]>
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 | |
a6255b7f DQ |
30 | /* |
31 | * Copyright (c) 2016 by Delphix. All rights reserved. | |
32 | */ | |
33 | ||
9babb374 BB |
34 | /* |
35 | * Fletcher Checksums | |
36 | * ------------------ | |
37 | * | |
38 | * ZFS's 2nd and 4th order Fletcher checksums are defined by the following | |
39 | * recurrence relations: | |
40 | * | |
41 | * a = a + f | |
42 | * i i-1 i-1 | |
43 | * | |
44 | * b = b + a | |
45 | * i i-1 i | |
46 | * | |
47 | * c = c + b (fletcher-4 only) | |
48 | * i i-1 i | |
49 | * | |
50 | * d = d + c (fletcher-4 only) | |
51 | * i i-1 i | |
52 | * | |
53 | * Where | |
54 | * a_0 = b_0 = c_0 = d_0 = 0 | |
55 | * and | |
56 | * f_0 .. f_(n-1) are the input data. | |
57 | * | |
58 | * Using standard techniques, these translate into the following series: | |
59 | * | |
60 | * __n_ __n_ | |
61 | * \ | \ | | |
62 | * a = > f b = > i * f | |
63 | * n /___| n - i n /___| n - i | |
64 | * i = 1 i = 1 | |
65 | * | |
66 | * | |
67 | * __n_ __n_ | |
68 | * \ | i*(i+1) \ | i*(i+1)*(i+2) | |
69 | * c = > ------- f d = > ------------- f | |
70 | * n /___| 2 n - i n /___| 6 n - i | |
71 | * i = 1 i = 1 | |
72 | * | |
73 | * For fletcher-2, the f_is are 64-bit, and [ab]_i are 64-bit accumulators. | |
74 | * Since the additions are done mod (2^64), errors in the high bits may not | |
75 | * be noticed. For this reason, fletcher-2 is deprecated. | |
76 | * | |
77 | * For fletcher-4, the f_is are 32-bit, and [abcd]_i are 64-bit accumulators. | |
78 | * A conservative estimate of how big the buffer can get before we overflow | |
79 | * can be estimated using f_i = 0xffffffff for all i: | |
80 | * | |
81 | * % bc | |
82 | * f=2^32-1;d=0; for (i = 1; d<2^64; i++) { d += f*i*(i+1)*(i+2)/6 }; (i-1)*4 | |
83 | * 2264 | |
84 | * quit | |
85 | * % | |
86 | * | |
87 | * So blocks of up to 2k will not overflow. Our largest block size is | |
88 | * 128k, which has 32k 4-byte words, so we can compute the largest possible | |
89 | * accumulators, then divide by 2^64 to figure the max amount of overflow: | |
90 | * | |
91 | * % bc | |
92 | * a=b=c=d=0; f=2^32-1; for (i=1; i<=32*1024; i++) { a+=f; b+=a; c+=b; d+=c } | |
93 | * a/2^64;b/2^64;c/2^64;d/2^64 | |
94 | * 0 | |
95 | * 0 | |
96 | * 1365 | |
97 | * 11186858 | |
98 | * quit | |
99 | * % | |
100 | * | |
101 | * So a and b cannot overflow. To make sure each bit of input has some | |
102 | * effect on the contents of c and d, we can look at what the factors of | |
103 | * the coefficients in the equations for c_n and d_n are. The number of 2s | |
104 | * in the factors determines the lowest set bit in the multiplier. Running | |
105 | * through the cases for n*(n+1)/2 reveals that the highest power of 2 is | |
106 | * 2^14, and for n*(n+1)*(n+2)/6 it is 2^15. So while some data may overflow | |
107 | * the 64-bit accumulators, every bit of every f_i effects every accumulator, | |
108 | * even for 128k blocks. | |
109 | * | |
110 | * If we wanted to make a stronger version of fletcher4 (fletcher4c?), | |
111 | * we could do our calculations mod (2^32 - 1) by adding in the carries | |
112 | * periodically, and store the number of carries in the top 32-bits. | |
113 | * | |
114 | * -------------------- | |
115 | * Checksum Performance | |
116 | * -------------------- | |
117 | * | |
118 | * There are two interesting components to checksum performance: cached and | |
119 | * uncached performance. With cached data, fletcher-2 is about four times | |
120 | * faster than fletcher-4. With uncached data, the performance difference is | |
121 | * negligible, since the cost of a cache fill dominates the processing time. | |
122 | * Even though fletcher-4 is slower than fletcher-2, it is still a pretty | |
123 | * efficient pass over the data. | |
124 | * | |
125 | * In normal operation, the data which is being checksummed is in a buffer | |
126 | * which has been filled either by: | |
127 | * | |
128 | * 1. a compression step, which will be mostly cached, or | |
129 | * 2. a bcopy() or copyin(), which will be uncached (because the | |
130 | * copy is cache-bypassing). | |
131 | * | |
132 | * For both cached and uncached data, both fletcher checksums are much faster | |
133 | * than sha-256, and slower than 'off', which doesn't touch the data at all. | |
134 | */ | |
34dc7c2f BB |
135 | |
136 | #include <sys/types.h> | |
137 | #include <sys/sysmacros.h> | |
138 | #include <sys/byteorder.h> | |
139 | #include <sys/spa.h> | |
fc897b24 | 140 | #include <sys/zio_checksum.h> |
1eeb4562 JX |
141 | #include <sys/zfs_context.h> |
142 | #include <zfs_fletcher.h> | |
143 | ||
2fe36b0b | 144 | #define FLETCHER_MIN_SIMD_SIZE 64 |
fc897b24 | 145 | |
5bf703b8 GN |
146 | static void fletcher_4_scalar_init(fletcher_4_ctx_t *ctx); |
147 | static void fletcher_4_scalar_fini(fletcher_4_ctx_t *ctx, zio_cksum_t *zcp); | |
148 | static void fletcher_4_scalar_native(fletcher_4_ctx_t *ctx, | |
149 | const void *buf, uint64_t size); | |
150 | static void fletcher_4_scalar_byteswap(fletcher_4_ctx_t *ctx, | |
151 | const void *buf, uint64_t size); | |
1eeb4562 JX |
152 | static boolean_t fletcher_4_scalar_valid(void); |
153 | ||
154 | static const fletcher_4_ops_t fletcher_4_scalar_ops = { | |
fc897b24 | 155 | .init_native = fletcher_4_scalar_init, |
5bf703b8 | 156 | .fini_native = fletcher_4_scalar_fini, |
fc897b24 GN |
157 | .compute_native = fletcher_4_scalar_native, |
158 | .init_byteswap = fletcher_4_scalar_init, | |
5bf703b8 | 159 | .fini_byteswap = fletcher_4_scalar_fini, |
1eeb4562 JX |
160 | .compute_byteswap = fletcher_4_scalar_byteswap, |
161 | .valid = fletcher_4_scalar_valid, | |
162 | .name = "scalar" | |
163 | }; | |
164 | ||
fc897b24 GN |
165 | static fletcher_4_ops_t fletcher_4_fastest_impl = { |
166 | .name = "fastest", | |
167 | .valid = fletcher_4_scalar_valid | |
168 | }; | |
169 | ||
170 | static const fletcher_4_ops_t *fletcher_4_impls[] = { | |
1eeb4562 | 171 | &fletcher_4_scalar_ops, |
7f319493 RD |
172 | &fletcher_4_superscalar_ops, |
173 | &fletcher_4_superscalar4_ops, | |
35a76a03 TS |
174 | #if defined(HAVE_SSE2) |
175 | &fletcher_4_sse2_ops, | |
176 | #endif | |
177 | #if defined(HAVE_SSE2) && defined(HAVE_SSSE3) | |
178 | &fletcher_4_ssse3_ops, | |
179 | #endif | |
1eeb4562 JX |
180 | #if defined(HAVE_AVX) && defined(HAVE_AVX2) |
181 | &fletcher_4_avx2_ops, | |
182 | #endif | |
70b258fc GN |
183 | #if defined(__x86_64) && defined(HAVE_AVX512F) |
184 | &fletcher_4_avx512f_ops, | |
185 | #endif | |
24cdeaf1 RD |
186 | #if defined(__aarch64__) |
187 | &fletcher_4_aarch64_neon_ops, | |
188 | #endif | |
1eeb4562 JX |
189 | }; |
190 | ||
fc897b24 GN |
191 | /* Hold all supported implementations */ |
192 | static uint32_t fletcher_4_supp_impls_cnt = 0; | |
193 | static fletcher_4_ops_t *fletcher_4_supp_impls[ARRAY_SIZE(fletcher_4_impls)]; | |
194 | ||
195 | /* Select fletcher4 implementation */ | |
196 | #define IMPL_FASTEST (UINT32_MAX) | |
197 | #define IMPL_CYCLE (UINT32_MAX - 1) | |
198 | #define IMPL_SCALAR (0) | |
199 | ||
200 | static uint32_t fletcher_4_impl_chosen = IMPL_FASTEST; | |
201 | ||
202 | #define IMPL_READ(i) (*(volatile uint32_t *) &(i)) | |
1eeb4562 JX |
203 | |
204 | static struct fletcher_4_impl_selector { | |
fc897b24 GN |
205 | const char *fis_name; |
206 | uint32_t fis_sel; | |
1eeb4562 | 207 | } fletcher_4_impl_selectors[] = { |
1eeb4562 | 208 | #if !defined(_KERNEL) |
fc897b24 | 209 | { "cycle", IMPL_CYCLE }, |
1eeb4562 | 210 | #endif |
fc897b24 GN |
211 | { "fastest", IMPL_FASTEST }, |
212 | { "scalar", IMPL_SCALAR } | |
1eeb4562 JX |
213 | }; |
214 | ||
46364cb2 | 215 | #if defined(_KERNEL) |
1eeb4562 | 216 | static kstat_t *fletcher_4_kstat; |
46364cb2 | 217 | #endif |
1eeb4562 | 218 | |
fc897b24 GN |
219 | static struct fletcher_4_kstat { |
220 | uint64_t native; | |
221 | uint64_t byteswap; | |
222 | } fletcher_4_stat_data[ARRAY_SIZE(fletcher_4_impls) + 1]; | |
223 | ||
224 | /* Indicate that benchmark has been completed */ | |
225 | static boolean_t fletcher_4_initialized = B_FALSE; | |
34dc7c2f | 226 | |
3c67d83a | 227 | /*ARGSUSED*/ |
34dc7c2f | 228 | void |
a6255b7f DQ |
229 | fletcher_init(zio_cksum_t *zcp) |
230 | { | |
231 | ZIO_SET_CHECKSUM(zcp, 0, 0, 0, 0); | |
232 | } | |
233 | ||
234 | int | |
235 | fletcher_2_incremental_native(void *buf, size_t size, void *data) | |
34dc7c2f | 236 | { |
a6255b7f DQ |
237 | zio_cksum_t *zcp = data; |
238 | ||
34dc7c2f BB |
239 | const uint64_t *ip = buf; |
240 | const uint64_t *ipend = ip + (size / sizeof (uint64_t)); | |
241 | uint64_t a0, b0, a1, b1; | |
242 | ||
a6255b7f DQ |
243 | a0 = zcp->zc_word[0]; |
244 | a1 = zcp->zc_word[1]; | |
245 | b0 = zcp->zc_word[2]; | |
246 | b1 = zcp->zc_word[3]; | |
247 | ||
248 | for (; ip < ipend; ip += 2) { | |
34dc7c2f BB |
249 | a0 += ip[0]; |
250 | a1 += ip[1]; | |
251 | b0 += a0; | |
252 | b1 += a1; | |
253 | } | |
254 | ||
255 | ZIO_SET_CHECKSUM(zcp, a0, a1, b0, b1); | |
a6255b7f | 256 | return (0); |
34dc7c2f BB |
257 | } |
258 | ||
3c67d83a | 259 | /*ARGSUSED*/ |
34dc7c2f | 260 | void |
a6255b7f | 261 | fletcher_2_native(const void *buf, uint64_t size, |
3c67d83a | 262 | const void *ctx_template, zio_cksum_t *zcp) |
34dc7c2f | 263 | { |
a6255b7f DQ |
264 | fletcher_init(zcp); |
265 | (void) fletcher_2_incremental_native((void *) buf, size, zcp); | |
266 | } | |
267 | ||
268 | int | |
269 | fletcher_2_incremental_byteswap(void *buf, size_t size, void *data) | |
270 | { | |
271 | zio_cksum_t *zcp = data; | |
272 | ||
34dc7c2f BB |
273 | const uint64_t *ip = buf; |
274 | const uint64_t *ipend = ip + (size / sizeof (uint64_t)); | |
275 | uint64_t a0, b0, a1, b1; | |
276 | ||
a6255b7f DQ |
277 | a0 = zcp->zc_word[0]; |
278 | a1 = zcp->zc_word[1]; | |
279 | b0 = zcp->zc_word[2]; | |
280 | b1 = zcp->zc_word[3]; | |
281 | ||
282 | for (; ip < ipend; ip += 2) { | |
34dc7c2f BB |
283 | a0 += BSWAP_64(ip[0]); |
284 | a1 += BSWAP_64(ip[1]); | |
285 | b0 += a0; | |
286 | b1 += a1; | |
287 | } | |
288 | ||
289 | ZIO_SET_CHECKSUM(zcp, a0, a1, b0, b1); | |
a6255b7f DQ |
290 | return (0); |
291 | } | |
292 | ||
293 | /*ARGSUSED*/ | |
294 | void | |
295 | fletcher_2_byteswap(const void *buf, uint64_t size, | |
296 | const void *ctx_template, zio_cksum_t *zcp) | |
297 | { | |
298 | fletcher_init(zcp); | |
299 | (void) fletcher_2_incremental_byteswap((void *) buf, size, zcp); | |
34dc7c2f BB |
300 | } |
301 | ||
fc897b24 | 302 | static void |
5bf703b8 | 303 | fletcher_4_scalar_init(fletcher_4_ctx_t *ctx) |
34dc7c2f | 304 | { |
5bf703b8 GN |
305 | ZIO_SET_CHECKSUM(&ctx->scalar, 0, 0, 0, 0); |
306 | } | |
307 | ||
308 | static void | |
309 | fletcher_4_scalar_fini(fletcher_4_ctx_t *ctx, zio_cksum_t *zcp) | |
310 | { | |
311 | memcpy(zcp, &ctx->scalar, sizeof (zio_cksum_t)); | |
34dc7c2f BB |
312 | } |
313 | ||
1eeb4562 | 314 | static void |
5bf703b8 GN |
315 | fletcher_4_scalar_native(fletcher_4_ctx_t *ctx, const void *buf, |
316 | uint64_t size) | |
34dc7c2f BB |
317 | { |
318 | const uint32_t *ip = buf; | |
319 | const uint32_t *ipend = ip + (size / sizeof (uint32_t)); | |
320 | uint64_t a, b, c, d; | |
321 | ||
5bf703b8 GN |
322 | a = ctx->scalar.zc_word[0]; |
323 | b = ctx->scalar.zc_word[1]; | |
324 | c = ctx->scalar.zc_word[2]; | |
325 | d = ctx->scalar.zc_word[3]; | |
1eeb4562 JX |
326 | |
327 | for (; ip < ipend; ip++) { | |
328 | a += ip[0]; | |
34dc7c2f BB |
329 | b += a; |
330 | c += b; | |
331 | d += c; | |
332 | } | |
333 | ||
5bf703b8 | 334 | ZIO_SET_CHECKSUM(&ctx->scalar, a, b, c, d); |
34dc7c2f BB |
335 | } |
336 | ||
1eeb4562 | 337 | static void |
5bf703b8 GN |
338 | fletcher_4_scalar_byteswap(fletcher_4_ctx_t *ctx, const void *buf, |
339 | uint64_t size) | |
34dc7c2f BB |
340 | { |
341 | const uint32_t *ip = buf; | |
342 | const uint32_t *ipend = ip + (size / sizeof (uint32_t)); | |
343 | uint64_t a, b, c, d; | |
344 | ||
5bf703b8 GN |
345 | a = ctx->scalar.zc_word[0]; |
346 | b = ctx->scalar.zc_word[1]; | |
347 | c = ctx->scalar.zc_word[2]; | |
348 | d = ctx->scalar.zc_word[3]; | |
34dc7c2f BB |
349 | |
350 | for (; ip < ipend; ip++) { | |
1eeb4562 | 351 | a += BSWAP_32(ip[0]); |
34dc7c2f BB |
352 | b += a; |
353 | c += b; | |
354 | d += c; | |
355 | } | |
356 | ||
5bf703b8 | 357 | ZIO_SET_CHECKSUM(&ctx->scalar, a, b, c, d); |
34dc7c2f BB |
358 | } |
359 | ||
1eeb4562 JX |
360 | static boolean_t |
361 | fletcher_4_scalar_valid(void) | |
362 | { | |
363 | return (B_TRUE); | |
364 | } | |
365 | ||
366 | int | |
367 | fletcher_4_impl_set(const char *val) | |
368 | { | |
fc897b24 GN |
369 | int err = -EINVAL; |
370 | uint32_t impl = IMPL_READ(fletcher_4_impl_chosen); | |
371 | size_t i, val_len; | |
1eeb4562 JX |
372 | |
373 | val_len = strlen(val); | |
374 | while ((val_len > 0) && !!isspace(val[val_len-1])) /* trim '\n' */ | |
375 | val_len--; | |
376 | ||
fc897b24 | 377 | /* check mandatory implementations */ |
1eeb4562 JX |
378 | for (i = 0; i < ARRAY_SIZE(fletcher_4_impl_selectors); i++) { |
379 | const char *name = fletcher_4_impl_selectors[i].fis_name; | |
380 | ||
381 | if (val_len == strlen(name) && | |
382 | strncmp(val, name, val_len) == 0) { | |
fc897b24 GN |
383 | impl = fletcher_4_impl_selectors[i].fis_sel; |
384 | err = 0; | |
1eeb4562 JX |
385 | break; |
386 | } | |
387 | } | |
1eeb4562 | 388 | |
fc897b24 GN |
389 | if (err != 0 && fletcher_4_initialized) { |
390 | /* check all supported implementations */ | |
391 | for (i = 0; i < fletcher_4_supp_impls_cnt; i++) { | |
392 | const char *name = fletcher_4_supp_impls[i]->name; | |
1eeb4562 | 393 | |
fc897b24 GN |
394 | if (val_len == strlen(name) && |
395 | strncmp(val, name, val_len) == 0) { | |
396 | impl = i; | |
397 | err = 0; | |
398 | break; | |
399 | } | |
400 | } | |
401 | } | |
1eeb4562 | 402 | |
fc897b24 GN |
403 | if (err == 0) { |
404 | atomic_swap_32(&fletcher_4_impl_chosen, impl); | |
405 | membar_producer(); | |
406 | } | |
407 | ||
408 | return (err); | |
1eeb4562 JX |
409 | } |
410 | ||
411 | static inline const fletcher_4_ops_t * | |
412 | fletcher_4_impl_get(void) | |
413 | { | |
fc897b24 GN |
414 | fletcher_4_ops_t *ops = NULL; |
415 | const uint32_t impl = IMPL_READ(fletcher_4_impl_chosen); | |
416 | ||
417 | switch (impl) { | |
418 | case IMPL_FASTEST: | |
419 | ASSERT(fletcher_4_initialized); | |
420 | ops = &fletcher_4_fastest_impl; | |
421 | break; | |
1eeb4562 | 422 | #if !defined(_KERNEL) |
fc897b24 GN |
423 | case IMPL_CYCLE: { |
424 | ASSERT(fletcher_4_initialized); | |
425 | ASSERT3U(fletcher_4_supp_impls_cnt, >, 0); | |
426 | ||
427 | static uint32_t cycle_count = 0; | |
428 | uint32_t idx = (++cycle_count) % fletcher_4_supp_impls_cnt; | |
429 | ops = fletcher_4_supp_impls[idx]; | |
1eeb4562 | 430 | } |
fc897b24 | 431 | break; |
1eeb4562 | 432 | #endif |
fc897b24 GN |
433 | default: |
434 | ASSERT3U(fletcher_4_supp_impls_cnt, >, 0); | |
435 | ASSERT3U(impl, <, fletcher_4_supp_impls_cnt); | |
436 | ||
437 | ops = fletcher_4_supp_impls[impl]; | |
438 | break; | |
439 | } | |
440 | ||
441 | ASSERT3P(ops, !=, NULL); | |
442 | ||
443 | return (ops); | |
444 | } | |
445 | ||
fc897b24 | 446 | static inline void |
5bf703b8 | 447 | fletcher_4_native_impl(const void *buf, uint64_t size, zio_cksum_t *zcp) |
fc897b24 | 448 | { |
5bf703b8 GN |
449 | fletcher_4_ctx_t ctx; |
450 | const fletcher_4_ops_t *ops = fletcher_4_impl_get(); | |
451 | ||
452 | ops->init_native(&ctx); | |
453 | ops->compute_native(&ctx, buf, size); | |
454 | ops->fini_native(&ctx, zcp); | |
1eeb4562 JX |
455 | } |
456 | ||
3c67d83a | 457 | /*ARGSUSED*/ |
1eeb4562 | 458 | void |
3c67d83a TH |
459 | fletcher_4_native(const void *buf, uint64_t size, |
460 | const void *ctx_template, zio_cksum_t *zcp) | |
1eeb4562 | 461 | { |
2fe36b0b | 462 | const uint64_t p2size = P2ALIGN(size, FLETCHER_MIN_SIMD_SIZE); |
0dab2e84 | 463 | |
fc897b24 GN |
464 | ASSERT(IS_P2ALIGNED(size, sizeof (uint32_t))); |
465 | ||
5bf703b8 | 466 | if (size == 0 || p2size == 0) { |
fc897b24 | 467 | ZIO_SET_CHECKSUM(zcp, 0, 0, 0, 0); |
5bf703b8 GN |
468 | |
469 | if (size > 0) | |
470 | fletcher_4_scalar_native((fletcher_4_ctx_t *)zcp, | |
471 | buf, size); | |
fc897b24 | 472 | } else { |
5bf703b8 | 473 | fletcher_4_native_impl(buf, p2size, zcp); |
1eeb4562 | 474 | |
fc897b24 | 475 | if (p2size < size) |
5bf703b8 GN |
476 | fletcher_4_scalar_native((fletcher_4_ctx_t *)zcp, |
477 | (char *)buf + p2size, size - p2size); | |
fc897b24 GN |
478 | } |
479 | } | |
480 | ||
481 | void | |
482 | fletcher_4_native_varsize(const void *buf, uint64_t size, zio_cksum_t *zcp) | |
483 | { | |
5bf703b8 GN |
484 | ZIO_SET_CHECKSUM(zcp, 0, 0, 0, 0); |
485 | fletcher_4_scalar_native((fletcher_4_ctx_t *)zcp, buf, size); | |
fc897b24 GN |
486 | } |
487 | ||
488 | static inline void | |
5bf703b8 | 489 | fletcher_4_byteswap_impl(const void *buf, uint64_t size, zio_cksum_t *zcp) |
fc897b24 | 490 | { |
5bf703b8 GN |
491 | fletcher_4_ctx_t ctx; |
492 | const fletcher_4_ops_t *ops = fletcher_4_impl_get(); | |
493 | ||
494 | ops->init_byteswap(&ctx); | |
495 | ops->compute_byteswap(&ctx, buf, size); | |
496 | ops->fini_byteswap(&ctx, zcp); | |
1eeb4562 JX |
497 | } |
498 | ||
3c67d83a | 499 | /*ARGSUSED*/ |
1eeb4562 | 500 | void |
3c67d83a TH |
501 | fletcher_4_byteswap(const void *buf, uint64_t size, |
502 | const void *ctx_template, zio_cksum_t *zcp) | |
1eeb4562 | 503 | { |
2fe36b0b | 504 | const uint64_t p2size = P2ALIGN(size, FLETCHER_MIN_SIMD_SIZE); |
0dab2e84 | 505 | |
fc897b24 GN |
506 | ASSERT(IS_P2ALIGNED(size, sizeof (uint32_t))); |
507 | ||
5bf703b8 | 508 | if (size == 0 || p2size == 0) { |
fc897b24 | 509 | ZIO_SET_CHECKSUM(zcp, 0, 0, 0, 0); |
5bf703b8 GN |
510 | |
511 | if (size > 0) | |
512 | fletcher_4_scalar_byteswap((fletcher_4_ctx_t *)zcp, | |
513 | buf, size); | |
fc897b24 | 514 | } else { |
5bf703b8 | 515 | fletcher_4_byteswap_impl(buf, p2size, zcp); |
1eeb4562 | 516 | |
fc897b24 | 517 | if (p2size < size) |
5bf703b8 GN |
518 | fletcher_4_scalar_byteswap((fletcher_4_ctx_t *)zcp, |
519 | (char *)buf + p2size, size - p2size); | |
fc897b24 | 520 | } |
1eeb4562 JX |
521 | } |
522 | ||
37f520db GN |
523 | /* Incremental Fletcher 4 */ |
524 | ||
5bf703b8 GN |
525 | #define ZFS_FLETCHER_4_INC_MAX_SIZE (8ULL << 20) |
526 | ||
37f520db GN |
527 | static inline void |
528 | fletcher_4_incremental_combine(zio_cksum_t *zcp, const uint64_t size, | |
529 | const zio_cksum_t *nzcp) | |
530 | { | |
531 | const uint64_t c1 = size / sizeof (uint32_t); | |
532 | const uint64_t c2 = c1 * (c1 + 1) / 2; | |
533 | const uint64_t c3 = c2 * (c1 + 2) / 3; | |
534 | ||
5bf703b8 GN |
535 | /* |
536 | * Value of 'c3' overflows on buffer sizes close to 16MiB. For that | |
537 | * reason we split incremental fletcher4 computation of large buffers | |
538 | * to steps of (ZFS_FLETCHER_4_INC_MAX_SIZE) size. | |
539 | */ | |
540 | ASSERT3U(size, <=, ZFS_FLETCHER_4_INC_MAX_SIZE); | |
541 | ||
37f520db GN |
542 | zcp->zc_word[3] += nzcp->zc_word[3] + c1 * zcp->zc_word[2] + |
543 | c2 * zcp->zc_word[1] + c3 * zcp->zc_word[0]; | |
544 | zcp->zc_word[2] += nzcp->zc_word[2] + c1 * zcp->zc_word[1] + | |
545 | c2 * zcp->zc_word[0]; | |
546 | zcp->zc_word[1] += nzcp->zc_word[1] + c1 * zcp->zc_word[0]; | |
547 | zcp->zc_word[0] += nzcp->zc_word[0]; | |
548 | } | |
549 | ||
550 | static inline void | |
551 | fletcher_4_incremental_impl(boolean_t native, const void *buf, uint64_t size, | |
552 | zio_cksum_t *zcp) | |
553 | { | |
37f520db GN |
554 | while (size > 0) { |
555 | zio_cksum_t nzc; | |
5bf703b8 | 556 | uint64_t len = MIN(size, ZFS_FLETCHER_4_INC_MAX_SIZE); |
37f520db GN |
557 | |
558 | if (native) | |
559 | fletcher_4_native(buf, len, NULL, &nzc); | |
560 | else | |
561 | fletcher_4_byteswap(buf, len, NULL, &nzc); | |
562 | ||
563 | fletcher_4_incremental_combine(zcp, len, &nzc); | |
564 | ||
565 | size -= len; | |
566 | buf += len; | |
567 | } | |
568 | } | |
569 | ||
a6255b7f DQ |
570 | int |
571 | fletcher_4_incremental_native(void *buf, size_t size, void *data) | |
37f520db | 572 | { |
a6255b7f | 573 | zio_cksum_t *zcp = data; |
5bf703b8 GN |
574 | /* Use scalar impl to directly update cksum of small blocks */ |
575 | if (size < SPA_MINBLOCKSIZE) | |
576 | fletcher_4_scalar_native((fletcher_4_ctx_t *)zcp, buf, size); | |
577 | else | |
578 | fletcher_4_incremental_impl(B_TRUE, buf, size, zcp); | |
a6255b7f | 579 | return (0); |
37f520db GN |
580 | } |
581 | ||
a6255b7f DQ |
582 | int |
583 | fletcher_4_incremental_byteswap(void *buf, size_t size, void *data) | |
37f520db | 584 | { |
a6255b7f | 585 | zio_cksum_t *zcp = data; |
5bf703b8 GN |
586 | /* Use scalar impl to directly update cksum of small blocks */ |
587 | if (size < SPA_MINBLOCKSIZE) | |
588 | fletcher_4_scalar_byteswap((fletcher_4_ctx_t *)zcp, buf, size); | |
589 | else | |
590 | fletcher_4_incremental_impl(B_FALSE, buf, size, zcp); | |
a6255b7f | 591 | return (0); |
37f520db GN |
592 | } |
593 | ||
46364cb2 | 594 | #if defined(_KERNEL) |
37f520db GN |
595 | /* Fletcher 4 kstats */ |
596 | ||
fc897b24 GN |
597 | static int |
598 | fletcher_4_kstat_headers(char *buf, size_t size) | |
1eeb4562 | 599 | { |
fc897b24 GN |
600 | ssize_t off = 0; |
601 | ||
602 | off += snprintf(buf + off, size, "%-17s", "implementation"); | |
603 | off += snprintf(buf + off, size - off, "%-15s", "native"); | |
604 | (void) snprintf(buf + off, size - off, "%-15s\n", "byteswap"); | |
605 | ||
606 | return (0); | |
1eeb4562 JX |
607 | } |
608 | ||
fc897b24 GN |
609 | static int |
610 | fletcher_4_kstat_data(char *buf, size_t size, void *data) | |
34dc7c2f | 611 | { |
fc897b24 GN |
612 | struct fletcher_4_kstat *fastest_stat = |
613 | &fletcher_4_stat_data[fletcher_4_supp_impls_cnt]; | |
02730c33 | 614 | struct fletcher_4_kstat *curr_stat = (struct fletcher_4_kstat *)data; |
fc897b24 GN |
615 | ssize_t off = 0; |
616 | ||
617 | if (curr_stat == fastest_stat) { | |
618 | off += snprintf(buf + off, size - off, "%-17s", "fastest"); | |
619 | off += snprintf(buf + off, size - off, "%-15s", | |
620 | fletcher_4_supp_impls[fastest_stat->native]->name); | |
621 | off += snprintf(buf + off, size - off, "%-15s\n", | |
622 | fletcher_4_supp_impls[fastest_stat->byteswap]->name); | |
623 | } else { | |
624 | ptrdiff_t id = curr_stat - fletcher_4_stat_data; | |
625 | ||
626 | off += snprintf(buf + off, size - off, "%-17s", | |
627 | fletcher_4_supp_impls[id]->name); | |
628 | off += snprintf(buf + off, size - off, "%-15llu", | |
02730c33 | 629 | (u_longlong_t)curr_stat->native); |
fc897b24 | 630 | off += snprintf(buf + off, size - off, "%-15llu\n", |
02730c33 | 631 | (u_longlong_t)curr_stat->byteswap); |
fc897b24 GN |
632 | } |
633 | ||
634 | return (0); | |
1eeb4562 | 635 | } |
34dc7c2f | 636 | |
fc897b24 GN |
637 | static void * |
638 | fletcher_4_kstat_addr(kstat_t *ksp, loff_t n) | |
1eeb4562 | 639 | { |
fc897b24 GN |
640 | if (n <= fletcher_4_supp_impls_cnt) |
641 | ksp->ks_private = (void *) (fletcher_4_stat_data + n); | |
642 | else | |
643 | ksp->ks_private = NULL; | |
644 | ||
645 | return (ksp->ks_private); | |
646 | } | |
46364cb2 | 647 | #endif |
fc897b24 GN |
648 | |
649 | #define FLETCHER_4_FASTEST_FN_COPY(type, src) \ | |
650 | { \ | |
651 | fletcher_4_fastest_impl.init_ ## type = src->init_ ## type; \ | |
652 | fletcher_4_fastest_impl.fini_ ## type = src->fini_ ## type; \ | |
653 | fletcher_4_fastest_impl.compute_ ## type = src->compute_ ## type; \ | |
654 | } | |
655 | ||
656 | #define FLETCHER_4_BENCH_NS (MSEC2NSEC(50)) /* 50ms */ | |
34dc7c2f | 657 | |
a6255b7f DQ |
658 | typedef void fletcher_checksum_func_t(const void *, uint64_t, const void *, |
659 | zio_cksum_t *); | |
660 | ||
fc897b24 GN |
661 | static void |
662 | fletcher_4_benchmark_impl(boolean_t native, char *data, uint64_t data_size) | |
663 | { | |
664 | ||
665 | struct fletcher_4_kstat *fastest_stat = | |
666 | &fletcher_4_stat_data[fletcher_4_supp_impls_cnt]; | |
667 | hrtime_t start; | |
668 | uint64_t run_bw, run_time_ns, best_run = 0; | |
669 | zio_cksum_t zc; | |
670 | uint32_t i, l, sel_save = IMPL_READ(fletcher_4_impl_chosen); | |
671 | ||
a6255b7f DQ |
672 | |
673 | fletcher_checksum_func_t *fletcher_4_test = native ? | |
674 | fletcher_4_native : fletcher_4_byteswap; | |
1eeb4562 | 675 | |
fc897b24 GN |
676 | for (i = 0; i < fletcher_4_supp_impls_cnt; i++) { |
677 | struct fletcher_4_kstat *stat = &fletcher_4_stat_data[i]; | |
678 | uint64_t run_count = 0; | |
1eeb4562 | 679 | |
fc897b24 GN |
680 | /* temporary set an implementation */ |
681 | fletcher_4_impl_chosen = i; | |
1eeb4562 JX |
682 | |
683 | kpreempt_disable(); | |
684 | start = gethrtime(); | |
1eeb4562 | 685 | do { |
fc897b24 | 686 | for (l = 0; l < 32; l++, run_count++) |
3c67d83a | 687 | fletcher_4_test(data, data_size, NULL, &zc); |
fc897b24 GN |
688 | |
689 | run_time_ns = gethrtime() - start; | |
690 | } while (run_time_ns < FLETCHER_4_BENCH_NS); | |
1eeb4562 JX |
691 | kpreempt_enable(); |
692 | ||
fc897b24 GN |
693 | run_bw = data_size * run_count * NANOSEC; |
694 | run_bw /= run_time_ns; /* B/s */ | |
695 | ||
696 | if (native) | |
697 | stat->native = run_bw; | |
698 | else | |
699 | stat->byteswap = run_bw; | |
700 | ||
701 | if (run_bw > best_run) { | |
702 | best_run = run_bw; | |
703 | ||
704 | if (native) { | |
705 | fastest_stat->native = i; | |
706 | FLETCHER_4_FASTEST_FN_COPY(native, | |
707 | fletcher_4_supp_impls[i]); | |
708 | } else { | |
709 | fastest_stat->byteswap = i; | |
710 | FLETCHER_4_FASTEST_FN_COPY(byteswap, | |
711 | fletcher_4_supp_impls[i]); | |
712 | } | |
1eeb4562 | 713 | } |
fc897b24 GN |
714 | } |
715 | ||
716 | /* restore original selection */ | |
717 | atomic_swap_32(&fletcher_4_impl_chosen, sel_save); | |
718 | } | |
1eeb4562 | 719 | |
fc897b24 GN |
720 | void |
721 | fletcher_4_init(void) | |
722 | { | |
723 | static const size_t data_size = 1 << SPA_OLD_MAXBLOCKSHIFT; /* 128kiB */ | |
724 | fletcher_4_ops_t *curr_impl; | |
725 | char *databuf; | |
726 | int i, c; | |
727 | ||
728 | /* move supported impl into fletcher_4_supp_impls */ | |
729 | for (i = 0, c = 0; i < ARRAY_SIZE(fletcher_4_impls); i++) { | |
02730c33 | 730 | curr_impl = (fletcher_4_ops_t *)fletcher_4_impls[i]; |
fc897b24 GN |
731 | |
732 | if (curr_impl->valid && curr_impl->valid()) | |
733 | fletcher_4_supp_impls[c++] = curr_impl; | |
34dc7c2f | 734 | } |
fc897b24 GN |
735 | membar_producer(); /* complete fletcher_4_supp_impls[] init */ |
736 | fletcher_4_supp_impls_cnt = c; /* number of supported impl */ | |
34dc7c2f | 737 | |
fc897b24 GN |
738 | #if !defined(_KERNEL) |
739 | /* Skip benchmarking and use last implementation as fastest */ | |
740 | memcpy(&fletcher_4_fastest_impl, | |
741 | fletcher_4_supp_impls[fletcher_4_supp_impls_cnt-1], | |
742 | sizeof (fletcher_4_fastest_impl)); | |
743 | fletcher_4_fastest_impl.name = "fastest"; | |
744 | membar_producer(); | |
1eeb4562 | 745 | |
fc897b24 | 746 | fletcher_4_initialized = B_TRUE; |
fc897b24 GN |
747 | return; |
748 | #endif | |
749 | /* Benchmark all supported implementations */ | |
750 | databuf = vmem_alloc(data_size, KM_SLEEP); | |
751 | for (i = 0; i < data_size / sizeof (uint64_t); i++) | |
752 | ((uint64_t *)databuf)[i] = (uintptr_t)(databuf+i); /* warm-up */ | |
753 | ||
754 | fletcher_4_benchmark_impl(B_FALSE, databuf, data_size); | |
755 | fletcher_4_benchmark_impl(B_TRUE, databuf, data_size); | |
756 | ||
757 | vmem_free(databuf, data_size); | |
758 | ||
46364cb2 | 759 | #if defined(_KERNEL) |
fc897b24 GN |
760 | /* install kstats for all implementations */ |
761 | fletcher_4_kstat = kstat_create("zfs", 0, "fletcher_4_bench", "misc", | |
02730c33 | 762 | KSTAT_TYPE_RAW, 0, KSTAT_FLAG_VIRTUAL); |
1eeb4562 | 763 | if (fletcher_4_kstat != NULL) { |
fc897b24 GN |
764 | fletcher_4_kstat->ks_data = NULL; |
765 | fletcher_4_kstat->ks_ndata = UINT32_MAX; | |
766 | kstat_set_raw_ops(fletcher_4_kstat, | |
767 | fletcher_4_kstat_headers, | |
768 | fletcher_4_kstat_data, | |
769 | fletcher_4_kstat_addr); | |
1eeb4562 JX |
770 | kstat_install(fletcher_4_kstat); |
771 | } | |
46364cb2 | 772 | #endif |
fc897b24 GN |
773 | |
774 | /* Finish initialization */ | |
775 | fletcher_4_initialized = B_TRUE; | |
1eeb4562 JX |
776 | } |
777 | ||
778 | void | |
779 | fletcher_4_fini(void) | |
780 | { | |
46364cb2 | 781 | #if defined(_KERNEL) |
1eeb4562 JX |
782 | if (fletcher_4_kstat != NULL) { |
783 | kstat_delete(fletcher_4_kstat); | |
784 | fletcher_4_kstat = NULL; | |
785 | } | |
46364cb2 | 786 | #endif |
34dc7c2f | 787 | } |
c28b2279 | 788 | |
2fe36b0b DQ |
789 | /* ABD adapters */ |
790 | ||
791 | static void | |
792 | abd_fletcher_4_init(zio_abd_checksum_data_t *cdp) | |
793 | { | |
794 | const fletcher_4_ops_t *ops = fletcher_4_impl_get(); | |
795 | cdp->acd_private = (void *) ops; | |
796 | ||
797 | if (cdp->acd_byteorder == ZIO_CHECKSUM_NATIVE) | |
798 | ops->init_native(cdp->acd_ctx); | |
799 | else | |
800 | ops->init_byteswap(cdp->acd_ctx); | |
801 | } | |
802 | ||
803 | static void | |
804 | abd_fletcher_4_fini(zio_abd_checksum_data_t *cdp) | |
805 | { | |
806 | fletcher_4_ops_t *ops = (fletcher_4_ops_t *)cdp->acd_private; | |
807 | ||
808 | ASSERT(ops); | |
809 | ||
810 | if (cdp->acd_byteorder == ZIO_CHECKSUM_NATIVE) | |
811 | ops->fini_native(cdp->acd_ctx, cdp->acd_zcp); | |
812 | else | |
813 | ops->fini_byteswap(cdp->acd_ctx, cdp->acd_zcp); | |
814 | } | |
815 | ||
816 | static void | |
817 | abd_fletcher_4_simd2scalar(boolean_t native, void *data, size_t size, | |
818 | zio_abd_checksum_data_t *cdp) | |
819 | { | |
820 | zio_cksum_t *zcp = cdp->acd_zcp; | |
821 | ||
822 | ASSERT3U(size, <, FLETCHER_MIN_SIMD_SIZE); | |
823 | ||
824 | abd_fletcher_4_fini(cdp); | |
825 | cdp->acd_private = (void *)&fletcher_4_scalar_ops; | |
826 | ||
827 | if (native) | |
828 | fletcher_4_incremental_native(data, size, zcp); | |
829 | else | |
830 | fletcher_4_incremental_byteswap(data, size, zcp); | |
831 | } | |
832 | ||
833 | static int | |
834 | abd_fletcher_4_iter(void *data, size_t size, void *private) | |
835 | { | |
836 | zio_abd_checksum_data_t *cdp = (zio_abd_checksum_data_t *)private; | |
837 | fletcher_4_ctx_t *ctx = cdp->acd_ctx; | |
838 | fletcher_4_ops_t *ops = (fletcher_4_ops_t *)cdp->acd_private; | |
839 | boolean_t native = cdp->acd_byteorder == ZIO_CHECKSUM_NATIVE; | |
840 | uint64_t asize = P2ALIGN(size, FLETCHER_MIN_SIMD_SIZE); | |
841 | ||
842 | ASSERT(IS_P2ALIGNED(size, sizeof (uint32_t))); | |
843 | ||
844 | if (asize > 0) { | |
845 | if (native) | |
846 | ops->compute_native(ctx, data, asize); | |
847 | else | |
848 | ops->compute_byteswap(ctx, data, asize); | |
849 | ||
850 | size -= asize; | |
851 | data = (char *)data + asize; | |
852 | } | |
853 | ||
854 | if (size > 0) { | |
855 | ASSERT3U(size, <, FLETCHER_MIN_SIMD_SIZE); | |
856 | /* At this point we have to switch to scalar impl */ | |
857 | abd_fletcher_4_simd2scalar(native, data, size, cdp); | |
858 | } | |
859 | ||
860 | return (0); | |
861 | } | |
862 | ||
863 | zio_abd_checksum_func_t fletcher_4_abd_ops = { | |
864 | .acf_init = abd_fletcher_4_init, | |
865 | .acf_fini = abd_fletcher_4_fini, | |
866 | .acf_iter = abd_fletcher_4_iter | |
867 | }; | |
868 | ||
869 | ||
93ce2b4c | 870 | #if defined(_KERNEL) |
9cc1844a | 871 | #include <linux/mod_compat.h> |
1eeb4562 JX |
872 | |
873 | static int | |
9cc1844a | 874 | fletcher_4_param_get(char *buffer, zfs_kernel_param_t *unused) |
1eeb4562 | 875 | { |
fc897b24 GN |
876 | const uint32_t impl = IMPL_READ(fletcher_4_impl_chosen); |
877 | char *fmt; | |
1eeb4562 JX |
878 | int i, cnt = 0; |
879 | ||
fc897b24 GN |
880 | /* list fastest */ |
881 | fmt = (impl == IMPL_FASTEST) ? "[%s] " : "%s "; | |
882 | cnt += sprintf(buffer + cnt, fmt, "fastest"); | |
1eeb4562 | 883 | |
fc897b24 GN |
884 | /* list all supported implementations */ |
885 | for (i = 0; i < fletcher_4_supp_impls_cnt; i++) { | |
886 | fmt = (i == impl) ? "[%s] " : "%s "; | |
887 | cnt += sprintf(buffer + cnt, fmt, | |
888 | fletcher_4_supp_impls[i]->name); | |
1eeb4562 JX |
889 | } |
890 | ||
891 | return (cnt); | |
892 | } | |
893 | ||
894 | static int | |
9cc1844a | 895 | fletcher_4_param_set(const char *val, zfs_kernel_param_t *unused) |
1eeb4562 JX |
896 | { |
897 | return (fletcher_4_impl_set(val)); | |
898 | } | |
899 | ||
900 | /* | |
901 | * Choose a fletcher 4 implementation in ZFS. | |
fc897b24 | 902 | * Users can choose "cycle" to exercise all implementations, but this is |
1eeb4562 JX |
903 | * for testing purpose therefore it can only be set in user space. |
904 | */ | |
905 | module_param_call(zfs_fletcher_4_impl, | |
906 | fletcher_4_param_set, fletcher_4_param_get, NULL, 0644); | |
fc897b24 | 907 | MODULE_PARM_DESC(zfs_fletcher_4_impl, "Select fletcher 4 implementation."); |
1eeb4562 | 908 | |
a6255b7f DQ |
909 | EXPORT_SYMBOL(fletcher_init); |
910 | EXPORT_SYMBOL(fletcher_2_incremental_native); | |
911 | EXPORT_SYMBOL(fletcher_2_incremental_byteswap); | |
1eeb4562 JX |
912 | EXPORT_SYMBOL(fletcher_4_init); |
913 | EXPORT_SYMBOL(fletcher_4_fini); | |
c28b2279 BB |
914 | EXPORT_SYMBOL(fletcher_2_native); |
915 | EXPORT_SYMBOL(fletcher_2_byteswap); | |
916 | EXPORT_SYMBOL(fletcher_4_native); | |
fc897b24 | 917 | EXPORT_SYMBOL(fletcher_4_native_varsize); |
c28b2279 BB |
918 | EXPORT_SYMBOL(fletcher_4_byteswap); |
919 | EXPORT_SYMBOL(fletcher_4_incremental_native); | |
920 | EXPORT_SYMBOL(fletcher_4_incremental_byteswap); | |
2fe36b0b | 921 | EXPORT_SYMBOL(fletcher_4_abd_ops); |
c28b2279 | 922 | #endif |