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.
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.
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]
22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
24 * Copyright (C) 2016 Gvozden Nešković. All rights reserved.
27 * Copyright 2013 Saso Kiselkov. All rights reserved.
31 * Copyright (c) 2016 by Delphix. All rights reserved.
38 * ZFS's 2nd and 4th order Fletcher checksums are defined by the following
39 * recurrence relations:
47 * c = c + b (fletcher-4 only)
50 * d = d + c (fletcher-4 only)
54 * a_0 = b_0 = c_0 = d_0 = 0
56 * f_0 .. f_(n-1) are the input data.
58 * Using standard techniques, these translate into the following series:
63 * n /___| n - i n /___| n - i
68 * \ | i*(i+1) \ | i*(i+1)*(i+2)
69 * c = > ------- f d = > ------------- f
70 * n /___| 2 n - i n /___| 6 n - i
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.
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:
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
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:
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
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.
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.
114 * --------------------
115 * Checksum Performance
116 * --------------------
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.
125 * In normal operation, the data which is being checksummed is in a buffer
126 * which has been filled either by:
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).
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.
136 #include <sys/types.h>
137 #include <sys/sysmacros.h>
138 #include <sys/byteorder.h>
140 #include <sys/simd.h>
141 #include <sys/zio_checksum.h>
142 #include <sys/zfs_context.h>
143 #include <zfs_fletcher.h>
145 #define FLETCHER_MIN_SIMD_SIZE 64
147 static void fletcher_4_scalar_init(fletcher_4_ctx_t *ctx);
148 static void fletcher_4_scalar_fini(fletcher_4_ctx_t *ctx, zio_cksum_t *zcp);
149 static void fletcher_4_scalar_native(fletcher_4_ctx_t *ctx,
150 const void *buf, uint64_t size);
151 static void fletcher_4_scalar_byteswap(fletcher_4_ctx_t *ctx,
152 const void *buf, uint64_t size);
153 static boolean_t fletcher_4_scalar_valid(void);
155 static const fletcher_4_ops_t fletcher_4_scalar_ops = {
156 .init_native = fletcher_4_scalar_init,
157 .fini_native = fletcher_4_scalar_fini,
158 .compute_native = fletcher_4_scalar_native,
159 .init_byteswap = fletcher_4_scalar_init,
160 .fini_byteswap = fletcher_4_scalar_fini,
161 .compute_byteswap = fletcher_4_scalar_byteswap,
162 .valid = fletcher_4_scalar_valid,
166 static fletcher_4_ops_t fletcher_4_fastest_impl = {
168 .valid = fletcher_4_scalar_valid
171 static const fletcher_4_ops_t *fletcher_4_impls[] = {
172 &fletcher_4_scalar_ops,
173 &fletcher_4_superscalar_ops,
174 &fletcher_4_superscalar4_ops,
175 #if defined(HAVE_SSE2)
176 &fletcher_4_sse2_ops,
178 #if defined(HAVE_SSE2) && defined(HAVE_SSSE3)
179 &fletcher_4_ssse3_ops,
181 #if defined(HAVE_AVX) && defined(HAVE_AVX2)
182 &fletcher_4_avx2_ops,
184 #if defined(__x86_64) && defined(HAVE_AVX512F)
185 &fletcher_4_avx512f_ops,
187 #if defined(__x86_64) && defined(HAVE_AVX512BW)
188 &fletcher_4_avx512bw_ops,
190 #if defined(__aarch64__) && !defined(__FreeBSD__)
191 &fletcher_4_aarch64_neon_ops,
195 /* Hold all supported implementations */
196 static uint32_t fletcher_4_supp_impls_cnt = 0;
197 static fletcher_4_ops_t *fletcher_4_supp_impls[ARRAY_SIZE(fletcher_4_impls)];
199 /* Select fletcher4 implementation */
200 #define IMPL_FASTEST (UINT32_MAX)
201 #define IMPL_CYCLE (UINT32_MAX - 1)
202 #define IMPL_SCALAR (0)
204 static uint32_t fletcher_4_impl_chosen = IMPL_FASTEST;
206 #define IMPL_READ(i) (*(volatile uint32_t *) &(i))
208 static struct fletcher_4_impl_selector {
209 const char *fis_name;
211 } fletcher_4_impl_selectors[] = {
212 { "cycle", IMPL_CYCLE },
213 { "fastest", IMPL_FASTEST },
214 { "scalar", IMPL_SCALAR }
218 static kstat_t *fletcher_4_kstat;
220 static struct fletcher_4_kstat {
223 } fletcher_4_stat_data[ARRAY_SIZE(fletcher_4_impls) + 1];
226 /* Indicate that benchmark has been completed */
227 static boolean_t fletcher_4_initialized = B_FALSE;
230 fletcher_init(zio_cksum_t *zcp)
232 ZIO_SET_CHECKSUM(zcp, 0, 0, 0, 0);
236 fletcher_2_incremental_native(void *buf, size_t size, void *data)
238 zio_cksum_t *zcp = data;
240 const uint64_t *ip = buf;
241 const uint64_t *ipend = ip + (size / sizeof (uint64_t));
242 uint64_t a0, b0, a1, b1;
244 a0 = zcp->zc_word[0];
245 a1 = zcp->zc_word[1];
246 b0 = zcp->zc_word[2];
247 b1 = zcp->zc_word[3];
249 for (; ip < ipend; ip += 2) {
256 ZIO_SET_CHECKSUM(zcp, a0, a1, b0, b1);
261 fletcher_2_native(const void *buf, uint64_t size,
262 const void *ctx_template, zio_cksum_t *zcp)
266 (void) fletcher_2_incremental_native((void *) buf, size, zcp);
270 fletcher_2_incremental_byteswap(void *buf, size_t size, void *data)
272 zio_cksum_t *zcp = data;
274 const uint64_t *ip = buf;
275 const uint64_t *ipend = ip + (size / sizeof (uint64_t));
276 uint64_t a0, b0, a1, b1;
278 a0 = zcp->zc_word[0];
279 a1 = zcp->zc_word[1];
280 b0 = zcp->zc_word[2];
281 b1 = zcp->zc_word[3];
283 for (; ip < ipend; ip += 2) {
284 a0 += BSWAP_64(ip[0]);
285 a1 += BSWAP_64(ip[1]);
290 ZIO_SET_CHECKSUM(zcp, a0, a1, b0, b1);
295 fletcher_2_byteswap(const void *buf, uint64_t size,
296 const void *ctx_template, zio_cksum_t *zcp)
300 (void) fletcher_2_incremental_byteswap((void *) buf, size, zcp);
304 fletcher_4_scalar_init(fletcher_4_ctx_t *ctx)
306 ZIO_SET_CHECKSUM(&ctx->scalar, 0, 0, 0, 0);
310 fletcher_4_scalar_fini(fletcher_4_ctx_t *ctx, zio_cksum_t *zcp)
312 memcpy(zcp, &ctx->scalar, sizeof (zio_cksum_t));
316 fletcher_4_scalar_native(fletcher_4_ctx_t *ctx, const void *buf,
319 const uint32_t *ip = buf;
320 const uint32_t *ipend = ip + (size / sizeof (uint32_t));
323 a = ctx->scalar.zc_word[0];
324 b = ctx->scalar.zc_word[1];
325 c = ctx->scalar.zc_word[2];
326 d = ctx->scalar.zc_word[3];
328 for (; ip < ipend; ip++) {
335 ZIO_SET_CHECKSUM(&ctx->scalar, a, b, c, d);
339 fletcher_4_scalar_byteswap(fletcher_4_ctx_t *ctx, const void *buf,
342 const uint32_t *ip = buf;
343 const uint32_t *ipend = ip + (size / sizeof (uint32_t));
346 a = ctx->scalar.zc_word[0];
347 b = ctx->scalar.zc_word[1];
348 c = ctx->scalar.zc_word[2];
349 d = ctx->scalar.zc_word[3];
351 for (; ip < ipend; ip++) {
352 a += BSWAP_32(ip[0]);
358 ZIO_SET_CHECKSUM(&ctx->scalar, a, b, c, d);
362 fletcher_4_scalar_valid(void)
368 fletcher_4_impl_set(const char *val)
371 uint32_t impl = IMPL_READ(fletcher_4_impl_chosen);
374 val_len = strlen(val);
375 while ((val_len > 0) && !!isspace(val[val_len-1])) /* trim '\n' */
378 /* check mandatory implementations */
379 for (i = 0; i < ARRAY_SIZE(fletcher_4_impl_selectors); i++) {
380 const char *name = fletcher_4_impl_selectors[i].fis_name;
382 if (val_len == strlen(name) &&
383 strncmp(val, name, val_len) == 0) {
384 impl = fletcher_4_impl_selectors[i].fis_sel;
390 if (err != 0 && fletcher_4_initialized) {
391 /* check all supported implementations */
392 for (i = 0; i < fletcher_4_supp_impls_cnt; i++) {
393 const char *name = fletcher_4_supp_impls[i]->name;
395 if (val_len == strlen(name) &&
396 strncmp(val, name, val_len) == 0) {
405 atomic_swap_32(&fletcher_4_impl_chosen, impl);
413 * Returns the Fletcher 4 operations for checksums. When a SIMD
414 * implementation is not allowed in the current context, then fallback
415 * to the fastest generic implementation.
417 static inline const fletcher_4_ops_t *
418 fletcher_4_impl_get(void)
421 return (&fletcher_4_superscalar4_ops);
423 const fletcher_4_ops_t *ops = NULL;
424 uint32_t impl = IMPL_READ(fletcher_4_impl_chosen);
428 ASSERT(fletcher_4_initialized);
429 ops = &fletcher_4_fastest_impl;
432 /* Cycle through supported implementations */
433 ASSERT(fletcher_4_initialized);
434 ASSERT3U(fletcher_4_supp_impls_cnt, >, 0);
435 static uint32_t cycle_count = 0;
436 uint32_t idx = (++cycle_count) % fletcher_4_supp_impls_cnt;
437 ops = fletcher_4_supp_impls[idx];
440 ASSERT3U(fletcher_4_supp_impls_cnt, >, 0);
441 ASSERT3U(impl, <, fletcher_4_supp_impls_cnt);
442 ops = fletcher_4_supp_impls[impl];
446 ASSERT3P(ops, !=, NULL);
452 fletcher_4_native_impl(const void *buf, uint64_t size, zio_cksum_t *zcp)
454 fletcher_4_ctx_t ctx;
455 const fletcher_4_ops_t *ops = fletcher_4_impl_get();
457 ops->init_native(&ctx);
458 ops->compute_native(&ctx, buf, size);
459 ops->fini_native(&ctx, zcp);
463 fletcher_4_native(const void *buf, uint64_t size,
464 const void *ctx_template, zio_cksum_t *zcp)
467 const uint64_t p2size = P2ALIGN(size, FLETCHER_MIN_SIMD_SIZE);
469 ASSERT(IS_P2ALIGNED(size, sizeof (uint32_t)));
471 if (size == 0 || p2size == 0) {
472 ZIO_SET_CHECKSUM(zcp, 0, 0, 0, 0);
475 fletcher_4_scalar_native((fletcher_4_ctx_t *)zcp,
478 fletcher_4_native_impl(buf, p2size, zcp);
481 fletcher_4_scalar_native((fletcher_4_ctx_t *)zcp,
482 (char *)buf + p2size, size - p2size);
487 fletcher_4_native_varsize(const void *buf, uint64_t size, zio_cksum_t *zcp)
489 ZIO_SET_CHECKSUM(zcp, 0, 0, 0, 0);
490 fletcher_4_scalar_native((fletcher_4_ctx_t *)zcp, buf, size);
494 fletcher_4_byteswap_impl(const void *buf, uint64_t size, zio_cksum_t *zcp)
496 fletcher_4_ctx_t ctx;
497 const fletcher_4_ops_t *ops = fletcher_4_impl_get();
499 ops->init_byteswap(&ctx);
500 ops->compute_byteswap(&ctx, buf, size);
501 ops->fini_byteswap(&ctx, zcp);
505 fletcher_4_byteswap(const void *buf, uint64_t size,
506 const void *ctx_template, zio_cksum_t *zcp)
509 const uint64_t p2size = P2ALIGN(size, FLETCHER_MIN_SIMD_SIZE);
511 ASSERT(IS_P2ALIGNED(size, sizeof (uint32_t)));
513 if (size == 0 || p2size == 0) {
514 ZIO_SET_CHECKSUM(zcp, 0, 0, 0, 0);
517 fletcher_4_scalar_byteswap((fletcher_4_ctx_t *)zcp,
520 fletcher_4_byteswap_impl(buf, p2size, zcp);
523 fletcher_4_scalar_byteswap((fletcher_4_ctx_t *)zcp,
524 (char *)buf + p2size, size - p2size);
528 /* Incremental Fletcher 4 */
530 #define ZFS_FLETCHER_4_INC_MAX_SIZE (8ULL << 20)
533 fletcher_4_incremental_combine(zio_cksum_t *zcp, const uint64_t size,
534 const zio_cksum_t *nzcp)
536 const uint64_t c1 = size / sizeof (uint32_t);
537 const uint64_t c2 = c1 * (c1 + 1) / 2;
538 const uint64_t c3 = c2 * (c1 + 2) / 3;
541 * Value of 'c3' overflows on buffer sizes close to 16MiB. For that
542 * reason we split incremental fletcher4 computation of large buffers
543 * to steps of (ZFS_FLETCHER_4_INC_MAX_SIZE) size.
545 ASSERT3U(size, <=, ZFS_FLETCHER_4_INC_MAX_SIZE);
547 zcp->zc_word[3] += nzcp->zc_word[3] + c1 * zcp->zc_word[2] +
548 c2 * zcp->zc_word[1] + c3 * zcp->zc_word[0];
549 zcp->zc_word[2] += nzcp->zc_word[2] + c1 * zcp->zc_word[1] +
550 c2 * zcp->zc_word[0];
551 zcp->zc_word[1] += nzcp->zc_word[1] + c1 * zcp->zc_word[0];
552 zcp->zc_word[0] += nzcp->zc_word[0];
556 fletcher_4_incremental_impl(boolean_t native, const void *buf, uint64_t size,
561 uint64_t len = MIN(size, ZFS_FLETCHER_4_INC_MAX_SIZE);
564 fletcher_4_native(buf, len, NULL, &nzc);
566 fletcher_4_byteswap(buf, len, NULL, &nzc);
568 fletcher_4_incremental_combine(zcp, len, &nzc);
576 fletcher_4_incremental_native(void *buf, size_t size, void *data)
578 zio_cksum_t *zcp = data;
579 /* Use scalar impl to directly update cksum of small blocks */
580 if (size < SPA_MINBLOCKSIZE)
581 fletcher_4_scalar_native((fletcher_4_ctx_t *)zcp, buf, size);
583 fletcher_4_incremental_impl(B_TRUE, buf, size, zcp);
588 fletcher_4_incremental_byteswap(void *buf, size_t size, void *data)
590 zio_cksum_t *zcp = data;
591 /* Use scalar impl to directly update cksum of small blocks */
592 if (size < SPA_MINBLOCKSIZE)
593 fletcher_4_scalar_byteswap((fletcher_4_ctx_t *)zcp, buf, size);
595 fletcher_4_incremental_impl(B_FALSE, buf, size, zcp);
604 fletcher_4_kstat_headers(char *buf, size_t size)
608 off += snprintf(buf + off, size, "%-17s", "implementation");
609 off += snprintf(buf + off, size - off, "%-15s", "native");
610 (void) snprintf(buf + off, size - off, "%-15s\n", "byteswap");
616 fletcher_4_kstat_data(char *buf, size_t size, void *data)
618 struct fletcher_4_kstat *fastest_stat =
619 &fletcher_4_stat_data[fletcher_4_supp_impls_cnt];
620 struct fletcher_4_kstat *curr_stat = (struct fletcher_4_kstat *)data;
623 if (curr_stat == fastest_stat) {
624 off += snprintf(buf + off, size - off, "%-17s", "fastest");
625 off += snprintf(buf + off, size - off, "%-15s",
626 fletcher_4_supp_impls[fastest_stat->native]->name);
627 off += snprintf(buf + off, size - off, "%-15s\n",
628 fletcher_4_supp_impls[fastest_stat->byteswap]->name);
630 ptrdiff_t id = curr_stat - fletcher_4_stat_data;
632 off += snprintf(buf + off, size - off, "%-17s",
633 fletcher_4_supp_impls[id]->name);
634 off += snprintf(buf + off, size - off, "%-15llu",
635 (u_longlong_t)curr_stat->native);
636 off += snprintf(buf + off, size - off, "%-15llu\n",
637 (u_longlong_t)curr_stat->byteswap);
644 fletcher_4_kstat_addr(kstat_t *ksp, loff_t n)
646 if (n <= fletcher_4_supp_impls_cnt)
647 ksp->ks_private = (void *) (fletcher_4_stat_data + n);
649 ksp->ks_private = NULL;
651 return (ksp->ks_private);
655 #define FLETCHER_4_FASTEST_FN_COPY(type, src) \
657 fletcher_4_fastest_impl.init_ ## type = src->init_ ## type; \
658 fletcher_4_fastest_impl.fini_ ## type = src->fini_ ## type; \
659 fletcher_4_fastest_impl.compute_ ## type = src->compute_ ## type; \
662 #define FLETCHER_4_BENCH_NS (MSEC2NSEC(1)) /* 1ms */
664 typedef void fletcher_checksum_func_t(const void *, uint64_t, const void *,
669 fletcher_4_benchmark_impl(boolean_t native, char *data, uint64_t data_size)
672 struct fletcher_4_kstat *fastest_stat =
673 &fletcher_4_stat_data[fletcher_4_supp_impls_cnt];
675 uint64_t run_bw, run_time_ns, best_run = 0;
677 uint32_t i, l, sel_save = IMPL_READ(fletcher_4_impl_chosen);
679 fletcher_checksum_func_t *fletcher_4_test = native ?
680 fletcher_4_native : fletcher_4_byteswap;
682 for (i = 0; i < fletcher_4_supp_impls_cnt; i++) {
683 struct fletcher_4_kstat *stat = &fletcher_4_stat_data[i];
684 uint64_t run_count = 0;
686 /* temporary set an implementation */
687 fletcher_4_impl_chosen = i;
692 for (l = 0; l < 32; l++, run_count++)
693 fletcher_4_test(data, data_size, NULL, &zc);
695 run_time_ns = gethrtime() - start;
696 } while (run_time_ns < FLETCHER_4_BENCH_NS);
699 run_bw = data_size * run_count * NANOSEC;
700 run_bw /= run_time_ns; /* B/s */
703 stat->native = run_bw;
705 stat->byteswap = run_bw;
707 if (run_bw > best_run) {
711 fastest_stat->native = i;
712 FLETCHER_4_FASTEST_FN_COPY(native,
713 fletcher_4_supp_impls[i]);
715 fastest_stat->byteswap = i;
716 FLETCHER_4_FASTEST_FN_COPY(byteswap,
717 fletcher_4_supp_impls[i]);
722 /* restore original selection */
723 atomic_swap_32(&fletcher_4_impl_chosen, sel_save);
728 * Initialize and benchmark all supported implementations.
731 fletcher_4_benchmark(void)
733 fletcher_4_ops_t *curr_impl;
736 /* Move supported implementations into fletcher_4_supp_impls */
737 for (i = 0, c = 0; i < ARRAY_SIZE(fletcher_4_impls); i++) {
738 curr_impl = (fletcher_4_ops_t *)fletcher_4_impls[i];
740 if (curr_impl->valid && curr_impl->valid())
741 fletcher_4_supp_impls[c++] = curr_impl;
743 membar_producer(); /* complete fletcher_4_supp_impls[] init */
744 fletcher_4_supp_impls_cnt = c; /* number of supported impl */
747 static const size_t data_size = 1 << SPA_OLD_MAXBLOCKSHIFT; /* 128kiB */
748 char *databuf = vmem_alloc(data_size, KM_SLEEP);
750 for (i = 0; i < data_size / sizeof (uint64_t); i++)
751 ((uint64_t *)databuf)[i] = (uintptr_t)(databuf+i); /* warm-up */
753 fletcher_4_benchmark_impl(B_FALSE, databuf, data_size);
754 fletcher_4_benchmark_impl(B_TRUE, databuf, data_size);
756 vmem_free(databuf, data_size);
759 * Skip the benchmark in user space to avoid impacting libzpool
760 * consumers (zdb, zhack, zinject, ztest). The last implementation
761 * is assumed to be the fastest and used by default.
763 memcpy(&fletcher_4_fastest_impl,
764 fletcher_4_supp_impls[fletcher_4_supp_impls_cnt - 1],
765 sizeof (fletcher_4_fastest_impl));
766 fletcher_4_fastest_impl.name = "fastest";
772 fletcher_4_init(void)
774 /* Determine the fastest available implementation. */
775 fletcher_4_benchmark();
778 /* Install kstats for all implementations */
779 fletcher_4_kstat = kstat_create("zfs", 0, "fletcher_4_bench", "misc",
780 KSTAT_TYPE_RAW, 0, KSTAT_FLAG_VIRTUAL);
781 if (fletcher_4_kstat != NULL) {
782 fletcher_4_kstat->ks_data = NULL;
783 fletcher_4_kstat->ks_ndata = UINT32_MAX;
784 kstat_set_raw_ops(fletcher_4_kstat,
785 fletcher_4_kstat_headers,
786 fletcher_4_kstat_data,
787 fletcher_4_kstat_addr);
788 kstat_install(fletcher_4_kstat);
792 /* Finish initialization */
793 fletcher_4_initialized = B_TRUE;
797 fletcher_4_fini(void)
800 if (fletcher_4_kstat != NULL) {
801 kstat_delete(fletcher_4_kstat);
802 fletcher_4_kstat = NULL;
810 abd_fletcher_4_init(zio_abd_checksum_data_t *cdp)
812 const fletcher_4_ops_t *ops = fletcher_4_impl_get();
813 cdp->acd_private = (void *) ops;
815 if (cdp->acd_byteorder == ZIO_CHECKSUM_NATIVE)
816 ops->init_native(cdp->acd_ctx);
818 ops->init_byteswap(cdp->acd_ctx);
822 abd_fletcher_4_fini(zio_abd_checksum_data_t *cdp)
824 fletcher_4_ops_t *ops = (fletcher_4_ops_t *)cdp->acd_private;
828 if (cdp->acd_byteorder == ZIO_CHECKSUM_NATIVE)
829 ops->fini_native(cdp->acd_ctx, cdp->acd_zcp);
831 ops->fini_byteswap(cdp->acd_ctx, cdp->acd_zcp);
835 abd_fletcher_4_simd2scalar(boolean_t native, void *data, size_t size,
836 zio_abd_checksum_data_t *cdp)
838 zio_cksum_t *zcp = cdp->acd_zcp;
840 ASSERT3U(size, <, FLETCHER_MIN_SIMD_SIZE);
842 abd_fletcher_4_fini(cdp);
843 cdp->acd_private = (void *)&fletcher_4_scalar_ops;
846 fletcher_4_incremental_native(data, size, zcp);
848 fletcher_4_incremental_byteswap(data, size, zcp);
852 abd_fletcher_4_iter(void *data, size_t size, void *private)
854 zio_abd_checksum_data_t *cdp = (zio_abd_checksum_data_t *)private;
855 fletcher_4_ctx_t *ctx = cdp->acd_ctx;
856 fletcher_4_ops_t *ops = (fletcher_4_ops_t *)cdp->acd_private;
857 boolean_t native = cdp->acd_byteorder == ZIO_CHECKSUM_NATIVE;
858 uint64_t asize = P2ALIGN(size, FLETCHER_MIN_SIMD_SIZE);
860 ASSERT(IS_P2ALIGNED(size, sizeof (uint32_t)));
864 ops->compute_native(ctx, data, asize);
866 ops->compute_byteswap(ctx, data, asize);
869 data = (char *)data + asize;
873 ASSERT3U(size, <, FLETCHER_MIN_SIMD_SIZE);
874 /* At this point we have to switch to scalar impl */
875 abd_fletcher_4_simd2scalar(native, data, size, cdp);
881 zio_abd_checksum_func_t fletcher_4_abd_ops = {
882 .acf_init = abd_fletcher_4_init,
883 .acf_fini = abd_fletcher_4_fini,
884 .acf_iter = abd_fletcher_4_iter
889 #define IMPL_FMT(impl, i) (((impl) == (i)) ? "[%s] " : "%s ")
891 #if defined(__linux__)
894 fletcher_4_param_get(char *buffer, zfs_kernel_param_t *unused)
896 const uint32_t impl = IMPL_READ(fletcher_4_impl_chosen);
901 fmt = IMPL_FMT(impl, IMPL_FASTEST);
902 cnt += sprintf(buffer + cnt, fmt, "fastest");
904 /* list all supported implementations */
905 for (uint32_t i = 0; i < fletcher_4_supp_impls_cnt; ++i) {
906 fmt = IMPL_FMT(impl, i);
907 cnt += sprintf(buffer + cnt, fmt,
908 fletcher_4_supp_impls[i]->name);
915 fletcher_4_param_set(const char *val, zfs_kernel_param_t *unused)
917 return (fletcher_4_impl_set(val));
922 #include <sys/sbuf.h>
925 fletcher_4_param(ZFS_MODULE_PARAM_ARGS)
929 if (req->newptr == NULL) {
930 const uint32_t impl = IMPL_READ(fletcher_4_impl_chosen);
931 const int init_buflen = 64;
935 s = sbuf_new_for_sysctl(NULL, NULL, init_buflen, req);
938 fmt = IMPL_FMT(impl, IMPL_FASTEST);
939 (void) sbuf_printf(s, fmt, "fastest");
941 /* list all supported implementations */
942 for (uint32_t i = 0; i < fletcher_4_supp_impls_cnt; ++i) {
943 fmt = IMPL_FMT(impl, i);
944 (void) sbuf_printf(s, fmt,
945 fletcher_4_supp_impls[i]->name);
948 err = sbuf_finish(s);
956 err = sysctl_handle_string(oidp, buf, sizeof (buf), req);
959 return (-fletcher_4_impl_set(buf));
967 * Choose a fletcher 4 implementation in ZFS.
968 * Users can choose "cycle" to exercise all implementations, but this is
969 * for testing purpose therefore it can only be set in user space.
972 ZFS_MODULE_VIRTUAL_PARAM_CALL(zfs, zfs_, fletcher_4_impl,
973 fletcher_4_param_set, fletcher_4_param_get, ZMOD_RW,
974 "Select fletcher 4 implementation.");
977 EXPORT_SYMBOL(fletcher_init);
978 EXPORT_SYMBOL(fletcher_2_incremental_native);
979 EXPORT_SYMBOL(fletcher_2_incremental_byteswap);
980 EXPORT_SYMBOL(fletcher_4_init);
981 EXPORT_SYMBOL(fletcher_4_fini);
982 EXPORT_SYMBOL(fletcher_2_native);
983 EXPORT_SYMBOL(fletcher_2_byteswap);
984 EXPORT_SYMBOL(fletcher_4_native);
985 EXPORT_SYMBOL(fletcher_4_native_varsize);
986 EXPORT_SYMBOL(fletcher_4_byteswap);
987 EXPORT_SYMBOL(fletcher_4_incremental_native);
988 EXPORT_SYMBOL(fletcher_4_incremental_byteswap);
989 EXPORT_SYMBOL(fletcher_4_abd_ops);