xref: /linux/drivers/infiniband/hw/mana/qp.c (revision 0e9b70c1e3623fa110fb6be553e644524228ef60)
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 *req = NULL;
17 	struct mana_cfg_rx_steer_resp resp = {};
18 	mana_handle_t *req_indir_tab;
19 	struct gdma_context *gc;
20 	struct gdma_dev *mdev;
21 	u32 req_buf_size;
22 	int i, err;
23 
24 	mdev = dev->gdma_dev;
25 	gc = mdev->gdma_context;
26 
27 	req_buf_size =
28 		sizeof(*req) + sizeof(mana_handle_t) * MANA_INDIRECT_TABLE_SIZE;
29 	req = kzalloc(req_buf_size, GFP_KERNEL);
30 	if (!req)
31 		return -ENOMEM;
32 
33 	mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size,
34 			     sizeof(resp));
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 = mdev->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 
50 	req_indir_tab = (mana_handle_t *)(req + 1);
51 	/* The ind table passed to the hardware must have
52 	 * MANA_INDIRECT_TABLE_SIZE entries. Adjust the verb
53 	 * ind_table to MANA_INDIRECT_TABLE_SIZE if required
54 	 */
55 	ibdev_dbg(&dev->ib_dev, "ind table size %u\n", 1 << log_ind_tbl_size);
56 	for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++) {
57 		req_indir_tab[i] = ind_table[i % (1 << log_ind_tbl_size)];
58 		ibdev_dbg(&dev->ib_dev, "index %u handle 0x%llx\n", i,
59 			  req_indir_tab[i]);
60 	}
61 
62 	req->update_hashkey = true;
63 	if (rx_hash_key_len)
64 		memcpy(req->hashkey, rx_hash_key, rx_hash_key_len);
65 	else
66 		netdev_rss_key_fill(req->hashkey, MANA_HASH_KEY_SIZE);
67 
68 	ibdev_dbg(&dev->ib_dev, "vport handle %llu default_rxobj 0x%llx\n",
69 		  req->vport, default_rxobj);
70 
71 	err = mana_gd_send_request(gc, req_buf_size, req, sizeof(resp), &resp);
72 	if (err) {
73 		netdev_err(ndev, "Failed to configure vPort RX: %d\n", err);
74 		goto out;
75 	}
76 
77 	if (resp.hdr.status) {
78 		netdev_err(ndev, "vPort RX configuration failed: 0x%x\n",
79 			   resp.hdr.status);
80 		err = -EPROTO;
81 		goto out;
82 	}
83 
84 	netdev_info(ndev, "Configured steering vPort %llu log_entries %u\n",
85 		    mpc->port_handle, log_ind_tbl_size);
86 
87 out:
88 	kfree(req);
89 	return err;
90 }
91 
92 static int mana_ib_create_qp_rss(struct ib_qp *ibqp, struct ib_pd *pd,
93 				 struct ib_qp_init_attr *attr,
94 				 struct ib_udata *udata)
95 {
96 	struct mana_ib_qp *qp = container_of(ibqp, struct mana_ib_qp, ibqp);
97 	struct mana_ib_dev *mdev =
98 		container_of(pd->device, struct mana_ib_dev, ib_dev);
99 	struct ib_rwq_ind_table *ind_tbl = attr->rwq_ind_tbl;
100 	struct mana_ib_create_qp_rss_resp resp = {};
101 	struct mana_ib_create_qp_rss ucmd = {};
102 	struct gdma_dev *gd = mdev->gdma_dev;
103 	mana_handle_t *mana_ind_table;
104 	struct mana_port_context *mpc;
105 	struct mana_context *mc;
106 	struct net_device *ndev;
107 	struct mana_ib_cq *cq;
108 	struct mana_ib_wq *wq;
109 	unsigned int ind_tbl_size;
110 	struct ib_cq *ibcq;
111 	struct ib_wq *ibwq;
112 	int i = 0;
113 	u32 port;
114 	int ret;
115 
116 	mc = gd->driver_data;
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 > MAX_SEND_BUFFERS_PER_QUEUE) {
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 	if (port < 1 || port > mc->num_ports) {
161 		ibdev_dbg(&mdev->ib_dev, "Invalid port %u in creating qp\n",
162 			  port);
163 		return -EINVAL;
164 	}
165 	ndev = mc->ports[port - 1];
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 	qp->port = port;
179 
180 	for (i = 0; i < ind_tbl_size; i++) {
181 		struct mana_obj_spec wq_spec = {};
182 		struct mana_obj_spec cq_spec = {};
183 
184 		ibwq = ind_tbl->ind_tbl[i];
185 		wq = container_of(ibwq, struct mana_ib_wq, ibwq);
186 
187 		ibcq = ibwq->cq;
188 		cq = container_of(ibcq, struct mana_ib_cq, ibcq);
189 
190 		wq_spec.gdma_region = wq->gdma_region;
191 		wq_spec.queue_size = wq->wq_buf_size;
192 
193 		cq_spec.gdma_region = cq->gdma_region;
194 		cq_spec.queue_size = cq->cqe * COMP_ENTRY_SIZE;
195 		cq_spec.modr_ctx_id = 0;
196 		cq_spec.attached_eq = GDMA_CQ_NO_EQ;
197 
198 		ret = mana_create_wq_obj(mpc, mpc->port_handle, GDMA_RQ,
199 					 &wq_spec, &cq_spec, &wq->rx_object);
200 		if (ret)
201 			goto fail;
202 
203 		/* The GDMA regions are now owned by the WQ object */
204 		wq->gdma_region = GDMA_INVALID_DMA_REGION;
205 		cq->gdma_region = GDMA_INVALID_DMA_REGION;
206 
207 		wq->id = wq_spec.queue_index;
208 		cq->id = cq_spec.queue_index;
209 
210 		ibdev_dbg(&mdev->ib_dev,
211 			  "ret %d rx_object 0x%llx wq id %llu cq id %llu\n",
212 			  ret, wq->rx_object, wq->id, cq->id);
213 
214 		resp.entries[i].cqid = cq->id;
215 		resp.entries[i].wqid = wq->id;
216 
217 		mana_ind_table[i] = wq->rx_object;
218 	}
219 	resp.num_entries = i;
220 
221 	ret = mana_ib_cfg_vport_steering(mdev, ndev, wq->rx_object,
222 					 mana_ind_table,
223 					 ind_tbl->log_ind_tbl_size,
224 					 ucmd.rx_hash_key_len,
225 					 ucmd.rx_hash_key);
226 	if (ret)
227 		goto fail;
228 
229 	ret = ib_copy_to_udata(udata, &resp, sizeof(resp));
230 	if (ret) {
231 		ibdev_dbg(&mdev->ib_dev,
232 			  "Failed to copy to udata create rss-qp, %d\n",
233 			  ret);
234 		goto fail;
235 	}
236 
237 	kfree(mana_ind_table);
238 
239 	return 0;
240 
241 fail:
242 	while (i-- > 0) {
243 		ibwq = ind_tbl->ind_tbl[i];
244 		wq = container_of(ibwq, struct mana_ib_wq, ibwq);
245 		mana_destroy_wq_obj(mpc, GDMA_RQ, wq->rx_object);
246 	}
247 
248 	kfree(mana_ind_table);
249 
250 	return ret;
251 }
252 
253 static int mana_ib_create_qp_raw(struct ib_qp *ibqp, struct ib_pd *ibpd,
254 				 struct ib_qp_init_attr *attr,
255 				 struct ib_udata *udata)
256 {
257 	struct mana_ib_pd *pd = container_of(ibpd, struct mana_ib_pd, ibpd);
258 	struct mana_ib_qp *qp = container_of(ibqp, struct mana_ib_qp, ibqp);
259 	struct mana_ib_dev *mdev =
260 		container_of(ibpd->device, struct mana_ib_dev, ib_dev);
261 	struct mana_ib_cq *send_cq =
262 		container_of(attr->send_cq, struct mana_ib_cq, ibcq);
263 	struct mana_ib_ucontext *mana_ucontext =
264 		rdma_udata_to_drv_context(udata, struct mana_ib_ucontext,
265 					  ibucontext);
266 	struct mana_ib_create_qp_resp resp = {};
267 	struct gdma_dev *gd = mdev->gdma_dev;
268 	struct mana_ib_create_qp ucmd = {};
269 	struct mana_obj_spec wq_spec = {};
270 	struct mana_obj_spec cq_spec = {};
271 	struct mana_port_context *mpc;
272 	struct mana_context *mc;
273 	struct net_device *ndev;
274 	struct ib_umem *umem;
275 	int err;
276 	u32 port;
277 
278 	mc = gd->driver_data;
279 
280 	if (!mana_ucontext || udata->inlen < sizeof(ucmd))
281 		return -EINVAL;
282 
283 	err = ib_copy_from_udata(&ucmd, udata, min(sizeof(ucmd), udata->inlen));
284 	if (err) {
285 		ibdev_dbg(&mdev->ib_dev,
286 			  "Failed to copy from udata create qp-raw, %d\n", err);
287 		return err;
288 	}
289 
290 	/* IB ports start with 1, MANA Ethernet ports start with 0 */
291 	port = ucmd.port;
292 	if (port < 1 || port > mc->num_ports)
293 		return -EINVAL;
294 
295 	if (attr->cap.max_send_wr > MAX_SEND_BUFFERS_PER_QUEUE) {
296 		ibdev_dbg(&mdev->ib_dev,
297 			  "Requested max_send_wr %d exceeding limit\n",
298 			  attr->cap.max_send_wr);
299 		return -EINVAL;
300 	}
301 
302 	if (attr->cap.max_send_sge > MAX_TX_WQE_SGL_ENTRIES) {
303 		ibdev_dbg(&mdev->ib_dev,
304 			  "Requested max_send_sge %d exceeding limit\n",
305 			  attr->cap.max_send_sge);
306 		return -EINVAL;
307 	}
308 
309 	ndev = mc->ports[port - 1];
310 	mpc = netdev_priv(ndev);
311 	ibdev_dbg(&mdev->ib_dev, "port %u ndev %p mpc %p\n", port, ndev, mpc);
312 
313 	err = mana_ib_cfg_vport(mdev, port - 1, pd, mana_ucontext->doorbell);
314 	if (err)
315 		return -ENODEV;
316 
317 	qp->port = port;
318 
319 	ibdev_dbg(&mdev->ib_dev, "ucmd sq_buf_addr 0x%llx port %u\n",
320 		  ucmd.sq_buf_addr, ucmd.port);
321 
322 	umem = ib_umem_get(ibpd->device, ucmd.sq_buf_addr, ucmd.sq_buf_size,
323 			   IB_ACCESS_LOCAL_WRITE);
324 	if (IS_ERR(umem)) {
325 		err = PTR_ERR(umem);
326 		ibdev_dbg(&mdev->ib_dev,
327 			  "Failed to get umem for create qp-raw, err %d\n",
328 			  err);
329 		goto err_free_vport;
330 	}
331 	qp->sq_umem = umem;
332 
333 	err = mana_ib_gd_create_dma_region(mdev, qp->sq_umem,
334 					   &qp->sq_gdma_region);
335 	if (err) {
336 		ibdev_dbg(&mdev->ib_dev,
337 			  "Failed to create dma region for create qp-raw, %d\n",
338 			  err);
339 		goto err_release_umem;
340 	}
341 
342 	ibdev_dbg(&mdev->ib_dev,
343 		  "mana_ib_gd_create_dma_region ret %d gdma_region 0x%llx\n",
344 		  err, qp->sq_gdma_region);
345 
346 	/* Create a WQ on the same port handle used by the Ethernet */
347 	wq_spec.gdma_region = qp->sq_gdma_region;
348 	wq_spec.queue_size = ucmd.sq_buf_size;
349 
350 	cq_spec.gdma_region = send_cq->gdma_region;
351 	cq_spec.queue_size = send_cq->cqe * COMP_ENTRY_SIZE;
352 	cq_spec.modr_ctx_id = 0;
353 	cq_spec.attached_eq = GDMA_CQ_NO_EQ;
354 
355 	err = mana_create_wq_obj(mpc, mpc->port_handle, GDMA_SQ, &wq_spec,
356 				 &cq_spec, &qp->tx_object);
357 	if (err) {
358 		ibdev_dbg(&mdev->ib_dev,
359 			  "Failed to create wq for create raw-qp, err %d\n",
360 			  err);
361 		goto err_destroy_dma_region;
362 	}
363 
364 	/* The GDMA regions are now owned by the WQ object */
365 	qp->sq_gdma_region = GDMA_INVALID_DMA_REGION;
366 	send_cq->gdma_region = GDMA_INVALID_DMA_REGION;
367 
368 	qp->sq_id = wq_spec.queue_index;
369 	send_cq->id = cq_spec.queue_index;
370 
371 	ibdev_dbg(&mdev->ib_dev,
372 		  "ret %d qp->tx_object 0x%llx sq id %llu cq id %llu\n", err,
373 		  qp->tx_object, qp->sq_id, send_cq->id);
374 
375 	resp.sqid = qp->sq_id;
376 	resp.cqid = send_cq->id;
377 	resp.tx_vp_offset = pd->tx_vp_offset;
378 
379 	err = ib_copy_to_udata(udata, &resp, sizeof(resp));
380 	if (err) {
381 		ibdev_dbg(&mdev->ib_dev,
382 			  "Failed copy udata for create qp-raw, %d\n",
383 			  err);
384 		goto err_destroy_wq_obj;
385 	}
386 
387 	return 0;
388 
389 err_destroy_wq_obj:
390 	mana_destroy_wq_obj(mpc, GDMA_SQ, qp->tx_object);
391 
392 err_destroy_dma_region:
393 	mana_ib_gd_destroy_dma_region(mdev, qp->sq_gdma_region);
394 
395 err_release_umem:
396 	ib_umem_release(umem);
397 
398 err_free_vport:
399 	mana_ib_uncfg_vport(mdev, pd, port - 1);
400 
401 	return err;
402 }
403 
404 int mana_ib_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *attr,
405 		      struct ib_udata *udata)
406 {
407 	switch (attr->qp_type) {
408 	case IB_QPT_RAW_PACKET:
409 		/* When rwq_ind_tbl is used, it's for creating WQs for RSS */
410 		if (attr->rwq_ind_tbl)
411 			return mana_ib_create_qp_rss(ibqp, ibqp->pd, attr,
412 						     udata);
413 
414 		return mana_ib_create_qp_raw(ibqp, ibqp->pd, attr, udata);
415 	default:
416 		/* Creating QP other than IB_QPT_RAW_PACKET is not supported */
417 		ibdev_dbg(ibqp->device, "Creating QP type %u not supported\n",
418 			  attr->qp_type);
419 	}
420 
421 	return -EINVAL;
422 }
423 
424 int mana_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
425 		      int attr_mask, struct ib_udata *udata)
426 {
427 	/* modify_qp is not supported by this version of the driver */
428 	return -EOPNOTSUPP;
429 }
430 
431 static int mana_ib_destroy_qp_rss(struct mana_ib_qp *qp,
432 				  struct ib_rwq_ind_table *ind_tbl,
433 				  struct ib_udata *udata)
434 {
435 	struct mana_ib_dev *mdev =
436 		container_of(qp->ibqp.device, struct mana_ib_dev, ib_dev);
437 	struct gdma_dev *gd = mdev->gdma_dev;
438 	struct mana_port_context *mpc;
439 	struct mana_context *mc;
440 	struct net_device *ndev;
441 	struct mana_ib_wq *wq;
442 	struct ib_wq *ibwq;
443 	int i;
444 
445 	mc = gd->driver_data;
446 	ndev = mc->ports[qp->port - 1];
447 	mpc = netdev_priv(ndev);
448 
449 	for (i = 0; i < (1 << ind_tbl->log_ind_tbl_size); i++) {
450 		ibwq = ind_tbl->ind_tbl[i];
451 		wq = container_of(ibwq, struct mana_ib_wq, ibwq);
452 		ibdev_dbg(&mdev->ib_dev, "destroying wq->rx_object %llu\n",
453 			  wq->rx_object);
454 		mana_destroy_wq_obj(mpc, GDMA_RQ, wq->rx_object);
455 	}
456 
457 	return 0;
458 }
459 
460 static int mana_ib_destroy_qp_raw(struct mana_ib_qp *qp, struct ib_udata *udata)
461 {
462 	struct mana_ib_dev *mdev =
463 		container_of(qp->ibqp.device, struct mana_ib_dev, ib_dev);
464 	struct gdma_dev *gd = mdev->gdma_dev;
465 	struct ib_pd *ibpd = qp->ibqp.pd;
466 	struct mana_port_context *mpc;
467 	struct mana_context *mc;
468 	struct net_device *ndev;
469 	struct mana_ib_pd *pd;
470 
471 	mc = gd->driver_data;
472 	ndev = mc->ports[qp->port - 1];
473 	mpc = netdev_priv(ndev);
474 	pd = container_of(ibpd, struct mana_ib_pd, ibpd);
475 
476 	mana_destroy_wq_obj(mpc, GDMA_SQ, qp->tx_object);
477 
478 	if (qp->sq_umem) {
479 		mana_ib_gd_destroy_dma_region(mdev, qp->sq_gdma_region);
480 		ib_umem_release(qp->sq_umem);
481 	}
482 
483 	mana_ib_uncfg_vport(mdev, pd, qp->port - 1);
484 
485 	return 0;
486 }
487 
488 int mana_ib_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
489 {
490 	struct mana_ib_qp *qp = container_of(ibqp, struct mana_ib_qp, ibqp);
491 
492 	switch (ibqp->qp_type) {
493 	case IB_QPT_RAW_PACKET:
494 		if (ibqp->rwq_ind_tbl)
495 			return mana_ib_destroy_qp_rss(qp, ibqp->rwq_ind_tbl,
496 						      udata);
497 
498 		return mana_ib_destroy_qp_raw(qp, udata);
499 
500 	default:
501 		ibdev_dbg(ibqp->device, "Unexpected QP type %u\n",
502 			  ibqp->qp_type);
503 	}
504 
505 	return -ENOENT;
506 }
507