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 (IS_ERR(rsp)) {
207 		rc = PTR_ERR(rsp);
208 		goto out;
209 	}
210 
211 	if (!rsp->prof_count[BAND_PROF_LEAF_LAYER]) {
212 		rc = -EIO;
213 		goto out;
214 	}
215 
216 	*leaf = rsp->prof_idx[BAND_PROF_LEAF_LAYER][0];
217 out:
218 	if (rc) {
219 		dev_warn(pfvf->dev,
220 			 "Failed to allocate ingress bandwidth policer\n");
221 	}
222 
223 	return rc;
224 }
225 
226 int cn10k_alloc_matchall_ipolicer(struct otx2_nic *pfvf)
227 {
228 	struct otx2_hw *hw = &pfvf->hw;
229 	int ret;
230 
231 	mutex_lock(&pfvf->mbox.lock);
232 
233 	ret = cn10k_alloc_leaf_profile(pfvf, &hw->matchall_ipolicer);
234 
235 	mutex_unlock(&pfvf->mbox.lock);
236 
237 	return ret;
238 }
239 
240 #define POLICER_TIMESTAMP	  1  /* 1 second */
241 #define MAX_RATE_EXP		  22 /* Valid rate exponent range: 0 - 22 */
242 
243 static void cn10k_get_ingress_burst_cfg(u32 burst, u32 *burst_exp,
244 					u32 *burst_mantissa)
245 {
246 	int tmp;
247 
248 	/* Burst is calculated as
249 	 * (1+[BURST_MANTISSA]/256)*2^[BURST_EXPONENT]
250 	 * This is the upper limit on number tokens (bytes) that
251 	 * can be accumulated in the bucket.
252 	 */
253 	*burst_exp = ilog2(burst);
254 	if (burst < 256) {
255 		/* No float: can't express mantissa in this case */
256 		*burst_mantissa = 0;
257 		return;
258 	}
259 
260 	if (*burst_exp > MAX_RATE_EXP)
261 		*burst_exp = MAX_RATE_EXP;
262 
263 	/* Calculate mantissa
264 	 * Find remaining bytes 'burst - 2^burst_exp'
265 	 * mantissa = (remaining bytes) / 2^ (burst_exp - 8)
266 	 */
267 	tmp = burst - rounddown_pow_of_two(burst);
268 	*burst_mantissa = tmp / (1UL << (*burst_exp - 8));
269 }
270 
271 static void cn10k_get_ingress_rate_cfg(u64 rate, u32 *rate_exp,
272 				       u32 *rate_mantissa, u32 *rdiv)
273 {
274 	u32 div = 0;
275 	u32 exp = 0;
276 	u64 tmp;
277 
278 	/* Figure out mantissa, exponent and divider from given max pkt rate
279 	 *
280 	 * To achieve desired rate HW adds
281 	 * (1+[RATE_MANTISSA]/256)*2^[RATE_EXPONENT] tokens (bytes) at every
282 	 * policer timeunit * 2^rdiv ie 2 * 2^rdiv usecs, to the token bucket.
283 	 * Here policer timeunit is 2 usecs and rate is in bits per sec.
284 	 * Since floating point cannot be used below algorithm uses 1000000
285 	 * scale factor to support rates upto 100Gbps.
286 	 */
287 	tmp = rate * 32 * 2;
288 	if (tmp < 256000000) {
289 		while (tmp < 256000000) {
290 			tmp = tmp * 2;
291 			div++;
292 		}
293 	} else {
294 		for (exp = 0; tmp >= 512000000 && exp <= MAX_RATE_EXP; exp++)
295 			tmp = tmp / 2;
296 
297 		if (exp > MAX_RATE_EXP)
298 			exp = MAX_RATE_EXP;
299 	}
300 
301 	*rate_mantissa = (tmp - 256000000) / 1000000;
302 	*rate_exp = exp;
303 	*rdiv = div;
304 }
305 
306 int cn10k_map_unmap_rq_policer(struct otx2_nic *pfvf, int rq_idx,
307 			       u16 policer, bool map)
308 {
309 	struct nix_cn10k_aq_enq_req *aq;
310 
311 	aq = otx2_mbox_alloc_msg_nix_cn10k_aq_enq(&pfvf->mbox);
312 	if (!aq)
313 		return -ENOMEM;
314 
315 	/* Enable policing and set the bandwidth profile (policer) index */
316 	if (map)
317 		aq->rq.policer_ena = 1;
318 	else
319 		aq->rq.policer_ena = 0;
320 	aq->rq_mask.policer_ena = 1;
321 
322 	aq->rq.band_prof_id = policer;
323 	aq->rq_mask.band_prof_id = GENMASK(9, 0);
324 
325 	/* Fill AQ info */
326 	aq->qidx = rq_idx;
327 	aq->ctype = NIX_AQ_CTYPE_RQ;
328 	aq->op = NIX_AQ_INSTOP_WRITE;
329 
330 	return otx2_sync_mbox_msg(&pfvf->mbox);
331 }
332 
333 int cn10k_free_leaf_profile(struct otx2_nic *pfvf, u16 leaf)
334 {
335 	struct nix_bandprof_free_req *req;
336 
337 	req = otx2_mbox_alloc_msg_nix_bandprof_free(&pfvf->mbox);
338 	if (!req)
339 		return -ENOMEM;
340 
341 	req->prof_count[BAND_PROF_LEAF_LAYER] = 1;
342 	req->prof_idx[BAND_PROF_LEAF_LAYER][0] = leaf;
343 
344 	return otx2_sync_mbox_msg(&pfvf->mbox);
345 }
346 
347 int cn10k_free_matchall_ipolicer(struct otx2_nic *pfvf)
348 {
349 	struct otx2_hw *hw = &pfvf->hw;
350 	int qidx, rc;
351 
352 	mutex_lock(&pfvf->mbox.lock);
353 
354 	/* Remove RQ's policer mapping */
355 	for (qidx = 0; qidx < hw->rx_queues; qidx++)
356 		cn10k_map_unmap_rq_policer(pfvf, qidx,
357 					   hw->matchall_ipolicer, false);
358 
359 	rc = cn10k_free_leaf_profile(pfvf, hw->matchall_ipolicer);
360 
361 	mutex_unlock(&pfvf->mbox.lock);
362 	return rc;
363 }
364 
365 int cn10k_set_ipolicer_rate(struct otx2_nic *pfvf, u16 profile,
366 			    u32 burst, u64 rate, bool pps)
367 {
368 	struct nix_cn10k_aq_enq_req *aq;
369 	u32 burst_exp, burst_mantissa;
370 	u32 rate_exp, rate_mantissa;
371 	u32 rdiv;
372 
373 	/* Get exponent and mantissa values for the desired rate */
374 	cn10k_get_ingress_burst_cfg(burst, &burst_exp, &burst_mantissa);
375 	cn10k_get_ingress_rate_cfg(rate, &rate_exp, &rate_mantissa, &rdiv);
376 
377 	/* Init bandwidth profile */
378 	aq = otx2_mbox_alloc_msg_nix_cn10k_aq_enq(&pfvf->mbox);
379 	if (!aq)
380 		return -ENOMEM;
381 
382 	/* Set initial color mode to blind */
383 	aq->prof.icolor = 0x03;
384 	aq->prof_mask.icolor = 0x03;
385 
386 	/* Set rate and burst values */
387 	aq->prof.cir_exponent = rate_exp;
388 	aq->prof_mask.cir_exponent = 0x1F;
389 
390 	aq->prof.cir_mantissa = rate_mantissa;
391 	aq->prof_mask.cir_mantissa = 0xFF;
392 
393 	aq->prof.cbs_exponent = burst_exp;
394 	aq->prof_mask.cbs_exponent = 0x1F;
395 
396 	aq->prof.cbs_mantissa = burst_mantissa;
397 	aq->prof_mask.cbs_mantissa = 0xFF;
398 
399 	aq->prof.rdiv = rdiv;
400 	aq->prof_mask.rdiv = 0xF;
401 
402 	if (pps) {
403 		/* The amount of decremented tokens is calculated according to
404 		 * the following equation:
405 		 * max([ LMODE ? 0 : (packet_length - LXPTR)] +
406 		 *	     ([ADJUST_MANTISSA]/256 - 1) * 2^[ADJUST_EXPONENT],
407 		 *	1/256)
408 		 * if LMODE is 1 then rate limiting will be based on
409 		 * PPS otherwise bps.
410 		 * The aim of the ADJUST value is to specify a token cost per
411 		 * packet in contrary to the packet length that specifies a
412 		 * cost per byte. To rate limit based on PPS adjust mantissa
413 		 * is set as 384 and exponent as 1 so that number of tokens
414 		 * decremented becomes 1 i.e, 1 token per packeet.
415 		 */
416 		aq->prof.adjust_exponent = 1;
417 		aq->prof_mask.adjust_exponent = 0x1F;
418 
419 		aq->prof.adjust_mantissa = 384;
420 		aq->prof_mask.adjust_mantissa = 0x1FF;
421 
422 		aq->prof.lmode = 0x1;
423 		aq->prof_mask.lmode = 0x1;
424 	}
425 
426 	/* Two rate three color marker
427 	 * With PEIR/EIR set to zero, color will be either green or red
428 	 */
429 	aq->prof.meter_algo = 2;
430 	aq->prof_mask.meter_algo = 0x3;
431 
432 	aq->prof.rc_action = NIX_RX_BAND_PROF_ACTIONRESULT_DROP;
433 	aq->prof_mask.rc_action = 0x3;
434 
435 	aq->prof.yc_action = NIX_RX_BAND_PROF_ACTIONRESULT_PASS;
436 	aq->prof_mask.yc_action = 0x3;
437 
438 	aq->prof.gc_action = NIX_RX_BAND_PROF_ACTIONRESULT_PASS;
439 	aq->prof_mask.gc_action = 0x3;
440 
441 	/* Setting exponent value as 24 and mantissa as 0 configures
442 	 * the bucket with zero values making bucket unused. Peak
443 	 * information rate and Excess information rate buckets are
444 	 * unused here.
445 	 */
446 	aq->prof.peir_exponent = 24;
447 	aq->prof_mask.peir_exponent = 0x1F;
448 
449 	aq->prof.peir_mantissa = 0;
450 	aq->prof_mask.peir_mantissa = 0xFF;
451 
452 	aq->prof.pebs_exponent = 24;
453 	aq->prof_mask.pebs_exponent = 0x1F;
454 
455 	aq->prof.pebs_mantissa = 0;
456 	aq->prof_mask.pebs_mantissa = 0xFF;
457 
458 	aq->prof.hl_en = 0;
459 	aq->prof_mask.hl_en = 1;
460 
461 	/* Fill AQ info */
462 	aq->qidx = profile;
463 	aq->ctype = NIX_AQ_CTYPE_BANDPROF;
464 	aq->op = NIX_AQ_INSTOP_WRITE;
465 
466 	return otx2_sync_mbox_msg(&pfvf->mbox);
467 }
468 
469 int cn10k_set_matchall_ipolicer_rate(struct otx2_nic *pfvf,
470 				     u32 burst, u64 rate)
471 {
472 	struct otx2_hw *hw = &pfvf->hw;
473 	int qidx, rc;
474 
475 	mutex_lock(&pfvf->mbox.lock);
476 
477 	rc = cn10k_set_ipolicer_rate(pfvf, hw->matchall_ipolicer, burst,
478 				     rate, false);
479 	if (rc)
480 		goto out;
481 
482 	for (qidx = 0; qidx < hw->rx_queues; qidx++) {
483 		rc = cn10k_map_unmap_rq_policer(pfvf, qidx,
484 						hw->matchall_ipolicer, true);
485 		if (rc)
486 			break;
487 	}
488 
489 out:
490 	mutex_unlock(&pfvf->mbox.lock);
491 	return rc;
492 }
493