xref: /linux/drivers/infiniband/hw/mana/main.c (revision 092e335082f22880207384ad736729c67d784665)
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 void mana_ib_uncfg_vport(struct mana_ib_dev *dev, struct mana_ib_pd *pd,
9 			 u32 port)
10 {
11 	struct mana_port_context *mpc;
12 	struct net_device *ndev;
13 
14 	ndev = mana_ib_get_netdev(&dev->ib_dev, port);
15 	mpc = netdev_priv(ndev);
16 
17 	mutex_lock(&pd->vport_mutex);
18 
19 	pd->vport_use_count--;
20 	WARN_ON(pd->vport_use_count < 0);
21 
22 	if (!pd->vport_use_count)
23 		mana_uncfg_vport(mpc);
24 
25 	mutex_unlock(&pd->vport_mutex);
26 }
27 
28 int mana_ib_cfg_vport(struct mana_ib_dev *dev, u32 port, struct mana_ib_pd *pd,
29 		      u32 doorbell_id)
30 {
31 	struct mana_port_context *mpc;
32 	struct net_device *ndev;
33 	int err;
34 
35 	ndev = mana_ib_get_netdev(&dev->ib_dev, port);
36 	mpc = netdev_priv(ndev);
37 
38 	mutex_lock(&pd->vport_mutex);
39 
40 	pd->vport_use_count++;
41 	if (pd->vport_use_count > 1) {
42 		ibdev_dbg(&dev->ib_dev,
43 			  "Skip as this PD is already configured vport\n");
44 		mutex_unlock(&pd->vport_mutex);
45 		return 0;
46 	}
47 
48 	err = mana_cfg_vport(mpc, pd->pdn, doorbell_id);
49 	if (err) {
50 		pd->vport_use_count--;
51 		mutex_unlock(&pd->vport_mutex);
52 
53 		ibdev_dbg(&dev->ib_dev, "Failed to configure vPort %d\n", err);
54 		return err;
55 	}
56 
57 	mutex_unlock(&pd->vport_mutex);
58 
59 	pd->tx_shortform_allowed = mpc->tx_shortform_allowed;
60 	pd->tx_vp_offset = mpc->tx_vp_offset;
61 
62 	ibdev_dbg(&dev->ib_dev, "vport handle %llx pdid %x doorbell_id %x\n",
63 		  mpc->port_handle, pd->pdn, doorbell_id);
64 
65 	return 0;
66 }
67 
68 int mana_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
69 {
70 	struct mana_ib_pd *pd = container_of(ibpd, struct mana_ib_pd, ibpd);
71 	struct ib_device *ibdev = ibpd->device;
72 	struct gdma_create_pd_resp resp = {};
73 	struct gdma_create_pd_req req = {};
74 	enum gdma_pd_flags flags = 0;
75 	struct mana_ib_dev *dev;
76 	struct gdma_context *gc;
77 	int err;
78 
79 	dev = container_of(ibdev, struct mana_ib_dev, ib_dev);
80 	gc = mdev_to_gc(dev);
81 
82 	mana_gd_init_req_hdr(&req.hdr, GDMA_CREATE_PD, sizeof(req),
83 			     sizeof(resp));
84 
85 	if (!udata)
86 		flags |= GDMA_PD_FLAG_ALLOW_GPA_MR;
87 
88 	req.flags = flags;
89 	err = mana_gd_send_request(gc, sizeof(req), &req,
90 				   sizeof(resp), &resp);
91 
92 	if (err || resp.hdr.status) {
93 		ibdev_dbg(&dev->ib_dev,
94 			  "Failed to get pd_id err %d status %u\n", err,
95 			  resp.hdr.status);
96 		if (!err)
97 			err = -EPROTO;
98 
99 		return err;
100 	}
101 
102 	pd->pd_handle = resp.pd_handle;
103 	pd->pdn = resp.pd_id;
104 	ibdev_dbg(&dev->ib_dev, "pd_handle 0x%llx pd_id %d\n",
105 		  pd->pd_handle, pd->pdn);
106 
107 	mutex_init(&pd->vport_mutex);
108 	pd->vport_use_count = 0;
109 	return 0;
110 }
111 
112 int mana_ib_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
113 {
114 	struct mana_ib_pd *pd = container_of(ibpd, struct mana_ib_pd, ibpd);
115 	struct ib_device *ibdev = ibpd->device;
116 	struct gdma_destory_pd_resp resp = {};
117 	struct gdma_destroy_pd_req req = {};
118 	struct mana_ib_dev *dev;
119 	struct gdma_context *gc;
120 	int err;
121 
122 	dev = container_of(ibdev, struct mana_ib_dev, ib_dev);
123 	gc = mdev_to_gc(dev);
124 
125 	mana_gd_init_req_hdr(&req.hdr, GDMA_DESTROY_PD, sizeof(req),
126 			     sizeof(resp));
127 
128 	req.pd_handle = pd->pd_handle;
129 	err = mana_gd_send_request(gc, sizeof(req), &req,
130 				   sizeof(resp), &resp);
131 
132 	if (err || resp.hdr.status) {
133 		ibdev_dbg(&dev->ib_dev,
134 			  "Failed to destroy pd_handle 0x%llx err %d status %u",
135 			  pd->pd_handle, err, resp.hdr.status);
136 		if (!err)
137 			err = -EPROTO;
138 	}
139 
140 	return err;
141 }
142 
143 static int mana_gd_destroy_doorbell_page(struct gdma_context *gc,
144 					 int doorbell_page)
145 {
146 	struct gdma_destroy_resource_range_req req = {};
147 	struct gdma_resp_hdr resp = {};
148 	int err;
149 
150 	mana_gd_init_req_hdr(&req.hdr, GDMA_DESTROY_RESOURCE_RANGE,
151 			     sizeof(req), sizeof(resp));
152 
153 	req.resource_type = GDMA_RESOURCE_DOORBELL_PAGE;
154 	req.num_resources = 1;
155 	req.allocated_resources = doorbell_page;
156 
157 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
158 	if (err || resp.status) {
159 		dev_err(gc->dev,
160 			"Failed to destroy doorbell page: ret %d, 0x%x\n",
161 			err, resp.status);
162 		return err ?: -EPROTO;
163 	}
164 
165 	return 0;
166 }
167 
168 static int mana_gd_allocate_doorbell_page(struct gdma_context *gc,
169 					  int *doorbell_page)
170 {
171 	struct gdma_allocate_resource_range_req req = {};
172 	struct gdma_allocate_resource_range_resp resp = {};
173 	int err;
174 
175 	mana_gd_init_req_hdr(&req.hdr, GDMA_ALLOCATE_RESOURCE_RANGE,
176 			     sizeof(req), sizeof(resp));
177 
178 	req.resource_type = GDMA_RESOURCE_DOORBELL_PAGE;
179 	req.num_resources = 1;
180 	req.alignment = PAGE_SIZE / MANA_PAGE_SIZE;
181 
182 	/* Have GDMA start searching from 0 */
183 	req.allocated_resources = 0;
184 
185 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
186 	if (err || resp.hdr.status) {
187 		dev_err(gc->dev,
188 			"Failed to allocate doorbell page: ret %d, 0x%x\n",
189 			err, resp.hdr.status);
190 		return err ?: -EPROTO;
191 	}
192 
193 	*doorbell_page = resp.allocated_resources;
194 
195 	return 0;
196 }
197 
198 int mana_ib_alloc_ucontext(struct ib_ucontext *ibcontext,
199 			   struct ib_udata *udata)
200 {
201 	struct mana_ib_ucontext *ucontext =
202 		container_of(ibcontext, struct mana_ib_ucontext, ibucontext);
203 	struct ib_device *ibdev = ibcontext->device;
204 	struct mana_ib_dev *mdev;
205 	struct gdma_context *gc;
206 	int doorbell_page;
207 	int ret;
208 
209 	mdev = container_of(ibdev, struct mana_ib_dev, ib_dev);
210 	gc = mdev_to_gc(mdev);
211 
212 	/* Allocate a doorbell page index */
213 	ret = mana_gd_allocate_doorbell_page(gc, &doorbell_page);
214 	if (ret) {
215 		ibdev_dbg(ibdev, "Failed to allocate doorbell page %d\n", ret);
216 		return ret;
217 	}
218 
219 	ibdev_dbg(ibdev, "Doorbell page allocated %d\n", doorbell_page);
220 
221 	ucontext->doorbell = doorbell_page;
222 
223 	return 0;
224 }
225 
226 void mana_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
227 {
228 	struct mana_ib_ucontext *mana_ucontext =
229 		container_of(ibcontext, struct mana_ib_ucontext, ibucontext);
230 	struct ib_device *ibdev = ibcontext->device;
231 	struct mana_ib_dev *mdev;
232 	struct gdma_context *gc;
233 	int ret;
234 
235 	mdev = container_of(ibdev, struct mana_ib_dev, ib_dev);
236 	gc = mdev_to_gc(mdev);
237 
238 	ret = mana_gd_destroy_doorbell_page(gc, mana_ucontext->doorbell);
239 	if (ret)
240 		ibdev_dbg(ibdev, "Failed to destroy doorbell page %d\n", ret);
241 }
242 
243 int mana_ib_create_kernel_queue(struct mana_ib_dev *mdev, u32 size, enum gdma_queue_type type,
244 				struct mana_ib_queue *queue)
245 {
246 	struct gdma_context *gc = mdev_to_gc(mdev);
247 	struct gdma_queue_spec spec = {};
248 	int err;
249 
250 	queue->id = INVALID_QUEUE_ID;
251 	queue->gdma_region = GDMA_INVALID_DMA_REGION;
252 	spec.type = type;
253 	spec.monitor_avl_buf = false;
254 	spec.queue_size = size;
255 	err = mana_gd_create_mana_wq_cq(&gc->mana_ib, &spec, &queue->kmem);
256 	if (err)
257 		return err;
258 	/* take ownership into mana_ib from mana */
259 	queue->gdma_region = queue->kmem->mem_info.dma_region_handle;
260 	queue->kmem->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
261 	return 0;
262 }
263 
264 int mana_ib_create_queue(struct mana_ib_dev *mdev, u64 addr, u32 size,
265 			 struct mana_ib_queue *queue)
266 {
267 	struct ib_umem *umem;
268 	int err;
269 
270 	queue->umem = NULL;
271 	queue->id = INVALID_QUEUE_ID;
272 	queue->gdma_region = GDMA_INVALID_DMA_REGION;
273 
274 	umem = ib_umem_get(&mdev->ib_dev, addr, size, IB_ACCESS_LOCAL_WRITE);
275 	if (IS_ERR(umem)) {
276 		err = PTR_ERR(umem);
277 		ibdev_dbg(&mdev->ib_dev, "Failed to get umem, %d\n", err);
278 		return err;
279 	}
280 
281 	err = mana_ib_create_zero_offset_dma_region(mdev, umem, &queue->gdma_region);
282 	if (err) {
283 		ibdev_dbg(&mdev->ib_dev, "Failed to create dma region, %d\n", err);
284 		goto free_umem;
285 	}
286 	queue->umem = umem;
287 
288 	ibdev_dbg(&mdev->ib_dev, "created dma region 0x%llx\n", queue->gdma_region);
289 
290 	return 0;
291 free_umem:
292 	ib_umem_release(umem);
293 	return err;
294 }
295 
296 void mana_ib_destroy_queue(struct mana_ib_dev *mdev, struct mana_ib_queue *queue)
297 {
298 	/* Ignore return code as there is not much we can do about it.
299 	 * The error message is printed inside.
300 	 */
301 	mana_ib_gd_destroy_dma_region(mdev, queue->gdma_region);
302 	ib_umem_release(queue->umem);
303 	if (queue->kmem)
304 		mana_gd_destroy_queue(mdev_to_gc(mdev), queue->kmem);
305 }
306 
307 static int
308 mana_ib_gd_first_dma_region(struct mana_ib_dev *dev,
309 			    struct gdma_context *gc,
310 			    struct gdma_create_dma_region_req *create_req,
311 			    size_t num_pages, mana_handle_t *gdma_region,
312 			    u32 expected_status)
313 {
314 	struct gdma_create_dma_region_resp create_resp = {};
315 	unsigned int create_req_msg_size;
316 	int err;
317 
318 	create_req_msg_size =
319 		struct_size(create_req, page_addr_list, num_pages);
320 	create_req->page_addr_list_len = num_pages;
321 
322 	err = mana_gd_send_request(gc, create_req_msg_size, create_req,
323 				   sizeof(create_resp), &create_resp);
324 	if (err || create_resp.hdr.status != expected_status) {
325 		ibdev_dbg(&dev->ib_dev,
326 			  "Failed to create DMA region: %d, 0x%x\n",
327 			  err, create_resp.hdr.status);
328 		if (!err)
329 			err = -EPROTO;
330 
331 		return err;
332 	}
333 
334 	*gdma_region = create_resp.dma_region_handle;
335 	ibdev_dbg(&dev->ib_dev, "Created DMA region handle 0x%llx\n",
336 		  *gdma_region);
337 
338 	return 0;
339 }
340 
341 static int
342 mana_ib_gd_add_dma_region(struct mana_ib_dev *dev, struct gdma_context *gc,
343 			  struct gdma_dma_region_add_pages_req *add_req,
344 			  unsigned int num_pages, u32 expected_status)
345 {
346 	unsigned int add_req_msg_size =
347 		struct_size(add_req, page_addr_list, num_pages);
348 	struct gdma_general_resp add_resp = {};
349 	int err;
350 
351 	mana_gd_init_req_hdr(&add_req->hdr, GDMA_DMA_REGION_ADD_PAGES,
352 			     add_req_msg_size, sizeof(add_resp));
353 	add_req->page_addr_list_len = num_pages;
354 
355 	err = mana_gd_send_request(gc, add_req_msg_size, add_req,
356 				   sizeof(add_resp), &add_resp);
357 	if (err || add_resp.hdr.status != expected_status) {
358 		ibdev_dbg(&dev->ib_dev,
359 			  "Failed to create DMA region: %d, 0x%x\n",
360 			  err, add_resp.hdr.status);
361 
362 		if (!err)
363 			err = -EPROTO;
364 
365 		return err;
366 	}
367 
368 	return 0;
369 }
370 
371 static int mana_ib_gd_create_dma_region(struct mana_ib_dev *dev, struct ib_umem *umem,
372 					mana_handle_t *gdma_region, unsigned long page_sz)
373 {
374 	struct gdma_dma_region_add_pages_req *add_req = NULL;
375 	size_t num_pages_processed = 0, num_pages_to_handle;
376 	struct gdma_create_dma_region_req *create_req;
377 	unsigned int create_req_msg_size;
378 	struct hw_channel_context *hwc;
379 	struct ib_block_iter biter;
380 	size_t max_pgs_add_cmd = 0;
381 	size_t max_pgs_create_cmd;
382 	struct gdma_context *gc;
383 	size_t num_pages_total;
384 	unsigned int tail = 0;
385 	u64 *page_addr_list;
386 	void *request_buf;
387 	int err = 0;
388 
389 	gc = mdev_to_gc(dev);
390 	hwc = gc->hwc.driver_data;
391 
392 	num_pages_total = ib_umem_num_dma_blocks(umem, page_sz);
393 
394 	max_pgs_create_cmd =
395 		(hwc->max_req_msg_size - sizeof(*create_req)) / sizeof(u64);
396 	num_pages_to_handle =
397 		min_t(size_t, num_pages_total, max_pgs_create_cmd);
398 	create_req_msg_size =
399 		struct_size(create_req, page_addr_list, num_pages_to_handle);
400 
401 	request_buf = kzalloc(hwc->max_req_msg_size, GFP_KERNEL);
402 	if (!request_buf)
403 		return -ENOMEM;
404 
405 	create_req = request_buf;
406 	mana_gd_init_req_hdr(&create_req->hdr, GDMA_CREATE_DMA_REGION,
407 			     create_req_msg_size,
408 			     sizeof(struct gdma_create_dma_region_resp));
409 
410 	create_req->length = umem->length;
411 	create_req->offset_in_page = ib_umem_dma_offset(umem, page_sz);
412 	create_req->gdma_page_type = order_base_2(page_sz) - MANA_PAGE_SHIFT;
413 	create_req->page_count = num_pages_total;
414 
415 	ibdev_dbg(&dev->ib_dev, "size_dma_region %lu num_pages_total %lu\n",
416 		  umem->length, num_pages_total);
417 
418 	ibdev_dbg(&dev->ib_dev, "page_sz %lu offset_in_page %u\n",
419 		  page_sz, create_req->offset_in_page);
420 
421 	ibdev_dbg(&dev->ib_dev, "num_pages_to_handle %lu, gdma_page_type %u",
422 		  num_pages_to_handle, create_req->gdma_page_type);
423 
424 	page_addr_list = create_req->page_addr_list;
425 	rdma_umem_for_each_dma_block(umem, &biter, page_sz) {
426 		u32 expected_status = 0;
427 
428 		page_addr_list[tail++] = rdma_block_iter_dma_address(&biter);
429 		if (tail < num_pages_to_handle)
430 			continue;
431 
432 		if (num_pages_processed + num_pages_to_handle <
433 		    num_pages_total)
434 			expected_status = GDMA_STATUS_MORE_ENTRIES;
435 
436 		if (!num_pages_processed) {
437 			/* First create message */
438 			err = mana_ib_gd_first_dma_region(dev, gc, create_req,
439 							  tail, gdma_region,
440 							  expected_status);
441 			if (err)
442 				goto out;
443 
444 			max_pgs_add_cmd = (hwc->max_req_msg_size -
445 				sizeof(*add_req)) / sizeof(u64);
446 
447 			add_req = request_buf;
448 			add_req->dma_region_handle = *gdma_region;
449 			add_req->reserved3 = 0;
450 			page_addr_list = add_req->page_addr_list;
451 		} else {
452 			/* Subsequent create messages */
453 			err = mana_ib_gd_add_dma_region(dev, gc, add_req, tail,
454 							expected_status);
455 			if (err)
456 				break;
457 		}
458 
459 		num_pages_processed += tail;
460 		tail = 0;
461 
462 		/* The remaining pages to create */
463 		num_pages_to_handle =
464 			min_t(size_t,
465 			      num_pages_total - num_pages_processed,
466 			      max_pgs_add_cmd);
467 	}
468 
469 	if (err)
470 		mana_ib_gd_destroy_dma_region(dev, *gdma_region);
471 
472 out:
473 	kfree(request_buf);
474 	return err;
475 }
476 
477 int mana_ib_create_dma_region(struct mana_ib_dev *dev, struct ib_umem *umem,
478 			      mana_handle_t *gdma_region, u64 virt)
479 {
480 	unsigned long page_sz;
481 
482 	page_sz = ib_umem_find_best_pgsz(umem, PAGE_SZ_BM, virt);
483 	if (!page_sz) {
484 		ibdev_dbg(&dev->ib_dev, "Failed to find page size.\n");
485 		return -EINVAL;
486 	}
487 
488 	return mana_ib_gd_create_dma_region(dev, umem, gdma_region, page_sz);
489 }
490 
491 int mana_ib_create_zero_offset_dma_region(struct mana_ib_dev *dev, struct ib_umem *umem,
492 					  mana_handle_t *gdma_region)
493 {
494 	unsigned long page_sz;
495 
496 	/* Hardware requires dma region to align to chosen page size */
497 	page_sz = ib_umem_find_best_pgoff(umem, PAGE_SZ_BM, 0);
498 	if (!page_sz) {
499 		ibdev_dbg(&dev->ib_dev, "Failed to find page size.\n");
500 		return -EINVAL;
501 	}
502 
503 	return mana_ib_gd_create_dma_region(dev, umem, gdma_region, page_sz);
504 }
505 
506 int mana_ib_gd_destroy_dma_region(struct mana_ib_dev *dev, u64 gdma_region)
507 {
508 	struct gdma_context *gc = mdev_to_gc(dev);
509 
510 	ibdev_dbg(&dev->ib_dev, "destroy dma region 0x%llx\n", gdma_region);
511 
512 	return mana_gd_destroy_dma_region(gc, gdma_region);
513 }
514 
515 int mana_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
516 {
517 	struct mana_ib_ucontext *mana_ucontext =
518 		container_of(ibcontext, struct mana_ib_ucontext, ibucontext);
519 	struct ib_device *ibdev = ibcontext->device;
520 	struct mana_ib_dev *mdev;
521 	struct gdma_context *gc;
522 	phys_addr_t pfn;
523 	pgprot_t prot;
524 	int ret;
525 
526 	mdev = container_of(ibdev, struct mana_ib_dev, ib_dev);
527 	gc = mdev_to_gc(mdev);
528 
529 	if (vma->vm_pgoff != 0) {
530 		ibdev_dbg(ibdev, "Unexpected vm_pgoff %lu\n", vma->vm_pgoff);
531 		return -EINVAL;
532 	}
533 
534 	/* Map to the page indexed by ucontext->doorbell */
535 	pfn = (gc->phys_db_page_base +
536 	       gc->db_page_size * mana_ucontext->doorbell) >>
537 	      PAGE_SHIFT;
538 	prot = pgprot_writecombine(vma->vm_page_prot);
539 
540 	ret = rdma_user_mmap_io(ibcontext, vma, pfn, PAGE_SIZE, prot,
541 				NULL);
542 	if (ret)
543 		ibdev_dbg(ibdev, "can't rdma_user_mmap_io ret %d\n", ret);
544 	else
545 		ibdev_dbg(ibdev, "mapped I/O pfn 0x%llx page_size %lu, ret %d\n",
546 			  pfn, PAGE_SIZE, ret);
547 
548 	return ret;
549 }
550 
551 int mana_ib_get_port_immutable(struct ib_device *ibdev, u32 port_num,
552 			       struct ib_port_immutable *immutable)
553 {
554 	struct ib_port_attr attr;
555 	int err;
556 
557 	err = ib_query_port(ibdev, port_num, &attr);
558 	if (err)
559 		return err;
560 
561 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
562 	immutable->gid_tbl_len = attr.gid_tbl_len;
563 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
564 	if (port_num == 1) {
565 		immutable->core_cap_flags |= RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
566 		immutable->max_mad_size = IB_MGMT_MAD_SIZE;
567 	}
568 
569 	return 0;
570 }
571 
572 int mana_ib_query_device(struct ib_device *ibdev, struct ib_device_attr *props,
573 			 struct ib_udata *uhw)
574 {
575 	struct mana_ib_dev *dev = container_of(ibdev,
576 			struct mana_ib_dev, ib_dev);
577 
578 	memset(props, 0, sizeof(*props));
579 	props->max_mr_size = MANA_IB_MAX_MR_SIZE;
580 	props->page_size_cap = PAGE_SZ_BM;
581 	props->max_qp = dev->adapter_caps.max_qp_count;
582 	props->max_qp_wr = dev->adapter_caps.max_qp_wr;
583 	props->device_cap_flags = IB_DEVICE_RC_RNR_NAK_GEN;
584 	props->max_send_sge = dev->adapter_caps.max_send_sge_count;
585 	props->max_recv_sge = dev->adapter_caps.max_recv_sge_count;
586 	props->max_sge_rd = dev->adapter_caps.max_recv_sge_count;
587 	props->max_cq = dev->adapter_caps.max_cq_count;
588 	props->max_cqe = dev->adapter_caps.max_qp_wr;
589 	props->max_mr = dev->adapter_caps.max_mr_count;
590 	props->max_pd = dev->adapter_caps.max_pd_count;
591 	props->max_qp_rd_atom = dev->adapter_caps.max_inbound_read_limit;
592 	props->max_res_rd_atom = props->max_qp_rd_atom * props->max_qp;
593 	props->max_qp_init_rd_atom = dev->adapter_caps.max_outbound_read_limit;
594 	props->atomic_cap = IB_ATOMIC_NONE;
595 	props->masked_atomic_cap = IB_ATOMIC_NONE;
596 	props->max_ah = INT_MAX;
597 	props->max_pkeys = 1;
598 	props->local_ca_ack_delay = MANA_CA_ACK_DELAY;
599 
600 	return 0;
601 }
602 
603 int mana_ib_query_port(struct ib_device *ibdev, u32 port,
604 		       struct ib_port_attr *props)
605 {
606 	struct net_device *ndev = mana_ib_get_netdev(ibdev, port);
607 
608 	if (!ndev)
609 		return -EINVAL;
610 
611 	memset(props, 0, sizeof(*props));
612 	props->max_mtu = IB_MTU_4096;
613 	props->active_mtu = ib_mtu_int_to_enum(ndev->mtu);
614 
615 	if (netif_carrier_ok(ndev) && netif_running(ndev)) {
616 		props->state = IB_PORT_ACTIVE;
617 		props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
618 	} else {
619 		props->state = IB_PORT_DOWN;
620 		props->phys_state = IB_PORT_PHYS_STATE_DISABLED;
621 	}
622 
623 	props->active_width = IB_WIDTH_4X;
624 	props->active_speed = IB_SPEED_EDR;
625 	props->pkey_tbl_len = 1;
626 	if (port == 1) {
627 		props->gid_tbl_len = 16;
628 		props->port_cap_flags = IB_PORT_CM_SUP;
629 		props->ip_gids = true;
630 	}
631 
632 	return 0;
633 }
634 
635 enum rdma_link_layer mana_ib_get_link_layer(struct ib_device *device, u32 port_num)
636 {
637 	return IB_LINK_LAYER_ETHERNET;
638 }
639 
640 int mana_ib_query_pkey(struct ib_device *ibdev, u32 port, u16 index, u16 *pkey)
641 {
642 	if (index != 0)
643 		return -EINVAL;
644 	*pkey = IB_DEFAULT_PKEY_FULL;
645 	return 0;
646 }
647 
648 int mana_ib_query_gid(struct ib_device *ibdev, u32 port, int index,
649 		      union ib_gid *gid)
650 {
651 	/* This version doesn't return GID properties */
652 	return 0;
653 }
654 
655 void mana_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
656 {
657 }
658 
659 int mana_ib_gd_query_adapter_caps(struct mana_ib_dev *dev)
660 {
661 	struct mana_ib_adapter_caps *caps = &dev->adapter_caps;
662 	struct mana_ib_query_adapter_caps_resp resp = {};
663 	struct mana_ib_query_adapter_caps_req req = {};
664 	int err;
665 
666 	mana_gd_init_req_hdr(&req.hdr, MANA_IB_GET_ADAPTER_CAP, sizeof(req),
667 			     sizeof(resp));
668 	req.hdr.resp.msg_version = GDMA_MESSAGE_V4;
669 	req.hdr.dev_id = dev->gdma_dev->dev_id;
670 
671 	err = mana_gd_send_request(mdev_to_gc(dev), sizeof(req),
672 				   &req, sizeof(resp), &resp);
673 
674 	if (err) {
675 		ibdev_err(&dev->ib_dev,
676 			  "Failed to query adapter caps err %d", err);
677 		return err;
678 	}
679 
680 	caps->max_sq_id = resp.max_sq_id;
681 	caps->max_rq_id = resp.max_rq_id;
682 	caps->max_cq_id = resp.max_cq_id;
683 	caps->max_qp_count = resp.max_qp_count;
684 	caps->max_cq_count = resp.max_cq_count;
685 	caps->max_mr_count = resp.max_mr_count;
686 	caps->max_pd_count = resp.max_pd_count;
687 	caps->max_inbound_read_limit = resp.max_inbound_read_limit;
688 	caps->max_outbound_read_limit = resp.max_outbound_read_limit;
689 	caps->mw_count = resp.mw_count;
690 	caps->max_srq_count = resp.max_srq_count;
691 	caps->max_qp_wr = min_t(u32,
692 				resp.max_requester_sq_size / GDMA_MAX_SQE_SIZE,
693 				resp.max_requester_rq_size / GDMA_MAX_RQE_SIZE);
694 	caps->max_inline_data_size = resp.max_inline_data_size;
695 	caps->max_send_sge_count = resp.max_send_sge_count;
696 	caps->max_recv_sge_count = resp.max_recv_sge_count;
697 	caps->feature_flags = resp.feature_flags;
698 
699 	return 0;
700 }
701 
702 static void
703 mana_ib_event_handler(void *ctx, struct gdma_queue *q, struct gdma_event *event)
704 {
705 	struct mana_ib_dev *mdev = (struct mana_ib_dev *)ctx;
706 	struct mana_ib_qp *qp;
707 	struct ib_event ev;
708 	u32 qpn;
709 
710 	switch (event->type) {
711 	case GDMA_EQE_RNIC_QP_FATAL:
712 		qpn = event->details[0];
713 		qp = mana_get_qp_ref(mdev, qpn, false);
714 		if (!qp)
715 			break;
716 		if (qp->ibqp.event_handler) {
717 			ev.device = qp->ibqp.device;
718 			ev.element.qp = &qp->ibqp;
719 			ev.event = IB_EVENT_QP_FATAL;
720 			qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
721 		}
722 		mana_put_qp_ref(qp);
723 		break;
724 	default:
725 		break;
726 	}
727 }
728 
729 int mana_ib_create_eqs(struct mana_ib_dev *mdev)
730 {
731 	struct gdma_context *gc = mdev_to_gc(mdev);
732 	struct gdma_queue_spec spec = {};
733 	int err, i;
734 
735 	spec.type = GDMA_EQ;
736 	spec.monitor_avl_buf = false;
737 	spec.queue_size = EQ_SIZE;
738 	spec.eq.callback = mana_ib_event_handler;
739 	spec.eq.context = mdev;
740 	spec.eq.log2_throttle_limit = LOG2_EQ_THROTTLE;
741 	spec.eq.msix_index = 0;
742 
743 	err = mana_gd_create_mana_eq(&gc->mana_ib, &spec, &mdev->fatal_err_eq);
744 	if (err)
745 		return err;
746 
747 	mdev->eqs = kcalloc(mdev->ib_dev.num_comp_vectors, sizeof(struct gdma_queue *),
748 			    GFP_KERNEL);
749 	if (!mdev->eqs) {
750 		err = -ENOMEM;
751 		goto destroy_fatal_eq;
752 	}
753 	spec.eq.callback = NULL;
754 	for (i = 0; i < mdev->ib_dev.num_comp_vectors; i++) {
755 		spec.eq.msix_index = (i + 1) % gc->num_msix_usable;
756 		err = mana_gd_create_mana_eq(mdev->gdma_dev, &spec, &mdev->eqs[i]);
757 		if (err)
758 			goto destroy_eqs;
759 	}
760 
761 	return 0;
762 
763 destroy_eqs:
764 	while (i-- > 0)
765 		mana_gd_destroy_queue(gc, mdev->eqs[i]);
766 	kfree(mdev->eqs);
767 destroy_fatal_eq:
768 	mana_gd_destroy_queue(gc, mdev->fatal_err_eq);
769 	return err;
770 }
771 
772 void mana_ib_destroy_eqs(struct mana_ib_dev *mdev)
773 {
774 	struct gdma_context *gc = mdev_to_gc(mdev);
775 	int i;
776 
777 	mana_gd_destroy_queue(gc, mdev->fatal_err_eq);
778 
779 	for (i = 0; i < mdev->ib_dev.num_comp_vectors; i++)
780 		mana_gd_destroy_queue(gc, mdev->eqs[i]);
781 
782 	kfree(mdev->eqs);
783 }
784 
785 int mana_ib_gd_create_rnic_adapter(struct mana_ib_dev *mdev)
786 {
787 	struct mana_rnic_create_adapter_resp resp = {};
788 	struct mana_rnic_create_adapter_req req = {};
789 	struct gdma_context *gc = mdev_to_gc(mdev);
790 	int err;
791 
792 	mana_gd_init_req_hdr(&req.hdr, MANA_IB_CREATE_ADAPTER, sizeof(req), sizeof(resp));
793 	req.hdr.req.msg_version = GDMA_MESSAGE_V2;
794 	req.hdr.dev_id = gc->mana_ib.dev_id;
795 	req.notify_eq_id = mdev->fatal_err_eq->id;
796 
797 	if (mdev->adapter_caps.feature_flags & MANA_IB_FEATURE_CLIENT_ERROR_CQE_SUPPORT)
798 		req.feature_flags |= MANA_IB_FEATURE_CLIENT_ERROR_CQE_REQUEST;
799 
800 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
801 	if (err) {
802 		ibdev_err(&mdev->ib_dev, "Failed to create RNIC adapter err %d", err);
803 		return err;
804 	}
805 	mdev->adapter_handle = resp.adapter;
806 
807 	return 0;
808 }
809 
810 int mana_ib_gd_destroy_rnic_adapter(struct mana_ib_dev *mdev)
811 {
812 	struct mana_rnic_destroy_adapter_resp resp = {};
813 	struct mana_rnic_destroy_adapter_req req = {};
814 	struct gdma_context *gc;
815 	int err;
816 
817 	gc = mdev_to_gc(mdev);
818 	mana_gd_init_req_hdr(&req.hdr, MANA_IB_DESTROY_ADAPTER, sizeof(req), sizeof(resp));
819 	req.hdr.dev_id = gc->mana_ib.dev_id;
820 	req.adapter = mdev->adapter_handle;
821 
822 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
823 	if (err) {
824 		ibdev_err(&mdev->ib_dev, "Failed to destroy RNIC adapter err %d", err);
825 		return err;
826 	}
827 
828 	return 0;
829 }
830 
831 int mana_ib_gd_add_gid(const struct ib_gid_attr *attr, void **context)
832 {
833 	struct mana_ib_dev *mdev = container_of(attr->device, struct mana_ib_dev, ib_dev);
834 	enum rdma_network_type ntype = rdma_gid_attr_network_type(attr);
835 	struct mana_rnic_config_addr_resp resp = {};
836 	struct gdma_context *gc = mdev_to_gc(mdev);
837 	struct mana_rnic_config_addr_req req = {};
838 	int err;
839 
840 	if (ntype != RDMA_NETWORK_IPV4 && ntype != RDMA_NETWORK_IPV6) {
841 		ibdev_dbg(&mdev->ib_dev, "Unsupported rdma network type %d", ntype);
842 		return -EINVAL;
843 	}
844 
845 	mana_gd_init_req_hdr(&req.hdr, MANA_IB_CONFIG_IP_ADDR, sizeof(req), sizeof(resp));
846 	req.hdr.dev_id = gc->mana_ib.dev_id;
847 	req.adapter = mdev->adapter_handle;
848 	req.op = ADDR_OP_ADD;
849 	req.sgid_type = (ntype == RDMA_NETWORK_IPV6) ? SGID_TYPE_IPV6 : SGID_TYPE_IPV4;
850 	copy_in_reverse(req.ip_addr, attr->gid.raw, sizeof(union ib_gid));
851 
852 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
853 	if (err) {
854 		ibdev_err(&mdev->ib_dev, "Failed to config IP addr err %d\n", err);
855 		return err;
856 	}
857 
858 	return 0;
859 }
860 
861 int mana_ib_gd_del_gid(const struct ib_gid_attr *attr, void **context)
862 {
863 	struct mana_ib_dev *mdev = container_of(attr->device, struct mana_ib_dev, ib_dev);
864 	enum rdma_network_type ntype = rdma_gid_attr_network_type(attr);
865 	struct mana_rnic_config_addr_resp resp = {};
866 	struct gdma_context *gc = mdev_to_gc(mdev);
867 	struct mana_rnic_config_addr_req req = {};
868 	int err;
869 
870 	if (ntype != RDMA_NETWORK_IPV4 && ntype != RDMA_NETWORK_IPV6) {
871 		ibdev_dbg(&mdev->ib_dev, "Unsupported rdma network type %d", ntype);
872 		return -EINVAL;
873 	}
874 
875 	mana_gd_init_req_hdr(&req.hdr, MANA_IB_CONFIG_IP_ADDR, sizeof(req), sizeof(resp));
876 	req.hdr.dev_id = gc->mana_ib.dev_id;
877 	req.adapter = mdev->adapter_handle;
878 	req.op = ADDR_OP_REMOVE;
879 	req.sgid_type = (ntype == RDMA_NETWORK_IPV6) ? SGID_TYPE_IPV6 : SGID_TYPE_IPV4;
880 	copy_in_reverse(req.ip_addr, attr->gid.raw, sizeof(union ib_gid));
881 
882 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
883 	if (err) {
884 		ibdev_err(&mdev->ib_dev, "Failed to config IP addr err %d\n", err);
885 		return err;
886 	}
887 
888 	return 0;
889 }
890 
891 int mana_ib_gd_config_mac(struct mana_ib_dev *mdev, enum mana_ib_addr_op op, u8 *mac)
892 {
893 	struct mana_rnic_config_mac_addr_resp resp = {};
894 	struct mana_rnic_config_mac_addr_req req = {};
895 	struct gdma_context *gc = mdev_to_gc(mdev);
896 	int err;
897 
898 	mana_gd_init_req_hdr(&req.hdr, MANA_IB_CONFIG_MAC_ADDR, sizeof(req), sizeof(resp));
899 	req.hdr.dev_id = gc->mana_ib.dev_id;
900 	req.adapter = mdev->adapter_handle;
901 	req.op = op;
902 	copy_in_reverse(req.mac_addr, mac, ETH_ALEN);
903 
904 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
905 	if (err) {
906 		ibdev_err(&mdev->ib_dev, "Failed to config Mac addr err %d", err);
907 		return err;
908 	}
909 
910 	return 0;
911 }
912 
913 int mana_ib_gd_create_cq(struct mana_ib_dev *mdev, struct mana_ib_cq *cq, u32 doorbell)
914 {
915 	struct gdma_context *gc = mdev_to_gc(mdev);
916 	struct mana_rnic_create_cq_resp resp = {};
917 	struct mana_rnic_create_cq_req req = {};
918 	int err;
919 
920 	mana_gd_init_req_hdr(&req.hdr, MANA_IB_CREATE_CQ, sizeof(req), sizeof(resp));
921 	req.hdr.dev_id = gc->mana_ib.dev_id;
922 	req.adapter = mdev->adapter_handle;
923 	req.gdma_region = cq->queue.gdma_region;
924 	req.eq_id = mdev->eqs[cq->comp_vector]->id;
925 	req.doorbell_page = doorbell;
926 
927 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
928 
929 	if (err) {
930 		ibdev_err(&mdev->ib_dev, "Failed to create cq err %d", err);
931 		return err;
932 	}
933 
934 	cq->queue.id  = resp.cq_id;
935 	cq->cq_handle = resp.cq_handle;
936 	/* The GDMA region is now owned by the CQ handle */
937 	cq->queue.gdma_region = GDMA_INVALID_DMA_REGION;
938 
939 	return 0;
940 }
941 
942 int mana_ib_gd_destroy_cq(struct mana_ib_dev *mdev, struct mana_ib_cq *cq)
943 {
944 	struct gdma_context *gc = mdev_to_gc(mdev);
945 	struct mana_rnic_destroy_cq_resp resp = {};
946 	struct mana_rnic_destroy_cq_req req = {};
947 	int err;
948 
949 	if (cq->cq_handle == INVALID_MANA_HANDLE)
950 		return 0;
951 
952 	mana_gd_init_req_hdr(&req.hdr, MANA_IB_DESTROY_CQ, sizeof(req), sizeof(resp));
953 	req.hdr.dev_id = gc->mana_ib.dev_id;
954 	req.adapter = mdev->adapter_handle;
955 	req.cq_handle = cq->cq_handle;
956 
957 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
958 
959 	if (err) {
960 		ibdev_err(&mdev->ib_dev, "Failed to destroy cq err %d", err);
961 		return err;
962 	}
963 
964 	return 0;
965 }
966 
967 int mana_ib_gd_create_rc_qp(struct mana_ib_dev *mdev, struct mana_ib_qp *qp,
968 			    struct ib_qp_init_attr *attr, u32 doorbell, u64 flags)
969 {
970 	struct mana_ib_cq *send_cq = container_of(qp->ibqp.send_cq, struct mana_ib_cq, ibcq);
971 	struct mana_ib_cq *recv_cq = container_of(qp->ibqp.recv_cq, struct mana_ib_cq, ibcq);
972 	struct mana_ib_pd *pd = container_of(qp->ibqp.pd, struct mana_ib_pd, ibpd);
973 	struct gdma_context *gc = mdev_to_gc(mdev);
974 	struct mana_rnic_create_qp_resp resp = {};
975 	struct mana_rnic_create_qp_req req = {};
976 	int err, i;
977 
978 	mana_gd_init_req_hdr(&req.hdr, MANA_IB_CREATE_RC_QP, sizeof(req), sizeof(resp));
979 	req.hdr.dev_id = gc->mana_ib.dev_id;
980 	req.adapter = mdev->adapter_handle;
981 	req.pd_handle = pd->pd_handle;
982 	req.send_cq_handle = send_cq->cq_handle;
983 	req.recv_cq_handle = recv_cq->cq_handle;
984 	for (i = 0; i < MANA_RC_QUEUE_TYPE_MAX; i++)
985 		req.dma_region[i] = qp->rc_qp.queues[i].gdma_region;
986 	req.doorbell_page = doorbell;
987 	req.max_send_wr = attr->cap.max_send_wr;
988 	req.max_recv_wr = attr->cap.max_recv_wr;
989 	req.max_send_sge = attr->cap.max_send_sge;
990 	req.max_recv_sge = attr->cap.max_recv_sge;
991 	req.flags = flags;
992 
993 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
994 	if (err) {
995 		ibdev_err(&mdev->ib_dev, "Failed to create rc qp err %d", err);
996 		return err;
997 	}
998 	qp->qp_handle = resp.rc_qp_handle;
999 	for (i = 0; i < MANA_RC_QUEUE_TYPE_MAX; i++) {
1000 		qp->rc_qp.queues[i].id = resp.queue_ids[i];
1001 		/* The GDMA regions are now owned by the RNIC QP handle */
1002 		qp->rc_qp.queues[i].gdma_region = GDMA_INVALID_DMA_REGION;
1003 	}
1004 	return 0;
1005 }
1006 
1007 int mana_ib_gd_destroy_rc_qp(struct mana_ib_dev *mdev, struct mana_ib_qp *qp)
1008 {
1009 	struct mana_rnic_destroy_rc_qp_resp resp = {0};
1010 	struct mana_rnic_destroy_rc_qp_req req = {0};
1011 	struct gdma_context *gc = mdev_to_gc(mdev);
1012 	int err;
1013 
1014 	mana_gd_init_req_hdr(&req.hdr, MANA_IB_DESTROY_RC_QP, sizeof(req), sizeof(resp));
1015 	req.hdr.dev_id = gc->mana_ib.dev_id;
1016 	req.adapter = mdev->adapter_handle;
1017 	req.rc_qp_handle = qp->qp_handle;
1018 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
1019 	if (err) {
1020 		ibdev_err(&mdev->ib_dev, "Failed to destroy rc qp err %d", err);
1021 		return err;
1022 	}
1023 	return 0;
1024 }
1025 
1026 int mana_ib_gd_create_ud_qp(struct mana_ib_dev *mdev, struct mana_ib_qp *qp,
1027 			    struct ib_qp_init_attr *attr, u32 doorbell, u32 type)
1028 {
1029 	struct mana_ib_cq *send_cq = container_of(qp->ibqp.send_cq, struct mana_ib_cq, ibcq);
1030 	struct mana_ib_cq *recv_cq = container_of(qp->ibqp.recv_cq, struct mana_ib_cq, ibcq);
1031 	struct mana_ib_pd *pd = container_of(qp->ibqp.pd, struct mana_ib_pd, ibpd);
1032 	struct gdma_context *gc = mdev_to_gc(mdev);
1033 	struct mana_rnic_create_udqp_resp resp = {};
1034 	struct mana_rnic_create_udqp_req req = {};
1035 	int err, i;
1036 
1037 	mana_gd_init_req_hdr(&req.hdr, MANA_IB_CREATE_UD_QP, sizeof(req), sizeof(resp));
1038 	req.hdr.dev_id = gc->mana_ib.dev_id;
1039 	req.adapter = mdev->adapter_handle;
1040 	req.pd_handle = pd->pd_handle;
1041 	req.send_cq_handle = send_cq->cq_handle;
1042 	req.recv_cq_handle = recv_cq->cq_handle;
1043 	for (i = 0; i < MANA_UD_QUEUE_TYPE_MAX; i++)
1044 		req.dma_region[i] = qp->ud_qp.queues[i].gdma_region;
1045 	req.doorbell_page = doorbell;
1046 	req.max_send_wr = attr->cap.max_send_wr;
1047 	req.max_recv_wr = attr->cap.max_recv_wr;
1048 	req.max_send_sge = attr->cap.max_send_sge;
1049 	req.max_recv_sge = attr->cap.max_recv_sge;
1050 	req.qp_type = type;
1051 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
1052 	if (err) {
1053 		ibdev_err(&mdev->ib_dev, "Failed to create ud qp err %d", err);
1054 		return err;
1055 	}
1056 	qp->qp_handle = resp.qp_handle;
1057 	for (i = 0; i < MANA_UD_QUEUE_TYPE_MAX; i++) {
1058 		qp->ud_qp.queues[i].id = resp.queue_ids[i];
1059 		/* The GDMA regions are now owned by the RNIC QP handle */
1060 		qp->ud_qp.queues[i].gdma_region = GDMA_INVALID_DMA_REGION;
1061 	}
1062 	return 0;
1063 }
1064 
1065 int mana_ib_gd_destroy_ud_qp(struct mana_ib_dev *mdev, struct mana_ib_qp *qp)
1066 {
1067 	struct mana_rnic_destroy_udqp_resp resp = {0};
1068 	struct mana_rnic_destroy_udqp_req req = {0};
1069 	struct gdma_context *gc = mdev_to_gc(mdev);
1070 	int err;
1071 
1072 	mana_gd_init_req_hdr(&req.hdr, MANA_IB_DESTROY_UD_QP, sizeof(req), sizeof(resp));
1073 	req.hdr.dev_id = gc->mana_ib.dev_id;
1074 	req.adapter = mdev->adapter_handle;
1075 	req.qp_handle = qp->qp_handle;
1076 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
1077 	if (err) {
1078 		ibdev_err(&mdev->ib_dev, "Failed to destroy ud qp err %d", err);
1079 		return err;
1080 	}
1081 	return 0;
1082 }
1083