1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /* Copyright (c) 2019 Mellanox Technologies. */
3
4 #include "en/params.h"
5 #include "en/txrx.h"
6 #include "en/port.h"
7 #include "en_accel/en_accel.h"
8 #include "accel/ipsec.h"
9 #include "fpga/ipsec.h"
10
mlx5e_rx_is_xdp(struct mlx5e_params * params,struct mlx5e_xsk_param * xsk)11 static bool mlx5e_rx_is_xdp(struct mlx5e_params *params,
12 struct mlx5e_xsk_param *xsk)
13 {
14 return params->xdp_prog || xsk;
15 }
16
mlx5e_get_linear_rq_headroom(struct mlx5e_params * params,struct mlx5e_xsk_param * xsk)17 u16 mlx5e_get_linear_rq_headroom(struct mlx5e_params *params,
18 struct mlx5e_xsk_param *xsk)
19 {
20 u16 headroom;
21
22 if (xsk)
23 return xsk->headroom;
24
25 headroom = NET_IP_ALIGN;
26 if (mlx5e_rx_is_xdp(params, xsk))
27 headroom += XDP_PACKET_HEADROOM;
28 else
29 headroom += MLX5_RX_HEADROOM;
30
31 return headroom;
32 }
33
mlx5e_rx_get_min_frag_sz(struct mlx5e_params * params,struct mlx5e_xsk_param * xsk)34 u32 mlx5e_rx_get_min_frag_sz(struct mlx5e_params *params,
35 struct mlx5e_xsk_param *xsk)
36 {
37 u32 hw_mtu = MLX5E_SW2HW_MTU(params, params->sw_mtu);
38 u16 linear_rq_headroom = mlx5e_get_linear_rq_headroom(params, xsk);
39
40 return linear_rq_headroom + hw_mtu;
41 }
42
mlx5e_rx_get_linear_frag_sz(struct mlx5e_params * params,struct mlx5e_xsk_param * xsk)43 static u32 mlx5e_rx_get_linear_frag_sz(struct mlx5e_params *params,
44 struct mlx5e_xsk_param *xsk)
45 {
46 u32 frag_sz = mlx5e_rx_get_min_frag_sz(params, xsk);
47
48 /* AF_XDP doesn't build SKBs in place. */
49 if (!xsk)
50 frag_sz = MLX5_SKB_FRAG_SZ(frag_sz);
51
52 /* XDP in mlx5e doesn't support multiple packets per page. AF_XDP is a
53 * special case. It can run with frames smaller than a page, as it
54 * doesn't allocate pages dynamically. However, here we pretend that
55 * fragments are page-sized: it allows to treat XSK frames like pages
56 * by redirecting alloc and free operations to XSK rings and by using
57 * the fact there are no multiple packets per "page" (which is a frame).
58 * The latter is important, because frames may come in a random order,
59 * and we will have trouble assemblying a real page of multiple frames.
60 */
61 if (mlx5e_rx_is_xdp(params, xsk))
62 frag_sz = max_t(u32, frag_sz, PAGE_SIZE);
63
64 /* Even if we can go with a smaller fragment size, we must not put
65 * multiple packets into a single frame.
66 */
67 if (xsk)
68 frag_sz = max_t(u32, frag_sz, xsk->chunk_size);
69
70 return frag_sz;
71 }
72
mlx5e_mpwqe_log_pkts_per_wqe(struct mlx5e_params * params,struct mlx5e_xsk_param * xsk)73 u8 mlx5e_mpwqe_log_pkts_per_wqe(struct mlx5e_params *params,
74 struct mlx5e_xsk_param *xsk)
75 {
76 u32 linear_frag_sz = mlx5e_rx_get_linear_frag_sz(params, xsk);
77
78 return MLX5_MPWRQ_LOG_WQE_SZ - order_base_2(linear_frag_sz);
79 }
80
mlx5e_rx_is_linear_skb(struct mlx5e_params * params,struct mlx5e_xsk_param * xsk)81 bool mlx5e_rx_is_linear_skb(struct mlx5e_params *params,
82 struct mlx5e_xsk_param *xsk)
83 {
84 /* AF_XDP allocates SKBs on XDP_PASS - ensure they don't occupy more
85 * than one page. For this, check both with and without xsk.
86 */
87 u32 linear_frag_sz = max(mlx5e_rx_get_linear_frag_sz(params, xsk),
88 mlx5e_rx_get_linear_frag_sz(params, NULL));
89
90 return params->packet_merge.type == MLX5E_PACKET_MERGE_NONE &&
91 linear_frag_sz <= PAGE_SIZE;
92 }
93
mlx5e_verify_rx_mpwqe_strides(struct mlx5_core_dev * mdev,u8 log_stride_sz,u8 log_num_strides)94 bool mlx5e_verify_rx_mpwqe_strides(struct mlx5_core_dev *mdev,
95 u8 log_stride_sz, u8 log_num_strides)
96 {
97 if (log_stride_sz + log_num_strides != MLX5_MPWRQ_LOG_WQE_SZ)
98 return false;
99
100 if (log_stride_sz < MLX5_MPWQE_LOG_STRIDE_SZ_BASE ||
101 log_stride_sz > MLX5_MPWQE_LOG_STRIDE_SZ_MAX)
102 return false;
103
104 if (log_num_strides > MLX5_MPWQE_LOG_NUM_STRIDES_MAX)
105 return false;
106
107 if (MLX5_CAP_GEN(mdev, ext_stride_num_range))
108 return log_num_strides >= MLX5_MPWQE_LOG_NUM_STRIDES_EXT_BASE;
109
110 return log_num_strides >= MLX5_MPWQE_LOG_NUM_STRIDES_BASE;
111 }
112
mlx5e_rx_mpwqe_is_linear_skb(struct mlx5_core_dev * mdev,struct mlx5e_params * params,struct mlx5e_xsk_param * xsk)113 bool mlx5e_rx_mpwqe_is_linear_skb(struct mlx5_core_dev *mdev,
114 struct mlx5e_params *params,
115 struct mlx5e_xsk_param *xsk)
116 {
117 s8 log_num_strides;
118 u8 log_stride_sz;
119
120 if (!mlx5e_rx_is_linear_skb(params, xsk))
121 return false;
122
123 log_stride_sz = order_base_2(mlx5e_rx_get_linear_frag_sz(params, xsk));
124 log_num_strides = MLX5_MPWRQ_LOG_WQE_SZ - log_stride_sz;
125
126 return mlx5e_verify_rx_mpwqe_strides(mdev, log_stride_sz, log_num_strides);
127 }
128
mlx5e_mpwqe_get_log_rq_size(struct mlx5e_params * params,struct mlx5e_xsk_param * xsk)129 u8 mlx5e_mpwqe_get_log_rq_size(struct mlx5e_params *params,
130 struct mlx5e_xsk_param *xsk)
131 {
132 u8 log_pkts_per_wqe = mlx5e_mpwqe_log_pkts_per_wqe(params, xsk);
133
134 /* Numbers are unsigned, don't subtract to avoid underflow. */
135 if (params->log_rq_mtu_frames <
136 log_pkts_per_wqe + MLX5E_PARAMS_MINIMUM_LOG_RQ_SIZE_MPW)
137 return MLX5E_PARAMS_MINIMUM_LOG_RQ_SIZE_MPW;
138
139 return params->log_rq_mtu_frames - log_pkts_per_wqe;
140 }
141
mlx5e_shampo_get_log_hd_entry_size(struct mlx5_core_dev * mdev,struct mlx5e_params * params)142 u8 mlx5e_shampo_get_log_hd_entry_size(struct mlx5_core_dev *mdev,
143 struct mlx5e_params *params)
144 {
145 return order_base_2(DIV_ROUND_UP(MLX5E_RX_MAX_HEAD, MLX5E_SHAMPO_WQ_BASE_HEAD_ENTRY_SIZE));
146 }
147
mlx5e_shampo_get_log_rsrv_size(struct mlx5_core_dev * mdev,struct mlx5e_params * params)148 u8 mlx5e_shampo_get_log_rsrv_size(struct mlx5_core_dev *mdev,
149 struct mlx5e_params *params)
150 {
151 return order_base_2(MLX5E_SHAMPO_WQ_RESRV_SIZE / MLX5E_SHAMPO_WQ_BASE_RESRV_SIZE);
152 }
153
mlx5e_shampo_get_log_pkt_per_rsrv(struct mlx5_core_dev * mdev,struct mlx5e_params * params)154 u8 mlx5e_shampo_get_log_pkt_per_rsrv(struct mlx5_core_dev *mdev,
155 struct mlx5e_params *params)
156 {
157 u32 resrv_size = BIT(mlx5e_shampo_get_log_rsrv_size(mdev, params)) *
158 PAGE_SIZE;
159
160 return order_base_2(DIV_ROUND_UP(resrv_size, params->sw_mtu));
161 }
162
mlx5e_mpwqe_get_log_stride_size(struct mlx5_core_dev * mdev,struct mlx5e_params * params,struct mlx5e_xsk_param * xsk)163 u8 mlx5e_mpwqe_get_log_stride_size(struct mlx5_core_dev *mdev,
164 struct mlx5e_params *params,
165 struct mlx5e_xsk_param *xsk)
166 {
167 if (mlx5e_rx_mpwqe_is_linear_skb(mdev, params, xsk))
168 return order_base_2(mlx5e_rx_get_linear_frag_sz(params, xsk));
169
170 return MLX5_MPWRQ_DEF_LOG_STRIDE_SZ(mdev);
171 }
172
mlx5e_mpwqe_get_log_num_strides(struct mlx5_core_dev * mdev,struct mlx5e_params * params,struct mlx5e_xsk_param * xsk)173 u8 mlx5e_mpwqe_get_log_num_strides(struct mlx5_core_dev *mdev,
174 struct mlx5e_params *params,
175 struct mlx5e_xsk_param *xsk)
176 {
177 return MLX5_MPWRQ_LOG_WQE_SZ -
178 mlx5e_mpwqe_get_log_stride_size(mdev, params, xsk);
179 }
180
mlx5e_get_rq_headroom(struct mlx5_core_dev * mdev,struct mlx5e_params * params,struct mlx5e_xsk_param * xsk)181 u16 mlx5e_get_rq_headroom(struct mlx5_core_dev *mdev,
182 struct mlx5e_params *params,
183 struct mlx5e_xsk_param *xsk)
184 {
185 bool is_linear_skb = (params->rq_wq_type == MLX5_WQ_TYPE_CYCLIC) ?
186 mlx5e_rx_is_linear_skb(params, xsk) :
187 mlx5e_rx_mpwqe_is_linear_skb(mdev, params, xsk);
188
189 return is_linear_skb || params->packet_merge.type == MLX5E_PACKET_MERGE_SHAMPO ?
190 mlx5e_get_linear_rq_headroom(params, xsk) : 0;
191 }
192
mlx5e_calc_sq_stop_room(struct mlx5_core_dev * mdev,struct mlx5e_params * params)193 u16 mlx5e_calc_sq_stop_room(struct mlx5_core_dev *mdev, struct mlx5e_params *params)
194 {
195 bool is_mpwqe = MLX5E_GET_PFLAG(params, MLX5E_PFLAG_SKB_TX_MPWQE);
196 u16 stop_room;
197
198 stop_room = mlx5e_tls_get_stop_room(mdev, params);
199 stop_room += mlx5e_stop_room_for_wqe(MLX5_SEND_WQE_MAX_WQEBBS);
200 if (is_mpwqe)
201 /* A MPWQE can take up to the maximum-sized WQE + all the normal
202 * stop room can be taken if a new packet breaks the active
203 * MPWQE session and allocates its WQEs right away.
204 */
205 stop_room += mlx5e_stop_room_for_wqe(MLX5_SEND_WQE_MAX_WQEBBS);
206
207 return stop_room;
208 }
209
mlx5e_validate_params(struct mlx5_core_dev * mdev,struct mlx5e_params * params)210 int mlx5e_validate_params(struct mlx5_core_dev *mdev, struct mlx5e_params *params)
211 {
212 size_t sq_size = 1 << params->log_sq_size;
213 u16 stop_room;
214
215 stop_room = mlx5e_calc_sq_stop_room(mdev, params);
216 if (stop_room >= sq_size) {
217 mlx5_core_err(mdev, "Stop room %u is bigger than the SQ size %zu\n",
218 stop_room, sq_size);
219 return -EINVAL;
220 }
221
222 return 0;
223 }
224
mlx5e_get_def_tx_moderation(u8 cq_period_mode)225 static struct dim_cq_moder mlx5e_get_def_tx_moderation(u8 cq_period_mode)
226 {
227 struct dim_cq_moder moder = {};
228
229 moder.cq_period_mode = cq_period_mode;
230 moder.pkts = MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_PKTS;
231 moder.usec = MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC;
232 if (cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE)
233 moder.usec = MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC_FROM_CQE;
234
235 return moder;
236 }
237
mlx5e_get_def_rx_moderation(u8 cq_period_mode)238 static struct dim_cq_moder mlx5e_get_def_rx_moderation(u8 cq_period_mode)
239 {
240 struct dim_cq_moder moder = {};
241
242 moder.cq_period_mode = cq_period_mode;
243 moder.pkts = MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_PKTS;
244 moder.usec = MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC;
245 if (cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE)
246 moder.usec = MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC_FROM_CQE;
247
248 return moder;
249 }
250
mlx5_to_net_dim_cq_period_mode(u8 cq_period_mode)251 static u8 mlx5_to_net_dim_cq_period_mode(u8 cq_period_mode)
252 {
253 return cq_period_mode == MLX5_CQ_PERIOD_MODE_START_FROM_CQE ?
254 DIM_CQ_PERIOD_MODE_START_FROM_CQE :
255 DIM_CQ_PERIOD_MODE_START_FROM_EQE;
256 }
257
mlx5e_reset_tx_moderation(struct mlx5e_params * params,u8 cq_period_mode)258 void mlx5e_reset_tx_moderation(struct mlx5e_params *params, u8 cq_period_mode)
259 {
260 if (params->tx_dim_enabled) {
261 u8 dim_period_mode = mlx5_to_net_dim_cq_period_mode(cq_period_mode);
262
263 params->tx_cq_moderation = net_dim_get_def_tx_moderation(dim_period_mode);
264 } else {
265 params->tx_cq_moderation = mlx5e_get_def_tx_moderation(cq_period_mode);
266 }
267 }
268
mlx5e_reset_rx_moderation(struct mlx5e_params * params,u8 cq_period_mode)269 void mlx5e_reset_rx_moderation(struct mlx5e_params *params, u8 cq_period_mode)
270 {
271 if (params->rx_dim_enabled) {
272 u8 dim_period_mode = mlx5_to_net_dim_cq_period_mode(cq_period_mode);
273
274 params->rx_cq_moderation = net_dim_get_def_rx_moderation(dim_period_mode);
275 } else {
276 params->rx_cq_moderation = mlx5e_get_def_rx_moderation(cq_period_mode);
277 }
278 }
279
mlx5e_set_tx_cq_mode_params(struct mlx5e_params * params,u8 cq_period_mode)280 void mlx5e_set_tx_cq_mode_params(struct mlx5e_params *params, u8 cq_period_mode)
281 {
282 mlx5e_reset_tx_moderation(params, cq_period_mode);
283 MLX5E_SET_PFLAG(params, MLX5E_PFLAG_TX_CQE_BASED_MODER,
284 params->tx_cq_moderation.cq_period_mode ==
285 MLX5_CQ_PERIOD_MODE_START_FROM_CQE);
286 }
287
mlx5e_set_rx_cq_mode_params(struct mlx5e_params * params,u8 cq_period_mode)288 void mlx5e_set_rx_cq_mode_params(struct mlx5e_params *params, u8 cq_period_mode)
289 {
290 mlx5e_reset_rx_moderation(params, cq_period_mode);
291 MLX5E_SET_PFLAG(params, MLX5E_PFLAG_RX_CQE_BASED_MODER,
292 params->rx_cq_moderation.cq_period_mode ==
293 MLX5_CQ_PERIOD_MODE_START_FROM_CQE);
294 }
295
slow_pci_heuristic(struct mlx5_core_dev * mdev)296 bool slow_pci_heuristic(struct mlx5_core_dev *mdev)
297 {
298 u32 link_speed = 0;
299 u32 pci_bw = 0;
300
301 mlx5e_port_max_linkspeed(mdev, &link_speed);
302 pci_bw = pcie_bandwidth_available(mdev->pdev, NULL, NULL, NULL);
303 mlx5_core_dbg_once(mdev, "Max link speed = %d, PCI BW = %d\n",
304 link_speed, pci_bw);
305
306 #define MLX5E_SLOW_PCI_RATIO (2)
307
308 return link_speed && pci_bw &&
309 link_speed > MLX5E_SLOW_PCI_RATIO * pci_bw;
310 }
311
mlx5e_striding_rq_possible(struct mlx5_core_dev * mdev,struct mlx5e_params * params)312 bool mlx5e_striding_rq_possible(struct mlx5_core_dev *mdev,
313 struct mlx5e_params *params)
314 {
315 if (!mlx5e_check_fragmented_striding_rq_cap(mdev))
316 return false;
317
318 if (mlx5_fpga_is_ipsec_device(mdev))
319 return false;
320
321 if (params->xdp_prog) {
322 /* XSK params are not considered here. If striding RQ is in use,
323 * and an XSK is being opened, mlx5e_rx_mpwqe_is_linear_skb will
324 * be called with the known XSK params.
325 */
326 if (!mlx5e_rx_mpwqe_is_linear_skb(mdev, params, NULL))
327 return false;
328 }
329
330 return true;
331 }
332
mlx5e_init_rq_type_params(struct mlx5_core_dev * mdev,struct mlx5e_params * params)333 void mlx5e_init_rq_type_params(struct mlx5_core_dev *mdev,
334 struct mlx5e_params *params)
335 {
336 params->log_rq_mtu_frames = is_kdump_kernel() ?
337 MLX5E_PARAMS_MINIMUM_LOG_RQ_SIZE :
338 MLX5E_PARAMS_DEFAULT_LOG_RQ_SIZE;
339
340 mlx5_core_info(mdev, "MLX5E: StrdRq(%d) RqSz(%ld) StrdSz(%ld) RxCqeCmprss(%d)\n",
341 params->rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ,
342 params->rq_wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ ?
343 BIT(mlx5e_mpwqe_get_log_rq_size(params, NULL)) :
344 BIT(params->log_rq_mtu_frames),
345 BIT(mlx5e_mpwqe_get_log_stride_size(mdev, params, NULL)),
346 MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_CQE_COMPRESS));
347 }
348
mlx5e_set_rq_type(struct mlx5_core_dev * mdev,struct mlx5e_params * params)349 void mlx5e_set_rq_type(struct mlx5_core_dev *mdev, struct mlx5e_params *params)
350 {
351 params->rq_wq_type = mlx5e_striding_rq_possible(mdev, params) &&
352 MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_STRIDING_RQ) ?
353 MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ :
354 MLX5_WQ_TYPE_CYCLIC;
355 }
356
mlx5e_build_rq_params(struct mlx5_core_dev * mdev,struct mlx5e_params * params)357 void mlx5e_build_rq_params(struct mlx5_core_dev *mdev,
358 struct mlx5e_params *params)
359 {
360 /* Prefer Striding RQ, unless any of the following holds:
361 * - Striding RQ configuration is not possible/supported.
362 * - Slow PCI heuristic.
363 * - Legacy RQ would use linear SKB while Striding RQ would use non-linear.
364 *
365 * No XSK params: checking the availability of striding RQ in general.
366 */
367 if (!slow_pci_heuristic(mdev) &&
368 mlx5e_striding_rq_possible(mdev, params) &&
369 (mlx5e_rx_mpwqe_is_linear_skb(mdev, params, NULL) ||
370 !mlx5e_rx_is_linear_skb(params, NULL)))
371 MLX5E_SET_PFLAG(params, MLX5E_PFLAG_RX_STRIDING_RQ, true);
372 mlx5e_set_rq_type(mdev, params);
373 mlx5e_init_rq_type_params(mdev, params);
374 }
375
376 /* Build queue parameters */
377
mlx5e_build_create_cq_param(struct mlx5e_create_cq_param * ccp,struct mlx5e_channel * c)378 void mlx5e_build_create_cq_param(struct mlx5e_create_cq_param *ccp, struct mlx5e_channel *c)
379 {
380 *ccp = (struct mlx5e_create_cq_param) {
381 .napi = &c->napi,
382 .ch_stats = c->stats,
383 .node = cpu_to_node(c->cpu),
384 .ix = c->ix,
385 };
386 }
387
388 #define DEFAULT_FRAG_SIZE (2048)
389
mlx5e_build_rq_frags_info(struct mlx5_core_dev * mdev,struct mlx5e_params * params,struct mlx5e_xsk_param * xsk,struct mlx5e_rq_frags_info * info)390 static void mlx5e_build_rq_frags_info(struct mlx5_core_dev *mdev,
391 struct mlx5e_params *params,
392 struct mlx5e_xsk_param *xsk,
393 struct mlx5e_rq_frags_info *info)
394 {
395 u32 byte_count = MLX5E_SW2HW_MTU(params, params->sw_mtu);
396 int frag_size_max = DEFAULT_FRAG_SIZE;
397 u32 buf_size = 0;
398 int i;
399
400 if (mlx5_fpga_is_ipsec_device(mdev))
401 byte_count += MLX5E_METADATA_ETHER_LEN;
402
403 if (mlx5e_rx_is_linear_skb(params, xsk)) {
404 int frag_stride;
405
406 frag_stride = mlx5e_rx_get_linear_frag_sz(params, xsk);
407 frag_stride = roundup_pow_of_two(frag_stride);
408
409 info->arr[0].frag_size = byte_count;
410 info->arr[0].frag_stride = frag_stride;
411 info->num_frags = 1;
412 info->wqe_bulk = PAGE_SIZE / frag_stride;
413 goto out;
414 }
415
416 if (byte_count > PAGE_SIZE +
417 (MLX5E_MAX_RX_FRAGS - 1) * frag_size_max)
418 frag_size_max = PAGE_SIZE;
419
420 i = 0;
421 while (buf_size < byte_count) {
422 int frag_size = byte_count - buf_size;
423
424 if (i < MLX5E_MAX_RX_FRAGS - 1)
425 frag_size = min(frag_size, frag_size_max);
426
427 info->arr[i].frag_size = frag_size;
428 info->arr[i].frag_stride = roundup_pow_of_two(frag_size);
429
430 buf_size += frag_size;
431 i++;
432 }
433 info->num_frags = i;
434 /* number of different wqes sharing a page */
435 info->wqe_bulk = 1 + (info->num_frags % 2);
436
437 out:
438 info->wqe_bulk = max_t(u8, info->wqe_bulk, 8);
439 info->log_num_frags = order_base_2(info->num_frags);
440 }
441
mlx5e_get_rqwq_log_stride(u8 wq_type,int ndsegs)442 static u8 mlx5e_get_rqwq_log_stride(u8 wq_type, int ndsegs)
443 {
444 int sz = sizeof(struct mlx5_wqe_data_seg) * ndsegs;
445
446 switch (wq_type) {
447 case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
448 sz += sizeof(struct mlx5e_rx_wqe_ll);
449 break;
450 default: /* MLX5_WQ_TYPE_CYCLIC */
451 sz += sizeof(struct mlx5e_rx_wqe_cyc);
452 }
453
454 return order_base_2(sz);
455 }
456
mlx5e_build_common_cq_param(struct mlx5_core_dev * mdev,struct mlx5e_cq_param * param)457 static void mlx5e_build_common_cq_param(struct mlx5_core_dev *mdev,
458 struct mlx5e_cq_param *param)
459 {
460 void *cqc = param->cqc;
461
462 MLX5_SET(cqc, cqc, uar_page, mdev->priv.uar->index);
463 if (MLX5_CAP_GEN(mdev, cqe_128_always) && cache_line_size() >= 128)
464 MLX5_SET(cqc, cqc, cqe_sz, CQE_STRIDE_128_PAD);
465 }
466
mlx5e_shampo_get_log_cq_size(struct mlx5_core_dev * mdev,struct mlx5e_params * params,struct mlx5e_xsk_param * xsk)467 static u32 mlx5e_shampo_get_log_cq_size(struct mlx5_core_dev *mdev,
468 struct mlx5e_params *params,
469 struct mlx5e_xsk_param *xsk)
470 {
471 int rsrv_size = BIT(mlx5e_shampo_get_log_rsrv_size(mdev, params)) * PAGE_SIZE;
472 u16 num_strides = BIT(mlx5e_mpwqe_get_log_num_strides(mdev, params, xsk));
473 int pkt_per_rsrv = BIT(mlx5e_shampo_get_log_pkt_per_rsrv(mdev, params));
474 u8 log_stride_sz = mlx5e_mpwqe_get_log_stride_size(mdev, params, xsk);
475 int wq_size = BIT(mlx5e_mpwqe_get_log_rq_size(params, xsk));
476 int wqe_size = BIT(log_stride_sz) * num_strides;
477
478 /* +1 is for the case that the pkt_per_rsrv dont consume the reservation
479 * so we get a filler cqe for the rest of the reservation.
480 */
481 return order_base_2((wqe_size / rsrv_size) * wq_size * (pkt_per_rsrv + 1));
482 }
483
mlx5e_build_rx_cq_param(struct mlx5_core_dev * mdev,struct mlx5e_params * params,struct mlx5e_xsk_param * xsk,struct mlx5e_cq_param * param)484 static void mlx5e_build_rx_cq_param(struct mlx5_core_dev *mdev,
485 struct mlx5e_params *params,
486 struct mlx5e_xsk_param *xsk,
487 struct mlx5e_cq_param *param)
488 {
489 bool hw_stridx = false;
490 void *cqc = param->cqc;
491 u8 log_cq_size;
492
493 switch (params->rq_wq_type) {
494 case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
495 hw_stridx = MLX5_CAP_GEN(mdev, mini_cqe_resp_stride_index);
496 if (params->packet_merge.type == MLX5E_PACKET_MERGE_SHAMPO)
497 log_cq_size = mlx5e_shampo_get_log_cq_size(mdev, params, xsk);
498 else
499 log_cq_size = mlx5e_mpwqe_get_log_rq_size(params, xsk) +
500 mlx5e_mpwqe_get_log_num_strides(mdev, params, xsk);
501 break;
502 default: /* MLX5_WQ_TYPE_CYCLIC */
503 log_cq_size = params->log_rq_mtu_frames;
504 }
505
506 MLX5_SET(cqc, cqc, log_cq_size, log_cq_size);
507 if (MLX5E_GET_PFLAG(params, MLX5E_PFLAG_RX_CQE_COMPRESS)) {
508 MLX5_SET(cqc, cqc, mini_cqe_res_format, hw_stridx ?
509 MLX5_CQE_FORMAT_CSUM_STRIDX : MLX5_CQE_FORMAT_CSUM);
510 MLX5_SET(cqc, cqc, cqe_comp_en, 1);
511 }
512
513 mlx5e_build_common_cq_param(mdev, param);
514 param->cq_period_mode = params->rx_cq_moderation.cq_period_mode;
515 }
516
rq_end_pad_mode(struct mlx5_core_dev * mdev,struct mlx5e_params * params)517 static u8 rq_end_pad_mode(struct mlx5_core_dev *mdev, struct mlx5e_params *params)
518 {
519 bool lro_en = params->packet_merge.type == MLX5E_PACKET_MERGE_LRO;
520 bool ro = pcie_relaxed_ordering_enabled(mdev->pdev) &&
521 MLX5_CAP_GEN(mdev, relaxed_ordering_write);
522
523 return ro && lro_en ?
524 MLX5_WQ_END_PAD_MODE_NONE : MLX5_WQ_END_PAD_MODE_ALIGN;
525 }
526
mlx5e_build_rq_param(struct mlx5_core_dev * mdev,struct mlx5e_params * params,struct mlx5e_xsk_param * xsk,u16 q_counter,struct mlx5e_rq_param * param)527 int mlx5e_build_rq_param(struct mlx5_core_dev *mdev,
528 struct mlx5e_params *params,
529 struct mlx5e_xsk_param *xsk,
530 u16 q_counter,
531 struct mlx5e_rq_param *param)
532 {
533 void *rqc = param->rqc;
534 void *wq = MLX5_ADDR_OF(rqc, rqc, wq);
535 int ndsegs = 1;
536
537 switch (params->rq_wq_type) {
538 case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ: {
539 u8 log_wqe_num_of_strides = mlx5e_mpwqe_get_log_num_strides(mdev, params, xsk);
540 u8 log_wqe_stride_size = mlx5e_mpwqe_get_log_stride_size(mdev, params, xsk);
541
542 if (!mlx5e_verify_rx_mpwqe_strides(mdev, log_wqe_stride_size,
543 log_wqe_num_of_strides)) {
544 mlx5_core_err(mdev,
545 "Bad RX MPWQE params: log_stride_size %u, log_num_strides %u\n",
546 log_wqe_stride_size, log_wqe_num_of_strides);
547 return -EINVAL;
548 }
549
550 MLX5_SET(wq, wq, log_wqe_num_of_strides,
551 log_wqe_num_of_strides - MLX5_MPWQE_LOG_NUM_STRIDES_BASE);
552 MLX5_SET(wq, wq, log_wqe_stride_size,
553 log_wqe_stride_size - MLX5_MPWQE_LOG_STRIDE_SZ_BASE);
554 MLX5_SET(wq, wq, log_wq_sz, mlx5e_mpwqe_get_log_rq_size(params, xsk));
555 if (params->packet_merge.type == MLX5E_PACKET_MERGE_SHAMPO) {
556 MLX5_SET(wq, wq, shampo_enable, true);
557 MLX5_SET(wq, wq, log_reservation_size,
558 mlx5e_shampo_get_log_rsrv_size(mdev, params));
559 MLX5_SET(wq, wq,
560 log_max_num_of_packets_per_reservation,
561 mlx5e_shampo_get_log_pkt_per_rsrv(mdev, params));
562 MLX5_SET(wq, wq, log_headers_entry_size,
563 mlx5e_shampo_get_log_hd_entry_size(mdev, params));
564 MLX5_SET(rqc, rqc, reservation_timeout,
565 params->packet_merge.timeout);
566 MLX5_SET(rqc, rqc, shampo_match_criteria_type,
567 params->packet_merge.shampo.match_criteria_type);
568 MLX5_SET(rqc, rqc, shampo_no_match_alignment_granularity,
569 params->packet_merge.shampo.alignment_granularity);
570 }
571 break;
572 }
573 default: /* MLX5_WQ_TYPE_CYCLIC */
574 MLX5_SET(wq, wq, log_wq_sz, params->log_rq_mtu_frames);
575 mlx5e_build_rq_frags_info(mdev, params, xsk, ¶m->frags_info);
576 ndsegs = param->frags_info.num_frags;
577 }
578
579 MLX5_SET(wq, wq, wq_type, params->rq_wq_type);
580 MLX5_SET(wq, wq, end_padding_mode, rq_end_pad_mode(mdev, params));
581 MLX5_SET(wq, wq, log_wq_stride,
582 mlx5e_get_rqwq_log_stride(params->rq_wq_type, ndsegs));
583 MLX5_SET(wq, wq, pd, mdev->mlx5e_res.hw_objs.pdn);
584 MLX5_SET(rqc, rqc, counter_set_id, q_counter);
585 MLX5_SET(rqc, rqc, vsd, params->vlan_strip_disable);
586 MLX5_SET(rqc, rqc, scatter_fcs, params->scatter_fcs_en);
587
588 param->wq.buf_numa_node = dev_to_node(mlx5_core_dma_dev(mdev));
589 mlx5e_build_rx_cq_param(mdev, params, xsk, ¶m->cqp);
590
591 return 0;
592 }
593
mlx5e_build_drop_rq_param(struct mlx5_core_dev * mdev,u16 q_counter,struct mlx5e_rq_param * param)594 void mlx5e_build_drop_rq_param(struct mlx5_core_dev *mdev,
595 u16 q_counter,
596 struct mlx5e_rq_param *param)
597 {
598 void *rqc = param->rqc;
599 void *wq = MLX5_ADDR_OF(rqc, rqc, wq);
600
601 MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
602 MLX5_SET(wq, wq, log_wq_stride,
603 mlx5e_get_rqwq_log_stride(MLX5_WQ_TYPE_CYCLIC, 1));
604 MLX5_SET(rqc, rqc, counter_set_id, q_counter);
605
606 param->wq.buf_numa_node = dev_to_node(mlx5_core_dma_dev(mdev));
607 }
608
mlx5e_build_tx_cq_param(struct mlx5_core_dev * mdev,struct mlx5e_params * params,struct mlx5e_cq_param * param)609 void mlx5e_build_tx_cq_param(struct mlx5_core_dev *mdev,
610 struct mlx5e_params *params,
611 struct mlx5e_cq_param *param)
612 {
613 void *cqc = param->cqc;
614
615 MLX5_SET(cqc, cqc, log_cq_size, params->log_sq_size);
616
617 mlx5e_build_common_cq_param(mdev, param);
618 param->cq_period_mode = params->tx_cq_moderation.cq_period_mode;
619 }
620
mlx5e_build_sq_param_common(struct mlx5_core_dev * mdev,struct mlx5e_sq_param * param)621 void mlx5e_build_sq_param_common(struct mlx5_core_dev *mdev,
622 struct mlx5e_sq_param *param)
623 {
624 void *sqc = param->sqc;
625 void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
626
627 MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
628 MLX5_SET(wq, wq, pd, mdev->mlx5e_res.hw_objs.pdn);
629
630 param->wq.buf_numa_node = dev_to_node(mlx5_core_dma_dev(mdev));
631 }
632
mlx5e_build_sq_param(struct mlx5_core_dev * mdev,struct mlx5e_params * params,struct mlx5e_sq_param * param)633 void mlx5e_build_sq_param(struct mlx5_core_dev *mdev,
634 struct mlx5e_params *params,
635 struct mlx5e_sq_param *param)
636 {
637 void *sqc = param->sqc;
638 void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
639 bool allow_swp;
640
641 allow_swp = mlx5_geneve_tx_allowed(mdev) ||
642 !!MLX5_IPSEC_DEV(mdev);
643 mlx5e_build_sq_param_common(mdev, param);
644 MLX5_SET(wq, wq, log_wq_sz, params->log_sq_size);
645 MLX5_SET(sqc, sqc, allow_swp, allow_swp);
646 param->is_mpw = MLX5E_GET_PFLAG(params, MLX5E_PFLAG_SKB_TX_MPWQE);
647 param->stop_room = mlx5e_calc_sq_stop_room(mdev, params);
648 mlx5e_build_tx_cq_param(mdev, params, ¶m->cqp);
649 }
650
mlx5e_build_ico_cq_param(struct mlx5_core_dev * mdev,u8 log_wq_size,struct mlx5e_cq_param * param)651 static void mlx5e_build_ico_cq_param(struct mlx5_core_dev *mdev,
652 u8 log_wq_size,
653 struct mlx5e_cq_param *param)
654 {
655 void *cqc = param->cqc;
656
657 MLX5_SET(cqc, cqc, log_cq_size, log_wq_size);
658
659 mlx5e_build_common_cq_param(mdev, param);
660
661 param->cq_period_mode = DIM_CQ_PERIOD_MODE_START_FROM_EQE;
662 }
663
mlx5e_get_rq_log_wq_sz(void * rqc)664 static u8 mlx5e_get_rq_log_wq_sz(void *rqc)
665 {
666 void *wq = MLX5_ADDR_OF(rqc, rqc, wq);
667
668 return MLX5_GET(wq, wq, log_wq_sz);
669 }
670
671 /* This function calculates the maximum number of headers entries that are needed
672 * per WQE, the formula is based on the size of the reservations and the
673 * restriction we have about max packets for reservation that is equal to max
674 * headers per reservation.
675 */
mlx5e_shampo_hd_per_wqe(struct mlx5_core_dev * mdev,struct mlx5e_params * params,struct mlx5e_rq_param * rq_param)676 u32 mlx5e_shampo_hd_per_wqe(struct mlx5_core_dev *mdev,
677 struct mlx5e_params *params,
678 struct mlx5e_rq_param *rq_param)
679 {
680 int resv_size = BIT(mlx5e_shampo_get_log_rsrv_size(mdev, params)) * PAGE_SIZE;
681 u16 num_strides = BIT(mlx5e_mpwqe_get_log_num_strides(mdev, params, NULL));
682 int pkt_per_resv = BIT(mlx5e_shampo_get_log_pkt_per_rsrv(mdev, params));
683 u8 log_stride_sz = mlx5e_mpwqe_get_log_stride_size(mdev, params, NULL);
684 int wqe_size = BIT(log_stride_sz) * num_strides;
685 u32 hd_per_wqe;
686
687 /* Assumption: hd_per_wqe % 8 == 0. */
688 hd_per_wqe = (wqe_size / resv_size) * pkt_per_resv;
689 mlx5_core_dbg(mdev, "%s hd_per_wqe = %d rsrv_size = %d wqe_size = %d pkt_per_resv = %d\n",
690 __func__, hd_per_wqe, resv_size, wqe_size, pkt_per_resv);
691 return hd_per_wqe;
692 }
693
694 /* This function calculates the maximum number of headers entries that are needed
695 * for the WQ, this value is uesed to allocate the header buffer in HW, thus
696 * must be a pow of 2.
697 */
mlx5e_shampo_hd_per_wq(struct mlx5_core_dev * mdev,struct mlx5e_params * params,struct mlx5e_rq_param * rq_param)698 u32 mlx5e_shampo_hd_per_wq(struct mlx5_core_dev *mdev,
699 struct mlx5e_params *params,
700 struct mlx5e_rq_param *rq_param)
701 {
702 void *wqc = MLX5_ADDR_OF(rqc, rq_param->rqc, wq);
703 int wq_size = BIT(MLX5_GET(wq, wqc, log_wq_sz));
704 u32 hd_per_wqe, hd_per_wq;
705
706 hd_per_wqe = mlx5e_shampo_hd_per_wqe(mdev, params, rq_param);
707 hd_per_wq = roundup_pow_of_two(hd_per_wqe * wq_size);
708 return hd_per_wq;
709 }
710
mlx5e_shampo_icosq_sz(struct mlx5_core_dev * mdev,struct mlx5e_params * params,struct mlx5e_rq_param * rq_param)711 static u32 mlx5e_shampo_icosq_sz(struct mlx5_core_dev *mdev,
712 struct mlx5e_params *params,
713 struct mlx5e_rq_param *rq_param)
714 {
715 int max_num_of_umr_per_wqe, max_hd_per_wqe, max_klm_per_umr, rest;
716 void *wqc = MLX5_ADDR_OF(rqc, rq_param->rqc, wq);
717 int wq_size = BIT(MLX5_GET(wq, wqc, log_wq_sz));
718 u32 wqebbs;
719
720 max_klm_per_umr = MLX5E_MAX_KLM_PER_WQE(mdev);
721 max_hd_per_wqe = mlx5e_shampo_hd_per_wqe(mdev, params, rq_param);
722 max_num_of_umr_per_wqe = max_hd_per_wqe / max_klm_per_umr;
723 rest = max_hd_per_wqe % max_klm_per_umr;
724 wqebbs = MLX5E_KLM_UMR_WQEBBS(max_klm_per_umr) * max_num_of_umr_per_wqe;
725 if (rest)
726 wqebbs += MLX5E_KLM_UMR_WQEBBS(rest);
727 wqebbs *= wq_size;
728 return wqebbs;
729 }
730
mlx5e_build_icosq_log_wq_sz(struct mlx5_core_dev * mdev,struct mlx5e_params * params,struct mlx5e_rq_param * rqp)731 static u8 mlx5e_build_icosq_log_wq_sz(struct mlx5_core_dev *mdev,
732 struct mlx5e_params *params,
733 struct mlx5e_rq_param *rqp)
734 {
735 u32 wqebbs;
736
737 /* MLX5_WQ_TYPE_CYCLIC */
738 if (params->rq_wq_type != MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ)
739 return MLX5E_PARAMS_MINIMUM_LOG_SQ_SIZE;
740
741 wqebbs = MLX5E_UMR_WQEBBS * BIT(mlx5e_get_rq_log_wq_sz(rqp->rqc));
742 if (params->packet_merge.type == MLX5E_PACKET_MERGE_SHAMPO)
743 wqebbs += mlx5e_shampo_icosq_sz(mdev, params, rqp);
744 return max_t(u8, MLX5E_PARAMS_MINIMUM_LOG_SQ_SIZE, order_base_2(wqebbs));
745 }
746
mlx5e_build_async_icosq_log_wq_sz(struct mlx5_core_dev * mdev)747 static u8 mlx5e_build_async_icosq_log_wq_sz(struct mlx5_core_dev *mdev)
748 {
749 if (mlx5e_accel_is_ktls_rx(mdev))
750 return MLX5E_PARAMS_DEFAULT_LOG_SQ_SIZE;
751
752 return MLX5E_PARAMS_MINIMUM_LOG_SQ_SIZE;
753 }
754
mlx5e_build_icosq_param(struct mlx5_core_dev * mdev,u8 log_wq_size,struct mlx5e_sq_param * param)755 static void mlx5e_build_icosq_param(struct mlx5_core_dev *mdev,
756 u8 log_wq_size,
757 struct mlx5e_sq_param *param)
758 {
759 void *sqc = param->sqc;
760 void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
761
762 mlx5e_build_sq_param_common(mdev, param);
763
764 MLX5_SET(wq, wq, log_wq_sz, log_wq_size);
765 MLX5_SET(sqc, sqc, reg_umr, MLX5_CAP_ETH(mdev, reg_umr_sq));
766 mlx5e_build_ico_cq_param(mdev, log_wq_size, ¶m->cqp);
767 }
768
mlx5e_build_async_icosq_param(struct mlx5_core_dev * mdev,u8 log_wq_size,struct mlx5e_sq_param * param)769 static void mlx5e_build_async_icosq_param(struct mlx5_core_dev *mdev,
770 u8 log_wq_size,
771 struct mlx5e_sq_param *param)
772 {
773 void *sqc = param->sqc;
774 void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
775
776 mlx5e_build_sq_param_common(mdev, param);
777 param->stop_room = mlx5e_stop_room_for_wqe(1); /* for XSK NOP */
778 param->is_tls = mlx5e_accel_is_ktls_rx(mdev);
779 if (param->is_tls)
780 param->stop_room += mlx5e_stop_room_for_wqe(1); /* for TLS RX resync NOP */
781 MLX5_SET(sqc, sqc, reg_umr, MLX5_CAP_ETH(mdev, reg_umr_sq));
782 MLX5_SET(wq, wq, log_wq_sz, log_wq_size);
783 mlx5e_build_ico_cq_param(mdev, log_wq_size, ¶m->cqp);
784 }
785
mlx5e_build_xdpsq_param(struct mlx5_core_dev * mdev,struct mlx5e_params * params,struct mlx5e_sq_param * param)786 void mlx5e_build_xdpsq_param(struct mlx5_core_dev *mdev,
787 struct mlx5e_params *params,
788 struct mlx5e_sq_param *param)
789 {
790 void *sqc = param->sqc;
791 void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
792
793 mlx5e_build_sq_param_common(mdev, param);
794 MLX5_SET(wq, wq, log_wq_sz, params->log_sq_size);
795 param->is_mpw = MLX5E_GET_PFLAG(params, MLX5E_PFLAG_XDP_TX_MPWQE);
796 mlx5e_build_tx_cq_param(mdev, params, ¶m->cqp);
797 }
798
mlx5e_build_channel_param(struct mlx5_core_dev * mdev,struct mlx5e_params * params,u16 q_counter,struct mlx5e_channel_param * cparam)799 int mlx5e_build_channel_param(struct mlx5_core_dev *mdev,
800 struct mlx5e_params *params,
801 u16 q_counter,
802 struct mlx5e_channel_param *cparam)
803 {
804 u8 icosq_log_wq_sz, async_icosq_log_wq_sz;
805 int err;
806
807 err = mlx5e_build_rq_param(mdev, params, NULL, q_counter, &cparam->rq);
808 if (err)
809 return err;
810
811 icosq_log_wq_sz = mlx5e_build_icosq_log_wq_sz(mdev, params, &cparam->rq);
812 async_icosq_log_wq_sz = mlx5e_build_async_icosq_log_wq_sz(mdev);
813
814 mlx5e_build_sq_param(mdev, params, &cparam->txq_sq);
815 mlx5e_build_xdpsq_param(mdev, params, &cparam->xdp_sq);
816 mlx5e_build_icosq_param(mdev, icosq_log_wq_sz, &cparam->icosq);
817 mlx5e_build_async_icosq_param(mdev, async_icosq_log_wq_sz, &cparam->async_icosq);
818
819 return 0;
820 }
821