xref: /linux/drivers/infiniband/sw/siw/siw_verbs.c (revision b77e0ce62d63a761ffb7f7245a215a49f5921c2f)
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 int 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 	return 0;
244 }
245 
246 void siw_qp_get_ref(struct ib_qp *base_qp)
247 {
248 	siw_qp_get(to_siw_qp(base_qp));
249 }
250 
251 void siw_qp_put_ref(struct ib_qp *base_qp)
252 {
253 	siw_qp_put(to_siw_qp(base_qp));
254 }
255 
256 static struct rdma_user_mmap_entry *
257 siw_mmap_entry_insert(struct siw_ucontext *uctx,
258 		      void *address, size_t length,
259 		      u64 *offset)
260 {
261 	struct siw_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
262 	int rv;
263 
264 	*offset = SIW_INVAL_UOBJ_KEY;
265 	if (!entry)
266 		return NULL;
267 
268 	entry->address = address;
269 
270 	rv = rdma_user_mmap_entry_insert(&uctx->base_ucontext,
271 					 &entry->rdma_entry,
272 					 length);
273 	if (rv) {
274 		kfree(entry);
275 		return NULL;
276 	}
277 
278 	*offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
279 
280 	return &entry->rdma_entry;
281 }
282 
283 /*
284  * siw_create_qp()
285  *
286  * Create QP of requested size on given device.
287  *
288  * @pd:		Protection Domain
289  * @attrs:	Initial QP attributes.
290  * @udata:	used to provide QP ID, SQ and RQ size back to user.
291  */
292 
293 struct ib_qp *siw_create_qp(struct ib_pd *pd,
294 			    struct ib_qp_init_attr *attrs,
295 			    struct ib_udata *udata)
296 {
297 	struct siw_qp *qp = NULL;
298 	struct ib_device *base_dev = pd->device;
299 	struct siw_device *sdev = to_siw_dev(base_dev);
300 	struct siw_ucontext *uctx =
301 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
302 					  base_ucontext);
303 	struct siw_cq *scq = NULL, *rcq = NULL;
304 	unsigned long flags;
305 	int num_sqe, num_rqe, rv = 0;
306 	size_t length;
307 
308 	siw_dbg(base_dev, "create new QP\n");
309 
310 	if (attrs->create_flags)
311 		return ERR_PTR(-EOPNOTSUPP);
312 
313 	if (atomic_inc_return(&sdev->num_qp) > SIW_MAX_QP) {
314 		siw_dbg(base_dev, "too many QP's\n");
315 		rv = -ENOMEM;
316 		goto err_out;
317 	}
318 	if (attrs->qp_type != IB_QPT_RC) {
319 		siw_dbg(base_dev, "only RC QP's supported\n");
320 		rv = -EOPNOTSUPP;
321 		goto err_out;
322 	}
323 	if ((attrs->cap.max_send_wr > SIW_MAX_QP_WR) ||
324 	    (attrs->cap.max_recv_wr > SIW_MAX_QP_WR) ||
325 	    (attrs->cap.max_send_sge > SIW_MAX_SGE) ||
326 	    (attrs->cap.max_recv_sge > SIW_MAX_SGE)) {
327 		siw_dbg(base_dev, "QP size error\n");
328 		rv = -EINVAL;
329 		goto err_out;
330 	}
331 	if (attrs->cap.max_inline_data > SIW_MAX_INLINE) {
332 		siw_dbg(base_dev, "max inline send: %d > %d\n",
333 			attrs->cap.max_inline_data, (int)SIW_MAX_INLINE);
334 		rv = -EINVAL;
335 		goto err_out;
336 	}
337 	/*
338 	 * NOTE: we allow for zero element SQ and RQ WQE's SGL's
339 	 * but not for a QP unable to hold any WQE (SQ + RQ)
340 	 */
341 	if (attrs->cap.max_send_wr + attrs->cap.max_recv_wr == 0) {
342 		siw_dbg(base_dev, "QP must have send or receive queue\n");
343 		rv = -EINVAL;
344 		goto err_out;
345 	}
346 	scq = to_siw_cq(attrs->send_cq);
347 	rcq = to_siw_cq(attrs->recv_cq);
348 
349 	if (!scq || (!rcq && !attrs->srq)) {
350 		siw_dbg(base_dev, "send CQ or receive CQ invalid\n");
351 		rv = -EINVAL;
352 		goto err_out;
353 	}
354 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
355 	if (!qp) {
356 		rv = -ENOMEM;
357 		goto err_out;
358 	}
359 	init_rwsem(&qp->state_lock);
360 	spin_lock_init(&qp->sq_lock);
361 	spin_lock_init(&qp->rq_lock);
362 	spin_lock_init(&qp->orq_lock);
363 
364 	rv = siw_qp_add(sdev, qp);
365 	if (rv)
366 		goto err_out;
367 
368 	num_sqe = attrs->cap.max_send_wr;
369 	num_rqe = attrs->cap.max_recv_wr;
370 
371 	/* All queue indices are derived from modulo operations
372 	 * on a free running 'get' (consumer) and 'put' (producer)
373 	 * unsigned counter. Having queue sizes at power of two
374 	 * avoids handling counter wrap around.
375 	 */
376 	if (num_sqe)
377 		num_sqe = roundup_pow_of_two(num_sqe);
378 	else {
379 		/* Zero sized SQ is not supported */
380 		rv = -EINVAL;
381 		goto err_out;
382 	}
383 	if (num_rqe)
384 		num_rqe = roundup_pow_of_two(num_rqe);
385 
386 	if (udata)
387 		qp->sendq = vmalloc_user(num_sqe * sizeof(struct siw_sqe));
388 	else
389 		qp->sendq = vzalloc(num_sqe * sizeof(struct siw_sqe));
390 
391 	if (qp->sendq == NULL) {
392 		rv = -ENOMEM;
393 		goto err_out_xa;
394 	}
395 	if (attrs->sq_sig_type != IB_SIGNAL_REQ_WR) {
396 		if (attrs->sq_sig_type == IB_SIGNAL_ALL_WR)
397 			qp->attrs.flags |= SIW_SIGNAL_ALL_WR;
398 		else {
399 			rv = -EINVAL;
400 			goto err_out_xa;
401 		}
402 	}
403 	qp->pd = pd;
404 	qp->scq = scq;
405 	qp->rcq = rcq;
406 
407 	if (attrs->srq) {
408 		/*
409 		 * SRQ support.
410 		 * Verbs 6.3.7: ignore RQ size, if SRQ present
411 		 * Verbs 6.3.5: do not check PD of SRQ against PD of QP
412 		 */
413 		qp->srq = to_siw_srq(attrs->srq);
414 		qp->attrs.rq_size = 0;
415 		siw_dbg(base_dev, "QP [%u]: SRQ attached\n",
416 			qp->base_qp.qp_num);
417 	} else if (num_rqe) {
418 		if (udata)
419 			qp->recvq =
420 				vmalloc_user(num_rqe * sizeof(struct siw_rqe));
421 		else
422 			qp->recvq = vzalloc(num_rqe * sizeof(struct siw_rqe));
423 
424 		if (qp->recvq == NULL) {
425 			rv = -ENOMEM;
426 			goto err_out_xa;
427 		}
428 		qp->attrs.rq_size = num_rqe;
429 	}
430 	qp->attrs.sq_size = num_sqe;
431 	qp->attrs.sq_max_sges = attrs->cap.max_send_sge;
432 	qp->attrs.rq_max_sges = attrs->cap.max_recv_sge;
433 
434 	/* Make those two tunables fixed for now. */
435 	qp->tx_ctx.gso_seg_limit = 1;
436 	qp->tx_ctx.zcopy_tx = zcopy_tx;
437 
438 	qp->attrs.state = SIW_QP_STATE_IDLE;
439 
440 	if (udata) {
441 		struct siw_uresp_create_qp uresp = {};
442 
443 		uresp.num_sqe = num_sqe;
444 		uresp.num_rqe = num_rqe;
445 		uresp.qp_id = qp_id(qp);
446 
447 		if (qp->sendq) {
448 			length = num_sqe * sizeof(struct siw_sqe);
449 			qp->sq_entry =
450 				siw_mmap_entry_insert(uctx, qp->sendq,
451 						      length, &uresp.sq_key);
452 			if (!qp->sq_entry) {
453 				rv = -ENOMEM;
454 				goto err_out_xa;
455 			}
456 		}
457 
458 		if (qp->recvq) {
459 			length = num_rqe * sizeof(struct siw_rqe);
460 			qp->rq_entry =
461 				siw_mmap_entry_insert(uctx, qp->recvq,
462 						      length, &uresp.rq_key);
463 			if (!qp->rq_entry) {
464 				uresp.sq_key = SIW_INVAL_UOBJ_KEY;
465 				rv = -ENOMEM;
466 				goto err_out_xa;
467 			}
468 		}
469 
470 		if (udata->outlen < sizeof(uresp)) {
471 			rv = -EINVAL;
472 			goto err_out_xa;
473 		}
474 		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
475 		if (rv)
476 			goto err_out_xa;
477 	}
478 	qp->tx_cpu = siw_get_tx_cpu(sdev);
479 	if (qp->tx_cpu < 0) {
480 		rv = -EINVAL;
481 		goto err_out_xa;
482 	}
483 	INIT_LIST_HEAD(&qp->devq);
484 	spin_lock_irqsave(&sdev->lock, flags);
485 	list_add_tail(&qp->devq, &sdev->qp_list);
486 	spin_unlock_irqrestore(&sdev->lock, flags);
487 
488 	return &qp->base_qp;
489 
490 err_out_xa:
491 	xa_erase(&sdev->qp_xa, qp_id(qp));
492 err_out:
493 	if (qp) {
494 		if (uctx) {
495 			rdma_user_mmap_entry_remove(qp->sq_entry);
496 			rdma_user_mmap_entry_remove(qp->rq_entry);
497 		}
498 		vfree(qp->sendq);
499 		vfree(qp->recvq);
500 		kfree(qp);
501 	}
502 	atomic_dec(&sdev->num_qp);
503 
504 	return ERR_PTR(rv);
505 }
506 
507 /*
508  * Minimum siw_query_qp() verb interface.
509  *
510  * @qp_attr_mask is not used but all available information is provided
511  */
512 int siw_query_qp(struct ib_qp *base_qp, struct ib_qp_attr *qp_attr,
513 		 int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
514 {
515 	struct siw_qp *qp;
516 	struct siw_device *sdev;
517 
518 	if (base_qp && qp_attr && qp_init_attr) {
519 		qp = to_siw_qp(base_qp);
520 		sdev = to_siw_dev(base_qp->device);
521 	} else {
522 		return -EINVAL;
523 	}
524 	qp_attr->cap.max_inline_data = SIW_MAX_INLINE;
525 	qp_attr->cap.max_send_wr = qp->attrs.sq_size;
526 	qp_attr->cap.max_send_sge = qp->attrs.sq_max_sges;
527 	qp_attr->cap.max_recv_wr = qp->attrs.rq_size;
528 	qp_attr->cap.max_recv_sge = qp->attrs.rq_max_sges;
529 	qp_attr->path_mtu = ib_mtu_int_to_enum(sdev->netdev->mtu);
530 	qp_attr->max_rd_atomic = qp->attrs.irq_size;
531 	qp_attr->max_dest_rd_atomic = qp->attrs.orq_size;
532 
533 	qp_attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE |
534 				   IB_ACCESS_REMOTE_WRITE |
535 				   IB_ACCESS_REMOTE_READ;
536 
537 	qp_init_attr->qp_type = base_qp->qp_type;
538 	qp_init_attr->send_cq = base_qp->send_cq;
539 	qp_init_attr->recv_cq = base_qp->recv_cq;
540 	qp_init_attr->srq = base_qp->srq;
541 
542 	qp_init_attr->cap = qp_attr->cap;
543 
544 	return 0;
545 }
546 
547 int siw_verbs_modify_qp(struct ib_qp *base_qp, struct ib_qp_attr *attr,
548 			int attr_mask, struct ib_udata *udata)
549 {
550 	struct siw_qp_attrs new_attrs;
551 	enum siw_qp_attr_mask siw_attr_mask = 0;
552 	struct siw_qp *qp = to_siw_qp(base_qp);
553 	int rv = 0;
554 
555 	if (!attr_mask)
556 		return 0;
557 
558 	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
559 		return -EOPNOTSUPP;
560 
561 	memset(&new_attrs, 0, sizeof(new_attrs));
562 
563 	if (attr_mask & IB_QP_ACCESS_FLAGS) {
564 		siw_attr_mask = SIW_QP_ATTR_ACCESS_FLAGS;
565 
566 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
567 			new_attrs.flags |= SIW_RDMA_READ_ENABLED;
568 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
569 			new_attrs.flags |= SIW_RDMA_WRITE_ENABLED;
570 		if (attr->qp_access_flags & IB_ACCESS_MW_BIND)
571 			new_attrs.flags |= SIW_RDMA_BIND_ENABLED;
572 	}
573 	if (attr_mask & IB_QP_STATE) {
574 		siw_dbg_qp(qp, "desired IB QP state: %s\n",
575 			   ib_qp_state_to_string[attr->qp_state]);
576 
577 		new_attrs.state = ib_qp_state_to_siw_qp_state[attr->qp_state];
578 
579 		if (new_attrs.state > SIW_QP_STATE_RTS)
580 			qp->tx_ctx.tx_suspend = 1;
581 
582 		siw_attr_mask |= SIW_QP_ATTR_STATE;
583 	}
584 	if (!siw_attr_mask)
585 		goto out;
586 
587 	down_write(&qp->state_lock);
588 
589 	rv = siw_qp_modify(qp, &new_attrs, siw_attr_mask);
590 
591 	up_write(&qp->state_lock);
592 out:
593 	return rv;
594 }
595 
596 int siw_destroy_qp(struct ib_qp *base_qp, struct ib_udata *udata)
597 {
598 	struct siw_qp *qp = to_siw_qp(base_qp);
599 	struct siw_ucontext *uctx =
600 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
601 					  base_ucontext);
602 	struct siw_qp_attrs qp_attrs;
603 
604 	siw_dbg_qp(qp, "state %d\n", qp->attrs.state);
605 
606 	/*
607 	 * Mark QP as in process of destruction to prevent from
608 	 * any async callbacks to RDMA core
609 	 */
610 	qp->attrs.flags |= SIW_QP_IN_DESTROY;
611 	qp->rx_stream.rx_suspend = 1;
612 
613 	if (uctx) {
614 		rdma_user_mmap_entry_remove(qp->sq_entry);
615 		rdma_user_mmap_entry_remove(qp->rq_entry);
616 	}
617 
618 	down_write(&qp->state_lock);
619 
620 	qp_attrs.state = SIW_QP_STATE_ERROR;
621 	siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE);
622 
623 	if (qp->cep) {
624 		siw_cep_put(qp->cep);
625 		qp->cep = NULL;
626 	}
627 	up_write(&qp->state_lock);
628 
629 	kfree(qp->tx_ctx.mpa_crc_hd);
630 	kfree(qp->rx_stream.mpa_crc_hd);
631 
632 	qp->scq = qp->rcq = NULL;
633 
634 	siw_qp_put(qp);
635 
636 	return 0;
637 }
638 
639 /*
640  * siw_copy_inline_sgl()
641  *
642  * Prepare sgl of inlined data for sending. For userland callers
643  * function checks if given buffer addresses and len's are within
644  * process context bounds.
645  * Data from all provided sge's are copied together into the wqe,
646  * referenced by a single sge.
647  */
648 static int siw_copy_inline_sgl(const struct ib_send_wr *core_wr,
649 			       struct siw_sqe *sqe)
650 {
651 	struct ib_sge *core_sge = core_wr->sg_list;
652 	void *kbuf = &sqe->sge[1];
653 	int num_sge = core_wr->num_sge, bytes = 0;
654 
655 	sqe->sge[0].laddr = (uintptr_t)kbuf;
656 	sqe->sge[0].lkey = 0;
657 
658 	while (num_sge--) {
659 		if (!core_sge->length) {
660 			core_sge++;
661 			continue;
662 		}
663 		bytes += core_sge->length;
664 		if (bytes > SIW_MAX_INLINE) {
665 			bytes = -EINVAL;
666 			break;
667 		}
668 		memcpy(kbuf, (void *)(uintptr_t)core_sge->addr,
669 		       core_sge->length);
670 
671 		kbuf += core_sge->length;
672 		core_sge++;
673 	}
674 	sqe->sge[0].length = bytes > 0 ? bytes : 0;
675 	sqe->num_sge = bytes > 0 ? 1 : 0;
676 
677 	return bytes;
678 }
679 
680 /* Complete SQ WR's without processing */
681 static int siw_sq_flush_wr(struct siw_qp *qp, const struct ib_send_wr *wr,
682 			   const struct ib_send_wr **bad_wr)
683 {
684 	struct siw_sqe sqe = {};
685 	int rv = 0;
686 
687 	while (wr) {
688 		sqe.id = wr->wr_id;
689 		sqe.opcode = wr->opcode;
690 		rv = siw_sqe_complete(qp, &sqe, 0, SIW_WC_WR_FLUSH_ERR);
691 		if (rv) {
692 			if (bad_wr)
693 				*bad_wr = wr;
694 			break;
695 		}
696 		wr = wr->next;
697 	}
698 	return rv;
699 }
700 
701 /* Complete RQ WR's without processing */
702 static int siw_rq_flush_wr(struct siw_qp *qp, const struct ib_recv_wr *wr,
703 			   const struct ib_recv_wr **bad_wr)
704 {
705 	struct siw_rqe rqe = {};
706 	int rv = 0;
707 
708 	while (wr) {
709 		rqe.id = wr->wr_id;
710 		rv = siw_rqe_complete(qp, &rqe, 0, 0, SIW_WC_WR_FLUSH_ERR);
711 		if (rv) {
712 			if (bad_wr)
713 				*bad_wr = wr;
714 			break;
715 		}
716 		wr = wr->next;
717 	}
718 	return rv;
719 }
720 
721 /*
722  * siw_post_send()
723  *
724  * Post a list of S-WR's to a SQ.
725  *
726  * @base_qp:	Base QP contained in siw QP
727  * @wr:		Null terminated list of user WR's
728  * @bad_wr:	Points to failing WR in case of synchronous failure.
729  */
730 int siw_post_send(struct ib_qp *base_qp, const struct ib_send_wr *wr,
731 		  const struct ib_send_wr **bad_wr)
732 {
733 	struct siw_qp *qp = to_siw_qp(base_qp);
734 	struct siw_wqe *wqe = tx_wqe(qp);
735 
736 	unsigned long flags;
737 	int rv = 0;
738 
739 	if (wr && !rdma_is_kernel_res(&qp->base_qp.res)) {
740 		siw_dbg_qp(qp, "wr must be empty for user mapped sq\n");
741 		*bad_wr = wr;
742 		return -EINVAL;
743 	}
744 
745 	/*
746 	 * Try to acquire QP state lock. Must be non-blocking
747 	 * to accommodate kernel clients needs.
748 	 */
749 	if (!down_read_trylock(&qp->state_lock)) {
750 		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
751 			/*
752 			 * ERROR state is final, so we can be sure
753 			 * this state will not change as long as the QP
754 			 * exists.
755 			 *
756 			 * This handles an ib_drain_sq() call with
757 			 * a concurrent request to set the QP state
758 			 * to ERROR.
759 			 */
760 			rv = siw_sq_flush_wr(qp, wr, bad_wr);
761 		} else {
762 			siw_dbg_qp(qp, "QP locked, state %d\n",
763 				   qp->attrs.state);
764 			*bad_wr = wr;
765 			rv = -ENOTCONN;
766 		}
767 		return rv;
768 	}
769 	if (unlikely(qp->attrs.state != SIW_QP_STATE_RTS)) {
770 		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
771 			/*
772 			 * Immediately flush this WR to CQ, if QP
773 			 * is in ERROR state. SQ is guaranteed to
774 			 * be empty, so WR complets in-order.
775 			 *
776 			 * Typically triggered by ib_drain_sq().
777 			 */
778 			rv = siw_sq_flush_wr(qp, wr, bad_wr);
779 		} else {
780 			siw_dbg_qp(qp, "QP out of state %d\n",
781 				   qp->attrs.state);
782 			*bad_wr = wr;
783 			rv = -ENOTCONN;
784 		}
785 		up_read(&qp->state_lock);
786 		return rv;
787 	}
788 	spin_lock_irqsave(&qp->sq_lock, flags);
789 
790 	while (wr) {
791 		u32 idx = qp->sq_put % qp->attrs.sq_size;
792 		struct siw_sqe *sqe = &qp->sendq[idx];
793 
794 		if (sqe->flags) {
795 			siw_dbg_qp(qp, "sq full\n");
796 			rv = -ENOMEM;
797 			break;
798 		}
799 		if (wr->num_sge > qp->attrs.sq_max_sges) {
800 			siw_dbg_qp(qp, "too many sge's: %d\n", wr->num_sge);
801 			rv = -EINVAL;
802 			break;
803 		}
804 		sqe->id = wr->wr_id;
805 
806 		if ((wr->send_flags & IB_SEND_SIGNALED) ||
807 		    (qp->attrs.flags & SIW_SIGNAL_ALL_WR))
808 			sqe->flags |= SIW_WQE_SIGNALLED;
809 
810 		if (wr->send_flags & IB_SEND_FENCE)
811 			sqe->flags |= SIW_WQE_READ_FENCE;
812 
813 		switch (wr->opcode) {
814 		case IB_WR_SEND:
815 		case IB_WR_SEND_WITH_INV:
816 			if (wr->send_flags & IB_SEND_SOLICITED)
817 				sqe->flags |= SIW_WQE_SOLICITED;
818 
819 			if (!(wr->send_flags & IB_SEND_INLINE)) {
820 				siw_copy_sgl(wr->sg_list, sqe->sge,
821 					     wr->num_sge);
822 				sqe->num_sge = wr->num_sge;
823 			} else {
824 				rv = siw_copy_inline_sgl(wr, sqe);
825 				if (rv <= 0) {
826 					rv = -EINVAL;
827 					break;
828 				}
829 				sqe->flags |= SIW_WQE_INLINE;
830 				sqe->num_sge = 1;
831 			}
832 			if (wr->opcode == IB_WR_SEND)
833 				sqe->opcode = SIW_OP_SEND;
834 			else {
835 				sqe->opcode = SIW_OP_SEND_REMOTE_INV;
836 				sqe->rkey = wr->ex.invalidate_rkey;
837 			}
838 			break;
839 
840 		case IB_WR_RDMA_READ_WITH_INV:
841 		case IB_WR_RDMA_READ:
842 			/*
843 			 * iWarp restricts RREAD sink to SGL containing
844 			 * 1 SGE only. we could relax to SGL with multiple
845 			 * elements referring the SAME ltag or even sending
846 			 * a private per-rreq tag referring to a checked
847 			 * local sgl with MULTIPLE ltag's.
848 			 */
849 			if (unlikely(wr->num_sge != 1)) {
850 				rv = -EINVAL;
851 				break;
852 			}
853 			siw_copy_sgl(wr->sg_list, &sqe->sge[0], 1);
854 			/*
855 			 * NOTE: zero length RREAD is allowed!
856 			 */
857 			sqe->raddr = rdma_wr(wr)->remote_addr;
858 			sqe->rkey = rdma_wr(wr)->rkey;
859 			sqe->num_sge = 1;
860 
861 			if (wr->opcode == IB_WR_RDMA_READ)
862 				sqe->opcode = SIW_OP_READ;
863 			else
864 				sqe->opcode = SIW_OP_READ_LOCAL_INV;
865 			break;
866 
867 		case IB_WR_RDMA_WRITE:
868 			if (!(wr->send_flags & IB_SEND_INLINE)) {
869 				siw_copy_sgl(wr->sg_list, &sqe->sge[0],
870 					     wr->num_sge);
871 				sqe->num_sge = wr->num_sge;
872 			} else {
873 				rv = siw_copy_inline_sgl(wr, sqe);
874 				if (unlikely(rv < 0)) {
875 					rv = -EINVAL;
876 					break;
877 				}
878 				sqe->flags |= SIW_WQE_INLINE;
879 				sqe->num_sge = 1;
880 			}
881 			sqe->raddr = rdma_wr(wr)->remote_addr;
882 			sqe->rkey = rdma_wr(wr)->rkey;
883 			sqe->opcode = SIW_OP_WRITE;
884 			break;
885 
886 		case IB_WR_REG_MR:
887 			sqe->base_mr = (uintptr_t)reg_wr(wr)->mr;
888 			sqe->rkey = reg_wr(wr)->key;
889 			sqe->access = reg_wr(wr)->access & IWARP_ACCESS_MASK;
890 			sqe->opcode = SIW_OP_REG_MR;
891 			break;
892 
893 		case IB_WR_LOCAL_INV:
894 			sqe->rkey = wr->ex.invalidate_rkey;
895 			sqe->opcode = SIW_OP_INVAL_STAG;
896 			break;
897 
898 		default:
899 			siw_dbg_qp(qp, "ib wr type %d unsupported\n",
900 				   wr->opcode);
901 			rv = -EINVAL;
902 			break;
903 		}
904 		siw_dbg_qp(qp, "opcode %d, flags 0x%x, wr_id 0x%pK\n",
905 			   sqe->opcode, sqe->flags,
906 			   (void *)(uintptr_t)sqe->id);
907 
908 		if (unlikely(rv < 0))
909 			break;
910 
911 		/* make SQE only valid after completely written */
912 		smp_wmb();
913 		sqe->flags |= SIW_WQE_VALID;
914 
915 		qp->sq_put++;
916 		wr = wr->next;
917 	}
918 
919 	/*
920 	 * Send directly if SQ processing is not in progress.
921 	 * Eventual immediate errors (rv < 0) do not affect the involved
922 	 * RI resources (Verbs, 8.3.1) and thus do not prevent from SQ
923 	 * processing, if new work is already pending. But rv must be passed
924 	 * to caller.
925 	 */
926 	if (wqe->wr_status != SIW_WR_IDLE) {
927 		spin_unlock_irqrestore(&qp->sq_lock, flags);
928 		goto skip_direct_sending;
929 	}
930 	rv = siw_activate_tx(qp);
931 	spin_unlock_irqrestore(&qp->sq_lock, flags);
932 
933 	if (rv <= 0)
934 		goto skip_direct_sending;
935 
936 	if (rdma_is_kernel_res(&qp->base_qp.res)) {
937 		rv = siw_sq_start(qp);
938 	} else {
939 		qp->tx_ctx.in_syscall = 1;
940 
941 		if (siw_qp_sq_process(qp) != 0 && !(qp->tx_ctx.tx_suspend))
942 			siw_qp_cm_drop(qp, 0);
943 
944 		qp->tx_ctx.in_syscall = 0;
945 	}
946 skip_direct_sending:
947 
948 	up_read(&qp->state_lock);
949 
950 	if (rv >= 0)
951 		return 0;
952 	/*
953 	 * Immediate error
954 	 */
955 	siw_dbg_qp(qp, "error %d\n", rv);
956 
957 	*bad_wr = wr;
958 	return rv;
959 }
960 
961 /*
962  * siw_post_receive()
963  *
964  * Post a list of R-WR's to a RQ.
965  *
966  * @base_qp:	Base QP contained in siw QP
967  * @wr:		Null terminated list of user WR's
968  * @bad_wr:	Points to failing WR in case of synchronous failure.
969  */
970 int siw_post_receive(struct ib_qp *base_qp, const struct ib_recv_wr *wr,
971 		     const struct ib_recv_wr **bad_wr)
972 {
973 	struct siw_qp *qp = to_siw_qp(base_qp);
974 	unsigned long flags;
975 	int rv = 0;
976 
977 	if (qp->srq || qp->attrs.rq_size == 0) {
978 		*bad_wr = wr;
979 		return -EINVAL;
980 	}
981 	if (!rdma_is_kernel_res(&qp->base_qp.res)) {
982 		siw_dbg_qp(qp, "no kernel post_recv for user mapped rq\n");
983 		*bad_wr = wr;
984 		return -EINVAL;
985 	}
986 
987 	/*
988 	 * Try to acquire QP state lock. Must be non-blocking
989 	 * to accommodate kernel clients needs.
990 	 */
991 	if (!down_read_trylock(&qp->state_lock)) {
992 		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
993 			/*
994 			 * ERROR state is final, so we can be sure
995 			 * this state will not change as long as the QP
996 			 * exists.
997 			 *
998 			 * This handles an ib_drain_rq() call with
999 			 * a concurrent request to set the QP state
1000 			 * to ERROR.
1001 			 */
1002 			rv = siw_rq_flush_wr(qp, wr, bad_wr);
1003 		} else {
1004 			siw_dbg_qp(qp, "QP locked, state %d\n",
1005 				   qp->attrs.state);
1006 			*bad_wr = wr;
1007 			rv = -ENOTCONN;
1008 		}
1009 		return rv;
1010 	}
1011 	if (qp->attrs.state > SIW_QP_STATE_RTS) {
1012 		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
1013 			/*
1014 			 * Immediately flush this WR to CQ, if QP
1015 			 * is in ERROR state. RQ is guaranteed to
1016 			 * be empty, so WR complets in-order.
1017 			 *
1018 			 * Typically triggered by ib_drain_rq().
1019 			 */
1020 			rv = siw_rq_flush_wr(qp, wr, bad_wr);
1021 		} else {
1022 			siw_dbg_qp(qp, "QP out of state %d\n",
1023 				   qp->attrs.state);
1024 			*bad_wr = wr;
1025 			rv = -ENOTCONN;
1026 		}
1027 		up_read(&qp->state_lock);
1028 		return rv;
1029 	}
1030 	/*
1031 	 * Serialize potentially multiple producers.
1032 	 * Not needed for single threaded consumer side.
1033 	 */
1034 	spin_lock_irqsave(&qp->rq_lock, flags);
1035 
1036 	while (wr) {
1037 		u32 idx = qp->rq_put % qp->attrs.rq_size;
1038 		struct siw_rqe *rqe = &qp->recvq[idx];
1039 
1040 		if (rqe->flags) {
1041 			siw_dbg_qp(qp, "RQ full\n");
1042 			rv = -ENOMEM;
1043 			break;
1044 		}
1045 		if (wr->num_sge > qp->attrs.rq_max_sges) {
1046 			siw_dbg_qp(qp, "too many sge's: %d\n", wr->num_sge);
1047 			rv = -EINVAL;
1048 			break;
1049 		}
1050 		rqe->id = wr->wr_id;
1051 		rqe->num_sge = wr->num_sge;
1052 		siw_copy_sgl(wr->sg_list, rqe->sge, wr->num_sge);
1053 
1054 		/* make sure RQE is completely written before valid */
1055 		smp_wmb();
1056 
1057 		rqe->flags = SIW_WQE_VALID;
1058 
1059 		qp->rq_put++;
1060 		wr = wr->next;
1061 	}
1062 	spin_unlock_irqrestore(&qp->rq_lock, flags);
1063 
1064 	up_read(&qp->state_lock);
1065 
1066 	if (rv < 0) {
1067 		siw_dbg_qp(qp, "error %d\n", rv);
1068 		*bad_wr = wr;
1069 	}
1070 	return rv > 0 ? 0 : rv;
1071 }
1072 
1073 int siw_destroy_cq(struct ib_cq *base_cq, struct ib_udata *udata)
1074 {
1075 	struct siw_cq *cq = to_siw_cq(base_cq);
1076 	struct siw_device *sdev = to_siw_dev(base_cq->device);
1077 	struct siw_ucontext *ctx =
1078 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
1079 					  base_ucontext);
1080 
1081 	siw_dbg_cq(cq, "free CQ resources\n");
1082 
1083 	siw_cq_flush(cq);
1084 
1085 	if (ctx)
1086 		rdma_user_mmap_entry_remove(cq->cq_entry);
1087 
1088 	atomic_dec(&sdev->num_cq);
1089 
1090 	vfree(cq->queue);
1091 	return 0;
1092 }
1093 
1094 /*
1095  * siw_create_cq()
1096  *
1097  * Populate CQ of requested size
1098  *
1099  * @base_cq: CQ as allocated by RDMA midlayer
1100  * @attr: Initial CQ attributes
1101  * @udata: relates to user context
1102  */
1103 
1104 int siw_create_cq(struct ib_cq *base_cq, const struct ib_cq_init_attr *attr,
1105 		  struct ib_udata *udata)
1106 {
1107 	struct siw_device *sdev = to_siw_dev(base_cq->device);
1108 	struct siw_cq *cq = to_siw_cq(base_cq);
1109 	int rv, size = attr->cqe;
1110 
1111 	if (attr->flags)
1112 		return -EOPNOTSUPP;
1113 
1114 	if (atomic_inc_return(&sdev->num_cq) > SIW_MAX_CQ) {
1115 		siw_dbg(base_cq->device, "too many CQ's\n");
1116 		rv = -ENOMEM;
1117 		goto err_out;
1118 	}
1119 	if (size < 1 || size > sdev->attrs.max_cqe) {
1120 		siw_dbg(base_cq->device, "CQ size error: %d\n", size);
1121 		rv = -EINVAL;
1122 		goto err_out;
1123 	}
1124 	size = roundup_pow_of_two(size);
1125 	cq->base_cq.cqe = size;
1126 	cq->num_cqe = size;
1127 
1128 	if (udata)
1129 		cq->queue = vmalloc_user(size * sizeof(struct siw_cqe) +
1130 					 sizeof(struct siw_cq_ctrl));
1131 	else
1132 		cq->queue = vzalloc(size * sizeof(struct siw_cqe) +
1133 				    sizeof(struct siw_cq_ctrl));
1134 
1135 	if (cq->queue == NULL) {
1136 		rv = -ENOMEM;
1137 		goto err_out;
1138 	}
1139 	get_random_bytes(&cq->id, 4);
1140 	siw_dbg(base_cq->device, "new CQ [%u]\n", cq->id);
1141 
1142 	spin_lock_init(&cq->lock);
1143 
1144 	cq->notify = (struct siw_cq_ctrl *)&cq->queue[size];
1145 
1146 	if (udata) {
1147 		struct siw_uresp_create_cq uresp = {};
1148 		struct siw_ucontext *ctx =
1149 			rdma_udata_to_drv_context(udata, struct siw_ucontext,
1150 						  base_ucontext);
1151 		size_t length = size * sizeof(struct siw_cqe) +
1152 			sizeof(struct siw_cq_ctrl);
1153 
1154 		cq->cq_entry =
1155 			siw_mmap_entry_insert(ctx, cq->queue,
1156 					      length, &uresp.cq_key);
1157 		if (!cq->cq_entry) {
1158 			rv = -ENOMEM;
1159 			goto err_out;
1160 		}
1161 
1162 		uresp.cq_id = cq->id;
1163 		uresp.num_cqe = size;
1164 
1165 		if (udata->outlen < sizeof(uresp)) {
1166 			rv = -EINVAL;
1167 			goto err_out;
1168 		}
1169 		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1170 		if (rv)
1171 			goto err_out;
1172 	}
1173 	return 0;
1174 
1175 err_out:
1176 	siw_dbg(base_cq->device, "CQ creation failed: %d", rv);
1177 
1178 	if (cq && cq->queue) {
1179 		struct siw_ucontext *ctx =
1180 			rdma_udata_to_drv_context(udata, struct siw_ucontext,
1181 						  base_ucontext);
1182 		if (ctx)
1183 			rdma_user_mmap_entry_remove(cq->cq_entry);
1184 		vfree(cq->queue);
1185 	}
1186 	atomic_dec(&sdev->num_cq);
1187 
1188 	return rv;
1189 }
1190 
1191 /*
1192  * siw_poll_cq()
1193  *
1194  * Reap CQ entries if available and copy work completion status into
1195  * array of WC's provided by caller. Returns number of reaped CQE's.
1196  *
1197  * @base_cq:	Base CQ contained in siw CQ.
1198  * @num_cqe:	Maximum number of CQE's to reap.
1199  * @wc:		Array of work completions to be filled by siw.
1200  */
1201 int siw_poll_cq(struct ib_cq *base_cq, int num_cqe, struct ib_wc *wc)
1202 {
1203 	struct siw_cq *cq = to_siw_cq(base_cq);
1204 	int i;
1205 
1206 	for (i = 0; i < num_cqe; i++) {
1207 		if (!siw_reap_cqe(cq, wc))
1208 			break;
1209 		wc++;
1210 	}
1211 	return i;
1212 }
1213 
1214 /*
1215  * siw_req_notify_cq()
1216  *
1217  * Request notification for new CQE's added to that CQ.
1218  * Defined flags:
1219  * o SIW_CQ_NOTIFY_SOLICITED lets siw trigger a notification
1220  *   event if a WQE with notification flag set enters the CQ
1221  * o SIW_CQ_NOTIFY_NEXT_COMP lets siw trigger a notification
1222  *   event if a WQE enters the CQ.
1223  * o IB_CQ_REPORT_MISSED_EVENTS: return value will provide the
1224  *   number of not reaped CQE's regardless of its notification
1225  *   type and current or new CQ notification settings.
1226  *
1227  * @base_cq:	Base CQ contained in siw CQ.
1228  * @flags:	Requested notification flags.
1229  */
1230 int siw_req_notify_cq(struct ib_cq *base_cq, enum ib_cq_notify_flags flags)
1231 {
1232 	struct siw_cq *cq = to_siw_cq(base_cq);
1233 
1234 	siw_dbg_cq(cq, "flags: 0x%02x\n", flags);
1235 
1236 	if ((flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED)
1237 		/*
1238 		 * Enable CQ event for next solicited completion.
1239 		 * and make it visible to all associated producers.
1240 		 */
1241 		smp_store_mb(cq->notify->flags, SIW_NOTIFY_SOLICITED);
1242 	else
1243 		/*
1244 		 * Enable CQ event for any signalled completion.
1245 		 * and make it visible to all associated producers.
1246 		 */
1247 		smp_store_mb(cq->notify->flags, SIW_NOTIFY_ALL);
1248 
1249 	if (flags & IB_CQ_REPORT_MISSED_EVENTS)
1250 		return cq->cq_put - cq->cq_get;
1251 
1252 	return 0;
1253 }
1254 
1255 /*
1256  * siw_dereg_mr()
1257  *
1258  * Release Memory Region.
1259  *
1260  * @base_mr: Base MR contained in siw MR.
1261  * @udata: points to user context, unused.
1262  */
1263 int siw_dereg_mr(struct ib_mr *base_mr, struct ib_udata *udata)
1264 {
1265 	struct siw_mr *mr = to_siw_mr(base_mr);
1266 	struct siw_device *sdev = to_siw_dev(base_mr->device);
1267 
1268 	siw_dbg_mem(mr->mem, "deregister MR\n");
1269 
1270 	atomic_dec(&sdev->num_mr);
1271 
1272 	siw_mr_drop_mem(mr);
1273 	kfree_rcu(mr, rcu);
1274 
1275 	return 0;
1276 }
1277 
1278 /*
1279  * siw_reg_user_mr()
1280  *
1281  * Register Memory Region.
1282  *
1283  * @pd:		Protection Domain
1284  * @start:	starting address of MR (virtual address)
1285  * @len:	len of MR
1286  * @rnic_va:	not used by siw
1287  * @rights:	MR access rights
1288  * @udata:	user buffer to communicate STag and Key.
1289  */
1290 struct ib_mr *siw_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,
1291 			      u64 rnic_va, int rights, struct ib_udata *udata)
1292 {
1293 	struct siw_mr *mr = NULL;
1294 	struct siw_umem *umem = NULL;
1295 	struct siw_ureq_reg_mr ureq;
1296 	struct siw_device *sdev = to_siw_dev(pd->device);
1297 
1298 	unsigned long mem_limit = rlimit(RLIMIT_MEMLOCK);
1299 	int rv;
1300 
1301 	siw_dbg_pd(pd, "start: 0x%pK, va: 0x%pK, len: %llu\n",
1302 		   (void *)(uintptr_t)start, (void *)(uintptr_t)rnic_va,
1303 		   (unsigned long long)len);
1304 
1305 	if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1306 		siw_dbg_pd(pd, "too many mr's\n");
1307 		rv = -ENOMEM;
1308 		goto err_out;
1309 	}
1310 	if (!len) {
1311 		rv = -EINVAL;
1312 		goto err_out;
1313 	}
1314 	if (mem_limit != RLIM_INFINITY) {
1315 		unsigned long num_pages =
1316 			(PAGE_ALIGN(len + (start & ~PAGE_MASK))) >> PAGE_SHIFT;
1317 		mem_limit >>= PAGE_SHIFT;
1318 
1319 		if (num_pages > mem_limit - current->mm->locked_vm) {
1320 			siw_dbg_pd(pd, "pages req %lu, max %lu, lock %lu\n",
1321 				   num_pages, mem_limit,
1322 				   current->mm->locked_vm);
1323 			rv = -ENOMEM;
1324 			goto err_out;
1325 		}
1326 	}
1327 	umem = siw_umem_get(start, len, ib_access_writable(rights));
1328 	if (IS_ERR(umem)) {
1329 		rv = PTR_ERR(umem);
1330 		siw_dbg_pd(pd, "getting user memory failed: %d\n", rv);
1331 		umem = NULL;
1332 		goto err_out;
1333 	}
1334 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1335 	if (!mr) {
1336 		rv = -ENOMEM;
1337 		goto err_out;
1338 	}
1339 	rv = siw_mr_add_mem(mr, pd, umem, start, len, rights);
1340 	if (rv)
1341 		goto err_out;
1342 
1343 	if (udata) {
1344 		struct siw_uresp_reg_mr uresp = {};
1345 		struct siw_mem *mem = mr->mem;
1346 
1347 		if (udata->inlen < sizeof(ureq)) {
1348 			rv = -EINVAL;
1349 			goto err_out;
1350 		}
1351 		rv = ib_copy_from_udata(&ureq, udata, sizeof(ureq));
1352 		if (rv)
1353 			goto err_out;
1354 
1355 		mr->base_mr.lkey |= ureq.stag_key;
1356 		mr->base_mr.rkey |= ureq.stag_key;
1357 		mem->stag |= ureq.stag_key;
1358 		uresp.stag = mem->stag;
1359 
1360 		if (udata->outlen < sizeof(uresp)) {
1361 			rv = -EINVAL;
1362 			goto err_out;
1363 		}
1364 		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1365 		if (rv)
1366 			goto err_out;
1367 	}
1368 	mr->mem->stag_valid = 1;
1369 
1370 	return &mr->base_mr;
1371 
1372 err_out:
1373 	atomic_dec(&sdev->num_mr);
1374 	if (mr) {
1375 		if (mr->mem)
1376 			siw_mr_drop_mem(mr);
1377 		kfree_rcu(mr, rcu);
1378 	} else {
1379 		if (umem)
1380 			siw_umem_release(umem, false);
1381 	}
1382 	return ERR_PTR(rv);
1383 }
1384 
1385 struct ib_mr *siw_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
1386 			   u32 max_sge)
1387 {
1388 	struct siw_device *sdev = to_siw_dev(pd->device);
1389 	struct siw_mr *mr = NULL;
1390 	struct siw_pbl *pbl = NULL;
1391 	int rv;
1392 
1393 	if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1394 		siw_dbg_pd(pd, "too many mr's\n");
1395 		rv = -ENOMEM;
1396 		goto err_out;
1397 	}
1398 	if (mr_type != IB_MR_TYPE_MEM_REG) {
1399 		siw_dbg_pd(pd, "mr type %d unsupported\n", mr_type);
1400 		rv = -EOPNOTSUPP;
1401 		goto err_out;
1402 	}
1403 	if (max_sge > SIW_MAX_SGE_PBL) {
1404 		siw_dbg_pd(pd, "too many sge's: %d\n", max_sge);
1405 		rv = -ENOMEM;
1406 		goto err_out;
1407 	}
1408 	pbl = siw_pbl_alloc(max_sge);
1409 	if (IS_ERR(pbl)) {
1410 		rv = PTR_ERR(pbl);
1411 		siw_dbg_pd(pd, "pbl allocation failed: %d\n", rv);
1412 		pbl = NULL;
1413 		goto err_out;
1414 	}
1415 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1416 	if (!mr) {
1417 		rv = -ENOMEM;
1418 		goto err_out;
1419 	}
1420 	rv = siw_mr_add_mem(mr, pd, pbl, 0, max_sge * PAGE_SIZE, 0);
1421 	if (rv)
1422 		goto err_out;
1423 
1424 	mr->mem->is_pbl = 1;
1425 
1426 	siw_dbg_pd(pd, "[MEM %u]: success\n", mr->mem->stag);
1427 
1428 	return &mr->base_mr;
1429 
1430 err_out:
1431 	atomic_dec(&sdev->num_mr);
1432 
1433 	if (!mr) {
1434 		kfree(pbl);
1435 	} else {
1436 		if (mr->mem)
1437 			siw_mr_drop_mem(mr);
1438 		kfree_rcu(mr, rcu);
1439 	}
1440 	siw_dbg_pd(pd, "failed: %d\n", rv);
1441 
1442 	return ERR_PTR(rv);
1443 }
1444 
1445 /* Just used to count number of pages being mapped */
1446 static int siw_set_pbl_page(struct ib_mr *base_mr, u64 buf_addr)
1447 {
1448 	return 0;
1449 }
1450 
1451 int siw_map_mr_sg(struct ib_mr *base_mr, struct scatterlist *sl, int num_sle,
1452 		  unsigned int *sg_off)
1453 {
1454 	struct scatterlist *slp;
1455 	struct siw_mr *mr = to_siw_mr(base_mr);
1456 	struct siw_mem *mem = mr->mem;
1457 	struct siw_pbl *pbl = mem->pbl;
1458 	struct siw_pble *pble;
1459 	unsigned long pbl_size;
1460 	int i, rv;
1461 
1462 	if (!pbl) {
1463 		siw_dbg_mem(mem, "no PBL allocated\n");
1464 		return -EINVAL;
1465 	}
1466 	pble = pbl->pbe;
1467 
1468 	if (pbl->max_buf < num_sle) {
1469 		siw_dbg_mem(mem, "too many SGE's: %d > %d\n",
1470 			    mem->pbl->max_buf, num_sle);
1471 		return -ENOMEM;
1472 	}
1473 	for_each_sg(sl, slp, num_sle, i) {
1474 		if (sg_dma_len(slp) == 0) {
1475 			siw_dbg_mem(mem, "empty SGE\n");
1476 			return -EINVAL;
1477 		}
1478 		if (i == 0) {
1479 			pble->addr = sg_dma_address(slp);
1480 			pble->size = sg_dma_len(slp);
1481 			pble->pbl_off = 0;
1482 			pbl_size = pble->size;
1483 			pbl->num_buf = 1;
1484 		} else {
1485 			/* Merge PBL entries if adjacent */
1486 			if (pble->addr + pble->size == sg_dma_address(slp)) {
1487 				pble->size += sg_dma_len(slp);
1488 			} else {
1489 				pble++;
1490 				pbl->num_buf++;
1491 				pble->addr = sg_dma_address(slp);
1492 				pble->size = sg_dma_len(slp);
1493 				pble->pbl_off = pbl_size;
1494 			}
1495 			pbl_size += sg_dma_len(slp);
1496 		}
1497 		siw_dbg_mem(mem,
1498 			"sge[%d], size %u, addr 0x%p, total %lu\n",
1499 			i, pble->size, (void *)(uintptr_t)pble->addr,
1500 			pbl_size);
1501 	}
1502 	rv = ib_sg_to_pages(base_mr, sl, num_sle, sg_off, siw_set_pbl_page);
1503 	if (rv > 0) {
1504 		mem->len = base_mr->length;
1505 		mem->va = base_mr->iova;
1506 		siw_dbg_mem(mem,
1507 			"%llu bytes, start 0x%pK, %u SLE to %u entries\n",
1508 			mem->len, (void *)(uintptr_t)mem->va, num_sle,
1509 			pbl->num_buf);
1510 	}
1511 	return rv;
1512 }
1513 
1514 /*
1515  * siw_get_dma_mr()
1516  *
1517  * Create a (empty) DMA memory region, where no umem is attached.
1518  */
1519 struct ib_mr *siw_get_dma_mr(struct ib_pd *pd, int rights)
1520 {
1521 	struct siw_device *sdev = to_siw_dev(pd->device);
1522 	struct siw_mr *mr = NULL;
1523 	int rv;
1524 
1525 	if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1526 		siw_dbg_pd(pd, "too many mr's\n");
1527 		rv = -ENOMEM;
1528 		goto err_out;
1529 	}
1530 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1531 	if (!mr) {
1532 		rv = -ENOMEM;
1533 		goto err_out;
1534 	}
1535 	rv = siw_mr_add_mem(mr, pd, NULL, 0, ULONG_MAX, rights);
1536 	if (rv)
1537 		goto err_out;
1538 
1539 	mr->mem->stag_valid = 1;
1540 
1541 	siw_dbg_pd(pd, "[MEM %u]: success\n", mr->mem->stag);
1542 
1543 	return &mr->base_mr;
1544 
1545 err_out:
1546 	if (rv)
1547 		kfree(mr);
1548 
1549 	atomic_dec(&sdev->num_mr);
1550 
1551 	return ERR_PTR(rv);
1552 }
1553 
1554 /*
1555  * siw_create_srq()
1556  *
1557  * Create Shared Receive Queue of attributes @init_attrs
1558  * within protection domain given by @pd.
1559  *
1560  * @base_srq:	Base SRQ contained in siw SRQ.
1561  * @init_attrs:	SRQ init attributes.
1562  * @udata:	points to user context
1563  */
1564 int siw_create_srq(struct ib_srq *base_srq,
1565 		   struct ib_srq_init_attr *init_attrs, struct ib_udata *udata)
1566 {
1567 	struct siw_srq *srq = to_siw_srq(base_srq);
1568 	struct ib_srq_attr *attrs = &init_attrs->attr;
1569 	struct siw_device *sdev = to_siw_dev(base_srq->device);
1570 	struct siw_ucontext *ctx =
1571 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
1572 					  base_ucontext);
1573 	int rv;
1574 
1575 	if (init_attrs->srq_type != IB_SRQT_BASIC)
1576 		return -EOPNOTSUPP;
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 = true;
1593 
1594 	srq->is_kernel_res = !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 = true;
1675 		} else {
1676 			srq->armed = false;
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 int 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 	return 0;
1728 }
1729 
1730 /*
1731  * siw_post_srq_recv()
1732  *
1733  * Post a list of receive queue elements to SRQ.
1734  * NOTE: The function does not check or lock a certain SRQ state
1735  *       during the post operation. The code simply trusts the
1736  *       RDMA core environment.
1737  *
1738  * @base_srq:	Base SRQ contained in siw SRQ
1739  * @wr:		List of R-WR's
1740  * @bad_wr:	Updated to failing WR if posting fails.
1741  */
1742 int siw_post_srq_recv(struct ib_srq *base_srq, const struct ib_recv_wr *wr,
1743 		      const struct ib_recv_wr **bad_wr)
1744 {
1745 	struct siw_srq *srq = to_siw_srq(base_srq);
1746 	unsigned long flags;
1747 	int rv = 0;
1748 
1749 	if (unlikely(!srq->is_kernel_res)) {
1750 		siw_dbg_pd(base_srq->pd,
1751 			   "[SRQ]: no kernel post_recv for mapped srq\n");
1752 		rv = -EINVAL;
1753 		goto out;
1754 	}
1755 	/*
1756 	 * Serialize potentially multiple producers.
1757 	 * Also needed to serialize potentially multiple
1758 	 * consumers.
1759 	 */
1760 	spin_lock_irqsave(&srq->lock, flags);
1761 
1762 	while (wr) {
1763 		u32 idx = srq->rq_put % srq->num_rqe;
1764 		struct siw_rqe *rqe = &srq->recvq[idx];
1765 
1766 		if (rqe->flags) {
1767 			siw_dbg_pd(base_srq->pd, "SRQ full\n");
1768 			rv = -ENOMEM;
1769 			break;
1770 		}
1771 		if (unlikely(wr->num_sge > srq->max_sge)) {
1772 			siw_dbg_pd(base_srq->pd,
1773 				   "[SRQ]: too many sge's: %d\n", wr->num_sge);
1774 			rv = -EINVAL;
1775 			break;
1776 		}
1777 		rqe->id = wr->wr_id;
1778 		rqe->num_sge = wr->num_sge;
1779 		siw_copy_sgl(wr->sg_list, rqe->sge, wr->num_sge);
1780 
1781 		/* Make sure S-RQE is completely written before valid */
1782 		smp_wmb();
1783 
1784 		rqe->flags = SIW_WQE_VALID;
1785 
1786 		srq->rq_put++;
1787 		wr = wr->next;
1788 	}
1789 	spin_unlock_irqrestore(&srq->lock, flags);
1790 out:
1791 	if (unlikely(rv < 0)) {
1792 		siw_dbg_pd(base_srq->pd, "[SRQ]: error %d\n", rv);
1793 		*bad_wr = wr;
1794 	}
1795 	return rv;
1796 }
1797 
1798 void siw_qp_event(struct siw_qp *qp, enum ib_event_type etype)
1799 {
1800 	struct ib_event event;
1801 	struct ib_qp *base_qp = &qp->base_qp;
1802 
1803 	/*
1804 	 * Do not report asynchronous errors on QP which gets
1805 	 * destroyed via verbs interface (siw_destroy_qp())
1806 	 */
1807 	if (qp->attrs.flags & SIW_QP_IN_DESTROY)
1808 		return;
1809 
1810 	event.event = etype;
1811 	event.device = base_qp->device;
1812 	event.element.qp = base_qp;
1813 
1814 	if (base_qp->event_handler) {
1815 		siw_dbg_qp(qp, "reporting event %d\n", etype);
1816 		base_qp->event_handler(&event, base_qp->qp_context);
1817 	}
1818 }
1819 
1820 void siw_cq_event(struct siw_cq *cq, enum ib_event_type etype)
1821 {
1822 	struct ib_event event;
1823 	struct ib_cq *base_cq = &cq->base_cq;
1824 
1825 	event.event = etype;
1826 	event.device = base_cq->device;
1827 	event.element.cq = base_cq;
1828 
1829 	if (base_cq->event_handler) {
1830 		siw_dbg_cq(cq, "reporting CQ event %d\n", etype);
1831 		base_cq->event_handler(&event, base_cq->cq_context);
1832 	}
1833 }
1834 
1835 void siw_srq_event(struct siw_srq *srq, enum ib_event_type etype)
1836 {
1837 	struct ib_event event;
1838 	struct ib_srq *base_srq = &srq->base_srq;
1839 
1840 	event.event = etype;
1841 	event.device = base_srq->device;
1842 	event.element.srq = base_srq;
1843 
1844 	if (base_srq->event_handler) {
1845 		siw_dbg_pd(srq->base_srq.pd,
1846 			   "reporting SRQ event %d\n", etype);
1847 		base_srq->event_handler(&event, base_srq->srq_context);
1848 	}
1849 }
1850 
1851 void siw_port_event(struct siw_device *sdev, u8 port, enum ib_event_type etype)
1852 {
1853 	struct ib_event event;
1854 
1855 	event.event = etype;
1856 	event.device = &sdev->base_dev;
1857 	event.element.port_num = port;
1858 
1859 	siw_dbg(&sdev->base_dev, "reporting port event %d\n", etype);
1860 
1861 	ib_dispatch_event(&event);
1862 }
1863