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