1 // SPDX-License-Identifier: GPL-2.0
2 /* Marvell RVU Ethernet driver
3  *
4  * Copyright (C) 2021 Marvell.
5  *
6  */
7 
8 #include "cn10k.h"
9 #include "otx2_reg.h"
10 #include "otx2_struct.h"
11 
12 static struct dev_hw_ops	otx2_hw_ops = {
13 	.sq_aq_init = otx2_sq_aq_init,
14 	.sqe_flush = otx2_sqe_flush,
15 	.aura_freeptr = otx2_aura_freeptr,
16 	.refill_pool_ptrs = otx2_refill_pool_ptrs,
17 };
18 
19 static struct dev_hw_ops cn10k_hw_ops = {
20 	.sq_aq_init = cn10k_sq_aq_init,
21 	.sqe_flush = cn10k_sqe_flush,
22 	.aura_freeptr = cn10k_aura_freeptr,
23 	.refill_pool_ptrs = cn10k_refill_pool_ptrs,
24 };
25 
26 int cn10k_lmtst_init(struct otx2_nic *pfvf)
27 {
28 
29 	struct lmtst_tbl_setup_req *req;
30 	struct otx2_lmt_info *lmt_info;
31 	int err, cpu;
32 
33 	if (!test_bit(CN10K_LMTST, &pfvf->hw.cap_flag)) {
34 		pfvf->hw_ops = &otx2_hw_ops;
35 		return 0;
36 	}
37 
38 	pfvf->hw_ops = &cn10k_hw_ops;
39 	/* Total LMTLINES = num_online_cpus() * 32 (For Burst flush).*/
40 	pfvf->tot_lmt_lines = (num_online_cpus() * LMT_BURST_SIZE);
41 	pfvf->hw.lmt_info = alloc_percpu(struct otx2_lmt_info);
42 
43 	mutex_lock(&pfvf->mbox.lock);
44 	req = otx2_mbox_alloc_msg_lmtst_tbl_setup(&pfvf->mbox);
45 	if (!req) {
46 		mutex_unlock(&pfvf->mbox.lock);
47 		return -ENOMEM;
48 	}
49 
50 	req->use_local_lmt_region = true;
51 
52 	err = qmem_alloc(pfvf->dev, &pfvf->dync_lmt, pfvf->tot_lmt_lines,
53 			 LMT_LINE_SIZE);
54 	if (err) {
55 		mutex_unlock(&pfvf->mbox.lock);
56 		return err;
57 	}
58 	pfvf->hw.lmt_base = (u64 *)pfvf->dync_lmt->base;
59 	req->lmt_iova = (u64)pfvf->dync_lmt->iova;
60 
61 	err = otx2_sync_mbox_msg(&pfvf->mbox);
62 	mutex_unlock(&pfvf->mbox.lock);
63 
64 	for_each_possible_cpu(cpu) {
65 		lmt_info = per_cpu_ptr(pfvf->hw.lmt_info, cpu);
66 		lmt_info->lmt_addr = ((u64)pfvf->hw.lmt_base +
67 				      (cpu * LMT_BURST_SIZE * LMT_LINE_SIZE));
68 		lmt_info->lmt_id = cpu * LMT_BURST_SIZE;
69 	}
70 
71 	return 0;
72 }
73 EXPORT_SYMBOL(cn10k_lmtst_init);
74 
75 int cn10k_sq_aq_init(void *dev, u16 qidx, u16 sqb_aura)
76 {
77 	struct nix_cn10k_aq_enq_req *aq;
78 	struct otx2_nic *pfvf = dev;
79 
80 	/* Get memory to put this msg */
81 	aq = otx2_mbox_alloc_msg_nix_cn10k_aq_enq(&pfvf->mbox);
82 	if (!aq)
83 		return -ENOMEM;
84 
85 	aq->sq.cq = pfvf->hw.rx_queues + qidx;
86 	aq->sq.max_sqe_size = NIX_MAXSQESZ_W16; /* 128 byte */
87 	aq->sq.cq_ena = 1;
88 	aq->sq.ena = 1;
89 	aq->sq.smq = otx2_get_smq_idx(pfvf, qidx);
90 	aq->sq.smq_rr_weight = mtu_to_dwrr_weight(pfvf, pfvf->tx_max_pktlen);
91 	aq->sq.default_chan = pfvf->hw.tx_chan_base;
92 	aq->sq.sqe_stype = NIX_STYPE_STF; /* Cache SQB */
93 	aq->sq.sqb_aura = sqb_aura;
94 	aq->sq.sq_int_ena = NIX_SQINT_BITS;
95 	aq->sq.qint_idx = 0;
96 	/* Due pipelining impact minimum 2000 unused SQ CQE's
97 	 * need to maintain to avoid CQ overflow.
98 	 */
99 	aq->sq.cq_limit = ((SEND_CQ_SKID * 256) / (pfvf->qset.sqe_cnt));
100 
101 	/* Fill AQ info */
102 	aq->qidx = qidx;
103 	aq->ctype = NIX_AQ_CTYPE_SQ;
104 	aq->op = NIX_AQ_INSTOP_INIT;
105 
106 	return otx2_sync_mbox_msg(&pfvf->mbox);
107 }
108 
109 #define NPA_MAX_BURST 16
110 int cn10k_refill_pool_ptrs(void *dev, struct otx2_cq_queue *cq)
111 {
112 	struct otx2_nic *pfvf = dev;
113 	int cnt = cq->pool_ptrs;
114 	u64 ptrs[NPA_MAX_BURST];
115 	dma_addr_t bufptr;
116 	int num_ptrs = 1;
117 
118 	/* Refill pool with new buffers */
119 	while (cq->pool_ptrs) {
120 		if (otx2_alloc_buffer(pfvf, cq, &bufptr)) {
121 			if (num_ptrs--)
122 				__cn10k_aura_freeptr(pfvf, cq->cq_idx, ptrs,
123 						     num_ptrs);
124 			break;
125 		}
126 		cq->pool_ptrs--;
127 		ptrs[num_ptrs] = (u64)bufptr + OTX2_HEAD_ROOM;
128 		num_ptrs++;
129 		if (num_ptrs == NPA_MAX_BURST || cq->pool_ptrs == 0) {
130 			__cn10k_aura_freeptr(pfvf, cq->cq_idx, ptrs,
131 					     num_ptrs);
132 			num_ptrs = 1;
133 		}
134 	}
135 	return cnt - cq->pool_ptrs;
136 }
137 
138 void cn10k_sqe_flush(void *dev, struct otx2_snd_queue *sq, int size, int qidx)
139 {
140 	struct otx2_lmt_info *lmt_info;
141 	struct otx2_nic *pfvf = dev;
142 	u64 val = 0, tar_addr = 0;
143 
144 	lmt_info = per_cpu_ptr(pfvf->hw.lmt_info, smp_processor_id());
145 	/* FIXME: val[0:10] LMT_ID.
146 	 * [12:15] no of LMTST - 1 in the burst.
147 	 * [19:63] data size of each LMTST in the burst except first.
148 	 */
149 	val = (lmt_info->lmt_id & 0x7FF);
150 	/* Target address for LMTST flush tells HW how many 128bit
151 	 * words are present.
152 	 * tar_addr[6:4] size of first LMTST - 1 in units of 128b.
153 	 */
154 	tar_addr |= sq->io_addr | (((size / 16) - 1) & 0x7) << 4;
155 	dma_wmb();
156 	memcpy((u64 *)lmt_info->lmt_addr, sq->sqe_base, size);
157 	cn10k_lmt_flush(val, tar_addr);
158 
159 	sq->head++;
160 	sq->head &= (sq->sqe_cnt - 1);
161 }
162 
163 int cn10k_free_all_ipolicers(struct otx2_nic *pfvf)
164 {
165 	struct nix_bandprof_free_req *req;
166 	int rc;
167 
168 	if (is_dev_otx2(pfvf->pdev))
169 		return 0;
170 
171 	mutex_lock(&pfvf->mbox.lock);
172 
173 	req = otx2_mbox_alloc_msg_nix_bandprof_free(&pfvf->mbox);
174 	if (!req) {
175 		rc =  -ENOMEM;
176 		goto out;
177 	}
178 
179 	/* Free all bandwidth profiles allocated */
180 	req->free_all = true;
181 
182 	rc = otx2_sync_mbox_msg(&pfvf->mbox);
183 out:
184 	mutex_unlock(&pfvf->mbox.lock);
185 	return rc;
186 }
187 
188 int cn10k_alloc_leaf_profile(struct otx2_nic *pfvf, u16 *leaf)
189 {
190 	struct nix_bandprof_alloc_req *req;
191 	struct nix_bandprof_alloc_rsp *rsp;
192 	int rc;
193 
194 	req = otx2_mbox_alloc_msg_nix_bandprof_alloc(&pfvf->mbox);
195 	if (!req)
196 		return  -ENOMEM;
197 
198 	req->prof_count[BAND_PROF_LEAF_LAYER] = 1;
199 
200 	rc = otx2_sync_mbox_msg(&pfvf->mbox);
201 	if (rc)
202 		goto out;
203 
204 	rsp = (struct  nix_bandprof_alloc_rsp *)
205 	       otx2_mbox_get_rsp(&pfvf->mbox.mbox, 0, &req->hdr);
206 	if (!rsp->prof_count[BAND_PROF_LEAF_LAYER]) {
207 		rc = -EIO;
208 		goto out;
209 	}
210 
211 	*leaf = rsp->prof_idx[BAND_PROF_LEAF_LAYER][0];
212 out:
213 	if (rc) {
214 		dev_warn(pfvf->dev,
215 			 "Failed to allocate ingress bandwidth policer\n");
216 	}
217 
218 	return rc;
219 }
220 
221 int cn10k_alloc_matchall_ipolicer(struct otx2_nic *pfvf)
222 {
223 	struct otx2_hw *hw = &pfvf->hw;
224 	int ret;
225 
226 	mutex_lock(&pfvf->mbox.lock);
227 
228 	ret = cn10k_alloc_leaf_profile(pfvf, &hw->matchall_ipolicer);
229 
230 	mutex_unlock(&pfvf->mbox.lock);
231 
232 	return ret;
233 }
234 
235 #define POLICER_TIMESTAMP	  1  /* 1 second */
236 #define MAX_RATE_EXP		  22 /* Valid rate exponent range: 0 - 22 */
237 
238 static void cn10k_get_ingress_burst_cfg(u32 burst, u32 *burst_exp,
239 					u32 *burst_mantissa)
240 {
241 	int tmp;
242 
243 	/* Burst is calculated as
244 	 * (1+[BURST_MANTISSA]/256)*2^[BURST_EXPONENT]
245 	 * This is the upper limit on number tokens (bytes) that
246 	 * can be accumulated in the bucket.
247 	 */
248 	*burst_exp = ilog2(burst);
249 	if (burst < 256) {
250 		/* No float: can't express mantissa in this case */
251 		*burst_mantissa = 0;
252 		return;
253 	}
254 
255 	if (*burst_exp > MAX_RATE_EXP)
256 		*burst_exp = MAX_RATE_EXP;
257 
258 	/* Calculate mantissa
259 	 * Find remaining bytes 'burst - 2^burst_exp'
260 	 * mantissa = (remaining bytes) / 2^ (burst_exp - 8)
261 	 */
262 	tmp = burst - rounddown_pow_of_two(burst);
263 	*burst_mantissa = tmp / (1UL << (*burst_exp - 8));
264 }
265 
266 static void cn10k_get_ingress_rate_cfg(u64 rate, u32 *rate_exp,
267 				       u32 *rate_mantissa, u32 *rdiv)
268 {
269 	u32 div = 0;
270 	u32 exp = 0;
271 	u64 tmp;
272 
273 	/* Figure out mantissa, exponent and divider from given max pkt rate
274 	 *
275 	 * To achieve desired rate HW adds
276 	 * (1+[RATE_MANTISSA]/256)*2^[RATE_EXPONENT] tokens (bytes) at every
277 	 * policer timeunit * 2^rdiv ie 2 * 2^rdiv usecs, to the token bucket.
278 	 * Here policer timeunit is 2 usecs and rate is in bits per sec.
279 	 * Since floating point cannot be used below algorithm uses 1000000
280 	 * scale factor to support rates upto 100Gbps.
281 	 */
282 	tmp = rate * 32 * 2;
283 	if (tmp < 256000000) {
284 		while (tmp < 256000000) {
285 			tmp = tmp * 2;
286 			div++;
287 		}
288 	} else {
289 		for (exp = 0; tmp >= 512000000 && exp <= MAX_RATE_EXP; exp++)
290 			tmp = tmp / 2;
291 
292 		if (exp > MAX_RATE_EXP)
293 			exp = MAX_RATE_EXP;
294 	}
295 
296 	*rate_mantissa = (tmp - 256000000) / 1000000;
297 	*rate_exp = exp;
298 	*rdiv = div;
299 }
300 
301 int cn10k_map_unmap_rq_policer(struct otx2_nic *pfvf, int rq_idx,
302 			       u16 policer, bool map)
303 {
304 	struct nix_cn10k_aq_enq_req *aq;
305 
306 	aq = otx2_mbox_alloc_msg_nix_cn10k_aq_enq(&pfvf->mbox);
307 	if (!aq)
308 		return -ENOMEM;
309 
310 	/* Enable policing and set the bandwidth profile (policer) index */
311 	if (map)
312 		aq->rq.policer_ena = 1;
313 	else
314 		aq->rq.policer_ena = 0;
315 	aq->rq_mask.policer_ena = 1;
316 
317 	aq->rq.band_prof_id = policer;
318 	aq->rq_mask.band_prof_id = GENMASK(9, 0);
319 
320 	/* Fill AQ info */
321 	aq->qidx = rq_idx;
322 	aq->ctype = NIX_AQ_CTYPE_RQ;
323 	aq->op = NIX_AQ_INSTOP_WRITE;
324 
325 	return otx2_sync_mbox_msg(&pfvf->mbox);
326 }
327 
328 int cn10k_free_leaf_profile(struct otx2_nic *pfvf, u16 leaf)
329 {
330 	struct nix_bandprof_free_req *req;
331 
332 	req = otx2_mbox_alloc_msg_nix_bandprof_free(&pfvf->mbox);
333 	if (!req)
334 		return -ENOMEM;
335 
336 	req->prof_count[BAND_PROF_LEAF_LAYER] = 1;
337 	req->prof_idx[BAND_PROF_LEAF_LAYER][0] = leaf;
338 
339 	return otx2_sync_mbox_msg(&pfvf->mbox);
340 }
341 
342 int cn10k_free_matchall_ipolicer(struct otx2_nic *pfvf)
343 {
344 	struct otx2_hw *hw = &pfvf->hw;
345 	int qidx, rc;
346 
347 	mutex_lock(&pfvf->mbox.lock);
348 
349 	/* Remove RQ's policer mapping */
350 	for (qidx = 0; qidx < hw->rx_queues; qidx++)
351 		cn10k_map_unmap_rq_policer(pfvf, qidx,
352 					   hw->matchall_ipolicer, false);
353 
354 	rc = cn10k_free_leaf_profile(pfvf, hw->matchall_ipolicer);
355 
356 	mutex_unlock(&pfvf->mbox.lock);
357 	return rc;
358 }
359 
360 int cn10k_set_ipolicer_rate(struct otx2_nic *pfvf, u16 profile,
361 			    u32 burst, u64 rate, bool pps)
362 {
363 	struct nix_cn10k_aq_enq_req *aq;
364 	u32 burst_exp, burst_mantissa;
365 	u32 rate_exp, rate_mantissa;
366 	u32 rdiv;
367 
368 	/* Get exponent and mantissa values for the desired rate */
369 	cn10k_get_ingress_burst_cfg(burst, &burst_exp, &burst_mantissa);
370 	cn10k_get_ingress_rate_cfg(rate, &rate_exp, &rate_mantissa, &rdiv);
371 
372 	/* Init bandwidth profile */
373 	aq = otx2_mbox_alloc_msg_nix_cn10k_aq_enq(&pfvf->mbox);
374 	if (!aq)
375 		return -ENOMEM;
376 
377 	/* Set initial color mode to blind */
378 	aq->prof.icolor = 0x03;
379 	aq->prof_mask.icolor = 0x03;
380 
381 	/* Set rate and burst values */
382 	aq->prof.cir_exponent = rate_exp;
383 	aq->prof_mask.cir_exponent = 0x1F;
384 
385 	aq->prof.cir_mantissa = rate_mantissa;
386 	aq->prof_mask.cir_mantissa = 0xFF;
387 
388 	aq->prof.cbs_exponent = burst_exp;
389 	aq->prof_mask.cbs_exponent = 0x1F;
390 
391 	aq->prof.cbs_mantissa = burst_mantissa;
392 	aq->prof_mask.cbs_mantissa = 0xFF;
393 
394 	aq->prof.rdiv = rdiv;
395 	aq->prof_mask.rdiv = 0xF;
396 
397 	if (pps) {
398 		/* The amount of decremented tokens is calculated according to
399 		 * the following equation:
400 		 * max([ LMODE ? 0 : (packet_length - LXPTR)] +
401 		 *	     ([ADJUST_MANTISSA]/256 - 1) * 2^[ADJUST_EXPONENT],
402 		 *	1/256)
403 		 * if LMODE is 1 then rate limiting will be based on
404 		 * PPS otherwise bps.
405 		 * The aim of the ADJUST value is to specify a token cost per
406 		 * packet in contrary to the packet length that specifies a
407 		 * cost per byte. To rate limit based on PPS adjust mantissa
408 		 * is set as 384 and exponent as 1 so that number of tokens
409 		 * decremented becomes 1 i.e, 1 token per packeet.
410 		 */
411 		aq->prof.adjust_exponent = 1;
412 		aq->prof_mask.adjust_exponent = 0x1F;
413 
414 		aq->prof.adjust_mantissa = 384;
415 		aq->prof_mask.adjust_mantissa = 0x1FF;
416 
417 		aq->prof.lmode = 0x1;
418 		aq->prof_mask.lmode = 0x1;
419 	}
420 
421 	/* Two rate three color marker
422 	 * With PEIR/EIR set to zero, color will be either green or red
423 	 */
424 	aq->prof.meter_algo = 2;
425 	aq->prof_mask.meter_algo = 0x3;
426 
427 	aq->prof.rc_action = NIX_RX_BAND_PROF_ACTIONRESULT_DROP;
428 	aq->prof_mask.rc_action = 0x3;
429 
430 	aq->prof.yc_action = NIX_RX_BAND_PROF_ACTIONRESULT_PASS;
431 	aq->prof_mask.yc_action = 0x3;
432 
433 	aq->prof.gc_action = NIX_RX_BAND_PROF_ACTIONRESULT_PASS;
434 	aq->prof_mask.gc_action = 0x3;
435 
436 	/* Setting exponent value as 24 and mantissa as 0 configures
437 	 * the bucket with zero values making bucket unused. Peak
438 	 * information rate and Excess information rate buckets are
439 	 * unused here.
440 	 */
441 	aq->prof.peir_exponent = 24;
442 	aq->prof_mask.peir_exponent = 0x1F;
443 
444 	aq->prof.peir_mantissa = 0;
445 	aq->prof_mask.peir_mantissa = 0xFF;
446 
447 	aq->prof.pebs_exponent = 24;
448 	aq->prof_mask.pebs_exponent = 0x1F;
449 
450 	aq->prof.pebs_mantissa = 0;
451 	aq->prof_mask.pebs_mantissa = 0xFF;
452 
453 	aq->prof.hl_en = 0;
454 	aq->prof_mask.hl_en = 1;
455 
456 	/* Fill AQ info */
457 	aq->qidx = profile;
458 	aq->ctype = NIX_AQ_CTYPE_BANDPROF;
459 	aq->op = NIX_AQ_INSTOP_WRITE;
460 
461 	return otx2_sync_mbox_msg(&pfvf->mbox);
462 }
463 
464 int cn10k_set_matchall_ipolicer_rate(struct otx2_nic *pfvf,
465 				     u32 burst, u64 rate)
466 {
467 	struct otx2_hw *hw = &pfvf->hw;
468 	int qidx, rc;
469 
470 	mutex_lock(&pfvf->mbox.lock);
471 
472 	rc = cn10k_set_ipolicer_rate(pfvf, hw->matchall_ipolicer, burst,
473 				     rate, false);
474 	if (rc)
475 		goto out;
476 
477 	for (qidx = 0; qidx < hw->rx_queues; qidx++) {
478 		rc = cn10k_map_unmap_rq_policer(pfvf, qidx,
479 						hw->matchall_ipolicer, true);
480 		if (rc)
481 			break;
482 	}
483 
484 out:
485 	mutex_unlock(&pfvf->mbox.lock);
486 	return rc;
487 }
488