
373 return;
374 case PTP_WRITEBACK:
375 /* Write back the same register state verbatim */
376 break;
377 case PTP_FUZZRET:
378 regs.rax = MODIFIED_BY_PTRACE;
379 break;
380 case PTP_FUZZHIGH:
381 regs.rax = MODIFIED_BY_PTRACE;
382 regs.orig_rax = regs.orig_rax | 0xffffffff00000000ULL;
383 break;
384 case PTP_INTNUM:
385 regs.rax = MODIFIED_BY_PTRACE;
386 regs.orig_rax = (int)regs.orig_rax;
387 break;
388 default:
389 crit("invalid ptrace_pass\n");
390 break;
391 }
392
393 ptrace(PTRACE_SETREGS, testpid, NULL, &regs);
394}
395
396static void syscall_numbering_tracer(pid_t testpid)
397{
398 int wstatus;
399
400 do {
401 pid_t wpid = waitpid(testpid, &wstatus, 0);
402 if (wpid < 0 && errno != EINTR)
403 break;
404 if (wpid != testpid)
405 continue;
406 if (!WIFSTOPPED(wstatus))
407 break; /* Thread exited? */
408
409 if (sh->probing_syscall && WSTOPSIG(wstatus) == SIGTRAP)
410 mess_with_syscall(testpid, pass: sh->ptrace_pass);
411 } while (sh->ptrace_pass != PTP_DONE &&
412 !ptrace(PTRACE_SYSCALL, testpid, NULL, NULL));
413
414 ptrace(PTRACE_DETACH, testpid, NULL, NULL);
415
416 /* Wait for the child process to terminate */
417 while (waitpid(testpid, &wstatus, 0) != testpid || !WIFEXITED(wstatus))
418 /* wait some more */;
419}
420
421static void test_traced_syscall_numbering(void)
422{
423 pid_t testpid;
424
425 /* Launch the test thread; this thread continues as the tracer thread */
426 testpid = fork();
427
428 if (testpid < 0) {
429 crit("Unable to launch tracer process\n");
430 } else if (testpid == 0) {
431 syscall_numbering_tracee();
432 _exit(0);
433 } else {
434 syscall_numbering_tracer(testpid);
435 }
436}
437
438int main(void)
439{
440 unsigned int nerr;
441
442 /*
443 * It is quite likely to get a segfault on a failure, so make
444 * sure the message gets out by setting stdout to nonbuffered.
445 */
446 setvbuf(stdout, NULL, _IONBF, 0);
447
448 /*
449 * Harmless file descriptor to work on...
450 */
451 nullfd = open("/dev/null", O_RDWR);
452 if (nullfd < 0) {
453 crit("Unable to open /dev/null: %s\n", strerror(errno));
454 }
455
456 /*
457 * Set up a block of shared memory...
458 */
459 sh = mmap(NULL, sysconf(_SC_PAGE_SIZE), PROT_READ|PROT_WRITE,
460 MAP_ANONYMOUS|MAP_SHARED, 0, 0);
461 if (sh == MAP_FAILED) {
462 crit("Unable to allocated shared memory block: %s\n",
463 strerror(errno));
464 }
465
466 with_x32 = test_x32();
467
468 run("Running tests without ptrace...\n");
469 test_syscall_numbering();
470
471 test_traced_syscall_numbering();
472
473 nerr = sh->nerr;
474 if (!nerr) {
475 ok("All system calls succeeded or failed as expected\n");
476 return 0;
477 } else {
478 fail("A total of %u system call%s had incorrect behavior\n",
479 nerr, nerr != 1 ? "s" : "");
480 return 1;
481 }
482}
483

source code of linux/tools/testing/selftests/x86/syscall_numbering.c