/src/openvswitch/lib/dpif-netdev-private-extract.c
Line | Count | Source (jump to first uncovered line) |
1 | | /* |
2 | | * Copyright (c) 2021 Intel. |
3 | | * |
4 | | * Licensed under the Apache License, Version 2.0 (the "License"); |
5 | | * you may not use this file except in compliance with the License. |
6 | | * You may obtain a copy of the License at: |
7 | | * |
8 | | * http://www.apache.org/licenses/LICENSE-2.0 |
9 | | * |
10 | | * Unless required by applicable law or agreed to in writing, software |
11 | | * distributed under the License is distributed on an "AS IS" BASIS, |
12 | | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
13 | | * See the License for the specific language governing permissions and |
14 | | * limitations under the License. |
15 | | */ |
16 | | |
17 | | #include <config.h> |
18 | | #include <errno.h> |
19 | | #include <stdint.h> |
20 | | #include <string.h> |
21 | | |
22 | | #include "cpu.h" |
23 | | #include "dp-packet.h" |
24 | | #include "dpif-netdev-private-dpcls.h" |
25 | | #include "dpif-netdev-private-extract.h" |
26 | | #include "dpif-netdev-private-thread.h" |
27 | | #include "flow.h" |
28 | | #include "openvswitch/vlog.h" |
29 | | #include "ovs-thread.h" |
30 | | #include "util.h" |
31 | | |
32 | | VLOG_DEFINE_THIS_MODULE(dpif_netdev_extract); |
33 | | |
34 | | /* Variable to hold the default MFEX implementation. */ |
35 | | static ATOMIC(miniflow_extract_func) default_mfex_func; |
36 | | |
37 | | #if MFEX_IMPL_AVX512_CHECK |
38 | | static int32_t |
39 | | avx512_isa_probe(bool needs_vbmi) |
40 | | { |
41 | | static enum ovs_cpu_isa isa_required[] = { |
42 | | OVS_CPU_ISA_X86_AVX512F, |
43 | | OVS_CPU_ISA_X86_AVX512BW, |
44 | | OVS_CPU_ISA_X86_BMI2, |
45 | | }; |
46 | | |
47 | | for (uint32_t i = 0; i < ARRAY_SIZE(isa_required); i++) { |
48 | | if (!cpu_has_isa(isa_required[i])) { |
49 | | return -ENOTSUP; |
50 | | } |
51 | | } |
52 | | |
53 | | if (needs_vbmi && !cpu_has_isa(OVS_CPU_ISA_X86_AVX512VBMI)) { |
54 | | return -ENOTSUP; |
55 | | } |
56 | | |
57 | | return 0; |
58 | | } |
59 | | |
60 | | /* Probe functions to check ISA requirements. */ |
61 | | static int32_t |
62 | | mfex_avx512_probe(void) |
63 | | { |
64 | | return avx512_isa_probe(false); |
65 | | } |
66 | | |
67 | | #if HAVE_AVX512VBMI |
68 | | static int32_t |
69 | | mfex_avx512_vbmi_probe(void) |
70 | | { |
71 | | return avx512_isa_probe(true); |
72 | | } |
73 | | #endif |
74 | | #endif |
75 | | |
76 | | /* Implementations of available extract options and |
77 | | * the implementations are always in order of preference. |
78 | | */ |
79 | | static struct dpif_miniflow_extract_impl mfex_impls[] = { |
80 | | |
81 | | [MFEX_IMPL_AUTOVALIDATOR] = { |
82 | | .probe = NULL, |
83 | | .extract_func = dpif_miniflow_extract_autovalidator, |
84 | | .name = "autovalidator", }, |
85 | | |
86 | | [MFEX_IMPL_SCALAR] = { |
87 | | .probe = NULL, |
88 | | .extract_func = NULL, |
89 | | .name = "scalar", }, |
90 | | |
91 | | [MFEX_IMPL_STUDY] = { |
92 | | .probe = NULL, |
93 | | .extract_func = mfex_study_traffic, |
94 | | .name = "study", }, |
95 | | |
96 | | /* Compile in implementations only if the compiler ISA checks pass. */ |
97 | | #if MFEX_IMPL_AVX512_CHECK |
98 | | #if HAVE_AVX512VBMI |
99 | | [MFEX_IMPL_VBMI_IPv4_UDP] = { |
100 | | .probe = mfex_avx512_vbmi_probe, |
101 | | .extract_func = mfex_avx512_vbmi_ip_udp, |
102 | | .name = "avx512_vbmi_ipv4_udp", }, |
103 | | #endif |
104 | | [MFEX_IMPL_IPv4_UDP] = { |
105 | | .probe = mfex_avx512_probe, |
106 | | .extract_func = mfex_avx512_ip_udp, |
107 | | .name = "avx512_ipv4_udp", }, |
108 | | #if HAVE_AVX512VBMI |
109 | | [MFEX_IMPL_VBMI_IPv4_TCP] = { |
110 | | .probe = mfex_avx512_vbmi_probe, |
111 | | .extract_func = mfex_avx512_vbmi_ip_tcp, |
112 | | .name = "avx512_vbmi_ipv4_tcp", }, |
113 | | #endif |
114 | | [MFEX_IMPL_IPv4_TCP] = { |
115 | | .probe = mfex_avx512_probe, |
116 | | .extract_func = mfex_avx512_ip_tcp, |
117 | | .name = "avx512_ipv4_tcp", }, |
118 | | #if HAVE_AVX512VBMI |
119 | | [MFEX_IMPL_VBMI_DOT1Q_IPv4_UDP] = { |
120 | | .probe = mfex_avx512_vbmi_probe, |
121 | | .extract_func = mfex_avx512_vbmi_dot1q_ip_udp, |
122 | | .name = "avx512_vbmi_dot1q_ipv4_udp", }, |
123 | | #endif |
124 | | [MFEX_IMPL_DOT1Q_IPv4_UDP] = { |
125 | | .probe = mfex_avx512_probe, |
126 | | .extract_func = mfex_avx512_dot1q_ip_udp, |
127 | | .name = "avx512_dot1q_ipv4_udp", }, |
128 | | #if HAVE_AVX512VBMI |
129 | | [MFEX_IMPL_VBMI_DOT1Q_IPv4_TCP] = { |
130 | | .probe = mfex_avx512_vbmi_probe, |
131 | | .extract_func = mfex_avx512_vbmi_dot1q_ip_tcp, |
132 | | .name = "avx512_vbmi_dot1q_ipv4_tcp", }, |
133 | | #endif |
134 | | [MFEX_IMPL_DOT1Q_IPv4_TCP] = { |
135 | | .probe = mfex_avx512_probe, |
136 | | .extract_func = mfex_avx512_dot1q_ip_tcp, |
137 | | .name = "avx512_dot1q_ipv4_tcp", |
138 | | }, |
139 | | #if HAVE_AVX512VBMI |
140 | | [MFEX_IMPL_VBMI_IPv6_UDP] = { |
141 | | .probe = mfex_avx512_vbmi_probe, |
142 | | .extract_func = mfex_avx512_vbmi_ipv6_udp, |
143 | | .name = "avx512_vbmi_ipv6_udp", |
144 | | }, |
145 | | #endif |
146 | | [MFEX_IMPL_IPv6_UDP] = { |
147 | | .probe = mfex_avx512_probe, |
148 | | .extract_func = mfex_avx512_ipv6_udp, |
149 | | .name = "avx512_ipv6_udp", |
150 | | }, |
151 | | #if HAVE_AVX512VBMI |
152 | | [MFEX_IMPL_VBMI_IPv6_TCP] = { |
153 | | .probe = mfex_avx512_vbmi_probe, |
154 | | .extract_func = mfex_avx512_vbmi_ipv6_tcp, |
155 | | .name = "avx512_vbmi_ipv6_tcp", |
156 | | }, |
157 | | #endif |
158 | | [MFEX_IMPL_IPv6_TCP] = { |
159 | | .probe = mfex_avx512_probe, |
160 | | .extract_func = mfex_avx512_ipv6_tcp, |
161 | | .name = "avx512_ipv6_tcp", |
162 | | }, |
163 | | #if HAVE_AVX512VBMI |
164 | | [MFEX_IMPL_VBMI_DOT1Q_IPv6_TCP] = { |
165 | | .probe = mfex_avx512_vbmi_probe, |
166 | | .extract_func = mfex_avx512_vbmi_dot1q_ipv6_tcp, |
167 | | .name = "avx512_vbmi_avx512_dot1q_ipv6_tcp", |
168 | | }, |
169 | | #endif |
170 | | [MFEX_IMPL_DOT1Q_IPv6_TCP] = { |
171 | | .probe = mfex_avx512_probe, |
172 | | .extract_func = mfex_avx512_dot1q_ipv6_tcp, |
173 | | .name = "avx512_dot1q_ipv6_tcp", |
174 | | }, |
175 | | #if HAVE_AVX512VBMI |
176 | | [MFEX_IMPL_VBMI_DOT1Q_IPv6_UDP] = { |
177 | | .probe = mfex_avx512_vbmi_probe, |
178 | | .extract_func = mfex_avx512_vbmi_dot1q_ipv6_udp, |
179 | | .name = "avx512_vbmi_avx512_dot1q_ipv6_udp", |
180 | | }, |
181 | | #endif |
182 | | [MFEX_IMPL_DOT1Q_IPv6_UDP] = { |
183 | | .probe = mfex_avx512_probe, |
184 | | .extract_func = mfex_avx512_dot1q_ipv6_udp, |
185 | | .name = "avx512_dot1q_ipv6_udp", |
186 | | }, |
187 | | #if HAVE_AVX512VBMI |
188 | | [MFEX_IMPL_VBMI_IPv4_NVGRE] = { |
189 | | .probe = mfex_avx512_vbmi_probe, |
190 | | .extract_func = mfex_avx512_vbmi_ip_nvgre, |
191 | | .name = "avx512_vbmi_ipv4_nvgre", }, |
192 | | #endif |
193 | | [MFEX_IMPL_IPv4_NVGRE] = { |
194 | | .probe = mfex_avx512_probe, |
195 | | .extract_func = mfex_avx512_ip_nvgre, |
196 | | .name = "avx512_ipv4_nvgre", }, |
197 | | #endif |
198 | | }; |
199 | | |
200 | | BUILD_ASSERT_DECL(MFEX_IMPL_MAX == ARRAY_SIZE(mfex_impls)); |
201 | | |
202 | | void |
203 | | dpif_miniflow_extract_init(void) |
204 | 0 | { |
205 | 0 | atomic_uintptr_t *mfex_func = (void *)&default_mfex_func; |
206 | | #ifdef MFEX_AUTOVALIDATOR_DEFAULT |
207 | | int mfex_idx = MFEX_IMPL_AUTOVALIDATOR; |
208 | | #else |
209 | 0 | int mfex_idx = MFEX_IMPL_SCALAR; |
210 | 0 | #endif |
211 | | |
212 | | /* Call probe on each impl, and cache the result. */ |
213 | 0 | for (int i = 0; i < MFEX_IMPL_MAX; i++) { |
214 | 0 | bool avail = true; |
215 | 0 | if (mfex_impls[i].probe) { |
216 | | /* Return zero is success, non-zero means error. */ |
217 | 0 | avail = (mfex_impls[i].probe() == 0); |
218 | 0 | } |
219 | 0 | VLOG_DBG("Miniflow Extract implementation '%s' %s available.", |
220 | 0 | mfex_impls[i].name, avail ? "is" : "is not"); |
221 | 0 | mfex_impls[i].available = avail; |
222 | 0 | } |
223 | | |
224 | | /* For the first call, this will be choosen based on the |
225 | | * compile time flag. |
226 | | */ |
227 | 0 | VLOG_INFO("Default MFEX Extract implementation is %s.\n", |
228 | 0 | mfex_impls[mfex_idx].name); |
229 | 0 | atomic_store_relaxed(mfex_func, (uintptr_t) mfex_impls |
230 | 0 | [mfex_idx].extract_func); |
231 | 0 | } |
232 | | |
233 | | miniflow_extract_func |
234 | | dp_mfex_impl_get_default(void) |
235 | 0 | { |
236 | 0 | miniflow_extract_func return_func; |
237 | 0 | atomic_uintptr_t *mfex_func = (void *)&default_mfex_func; |
238 | |
|
239 | 0 | atomic_read_relaxed(mfex_func, (uintptr_t *) &return_func); |
240 | |
|
241 | 0 | return return_func; |
242 | 0 | } |
243 | | |
244 | | int |
245 | | dp_mfex_impl_set_default_by_name(const char *name) |
246 | 0 | { |
247 | 0 | miniflow_extract_func new_default; |
248 | 0 | atomic_uintptr_t *mfex_func = (void *)&default_mfex_func; |
249 | |
|
250 | 0 | int err = dp_mfex_impl_get_by_name(name, &new_default); |
251 | |
|
252 | 0 | if (!err) { |
253 | 0 | atomic_store_relaxed(mfex_func, (uintptr_t) new_default); |
254 | 0 | } |
255 | |
|
256 | 0 | return err; |
257 | |
|
258 | 0 | } |
259 | | |
260 | | void |
261 | | dp_mfex_impl_get(struct ds *reply, struct dp_netdev_pmd_thread **pmd_list, |
262 | | size_t pmd_list_size) |
263 | 0 | { |
264 | | /* Add all MFEX functions to reply string. */ |
265 | 0 | ds_put_cstr(reply, "Available MFEX implementations:\n"); |
266 | |
|
267 | 0 | for (int i = 0; i < MFEX_IMPL_MAX; i++) { |
268 | 0 | ds_put_format(reply, " %s (available: %s pmds: ", |
269 | 0 | mfex_impls[i].name, mfex_impls[i].available ? |
270 | 0 | "True" : "False"); |
271 | |
|
272 | 0 | for (size_t j = 0; j < pmd_list_size; j++) { |
273 | 0 | struct dp_netdev_pmd_thread *pmd = pmd_list[j]; |
274 | 0 | if (pmd->core_id == NON_PMD_CORE_ID) { |
275 | 0 | continue; |
276 | 0 | } |
277 | | |
278 | 0 | if (pmd->miniflow_extract_opt == mfex_impls[i].extract_func) { |
279 | 0 | ds_put_format(reply, "%u,", pmd->core_id); |
280 | 0 | } |
281 | 0 | } |
282 | |
|
283 | 0 | ds_chomp(reply, ','); |
284 | |
|
285 | 0 | if (ds_last(reply) == ' ') { |
286 | 0 | ds_put_cstr(reply, "none"); |
287 | 0 | } |
288 | |
|
289 | 0 | ds_put_cstr(reply, ")\n"); |
290 | 0 | } |
291 | |
|
292 | 0 | } |
293 | | |
294 | | /* This function checks all available MFEX implementations, and selects and |
295 | | * returns the function pointer to the one requested by "name". If nothing |
296 | | * is found it returns error. |
297 | | */ |
298 | | int |
299 | | dp_mfex_impl_get_by_name(const char *name, miniflow_extract_func *out_func) |
300 | 0 | { |
301 | 0 | if (!name || !out_func) { |
302 | 0 | return -EINVAL; |
303 | 0 | } |
304 | | |
305 | 0 | for (int i = 0; i < MFEX_IMPL_MAX; i++) { |
306 | 0 | if (strcmp(mfex_impls[i].name, name) == 0) { |
307 | | /* Check available is set before exec. */ |
308 | 0 | if (!mfex_impls[i].available) { |
309 | 0 | *out_func = NULL; |
310 | 0 | return -ENODEV; |
311 | 0 | } |
312 | | |
313 | 0 | *out_func = mfex_impls[i].extract_func; |
314 | 0 | return 0; |
315 | 0 | } |
316 | 0 | } |
317 | | |
318 | 0 | return -ENOENT; |
319 | 0 | } |
320 | | |
321 | | struct dpif_miniflow_extract_impl * |
322 | 0 | dpif_mfex_impl_info_get(void) { |
323 | |
|
324 | 0 | return mfex_impls; |
325 | |
|
326 | 0 | } |
327 | | |
328 | | uint32_t |
329 | | dpif_miniflow_extract_autovalidator(struct dp_packet_batch *packets, |
330 | | struct netdev_flow_key *keys, |
331 | | uint32_t keys_size, odp_port_t in_port, |
332 | | struct dp_netdev_pmd_thread *pmd_handle) |
333 | 0 | { |
334 | 0 | const size_t cnt = dp_packet_batch_size(packets); |
335 | 0 | uint16_t good_l2_5_ofs[NETDEV_MAX_BURST]; |
336 | 0 | uint16_t good_l3_ofs[NETDEV_MAX_BURST]; |
337 | 0 | uint16_t good_l4_ofs[NETDEV_MAX_BURST]; |
338 | 0 | uint16_t good_l2_pad_size[NETDEV_MAX_BURST]; |
339 | 0 | struct dp_packet *packet; |
340 | 0 | struct dp_netdev_pmd_thread *pmd = pmd_handle; |
341 | 0 | struct netdev_flow_key test_keys[NETDEV_MAX_BURST]; |
342 | |
|
343 | 0 | if (keys_size < cnt) { |
344 | 0 | atomic_store_relaxed(&pmd->miniflow_extract_opt, NULL); |
345 | 0 | VLOG_ERR("Invalid key size supplied, Key_size: %d less than" |
346 | 0 | "batch_size: %" PRIuSIZE"\n", keys_size, cnt); |
347 | 0 | VLOG_ERR("Autovalidatior is disabled.\n"); |
348 | 0 | return 0; |
349 | 0 | } |
350 | | |
351 | | /* Run scalar miniflow_extract to get default result. */ |
352 | 0 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets) { |
353 | 0 | pkt_metadata_init(&packet->md, in_port); |
354 | 0 | miniflow_extract(packet, &keys[i].mf); |
355 | | |
356 | | /* Store known good metadata to compare with optimized metadata. */ |
357 | 0 | good_l2_5_ofs[i] = packet->l2_5_ofs; |
358 | 0 | good_l3_ofs[i] = packet->l3_ofs; |
359 | 0 | good_l4_ofs[i] = packet->l4_ofs; |
360 | 0 | good_l2_pad_size[i] = packet->l2_pad_size; |
361 | 0 | } |
362 | |
|
363 | 0 | uint32_t batch_failed = 0; |
364 | | /* Iterate through each version of miniflow implementations. */ |
365 | 0 | for (int j = MFEX_IMPL_START_IDX; j < MFEX_IMPL_MAX; j++) { |
366 | 0 | if (!mfex_impls[j].available) { |
367 | 0 | continue; |
368 | 0 | } |
369 | | /* Reset keys and offsets before each implementation. */ |
370 | 0 | memset(test_keys, 0, keys_size * sizeof(struct netdev_flow_key)); |
371 | 0 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets) { |
372 | 0 | dp_packet_reset_offsets(packet); |
373 | 0 | } |
374 | | /* Call optimized miniflow for each batch of packet. */ |
375 | 0 | uint32_t hit_mask = mfex_impls[j].extract_func(packets, test_keys, |
376 | 0 | keys_size, in_port, |
377 | 0 | pmd_handle); |
378 | | |
379 | | /* Do a miniflow compare for bits, blocks and offsets for all the |
380 | | * classified packets in the hitmask marked by set bits. */ |
381 | 0 | while (hit_mask) { |
382 | | /* Index for the set bit. */ |
383 | 0 | uint32_t i = raw_ctz(hit_mask); |
384 | | /* Set the index in hitmask to Zero. */ |
385 | 0 | hit_mask &= (hit_mask - 1); |
386 | |
|
387 | 0 | uint32_t failed = 0; |
388 | |
|
389 | 0 | struct ds log_msg = DS_EMPTY_INITIALIZER; |
390 | 0 | ds_put_format(&log_msg, "MFEX autovalidator pkt %d\n", i); |
391 | | |
392 | | /* Check miniflow bits are equal. */ |
393 | 0 | if ((keys[i].mf.map.bits[0] != test_keys[i].mf.map.bits[0]) || |
394 | 0 | (keys[i].mf.map.bits[1] != test_keys[i].mf.map.bits[1])) { |
395 | 0 | ds_put_format(&log_msg, "Autovalidation map failed\n" |
396 | 0 | "Good: 0x%llx 0x%llx Test: 0x%llx 0x%llx\n", |
397 | 0 | keys[i].mf.map.bits[0], |
398 | 0 | keys[i].mf.map.bits[1], |
399 | 0 | test_keys[i].mf.map.bits[0], |
400 | 0 | test_keys[i].mf.map.bits[1]); |
401 | 0 | failed = 1; |
402 | 0 | } |
403 | |
|
404 | 0 | if (!miniflow_equal(&keys[i].mf, &test_keys[i].mf)) { |
405 | 0 | uint32_t block_cnt = miniflow_n_values(&keys[i].mf); |
406 | 0 | uint32_t test_block_cnt = miniflow_n_values(&test_keys[i].mf); |
407 | |
|
408 | 0 | ds_put_format(&log_msg, "Autovalidation blocks failed\n" |
409 | 0 | "Good hex:\n"); |
410 | 0 | ds_put_hex_dump(&log_msg, &keys[i].buf, block_cnt * 8, 0, |
411 | 0 | false); |
412 | 0 | ds_put_format(&log_msg, "Test hex:\n"); |
413 | 0 | ds_put_hex_dump(&log_msg, &test_keys[i].buf, |
414 | 0 | test_block_cnt * 8, 0, false); |
415 | 0 | failed = 1; |
416 | 0 | } |
417 | |
|
418 | 0 | packet = packets->packets[i]; |
419 | 0 | if ((packet->l2_pad_size != good_l2_pad_size[i]) || |
420 | 0 | (packet->l2_5_ofs != good_l2_5_ofs[i]) || |
421 | 0 | (packet->l3_ofs != good_l3_ofs[i]) || |
422 | 0 | (packet->l4_ofs != good_l4_ofs[i])) { |
423 | 0 | ds_put_format(&log_msg, |
424 | 0 | "Autovalidation packet offsets failed\n"); |
425 | 0 | ds_put_format(&log_msg, "Good offsets: " |
426 | 0 | "l2_pad_size: %"PRIu16", l2_5_ofs: %"PRIu16", " |
427 | 0 | "l3_ofs: %"PRIu16", l4_ofs: %"PRIu16"\n", |
428 | 0 | good_l2_pad_size[i], good_l2_5_ofs[i], |
429 | 0 | good_l3_ofs[i], good_l4_ofs[i]); |
430 | 0 | ds_put_format(&log_msg, "Test offsets: " |
431 | 0 | "l2_pad_size: %"PRIu16", l2_5_ofs: %"PRIu16", " |
432 | 0 | "l3_ofs: %"PRIu16", l4_ofs: %"PRIu16"\n", |
433 | 0 | packet->l2_pad_size, packet->l2_5_ofs, |
434 | 0 | packet->l3_ofs, packet->l4_ofs); |
435 | 0 | failed = 1; |
436 | 0 | } |
437 | |
|
438 | 0 | if (failed) { |
439 | 0 | VLOG_ERR("Autovalidation for %s failed in pkt %d," |
440 | 0 | " disabling.", mfex_impls[j].name, i); |
441 | 0 | VLOG_ERR("Autovalidation failure details:\n%s", |
442 | 0 | ds_cstr(&log_msg)); |
443 | 0 | batch_failed = 1; |
444 | 0 | } |
445 | 0 | ds_destroy(&log_msg); |
446 | 0 | } |
447 | 0 | } |
448 | | |
449 | | /* Having dumped the debug info for the batch, disable autovalidator. */ |
450 | 0 | if (batch_failed) { |
451 | 0 | atomic_store_relaxed(&pmd->miniflow_extract_opt, NULL); |
452 | 0 | } |
453 | | |
454 | | /* Preserve packet correctness by storing back the good offsets in |
455 | | * packets back. */ |
456 | 0 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets) { |
457 | 0 | packet->l2_5_ofs = good_l2_5_ofs[i]; |
458 | 0 | packet->l3_ofs = good_l3_ofs[i]; |
459 | 0 | packet->l4_ofs = good_l4_ofs[i]; |
460 | 0 | packet->l2_pad_size = good_l2_pad_size[i]; |
461 | 0 | } |
462 | | |
463 | | /* Returning zero implies no packets were hit by autovalidation. This |
464 | | * simplifies unit-tests as changing --enable-mfex-default-autovalidator |
465 | | * would pass/fail. By always returning zero, autovalidator is a little |
466 | | * slower, but we gain consistency in testing. The auto-validator is only |
467 | | * meant to test different implementaions against a batch of packets |
468 | | * without incrementing hit counters. |
469 | | */ |
470 | 0 | return 0; |
471 | 0 | } |