xref: /linux/drivers/infiniband/hw/mana/qp.c (revision 7255fcc80d4b525cc10cfaaf7f485830d4ed2000)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (c) 2022, Microsoft Corporation. All rights reserved.
4  */
5 
6 #include "mana_ib.h"
7 
8 static int mana_ib_cfg_vport_steering(struct mana_ib_dev *dev,
9 				      struct net_device *ndev,
10 				      mana_handle_t default_rxobj,
11 				      mana_handle_t ind_table[],
12 				      u32 log_ind_tbl_size, u32 rx_hash_key_len,
13 				      u8 *rx_hash_key)
14 {
15 	struct mana_port_context *mpc = netdev_priv(ndev);
16 	struct mana_cfg_rx_steer_req_v2 *req;
17 	struct mana_cfg_rx_steer_resp resp = {};
18 	mana_handle_t *req_indir_tab;
19 	struct gdma_context *gc;
20 	u32 req_buf_size;
21 	int i, err;
22 
23 	gc = mdev_to_gc(dev);
24 
25 	req_buf_size =
26 		sizeof(*req) + sizeof(mana_handle_t) * MANA_INDIRECT_TABLE_SIZE;
27 	req = kzalloc(req_buf_size, GFP_KERNEL);
28 	if (!req)
29 		return -ENOMEM;
30 
31 	mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size,
32 			     sizeof(resp));
33 
34 	req->hdr.req.msg_version = GDMA_MESSAGE_V2;
35 
36 	req->vport = mpc->port_handle;
37 	req->rx_enable = 1;
38 	req->update_default_rxobj = 1;
39 	req->default_rxobj = default_rxobj;
40 	req->hdr.dev_id = gc->mana.dev_id;
41 
42 	/* If there are more than 1 entries in indirection table, enable RSS */
43 	if (log_ind_tbl_size)
44 		req->rss_enable = true;
45 
46 	req->num_indir_entries = MANA_INDIRECT_TABLE_SIZE;
47 	req->indir_tab_offset = sizeof(*req);
48 	req->update_indir_tab = true;
49 	req->cqe_coalescing_enable = 1;
50 
51 	req_indir_tab = (mana_handle_t *)(req + 1);
52 	/* The ind table passed to the hardware must have
53 	 * MANA_INDIRECT_TABLE_SIZE entries. Adjust the verb
54 	 * ind_table to MANA_INDIRECT_TABLE_SIZE if required
55 	 */
56 	ibdev_dbg(&dev->ib_dev, "ind table size %u\n", 1 << log_ind_tbl_size);
57 	for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++) {
58 		req_indir_tab[i] = ind_table[i % (1 << log_ind_tbl_size)];
59 		ibdev_dbg(&dev->ib_dev, "index %u handle 0x%llx\n", i,
60 			  req_indir_tab[i]);
61 	}
62 
63 	req->update_hashkey = true;
64 	if (rx_hash_key_len)
65 		memcpy(req->hashkey, rx_hash_key, rx_hash_key_len);
66 	else
67 		netdev_rss_key_fill(req->hashkey, MANA_HASH_KEY_SIZE);
68 
69 	ibdev_dbg(&dev->ib_dev, "vport handle %llu default_rxobj 0x%llx\n",
70 		  req->vport, default_rxobj);
71 
72 	err = mana_gd_send_request(gc, req_buf_size, req, sizeof(resp), &resp);
73 	if (err) {
74 		netdev_err(ndev, "Failed to configure vPort RX: %d\n", err);
75 		goto out;
76 	}
77 
78 	if (resp.hdr.status) {
79 		netdev_err(ndev, "vPort RX configuration failed: 0x%x\n",
80 			   resp.hdr.status);
81 		err = -EPROTO;
82 		goto out;
83 	}
84 
85 	netdev_info(ndev, "Configured steering vPort %llu log_entries %u\n",
86 		    mpc->port_handle, log_ind_tbl_size);
87 
88 out:
89 	kfree(req);
90 	return err;
91 }
92 
93 static int mana_ib_create_qp_rss(struct ib_qp *ibqp, struct ib_pd *pd,
94 				 struct ib_qp_init_attr *attr,
95 				 struct ib_udata *udata)
96 {
97 	struct mana_ib_qp *qp = container_of(ibqp, struct mana_ib_qp, ibqp);
98 	struct mana_ib_dev *mdev =
99 		container_of(pd->device, struct mana_ib_dev, ib_dev);
100 	struct gdma_context *gc = mdev_to_gc(mdev);
101 	struct ib_rwq_ind_table *ind_tbl = attr->rwq_ind_tbl;
102 	struct mana_ib_create_qp_rss_resp resp = {};
103 	struct mana_ib_create_qp_rss ucmd = {};
104 	struct gdma_queue **gdma_cq_allocated;
105 	mana_handle_t *mana_ind_table;
106 	struct mana_port_context *mpc;
107 	unsigned int ind_tbl_size;
108 	struct net_device *ndev;
109 	struct mana_ib_cq *cq;
110 	struct mana_ib_wq *wq;
111 	struct mana_eq *eq;
112 	struct ib_cq *ibcq;
113 	struct ib_wq *ibwq;
114 	int i = 0;
115 	u32 port;
116 	int ret;
117 
118 	if (!udata || udata->inlen < sizeof(ucmd))
119 		return -EINVAL;
120 
121 	ret = ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen));
122 	if (ret) {
123 		ibdev_dbg(&mdev->ib_dev,
124 			  "Failed copy from udata for create rss-qp, err %d\n",
125 			  ret);
126 		return ret;
127 	}
128 
129 	if (attr->cap.max_recv_wr > mdev->adapter_caps.max_qp_wr) {
130 		ibdev_dbg(&mdev->ib_dev,
131 			  "Requested max_recv_wr %d exceeding limit\n",
132 			  attr->cap.max_recv_wr);
133 		return -EINVAL;
134 	}
135 
136 	if (attr->cap.max_recv_sge > MAX_RX_WQE_SGL_ENTRIES) {
137 		ibdev_dbg(&mdev->ib_dev,
138 			  "Requested max_recv_sge %d exceeding limit\n",
139 			  attr->cap.max_recv_sge);
140 		return -EINVAL;
141 	}
142 
143 	ind_tbl_size = 1 << ind_tbl->log_ind_tbl_size;
144 	if (ind_tbl_size > MANA_INDIRECT_TABLE_SIZE) {
145 		ibdev_dbg(&mdev->ib_dev,
146 			  "Indirect table size %d exceeding limit\n",
147 			  ind_tbl_size);
148 		return -EINVAL;
149 	}
150 
151 	if (ucmd.rx_hash_function != MANA_IB_RX_HASH_FUNC_TOEPLITZ) {
152 		ibdev_dbg(&mdev->ib_dev,
153 			  "RX Hash function is not supported, %d\n",
154 			  ucmd.rx_hash_function);
155 		return -EINVAL;
156 	}
157 
158 	/* IB ports start with 1, MANA start with 0 */
159 	port = ucmd.port;
160 	ndev = mana_ib_get_netdev(pd->device, port);
161 	if (!ndev) {
162 		ibdev_dbg(&mdev->ib_dev, "Invalid port %u in creating qp\n",
163 			  port);
164 		return -EINVAL;
165 	}
166 	mpc = netdev_priv(ndev);
167 
168 	ibdev_dbg(&mdev->ib_dev, "rx_hash_function %d port %d\n",
169 		  ucmd.rx_hash_function, port);
170 
171 	mana_ind_table = kcalloc(ind_tbl_size, sizeof(mana_handle_t),
172 				 GFP_KERNEL);
173 	if (!mana_ind_table) {
174 		ret = -ENOMEM;
175 		goto fail;
176 	}
177 
178 	gdma_cq_allocated = kcalloc(ind_tbl_size, sizeof(*gdma_cq_allocated),
179 				    GFP_KERNEL);
180 	if (!gdma_cq_allocated) {
181 		ret = -ENOMEM;
182 		goto fail;
183 	}
184 
185 	qp->port = port;
186 
187 	for (i = 0; i < ind_tbl_size; i++) {
188 		struct mana_obj_spec wq_spec = {};
189 		struct mana_obj_spec cq_spec = {};
190 
191 		ibwq = ind_tbl->ind_tbl[i];
192 		wq = container_of(ibwq, struct mana_ib_wq, ibwq);
193 
194 		ibcq = ibwq->cq;
195 		cq = container_of(ibcq, struct mana_ib_cq, ibcq);
196 
197 		wq_spec.gdma_region = wq->gdma_region;
198 		wq_spec.queue_size = wq->wq_buf_size;
199 
200 		cq_spec.gdma_region = cq->gdma_region;
201 		cq_spec.queue_size = cq->cqe * COMP_ENTRY_SIZE;
202 		cq_spec.modr_ctx_id = 0;
203 		eq = &mpc->ac->eqs[cq->comp_vector % gc->max_num_queues];
204 		cq_spec.attached_eq = eq->eq->id;
205 
206 		ret = mana_create_wq_obj(mpc, mpc->port_handle, GDMA_RQ,
207 					 &wq_spec, &cq_spec, &wq->rx_object);
208 		if (ret) {
209 			/* Do cleanup starting with index i-1 */
210 			i--;
211 			goto fail;
212 		}
213 
214 		/* The GDMA regions are now owned by the WQ object */
215 		wq->gdma_region = GDMA_INVALID_DMA_REGION;
216 		cq->gdma_region = GDMA_INVALID_DMA_REGION;
217 
218 		wq->id = wq_spec.queue_index;
219 		cq->id = cq_spec.queue_index;
220 
221 		ibdev_dbg(&mdev->ib_dev,
222 			  "ret %d rx_object 0x%llx wq id %llu cq id %llu\n",
223 			  ret, wq->rx_object, wq->id, cq->id);
224 
225 		resp.entries[i].cqid = cq->id;
226 		resp.entries[i].wqid = wq->id;
227 
228 		mana_ind_table[i] = wq->rx_object;
229 
230 		/* Create CQ table entry */
231 		ret = mana_ib_install_cq_cb(mdev, cq);
232 		if (ret)
233 			goto fail;
234 
235 		gdma_cq_allocated[i] = gc->cq_table[cq->id];
236 	}
237 	resp.num_entries = i;
238 
239 	ret = mana_ib_cfg_vport_steering(mdev, ndev, wq->rx_object,
240 					 mana_ind_table,
241 					 ind_tbl->log_ind_tbl_size,
242 					 ucmd.rx_hash_key_len,
243 					 ucmd.rx_hash_key);
244 	if (ret)
245 		goto fail;
246 
247 	ret = ib_copy_to_udata(udata, &resp, sizeof(resp));
248 	if (ret) {
249 		ibdev_dbg(&mdev->ib_dev,
250 			  "Failed to copy to udata create rss-qp, %d\n",
251 			  ret);
252 		goto fail;
253 	}
254 
255 	kfree(gdma_cq_allocated);
256 	kfree(mana_ind_table);
257 
258 	return 0;
259 
260 fail:
261 	while (i-- > 0) {
262 		ibwq = ind_tbl->ind_tbl[i];
263 		ibcq = ibwq->cq;
264 		wq = container_of(ibwq, struct mana_ib_wq, ibwq);
265 		cq = container_of(ibcq, struct mana_ib_cq, ibcq);
266 
267 		gc->cq_table[cq->id] = NULL;
268 		kfree(gdma_cq_allocated[i]);
269 
270 		mana_destroy_wq_obj(mpc, GDMA_RQ, wq->rx_object);
271 	}
272 
273 	kfree(gdma_cq_allocated);
274 	kfree(mana_ind_table);
275 
276 	return ret;
277 }
278 
279 static int mana_ib_create_qp_raw(struct ib_qp *ibqp, struct ib_pd *ibpd,
280 				 struct ib_qp_init_attr *attr,
281 				 struct ib_udata *udata)
282 {
283 	struct mana_ib_pd *pd = container_of(ibpd, struct mana_ib_pd, ibpd);
284 	struct mana_ib_qp *qp = container_of(ibqp, struct mana_ib_qp, ibqp);
285 	struct mana_ib_dev *mdev =
286 		container_of(ibpd->device, struct mana_ib_dev, ib_dev);
287 	struct mana_ib_cq *send_cq =
288 		container_of(attr->send_cq, struct mana_ib_cq, ibcq);
289 	struct mana_ib_ucontext *mana_ucontext =
290 		rdma_udata_to_drv_context(udata, struct mana_ib_ucontext,
291 					  ibucontext);
292 	struct gdma_context *gc = mdev_to_gc(mdev);
293 	struct mana_ib_create_qp_resp resp = {};
294 	struct mana_ib_create_qp ucmd = {};
295 	struct gdma_queue *gdma_cq = NULL;
296 	struct mana_obj_spec wq_spec = {};
297 	struct mana_obj_spec cq_spec = {};
298 	struct mana_port_context *mpc;
299 	struct net_device *ndev;
300 	struct ib_umem *umem;
301 	struct mana_eq *eq;
302 	int eq_vec;
303 	u32 port;
304 	int err;
305 
306 	if (!mana_ucontext || udata->inlen < sizeof(ucmd))
307 		return -EINVAL;
308 
309 	err = ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen));
310 	if (err) {
311 		ibdev_dbg(&mdev->ib_dev,
312 			  "Failed to copy from udata create qp-raw, %d\n", err);
313 		return err;
314 	}
315 
316 	if (attr->cap.max_send_wr > mdev->adapter_caps.max_qp_wr) {
317 		ibdev_dbg(&mdev->ib_dev,
318 			  "Requested max_send_wr %d exceeding limit\n",
319 			  attr->cap.max_send_wr);
320 		return -EINVAL;
321 	}
322 
323 	if (attr->cap.max_send_sge > MAX_TX_WQE_SGL_ENTRIES) {
324 		ibdev_dbg(&mdev->ib_dev,
325 			  "Requested max_send_sge %d exceeding limit\n",
326 			  attr->cap.max_send_sge);
327 		return -EINVAL;
328 	}
329 
330 	port = ucmd.port;
331 	ndev = mana_ib_get_netdev(ibpd->device, port);
332 	if (!ndev) {
333 		ibdev_dbg(&mdev->ib_dev, "Invalid port %u in creating qp\n",
334 			  port);
335 		return -EINVAL;
336 	}
337 	mpc = netdev_priv(ndev);
338 	ibdev_dbg(&mdev->ib_dev, "port %u ndev %p mpc %p\n", port, ndev, mpc);
339 
340 	err = mana_ib_cfg_vport(mdev, port, pd, mana_ucontext->doorbell);
341 	if (err)
342 		return -ENODEV;
343 
344 	qp->port = port;
345 
346 	ibdev_dbg(&mdev->ib_dev, "ucmd sq_buf_addr 0x%llx port %u\n",
347 		  ucmd.sq_buf_addr, ucmd.port);
348 
349 	umem = ib_umem_get(ibpd->device, ucmd.sq_buf_addr, ucmd.sq_buf_size,
350 			   IB_ACCESS_LOCAL_WRITE);
351 	if (IS_ERR(umem)) {
352 		err = PTR_ERR(umem);
353 		ibdev_dbg(&mdev->ib_dev,
354 			  "Failed to get umem for create qp-raw, err %d\n",
355 			  err);
356 		goto err_free_vport;
357 	}
358 	qp->sq_umem = umem;
359 
360 	err = mana_ib_create_zero_offset_dma_region(mdev, qp->sq_umem,
361 						    &qp->sq_gdma_region);
362 	if (err) {
363 		ibdev_dbg(&mdev->ib_dev,
364 			  "Failed to create dma region for create qp-raw, %d\n",
365 			  err);
366 		goto err_release_umem;
367 	}
368 
369 	ibdev_dbg(&mdev->ib_dev,
370 		  "create_dma_region ret %d gdma_region 0x%llx\n",
371 		  err, qp->sq_gdma_region);
372 
373 	/* Create a WQ on the same port handle used by the Ethernet */
374 	wq_spec.gdma_region = qp->sq_gdma_region;
375 	wq_spec.queue_size = ucmd.sq_buf_size;
376 
377 	cq_spec.gdma_region = send_cq->gdma_region;
378 	cq_spec.queue_size = send_cq->cqe * COMP_ENTRY_SIZE;
379 	cq_spec.modr_ctx_id = 0;
380 	eq_vec = send_cq->comp_vector % gc->max_num_queues;
381 	eq = &mpc->ac->eqs[eq_vec];
382 	cq_spec.attached_eq = eq->eq->id;
383 
384 	err = mana_create_wq_obj(mpc, mpc->port_handle, GDMA_SQ, &wq_spec,
385 				 &cq_spec, &qp->tx_object);
386 	if (err) {
387 		ibdev_dbg(&mdev->ib_dev,
388 			  "Failed to create wq for create raw-qp, err %d\n",
389 			  err);
390 		goto err_destroy_dma_region;
391 	}
392 
393 	/* The GDMA regions are now owned by the WQ object */
394 	qp->sq_gdma_region = GDMA_INVALID_DMA_REGION;
395 	send_cq->gdma_region = GDMA_INVALID_DMA_REGION;
396 
397 	qp->sq_id = wq_spec.queue_index;
398 	send_cq->id = cq_spec.queue_index;
399 
400 	/* Create CQ table entry */
401 	err = mana_ib_install_cq_cb(mdev, send_cq);
402 	if (err)
403 		goto err_destroy_wq_obj;
404 
405 	ibdev_dbg(&mdev->ib_dev,
406 		  "ret %d qp->tx_object 0x%llx sq id %llu cq id %llu\n", err,
407 		  qp->tx_object, qp->sq_id, send_cq->id);
408 
409 	resp.sqid = qp->sq_id;
410 	resp.cqid = send_cq->id;
411 	resp.tx_vp_offset = pd->tx_vp_offset;
412 
413 	err = ib_copy_to_udata(udata, &resp, sizeof(resp));
414 	if (err) {
415 		ibdev_dbg(&mdev->ib_dev,
416 			  "Failed copy udata for create qp-raw, %d\n",
417 			  err);
418 		goto err_release_gdma_cq;
419 	}
420 
421 	return 0;
422 
423 err_release_gdma_cq:
424 	kfree(gdma_cq);
425 	gc->cq_table[send_cq->id] = NULL;
426 
427 err_destroy_wq_obj:
428 	mana_destroy_wq_obj(mpc, GDMA_SQ, qp->tx_object);
429 
430 err_destroy_dma_region:
431 	mana_ib_gd_destroy_dma_region(mdev, qp->sq_gdma_region);
432 
433 err_release_umem:
434 	ib_umem_release(umem);
435 
436 err_free_vport:
437 	mana_ib_uncfg_vport(mdev, pd, port);
438 
439 	return err;
440 }
441 
442 int mana_ib_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *attr,
443 		      struct ib_udata *udata)
444 {
445 	switch (attr->qp_type) {
446 	case IB_QPT_RAW_PACKET:
447 		/* When rwq_ind_tbl is used, it's for creating WQs for RSS */
448 		if (attr->rwq_ind_tbl)
449 			return mana_ib_create_qp_rss(ibqp, ibqp->pd, attr,
450 						     udata);
451 
452 		return mana_ib_create_qp_raw(ibqp, ibqp->pd, attr, udata);
453 	default:
454 		/* Creating QP other than IB_QPT_RAW_PACKET is not supported */
455 		ibdev_dbg(ibqp->device, "Creating QP type %u not supported\n",
456 			  attr->qp_type);
457 	}
458 
459 	return -EINVAL;
460 }
461 
462 int mana_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
463 		      int attr_mask, struct ib_udata *udata)
464 {
465 	/* modify_qp is not supported by this version of the driver */
466 	return -EOPNOTSUPP;
467 }
468 
469 static int mana_ib_destroy_qp_rss(struct mana_ib_qp *qp,
470 				  struct ib_rwq_ind_table *ind_tbl,
471 				  struct ib_udata *udata)
472 {
473 	struct mana_ib_dev *mdev =
474 		container_of(qp->ibqp.device, struct mana_ib_dev, ib_dev);
475 	struct mana_port_context *mpc;
476 	struct net_device *ndev;
477 	struct mana_ib_wq *wq;
478 	struct ib_wq *ibwq;
479 	int i;
480 
481 	ndev = mana_ib_get_netdev(qp->ibqp.device, qp->port);
482 	mpc = netdev_priv(ndev);
483 
484 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
485 		ibwq = ind_tbl->ind_tbl[i];
486 		wq = container_of(ibwq, struct mana_ib_wq, ibwq);
487 		ibdev_dbg(&mdev->ib_dev, "destroying wq->rx_object %llu\n",
488 			  wq->rx_object);
489 		mana_destroy_wq_obj(mpc, GDMA_RQ, wq->rx_object);
490 	}
491 
492 	return 0;
493 }
494 
495 static int mana_ib_destroy_qp_raw(struct mana_ib_qp *qp, struct ib_udata *udata)
496 {
497 	struct mana_ib_dev *mdev =
498 		container_of(qp->ibqp.device, struct mana_ib_dev, ib_dev);
499 	struct ib_pd *ibpd = qp->ibqp.pd;
500 	struct mana_port_context *mpc;
501 	struct net_device *ndev;
502 	struct mana_ib_pd *pd;
503 
504 	ndev = mana_ib_get_netdev(qp->ibqp.device, qp->port);
505 	mpc = netdev_priv(ndev);
506 	pd = container_of(ibpd, struct mana_ib_pd, ibpd);
507 
508 	mana_destroy_wq_obj(mpc, GDMA_SQ, qp->tx_object);
509 
510 	if (qp->sq_umem) {
511 		mana_ib_gd_destroy_dma_region(mdev, qp->sq_gdma_region);
512 		ib_umem_release(qp->sq_umem);
513 	}
514 
515 	mana_ib_uncfg_vport(mdev, pd, qp->port);
516 
517 	return 0;
518 }
519 
520 int mana_ib_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
521 {
522 	struct mana_ib_qp *qp = container_of(ibqp, struct mana_ib_qp, ibqp);
523 
524 	switch (ibqp->qp_type) {
525 	case IB_QPT_RAW_PACKET:
526 		if (ibqp->rwq_ind_tbl)
527 			return mana_ib_destroy_qp_rss(qp, ibqp->rwq_ind_tbl,
528 						      udata);
529 
530 		return mana_ib_destroy_qp_raw(qp, udata);
531 
532 	default:
533 		ibdev_dbg(ibqp->device, "Unexpected QP type %u\n",
534 			  ibqp->qp_type);
535 	}
536 
537 	return -ENOENT;
538 }
539