xref: /linux/drivers/infiniband/sw/siw/siw_verbs.c (revision 2634682fdffd9ba6e74b76be8aa91cf8b2e05c41)
1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
2 
3 /* Authors: Bernard Metzler <bmt@zurich.ibm.com> */
4 /* Copyright (c) 2008-2019, IBM Corporation */
5 
6 #include <linux/errno.h>
7 #include <linux/types.h>
8 #include <linux/uaccess.h>
9 #include <linux/vmalloc.h>
10 #include <linux/xarray.h>
11 
12 #include <rdma/iw_cm.h>
13 #include <rdma/ib_verbs.h>
14 #include <rdma/ib_user_verbs.h>
15 #include <rdma/uverbs_ioctl.h>
16 
17 #include "siw.h"
18 #include "siw_verbs.h"
19 #include "siw_mem.h"
20 
21 static int ib_qp_state_to_siw_qp_state[IB_QPS_ERR + 1] = {
22 	[IB_QPS_RESET] = SIW_QP_STATE_IDLE,
23 	[IB_QPS_INIT] = SIW_QP_STATE_IDLE,
24 	[IB_QPS_RTR] = SIW_QP_STATE_RTR,
25 	[IB_QPS_RTS] = SIW_QP_STATE_RTS,
26 	[IB_QPS_SQD] = SIW_QP_STATE_CLOSING,
27 	[IB_QPS_SQE] = SIW_QP_STATE_TERMINATE,
28 	[IB_QPS_ERR] = SIW_QP_STATE_ERROR
29 };
30 
31 static char ib_qp_state_to_string[IB_QPS_ERR + 1][sizeof("RESET")] = {
32 	[IB_QPS_RESET] = "RESET", [IB_QPS_INIT] = "INIT", [IB_QPS_RTR] = "RTR",
33 	[IB_QPS_RTS] = "RTS",     [IB_QPS_SQD] = "SQD",   [IB_QPS_SQE] = "SQE",
34 	[IB_QPS_ERR] = "ERR"
35 };
36 
37 void siw_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
38 {
39 	struct siw_user_mmap_entry *entry = to_siw_mmap_entry(rdma_entry);
40 
41 	kfree(entry);
42 }
43 
44 int siw_mmap(struct ib_ucontext *ctx, struct vm_area_struct *vma)
45 {
46 	struct siw_ucontext *uctx = to_siw_ctx(ctx);
47 	size_t size = vma->vm_end - vma->vm_start;
48 	struct rdma_user_mmap_entry *rdma_entry;
49 	struct siw_user_mmap_entry *entry;
50 	int rv = -EINVAL;
51 
52 	/*
53 	 * Must be page aligned
54 	 */
55 	if (vma->vm_start & (PAGE_SIZE - 1)) {
56 		pr_warn("siw: mmap not page aligned\n");
57 		return -EINVAL;
58 	}
59 	rdma_entry = rdma_user_mmap_entry_get(&uctx->base_ucontext, vma);
60 	if (!rdma_entry) {
61 		siw_dbg(&uctx->sdev->base_dev, "mmap lookup failed: %lu, %#zx\n",
62 			vma->vm_pgoff, size);
63 		return -EINVAL;
64 	}
65 	entry = to_siw_mmap_entry(rdma_entry);
66 
67 	rv = remap_vmalloc_range(vma, entry->address, 0);
68 	if (rv) {
69 		pr_warn("remap_vmalloc_range failed: %lu, %zu\n", vma->vm_pgoff,
70 			size);
71 		goto out;
72 	}
73 out:
74 	rdma_user_mmap_entry_put(rdma_entry);
75 
76 	return rv;
77 }
78 
79 int siw_alloc_ucontext(struct ib_ucontext *base_ctx, struct ib_udata *udata)
80 {
81 	struct siw_device *sdev = to_siw_dev(base_ctx->device);
82 	struct siw_ucontext *ctx = to_siw_ctx(base_ctx);
83 	struct siw_uresp_alloc_ctx uresp = {};
84 	int rv;
85 
86 	if (atomic_inc_return(&sdev->num_ctx) > SIW_MAX_CONTEXT) {
87 		rv = -ENOMEM;
88 		goto err_out;
89 	}
90 	ctx->sdev = sdev;
91 
92 	uresp.dev_id = sdev->vendor_part_id;
93 
94 	if (udata->outlen < sizeof(uresp)) {
95 		rv = -EINVAL;
96 		goto err_out;
97 	}
98 	rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
99 	if (rv)
100 		goto err_out;
101 
102 	siw_dbg(base_ctx->device, "success. now %d context(s)\n",
103 		atomic_read(&sdev->num_ctx));
104 
105 	return 0;
106 
107 err_out:
108 	atomic_dec(&sdev->num_ctx);
109 	siw_dbg(base_ctx->device, "failure %d. now %d context(s)\n", rv,
110 		atomic_read(&sdev->num_ctx));
111 
112 	return rv;
113 }
114 
115 void siw_dealloc_ucontext(struct ib_ucontext *base_ctx)
116 {
117 	struct siw_ucontext *uctx = to_siw_ctx(base_ctx);
118 
119 	atomic_dec(&uctx->sdev->num_ctx);
120 }
121 
122 int siw_query_device(struct ib_device *base_dev, struct ib_device_attr *attr,
123 		     struct ib_udata *udata)
124 {
125 	struct siw_device *sdev = to_siw_dev(base_dev);
126 
127 	if (udata->inlen || udata->outlen)
128 		return -EINVAL;
129 
130 	memset(attr, 0, sizeof(*attr));
131 
132 	/* Revisit atomic caps if RFC 7306 gets supported */
133 	attr->atomic_cap = 0;
134 	attr->device_cap_flags =
135 		IB_DEVICE_MEM_MGT_EXTENSIONS | IB_DEVICE_ALLOW_USER_UNREG;
136 	attr->max_cq = sdev->attrs.max_cq;
137 	attr->max_cqe = sdev->attrs.max_cqe;
138 	attr->max_fast_reg_page_list_len = SIW_MAX_SGE_PBL;
139 	attr->max_mr = sdev->attrs.max_mr;
140 	attr->max_mw = sdev->attrs.max_mw;
141 	attr->max_mr_size = ~0ull;
142 	attr->max_pd = sdev->attrs.max_pd;
143 	attr->max_qp = sdev->attrs.max_qp;
144 	attr->max_qp_init_rd_atom = sdev->attrs.max_ird;
145 	attr->max_qp_rd_atom = sdev->attrs.max_ord;
146 	attr->max_qp_wr = sdev->attrs.max_qp_wr;
147 	attr->max_recv_sge = sdev->attrs.max_sge;
148 	attr->max_res_rd_atom = sdev->attrs.max_qp * sdev->attrs.max_ird;
149 	attr->max_send_sge = sdev->attrs.max_sge;
150 	attr->max_sge_rd = sdev->attrs.max_sge_rd;
151 	attr->max_srq = sdev->attrs.max_srq;
152 	attr->max_srq_sge = sdev->attrs.max_srq_sge;
153 	attr->max_srq_wr = sdev->attrs.max_srq_wr;
154 	attr->page_size_cap = PAGE_SIZE;
155 	attr->vendor_id = SIW_VENDOR_ID;
156 	attr->vendor_part_id = sdev->vendor_part_id;
157 
158 	memcpy(&attr->sys_image_guid, sdev->netdev->dev_addr, 6);
159 
160 	return 0;
161 }
162 
163 int siw_query_port(struct ib_device *base_dev, u8 port,
164 		   struct ib_port_attr *attr)
165 {
166 	struct siw_device *sdev = to_siw_dev(base_dev);
167 	int rv;
168 
169 	memset(attr, 0, sizeof(*attr));
170 
171 	rv = ib_get_eth_speed(base_dev, port, &attr->active_speed,
172 			 &attr->active_width);
173 	attr->gid_tbl_len = 1;
174 	attr->max_msg_sz = -1;
175 	attr->max_mtu = ib_mtu_int_to_enum(sdev->netdev->mtu);
176 	attr->active_mtu = ib_mtu_int_to_enum(sdev->netdev->mtu);
177 	attr->phys_state = sdev->state == IB_PORT_ACTIVE ?
178 		IB_PORT_PHYS_STATE_LINK_UP : IB_PORT_PHYS_STATE_DISABLED;
179 	attr->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_DEVICE_MGMT_SUP;
180 	attr->state = sdev->state;
181 	/*
182 	 * All zero
183 	 *
184 	 * attr->lid = 0;
185 	 * attr->bad_pkey_cntr = 0;
186 	 * attr->qkey_viol_cntr = 0;
187 	 * attr->sm_lid = 0;
188 	 * attr->lmc = 0;
189 	 * attr->max_vl_num = 0;
190 	 * attr->sm_sl = 0;
191 	 * attr->subnet_timeout = 0;
192 	 * attr->init_type_repy = 0;
193 	 */
194 	return rv;
195 }
196 
197 int siw_get_port_immutable(struct ib_device *base_dev, u8 port,
198 			   struct ib_port_immutable *port_immutable)
199 {
200 	struct ib_port_attr attr;
201 	int rv = siw_query_port(base_dev, port, &attr);
202 
203 	if (rv)
204 		return rv;
205 
206 	port_immutable->gid_tbl_len = attr.gid_tbl_len;
207 	port_immutable->core_cap_flags = RDMA_CORE_PORT_IWARP;
208 
209 	return 0;
210 }
211 
212 int siw_query_gid(struct ib_device *base_dev, u8 port, int idx,
213 		  union ib_gid *gid)
214 {
215 	struct siw_device *sdev = to_siw_dev(base_dev);
216 
217 	/* subnet_prefix == interface_id == 0; */
218 	memset(gid, 0, sizeof(*gid));
219 	memcpy(&gid->raw[0], sdev->netdev->dev_addr, 6);
220 
221 	return 0;
222 }
223 
224 int siw_alloc_pd(struct ib_pd *pd, struct ib_udata *udata)
225 {
226 	struct siw_device *sdev = to_siw_dev(pd->device);
227 
228 	if (atomic_inc_return(&sdev->num_pd) > SIW_MAX_PD) {
229 		atomic_dec(&sdev->num_pd);
230 		return -ENOMEM;
231 	}
232 	siw_dbg_pd(pd, "now %d PD's(s)\n", atomic_read(&sdev->num_pd));
233 
234 	return 0;
235 }
236 
237 void siw_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
238 {
239 	struct siw_device *sdev = to_siw_dev(pd->device);
240 
241 	siw_dbg_pd(pd, "free PD\n");
242 	atomic_dec(&sdev->num_pd);
243 }
244 
245 void siw_qp_get_ref(struct ib_qp *base_qp)
246 {
247 	siw_qp_get(to_siw_qp(base_qp));
248 }
249 
250 void siw_qp_put_ref(struct ib_qp *base_qp)
251 {
252 	siw_qp_put(to_siw_qp(base_qp));
253 }
254 
255 static struct rdma_user_mmap_entry *
256 siw_mmap_entry_insert(struct siw_ucontext *uctx,
257 		      void *address, size_t length,
258 		      u64 *offset)
259 {
260 	struct siw_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
261 	int rv;
262 
263 	*offset = SIW_INVAL_UOBJ_KEY;
264 	if (!entry)
265 		return NULL;
266 
267 	entry->address = address;
268 
269 	rv = rdma_user_mmap_entry_insert(&uctx->base_ucontext,
270 					 &entry->rdma_entry,
271 					 length);
272 	if (rv) {
273 		kfree(entry);
274 		return NULL;
275 	}
276 
277 	*offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
278 
279 	return &entry->rdma_entry;
280 }
281 
282 /*
283  * siw_create_qp()
284  *
285  * Create QP of requested size on given device.
286  *
287  * @pd:		Protection Domain
288  * @attrs:	Initial QP attributes.
289  * @udata:	used to provide QP ID, SQ and RQ size back to user.
290  */
291 
292 struct ib_qp *siw_create_qp(struct ib_pd *pd,
293 			    struct ib_qp_init_attr *attrs,
294 			    struct ib_udata *udata)
295 {
296 	struct siw_qp *qp = NULL;
297 	struct ib_device *base_dev = pd->device;
298 	struct siw_device *sdev = to_siw_dev(base_dev);
299 	struct siw_ucontext *uctx =
300 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
301 					  base_ucontext);
302 	struct siw_cq *scq = NULL, *rcq = NULL;
303 	unsigned long flags;
304 	int num_sqe, num_rqe, rv = 0;
305 	size_t length;
306 
307 	siw_dbg(base_dev, "create new QP\n");
308 
309 	if (atomic_inc_return(&sdev->num_qp) > SIW_MAX_QP) {
310 		siw_dbg(base_dev, "too many QP's\n");
311 		rv = -ENOMEM;
312 		goto err_out;
313 	}
314 	if (attrs->qp_type != IB_QPT_RC) {
315 		siw_dbg(base_dev, "only RC QP's supported\n");
316 		rv = -EOPNOTSUPP;
317 		goto err_out;
318 	}
319 	if ((attrs->cap.max_send_wr > SIW_MAX_QP_WR) ||
320 	    (attrs->cap.max_recv_wr > SIW_MAX_QP_WR) ||
321 	    (attrs->cap.max_send_sge > SIW_MAX_SGE) ||
322 	    (attrs->cap.max_recv_sge > SIW_MAX_SGE)) {
323 		siw_dbg(base_dev, "QP size error\n");
324 		rv = -EINVAL;
325 		goto err_out;
326 	}
327 	if (attrs->cap.max_inline_data > SIW_MAX_INLINE) {
328 		siw_dbg(base_dev, "max inline send: %d > %d\n",
329 			attrs->cap.max_inline_data, (int)SIW_MAX_INLINE);
330 		rv = -EINVAL;
331 		goto err_out;
332 	}
333 	/*
334 	 * NOTE: we allow for zero element SQ and RQ WQE's SGL's
335 	 * but not for a QP unable to hold any WQE (SQ + RQ)
336 	 */
337 	if (attrs->cap.max_send_wr + attrs->cap.max_recv_wr == 0) {
338 		siw_dbg(base_dev, "QP must have send or receive queue\n");
339 		rv = -EINVAL;
340 		goto err_out;
341 	}
342 	scq = to_siw_cq(attrs->send_cq);
343 	rcq = to_siw_cq(attrs->recv_cq);
344 
345 	if (!scq || (!rcq && !attrs->srq)) {
346 		siw_dbg(base_dev, "send CQ or receive CQ invalid\n");
347 		rv = -EINVAL;
348 		goto err_out;
349 	}
350 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
351 	if (!qp) {
352 		rv = -ENOMEM;
353 		goto err_out;
354 	}
355 	init_rwsem(&qp->state_lock);
356 	spin_lock_init(&qp->sq_lock);
357 	spin_lock_init(&qp->rq_lock);
358 	spin_lock_init(&qp->orq_lock);
359 
360 	rv = siw_qp_add(sdev, qp);
361 	if (rv)
362 		goto err_out;
363 
364 	/* All queue indices are derived from modulo operations
365 	 * on a free running 'get' (consumer) and 'put' (producer)
366 	 * unsigned counter. Having queue sizes at power of two
367 	 * avoids handling counter wrap around.
368 	 */
369 	num_sqe = roundup_pow_of_two(attrs->cap.max_send_wr);
370 	num_rqe = roundup_pow_of_two(attrs->cap.max_recv_wr);
371 
372 	if (udata)
373 		qp->sendq = vmalloc_user(num_sqe * sizeof(struct siw_sqe));
374 	else
375 		qp->sendq = vzalloc(num_sqe * sizeof(struct siw_sqe));
376 
377 	if (qp->sendq == NULL) {
378 		siw_dbg(base_dev, "SQ size %d alloc failed\n", num_sqe);
379 		rv = -ENOMEM;
380 		goto err_out_xa;
381 	}
382 	if (attrs->sq_sig_type != IB_SIGNAL_REQ_WR) {
383 		if (attrs->sq_sig_type == IB_SIGNAL_ALL_WR)
384 			qp->attrs.flags |= SIW_SIGNAL_ALL_WR;
385 		else {
386 			rv = -EINVAL;
387 			goto err_out_xa;
388 		}
389 	}
390 	qp->pd = pd;
391 	qp->scq = scq;
392 	qp->rcq = rcq;
393 
394 	if (attrs->srq) {
395 		/*
396 		 * SRQ support.
397 		 * Verbs 6.3.7: ignore RQ size, if SRQ present
398 		 * Verbs 6.3.5: do not check PD of SRQ against PD of QP
399 		 */
400 		qp->srq = to_siw_srq(attrs->srq);
401 		qp->attrs.rq_size = 0;
402 		siw_dbg(base_dev, "QP [%u]: SRQ attached\n",
403 			qp->base_qp.qp_num);
404 	} else if (num_rqe) {
405 		if (udata)
406 			qp->recvq =
407 				vmalloc_user(num_rqe * sizeof(struct siw_rqe));
408 		else
409 			qp->recvq = vzalloc(num_rqe * sizeof(struct siw_rqe));
410 
411 		if (qp->recvq == NULL) {
412 			siw_dbg(base_dev, "RQ size %d alloc failed\n", num_rqe);
413 			rv = -ENOMEM;
414 			goto err_out_xa;
415 		}
416 		qp->attrs.rq_size = num_rqe;
417 	}
418 	qp->attrs.sq_size = num_sqe;
419 	qp->attrs.sq_max_sges = attrs->cap.max_send_sge;
420 	qp->attrs.rq_max_sges = attrs->cap.max_recv_sge;
421 
422 	/* Make those two tunables fixed for now. */
423 	qp->tx_ctx.gso_seg_limit = 1;
424 	qp->tx_ctx.zcopy_tx = zcopy_tx;
425 
426 	qp->attrs.state = SIW_QP_STATE_IDLE;
427 
428 	if (udata) {
429 		struct siw_uresp_create_qp uresp = {};
430 
431 		uresp.num_sqe = num_sqe;
432 		uresp.num_rqe = num_rqe;
433 		uresp.qp_id = qp_id(qp);
434 
435 		if (qp->sendq) {
436 			length = num_sqe * sizeof(struct siw_sqe);
437 			qp->sq_entry =
438 				siw_mmap_entry_insert(uctx, qp->sendq,
439 						      length, &uresp.sq_key);
440 			if (!qp->sq_entry) {
441 				rv = -ENOMEM;
442 				goto err_out_xa;
443 			}
444 		}
445 
446 		if (qp->recvq) {
447 			length = num_rqe * sizeof(struct siw_rqe);
448 			qp->rq_entry =
449 				siw_mmap_entry_insert(uctx, qp->recvq,
450 						      length, &uresp.rq_key);
451 			if (!qp->rq_entry) {
452 				uresp.sq_key = SIW_INVAL_UOBJ_KEY;
453 				rv = -ENOMEM;
454 				goto err_out_xa;
455 			}
456 		}
457 
458 		if (udata->outlen < sizeof(uresp)) {
459 			rv = -EINVAL;
460 			goto err_out_xa;
461 		}
462 		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
463 		if (rv)
464 			goto err_out_xa;
465 	}
466 	qp->tx_cpu = siw_get_tx_cpu(sdev);
467 	if (qp->tx_cpu < 0) {
468 		rv = -EINVAL;
469 		goto err_out_xa;
470 	}
471 	INIT_LIST_HEAD(&qp->devq);
472 	spin_lock_irqsave(&sdev->lock, flags);
473 	list_add_tail(&qp->devq, &sdev->qp_list);
474 	spin_unlock_irqrestore(&sdev->lock, flags);
475 
476 	return &qp->base_qp;
477 
478 err_out_xa:
479 	xa_erase(&sdev->qp_xa, qp_id(qp));
480 err_out:
481 	if (qp) {
482 		if (uctx) {
483 			rdma_user_mmap_entry_remove(qp->sq_entry);
484 			rdma_user_mmap_entry_remove(qp->rq_entry);
485 		}
486 		vfree(qp->sendq);
487 		vfree(qp->recvq);
488 		kfree(qp);
489 	}
490 	atomic_dec(&sdev->num_qp);
491 
492 	return ERR_PTR(rv);
493 }
494 
495 /*
496  * Minimum siw_query_qp() verb interface.
497  *
498  * @qp_attr_mask is not used but all available information is provided
499  */
500 int siw_query_qp(struct ib_qp *base_qp, struct ib_qp_attr *qp_attr,
501 		 int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
502 {
503 	struct siw_qp *qp;
504 	struct siw_device *sdev;
505 
506 	if (base_qp && qp_attr && qp_init_attr) {
507 		qp = to_siw_qp(base_qp);
508 		sdev = to_siw_dev(base_qp->device);
509 	} else {
510 		return -EINVAL;
511 	}
512 	qp_attr->cap.max_inline_data = SIW_MAX_INLINE;
513 	qp_attr->cap.max_send_wr = qp->attrs.sq_size;
514 	qp_attr->cap.max_send_sge = qp->attrs.sq_max_sges;
515 	qp_attr->cap.max_recv_wr = qp->attrs.rq_size;
516 	qp_attr->cap.max_recv_sge = qp->attrs.rq_max_sges;
517 	qp_attr->path_mtu = ib_mtu_int_to_enum(sdev->netdev->mtu);
518 	qp_attr->max_rd_atomic = qp->attrs.irq_size;
519 	qp_attr->max_dest_rd_atomic = qp->attrs.orq_size;
520 
521 	qp_attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE |
522 				   IB_ACCESS_REMOTE_WRITE |
523 				   IB_ACCESS_REMOTE_READ;
524 
525 	qp_init_attr->qp_type = base_qp->qp_type;
526 	qp_init_attr->send_cq = base_qp->send_cq;
527 	qp_init_attr->recv_cq = base_qp->recv_cq;
528 	qp_init_attr->srq = base_qp->srq;
529 
530 	qp_init_attr->cap = qp_attr->cap;
531 
532 	return 0;
533 }
534 
535 int siw_verbs_modify_qp(struct ib_qp *base_qp, struct ib_qp_attr *attr,
536 			int attr_mask, struct ib_udata *udata)
537 {
538 	struct siw_qp_attrs new_attrs;
539 	enum siw_qp_attr_mask siw_attr_mask = 0;
540 	struct siw_qp *qp = to_siw_qp(base_qp);
541 	int rv = 0;
542 
543 	if (!attr_mask)
544 		return 0;
545 
546 	memset(&new_attrs, 0, sizeof(new_attrs));
547 
548 	if (attr_mask & IB_QP_ACCESS_FLAGS) {
549 		siw_attr_mask = SIW_QP_ATTR_ACCESS_FLAGS;
550 
551 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
552 			new_attrs.flags |= SIW_RDMA_READ_ENABLED;
553 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
554 			new_attrs.flags |= SIW_RDMA_WRITE_ENABLED;
555 		if (attr->qp_access_flags & IB_ACCESS_MW_BIND)
556 			new_attrs.flags |= SIW_RDMA_BIND_ENABLED;
557 	}
558 	if (attr_mask & IB_QP_STATE) {
559 		siw_dbg_qp(qp, "desired IB QP state: %s\n",
560 			   ib_qp_state_to_string[attr->qp_state]);
561 
562 		new_attrs.state = ib_qp_state_to_siw_qp_state[attr->qp_state];
563 
564 		if (new_attrs.state > SIW_QP_STATE_RTS)
565 			qp->tx_ctx.tx_suspend = 1;
566 
567 		siw_attr_mask |= SIW_QP_ATTR_STATE;
568 	}
569 	if (!siw_attr_mask)
570 		goto out;
571 
572 	down_write(&qp->state_lock);
573 
574 	rv = siw_qp_modify(qp, &new_attrs, siw_attr_mask);
575 
576 	up_write(&qp->state_lock);
577 out:
578 	return rv;
579 }
580 
581 int siw_destroy_qp(struct ib_qp *base_qp, struct ib_udata *udata)
582 {
583 	struct siw_qp *qp = to_siw_qp(base_qp);
584 	struct siw_ucontext *uctx =
585 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
586 					  base_ucontext);
587 	struct siw_qp_attrs qp_attrs;
588 
589 	siw_dbg_qp(qp, "state %d\n", qp->attrs.state);
590 
591 	/*
592 	 * Mark QP as in process of destruction to prevent from
593 	 * any async callbacks to RDMA core
594 	 */
595 	qp->attrs.flags |= SIW_QP_IN_DESTROY;
596 	qp->rx_stream.rx_suspend = 1;
597 
598 	if (uctx) {
599 		rdma_user_mmap_entry_remove(qp->sq_entry);
600 		rdma_user_mmap_entry_remove(qp->rq_entry);
601 	}
602 
603 	down_write(&qp->state_lock);
604 
605 	qp_attrs.state = SIW_QP_STATE_ERROR;
606 	siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE);
607 
608 	if (qp->cep) {
609 		siw_cep_put(qp->cep);
610 		qp->cep = NULL;
611 	}
612 	up_write(&qp->state_lock);
613 
614 	kfree(qp->tx_ctx.mpa_crc_hd);
615 	kfree(qp->rx_stream.mpa_crc_hd);
616 
617 	qp->scq = qp->rcq = NULL;
618 
619 	siw_qp_put(qp);
620 
621 	return 0;
622 }
623 
624 /*
625  * siw_copy_inline_sgl()
626  *
627  * Prepare sgl of inlined data for sending. For userland callers
628  * function checks if given buffer addresses and len's are within
629  * process context bounds.
630  * Data from all provided sge's are copied together into the wqe,
631  * referenced by a single sge.
632  */
633 static int siw_copy_inline_sgl(const struct ib_send_wr *core_wr,
634 			       struct siw_sqe *sqe)
635 {
636 	struct ib_sge *core_sge = core_wr->sg_list;
637 	void *kbuf = &sqe->sge[1];
638 	int num_sge = core_wr->num_sge, bytes = 0;
639 
640 	sqe->sge[0].laddr = (uintptr_t)kbuf;
641 	sqe->sge[0].lkey = 0;
642 
643 	while (num_sge--) {
644 		if (!core_sge->length) {
645 			core_sge++;
646 			continue;
647 		}
648 		bytes += core_sge->length;
649 		if (bytes > SIW_MAX_INLINE) {
650 			bytes = -EINVAL;
651 			break;
652 		}
653 		memcpy(kbuf, (void *)(uintptr_t)core_sge->addr,
654 		       core_sge->length);
655 
656 		kbuf += core_sge->length;
657 		core_sge++;
658 	}
659 	sqe->sge[0].length = bytes > 0 ? bytes : 0;
660 	sqe->num_sge = bytes > 0 ? 1 : 0;
661 
662 	return bytes;
663 }
664 
665 /* Complete SQ WR's without processing */
666 static int siw_sq_flush_wr(struct siw_qp *qp, const struct ib_send_wr *wr,
667 			   const struct ib_send_wr **bad_wr)
668 {
669 	struct siw_sqe sqe = {};
670 	int rv = 0;
671 
672 	while (wr) {
673 		sqe.id = wr->wr_id;
674 		sqe.opcode = wr->opcode;
675 		rv = siw_sqe_complete(qp, &sqe, 0, SIW_WC_WR_FLUSH_ERR);
676 		if (rv) {
677 			if (bad_wr)
678 				*bad_wr = wr;
679 			break;
680 		}
681 		wr = wr->next;
682 	}
683 	return rv;
684 }
685 
686 /* Complete RQ WR's without processing */
687 static int siw_rq_flush_wr(struct siw_qp *qp, const struct ib_recv_wr *wr,
688 			   const struct ib_recv_wr **bad_wr)
689 {
690 	struct siw_rqe rqe = {};
691 	int rv = 0;
692 
693 	while (wr) {
694 		rqe.id = wr->wr_id;
695 		rv = siw_rqe_complete(qp, &rqe, 0, 0, SIW_WC_WR_FLUSH_ERR);
696 		if (rv) {
697 			if (bad_wr)
698 				*bad_wr = wr;
699 			break;
700 		}
701 		wr = wr->next;
702 	}
703 	return rv;
704 }
705 
706 /*
707  * siw_post_send()
708  *
709  * Post a list of S-WR's to a SQ.
710  *
711  * @base_qp:	Base QP contained in siw QP
712  * @wr:		Null terminated list of user WR's
713  * @bad_wr:	Points to failing WR in case of synchronous failure.
714  */
715 int siw_post_send(struct ib_qp *base_qp, const struct ib_send_wr *wr,
716 		  const struct ib_send_wr **bad_wr)
717 {
718 	struct siw_qp *qp = to_siw_qp(base_qp);
719 	struct siw_wqe *wqe = tx_wqe(qp);
720 
721 	unsigned long flags;
722 	int rv = 0;
723 
724 	if (wr && !rdma_is_kernel_res(&qp->base_qp.res)) {
725 		siw_dbg_qp(qp, "wr must be empty for user mapped sq\n");
726 		*bad_wr = wr;
727 		return -EINVAL;
728 	}
729 
730 	/*
731 	 * Try to acquire QP state lock. Must be non-blocking
732 	 * to accommodate kernel clients needs.
733 	 */
734 	if (!down_read_trylock(&qp->state_lock)) {
735 		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
736 			/*
737 			 * ERROR state is final, so we can be sure
738 			 * this state will not change as long as the QP
739 			 * exists.
740 			 *
741 			 * This handles an ib_drain_sq() call with
742 			 * a concurrent request to set the QP state
743 			 * to ERROR.
744 			 */
745 			rv = siw_sq_flush_wr(qp, wr, bad_wr);
746 		} else {
747 			siw_dbg_qp(qp, "QP locked, state %d\n",
748 				   qp->attrs.state);
749 			*bad_wr = wr;
750 			rv = -ENOTCONN;
751 		}
752 		return rv;
753 	}
754 	if (unlikely(qp->attrs.state != SIW_QP_STATE_RTS)) {
755 		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
756 			/*
757 			 * Immediately flush this WR to CQ, if QP
758 			 * is in ERROR state. SQ is guaranteed to
759 			 * be empty, so WR complets in-order.
760 			 *
761 			 * Typically triggered by ib_drain_sq().
762 			 */
763 			rv = siw_sq_flush_wr(qp, wr, bad_wr);
764 		} else {
765 			siw_dbg_qp(qp, "QP out of state %d\n",
766 				   qp->attrs.state);
767 			*bad_wr = wr;
768 			rv = -ENOTCONN;
769 		}
770 		up_read(&qp->state_lock);
771 		return rv;
772 	}
773 	spin_lock_irqsave(&qp->sq_lock, flags);
774 
775 	while (wr) {
776 		u32 idx = qp->sq_put % qp->attrs.sq_size;
777 		struct siw_sqe *sqe = &qp->sendq[idx];
778 
779 		if (sqe->flags) {
780 			siw_dbg_qp(qp, "sq full\n");
781 			rv = -ENOMEM;
782 			break;
783 		}
784 		if (wr->num_sge > qp->attrs.sq_max_sges) {
785 			siw_dbg_qp(qp, "too many sge's: %d\n", wr->num_sge);
786 			rv = -EINVAL;
787 			break;
788 		}
789 		sqe->id = wr->wr_id;
790 
791 		if ((wr->send_flags & IB_SEND_SIGNALED) ||
792 		    (qp->attrs.flags & SIW_SIGNAL_ALL_WR))
793 			sqe->flags |= SIW_WQE_SIGNALLED;
794 
795 		if (wr->send_flags & IB_SEND_FENCE)
796 			sqe->flags |= SIW_WQE_READ_FENCE;
797 
798 		switch (wr->opcode) {
799 		case IB_WR_SEND:
800 		case IB_WR_SEND_WITH_INV:
801 			if (wr->send_flags & IB_SEND_SOLICITED)
802 				sqe->flags |= SIW_WQE_SOLICITED;
803 
804 			if (!(wr->send_flags & IB_SEND_INLINE)) {
805 				siw_copy_sgl(wr->sg_list, sqe->sge,
806 					     wr->num_sge);
807 				sqe->num_sge = wr->num_sge;
808 			} else {
809 				rv = siw_copy_inline_sgl(wr, sqe);
810 				if (rv <= 0) {
811 					rv = -EINVAL;
812 					break;
813 				}
814 				sqe->flags |= SIW_WQE_INLINE;
815 				sqe->num_sge = 1;
816 			}
817 			if (wr->opcode == IB_WR_SEND)
818 				sqe->opcode = SIW_OP_SEND;
819 			else {
820 				sqe->opcode = SIW_OP_SEND_REMOTE_INV;
821 				sqe->rkey = wr->ex.invalidate_rkey;
822 			}
823 			break;
824 
825 		case IB_WR_RDMA_READ_WITH_INV:
826 		case IB_WR_RDMA_READ:
827 			/*
828 			 * iWarp restricts RREAD sink to SGL containing
829 			 * 1 SGE only. we could relax to SGL with multiple
830 			 * elements referring the SAME ltag or even sending
831 			 * a private per-rreq tag referring to a checked
832 			 * local sgl with MULTIPLE ltag's.
833 			 */
834 			if (unlikely(wr->num_sge != 1)) {
835 				rv = -EINVAL;
836 				break;
837 			}
838 			siw_copy_sgl(wr->sg_list, &sqe->sge[0], 1);
839 			/*
840 			 * NOTE: zero length RREAD is allowed!
841 			 */
842 			sqe->raddr = rdma_wr(wr)->remote_addr;
843 			sqe->rkey = rdma_wr(wr)->rkey;
844 			sqe->num_sge = 1;
845 
846 			if (wr->opcode == IB_WR_RDMA_READ)
847 				sqe->opcode = SIW_OP_READ;
848 			else
849 				sqe->opcode = SIW_OP_READ_LOCAL_INV;
850 			break;
851 
852 		case IB_WR_RDMA_WRITE:
853 			if (!(wr->send_flags & IB_SEND_INLINE)) {
854 				siw_copy_sgl(wr->sg_list, &sqe->sge[0],
855 					     wr->num_sge);
856 				sqe->num_sge = wr->num_sge;
857 			} else {
858 				rv = siw_copy_inline_sgl(wr, sqe);
859 				if (unlikely(rv < 0)) {
860 					rv = -EINVAL;
861 					break;
862 				}
863 				sqe->flags |= SIW_WQE_INLINE;
864 				sqe->num_sge = 1;
865 			}
866 			sqe->raddr = rdma_wr(wr)->remote_addr;
867 			sqe->rkey = rdma_wr(wr)->rkey;
868 			sqe->opcode = SIW_OP_WRITE;
869 			break;
870 
871 		case IB_WR_REG_MR:
872 			sqe->base_mr = (uintptr_t)reg_wr(wr)->mr;
873 			sqe->rkey = reg_wr(wr)->key;
874 			sqe->access = reg_wr(wr)->access & IWARP_ACCESS_MASK;
875 			sqe->opcode = SIW_OP_REG_MR;
876 			break;
877 
878 		case IB_WR_LOCAL_INV:
879 			sqe->rkey = wr->ex.invalidate_rkey;
880 			sqe->opcode = SIW_OP_INVAL_STAG;
881 			break;
882 
883 		default:
884 			siw_dbg_qp(qp, "ib wr type %d unsupported\n",
885 				   wr->opcode);
886 			rv = -EINVAL;
887 			break;
888 		}
889 		siw_dbg_qp(qp, "opcode %d, flags 0x%x, wr_id 0x%pK\n",
890 			   sqe->opcode, sqe->flags,
891 			   (void *)(uintptr_t)sqe->id);
892 
893 		if (unlikely(rv < 0))
894 			break;
895 
896 		/* make SQE only valid after completely written */
897 		smp_wmb();
898 		sqe->flags |= SIW_WQE_VALID;
899 
900 		qp->sq_put++;
901 		wr = wr->next;
902 	}
903 
904 	/*
905 	 * Send directly if SQ processing is not in progress.
906 	 * Eventual immediate errors (rv < 0) do not affect the involved
907 	 * RI resources (Verbs, 8.3.1) and thus do not prevent from SQ
908 	 * processing, if new work is already pending. But rv must be passed
909 	 * to caller.
910 	 */
911 	if (wqe->wr_status != SIW_WR_IDLE) {
912 		spin_unlock_irqrestore(&qp->sq_lock, flags);
913 		goto skip_direct_sending;
914 	}
915 	rv = siw_activate_tx(qp);
916 	spin_unlock_irqrestore(&qp->sq_lock, flags);
917 
918 	if (rv <= 0)
919 		goto skip_direct_sending;
920 
921 	if (rdma_is_kernel_res(&qp->base_qp.res)) {
922 		rv = siw_sq_start(qp);
923 	} else {
924 		qp->tx_ctx.in_syscall = 1;
925 
926 		if (siw_qp_sq_process(qp) != 0 && !(qp->tx_ctx.tx_suspend))
927 			siw_qp_cm_drop(qp, 0);
928 
929 		qp->tx_ctx.in_syscall = 0;
930 	}
931 skip_direct_sending:
932 
933 	up_read(&qp->state_lock);
934 
935 	if (rv >= 0)
936 		return 0;
937 	/*
938 	 * Immediate error
939 	 */
940 	siw_dbg_qp(qp, "error %d\n", rv);
941 
942 	*bad_wr = wr;
943 	return rv;
944 }
945 
946 /*
947  * siw_post_receive()
948  *
949  * Post a list of R-WR's to a RQ.
950  *
951  * @base_qp:	Base QP contained in siw QP
952  * @wr:		Null terminated list of user WR's
953  * @bad_wr:	Points to failing WR in case of synchronous failure.
954  */
955 int siw_post_receive(struct ib_qp *base_qp, const struct ib_recv_wr *wr,
956 		     const struct ib_recv_wr **bad_wr)
957 {
958 	struct siw_qp *qp = to_siw_qp(base_qp);
959 	unsigned long flags;
960 	int rv = 0;
961 
962 	if (qp->srq) {
963 		*bad_wr = wr;
964 		return -EOPNOTSUPP; /* what else from errno.h? */
965 	}
966 	if (!rdma_is_kernel_res(&qp->base_qp.res)) {
967 		siw_dbg_qp(qp, "no kernel post_recv for user mapped rq\n");
968 		*bad_wr = wr;
969 		return -EINVAL;
970 	}
971 
972 	/*
973 	 * Try to acquire QP state lock. Must be non-blocking
974 	 * to accommodate kernel clients needs.
975 	 */
976 	if (!down_read_trylock(&qp->state_lock)) {
977 		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
978 			/*
979 			 * ERROR state is final, so we can be sure
980 			 * this state will not change as long as the QP
981 			 * exists.
982 			 *
983 			 * This handles an ib_drain_rq() call with
984 			 * a concurrent request to set the QP state
985 			 * to ERROR.
986 			 */
987 			rv = siw_rq_flush_wr(qp, wr, bad_wr);
988 		} else {
989 			siw_dbg_qp(qp, "QP locked, state %d\n",
990 				   qp->attrs.state);
991 			*bad_wr = wr;
992 			rv = -ENOTCONN;
993 		}
994 		return rv;
995 	}
996 	if (qp->attrs.state > SIW_QP_STATE_RTS) {
997 		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
998 			/*
999 			 * Immediately flush this WR to CQ, if QP
1000 			 * is in ERROR state. RQ is guaranteed to
1001 			 * be empty, so WR complets in-order.
1002 			 *
1003 			 * Typically triggered by ib_drain_rq().
1004 			 */
1005 			rv = siw_rq_flush_wr(qp, wr, bad_wr);
1006 		} else {
1007 			siw_dbg_qp(qp, "QP out of state %d\n",
1008 				   qp->attrs.state);
1009 			*bad_wr = wr;
1010 			rv = -ENOTCONN;
1011 		}
1012 		up_read(&qp->state_lock);
1013 		return rv;
1014 	}
1015 	/*
1016 	 * Serialize potentially multiple producers.
1017 	 * Not needed for single threaded consumer side.
1018 	 */
1019 	spin_lock_irqsave(&qp->rq_lock, flags);
1020 
1021 	while (wr) {
1022 		u32 idx = qp->rq_put % qp->attrs.rq_size;
1023 		struct siw_rqe *rqe = &qp->recvq[idx];
1024 
1025 		if (rqe->flags) {
1026 			siw_dbg_qp(qp, "RQ full\n");
1027 			rv = -ENOMEM;
1028 			break;
1029 		}
1030 		if (wr->num_sge > qp->attrs.rq_max_sges) {
1031 			siw_dbg_qp(qp, "too many sge's: %d\n", wr->num_sge);
1032 			rv = -EINVAL;
1033 			break;
1034 		}
1035 		rqe->id = wr->wr_id;
1036 		rqe->num_sge = wr->num_sge;
1037 		siw_copy_sgl(wr->sg_list, rqe->sge, wr->num_sge);
1038 
1039 		/* make sure RQE is completely written before valid */
1040 		smp_wmb();
1041 
1042 		rqe->flags = SIW_WQE_VALID;
1043 
1044 		qp->rq_put++;
1045 		wr = wr->next;
1046 	}
1047 	spin_unlock_irqrestore(&qp->rq_lock, flags);
1048 
1049 	up_read(&qp->state_lock);
1050 
1051 	if (rv < 0) {
1052 		siw_dbg_qp(qp, "error %d\n", rv);
1053 		*bad_wr = wr;
1054 	}
1055 	return rv > 0 ? 0 : rv;
1056 }
1057 
1058 void siw_destroy_cq(struct ib_cq *base_cq, struct ib_udata *udata)
1059 {
1060 	struct siw_cq *cq = to_siw_cq(base_cq);
1061 	struct siw_device *sdev = to_siw_dev(base_cq->device);
1062 	struct siw_ucontext *ctx =
1063 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
1064 					  base_ucontext);
1065 
1066 	siw_dbg_cq(cq, "free CQ resources\n");
1067 
1068 	siw_cq_flush(cq);
1069 
1070 	if (ctx)
1071 		rdma_user_mmap_entry_remove(cq->cq_entry);
1072 
1073 	atomic_dec(&sdev->num_cq);
1074 
1075 	vfree(cq->queue);
1076 }
1077 
1078 /*
1079  * siw_create_cq()
1080  *
1081  * Populate CQ of requested size
1082  *
1083  * @base_cq: CQ as allocated by RDMA midlayer
1084  * @attr: Initial CQ attributes
1085  * @udata: relates to user context
1086  */
1087 
1088 int siw_create_cq(struct ib_cq *base_cq, const struct ib_cq_init_attr *attr,
1089 		  struct ib_udata *udata)
1090 {
1091 	struct siw_device *sdev = to_siw_dev(base_cq->device);
1092 	struct siw_cq *cq = to_siw_cq(base_cq);
1093 	int rv, size = attr->cqe;
1094 
1095 	if (atomic_inc_return(&sdev->num_cq) > SIW_MAX_CQ) {
1096 		siw_dbg(base_cq->device, "too many CQ's\n");
1097 		rv = -ENOMEM;
1098 		goto err_out;
1099 	}
1100 	if (size < 1 || size > sdev->attrs.max_cqe) {
1101 		siw_dbg(base_cq->device, "CQ size error: %d\n", size);
1102 		rv = -EINVAL;
1103 		goto err_out;
1104 	}
1105 	size = roundup_pow_of_two(size);
1106 	cq->base_cq.cqe = size;
1107 	cq->num_cqe = size;
1108 
1109 	if (udata)
1110 		cq->queue = vmalloc_user(size * sizeof(struct siw_cqe) +
1111 					 sizeof(struct siw_cq_ctrl));
1112 	else
1113 		cq->queue = vzalloc(size * sizeof(struct siw_cqe) +
1114 				    sizeof(struct siw_cq_ctrl));
1115 
1116 	if (cq->queue == NULL) {
1117 		rv = -ENOMEM;
1118 		goto err_out;
1119 	}
1120 	get_random_bytes(&cq->id, 4);
1121 	siw_dbg(base_cq->device, "new CQ [%u]\n", cq->id);
1122 
1123 	spin_lock_init(&cq->lock);
1124 
1125 	cq->notify = (struct siw_cq_ctrl *)&cq->queue[size];
1126 
1127 	if (udata) {
1128 		struct siw_uresp_create_cq uresp = {};
1129 		struct siw_ucontext *ctx =
1130 			rdma_udata_to_drv_context(udata, struct siw_ucontext,
1131 						  base_ucontext);
1132 		size_t length = size * sizeof(struct siw_cqe) +
1133 			sizeof(struct siw_cq_ctrl);
1134 
1135 		cq->cq_entry =
1136 			siw_mmap_entry_insert(ctx, cq->queue,
1137 					      length, &uresp.cq_key);
1138 		if (!cq->cq_entry) {
1139 			rv = -ENOMEM;
1140 			goto err_out;
1141 		}
1142 
1143 		uresp.cq_id = cq->id;
1144 		uresp.num_cqe = size;
1145 
1146 		if (udata->outlen < sizeof(uresp)) {
1147 			rv = -EINVAL;
1148 			goto err_out;
1149 		}
1150 		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1151 		if (rv)
1152 			goto err_out;
1153 	}
1154 	return 0;
1155 
1156 err_out:
1157 	siw_dbg(base_cq->device, "CQ creation failed: %d", rv);
1158 
1159 	if (cq && cq->queue) {
1160 		struct siw_ucontext *ctx =
1161 			rdma_udata_to_drv_context(udata, struct siw_ucontext,
1162 						  base_ucontext);
1163 		if (ctx)
1164 			rdma_user_mmap_entry_remove(cq->cq_entry);
1165 		vfree(cq->queue);
1166 	}
1167 	atomic_dec(&sdev->num_cq);
1168 
1169 	return rv;
1170 }
1171 
1172 /*
1173  * siw_poll_cq()
1174  *
1175  * Reap CQ entries if available and copy work completion status into
1176  * array of WC's provided by caller. Returns number of reaped CQE's.
1177  *
1178  * @base_cq:	Base CQ contained in siw CQ.
1179  * @num_cqe:	Maximum number of CQE's to reap.
1180  * @wc:		Array of work completions to be filled by siw.
1181  */
1182 int siw_poll_cq(struct ib_cq *base_cq, int num_cqe, struct ib_wc *wc)
1183 {
1184 	struct siw_cq *cq = to_siw_cq(base_cq);
1185 	int i;
1186 
1187 	for (i = 0; i < num_cqe; i++) {
1188 		if (!siw_reap_cqe(cq, wc))
1189 			break;
1190 		wc++;
1191 	}
1192 	return i;
1193 }
1194 
1195 /*
1196  * siw_req_notify_cq()
1197  *
1198  * Request notification for new CQE's added to that CQ.
1199  * Defined flags:
1200  * o SIW_CQ_NOTIFY_SOLICITED lets siw trigger a notification
1201  *   event if a WQE with notification flag set enters the CQ
1202  * o SIW_CQ_NOTIFY_NEXT_COMP lets siw trigger a notification
1203  *   event if a WQE enters the CQ.
1204  * o IB_CQ_REPORT_MISSED_EVENTS: return value will provide the
1205  *   number of not reaped CQE's regardless of its notification
1206  *   type and current or new CQ notification settings.
1207  *
1208  * @base_cq:	Base CQ contained in siw CQ.
1209  * @flags:	Requested notification flags.
1210  */
1211 int siw_req_notify_cq(struct ib_cq *base_cq, enum ib_cq_notify_flags flags)
1212 {
1213 	struct siw_cq *cq = to_siw_cq(base_cq);
1214 
1215 	siw_dbg_cq(cq, "flags: 0x%02x\n", flags);
1216 
1217 	if ((flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED)
1218 		/*
1219 		 * Enable CQ event for next solicited completion.
1220 		 * and make it visible to all associated producers.
1221 		 */
1222 		smp_store_mb(cq->notify->flags, SIW_NOTIFY_SOLICITED);
1223 	else
1224 		/*
1225 		 * Enable CQ event for any signalled completion.
1226 		 * and make it visible to all associated producers.
1227 		 */
1228 		smp_store_mb(cq->notify->flags, SIW_NOTIFY_ALL);
1229 
1230 	if (flags & IB_CQ_REPORT_MISSED_EVENTS)
1231 		return cq->cq_put - cq->cq_get;
1232 
1233 	return 0;
1234 }
1235 
1236 /*
1237  * siw_dereg_mr()
1238  *
1239  * Release Memory Region.
1240  *
1241  * @base_mr: Base MR contained in siw MR.
1242  * @udata: points to user context, unused.
1243  */
1244 int siw_dereg_mr(struct ib_mr *base_mr, struct ib_udata *udata)
1245 {
1246 	struct siw_mr *mr = to_siw_mr(base_mr);
1247 	struct siw_device *sdev = to_siw_dev(base_mr->device);
1248 
1249 	siw_dbg_mem(mr->mem, "deregister MR\n");
1250 
1251 	atomic_dec(&sdev->num_mr);
1252 
1253 	siw_mr_drop_mem(mr);
1254 	kfree_rcu(mr, rcu);
1255 
1256 	return 0;
1257 }
1258 
1259 /*
1260  * siw_reg_user_mr()
1261  *
1262  * Register Memory Region.
1263  *
1264  * @pd:		Protection Domain
1265  * @start:	starting address of MR (virtual address)
1266  * @len:	len of MR
1267  * @rnic_va:	not used by siw
1268  * @rights:	MR access rights
1269  * @udata:	user buffer to communicate STag and Key.
1270  */
1271 struct ib_mr *siw_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,
1272 			      u64 rnic_va, int rights, struct ib_udata *udata)
1273 {
1274 	struct siw_mr *mr = NULL;
1275 	struct siw_umem *umem = NULL;
1276 	struct siw_ureq_reg_mr ureq;
1277 	struct siw_device *sdev = to_siw_dev(pd->device);
1278 
1279 	unsigned long mem_limit = rlimit(RLIMIT_MEMLOCK);
1280 	int rv;
1281 
1282 	siw_dbg_pd(pd, "start: 0x%pK, va: 0x%pK, len: %llu\n",
1283 		   (void *)(uintptr_t)start, (void *)(uintptr_t)rnic_va,
1284 		   (unsigned long long)len);
1285 
1286 	if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1287 		siw_dbg_pd(pd, "too many mr's\n");
1288 		rv = -ENOMEM;
1289 		goto err_out;
1290 	}
1291 	if (!len) {
1292 		rv = -EINVAL;
1293 		goto err_out;
1294 	}
1295 	if (mem_limit != RLIM_INFINITY) {
1296 		unsigned long num_pages =
1297 			(PAGE_ALIGN(len + (start & ~PAGE_MASK))) >> PAGE_SHIFT;
1298 		mem_limit >>= PAGE_SHIFT;
1299 
1300 		if (num_pages > mem_limit - current->mm->locked_vm) {
1301 			siw_dbg_pd(pd, "pages req %lu, max %lu, lock %lu\n",
1302 				   num_pages, mem_limit,
1303 				   current->mm->locked_vm);
1304 			rv = -ENOMEM;
1305 			goto err_out;
1306 		}
1307 	}
1308 	umem = siw_umem_get(start, len, ib_access_writable(rights));
1309 	if (IS_ERR(umem)) {
1310 		rv = PTR_ERR(umem);
1311 		siw_dbg_pd(pd, "getting user memory failed: %d\n", rv);
1312 		umem = NULL;
1313 		goto err_out;
1314 	}
1315 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1316 	if (!mr) {
1317 		rv = -ENOMEM;
1318 		goto err_out;
1319 	}
1320 	rv = siw_mr_add_mem(mr, pd, umem, start, len, rights);
1321 	if (rv)
1322 		goto err_out;
1323 
1324 	if (udata) {
1325 		struct siw_uresp_reg_mr uresp = {};
1326 		struct siw_mem *mem = mr->mem;
1327 
1328 		if (udata->inlen < sizeof(ureq)) {
1329 			rv = -EINVAL;
1330 			goto err_out;
1331 		}
1332 		rv = ib_copy_from_udata(&ureq, udata, sizeof(ureq));
1333 		if (rv)
1334 			goto err_out;
1335 
1336 		mr->base_mr.lkey |= ureq.stag_key;
1337 		mr->base_mr.rkey |= ureq.stag_key;
1338 		mem->stag |= ureq.stag_key;
1339 		uresp.stag = mem->stag;
1340 
1341 		if (udata->outlen < sizeof(uresp)) {
1342 			rv = -EINVAL;
1343 			goto err_out;
1344 		}
1345 		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1346 		if (rv)
1347 			goto err_out;
1348 	}
1349 	mr->mem->stag_valid = 1;
1350 
1351 	return &mr->base_mr;
1352 
1353 err_out:
1354 	atomic_dec(&sdev->num_mr);
1355 	if (mr) {
1356 		if (mr->mem)
1357 			siw_mr_drop_mem(mr);
1358 		kfree_rcu(mr, rcu);
1359 	} else {
1360 		if (umem)
1361 			siw_umem_release(umem, false);
1362 	}
1363 	return ERR_PTR(rv);
1364 }
1365 
1366 struct ib_mr *siw_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
1367 			   u32 max_sge)
1368 {
1369 	struct siw_device *sdev = to_siw_dev(pd->device);
1370 	struct siw_mr *mr = NULL;
1371 	struct siw_pbl *pbl = NULL;
1372 	int rv;
1373 
1374 	if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1375 		siw_dbg_pd(pd, "too many mr's\n");
1376 		rv = -ENOMEM;
1377 		goto err_out;
1378 	}
1379 	if (mr_type != IB_MR_TYPE_MEM_REG) {
1380 		siw_dbg_pd(pd, "mr type %d unsupported\n", mr_type);
1381 		rv = -EOPNOTSUPP;
1382 		goto err_out;
1383 	}
1384 	if (max_sge > SIW_MAX_SGE_PBL) {
1385 		siw_dbg_pd(pd, "too many sge's: %d\n", max_sge);
1386 		rv = -ENOMEM;
1387 		goto err_out;
1388 	}
1389 	pbl = siw_pbl_alloc(max_sge);
1390 	if (IS_ERR(pbl)) {
1391 		rv = PTR_ERR(pbl);
1392 		siw_dbg_pd(pd, "pbl allocation failed: %d\n", rv);
1393 		pbl = NULL;
1394 		goto err_out;
1395 	}
1396 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1397 	if (!mr) {
1398 		rv = -ENOMEM;
1399 		goto err_out;
1400 	}
1401 	rv = siw_mr_add_mem(mr, pd, pbl, 0, max_sge * PAGE_SIZE, 0);
1402 	if (rv)
1403 		goto err_out;
1404 
1405 	mr->mem->is_pbl = 1;
1406 
1407 	siw_dbg_pd(pd, "[MEM %u]: success\n", mr->mem->stag);
1408 
1409 	return &mr->base_mr;
1410 
1411 err_out:
1412 	atomic_dec(&sdev->num_mr);
1413 
1414 	if (!mr) {
1415 		kfree(pbl);
1416 	} else {
1417 		if (mr->mem)
1418 			siw_mr_drop_mem(mr);
1419 		kfree_rcu(mr, rcu);
1420 	}
1421 	siw_dbg_pd(pd, "failed: %d\n", rv);
1422 
1423 	return ERR_PTR(rv);
1424 }
1425 
1426 /* Just used to count number of pages being mapped */
1427 static int siw_set_pbl_page(struct ib_mr *base_mr, u64 buf_addr)
1428 {
1429 	return 0;
1430 }
1431 
1432 int siw_map_mr_sg(struct ib_mr *base_mr, struct scatterlist *sl, int num_sle,
1433 		  unsigned int *sg_off)
1434 {
1435 	struct scatterlist *slp;
1436 	struct siw_mr *mr = to_siw_mr(base_mr);
1437 	struct siw_mem *mem = mr->mem;
1438 	struct siw_pbl *pbl = mem->pbl;
1439 	struct siw_pble *pble;
1440 	unsigned long pbl_size;
1441 	int i, rv;
1442 
1443 	if (!pbl) {
1444 		siw_dbg_mem(mem, "no PBL allocated\n");
1445 		return -EINVAL;
1446 	}
1447 	pble = pbl->pbe;
1448 
1449 	if (pbl->max_buf < num_sle) {
1450 		siw_dbg_mem(mem, "too many SGE's: %d > %d\n",
1451 			    mem->pbl->max_buf, num_sle);
1452 		return -ENOMEM;
1453 	}
1454 	for_each_sg(sl, slp, num_sle, i) {
1455 		if (sg_dma_len(slp) == 0) {
1456 			siw_dbg_mem(mem, "empty SGE\n");
1457 			return -EINVAL;
1458 		}
1459 		if (i == 0) {
1460 			pble->addr = sg_dma_address(slp);
1461 			pble->size = sg_dma_len(slp);
1462 			pble->pbl_off = 0;
1463 			pbl_size = pble->size;
1464 			pbl->num_buf = 1;
1465 		} else {
1466 			/* Merge PBL entries if adjacent */
1467 			if (pble->addr + pble->size == sg_dma_address(slp)) {
1468 				pble->size += sg_dma_len(slp);
1469 			} else {
1470 				pble++;
1471 				pbl->num_buf++;
1472 				pble->addr = sg_dma_address(slp);
1473 				pble->size = sg_dma_len(slp);
1474 				pble->pbl_off = pbl_size;
1475 			}
1476 			pbl_size += sg_dma_len(slp);
1477 		}
1478 		siw_dbg_mem(mem,
1479 			"sge[%d], size %u, addr 0x%p, total %lu\n",
1480 			i, pble->size, (void *)(uintptr_t)pble->addr,
1481 			pbl_size);
1482 	}
1483 	rv = ib_sg_to_pages(base_mr, sl, num_sle, sg_off, siw_set_pbl_page);
1484 	if (rv > 0) {
1485 		mem->len = base_mr->length;
1486 		mem->va = base_mr->iova;
1487 		siw_dbg_mem(mem,
1488 			"%llu bytes, start 0x%pK, %u SLE to %u entries\n",
1489 			mem->len, (void *)(uintptr_t)mem->va, num_sle,
1490 			pbl->num_buf);
1491 	}
1492 	return rv;
1493 }
1494 
1495 /*
1496  * siw_get_dma_mr()
1497  *
1498  * Create a (empty) DMA memory region, where no umem is attached.
1499  */
1500 struct ib_mr *siw_get_dma_mr(struct ib_pd *pd, int rights)
1501 {
1502 	struct siw_device *sdev = to_siw_dev(pd->device);
1503 	struct siw_mr *mr = NULL;
1504 	int rv;
1505 
1506 	if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1507 		siw_dbg_pd(pd, "too many mr's\n");
1508 		rv = -ENOMEM;
1509 		goto err_out;
1510 	}
1511 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1512 	if (!mr) {
1513 		rv = -ENOMEM;
1514 		goto err_out;
1515 	}
1516 	rv = siw_mr_add_mem(mr, pd, NULL, 0, ULONG_MAX, rights);
1517 	if (rv)
1518 		goto err_out;
1519 
1520 	mr->mem->stag_valid = 1;
1521 
1522 	siw_dbg_pd(pd, "[MEM %u]: success\n", mr->mem->stag);
1523 
1524 	return &mr->base_mr;
1525 
1526 err_out:
1527 	if (rv)
1528 		kfree(mr);
1529 
1530 	atomic_dec(&sdev->num_mr);
1531 
1532 	return ERR_PTR(rv);
1533 }
1534 
1535 /*
1536  * siw_create_srq()
1537  *
1538  * Create Shared Receive Queue of attributes @init_attrs
1539  * within protection domain given by @pd.
1540  *
1541  * @base_srq:	Base SRQ contained in siw SRQ.
1542  * @init_attrs:	SRQ init attributes.
1543  * @udata:	points to user context
1544  */
1545 int siw_create_srq(struct ib_srq *base_srq,
1546 		   struct ib_srq_init_attr *init_attrs, struct ib_udata *udata)
1547 {
1548 	struct siw_srq *srq = to_siw_srq(base_srq);
1549 	struct ib_srq_attr *attrs = &init_attrs->attr;
1550 	struct siw_device *sdev = to_siw_dev(base_srq->device);
1551 	struct siw_ucontext *ctx =
1552 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
1553 					  base_ucontext);
1554 	int rv;
1555 
1556 	if (atomic_inc_return(&sdev->num_srq) > SIW_MAX_SRQ) {
1557 		siw_dbg_pd(base_srq->pd, "too many SRQ's\n");
1558 		rv = -ENOMEM;
1559 		goto err_out;
1560 	}
1561 	if (attrs->max_wr == 0 || attrs->max_wr > SIW_MAX_SRQ_WR ||
1562 	    attrs->max_sge > SIW_MAX_SGE || attrs->srq_limit > attrs->max_wr) {
1563 		rv = -EINVAL;
1564 		goto err_out;
1565 	}
1566 	srq->max_sge = attrs->max_sge;
1567 	srq->num_rqe = roundup_pow_of_two(attrs->max_wr);
1568 	srq->limit = attrs->srq_limit;
1569 	if (srq->limit)
1570 		srq->armed = true;
1571 
1572 	srq->is_kernel_res = !udata;
1573 
1574 	if (udata)
1575 		srq->recvq =
1576 			vmalloc_user(srq->num_rqe * sizeof(struct siw_rqe));
1577 	else
1578 		srq->recvq = vzalloc(srq->num_rqe * sizeof(struct siw_rqe));
1579 
1580 	if (srq->recvq == NULL) {
1581 		rv = -ENOMEM;
1582 		goto err_out;
1583 	}
1584 	if (udata) {
1585 		struct siw_uresp_create_srq uresp = {};
1586 		size_t length = srq->num_rqe * sizeof(struct siw_rqe);
1587 
1588 		srq->srq_entry =
1589 			siw_mmap_entry_insert(ctx, srq->recvq,
1590 					      length, &uresp.srq_key);
1591 		if (!srq->srq_entry) {
1592 			rv = -ENOMEM;
1593 			goto err_out;
1594 		}
1595 
1596 		uresp.num_rqe = srq->num_rqe;
1597 
1598 		if (udata->outlen < sizeof(uresp)) {
1599 			rv = -EINVAL;
1600 			goto err_out;
1601 		}
1602 		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1603 		if (rv)
1604 			goto err_out;
1605 	}
1606 	spin_lock_init(&srq->lock);
1607 
1608 	siw_dbg_pd(base_srq->pd, "[SRQ]: success\n");
1609 
1610 	return 0;
1611 
1612 err_out:
1613 	if (srq->recvq) {
1614 		if (ctx)
1615 			rdma_user_mmap_entry_remove(srq->srq_entry);
1616 		vfree(srq->recvq);
1617 	}
1618 	atomic_dec(&sdev->num_srq);
1619 
1620 	return rv;
1621 }
1622 
1623 /*
1624  * siw_modify_srq()
1625  *
1626  * Modify SRQ. The caller may resize SRQ and/or set/reset notification
1627  * limit and (re)arm IB_EVENT_SRQ_LIMIT_REACHED notification.
1628  *
1629  * NOTE: it is unclear if RDMA core allows for changing the MAX_SGE
1630  * parameter. siw_modify_srq() does not check the attrs->max_sge param.
1631  */
1632 int siw_modify_srq(struct ib_srq *base_srq, struct ib_srq_attr *attrs,
1633 		   enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
1634 {
1635 	struct siw_srq *srq = to_siw_srq(base_srq);
1636 	unsigned long flags;
1637 	int rv = 0;
1638 
1639 	spin_lock_irqsave(&srq->lock, flags);
1640 
1641 	if (attr_mask & IB_SRQ_MAX_WR) {
1642 		/* resize request not yet supported */
1643 		rv = -EOPNOTSUPP;
1644 		goto out;
1645 	}
1646 	if (attr_mask & IB_SRQ_LIMIT) {
1647 		if (attrs->srq_limit) {
1648 			if (unlikely(attrs->srq_limit > srq->num_rqe)) {
1649 				rv = -EINVAL;
1650 				goto out;
1651 			}
1652 			srq->armed = true;
1653 		} else {
1654 			srq->armed = false;
1655 		}
1656 		srq->limit = attrs->srq_limit;
1657 	}
1658 out:
1659 	spin_unlock_irqrestore(&srq->lock, flags);
1660 
1661 	return rv;
1662 }
1663 
1664 /*
1665  * siw_query_srq()
1666  *
1667  * Query SRQ attributes.
1668  */
1669 int siw_query_srq(struct ib_srq *base_srq, struct ib_srq_attr *attrs)
1670 {
1671 	struct siw_srq *srq = to_siw_srq(base_srq);
1672 	unsigned long flags;
1673 
1674 	spin_lock_irqsave(&srq->lock, flags);
1675 
1676 	attrs->max_wr = srq->num_rqe;
1677 	attrs->max_sge = srq->max_sge;
1678 	attrs->srq_limit = srq->limit;
1679 
1680 	spin_unlock_irqrestore(&srq->lock, flags);
1681 
1682 	return 0;
1683 }
1684 
1685 /*
1686  * siw_destroy_srq()
1687  *
1688  * Destroy SRQ.
1689  * It is assumed that the SRQ is not referenced by any
1690  * QP anymore - the code trusts the RDMA core environment to keep track
1691  * of QP references.
1692  */
1693 void siw_destroy_srq(struct ib_srq *base_srq, struct ib_udata *udata)
1694 {
1695 	struct siw_srq *srq = to_siw_srq(base_srq);
1696 	struct siw_device *sdev = to_siw_dev(base_srq->device);
1697 	struct siw_ucontext *ctx =
1698 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
1699 					  base_ucontext);
1700 
1701 	if (ctx)
1702 		rdma_user_mmap_entry_remove(srq->srq_entry);
1703 	vfree(srq->recvq);
1704 	atomic_dec(&sdev->num_srq);
1705 }
1706 
1707 /*
1708  * siw_post_srq_recv()
1709  *
1710  * Post a list of receive queue elements to SRQ.
1711  * NOTE: The function does not check or lock a certain SRQ state
1712  *       during the post operation. The code simply trusts the
1713  *       RDMA core environment.
1714  *
1715  * @base_srq:	Base SRQ contained in siw SRQ
1716  * @wr:		List of R-WR's
1717  * @bad_wr:	Updated to failing WR if posting fails.
1718  */
1719 int siw_post_srq_recv(struct ib_srq *base_srq, const struct ib_recv_wr *wr,
1720 		      const struct ib_recv_wr **bad_wr)
1721 {
1722 	struct siw_srq *srq = to_siw_srq(base_srq);
1723 	unsigned long flags;
1724 	int rv = 0;
1725 
1726 	if (unlikely(!srq->is_kernel_res)) {
1727 		siw_dbg_pd(base_srq->pd,
1728 			   "[SRQ]: no kernel post_recv for mapped srq\n");
1729 		rv = -EINVAL;
1730 		goto out;
1731 	}
1732 	/*
1733 	 * Serialize potentially multiple producers.
1734 	 * Also needed to serialize potentially multiple
1735 	 * consumers.
1736 	 */
1737 	spin_lock_irqsave(&srq->lock, flags);
1738 
1739 	while (wr) {
1740 		u32 idx = srq->rq_put % srq->num_rqe;
1741 		struct siw_rqe *rqe = &srq->recvq[idx];
1742 
1743 		if (rqe->flags) {
1744 			siw_dbg_pd(base_srq->pd, "SRQ full\n");
1745 			rv = -ENOMEM;
1746 			break;
1747 		}
1748 		if (unlikely(wr->num_sge > srq->max_sge)) {
1749 			siw_dbg_pd(base_srq->pd,
1750 				   "[SRQ]: too many sge's: %d\n", wr->num_sge);
1751 			rv = -EINVAL;
1752 			break;
1753 		}
1754 		rqe->id = wr->wr_id;
1755 		rqe->num_sge = wr->num_sge;
1756 		siw_copy_sgl(wr->sg_list, rqe->sge, wr->num_sge);
1757 
1758 		/* Make sure S-RQE is completely written before valid */
1759 		smp_wmb();
1760 
1761 		rqe->flags = SIW_WQE_VALID;
1762 
1763 		srq->rq_put++;
1764 		wr = wr->next;
1765 	}
1766 	spin_unlock_irqrestore(&srq->lock, flags);
1767 out:
1768 	if (unlikely(rv < 0)) {
1769 		siw_dbg_pd(base_srq->pd, "[SRQ]: error %d\n", rv);
1770 		*bad_wr = wr;
1771 	}
1772 	return rv;
1773 }
1774 
1775 void siw_qp_event(struct siw_qp *qp, enum ib_event_type etype)
1776 {
1777 	struct ib_event event;
1778 	struct ib_qp *base_qp = &qp->base_qp;
1779 
1780 	/*
1781 	 * Do not report asynchronous errors on QP which gets
1782 	 * destroyed via verbs interface (siw_destroy_qp())
1783 	 */
1784 	if (qp->attrs.flags & SIW_QP_IN_DESTROY)
1785 		return;
1786 
1787 	event.event = etype;
1788 	event.device = base_qp->device;
1789 	event.element.qp = base_qp;
1790 
1791 	if (base_qp->event_handler) {
1792 		siw_dbg_qp(qp, "reporting event %d\n", etype);
1793 		base_qp->event_handler(&event, base_qp->qp_context);
1794 	}
1795 }
1796 
1797 void siw_cq_event(struct siw_cq *cq, enum ib_event_type etype)
1798 {
1799 	struct ib_event event;
1800 	struct ib_cq *base_cq = &cq->base_cq;
1801 
1802 	event.event = etype;
1803 	event.device = base_cq->device;
1804 	event.element.cq = base_cq;
1805 
1806 	if (base_cq->event_handler) {
1807 		siw_dbg_cq(cq, "reporting CQ event %d\n", etype);
1808 		base_cq->event_handler(&event, base_cq->cq_context);
1809 	}
1810 }
1811 
1812 void siw_srq_event(struct siw_srq *srq, enum ib_event_type etype)
1813 {
1814 	struct ib_event event;
1815 	struct ib_srq *base_srq = &srq->base_srq;
1816 
1817 	event.event = etype;
1818 	event.device = base_srq->device;
1819 	event.element.srq = base_srq;
1820 
1821 	if (base_srq->event_handler) {
1822 		siw_dbg_pd(srq->base_srq.pd,
1823 			   "reporting SRQ event %d\n", etype);
1824 		base_srq->event_handler(&event, base_srq->srq_context);
1825 	}
1826 }
1827 
1828 void siw_port_event(struct siw_device *sdev, u8 port, enum ib_event_type etype)
1829 {
1830 	struct ib_event event;
1831 
1832 	event.event = etype;
1833 	event.device = &sdev->base_dev;
1834 	event.element.port_num = port;
1835 
1836 	siw_dbg(&sdev->base_dev, "reporting port event %d\n", etype);
1837 
1838 	ib_dispatch_event(&event);
1839 }
1840