1 // SPDX-License-Identifier: GPL-2.0
2 
3 /* Authors: Cheng Xu <chengyou@linux.alibaba.com> */
4 /*          Kai Shen <kaishen@linux.alibaba.com> */
5 /* Copyright (c) 2020-2022, Alibaba Group. */
6 
7 /* Authors: Bernard Metzler <bmt@zurich.ibm.com> */
8 /* Copyright (c) 2008-2019, IBM Corporation */
9 
10 /* Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved. */
11 
12 #include <linux/vmalloc.h>
13 #include <net/addrconf.h>
14 #include <rdma/erdma-abi.h>
15 #include <rdma/ib_umem.h>
16 #include <rdma/uverbs_ioctl.h>
17 
18 #include "erdma.h"
19 #include "erdma_cm.h"
20 #include "erdma_verbs.h"
21 
22 static int create_qp_cmd(struct erdma_dev *dev, struct erdma_qp *qp)
23 {
24 	struct erdma_cmdq_create_qp_req req;
25 	struct erdma_pd *pd = to_epd(qp->ibqp.pd);
26 	struct erdma_uqp *user_qp;
27 	u64 resp0, resp1;
28 	int err;
29 
30 	erdma_cmdq_build_reqhdr(&req.hdr, CMDQ_SUBMOD_RDMA,
31 				CMDQ_OPCODE_CREATE_QP);
32 
33 	req.cfg0 = FIELD_PREP(ERDMA_CMD_CREATE_QP_SQ_DEPTH_MASK,
34 			      ilog2(qp->attrs.sq_size)) |
35 		   FIELD_PREP(ERDMA_CMD_CREATE_QP_QPN_MASK, QP_ID(qp));
36 	req.cfg1 = FIELD_PREP(ERDMA_CMD_CREATE_QP_RQ_DEPTH_MASK,
37 			      ilog2(qp->attrs.rq_size)) |
38 		   FIELD_PREP(ERDMA_CMD_CREATE_QP_PD_MASK, pd->pdn);
39 
40 	if (rdma_is_kernel_res(&qp->ibqp.res)) {
41 		u32 pgsz_range = ilog2(SZ_1M) - PAGE_SHIFT;
42 
43 		req.sq_cqn_mtt_cfg =
44 			FIELD_PREP(ERDMA_CMD_CREATE_QP_PAGE_SIZE_MASK,
45 				   pgsz_range) |
46 			FIELD_PREP(ERDMA_CMD_CREATE_QP_CQN_MASK, qp->scq->cqn);
47 		req.rq_cqn_mtt_cfg =
48 			FIELD_PREP(ERDMA_CMD_CREATE_QP_PAGE_SIZE_MASK,
49 				   pgsz_range) |
50 			FIELD_PREP(ERDMA_CMD_CREATE_QP_CQN_MASK, qp->rcq->cqn);
51 
52 		req.sq_mtt_cfg =
53 			FIELD_PREP(ERDMA_CMD_CREATE_QP_PAGE_OFFSET_MASK, 0) |
54 			FIELD_PREP(ERDMA_CMD_CREATE_QP_MTT_CNT_MASK, 1) |
55 			FIELD_PREP(ERDMA_CMD_CREATE_QP_MTT_TYPE_MASK,
56 				   ERDMA_MR_INLINE_MTT);
57 		req.rq_mtt_cfg = req.sq_mtt_cfg;
58 
59 		req.rq_buf_addr = qp->kern_qp.rq_buf_dma_addr;
60 		req.sq_buf_addr = qp->kern_qp.sq_buf_dma_addr;
61 		req.sq_db_info_dma_addr = qp->kern_qp.sq_buf_dma_addr +
62 					  (qp->attrs.sq_size << SQEBB_SHIFT);
63 		req.rq_db_info_dma_addr = qp->kern_qp.rq_buf_dma_addr +
64 					  (qp->attrs.rq_size << RQE_SHIFT);
65 	} else {
66 		user_qp = &qp->user_qp;
67 		req.sq_cqn_mtt_cfg = FIELD_PREP(
68 			ERDMA_CMD_CREATE_QP_PAGE_SIZE_MASK,
69 			ilog2(user_qp->sq_mtt.page_size) - PAGE_SHIFT);
70 		req.sq_cqn_mtt_cfg |=
71 			FIELD_PREP(ERDMA_CMD_CREATE_QP_CQN_MASK, qp->scq->cqn);
72 
73 		req.rq_cqn_mtt_cfg = FIELD_PREP(
74 			ERDMA_CMD_CREATE_QP_PAGE_SIZE_MASK,
75 			ilog2(user_qp->rq_mtt.page_size) - PAGE_SHIFT);
76 		req.rq_cqn_mtt_cfg |=
77 			FIELD_PREP(ERDMA_CMD_CREATE_QP_CQN_MASK, qp->rcq->cqn);
78 
79 		req.sq_mtt_cfg = user_qp->sq_mtt.page_offset;
80 		req.sq_mtt_cfg |= FIELD_PREP(ERDMA_CMD_CREATE_QP_MTT_CNT_MASK,
81 					     user_qp->sq_mtt.mtt_nents) |
82 				  FIELD_PREP(ERDMA_CMD_CREATE_QP_MTT_TYPE_MASK,
83 					     user_qp->sq_mtt.mtt_type);
84 
85 		req.rq_mtt_cfg = user_qp->rq_mtt.page_offset;
86 		req.rq_mtt_cfg |= FIELD_PREP(ERDMA_CMD_CREATE_QP_MTT_CNT_MASK,
87 					     user_qp->rq_mtt.mtt_nents) |
88 				  FIELD_PREP(ERDMA_CMD_CREATE_QP_MTT_TYPE_MASK,
89 					     user_qp->rq_mtt.mtt_type);
90 
91 		req.sq_buf_addr = user_qp->sq_mtt.mtt_entry[0];
92 		req.rq_buf_addr = user_qp->rq_mtt.mtt_entry[0];
93 
94 		req.sq_db_info_dma_addr = user_qp->sq_db_info_dma_addr;
95 		req.rq_db_info_dma_addr = user_qp->rq_db_info_dma_addr;
96 	}
97 
98 	err = erdma_post_cmd_wait(&dev->cmdq, &req, sizeof(req), &resp0,
99 				  &resp1);
100 	if (!err)
101 		qp->attrs.cookie =
102 			FIELD_GET(ERDMA_CMDQ_CREATE_QP_RESP_COOKIE_MASK, resp0);
103 
104 	return err;
105 }
106 
107 static int regmr_cmd(struct erdma_dev *dev, struct erdma_mr *mr)
108 {
109 	struct erdma_cmdq_reg_mr_req req;
110 	struct erdma_pd *pd = to_epd(mr->ibmr.pd);
111 	u64 *phy_addr;
112 	int i;
113 
114 	erdma_cmdq_build_reqhdr(&req.hdr, CMDQ_SUBMOD_RDMA, CMDQ_OPCODE_REG_MR);
115 
116 	req.cfg0 = FIELD_PREP(ERDMA_CMD_MR_VALID_MASK, mr->valid) |
117 		   FIELD_PREP(ERDMA_CMD_MR_KEY_MASK, mr->ibmr.lkey & 0xFF) |
118 		   FIELD_PREP(ERDMA_CMD_MR_MPT_IDX_MASK, mr->ibmr.lkey >> 8);
119 	req.cfg1 = FIELD_PREP(ERDMA_CMD_REGMR_PD_MASK, pd->pdn) |
120 		   FIELD_PREP(ERDMA_CMD_REGMR_TYPE_MASK, mr->type) |
121 		   FIELD_PREP(ERDMA_CMD_REGMR_RIGHT_MASK, mr->access) |
122 		   FIELD_PREP(ERDMA_CMD_REGMR_ACC_MODE_MASK, 0);
123 	req.cfg2 = FIELD_PREP(ERDMA_CMD_REGMR_PAGESIZE_MASK,
124 			      ilog2(mr->mem.page_size)) |
125 		   FIELD_PREP(ERDMA_CMD_REGMR_MTT_TYPE_MASK, mr->mem.mtt_type) |
126 		   FIELD_PREP(ERDMA_CMD_REGMR_MTT_CNT_MASK, mr->mem.page_cnt);
127 
128 	if (mr->type == ERDMA_MR_TYPE_DMA)
129 		goto post_cmd;
130 
131 	if (mr->type == ERDMA_MR_TYPE_NORMAL) {
132 		req.start_va = mr->mem.va;
133 		req.size = mr->mem.len;
134 	}
135 
136 	if (mr->type == ERDMA_MR_TYPE_FRMR ||
137 	    mr->mem.mtt_type == ERDMA_MR_INDIRECT_MTT) {
138 		phy_addr = req.phy_addr;
139 		*phy_addr = mr->mem.mtt_entry[0];
140 	} else {
141 		phy_addr = req.phy_addr;
142 		for (i = 0; i < mr->mem.mtt_nents; i++)
143 			*phy_addr++ = mr->mem.mtt_entry[i];
144 	}
145 
146 post_cmd:
147 	return erdma_post_cmd_wait(&dev->cmdq, &req, sizeof(req), NULL, NULL);
148 }
149 
150 static int create_cq_cmd(struct erdma_dev *dev, struct erdma_cq *cq)
151 {
152 	struct erdma_cmdq_create_cq_req req;
153 	u32 page_size;
154 	struct erdma_mem *mtt;
155 
156 	erdma_cmdq_build_reqhdr(&req.hdr, CMDQ_SUBMOD_RDMA,
157 				CMDQ_OPCODE_CREATE_CQ);
158 
159 	req.cfg0 = FIELD_PREP(ERDMA_CMD_CREATE_CQ_CQN_MASK, cq->cqn) |
160 		   FIELD_PREP(ERDMA_CMD_CREATE_CQ_DEPTH_MASK, ilog2(cq->depth));
161 	req.cfg1 = FIELD_PREP(ERDMA_CMD_CREATE_CQ_EQN_MASK, cq->assoc_eqn);
162 
163 	if (rdma_is_kernel_res(&cq->ibcq.res)) {
164 		page_size = SZ_32M;
165 		req.cfg0 |= FIELD_PREP(ERDMA_CMD_CREATE_CQ_PAGESIZE_MASK,
166 				       ilog2(page_size) - PAGE_SHIFT);
167 		req.qbuf_addr_l = lower_32_bits(cq->kern_cq.qbuf_dma_addr);
168 		req.qbuf_addr_h = upper_32_bits(cq->kern_cq.qbuf_dma_addr);
169 
170 		req.cfg1 |= FIELD_PREP(ERDMA_CMD_CREATE_CQ_MTT_CNT_MASK, 1) |
171 			    FIELD_PREP(ERDMA_CMD_CREATE_CQ_MTT_TYPE_MASK,
172 				       ERDMA_MR_INLINE_MTT);
173 
174 		req.first_page_offset = 0;
175 		req.cq_db_info_addr =
176 			cq->kern_cq.qbuf_dma_addr + (cq->depth << CQE_SHIFT);
177 	} else {
178 		mtt = &cq->user_cq.qbuf_mtt;
179 		req.cfg0 |= FIELD_PREP(ERDMA_CMD_CREATE_CQ_PAGESIZE_MASK,
180 				       ilog2(mtt->page_size) - PAGE_SHIFT);
181 		if (mtt->mtt_nents == 1) {
182 			req.qbuf_addr_l = lower_32_bits(*(u64 *)mtt->mtt_buf);
183 			req.qbuf_addr_h = upper_32_bits(*(u64 *)mtt->mtt_buf);
184 		} else {
185 			req.qbuf_addr_l = lower_32_bits(mtt->mtt_entry[0]);
186 			req.qbuf_addr_h = upper_32_bits(mtt->mtt_entry[0]);
187 		}
188 		req.cfg1 |= FIELD_PREP(ERDMA_CMD_CREATE_CQ_MTT_CNT_MASK,
189 				       mtt->mtt_nents);
190 		req.cfg1 |= FIELD_PREP(ERDMA_CMD_CREATE_CQ_MTT_TYPE_MASK,
191 				       mtt->mtt_type);
192 
193 		req.first_page_offset = mtt->page_offset;
194 		req.cq_db_info_addr = cq->user_cq.db_info_dma_addr;
195 	}
196 
197 	return erdma_post_cmd_wait(&dev->cmdq, &req, sizeof(req), NULL, NULL);
198 }
199 
200 static int erdma_alloc_idx(struct erdma_resource_cb *res_cb)
201 {
202 	int idx;
203 	unsigned long flags;
204 
205 	spin_lock_irqsave(&res_cb->lock, flags);
206 	idx = find_next_zero_bit(res_cb->bitmap, res_cb->max_cap,
207 				 res_cb->next_alloc_idx);
208 	if (idx == res_cb->max_cap) {
209 		idx = find_first_zero_bit(res_cb->bitmap, res_cb->max_cap);
210 		if (idx == res_cb->max_cap) {
211 			res_cb->next_alloc_idx = 1;
212 			spin_unlock_irqrestore(&res_cb->lock, flags);
213 			return -ENOSPC;
214 		}
215 	}
216 
217 	set_bit(idx, res_cb->bitmap);
218 	res_cb->next_alloc_idx = idx + 1;
219 	spin_unlock_irqrestore(&res_cb->lock, flags);
220 
221 	return idx;
222 }
223 
224 static inline void erdma_free_idx(struct erdma_resource_cb *res_cb, u32 idx)
225 {
226 	unsigned long flags;
227 	u32 used;
228 
229 	spin_lock_irqsave(&res_cb->lock, flags);
230 	used = __test_and_clear_bit(idx, res_cb->bitmap);
231 	spin_unlock_irqrestore(&res_cb->lock, flags);
232 	WARN_ON(!used);
233 }
234 
235 static struct rdma_user_mmap_entry *
236 erdma_user_mmap_entry_insert(struct erdma_ucontext *uctx, void *address,
237 			     u32 size, u8 mmap_flag, u64 *mmap_offset)
238 {
239 	struct erdma_user_mmap_entry *entry =
240 		kzalloc(sizeof(*entry), GFP_KERNEL);
241 	int ret;
242 
243 	if (!entry)
244 		return NULL;
245 
246 	entry->address = (u64)address;
247 	entry->mmap_flag = mmap_flag;
248 
249 	size = PAGE_ALIGN(size);
250 
251 	ret = rdma_user_mmap_entry_insert(&uctx->ibucontext, &entry->rdma_entry,
252 					  size);
253 	if (ret) {
254 		kfree(entry);
255 		return NULL;
256 	}
257 
258 	*mmap_offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
259 
260 	return &entry->rdma_entry;
261 }
262 
263 int erdma_query_device(struct ib_device *ibdev, struct ib_device_attr *attr,
264 		       struct ib_udata *unused)
265 {
266 	struct erdma_dev *dev = to_edev(ibdev);
267 
268 	memset(attr, 0, sizeof(*attr));
269 
270 	attr->max_mr_size = dev->attrs.max_mr_size;
271 	attr->vendor_id = PCI_VENDOR_ID_ALIBABA;
272 	attr->vendor_part_id = dev->pdev->device;
273 	attr->hw_ver = dev->pdev->revision;
274 	attr->max_qp = dev->attrs.max_qp - 1;
275 	attr->max_qp_wr = min(dev->attrs.max_send_wr, dev->attrs.max_recv_wr);
276 	attr->max_qp_rd_atom = dev->attrs.max_ord;
277 	attr->max_qp_init_rd_atom = dev->attrs.max_ird;
278 	attr->max_res_rd_atom = dev->attrs.max_qp * dev->attrs.max_ird;
279 	attr->device_cap_flags = IB_DEVICE_MEM_MGT_EXTENSIONS;
280 	attr->kernel_cap_flags = IBK_LOCAL_DMA_LKEY;
281 	ibdev->local_dma_lkey = dev->attrs.local_dma_key;
282 	attr->max_send_sge = dev->attrs.max_send_sge;
283 	attr->max_recv_sge = dev->attrs.max_recv_sge;
284 	attr->max_sge_rd = dev->attrs.max_sge_rd;
285 	attr->max_cq = dev->attrs.max_cq - 1;
286 	attr->max_cqe = dev->attrs.max_cqe;
287 	attr->max_mr = dev->attrs.max_mr;
288 	attr->max_pd = dev->attrs.max_pd;
289 	attr->max_mw = dev->attrs.max_mw;
290 	attr->max_fast_reg_page_list_len = ERDMA_MAX_FRMR_PA;
291 	attr->page_size_cap = ERDMA_PAGE_SIZE_SUPPORT;
292 	attr->fw_ver = dev->attrs.fw_version;
293 
294 	if (dev->netdev)
295 		addrconf_addr_eui48((u8 *)&attr->sys_image_guid,
296 				    dev->netdev->dev_addr);
297 
298 	return 0;
299 }
300 
301 int erdma_query_gid(struct ib_device *ibdev, u32 port, int idx,
302 		    union ib_gid *gid)
303 {
304 	struct erdma_dev *dev = to_edev(ibdev);
305 
306 	memset(gid, 0, sizeof(*gid));
307 	ether_addr_copy(gid->raw, dev->attrs.peer_addr);
308 
309 	return 0;
310 }
311 
312 int erdma_query_port(struct ib_device *ibdev, u32 port,
313 		     struct ib_port_attr *attr)
314 {
315 	struct erdma_dev *dev = to_edev(ibdev);
316 	struct net_device *ndev = dev->netdev;
317 
318 	memset(attr, 0, sizeof(*attr));
319 
320 	attr->gid_tbl_len = 1;
321 	attr->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_DEVICE_MGMT_SUP;
322 	attr->max_msg_sz = -1;
323 
324 	if (!ndev)
325 		goto out;
326 
327 	ib_get_eth_speed(ibdev, port, &attr->active_speed, &attr->active_width);
328 	attr->max_mtu = ib_mtu_int_to_enum(ndev->mtu);
329 	attr->active_mtu = ib_mtu_int_to_enum(ndev->mtu);
330 	if (netif_running(ndev) && netif_carrier_ok(ndev))
331 		dev->state = IB_PORT_ACTIVE;
332 	else
333 		dev->state = IB_PORT_DOWN;
334 	attr->state = dev->state;
335 
336 out:
337 	if (dev->state == IB_PORT_ACTIVE)
338 		attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
339 	else
340 		attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
341 
342 	return 0;
343 }
344 
345 int erdma_get_port_immutable(struct ib_device *ibdev, u32 port,
346 			     struct ib_port_immutable *port_immutable)
347 {
348 	port_immutable->gid_tbl_len = 1;
349 	port_immutable->core_cap_flags = RDMA_CORE_PORT_IWARP;
350 
351 	return 0;
352 }
353 
354 int erdma_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
355 {
356 	struct erdma_pd *pd = to_epd(ibpd);
357 	struct erdma_dev *dev = to_edev(ibpd->device);
358 	int pdn;
359 
360 	pdn = erdma_alloc_idx(&dev->res_cb[ERDMA_RES_TYPE_PD]);
361 	if (pdn < 0)
362 		return pdn;
363 
364 	pd->pdn = pdn;
365 
366 	return 0;
367 }
368 
369 int erdma_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
370 {
371 	struct erdma_pd *pd = to_epd(ibpd);
372 	struct erdma_dev *dev = to_edev(ibpd->device);
373 
374 	erdma_free_idx(&dev->res_cb[ERDMA_RES_TYPE_PD], pd->pdn);
375 
376 	return 0;
377 }
378 
379 static int erdma_qp_validate_cap(struct erdma_dev *dev,
380 				 struct ib_qp_init_attr *attrs)
381 {
382 	if ((attrs->cap.max_send_wr > dev->attrs.max_send_wr) ||
383 	    (attrs->cap.max_recv_wr > dev->attrs.max_recv_wr) ||
384 	    (attrs->cap.max_send_sge > dev->attrs.max_send_sge) ||
385 	    (attrs->cap.max_recv_sge > dev->attrs.max_recv_sge) ||
386 	    (attrs->cap.max_inline_data > ERDMA_MAX_INLINE) ||
387 	    !attrs->cap.max_send_wr || !attrs->cap.max_recv_wr) {
388 		return -EINVAL;
389 	}
390 
391 	return 0;
392 }
393 
394 static int erdma_qp_validate_attr(struct erdma_dev *dev,
395 				  struct ib_qp_init_attr *attrs)
396 {
397 	if (attrs->qp_type != IB_QPT_RC)
398 		return -EOPNOTSUPP;
399 
400 	if (attrs->srq)
401 		return -EOPNOTSUPP;
402 
403 	if (!attrs->send_cq || !attrs->recv_cq)
404 		return -EOPNOTSUPP;
405 
406 	return 0;
407 }
408 
409 static void free_kernel_qp(struct erdma_qp *qp)
410 {
411 	struct erdma_dev *dev = qp->dev;
412 
413 	vfree(qp->kern_qp.swr_tbl);
414 	vfree(qp->kern_qp.rwr_tbl);
415 
416 	if (qp->kern_qp.sq_buf)
417 		dma_free_coherent(
418 			&dev->pdev->dev,
419 			WARPPED_BUFSIZE(qp->attrs.sq_size << SQEBB_SHIFT),
420 			qp->kern_qp.sq_buf, qp->kern_qp.sq_buf_dma_addr);
421 
422 	if (qp->kern_qp.rq_buf)
423 		dma_free_coherent(
424 			&dev->pdev->dev,
425 			WARPPED_BUFSIZE(qp->attrs.rq_size << RQE_SHIFT),
426 			qp->kern_qp.rq_buf, qp->kern_qp.rq_buf_dma_addr);
427 }
428 
429 static int init_kernel_qp(struct erdma_dev *dev, struct erdma_qp *qp,
430 			  struct ib_qp_init_attr *attrs)
431 {
432 	struct erdma_kqp *kqp = &qp->kern_qp;
433 	int size;
434 
435 	if (attrs->sq_sig_type == IB_SIGNAL_ALL_WR)
436 		kqp->sig_all = 1;
437 
438 	kqp->sq_pi = 0;
439 	kqp->sq_ci = 0;
440 	kqp->rq_pi = 0;
441 	kqp->rq_ci = 0;
442 	kqp->hw_sq_db =
443 		dev->func_bar + (ERDMA_SDB_SHARED_PAGE_INDEX << PAGE_SHIFT);
444 	kqp->hw_rq_db = dev->func_bar + ERDMA_BAR_RQDB_SPACE_OFFSET;
445 
446 	kqp->swr_tbl = vmalloc(qp->attrs.sq_size * sizeof(u64));
447 	kqp->rwr_tbl = vmalloc(qp->attrs.rq_size * sizeof(u64));
448 	if (!kqp->swr_tbl || !kqp->rwr_tbl)
449 		goto err_out;
450 
451 	size = (qp->attrs.sq_size << SQEBB_SHIFT) + ERDMA_EXTRA_BUFFER_SIZE;
452 	kqp->sq_buf = dma_alloc_coherent(&dev->pdev->dev, size,
453 					 &kqp->sq_buf_dma_addr, GFP_KERNEL);
454 	if (!kqp->sq_buf)
455 		goto err_out;
456 
457 	size = (qp->attrs.rq_size << RQE_SHIFT) + ERDMA_EXTRA_BUFFER_SIZE;
458 	kqp->rq_buf = dma_alloc_coherent(&dev->pdev->dev, size,
459 					 &kqp->rq_buf_dma_addr, GFP_KERNEL);
460 	if (!kqp->rq_buf)
461 		goto err_out;
462 
463 	kqp->sq_db_info = kqp->sq_buf + (qp->attrs.sq_size << SQEBB_SHIFT);
464 	kqp->rq_db_info = kqp->rq_buf + (qp->attrs.rq_size << RQE_SHIFT);
465 
466 	return 0;
467 
468 err_out:
469 	free_kernel_qp(qp);
470 	return -ENOMEM;
471 }
472 
473 static int get_mtt_entries(struct erdma_dev *dev, struct erdma_mem *mem,
474 			   u64 start, u64 len, int access, u64 virt,
475 			   unsigned long req_page_size, u8 force_indirect_mtt)
476 {
477 	struct ib_block_iter biter;
478 	uint64_t *phy_addr = NULL;
479 	int ret = 0;
480 
481 	mem->umem = ib_umem_get(&dev->ibdev, start, len, access);
482 	if (IS_ERR(mem->umem)) {
483 		ret = PTR_ERR(mem->umem);
484 		mem->umem = NULL;
485 		return ret;
486 	}
487 
488 	mem->va = virt;
489 	mem->len = len;
490 	mem->page_size = ib_umem_find_best_pgsz(mem->umem, req_page_size, virt);
491 	mem->page_offset = start & (mem->page_size - 1);
492 	mem->mtt_nents = ib_umem_num_dma_blocks(mem->umem, mem->page_size);
493 	mem->page_cnt = mem->mtt_nents;
494 
495 	if (mem->page_cnt > ERDMA_MAX_INLINE_MTT_ENTRIES ||
496 	    force_indirect_mtt) {
497 		mem->mtt_type = ERDMA_MR_INDIRECT_MTT;
498 		mem->mtt_buf =
499 			alloc_pages_exact(MTT_SIZE(mem->page_cnt), GFP_KERNEL);
500 		if (!mem->mtt_buf) {
501 			ret = -ENOMEM;
502 			goto error_ret;
503 		}
504 		phy_addr = mem->mtt_buf;
505 	} else {
506 		mem->mtt_type = ERDMA_MR_INLINE_MTT;
507 		phy_addr = mem->mtt_entry;
508 	}
509 
510 	rdma_umem_for_each_dma_block(mem->umem, &biter, mem->page_size) {
511 		*phy_addr = rdma_block_iter_dma_address(&biter);
512 		phy_addr++;
513 	}
514 
515 	if (mem->mtt_type == ERDMA_MR_INDIRECT_MTT) {
516 		mem->mtt_entry[0] =
517 			dma_map_single(&dev->pdev->dev, mem->mtt_buf,
518 				       MTT_SIZE(mem->page_cnt), DMA_TO_DEVICE);
519 		if (dma_mapping_error(&dev->pdev->dev, mem->mtt_entry[0])) {
520 			free_pages_exact(mem->mtt_buf, MTT_SIZE(mem->page_cnt));
521 			mem->mtt_buf = NULL;
522 			ret = -ENOMEM;
523 			goto error_ret;
524 		}
525 	}
526 
527 	return 0;
528 
529 error_ret:
530 	if (mem->umem) {
531 		ib_umem_release(mem->umem);
532 		mem->umem = NULL;
533 	}
534 
535 	return ret;
536 }
537 
538 static void put_mtt_entries(struct erdma_dev *dev, struct erdma_mem *mem)
539 {
540 	if (mem->mtt_buf) {
541 		dma_unmap_single(&dev->pdev->dev, mem->mtt_entry[0],
542 				 MTT_SIZE(mem->page_cnt), DMA_TO_DEVICE);
543 		free_pages_exact(mem->mtt_buf, MTT_SIZE(mem->page_cnt));
544 	}
545 
546 	if (mem->umem) {
547 		ib_umem_release(mem->umem);
548 		mem->umem = NULL;
549 	}
550 }
551 
552 static int erdma_map_user_dbrecords(struct erdma_ucontext *ctx,
553 				    u64 dbrecords_va,
554 				    struct erdma_user_dbrecords_page **dbr_page,
555 				    dma_addr_t *dma_addr)
556 {
557 	struct erdma_user_dbrecords_page *page = NULL;
558 	int rv = 0;
559 
560 	mutex_lock(&ctx->dbrecords_page_mutex);
561 
562 	list_for_each_entry(page, &ctx->dbrecords_page_list, list)
563 		if (page->va == (dbrecords_va & PAGE_MASK))
564 			goto found;
565 
566 	page = kmalloc(sizeof(*page), GFP_KERNEL);
567 	if (!page) {
568 		rv = -ENOMEM;
569 		goto out;
570 	}
571 
572 	page->va = (dbrecords_va & PAGE_MASK);
573 	page->refcnt = 0;
574 
575 	page->umem = ib_umem_get(ctx->ibucontext.device,
576 				 dbrecords_va & PAGE_MASK, PAGE_SIZE, 0);
577 	if (IS_ERR(page->umem)) {
578 		rv = PTR_ERR(page->umem);
579 		kfree(page);
580 		goto out;
581 	}
582 
583 	list_add(&page->list, &ctx->dbrecords_page_list);
584 
585 found:
586 	*dma_addr = sg_dma_address(page->umem->sgt_append.sgt.sgl) +
587 		    (dbrecords_va & ~PAGE_MASK);
588 	*dbr_page = page;
589 	page->refcnt++;
590 
591 out:
592 	mutex_unlock(&ctx->dbrecords_page_mutex);
593 	return rv;
594 }
595 
596 static void
597 erdma_unmap_user_dbrecords(struct erdma_ucontext *ctx,
598 			   struct erdma_user_dbrecords_page **dbr_page)
599 {
600 	if (!ctx || !(*dbr_page))
601 		return;
602 
603 	mutex_lock(&ctx->dbrecords_page_mutex);
604 	if (--(*dbr_page)->refcnt == 0) {
605 		list_del(&(*dbr_page)->list);
606 		ib_umem_release((*dbr_page)->umem);
607 		kfree(*dbr_page);
608 	}
609 
610 	*dbr_page = NULL;
611 	mutex_unlock(&ctx->dbrecords_page_mutex);
612 }
613 
614 static int init_user_qp(struct erdma_qp *qp, struct erdma_ucontext *uctx,
615 			u64 va, u32 len, u64 db_info_va)
616 {
617 	dma_addr_t db_info_dma_addr;
618 	u32 rq_offset;
619 	int ret;
620 
621 	if (len < (PAGE_ALIGN(qp->attrs.sq_size * SQEBB_SIZE) +
622 		   qp->attrs.rq_size * RQE_SIZE))
623 		return -EINVAL;
624 
625 	ret = get_mtt_entries(qp->dev, &qp->user_qp.sq_mtt, va,
626 			      qp->attrs.sq_size << SQEBB_SHIFT, 0, va,
627 			      (SZ_1M - SZ_4K), 1);
628 	if (ret)
629 		return ret;
630 
631 	rq_offset = PAGE_ALIGN(qp->attrs.sq_size << SQEBB_SHIFT);
632 	qp->user_qp.rq_offset = rq_offset;
633 
634 	ret = get_mtt_entries(qp->dev, &qp->user_qp.rq_mtt, va + rq_offset,
635 			      qp->attrs.rq_size << RQE_SHIFT, 0, va + rq_offset,
636 			      (SZ_1M - SZ_4K), 1);
637 	if (ret)
638 		goto put_sq_mtt;
639 
640 	ret = erdma_map_user_dbrecords(uctx, db_info_va,
641 				       &qp->user_qp.user_dbr_page,
642 				       &db_info_dma_addr);
643 	if (ret)
644 		goto put_rq_mtt;
645 
646 	qp->user_qp.sq_db_info_dma_addr = db_info_dma_addr;
647 	qp->user_qp.rq_db_info_dma_addr = db_info_dma_addr + ERDMA_DB_SIZE;
648 
649 	return 0;
650 
651 put_rq_mtt:
652 	put_mtt_entries(qp->dev, &qp->user_qp.rq_mtt);
653 
654 put_sq_mtt:
655 	put_mtt_entries(qp->dev, &qp->user_qp.sq_mtt);
656 
657 	return ret;
658 }
659 
660 static void free_user_qp(struct erdma_qp *qp, struct erdma_ucontext *uctx)
661 {
662 	put_mtt_entries(qp->dev, &qp->user_qp.sq_mtt);
663 	put_mtt_entries(qp->dev, &qp->user_qp.rq_mtt);
664 	erdma_unmap_user_dbrecords(uctx, &qp->user_qp.user_dbr_page);
665 }
666 
667 int erdma_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *attrs,
668 		    struct ib_udata *udata)
669 {
670 	struct erdma_qp *qp = to_eqp(ibqp);
671 	struct erdma_dev *dev = to_edev(ibqp->device);
672 	struct erdma_ucontext *uctx = rdma_udata_to_drv_context(
673 		udata, struct erdma_ucontext, ibucontext);
674 	struct erdma_ureq_create_qp ureq;
675 	struct erdma_uresp_create_qp uresp;
676 	int ret;
677 
678 	ret = erdma_qp_validate_cap(dev, attrs);
679 	if (ret)
680 		goto err_out;
681 
682 	ret = erdma_qp_validate_attr(dev, attrs);
683 	if (ret)
684 		goto err_out;
685 
686 	qp->scq = to_ecq(attrs->send_cq);
687 	qp->rcq = to_ecq(attrs->recv_cq);
688 	qp->dev = dev;
689 	qp->attrs.cc = dev->attrs.cc;
690 
691 	init_rwsem(&qp->state_lock);
692 	kref_init(&qp->ref);
693 	init_completion(&qp->safe_free);
694 
695 	ret = xa_alloc_cyclic(&dev->qp_xa, &qp->ibqp.qp_num, qp,
696 			      XA_LIMIT(1, dev->attrs.max_qp - 1),
697 			      &dev->next_alloc_qpn, GFP_KERNEL);
698 	if (ret < 0) {
699 		ret = -ENOMEM;
700 		goto err_out;
701 	}
702 
703 	qp->attrs.sq_size = roundup_pow_of_two(attrs->cap.max_send_wr *
704 					       ERDMA_MAX_WQEBB_PER_SQE);
705 	qp->attrs.rq_size = roundup_pow_of_two(attrs->cap.max_recv_wr);
706 
707 	if (uctx) {
708 		ret = ib_copy_from_udata(&ureq, udata,
709 					 min(sizeof(ureq), udata->inlen));
710 		if (ret)
711 			goto err_out_xa;
712 
713 		ret = init_user_qp(qp, uctx, ureq.qbuf_va, ureq.qbuf_len,
714 				   ureq.db_record_va);
715 		if (ret)
716 			goto err_out_xa;
717 
718 		memset(&uresp, 0, sizeof(uresp));
719 
720 		uresp.num_sqe = qp->attrs.sq_size;
721 		uresp.num_rqe = qp->attrs.rq_size;
722 		uresp.qp_id = QP_ID(qp);
723 		uresp.rq_offset = qp->user_qp.rq_offset;
724 
725 		ret = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
726 		if (ret)
727 			goto err_out_cmd;
728 	} else {
729 		init_kernel_qp(dev, qp, attrs);
730 	}
731 
732 	qp->attrs.max_send_sge = attrs->cap.max_send_sge;
733 	qp->attrs.max_recv_sge = attrs->cap.max_recv_sge;
734 	qp->attrs.state = ERDMA_QP_STATE_IDLE;
735 
736 	ret = create_qp_cmd(dev, qp);
737 	if (ret)
738 		goto err_out_cmd;
739 
740 	spin_lock_init(&qp->lock);
741 
742 	return 0;
743 
744 err_out_cmd:
745 	if (uctx)
746 		free_user_qp(qp, uctx);
747 	else
748 		free_kernel_qp(qp);
749 err_out_xa:
750 	xa_erase(&dev->qp_xa, QP_ID(qp));
751 err_out:
752 	return ret;
753 }
754 
755 static int erdma_create_stag(struct erdma_dev *dev, u32 *stag)
756 {
757 	int stag_idx;
758 
759 	stag_idx = erdma_alloc_idx(&dev->res_cb[ERDMA_RES_TYPE_STAG_IDX]);
760 	if (stag_idx < 0)
761 		return stag_idx;
762 
763 	/* For now, we always let key field be zero. */
764 	*stag = (stag_idx << 8);
765 
766 	return 0;
767 }
768 
769 struct ib_mr *erdma_get_dma_mr(struct ib_pd *ibpd, int acc)
770 {
771 	struct erdma_dev *dev = to_edev(ibpd->device);
772 	struct erdma_mr *mr;
773 	u32 stag;
774 	int ret;
775 
776 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
777 	if (!mr)
778 		return ERR_PTR(-ENOMEM);
779 
780 	ret = erdma_create_stag(dev, &stag);
781 	if (ret)
782 		goto out_free;
783 
784 	mr->type = ERDMA_MR_TYPE_DMA;
785 
786 	mr->ibmr.lkey = stag;
787 	mr->ibmr.rkey = stag;
788 	mr->ibmr.pd = ibpd;
789 	mr->access = ERDMA_MR_ACC_LR | to_erdma_access_flags(acc);
790 	ret = regmr_cmd(dev, mr);
791 	if (ret)
792 		goto out_remove_stag;
793 
794 	return &mr->ibmr;
795 
796 out_remove_stag:
797 	erdma_free_idx(&dev->res_cb[ERDMA_RES_TYPE_STAG_IDX],
798 		       mr->ibmr.lkey >> 8);
799 
800 out_free:
801 	kfree(mr);
802 
803 	return ERR_PTR(ret);
804 }
805 
806 struct ib_mr *erdma_ib_alloc_mr(struct ib_pd *ibpd, enum ib_mr_type mr_type,
807 				u32 max_num_sg)
808 {
809 	struct erdma_mr *mr;
810 	struct erdma_dev *dev = to_edev(ibpd->device);
811 	int ret;
812 	u32 stag;
813 
814 	if (mr_type != IB_MR_TYPE_MEM_REG)
815 		return ERR_PTR(-EOPNOTSUPP);
816 
817 	if (max_num_sg > ERDMA_MR_MAX_MTT_CNT)
818 		return ERR_PTR(-EINVAL);
819 
820 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
821 	if (!mr)
822 		return ERR_PTR(-ENOMEM);
823 
824 	ret = erdma_create_stag(dev, &stag);
825 	if (ret)
826 		goto out_free;
827 
828 	mr->type = ERDMA_MR_TYPE_FRMR;
829 
830 	mr->ibmr.lkey = stag;
831 	mr->ibmr.rkey = stag;
832 	mr->ibmr.pd = ibpd;
833 	/* update it in FRMR. */
834 	mr->access = ERDMA_MR_ACC_LR | ERDMA_MR_ACC_LW | ERDMA_MR_ACC_RR |
835 		     ERDMA_MR_ACC_RW;
836 
837 	mr->mem.page_size = PAGE_SIZE; /* update it later. */
838 	mr->mem.page_cnt = max_num_sg;
839 	mr->mem.mtt_type = ERDMA_MR_INDIRECT_MTT;
840 	mr->mem.mtt_buf =
841 		alloc_pages_exact(MTT_SIZE(mr->mem.page_cnt), GFP_KERNEL);
842 	if (!mr->mem.mtt_buf) {
843 		ret = -ENOMEM;
844 		goto out_remove_stag;
845 	}
846 
847 	mr->mem.mtt_entry[0] =
848 		dma_map_single(&dev->pdev->dev, mr->mem.mtt_buf,
849 			       MTT_SIZE(mr->mem.page_cnt), DMA_TO_DEVICE);
850 	if (dma_mapping_error(&dev->pdev->dev, mr->mem.mtt_entry[0])) {
851 		ret = -ENOMEM;
852 		goto out_free_mtt;
853 	}
854 
855 	ret = regmr_cmd(dev, mr);
856 	if (ret)
857 		goto out_dma_unmap;
858 
859 	return &mr->ibmr;
860 
861 out_dma_unmap:
862 	dma_unmap_single(&dev->pdev->dev, mr->mem.mtt_entry[0],
863 			 MTT_SIZE(mr->mem.page_cnt), DMA_TO_DEVICE);
864 out_free_mtt:
865 	free_pages_exact(mr->mem.mtt_buf, MTT_SIZE(mr->mem.page_cnt));
866 
867 out_remove_stag:
868 	erdma_free_idx(&dev->res_cb[ERDMA_RES_TYPE_STAG_IDX],
869 		       mr->ibmr.lkey >> 8);
870 
871 out_free:
872 	kfree(mr);
873 
874 	return ERR_PTR(ret);
875 }
876 
877 static int erdma_set_page(struct ib_mr *ibmr, u64 addr)
878 {
879 	struct erdma_mr *mr = to_emr(ibmr);
880 
881 	if (mr->mem.mtt_nents >= mr->mem.page_cnt)
882 		return -1;
883 
884 	*((u64 *)mr->mem.mtt_buf + mr->mem.mtt_nents) = addr;
885 	mr->mem.mtt_nents++;
886 
887 	return 0;
888 }
889 
890 int erdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
891 		    unsigned int *sg_offset)
892 {
893 	struct erdma_mr *mr = to_emr(ibmr);
894 	int num;
895 
896 	mr->mem.mtt_nents = 0;
897 
898 	num = ib_sg_to_pages(&mr->ibmr, sg, sg_nents, sg_offset,
899 			     erdma_set_page);
900 
901 	return num;
902 }
903 
904 struct ib_mr *erdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
905 				u64 virt, int access, struct ib_udata *udata)
906 {
907 	struct erdma_mr *mr = NULL;
908 	struct erdma_dev *dev = to_edev(ibpd->device);
909 	u32 stag;
910 	int ret;
911 
912 	if (!len || len > dev->attrs.max_mr_size)
913 		return ERR_PTR(-EINVAL);
914 
915 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
916 	if (!mr)
917 		return ERR_PTR(-ENOMEM);
918 
919 	ret = get_mtt_entries(dev, &mr->mem, start, len, access, virt,
920 			      SZ_2G - SZ_4K, 0);
921 	if (ret)
922 		goto err_out_free;
923 
924 	ret = erdma_create_stag(dev, &stag);
925 	if (ret)
926 		goto err_out_put_mtt;
927 
928 	mr->ibmr.lkey = mr->ibmr.rkey = stag;
929 	mr->ibmr.pd = ibpd;
930 	mr->mem.va = virt;
931 	mr->mem.len = len;
932 	mr->access = ERDMA_MR_ACC_LR | to_erdma_access_flags(access);
933 	mr->valid = 1;
934 	mr->type = ERDMA_MR_TYPE_NORMAL;
935 
936 	ret = regmr_cmd(dev, mr);
937 	if (ret)
938 		goto err_out_mr;
939 
940 	return &mr->ibmr;
941 
942 err_out_mr:
943 	erdma_free_idx(&dev->res_cb[ERDMA_RES_TYPE_STAG_IDX],
944 		       mr->ibmr.lkey >> 8);
945 
946 err_out_put_mtt:
947 	put_mtt_entries(dev, &mr->mem);
948 
949 err_out_free:
950 	kfree(mr);
951 
952 	return ERR_PTR(ret);
953 }
954 
955 int erdma_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
956 {
957 	struct erdma_mr *mr;
958 	struct erdma_dev *dev = to_edev(ibmr->device);
959 	struct erdma_cmdq_dereg_mr_req req;
960 	int ret;
961 
962 	mr = to_emr(ibmr);
963 
964 	erdma_cmdq_build_reqhdr(&req.hdr, CMDQ_SUBMOD_RDMA,
965 				CMDQ_OPCODE_DEREG_MR);
966 
967 	req.cfg = FIELD_PREP(ERDMA_CMD_MR_MPT_IDX_MASK, ibmr->lkey >> 8) |
968 		  FIELD_PREP(ERDMA_CMD_MR_KEY_MASK, ibmr->lkey & 0xFF);
969 
970 	ret = erdma_post_cmd_wait(&dev->cmdq, &req, sizeof(req), NULL, NULL);
971 	if (ret)
972 		return ret;
973 
974 	erdma_free_idx(&dev->res_cb[ERDMA_RES_TYPE_STAG_IDX], ibmr->lkey >> 8);
975 
976 	put_mtt_entries(dev, &mr->mem);
977 
978 	kfree(mr);
979 	return 0;
980 }
981 
982 int erdma_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
983 {
984 	struct erdma_cq *cq = to_ecq(ibcq);
985 	struct erdma_dev *dev = to_edev(ibcq->device);
986 	struct erdma_ucontext *ctx = rdma_udata_to_drv_context(
987 		udata, struct erdma_ucontext, ibucontext);
988 	int err;
989 	struct erdma_cmdq_destroy_cq_req req;
990 
991 	erdma_cmdq_build_reqhdr(&req.hdr, CMDQ_SUBMOD_RDMA,
992 				CMDQ_OPCODE_DESTROY_CQ);
993 	req.cqn = cq->cqn;
994 
995 	err = erdma_post_cmd_wait(&dev->cmdq, &req, sizeof(req), NULL, NULL);
996 	if (err)
997 		return err;
998 
999 	if (rdma_is_kernel_res(&cq->ibcq.res)) {
1000 		dma_free_coherent(&dev->pdev->dev,
1001 				  WARPPED_BUFSIZE(cq->depth << CQE_SHIFT),
1002 				  cq->kern_cq.qbuf, cq->kern_cq.qbuf_dma_addr);
1003 	} else {
1004 		erdma_unmap_user_dbrecords(ctx, &cq->user_cq.user_dbr_page);
1005 		put_mtt_entries(dev, &cq->user_cq.qbuf_mtt);
1006 	}
1007 
1008 	xa_erase(&dev->cq_xa, cq->cqn);
1009 
1010 	return 0;
1011 }
1012 
1013 int erdma_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
1014 {
1015 	struct erdma_qp *qp = to_eqp(ibqp);
1016 	struct erdma_dev *dev = to_edev(ibqp->device);
1017 	struct erdma_ucontext *ctx = rdma_udata_to_drv_context(
1018 		udata, struct erdma_ucontext, ibucontext);
1019 	struct erdma_qp_attrs qp_attrs;
1020 	int err;
1021 	struct erdma_cmdq_destroy_qp_req req;
1022 
1023 	down_write(&qp->state_lock);
1024 	qp_attrs.state = ERDMA_QP_STATE_ERROR;
1025 	erdma_modify_qp_internal(qp, &qp_attrs, ERDMA_QP_ATTR_STATE);
1026 	up_write(&qp->state_lock);
1027 
1028 	erdma_cmdq_build_reqhdr(&req.hdr, CMDQ_SUBMOD_RDMA,
1029 				CMDQ_OPCODE_DESTROY_QP);
1030 	req.qpn = QP_ID(qp);
1031 
1032 	err = erdma_post_cmd_wait(&dev->cmdq, &req, sizeof(req), NULL, NULL);
1033 	if (err)
1034 		return err;
1035 
1036 	erdma_qp_put(qp);
1037 	wait_for_completion(&qp->safe_free);
1038 
1039 	if (rdma_is_kernel_res(&qp->ibqp.res)) {
1040 		vfree(qp->kern_qp.swr_tbl);
1041 		vfree(qp->kern_qp.rwr_tbl);
1042 		dma_free_coherent(
1043 			&dev->pdev->dev,
1044 			WARPPED_BUFSIZE(qp->attrs.rq_size << RQE_SHIFT),
1045 			qp->kern_qp.rq_buf, qp->kern_qp.rq_buf_dma_addr);
1046 		dma_free_coherent(
1047 			&dev->pdev->dev,
1048 			WARPPED_BUFSIZE(qp->attrs.sq_size << SQEBB_SHIFT),
1049 			qp->kern_qp.sq_buf, qp->kern_qp.sq_buf_dma_addr);
1050 	} else {
1051 		put_mtt_entries(dev, &qp->user_qp.sq_mtt);
1052 		put_mtt_entries(dev, &qp->user_qp.rq_mtt);
1053 		erdma_unmap_user_dbrecords(ctx, &qp->user_qp.user_dbr_page);
1054 	}
1055 
1056 	if (qp->cep)
1057 		erdma_cep_put(qp->cep);
1058 	xa_erase(&dev->qp_xa, QP_ID(qp));
1059 
1060 	return 0;
1061 }
1062 
1063 void erdma_qp_get_ref(struct ib_qp *ibqp)
1064 {
1065 	erdma_qp_get(to_eqp(ibqp));
1066 }
1067 
1068 void erdma_qp_put_ref(struct ib_qp *ibqp)
1069 {
1070 	erdma_qp_put(to_eqp(ibqp));
1071 }
1072 
1073 int erdma_mmap(struct ib_ucontext *ctx, struct vm_area_struct *vma)
1074 {
1075 	struct rdma_user_mmap_entry *rdma_entry;
1076 	struct erdma_user_mmap_entry *entry;
1077 	pgprot_t prot;
1078 	int err;
1079 
1080 	rdma_entry = rdma_user_mmap_entry_get(ctx, vma);
1081 	if (!rdma_entry)
1082 		return -EINVAL;
1083 
1084 	entry = to_emmap(rdma_entry);
1085 
1086 	switch (entry->mmap_flag) {
1087 	case ERDMA_MMAP_IO_NC:
1088 		/* map doorbell. */
1089 		prot = pgprot_device(vma->vm_page_prot);
1090 		break;
1091 	default:
1092 		return -EINVAL;
1093 	}
1094 
1095 	err = rdma_user_mmap_io(ctx, vma, PFN_DOWN(entry->address), PAGE_SIZE,
1096 				prot, rdma_entry);
1097 
1098 	rdma_user_mmap_entry_put(rdma_entry);
1099 	return err;
1100 }
1101 
1102 void erdma_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
1103 {
1104 	struct erdma_user_mmap_entry *entry = to_emmap(rdma_entry);
1105 
1106 	kfree(entry);
1107 }
1108 
1109 #define ERDMA_SDB_PAGE 0
1110 #define ERDMA_SDB_ENTRY 1
1111 #define ERDMA_SDB_SHARED 2
1112 
1113 static void alloc_db_resources(struct erdma_dev *dev,
1114 			       struct erdma_ucontext *ctx)
1115 {
1116 	u32 bitmap_idx;
1117 	struct erdma_devattr *attrs = &dev->attrs;
1118 
1119 	if (attrs->disable_dwqe)
1120 		goto alloc_normal_db;
1121 
1122 	/* Try to alloc independent SDB page. */
1123 	spin_lock(&dev->db_bitmap_lock);
1124 	bitmap_idx = find_first_zero_bit(dev->sdb_page, attrs->dwqe_pages);
1125 	if (bitmap_idx != attrs->dwqe_pages) {
1126 		set_bit(bitmap_idx, dev->sdb_page);
1127 		spin_unlock(&dev->db_bitmap_lock);
1128 
1129 		ctx->sdb_type = ERDMA_SDB_PAGE;
1130 		ctx->sdb_idx = bitmap_idx;
1131 		ctx->sdb_page_idx = bitmap_idx;
1132 		ctx->sdb = dev->func_bar_addr + ERDMA_BAR_SQDB_SPACE_OFFSET +
1133 			   (bitmap_idx << PAGE_SHIFT);
1134 		ctx->sdb_page_off = 0;
1135 
1136 		return;
1137 	}
1138 
1139 	bitmap_idx = find_first_zero_bit(dev->sdb_entry, attrs->dwqe_entries);
1140 	if (bitmap_idx != attrs->dwqe_entries) {
1141 		set_bit(bitmap_idx, dev->sdb_entry);
1142 		spin_unlock(&dev->db_bitmap_lock);
1143 
1144 		ctx->sdb_type = ERDMA_SDB_ENTRY;
1145 		ctx->sdb_idx = bitmap_idx;
1146 		ctx->sdb_page_idx = attrs->dwqe_pages +
1147 				    bitmap_idx / ERDMA_DWQE_TYPE1_CNT_PER_PAGE;
1148 		ctx->sdb_page_off = bitmap_idx % ERDMA_DWQE_TYPE1_CNT_PER_PAGE;
1149 
1150 		ctx->sdb = dev->func_bar_addr + ERDMA_BAR_SQDB_SPACE_OFFSET +
1151 			   (ctx->sdb_page_idx << PAGE_SHIFT);
1152 
1153 		return;
1154 	}
1155 
1156 	spin_unlock(&dev->db_bitmap_lock);
1157 
1158 alloc_normal_db:
1159 	ctx->sdb_type = ERDMA_SDB_SHARED;
1160 	ctx->sdb_idx = 0;
1161 	ctx->sdb_page_idx = ERDMA_SDB_SHARED_PAGE_INDEX;
1162 	ctx->sdb_page_off = 0;
1163 
1164 	ctx->sdb = dev->func_bar_addr + (ctx->sdb_page_idx << PAGE_SHIFT);
1165 }
1166 
1167 static void erdma_uctx_user_mmap_entries_remove(struct erdma_ucontext *uctx)
1168 {
1169 	rdma_user_mmap_entry_remove(uctx->sq_db_mmap_entry);
1170 	rdma_user_mmap_entry_remove(uctx->rq_db_mmap_entry);
1171 	rdma_user_mmap_entry_remove(uctx->cq_db_mmap_entry);
1172 }
1173 
1174 int erdma_alloc_ucontext(struct ib_ucontext *ibctx, struct ib_udata *udata)
1175 {
1176 	struct erdma_ucontext *ctx = to_ectx(ibctx);
1177 	struct erdma_dev *dev = to_edev(ibctx->device);
1178 	int ret;
1179 	struct erdma_uresp_alloc_ctx uresp = {};
1180 
1181 	if (atomic_inc_return(&dev->num_ctx) > ERDMA_MAX_CONTEXT) {
1182 		ret = -ENOMEM;
1183 		goto err_out;
1184 	}
1185 
1186 	INIT_LIST_HEAD(&ctx->dbrecords_page_list);
1187 	mutex_init(&ctx->dbrecords_page_mutex);
1188 
1189 	alloc_db_resources(dev, ctx);
1190 
1191 	ctx->rdb = dev->func_bar_addr + ERDMA_BAR_RQDB_SPACE_OFFSET;
1192 	ctx->cdb = dev->func_bar_addr + ERDMA_BAR_CQDB_SPACE_OFFSET;
1193 
1194 	if (udata->outlen < sizeof(uresp)) {
1195 		ret = -EINVAL;
1196 		goto err_out;
1197 	}
1198 
1199 	ctx->sq_db_mmap_entry = erdma_user_mmap_entry_insert(
1200 		ctx, (void *)ctx->sdb, PAGE_SIZE, ERDMA_MMAP_IO_NC, &uresp.sdb);
1201 	if (!ctx->sq_db_mmap_entry) {
1202 		ret = -ENOMEM;
1203 		goto err_out;
1204 	}
1205 
1206 	ctx->rq_db_mmap_entry = erdma_user_mmap_entry_insert(
1207 		ctx, (void *)ctx->rdb, PAGE_SIZE, ERDMA_MMAP_IO_NC, &uresp.rdb);
1208 	if (!ctx->rq_db_mmap_entry) {
1209 		ret = -EINVAL;
1210 		goto err_out;
1211 	}
1212 
1213 	ctx->cq_db_mmap_entry = erdma_user_mmap_entry_insert(
1214 		ctx, (void *)ctx->cdb, PAGE_SIZE, ERDMA_MMAP_IO_NC, &uresp.cdb);
1215 	if (!ctx->cq_db_mmap_entry) {
1216 		ret = -EINVAL;
1217 		goto err_out;
1218 	}
1219 
1220 	uresp.dev_id = dev->pdev->device;
1221 	uresp.sdb_type = ctx->sdb_type;
1222 	uresp.sdb_offset = ctx->sdb_page_off;
1223 
1224 	ret = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1225 	if (ret)
1226 		goto err_out;
1227 
1228 	return 0;
1229 
1230 err_out:
1231 	erdma_uctx_user_mmap_entries_remove(ctx);
1232 	atomic_dec(&dev->num_ctx);
1233 	return ret;
1234 }
1235 
1236 void erdma_dealloc_ucontext(struct ib_ucontext *ibctx)
1237 {
1238 	struct erdma_ucontext *ctx = to_ectx(ibctx);
1239 	struct erdma_dev *dev = to_edev(ibctx->device);
1240 
1241 	spin_lock(&dev->db_bitmap_lock);
1242 	if (ctx->sdb_type == ERDMA_SDB_PAGE)
1243 		clear_bit(ctx->sdb_idx, dev->sdb_page);
1244 	else if (ctx->sdb_type == ERDMA_SDB_ENTRY)
1245 		clear_bit(ctx->sdb_idx, dev->sdb_entry);
1246 
1247 	erdma_uctx_user_mmap_entries_remove(ctx);
1248 
1249 	spin_unlock(&dev->db_bitmap_lock);
1250 
1251 	atomic_dec(&dev->num_ctx);
1252 }
1253 
1254 static int ib_qp_state_to_erdma_qp_state[IB_QPS_ERR + 1] = {
1255 	[IB_QPS_RESET] = ERDMA_QP_STATE_IDLE,
1256 	[IB_QPS_INIT] = ERDMA_QP_STATE_IDLE,
1257 	[IB_QPS_RTR] = ERDMA_QP_STATE_RTR,
1258 	[IB_QPS_RTS] = ERDMA_QP_STATE_RTS,
1259 	[IB_QPS_SQD] = ERDMA_QP_STATE_CLOSING,
1260 	[IB_QPS_SQE] = ERDMA_QP_STATE_TERMINATE,
1261 	[IB_QPS_ERR] = ERDMA_QP_STATE_ERROR
1262 };
1263 
1264 int erdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask,
1265 		    struct ib_udata *udata)
1266 {
1267 	struct erdma_qp_attrs new_attrs;
1268 	enum erdma_qp_attr_mask erdma_attr_mask = 0;
1269 	struct erdma_qp *qp = to_eqp(ibqp);
1270 	int ret = 0;
1271 
1272 	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1273 		return -EOPNOTSUPP;
1274 
1275 	memset(&new_attrs, 0, sizeof(new_attrs));
1276 
1277 	if (attr_mask & IB_QP_STATE) {
1278 		new_attrs.state = ib_qp_state_to_erdma_qp_state[attr->qp_state];
1279 
1280 		erdma_attr_mask |= ERDMA_QP_ATTR_STATE;
1281 	}
1282 
1283 	down_write(&qp->state_lock);
1284 
1285 	ret = erdma_modify_qp_internal(qp, &new_attrs, erdma_attr_mask);
1286 
1287 	up_write(&qp->state_lock);
1288 
1289 	return ret;
1290 }
1291 
1292 int erdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
1293 		   int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
1294 {
1295 	struct erdma_qp *qp;
1296 	struct erdma_dev *dev;
1297 
1298 	if (ibqp && qp_attr && qp_init_attr) {
1299 		qp = to_eqp(ibqp);
1300 		dev = to_edev(ibqp->device);
1301 	} else {
1302 		return -EINVAL;
1303 	}
1304 
1305 	qp_attr->cap.max_inline_data = ERDMA_MAX_INLINE;
1306 	qp_init_attr->cap.max_inline_data = ERDMA_MAX_INLINE;
1307 
1308 	qp_attr->cap.max_send_wr = qp->attrs.sq_size;
1309 	qp_attr->cap.max_recv_wr = qp->attrs.rq_size;
1310 	qp_attr->cap.max_send_sge = qp->attrs.max_send_sge;
1311 	qp_attr->cap.max_recv_sge = qp->attrs.max_recv_sge;
1312 
1313 	qp_attr->path_mtu = ib_mtu_int_to_enum(dev->netdev->mtu);
1314 	qp_attr->max_rd_atomic = qp->attrs.irq_size;
1315 	qp_attr->max_dest_rd_atomic = qp->attrs.orq_size;
1316 
1317 	qp_attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE |
1318 				   IB_ACCESS_REMOTE_WRITE |
1319 				   IB_ACCESS_REMOTE_READ;
1320 
1321 	qp_init_attr->cap = qp_attr->cap;
1322 
1323 	return 0;
1324 }
1325 
1326 static int erdma_init_user_cq(struct erdma_ucontext *ctx, struct erdma_cq *cq,
1327 			      struct erdma_ureq_create_cq *ureq)
1328 {
1329 	int ret;
1330 	struct erdma_dev *dev = to_edev(cq->ibcq.device);
1331 
1332 	ret = get_mtt_entries(dev, &cq->user_cq.qbuf_mtt, ureq->qbuf_va,
1333 			      ureq->qbuf_len, 0, ureq->qbuf_va, SZ_64M - SZ_4K,
1334 			      1);
1335 	if (ret)
1336 		return ret;
1337 
1338 	ret = erdma_map_user_dbrecords(ctx, ureq->db_record_va,
1339 				       &cq->user_cq.user_dbr_page,
1340 				       &cq->user_cq.db_info_dma_addr);
1341 	if (ret)
1342 		put_mtt_entries(dev, &cq->user_cq.qbuf_mtt);
1343 
1344 	return ret;
1345 }
1346 
1347 static int erdma_init_kernel_cq(struct erdma_cq *cq)
1348 {
1349 	struct erdma_dev *dev = to_edev(cq->ibcq.device);
1350 
1351 	cq->kern_cq.qbuf =
1352 		dma_alloc_coherent(&dev->pdev->dev,
1353 				   WARPPED_BUFSIZE(cq->depth << CQE_SHIFT),
1354 				   &cq->kern_cq.qbuf_dma_addr, GFP_KERNEL);
1355 	if (!cq->kern_cq.qbuf)
1356 		return -ENOMEM;
1357 
1358 	cq->kern_cq.db_record =
1359 		(u64 *)(cq->kern_cq.qbuf + (cq->depth << CQE_SHIFT));
1360 	spin_lock_init(&cq->kern_cq.lock);
1361 	/* use default cqdb addr */
1362 	cq->kern_cq.db = dev->func_bar + ERDMA_BAR_CQDB_SPACE_OFFSET;
1363 
1364 	return 0;
1365 }
1366 
1367 int erdma_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
1368 		    struct ib_udata *udata)
1369 {
1370 	struct erdma_cq *cq = to_ecq(ibcq);
1371 	struct erdma_dev *dev = to_edev(ibcq->device);
1372 	unsigned int depth = attr->cqe;
1373 	int ret;
1374 	struct erdma_ucontext *ctx = rdma_udata_to_drv_context(
1375 		udata, struct erdma_ucontext, ibucontext);
1376 
1377 	if (depth > dev->attrs.max_cqe)
1378 		return -EINVAL;
1379 
1380 	depth = roundup_pow_of_two(depth);
1381 	cq->ibcq.cqe = depth;
1382 	cq->depth = depth;
1383 	cq->assoc_eqn = attr->comp_vector + 1;
1384 
1385 	ret = xa_alloc_cyclic(&dev->cq_xa, &cq->cqn, cq,
1386 			      XA_LIMIT(1, dev->attrs.max_cq - 1),
1387 			      &dev->next_alloc_cqn, GFP_KERNEL);
1388 	if (ret < 0)
1389 		return ret;
1390 
1391 	if (!rdma_is_kernel_res(&ibcq->res)) {
1392 		struct erdma_ureq_create_cq ureq;
1393 		struct erdma_uresp_create_cq uresp;
1394 
1395 		ret = ib_copy_from_udata(&ureq, udata,
1396 					 min(udata->inlen, sizeof(ureq)));
1397 		if (ret)
1398 			goto err_out_xa;
1399 
1400 		ret = erdma_init_user_cq(ctx, cq, &ureq);
1401 		if (ret)
1402 			goto err_out_xa;
1403 
1404 		uresp.cq_id = cq->cqn;
1405 		uresp.num_cqe = depth;
1406 
1407 		ret = ib_copy_to_udata(udata, &uresp,
1408 				       min(sizeof(uresp), udata->outlen));
1409 		if (ret)
1410 			goto err_free_res;
1411 	} else {
1412 		ret = erdma_init_kernel_cq(cq);
1413 		if (ret)
1414 			goto err_out_xa;
1415 	}
1416 
1417 	ret = create_cq_cmd(dev, cq);
1418 	if (ret)
1419 		goto err_free_res;
1420 
1421 	return 0;
1422 
1423 err_free_res:
1424 	if (!rdma_is_kernel_res(&ibcq->res)) {
1425 		erdma_unmap_user_dbrecords(ctx, &cq->user_cq.user_dbr_page);
1426 		put_mtt_entries(dev, &cq->user_cq.qbuf_mtt);
1427 	} else {
1428 		dma_free_coherent(&dev->pdev->dev,
1429 				  WARPPED_BUFSIZE(depth << CQE_SHIFT),
1430 				  cq->kern_cq.qbuf, cq->kern_cq.qbuf_dma_addr);
1431 	}
1432 
1433 err_out_xa:
1434 	xa_erase(&dev->cq_xa, cq->cqn);
1435 
1436 	return ret;
1437 }
1438 
1439 void erdma_set_mtu(struct erdma_dev *dev, u32 mtu)
1440 {
1441 	struct erdma_cmdq_config_mtu_req req;
1442 
1443 	erdma_cmdq_build_reqhdr(&req.hdr, CMDQ_SUBMOD_COMMON,
1444 				CMDQ_OPCODE_CONF_MTU);
1445 	req.mtu = mtu;
1446 
1447 	erdma_post_cmd_wait(&dev->cmdq, &req, sizeof(req), NULL, NULL);
1448 }
1449 
1450 void erdma_port_event(struct erdma_dev *dev, enum ib_event_type reason)
1451 {
1452 	struct ib_event event;
1453 
1454 	event.device = &dev->ibdev;
1455 	event.element.port_num = 1;
1456 	event.event = reason;
1457 
1458 	ib_dispatch_event(&event);
1459 }
1460