xref: /linux/drivers/infiniband/hw/mana/qp.c (revision f14aa5ea415b8add245e976bfab96a12986c6843)
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 	struct gdma_context *gc;
19 	u32 req_buf_size;
20 	int i, err;
21 
22 	gc = mdev_to_gc(dev);
23 
24 	req_buf_size = struct_size(req, indir_tab, MANA_INDIRECT_TABLE_SIZE);
25 	req = kzalloc(req_buf_size, GFP_KERNEL);
26 	if (!req)
27 		return -ENOMEM;
28 
29 	mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size,
30 			     sizeof(resp));
31 
32 	req->hdr.req.msg_version = GDMA_MESSAGE_V2;
33 
34 	req->vport = mpc->port_handle;
35 	req->rx_enable = 1;
36 	req->update_default_rxobj = 1;
37 	req->default_rxobj = default_rxobj;
38 	req->hdr.dev_id = gc->mana.dev_id;
39 
40 	/* If there are more than 1 entries in indirection table, enable RSS */
41 	if (log_ind_tbl_size)
42 		req->rss_enable = true;
43 
44 	req->num_indir_entries = MANA_INDIRECT_TABLE_SIZE;
45 	req->indir_tab_offset = offsetof(struct mana_cfg_rx_steer_req_v2,
46 					 indir_tab);
47 	req->update_indir_tab = true;
48 	req->cqe_coalescing_enable = 1;
49 
50 	/* The ind table passed to the hardware must have
51 	 * MANA_INDIRECT_TABLE_SIZE entries. Adjust the verb
52 	 * ind_table to MANA_INDIRECT_TABLE_SIZE if required
53 	 */
54 	ibdev_dbg(&dev->ib_dev, "ind table size %u\n", 1 << log_ind_tbl_size);
55 	for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++) {
56 		req->indir_tab[i] = ind_table[i % (1 << log_ind_tbl_size)];
57 		ibdev_dbg(&dev->ib_dev, "index %u handle 0x%llx\n", i,
58 			  req->indir_tab[i]);
59 	}
60 
61 	req->update_hashkey = true;
62 	if (rx_hash_key_len)
63 		memcpy(req->hashkey, rx_hash_key, rx_hash_key_len);
64 	else
65 		netdev_rss_key_fill(req->hashkey, MANA_HASH_KEY_SIZE);
66 
67 	ibdev_dbg(&dev->ib_dev, "vport handle %llu default_rxobj 0x%llx\n",
68 		  req->vport, default_rxobj);
69 
70 	err = mana_gd_send_request(gc, req_buf_size, req, sizeof(resp), &resp);
71 	if (err) {
72 		netdev_err(ndev, "Failed to configure vPort RX: %d\n", err);
73 		goto out;
74 	}
75 
76 	if (resp.hdr.status) {
77 		netdev_err(ndev, "vPort RX configuration failed: 0x%x\n",
78 			   resp.hdr.status);
79 		err = -EPROTO;
80 		goto out;
81 	}
82 
83 	netdev_info(ndev, "Configured steering vPort %llu log_entries %u\n",
84 		    mpc->port_handle, log_ind_tbl_size);
85 
86 out:
87 	kfree(req);
88 	return err;
89 }
90 
91 static int mana_ib_create_qp_rss(struct ib_qp *ibqp, struct ib_pd *pd,
92 				 struct ib_qp_init_attr *attr,
93 				 struct ib_udata *udata)
94 {
95 	struct mana_ib_qp *qp = container_of(ibqp, struct mana_ib_qp, ibqp);
96 	struct mana_ib_dev *mdev =
97 		container_of(pd->device, struct mana_ib_dev, ib_dev);
98 	struct ib_rwq_ind_table *ind_tbl = attr->rwq_ind_tbl;
99 	struct mana_ib_create_qp_rss_resp resp = {};
100 	struct mana_ib_create_qp_rss ucmd = {};
101 	mana_handle_t *mana_ind_table;
102 	struct mana_port_context *mpc;
103 	unsigned int ind_tbl_size;
104 	struct net_device *ndev;
105 	struct mana_ib_cq *cq;
106 	struct mana_ib_wq *wq;
107 	struct mana_eq *eq;
108 	struct ib_cq *ibcq;
109 	struct ib_wq *ibwq;
110 	int i = 0;
111 	u32 port;
112 	int ret;
113 
114 	if (!udata || udata->inlen < sizeof(ucmd))
115 		return -EINVAL;
116 
117 	ret = ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen));
118 	if (ret) {
119 		ibdev_dbg(&mdev->ib_dev,
120 			  "Failed copy from udata for create rss-qp, err %d\n",
121 			  ret);
122 		return ret;
123 	}
124 
125 	if (attr->cap.max_recv_wr > mdev->adapter_caps.max_qp_wr) {
126 		ibdev_dbg(&mdev->ib_dev,
127 			  "Requested max_recv_wr %d exceeding limit\n",
128 			  attr->cap.max_recv_wr);
129 		return -EINVAL;
130 	}
131 
132 	if (attr->cap.max_recv_sge > MAX_RX_WQE_SGL_ENTRIES) {
133 		ibdev_dbg(&mdev->ib_dev,
134 			  "Requested max_recv_sge %d exceeding limit\n",
135 			  attr->cap.max_recv_sge);
136 		return -EINVAL;
137 	}
138 
139 	ind_tbl_size = 1 << ind_tbl->log_ind_tbl_size;
140 	if (ind_tbl_size > MANA_INDIRECT_TABLE_SIZE) {
141 		ibdev_dbg(&mdev->ib_dev,
142 			  "Indirect table size %d exceeding limit\n",
143 			  ind_tbl_size);
144 		return -EINVAL;
145 	}
146 
147 	if (ucmd.rx_hash_function != MANA_IB_RX_HASH_FUNC_TOEPLITZ) {
148 		ibdev_dbg(&mdev->ib_dev,
149 			  "RX Hash function is not supported, %d\n",
150 			  ucmd.rx_hash_function);
151 		return -EINVAL;
152 	}
153 
154 	/* IB ports start with 1, MANA start with 0 */
155 	port = ucmd.port;
156 	ndev = mana_ib_get_netdev(pd->device, port);
157 	if (!ndev) {
158 		ibdev_dbg(&mdev->ib_dev, "Invalid port %u in creating qp\n",
159 			  port);
160 		return -EINVAL;
161 	}
162 	mpc = netdev_priv(ndev);
163 
164 	ibdev_dbg(&mdev->ib_dev, "rx_hash_function %d port %d\n",
165 		  ucmd.rx_hash_function, port);
166 
167 	mana_ind_table = kcalloc(ind_tbl_size, sizeof(mana_handle_t),
168 				 GFP_KERNEL);
169 	if (!mana_ind_table) {
170 		ret = -ENOMEM;
171 		goto fail;
172 	}
173 
174 	qp->port = port;
175 
176 	for (i = 0; i < ind_tbl_size; i++) {
177 		struct mana_obj_spec wq_spec = {};
178 		struct mana_obj_spec cq_spec = {};
179 
180 		ibwq = ind_tbl->ind_tbl[i];
181 		wq = container_of(ibwq, struct mana_ib_wq, ibwq);
182 
183 		ibcq = ibwq->cq;
184 		cq = container_of(ibcq, struct mana_ib_cq, ibcq);
185 
186 		wq_spec.gdma_region = wq->queue.gdma_region;
187 		wq_spec.queue_size = wq->wq_buf_size;
188 
189 		cq_spec.gdma_region = cq->queue.gdma_region;
190 		cq_spec.queue_size = cq->cqe * COMP_ENTRY_SIZE;
191 		cq_spec.modr_ctx_id = 0;
192 		eq = &mpc->ac->eqs[cq->comp_vector];
193 		cq_spec.attached_eq = eq->eq->id;
194 
195 		ret = mana_create_wq_obj(mpc, mpc->port_handle, GDMA_RQ,
196 					 &wq_spec, &cq_spec, &wq->rx_object);
197 		if (ret) {
198 			/* Do cleanup starting with index i-1 */
199 			i--;
200 			goto fail;
201 		}
202 
203 		/* The GDMA regions are now owned by the WQ object */
204 		wq->queue.gdma_region = GDMA_INVALID_DMA_REGION;
205 		cq->queue.gdma_region = GDMA_INVALID_DMA_REGION;
206 
207 		wq->queue.id = wq_spec.queue_index;
208 		cq->queue.id = cq_spec.queue_index;
209 
210 		ibdev_dbg(&mdev->ib_dev,
211 			  "rx_object 0x%llx wq id %llu cq id %llu\n",
212 			  wq->rx_object, wq->queue.id, cq->queue.id);
213 
214 		resp.entries[i].cqid = cq->queue.id;
215 		resp.entries[i].wqid = wq->queue.id;
216 
217 		mana_ind_table[i] = wq->rx_object;
218 
219 		/* Create CQ table entry */
220 		ret = mana_ib_install_cq_cb(mdev, cq);
221 		if (ret)
222 			goto fail;
223 	}
224 	resp.num_entries = i;
225 
226 	ret = mana_ib_cfg_vport_steering(mdev, ndev, wq->rx_object,
227 					 mana_ind_table,
228 					 ind_tbl->log_ind_tbl_size,
229 					 ucmd.rx_hash_key_len,
230 					 ucmd.rx_hash_key);
231 	if (ret)
232 		goto fail;
233 
234 	ret = ib_copy_to_udata(udata, &resp, sizeof(resp));
235 	if (ret) {
236 		ibdev_dbg(&mdev->ib_dev,
237 			  "Failed to copy to udata create rss-qp, %d\n",
238 			  ret);
239 		goto fail;
240 	}
241 
242 	kfree(mana_ind_table);
243 
244 	return 0;
245 
246 fail:
247 	while (i-- > 0) {
248 		ibwq = ind_tbl->ind_tbl[i];
249 		ibcq = ibwq->cq;
250 		wq = container_of(ibwq, struct mana_ib_wq, ibwq);
251 		cq = container_of(ibcq, struct mana_ib_cq, ibcq);
252 
253 		mana_ib_remove_cq_cb(mdev, cq);
254 		mana_destroy_wq_obj(mpc, GDMA_RQ, wq->rx_object);
255 	}
256 
257 	kfree(mana_ind_table);
258 
259 	return ret;
260 }
261 
262 static int mana_ib_create_qp_raw(struct ib_qp *ibqp, struct ib_pd *ibpd,
263 				 struct ib_qp_init_attr *attr,
264 				 struct ib_udata *udata)
265 {
266 	struct mana_ib_pd *pd = container_of(ibpd, struct mana_ib_pd, ibpd);
267 	struct mana_ib_qp *qp = container_of(ibqp, struct mana_ib_qp, ibqp);
268 	struct mana_ib_dev *mdev =
269 		container_of(ibpd->device, struct mana_ib_dev, ib_dev);
270 	struct mana_ib_cq *send_cq =
271 		container_of(attr->send_cq, struct mana_ib_cq, ibcq);
272 	struct mana_ib_ucontext *mana_ucontext =
273 		rdma_udata_to_drv_context(udata, struct mana_ib_ucontext,
274 					  ibucontext);
275 	struct mana_ib_create_qp_resp resp = {};
276 	struct mana_ib_create_qp ucmd = {};
277 	struct mana_obj_spec wq_spec = {};
278 	struct mana_obj_spec cq_spec = {};
279 	struct mana_port_context *mpc;
280 	struct net_device *ndev;
281 	struct mana_eq *eq;
282 	int eq_vec;
283 	u32 port;
284 	int err;
285 
286 	if (!mana_ucontext || udata->inlen < sizeof(ucmd))
287 		return -EINVAL;
288 
289 	err = ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen));
290 	if (err) {
291 		ibdev_dbg(&mdev->ib_dev,
292 			  "Failed to copy from udata create qp-raw, %d\n", err);
293 		return err;
294 	}
295 
296 	if (attr->cap.max_send_wr > mdev->adapter_caps.max_qp_wr) {
297 		ibdev_dbg(&mdev->ib_dev,
298 			  "Requested max_send_wr %d exceeding limit\n",
299 			  attr->cap.max_send_wr);
300 		return -EINVAL;
301 	}
302 
303 	if (attr->cap.max_send_sge > MAX_TX_WQE_SGL_ENTRIES) {
304 		ibdev_dbg(&mdev->ib_dev,
305 			  "Requested max_send_sge %d exceeding limit\n",
306 			  attr->cap.max_send_sge);
307 		return -EINVAL;
308 	}
309 
310 	port = ucmd.port;
311 	ndev = mana_ib_get_netdev(ibpd->device, port);
312 	if (!ndev) {
313 		ibdev_dbg(&mdev->ib_dev, "Invalid port %u in creating qp\n",
314 			  port);
315 		return -EINVAL;
316 	}
317 	mpc = netdev_priv(ndev);
318 	ibdev_dbg(&mdev->ib_dev, "port %u ndev %p mpc %p\n", port, ndev, mpc);
319 
320 	err = mana_ib_cfg_vport(mdev, port, pd, mana_ucontext->doorbell);
321 	if (err)
322 		return -ENODEV;
323 
324 	qp->port = port;
325 
326 	ibdev_dbg(&mdev->ib_dev, "ucmd sq_buf_addr 0x%llx port %u\n",
327 		  ucmd.sq_buf_addr, ucmd.port);
328 
329 	err = mana_ib_create_queue(mdev, ucmd.sq_buf_addr, ucmd.sq_buf_size, &qp->raw_sq);
330 	if (err) {
331 		ibdev_dbg(&mdev->ib_dev,
332 			  "Failed to create queue for create qp-raw, err %d\n", err);
333 		goto err_free_vport;
334 	}
335 
336 	/* Create a WQ on the same port handle used by the Ethernet */
337 	wq_spec.gdma_region = qp->raw_sq.gdma_region;
338 	wq_spec.queue_size = ucmd.sq_buf_size;
339 
340 	cq_spec.gdma_region = send_cq->queue.gdma_region;
341 	cq_spec.queue_size = send_cq->cqe * COMP_ENTRY_SIZE;
342 	cq_spec.modr_ctx_id = 0;
343 	eq_vec = send_cq->comp_vector;
344 	eq = &mpc->ac->eqs[eq_vec];
345 	cq_spec.attached_eq = eq->eq->id;
346 
347 	err = mana_create_wq_obj(mpc, mpc->port_handle, GDMA_SQ, &wq_spec,
348 				 &cq_spec, &qp->qp_handle);
349 	if (err) {
350 		ibdev_dbg(&mdev->ib_dev,
351 			  "Failed to create wq for create raw-qp, err %d\n",
352 			  err);
353 		goto err_destroy_queue;
354 	}
355 
356 	/* The GDMA regions are now owned by the WQ object */
357 	qp->raw_sq.gdma_region = GDMA_INVALID_DMA_REGION;
358 	send_cq->queue.gdma_region = GDMA_INVALID_DMA_REGION;
359 
360 	qp->raw_sq.id = wq_spec.queue_index;
361 	send_cq->queue.id = cq_spec.queue_index;
362 
363 	/* Create CQ table entry */
364 	err = mana_ib_install_cq_cb(mdev, send_cq);
365 	if (err)
366 		goto err_destroy_wq_obj;
367 
368 	ibdev_dbg(&mdev->ib_dev,
369 		  "qp->qp_handle 0x%llx sq id %llu cq id %llu\n",
370 		  qp->qp_handle, qp->raw_sq.id, send_cq->queue.id);
371 
372 	resp.sqid = qp->raw_sq.id;
373 	resp.cqid = send_cq->queue.id;
374 	resp.tx_vp_offset = pd->tx_vp_offset;
375 
376 	err = ib_copy_to_udata(udata, &resp, sizeof(resp));
377 	if (err) {
378 		ibdev_dbg(&mdev->ib_dev,
379 			  "Failed copy udata for create qp-raw, %d\n",
380 			  err);
381 		goto err_remove_cq_cb;
382 	}
383 
384 	return 0;
385 
386 err_remove_cq_cb:
387 	mana_ib_remove_cq_cb(mdev, send_cq);
388 
389 err_destroy_wq_obj:
390 	mana_destroy_wq_obj(mpc, GDMA_SQ, qp->qp_handle);
391 
392 err_destroy_queue:
393 	mana_ib_destroy_queue(mdev, &qp->raw_sq);
394 
395 err_free_vport:
396 	mana_ib_uncfg_vport(mdev, pd, port);
397 
398 	return err;
399 }
400 
401 int mana_ib_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *attr,
402 		      struct ib_udata *udata)
403 {
404 	switch (attr->qp_type) {
405 	case IB_QPT_RAW_PACKET:
406 		/* When rwq_ind_tbl is used, it's for creating WQs for RSS */
407 		if (attr->rwq_ind_tbl)
408 			return mana_ib_create_qp_rss(ibqp, ibqp->pd, attr,
409 						     udata);
410 
411 		return mana_ib_create_qp_raw(ibqp, ibqp->pd, attr, udata);
412 	default:
413 		/* Creating QP other than IB_QPT_RAW_PACKET is not supported */
414 		ibdev_dbg(ibqp->device, "Creating QP type %u not supported\n",
415 			  attr->qp_type);
416 	}
417 
418 	return -EINVAL;
419 }
420 
421 int mana_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
422 		      int attr_mask, struct ib_udata *udata)
423 {
424 	/* modify_qp is not supported by this version of the driver */
425 	return -EOPNOTSUPP;
426 }
427 
428 static int mana_ib_destroy_qp_rss(struct mana_ib_qp *qp,
429 				  struct ib_rwq_ind_table *ind_tbl,
430 				  struct ib_udata *udata)
431 {
432 	struct mana_ib_dev *mdev =
433 		container_of(qp->ibqp.device, struct mana_ib_dev, ib_dev);
434 	struct mana_port_context *mpc;
435 	struct net_device *ndev;
436 	struct mana_ib_wq *wq;
437 	struct ib_wq *ibwq;
438 	int i;
439 
440 	ndev = mana_ib_get_netdev(qp->ibqp.device, qp->port);
441 	mpc = netdev_priv(ndev);
442 
443 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
444 		ibwq = ind_tbl->ind_tbl[i];
445 		wq = container_of(ibwq, struct mana_ib_wq, ibwq);
446 		ibdev_dbg(&mdev->ib_dev, "destroying wq->rx_object %llu\n",
447 			  wq->rx_object);
448 		mana_destroy_wq_obj(mpc, GDMA_RQ, wq->rx_object);
449 	}
450 
451 	return 0;
452 }
453 
454 static int mana_ib_destroy_qp_raw(struct mana_ib_qp *qp, struct ib_udata *udata)
455 {
456 	struct mana_ib_dev *mdev =
457 		container_of(qp->ibqp.device, struct mana_ib_dev, ib_dev);
458 	struct ib_pd *ibpd = qp->ibqp.pd;
459 	struct mana_port_context *mpc;
460 	struct net_device *ndev;
461 	struct mana_ib_pd *pd;
462 
463 	ndev = mana_ib_get_netdev(qp->ibqp.device, qp->port);
464 	mpc = netdev_priv(ndev);
465 	pd = container_of(ibpd, struct mana_ib_pd, ibpd);
466 
467 	mana_destroy_wq_obj(mpc, GDMA_SQ, qp->qp_handle);
468 
469 	mana_ib_destroy_queue(mdev, &qp->raw_sq);
470 
471 	mana_ib_uncfg_vport(mdev, pd, qp->port);
472 
473 	return 0;
474 }
475 
476 int mana_ib_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
477 {
478 	struct mana_ib_qp *qp = container_of(ibqp, struct mana_ib_qp, ibqp);
479 
480 	switch (ibqp->qp_type) {
481 	case IB_QPT_RAW_PACKET:
482 		if (ibqp->rwq_ind_tbl)
483 			return mana_ib_destroy_qp_rss(qp, ibqp->rwq_ind_tbl,
484 						      udata);
485 
486 		return mana_ib_destroy_qp_raw(qp, udata);
487 
488 	default:
489 		ibdev_dbg(ibqp->device, "Unexpected QP type %u\n",
490 			  ibqp->qp_type);
491 	}
492 
493 	return -ENOENT;
494 }
495