]> git.ipfire.org Git - thirdparty/systemd.git/blob - src/basic/random-util.c
random-seed: move pool size determination to random-util.[ch]
[thirdparty/systemd.git] / src / basic / random-util.c
1 /* SPDX-License-Identifier: LGPL-2.1+ */
2
3 #if defined(__i386__) || defined(__x86_64__)
4 #include <cpuid.h>
5 #endif
6
7 #include <elf.h>
8 #include <errno.h>
9 #include <fcntl.h>
10 #include <stdbool.h>
11 #include <stdint.h>
12 #include <stdlib.h>
13 #include <string.h>
14 #include <sys/time.h>
15
16 #if HAVE_SYS_AUXV_H
17 # include <sys/auxv.h>
18 #endif
19
20 #if USE_SYS_RANDOM_H
21 # include <sys/random.h>
22 #else
23 # include <linux/random.h>
24 #endif
25
26 #include "alloc-util.h"
27 #include "fd-util.h"
28 #include "fileio.h"
29 #include "io-util.h"
30 #include "missing.h"
31 #include "parse-util.h"
32 #include "random-util.h"
33 #include "siphash24.h"
34 #include "time-util.h"
35
36 int rdrand(unsigned long *ret) {
37
38 /* So, you are a "security researcher", and you wonder why we bother with using raw RDRAND here,
39 * instead of sticking to /dev/urandom or getrandom()?
40 *
41 * Here's why: early boot. On Linux, during early boot the random pool that backs /dev/urandom and
42 * getrandom() is generally not initialized yet. It is very common that initialization of the random
43 * pool takes a longer time (up to many minutes), in particular on embedded devices that have no
44 * explicit hardware random generator, as well as in virtualized environments such as major cloud
45 * installations that do not provide virtio-rng or a similar mechanism.
46 *
47 * In such an environment using getrandom() synchronously means we'd block the entire system boot-up
48 * until the pool is initialized, i.e. *very* long. Using getrandom() asynchronously (GRND_NONBLOCK)
49 * would mean acquiring randomness during early boot would simply fail. Using /dev/urandom would mean
50 * generating many kmsg log messages about our use of it before the random pool is properly
51 * initialized. Neither of these outcomes is desirable.
52 *
53 * Thus, for very specific purposes we use RDRAND instead of either of these three options. RDRAND
54 * provides us quickly and relatively reliably with random values, without having to delay boot,
55 * without triggering warning messages in kmsg.
56 *
57 * Note that we use RDRAND only under very specific circumstances, when the requirements on the
58 * quality of the returned entropy permit it. Specifically, here are some cases where we *do* use
59 * RDRAND:
60 *
61 * • UUID generation: UUIDs are supposed to be universally unique but are not cryptographic
62 * key material. The quality and trust level of RDRAND should hence be OK: UUIDs should be
63 * generated in a way that is reliably unique, but they do not require ultimate trust into
64 * the entropy generator. systemd generates a number of UUIDs during early boot, including
65 * 'invocation IDs' for every unit spawned that identify the specific invocation of the
66 * service globally, and a number of others. Other alternatives for generating these UUIDs
67 * have been considered, but don't really work: for example, hashing uuids from a local
68 * system identifier combined with a counter falls flat because during early boot disk
69 * storage is not yet available (think: initrd) and thus a system-specific ID cannot be
70 * stored or retrieved yet.
71 *
72 * • Hash table seed generation: systemd uses many hash tables internally. Hash tables are
73 * generally assumed to have O(1) access complexity, but can deteriorate to prohibitive
74 * O(n) access complexity if an attacker manages to trigger a large number of hash
75 * collisions. Thus, systemd (as any software employing hash tables should) uses seeded
76 * hash functions for its hash tables, with a seed generated randomly. The hash tables
77 * systemd employs watch the fill level closely and reseed if necessary. This allows use of
78 * a low quality RNG initially, as long as it improves should a hash table be under attack:
79 * the attacker after all needs to to trigger many collisions to exploit it for the purpose
80 * of DoS, but if doing so improves the seed the attack surface is reduced as the attack
81 * takes place.
82 *
83 * Some cases where we do NOT use RDRAND are:
84 *
85 * • Generation of cryptographic key material 🔑
86 *
87 * • Generation of cryptographic salt values 🧂
88 *
89 * This function returns:
90 *
91 * -EOPNOTSUPP → RDRAND is not available on this system 😔
92 * -EAGAIN → The operation failed this time, but is likely to work if you try again a few
93 * times ♻
94 * -EUCLEAN → We got some random value, but it looked strange, so we refused using it.
95 * This failure might or might not be temporary. 😕
96 */
97
98 #if defined(__i386__) || defined(__x86_64__)
99 static int have_rdrand = -1;
100 unsigned long v;
101 uint8_t success;
102
103 if (have_rdrand < 0) {
104 uint32_t eax, ebx, ecx, edx;
105
106 /* Check if RDRAND is supported by the CPU */
107 if (__get_cpuid(1, &eax, &ebx, &ecx, &edx) == 0) {
108 have_rdrand = false;
109 return -EOPNOTSUPP;
110 }
111
112 /* Compat with old gcc where bit_RDRND didn't exist yet */
113 #ifndef bit_RDRND
114 #define bit_RDRND (1U << 30)
115 #endif
116
117 have_rdrand = !!(ecx & bit_RDRND);
118 }
119
120 if (have_rdrand == 0)
121 return -EOPNOTSUPP;
122
123 asm volatile("rdrand %0;"
124 "setc %1"
125 : "=r" (v),
126 "=qm" (success));
127 msan_unpoison(&success, sizeof(success));
128 if (!success)
129 return -EAGAIN;
130
131 /* Apparently on some AMD CPUs RDRAND will sometimes (after a suspend/resume cycle?) report success
132 * via the carry flag but nonetheless return the same fixed value -1 in all cases. This appears to be
133 * a bad bug in the CPU or firmware. Let's deal with that and work-around this by explicitly checking
134 * for this special value (and also 0, just to be sure) and filtering it out. This is a work-around
135 * only however and something AMD really should fix properly. The Linux kernel should probably work
136 * around this issue by turning off RDRAND altogether on those CPUs. See:
137 * https://github.com/systemd/systemd/issues/11810 */
138 if (v == 0 || v == ULONG_MAX)
139 return log_debug_errno(SYNTHETIC_ERRNO(EUCLEAN),
140 "RDRAND returned suspicious value %lx, assuming bad hardware RNG, not using value.", v);
141
142 *ret = v;
143 return 0;
144 #else
145 return -EOPNOTSUPP;
146 #endif
147 }
148
149 int genuine_random_bytes(void *p, size_t n, RandomFlags flags) {
150 static int have_syscall = -1;
151 _cleanup_close_ int fd = -1;
152 bool got_some = false;
153 int r;
154
155 /* Gathers some high-quality randomness from the kernel (or potentially mid-quality randomness from
156 * the CPU if the RANDOM_ALLOW_RDRAND flag is set). This call won't block, unless the RANDOM_BLOCK
157 * flag is set. If RANDOM_MAY_FAIL is set, an error is returned if the random pool is not
158 * initialized. Otherwise it will always return some data from the kernel, regardless of whether the
159 * random pool is fully initialized or not. If RANDOM_EXTEND_WITH_PSEUDO is set, and some but not
160 * enough better quality randomness could be acquired, the rest is filled up with low quality
161 * randomness.
162 *
163 * Of course, when creating cryptographic key material you really shouldn't use RANDOM_ALLOW_DRDRAND
164 * or even RANDOM_EXTEND_WITH_PSEUDO.
165 *
166 * When generating UUIDs it's fine to use RANDOM_ALLOW_RDRAND but not OK to use
167 * RANDOM_EXTEND_WITH_PSEUDO. In fact RANDOM_EXTEND_WITH_PSEUDO is only really fine when invoked via
168 * an "all bets are off" wrapper, such as random_bytes(), see below. */
169
170 if (n == 0)
171 return 0;
172
173 if (FLAGS_SET(flags, RANDOM_ALLOW_RDRAND))
174 /* Try x86-64' RDRAND intrinsic if we have it. We only use it if high quality randomness is
175 * not required, as we don't trust it (who does?). Note that we only do a single iteration of
176 * RDRAND here, even though the Intel docs suggest calling this in a tight loop of 10
177 * invocations or so. That's because we don't really care about the quality here. We
178 * generally prefer using RDRAND if the caller allows us to, since this way we won't upset
179 * the kernel's random subsystem by accessing it before the pool is initialized (after all it
180 * will kmsg log about every attempt to do so)..*/
181 for (;;) {
182 unsigned long u;
183 size_t m;
184
185 if (rdrand(&u) < 0) {
186 if (got_some && FLAGS_SET(flags, RANDOM_EXTEND_WITH_PSEUDO)) {
187 /* Fill in the remaining bytes using pseudo-random values */
188 pseudo_random_bytes(p, n);
189 return 0;
190 }
191
192 /* OK, this didn't work, let's go to getrandom() + /dev/urandom instead */
193 break;
194 }
195
196 m = MIN(sizeof(u), n);
197 memcpy(p, &u, m);
198
199 p = (uint8_t*) p + m;
200 n -= m;
201
202 if (n == 0)
203 return 0; /* Yay, success! */
204
205 got_some = true;
206 }
207
208 /* Use the getrandom() syscall unless we know we don't have it. */
209 if (have_syscall != 0 && !HAS_FEATURE_MEMORY_SANITIZER) {
210
211 for (;;) {
212 r = getrandom(p, n, FLAGS_SET(flags, RANDOM_BLOCK) ? 0 : GRND_NONBLOCK);
213 if (r > 0) {
214 have_syscall = true;
215
216 if ((size_t) r == n)
217 return 0; /* Yay, success! */
218
219 assert((size_t) r < n);
220 p = (uint8_t*) p + r;
221 n -= r;
222
223 if (FLAGS_SET(flags, RANDOM_EXTEND_WITH_PSEUDO)) {
224 /* Fill in the remaining bytes using pseudo-random values */
225 pseudo_random_bytes(p, n);
226 return 0;
227 }
228
229 got_some = true;
230
231 /* Hmm, we didn't get enough good data but the caller insists on good data? Then try again */
232 if (FLAGS_SET(flags, RANDOM_BLOCK))
233 continue;
234
235 /* Fill in the rest with /dev/urandom */
236 break;
237
238 } else if (r == 0) {
239 have_syscall = true;
240 return -EIO;
241
242 } else if (errno == ENOSYS) {
243 /* We lack the syscall, continue with reading from /dev/urandom. */
244 have_syscall = false;
245 break;
246
247 } else if (errno == EAGAIN) {
248 /* The kernel has no entropy whatsoever. Let's remember to use the syscall
249 * the next time again though.
250 *
251 * If RANDOM_MAY_FAIL is set, return an error so that random_bytes() can
252 * produce some pseudo-random bytes instead. Otherwise, fall back to
253 * /dev/urandom, which we know is empty, but the kernel will produce some
254 * bytes for us on a best-effort basis. */
255 have_syscall = true;
256
257 if (got_some && FLAGS_SET(flags, RANDOM_EXTEND_WITH_PSEUDO)) {
258 /* Fill in the remaining bytes using pseudorandom values */
259 pseudo_random_bytes(p, n);
260 return 0;
261 }
262
263 if (FLAGS_SET(flags, RANDOM_MAY_FAIL))
264 return -ENODATA;
265
266 /* Use /dev/urandom instead */
267 break;
268 } else
269 return -errno;
270 }
271 }
272
273 fd = open("/dev/urandom", O_RDONLY|O_CLOEXEC|O_NOCTTY);
274 if (fd < 0)
275 return errno == ENOENT ? -ENOSYS : -errno;
276
277 return loop_read_exact(fd, p, n, true);
278 }
279
280 void initialize_srand(void) {
281 static bool srand_called = false;
282 unsigned x;
283 #if HAVE_SYS_AUXV_H
284 const void *auxv;
285 #endif
286 unsigned long k;
287
288 if (srand_called)
289 return;
290
291 #if HAVE_SYS_AUXV_H
292 /* The kernel provides us with 16 bytes of entropy in auxv, so let's try to make use of that to seed
293 * the pseudo-random generator. It's better than nothing... But let's first hash it to make it harder
294 * to recover the original value by watching any pseudo-random bits we generate. After all the
295 * AT_RANDOM data might be used by other stuff too (in particular: ASLR), and we probably shouldn't
296 * leak the seed for that. */
297
298 auxv = ULONG_TO_PTR(getauxval(AT_RANDOM));
299 if (auxv) {
300 static const uint8_t auxval_hash_key[16] = {
301 0x92, 0x6e, 0xfe, 0x1b, 0xcf, 0x00, 0x52, 0x9c, 0xcc, 0x42, 0xcf, 0xdc, 0x94, 0x1f, 0x81, 0x0f
302 };
303
304 x = (unsigned) siphash24(auxv, 16, auxval_hash_key);
305 } else
306 #endif
307 x = 0;
308
309 x ^= (unsigned) now(CLOCK_REALTIME);
310 x ^= (unsigned) gettid();
311
312 if (rdrand(&k) >= 0)
313 x ^= (unsigned) k;
314
315 srand(x);
316 srand_called = true;
317 }
318
319 /* INT_MAX gives us only 31 bits, so use 24 out of that. */
320 #if RAND_MAX >= INT_MAX
321 # define RAND_STEP 3
322 #else
323 /* SHORT_INT_MAX or lower gives at most 15 bits, we just just 8 out of that. */
324 # define RAND_STEP 1
325 #endif
326
327 void pseudo_random_bytes(void *p, size_t n) {
328 uint8_t *q;
329
330 /* This returns pseudo-random data using libc's rand() function. You probably never want to call this
331 * directly, because why would you use this if you can get better stuff cheaply? Use random_bytes()
332 * instead, see below: it will fall back to this function if there's nothing better to get, but only
333 * then. */
334
335 initialize_srand();
336
337 for (q = p; q < (uint8_t*) p + n; q += RAND_STEP) {
338 unsigned rr;
339
340 rr = (unsigned) rand();
341
342 #if RAND_STEP >= 3
343 if ((size_t) (q - (uint8_t*) p + 2) < n)
344 q[2] = rr >> 16;
345 #endif
346 #if RAND_STEP >= 2
347 if ((size_t) (q - (uint8_t*) p + 1) < n)
348 q[1] = rr >> 8;
349 #endif
350 q[0] = rr;
351 }
352 }
353
354 void random_bytes(void *p, size_t n) {
355
356 /* This returns high quality randomness if we can get it cheaply. If we can't because for some reason
357 * it is not available we'll try some crappy fallbacks.
358 *
359 * What this function will do:
360 *
361 * • This function will preferably use the CPU's RDRAND operation, if it is available, in
362 * order to return "mid-quality" random values cheaply.
363 *
364 * • Use getrandom() with GRND_NONBLOCK, to return high-quality random values if they are
365 * cheaply available.
366 *
367 * • This function will return pseudo-random data, generated via libc rand() if nothing
368 * better is available.
369 *
370 * • This function will work fine in early boot
371 *
372 * • This function will always succeed
373 *
374 * What this function won't do:
375 *
376 * • This function will never fail: it will give you randomness no matter what. It might not
377 * be high quality, but it will return some, possibly generated via libc's rand() call.
378 *
379 * • This function will never block: if the only way to get good randomness is a blocking,
380 * synchronous getrandom() we'll instead provide you with pseudo-random data.
381 *
382 * This function is hence great for things like seeding hash tables, generating random numeric UNIX
383 * user IDs (that are checked for collisions before use) and such.
384 *
385 * This function is hence not useful for generating UUIDs or cryptographic key material.
386 */
387
388 if (genuine_random_bytes(p, n, RANDOM_EXTEND_WITH_PSEUDO|RANDOM_MAY_FAIL|RANDOM_ALLOW_RDRAND) >= 0)
389 return;
390
391 /* If for some reason some user made /dev/urandom unavailable to us, or the kernel has no entropy, use a PRNG instead. */
392 pseudo_random_bytes(p, n);
393 }
394
395 size_t random_pool_size(void) {
396 _cleanup_free_ char *s = NULL;
397 int r;
398
399 /* Read pool size, if possible */
400 r = read_one_line_file("/proc/sys/kernel/random/poolsize", &s);
401 if (r < 0)
402 log_debug_errno(r, "Failed to read pool size from kernel: %m");
403 else {
404 unsigned sz;
405
406 r = safe_atou(s, &sz);
407 if (r < 0)
408 log_debug_errno(r, "Failed to parse pool size: %s", s);
409 else
410 /* poolsize is in bits on 2.6, but we want bytes */
411 return CLAMP(sz / 8, RANDOM_POOL_SIZE_MIN, RANDOM_POOL_SIZE_MAX);
412 }
413
414 /* Use the minimum as default, if we can't retrieve the correct value */
415 return RANDOM_POOL_SIZE_MIN;
416 }