xref: /linux/drivers/infiniband/hw/mana/main.c (revision be239684b18e1cdcafcf8c7face4a2f562c745ad)
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 gdma_dev *gd = &dev->gdma_dev->gdma_context->mana;
12 	struct mana_port_context *mpc;
13 	struct net_device *ndev;
14 	struct mana_context *mc;
15 
16 	mc = gd->driver_data;
17 	ndev = mc->ports[port];
18 	mpc = netdev_priv(ndev);
19 
20 	mutex_lock(&pd->vport_mutex);
21 
22 	pd->vport_use_count--;
23 	WARN_ON(pd->vport_use_count < 0);
24 
25 	if (!pd->vport_use_count)
26 		mana_uncfg_vport(mpc);
27 
28 	mutex_unlock(&pd->vport_mutex);
29 }
30 
31 int mana_ib_cfg_vport(struct mana_ib_dev *dev, u32 port, struct mana_ib_pd *pd,
32 		      u32 doorbell_id)
33 {
34 	struct gdma_dev *mdev = &dev->gdma_dev->gdma_context->mana;
35 	struct mana_port_context *mpc;
36 	struct mana_context *mc;
37 	struct net_device *ndev;
38 	int err;
39 
40 	mc = mdev->driver_data;
41 	ndev = mc->ports[port];
42 	mpc = netdev_priv(ndev);
43 
44 	mutex_lock(&pd->vport_mutex);
45 
46 	pd->vport_use_count++;
47 	if (pd->vport_use_count > 1) {
48 		ibdev_dbg(&dev->ib_dev,
49 			  "Skip as this PD is already configured vport\n");
50 		mutex_unlock(&pd->vport_mutex);
51 		return 0;
52 	}
53 
54 	err = mana_cfg_vport(mpc, pd->pdn, doorbell_id);
55 	if (err) {
56 		pd->vport_use_count--;
57 		mutex_unlock(&pd->vport_mutex);
58 
59 		ibdev_dbg(&dev->ib_dev, "Failed to configure vPort %d\n", err);
60 		return err;
61 	}
62 
63 	mutex_unlock(&pd->vport_mutex);
64 
65 	pd->tx_shortform_allowed = mpc->tx_shortform_allowed;
66 	pd->tx_vp_offset = mpc->tx_vp_offset;
67 
68 	ibdev_dbg(&dev->ib_dev, "vport handle %llx pdid %x doorbell_id %x\n",
69 		  mpc->port_handle, pd->pdn, doorbell_id);
70 
71 	return 0;
72 }
73 
74 int mana_ib_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
75 {
76 	struct mana_ib_pd *pd = container_of(ibpd, struct mana_ib_pd, ibpd);
77 	struct ib_device *ibdev = ibpd->device;
78 	struct gdma_create_pd_resp resp = {};
79 	struct gdma_create_pd_req req = {};
80 	enum gdma_pd_flags flags = 0;
81 	struct mana_ib_dev *dev;
82 	struct gdma_dev *mdev;
83 	int err;
84 
85 	dev = container_of(ibdev, struct mana_ib_dev, ib_dev);
86 	mdev = dev->gdma_dev;
87 
88 	mana_gd_init_req_hdr(&req.hdr, GDMA_CREATE_PD, sizeof(req),
89 			     sizeof(resp));
90 
91 	req.flags = flags;
92 	err = mana_gd_send_request(mdev->gdma_context, sizeof(req), &req,
93 				   sizeof(resp), &resp);
94 
95 	if (err || resp.hdr.status) {
96 		ibdev_dbg(&dev->ib_dev,
97 			  "Failed to get pd_id err %d status %u\n", err,
98 			  resp.hdr.status);
99 		if (!err)
100 			err = -EPROTO;
101 
102 		return err;
103 	}
104 
105 	pd->pd_handle = resp.pd_handle;
106 	pd->pdn = resp.pd_id;
107 	ibdev_dbg(&dev->ib_dev, "pd_handle 0x%llx pd_id %d\n",
108 		  pd->pd_handle, pd->pdn);
109 
110 	mutex_init(&pd->vport_mutex);
111 	pd->vport_use_count = 0;
112 	return 0;
113 }
114 
115 int mana_ib_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
116 {
117 	struct mana_ib_pd *pd = container_of(ibpd, struct mana_ib_pd, ibpd);
118 	struct ib_device *ibdev = ibpd->device;
119 	struct gdma_destory_pd_resp resp = {};
120 	struct gdma_destroy_pd_req req = {};
121 	struct mana_ib_dev *dev;
122 	struct gdma_dev *mdev;
123 	int err;
124 
125 	dev = container_of(ibdev, struct mana_ib_dev, ib_dev);
126 	mdev = dev->gdma_dev;
127 
128 	mana_gd_init_req_hdr(&req.hdr, GDMA_DESTROY_PD, sizeof(req),
129 			     sizeof(resp));
130 
131 	req.pd_handle = pd->pd_handle;
132 	err = mana_gd_send_request(mdev->gdma_context, sizeof(req), &req,
133 				   sizeof(resp), &resp);
134 
135 	if (err || resp.hdr.status) {
136 		ibdev_dbg(&dev->ib_dev,
137 			  "Failed to destroy pd_handle 0x%llx err %d status %u",
138 			  pd->pd_handle, err, resp.hdr.status);
139 		if (!err)
140 			err = -EPROTO;
141 	}
142 
143 	return err;
144 }
145 
146 static int mana_gd_destroy_doorbell_page(struct gdma_context *gc,
147 					 int doorbell_page)
148 {
149 	struct gdma_destroy_resource_range_req req = {};
150 	struct gdma_resp_hdr resp = {};
151 	int err;
152 
153 	mana_gd_init_req_hdr(&req.hdr, GDMA_DESTROY_RESOURCE_RANGE,
154 			     sizeof(req), sizeof(resp));
155 
156 	req.resource_type = GDMA_RESOURCE_DOORBELL_PAGE;
157 	req.num_resources = 1;
158 	req.allocated_resources = doorbell_page;
159 
160 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
161 	if (err || resp.status) {
162 		dev_err(gc->dev,
163 			"Failed to destroy doorbell page: ret %d, 0x%x\n",
164 			err, resp.status);
165 		return err ?: -EPROTO;
166 	}
167 
168 	return 0;
169 }
170 
171 static int mana_gd_allocate_doorbell_page(struct gdma_context *gc,
172 					  int *doorbell_page)
173 {
174 	struct gdma_allocate_resource_range_req req = {};
175 	struct gdma_allocate_resource_range_resp resp = {};
176 	int err;
177 
178 	mana_gd_init_req_hdr(&req.hdr, GDMA_ALLOCATE_RESOURCE_RANGE,
179 			     sizeof(req), sizeof(resp));
180 
181 	req.resource_type = GDMA_RESOURCE_DOORBELL_PAGE;
182 	req.num_resources = 1;
183 	req.alignment = 1;
184 
185 	/* Have GDMA start searching from 0 */
186 	req.allocated_resources = 0;
187 
188 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
189 	if (err || resp.hdr.status) {
190 		dev_err(gc->dev,
191 			"Failed to allocate doorbell page: ret %d, 0x%x\n",
192 			err, resp.hdr.status);
193 		return err ?: -EPROTO;
194 	}
195 
196 	*doorbell_page = resp.allocated_resources;
197 
198 	return 0;
199 }
200 
201 int mana_ib_alloc_ucontext(struct ib_ucontext *ibcontext,
202 			   struct ib_udata *udata)
203 {
204 	struct mana_ib_ucontext *ucontext =
205 		container_of(ibcontext, struct mana_ib_ucontext, ibucontext);
206 	struct ib_device *ibdev = ibcontext->device;
207 	struct mana_ib_dev *mdev;
208 	struct gdma_context *gc;
209 	struct gdma_dev *dev;
210 	int doorbell_page;
211 	int ret;
212 
213 	mdev = container_of(ibdev, struct mana_ib_dev, ib_dev);
214 	dev = mdev->gdma_dev;
215 	gc = dev->gdma_context;
216 
217 	/* Allocate a doorbell page index */
218 	ret = mana_gd_allocate_doorbell_page(gc, &doorbell_page);
219 	if (ret) {
220 		ibdev_dbg(ibdev, "Failed to allocate doorbell page %d\n", ret);
221 		return ret;
222 	}
223 
224 	ibdev_dbg(ibdev, "Doorbell page allocated %d\n", doorbell_page);
225 
226 	ucontext->doorbell = doorbell_page;
227 
228 	return 0;
229 }
230 
231 void mana_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
232 {
233 	struct mana_ib_ucontext *mana_ucontext =
234 		container_of(ibcontext, struct mana_ib_ucontext, ibucontext);
235 	struct ib_device *ibdev = ibcontext->device;
236 	struct mana_ib_dev *mdev;
237 	struct gdma_context *gc;
238 	int ret;
239 
240 	mdev = container_of(ibdev, struct mana_ib_dev, ib_dev);
241 	gc = mdev->gdma_dev->gdma_context;
242 
243 	ret = mana_gd_destroy_doorbell_page(gc, mana_ucontext->doorbell);
244 	if (ret)
245 		ibdev_dbg(ibdev, "Failed to destroy doorbell page %d\n", ret);
246 }
247 
248 static int
249 mana_ib_gd_first_dma_region(struct mana_ib_dev *dev,
250 			    struct gdma_context *gc,
251 			    struct gdma_create_dma_region_req *create_req,
252 			    size_t num_pages, mana_handle_t *gdma_region,
253 			    u32 expected_status)
254 {
255 	struct gdma_create_dma_region_resp create_resp = {};
256 	unsigned int create_req_msg_size;
257 	int err;
258 
259 	create_req_msg_size =
260 		struct_size(create_req, page_addr_list, num_pages);
261 	create_req->page_addr_list_len = num_pages;
262 
263 	err = mana_gd_send_request(gc, create_req_msg_size, create_req,
264 				   sizeof(create_resp), &create_resp);
265 	if (err || create_resp.hdr.status != expected_status) {
266 		ibdev_dbg(&dev->ib_dev,
267 			  "Failed to create DMA region: %d, 0x%x\n",
268 			  err, create_resp.hdr.status);
269 		if (!err)
270 			err = -EPROTO;
271 
272 		return err;
273 	}
274 
275 	*gdma_region = create_resp.dma_region_handle;
276 	ibdev_dbg(&dev->ib_dev, "Created DMA region handle 0x%llx\n",
277 		  *gdma_region);
278 
279 	return 0;
280 }
281 
282 static int
283 mana_ib_gd_add_dma_region(struct mana_ib_dev *dev, struct gdma_context *gc,
284 			  struct gdma_dma_region_add_pages_req *add_req,
285 			  unsigned int num_pages, u32 expected_status)
286 {
287 	unsigned int add_req_msg_size =
288 		struct_size(add_req, page_addr_list, num_pages);
289 	struct gdma_general_resp add_resp = {};
290 	int err;
291 
292 	mana_gd_init_req_hdr(&add_req->hdr, GDMA_DMA_REGION_ADD_PAGES,
293 			     add_req_msg_size, sizeof(add_resp));
294 	add_req->page_addr_list_len = num_pages;
295 
296 	err = mana_gd_send_request(gc, add_req_msg_size, add_req,
297 				   sizeof(add_resp), &add_resp);
298 	if (err || add_resp.hdr.status != expected_status) {
299 		ibdev_dbg(&dev->ib_dev,
300 			  "Failed to create DMA region: %d, 0x%x\n",
301 			  err, add_resp.hdr.status);
302 
303 		if (!err)
304 			err = -EPROTO;
305 
306 		return err;
307 	}
308 
309 	return 0;
310 }
311 
312 int mana_ib_gd_create_dma_region(struct mana_ib_dev *dev, struct ib_umem *umem,
313 				 mana_handle_t *gdma_region)
314 {
315 	struct gdma_dma_region_add_pages_req *add_req = NULL;
316 	size_t num_pages_processed = 0, num_pages_to_handle;
317 	struct gdma_create_dma_region_req *create_req;
318 	unsigned int create_req_msg_size;
319 	struct hw_channel_context *hwc;
320 	struct ib_block_iter biter;
321 	size_t max_pgs_add_cmd = 0;
322 	size_t max_pgs_create_cmd;
323 	struct gdma_context *gc;
324 	size_t num_pages_total;
325 	struct gdma_dev *mdev;
326 	unsigned long page_sz;
327 	unsigned int tail = 0;
328 	u64 *page_addr_list;
329 	void *request_buf;
330 	int err;
331 
332 	mdev = dev->gdma_dev;
333 	gc = mdev->gdma_context;
334 	hwc = gc->hwc.driver_data;
335 
336 	/* Hardware requires dma region to align to chosen page size */
337 	page_sz = ib_umem_find_best_pgsz(umem, PAGE_SZ_BM, 0);
338 	if (!page_sz) {
339 		ibdev_dbg(&dev->ib_dev, "failed to find page size.\n");
340 		return -ENOMEM;
341 	}
342 	num_pages_total = ib_umem_num_dma_blocks(umem, page_sz);
343 
344 	max_pgs_create_cmd =
345 		(hwc->max_req_msg_size - sizeof(*create_req)) / sizeof(u64);
346 	num_pages_to_handle =
347 		min_t(size_t, num_pages_total, max_pgs_create_cmd);
348 	create_req_msg_size =
349 		struct_size(create_req, page_addr_list, num_pages_to_handle);
350 
351 	request_buf = kzalloc(hwc->max_req_msg_size, GFP_KERNEL);
352 	if (!request_buf)
353 		return -ENOMEM;
354 
355 	create_req = request_buf;
356 	mana_gd_init_req_hdr(&create_req->hdr, GDMA_CREATE_DMA_REGION,
357 			     create_req_msg_size,
358 			     sizeof(struct gdma_create_dma_region_resp));
359 
360 	create_req->length = umem->length;
361 	create_req->offset_in_page = umem->address & (page_sz - 1);
362 	create_req->gdma_page_type = order_base_2(page_sz) - PAGE_SHIFT;
363 	create_req->page_count = num_pages_total;
364 
365 	ibdev_dbg(&dev->ib_dev, "size_dma_region %lu num_pages_total %lu\n",
366 		  umem->length, num_pages_total);
367 
368 	ibdev_dbg(&dev->ib_dev, "page_sz %lu offset_in_page %u\n",
369 		  page_sz, create_req->offset_in_page);
370 
371 	ibdev_dbg(&dev->ib_dev, "num_pages_to_handle %lu, gdma_page_type %u",
372 		  num_pages_to_handle, create_req->gdma_page_type);
373 
374 	page_addr_list = create_req->page_addr_list;
375 	rdma_umem_for_each_dma_block(umem, &biter, page_sz) {
376 		u32 expected_status = 0;
377 
378 		page_addr_list[tail++] = rdma_block_iter_dma_address(&biter);
379 		if (tail < num_pages_to_handle)
380 			continue;
381 
382 		if (num_pages_processed + num_pages_to_handle <
383 		    num_pages_total)
384 			expected_status = GDMA_STATUS_MORE_ENTRIES;
385 
386 		if (!num_pages_processed) {
387 			/* First create message */
388 			err = mana_ib_gd_first_dma_region(dev, gc, create_req,
389 							  tail, gdma_region,
390 							  expected_status);
391 			if (err)
392 				goto out;
393 
394 			max_pgs_add_cmd = (hwc->max_req_msg_size -
395 				sizeof(*add_req)) / sizeof(u64);
396 
397 			add_req = request_buf;
398 			add_req->dma_region_handle = *gdma_region;
399 			add_req->reserved3 = 0;
400 			page_addr_list = add_req->page_addr_list;
401 		} else {
402 			/* Subsequent create messages */
403 			err = mana_ib_gd_add_dma_region(dev, gc, add_req, tail,
404 							expected_status);
405 			if (err)
406 				break;
407 		}
408 
409 		num_pages_processed += tail;
410 		tail = 0;
411 
412 		/* The remaining pages to create */
413 		num_pages_to_handle =
414 			min_t(size_t,
415 			      num_pages_total - num_pages_processed,
416 			      max_pgs_add_cmd);
417 	}
418 
419 	if (err)
420 		mana_ib_gd_destroy_dma_region(dev, *gdma_region);
421 
422 out:
423 	kfree(request_buf);
424 	return err;
425 }
426 
427 int mana_ib_gd_destroy_dma_region(struct mana_ib_dev *dev, u64 gdma_region)
428 {
429 	struct gdma_dev *mdev = dev->gdma_dev;
430 	struct gdma_context *gc;
431 
432 	gc = mdev->gdma_context;
433 	ibdev_dbg(&dev->ib_dev, "destroy dma region 0x%llx\n", gdma_region);
434 
435 	return mana_gd_destroy_dma_region(gc, gdma_region);
436 }
437 
438 int mana_ib_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
439 {
440 	struct mana_ib_ucontext *mana_ucontext =
441 		container_of(ibcontext, struct mana_ib_ucontext, ibucontext);
442 	struct ib_device *ibdev = ibcontext->device;
443 	struct mana_ib_dev *mdev;
444 	struct gdma_context *gc;
445 	phys_addr_t pfn;
446 	pgprot_t prot;
447 	int ret;
448 
449 	mdev = container_of(ibdev, struct mana_ib_dev, ib_dev);
450 	gc = mdev->gdma_dev->gdma_context;
451 
452 	if (vma->vm_pgoff != 0) {
453 		ibdev_dbg(ibdev, "Unexpected vm_pgoff %lu\n", vma->vm_pgoff);
454 		return -EINVAL;
455 	}
456 
457 	/* Map to the page indexed by ucontext->doorbell */
458 	pfn = (gc->phys_db_page_base +
459 	       gc->db_page_size * mana_ucontext->doorbell) >>
460 	      PAGE_SHIFT;
461 	prot = pgprot_writecombine(vma->vm_page_prot);
462 
463 	ret = rdma_user_mmap_io(ibcontext, vma, pfn, gc->db_page_size, prot,
464 				NULL);
465 	if (ret)
466 		ibdev_dbg(ibdev, "can't rdma_user_mmap_io ret %d\n", ret);
467 	else
468 		ibdev_dbg(ibdev, "mapped I/O pfn 0x%llx page_size %u, ret %d\n",
469 			  pfn, gc->db_page_size, ret);
470 
471 	return ret;
472 }
473 
474 int mana_ib_get_port_immutable(struct ib_device *ibdev, u32 port_num,
475 			       struct ib_port_immutable *immutable)
476 {
477 	/*
478 	 * This version only support RAW_PACKET
479 	 * other values need to be filled for other types
480 	 */
481 	immutable->core_cap_flags = RDMA_CORE_PORT_RAW_PACKET;
482 
483 	return 0;
484 }
485 
486 int mana_ib_query_device(struct ib_device *ibdev, struct ib_device_attr *props,
487 			 struct ib_udata *uhw)
488 {
489 	struct mana_ib_dev *dev = container_of(ibdev,
490 			struct mana_ib_dev, ib_dev);
491 
492 	props->max_qp = dev->adapter_caps.max_qp_count;
493 	props->max_qp_wr = dev->adapter_caps.max_qp_wr;
494 	props->max_cq = dev->adapter_caps.max_cq_count;
495 	props->max_cqe = dev->adapter_caps.max_qp_wr;
496 	props->max_mr = dev->adapter_caps.max_mr_count;
497 	props->max_mr_size = MANA_IB_MAX_MR_SIZE;
498 	props->max_send_sge = dev->adapter_caps.max_send_sge_count;
499 	props->max_recv_sge = dev->adapter_caps.max_recv_sge_count;
500 
501 	return 0;
502 }
503 
504 int mana_ib_query_port(struct ib_device *ibdev, u32 port,
505 		       struct ib_port_attr *props)
506 {
507 	/* This version doesn't return port properties */
508 	return 0;
509 }
510 
511 int mana_ib_query_gid(struct ib_device *ibdev, u32 port, int index,
512 		      union ib_gid *gid)
513 {
514 	/* This version doesn't return GID properties */
515 	return 0;
516 }
517 
518 void mana_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
519 {
520 }
521 
522 int mana_ib_gd_query_adapter_caps(struct mana_ib_dev *dev)
523 {
524 	struct mana_ib_adapter_caps *caps = &dev->adapter_caps;
525 	struct mana_ib_query_adapter_caps_resp resp = {};
526 	struct mana_ib_query_adapter_caps_req req = {};
527 	int err;
528 
529 	mana_gd_init_req_hdr(&req.hdr, MANA_IB_GET_ADAPTER_CAP, sizeof(req),
530 			     sizeof(resp));
531 	req.hdr.resp.msg_version = GDMA_MESSAGE_V3;
532 	req.hdr.dev_id = dev->gdma_dev->dev_id;
533 
534 	err = mana_gd_send_request(dev->gdma_dev->gdma_context, sizeof(req),
535 				   &req, sizeof(resp), &resp);
536 
537 	if (err) {
538 		ibdev_err(&dev->ib_dev,
539 			  "Failed to query adapter caps err %d", err);
540 		return err;
541 	}
542 
543 	caps->max_sq_id = resp.max_sq_id;
544 	caps->max_rq_id = resp.max_rq_id;
545 	caps->max_cq_id = resp.max_cq_id;
546 	caps->max_qp_count = resp.max_qp_count;
547 	caps->max_cq_count = resp.max_cq_count;
548 	caps->max_mr_count = resp.max_mr_count;
549 	caps->max_pd_count = resp.max_pd_count;
550 	caps->max_inbound_read_limit = resp.max_inbound_read_limit;
551 	caps->max_outbound_read_limit = resp.max_outbound_read_limit;
552 	caps->mw_count = resp.mw_count;
553 	caps->max_srq_count = resp.max_srq_count;
554 	caps->max_qp_wr = min_t(u32,
555 				resp.max_requester_sq_size / GDMA_MAX_SQE_SIZE,
556 				resp.max_requester_rq_size / GDMA_MAX_RQE_SIZE);
557 	caps->max_inline_data_size = resp.max_inline_data_size;
558 	caps->max_send_sge_count = resp.max_send_sge_count;
559 	caps->max_recv_sge_count = resp.max_recv_sge_count;
560 
561 	return 0;
562 }
563