1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause
3 *
4 * Copyright (c) 2015-2024 Amazon.com, Inc. or its affiliates.
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 *
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 *
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 *
18 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
19 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
20 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
21 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
22 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
23 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
24 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
28 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29 */
30
31 #include <sys/param.h>
32 #include "opt_rss.h"
33
34 #include "ena_rss.h"
35 #include "ena_sysctl.h"
36
37 static void ena_sysctl_add_wd(struct ena_adapter *);
38 static void ena_sysctl_add_stats(struct ena_adapter *);
39 static void ena_sysctl_add_eni_metrics(struct ena_adapter *);
40 static void ena_sysctl_add_customer_metrics(struct ena_adapter *);
41 static void ena_sysctl_add_srd_info(struct ena_adapter *);
42 static void ena_sysctl_add_tuneables(struct ena_adapter *);
43 static void ena_sysctl_add_irq_affinity(struct ena_adapter *);
44 /* Kernel option RSS prevents manipulation of key hash and indirection table. */
45 #ifndef RSS
46 static void ena_sysctl_add_rss(struct ena_adapter *);
47 #endif
48 static int ena_sysctl_buf_ring_size(SYSCTL_HANDLER_ARGS);
49 static int ena_sysctl_rx_queue_size(SYSCTL_HANDLER_ARGS);
50 static int ena_sysctl_io_queues_nb(SYSCTL_HANDLER_ARGS);
51 static int ena_sysctl_irq_base_cpu(SYSCTL_HANDLER_ARGS);
52 static int ena_sysctl_irq_cpu_stride(SYSCTL_HANDLER_ARGS);
53 static int ena_sysctl_metrics_interval(SYSCTL_HANDLER_ARGS);
54 #ifndef RSS
55 static int ena_sysctl_rss_key(SYSCTL_HANDLER_ARGS);
56 static int ena_sysctl_rss_indir_table(SYSCTL_HANDLER_ARGS);
57 #endif
58
59 /* Limit max ENA sample rate to be an hour. */
60 #define ENA_METRICS_MAX_SAMPLE_INTERVAL 3600
61 #define ENA_HASH_KEY_MSG_SIZE (ENA_HASH_KEY_SIZE * 2 + 1)
62
63 #define SYSCTL_GSTRING_LEN 128
64
65 #define ENA_METRIC_ENI_ENTRY(stat, desc) { \
66 .name = #stat, \
67 .description = #desc, \
68 }
69
70 #define ENA_STAT_ENTRY(stat, desc, stat_type) { \
71 .name = #stat, \
72 .description = #desc, \
73 .stat_offset = offsetof(struct ena_admin_##stat_type, stat) / sizeof(u64), \
74 }
75
76 #define ENA_STAT_ENA_SRD_ENTRY(stat, desc) \
77 ENA_STAT_ENTRY(stat, desc, ena_srd_stats)
78
79 struct ena_hw_metrics {
80 char name[SYSCTL_GSTRING_LEN];
81 char description[SYSCTL_GSTRING_LEN];
82 };
83
84 struct ena_srd_metrics {
85 char name[SYSCTL_GSTRING_LEN];
86 char description[SYSCTL_GSTRING_LEN];
87 int stat_offset;
88 };
89
90 static const struct ena_srd_metrics ena_srd_stats_strings[] = {
91 ENA_STAT_ENA_SRD_ENTRY(
92 ena_srd_tx_pkts, Number of packets transmitted over ENA SRD),
93 ENA_STAT_ENA_SRD_ENTRY(
94 ena_srd_eligible_tx_pkts, Number of packets transmitted or could
95 have been transmitted over ENA SRD),
96 ENA_STAT_ENA_SRD_ENTRY(
97 ena_srd_rx_pkts, Number of packets received over ENA SRD),
98 ENA_STAT_ENA_SRD_ENTRY(
99 ena_srd_resource_utilization, Percentage of the ENA SRD resources
100 that are in use),
101 };
102
103 static const struct ena_hw_metrics ena_hw_stats_strings[] = {
104 ENA_METRIC_ENI_ENTRY(
105 bw_in_allowance_exceeded, Inbound BW allowance exceeded),
106 ENA_METRIC_ENI_ENTRY(
107 bw_out_allowance_exceeded, Outbound BW allowance exceeded),
108 ENA_METRIC_ENI_ENTRY(
109 pps_allowance_exceeded, PPS allowance exceeded),
110 ENA_METRIC_ENI_ENTRY(
111 conntrack_allowance_exceeded, Connection tracking allowance exceeded),
112 ENA_METRIC_ENI_ENTRY(
113 linklocal_allowance_exceeded, Linklocal packet rate allowance),
114 ENA_METRIC_ENI_ENTRY(
115 conntrack_allowance_available, Number of available conntracks),
116 };
117
118 #ifndef ARRAY_SIZE
119 #define ARRAY_SIZE(a) (sizeof(a) / sizeof(a[0]))
120 #endif
121
122 #define ENA_CUSTOMER_METRICS_ARRAY_SIZE ARRAY_SIZE(ena_hw_stats_strings)
123 #define ENA_SRD_METRICS_ARRAY_SIZE ARRAY_SIZE(ena_srd_stats_strings)
124
125 static SYSCTL_NODE(_hw, OID_AUTO, ena, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
126 "ENA driver parameters");
127
128 /*
129 * Logging level for changing verbosity of the output
130 */
131 int ena_log_level = ENA_INFO;
132 SYSCTL_INT(_hw_ena, OID_AUTO, log_level, CTLFLAG_RWTUN, &ena_log_level, 0,
133 "Logging level indicating verbosity of the logs");
134
135 SYSCTL_CONST_STRING(_hw_ena, OID_AUTO, driver_version, CTLFLAG_RD,
136 ENA_DRV_MODULE_VERSION, "ENA driver version");
137
138 /*
139 * Use 9k mbufs for the Rx buffers. Default to 0 (use page size mbufs instead).
140 * Using 9k mbufs in low memory conditions might cause allocation to take a lot
141 * of time and lead to the OS instability as it needs to look for the contiguous
142 * pages.
143 * However, page size mbufs has a bit smaller throughput than 9k mbufs, so if
144 * the network performance is the priority, the 9k mbufs can be used.
145 */
146 int ena_enable_9k_mbufs = 0;
147 SYSCTL_INT(_hw_ena, OID_AUTO, enable_9k_mbufs, CTLFLAG_RDTUN,
148 &ena_enable_9k_mbufs, 0, "Use 9 kB mbufs for Rx descriptors");
149
150 /*
151 * Force the driver to use large or regular LLQ (Low Latency Queue) header size.
152 * Defaults to ENA_LLQ_HEADER_SIZE_POLICY_DEFAULT. This option may be
153 * important for platforms, which often handle packet headers on Tx with total
154 * header size greater than 96B, as it may reduce the latency.
155 * It also reduces the maximum Tx queue size by half, so it may cause more Tx
156 * packet drops.
157 */
158 int ena_force_large_llq_header = ENA_LLQ_HEADER_SIZE_POLICY_DEFAULT;
159 SYSCTL_INT(_hw_ena, OID_AUTO, force_large_llq_header, CTLFLAG_RDTUN,
160 &ena_force_large_llq_header, 0,
161 "Change default LLQ entry size received from the device\n");
162
163 int ena_rss_table_size = ENA_RX_RSS_TABLE_SIZE;
164
ena_sysctl_allocate_customer_metrics_buffer(struct ena_adapter * adapter)165 int ena_sysctl_allocate_customer_metrics_buffer(struct ena_adapter *adapter)
166 {
167 int rc = 0;
168
169 adapter->customer_metrics_array = malloc((sizeof(u64) * ENA_CUSTOMER_METRICS_ARRAY_SIZE),
170 M_DEVBUF, M_NOWAIT | M_ZERO);
171 if (unlikely(adapter->customer_metrics_array == NULL))
172 rc = ENOMEM;
173
174 return rc;
175 }
176 void
ena_sysctl_add_nodes(struct ena_adapter * adapter)177 ena_sysctl_add_nodes(struct ena_adapter *adapter)
178 {
179 struct ena_com_dev *dev = adapter->ena_dev;
180
181 if (ena_com_get_cap(dev, ENA_ADMIN_CUSTOMER_METRICS))
182 ena_sysctl_add_customer_metrics(adapter);
183 else if (ena_com_get_cap(dev, ENA_ADMIN_ENI_STATS))
184 ena_sysctl_add_eni_metrics(adapter);
185
186 if (ena_com_get_cap(adapter->ena_dev, ENA_ADMIN_ENA_SRD_INFO))
187 ena_sysctl_add_srd_info(adapter);
188
189 ena_sysctl_add_wd(adapter);
190 ena_sysctl_add_stats(adapter);
191 ena_sysctl_add_tuneables(adapter);
192 ena_sysctl_add_irq_affinity(adapter);
193 #ifndef RSS
194 ena_sysctl_add_rss(adapter);
195 #endif
196 }
197
198 static void
ena_sysctl_add_wd(struct ena_adapter * adapter)199 ena_sysctl_add_wd(struct ena_adapter *adapter)
200 {
201 device_t dev;
202
203 struct sysctl_ctx_list *ctx;
204 struct sysctl_oid *tree;
205 struct sysctl_oid_list *child;
206
207 dev = adapter->pdev;
208
209 ctx = device_get_sysctl_ctx(dev);
210 tree = device_get_sysctl_tree(dev);
211 child = SYSCTL_CHILDREN(tree);
212
213 /* Sysctl calls for Watchdog service */
214 SYSCTL_ADD_INT(ctx, child, OID_AUTO, "wd_active", CTLFLAG_RWTUN,
215 &adapter->wd_active, 0, "Watchdog is active");
216
217 SYSCTL_ADD_QUAD(ctx, child, OID_AUTO, "keep_alive_timeout",
218 CTLFLAG_RWTUN, &adapter->keep_alive_timeout,
219 "Timeout for Keep Alive messages");
220
221 SYSCTL_ADD_QUAD(ctx, child, OID_AUTO, "missing_tx_timeout",
222 CTLFLAG_RWTUN, &adapter->missing_tx_timeout,
223 "Timeout for TX completion");
224
225 SYSCTL_ADD_U32(ctx, child, OID_AUTO, "missing_tx_max_queues",
226 CTLFLAG_RWTUN, &adapter->missing_tx_max_queues, 0,
227 "Number of TX queues to check per run");
228
229 SYSCTL_ADD_U32(ctx, child, OID_AUTO, "missing_tx_threshold",
230 CTLFLAG_RWTUN, &adapter->missing_tx_threshold, 0,
231 "Max number of timeouted packets");
232 }
233
234 static void
ena_sysctl_add_stats(struct ena_adapter * adapter)235 ena_sysctl_add_stats(struct ena_adapter *adapter)
236 {
237 device_t dev;
238
239 struct ena_ring *tx_ring;
240 struct ena_ring *rx_ring;
241
242 struct ena_hw_stats *hw_stats;
243 struct ena_stats_dev *dev_stats;
244 struct ena_stats_tx *tx_stats;
245 struct ena_stats_rx *rx_stats;
246 struct ena_com_stats_admin *admin_stats;
247
248 struct sysctl_ctx_list *ctx;
249 struct sysctl_oid *tree;
250 struct sysctl_oid_list *child;
251
252 struct sysctl_oid *queue_node, *tx_node, *rx_node, *hw_node;
253 struct sysctl_oid *admin_node;
254 struct sysctl_oid_list *queue_list, *tx_list, *rx_list, *hw_list;
255 struct sysctl_oid_list *admin_list;
256
257 #define QUEUE_NAME_LEN 32
258 char namebuf[QUEUE_NAME_LEN];
259 int i;
260
261 dev = adapter->pdev;
262
263 ctx = device_get_sysctl_ctx(dev);
264 tree = device_get_sysctl_tree(dev);
265 child = SYSCTL_CHILDREN(tree);
266
267 tx_ring = adapter->tx_ring;
268 rx_ring = adapter->rx_ring;
269
270 hw_stats = &adapter->hw_stats;
271 dev_stats = &adapter->dev_stats;
272 admin_stats = &adapter->ena_dev->admin_queue.stats;
273
274 SYSCTL_ADD_COUNTER_U64(ctx, child, OID_AUTO, "wd_expired", CTLFLAG_RD,
275 &dev_stats->wd_expired, "Watchdog expiry count");
276 SYSCTL_ADD_COUNTER_U64(ctx, child, OID_AUTO, "interface_up", CTLFLAG_RD,
277 &dev_stats->interface_up, "Network interface up count");
278 SYSCTL_ADD_COUNTER_U64(ctx, child, OID_AUTO, "interface_down", CTLFLAG_RD,
279 &dev_stats->interface_down, "Network interface down count");
280 SYSCTL_ADD_COUNTER_U64(ctx, child, OID_AUTO, "admin_q_pause", CTLFLAG_RD,
281 &dev_stats->admin_q_pause, "Admin queue pauses");
282 SYSCTL_ADD_COUNTER_U64(ctx, child, OID_AUTO, "os_trigger", CTLFLAG_RD,
283 &dev_stats->os_trigger, "OS trigger count");
284 SYSCTL_ADD_COUNTER_U64(ctx, child, OID_AUTO, "missing_tx_cmpl", CTLFLAG_RD,
285 &dev_stats->missing_tx_cmpl, "Missing TX completions resets count");
286 SYSCTL_ADD_COUNTER_U64(ctx, child, OID_AUTO, "bad_rx_req_id", CTLFLAG_RD,
287 &dev_stats->bad_rx_req_id, "Bad RX req id count");
288 SYSCTL_ADD_COUNTER_U64(ctx, child, OID_AUTO, "bad_tx_req_id", CTLFLAG_RD,
289 &dev_stats->bad_tx_req_id, "Bad TX req id count");
290 SYSCTL_ADD_COUNTER_U64(ctx, child, OID_AUTO, "bad_rx_desc_num", CTLFLAG_RD,
291 &dev_stats->bad_rx_desc_num, "Bad RX descriptors number count");
292 SYSCTL_ADD_COUNTER_U64(ctx, child, OID_AUTO, "invalid_state", CTLFLAG_RD,
293 &dev_stats->invalid_state, "Driver invalid state count");
294 SYSCTL_ADD_COUNTER_U64(ctx, child, OID_AUTO, "missing_intr", CTLFLAG_RD,
295 &dev_stats->missing_intr, "Missing interrupt count");
296 SYSCTL_ADD_COUNTER_U64(ctx, child, OID_AUTO, "tx_desc_malformed", CTLFLAG_RD,
297 &dev_stats->tx_desc_malformed, "TX descriptors malformed count");
298 SYSCTL_ADD_COUNTER_U64(ctx, child, OID_AUTO, "rx_desc_malformed", CTLFLAG_RD,
299 &dev_stats->rx_desc_malformed, "RX descriptors malformed count");
300 SYSCTL_ADD_COUNTER_U64(ctx, child, OID_AUTO, "missing_admin_interrupt", CTLFLAG_RD,
301 &dev_stats->missing_admin_interrupt, "Missing admin interrupts count");
302 SYSCTL_ADD_COUNTER_U64(ctx, child, OID_AUTO, "admin_to", CTLFLAG_RD,
303 &dev_stats->admin_to, "Admin queue timeouts count");
304 SYSCTL_ADD_COUNTER_U64(ctx, child, OID_AUTO, "device_request_reset", CTLFLAG_RD,
305 &dev_stats->device_request_reset, "Device reset requests count");
306 SYSCTL_ADD_COUNTER_U64(ctx, child, OID_AUTO, "total_resets", CTLFLAG_RD,
307 &dev_stats->total_resets, "Total resets count");
308
309 for (i = 0; i < adapter->num_io_queues; ++i, ++tx_ring, ++rx_ring) {
310 snprintf(namebuf, QUEUE_NAME_LEN, "queue%d", i);
311
312 queue_node = SYSCTL_ADD_NODE(ctx, child, OID_AUTO, namebuf,
313 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
314 queue_list = SYSCTL_CHILDREN(queue_node);
315
316 adapter->que[i].oid = queue_node;
317
318 #ifdef RSS
319 /* Common stats */
320 SYSCTL_ADD_INT(ctx, queue_list, OID_AUTO, "cpu", CTLFLAG_RD,
321 &adapter->que[i].cpu, 0, "CPU affinity");
322 SYSCTL_ADD_INT(ctx, queue_list, OID_AUTO, "domain", CTLFLAG_RD,
323 &adapter->que[i].domain, 0, "NUMA domain");
324 #endif
325
326 /* TX specific stats */
327 tx_node = SYSCTL_ADD_NODE(ctx, queue_list, OID_AUTO, "tx_ring",
328 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "TX ring");
329 tx_list = SYSCTL_CHILDREN(tx_node);
330
331 tx_stats = &tx_ring->tx_stats;
332
333 SYSCTL_ADD_COUNTER_U64(ctx, tx_list, OID_AUTO, "count",
334 CTLFLAG_RD, &tx_stats->cnt, "Packets sent");
335 SYSCTL_ADD_COUNTER_U64(ctx, tx_list, OID_AUTO, "bytes",
336 CTLFLAG_RD, &tx_stats->bytes, "Bytes sent");
337 SYSCTL_ADD_COUNTER_U64(ctx, tx_list, OID_AUTO,
338 "prepare_ctx_err", CTLFLAG_RD, &tx_stats->prepare_ctx_err,
339 "TX buffer preparation failures");
340 SYSCTL_ADD_COUNTER_U64(ctx, tx_list, OID_AUTO,
341 "dma_mapping_err", CTLFLAG_RD, &tx_stats->dma_mapping_err,
342 "DMA mapping failures");
343 SYSCTL_ADD_COUNTER_U64(ctx, tx_list, OID_AUTO, "doorbells",
344 CTLFLAG_RD, &tx_stats->doorbells, "Queue doorbells");
345 SYSCTL_ADD_COUNTER_U64(ctx, tx_list, OID_AUTO,
346 "missing_tx_comp", CTLFLAG_RD, &tx_stats->missing_tx_comp,
347 "TX completions missed");
348 SYSCTL_ADD_COUNTER_U64(ctx, tx_list, OID_AUTO, "bad_req_id",
349 CTLFLAG_RD, &tx_stats->bad_req_id, "Bad request id count");
350 SYSCTL_ADD_COUNTER_U64(ctx, tx_list, OID_AUTO, "mbuf_collapses",
351 CTLFLAG_RD, &tx_stats->collapse, "Mbuf collapse count");
352 SYSCTL_ADD_COUNTER_U64(ctx, tx_list, OID_AUTO,
353 "mbuf_collapse_err", CTLFLAG_RD, &tx_stats->collapse_err,
354 "Mbuf collapse failures");
355 SYSCTL_ADD_COUNTER_U64(ctx, tx_list, OID_AUTO, "queue_wakeups",
356 CTLFLAG_RD, &tx_stats->queue_wakeup, "Queue wakeups");
357 SYSCTL_ADD_COUNTER_U64(ctx, tx_list, OID_AUTO, "queue_stops",
358 CTLFLAG_RD, &tx_stats->queue_stop, "Queue stops");
359 SYSCTL_ADD_COUNTER_U64(ctx, tx_list, OID_AUTO,
360 "llq_buffer_copy", CTLFLAG_RD, &tx_stats->llq_buffer_copy,
361 "Header copies for llq transaction");
362 SYSCTL_ADD_COUNTER_U64(ctx, tx_list, OID_AUTO,
363 "unmask_interrupt_num", CTLFLAG_RD,
364 &tx_stats->unmask_interrupt_num,
365 "Unmasked interrupt count");
366
367 /* RX specific stats */
368 rx_node = SYSCTL_ADD_NODE(ctx, queue_list, OID_AUTO, "rx_ring",
369 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "RX ring");
370 rx_list = SYSCTL_CHILDREN(rx_node);
371
372 rx_stats = &rx_ring->rx_stats;
373
374 SYSCTL_ADD_COUNTER_U64(ctx, rx_list, OID_AUTO, "count",
375 CTLFLAG_RD, &rx_stats->cnt, "Packets received");
376 SYSCTL_ADD_COUNTER_U64(ctx, rx_list, OID_AUTO, "bytes",
377 CTLFLAG_RD, &rx_stats->bytes, "Bytes received");
378 SYSCTL_ADD_COUNTER_U64(ctx, rx_list, OID_AUTO, "refil_partial",
379 CTLFLAG_RD, &rx_stats->refil_partial,
380 "Partial refilled mbufs");
381 SYSCTL_ADD_COUNTER_U64(ctx, rx_list, OID_AUTO, "csum_bad",
382 CTLFLAG_RD, &rx_stats->csum_bad, "Bad RX checksum");
383 SYSCTL_ADD_COUNTER_U64(ctx, rx_list, OID_AUTO,
384 "mbuf_alloc_fail", CTLFLAG_RD, &rx_stats->mbuf_alloc_fail,
385 "Failed mbuf allocs");
386 SYSCTL_ADD_COUNTER_U64(ctx, rx_list, OID_AUTO,
387 "mjum_alloc_fail", CTLFLAG_RD, &rx_stats->mjum_alloc_fail,
388 "Failed jumbo mbuf allocs");
389 SYSCTL_ADD_COUNTER_U64(ctx, rx_list, OID_AUTO,
390 "dma_mapping_err", CTLFLAG_RD, &rx_stats->dma_mapping_err,
391 "DMA mapping errors");
392 SYSCTL_ADD_COUNTER_U64(ctx, rx_list, OID_AUTO, "bad_desc_num",
393 CTLFLAG_RD, &rx_stats->bad_desc_num,
394 "Bad descriptor count");
395 SYSCTL_ADD_COUNTER_U64(ctx, rx_list, OID_AUTO, "bad_req_id",
396 CTLFLAG_RD, &rx_stats->bad_req_id, "Bad request id count");
397 SYSCTL_ADD_COUNTER_U64(ctx, rx_list, OID_AUTO, "empty_rx_ring",
398 CTLFLAG_RD, &rx_stats->empty_rx_ring,
399 "RX descriptors depletion count");
400 SYSCTL_ADD_COUNTER_U64(ctx, rx_list, OID_AUTO, "csum_good",
401 CTLFLAG_RD, &rx_stats->csum_good,
402 "Valid RX checksum calculations");
403 }
404
405 /* Stats read from device */
406 hw_node = SYSCTL_ADD_NODE(ctx, child, OID_AUTO, "hw_stats",
407 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Statistics from hardware");
408 hw_list = SYSCTL_CHILDREN(hw_node);
409
410 SYSCTL_ADD_COUNTER_U64(ctx, hw_list, OID_AUTO, "rx_packets", CTLFLAG_RD,
411 &hw_stats->rx_packets, "Packets received");
412 SYSCTL_ADD_COUNTER_U64(ctx, hw_list, OID_AUTO, "tx_packets", CTLFLAG_RD,
413 &hw_stats->tx_packets, "Packets transmitted");
414 SYSCTL_ADD_COUNTER_U64(ctx, hw_list, OID_AUTO, "rx_bytes", CTLFLAG_RD,
415 &hw_stats->rx_bytes, "Bytes received");
416 SYSCTL_ADD_COUNTER_U64(ctx, hw_list, OID_AUTO, "tx_bytes", CTLFLAG_RD,
417 &hw_stats->tx_bytes, "Bytes transmitted");
418 SYSCTL_ADD_COUNTER_U64(ctx, hw_list, OID_AUTO, "rx_drops", CTLFLAG_RD,
419 &hw_stats->rx_drops, "Receive packet drops");
420 SYSCTL_ADD_COUNTER_U64(ctx, hw_list, OID_AUTO, "tx_drops", CTLFLAG_RD,
421 &hw_stats->tx_drops, "Transmit packet drops");
422
423 /* ENA Admin queue stats */
424 admin_node = SYSCTL_ADD_NODE(ctx, child, OID_AUTO, "admin_stats",
425 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "ENA Admin Queue statistics");
426 admin_list = SYSCTL_CHILDREN(admin_node);
427
428 SYSCTL_ADD_U64(ctx, admin_list, OID_AUTO, "aborted_cmd", CTLFLAG_RD,
429 &admin_stats->aborted_cmd, 0, "Aborted commands");
430 SYSCTL_ADD_U64(ctx, admin_list, OID_AUTO, "sumbitted_cmd", CTLFLAG_RD,
431 &admin_stats->submitted_cmd, 0, "Submitted commands");
432 SYSCTL_ADD_U64(ctx, admin_list, OID_AUTO, "completed_cmd", CTLFLAG_RD,
433 &admin_stats->completed_cmd, 0, "Completed commands");
434 SYSCTL_ADD_U64(ctx, admin_list, OID_AUTO, "out_of_space", CTLFLAG_RD,
435 &admin_stats->out_of_space, 0, "Queue out of space");
436 SYSCTL_ADD_U64(ctx, admin_list, OID_AUTO, "no_completion", CTLFLAG_RD,
437 &admin_stats->no_completion, 0, "Commands not completed");
438 }
439
440 static void
ena_sysctl_add_srd_info(struct ena_adapter * adapter)441 ena_sysctl_add_srd_info(struct ena_adapter *adapter)
442 {
443 device_t dev;
444
445 struct sysctl_oid *ena_srd_info;
446 struct sysctl_oid_list *srd_list;
447
448 struct sysctl_ctx_list *ctx;
449 struct sysctl_oid *tree;
450 struct sysctl_oid_list *child;
451
452 struct ena_admin_ena_srd_stats *srd_stats_ptr;
453 struct ena_srd_metrics cur_stat_strings;
454
455 int i;
456
457 dev = adapter->pdev;
458
459 ctx = device_get_sysctl_ctx(dev);
460 tree = device_get_sysctl_tree(dev);
461 child = SYSCTL_CHILDREN(tree);
462
463 ena_srd_info = SYSCTL_ADD_NODE(ctx, child, OID_AUTO, "ena_srd_info",
464 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "ENA's SRD information");
465 srd_list = SYSCTL_CHILDREN(ena_srd_info);
466
467 SYSCTL_ADD_U64(ctx, srd_list, OID_AUTO, "ena_srd_mode",
468 CTLFLAG_RD, &adapter->ena_srd_info.flags, 0,
469 "Describes which ENA-express features are enabled");
470
471 srd_stats_ptr = &adapter->ena_srd_info.ena_srd_stats;
472
473 for (i = 0 ; i < ENA_SRD_METRICS_ARRAY_SIZE; i++) {
474 cur_stat_strings = ena_srd_stats_strings[i];
475 SYSCTL_ADD_U64(ctx, srd_list, OID_AUTO, cur_stat_strings.name,
476 CTLFLAG_RD, (u64 *)srd_stats_ptr + cur_stat_strings.stat_offset,
477 0, cur_stat_strings.description);
478 }
479 }
480
481 static void
ena_sysctl_add_customer_metrics(struct ena_adapter * adapter)482 ena_sysctl_add_customer_metrics(struct ena_adapter *adapter)
483 {
484 device_t dev;
485 struct ena_com_dev *ena_dev;
486
487 struct sysctl_ctx_list *ctx;
488 struct sysctl_oid *tree;
489 struct sysctl_oid_list *child;
490
491 struct sysctl_oid *customer_metric;
492 struct sysctl_oid_list *customer_list;
493
494 int i;
495
496 dev = adapter->pdev;
497 ena_dev = adapter->ena_dev;
498
499 ctx = device_get_sysctl_ctx(dev);
500 tree = device_get_sysctl_tree(dev);
501 child = SYSCTL_CHILDREN(tree);
502 customer_metric = SYSCTL_ADD_NODE(ctx, child, OID_AUTO, "customer_metrics",
503 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "ENA's customer metrics");
504 customer_list = SYSCTL_CHILDREN(customer_metric);
505
506 for (i = 0; i < ENA_CUSTOMER_METRICS_ARRAY_SIZE; i++) {
507 if (ena_com_get_customer_metric_support(ena_dev, i)) {
508 SYSCTL_ADD_U64(ctx, customer_list, OID_AUTO, ena_hw_stats_strings[i].name,
509 CTLFLAG_RD, &adapter->customer_metrics_array[i], 0,
510 ena_hw_stats_strings[i].description);
511 }
512 }
513 }
514
515 static void
ena_sysctl_add_eni_metrics(struct ena_adapter * adapter)516 ena_sysctl_add_eni_metrics(struct ena_adapter *adapter)
517 {
518 device_t dev;
519 struct ena_admin_eni_stats *eni_metrics;
520
521 struct sysctl_ctx_list *ctx;
522 struct sysctl_oid *tree;
523 struct sysctl_oid_list *child;
524
525 struct sysctl_oid *eni_node;
526 struct sysctl_oid_list *eni_list;
527
528 dev = adapter->pdev;
529
530 ctx = device_get_sysctl_ctx(dev);
531 tree = device_get_sysctl_tree(dev);
532 child = SYSCTL_CHILDREN(tree);
533
534 eni_node = SYSCTL_ADD_NODE(ctx, child, OID_AUTO, "eni_metrics",
535 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "ENA's ENI metrics");
536 eni_list = SYSCTL_CHILDREN(eni_node);
537
538 eni_metrics = &adapter->eni_metrics;
539
540 SYSCTL_ADD_U64(ctx, eni_list, OID_AUTO, "bw_in_allowance_exceeded",
541 CTLFLAG_RD, &eni_metrics->bw_in_allowance_exceeded, 0,
542 "Inbound BW allowance exceeded");
543 SYSCTL_ADD_U64(ctx, eni_list, OID_AUTO, "bw_out_allowance_exceeded",
544 CTLFLAG_RD, &eni_metrics->bw_out_allowance_exceeded, 0,
545 "Outbound BW allowance exceeded");
546 SYSCTL_ADD_U64(ctx, eni_list, OID_AUTO, "pps_allowance_exceeded",
547 CTLFLAG_RD, &eni_metrics->pps_allowance_exceeded, 0,
548 "PPS allowance exceeded");
549 SYSCTL_ADD_U64(ctx, eni_list, OID_AUTO, "conntrack_allowance_exceeded",
550 CTLFLAG_RD, &eni_metrics->conntrack_allowance_exceeded, 0,
551 "Connection tracking allowance exceeded");
552 SYSCTL_ADD_U64(ctx, eni_list, OID_AUTO, "linklocal_allowance_exceeded",
553 CTLFLAG_RD, &eni_metrics->linklocal_allowance_exceeded, 0,
554 "Linklocal packet rate allowance exceeded");
555 }
556
557 static void
ena_sysctl_add_tuneables(struct ena_adapter * adapter)558 ena_sysctl_add_tuneables(struct ena_adapter *adapter)
559 {
560 device_t dev;
561
562 struct sysctl_ctx_list *ctx;
563 struct sysctl_oid *tree;
564 struct sysctl_oid_list *child;
565
566 dev = adapter->pdev;
567
568 ctx = device_get_sysctl_ctx(dev);
569 tree = device_get_sysctl_tree(dev);
570 child = SYSCTL_CHILDREN(tree);
571
572 /* Tuneable number of buffers in the buf-ring (drbr) */
573 SYSCTL_ADD_PROC(ctx, child, OID_AUTO, "buf_ring_size",
574 CTLTYPE_U32 | CTLFLAG_RW | CTLFLAG_MPSAFE, adapter, 0,
575 ena_sysctl_buf_ring_size, "I",
576 "Size of the Tx buffer ring (drbr).");
577
578 /* Tuneable number of the Rx ring size */
579 SYSCTL_ADD_PROC(ctx, child, OID_AUTO, "rx_queue_size",
580 CTLTYPE_U32 | CTLFLAG_RW | CTLFLAG_MPSAFE, adapter, 0,
581 ena_sysctl_rx_queue_size, "I",
582 "Size of the Rx ring. The size should be a power of 2.");
583
584 /* Tuneable number of IO queues */
585 SYSCTL_ADD_PROC(ctx, child, OID_AUTO, "io_queues_nb",
586 CTLTYPE_U32 | CTLFLAG_RW | CTLFLAG_MPSAFE, adapter, 0,
587 ena_sysctl_io_queues_nb, "I", "Number of IO queues.");
588
589 /*
590 * Tuneable, which determines how often ENA metrics will be read.
591 * 0 means it's turned off. Maximum allowed value is limited by:
592 * ENA_METRICS_MAX_SAMPLE_INTERVAL.
593 */
594 SYSCTL_ADD_PROC(ctx, child, OID_AUTO, "stats_sample_interval",
595 CTLTYPE_U16 | CTLFLAG_RW | CTLFLAG_MPSAFE, adapter, 0,
596 ena_sysctl_metrics_interval, "SU",
597 "Interval in seconds for updating Netword interface metrics. 0 turns off the update.");
598 }
599
600 /* Kernel option RSS prevents manipulation of key hash and indirection table. */
601 #ifndef RSS
602 static void
ena_sysctl_add_rss(struct ena_adapter * adapter)603 ena_sysctl_add_rss(struct ena_adapter *adapter)
604 {
605 device_t dev;
606
607 struct sysctl_ctx_list *ctx;
608 struct sysctl_oid *tree;
609 struct sysctl_oid_list *child;
610
611 dev = adapter->pdev;
612
613 ctx = device_get_sysctl_ctx(dev);
614 tree = device_get_sysctl_tree(dev);
615 child = SYSCTL_CHILDREN(tree);
616
617 /* RSS options */
618 tree = SYSCTL_ADD_NODE(ctx, child, OID_AUTO, "rss",
619 CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, "Receive Side Scaling options.");
620 child = SYSCTL_CHILDREN(tree);
621
622 /* RSS hash key */
623 SYSCTL_ADD_PROC(ctx, child, OID_AUTO, "key",
624 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, adapter, 0,
625 ena_sysctl_rss_key, "A", "RSS key.");
626
627 /* Tuneable RSS indirection table */
628 SYSCTL_ADD_PROC(ctx, child, OID_AUTO, "indir_table",
629 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, adapter, 0,
630 ena_sysctl_rss_indir_table, "A", "RSS indirection table.");
631
632 /* RSS indirection table size */
633 SYSCTL_ADD_INT(ctx, child, OID_AUTO, "indir_table_size",
634 CTLFLAG_RD | CTLFLAG_MPSAFE, &ena_rss_table_size, 0,
635 "RSS indirection table size.");
636 }
637 #endif /* RSS */
638
639 static void
ena_sysctl_add_irq_affinity(struct ena_adapter * adapter)640 ena_sysctl_add_irq_affinity(struct ena_adapter *adapter)
641 {
642 device_t dev;
643
644 struct sysctl_ctx_list *ctx;
645 struct sysctl_oid *tree;
646 struct sysctl_oid_list *child;
647
648 dev = adapter->pdev;
649
650 ctx = device_get_sysctl_ctx(dev);
651 tree = device_get_sysctl_tree(dev);
652 child = SYSCTL_CHILDREN(tree);
653
654 tree = SYSCTL_ADD_NODE(ctx, child, OID_AUTO, "irq_affinity",
655 CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, "Decide base CPU and stride for irqs affinity.");
656 child = SYSCTL_CHILDREN(tree);
657
658 /* Add base cpu leaf */
659 SYSCTL_ADD_PROC(ctx, child, OID_AUTO, "base_cpu",
660 CTLTYPE_S32 | CTLFLAG_RW | CTLFLAG_MPSAFE, adapter, 0,
661 ena_sysctl_irq_base_cpu, "I", "Base cpu index for setting irq affinity.");
662
663 /* Add cpu stride leaf */
664 SYSCTL_ADD_PROC(ctx, child, OID_AUTO, "cpu_stride",
665 CTLTYPE_S32 | CTLFLAG_RW | CTLFLAG_MPSAFE, adapter, 0,
666 ena_sysctl_irq_cpu_stride, "I", "Distance between irqs when setting affinity.");
667 }
668
669
670 /*
671 * ena_sysctl_update_queue_node_nb - Register/unregister sysctl queue nodes.
672 *
673 * Whether the nodes are registered or unregistered depends on a delta between
674 * the `old` and `new` parameters, representing the number of queues.
675 *
676 * This function is used to hide sysctl attributes for queue nodes which aren't
677 * currently used by the HW (e.g. after a call to `ena_sysctl_io_queues_nb`).
678 *
679 * NOTE:
680 * All unregistered nodes must be registered again at detach, i.e. by a call to
681 * this function.
682 */
683 void
ena_sysctl_update_queue_node_nb(struct ena_adapter * adapter,int old,int new)684 ena_sysctl_update_queue_node_nb(struct ena_adapter *adapter, int old, int new)
685 {
686 struct sysctl_oid *oid;
687 int min, max, i;
688
689 min = MIN(old, new);
690 max = MIN(MAX(old, new), adapter->max_num_io_queues);
691
692 for (i = min; i < max; ++i) {
693 oid = adapter->que[i].oid;
694
695 sysctl_wlock();
696 if (old > new)
697 sysctl_unregister_oid(oid);
698 else
699 sysctl_register_oid(oid);
700 sysctl_wunlock();
701 }
702 }
703
704 static int
ena_sysctl_buf_ring_size(SYSCTL_HANDLER_ARGS)705 ena_sysctl_buf_ring_size(SYSCTL_HANDLER_ARGS)
706 {
707 struct ena_adapter *adapter = arg1;
708 uint32_t val;
709 int error;
710
711 ENA_LOCK_LOCK();
712 if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_DEVICE_RUNNING, adapter))) {
713 error = EINVAL;
714 goto unlock;
715 }
716
717 val = 0;
718 error = sysctl_wire_old_buffer(req, sizeof(val));
719 if (error == 0) {
720 val = adapter->buf_ring_size;
721 error = sysctl_handle_32(oidp, &val, 0, req);
722 }
723 if (error != 0 || req->newptr == NULL)
724 goto unlock;
725
726 if (!powerof2(val) || val == 0) {
727 ena_log(adapter->pdev, ERR,
728 "Requested new Tx buffer ring size (%u) is not a power of 2\n",
729 val);
730 error = EINVAL;
731 goto unlock;
732 }
733
734 if (val != adapter->buf_ring_size) {
735 ena_log(adapter->pdev, INFO,
736 "Requested new Tx buffer ring size: %d. Old size: %d\n",
737 val, adapter->buf_ring_size);
738
739 error = ena_update_buf_ring_size(adapter, val);
740 } else {
741 ena_log(adapter->pdev, ERR,
742 "New Tx buffer ring size is the same as already used: %u\n",
743 adapter->buf_ring_size);
744 }
745
746 unlock:
747 ENA_LOCK_UNLOCK();
748
749 return (error);
750 }
751
752 static int
ena_sysctl_rx_queue_size(SYSCTL_HANDLER_ARGS)753 ena_sysctl_rx_queue_size(SYSCTL_HANDLER_ARGS)
754 {
755 struct ena_adapter *adapter = arg1;
756 uint32_t val;
757 int error;
758
759 ENA_LOCK_LOCK();
760 if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_DEVICE_RUNNING, adapter))) {
761 error = EINVAL;
762 goto unlock;
763 }
764
765 val = 0;
766 error = sysctl_wire_old_buffer(req, sizeof(val));
767 if (error == 0) {
768 val = adapter->requested_rx_ring_size;
769 error = sysctl_handle_32(oidp, &val, 0, req);
770 }
771 if (error != 0 || req->newptr == NULL)
772 goto unlock;
773
774 if (val < ENA_MIN_RING_SIZE || val > adapter->max_rx_ring_size) {
775 ena_log(adapter->pdev, ERR,
776 "Requested new Rx queue size (%u) is out of range: [%u, %u]\n",
777 val, ENA_MIN_RING_SIZE, adapter->max_rx_ring_size);
778 error = EINVAL;
779 goto unlock;
780 }
781
782 /* Check if the parameter is power of 2 */
783 if (!powerof2(val)) {
784 ena_log(adapter->pdev, ERR,
785 "Requested new Rx queue size (%u) is not a power of 2\n",
786 val);
787 error = EINVAL;
788 goto unlock;
789 }
790
791 if (val != adapter->requested_rx_ring_size) {
792 ena_log(adapter->pdev, INFO,
793 "Requested new Rx queue size: %u. Old size: %u\n", val,
794 adapter->requested_rx_ring_size);
795
796 error = ena_update_queue_size(adapter,
797 adapter->requested_tx_ring_size, val);
798 } else {
799 ena_log(adapter->pdev, ERR,
800 "New Rx queue size is the same as already used: %u\n",
801 adapter->requested_rx_ring_size);
802 }
803
804 unlock:
805 ENA_LOCK_UNLOCK();
806
807 return (error);
808 }
809
810 /*
811 * Change number of effectively used IO queues adapter->num_io_queues
812 */
813 static int
ena_sysctl_io_queues_nb(SYSCTL_HANDLER_ARGS)814 ena_sysctl_io_queues_nb(SYSCTL_HANDLER_ARGS)
815 {
816 struct ena_adapter *adapter = arg1;
817 uint32_t old_num_queues, tmp = 0;
818 int error;
819
820 ENA_LOCK_LOCK();
821 if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_DEVICE_RUNNING, adapter))) {
822 error = EINVAL;
823 goto unlock;
824 }
825
826 error = sysctl_wire_old_buffer(req, sizeof(tmp));
827 if (error == 0) {
828 tmp = adapter->num_io_queues;
829 error = sysctl_handle_int(oidp, &tmp, 0, req);
830 }
831 if (error != 0 || req->newptr == NULL)
832 goto unlock;
833
834 if (tmp == 0) {
835 ena_log(adapter->pdev, ERR,
836 "Requested number of IO queues is zero\n");
837 error = EINVAL;
838 goto unlock;
839 }
840
841 /*
842 * The adapter::max_num_io_queues is the HW capability. The system
843 * resources availability may potentially be a tighter limit. Therefore
844 * the relation `adapter::max_num_io_queues >= adapter::msix_vecs`
845 * always holds true, while the `adapter::msix_vecs` is variable across
846 * device reset (`ena_destroy_device()` + `ena_restore_device()`).
847 */
848 if (tmp > (adapter->msix_vecs - ENA_ADMIN_MSIX_VEC)) {
849 ena_log(adapter->pdev, ERR,
850 "Requested number of IO queues is higher than maximum allowed (%u)\n",
851 adapter->msix_vecs - ENA_ADMIN_MSIX_VEC);
852 error = EINVAL;
853 goto unlock;
854 }
855 if (tmp == adapter->num_io_queues) {
856 ena_log(adapter->pdev, ERR,
857 "Requested number of IO queues is equal to current value "
858 "(%u)\n",
859 adapter->num_io_queues);
860 } else {
861 ena_log(adapter->pdev, INFO,
862 "Requested new number of IO queues: %u, current value: "
863 "%u\n",
864 tmp, adapter->num_io_queues);
865
866 old_num_queues = adapter->num_io_queues;
867 error = ena_update_io_queue_nb(adapter, tmp);
868 if (error != 0)
869 return (error);
870
871 ena_sysctl_update_queue_node_nb(adapter, old_num_queues, tmp);
872 }
873
874 unlock:
875 ENA_LOCK_UNLOCK();
876
877 return (error);
878 }
879
880 static int
ena_sysctl_metrics_interval(SYSCTL_HANDLER_ARGS)881 ena_sysctl_metrics_interval(SYSCTL_HANDLER_ARGS)
882 {
883 struct ena_adapter *adapter = arg1;
884 uint16_t interval;
885 int error;
886
887 ENA_LOCK_LOCK();
888 if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_DEVICE_RUNNING, adapter))) {
889 error = EINVAL;
890 goto unlock;
891 }
892
893 error = sysctl_wire_old_buffer(req, sizeof(interval));
894 if (error == 0) {
895 interval = adapter->metrics_sample_interval;
896 error = sysctl_handle_16(oidp, &interval, 0, req);
897 }
898 if (error != 0 || req->newptr == NULL)
899 goto unlock;
900
901 if (interval > ENA_METRICS_MAX_SAMPLE_INTERVAL) {
902 ena_log(adapter->pdev, ERR,
903 "ENA metrics update interval is out of range - maximum allowed value: %d seconds\n",
904 ENA_METRICS_MAX_SAMPLE_INTERVAL);
905 error = EINVAL;
906 goto unlock;
907 }
908
909 if (interval == 0) {
910 ena_log(adapter->pdev, INFO,
911 "ENA metrics update is now turned off\n");
912 bzero(&adapter->eni_metrics, sizeof(adapter->eni_metrics));
913 } else {
914 ena_log(adapter->pdev, INFO,
915 "ENA metrics update interval is set to: %" PRIu16
916 " seconds\n",
917 interval);
918 }
919
920 adapter->metrics_sample_interval = interval;
921
922 unlock:
923 ENA_LOCK_UNLOCK();
924
925 return (0);
926 }
927
928 static int
ena_sysctl_irq_base_cpu(SYSCTL_HANDLER_ARGS)929 ena_sysctl_irq_base_cpu(SYSCTL_HANDLER_ARGS)
930 {
931 struct ena_adapter *adapter = arg1;
932 int irq_base_cpu = 0;
933 int error;
934
935 ENA_LOCK_LOCK();
936 if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_DEVICE_RUNNING, adapter))) {
937 error = ENODEV;
938 goto unlock;
939 }
940
941 error = sysctl_wire_old_buffer(req, sizeof(irq_base_cpu));
942 if (error == 0) {
943 irq_base_cpu = adapter->irq_cpu_base;
944 error = sysctl_handle_int(oidp, &irq_base_cpu, 0, req);
945 }
946 if (error != 0 || req->newptr == NULL)
947 goto unlock;
948
949 if (irq_base_cpu <= ENA_BASE_CPU_UNSPECIFIED) {
950 ena_log(adapter->pdev, ERR,
951 "Requested base CPU is less than zero.\n");
952 error = EINVAL;
953 goto unlock;
954 }
955
956 if (irq_base_cpu > mp_ncpus) {
957 ena_log(adapter->pdev, INFO,
958 "Requested base CPU is larger than the number of available CPUs. \n");
959 error = EINVAL;
960 goto unlock;
961
962 }
963
964 if (irq_base_cpu == adapter->irq_cpu_base) {
965 ena_log(adapter->pdev, INFO,
966 "Requested IRQ base CPU is equal to current value "
967 "(%d)\n",
968 adapter->irq_cpu_base);
969 goto unlock;
970 }
971
972 ena_log(adapter->pdev, INFO,
973 "Requested new IRQ base CPU: %d, current value: %d\n",
974 irq_base_cpu, adapter->irq_cpu_base);
975
976 error = ena_update_base_cpu(adapter, irq_base_cpu);
977
978 unlock:
979 ENA_LOCK_UNLOCK();
980
981 return (error);
982 }
983
984 static int
ena_sysctl_irq_cpu_stride(SYSCTL_HANDLER_ARGS)985 ena_sysctl_irq_cpu_stride(SYSCTL_HANDLER_ARGS)
986 {
987 struct ena_adapter *adapter = arg1;
988 int32_t irq_cpu_stride = 0;
989 int error;
990
991 ENA_LOCK_LOCK();
992 if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_DEVICE_RUNNING, adapter))) {
993 error = ENODEV;
994 goto unlock;
995 }
996
997 error = sysctl_wire_old_buffer(req, sizeof(irq_cpu_stride));
998 if (error == 0) {
999 irq_cpu_stride = adapter->irq_cpu_stride;
1000 error = sysctl_handle_int(oidp, &irq_cpu_stride, 0, req);
1001 }
1002 if (error != 0 || req->newptr == NULL)
1003 goto unlock;
1004
1005 if (irq_cpu_stride < 0) {
1006 ena_log(adapter->pdev, ERR,
1007 "Requested IRQ stride is less than zero.\n");
1008 error = EINVAL;
1009 goto unlock;
1010 }
1011
1012 if (irq_cpu_stride > mp_ncpus) {
1013 ena_log(adapter->pdev, INFO,
1014 "Warning: Requested IRQ stride is larger than the number of available CPUs.\n");
1015 }
1016
1017 if (irq_cpu_stride == adapter->irq_cpu_stride) {
1018 ena_log(adapter->pdev, INFO,
1019 "Requested IRQ CPU stride is equal to current value "
1020 "(%u)\n",
1021 adapter->irq_cpu_stride);
1022 goto unlock;
1023 }
1024
1025 ena_log(adapter->pdev, INFO,
1026 "Requested new IRQ CPU stride: %u, current value: %u\n",
1027 irq_cpu_stride, adapter->irq_cpu_stride);
1028
1029 error = ena_update_cpu_stride(adapter, irq_cpu_stride);
1030 if (error != 0)
1031 goto unlock;
1032
1033 unlock:
1034 ENA_LOCK_UNLOCK();
1035
1036 return (error);
1037 }
1038
1039 #ifndef RSS
1040 /*
1041 * Change the Receive Side Scaling hash key.
1042 */
1043 static int
ena_sysctl_rss_key(SYSCTL_HANDLER_ARGS)1044 ena_sysctl_rss_key(SYSCTL_HANDLER_ARGS)
1045 {
1046 struct ena_adapter *adapter = arg1;
1047 struct ena_com_dev *ena_dev = adapter->ena_dev;
1048 enum ena_admin_hash_functions ena_func;
1049 char msg[ENA_HASH_KEY_MSG_SIZE];
1050 char elem[3] = { 0 };
1051 char *endp;
1052 u8 rss_key[ENA_HASH_KEY_SIZE];
1053 int error, i;
1054
1055 ENA_LOCK_LOCK();
1056 if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_DEVICE_RUNNING, adapter))) {
1057 error = EINVAL;
1058 goto unlock;
1059 }
1060
1061 if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_RSS_ACTIVE, adapter))) {
1062 error = ENOTSUP;
1063 goto unlock;
1064 }
1065
1066 error = sysctl_wire_old_buffer(req, sizeof(msg));
1067 if (error != 0)
1068 goto unlock;
1069
1070 error = ena_com_get_hash_function(adapter->ena_dev, &ena_func);
1071 if (error != 0) {
1072 device_printf(adapter->pdev, "Cannot get hash function\n");
1073 goto unlock;
1074 }
1075
1076 if (ena_func != ENA_ADMIN_TOEPLITZ) {
1077 error = EINVAL;
1078 device_printf(adapter->pdev, "Unsupported hash algorithm\n");
1079 goto unlock;
1080 }
1081
1082 error = ena_rss_get_hash_key(ena_dev, rss_key);
1083 if (error != 0) {
1084 device_printf(adapter->pdev, "Cannot get hash key\n");
1085 goto unlock;
1086 }
1087
1088 for (i = 0; i < ENA_HASH_KEY_SIZE; ++i)
1089 snprintf(&msg[i * 2], 3, "%02x", rss_key[i]);
1090
1091 error = sysctl_handle_string(oidp, msg, sizeof(msg), req);
1092 if (error != 0 || req->newptr == NULL)
1093 goto unlock;
1094
1095 if (strlen(msg) != sizeof(msg) - 1) {
1096 error = EINVAL;
1097 device_printf(adapter->pdev, "Invalid key size\n");
1098 goto unlock;
1099 }
1100
1101 for (i = 0; i < ENA_HASH_KEY_SIZE; ++i) {
1102 strncpy(elem, &msg[i * 2], 2);
1103 rss_key[i] = strtol(elem, &endp, 16);
1104
1105 /* Both hex nibbles in the string must be valid to continue. */
1106 if (endp == elem || *endp != '\0' || rss_key[i] < 0) {
1107 error = EINVAL;
1108 device_printf(adapter->pdev,
1109 "Invalid key hex value: '%c'\n", *endp);
1110 goto unlock;
1111 }
1112 }
1113
1114 error = ena_rss_set_hash(ena_dev, rss_key);
1115 if (error != 0)
1116 device_printf(adapter->pdev, "Cannot fill hash key\n");
1117
1118 unlock:
1119 ENA_LOCK_UNLOCK();
1120
1121 return (error);
1122 }
1123
1124 /*
1125 * Change the Receive Side Scaling indirection table.
1126 *
1127 * The sysctl entry string consists of one or more `x:y` keypairs, where
1128 * x stands for the table index and y for its new value.
1129 * Table indices that don't need to be updated can be omitted from the string
1130 * and will retain their existing values. If an index is entered more than once,
1131 * the last value is used.
1132 *
1133 * Example:
1134 * To update two selected indices in the RSS indirection table, e.g. setting
1135 * index 0 to queue 5 and then index 5 to queue 0, the below command should be
1136 * used:
1137 * sysctl dev.ena.0.rss.indir_table="0:5 5:0"
1138 */
1139 static int
ena_sysctl_rss_indir_table(SYSCTL_HANDLER_ARGS)1140 ena_sysctl_rss_indir_table(SYSCTL_HANDLER_ARGS)
1141 {
1142 int num_queues, error;
1143 struct ena_adapter *adapter = arg1;
1144 struct ena_indir *indir;
1145 char *msg, *buf, *endp;
1146 uint32_t idx, value;
1147
1148 ENA_LOCK_LOCK();
1149 if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_DEVICE_RUNNING, adapter))) {
1150 error = EINVAL;
1151 goto unlock;
1152 }
1153
1154 if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_RSS_ACTIVE, adapter))) {
1155 error = ENOTSUP;
1156 goto unlock;
1157 }
1158
1159 indir = adapter->rss_indir;
1160 msg = indir->sysctl_buf;
1161
1162 if (unlikely(indir == NULL)) {
1163 error = ENOTSUP;
1164 goto unlock;
1165 }
1166
1167 error = sysctl_handle_string(oidp, msg, sizeof(indir->sysctl_buf), req);
1168 if (error != 0 || req->newptr == NULL)
1169 goto unlock;
1170
1171 num_queues = adapter->num_io_queues;
1172
1173 /*
1174 * This sysctl expects msg to be a list of `x:y` record pairs,
1175 * where x is the indirection table index and y is its value.
1176 */
1177 for (buf = msg; *buf != '\0'; buf = endp) {
1178 idx = strtol(buf, &endp, 10);
1179
1180 if (endp == buf || idx < 0) {
1181 device_printf(adapter->pdev, "Invalid index: %s\n",
1182 buf);
1183 error = EINVAL;
1184 break;
1185 }
1186
1187 if (idx >= ENA_RX_RSS_TABLE_SIZE) {
1188 device_printf(adapter->pdev, "Index %d out of range\n",
1189 idx);
1190 error = ERANGE;
1191 break;
1192 }
1193
1194 buf = endp;
1195
1196 if (*buf++ != ':') {
1197 device_printf(adapter->pdev, "Missing ':' separator\n");
1198 error = EINVAL;
1199 break;
1200 }
1201
1202 value = strtol(buf, &endp, 10);
1203
1204 if (endp == buf || value < 0) {
1205 device_printf(adapter->pdev, "Invalid value: %s\n",
1206 buf);
1207 error = EINVAL;
1208 break;
1209 }
1210
1211 if (value >= num_queues) {
1212 device_printf(adapter->pdev, "Value %d out of range\n",
1213 value);
1214 error = ERANGE;
1215 break;
1216 }
1217
1218 indir->table[idx] = value;
1219 }
1220
1221 if (error != 0) /* Reload indirection table with last good data. */
1222 ena_rss_indir_get(adapter, indir->table);
1223
1224 /* At this point msg has been clobbered by sysctl_handle_string. */
1225 ena_rss_copy_indir_buf(msg, indir->table);
1226
1227 if (error == 0)
1228 error = ena_rss_indir_set(adapter, indir->table);
1229
1230 unlock:
1231 ENA_LOCK_UNLOCK();
1232
1233 return (error);
1234 }
1235 #endif /* RSS */
1236