]>
Commit | Line | Data |
---|---|---|
ab9f4b0b GN |
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 | /* | |
22 | * Copyright (C) 2016 Gvozden Nešković. All rights reserved. | |
23 | */ | |
24 | ||
25 | #include <sys/zfs_context.h> | |
26 | #include <sys/types.h> | |
27 | #include <sys/zio.h> | |
28 | #include <sys/debug.h> | |
29 | #include <sys/zfs_debug.h> | |
30 | ||
31 | #include <sys/vdev_raidz.h> | |
32 | #include <sys/vdev_raidz_impl.h> | |
33 | ||
c9187d86 GN |
34 | extern boolean_t raidz_will_scalar_work(void); |
35 | ||
36 | /* Opaque implementation with NULL methods to represent original methods */ | |
37 | static const raidz_impl_ops_t vdev_raidz_original_impl = { | |
38 | .name = "original", | |
39 | .is_supported = raidz_will_scalar_work, | |
40 | }; | |
41 | ||
42 | /* RAIDZ parity op that contain the fastest methods */ | |
43 | static raidz_impl_ops_t vdev_raidz_fastest_impl = { | |
44 | .name = "fastest" | |
45 | }; | |
46 | ||
ab9f4b0b GN |
47 | /* All compiled in implementations */ |
48 | const raidz_impl_ops_t *raidz_all_maths[] = { | |
c9187d86 | 49 | &vdev_raidz_original_impl, |
ab9f4b0b | 50 | &vdev_raidz_scalar_impl, |
ae25d222 GN |
51 | #if defined(__x86_64) && defined(HAVE_SSE2) /* only x86_64 for now */ |
52 | &vdev_raidz_sse2_impl, | |
53 | #endif | |
ab9f4b0b | 54 | #if defined(__x86_64) && defined(HAVE_SSSE3) /* only x86_64 for now */ |
ae25d222 | 55 | &vdev_raidz_ssse3_impl, |
ab9f4b0b GN |
56 | #endif |
57 | #if defined(__x86_64) && defined(HAVE_AVX2) /* only x86_64 for now */ | |
62a65a65 RD |
58 | &vdev_raidz_avx2_impl, |
59 | #endif | |
7f547f85 | 60 | #if defined(__x86_64) && defined(HAVE_AVX512F) /* only x86_64 for now */ |
65d71d42 | 61 | &vdev_raidz_avx512f_impl, |
7f547f85 RD |
62 | #endif |
63 | #if defined(__x86_64) && defined(HAVE_AVX512BW) /* only x86_64 for now */ | |
01017962 | 64 | &vdev_raidz_avx512bw_impl, |
7f547f85 | 65 | #endif |
62a65a65 | 66 | #if defined(__aarch64__) |
88cc2352 RD |
67 | &vdev_raidz_aarch64_neon_impl, |
68 | &vdev_raidz_aarch64_neonx2_impl, | |
ab9f4b0b GN |
69 | #endif |
70 | }; | |
71 | ||
72 | /* Indicate that benchmark has been completed */ | |
73 | static boolean_t raidz_math_initialized = B_FALSE; | |
74 | ||
75 | /* Select raidz implementation */ | |
c9187d86 GN |
76 | #define IMPL_FASTEST (UINT32_MAX) |
77 | #define IMPL_CYCLE (UINT32_MAX - 1) | |
78 | #define IMPL_ORIGINAL (0) | |
79 | #define IMPL_SCALAR (1) | |
80 | ||
81 | #define RAIDZ_IMPL_READ(i) (*(volatile uint32_t *) &(i)) | |
82 | ||
83 | static uint32_t zfs_vdev_raidz_impl = IMPL_SCALAR; | |
84 | static uint32_t user_sel_impl = IMPL_FASTEST; | |
ab9f4b0b GN |
85 | |
86 | /* Hold all supported implementations */ | |
c9187d86 GN |
87 | static size_t raidz_supp_impl_cnt = 0; |
88 | static raidz_impl_ops_t *raidz_supp_impl[ARRAY_SIZE(raidz_all_maths)]; | |
ab9f4b0b GN |
89 | |
90 | /* | |
26a08b5c GN |
91 | * kstats values for supported implementations |
92 | * Values represent per disk throughput of 8 disk+parity raidz vdev [B/s] | |
ab9f4b0b GN |
93 | */ |
94 | static raidz_impl_kstat_t raidz_impl_kstats[ARRAY_SIZE(raidz_all_maths) + 1]; | |
95 | ||
96 | /* kstat for benchmarked implementations */ | |
97 | static kstat_t *raidz_math_kstat = NULL; | |
98 | ||
99 | /* | |
100 | * Selects the raidz operation for raidz_map | |
101 | * If rm_ops is set to NULL original raidz implementation will be used | |
102 | */ | |
c9187d86 GN |
103 | raidz_impl_ops_t * |
104 | vdev_raidz_math_get_ops() | |
ab9f4b0b | 105 | { |
c9187d86 GN |
106 | raidz_impl_ops_t *ops = NULL; |
107 | const uint32_t impl = RAIDZ_IMPL_READ(zfs_vdev_raidz_impl); | |
108 | ||
109 | switch (impl) { | |
110 | case IMPL_FASTEST: | |
111 | ASSERT(raidz_math_initialized); | |
112 | ops = &vdev_raidz_fastest_impl; | |
113 | break; | |
ab9f4b0b | 114 | #if !defined(_KERNEL) |
c9187d86 GN |
115 | case IMPL_CYCLE: |
116 | { | |
117 | ASSERT(raidz_math_initialized); | |
118 | ASSERT3U(raidz_supp_impl_cnt, >, 0); | |
119 | /* Cycle through all supported implementations */ | |
ab9f4b0b | 120 | static size_t cycle_impl_idx = 0; |
c9187d86 GN |
121 | size_t idx = (++cycle_impl_idx) % raidz_supp_impl_cnt; |
122 | ops = raidz_supp_impl[idx]; | |
ab9f4b0b | 123 | } |
c9187d86 | 124 | break; |
ab9f4b0b | 125 | #endif |
c9187d86 | 126 | case IMPL_ORIGINAL: |
02730c33 | 127 | ops = (raidz_impl_ops_t *)&vdev_raidz_original_impl; |
c9187d86 GN |
128 | break; |
129 | case IMPL_SCALAR: | |
02730c33 | 130 | ops = (raidz_impl_ops_t *)&vdev_raidz_scalar_impl; |
c9187d86 GN |
131 | break; |
132 | default: | |
c9187d86 GN |
133 | ASSERT3U(impl, <, raidz_supp_impl_cnt); |
134 | ASSERT3U(raidz_supp_impl_cnt, >, 0); | |
d99a0153 CW |
135 | if (impl < ARRAY_SIZE(raidz_all_maths)) |
136 | ops = raidz_supp_impl[impl]; | |
c9187d86 GN |
137 | break; |
138 | } | |
ab9f4b0b | 139 | |
c9187d86 GN |
140 | ASSERT3P(ops, !=, NULL); |
141 | ||
142 | return (ops); | |
ab9f4b0b GN |
143 | } |
144 | ||
145 | /* | |
146 | * Select parity generation method for raidz_map | |
147 | */ | |
c9187d86 | 148 | int |
ab9f4b0b GN |
149 | vdev_raidz_math_generate(raidz_map_t *rm) |
150 | { | |
151 | raidz_gen_f gen_parity = NULL; | |
152 | ||
153 | switch (raidz_parity(rm)) { | |
154 | case 1: | |
155 | gen_parity = rm->rm_ops->gen[RAIDZ_GEN_P]; | |
156 | break; | |
157 | case 2: | |
158 | gen_parity = rm->rm_ops->gen[RAIDZ_GEN_PQ]; | |
159 | break; | |
160 | case 3: | |
161 | gen_parity = rm->rm_ops->gen[RAIDZ_GEN_PQR]; | |
162 | break; | |
163 | default: | |
164 | gen_parity = NULL; | |
165 | cmn_err(CE_PANIC, "invalid RAID-Z configuration %d", | |
02730c33 | 166 | raidz_parity(rm)); |
ab9f4b0b GN |
167 | break; |
168 | } | |
169 | ||
c9187d86 GN |
170 | /* if method is NULL execute the original implementation */ |
171 | if (gen_parity == NULL) | |
172 | return (RAIDZ_ORIGINAL_IMPL); | |
ab9f4b0b GN |
173 | |
174 | gen_parity(rm); | |
c9187d86 GN |
175 | |
176 | return (0); | |
ab9f4b0b GN |
177 | } |
178 | ||
179 | static raidz_rec_f | |
c9187d86 | 180 | reconstruct_fun_p_sel(raidz_map_t *rm, const int *parity_valid, |
4ea3f864 | 181 | const int nbaddata) |
ab9f4b0b GN |
182 | { |
183 | if (nbaddata == 1 && parity_valid[CODE_P]) { | |
184 | return (rm->rm_ops->rec[RAIDZ_REC_P]); | |
185 | } | |
186 | return ((raidz_rec_f) NULL); | |
187 | } | |
188 | ||
189 | static raidz_rec_f | |
c9187d86 | 190 | reconstruct_fun_pq_sel(raidz_map_t *rm, const int *parity_valid, |
4ea3f864 | 191 | const int nbaddata) |
ab9f4b0b GN |
192 | { |
193 | if (nbaddata == 1) { | |
194 | if (parity_valid[CODE_P]) { | |
195 | return (rm->rm_ops->rec[RAIDZ_REC_P]); | |
196 | } else if (parity_valid[CODE_Q]) { | |
197 | return (rm->rm_ops->rec[RAIDZ_REC_Q]); | |
198 | } | |
199 | } else if (nbaddata == 2 && | |
02730c33 | 200 | parity_valid[CODE_P] && parity_valid[CODE_Q]) { |
ab9f4b0b GN |
201 | return (rm->rm_ops->rec[RAIDZ_REC_PQ]); |
202 | } | |
203 | return ((raidz_rec_f) NULL); | |
204 | } | |
205 | ||
206 | static raidz_rec_f | |
c9187d86 | 207 | reconstruct_fun_pqr_sel(raidz_map_t *rm, const int *parity_valid, |
4ea3f864 | 208 | const int nbaddata) |
ab9f4b0b GN |
209 | { |
210 | if (nbaddata == 1) { | |
211 | if (parity_valid[CODE_P]) { | |
212 | return (rm->rm_ops->rec[RAIDZ_REC_P]); | |
213 | } else if (parity_valid[CODE_Q]) { | |
214 | return (rm->rm_ops->rec[RAIDZ_REC_Q]); | |
215 | } else if (parity_valid[CODE_R]) { | |
216 | return (rm->rm_ops->rec[RAIDZ_REC_R]); | |
217 | } | |
218 | } else if (nbaddata == 2) { | |
219 | if (parity_valid[CODE_P] && parity_valid[CODE_Q]) { | |
220 | return (rm->rm_ops->rec[RAIDZ_REC_PQ]); | |
221 | } else if (parity_valid[CODE_P] && parity_valid[CODE_R]) { | |
222 | return (rm->rm_ops->rec[RAIDZ_REC_PR]); | |
223 | } else if (parity_valid[CODE_Q] && parity_valid[CODE_R]) { | |
224 | return (rm->rm_ops->rec[RAIDZ_REC_QR]); | |
225 | } | |
226 | } else if (nbaddata == 3 && | |
02730c33 BB |
227 | parity_valid[CODE_P] && parity_valid[CODE_Q] && |
228 | parity_valid[CODE_R]) { | |
ab9f4b0b GN |
229 | return (rm->rm_ops->rec[RAIDZ_REC_PQR]); |
230 | } | |
231 | return ((raidz_rec_f) NULL); | |
232 | } | |
233 | ||
234 | /* | |
235 | * Select data reconstruction method for raidz_map | |
236 | * @parity_valid - Parity validity flag | |
237 | * @dt - Failed data index array | |
238 | * @nbaddata - Number of failed data columns | |
239 | */ | |
240 | int | |
241 | vdev_raidz_math_reconstruct(raidz_map_t *rm, const int *parity_valid, | |
4ea3f864 | 242 | const int *dt, const int nbaddata) |
ab9f4b0b | 243 | { |
cbf484f8 | 244 | raidz_rec_f rec_fn = NULL; |
ab9f4b0b GN |
245 | |
246 | switch (raidz_parity(rm)) { | |
c9187d86 | 247 | case PARITY_P: |
cbf484f8 | 248 | rec_fn = reconstruct_fun_p_sel(rm, parity_valid, nbaddata); |
c9187d86 GN |
249 | break; |
250 | case PARITY_PQ: | |
cbf484f8 | 251 | rec_fn = reconstruct_fun_pq_sel(rm, parity_valid, nbaddata); |
c9187d86 GN |
252 | break; |
253 | case PARITY_PQR: | |
cbf484f8 | 254 | rec_fn = reconstruct_fun_pqr_sel(rm, parity_valid, nbaddata); |
c9187d86 GN |
255 | break; |
256 | default: | |
257 | cmn_err(CE_PANIC, "invalid RAID-Z configuration %d", | |
258 | raidz_parity(rm)); | |
259 | break; | |
ab9f4b0b GN |
260 | } |
261 | ||
cbf484f8 | 262 | if (rec_fn == NULL) |
c9187d86 GN |
263 | return (RAIDZ_ORIGINAL_IMPL); |
264 | else | |
cbf484f8 | 265 | return (rec_fn(rm, dt)); |
ab9f4b0b GN |
266 | } |
267 | ||
268 | const char *raidz_gen_name[] = { | |
269 | "gen_p", "gen_pq", "gen_pqr" | |
270 | }; | |
271 | const char *raidz_rec_name[] = { | |
272 | "rec_p", "rec_q", "rec_r", | |
273 | "rec_pq", "rec_pr", "rec_qr", "rec_pqr" | |
274 | }; | |
275 | ||
26a08b5c GN |
276 | #define RAIDZ_KSTAT_LINE_LEN (17 + 10*12 + 1) |
277 | ||
278 | static int | |
279 | raidz_math_kstat_headers(char *buf, size_t size) | |
ab9f4b0b GN |
280 | { |
281 | int i; | |
26a08b5c GN |
282 | ssize_t off; |
283 | ||
284 | ASSERT3U(size, >=, RAIDZ_KSTAT_LINE_LEN); | |
285 | ||
286 | off = snprintf(buf, size, "%-17s", "implementation"); | |
287 | ||
288 | for (i = 0; i < ARRAY_SIZE(raidz_gen_name); i++) | |
62a65a65 | 289 | off += snprintf(buf + off, size - off, "%-16s", |
26a08b5c GN |
290 | raidz_gen_name[i]); |
291 | ||
292 | for (i = 0; i < ARRAY_SIZE(raidz_rec_name); i++) | |
62a65a65 | 293 | off += snprintf(buf + off, size - off, "%-16s", |
26a08b5c GN |
294 | raidz_rec_name[i]); |
295 | ||
296 | (void) snprintf(buf + off, size - off, "\n"); | |
297 | ||
298 | return (0); | |
299 | } | |
300 | ||
301 | static int | |
302 | raidz_math_kstat_data(char *buf, size_t size, void *data) | |
303 | { | |
02730c33 BB |
304 | raidz_impl_kstat_t *fstat = &raidz_impl_kstats[raidz_supp_impl_cnt]; |
305 | raidz_impl_kstat_t *cstat = (raidz_impl_kstat_t *)data; | |
26a08b5c GN |
306 | ssize_t off = 0; |
307 | int i; | |
308 | ||
309 | ASSERT3U(size, >=, RAIDZ_KSTAT_LINE_LEN); | |
ab9f4b0b | 310 | |
26a08b5c GN |
311 | if (cstat == fstat) { |
312 | off += snprintf(buf + off, size - off, "%-17s", "fastest"); | |
ab9f4b0b | 313 | |
26a08b5c GN |
314 | for (i = 0; i < ARRAY_SIZE(raidz_gen_name); i++) { |
315 | int id = fstat->gen[i]; | |
62a65a65 | 316 | off += snprintf(buf + off, size - off, "%-16s", |
26a08b5c GN |
317 | raidz_supp_impl[id]->name); |
318 | } | |
319 | for (i = 0; i < ARRAY_SIZE(raidz_rec_name); i++) { | |
320 | int id = fstat->rec[i]; | |
62a65a65 | 321 | off += snprintf(buf + off, size - off, "%-16s", |
26a08b5c GN |
322 | raidz_supp_impl[id]->name); |
323 | } | |
324 | } else { | |
325 | ptrdiff_t id = cstat - raidz_impl_kstats; | |
326 | ||
327 | off += snprintf(buf + off, size - off, "%-17s", | |
328 | raidz_supp_impl[id]->name); | |
329 | ||
330 | for (i = 0; i < ARRAY_SIZE(raidz_gen_name); i++) | |
62a65a65 | 331 | off += snprintf(buf + off, size - off, "%-16llu", |
02730c33 | 332 | (u_longlong_t)cstat->gen[i]); |
26a08b5c GN |
333 | |
334 | for (i = 0; i < ARRAY_SIZE(raidz_rec_name); i++) | |
62a65a65 | 335 | off += snprintf(buf + off, size - off, "%-16llu", |
02730c33 | 336 | (u_longlong_t)cstat->rec[i]); |
ab9f4b0b | 337 | } |
26a08b5c GN |
338 | |
339 | (void) snprintf(buf + off, size - off, "\n"); | |
340 | ||
341 | return (0); | |
342 | } | |
343 | ||
344 | static void * | |
345 | raidz_math_kstat_addr(kstat_t *ksp, loff_t n) | |
346 | { | |
347 | if (n <= raidz_supp_impl_cnt) | |
348 | ksp->ks_private = (void *) (raidz_impl_kstats + n); | |
349 | else | |
350 | ksp->ks_private = NULL; | |
351 | ||
352 | return (ksp->ks_private); | |
ab9f4b0b GN |
353 | } |
354 | ||
355 | #define BENCH_D_COLS (8ULL) | |
356 | #define BENCH_COLS (BENCH_D_COLS + PARITY_PQR) | |
590c9a09 | 357 | #define BENCH_ZIO_SIZE (1ULL << SPA_OLD_MAXBLOCKSHIFT) /* 128 kiB */ |
ab9f4b0b GN |
358 | #define BENCH_NS MSEC2NSEC(25) /* 25ms */ |
359 | ||
360 | typedef void (*benchmark_fn)(raidz_map_t *rm, const int fn); | |
361 | ||
362 | static void | |
363 | benchmark_gen_impl(raidz_map_t *rm, const int fn) | |
364 | { | |
365 | (void) fn; | |
366 | vdev_raidz_generate_parity(rm); | |
367 | } | |
368 | ||
369 | static void | |
370 | benchmark_rec_impl(raidz_map_t *rm, const int fn) | |
371 | { | |
372 | static const int rec_tgt[7][3] = { | |
373 | {1, 2, 3}, /* rec_p: bad QR & D[0] */ | |
374 | {0, 2, 3}, /* rec_q: bad PR & D[0] */ | |
375 | {0, 1, 3}, /* rec_r: bad PQ & D[0] */ | |
376 | {2, 3, 4}, /* rec_pq: bad R & D[0][1] */ | |
377 | {1, 3, 4}, /* rec_pr: bad Q & D[0][1] */ | |
378 | {0, 3, 4}, /* rec_qr: bad P & D[0][1] */ | |
379 | {3, 4, 5} /* rec_pqr: bad & D[0][1][2] */ | |
380 | }; | |
381 | ||
382 | vdev_raidz_reconstruct(rm, rec_tgt[fn], 3); | |
383 | } | |
384 | ||
385 | /* | |
386 | * Benchmarking of all supported implementations (raidz_supp_impl_cnt) | |
387 | * is performed by setting the rm_ops pointer and calling the top level | |
388 | * generate/reconstruct methods of bench_rm. | |
389 | */ | |
390 | static void | |
391 | benchmark_raidz_impl(raidz_map_t *bench_rm, const int fn, benchmark_fn bench_fn) | |
392 | { | |
393 | uint64_t run_cnt, speed, best_speed = 0; | |
394 | hrtime_t t_start, t_diff; | |
395 | raidz_impl_ops_t *curr_impl; | |
02730c33 | 396 | raidz_impl_kstat_t *fstat = &raidz_impl_kstats[raidz_supp_impl_cnt]; |
ab9f4b0b GN |
397 | int impl, i; |
398 | ||
c9187d86 | 399 | for (impl = 0; impl < raidz_supp_impl_cnt; impl++) { |
ab9f4b0b GN |
400 | /* set an implementation to benchmark */ |
401 | curr_impl = raidz_supp_impl[impl]; | |
402 | bench_rm->rm_ops = curr_impl; | |
403 | ||
404 | run_cnt = 0; | |
405 | t_start = gethrtime(); | |
406 | ||
407 | do { | |
408 | for (i = 0; i < 25; i++, run_cnt++) | |
409 | bench_fn(bench_rm, fn); | |
410 | ||
411 | t_diff = gethrtime() - t_start; | |
412 | } while (t_diff < BENCH_NS); | |
413 | ||
414 | speed = run_cnt * BENCH_ZIO_SIZE * NANOSEC; | |
415 | speed /= (t_diff * BENCH_COLS); | |
416 | ||
417 | if (bench_fn == benchmark_gen_impl) | |
26a08b5c | 418 | raidz_impl_kstats[impl].gen[fn] = speed; |
ab9f4b0b | 419 | else |
26a08b5c | 420 | raidz_impl_kstats[impl].rec[fn] = speed; |
ab9f4b0b | 421 | |
c9187d86 GN |
422 | /* Update fastest implementation method */ |
423 | if (speed > best_speed) { | |
ab9f4b0b GN |
424 | best_speed = speed; |
425 | ||
c9187d86 | 426 | if (bench_fn == benchmark_gen_impl) { |
26a08b5c | 427 | fstat->gen[fn] = impl; |
ab9f4b0b GN |
428 | vdev_raidz_fastest_impl.gen[fn] = |
429 | curr_impl->gen[fn]; | |
c9187d86 | 430 | } else { |
26a08b5c | 431 | fstat->rec[fn] = impl; |
ab9f4b0b GN |
432 | vdev_raidz_fastest_impl.rec[fn] = |
433 | curr_impl->rec[fn]; | |
c9187d86 | 434 | } |
ab9f4b0b GN |
435 | } |
436 | } | |
437 | } | |
438 | ||
439 | void | |
440 | vdev_raidz_math_init(void) | |
441 | { | |
442 | raidz_impl_ops_t *curr_impl; | |
443 | zio_t *bench_zio = NULL; | |
444 | raidz_map_t *bench_rm = NULL; | |
445 | uint64_t bench_parity; | |
446 | int i, c, fn; | |
447 | ||
ab9f4b0b GN |
448 | /* move supported impl into raidz_supp_impl */ |
449 | for (i = 0, c = 0; i < ARRAY_SIZE(raidz_all_maths); i++) { | |
02730c33 | 450 | curr_impl = (raidz_impl_ops_t *)raidz_all_maths[i]; |
ab9f4b0b GN |
451 | |
452 | /* initialize impl */ | |
453 | if (curr_impl->init) | |
454 | curr_impl->init(); | |
455 | ||
26a08b5c | 456 | if (curr_impl->is_supported()) |
02730c33 | 457 | raidz_supp_impl[c++] = (raidz_impl_ops_t *)curr_impl; |
ab9f4b0b | 458 | } |
c9187d86 | 459 | membar_producer(); /* complete raidz_supp_impl[] init */ |
ab9f4b0b | 460 | raidz_supp_impl_cnt = c; /* number of supported impl */ |
ab9f4b0b | 461 | |
ab9f4b0b | 462 | #if !defined(_KERNEL) |
c9187d86 | 463 | /* Skip benchmarking and use last implementation as fastest */ |
ab9f4b0b GN |
464 | memcpy(&vdev_raidz_fastest_impl, raidz_supp_impl[raidz_supp_impl_cnt-1], |
465 | sizeof (vdev_raidz_fastest_impl)); | |
c9187d86 | 466 | strcpy(vdev_raidz_fastest_impl.name, "fastest"); |
ab9f4b0b GN |
467 | |
468 | raidz_math_initialized = B_TRUE; | |
469 | ||
470 | /* Use 'cycle' math selection method for userspace */ | |
471 | VERIFY0(vdev_raidz_impl_set("cycle")); | |
472 | return; | |
473 | #endif | |
474 | ||
a6255b7f | 475 | /* Fake an zio and run the benchmark on a warmed up buffer */ |
ab9f4b0b GN |
476 | bench_zio = kmem_zalloc(sizeof (zio_t), KM_SLEEP); |
477 | bench_zio->io_offset = 0; | |
478 | bench_zio->io_size = BENCH_ZIO_SIZE; /* only data columns */ | |
a6255b7f DQ |
479 | bench_zio->io_abd = abd_alloc_linear(BENCH_ZIO_SIZE, B_TRUE); |
480 | memset(abd_to_buf(bench_zio->io_abd), 0xAA, BENCH_ZIO_SIZE); | |
ab9f4b0b GN |
481 | |
482 | /* Benchmark parity generation methods */ | |
483 | for (fn = 0; fn < RAIDZ_GEN_NUM; fn++) { | |
484 | bench_parity = fn + 1; | |
485 | /* New raidz_map is needed for each generate_p/q/r */ | |
c9187d86 | 486 | bench_rm = vdev_raidz_map_alloc(bench_zio, SPA_MINBLOCKSHIFT, |
ab9f4b0b GN |
487 | BENCH_D_COLS + bench_parity, bench_parity); |
488 | ||
489 | benchmark_raidz_impl(bench_rm, fn, benchmark_gen_impl); | |
490 | ||
491 | vdev_raidz_map_free(bench_rm); | |
492 | } | |
493 | ||
494 | /* Benchmark data reconstruction methods */ | |
c9187d86 GN |
495 | bench_rm = vdev_raidz_map_alloc(bench_zio, SPA_MINBLOCKSHIFT, |
496 | BENCH_COLS, PARITY_PQR); | |
ab9f4b0b GN |
497 | |
498 | for (fn = 0; fn < RAIDZ_REC_NUM; fn++) | |
499 | benchmark_raidz_impl(bench_rm, fn, benchmark_rec_impl); | |
500 | ||
501 | vdev_raidz_map_free(bench_rm); | |
502 | ||
503 | /* cleanup the bench zio */ | |
a6255b7f | 504 | abd_free(bench_zio->io_abd); |
ab9f4b0b GN |
505 | kmem_free(bench_zio, sizeof (zio_t)); |
506 | ||
507 | /* install kstats for all impl */ | |
26a08b5c | 508 | raidz_math_kstat = kstat_create("zfs", 0, "vdev_raidz_bench", "misc", |
02730c33 | 509 | KSTAT_TYPE_RAW, 0, KSTAT_FLAG_VIRTUAL); |
ab9f4b0b GN |
510 | |
511 | if (raidz_math_kstat != NULL) { | |
26a08b5c GN |
512 | raidz_math_kstat->ks_data = NULL; |
513 | raidz_math_kstat->ks_ndata = UINT32_MAX; | |
514 | kstat_set_raw_ops(raidz_math_kstat, | |
515 | raidz_math_kstat_headers, | |
516 | raidz_math_kstat_data, | |
517 | raidz_math_kstat_addr); | |
ab9f4b0b GN |
518 | kstat_install(raidz_math_kstat); |
519 | } | |
520 | ||
521 | /* Finish initialization */ | |
c9187d86 | 522 | atomic_swap_32(&zfs_vdev_raidz_impl, user_sel_impl); |
ab9f4b0b | 523 | raidz_math_initialized = B_TRUE; |
ab9f4b0b GN |
524 | } |
525 | ||
526 | void | |
527 | vdev_raidz_math_fini(void) | |
528 | { | |
529 | raidz_impl_ops_t const *curr_impl; | |
530 | int i; | |
531 | ||
532 | if (raidz_math_kstat != NULL) { | |
533 | kstat_delete(raidz_math_kstat); | |
534 | raidz_math_kstat = NULL; | |
535 | } | |
536 | ||
ab9f4b0b GN |
537 | /* fini impl */ |
538 | for (i = 0; i < ARRAY_SIZE(raidz_all_maths); i++) { | |
539 | curr_impl = raidz_all_maths[i]; | |
ab9f4b0b GN |
540 | if (curr_impl->fini) |
541 | curr_impl->fini(); | |
542 | } | |
543 | } | |
544 | ||
c9187d86 | 545 | static const struct { |
02730c33 | 546 | char *name; |
c9187d86 | 547 | uint32_t sel; |
ab9f4b0b | 548 | } math_impl_opts[] = { |
ab9f4b0b | 549 | #if !defined(_KERNEL) |
c9187d86 | 550 | { "cycle", IMPL_CYCLE }, |
ab9f4b0b | 551 | #endif |
c9187d86 GN |
552 | { "fastest", IMPL_FASTEST }, |
553 | { "original", IMPL_ORIGINAL }, | |
554 | { "scalar", IMPL_SCALAR } | |
ab9f4b0b GN |
555 | }; |
556 | ||
557 | /* | |
558 | * Function sets desired raidz implementation. | |
c9187d86 GN |
559 | * |
560 | * If we are called before init(), user preference will be saved in | |
561 | * user_sel_impl, and applied in later init() call. This occurs when module | |
562 | * parameter is specified on module load. Otherwise, directly update | |
563 | * zfs_vdev_raidz_impl. | |
ab9f4b0b GN |
564 | * |
565 | * @val Name of raidz implementation to use | |
566 | * @param Unused. | |
567 | */ | |
9cc1844a GN |
568 | int |
569 | vdev_raidz_impl_set(const char *val) | |
ab9f4b0b | 570 | { |
c9187d86 GN |
571 | int err = -EINVAL; |
572 | char req_name[RAIDZ_IMPL_NAME_MAX]; | |
573 | uint32_t impl = RAIDZ_IMPL_READ(user_sel_impl); | |
ab9f4b0b GN |
574 | size_t i; |
575 | ||
c9187d86 GN |
576 | /* sanitize input */ |
577 | i = strnlen(val, RAIDZ_IMPL_NAME_MAX); | |
578 | if (i == 0 || i == RAIDZ_IMPL_NAME_MAX) | |
579 | return (err); | |
580 | ||
581 | strlcpy(req_name, val, RAIDZ_IMPL_NAME_MAX); | |
582 | while (i > 0 && !!isspace(req_name[i-1])) | |
583 | i--; | |
584 | req_name[i] = '\0'; | |
585 | ||
ab9f4b0b GN |
586 | /* Check mandatory options */ |
587 | for (i = 0; i < ARRAY_SIZE(math_impl_opts); i++) { | |
c9187d86 GN |
588 | if (strcmp(req_name, math_impl_opts[i].name) == 0) { |
589 | impl = math_impl_opts[i].sel; | |
590 | err = 0; | |
591 | break; | |
ab9f4b0b GN |
592 | } |
593 | } | |
594 | ||
c9187d86 GN |
595 | /* check all supported impl if init() was already called */ |
596 | if (err != 0 && raidz_math_initialized) { | |
597 | /* check all supported implementations */ | |
598 | for (i = 0; i < raidz_supp_impl_cnt; i++) { | |
599 | if (strcmp(req_name, raidz_supp_impl[i]->name) == 0) { | |
600 | impl = i; | |
601 | err = 0; | |
602 | break; | |
603 | } | |
ab9f4b0b GN |
604 | } |
605 | } | |
606 | ||
c9187d86 GN |
607 | if (err == 0) { |
608 | if (raidz_math_initialized) | |
609 | atomic_swap_32(&zfs_vdev_raidz_impl, impl); | |
610 | else | |
611 | atomic_swap_32(&user_sel_impl, impl); | |
612 | } | |
613 | ||
614 | return (err); | |
ab9f4b0b GN |
615 | } |
616 | ||
9cc1844a GN |
617 | #if defined(_KERNEL) && defined(HAVE_SPL) |
618 | #include <linux/mod_compat.h> | |
ab9f4b0b | 619 | |
9cc1844a GN |
620 | static int |
621 | zfs_vdev_raidz_impl_set(const char *val, zfs_kernel_param_t *kp) | |
622 | { | |
623 | return (vdev_raidz_impl_set(val)); | |
ab9f4b0b GN |
624 | } |
625 | ||
ab9f4b0b | 626 | static int |
9cc1844a | 627 | zfs_vdev_raidz_impl_get(char *buffer, zfs_kernel_param_t *kp) |
ab9f4b0b GN |
628 | { |
629 | int i, cnt = 0; | |
630 | char *fmt; | |
c9187d86 | 631 | const uint32_t impl = RAIDZ_IMPL_READ(zfs_vdev_raidz_impl); |
ab9f4b0b GN |
632 | |
633 | ASSERT(raidz_math_initialized); | |
634 | ||
ab9f4b0b | 635 | /* list mandatory options */ |
c9187d86 GN |
636 | for (i = 0; i < ARRAY_SIZE(math_impl_opts) - 2; i++) { |
637 | fmt = (impl == math_impl_opts[i].sel) ? "[%s] " : "%s "; | |
ab9f4b0b GN |
638 | cnt += sprintf(buffer + cnt, fmt, math_impl_opts[i].name); |
639 | } | |
640 | ||
641 | /* list all supported implementations */ | |
642 | for (i = 0; i < raidz_supp_impl_cnt; i++) { | |
c9187d86 | 643 | fmt = (i == impl) ? "[%s] " : "%s "; |
ab9f4b0b GN |
644 | cnt += sprintf(buffer + cnt, fmt, raidz_supp_impl[i]->name); |
645 | } | |
646 | ||
ab9f4b0b GN |
647 | return (cnt); |
648 | } | |
649 | ||
650 | module_param_call(zfs_vdev_raidz_impl, zfs_vdev_raidz_impl_set, | |
4ea3f864 | 651 | zfs_vdev_raidz_impl_get, NULL, 0644); |
ab9f4b0b GN |
652 | MODULE_PARM_DESC(zfs_vdev_raidz_impl, "Select raidz implementation."); |
653 | #endif |