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_fmr = sdev->attrs.max_fmr;
140 	attr->max_mr = sdev->attrs.max_mr;
141 	attr->max_mw = sdev->attrs.max_mw;
142 	attr->max_mr_size = ~0ull;
143 	attr->max_pd = sdev->attrs.max_pd;
144 	attr->max_qp = sdev->attrs.max_qp;
145 	attr->max_qp_init_rd_atom = sdev->attrs.max_ird;
146 	attr->max_qp_rd_atom = sdev->attrs.max_ord;
147 	attr->max_qp_wr = sdev->attrs.max_qp_wr;
148 	attr->max_recv_sge = sdev->attrs.max_sge;
149 	attr->max_res_rd_atom = sdev->attrs.max_qp * sdev->attrs.max_ird;
150 	attr->max_send_sge = sdev->attrs.max_sge;
151 	attr->max_sge_rd = sdev->attrs.max_sge_rd;
152 	attr->max_srq = sdev->attrs.max_srq;
153 	attr->max_srq_sge = sdev->attrs.max_srq_sge;
154 	attr->max_srq_wr = sdev->attrs.max_srq_wr;
155 	attr->page_size_cap = PAGE_SIZE;
156 	attr->vendor_id = SIW_VENDOR_ID;
157 	attr->vendor_part_id = sdev->vendor_part_id;
158 
159 	memcpy(&attr->sys_image_guid, sdev->netdev->dev_addr, 6);
160 
161 	return 0;
162 }
163 
164 int siw_query_port(struct ib_device *base_dev, u8 port,
165 		   struct ib_port_attr *attr)
166 {
167 	struct siw_device *sdev = to_siw_dev(base_dev);
168 
169 	memset(attr, 0, sizeof(*attr));
170 
171 	attr->active_mtu = attr->max_mtu;
172 	attr->active_speed = 2;
173 	attr->active_width = 2;
174 	attr->gid_tbl_len = 1;
175 	attr->max_msg_sz = -1;
176 	attr->max_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->pkey_tbl_len = 1;
180 	attr->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_DEVICE_MGMT_SUP;
181 	attr->state = sdev->state;
182 	/*
183 	 * All zero
184 	 *
185 	 * attr->lid = 0;
186 	 * attr->bad_pkey_cntr = 0;
187 	 * attr->qkey_viol_cntr = 0;
188 	 * attr->sm_lid = 0;
189 	 * attr->lmc = 0;
190 	 * attr->max_vl_num = 0;
191 	 * attr->sm_sl = 0;
192 	 * attr->subnet_timeout = 0;
193 	 * attr->init_type_repy = 0;
194 	 */
195 	return 0;
196 }
197 
198 int siw_get_port_immutable(struct ib_device *base_dev, u8 port,
199 			   struct ib_port_immutable *port_immutable)
200 {
201 	struct ib_port_attr attr;
202 	int rv = siw_query_port(base_dev, port, &attr);
203 
204 	if (rv)
205 		return rv;
206 
207 	port_immutable->pkey_tbl_len = attr.pkey_tbl_len;
208 	port_immutable->gid_tbl_len = attr.gid_tbl_len;
209 	port_immutable->core_cap_flags = RDMA_CORE_PORT_IWARP;
210 
211 	return 0;
212 }
213 
214 int siw_query_pkey(struct ib_device *base_dev, u8 port, u16 idx, u16 *pkey)
215 {
216 	/* Report the default pkey */
217 	*pkey = 0xffff;
218 	return 0;
219 }
220 
221 int siw_query_gid(struct ib_device *base_dev, u8 port, int idx,
222 		  union ib_gid *gid)
223 {
224 	struct siw_device *sdev = to_siw_dev(base_dev);
225 
226 	/* subnet_prefix == interface_id == 0; */
227 	memset(gid, 0, sizeof(*gid));
228 	memcpy(&gid->raw[0], sdev->netdev->dev_addr, 6);
229 
230 	return 0;
231 }
232 
233 int siw_alloc_pd(struct ib_pd *pd, struct ib_udata *udata)
234 {
235 	struct siw_device *sdev = to_siw_dev(pd->device);
236 
237 	if (atomic_inc_return(&sdev->num_pd) > SIW_MAX_PD) {
238 		atomic_dec(&sdev->num_pd);
239 		return -ENOMEM;
240 	}
241 	siw_dbg_pd(pd, "now %d PD's(s)\n", atomic_read(&sdev->num_pd));
242 
243 	return 0;
244 }
245 
246 void siw_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
247 {
248 	struct siw_device *sdev = to_siw_dev(pd->device);
249 
250 	siw_dbg_pd(pd, "free PD\n");
251 	atomic_dec(&sdev->num_pd);
252 }
253 
254 void siw_qp_get_ref(struct ib_qp *base_qp)
255 {
256 	siw_qp_get(to_siw_qp(base_qp));
257 }
258 
259 void siw_qp_put_ref(struct ib_qp *base_qp)
260 {
261 	siw_qp_put(to_siw_qp(base_qp));
262 }
263 
264 static struct rdma_user_mmap_entry *
265 siw_mmap_entry_insert(struct siw_ucontext *uctx,
266 		      void *address, size_t length,
267 		      u64 *offset)
268 {
269 	struct siw_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
270 	int rv;
271 
272 	*offset = SIW_INVAL_UOBJ_KEY;
273 	if (!entry)
274 		return NULL;
275 
276 	entry->address = address;
277 
278 	rv = rdma_user_mmap_entry_insert(&uctx->base_ucontext,
279 					 &entry->rdma_entry,
280 					 length);
281 	if (rv) {
282 		kfree(entry);
283 		return NULL;
284 	}
285 
286 	*offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
287 
288 	return &entry->rdma_entry;
289 }
290 
291 /*
292  * siw_create_qp()
293  *
294  * Create QP of requested size on given device.
295  *
296  * @pd:		Protection Domain
297  * @attrs:	Initial QP attributes.
298  * @udata:	used to provide QP ID, SQ and RQ size back to user.
299  */
300 
301 struct ib_qp *siw_create_qp(struct ib_pd *pd,
302 			    struct ib_qp_init_attr *attrs,
303 			    struct ib_udata *udata)
304 {
305 	struct siw_qp *qp = NULL;
306 	struct siw_base_qp *siw_base_qp = NULL;
307 	struct ib_device *base_dev = pd->device;
308 	struct siw_device *sdev = to_siw_dev(base_dev);
309 	struct siw_ucontext *uctx =
310 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
311 					  base_ucontext);
312 	struct siw_cq *scq = NULL, *rcq = NULL;
313 	unsigned long flags;
314 	int num_sqe, num_rqe, rv = 0;
315 	size_t length;
316 
317 	siw_dbg(base_dev, "create new QP\n");
318 
319 	if (atomic_inc_return(&sdev->num_qp) > SIW_MAX_QP) {
320 		siw_dbg(base_dev, "too many QP's\n");
321 		rv = -ENOMEM;
322 		goto err_out;
323 	}
324 	if (attrs->qp_type != IB_QPT_RC) {
325 		siw_dbg(base_dev, "only RC QP's supported\n");
326 		rv = -EINVAL;
327 		goto err_out;
328 	}
329 	if ((attrs->cap.max_send_wr > SIW_MAX_QP_WR) ||
330 	    (attrs->cap.max_recv_wr > SIW_MAX_QP_WR) ||
331 	    (attrs->cap.max_send_sge > SIW_MAX_SGE) ||
332 	    (attrs->cap.max_recv_sge > SIW_MAX_SGE)) {
333 		siw_dbg(base_dev, "QP size error\n");
334 		rv = -EINVAL;
335 		goto err_out;
336 	}
337 	if (attrs->cap.max_inline_data > SIW_MAX_INLINE) {
338 		siw_dbg(base_dev, "max inline send: %d > %d\n",
339 			attrs->cap.max_inline_data, (int)SIW_MAX_INLINE);
340 		rv = -EINVAL;
341 		goto err_out;
342 	}
343 	/*
344 	 * NOTE: we allow for zero element SQ and RQ WQE's SGL's
345 	 * but not for a QP unable to hold any WQE (SQ + RQ)
346 	 */
347 	if (attrs->cap.max_send_wr + attrs->cap.max_recv_wr == 0) {
348 		siw_dbg(base_dev, "QP must have send or receive queue\n");
349 		rv = -EINVAL;
350 		goto err_out;
351 	}
352 	scq = to_siw_cq(attrs->send_cq);
353 	rcq = to_siw_cq(attrs->recv_cq);
354 
355 	if (!scq || (!rcq && !attrs->srq)) {
356 		siw_dbg(base_dev, "send CQ or receive CQ invalid\n");
357 		rv = -EINVAL;
358 		goto err_out;
359 	}
360 	siw_base_qp = kzalloc(sizeof(*siw_base_qp), GFP_KERNEL);
361 	if (!siw_base_qp) {
362 		rv = -ENOMEM;
363 		goto err_out;
364 	}
365 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
366 	if (!qp) {
367 		rv = -ENOMEM;
368 		goto err_out;
369 	}
370 	siw_base_qp->qp = qp;
371 	qp->ib_qp = &siw_base_qp->base_qp;
372 
373 	init_rwsem(&qp->state_lock);
374 	spin_lock_init(&qp->sq_lock);
375 	spin_lock_init(&qp->rq_lock);
376 	spin_lock_init(&qp->orq_lock);
377 
378 	qp->kernel_verbs = !udata;
379 
380 	rv = siw_qp_add(sdev, qp);
381 	if (rv)
382 		goto err_out;
383 
384 	/* All queue indices are derived from modulo operations
385 	 * on a free running 'get' (consumer) and 'put' (producer)
386 	 * unsigned counter. Having queue sizes at power of two
387 	 * avoids handling counter wrap around.
388 	 */
389 	num_sqe = roundup_pow_of_two(attrs->cap.max_send_wr);
390 	num_rqe = roundup_pow_of_two(attrs->cap.max_recv_wr);
391 
392 	if (qp->kernel_verbs)
393 		qp->sendq = vzalloc(num_sqe * sizeof(struct siw_sqe));
394 	else
395 		qp->sendq = vmalloc_user(num_sqe * sizeof(struct siw_sqe));
396 
397 	if (qp->sendq == NULL) {
398 		siw_dbg(base_dev, "SQ size %d alloc failed\n", num_sqe);
399 		rv = -ENOMEM;
400 		goto err_out_xa;
401 	}
402 	if (attrs->sq_sig_type != IB_SIGNAL_REQ_WR) {
403 		if (attrs->sq_sig_type == IB_SIGNAL_ALL_WR)
404 			qp->attrs.flags |= SIW_SIGNAL_ALL_WR;
405 		else {
406 			rv = -EINVAL;
407 			goto err_out_xa;
408 		}
409 	}
410 	qp->pd = pd;
411 	qp->scq = scq;
412 	qp->rcq = rcq;
413 
414 	if (attrs->srq) {
415 		/*
416 		 * SRQ support.
417 		 * Verbs 6.3.7: ignore RQ size, if SRQ present
418 		 * Verbs 6.3.5: do not check PD of SRQ against PD of QP
419 		 */
420 		qp->srq = to_siw_srq(attrs->srq);
421 		qp->attrs.rq_size = 0;
422 		siw_dbg(base_dev, "QP [%u]: SRQ attached\n", qp->qp_num);
423 	} else if (num_rqe) {
424 		if (qp->kernel_verbs)
425 			qp->recvq = vzalloc(num_rqe * sizeof(struct siw_rqe));
426 		else
427 			qp->recvq =
428 				vmalloc_user(num_rqe * sizeof(struct siw_rqe));
429 
430 		if (qp->recvq == NULL) {
431 			siw_dbg(base_dev, "RQ size %d alloc failed\n", num_rqe);
432 			rv = -ENOMEM;
433 			goto err_out_xa;
434 		}
435 		qp->attrs.rq_size = num_rqe;
436 	}
437 	qp->attrs.sq_size = num_sqe;
438 	qp->attrs.sq_max_sges = attrs->cap.max_send_sge;
439 	qp->attrs.rq_max_sges = attrs->cap.max_recv_sge;
440 
441 	/* Make those two tunables fixed for now. */
442 	qp->tx_ctx.gso_seg_limit = 1;
443 	qp->tx_ctx.zcopy_tx = zcopy_tx;
444 
445 	qp->attrs.state = SIW_QP_STATE_IDLE;
446 
447 	if (udata) {
448 		struct siw_uresp_create_qp uresp = {};
449 
450 		uresp.num_sqe = num_sqe;
451 		uresp.num_rqe = num_rqe;
452 		uresp.qp_id = qp_id(qp);
453 
454 		if (qp->sendq) {
455 			length = num_sqe * sizeof(struct siw_sqe);
456 			qp->sq_entry =
457 				siw_mmap_entry_insert(uctx, qp->sendq,
458 						      length, &uresp.sq_key);
459 			if (!qp->sq_entry) {
460 				rv = -ENOMEM;
461 				goto err_out_xa;
462 			}
463 		}
464 
465 		if (qp->recvq) {
466 			length = num_rqe * sizeof(struct siw_rqe);
467 			qp->rq_entry =
468 				siw_mmap_entry_insert(uctx, qp->recvq,
469 						      length, &uresp.rq_key);
470 			if (!qp->rq_entry) {
471 				uresp.sq_key = SIW_INVAL_UOBJ_KEY;
472 				rv = -ENOMEM;
473 				goto err_out_xa;
474 			}
475 		}
476 
477 		if (udata->outlen < sizeof(uresp)) {
478 			rv = -EINVAL;
479 			goto err_out_xa;
480 		}
481 		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
482 		if (rv)
483 			goto err_out_xa;
484 	}
485 	qp->tx_cpu = siw_get_tx_cpu(sdev);
486 	if (qp->tx_cpu < 0) {
487 		rv = -EINVAL;
488 		goto err_out_xa;
489 	}
490 	INIT_LIST_HEAD(&qp->devq);
491 	spin_lock_irqsave(&sdev->lock, flags);
492 	list_add_tail(&qp->devq, &sdev->qp_list);
493 	spin_unlock_irqrestore(&sdev->lock, flags);
494 
495 	return qp->ib_qp;
496 
497 err_out_xa:
498 	xa_erase(&sdev->qp_xa, qp_id(qp));
499 err_out:
500 	kfree(siw_base_qp);
501 
502 	if (qp) {
503 		if (uctx) {
504 			rdma_user_mmap_entry_remove(qp->sq_entry);
505 			rdma_user_mmap_entry_remove(qp->rq_entry);
506 		}
507 		vfree(qp->sendq);
508 		vfree(qp->recvq);
509 		kfree(qp);
510 	}
511 	atomic_dec(&sdev->num_qp);
512 
513 	return ERR_PTR(rv);
514 }
515 
516 /*
517  * Minimum siw_query_qp() verb interface.
518  *
519  * @qp_attr_mask is not used but all available information is provided
520  */
521 int siw_query_qp(struct ib_qp *base_qp, struct ib_qp_attr *qp_attr,
522 		 int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
523 {
524 	struct siw_qp *qp;
525 	struct siw_device *sdev;
526 
527 	if (base_qp && qp_attr && qp_init_attr) {
528 		qp = to_siw_qp(base_qp);
529 		sdev = to_siw_dev(base_qp->device);
530 	} else {
531 		return -EINVAL;
532 	}
533 	qp_attr->cap.max_inline_data = SIW_MAX_INLINE;
534 	qp_attr->cap.max_send_wr = qp->attrs.sq_size;
535 	qp_attr->cap.max_send_sge = qp->attrs.sq_max_sges;
536 	qp_attr->cap.max_recv_wr = qp->attrs.rq_size;
537 	qp_attr->cap.max_recv_sge = qp->attrs.rq_max_sges;
538 	qp_attr->path_mtu = ib_mtu_int_to_enum(sdev->netdev->mtu);
539 	qp_attr->max_rd_atomic = qp->attrs.irq_size;
540 	qp_attr->max_dest_rd_atomic = qp->attrs.orq_size;
541 
542 	qp_attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE |
543 				   IB_ACCESS_REMOTE_WRITE |
544 				   IB_ACCESS_REMOTE_READ;
545 
546 	qp_init_attr->qp_type = base_qp->qp_type;
547 	qp_init_attr->send_cq = base_qp->send_cq;
548 	qp_init_attr->recv_cq = base_qp->recv_cq;
549 	qp_init_attr->srq = base_qp->srq;
550 
551 	qp_init_attr->cap = qp_attr->cap;
552 
553 	return 0;
554 }
555 
556 int siw_verbs_modify_qp(struct ib_qp *base_qp, struct ib_qp_attr *attr,
557 			int attr_mask, struct ib_udata *udata)
558 {
559 	struct siw_qp_attrs new_attrs;
560 	enum siw_qp_attr_mask siw_attr_mask = 0;
561 	struct siw_qp *qp = to_siw_qp(base_qp);
562 	int rv = 0;
563 
564 	if (!attr_mask)
565 		return 0;
566 
567 	memset(&new_attrs, 0, sizeof(new_attrs));
568 
569 	if (attr_mask & IB_QP_ACCESS_FLAGS) {
570 		siw_attr_mask = SIW_QP_ATTR_ACCESS_FLAGS;
571 
572 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
573 			new_attrs.flags |= SIW_RDMA_READ_ENABLED;
574 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
575 			new_attrs.flags |= SIW_RDMA_WRITE_ENABLED;
576 		if (attr->qp_access_flags & IB_ACCESS_MW_BIND)
577 			new_attrs.flags |= SIW_RDMA_BIND_ENABLED;
578 	}
579 	if (attr_mask & IB_QP_STATE) {
580 		siw_dbg_qp(qp, "desired IB QP state: %s\n",
581 			   ib_qp_state_to_string[attr->qp_state]);
582 
583 		new_attrs.state = ib_qp_state_to_siw_qp_state[attr->qp_state];
584 
585 		if (new_attrs.state > SIW_QP_STATE_RTS)
586 			qp->tx_ctx.tx_suspend = 1;
587 
588 		siw_attr_mask |= SIW_QP_ATTR_STATE;
589 	}
590 	if (!siw_attr_mask)
591 		goto out;
592 
593 	down_write(&qp->state_lock);
594 
595 	rv = siw_qp_modify(qp, &new_attrs, siw_attr_mask);
596 
597 	up_write(&qp->state_lock);
598 out:
599 	return rv;
600 }
601 
602 int siw_destroy_qp(struct ib_qp *base_qp, struct ib_udata *udata)
603 {
604 	struct siw_qp *qp = to_siw_qp(base_qp);
605 	struct siw_ucontext *uctx =
606 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
607 					  base_ucontext);
608 	struct siw_qp_attrs qp_attrs;
609 
610 	siw_dbg_qp(qp, "state %d\n", qp->attrs.state);
611 
612 	/*
613 	 * Mark QP as in process of destruction to prevent from
614 	 * any async callbacks to RDMA core
615 	 */
616 	qp->attrs.flags |= SIW_QP_IN_DESTROY;
617 	qp->rx_stream.rx_suspend = 1;
618 
619 	if (uctx) {
620 		rdma_user_mmap_entry_remove(qp->sq_entry);
621 		rdma_user_mmap_entry_remove(qp->rq_entry);
622 	}
623 
624 	down_write(&qp->state_lock);
625 
626 	qp_attrs.state = SIW_QP_STATE_ERROR;
627 	siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE);
628 
629 	if (qp->cep) {
630 		siw_cep_put(qp->cep);
631 		qp->cep = NULL;
632 	}
633 	up_write(&qp->state_lock);
634 
635 	kfree(qp->tx_ctx.mpa_crc_hd);
636 	kfree(qp->rx_stream.mpa_crc_hd);
637 
638 	qp->scq = qp->rcq = NULL;
639 
640 	siw_qp_put(qp);
641 
642 	return 0;
643 }
644 
645 /*
646  * siw_copy_inline_sgl()
647  *
648  * Prepare sgl of inlined data for sending. For userland callers
649  * function checks if given buffer addresses and len's are within
650  * process context bounds.
651  * Data from all provided sge's are copied together into the wqe,
652  * referenced by a single sge.
653  */
654 static int siw_copy_inline_sgl(const struct ib_send_wr *core_wr,
655 			       struct siw_sqe *sqe)
656 {
657 	struct ib_sge *core_sge = core_wr->sg_list;
658 	void *kbuf = &sqe->sge[1];
659 	int num_sge = core_wr->num_sge, bytes = 0;
660 
661 	sqe->sge[0].laddr = (uintptr_t)kbuf;
662 	sqe->sge[0].lkey = 0;
663 
664 	while (num_sge--) {
665 		if (!core_sge->length) {
666 			core_sge++;
667 			continue;
668 		}
669 		bytes += core_sge->length;
670 		if (bytes > SIW_MAX_INLINE) {
671 			bytes = -EINVAL;
672 			break;
673 		}
674 		memcpy(kbuf, (void *)(uintptr_t)core_sge->addr,
675 		       core_sge->length);
676 
677 		kbuf += core_sge->length;
678 		core_sge++;
679 	}
680 	sqe->sge[0].length = bytes > 0 ? bytes : 0;
681 	sqe->num_sge = bytes > 0 ? 1 : 0;
682 
683 	return bytes;
684 }
685 
686 /* Complete SQ WR's without processing */
687 static int siw_sq_flush_wr(struct siw_qp *qp, const struct ib_send_wr *wr,
688 			   const struct ib_send_wr **bad_wr)
689 {
690 	struct siw_sqe sqe = {};
691 	int rv = 0;
692 
693 	while (wr) {
694 		sqe.id = wr->wr_id;
695 		sqe.opcode = wr->opcode;
696 		rv = siw_sqe_complete(qp, &sqe, 0, SIW_WC_WR_FLUSH_ERR);
697 		if (rv) {
698 			if (bad_wr)
699 				*bad_wr = wr;
700 			break;
701 		}
702 		wr = wr->next;
703 	}
704 	return rv;
705 }
706 
707 /* Complete RQ WR's without processing */
708 static int siw_rq_flush_wr(struct siw_qp *qp, const struct ib_recv_wr *wr,
709 			   const struct ib_recv_wr **bad_wr)
710 {
711 	struct siw_rqe rqe = {};
712 	int rv = 0;
713 
714 	while (wr) {
715 		rqe.id = wr->wr_id;
716 		rv = siw_rqe_complete(qp, &rqe, 0, 0, SIW_WC_WR_FLUSH_ERR);
717 		if (rv) {
718 			if (bad_wr)
719 				*bad_wr = wr;
720 			break;
721 		}
722 		wr = wr->next;
723 	}
724 	return rv;
725 }
726 
727 /*
728  * siw_post_send()
729  *
730  * Post a list of S-WR's to a SQ.
731  *
732  * @base_qp:	Base QP contained in siw QP
733  * @wr:		Null terminated list of user WR's
734  * @bad_wr:	Points to failing WR in case of synchronous failure.
735  */
736 int siw_post_send(struct ib_qp *base_qp, const struct ib_send_wr *wr,
737 		  const struct ib_send_wr **bad_wr)
738 {
739 	struct siw_qp *qp = to_siw_qp(base_qp);
740 	struct siw_wqe *wqe = tx_wqe(qp);
741 
742 	unsigned long flags;
743 	int rv = 0;
744 
745 	if (wr && !qp->kernel_verbs) {
746 		siw_dbg_qp(qp, "wr must be empty for user mapped sq\n");
747 		*bad_wr = wr;
748 		return -EINVAL;
749 	}
750 
751 	/*
752 	 * Try to acquire QP state lock. Must be non-blocking
753 	 * to accommodate kernel clients needs.
754 	 */
755 	if (!down_read_trylock(&qp->state_lock)) {
756 		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
757 			/*
758 			 * ERROR state is final, so we can be sure
759 			 * this state will not change as long as the QP
760 			 * exists.
761 			 *
762 			 * This handles an ib_drain_sq() call with
763 			 * a concurrent request to set the QP state
764 			 * to ERROR.
765 			 */
766 			rv = siw_sq_flush_wr(qp, wr, bad_wr);
767 		} else {
768 			siw_dbg_qp(qp, "QP locked, state %d\n",
769 				   qp->attrs.state);
770 			*bad_wr = wr;
771 			rv = -ENOTCONN;
772 		}
773 		return rv;
774 	}
775 	if (unlikely(qp->attrs.state != SIW_QP_STATE_RTS)) {
776 		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
777 			/*
778 			 * Immediately flush this WR to CQ, if QP
779 			 * is in ERROR state. SQ is guaranteed to
780 			 * be empty, so WR complets in-order.
781 			 *
782 			 * Typically triggered by ib_drain_sq().
783 			 */
784 			rv = siw_sq_flush_wr(qp, wr, bad_wr);
785 		} else {
786 			siw_dbg_qp(qp, "QP out of state %d\n",
787 				   qp->attrs.state);
788 			*bad_wr = wr;
789 			rv = -ENOTCONN;
790 		}
791 		up_read(&qp->state_lock);
792 		return rv;
793 	}
794 	spin_lock_irqsave(&qp->sq_lock, flags);
795 
796 	while (wr) {
797 		u32 idx = qp->sq_put % qp->attrs.sq_size;
798 		struct siw_sqe *sqe = &qp->sendq[idx];
799 
800 		if (sqe->flags) {
801 			siw_dbg_qp(qp, "sq full\n");
802 			rv = -ENOMEM;
803 			break;
804 		}
805 		if (wr->num_sge > qp->attrs.sq_max_sges) {
806 			siw_dbg_qp(qp, "too many sge's: %d\n", wr->num_sge);
807 			rv = -EINVAL;
808 			break;
809 		}
810 		sqe->id = wr->wr_id;
811 
812 		if ((wr->send_flags & IB_SEND_SIGNALED) ||
813 		    (qp->attrs.flags & SIW_SIGNAL_ALL_WR))
814 			sqe->flags |= SIW_WQE_SIGNALLED;
815 
816 		if (wr->send_flags & IB_SEND_FENCE)
817 			sqe->flags |= SIW_WQE_READ_FENCE;
818 
819 		switch (wr->opcode) {
820 		case IB_WR_SEND:
821 		case IB_WR_SEND_WITH_INV:
822 			if (wr->send_flags & IB_SEND_SOLICITED)
823 				sqe->flags |= SIW_WQE_SOLICITED;
824 
825 			if (!(wr->send_flags & IB_SEND_INLINE)) {
826 				siw_copy_sgl(wr->sg_list, sqe->sge,
827 					     wr->num_sge);
828 				sqe->num_sge = wr->num_sge;
829 			} else {
830 				rv = siw_copy_inline_sgl(wr, sqe);
831 				if (rv <= 0) {
832 					rv = -EINVAL;
833 					break;
834 				}
835 				sqe->flags |= SIW_WQE_INLINE;
836 				sqe->num_sge = 1;
837 			}
838 			if (wr->opcode == IB_WR_SEND)
839 				sqe->opcode = SIW_OP_SEND;
840 			else {
841 				sqe->opcode = SIW_OP_SEND_REMOTE_INV;
842 				sqe->rkey = wr->ex.invalidate_rkey;
843 			}
844 			break;
845 
846 		case IB_WR_RDMA_READ_WITH_INV:
847 		case IB_WR_RDMA_READ:
848 			/*
849 			 * iWarp restricts RREAD sink to SGL containing
850 			 * 1 SGE only. we could relax to SGL with multiple
851 			 * elements referring the SAME ltag or even sending
852 			 * a private per-rreq tag referring to a checked
853 			 * local sgl with MULTIPLE ltag's.
854 			 */
855 			if (unlikely(wr->num_sge != 1)) {
856 				rv = -EINVAL;
857 				break;
858 			}
859 			siw_copy_sgl(wr->sg_list, &sqe->sge[0], 1);
860 			/*
861 			 * NOTE: zero length RREAD is allowed!
862 			 */
863 			sqe->raddr = rdma_wr(wr)->remote_addr;
864 			sqe->rkey = rdma_wr(wr)->rkey;
865 			sqe->num_sge = 1;
866 
867 			if (wr->opcode == IB_WR_RDMA_READ)
868 				sqe->opcode = SIW_OP_READ;
869 			else
870 				sqe->opcode = SIW_OP_READ_LOCAL_INV;
871 			break;
872 
873 		case IB_WR_RDMA_WRITE:
874 			if (!(wr->send_flags & IB_SEND_INLINE)) {
875 				siw_copy_sgl(wr->sg_list, &sqe->sge[0],
876 					     wr->num_sge);
877 				sqe->num_sge = wr->num_sge;
878 			} else {
879 				rv = siw_copy_inline_sgl(wr, sqe);
880 				if (unlikely(rv < 0)) {
881 					rv = -EINVAL;
882 					break;
883 				}
884 				sqe->flags |= SIW_WQE_INLINE;
885 				sqe->num_sge = 1;
886 			}
887 			sqe->raddr = rdma_wr(wr)->remote_addr;
888 			sqe->rkey = rdma_wr(wr)->rkey;
889 			sqe->opcode = SIW_OP_WRITE;
890 			break;
891 
892 		case IB_WR_REG_MR:
893 			sqe->base_mr = (uintptr_t)reg_wr(wr)->mr;
894 			sqe->rkey = reg_wr(wr)->key;
895 			sqe->access = reg_wr(wr)->access & IWARP_ACCESS_MASK;
896 			sqe->opcode = SIW_OP_REG_MR;
897 			break;
898 
899 		case IB_WR_LOCAL_INV:
900 			sqe->rkey = wr->ex.invalidate_rkey;
901 			sqe->opcode = SIW_OP_INVAL_STAG;
902 			break;
903 
904 		default:
905 			siw_dbg_qp(qp, "ib wr type %d unsupported\n",
906 				   wr->opcode);
907 			rv = -EINVAL;
908 			break;
909 		}
910 		siw_dbg_qp(qp, "opcode %d, flags 0x%x, wr_id 0x%pK\n",
911 			   sqe->opcode, sqe->flags,
912 			   (void *)(uintptr_t)sqe->id);
913 
914 		if (unlikely(rv < 0))
915 			break;
916 
917 		/* make SQE only valid after completely written */
918 		smp_wmb();
919 		sqe->flags |= SIW_WQE_VALID;
920 
921 		qp->sq_put++;
922 		wr = wr->next;
923 	}
924 
925 	/*
926 	 * Send directly if SQ processing is not in progress.
927 	 * Eventual immediate errors (rv < 0) do not affect the involved
928 	 * RI resources (Verbs, 8.3.1) and thus do not prevent from SQ
929 	 * processing, if new work is already pending. But rv must be passed
930 	 * to caller.
931 	 */
932 	if (wqe->wr_status != SIW_WR_IDLE) {
933 		spin_unlock_irqrestore(&qp->sq_lock, flags);
934 		goto skip_direct_sending;
935 	}
936 	rv = siw_activate_tx(qp);
937 	spin_unlock_irqrestore(&qp->sq_lock, flags);
938 
939 	if (rv <= 0)
940 		goto skip_direct_sending;
941 
942 	if (qp->kernel_verbs) {
943 		rv = siw_sq_start(qp);
944 	} else {
945 		qp->tx_ctx.in_syscall = 1;
946 
947 		if (siw_qp_sq_process(qp) != 0 && !(qp->tx_ctx.tx_suspend))
948 			siw_qp_cm_drop(qp, 0);
949 
950 		qp->tx_ctx.in_syscall = 0;
951 	}
952 skip_direct_sending:
953 
954 	up_read(&qp->state_lock);
955 
956 	if (rv >= 0)
957 		return 0;
958 	/*
959 	 * Immediate error
960 	 */
961 	siw_dbg_qp(qp, "error %d\n", rv);
962 
963 	*bad_wr = wr;
964 	return rv;
965 }
966 
967 /*
968  * siw_post_receive()
969  *
970  * Post a list of R-WR's to a RQ.
971  *
972  * @base_qp:	Base QP contained in siw QP
973  * @wr:		Null terminated list of user WR's
974  * @bad_wr:	Points to failing WR in case of synchronous failure.
975  */
976 int siw_post_receive(struct ib_qp *base_qp, const struct ib_recv_wr *wr,
977 		     const struct ib_recv_wr **bad_wr)
978 {
979 	struct siw_qp *qp = to_siw_qp(base_qp);
980 	unsigned long flags;
981 	int rv = 0;
982 
983 	if (qp->srq) {
984 		*bad_wr = wr;
985 		return -EOPNOTSUPP; /* what else from errno.h? */
986 	}
987 	if (!qp->kernel_verbs) {
988 		siw_dbg_qp(qp, "no kernel post_recv for user mapped sq\n");
989 		*bad_wr = wr;
990 		return -EINVAL;
991 	}
992 
993 	/*
994 	 * Try to acquire QP state lock. Must be non-blocking
995 	 * to accommodate kernel clients needs.
996 	 */
997 	if (!down_read_trylock(&qp->state_lock)) {
998 		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
999 			/*
1000 			 * ERROR state is final, so we can be sure
1001 			 * this state will not change as long as the QP
1002 			 * exists.
1003 			 *
1004 			 * This handles an ib_drain_rq() call with
1005 			 * a concurrent request to set the QP state
1006 			 * to ERROR.
1007 			 */
1008 			rv = siw_rq_flush_wr(qp, wr, bad_wr);
1009 		} else {
1010 			siw_dbg_qp(qp, "QP locked, state %d\n",
1011 				   qp->attrs.state);
1012 			*bad_wr = wr;
1013 			rv = -ENOTCONN;
1014 		}
1015 		return rv;
1016 	}
1017 	if (qp->attrs.state > SIW_QP_STATE_RTS) {
1018 		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
1019 			/*
1020 			 * Immediately flush this WR to CQ, if QP
1021 			 * is in ERROR state. RQ is guaranteed to
1022 			 * be empty, so WR complets in-order.
1023 			 *
1024 			 * Typically triggered by ib_drain_rq().
1025 			 */
1026 			rv = siw_rq_flush_wr(qp, wr, bad_wr);
1027 		} else {
1028 			siw_dbg_qp(qp, "QP out of state %d\n",
1029 				   qp->attrs.state);
1030 			*bad_wr = wr;
1031 			rv = -ENOTCONN;
1032 		}
1033 		up_read(&qp->state_lock);
1034 		return rv;
1035 	}
1036 	/*
1037 	 * Serialize potentially multiple producers.
1038 	 * Not needed for single threaded consumer side.
1039 	 */
1040 	spin_lock_irqsave(&qp->rq_lock, flags);
1041 
1042 	while (wr) {
1043 		u32 idx = qp->rq_put % qp->attrs.rq_size;
1044 		struct siw_rqe *rqe = &qp->recvq[idx];
1045 
1046 		if (rqe->flags) {
1047 			siw_dbg_qp(qp, "RQ full\n");
1048 			rv = -ENOMEM;
1049 			break;
1050 		}
1051 		if (wr->num_sge > qp->attrs.rq_max_sges) {
1052 			siw_dbg_qp(qp, "too many sge's: %d\n", wr->num_sge);
1053 			rv = -EINVAL;
1054 			break;
1055 		}
1056 		rqe->id = wr->wr_id;
1057 		rqe->num_sge = wr->num_sge;
1058 		siw_copy_sgl(wr->sg_list, rqe->sge, wr->num_sge);
1059 
1060 		/* make sure RQE is completely written before valid */
1061 		smp_wmb();
1062 
1063 		rqe->flags = SIW_WQE_VALID;
1064 
1065 		qp->rq_put++;
1066 		wr = wr->next;
1067 	}
1068 	spin_unlock_irqrestore(&qp->rq_lock, flags);
1069 
1070 	up_read(&qp->state_lock);
1071 
1072 	if (rv < 0) {
1073 		siw_dbg_qp(qp, "error %d\n", rv);
1074 		*bad_wr = wr;
1075 	}
1076 	return rv > 0 ? 0 : rv;
1077 }
1078 
1079 void siw_destroy_cq(struct ib_cq *base_cq, struct ib_udata *udata)
1080 {
1081 	struct siw_cq *cq = to_siw_cq(base_cq);
1082 	struct siw_device *sdev = to_siw_dev(base_cq->device);
1083 	struct siw_ucontext *ctx =
1084 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
1085 					  base_ucontext);
1086 
1087 	siw_dbg_cq(cq, "free CQ resources\n");
1088 
1089 	siw_cq_flush(cq);
1090 
1091 	if (ctx)
1092 		rdma_user_mmap_entry_remove(cq->cq_entry);
1093 
1094 	atomic_dec(&sdev->num_cq);
1095 
1096 	vfree(cq->queue);
1097 }
1098 
1099 /*
1100  * siw_create_cq()
1101  *
1102  * Populate CQ of requested size
1103  *
1104  * @base_cq: CQ as allocated by RDMA midlayer
1105  * @attr: Initial CQ attributes
1106  * @udata: relates to user context
1107  */
1108 
1109 int siw_create_cq(struct ib_cq *base_cq, const struct ib_cq_init_attr *attr,
1110 		  struct ib_udata *udata)
1111 {
1112 	struct siw_device *sdev = to_siw_dev(base_cq->device);
1113 	struct siw_cq *cq = to_siw_cq(base_cq);
1114 	int rv, size = attr->cqe;
1115 
1116 	if (atomic_inc_return(&sdev->num_cq) > SIW_MAX_CQ) {
1117 		siw_dbg(base_cq->device, "too many CQ's\n");
1118 		rv = -ENOMEM;
1119 		goto err_out;
1120 	}
1121 	if (size < 1 || size > sdev->attrs.max_cqe) {
1122 		siw_dbg(base_cq->device, "CQ size error: %d\n", size);
1123 		rv = -EINVAL;
1124 		goto err_out;
1125 	}
1126 	size = roundup_pow_of_two(size);
1127 	cq->base_cq.cqe = size;
1128 	cq->num_cqe = size;
1129 
1130 	if (!udata) {
1131 		cq->kernel_verbs = 1;
1132 		cq->queue = vzalloc(size * sizeof(struct siw_cqe) +
1133 				    sizeof(struct siw_cq_ctrl));
1134 	} else {
1135 		cq->queue = vmalloc_user(size * sizeof(struct siw_cqe) +
1136 					 sizeof(struct siw_cq_ctrl));
1137 	}
1138 	if (cq->queue == NULL) {
1139 		rv = -ENOMEM;
1140 		goto err_out;
1141 	}
1142 	get_random_bytes(&cq->id, 4);
1143 	siw_dbg(base_cq->device, "new CQ [%u]\n", cq->id);
1144 
1145 	spin_lock_init(&cq->lock);
1146 
1147 	cq->notify = (struct siw_cq_ctrl *)&cq->queue[size];
1148 
1149 	if (udata) {
1150 		struct siw_uresp_create_cq uresp = {};
1151 		struct siw_ucontext *ctx =
1152 			rdma_udata_to_drv_context(udata, struct siw_ucontext,
1153 						  base_ucontext);
1154 		size_t length = size * sizeof(struct siw_cqe) +
1155 			sizeof(struct siw_cq_ctrl);
1156 
1157 		cq->cq_entry =
1158 			siw_mmap_entry_insert(ctx, cq->queue,
1159 					      length, &uresp.cq_key);
1160 		if (!cq->cq_entry) {
1161 			rv = -ENOMEM;
1162 			goto err_out;
1163 		}
1164 
1165 		uresp.cq_id = cq->id;
1166 		uresp.num_cqe = size;
1167 
1168 		if (udata->outlen < sizeof(uresp)) {
1169 			rv = -EINVAL;
1170 			goto err_out;
1171 		}
1172 		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1173 		if (rv)
1174 			goto err_out;
1175 	}
1176 	return 0;
1177 
1178 err_out:
1179 	siw_dbg(base_cq->device, "CQ creation failed: %d", rv);
1180 
1181 	if (cq && cq->queue) {
1182 		struct siw_ucontext *ctx =
1183 			rdma_udata_to_drv_context(udata, struct siw_ucontext,
1184 						  base_ucontext);
1185 		if (ctx)
1186 			rdma_user_mmap_entry_remove(cq->cq_entry);
1187 		vfree(cq->queue);
1188 	}
1189 	atomic_dec(&sdev->num_cq);
1190 
1191 	return rv;
1192 }
1193 
1194 /*
1195  * siw_poll_cq()
1196  *
1197  * Reap CQ entries if available and copy work completion status into
1198  * array of WC's provided by caller. Returns number of reaped CQE's.
1199  *
1200  * @base_cq:	Base CQ contained in siw CQ.
1201  * @num_cqe:	Maximum number of CQE's to reap.
1202  * @wc:		Array of work completions to be filled by siw.
1203  */
1204 int siw_poll_cq(struct ib_cq *base_cq, int num_cqe, struct ib_wc *wc)
1205 {
1206 	struct siw_cq *cq = to_siw_cq(base_cq);
1207 	int i;
1208 
1209 	for (i = 0; i < num_cqe; i++) {
1210 		if (!siw_reap_cqe(cq, wc))
1211 			break;
1212 		wc++;
1213 	}
1214 	return i;
1215 }
1216 
1217 /*
1218  * siw_req_notify_cq()
1219  *
1220  * Request notification for new CQE's added to that CQ.
1221  * Defined flags:
1222  * o SIW_CQ_NOTIFY_SOLICITED lets siw trigger a notification
1223  *   event if a WQE with notification flag set enters the CQ
1224  * o SIW_CQ_NOTIFY_NEXT_COMP lets siw trigger a notification
1225  *   event if a WQE enters the CQ.
1226  * o IB_CQ_REPORT_MISSED_EVENTS: return value will provide the
1227  *   number of not reaped CQE's regardless of its notification
1228  *   type and current or new CQ notification settings.
1229  *
1230  * @base_cq:	Base CQ contained in siw CQ.
1231  * @flags:	Requested notification flags.
1232  */
1233 int siw_req_notify_cq(struct ib_cq *base_cq, enum ib_cq_notify_flags flags)
1234 {
1235 	struct siw_cq *cq = to_siw_cq(base_cq);
1236 
1237 	siw_dbg_cq(cq, "flags: 0x%02x\n", flags);
1238 
1239 	if ((flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED)
1240 		/*
1241 		 * Enable CQ event for next solicited completion.
1242 		 * and make it visible to all associated producers.
1243 		 */
1244 		smp_store_mb(cq->notify->flags, SIW_NOTIFY_SOLICITED);
1245 	else
1246 		/*
1247 		 * Enable CQ event for any signalled completion.
1248 		 * and make it visible to all associated producers.
1249 		 */
1250 		smp_store_mb(cq->notify->flags, SIW_NOTIFY_ALL);
1251 
1252 	if (flags & IB_CQ_REPORT_MISSED_EVENTS)
1253 		return cq->cq_put - cq->cq_get;
1254 
1255 	return 0;
1256 }
1257 
1258 /*
1259  * siw_dereg_mr()
1260  *
1261  * Release Memory Region.
1262  *
1263  * @base_mr: Base MR contained in siw MR.
1264  * @udata: points to user context, unused.
1265  */
1266 int siw_dereg_mr(struct ib_mr *base_mr, struct ib_udata *udata)
1267 {
1268 	struct siw_mr *mr = to_siw_mr(base_mr);
1269 	struct siw_device *sdev = to_siw_dev(base_mr->device);
1270 
1271 	siw_dbg_mem(mr->mem, "deregister MR\n");
1272 
1273 	atomic_dec(&sdev->num_mr);
1274 
1275 	siw_mr_drop_mem(mr);
1276 	kfree_rcu(mr, rcu);
1277 
1278 	return 0;
1279 }
1280 
1281 /*
1282  * siw_reg_user_mr()
1283  *
1284  * Register Memory Region.
1285  *
1286  * @pd:		Protection Domain
1287  * @start:	starting address of MR (virtual address)
1288  * @len:	len of MR
1289  * @rnic_va:	not used by siw
1290  * @rights:	MR access rights
1291  * @udata:	user buffer to communicate STag and Key.
1292  */
1293 struct ib_mr *siw_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,
1294 			      u64 rnic_va, int rights, struct ib_udata *udata)
1295 {
1296 	struct siw_mr *mr = NULL;
1297 	struct siw_umem *umem = NULL;
1298 	struct siw_ureq_reg_mr ureq;
1299 	struct siw_device *sdev = to_siw_dev(pd->device);
1300 
1301 	unsigned long mem_limit = rlimit(RLIMIT_MEMLOCK);
1302 	int rv;
1303 
1304 	siw_dbg_pd(pd, "start: 0x%pK, va: 0x%pK, len: %llu\n",
1305 		   (void *)(uintptr_t)start, (void *)(uintptr_t)rnic_va,
1306 		   (unsigned long long)len);
1307 
1308 	if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1309 		siw_dbg_pd(pd, "too many mr's\n");
1310 		rv = -ENOMEM;
1311 		goto err_out;
1312 	}
1313 	if (!len) {
1314 		rv = -EINVAL;
1315 		goto err_out;
1316 	}
1317 	if (mem_limit != RLIM_INFINITY) {
1318 		unsigned long num_pages =
1319 			(PAGE_ALIGN(len + (start & ~PAGE_MASK))) >> PAGE_SHIFT;
1320 		mem_limit >>= PAGE_SHIFT;
1321 
1322 		if (num_pages > mem_limit - current->mm->locked_vm) {
1323 			siw_dbg_pd(pd, "pages req %lu, max %lu, lock %lu\n",
1324 				   num_pages, mem_limit,
1325 				   current->mm->locked_vm);
1326 			rv = -ENOMEM;
1327 			goto err_out;
1328 		}
1329 	}
1330 	umem = siw_umem_get(start, len, ib_access_writable(rights));
1331 	if (IS_ERR(umem)) {
1332 		rv = PTR_ERR(umem);
1333 		siw_dbg_pd(pd, "getting user memory failed: %d\n", rv);
1334 		umem = NULL;
1335 		goto err_out;
1336 	}
1337 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1338 	if (!mr) {
1339 		rv = -ENOMEM;
1340 		goto err_out;
1341 	}
1342 	rv = siw_mr_add_mem(mr, pd, umem, start, len, rights);
1343 	if (rv)
1344 		goto err_out;
1345 
1346 	if (udata) {
1347 		struct siw_uresp_reg_mr uresp = {};
1348 		struct siw_mem *mem = mr->mem;
1349 
1350 		if (udata->inlen < sizeof(ureq)) {
1351 			rv = -EINVAL;
1352 			goto err_out;
1353 		}
1354 		rv = ib_copy_from_udata(&ureq, udata, sizeof(ureq));
1355 		if (rv)
1356 			goto err_out;
1357 
1358 		mr->base_mr.lkey |= ureq.stag_key;
1359 		mr->base_mr.rkey |= ureq.stag_key;
1360 		mem->stag |= ureq.stag_key;
1361 		uresp.stag = mem->stag;
1362 
1363 		if (udata->outlen < sizeof(uresp)) {
1364 			rv = -EINVAL;
1365 			goto err_out;
1366 		}
1367 		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1368 		if (rv)
1369 			goto err_out;
1370 	}
1371 	mr->mem->stag_valid = 1;
1372 
1373 	return &mr->base_mr;
1374 
1375 err_out:
1376 	atomic_dec(&sdev->num_mr);
1377 	if (mr) {
1378 		if (mr->mem)
1379 			siw_mr_drop_mem(mr);
1380 		kfree_rcu(mr, rcu);
1381 	} else {
1382 		if (umem)
1383 			siw_umem_release(umem, false);
1384 	}
1385 	return ERR_PTR(rv);
1386 }
1387 
1388 struct ib_mr *siw_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
1389 			   u32 max_sge, struct ib_udata *udata)
1390 {
1391 	struct siw_device *sdev = to_siw_dev(pd->device);
1392 	struct siw_mr *mr = NULL;
1393 	struct siw_pbl *pbl = NULL;
1394 	int rv;
1395 
1396 	if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1397 		siw_dbg_pd(pd, "too many mr's\n");
1398 		rv = -ENOMEM;
1399 		goto err_out;
1400 	}
1401 	if (mr_type != IB_MR_TYPE_MEM_REG) {
1402 		siw_dbg_pd(pd, "mr type %d unsupported\n", mr_type);
1403 		rv = -EOPNOTSUPP;
1404 		goto err_out;
1405 	}
1406 	if (max_sge > SIW_MAX_SGE_PBL) {
1407 		siw_dbg_pd(pd, "too many sge's: %d\n", max_sge);
1408 		rv = -ENOMEM;
1409 		goto err_out;
1410 	}
1411 	pbl = siw_pbl_alloc(max_sge);
1412 	if (IS_ERR(pbl)) {
1413 		rv = PTR_ERR(pbl);
1414 		siw_dbg_pd(pd, "pbl allocation failed: %d\n", rv);
1415 		pbl = NULL;
1416 		goto err_out;
1417 	}
1418 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1419 	if (!mr) {
1420 		rv = -ENOMEM;
1421 		goto err_out;
1422 	}
1423 	rv = siw_mr_add_mem(mr, pd, pbl, 0, max_sge * PAGE_SIZE, 0);
1424 	if (rv)
1425 		goto err_out;
1426 
1427 	mr->mem->is_pbl = 1;
1428 
1429 	siw_dbg_pd(pd, "[MEM %u]: success\n", mr->mem->stag);
1430 
1431 	return &mr->base_mr;
1432 
1433 err_out:
1434 	atomic_dec(&sdev->num_mr);
1435 
1436 	if (!mr) {
1437 		kfree(pbl);
1438 	} else {
1439 		if (mr->mem)
1440 			siw_mr_drop_mem(mr);
1441 		kfree_rcu(mr, rcu);
1442 	}
1443 	siw_dbg_pd(pd, "failed: %d\n", rv);
1444 
1445 	return ERR_PTR(rv);
1446 }
1447 
1448 /* Just used to count number of pages being mapped */
1449 static int siw_set_pbl_page(struct ib_mr *base_mr, u64 buf_addr)
1450 {
1451 	return 0;
1452 }
1453 
1454 int siw_map_mr_sg(struct ib_mr *base_mr, struct scatterlist *sl, int num_sle,
1455 		  unsigned int *sg_off)
1456 {
1457 	struct scatterlist *slp;
1458 	struct siw_mr *mr = to_siw_mr(base_mr);
1459 	struct siw_mem *mem = mr->mem;
1460 	struct siw_pbl *pbl = mem->pbl;
1461 	struct siw_pble *pble;
1462 	unsigned long pbl_size;
1463 	int i, rv;
1464 
1465 	if (!pbl) {
1466 		siw_dbg_mem(mem, "no PBL allocated\n");
1467 		return -EINVAL;
1468 	}
1469 	pble = pbl->pbe;
1470 
1471 	if (pbl->max_buf < num_sle) {
1472 		siw_dbg_mem(mem, "too many SGE's: %d > %d\n",
1473 			    mem->pbl->max_buf, num_sle);
1474 		return -ENOMEM;
1475 	}
1476 	for_each_sg(sl, slp, num_sle, i) {
1477 		if (sg_dma_len(slp) == 0) {
1478 			siw_dbg_mem(mem, "empty SGE\n");
1479 			return -EINVAL;
1480 		}
1481 		if (i == 0) {
1482 			pble->addr = sg_dma_address(slp);
1483 			pble->size = sg_dma_len(slp);
1484 			pble->pbl_off = 0;
1485 			pbl_size = pble->size;
1486 			pbl->num_buf = 1;
1487 		} else {
1488 			/* Merge PBL entries if adjacent */
1489 			if (pble->addr + pble->size == sg_dma_address(slp)) {
1490 				pble->size += sg_dma_len(slp);
1491 			} else {
1492 				pble++;
1493 				pbl->num_buf++;
1494 				pble->addr = sg_dma_address(slp);
1495 				pble->size = sg_dma_len(slp);
1496 				pble->pbl_off = pbl_size;
1497 			}
1498 			pbl_size += sg_dma_len(slp);
1499 		}
1500 		siw_dbg_mem(mem,
1501 			"sge[%d], size %u, addr 0x%p, total %lu\n",
1502 			i, pble->size, (void *)(uintptr_t)pble->addr,
1503 			pbl_size);
1504 	}
1505 	rv = ib_sg_to_pages(base_mr, sl, num_sle, sg_off, siw_set_pbl_page);
1506 	if (rv > 0) {
1507 		mem->len = base_mr->length;
1508 		mem->va = base_mr->iova;
1509 		siw_dbg_mem(mem,
1510 			"%llu bytes, start 0x%pK, %u SLE to %u entries\n",
1511 			mem->len, (void *)(uintptr_t)mem->va, num_sle,
1512 			pbl->num_buf);
1513 	}
1514 	return rv;
1515 }
1516 
1517 /*
1518  * siw_get_dma_mr()
1519  *
1520  * Create a (empty) DMA memory region, where no umem is attached.
1521  */
1522 struct ib_mr *siw_get_dma_mr(struct ib_pd *pd, int rights)
1523 {
1524 	struct siw_device *sdev = to_siw_dev(pd->device);
1525 	struct siw_mr *mr = NULL;
1526 	int rv;
1527 
1528 	if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1529 		siw_dbg_pd(pd, "too many mr's\n");
1530 		rv = -ENOMEM;
1531 		goto err_out;
1532 	}
1533 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1534 	if (!mr) {
1535 		rv = -ENOMEM;
1536 		goto err_out;
1537 	}
1538 	rv = siw_mr_add_mem(mr, pd, NULL, 0, ULONG_MAX, rights);
1539 	if (rv)
1540 		goto err_out;
1541 
1542 	mr->mem->stag_valid = 1;
1543 
1544 	siw_dbg_pd(pd, "[MEM %u]: success\n", mr->mem->stag);
1545 
1546 	return &mr->base_mr;
1547 
1548 err_out:
1549 	if (rv)
1550 		kfree(mr);
1551 
1552 	atomic_dec(&sdev->num_mr);
1553 
1554 	return ERR_PTR(rv);
1555 }
1556 
1557 /*
1558  * siw_create_srq()
1559  *
1560  * Create Shared Receive Queue of attributes @init_attrs
1561  * within protection domain given by @pd.
1562  *
1563  * @base_srq:	Base SRQ contained in siw SRQ.
1564  * @init_attrs:	SRQ init attributes.
1565  * @udata:	points to user context
1566  */
1567 int siw_create_srq(struct ib_srq *base_srq,
1568 		   struct ib_srq_init_attr *init_attrs, struct ib_udata *udata)
1569 {
1570 	struct siw_srq *srq = to_siw_srq(base_srq);
1571 	struct ib_srq_attr *attrs = &init_attrs->attr;
1572 	struct siw_device *sdev = to_siw_dev(base_srq->device);
1573 	struct siw_ucontext *ctx =
1574 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
1575 					  base_ucontext);
1576 	int rv;
1577 
1578 	if (atomic_inc_return(&sdev->num_srq) > SIW_MAX_SRQ) {
1579 		siw_dbg_pd(base_srq->pd, "too many SRQ's\n");
1580 		rv = -ENOMEM;
1581 		goto err_out;
1582 	}
1583 	if (attrs->max_wr == 0 || attrs->max_wr > SIW_MAX_SRQ_WR ||
1584 	    attrs->max_sge > SIW_MAX_SGE || attrs->srq_limit > attrs->max_wr) {
1585 		rv = -EINVAL;
1586 		goto err_out;
1587 	}
1588 	srq->max_sge = attrs->max_sge;
1589 	srq->num_rqe = roundup_pow_of_two(attrs->max_wr);
1590 	srq->limit = attrs->srq_limit;
1591 	if (srq->limit)
1592 		srq->armed = 1;
1593 
1594 	srq->kernel_verbs = !udata;
1595 
1596 	if (udata)
1597 		srq->recvq =
1598 			vmalloc_user(srq->num_rqe * sizeof(struct siw_rqe));
1599 	else
1600 		srq->recvq = vzalloc(srq->num_rqe * sizeof(struct siw_rqe));
1601 
1602 	if (srq->recvq == NULL) {
1603 		rv = -ENOMEM;
1604 		goto err_out;
1605 	}
1606 	if (udata) {
1607 		struct siw_uresp_create_srq uresp = {};
1608 		size_t length = srq->num_rqe * sizeof(struct siw_rqe);
1609 
1610 		srq->srq_entry =
1611 			siw_mmap_entry_insert(ctx, srq->recvq,
1612 					      length, &uresp.srq_key);
1613 		if (!srq->srq_entry) {
1614 			rv = -ENOMEM;
1615 			goto err_out;
1616 		}
1617 
1618 		uresp.num_rqe = srq->num_rqe;
1619 
1620 		if (udata->outlen < sizeof(uresp)) {
1621 			rv = -EINVAL;
1622 			goto err_out;
1623 		}
1624 		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1625 		if (rv)
1626 			goto err_out;
1627 	}
1628 	spin_lock_init(&srq->lock);
1629 
1630 	siw_dbg_pd(base_srq->pd, "[SRQ]: success\n");
1631 
1632 	return 0;
1633 
1634 err_out:
1635 	if (srq->recvq) {
1636 		if (ctx)
1637 			rdma_user_mmap_entry_remove(srq->srq_entry);
1638 		vfree(srq->recvq);
1639 	}
1640 	atomic_dec(&sdev->num_srq);
1641 
1642 	return rv;
1643 }
1644 
1645 /*
1646  * siw_modify_srq()
1647  *
1648  * Modify SRQ. The caller may resize SRQ and/or set/reset notification
1649  * limit and (re)arm IB_EVENT_SRQ_LIMIT_REACHED notification.
1650  *
1651  * NOTE: it is unclear if RDMA core allows for changing the MAX_SGE
1652  * parameter. siw_modify_srq() does not check the attrs->max_sge param.
1653  */
1654 int siw_modify_srq(struct ib_srq *base_srq, struct ib_srq_attr *attrs,
1655 		   enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
1656 {
1657 	struct siw_srq *srq = to_siw_srq(base_srq);
1658 	unsigned long flags;
1659 	int rv = 0;
1660 
1661 	spin_lock_irqsave(&srq->lock, flags);
1662 
1663 	if (attr_mask & IB_SRQ_MAX_WR) {
1664 		/* resize request not yet supported */
1665 		rv = -EOPNOTSUPP;
1666 		goto out;
1667 	}
1668 	if (attr_mask & IB_SRQ_LIMIT) {
1669 		if (attrs->srq_limit) {
1670 			if (unlikely(attrs->srq_limit > srq->num_rqe)) {
1671 				rv = -EINVAL;
1672 				goto out;
1673 			}
1674 			srq->armed = 1;
1675 		} else {
1676 			srq->armed = 0;
1677 		}
1678 		srq->limit = attrs->srq_limit;
1679 	}
1680 out:
1681 	spin_unlock_irqrestore(&srq->lock, flags);
1682 
1683 	return rv;
1684 }
1685 
1686 /*
1687  * siw_query_srq()
1688  *
1689  * Query SRQ attributes.
1690  */
1691 int siw_query_srq(struct ib_srq *base_srq, struct ib_srq_attr *attrs)
1692 {
1693 	struct siw_srq *srq = to_siw_srq(base_srq);
1694 	unsigned long flags;
1695 
1696 	spin_lock_irqsave(&srq->lock, flags);
1697 
1698 	attrs->max_wr = srq->num_rqe;
1699 	attrs->max_sge = srq->max_sge;
1700 	attrs->srq_limit = srq->limit;
1701 
1702 	spin_unlock_irqrestore(&srq->lock, flags);
1703 
1704 	return 0;
1705 }
1706 
1707 /*
1708  * siw_destroy_srq()
1709  *
1710  * Destroy SRQ.
1711  * It is assumed that the SRQ is not referenced by any
1712  * QP anymore - the code trusts the RDMA core environment to keep track
1713  * of QP references.
1714  */
1715 void siw_destroy_srq(struct ib_srq *base_srq, struct ib_udata *udata)
1716 {
1717 	struct siw_srq *srq = to_siw_srq(base_srq);
1718 	struct siw_device *sdev = to_siw_dev(base_srq->device);
1719 	struct siw_ucontext *ctx =
1720 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
1721 					  base_ucontext);
1722 
1723 	if (ctx)
1724 		rdma_user_mmap_entry_remove(srq->srq_entry);
1725 	vfree(srq->recvq);
1726 	atomic_dec(&sdev->num_srq);
1727 }
1728 
1729 /*
1730  * siw_post_srq_recv()
1731  *
1732  * Post a list of receive queue elements to SRQ.
1733  * NOTE: The function does not check or lock a certain SRQ state
1734  *       during the post operation. The code simply trusts the
1735  *       RDMA core environment.
1736  *
1737  * @base_srq:	Base SRQ contained in siw SRQ
1738  * @wr:		List of R-WR's
1739  * @bad_wr:	Updated to failing WR if posting fails.
1740  */
1741 int siw_post_srq_recv(struct ib_srq *base_srq, const struct ib_recv_wr *wr,
1742 		      const struct ib_recv_wr **bad_wr)
1743 {
1744 	struct siw_srq *srq = to_siw_srq(base_srq);
1745 	unsigned long flags;
1746 	int rv = 0;
1747 
1748 	if (unlikely(!srq->kernel_verbs)) {
1749 		siw_dbg_pd(base_srq->pd,
1750 			   "[SRQ]: no kernel post_recv for mapped srq\n");
1751 		rv = -EINVAL;
1752 		goto out;
1753 	}
1754 	/*
1755 	 * Serialize potentially multiple producers.
1756 	 * Also needed to serialize potentially multiple
1757 	 * consumers.
1758 	 */
1759 	spin_lock_irqsave(&srq->lock, flags);
1760 
1761 	while (wr) {
1762 		u32 idx = srq->rq_put % srq->num_rqe;
1763 		struct siw_rqe *rqe = &srq->recvq[idx];
1764 
1765 		if (rqe->flags) {
1766 			siw_dbg_pd(base_srq->pd, "SRQ full\n");
1767 			rv = -ENOMEM;
1768 			break;
1769 		}
1770 		if (unlikely(wr->num_sge > srq->max_sge)) {
1771 			siw_dbg_pd(base_srq->pd,
1772 				   "[SRQ]: too many sge's: %d\n", wr->num_sge);
1773 			rv = -EINVAL;
1774 			break;
1775 		}
1776 		rqe->id = wr->wr_id;
1777 		rqe->num_sge = wr->num_sge;
1778 		siw_copy_sgl(wr->sg_list, rqe->sge, wr->num_sge);
1779 
1780 		/* Make sure S-RQE is completely written before valid */
1781 		smp_wmb();
1782 
1783 		rqe->flags = SIW_WQE_VALID;
1784 
1785 		srq->rq_put++;
1786 		wr = wr->next;
1787 	}
1788 	spin_unlock_irqrestore(&srq->lock, flags);
1789 out:
1790 	if (unlikely(rv < 0)) {
1791 		siw_dbg_pd(base_srq->pd, "[SRQ]: error %d\n", rv);
1792 		*bad_wr = wr;
1793 	}
1794 	return rv;
1795 }
1796 
1797 void siw_qp_event(struct siw_qp *qp, enum ib_event_type etype)
1798 {
1799 	struct ib_event event;
1800 	struct ib_qp *base_qp = qp->ib_qp;
1801 
1802 	/*
1803 	 * Do not report asynchronous errors on QP which gets
1804 	 * destroyed via verbs interface (siw_destroy_qp())
1805 	 */
1806 	if (qp->attrs.flags & SIW_QP_IN_DESTROY)
1807 		return;
1808 
1809 	event.event = etype;
1810 	event.device = base_qp->device;
1811 	event.element.qp = base_qp;
1812 
1813 	if (base_qp->event_handler) {
1814 		siw_dbg_qp(qp, "reporting event %d\n", etype);
1815 		base_qp->event_handler(&event, base_qp->qp_context);
1816 	}
1817 }
1818 
1819 void siw_cq_event(struct siw_cq *cq, enum ib_event_type etype)
1820 {
1821 	struct ib_event event;
1822 	struct ib_cq *base_cq = &cq->base_cq;
1823 
1824 	event.event = etype;
1825 	event.device = base_cq->device;
1826 	event.element.cq = base_cq;
1827 
1828 	if (base_cq->event_handler) {
1829 		siw_dbg_cq(cq, "reporting CQ event %d\n", etype);
1830 		base_cq->event_handler(&event, base_cq->cq_context);
1831 	}
1832 }
1833 
1834 void siw_srq_event(struct siw_srq *srq, enum ib_event_type etype)
1835 {
1836 	struct ib_event event;
1837 	struct ib_srq *base_srq = &srq->base_srq;
1838 
1839 	event.event = etype;
1840 	event.device = base_srq->device;
1841 	event.element.srq = base_srq;
1842 
1843 	if (base_srq->event_handler) {
1844 		siw_dbg_pd(srq->base_srq.pd,
1845 			   "reporting SRQ event %d\n", etype);
1846 		base_srq->event_handler(&event, base_srq->srq_context);
1847 	}
1848 }
1849 
1850 void siw_port_event(struct siw_device *sdev, u8 port, enum ib_event_type etype)
1851 {
1852 	struct ib_event event;
1853 
1854 	event.event = etype;
1855 	event.device = &sdev->base_dev;
1856 	event.element.port_num = port;
1857 
1858 	siw_dbg(&sdev->base_dev, "reporting port event %d\n", etype);
1859 
1860 	ib_dispatch_event(&event);
1861 }
1862