Line data Source code
1 : /*
2 : * SPDX-License-Identifier: Apache-2.0
3 : * Copyright(c) 2021 Cisco Systems, Inc.
4 : */
5 :
6 : #include <vlib/vlib.h>
7 : #include <vppinfra/time.h>
8 : #include <vppinfra/cache.h>
9 : #include <vppinfra/error.h>
10 : #include <vnet/ethernet/ethernet.h>
11 : #include <vnet/ip/ip.h>
12 : #include <vnet/gso/gso.h>
13 : #include <vnet/gso/hdr_offset_parser.h>
14 : #include <vnet/tcp/tcp_packet.h>
15 :
16 : #define MAX_GSO_PACKET_SIZE (TCP_MAX_GSO_SZ - 1)
17 : #define MIN_GSO_SEGMENT_SIZE 128
18 : #define MAX_GSO_SEGMENT_SIZE 2048
19 : #define DEFAULT_GSO_SEGMENT_SIZE 1448
20 :
21 : typedef struct _gso_test_data
22 : {
23 : const char *name;
24 : const char *description;
25 : u8 *data;
26 : u32 data_size;
27 : u32 l4_hdr_len;
28 : u8 is_l2;
29 : u8 is_ip6;
30 : struct _gso_test_data *next;
31 : } gso_test_data_t;
32 :
33 : typedef struct
34 : {
35 : int verbose;
36 :
37 : char *gso_name;
38 : u32 warmup_rounds;
39 : u32 rounds;
40 : u32 n_buffers;
41 : u32 buffer_size;
42 : u32 packet_size;
43 : u32 gso_size;
44 : gso_test_data_t *gso_test_data;
45 : } gso_test_main_t;
46 :
47 : gso_test_main_t gso_test_main;
48 :
49 : #define GSO_TEST_REGISTER_DATA(x, ...) \
50 : __VA_ARGS__ gso_test_data_t __gso_test_data_##x; \
51 : static void __clib_constructor __gso_test_data_fn_##x (void) \
52 : { \
53 : gso_test_main_t *gtm = &gso_test_main; \
54 : __gso_test_data_##x.next = gtm->gso_test_data; \
55 : gtm->gso_test_data = &__gso_test_data_##x; \
56 : } \
57 : __VA_ARGS__ gso_test_data_t __gso_test_data_##x
58 :
59 : // ipv4
60 : u8 gso_ipv4_tcp_data[64] = {
61 : 0x02, 0xfe, 0x39, 0xe5, 0x09, 0x8f, 0x02, 0xfe, 0x2d, 0x18, 0x63, 0x18, 0x08,
62 : 0x00, 0x45, 0x00, 0x05, 0xdc, 0xdb, 0x42, 0x40, 0x00, 0x40, 0x06, 0xc4, 0x85,
63 : 0xc0, 0xa8, 0x0a, 0x02, 0xc0, 0xa8, 0x0a, 0x01, 0xd8, 0xde, 0x14, 0x51, 0x34,
64 : 0x93, 0xa8, 0x1b, 0x7b, 0xef, 0x2e, 0x7e, 0x80, 0x10, 0x00, 0xe5, 0xc7, 0x03,
65 : 0x00, 0x00, 0x01, 0x01, 0x08, 0x0a, 0xce, 0xaa, 0x00, 0x2f, 0xf2, 0xc3
66 : };
67 :
68 559 : GSO_TEST_REGISTER_DATA (gso_ipv4_tcp, static) = {
69 : .name = "ipv4-tcp",
70 : .description = "IPv4 TCP",
71 : .data = gso_ipv4_tcp_data,
72 : .data_size = sizeof (gso_ipv4_tcp_data),
73 : .l4_hdr_len = sizeof (tcp_header_t),
74 : .is_l2 = 1,
75 : .is_ip6 = 0,
76 : };
77 :
78 : // ipv6
79 : u8 gso_ipv6_tcp_data[] = {
80 : 0x02, 0xfe, 0x39, 0xe5, 0x09, 0x8f, 0x02, 0xfe, 0x2d, 0x18, 0x63, 0x18,
81 : 0x08, 0x00, 0x60, 0x0d, 0xf4, 0x97, 0x00, 0x40, 0x06, 0x40, 0xfd, 0x01,
82 : 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
83 : 0x10, 0x00, 0xfd, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
84 : 0x00, 0x00, 0x00, 0x00, 0x10, 0x01, 0xd8, 0xde, 0x14, 0x51, 0x34, 0x93,
85 : 0xa8, 0x1b, 0x7b, 0xef, 0x2e, 0x7e, 0x80, 0x10, 0x00, 0xe5, 0xc7, 0x03,
86 : 0x00, 0x00, 0x01, 0x01, 0x08, 0x0a, 0xce, 0xaa, 0x00, 0x2f, 0xf2, 0xc3
87 : };
88 :
89 559 : GSO_TEST_REGISTER_DATA (gso_ipv6_tcp, static) = {
90 : .name = "ipv6-tcp",
91 : .description = "IPv6 TCP",
92 : .data = gso_ipv6_tcp_data,
93 : .data_size = sizeof (gso_ipv6_tcp_data),
94 : .l4_hdr_len = sizeof (tcp_header_t),
95 : .is_l2 = 1,
96 : .is_ip6 = 1,
97 : };
98 :
99 : static u32
100 0 : fill_buffers (vlib_main_t *vm, u32 *buffer_indices, u8 *data, u32 data_size,
101 : u32 n_buffers, u32 buffer_size, u32 packet_size, u32 gso_size,
102 : u32 l4_hdr_len)
103 : {
104 : u32 i;
105 :
106 0 : for (i = 0; i < n_buffers; i++)
107 : {
108 0 : u64 seed = clib_cpu_time_now ();
109 0 : vlib_buffer_t *b = vlib_get_buffer (vm, buffer_indices[i]);
110 0 : u32 len = 0;
111 0 : u32 remaining_data =
112 0 : (packet_size > buffer_size) ? (packet_size - buffer_size) : 0;
113 :
114 0 : clib_memcpy_fast (b->data, data, data_size);
115 0 : b->current_data = 0;
116 :
117 0 : for (u32 j = data_size; j < buffer_size; j += 8)
118 0 : *(u64 *) (b->data + j) = 1 + random_u64 (&seed);
119 0 : b->current_length = buffer_size;
120 :
121 0 : if (remaining_data)
122 : {
123 0 : vlib_buffer_t *pb = b;
124 : u32 n_alloc,
125 0 : n_bufs = ((remaining_data + buffer_size - 1) / buffer_size);
126 0 : u32 *buffers = 0;
127 : u32 fill_data_size;
128 0 : u32 k = 0;
129 :
130 0 : vec_validate (buffers, n_bufs - 1);
131 0 : n_alloc = vlib_buffer_alloc (vm, buffers, n_bufs);
132 0 : if (n_alloc < n_bufs)
133 : {
134 0 : vlib_buffer_free (vm, buffers, n_alloc);
135 0 : vlib_cli_output (
136 : vm, "vlib buffer alloc failed at %u requested %u actual %u", i,
137 : n_bufs, n_alloc);
138 0 : return i;
139 : }
140 :
141 : do
142 : {
143 0 : pb->next_buffer = buffers[k];
144 0 : pb->flags |= VLIB_BUFFER_NEXT_PRESENT;
145 0 : pb = vlib_get_buffer (vm, buffers[k]);
146 0 : pb->current_data = 0;
147 0 : fill_data_size = clib_min (buffer_size, remaining_data);
148 0 : remaining_data -= fill_data_size;
149 0 : for (u32 l = 0; l < fill_data_size; l += 8)
150 0 : *(u64 *) (pb->data + l) = 1 + random_u64 (&seed);
151 0 : pb->current_length = fill_data_size;
152 0 : k++;
153 0 : len += fill_data_size;
154 : }
155 0 : while (k < n_bufs);
156 0 : b->flags |= VNET_BUFFER_F_GSO;
157 0 : vnet_buffer2 (b)->gso_size = gso_size;
158 0 : vnet_buffer2 (b)->gso_l4_hdr_sz = l4_hdr_len;
159 : }
160 0 : b->total_length_not_including_first_buffer = len;
161 0 : b->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID;
162 : }
163 0 : return i;
164 : }
165 :
166 : static_always_inline u32
167 0 : gso_segment_buffer_test (vlib_main_t *vm, u32 bi,
168 : vnet_interface_per_thread_data_t *ptd, u8 is_l2,
169 : u8 is_ip6)
170 : {
171 0 : vlib_buffer_t *b = vlib_get_buffer (vm, bi);
172 0 : generic_header_offset_t gho = { 0 };
173 0 : u32 n_tx_bytes = 0;
174 :
175 0 : if (PREDICT_TRUE (b->flags & VNET_BUFFER_F_GSO))
176 : {
177 0 : vnet_generic_header_offset_parser (b, &gho, is_l2, !is_ip6, is_ip6);
178 0 : n_tx_bytes = gso_segment_buffer_inline (vm, ptd, b, &gho, is_l2, is_ip6);
179 : }
180 :
181 0 : return n_tx_bytes;
182 : }
183 :
184 : static clib_error_t *
185 0 : test_gso_perf (vlib_main_t *vm, gso_test_main_t *gtm)
186 : {
187 0 : clib_error_t *err = 0;
188 0 : vnet_interface_per_thread_data_t *ptd = 0;
189 0 : u32 packet_size = MAX_GSO_PACKET_SIZE;
190 0 : u32 buffer_size = vlib_buffer_get_default_data_size (vm);
191 : u32 gso_size;
192 : u32 n_buffers, warmup_rounds, rounds;
193 0 : u32 *buffer_indices = 0;
194 0 : u64 t0, t1, t2[10] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
195 0 : gso_test_data_t *gso_test_data = gtm->gso_test_data;
196 : int i, j, k;
197 :
198 0 : if (gtm->buffer_size > buffer_size)
199 0 : return clib_error_return (0, "buffer size must be <= %u", buffer_size);
200 :
201 0 : if (gtm->packet_size > packet_size)
202 0 : return clib_error_return (0, "gso packet size must be <= %u", packet_size);
203 :
204 0 : if ((gtm->gso_size > MAX_GSO_SEGMENT_SIZE) ||
205 0 : (gtm->gso_size < MIN_GSO_SEGMENT_SIZE))
206 0 : return clib_error_return (
207 : 0, "gso segment size must be in between %u >= and <= %u",
208 : MIN_GSO_SEGMENT_SIZE, MAX_GSO_SEGMENT_SIZE);
209 :
210 0 : rounds = gtm->rounds ? gtm->rounds : 256;
211 0 : n_buffers = gtm->n_buffers ? gtm->n_buffers : 256;
212 0 : warmup_rounds = gtm->warmup_rounds ? gtm->warmup_rounds : 256;
213 0 : buffer_size = gtm->buffer_size ? gtm->buffer_size : buffer_size;
214 0 : gso_size = gtm->gso_size;
215 0 : packet_size = gtm->packet_size ? gtm->packet_size : packet_size;
216 :
217 0 : vec_validate_aligned (ptd, n_buffers - 1, CLIB_CACHE_LINE_BYTES);
218 0 : vec_validate_aligned (buffer_indices, n_buffers - 1, CLIB_CACHE_LINE_BYTES);
219 :
220 0 : vlib_cli_output (vm,
221 : "GSO Segmentation: packet-size %u gso-size %u buffer-size "
222 : "%u n_buffers %u rounds %u "
223 : "warmup-rounds %u",
224 : packet_size, gso_size, buffer_size, n_buffers, rounds,
225 : warmup_rounds);
226 0 : vlib_cli_output (vm, " cpu-freq %.2f GHz",
227 0 : (f64) vm->clib_time.clocks_per_second * 1e-9);
228 :
229 0 : while (gso_test_data)
230 : {
231 0 : u32 n_filled = 0;
232 0 : u32 n_alloc = vlib_buffer_alloc (vm, buffer_indices, n_buffers);
233 0 : if (n_alloc != n_buffers)
234 : {
235 0 : vlib_cli_output (vm, " Test: %s FAILED", gso_test_data->description);
236 0 : err = clib_error_return (0, "buffer alloc failure");
237 0 : vlib_buffer_free (vm, buffer_indices, n_alloc);
238 0 : goto done;
239 : }
240 : n_filled =
241 0 : fill_buffers (vm, buffer_indices, gso_test_data->data,
242 : gso_test_data->data_size, n_buffers, buffer_size,
243 : packet_size, gso_size, gso_test_data->l4_hdr_len);
244 :
245 0 : u8 is_l2 = gso_test_data->is_l2;
246 0 : u8 is_ip6 = gso_test_data->is_ip6;
247 :
248 0 : for (k = 0; k < warmup_rounds; k++)
249 : {
250 0 : for (j = 0; j < n_filled; j++)
251 0 : gso_segment_buffer_test (vm, buffer_indices[j], &ptd[j], is_l2,
252 : is_ip6);
253 0 : for (j = 0; j < n_filled; j++)
254 : {
255 0 : vlib_buffer_free (vm, ptd[j].split_buffers,
256 0 : vec_len (ptd[j].split_buffers));
257 0 : vec_free (ptd[j].split_buffers);
258 : }
259 : }
260 :
261 0 : for (i = 0; i < 10; i++)
262 : {
263 0 : for (k = 0; k < rounds; k++)
264 : {
265 0 : t0 = clib_cpu_time_now ();
266 0 : for (j = 0; j < n_filled; j++)
267 0 : gso_segment_buffer_test (vm, buffer_indices[j], &ptd[j], is_l2,
268 : is_ip6);
269 0 : t1 = clib_cpu_time_now ();
270 0 : t2[i] += (t1 - t0);
271 0 : for (j = 0; j < n_filled; j++)
272 : {
273 0 : vlib_buffer_free (vm, ptd[j].split_buffers,
274 0 : vec_len (ptd[j].split_buffers));
275 0 : vec_free (ptd[j].split_buffers);
276 : }
277 : }
278 : }
279 :
280 0 : vlib_cli_output (
281 : vm, "===========================================================");
282 0 : vlib_cli_output (vm, " Test: %s", gso_test_data->description);
283 0 : vlib_cli_output (
284 : vm, "===========================================================");
285 0 : for (i = 0; i < 10; i++)
286 : {
287 : // ticks per packet
288 0 : f64 tpp1 = (f64) (t2[i]) / (n_filled * rounds);
289 : // ticks per Byte
290 0 : f64 tpB1 = (f64) (t2[i]) / (n_filled * rounds * packet_size);
291 : // Packets per second
292 0 : f64 Kpps1 = vm->clib_time.clocks_per_second * 1e-3 / tpp1;
293 : // Throughput Giga-bits per second
294 0 : f64 Gbps1 = vm->clib_time.clocks_per_second * 8 * 1e-9 / tpB1;
295 :
296 0 : vlib_cli_output (
297 : vm, "%-2u: %.03f ticks/packet, %.02f Kpps, %.02f Gbps\n", i + 1,
298 : tpp1, Kpps1, Gbps1);
299 : }
300 0 : if (n_alloc)
301 0 : vlib_buffer_free (vm, buffer_indices, n_alloc);
302 0 : clib_memset (t2, 0, sizeof (t2));
303 0 : gso_test_data = gso_test_data->next;
304 : }
305 :
306 0 : done:
307 :
308 0 : vec_free (ptd);
309 0 : vec_free (buffer_indices);
310 0 : return err;
311 : }
312 :
313 : static clib_error_t *
314 0 : test_gso_command_fn (vlib_main_t *vm, unformat_input_t *input,
315 : vlib_cli_command_t *cmd)
316 : {
317 0 : gso_test_main_t *gtm = &gso_test_main;
318 0 : clib_error_t *err = 0;
319 : f64 end, start, total_time;
320 :
321 0 : gtm->gso_size = DEFAULT_GSO_SEGMENT_SIZE;
322 0 : gtm->warmup_rounds = 0;
323 0 : gtm->rounds = 0;
324 0 : gtm->n_buffers = 0;
325 0 : gtm->buffer_size = 0;
326 0 : gtm->packet_size = 0;
327 :
328 0 : while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
329 : {
330 0 : if (unformat (input, "verbose"))
331 0 : gtm->verbose = 1;
332 0 : else if (unformat (input, "detail"))
333 0 : gtm->verbose = 2;
334 0 : else if (unformat (input, "buffers %u", >m->n_buffers))
335 : ;
336 0 : else if (unformat (input, "buffer-size %u", >m->buffer_size))
337 : ;
338 0 : else if (unformat (input, "packet-size %u", >m->packet_size))
339 : ;
340 0 : else if (unformat (input, "gso-size %u", >m->gso_size))
341 : ;
342 0 : else if (unformat (input, "rounds %u", >m->rounds))
343 : ;
344 0 : else if (unformat (input, "warmup-rounds %u", >m->warmup_rounds))
345 : ;
346 : else
347 : {
348 0 : return clib_error_return (0, "unknown input '%U'",
349 : format_unformat_error, input);
350 : }
351 : }
352 :
353 0 : start = clib_cpu_time_now ();
354 0 : err = test_gso_perf (vm, gtm);
355 0 : end = clib_cpu_time_now ();
356 :
357 0 : total_time = (f64) (end - start) / vm->clib_time.clocks_per_second;
358 0 : vlib_cli_output (vm, "Total Time Test Took %.02f seconds", total_time);
359 :
360 0 : return err;
361 : }
362 :
363 16239 : VLIB_CLI_COMMAND (test_gso_command, static) = {
364 : .path = "test gso",
365 : .short_help = "test gso [buffers <n>] [buffer-size <size>] [packet-size "
366 : "<size>] [gso-size <size>] [rounds <n>] "
367 : "[warmup-rounds <n>]",
368 : .function = test_gso_command_fn,
369 : };
370 :
371 : static clib_error_t *
372 559 : gso_test_init (vlib_main_t *vm)
373 : {
374 559 : return (0);
375 : }
376 :
377 2799 : VLIB_INIT_FUNCTION (gso_test_init);
|