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 static u32 siw_create_uobj(struct siw_ucontext *uctx, void *vaddr, u32 size)
38 {
39 	struct siw_uobj *uobj;
40 	struct xa_limit limit = XA_LIMIT(0, SIW_UOBJ_MAX_KEY);
41 	u32 key;
42 
43 	uobj = kzalloc(sizeof(*uobj), GFP_KERNEL);
44 	if (!uobj)
45 		return SIW_INVAL_UOBJ_KEY;
46 
47 	if (xa_alloc_cyclic(&uctx->xa, &key, uobj, limit, &uctx->uobj_nextkey,
48 			    GFP_KERNEL) < 0) {
49 		kfree(uobj);
50 		return SIW_INVAL_UOBJ_KEY;
51 	}
52 	uobj->size = PAGE_ALIGN(size);
53 	uobj->addr = vaddr;
54 
55 	return key;
56 }
57 
58 static struct siw_uobj *siw_get_uobj(struct siw_ucontext *uctx,
59 				     unsigned long off, u32 size)
60 {
61 	struct siw_uobj *uobj = xa_load(&uctx->xa, off);
62 
63 	if (uobj && uobj->size == size)
64 		return uobj;
65 
66 	return NULL;
67 }
68 
69 int siw_mmap(struct ib_ucontext *ctx, struct vm_area_struct *vma)
70 {
71 	struct siw_ucontext *uctx = to_siw_ctx(ctx);
72 	struct siw_uobj *uobj;
73 	unsigned long off = vma->vm_pgoff;
74 	int size = vma->vm_end - vma->vm_start;
75 	int rv = -EINVAL;
76 
77 	/*
78 	 * Must be page aligned
79 	 */
80 	if (vma->vm_start & (PAGE_SIZE - 1)) {
81 		pr_warn("siw: mmap not page aligned\n");
82 		goto out;
83 	}
84 	uobj = siw_get_uobj(uctx, off, size);
85 	if (!uobj) {
86 		siw_dbg(&uctx->sdev->base_dev, "mmap lookup failed: %lu, %u\n",
87 			off, size);
88 		goto out;
89 	}
90 	rv = remap_vmalloc_range(vma, uobj->addr, 0);
91 	if (rv)
92 		pr_warn("remap_vmalloc_range failed: %lu, %u\n", off, size);
93 out:
94 	return rv;
95 }
96 
97 int siw_alloc_ucontext(struct ib_ucontext *base_ctx, struct ib_udata *udata)
98 {
99 	struct siw_device *sdev = to_siw_dev(base_ctx->device);
100 	struct siw_ucontext *ctx = to_siw_ctx(base_ctx);
101 	struct siw_uresp_alloc_ctx uresp = {};
102 	int rv;
103 
104 	if (atomic_inc_return(&sdev->num_ctx) > SIW_MAX_CONTEXT) {
105 		rv = -ENOMEM;
106 		goto err_out;
107 	}
108 	xa_init_flags(&ctx->xa, XA_FLAGS_ALLOC);
109 	ctx->uobj_nextkey = 0;
110 	ctx->sdev = sdev;
111 
112 	uresp.dev_id = sdev->vendor_part_id;
113 
114 	if (udata->outlen < sizeof(uresp)) {
115 		rv = -EINVAL;
116 		goto err_out;
117 	}
118 	rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
119 	if (rv)
120 		goto err_out;
121 
122 	siw_dbg(base_ctx->device, "success. now %d context(s)\n",
123 		atomic_read(&sdev->num_ctx));
124 
125 	return 0;
126 
127 err_out:
128 	atomic_dec(&sdev->num_ctx);
129 	siw_dbg(base_ctx->device, "failure %d. now %d context(s)\n", rv,
130 		atomic_read(&sdev->num_ctx));
131 
132 	return rv;
133 }
134 
135 void siw_dealloc_ucontext(struct ib_ucontext *base_ctx)
136 {
137 	struct siw_ucontext *uctx = to_siw_ctx(base_ctx);
138 	void *entry;
139 	unsigned long index;
140 
141 	/*
142 	 * Make sure all user mmap objects are gone. Since QP, CQ
143 	 * and SRQ destroy routines destroy related objects, nothing
144 	 * should be found here.
145 	 */
146 	xa_for_each(&uctx->xa, index, entry) {
147 		kfree(xa_erase(&uctx->xa, index));
148 		pr_warn("siw: dropping orphaned uobj at %lu\n", index);
149 	}
150 	xa_destroy(&uctx->xa);
151 	atomic_dec(&uctx->sdev->num_ctx);
152 }
153 
154 int siw_query_device(struct ib_device *base_dev, struct ib_device_attr *attr,
155 		     struct ib_udata *udata)
156 {
157 	struct siw_device *sdev = to_siw_dev(base_dev);
158 
159 	if (udata->inlen || udata->outlen)
160 		return -EINVAL;
161 
162 	memset(attr, 0, sizeof(*attr));
163 
164 	/* Revisit atomic caps if RFC 7306 gets supported */
165 	attr->atomic_cap = 0;
166 	attr->device_cap_flags =
167 		IB_DEVICE_MEM_MGT_EXTENSIONS | IB_DEVICE_ALLOW_USER_UNREG;
168 	attr->max_cq = sdev->attrs.max_cq;
169 	attr->max_cqe = sdev->attrs.max_cqe;
170 	attr->max_fast_reg_page_list_len = SIW_MAX_SGE_PBL;
171 	attr->max_fmr = sdev->attrs.max_fmr;
172 	attr->max_mr = sdev->attrs.max_mr;
173 	attr->max_mw = sdev->attrs.max_mw;
174 	attr->max_mr_size = ~0ull;
175 	attr->max_pd = sdev->attrs.max_pd;
176 	attr->max_qp = sdev->attrs.max_qp;
177 	attr->max_qp_init_rd_atom = sdev->attrs.max_ird;
178 	attr->max_qp_rd_atom = sdev->attrs.max_ord;
179 	attr->max_qp_wr = sdev->attrs.max_qp_wr;
180 	attr->max_recv_sge = sdev->attrs.max_sge;
181 	attr->max_res_rd_atom = sdev->attrs.max_qp * sdev->attrs.max_ird;
182 	attr->max_send_sge = sdev->attrs.max_sge;
183 	attr->max_sge_rd = sdev->attrs.max_sge_rd;
184 	attr->max_srq = sdev->attrs.max_srq;
185 	attr->max_srq_sge = sdev->attrs.max_srq_sge;
186 	attr->max_srq_wr = sdev->attrs.max_srq_wr;
187 	attr->page_size_cap = PAGE_SIZE;
188 	attr->vendor_id = SIW_VENDOR_ID;
189 	attr->vendor_part_id = sdev->vendor_part_id;
190 
191 	memcpy(&attr->sys_image_guid, sdev->netdev->dev_addr, 6);
192 
193 	return 0;
194 }
195 
196 int siw_query_port(struct ib_device *base_dev, u8 port,
197 		   struct ib_port_attr *attr)
198 {
199 	struct siw_device *sdev = to_siw_dev(base_dev);
200 
201 	memset(attr, 0, sizeof(*attr));
202 
203 	attr->active_mtu = attr->max_mtu;
204 	attr->active_speed = 2;
205 	attr->active_width = 2;
206 	attr->gid_tbl_len = 1;
207 	attr->max_msg_sz = -1;
208 	attr->max_mtu = ib_mtu_int_to_enum(sdev->netdev->mtu);
209 	attr->phys_state = sdev->state == IB_PORT_ACTIVE ? 5 : 3;
210 	attr->pkey_tbl_len = 1;
211 	attr->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_DEVICE_MGMT_SUP;
212 	attr->state = sdev->state;
213 	/*
214 	 * All zero
215 	 *
216 	 * attr->lid = 0;
217 	 * attr->bad_pkey_cntr = 0;
218 	 * attr->qkey_viol_cntr = 0;
219 	 * attr->sm_lid = 0;
220 	 * attr->lmc = 0;
221 	 * attr->max_vl_num = 0;
222 	 * attr->sm_sl = 0;
223 	 * attr->subnet_timeout = 0;
224 	 * attr->init_type_repy = 0;
225 	 */
226 	return 0;
227 }
228 
229 int siw_get_port_immutable(struct ib_device *base_dev, u8 port,
230 			   struct ib_port_immutable *port_immutable)
231 {
232 	struct ib_port_attr attr;
233 	int rv = siw_query_port(base_dev, port, &attr);
234 
235 	if (rv)
236 		return rv;
237 
238 	port_immutable->pkey_tbl_len = attr.pkey_tbl_len;
239 	port_immutable->gid_tbl_len = attr.gid_tbl_len;
240 	port_immutable->core_cap_flags = RDMA_CORE_PORT_IWARP;
241 
242 	return 0;
243 }
244 
245 int siw_query_pkey(struct ib_device *base_dev, u8 port, u16 idx, u16 *pkey)
246 {
247 	/* Report the default pkey */
248 	*pkey = 0xffff;
249 	return 0;
250 }
251 
252 int siw_query_gid(struct ib_device *base_dev, u8 port, int idx,
253 		  union ib_gid *gid)
254 {
255 	struct siw_device *sdev = to_siw_dev(base_dev);
256 
257 	/* subnet_prefix == interface_id == 0; */
258 	memset(gid, 0, sizeof(*gid));
259 	memcpy(&gid->raw[0], sdev->netdev->dev_addr, 6);
260 
261 	return 0;
262 }
263 
264 int siw_alloc_pd(struct ib_pd *pd, struct ib_udata *udata)
265 {
266 	struct siw_device *sdev = to_siw_dev(pd->device);
267 
268 	if (atomic_inc_return(&sdev->num_pd) > SIW_MAX_PD) {
269 		atomic_dec(&sdev->num_pd);
270 		return -ENOMEM;
271 	}
272 	siw_dbg_pd(pd, "now %d PD's(s)\n", atomic_read(&sdev->num_pd));
273 
274 	return 0;
275 }
276 
277 void siw_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
278 {
279 	struct siw_device *sdev = to_siw_dev(pd->device);
280 
281 	siw_dbg_pd(pd, "free PD\n");
282 	atomic_dec(&sdev->num_pd);
283 }
284 
285 void siw_qp_get_ref(struct ib_qp *base_qp)
286 {
287 	siw_qp_get(to_siw_qp(base_qp));
288 }
289 
290 void siw_qp_put_ref(struct ib_qp *base_qp)
291 {
292 	siw_qp_put(to_siw_qp(base_qp));
293 }
294 
295 /*
296  * siw_create_qp()
297  *
298  * Create QP of requested size on given device.
299  *
300  * @pd:		Protection Domain
301  * @attrs:	Initial QP attributes.
302  * @udata:	used to provide QP ID, SQ and RQ size back to user.
303  */
304 
305 struct ib_qp *siw_create_qp(struct ib_pd *pd,
306 			    struct ib_qp_init_attr *attrs,
307 			    struct ib_udata *udata)
308 {
309 	struct siw_qp *qp = NULL;
310 	struct siw_base_qp *siw_base_qp = NULL;
311 	struct ib_device *base_dev = pd->device;
312 	struct siw_device *sdev = to_siw_dev(base_dev);
313 	struct siw_ucontext *uctx =
314 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
315 					  base_ucontext);
316 	struct siw_cq *scq = NULL, *rcq = NULL;
317 	unsigned long flags;
318 	int num_sqe, num_rqe, rv = 0;
319 
320 	siw_dbg(base_dev, "create new QP\n");
321 
322 	if (atomic_inc_return(&sdev->num_qp) > SIW_MAX_QP) {
323 		siw_dbg(base_dev, "too many QP's\n");
324 		rv = -ENOMEM;
325 		goto err_out;
326 	}
327 	if (attrs->qp_type != IB_QPT_RC) {
328 		siw_dbg(base_dev, "only RC QP's supported\n");
329 		rv = -EINVAL;
330 		goto err_out;
331 	}
332 	if ((attrs->cap.max_send_wr > SIW_MAX_QP_WR) ||
333 	    (attrs->cap.max_recv_wr > SIW_MAX_QP_WR) ||
334 	    (attrs->cap.max_send_sge > SIW_MAX_SGE) ||
335 	    (attrs->cap.max_recv_sge > SIW_MAX_SGE)) {
336 		siw_dbg(base_dev, "QP size error\n");
337 		rv = -EINVAL;
338 		goto err_out;
339 	}
340 	if (attrs->cap.max_inline_data > SIW_MAX_INLINE) {
341 		siw_dbg(base_dev, "max inline send: %d > %d\n",
342 			attrs->cap.max_inline_data, (int)SIW_MAX_INLINE);
343 		rv = -EINVAL;
344 		goto err_out;
345 	}
346 	/*
347 	 * NOTE: we allow for zero element SQ and RQ WQE's SGL's
348 	 * but not for a QP unable to hold any WQE (SQ + RQ)
349 	 */
350 	if (attrs->cap.max_send_wr + attrs->cap.max_recv_wr == 0) {
351 		siw_dbg(base_dev, "QP must have send or receive queue\n");
352 		rv = -EINVAL;
353 		goto err_out;
354 	}
355 	scq = to_siw_cq(attrs->send_cq);
356 	rcq = to_siw_cq(attrs->recv_cq);
357 
358 	if (!scq || (!rcq && !attrs->srq)) {
359 		siw_dbg(base_dev, "send CQ or receive CQ invalid\n");
360 		rv = -EINVAL;
361 		goto err_out;
362 	}
363 	siw_base_qp = kzalloc(sizeof(*siw_base_qp), GFP_KERNEL);
364 	if (!siw_base_qp) {
365 		rv = -ENOMEM;
366 		goto err_out;
367 	}
368 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
369 	if (!qp) {
370 		rv = -ENOMEM;
371 		goto err_out;
372 	}
373 	siw_base_qp->qp = qp;
374 	qp->ib_qp = &siw_base_qp->base_qp;
375 
376 	init_rwsem(&qp->state_lock);
377 	spin_lock_init(&qp->sq_lock);
378 	spin_lock_init(&qp->rq_lock);
379 	spin_lock_init(&qp->orq_lock);
380 
381 	qp->kernel_verbs = !udata;
382 	qp->xa_sq_index = SIW_INVAL_UOBJ_KEY;
383 	qp->xa_rq_index = SIW_INVAL_UOBJ_KEY;
384 
385 	rv = siw_qp_add(sdev, qp);
386 	if (rv)
387 		goto err_out;
388 
389 	/* All queue indices are derived from modulo operations
390 	 * on a free running 'get' (consumer) and 'put' (producer)
391 	 * unsigned counter. Having queue sizes at power of two
392 	 * avoids handling counter wrap around.
393 	 */
394 	num_sqe = roundup_pow_of_two(attrs->cap.max_send_wr);
395 	num_rqe = roundup_pow_of_two(attrs->cap.max_recv_wr);
396 
397 	if (qp->kernel_verbs)
398 		qp->sendq = vzalloc(num_sqe * sizeof(struct siw_sqe));
399 	else
400 		qp->sendq = vmalloc_user(num_sqe * sizeof(struct siw_sqe));
401 
402 	if (qp->sendq == NULL) {
403 		siw_dbg(base_dev, "SQ size %d alloc failed\n", num_sqe);
404 		rv = -ENOMEM;
405 		goto err_out_xa;
406 	}
407 	if (attrs->sq_sig_type != IB_SIGNAL_REQ_WR) {
408 		if (attrs->sq_sig_type == IB_SIGNAL_ALL_WR)
409 			qp->attrs.flags |= SIW_SIGNAL_ALL_WR;
410 		else {
411 			rv = -EINVAL;
412 			goto err_out_xa;
413 		}
414 	}
415 	qp->pd = pd;
416 	qp->scq = scq;
417 	qp->rcq = rcq;
418 
419 	if (attrs->srq) {
420 		/*
421 		 * SRQ support.
422 		 * Verbs 6.3.7: ignore RQ size, if SRQ present
423 		 * Verbs 6.3.5: do not check PD of SRQ against PD of QP
424 		 */
425 		qp->srq = to_siw_srq(attrs->srq);
426 		qp->attrs.rq_size = 0;
427 		siw_dbg(base_dev, "QP [%u]: SRQ attached\n", qp->qp_num);
428 	} else if (num_rqe) {
429 		if (qp->kernel_verbs)
430 			qp->recvq = vzalloc(num_rqe * sizeof(struct siw_rqe));
431 		else
432 			qp->recvq =
433 				vmalloc_user(num_rqe * sizeof(struct siw_rqe));
434 
435 		if (qp->recvq == NULL) {
436 			siw_dbg(base_dev, "RQ size %d alloc failed\n", num_rqe);
437 			rv = -ENOMEM;
438 			goto err_out_xa;
439 		}
440 		qp->attrs.rq_size = num_rqe;
441 	}
442 	qp->attrs.sq_size = num_sqe;
443 	qp->attrs.sq_max_sges = attrs->cap.max_send_sge;
444 	qp->attrs.rq_max_sges = attrs->cap.max_recv_sge;
445 
446 	/* Make those two tunables fixed for now. */
447 	qp->tx_ctx.gso_seg_limit = 1;
448 	qp->tx_ctx.zcopy_tx = zcopy_tx;
449 
450 	qp->attrs.state = SIW_QP_STATE_IDLE;
451 
452 	if (udata) {
453 		struct siw_uresp_create_qp uresp = {};
454 
455 		uresp.num_sqe = num_sqe;
456 		uresp.num_rqe = num_rqe;
457 		uresp.qp_id = qp_id(qp);
458 
459 		if (qp->sendq) {
460 			qp->xa_sq_index =
461 				siw_create_uobj(uctx, qp->sendq,
462 					num_sqe * sizeof(struct siw_sqe));
463 		}
464 		if (qp->recvq) {
465 			qp->xa_rq_index =
466 				 siw_create_uobj(uctx, qp->recvq,
467 					num_rqe * sizeof(struct siw_rqe));
468 		}
469 		if (qp->xa_sq_index == SIW_INVAL_UOBJ_KEY ||
470 		    qp->xa_rq_index == SIW_INVAL_UOBJ_KEY) {
471 			rv = -ENOMEM;
472 			goto err_out_xa;
473 		}
474 		uresp.sq_key = qp->xa_sq_index << PAGE_SHIFT;
475 		uresp.rq_key = qp->xa_rq_index << PAGE_SHIFT;
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 (qp->xa_sq_index != SIW_INVAL_UOBJ_KEY)
504 			kfree(xa_erase(&uctx->xa, qp->xa_sq_index));
505 		if (qp->xa_rq_index != SIW_INVAL_UOBJ_KEY)
506 			kfree(xa_erase(&uctx->xa, qp->xa_rq_index));
507 
508 		vfree(qp->sendq);
509 		vfree(qp->recvq);
510 		kfree(qp);
511 	}
512 	atomic_dec(&sdev->num_qp);
513 
514 	return ERR_PTR(rv);
515 }
516 
517 /*
518  * Minimum siw_query_qp() verb interface.
519  *
520  * @qp_attr_mask is not used but all available information is provided
521  */
522 int siw_query_qp(struct ib_qp *base_qp, struct ib_qp_attr *qp_attr,
523 		 int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
524 {
525 	struct siw_qp *qp;
526 	struct siw_device *sdev;
527 
528 	if (base_qp && qp_attr && qp_init_attr) {
529 		qp = to_siw_qp(base_qp);
530 		sdev = to_siw_dev(base_qp->device);
531 	} else {
532 		return -EINVAL;
533 	}
534 	qp_attr->cap.max_inline_data = SIW_MAX_INLINE;
535 	qp_attr->cap.max_send_wr = qp->attrs.sq_size;
536 	qp_attr->cap.max_send_sge = qp->attrs.sq_max_sges;
537 	qp_attr->cap.max_recv_wr = qp->attrs.rq_size;
538 	qp_attr->cap.max_recv_sge = qp->attrs.rq_max_sges;
539 	qp_attr->path_mtu = ib_mtu_int_to_enum(sdev->netdev->mtu);
540 	qp_attr->max_rd_atomic = qp->attrs.irq_size;
541 	qp_attr->max_dest_rd_atomic = qp->attrs.orq_size;
542 
543 	qp_attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE |
544 				   IB_ACCESS_REMOTE_WRITE |
545 				   IB_ACCESS_REMOTE_READ;
546 
547 	qp_init_attr->qp_type = base_qp->qp_type;
548 	qp_init_attr->send_cq = base_qp->send_cq;
549 	qp_init_attr->recv_cq = base_qp->recv_cq;
550 	qp_init_attr->srq = base_qp->srq;
551 
552 	qp_init_attr->cap = qp_attr->cap;
553 
554 	return 0;
555 }
556 
557 int siw_verbs_modify_qp(struct ib_qp *base_qp, struct ib_qp_attr *attr,
558 			int attr_mask, struct ib_udata *udata)
559 {
560 	struct siw_qp_attrs new_attrs;
561 	enum siw_qp_attr_mask siw_attr_mask = 0;
562 	struct siw_qp *qp = to_siw_qp(base_qp);
563 	int rv = 0;
564 
565 	if (!attr_mask)
566 		return 0;
567 
568 	memset(&new_attrs, 0, sizeof(new_attrs));
569 
570 	if (attr_mask & IB_QP_ACCESS_FLAGS) {
571 		siw_attr_mask = SIW_QP_ATTR_ACCESS_FLAGS;
572 
573 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
574 			new_attrs.flags |= SIW_RDMA_READ_ENABLED;
575 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
576 			new_attrs.flags |= SIW_RDMA_WRITE_ENABLED;
577 		if (attr->qp_access_flags & IB_ACCESS_MW_BIND)
578 			new_attrs.flags |= SIW_RDMA_BIND_ENABLED;
579 	}
580 	if (attr_mask & IB_QP_STATE) {
581 		siw_dbg_qp(qp, "desired IB QP state: %s\n",
582 			   ib_qp_state_to_string[attr->qp_state]);
583 
584 		new_attrs.state = ib_qp_state_to_siw_qp_state[attr->qp_state];
585 
586 		if (new_attrs.state > SIW_QP_STATE_RTS)
587 			qp->tx_ctx.tx_suspend = 1;
588 
589 		siw_attr_mask |= SIW_QP_ATTR_STATE;
590 	}
591 	if (!siw_attr_mask)
592 		goto out;
593 
594 	down_write(&qp->state_lock);
595 
596 	rv = siw_qp_modify(qp, &new_attrs, siw_attr_mask);
597 
598 	up_write(&qp->state_lock);
599 out:
600 	return rv;
601 }
602 
603 int siw_destroy_qp(struct ib_qp *base_qp, struct ib_udata *udata)
604 {
605 	struct siw_qp *qp = to_siw_qp(base_qp);
606 	struct siw_base_qp *siw_base_qp = to_siw_base_qp(base_qp);
607 	struct siw_ucontext *uctx =
608 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
609 					  base_ucontext);
610 	struct siw_qp_attrs qp_attrs;
611 
612 	siw_dbg_qp(qp, "state %d\n", qp->attrs.state);
613 
614 	/*
615 	 * Mark QP as in process of destruction to prevent from
616 	 * any async callbacks to RDMA core
617 	 */
618 	qp->attrs.flags |= SIW_QP_IN_DESTROY;
619 	qp->rx_stream.rx_suspend = 1;
620 
621 	if (uctx && qp->xa_sq_index != SIW_INVAL_UOBJ_KEY)
622 		kfree(xa_erase(&uctx->xa, qp->xa_sq_index));
623 	if (uctx && qp->xa_rq_index != SIW_INVAL_UOBJ_KEY)
624 		kfree(xa_erase(&uctx->xa, qp->xa_rq_index));
625 
626 	down_write(&qp->state_lock);
627 
628 	qp_attrs.state = SIW_QP_STATE_ERROR;
629 	siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE);
630 
631 	if (qp->cep) {
632 		siw_cep_put(qp->cep);
633 		qp->cep = NULL;
634 	}
635 	up_write(&qp->state_lock);
636 
637 	kfree(qp->tx_ctx.mpa_crc_hd);
638 	kfree(qp->rx_stream.mpa_crc_hd);
639 
640 	qp->scq = qp->rcq = NULL;
641 
642 	siw_qp_put(qp);
643 	kfree(siw_base_qp);
644 
645 	return 0;
646 }
647 
648 /*
649  * siw_copy_inline_sgl()
650  *
651  * Prepare sgl of inlined data for sending. For userland callers
652  * function checks if given buffer addresses and len's are within
653  * process context bounds.
654  * Data from all provided sge's are copied together into the wqe,
655  * referenced by a single sge.
656  */
657 static int siw_copy_inline_sgl(const struct ib_send_wr *core_wr,
658 			       struct siw_sqe *sqe)
659 {
660 	struct ib_sge *core_sge = core_wr->sg_list;
661 	void *kbuf = &sqe->sge[1];
662 	int num_sge = core_wr->num_sge, bytes = 0;
663 
664 	sqe->sge[0].laddr = (uintptr_t)kbuf;
665 	sqe->sge[0].lkey = 0;
666 
667 	while (num_sge--) {
668 		if (!core_sge->length) {
669 			core_sge++;
670 			continue;
671 		}
672 		bytes += core_sge->length;
673 		if (bytes > SIW_MAX_INLINE) {
674 			bytes = -EINVAL;
675 			break;
676 		}
677 		memcpy(kbuf, (void *)(uintptr_t)core_sge->addr,
678 		       core_sge->length);
679 
680 		kbuf += core_sge->length;
681 		core_sge++;
682 	}
683 	sqe->sge[0].length = bytes > 0 ? bytes : 0;
684 	sqe->num_sge = bytes > 0 ? 1 : 0;
685 
686 	return bytes;
687 }
688 
689 /*
690  * siw_post_send()
691  *
692  * Post a list of S-WR's to a SQ.
693  *
694  * @base_qp:	Base QP contained in siw QP
695  * @wr:		Null terminated list of user WR's
696  * @bad_wr:	Points to failing WR in case of synchronous failure.
697  */
698 int siw_post_send(struct ib_qp *base_qp, const struct ib_send_wr *wr,
699 		  const struct ib_send_wr **bad_wr)
700 {
701 	struct siw_qp *qp = to_siw_qp(base_qp);
702 	struct siw_wqe *wqe = tx_wqe(qp);
703 
704 	unsigned long flags;
705 	int rv = 0;
706 
707 	/*
708 	 * Try to acquire QP state lock. Must be non-blocking
709 	 * to accommodate kernel clients needs.
710 	 */
711 	if (!down_read_trylock(&qp->state_lock)) {
712 		*bad_wr = wr;
713 		siw_dbg_qp(qp, "QP locked, state %d\n", qp->attrs.state);
714 		return -ENOTCONN;
715 	}
716 	if (unlikely(qp->attrs.state != SIW_QP_STATE_RTS)) {
717 		up_read(&qp->state_lock);
718 		*bad_wr = wr;
719 		siw_dbg_qp(qp, "QP out of state %d\n", qp->attrs.state);
720 		return -ENOTCONN;
721 	}
722 	if (wr && !qp->kernel_verbs) {
723 		siw_dbg_qp(qp, "wr must be empty for user mapped sq\n");
724 		up_read(&qp->state_lock);
725 		*bad_wr = wr;
726 		return -EINVAL;
727 	}
728 	spin_lock_irqsave(&qp->sq_lock, flags);
729 
730 	while (wr) {
731 		u32 idx = qp->sq_put % qp->attrs.sq_size;
732 		struct siw_sqe *sqe = &qp->sendq[idx];
733 
734 		if (sqe->flags) {
735 			siw_dbg_qp(qp, "sq full\n");
736 			rv = -ENOMEM;
737 			break;
738 		}
739 		if (wr->num_sge > qp->attrs.sq_max_sges) {
740 			siw_dbg_qp(qp, "too many sge's: %d\n", wr->num_sge);
741 			rv = -EINVAL;
742 			break;
743 		}
744 		sqe->id = wr->wr_id;
745 
746 		if ((wr->send_flags & IB_SEND_SIGNALED) ||
747 		    (qp->attrs.flags & SIW_SIGNAL_ALL_WR))
748 			sqe->flags |= SIW_WQE_SIGNALLED;
749 
750 		if (wr->send_flags & IB_SEND_FENCE)
751 			sqe->flags |= SIW_WQE_READ_FENCE;
752 
753 		switch (wr->opcode) {
754 		case IB_WR_SEND:
755 		case IB_WR_SEND_WITH_INV:
756 			if (wr->send_flags & IB_SEND_SOLICITED)
757 				sqe->flags |= SIW_WQE_SOLICITED;
758 
759 			if (!(wr->send_flags & IB_SEND_INLINE)) {
760 				siw_copy_sgl(wr->sg_list, sqe->sge,
761 					     wr->num_sge);
762 				sqe->num_sge = wr->num_sge;
763 			} else {
764 				rv = siw_copy_inline_sgl(wr, sqe);
765 				if (rv <= 0) {
766 					rv = -EINVAL;
767 					break;
768 				}
769 				sqe->flags |= SIW_WQE_INLINE;
770 				sqe->num_sge = 1;
771 			}
772 			if (wr->opcode == IB_WR_SEND)
773 				sqe->opcode = SIW_OP_SEND;
774 			else {
775 				sqe->opcode = SIW_OP_SEND_REMOTE_INV;
776 				sqe->rkey = wr->ex.invalidate_rkey;
777 			}
778 			break;
779 
780 		case IB_WR_RDMA_READ_WITH_INV:
781 		case IB_WR_RDMA_READ:
782 			/*
783 			 * iWarp restricts RREAD sink to SGL containing
784 			 * 1 SGE only. we could relax to SGL with multiple
785 			 * elements referring the SAME ltag or even sending
786 			 * a private per-rreq tag referring to a checked
787 			 * local sgl with MULTIPLE ltag's.
788 			 */
789 			if (unlikely(wr->num_sge != 1)) {
790 				rv = -EINVAL;
791 				break;
792 			}
793 			siw_copy_sgl(wr->sg_list, &sqe->sge[0], 1);
794 			/*
795 			 * NOTE: zero length RREAD is allowed!
796 			 */
797 			sqe->raddr = rdma_wr(wr)->remote_addr;
798 			sqe->rkey = rdma_wr(wr)->rkey;
799 			sqe->num_sge = 1;
800 
801 			if (wr->opcode == IB_WR_RDMA_READ)
802 				sqe->opcode = SIW_OP_READ;
803 			else
804 				sqe->opcode = SIW_OP_READ_LOCAL_INV;
805 			break;
806 
807 		case IB_WR_RDMA_WRITE:
808 			if (!(wr->send_flags & IB_SEND_INLINE)) {
809 				siw_copy_sgl(wr->sg_list, &sqe->sge[0],
810 					     wr->num_sge);
811 				sqe->num_sge = wr->num_sge;
812 			} else {
813 				rv = siw_copy_inline_sgl(wr, sqe);
814 				if (unlikely(rv < 0)) {
815 					rv = -EINVAL;
816 					break;
817 				}
818 				sqe->flags |= SIW_WQE_INLINE;
819 				sqe->num_sge = 1;
820 			}
821 			sqe->raddr = rdma_wr(wr)->remote_addr;
822 			sqe->rkey = rdma_wr(wr)->rkey;
823 			sqe->opcode = SIW_OP_WRITE;
824 			break;
825 
826 		case IB_WR_REG_MR:
827 			sqe->base_mr = (uintptr_t)reg_wr(wr)->mr;
828 			sqe->rkey = reg_wr(wr)->key;
829 			sqe->access = reg_wr(wr)->access & IWARP_ACCESS_MASK;
830 			sqe->opcode = SIW_OP_REG_MR;
831 			break;
832 
833 		case IB_WR_LOCAL_INV:
834 			sqe->rkey = wr->ex.invalidate_rkey;
835 			sqe->opcode = SIW_OP_INVAL_STAG;
836 			break;
837 
838 		default:
839 			siw_dbg_qp(qp, "ib wr type %d unsupported\n",
840 				   wr->opcode);
841 			rv = -EINVAL;
842 			break;
843 		}
844 		siw_dbg_qp(qp, "opcode %d, flags 0x%x, wr_id 0x%pK\n",
845 			   sqe->opcode, sqe->flags,
846 			   (void *)(uintptr_t)sqe->id);
847 
848 		if (unlikely(rv < 0))
849 			break;
850 
851 		/* make SQE only valid after completely written */
852 		smp_wmb();
853 		sqe->flags |= SIW_WQE_VALID;
854 
855 		qp->sq_put++;
856 		wr = wr->next;
857 	}
858 
859 	/*
860 	 * Send directly if SQ processing is not in progress.
861 	 * Eventual immediate errors (rv < 0) do not affect the involved
862 	 * RI resources (Verbs, 8.3.1) and thus do not prevent from SQ
863 	 * processing, if new work is already pending. But rv must be passed
864 	 * to caller.
865 	 */
866 	if (wqe->wr_status != SIW_WR_IDLE) {
867 		spin_unlock_irqrestore(&qp->sq_lock, flags);
868 		goto skip_direct_sending;
869 	}
870 	rv = siw_activate_tx(qp);
871 	spin_unlock_irqrestore(&qp->sq_lock, flags);
872 
873 	if (rv <= 0)
874 		goto skip_direct_sending;
875 
876 	if (qp->kernel_verbs) {
877 		rv = siw_sq_start(qp);
878 	} else {
879 		qp->tx_ctx.in_syscall = 1;
880 
881 		if (siw_qp_sq_process(qp) != 0 && !(qp->tx_ctx.tx_suspend))
882 			siw_qp_cm_drop(qp, 0);
883 
884 		qp->tx_ctx.in_syscall = 0;
885 	}
886 skip_direct_sending:
887 
888 	up_read(&qp->state_lock);
889 
890 	if (rv >= 0)
891 		return 0;
892 	/*
893 	 * Immediate error
894 	 */
895 	siw_dbg_qp(qp, "error %d\n", rv);
896 
897 	*bad_wr = wr;
898 	return rv;
899 }
900 
901 /*
902  * siw_post_receive()
903  *
904  * Post a list of R-WR's to a RQ.
905  *
906  * @base_qp:	Base QP contained in siw QP
907  * @wr:		Null terminated list of user WR's
908  * @bad_wr:	Points to failing WR in case of synchronous failure.
909  */
910 int siw_post_receive(struct ib_qp *base_qp, const struct ib_recv_wr *wr,
911 		     const struct ib_recv_wr **bad_wr)
912 {
913 	struct siw_qp *qp = to_siw_qp(base_qp);
914 	unsigned long flags;
915 	int rv = 0;
916 
917 	if (qp->srq) {
918 		*bad_wr = wr;
919 		return -EOPNOTSUPP; /* what else from errno.h? */
920 	}
921 	/*
922 	 * Try to acquire QP state lock. Must be non-blocking
923 	 * to accommodate kernel clients needs.
924 	 */
925 	if (!down_read_trylock(&qp->state_lock)) {
926 		*bad_wr = wr;
927 		return -ENOTCONN;
928 	}
929 	if (!qp->kernel_verbs) {
930 		siw_dbg_qp(qp, "no kernel post_recv for user mapped sq\n");
931 		up_read(&qp->state_lock);
932 		*bad_wr = wr;
933 		return -EINVAL;
934 	}
935 	if (qp->attrs.state > SIW_QP_STATE_RTS) {
936 		up_read(&qp->state_lock);
937 		*bad_wr = wr;
938 		return -EINVAL;
939 	}
940 	/*
941 	 * Serialize potentially multiple producers.
942 	 * Not needed for single threaded consumer side.
943 	 */
944 	spin_lock_irqsave(&qp->rq_lock, flags);
945 
946 	while (wr) {
947 		u32 idx = qp->rq_put % qp->attrs.rq_size;
948 		struct siw_rqe *rqe = &qp->recvq[idx];
949 
950 		if (rqe->flags) {
951 			siw_dbg_qp(qp, "RQ full\n");
952 			rv = -ENOMEM;
953 			break;
954 		}
955 		if (wr->num_sge > qp->attrs.rq_max_sges) {
956 			siw_dbg_qp(qp, "too many sge's: %d\n", wr->num_sge);
957 			rv = -EINVAL;
958 			break;
959 		}
960 		rqe->id = wr->wr_id;
961 		rqe->num_sge = wr->num_sge;
962 		siw_copy_sgl(wr->sg_list, rqe->sge, wr->num_sge);
963 
964 		/* make sure RQE is completely written before valid */
965 		smp_wmb();
966 
967 		rqe->flags = SIW_WQE_VALID;
968 
969 		qp->rq_put++;
970 		wr = wr->next;
971 	}
972 	spin_unlock_irqrestore(&qp->rq_lock, flags);
973 
974 	up_read(&qp->state_lock);
975 
976 	if (rv < 0) {
977 		siw_dbg_qp(qp, "error %d\n", rv);
978 		*bad_wr = wr;
979 	}
980 	return rv > 0 ? 0 : rv;
981 }
982 
983 void siw_destroy_cq(struct ib_cq *base_cq, struct ib_udata *udata)
984 {
985 	struct siw_cq *cq = to_siw_cq(base_cq);
986 	struct siw_device *sdev = to_siw_dev(base_cq->device);
987 	struct siw_ucontext *ctx =
988 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
989 					  base_ucontext);
990 
991 	siw_dbg_cq(cq, "free CQ resources\n");
992 
993 	siw_cq_flush(cq);
994 
995 	if (ctx && cq->xa_cq_index != SIW_INVAL_UOBJ_KEY)
996 		kfree(xa_erase(&ctx->xa, cq->xa_cq_index));
997 
998 	atomic_dec(&sdev->num_cq);
999 
1000 	vfree(cq->queue);
1001 }
1002 
1003 /*
1004  * siw_create_cq()
1005  *
1006  * Populate CQ of requested size
1007  *
1008  * @base_cq: CQ as allocated by RDMA midlayer
1009  * @attr: Initial CQ attributes
1010  * @udata: relates to user context
1011  */
1012 
1013 int siw_create_cq(struct ib_cq *base_cq, const struct ib_cq_init_attr *attr,
1014 		  struct ib_udata *udata)
1015 {
1016 	struct siw_device *sdev = to_siw_dev(base_cq->device);
1017 	struct siw_cq *cq = to_siw_cq(base_cq);
1018 	int rv, size = attr->cqe;
1019 
1020 	if (atomic_inc_return(&sdev->num_cq) > SIW_MAX_CQ) {
1021 		siw_dbg(base_cq->device, "too many CQ's\n");
1022 		rv = -ENOMEM;
1023 		goto err_out;
1024 	}
1025 	if (size < 1 || size > sdev->attrs.max_cqe) {
1026 		siw_dbg(base_cq->device, "CQ size error: %d\n", size);
1027 		rv = -EINVAL;
1028 		goto err_out;
1029 	}
1030 	size = roundup_pow_of_two(size);
1031 	cq->base_cq.cqe = size;
1032 	cq->num_cqe = size;
1033 	cq->xa_cq_index = SIW_INVAL_UOBJ_KEY;
1034 
1035 	if (!udata) {
1036 		cq->kernel_verbs = 1;
1037 		cq->queue = vzalloc(size * sizeof(struct siw_cqe) +
1038 				    sizeof(struct siw_cq_ctrl));
1039 	} else {
1040 		cq->queue = vmalloc_user(size * sizeof(struct siw_cqe) +
1041 					 sizeof(struct siw_cq_ctrl));
1042 	}
1043 	if (cq->queue == NULL) {
1044 		rv = -ENOMEM;
1045 		goto err_out;
1046 	}
1047 	get_random_bytes(&cq->id, 4);
1048 	siw_dbg(base_cq->device, "new CQ [%u]\n", cq->id);
1049 
1050 	spin_lock_init(&cq->lock);
1051 
1052 	cq->notify = (struct siw_cq_ctrl *)&cq->queue[size];
1053 
1054 	if (udata) {
1055 		struct siw_uresp_create_cq uresp = {};
1056 		struct siw_ucontext *ctx =
1057 			rdma_udata_to_drv_context(udata, struct siw_ucontext,
1058 						  base_ucontext);
1059 
1060 		cq->xa_cq_index =
1061 			siw_create_uobj(ctx, cq->queue,
1062 					size * sizeof(struct siw_cqe) +
1063 						sizeof(struct siw_cq_ctrl));
1064 		if (cq->xa_cq_index == SIW_INVAL_UOBJ_KEY) {
1065 			rv = -ENOMEM;
1066 			goto err_out;
1067 		}
1068 		uresp.cq_key = cq->xa_cq_index << PAGE_SHIFT;
1069 		uresp.cq_id = cq->id;
1070 		uresp.num_cqe = size;
1071 
1072 		if (udata->outlen < sizeof(uresp)) {
1073 			rv = -EINVAL;
1074 			goto err_out;
1075 		}
1076 		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1077 		if (rv)
1078 			goto err_out;
1079 	}
1080 	return 0;
1081 
1082 err_out:
1083 	siw_dbg(base_cq->device, "CQ creation failed: %d", rv);
1084 
1085 	if (cq && cq->queue) {
1086 		struct siw_ucontext *ctx =
1087 			rdma_udata_to_drv_context(udata, struct siw_ucontext,
1088 						  base_ucontext);
1089 		if (cq->xa_cq_index != SIW_INVAL_UOBJ_KEY)
1090 			kfree(xa_erase(&ctx->xa, cq->xa_cq_index));
1091 		vfree(cq->queue);
1092 	}
1093 	atomic_dec(&sdev->num_cq);
1094 
1095 	return rv;
1096 }
1097 
1098 /*
1099  * siw_poll_cq()
1100  *
1101  * Reap CQ entries if available and copy work completion status into
1102  * array of WC's provided by caller. Returns number of reaped CQE's.
1103  *
1104  * @base_cq:	Base CQ contained in siw CQ.
1105  * @num_cqe:	Maximum number of CQE's to reap.
1106  * @wc:		Array of work completions to be filled by siw.
1107  */
1108 int siw_poll_cq(struct ib_cq *base_cq, int num_cqe, struct ib_wc *wc)
1109 {
1110 	struct siw_cq *cq = to_siw_cq(base_cq);
1111 	int i;
1112 
1113 	for (i = 0; i < num_cqe; i++) {
1114 		if (!siw_reap_cqe(cq, wc))
1115 			break;
1116 		wc++;
1117 	}
1118 	return i;
1119 }
1120 
1121 /*
1122  * siw_req_notify_cq()
1123  *
1124  * Request notification for new CQE's added to that CQ.
1125  * Defined flags:
1126  * o SIW_CQ_NOTIFY_SOLICITED lets siw trigger a notification
1127  *   event if a WQE with notification flag set enters the CQ
1128  * o SIW_CQ_NOTIFY_NEXT_COMP lets siw trigger a notification
1129  *   event if a WQE enters the CQ.
1130  * o IB_CQ_REPORT_MISSED_EVENTS: return value will provide the
1131  *   number of not reaped CQE's regardless of its notification
1132  *   type and current or new CQ notification settings.
1133  *
1134  * @base_cq:	Base CQ contained in siw CQ.
1135  * @flags:	Requested notification flags.
1136  */
1137 int siw_req_notify_cq(struct ib_cq *base_cq, enum ib_cq_notify_flags flags)
1138 {
1139 	struct siw_cq *cq = to_siw_cq(base_cq);
1140 
1141 	siw_dbg_cq(cq, "flags: 0x%02x\n", flags);
1142 
1143 	if ((flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED)
1144 		/*
1145 		 * Enable CQ event for next solicited completion.
1146 		 * and make it visible to all associated producers.
1147 		 */
1148 		smp_store_mb(cq->notify->flags, SIW_NOTIFY_SOLICITED);
1149 	else
1150 		/*
1151 		 * Enable CQ event for any signalled completion.
1152 		 * and make it visible to all associated producers.
1153 		 */
1154 		smp_store_mb(cq->notify->flags, SIW_NOTIFY_ALL);
1155 
1156 	if (flags & IB_CQ_REPORT_MISSED_EVENTS)
1157 		return cq->cq_put - cq->cq_get;
1158 
1159 	return 0;
1160 }
1161 
1162 /*
1163  * siw_dereg_mr()
1164  *
1165  * Release Memory Region.
1166  *
1167  * @base_mr: Base MR contained in siw MR.
1168  * @udata: points to user context, unused.
1169  */
1170 int siw_dereg_mr(struct ib_mr *base_mr, struct ib_udata *udata)
1171 {
1172 	struct siw_mr *mr = to_siw_mr(base_mr);
1173 	struct siw_device *sdev = to_siw_dev(base_mr->device);
1174 
1175 	siw_dbg_mem(mr->mem, "deregister MR\n");
1176 
1177 	atomic_dec(&sdev->num_mr);
1178 
1179 	siw_mr_drop_mem(mr);
1180 	kfree_rcu(mr, rcu);
1181 
1182 	return 0;
1183 }
1184 
1185 /*
1186  * siw_reg_user_mr()
1187  *
1188  * Register Memory Region.
1189  *
1190  * @pd:		Protection Domain
1191  * @start:	starting address of MR (virtual address)
1192  * @len:	len of MR
1193  * @rnic_va:	not used by siw
1194  * @rights:	MR access rights
1195  * @udata:	user buffer to communicate STag and Key.
1196  */
1197 struct ib_mr *siw_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,
1198 			      u64 rnic_va, int rights, struct ib_udata *udata)
1199 {
1200 	struct siw_mr *mr = NULL;
1201 	struct siw_umem *umem = NULL;
1202 	struct siw_ureq_reg_mr ureq;
1203 	struct siw_device *sdev = to_siw_dev(pd->device);
1204 
1205 	unsigned long mem_limit = rlimit(RLIMIT_MEMLOCK);
1206 	int rv;
1207 
1208 	siw_dbg_pd(pd, "start: 0x%pK, va: 0x%pK, len: %llu\n",
1209 		   (void *)(uintptr_t)start, (void *)(uintptr_t)rnic_va,
1210 		   (unsigned long long)len);
1211 
1212 	if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1213 		siw_dbg_pd(pd, "too many mr's\n");
1214 		rv = -ENOMEM;
1215 		goto err_out;
1216 	}
1217 	if (!len) {
1218 		rv = -EINVAL;
1219 		goto err_out;
1220 	}
1221 	if (mem_limit != RLIM_INFINITY) {
1222 		unsigned long num_pages =
1223 			(PAGE_ALIGN(len + (start & ~PAGE_MASK))) >> PAGE_SHIFT;
1224 		mem_limit >>= PAGE_SHIFT;
1225 
1226 		if (num_pages > mem_limit - current->mm->locked_vm) {
1227 			siw_dbg_pd(pd, "pages req %lu, max %lu, lock %lu\n",
1228 				   num_pages, mem_limit,
1229 				   current->mm->locked_vm);
1230 			rv = -ENOMEM;
1231 			goto err_out;
1232 		}
1233 	}
1234 	umem = siw_umem_get(start, len, ib_access_writable(rights));
1235 	if (IS_ERR(umem)) {
1236 		rv = PTR_ERR(umem);
1237 		siw_dbg_pd(pd, "getting user memory failed: %d\n", rv);
1238 		umem = NULL;
1239 		goto err_out;
1240 	}
1241 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1242 	if (!mr) {
1243 		rv = -ENOMEM;
1244 		goto err_out;
1245 	}
1246 	rv = siw_mr_add_mem(mr, pd, umem, start, len, rights);
1247 	if (rv)
1248 		goto err_out;
1249 
1250 	if (udata) {
1251 		struct siw_uresp_reg_mr uresp = {};
1252 		struct siw_mem *mem = mr->mem;
1253 
1254 		if (udata->inlen < sizeof(ureq)) {
1255 			rv = -EINVAL;
1256 			goto err_out;
1257 		}
1258 		rv = ib_copy_from_udata(&ureq, udata, sizeof(ureq));
1259 		if (rv)
1260 			goto err_out;
1261 
1262 		mr->base_mr.lkey |= ureq.stag_key;
1263 		mr->base_mr.rkey |= ureq.stag_key;
1264 		mem->stag |= ureq.stag_key;
1265 		uresp.stag = mem->stag;
1266 
1267 		if (udata->outlen < sizeof(uresp)) {
1268 			rv = -EINVAL;
1269 			goto err_out;
1270 		}
1271 		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1272 		if (rv)
1273 			goto err_out;
1274 	}
1275 	mr->mem->stag_valid = 1;
1276 
1277 	return &mr->base_mr;
1278 
1279 err_out:
1280 	atomic_dec(&sdev->num_mr);
1281 	if (mr) {
1282 		if (mr->mem)
1283 			siw_mr_drop_mem(mr);
1284 		kfree_rcu(mr, rcu);
1285 	} else {
1286 		if (umem)
1287 			siw_umem_release(umem, false);
1288 	}
1289 	return ERR_PTR(rv);
1290 }
1291 
1292 struct ib_mr *siw_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
1293 			   u32 max_sge, struct ib_udata *udata)
1294 {
1295 	struct siw_device *sdev = to_siw_dev(pd->device);
1296 	struct siw_mr *mr = NULL;
1297 	struct siw_pbl *pbl = NULL;
1298 	int rv;
1299 
1300 	if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1301 		siw_dbg_pd(pd, "too many mr's\n");
1302 		rv = -ENOMEM;
1303 		goto err_out;
1304 	}
1305 	if (mr_type != IB_MR_TYPE_MEM_REG) {
1306 		siw_dbg_pd(pd, "mr type %d unsupported\n", mr_type);
1307 		rv = -EOPNOTSUPP;
1308 		goto err_out;
1309 	}
1310 	if (max_sge > SIW_MAX_SGE_PBL) {
1311 		siw_dbg_pd(pd, "too many sge's: %d\n", max_sge);
1312 		rv = -ENOMEM;
1313 		goto err_out;
1314 	}
1315 	pbl = siw_pbl_alloc(max_sge);
1316 	if (IS_ERR(pbl)) {
1317 		rv = PTR_ERR(pbl);
1318 		siw_dbg_pd(pd, "pbl allocation failed: %d\n", rv);
1319 		pbl = NULL;
1320 		goto err_out;
1321 	}
1322 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1323 	if (!mr) {
1324 		rv = -ENOMEM;
1325 		goto err_out;
1326 	}
1327 	rv = siw_mr_add_mem(mr, pd, pbl, 0, max_sge * PAGE_SIZE, 0);
1328 	if (rv)
1329 		goto err_out;
1330 
1331 	mr->mem->is_pbl = 1;
1332 
1333 	siw_dbg_pd(pd, "[MEM %u]: success\n", mr->mem->stag);
1334 
1335 	return &mr->base_mr;
1336 
1337 err_out:
1338 	atomic_dec(&sdev->num_mr);
1339 
1340 	if (!mr) {
1341 		kfree(pbl);
1342 	} else {
1343 		if (mr->mem)
1344 			siw_mr_drop_mem(mr);
1345 		kfree_rcu(mr, rcu);
1346 	}
1347 	siw_dbg_pd(pd, "failed: %d\n", rv);
1348 
1349 	return ERR_PTR(rv);
1350 }
1351 
1352 /* Just used to count number of pages being mapped */
1353 static int siw_set_pbl_page(struct ib_mr *base_mr, u64 buf_addr)
1354 {
1355 	return 0;
1356 }
1357 
1358 int siw_map_mr_sg(struct ib_mr *base_mr, struct scatterlist *sl, int num_sle,
1359 		  unsigned int *sg_off)
1360 {
1361 	struct scatterlist *slp;
1362 	struct siw_mr *mr = to_siw_mr(base_mr);
1363 	struct siw_mem *mem = mr->mem;
1364 	struct siw_pbl *pbl = mem->pbl;
1365 	struct siw_pble *pble;
1366 	unsigned long pbl_size;
1367 	int i, rv;
1368 
1369 	if (!pbl) {
1370 		siw_dbg_mem(mem, "no PBL allocated\n");
1371 		return -EINVAL;
1372 	}
1373 	pble = pbl->pbe;
1374 
1375 	if (pbl->max_buf < num_sle) {
1376 		siw_dbg_mem(mem, "too many SGE's: %d > %d\n",
1377 			    mem->pbl->max_buf, num_sle);
1378 		return -ENOMEM;
1379 	}
1380 	for_each_sg(sl, slp, num_sle, i) {
1381 		if (sg_dma_len(slp) == 0) {
1382 			siw_dbg_mem(mem, "empty SGE\n");
1383 			return -EINVAL;
1384 		}
1385 		if (i == 0) {
1386 			pble->addr = sg_dma_address(slp);
1387 			pble->size = sg_dma_len(slp);
1388 			pble->pbl_off = 0;
1389 			pbl_size = pble->size;
1390 			pbl->num_buf = 1;
1391 		} else {
1392 			/* Merge PBL entries if adjacent */
1393 			if (pble->addr + pble->size == sg_dma_address(slp)) {
1394 				pble->size += sg_dma_len(slp);
1395 			} else {
1396 				pble++;
1397 				pbl->num_buf++;
1398 				pble->addr = sg_dma_address(slp);
1399 				pble->size = sg_dma_len(slp);
1400 				pble->pbl_off = pbl_size;
1401 			}
1402 			pbl_size += sg_dma_len(slp);
1403 		}
1404 		siw_dbg_mem(mem,
1405 			"sge[%d], size %u, addr 0x%p, total %lu\n",
1406 			i, pble->size, (void *)(uintptr_t)pble->addr,
1407 			pbl_size);
1408 	}
1409 	rv = ib_sg_to_pages(base_mr, sl, num_sle, sg_off, siw_set_pbl_page);
1410 	if (rv > 0) {
1411 		mem->len = base_mr->length;
1412 		mem->va = base_mr->iova;
1413 		siw_dbg_mem(mem,
1414 			"%llu bytes, start 0x%pK, %u SLE to %u entries\n",
1415 			mem->len, (void *)(uintptr_t)mem->va, num_sle,
1416 			pbl->num_buf);
1417 	}
1418 	return rv;
1419 }
1420 
1421 /*
1422  * siw_get_dma_mr()
1423  *
1424  * Create a (empty) DMA memory region, where no umem is attached.
1425  */
1426 struct ib_mr *siw_get_dma_mr(struct ib_pd *pd, int rights)
1427 {
1428 	struct siw_device *sdev = to_siw_dev(pd->device);
1429 	struct siw_mr *mr = NULL;
1430 	int rv;
1431 
1432 	if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1433 		siw_dbg_pd(pd, "too many mr's\n");
1434 		rv = -ENOMEM;
1435 		goto err_out;
1436 	}
1437 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1438 	if (!mr) {
1439 		rv = -ENOMEM;
1440 		goto err_out;
1441 	}
1442 	rv = siw_mr_add_mem(mr, pd, NULL, 0, ULONG_MAX, rights);
1443 	if (rv)
1444 		goto err_out;
1445 
1446 	mr->mem->stag_valid = 1;
1447 
1448 	siw_dbg_pd(pd, "[MEM %u]: success\n", mr->mem->stag);
1449 
1450 	return &mr->base_mr;
1451 
1452 err_out:
1453 	if (rv)
1454 		kfree(mr);
1455 
1456 	atomic_dec(&sdev->num_mr);
1457 
1458 	return ERR_PTR(rv);
1459 }
1460 
1461 /*
1462  * siw_create_srq()
1463  *
1464  * Create Shared Receive Queue of attributes @init_attrs
1465  * within protection domain given by @pd.
1466  *
1467  * @base_srq:	Base SRQ contained in siw SRQ.
1468  * @init_attrs:	SRQ init attributes.
1469  * @udata:	points to user context
1470  */
1471 int siw_create_srq(struct ib_srq *base_srq,
1472 		   struct ib_srq_init_attr *init_attrs, struct ib_udata *udata)
1473 {
1474 	struct siw_srq *srq = to_siw_srq(base_srq);
1475 	struct ib_srq_attr *attrs = &init_attrs->attr;
1476 	struct siw_device *sdev = to_siw_dev(base_srq->device);
1477 	struct siw_ucontext *ctx =
1478 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
1479 					  base_ucontext);
1480 	int rv;
1481 
1482 	if (atomic_inc_return(&sdev->num_srq) > SIW_MAX_SRQ) {
1483 		siw_dbg_pd(base_srq->pd, "too many SRQ's\n");
1484 		rv = -ENOMEM;
1485 		goto err_out;
1486 	}
1487 	if (attrs->max_wr == 0 || attrs->max_wr > SIW_MAX_SRQ_WR ||
1488 	    attrs->max_sge > SIW_MAX_SGE || attrs->srq_limit > attrs->max_wr) {
1489 		rv = -EINVAL;
1490 		goto err_out;
1491 	}
1492 	srq->max_sge = attrs->max_sge;
1493 	srq->num_rqe = roundup_pow_of_two(attrs->max_wr);
1494 	srq->xa_srq_index = SIW_INVAL_UOBJ_KEY;
1495 	srq->limit = attrs->srq_limit;
1496 	if (srq->limit)
1497 		srq->armed = 1;
1498 
1499 	srq->kernel_verbs = !udata;
1500 
1501 	if (udata)
1502 		srq->recvq =
1503 			vmalloc_user(srq->num_rqe * sizeof(struct siw_rqe));
1504 	else
1505 		srq->recvq = vzalloc(srq->num_rqe * sizeof(struct siw_rqe));
1506 
1507 	if (srq->recvq == NULL) {
1508 		rv = -ENOMEM;
1509 		goto err_out;
1510 	}
1511 	if (udata) {
1512 		struct siw_uresp_create_srq uresp = {};
1513 
1514 		srq->xa_srq_index = siw_create_uobj(
1515 			ctx, srq->recvq, srq->num_rqe * sizeof(struct siw_rqe));
1516 
1517 		if (srq->xa_srq_index == SIW_INVAL_UOBJ_KEY) {
1518 			rv = -ENOMEM;
1519 			goto err_out;
1520 		}
1521 		uresp.srq_key = srq->xa_srq_index;
1522 		uresp.num_rqe = srq->num_rqe;
1523 
1524 		if (udata->outlen < sizeof(uresp)) {
1525 			rv = -EINVAL;
1526 			goto err_out;
1527 		}
1528 		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1529 		if (rv)
1530 			goto err_out;
1531 	}
1532 	spin_lock_init(&srq->lock);
1533 
1534 	siw_dbg_pd(base_srq->pd, "[SRQ]: success\n");
1535 
1536 	return 0;
1537 
1538 err_out:
1539 	if (srq->recvq) {
1540 		if (ctx && srq->xa_srq_index != SIW_INVAL_UOBJ_KEY)
1541 			kfree(xa_erase(&ctx->xa, srq->xa_srq_index));
1542 		vfree(srq->recvq);
1543 	}
1544 	atomic_dec(&sdev->num_srq);
1545 
1546 	return rv;
1547 }
1548 
1549 /*
1550  * siw_modify_srq()
1551  *
1552  * Modify SRQ. The caller may resize SRQ and/or set/reset notification
1553  * limit and (re)arm IB_EVENT_SRQ_LIMIT_REACHED notification.
1554  *
1555  * NOTE: it is unclear if RDMA core allows for changing the MAX_SGE
1556  * parameter. siw_modify_srq() does not check the attrs->max_sge param.
1557  */
1558 int siw_modify_srq(struct ib_srq *base_srq, struct ib_srq_attr *attrs,
1559 		   enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
1560 {
1561 	struct siw_srq *srq = to_siw_srq(base_srq);
1562 	unsigned long flags;
1563 	int rv = 0;
1564 
1565 	spin_lock_irqsave(&srq->lock, flags);
1566 
1567 	if (attr_mask & IB_SRQ_MAX_WR) {
1568 		/* resize request not yet supported */
1569 		rv = -EOPNOTSUPP;
1570 		goto out;
1571 	}
1572 	if (attr_mask & IB_SRQ_LIMIT) {
1573 		if (attrs->srq_limit) {
1574 			if (unlikely(attrs->srq_limit > srq->num_rqe)) {
1575 				rv = -EINVAL;
1576 				goto out;
1577 			}
1578 			srq->armed = 1;
1579 		} else {
1580 			srq->armed = 0;
1581 		}
1582 		srq->limit = attrs->srq_limit;
1583 	}
1584 out:
1585 	spin_unlock_irqrestore(&srq->lock, flags);
1586 
1587 	return rv;
1588 }
1589 
1590 /*
1591  * siw_query_srq()
1592  *
1593  * Query SRQ attributes.
1594  */
1595 int siw_query_srq(struct ib_srq *base_srq, struct ib_srq_attr *attrs)
1596 {
1597 	struct siw_srq *srq = to_siw_srq(base_srq);
1598 	unsigned long flags;
1599 
1600 	spin_lock_irqsave(&srq->lock, flags);
1601 
1602 	attrs->max_wr = srq->num_rqe;
1603 	attrs->max_sge = srq->max_sge;
1604 	attrs->srq_limit = srq->limit;
1605 
1606 	spin_unlock_irqrestore(&srq->lock, flags);
1607 
1608 	return 0;
1609 }
1610 
1611 /*
1612  * siw_destroy_srq()
1613  *
1614  * Destroy SRQ.
1615  * It is assumed that the SRQ is not referenced by any
1616  * QP anymore - the code trusts the RDMA core environment to keep track
1617  * of QP references.
1618  */
1619 void siw_destroy_srq(struct ib_srq *base_srq, struct ib_udata *udata)
1620 {
1621 	struct siw_srq *srq = to_siw_srq(base_srq);
1622 	struct siw_device *sdev = to_siw_dev(base_srq->device);
1623 	struct siw_ucontext *ctx =
1624 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
1625 					  base_ucontext);
1626 
1627 	if (ctx && srq->xa_srq_index != SIW_INVAL_UOBJ_KEY)
1628 		kfree(xa_erase(&ctx->xa, srq->xa_srq_index));
1629 
1630 	vfree(srq->recvq);
1631 	atomic_dec(&sdev->num_srq);
1632 }
1633 
1634 /*
1635  * siw_post_srq_recv()
1636  *
1637  * Post a list of receive queue elements to SRQ.
1638  * NOTE: The function does not check or lock a certain SRQ state
1639  *       during the post operation. The code simply trusts the
1640  *       RDMA core environment.
1641  *
1642  * @base_srq:	Base SRQ contained in siw SRQ
1643  * @wr:		List of R-WR's
1644  * @bad_wr:	Updated to failing WR if posting fails.
1645  */
1646 int siw_post_srq_recv(struct ib_srq *base_srq, const struct ib_recv_wr *wr,
1647 		      const struct ib_recv_wr **bad_wr)
1648 {
1649 	struct siw_srq *srq = to_siw_srq(base_srq);
1650 	unsigned long flags;
1651 	int rv = 0;
1652 
1653 	if (unlikely(!srq->kernel_verbs)) {
1654 		siw_dbg_pd(base_srq->pd,
1655 			   "[SRQ]: no kernel post_recv for mapped srq\n");
1656 		rv = -EINVAL;
1657 		goto out;
1658 	}
1659 	/*
1660 	 * Serialize potentially multiple producers.
1661 	 * Also needed to serialize potentially multiple
1662 	 * consumers.
1663 	 */
1664 	spin_lock_irqsave(&srq->lock, flags);
1665 
1666 	while (wr) {
1667 		u32 idx = srq->rq_put % srq->num_rqe;
1668 		struct siw_rqe *rqe = &srq->recvq[idx];
1669 
1670 		if (rqe->flags) {
1671 			siw_dbg_pd(base_srq->pd, "SRQ full\n");
1672 			rv = -ENOMEM;
1673 			break;
1674 		}
1675 		if (unlikely(wr->num_sge > srq->max_sge)) {
1676 			siw_dbg_pd(base_srq->pd,
1677 				   "[SRQ]: too many sge's: %d\n", wr->num_sge);
1678 			rv = -EINVAL;
1679 			break;
1680 		}
1681 		rqe->id = wr->wr_id;
1682 		rqe->num_sge = wr->num_sge;
1683 		siw_copy_sgl(wr->sg_list, rqe->sge, wr->num_sge);
1684 
1685 		/* Make sure S-RQE is completely written before valid */
1686 		smp_wmb();
1687 
1688 		rqe->flags = SIW_WQE_VALID;
1689 
1690 		srq->rq_put++;
1691 		wr = wr->next;
1692 	}
1693 	spin_unlock_irqrestore(&srq->lock, flags);
1694 out:
1695 	if (unlikely(rv < 0)) {
1696 		siw_dbg_pd(base_srq->pd, "[SRQ]: error %d\n", rv);
1697 		*bad_wr = wr;
1698 	}
1699 	return rv;
1700 }
1701 
1702 void siw_qp_event(struct siw_qp *qp, enum ib_event_type etype)
1703 {
1704 	struct ib_event event;
1705 	struct ib_qp *base_qp = qp->ib_qp;
1706 
1707 	/*
1708 	 * Do not report asynchronous errors on QP which gets
1709 	 * destroyed via verbs interface (siw_destroy_qp())
1710 	 */
1711 	if (qp->attrs.flags & SIW_QP_IN_DESTROY)
1712 		return;
1713 
1714 	event.event = etype;
1715 	event.device = base_qp->device;
1716 	event.element.qp = base_qp;
1717 
1718 	if (base_qp->event_handler) {
1719 		siw_dbg_qp(qp, "reporting event %d\n", etype);
1720 		base_qp->event_handler(&event, base_qp->qp_context);
1721 	}
1722 }
1723 
1724 void siw_cq_event(struct siw_cq *cq, enum ib_event_type etype)
1725 {
1726 	struct ib_event event;
1727 	struct ib_cq *base_cq = &cq->base_cq;
1728 
1729 	event.event = etype;
1730 	event.device = base_cq->device;
1731 	event.element.cq = base_cq;
1732 
1733 	if (base_cq->event_handler) {
1734 		siw_dbg_cq(cq, "reporting CQ event %d\n", etype);
1735 		base_cq->event_handler(&event, base_cq->cq_context);
1736 	}
1737 }
1738 
1739 void siw_srq_event(struct siw_srq *srq, enum ib_event_type etype)
1740 {
1741 	struct ib_event event;
1742 	struct ib_srq *base_srq = &srq->base_srq;
1743 
1744 	event.event = etype;
1745 	event.device = base_srq->device;
1746 	event.element.srq = base_srq;
1747 
1748 	if (base_srq->event_handler) {
1749 		siw_dbg_pd(srq->base_srq.pd,
1750 			   "reporting SRQ event %d\n", etype);
1751 		base_srq->event_handler(&event, base_srq->srq_context);
1752 	}
1753 }
1754 
1755 void siw_port_event(struct siw_device *sdev, u8 port, enum ib_event_type etype)
1756 {
1757 	struct ib_event event;
1758 
1759 	event.event = etype;
1760 	event.device = &sdev->base_dev;
1761 	event.element.port_num = port;
1762 
1763 	siw_dbg(&sdev->base_dev, "reporting port event %d\n", etype);
1764 
1765 	ib_dispatch_event(&event);
1766 }
1767