]>
Commit | Line | Data |
---|---|---|
34dc7c2f BB |
1 | /* |
2 | * CDDL HEADER START | |
3 | * | |
4 | * The contents of this file are subject to the terms of the | |
5 | * Common Development and Distribution License (the "License"). | |
6 | * You may not use this file except in compliance with the License. | |
7 | * | |
8 | * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE | |
9 | * or http://www.opensolaris.org/os/licensing. | |
10 | * See the License for the specific language governing permissions | |
11 | * and limitations under the License. | |
12 | * | |
13 | * When distributing Covered Code, include this CDDL HEADER in each | |
14 | * file and include the License file at usr/src/OPENSOLARIS.LICENSE. | |
15 | * If applicable, add the following below this CDDL HEADER, with the | |
16 | * fields enclosed by brackets "[]" replaced with your own identifying | |
17 | * information: Portions Copyright [yyyy] [name of copyright owner] | |
18 | * | |
19 | * CDDL HEADER END | |
20 | */ | |
21 | /* | |
9babb374 | 22 | * Copyright 2009 Sun Microsystems, Inc. All rights reserved. |
34dc7c2f | 23 | * Use is subject to license terms. |
fc897b24 | 24 | * Copyright (C) 2016 Gvozden Nešković. All rights reserved. |
34dc7c2f | 25 | */ |
3c67d83a TH |
26 | /* |
27 | * Copyright 2013 Saso Kiselkov. All rights reserved. | |
28 | */ | |
34dc7c2f | 29 | |
9babb374 BB |
30 | /* |
31 | * Fletcher Checksums | |
32 | * ------------------ | |
33 | * | |
34 | * ZFS's 2nd and 4th order Fletcher checksums are defined by the following | |
35 | * recurrence relations: | |
36 | * | |
37 | * a = a + f | |
38 | * i i-1 i-1 | |
39 | * | |
40 | * b = b + a | |
41 | * i i-1 i | |
42 | * | |
43 | * c = c + b (fletcher-4 only) | |
44 | * i i-1 i | |
45 | * | |
46 | * d = d + c (fletcher-4 only) | |
47 | * i i-1 i | |
48 | * | |
49 | * Where | |
50 | * a_0 = b_0 = c_0 = d_0 = 0 | |
51 | * and | |
52 | * f_0 .. f_(n-1) are the input data. | |
53 | * | |
54 | * Using standard techniques, these translate into the following series: | |
55 | * | |
56 | * __n_ __n_ | |
57 | * \ | \ | | |
58 | * a = > f b = > i * f | |
59 | * n /___| n - i n /___| n - i | |
60 | * i = 1 i = 1 | |
61 | * | |
62 | * | |
63 | * __n_ __n_ | |
64 | * \ | i*(i+1) \ | i*(i+1)*(i+2) | |
65 | * c = > ------- f d = > ------------- f | |
66 | * n /___| 2 n - i n /___| 6 n - i | |
67 | * i = 1 i = 1 | |
68 | * | |
69 | * For fletcher-2, the f_is are 64-bit, and [ab]_i are 64-bit accumulators. | |
70 | * Since the additions are done mod (2^64), errors in the high bits may not | |
71 | * be noticed. For this reason, fletcher-2 is deprecated. | |
72 | * | |
73 | * For fletcher-4, the f_is are 32-bit, and [abcd]_i are 64-bit accumulators. | |
74 | * A conservative estimate of how big the buffer can get before we overflow | |
75 | * can be estimated using f_i = 0xffffffff for all i: | |
76 | * | |
77 | * % bc | |
78 | * f=2^32-1;d=0; for (i = 1; d<2^64; i++) { d += f*i*(i+1)*(i+2)/6 }; (i-1)*4 | |
79 | * 2264 | |
80 | * quit | |
81 | * % | |
82 | * | |
83 | * So blocks of up to 2k will not overflow. Our largest block size is | |
84 | * 128k, which has 32k 4-byte words, so we can compute the largest possible | |
85 | * accumulators, then divide by 2^64 to figure the max amount of overflow: | |
86 | * | |
87 | * % bc | |
88 | * a=b=c=d=0; f=2^32-1; for (i=1; i<=32*1024; i++) { a+=f; b+=a; c+=b; d+=c } | |
89 | * a/2^64;b/2^64;c/2^64;d/2^64 | |
90 | * 0 | |
91 | * 0 | |
92 | * 1365 | |
93 | * 11186858 | |
94 | * quit | |
95 | * % | |
96 | * | |
97 | * So a and b cannot overflow. To make sure each bit of input has some | |
98 | * effect on the contents of c and d, we can look at what the factors of | |
99 | * the coefficients in the equations for c_n and d_n are. The number of 2s | |
100 | * in the factors determines the lowest set bit in the multiplier. Running | |
101 | * through the cases for n*(n+1)/2 reveals that the highest power of 2 is | |
102 | * 2^14, and for n*(n+1)*(n+2)/6 it is 2^15. So while some data may overflow | |
103 | * the 64-bit accumulators, every bit of every f_i effects every accumulator, | |
104 | * even for 128k blocks. | |
105 | * | |
106 | * If we wanted to make a stronger version of fletcher4 (fletcher4c?), | |
107 | * we could do our calculations mod (2^32 - 1) by adding in the carries | |
108 | * periodically, and store the number of carries in the top 32-bits. | |
109 | * | |
110 | * -------------------- | |
111 | * Checksum Performance | |
112 | * -------------------- | |
113 | * | |
114 | * There are two interesting components to checksum performance: cached and | |
115 | * uncached performance. With cached data, fletcher-2 is about four times | |
116 | * faster than fletcher-4. With uncached data, the performance difference is | |
117 | * negligible, since the cost of a cache fill dominates the processing time. | |
118 | * Even though fletcher-4 is slower than fletcher-2, it is still a pretty | |
119 | * efficient pass over the data. | |
120 | * | |
121 | * In normal operation, the data which is being checksummed is in a buffer | |
122 | * which has been filled either by: | |
123 | * | |
124 | * 1. a compression step, which will be mostly cached, or | |
125 | * 2. a bcopy() or copyin(), which will be uncached (because the | |
126 | * copy is cache-bypassing). | |
127 | * | |
128 | * For both cached and uncached data, both fletcher checksums are much faster | |
129 | * than sha-256, and slower than 'off', which doesn't touch the data at all. | |
130 | */ | |
34dc7c2f BB |
131 | |
132 | #include <sys/types.h> | |
133 | #include <sys/sysmacros.h> | |
134 | #include <sys/byteorder.h> | |
135 | #include <sys/spa.h> | |
fc897b24 | 136 | #include <sys/zio_checksum.h> |
1eeb4562 JX |
137 | #include <sys/zfs_context.h> |
138 | #include <zfs_fletcher.h> | |
139 | ||
fc897b24 | 140 | |
5bf703b8 GN |
141 | static void fletcher_4_scalar_init(fletcher_4_ctx_t *ctx); |
142 | static void fletcher_4_scalar_fini(fletcher_4_ctx_t *ctx, zio_cksum_t *zcp); | |
143 | static void fletcher_4_scalar_native(fletcher_4_ctx_t *ctx, | |
144 | const void *buf, uint64_t size); | |
145 | static void fletcher_4_scalar_byteswap(fletcher_4_ctx_t *ctx, | |
146 | const void *buf, uint64_t size); | |
1eeb4562 JX |
147 | static boolean_t fletcher_4_scalar_valid(void); |
148 | ||
149 | static const fletcher_4_ops_t fletcher_4_scalar_ops = { | |
fc897b24 | 150 | .init_native = fletcher_4_scalar_init, |
5bf703b8 | 151 | .fini_native = fletcher_4_scalar_fini, |
fc897b24 GN |
152 | .compute_native = fletcher_4_scalar_native, |
153 | .init_byteswap = fletcher_4_scalar_init, | |
5bf703b8 | 154 | .fini_byteswap = fletcher_4_scalar_fini, |
1eeb4562 JX |
155 | .compute_byteswap = fletcher_4_scalar_byteswap, |
156 | .valid = fletcher_4_scalar_valid, | |
157 | .name = "scalar" | |
158 | }; | |
159 | ||
fc897b24 GN |
160 | static fletcher_4_ops_t fletcher_4_fastest_impl = { |
161 | .name = "fastest", | |
162 | .valid = fletcher_4_scalar_valid | |
163 | }; | |
164 | ||
165 | static const fletcher_4_ops_t *fletcher_4_impls[] = { | |
1eeb4562 | 166 | &fletcher_4_scalar_ops, |
7f319493 RD |
167 | &fletcher_4_superscalar_ops, |
168 | &fletcher_4_superscalar4_ops, | |
35a76a03 TS |
169 | #if defined(HAVE_SSE2) |
170 | &fletcher_4_sse2_ops, | |
171 | #endif | |
172 | #if defined(HAVE_SSE2) && defined(HAVE_SSSE3) | |
173 | &fletcher_4_ssse3_ops, | |
174 | #endif | |
1eeb4562 JX |
175 | #if defined(HAVE_AVX) && defined(HAVE_AVX2) |
176 | &fletcher_4_avx2_ops, | |
177 | #endif | |
70b258fc GN |
178 | #if defined(__x86_64) && defined(HAVE_AVX512F) |
179 | &fletcher_4_avx512f_ops, | |
180 | #endif | |
24cdeaf1 RD |
181 | #if defined(__aarch64__) |
182 | &fletcher_4_aarch64_neon_ops, | |
183 | #endif | |
1eeb4562 JX |
184 | }; |
185 | ||
fc897b24 GN |
186 | /* Hold all supported implementations */ |
187 | static uint32_t fletcher_4_supp_impls_cnt = 0; | |
188 | static fletcher_4_ops_t *fletcher_4_supp_impls[ARRAY_SIZE(fletcher_4_impls)]; | |
189 | ||
190 | /* Select fletcher4 implementation */ | |
191 | #define IMPL_FASTEST (UINT32_MAX) | |
192 | #define IMPL_CYCLE (UINT32_MAX - 1) | |
193 | #define IMPL_SCALAR (0) | |
194 | ||
195 | static uint32_t fletcher_4_impl_chosen = IMPL_FASTEST; | |
196 | ||
197 | #define IMPL_READ(i) (*(volatile uint32_t *) &(i)) | |
1eeb4562 JX |
198 | |
199 | static struct fletcher_4_impl_selector { | |
fc897b24 GN |
200 | const char *fis_name; |
201 | uint32_t fis_sel; | |
1eeb4562 | 202 | } fletcher_4_impl_selectors[] = { |
1eeb4562 | 203 | #if !defined(_KERNEL) |
fc897b24 | 204 | { "cycle", IMPL_CYCLE }, |
1eeb4562 | 205 | #endif |
fc897b24 GN |
206 | { "fastest", IMPL_FASTEST }, |
207 | { "scalar", IMPL_SCALAR } | |
1eeb4562 JX |
208 | }; |
209 | ||
1eeb4562 JX |
210 | static kstat_t *fletcher_4_kstat; |
211 | ||
fc897b24 GN |
212 | static struct fletcher_4_kstat { |
213 | uint64_t native; | |
214 | uint64_t byteswap; | |
215 | } fletcher_4_stat_data[ARRAY_SIZE(fletcher_4_impls) + 1]; | |
216 | ||
217 | /* Indicate that benchmark has been completed */ | |
218 | static boolean_t fletcher_4_initialized = B_FALSE; | |
34dc7c2f | 219 | |
3c67d83a | 220 | /*ARGSUSED*/ |
34dc7c2f | 221 | void |
3c67d83a TH |
222 | fletcher_2_native(const void *buf, uint64_t size, |
223 | const void *ctx_template, zio_cksum_t *zcp) | |
34dc7c2f BB |
224 | { |
225 | const uint64_t *ip = buf; | |
226 | const uint64_t *ipend = ip + (size / sizeof (uint64_t)); | |
227 | uint64_t a0, b0, a1, b1; | |
228 | ||
229 | for (a0 = b0 = a1 = b1 = 0; ip < ipend; ip += 2) { | |
230 | a0 += ip[0]; | |
231 | a1 += ip[1]; | |
232 | b0 += a0; | |
233 | b1 += a1; | |
234 | } | |
235 | ||
236 | ZIO_SET_CHECKSUM(zcp, a0, a1, b0, b1); | |
237 | } | |
238 | ||
3c67d83a | 239 | /*ARGSUSED*/ |
34dc7c2f | 240 | void |
3c67d83a TH |
241 | fletcher_2_byteswap(const void *buf, uint64_t size, |
242 | const void *ctx_template, zio_cksum_t *zcp) | |
34dc7c2f BB |
243 | { |
244 | const uint64_t *ip = buf; | |
245 | const uint64_t *ipend = ip + (size / sizeof (uint64_t)); | |
246 | uint64_t a0, b0, a1, b1; | |
247 | ||
248 | for (a0 = b0 = a1 = b1 = 0; ip < ipend; ip += 2) { | |
249 | a0 += BSWAP_64(ip[0]); | |
250 | a1 += BSWAP_64(ip[1]); | |
251 | b0 += a0; | |
252 | b1 += a1; | |
253 | } | |
254 | ||
255 | ZIO_SET_CHECKSUM(zcp, a0, a1, b0, b1); | |
256 | } | |
257 | ||
fc897b24 | 258 | static void |
5bf703b8 | 259 | fletcher_4_scalar_init(fletcher_4_ctx_t *ctx) |
34dc7c2f | 260 | { |
5bf703b8 GN |
261 | ZIO_SET_CHECKSUM(&ctx->scalar, 0, 0, 0, 0); |
262 | } | |
263 | ||
264 | static void | |
265 | fletcher_4_scalar_fini(fletcher_4_ctx_t *ctx, zio_cksum_t *zcp) | |
266 | { | |
267 | memcpy(zcp, &ctx->scalar, sizeof (zio_cksum_t)); | |
34dc7c2f BB |
268 | } |
269 | ||
1eeb4562 | 270 | static void |
5bf703b8 GN |
271 | fletcher_4_scalar_native(fletcher_4_ctx_t *ctx, const void *buf, |
272 | uint64_t size) | |
34dc7c2f BB |
273 | { |
274 | const uint32_t *ip = buf; | |
275 | const uint32_t *ipend = ip + (size / sizeof (uint32_t)); | |
276 | uint64_t a, b, c, d; | |
277 | ||
5bf703b8 GN |
278 | a = ctx->scalar.zc_word[0]; |
279 | b = ctx->scalar.zc_word[1]; | |
280 | c = ctx->scalar.zc_word[2]; | |
281 | d = ctx->scalar.zc_word[3]; | |
1eeb4562 JX |
282 | |
283 | for (; ip < ipend; ip++) { | |
284 | a += ip[0]; | |
34dc7c2f BB |
285 | b += a; |
286 | c += b; | |
287 | d += c; | |
288 | } | |
289 | ||
5bf703b8 | 290 | ZIO_SET_CHECKSUM(&ctx->scalar, a, b, c, d); |
34dc7c2f BB |
291 | } |
292 | ||
1eeb4562 | 293 | static void |
5bf703b8 GN |
294 | fletcher_4_scalar_byteswap(fletcher_4_ctx_t *ctx, const void *buf, |
295 | uint64_t size) | |
34dc7c2f BB |
296 | { |
297 | const uint32_t *ip = buf; | |
298 | const uint32_t *ipend = ip + (size / sizeof (uint32_t)); | |
299 | uint64_t a, b, c, d; | |
300 | ||
5bf703b8 GN |
301 | a = ctx->scalar.zc_word[0]; |
302 | b = ctx->scalar.zc_word[1]; | |
303 | c = ctx->scalar.zc_word[2]; | |
304 | d = ctx->scalar.zc_word[3]; | |
34dc7c2f BB |
305 | |
306 | for (; ip < ipend; ip++) { | |
1eeb4562 | 307 | a += BSWAP_32(ip[0]); |
34dc7c2f BB |
308 | b += a; |
309 | c += b; | |
310 | d += c; | |
311 | } | |
312 | ||
5bf703b8 | 313 | ZIO_SET_CHECKSUM(&ctx->scalar, a, b, c, d); |
34dc7c2f BB |
314 | } |
315 | ||
1eeb4562 JX |
316 | static boolean_t |
317 | fletcher_4_scalar_valid(void) | |
318 | { | |
319 | return (B_TRUE); | |
320 | } | |
321 | ||
322 | int | |
323 | fletcher_4_impl_set(const char *val) | |
324 | { | |
fc897b24 GN |
325 | int err = -EINVAL; |
326 | uint32_t impl = IMPL_READ(fletcher_4_impl_chosen); | |
327 | size_t i, val_len; | |
1eeb4562 JX |
328 | |
329 | val_len = strlen(val); | |
330 | while ((val_len > 0) && !!isspace(val[val_len-1])) /* trim '\n' */ | |
331 | val_len--; | |
332 | ||
fc897b24 | 333 | /* check mandatory implementations */ |
1eeb4562 JX |
334 | for (i = 0; i < ARRAY_SIZE(fletcher_4_impl_selectors); i++) { |
335 | const char *name = fletcher_4_impl_selectors[i].fis_name; | |
336 | ||
337 | if (val_len == strlen(name) && | |
338 | strncmp(val, name, val_len) == 0) { | |
fc897b24 GN |
339 | impl = fletcher_4_impl_selectors[i].fis_sel; |
340 | err = 0; | |
1eeb4562 JX |
341 | break; |
342 | } | |
343 | } | |
1eeb4562 | 344 | |
fc897b24 GN |
345 | if (err != 0 && fletcher_4_initialized) { |
346 | /* check all supported implementations */ | |
347 | for (i = 0; i < fletcher_4_supp_impls_cnt; i++) { | |
348 | const char *name = fletcher_4_supp_impls[i]->name; | |
1eeb4562 | 349 | |
fc897b24 GN |
350 | if (val_len == strlen(name) && |
351 | strncmp(val, name, val_len) == 0) { | |
352 | impl = i; | |
353 | err = 0; | |
354 | break; | |
355 | } | |
356 | } | |
357 | } | |
1eeb4562 | 358 | |
fc897b24 GN |
359 | if (err == 0) { |
360 | atomic_swap_32(&fletcher_4_impl_chosen, impl); | |
361 | membar_producer(); | |
362 | } | |
363 | ||
364 | return (err); | |
1eeb4562 JX |
365 | } |
366 | ||
367 | static inline const fletcher_4_ops_t * | |
368 | fletcher_4_impl_get(void) | |
369 | { | |
fc897b24 GN |
370 | fletcher_4_ops_t *ops = NULL; |
371 | const uint32_t impl = IMPL_READ(fletcher_4_impl_chosen); | |
372 | ||
373 | switch (impl) { | |
374 | case IMPL_FASTEST: | |
375 | ASSERT(fletcher_4_initialized); | |
376 | ops = &fletcher_4_fastest_impl; | |
377 | break; | |
1eeb4562 | 378 | #if !defined(_KERNEL) |
fc897b24 GN |
379 | case IMPL_CYCLE: { |
380 | ASSERT(fletcher_4_initialized); | |
381 | ASSERT3U(fletcher_4_supp_impls_cnt, >, 0); | |
382 | ||
383 | static uint32_t cycle_count = 0; | |
384 | uint32_t idx = (++cycle_count) % fletcher_4_supp_impls_cnt; | |
385 | ops = fletcher_4_supp_impls[idx]; | |
1eeb4562 | 386 | } |
fc897b24 | 387 | break; |
1eeb4562 | 388 | #endif |
fc897b24 GN |
389 | default: |
390 | ASSERT3U(fletcher_4_supp_impls_cnt, >, 0); | |
391 | ASSERT3U(impl, <, fletcher_4_supp_impls_cnt); | |
392 | ||
393 | ops = fletcher_4_supp_impls[impl]; | |
394 | break; | |
395 | } | |
396 | ||
397 | ASSERT3P(ops, !=, NULL); | |
398 | ||
399 | return (ops); | |
400 | } | |
401 | ||
fc897b24 | 402 | static inline void |
5bf703b8 | 403 | fletcher_4_native_impl(const void *buf, uint64_t size, zio_cksum_t *zcp) |
fc897b24 | 404 | { |
5bf703b8 GN |
405 | fletcher_4_ctx_t ctx; |
406 | const fletcher_4_ops_t *ops = fletcher_4_impl_get(); | |
407 | ||
408 | ops->init_native(&ctx); | |
409 | ops->compute_native(&ctx, buf, size); | |
410 | ops->fini_native(&ctx, zcp); | |
1eeb4562 JX |
411 | } |
412 | ||
3c67d83a | 413 | /*ARGSUSED*/ |
1eeb4562 | 414 | void |
3c67d83a TH |
415 | fletcher_4_native(const void *buf, uint64_t size, |
416 | const void *ctx_template, zio_cksum_t *zcp) | |
1eeb4562 | 417 | { |
5bf703b8 | 418 | const uint64_t p2size = P2ALIGN(size, 64); |
0dab2e84 | 419 | |
fc897b24 GN |
420 | ASSERT(IS_P2ALIGNED(size, sizeof (uint32_t))); |
421 | ||
5bf703b8 | 422 | if (size == 0 || p2size == 0) { |
fc897b24 | 423 | ZIO_SET_CHECKSUM(zcp, 0, 0, 0, 0); |
5bf703b8 GN |
424 | |
425 | if (size > 0) | |
426 | fletcher_4_scalar_native((fletcher_4_ctx_t *)zcp, | |
427 | buf, size); | |
fc897b24 | 428 | } else { |
5bf703b8 | 429 | fletcher_4_native_impl(buf, p2size, zcp); |
1eeb4562 | 430 | |
fc897b24 | 431 | if (p2size < size) |
5bf703b8 GN |
432 | fletcher_4_scalar_native((fletcher_4_ctx_t *)zcp, |
433 | (char *)buf + p2size, size - p2size); | |
fc897b24 GN |
434 | } |
435 | } | |
436 | ||
437 | void | |
438 | fletcher_4_native_varsize(const void *buf, uint64_t size, zio_cksum_t *zcp) | |
439 | { | |
5bf703b8 GN |
440 | ZIO_SET_CHECKSUM(zcp, 0, 0, 0, 0); |
441 | fletcher_4_scalar_native((fletcher_4_ctx_t *)zcp, buf, size); | |
fc897b24 GN |
442 | } |
443 | ||
444 | static inline void | |
5bf703b8 | 445 | fletcher_4_byteswap_impl(const void *buf, uint64_t size, zio_cksum_t *zcp) |
fc897b24 | 446 | { |
5bf703b8 GN |
447 | fletcher_4_ctx_t ctx; |
448 | const fletcher_4_ops_t *ops = fletcher_4_impl_get(); | |
449 | ||
450 | ops->init_byteswap(&ctx); | |
451 | ops->compute_byteswap(&ctx, buf, size); | |
452 | ops->fini_byteswap(&ctx, zcp); | |
1eeb4562 JX |
453 | } |
454 | ||
3c67d83a | 455 | /*ARGSUSED*/ |
1eeb4562 | 456 | void |
3c67d83a TH |
457 | fletcher_4_byteswap(const void *buf, uint64_t size, |
458 | const void *ctx_template, zio_cksum_t *zcp) | |
1eeb4562 | 459 | { |
5bf703b8 | 460 | const uint64_t p2size = P2ALIGN(size, 64); |
0dab2e84 | 461 | |
fc897b24 GN |
462 | ASSERT(IS_P2ALIGNED(size, sizeof (uint32_t))); |
463 | ||
5bf703b8 | 464 | if (size == 0 || p2size == 0) { |
fc897b24 | 465 | ZIO_SET_CHECKSUM(zcp, 0, 0, 0, 0); |
5bf703b8 GN |
466 | |
467 | if (size > 0) | |
468 | fletcher_4_scalar_byteswap((fletcher_4_ctx_t *)zcp, | |
469 | buf, size); | |
fc897b24 | 470 | } else { |
5bf703b8 | 471 | fletcher_4_byteswap_impl(buf, p2size, zcp); |
1eeb4562 | 472 | |
fc897b24 | 473 | if (p2size < size) |
5bf703b8 GN |
474 | fletcher_4_scalar_byteswap((fletcher_4_ctx_t *)zcp, |
475 | (char *)buf + p2size, size - p2size); | |
fc897b24 | 476 | } |
1eeb4562 JX |
477 | } |
478 | ||
37f520db GN |
479 | /* Incremental Fletcher 4 */ |
480 | ||
5bf703b8 GN |
481 | #define ZFS_FLETCHER_4_INC_MAX_SIZE (8ULL << 20) |
482 | ||
37f520db GN |
483 | static inline void |
484 | fletcher_4_incremental_combine(zio_cksum_t *zcp, const uint64_t size, | |
485 | const zio_cksum_t *nzcp) | |
486 | { | |
487 | const uint64_t c1 = size / sizeof (uint32_t); | |
488 | const uint64_t c2 = c1 * (c1 + 1) / 2; | |
489 | const uint64_t c3 = c2 * (c1 + 2) / 3; | |
490 | ||
5bf703b8 GN |
491 | /* |
492 | * Value of 'c3' overflows on buffer sizes close to 16MiB. For that | |
493 | * reason we split incremental fletcher4 computation of large buffers | |
494 | * to steps of (ZFS_FLETCHER_4_INC_MAX_SIZE) size. | |
495 | */ | |
496 | ASSERT3U(size, <=, ZFS_FLETCHER_4_INC_MAX_SIZE); | |
497 | ||
37f520db GN |
498 | zcp->zc_word[3] += nzcp->zc_word[3] + c1 * zcp->zc_word[2] + |
499 | c2 * zcp->zc_word[1] + c3 * zcp->zc_word[0]; | |
500 | zcp->zc_word[2] += nzcp->zc_word[2] + c1 * zcp->zc_word[1] + | |
501 | c2 * zcp->zc_word[0]; | |
502 | zcp->zc_word[1] += nzcp->zc_word[1] + c1 * zcp->zc_word[0]; | |
503 | zcp->zc_word[0] += nzcp->zc_word[0]; | |
504 | } | |
505 | ||
506 | static inline void | |
507 | fletcher_4_incremental_impl(boolean_t native, const void *buf, uint64_t size, | |
508 | zio_cksum_t *zcp) | |
509 | { | |
37f520db GN |
510 | while (size > 0) { |
511 | zio_cksum_t nzc; | |
5bf703b8 | 512 | uint64_t len = MIN(size, ZFS_FLETCHER_4_INC_MAX_SIZE); |
37f520db GN |
513 | |
514 | if (native) | |
515 | fletcher_4_native(buf, len, NULL, &nzc); | |
516 | else | |
517 | fletcher_4_byteswap(buf, len, NULL, &nzc); | |
518 | ||
519 | fletcher_4_incremental_combine(zcp, len, &nzc); | |
520 | ||
521 | size -= len; | |
522 | buf += len; | |
523 | } | |
524 | } | |
525 | ||
526 | void | |
527 | fletcher_4_incremental_native(const void *buf, uint64_t size, zio_cksum_t *zcp) | |
528 | { | |
5bf703b8 GN |
529 | /* Use scalar impl to directly update cksum of small blocks */ |
530 | if (size < SPA_MINBLOCKSIZE) | |
531 | fletcher_4_scalar_native((fletcher_4_ctx_t *)zcp, buf, size); | |
532 | else | |
533 | fletcher_4_incremental_impl(B_TRUE, buf, size, zcp); | |
37f520db GN |
534 | } |
535 | ||
536 | void | |
537 | fletcher_4_incremental_byteswap(const void *buf, uint64_t size, | |
538 | zio_cksum_t *zcp) | |
539 | { | |
5bf703b8 GN |
540 | /* Use scalar impl to directly update cksum of small blocks */ |
541 | if (size < SPA_MINBLOCKSIZE) | |
542 | fletcher_4_scalar_byteswap((fletcher_4_ctx_t *)zcp, buf, size); | |
543 | else | |
544 | fletcher_4_incremental_impl(B_FALSE, buf, size, zcp); | |
37f520db GN |
545 | } |
546 | ||
547 | ||
548 | /* Fletcher 4 kstats */ | |
549 | ||
fc897b24 GN |
550 | static int |
551 | fletcher_4_kstat_headers(char *buf, size_t size) | |
1eeb4562 | 552 | { |
fc897b24 GN |
553 | ssize_t off = 0; |
554 | ||
555 | off += snprintf(buf + off, size, "%-17s", "implementation"); | |
556 | off += snprintf(buf + off, size - off, "%-15s", "native"); | |
557 | (void) snprintf(buf + off, size - off, "%-15s\n", "byteswap"); | |
558 | ||
559 | return (0); | |
1eeb4562 JX |
560 | } |
561 | ||
fc897b24 GN |
562 | static int |
563 | fletcher_4_kstat_data(char *buf, size_t size, void *data) | |
34dc7c2f | 564 | { |
fc897b24 GN |
565 | struct fletcher_4_kstat *fastest_stat = |
566 | &fletcher_4_stat_data[fletcher_4_supp_impls_cnt]; | |
567 | struct fletcher_4_kstat *curr_stat = (struct fletcher_4_kstat *) data; | |
568 | ssize_t off = 0; | |
569 | ||
570 | if (curr_stat == fastest_stat) { | |
571 | off += snprintf(buf + off, size - off, "%-17s", "fastest"); | |
572 | off += snprintf(buf + off, size - off, "%-15s", | |
573 | fletcher_4_supp_impls[fastest_stat->native]->name); | |
574 | off += snprintf(buf + off, size - off, "%-15s\n", | |
575 | fletcher_4_supp_impls[fastest_stat->byteswap]->name); | |
576 | } else { | |
577 | ptrdiff_t id = curr_stat - fletcher_4_stat_data; | |
578 | ||
579 | off += snprintf(buf + off, size - off, "%-17s", | |
580 | fletcher_4_supp_impls[id]->name); | |
581 | off += snprintf(buf + off, size - off, "%-15llu", | |
582 | (u_longlong_t) curr_stat->native); | |
583 | off += snprintf(buf + off, size - off, "%-15llu\n", | |
584 | (u_longlong_t) curr_stat->byteswap); | |
585 | } | |
586 | ||
587 | return (0); | |
1eeb4562 | 588 | } |
34dc7c2f | 589 | |
fc897b24 GN |
590 | static void * |
591 | fletcher_4_kstat_addr(kstat_t *ksp, loff_t n) | |
1eeb4562 | 592 | { |
fc897b24 GN |
593 | if (n <= fletcher_4_supp_impls_cnt) |
594 | ksp->ks_private = (void *) (fletcher_4_stat_data + n); | |
595 | else | |
596 | ksp->ks_private = NULL; | |
597 | ||
598 | return (ksp->ks_private); | |
599 | } | |
600 | ||
601 | #define FLETCHER_4_FASTEST_FN_COPY(type, src) \ | |
602 | { \ | |
603 | fletcher_4_fastest_impl.init_ ## type = src->init_ ## type; \ | |
604 | fletcher_4_fastest_impl.fini_ ## type = src->fini_ ## type; \ | |
605 | fletcher_4_fastest_impl.compute_ ## type = src->compute_ ## type; \ | |
606 | } | |
607 | ||
608 | #define FLETCHER_4_BENCH_NS (MSEC2NSEC(50)) /* 50ms */ | |
34dc7c2f | 609 | |
fc897b24 GN |
610 | static void |
611 | fletcher_4_benchmark_impl(boolean_t native, char *data, uint64_t data_size) | |
612 | { | |
613 | ||
614 | struct fletcher_4_kstat *fastest_stat = | |
615 | &fletcher_4_stat_data[fletcher_4_supp_impls_cnt]; | |
616 | hrtime_t start; | |
617 | uint64_t run_bw, run_time_ns, best_run = 0; | |
618 | zio_cksum_t zc; | |
619 | uint32_t i, l, sel_save = IMPL_READ(fletcher_4_impl_chosen); | |
620 | ||
621 | zio_checksum_func_t *fletcher_4_test = native ? fletcher_4_native : | |
622 | fletcher_4_byteswap; | |
1eeb4562 | 623 | |
fc897b24 GN |
624 | for (i = 0; i < fletcher_4_supp_impls_cnt; i++) { |
625 | struct fletcher_4_kstat *stat = &fletcher_4_stat_data[i]; | |
626 | uint64_t run_count = 0; | |
1eeb4562 | 627 | |
fc897b24 GN |
628 | /* temporary set an implementation */ |
629 | fletcher_4_impl_chosen = i; | |
1eeb4562 JX |
630 | |
631 | kpreempt_disable(); | |
632 | start = gethrtime(); | |
1eeb4562 | 633 | do { |
fc897b24 | 634 | for (l = 0; l < 32; l++, run_count++) |
3c67d83a | 635 | fletcher_4_test(data, data_size, NULL, &zc); |
fc897b24 GN |
636 | |
637 | run_time_ns = gethrtime() - start; | |
638 | } while (run_time_ns < FLETCHER_4_BENCH_NS); | |
1eeb4562 JX |
639 | kpreempt_enable(); |
640 | ||
fc897b24 GN |
641 | run_bw = data_size * run_count * NANOSEC; |
642 | run_bw /= run_time_ns; /* B/s */ | |
643 | ||
644 | if (native) | |
645 | stat->native = run_bw; | |
646 | else | |
647 | stat->byteswap = run_bw; | |
648 | ||
649 | if (run_bw > best_run) { | |
650 | best_run = run_bw; | |
651 | ||
652 | if (native) { | |
653 | fastest_stat->native = i; | |
654 | FLETCHER_4_FASTEST_FN_COPY(native, | |
655 | fletcher_4_supp_impls[i]); | |
656 | } else { | |
657 | fastest_stat->byteswap = i; | |
658 | FLETCHER_4_FASTEST_FN_COPY(byteswap, | |
659 | fletcher_4_supp_impls[i]); | |
660 | } | |
1eeb4562 | 661 | } |
fc897b24 GN |
662 | } |
663 | ||
664 | /* restore original selection */ | |
665 | atomic_swap_32(&fletcher_4_impl_chosen, sel_save); | |
666 | } | |
1eeb4562 | 667 | |
fc897b24 GN |
668 | void |
669 | fletcher_4_init(void) | |
670 | { | |
671 | static const size_t data_size = 1 << SPA_OLD_MAXBLOCKSHIFT; /* 128kiB */ | |
672 | fletcher_4_ops_t *curr_impl; | |
673 | char *databuf; | |
674 | int i, c; | |
675 | ||
676 | /* move supported impl into fletcher_4_supp_impls */ | |
677 | for (i = 0, c = 0; i < ARRAY_SIZE(fletcher_4_impls); i++) { | |
678 | curr_impl = (fletcher_4_ops_t *) fletcher_4_impls[i]; | |
679 | ||
680 | if (curr_impl->valid && curr_impl->valid()) | |
681 | fletcher_4_supp_impls[c++] = curr_impl; | |
34dc7c2f | 682 | } |
fc897b24 GN |
683 | membar_producer(); /* complete fletcher_4_supp_impls[] init */ |
684 | fletcher_4_supp_impls_cnt = c; /* number of supported impl */ | |
34dc7c2f | 685 | |
fc897b24 GN |
686 | #if !defined(_KERNEL) |
687 | /* Skip benchmarking and use last implementation as fastest */ | |
688 | memcpy(&fletcher_4_fastest_impl, | |
689 | fletcher_4_supp_impls[fletcher_4_supp_impls_cnt-1], | |
690 | sizeof (fletcher_4_fastest_impl)); | |
691 | fletcher_4_fastest_impl.name = "fastest"; | |
692 | membar_producer(); | |
1eeb4562 | 693 | |
fc897b24 | 694 | fletcher_4_initialized = B_TRUE; |
fc897b24 GN |
695 | return; |
696 | #endif | |
697 | /* Benchmark all supported implementations */ | |
698 | databuf = vmem_alloc(data_size, KM_SLEEP); | |
699 | for (i = 0; i < data_size / sizeof (uint64_t); i++) | |
700 | ((uint64_t *)databuf)[i] = (uintptr_t)(databuf+i); /* warm-up */ | |
701 | ||
702 | fletcher_4_benchmark_impl(B_FALSE, databuf, data_size); | |
703 | fletcher_4_benchmark_impl(B_TRUE, databuf, data_size); | |
704 | ||
705 | vmem_free(databuf, data_size); | |
706 | ||
707 | /* install kstats for all implementations */ | |
708 | fletcher_4_kstat = kstat_create("zfs", 0, "fletcher_4_bench", "misc", | |
709 | KSTAT_TYPE_RAW, 0, KSTAT_FLAG_VIRTUAL); | |
1eeb4562 | 710 | if (fletcher_4_kstat != NULL) { |
fc897b24 GN |
711 | fletcher_4_kstat->ks_data = NULL; |
712 | fletcher_4_kstat->ks_ndata = UINT32_MAX; | |
713 | kstat_set_raw_ops(fletcher_4_kstat, | |
714 | fletcher_4_kstat_headers, | |
715 | fletcher_4_kstat_data, | |
716 | fletcher_4_kstat_addr); | |
1eeb4562 JX |
717 | kstat_install(fletcher_4_kstat); |
718 | } | |
fc897b24 GN |
719 | |
720 | /* Finish initialization */ | |
721 | fletcher_4_initialized = B_TRUE; | |
1eeb4562 JX |
722 | } |
723 | ||
724 | void | |
725 | fletcher_4_fini(void) | |
726 | { | |
1eeb4562 JX |
727 | if (fletcher_4_kstat != NULL) { |
728 | kstat_delete(fletcher_4_kstat); | |
729 | fletcher_4_kstat = NULL; | |
730 | } | |
34dc7c2f | 731 | } |
c28b2279 BB |
732 | |
733 | #if defined(_KERNEL) && defined(HAVE_SPL) | |
9cc1844a | 734 | #include <linux/mod_compat.h> |
1eeb4562 JX |
735 | |
736 | static int | |
9cc1844a | 737 | fletcher_4_param_get(char *buffer, zfs_kernel_param_t *unused) |
1eeb4562 | 738 | { |
fc897b24 GN |
739 | const uint32_t impl = IMPL_READ(fletcher_4_impl_chosen); |
740 | char *fmt; | |
1eeb4562 JX |
741 | int i, cnt = 0; |
742 | ||
fc897b24 GN |
743 | /* list fastest */ |
744 | fmt = (impl == IMPL_FASTEST) ? "[%s] " : "%s "; | |
745 | cnt += sprintf(buffer + cnt, fmt, "fastest"); | |
1eeb4562 | 746 | |
fc897b24 GN |
747 | /* list all supported implementations */ |
748 | for (i = 0; i < fletcher_4_supp_impls_cnt; i++) { | |
749 | fmt = (i == impl) ? "[%s] " : "%s "; | |
750 | cnt += sprintf(buffer + cnt, fmt, | |
751 | fletcher_4_supp_impls[i]->name); | |
1eeb4562 JX |
752 | } |
753 | ||
754 | return (cnt); | |
755 | } | |
756 | ||
757 | static int | |
9cc1844a | 758 | fletcher_4_param_set(const char *val, zfs_kernel_param_t *unused) |
1eeb4562 JX |
759 | { |
760 | return (fletcher_4_impl_set(val)); | |
761 | } | |
762 | ||
763 | /* | |
764 | * Choose a fletcher 4 implementation in ZFS. | |
fc897b24 | 765 | * Users can choose "cycle" to exercise all implementations, but this is |
1eeb4562 JX |
766 | * for testing purpose therefore it can only be set in user space. |
767 | */ | |
768 | module_param_call(zfs_fletcher_4_impl, | |
769 | fletcher_4_param_set, fletcher_4_param_get, NULL, 0644); | |
fc897b24 | 770 | MODULE_PARM_DESC(zfs_fletcher_4_impl, "Select fletcher 4 implementation."); |
1eeb4562 JX |
771 | |
772 | EXPORT_SYMBOL(fletcher_4_init); | |
773 | EXPORT_SYMBOL(fletcher_4_fini); | |
c28b2279 BB |
774 | EXPORT_SYMBOL(fletcher_2_native); |
775 | EXPORT_SYMBOL(fletcher_2_byteswap); | |
776 | EXPORT_SYMBOL(fletcher_4_native); | |
fc897b24 | 777 | EXPORT_SYMBOL(fletcher_4_native_varsize); |
c28b2279 BB |
778 | EXPORT_SYMBOL(fletcher_4_byteswap); |
779 | EXPORT_SYMBOL(fletcher_4_incremental_native); | |
780 | EXPORT_SYMBOL(fletcher_4_incremental_byteswap); | |
781 | #endif |