xref: /linux/drivers/infiniband/sw/siw/siw_verbs.c (revision 02680c23d7b3febe45ea3d4f9818c2b2dc89020a)
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, u32 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, u32 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, u32 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 	unsigned long flags;
304 	int num_sqe, num_rqe, rv = 0;
305 	size_t length;
306 
307 	siw_dbg(base_dev, "create new QP\n");
308 
309 	if (attrs->create_flags)
310 		return ERR_PTR(-EOPNOTSUPP);
311 
312 	if (atomic_inc_return(&sdev->num_qp) > SIW_MAX_QP) {
313 		siw_dbg(base_dev, "too many QP's\n");
314 		rv = -ENOMEM;
315 		goto err_out;
316 	}
317 	if (attrs->qp_type != IB_QPT_RC) {
318 		siw_dbg(base_dev, "only RC QP's supported\n");
319 		rv = -EOPNOTSUPP;
320 		goto err_out;
321 	}
322 	if ((attrs->cap.max_send_wr > SIW_MAX_QP_WR) ||
323 	    (attrs->cap.max_recv_wr > SIW_MAX_QP_WR) ||
324 	    (attrs->cap.max_send_sge > SIW_MAX_SGE) ||
325 	    (attrs->cap.max_recv_sge > SIW_MAX_SGE)) {
326 		siw_dbg(base_dev, "QP size error\n");
327 		rv = -EINVAL;
328 		goto err_out;
329 	}
330 	if (attrs->cap.max_inline_data > SIW_MAX_INLINE) {
331 		siw_dbg(base_dev, "max inline send: %d > %d\n",
332 			attrs->cap.max_inline_data, (int)SIW_MAX_INLINE);
333 		rv = -EINVAL;
334 		goto err_out;
335 	}
336 	/*
337 	 * NOTE: we allow for zero element SQ and RQ WQE's SGL's
338 	 * but not for a QP unable to hold any WQE (SQ + RQ)
339 	 */
340 	if (attrs->cap.max_send_wr + attrs->cap.max_recv_wr == 0) {
341 		siw_dbg(base_dev, "QP must have send or receive queue\n");
342 		rv = -EINVAL;
343 		goto err_out;
344 	}
345 
346 	if (!attrs->send_cq || (!attrs->recv_cq && !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 	num_sqe = attrs->cap.max_send_wr;
366 	num_rqe = attrs->cap.max_recv_wr;
367 
368 	/* All queue indices are derived from modulo operations
369 	 * on a free running 'get' (consumer) and 'put' (producer)
370 	 * unsigned counter. Having queue sizes at power of two
371 	 * avoids handling counter wrap around.
372 	 */
373 	if (num_sqe)
374 		num_sqe = roundup_pow_of_two(num_sqe);
375 	else {
376 		/* Zero sized SQ is not supported */
377 		rv = -EINVAL;
378 		goto err_out_xa;
379 	}
380 	if (num_rqe)
381 		num_rqe = roundup_pow_of_two(num_rqe);
382 
383 	if (udata)
384 		qp->sendq = vmalloc_user(num_sqe * sizeof(struct siw_sqe));
385 	else
386 		qp->sendq = vzalloc(num_sqe * sizeof(struct siw_sqe));
387 
388 	if (qp->sendq == NULL) {
389 		rv = -ENOMEM;
390 		goto err_out_xa;
391 	}
392 	if (attrs->sq_sig_type != IB_SIGNAL_REQ_WR) {
393 		if (attrs->sq_sig_type == IB_SIGNAL_ALL_WR)
394 			qp->attrs.flags |= SIW_SIGNAL_ALL_WR;
395 		else {
396 			rv = -EINVAL;
397 			goto err_out_xa;
398 		}
399 	}
400 	qp->pd = pd;
401 	qp->scq = to_siw_cq(attrs->send_cq);
402 	qp->rcq = to_siw_cq(attrs->recv_cq);
403 
404 	if (attrs->srq) {
405 		/*
406 		 * SRQ support.
407 		 * Verbs 6.3.7: ignore RQ size, if SRQ present
408 		 * Verbs 6.3.5: do not check PD of SRQ against PD of QP
409 		 */
410 		qp->srq = to_siw_srq(attrs->srq);
411 		qp->attrs.rq_size = 0;
412 		siw_dbg(base_dev, "QP [%u]: SRQ attached\n",
413 			qp->base_qp.qp_num);
414 	} else if (num_rqe) {
415 		if (udata)
416 			qp->recvq =
417 				vmalloc_user(num_rqe * sizeof(struct siw_rqe));
418 		else
419 			qp->recvq = vzalloc(num_rqe * sizeof(struct siw_rqe));
420 
421 		if (qp->recvq == NULL) {
422 			rv = -ENOMEM;
423 			goto err_out_xa;
424 		}
425 		qp->attrs.rq_size = num_rqe;
426 	}
427 	qp->attrs.sq_size = num_sqe;
428 	qp->attrs.sq_max_sges = attrs->cap.max_send_sge;
429 	qp->attrs.rq_max_sges = attrs->cap.max_recv_sge;
430 
431 	/* Make those two tunables fixed for now. */
432 	qp->tx_ctx.gso_seg_limit = 1;
433 	qp->tx_ctx.zcopy_tx = zcopy_tx;
434 
435 	qp->attrs.state = SIW_QP_STATE_IDLE;
436 
437 	if (udata) {
438 		struct siw_uresp_create_qp uresp = {};
439 
440 		uresp.num_sqe = num_sqe;
441 		uresp.num_rqe = num_rqe;
442 		uresp.qp_id = qp_id(qp);
443 
444 		if (qp->sendq) {
445 			length = num_sqe * sizeof(struct siw_sqe);
446 			qp->sq_entry =
447 				siw_mmap_entry_insert(uctx, qp->sendq,
448 						      length, &uresp.sq_key);
449 			if (!qp->sq_entry) {
450 				rv = -ENOMEM;
451 				goto err_out_xa;
452 			}
453 		}
454 
455 		if (qp->recvq) {
456 			length = num_rqe * sizeof(struct siw_rqe);
457 			qp->rq_entry =
458 				siw_mmap_entry_insert(uctx, qp->recvq,
459 						      length, &uresp.rq_key);
460 			if (!qp->rq_entry) {
461 				uresp.sq_key = SIW_INVAL_UOBJ_KEY;
462 				rv = -ENOMEM;
463 				goto err_out_xa;
464 			}
465 		}
466 
467 		if (udata->outlen < sizeof(uresp)) {
468 			rv = -EINVAL;
469 			goto err_out_xa;
470 		}
471 		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
472 		if (rv)
473 			goto err_out_xa;
474 	}
475 	qp->tx_cpu = siw_get_tx_cpu(sdev);
476 	if (qp->tx_cpu < 0) {
477 		rv = -EINVAL;
478 		goto err_out_xa;
479 	}
480 	INIT_LIST_HEAD(&qp->devq);
481 	spin_lock_irqsave(&sdev->lock, flags);
482 	list_add_tail(&qp->devq, &sdev->qp_list);
483 	spin_unlock_irqrestore(&sdev->lock, flags);
484 
485 	return &qp->base_qp;
486 
487 err_out_xa:
488 	xa_erase(&sdev->qp_xa, qp_id(qp));
489 err_out:
490 	if (qp) {
491 		if (uctx) {
492 			rdma_user_mmap_entry_remove(qp->sq_entry);
493 			rdma_user_mmap_entry_remove(qp->rq_entry);
494 		}
495 		vfree(qp->sendq);
496 		vfree(qp->recvq);
497 		kfree(qp);
498 	}
499 	atomic_dec(&sdev->num_qp);
500 
501 	return ERR_PTR(rv);
502 }
503 
504 /*
505  * Minimum siw_query_qp() verb interface.
506  *
507  * @qp_attr_mask is not used but all available information is provided
508  */
509 int siw_query_qp(struct ib_qp *base_qp, struct ib_qp_attr *qp_attr,
510 		 int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
511 {
512 	struct siw_qp *qp;
513 	struct siw_device *sdev;
514 
515 	if (base_qp && qp_attr && qp_init_attr) {
516 		qp = to_siw_qp(base_qp);
517 		sdev = to_siw_dev(base_qp->device);
518 	} else {
519 		return -EINVAL;
520 	}
521 	qp_attr->cap.max_inline_data = SIW_MAX_INLINE;
522 	qp_attr->cap.max_send_wr = qp->attrs.sq_size;
523 	qp_attr->cap.max_send_sge = qp->attrs.sq_max_sges;
524 	qp_attr->cap.max_recv_wr = qp->attrs.rq_size;
525 	qp_attr->cap.max_recv_sge = qp->attrs.rq_max_sges;
526 	qp_attr->path_mtu = ib_mtu_int_to_enum(sdev->netdev->mtu);
527 	qp_attr->max_rd_atomic = qp->attrs.irq_size;
528 	qp_attr->max_dest_rd_atomic = qp->attrs.orq_size;
529 
530 	qp_attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE |
531 				   IB_ACCESS_REMOTE_WRITE |
532 				   IB_ACCESS_REMOTE_READ;
533 
534 	qp_init_attr->qp_type = base_qp->qp_type;
535 	qp_init_attr->send_cq = base_qp->send_cq;
536 	qp_init_attr->recv_cq = base_qp->recv_cq;
537 	qp_init_attr->srq = base_qp->srq;
538 
539 	qp_init_attr->cap = qp_attr->cap;
540 
541 	return 0;
542 }
543 
544 int siw_verbs_modify_qp(struct ib_qp *base_qp, struct ib_qp_attr *attr,
545 			int attr_mask, struct ib_udata *udata)
546 {
547 	struct siw_qp_attrs new_attrs;
548 	enum siw_qp_attr_mask siw_attr_mask = 0;
549 	struct siw_qp *qp = to_siw_qp(base_qp);
550 	int rv = 0;
551 
552 	if (!attr_mask)
553 		return 0;
554 
555 	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
556 		return -EOPNOTSUPP;
557 
558 	memset(&new_attrs, 0, sizeof(new_attrs));
559 
560 	if (attr_mask & IB_QP_ACCESS_FLAGS) {
561 		siw_attr_mask = SIW_QP_ATTR_ACCESS_FLAGS;
562 
563 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
564 			new_attrs.flags |= SIW_RDMA_READ_ENABLED;
565 		if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
566 			new_attrs.flags |= SIW_RDMA_WRITE_ENABLED;
567 		if (attr->qp_access_flags & IB_ACCESS_MW_BIND)
568 			new_attrs.flags |= SIW_RDMA_BIND_ENABLED;
569 	}
570 	if (attr_mask & IB_QP_STATE) {
571 		siw_dbg_qp(qp, "desired IB QP state: %s\n",
572 			   ib_qp_state_to_string[attr->qp_state]);
573 
574 		new_attrs.state = ib_qp_state_to_siw_qp_state[attr->qp_state];
575 
576 		if (new_attrs.state > SIW_QP_STATE_RTS)
577 			qp->tx_ctx.tx_suspend = 1;
578 
579 		siw_attr_mask |= SIW_QP_ATTR_STATE;
580 	}
581 	if (!siw_attr_mask)
582 		goto out;
583 
584 	down_write(&qp->state_lock);
585 
586 	rv = siw_qp_modify(qp, &new_attrs, siw_attr_mask);
587 
588 	up_write(&qp->state_lock);
589 out:
590 	return rv;
591 }
592 
593 int siw_destroy_qp(struct ib_qp *base_qp, struct ib_udata *udata)
594 {
595 	struct siw_qp *qp = to_siw_qp(base_qp);
596 	struct siw_ucontext *uctx =
597 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
598 					  base_ucontext);
599 	struct siw_qp_attrs qp_attrs;
600 
601 	siw_dbg_qp(qp, "state %d\n", qp->attrs.state);
602 
603 	/*
604 	 * Mark QP as in process of destruction to prevent from
605 	 * any async callbacks to RDMA core
606 	 */
607 	qp->attrs.flags |= SIW_QP_IN_DESTROY;
608 	qp->rx_stream.rx_suspend = 1;
609 
610 	if (uctx) {
611 		rdma_user_mmap_entry_remove(qp->sq_entry);
612 		rdma_user_mmap_entry_remove(qp->rq_entry);
613 	}
614 
615 	down_write(&qp->state_lock);
616 
617 	qp_attrs.state = SIW_QP_STATE_ERROR;
618 	siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE);
619 
620 	if (qp->cep) {
621 		siw_cep_put(qp->cep);
622 		qp->cep = NULL;
623 	}
624 	up_write(&qp->state_lock);
625 
626 	kfree(qp->tx_ctx.mpa_crc_hd);
627 	kfree(qp->rx_stream.mpa_crc_hd);
628 
629 	qp->scq = qp->rcq = NULL;
630 
631 	siw_qp_put(qp);
632 
633 	return 0;
634 }
635 
636 /*
637  * siw_copy_inline_sgl()
638  *
639  * Prepare sgl of inlined data for sending. For userland callers
640  * function checks if given buffer addresses and len's are within
641  * process context bounds.
642  * Data from all provided sge's are copied together into the wqe,
643  * referenced by a single sge.
644  */
645 static int siw_copy_inline_sgl(const struct ib_send_wr *core_wr,
646 			       struct siw_sqe *sqe)
647 {
648 	struct ib_sge *core_sge = core_wr->sg_list;
649 	void *kbuf = &sqe->sge[1];
650 	int num_sge = core_wr->num_sge, bytes = 0;
651 
652 	sqe->sge[0].laddr = (uintptr_t)kbuf;
653 	sqe->sge[0].lkey = 0;
654 
655 	while (num_sge--) {
656 		if (!core_sge->length) {
657 			core_sge++;
658 			continue;
659 		}
660 		bytes += core_sge->length;
661 		if (bytes > SIW_MAX_INLINE) {
662 			bytes = -EINVAL;
663 			break;
664 		}
665 		memcpy(kbuf, (void *)(uintptr_t)core_sge->addr,
666 		       core_sge->length);
667 
668 		kbuf += core_sge->length;
669 		core_sge++;
670 	}
671 	sqe->sge[0].length = bytes > 0 ? bytes : 0;
672 	sqe->num_sge = bytes > 0 ? 1 : 0;
673 
674 	return bytes;
675 }
676 
677 /* Complete SQ WR's without processing */
678 static int siw_sq_flush_wr(struct siw_qp *qp, const struct ib_send_wr *wr,
679 			   const struct ib_send_wr **bad_wr)
680 {
681 	struct siw_sqe sqe = {};
682 	int rv = 0;
683 
684 	while (wr) {
685 		sqe.id = wr->wr_id;
686 		sqe.opcode = wr->opcode;
687 		rv = siw_sqe_complete(qp, &sqe, 0, SIW_WC_WR_FLUSH_ERR);
688 		if (rv) {
689 			if (bad_wr)
690 				*bad_wr = wr;
691 			break;
692 		}
693 		wr = wr->next;
694 	}
695 	return rv;
696 }
697 
698 /* Complete RQ WR's without processing */
699 static int siw_rq_flush_wr(struct siw_qp *qp, const struct ib_recv_wr *wr,
700 			   const struct ib_recv_wr **bad_wr)
701 {
702 	struct siw_rqe rqe = {};
703 	int rv = 0;
704 
705 	while (wr) {
706 		rqe.id = wr->wr_id;
707 		rv = siw_rqe_complete(qp, &rqe, 0, 0, SIW_WC_WR_FLUSH_ERR);
708 		if (rv) {
709 			if (bad_wr)
710 				*bad_wr = wr;
711 			break;
712 		}
713 		wr = wr->next;
714 	}
715 	return rv;
716 }
717 
718 /*
719  * siw_post_send()
720  *
721  * Post a list of S-WR's to a SQ.
722  *
723  * @base_qp:	Base QP contained in siw QP
724  * @wr:		Null terminated list of user WR's
725  * @bad_wr:	Points to failing WR in case of synchronous failure.
726  */
727 int siw_post_send(struct ib_qp *base_qp, const struct ib_send_wr *wr,
728 		  const struct ib_send_wr **bad_wr)
729 {
730 	struct siw_qp *qp = to_siw_qp(base_qp);
731 	struct siw_wqe *wqe = tx_wqe(qp);
732 
733 	unsigned long flags;
734 	int rv = 0;
735 
736 	if (wr && !rdma_is_kernel_res(&qp->base_qp.res)) {
737 		siw_dbg_qp(qp, "wr must be empty for user mapped sq\n");
738 		*bad_wr = wr;
739 		return -EINVAL;
740 	}
741 
742 	/*
743 	 * Try to acquire QP state lock. Must be non-blocking
744 	 * to accommodate kernel clients needs.
745 	 */
746 	if (!down_read_trylock(&qp->state_lock)) {
747 		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
748 			/*
749 			 * ERROR state is final, so we can be sure
750 			 * this state will not change as long as the QP
751 			 * exists.
752 			 *
753 			 * This handles an ib_drain_sq() call with
754 			 * a concurrent request to set the QP state
755 			 * to ERROR.
756 			 */
757 			rv = siw_sq_flush_wr(qp, wr, bad_wr);
758 		} else {
759 			siw_dbg_qp(qp, "QP locked, state %d\n",
760 				   qp->attrs.state);
761 			*bad_wr = wr;
762 			rv = -ENOTCONN;
763 		}
764 		return rv;
765 	}
766 	if (unlikely(qp->attrs.state != SIW_QP_STATE_RTS)) {
767 		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
768 			/*
769 			 * Immediately flush this WR to CQ, if QP
770 			 * is in ERROR state. SQ is guaranteed to
771 			 * be empty, so WR complets in-order.
772 			 *
773 			 * Typically triggered by ib_drain_sq().
774 			 */
775 			rv = siw_sq_flush_wr(qp, wr, bad_wr);
776 		} else {
777 			siw_dbg_qp(qp, "QP out of state %d\n",
778 				   qp->attrs.state);
779 			*bad_wr = wr;
780 			rv = -ENOTCONN;
781 		}
782 		up_read(&qp->state_lock);
783 		return rv;
784 	}
785 	spin_lock_irqsave(&qp->sq_lock, flags);
786 
787 	while (wr) {
788 		u32 idx = qp->sq_put % qp->attrs.sq_size;
789 		struct siw_sqe *sqe = &qp->sendq[idx];
790 
791 		if (sqe->flags) {
792 			siw_dbg_qp(qp, "sq full\n");
793 			rv = -ENOMEM;
794 			break;
795 		}
796 		if (wr->num_sge > qp->attrs.sq_max_sges) {
797 			siw_dbg_qp(qp, "too many sge's: %d\n", wr->num_sge);
798 			rv = -EINVAL;
799 			break;
800 		}
801 		sqe->id = wr->wr_id;
802 
803 		if ((wr->send_flags & IB_SEND_SIGNALED) ||
804 		    (qp->attrs.flags & SIW_SIGNAL_ALL_WR))
805 			sqe->flags |= SIW_WQE_SIGNALLED;
806 
807 		if (wr->send_flags & IB_SEND_FENCE)
808 			sqe->flags |= SIW_WQE_READ_FENCE;
809 
810 		switch (wr->opcode) {
811 		case IB_WR_SEND:
812 		case IB_WR_SEND_WITH_INV:
813 			if (wr->send_flags & IB_SEND_SOLICITED)
814 				sqe->flags |= SIW_WQE_SOLICITED;
815 
816 			if (!(wr->send_flags & IB_SEND_INLINE)) {
817 				siw_copy_sgl(wr->sg_list, sqe->sge,
818 					     wr->num_sge);
819 				sqe->num_sge = wr->num_sge;
820 			} else {
821 				rv = siw_copy_inline_sgl(wr, sqe);
822 				if (rv <= 0) {
823 					rv = -EINVAL;
824 					break;
825 				}
826 				sqe->flags |= SIW_WQE_INLINE;
827 				sqe->num_sge = 1;
828 			}
829 			if (wr->opcode == IB_WR_SEND)
830 				sqe->opcode = SIW_OP_SEND;
831 			else {
832 				sqe->opcode = SIW_OP_SEND_REMOTE_INV;
833 				sqe->rkey = wr->ex.invalidate_rkey;
834 			}
835 			break;
836 
837 		case IB_WR_RDMA_READ_WITH_INV:
838 		case IB_WR_RDMA_READ:
839 			/*
840 			 * iWarp restricts RREAD sink to SGL containing
841 			 * 1 SGE only. we could relax to SGL with multiple
842 			 * elements referring the SAME ltag or even sending
843 			 * a private per-rreq tag referring to a checked
844 			 * local sgl with MULTIPLE ltag's.
845 			 */
846 			if (unlikely(wr->num_sge != 1)) {
847 				rv = -EINVAL;
848 				break;
849 			}
850 			siw_copy_sgl(wr->sg_list, &sqe->sge[0], 1);
851 			/*
852 			 * NOTE: zero length RREAD is allowed!
853 			 */
854 			sqe->raddr = rdma_wr(wr)->remote_addr;
855 			sqe->rkey = rdma_wr(wr)->rkey;
856 			sqe->num_sge = 1;
857 
858 			if (wr->opcode == IB_WR_RDMA_READ)
859 				sqe->opcode = SIW_OP_READ;
860 			else
861 				sqe->opcode = SIW_OP_READ_LOCAL_INV;
862 			break;
863 
864 		case IB_WR_RDMA_WRITE:
865 			if (!(wr->send_flags & IB_SEND_INLINE)) {
866 				siw_copy_sgl(wr->sg_list, &sqe->sge[0],
867 					     wr->num_sge);
868 				sqe->num_sge = wr->num_sge;
869 			} else {
870 				rv = siw_copy_inline_sgl(wr, sqe);
871 				if (unlikely(rv < 0)) {
872 					rv = -EINVAL;
873 					break;
874 				}
875 				sqe->flags |= SIW_WQE_INLINE;
876 				sqe->num_sge = 1;
877 			}
878 			sqe->raddr = rdma_wr(wr)->remote_addr;
879 			sqe->rkey = rdma_wr(wr)->rkey;
880 			sqe->opcode = SIW_OP_WRITE;
881 			break;
882 
883 		case IB_WR_REG_MR:
884 			sqe->base_mr = (uintptr_t)reg_wr(wr)->mr;
885 			sqe->rkey = reg_wr(wr)->key;
886 			sqe->access = reg_wr(wr)->access & IWARP_ACCESS_MASK;
887 			sqe->opcode = SIW_OP_REG_MR;
888 			break;
889 
890 		case IB_WR_LOCAL_INV:
891 			sqe->rkey = wr->ex.invalidate_rkey;
892 			sqe->opcode = SIW_OP_INVAL_STAG;
893 			break;
894 
895 		default:
896 			siw_dbg_qp(qp, "ib wr type %d unsupported\n",
897 				   wr->opcode);
898 			rv = -EINVAL;
899 			break;
900 		}
901 		siw_dbg_qp(qp, "opcode %d, flags 0x%x, wr_id 0x%pK\n",
902 			   sqe->opcode, sqe->flags,
903 			   (void *)(uintptr_t)sqe->id);
904 
905 		if (unlikely(rv < 0))
906 			break;
907 
908 		/* make SQE only valid after completely written */
909 		smp_wmb();
910 		sqe->flags |= SIW_WQE_VALID;
911 
912 		qp->sq_put++;
913 		wr = wr->next;
914 	}
915 
916 	/*
917 	 * Send directly if SQ processing is not in progress.
918 	 * Eventual immediate errors (rv < 0) do not affect the involved
919 	 * RI resources (Verbs, 8.3.1) and thus do not prevent from SQ
920 	 * processing, if new work is already pending. But rv must be passed
921 	 * to caller.
922 	 */
923 	if (wqe->wr_status != SIW_WR_IDLE) {
924 		spin_unlock_irqrestore(&qp->sq_lock, flags);
925 		goto skip_direct_sending;
926 	}
927 	rv = siw_activate_tx(qp);
928 	spin_unlock_irqrestore(&qp->sq_lock, flags);
929 
930 	if (rv <= 0)
931 		goto skip_direct_sending;
932 
933 	if (rdma_is_kernel_res(&qp->base_qp.res)) {
934 		rv = siw_sq_start(qp);
935 	} else {
936 		qp->tx_ctx.in_syscall = 1;
937 
938 		if (siw_qp_sq_process(qp) != 0 && !(qp->tx_ctx.tx_suspend))
939 			siw_qp_cm_drop(qp, 0);
940 
941 		qp->tx_ctx.in_syscall = 0;
942 	}
943 skip_direct_sending:
944 
945 	up_read(&qp->state_lock);
946 
947 	if (rv >= 0)
948 		return 0;
949 	/*
950 	 * Immediate error
951 	 */
952 	siw_dbg_qp(qp, "error %d\n", rv);
953 
954 	*bad_wr = wr;
955 	return rv;
956 }
957 
958 /*
959  * siw_post_receive()
960  *
961  * Post a list of R-WR's to a RQ.
962  *
963  * @base_qp:	Base QP contained in siw QP
964  * @wr:		Null terminated list of user WR's
965  * @bad_wr:	Points to failing WR in case of synchronous failure.
966  */
967 int siw_post_receive(struct ib_qp *base_qp, const struct ib_recv_wr *wr,
968 		     const struct ib_recv_wr **bad_wr)
969 {
970 	struct siw_qp *qp = to_siw_qp(base_qp);
971 	unsigned long flags;
972 	int rv = 0;
973 
974 	if (qp->srq || qp->attrs.rq_size == 0) {
975 		*bad_wr = wr;
976 		return -EINVAL;
977 	}
978 	if (!rdma_is_kernel_res(&qp->base_qp.res)) {
979 		siw_dbg_qp(qp, "no kernel post_recv for user mapped rq\n");
980 		*bad_wr = wr;
981 		return -EINVAL;
982 	}
983 
984 	/*
985 	 * Try to acquire QP state lock. Must be non-blocking
986 	 * to accommodate kernel clients needs.
987 	 */
988 	if (!down_read_trylock(&qp->state_lock)) {
989 		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
990 			/*
991 			 * ERROR state is final, so we can be sure
992 			 * this state will not change as long as the QP
993 			 * exists.
994 			 *
995 			 * This handles an ib_drain_rq() call with
996 			 * a concurrent request to set the QP state
997 			 * to ERROR.
998 			 */
999 			rv = siw_rq_flush_wr(qp, wr, bad_wr);
1000 		} else {
1001 			siw_dbg_qp(qp, "QP locked, state %d\n",
1002 				   qp->attrs.state);
1003 			*bad_wr = wr;
1004 			rv = -ENOTCONN;
1005 		}
1006 		return rv;
1007 	}
1008 	if (qp->attrs.state > SIW_QP_STATE_RTS) {
1009 		if (qp->attrs.state == SIW_QP_STATE_ERROR) {
1010 			/*
1011 			 * Immediately flush this WR to CQ, if QP
1012 			 * is in ERROR state. RQ is guaranteed to
1013 			 * be empty, so WR complets in-order.
1014 			 *
1015 			 * Typically triggered by ib_drain_rq().
1016 			 */
1017 			rv = siw_rq_flush_wr(qp, wr, bad_wr);
1018 		} else {
1019 			siw_dbg_qp(qp, "QP out of state %d\n",
1020 				   qp->attrs.state);
1021 			*bad_wr = wr;
1022 			rv = -ENOTCONN;
1023 		}
1024 		up_read(&qp->state_lock);
1025 		return rv;
1026 	}
1027 	/*
1028 	 * Serialize potentially multiple producers.
1029 	 * Not needed for single threaded consumer side.
1030 	 */
1031 	spin_lock_irqsave(&qp->rq_lock, flags);
1032 
1033 	while (wr) {
1034 		u32 idx = qp->rq_put % qp->attrs.rq_size;
1035 		struct siw_rqe *rqe = &qp->recvq[idx];
1036 
1037 		if (rqe->flags) {
1038 			siw_dbg_qp(qp, "RQ full\n");
1039 			rv = -ENOMEM;
1040 			break;
1041 		}
1042 		if (wr->num_sge > qp->attrs.rq_max_sges) {
1043 			siw_dbg_qp(qp, "too many sge's: %d\n", wr->num_sge);
1044 			rv = -EINVAL;
1045 			break;
1046 		}
1047 		rqe->id = wr->wr_id;
1048 		rqe->num_sge = wr->num_sge;
1049 		siw_copy_sgl(wr->sg_list, rqe->sge, wr->num_sge);
1050 
1051 		/* make sure RQE is completely written before valid */
1052 		smp_wmb();
1053 
1054 		rqe->flags = SIW_WQE_VALID;
1055 
1056 		qp->rq_put++;
1057 		wr = wr->next;
1058 	}
1059 	spin_unlock_irqrestore(&qp->rq_lock, flags);
1060 
1061 	up_read(&qp->state_lock);
1062 
1063 	if (rv < 0) {
1064 		siw_dbg_qp(qp, "error %d\n", rv);
1065 		*bad_wr = wr;
1066 	}
1067 	return rv > 0 ? 0 : rv;
1068 }
1069 
1070 int siw_destroy_cq(struct ib_cq *base_cq, struct ib_udata *udata)
1071 {
1072 	struct siw_cq *cq = to_siw_cq(base_cq);
1073 	struct siw_device *sdev = to_siw_dev(base_cq->device);
1074 	struct siw_ucontext *ctx =
1075 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
1076 					  base_ucontext);
1077 
1078 	siw_dbg_cq(cq, "free CQ resources\n");
1079 
1080 	siw_cq_flush(cq);
1081 
1082 	if (ctx)
1083 		rdma_user_mmap_entry_remove(cq->cq_entry);
1084 
1085 	atomic_dec(&sdev->num_cq);
1086 
1087 	vfree(cq->queue);
1088 	return 0;
1089 }
1090 
1091 /*
1092  * siw_create_cq()
1093  *
1094  * Populate CQ of requested size
1095  *
1096  * @base_cq: CQ as allocated by RDMA midlayer
1097  * @attr: Initial CQ attributes
1098  * @udata: relates to user context
1099  */
1100 
1101 int siw_create_cq(struct ib_cq *base_cq, const struct ib_cq_init_attr *attr,
1102 		  struct ib_udata *udata)
1103 {
1104 	struct siw_device *sdev = to_siw_dev(base_cq->device);
1105 	struct siw_cq *cq = to_siw_cq(base_cq);
1106 	int rv, size = attr->cqe;
1107 
1108 	if (attr->flags)
1109 		return -EOPNOTSUPP;
1110 
1111 	if (atomic_inc_return(&sdev->num_cq) > SIW_MAX_CQ) {
1112 		siw_dbg(base_cq->device, "too many CQ's\n");
1113 		rv = -ENOMEM;
1114 		goto err_out;
1115 	}
1116 	if (size < 1 || size > sdev->attrs.max_cqe) {
1117 		siw_dbg(base_cq->device, "CQ size error: %d\n", size);
1118 		rv = -EINVAL;
1119 		goto err_out;
1120 	}
1121 	size = roundup_pow_of_two(size);
1122 	cq->base_cq.cqe = size;
1123 	cq->num_cqe = size;
1124 
1125 	if (udata)
1126 		cq->queue = vmalloc_user(size * sizeof(struct siw_cqe) +
1127 					 sizeof(struct siw_cq_ctrl));
1128 	else
1129 		cq->queue = vzalloc(size * sizeof(struct siw_cqe) +
1130 				    sizeof(struct siw_cq_ctrl));
1131 
1132 	if (cq->queue == NULL) {
1133 		rv = -ENOMEM;
1134 		goto err_out;
1135 	}
1136 	get_random_bytes(&cq->id, 4);
1137 	siw_dbg(base_cq->device, "new CQ [%u]\n", cq->id);
1138 
1139 	spin_lock_init(&cq->lock);
1140 
1141 	cq->notify = (struct siw_cq_ctrl *)&cq->queue[size];
1142 
1143 	if (udata) {
1144 		struct siw_uresp_create_cq uresp = {};
1145 		struct siw_ucontext *ctx =
1146 			rdma_udata_to_drv_context(udata, struct siw_ucontext,
1147 						  base_ucontext);
1148 		size_t length = size * sizeof(struct siw_cqe) +
1149 			sizeof(struct siw_cq_ctrl);
1150 
1151 		cq->cq_entry =
1152 			siw_mmap_entry_insert(ctx, cq->queue,
1153 					      length, &uresp.cq_key);
1154 		if (!cq->cq_entry) {
1155 			rv = -ENOMEM;
1156 			goto err_out;
1157 		}
1158 
1159 		uresp.cq_id = cq->id;
1160 		uresp.num_cqe = size;
1161 
1162 		if (udata->outlen < sizeof(uresp)) {
1163 			rv = -EINVAL;
1164 			goto err_out;
1165 		}
1166 		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1167 		if (rv)
1168 			goto err_out;
1169 	}
1170 	return 0;
1171 
1172 err_out:
1173 	siw_dbg(base_cq->device, "CQ creation failed: %d", rv);
1174 
1175 	if (cq && cq->queue) {
1176 		struct siw_ucontext *ctx =
1177 			rdma_udata_to_drv_context(udata, struct siw_ucontext,
1178 						  base_ucontext);
1179 		if (ctx)
1180 			rdma_user_mmap_entry_remove(cq->cq_entry);
1181 		vfree(cq->queue);
1182 	}
1183 	atomic_dec(&sdev->num_cq);
1184 
1185 	return rv;
1186 }
1187 
1188 /*
1189  * siw_poll_cq()
1190  *
1191  * Reap CQ entries if available and copy work completion status into
1192  * array of WC's provided by caller. Returns number of reaped CQE's.
1193  *
1194  * @base_cq:	Base CQ contained in siw CQ.
1195  * @num_cqe:	Maximum number of CQE's to reap.
1196  * @wc:		Array of work completions to be filled by siw.
1197  */
1198 int siw_poll_cq(struct ib_cq *base_cq, int num_cqe, struct ib_wc *wc)
1199 {
1200 	struct siw_cq *cq = to_siw_cq(base_cq);
1201 	int i;
1202 
1203 	for (i = 0; i < num_cqe; i++) {
1204 		if (!siw_reap_cqe(cq, wc))
1205 			break;
1206 		wc++;
1207 	}
1208 	return i;
1209 }
1210 
1211 /*
1212  * siw_req_notify_cq()
1213  *
1214  * Request notification for new CQE's added to that CQ.
1215  * Defined flags:
1216  * o SIW_CQ_NOTIFY_SOLICITED lets siw trigger a notification
1217  *   event if a WQE with notification flag set enters the CQ
1218  * o SIW_CQ_NOTIFY_NEXT_COMP lets siw trigger a notification
1219  *   event if a WQE enters the CQ.
1220  * o IB_CQ_REPORT_MISSED_EVENTS: return value will provide the
1221  *   number of not reaped CQE's regardless of its notification
1222  *   type and current or new CQ notification settings.
1223  *
1224  * @base_cq:	Base CQ contained in siw CQ.
1225  * @flags:	Requested notification flags.
1226  */
1227 int siw_req_notify_cq(struct ib_cq *base_cq, enum ib_cq_notify_flags flags)
1228 {
1229 	struct siw_cq *cq = to_siw_cq(base_cq);
1230 
1231 	siw_dbg_cq(cq, "flags: 0x%02x\n", flags);
1232 
1233 	if ((flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED)
1234 		/*
1235 		 * Enable CQ event for next solicited completion.
1236 		 * and make it visible to all associated producers.
1237 		 */
1238 		smp_store_mb(cq->notify->flags, SIW_NOTIFY_SOLICITED);
1239 	else
1240 		/*
1241 		 * Enable CQ event for any signalled completion.
1242 		 * and make it visible to all associated producers.
1243 		 */
1244 		smp_store_mb(cq->notify->flags, SIW_NOTIFY_ALL);
1245 
1246 	if (flags & IB_CQ_REPORT_MISSED_EVENTS)
1247 		return cq->cq_put - cq->cq_get;
1248 
1249 	return 0;
1250 }
1251 
1252 /*
1253  * siw_dereg_mr()
1254  *
1255  * Release Memory Region.
1256  *
1257  * @base_mr: Base MR contained in siw MR.
1258  * @udata: points to user context, unused.
1259  */
1260 int siw_dereg_mr(struct ib_mr *base_mr, struct ib_udata *udata)
1261 {
1262 	struct siw_mr *mr = to_siw_mr(base_mr);
1263 	struct siw_device *sdev = to_siw_dev(base_mr->device);
1264 
1265 	siw_dbg_mem(mr->mem, "deregister MR\n");
1266 
1267 	atomic_dec(&sdev->num_mr);
1268 
1269 	siw_mr_drop_mem(mr);
1270 	kfree_rcu(mr, rcu);
1271 
1272 	return 0;
1273 }
1274 
1275 /*
1276  * siw_reg_user_mr()
1277  *
1278  * Register Memory Region.
1279  *
1280  * @pd:		Protection Domain
1281  * @start:	starting address of MR (virtual address)
1282  * @len:	len of MR
1283  * @rnic_va:	not used by siw
1284  * @rights:	MR access rights
1285  * @udata:	user buffer to communicate STag and Key.
1286  */
1287 struct ib_mr *siw_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,
1288 			      u64 rnic_va, int rights, struct ib_udata *udata)
1289 {
1290 	struct siw_mr *mr = NULL;
1291 	struct siw_umem *umem = NULL;
1292 	struct siw_ureq_reg_mr ureq;
1293 	struct siw_device *sdev = to_siw_dev(pd->device);
1294 
1295 	unsigned long mem_limit = rlimit(RLIMIT_MEMLOCK);
1296 	int rv;
1297 
1298 	siw_dbg_pd(pd, "start: 0x%pK, va: 0x%pK, len: %llu\n",
1299 		   (void *)(uintptr_t)start, (void *)(uintptr_t)rnic_va,
1300 		   (unsigned long long)len);
1301 
1302 	if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1303 		siw_dbg_pd(pd, "too many mr's\n");
1304 		rv = -ENOMEM;
1305 		goto err_out;
1306 	}
1307 	if (!len) {
1308 		rv = -EINVAL;
1309 		goto err_out;
1310 	}
1311 	if (mem_limit != RLIM_INFINITY) {
1312 		unsigned long num_pages =
1313 			(PAGE_ALIGN(len + (start & ~PAGE_MASK))) >> PAGE_SHIFT;
1314 		mem_limit >>= PAGE_SHIFT;
1315 
1316 		if (num_pages > mem_limit - current->mm->locked_vm) {
1317 			siw_dbg_pd(pd, "pages req %lu, max %lu, lock %lu\n",
1318 				   num_pages, mem_limit,
1319 				   current->mm->locked_vm);
1320 			rv = -ENOMEM;
1321 			goto err_out;
1322 		}
1323 	}
1324 	umem = siw_umem_get(start, len, ib_access_writable(rights));
1325 	if (IS_ERR(umem)) {
1326 		rv = PTR_ERR(umem);
1327 		siw_dbg_pd(pd, "getting user memory failed: %d\n", rv);
1328 		umem = NULL;
1329 		goto err_out;
1330 	}
1331 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1332 	if (!mr) {
1333 		rv = -ENOMEM;
1334 		goto err_out;
1335 	}
1336 	rv = siw_mr_add_mem(mr, pd, umem, start, len, rights);
1337 	if (rv)
1338 		goto err_out;
1339 
1340 	if (udata) {
1341 		struct siw_uresp_reg_mr uresp = {};
1342 		struct siw_mem *mem = mr->mem;
1343 
1344 		if (udata->inlen < sizeof(ureq)) {
1345 			rv = -EINVAL;
1346 			goto err_out;
1347 		}
1348 		rv = ib_copy_from_udata(&ureq, udata, sizeof(ureq));
1349 		if (rv)
1350 			goto err_out;
1351 
1352 		mr->base_mr.lkey |= ureq.stag_key;
1353 		mr->base_mr.rkey |= ureq.stag_key;
1354 		mem->stag |= ureq.stag_key;
1355 		uresp.stag = mem->stag;
1356 
1357 		if (udata->outlen < sizeof(uresp)) {
1358 			rv = -EINVAL;
1359 			goto err_out;
1360 		}
1361 		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1362 		if (rv)
1363 			goto err_out;
1364 	}
1365 	mr->mem->stag_valid = 1;
1366 
1367 	return &mr->base_mr;
1368 
1369 err_out:
1370 	atomic_dec(&sdev->num_mr);
1371 	if (mr) {
1372 		if (mr->mem)
1373 			siw_mr_drop_mem(mr);
1374 		kfree_rcu(mr, rcu);
1375 	} else {
1376 		if (umem)
1377 			siw_umem_release(umem, false);
1378 	}
1379 	return ERR_PTR(rv);
1380 }
1381 
1382 struct ib_mr *siw_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
1383 			   u32 max_sge)
1384 {
1385 	struct siw_device *sdev = to_siw_dev(pd->device);
1386 	struct siw_mr *mr = NULL;
1387 	struct siw_pbl *pbl = NULL;
1388 	int rv;
1389 
1390 	if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1391 		siw_dbg_pd(pd, "too many mr's\n");
1392 		rv = -ENOMEM;
1393 		goto err_out;
1394 	}
1395 	if (mr_type != IB_MR_TYPE_MEM_REG) {
1396 		siw_dbg_pd(pd, "mr type %d unsupported\n", mr_type);
1397 		rv = -EOPNOTSUPP;
1398 		goto err_out;
1399 	}
1400 	if (max_sge > SIW_MAX_SGE_PBL) {
1401 		siw_dbg_pd(pd, "too many sge's: %d\n", max_sge);
1402 		rv = -ENOMEM;
1403 		goto err_out;
1404 	}
1405 	pbl = siw_pbl_alloc(max_sge);
1406 	if (IS_ERR(pbl)) {
1407 		rv = PTR_ERR(pbl);
1408 		siw_dbg_pd(pd, "pbl allocation failed: %d\n", rv);
1409 		pbl = NULL;
1410 		goto err_out;
1411 	}
1412 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1413 	if (!mr) {
1414 		rv = -ENOMEM;
1415 		goto err_out;
1416 	}
1417 	rv = siw_mr_add_mem(mr, pd, pbl, 0, max_sge * PAGE_SIZE, 0);
1418 	if (rv)
1419 		goto err_out;
1420 
1421 	mr->mem->is_pbl = 1;
1422 
1423 	siw_dbg_pd(pd, "[MEM %u]: success\n", mr->mem->stag);
1424 
1425 	return &mr->base_mr;
1426 
1427 err_out:
1428 	atomic_dec(&sdev->num_mr);
1429 
1430 	if (!mr) {
1431 		kfree(pbl);
1432 	} else {
1433 		if (mr->mem)
1434 			siw_mr_drop_mem(mr);
1435 		kfree_rcu(mr, rcu);
1436 	}
1437 	siw_dbg_pd(pd, "failed: %d\n", rv);
1438 
1439 	return ERR_PTR(rv);
1440 }
1441 
1442 /* Just used to count number of pages being mapped */
1443 static int siw_set_pbl_page(struct ib_mr *base_mr, u64 buf_addr)
1444 {
1445 	return 0;
1446 }
1447 
1448 int siw_map_mr_sg(struct ib_mr *base_mr, struct scatterlist *sl, int num_sle,
1449 		  unsigned int *sg_off)
1450 {
1451 	struct scatterlist *slp;
1452 	struct siw_mr *mr = to_siw_mr(base_mr);
1453 	struct siw_mem *mem = mr->mem;
1454 	struct siw_pbl *pbl = mem->pbl;
1455 	struct siw_pble *pble;
1456 	unsigned long pbl_size;
1457 	int i, rv;
1458 
1459 	if (!pbl) {
1460 		siw_dbg_mem(mem, "no PBL allocated\n");
1461 		return -EINVAL;
1462 	}
1463 	pble = pbl->pbe;
1464 
1465 	if (pbl->max_buf < num_sle) {
1466 		siw_dbg_mem(mem, "too many SGE's: %d > %d\n",
1467 			    mem->pbl->max_buf, num_sle);
1468 		return -ENOMEM;
1469 	}
1470 	for_each_sg(sl, slp, num_sle, i) {
1471 		if (sg_dma_len(slp) == 0) {
1472 			siw_dbg_mem(mem, "empty SGE\n");
1473 			return -EINVAL;
1474 		}
1475 		if (i == 0) {
1476 			pble->addr = sg_dma_address(slp);
1477 			pble->size = sg_dma_len(slp);
1478 			pble->pbl_off = 0;
1479 			pbl_size = pble->size;
1480 			pbl->num_buf = 1;
1481 		} else {
1482 			/* Merge PBL entries if adjacent */
1483 			if (pble->addr + pble->size == sg_dma_address(slp)) {
1484 				pble->size += sg_dma_len(slp);
1485 			} else {
1486 				pble++;
1487 				pbl->num_buf++;
1488 				pble->addr = sg_dma_address(slp);
1489 				pble->size = sg_dma_len(slp);
1490 				pble->pbl_off = pbl_size;
1491 			}
1492 			pbl_size += sg_dma_len(slp);
1493 		}
1494 		siw_dbg_mem(mem,
1495 			"sge[%d], size %u, addr 0x%p, total %lu\n",
1496 			i, pble->size, (void *)(uintptr_t)pble->addr,
1497 			pbl_size);
1498 	}
1499 	rv = ib_sg_to_pages(base_mr, sl, num_sle, sg_off, siw_set_pbl_page);
1500 	if (rv > 0) {
1501 		mem->len = base_mr->length;
1502 		mem->va = base_mr->iova;
1503 		siw_dbg_mem(mem,
1504 			"%llu bytes, start 0x%pK, %u SLE to %u entries\n",
1505 			mem->len, (void *)(uintptr_t)mem->va, num_sle,
1506 			pbl->num_buf);
1507 	}
1508 	return rv;
1509 }
1510 
1511 /*
1512  * siw_get_dma_mr()
1513  *
1514  * Create a (empty) DMA memory region, where no umem is attached.
1515  */
1516 struct ib_mr *siw_get_dma_mr(struct ib_pd *pd, int rights)
1517 {
1518 	struct siw_device *sdev = to_siw_dev(pd->device);
1519 	struct siw_mr *mr = NULL;
1520 	int rv;
1521 
1522 	if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
1523 		siw_dbg_pd(pd, "too many mr's\n");
1524 		rv = -ENOMEM;
1525 		goto err_out;
1526 	}
1527 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1528 	if (!mr) {
1529 		rv = -ENOMEM;
1530 		goto err_out;
1531 	}
1532 	rv = siw_mr_add_mem(mr, pd, NULL, 0, ULONG_MAX, rights);
1533 	if (rv)
1534 		goto err_out;
1535 
1536 	mr->mem->stag_valid = 1;
1537 
1538 	siw_dbg_pd(pd, "[MEM %u]: success\n", mr->mem->stag);
1539 
1540 	return &mr->base_mr;
1541 
1542 err_out:
1543 	if (rv)
1544 		kfree(mr);
1545 
1546 	atomic_dec(&sdev->num_mr);
1547 
1548 	return ERR_PTR(rv);
1549 }
1550 
1551 /*
1552  * siw_create_srq()
1553  *
1554  * Create Shared Receive Queue of attributes @init_attrs
1555  * within protection domain given by @pd.
1556  *
1557  * @base_srq:	Base SRQ contained in siw SRQ.
1558  * @init_attrs:	SRQ init attributes.
1559  * @udata:	points to user context
1560  */
1561 int siw_create_srq(struct ib_srq *base_srq,
1562 		   struct ib_srq_init_attr *init_attrs, struct ib_udata *udata)
1563 {
1564 	struct siw_srq *srq = to_siw_srq(base_srq);
1565 	struct ib_srq_attr *attrs = &init_attrs->attr;
1566 	struct siw_device *sdev = to_siw_dev(base_srq->device);
1567 	struct siw_ucontext *ctx =
1568 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
1569 					  base_ucontext);
1570 	int rv;
1571 
1572 	if (init_attrs->srq_type != IB_SRQT_BASIC)
1573 		return -EOPNOTSUPP;
1574 
1575 	if (atomic_inc_return(&sdev->num_srq) > SIW_MAX_SRQ) {
1576 		siw_dbg_pd(base_srq->pd, "too many SRQ's\n");
1577 		rv = -ENOMEM;
1578 		goto err_out;
1579 	}
1580 	if (attrs->max_wr == 0 || attrs->max_wr > SIW_MAX_SRQ_WR ||
1581 	    attrs->max_sge > SIW_MAX_SGE || attrs->srq_limit > attrs->max_wr) {
1582 		rv = -EINVAL;
1583 		goto err_out;
1584 	}
1585 	srq->max_sge = attrs->max_sge;
1586 	srq->num_rqe = roundup_pow_of_two(attrs->max_wr);
1587 	srq->limit = attrs->srq_limit;
1588 	if (srq->limit)
1589 		srq->armed = true;
1590 
1591 	srq->is_kernel_res = !udata;
1592 
1593 	if (udata)
1594 		srq->recvq =
1595 			vmalloc_user(srq->num_rqe * sizeof(struct siw_rqe));
1596 	else
1597 		srq->recvq = vzalloc(srq->num_rqe * sizeof(struct siw_rqe));
1598 
1599 	if (srq->recvq == NULL) {
1600 		rv = -ENOMEM;
1601 		goto err_out;
1602 	}
1603 	if (udata) {
1604 		struct siw_uresp_create_srq uresp = {};
1605 		size_t length = srq->num_rqe * sizeof(struct siw_rqe);
1606 
1607 		srq->srq_entry =
1608 			siw_mmap_entry_insert(ctx, srq->recvq,
1609 					      length, &uresp.srq_key);
1610 		if (!srq->srq_entry) {
1611 			rv = -ENOMEM;
1612 			goto err_out;
1613 		}
1614 
1615 		uresp.num_rqe = srq->num_rqe;
1616 
1617 		if (udata->outlen < sizeof(uresp)) {
1618 			rv = -EINVAL;
1619 			goto err_out;
1620 		}
1621 		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1622 		if (rv)
1623 			goto err_out;
1624 	}
1625 	spin_lock_init(&srq->lock);
1626 
1627 	siw_dbg_pd(base_srq->pd, "[SRQ]: success\n");
1628 
1629 	return 0;
1630 
1631 err_out:
1632 	if (srq->recvq) {
1633 		if (ctx)
1634 			rdma_user_mmap_entry_remove(srq->srq_entry);
1635 		vfree(srq->recvq);
1636 	}
1637 	atomic_dec(&sdev->num_srq);
1638 
1639 	return rv;
1640 }
1641 
1642 /*
1643  * siw_modify_srq()
1644  *
1645  * Modify SRQ. The caller may resize SRQ and/or set/reset notification
1646  * limit and (re)arm IB_EVENT_SRQ_LIMIT_REACHED notification.
1647  *
1648  * NOTE: it is unclear if RDMA core allows for changing the MAX_SGE
1649  * parameter. siw_modify_srq() does not check the attrs->max_sge param.
1650  */
1651 int siw_modify_srq(struct ib_srq *base_srq, struct ib_srq_attr *attrs,
1652 		   enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
1653 {
1654 	struct siw_srq *srq = to_siw_srq(base_srq);
1655 	unsigned long flags;
1656 	int rv = 0;
1657 
1658 	spin_lock_irqsave(&srq->lock, flags);
1659 
1660 	if (attr_mask & IB_SRQ_MAX_WR) {
1661 		/* resize request not yet supported */
1662 		rv = -EOPNOTSUPP;
1663 		goto out;
1664 	}
1665 	if (attr_mask & IB_SRQ_LIMIT) {
1666 		if (attrs->srq_limit) {
1667 			if (unlikely(attrs->srq_limit > srq->num_rqe)) {
1668 				rv = -EINVAL;
1669 				goto out;
1670 			}
1671 			srq->armed = true;
1672 		} else {
1673 			srq->armed = false;
1674 		}
1675 		srq->limit = attrs->srq_limit;
1676 	}
1677 out:
1678 	spin_unlock_irqrestore(&srq->lock, flags);
1679 
1680 	return rv;
1681 }
1682 
1683 /*
1684  * siw_query_srq()
1685  *
1686  * Query SRQ attributes.
1687  */
1688 int siw_query_srq(struct ib_srq *base_srq, struct ib_srq_attr *attrs)
1689 {
1690 	struct siw_srq *srq = to_siw_srq(base_srq);
1691 	unsigned long flags;
1692 
1693 	spin_lock_irqsave(&srq->lock, flags);
1694 
1695 	attrs->max_wr = srq->num_rqe;
1696 	attrs->max_sge = srq->max_sge;
1697 	attrs->srq_limit = srq->limit;
1698 
1699 	spin_unlock_irqrestore(&srq->lock, flags);
1700 
1701 	return 0;
1702 }
1703 
1704 /*
1705  * siw_destroy_srq()
1706  *
1707  * Destroy SRQ.
1708  * It is assumed that the SRQ is not referenced by any
1709  * QP anymore - the code trusts the RDMA core environment to keep track
1710  * of QP references.
1711  */
1712 int siw_destroy_srq(struct ib_srq *base_srq, struct ib_udata *udata)
1713 {
1714 	struct siw_srq *srq = to_siw_srq(base_srq);
1715 	struct siw_device *sdev = to_siw_dev(base_srq->device);
1716 	struct siw_ucontext *ctx =
1717 		rdma_udata_to_drv_context(udata, struct siw_ucontext,
1718 					  base_ucontext);
1719 
1720 	if (ctx)
1721 		rdma_user_mmap_entry_remove(srq->srq_entry);
1722 	vfree(srq->recvq);
1723 	atomic_dec(&sdev->num_srq);
1724 	return 0;
1725 }
1726 
1727 /*
1728  * siw_post_srq_recv()
1729  *
1730  * Post a list of receive queue elements to SRQ.
1731  * NOTE: The function does not check or lock a certain SRQ state
1732  *       during the post operation. The code simply trusts the
1733  *       RDMA core environment.
1734  *
1735  * @base_srq:	Base SRQ contained in siw SRQ
1736  * @wr:		List of R-WR's
1737  * @bad_wr:	Updated to failing WR if posting fails.
1738  */
1739 int siw_post_srq_recv(struct ib_srq *base_srq, const struct ib_recv_wr *wr,
1740 		      const struct ib_recv_wr **bad_wr)
1741 {
1742 	struct siw_srq *srq = to_siw_srq(base_srq);
1743 	unsigned long flags;
1744 	int rv = 0;
1745 
1746 	if (unlikely(!srq->is_kernel_res)) {
1747 		siw_dbg_pd(base_srq->pd,
1748 			   "[SRQ]: no kernel post_recv for mapped srq\n");
1749 		rv = -EINVAL;
1750 		goto out;
1751 	}
1752 	/*
1753 	 * Serialize potentially multiple producers.
1754 	 * Also needed to serialize potentially multiple
1755 	 * consumers.
1756 	 */
1757 	spin_lock_irqsave(&srq->lock, flags);
1758 
1759 	while (wr) {
1760 		u32 idx = srq->rq_put % srq->num_rqe;
1761 		struct siw_rqe *rqe = &srq->recvq[idx];
1762 
1763 		if (rqe->flags) {
1764 			siw_dbg_pd(base_srq->pd, "SRQ full\n");
1765 			rv = -ENOMEM;
1766 			break;
1767 		}
1768 		if (unlikely(wr->num_sge > srq->max_sge)) {
1769 			siw_dbg_pd(base_srq->pd,
1770 				   "[SRQ]: too many sge's: %d\n", wr->num_sge);
1771 			rv = -EINVAL;
1772 			break;
1773 		}
1774 		rqe->id = wr->wr_id;
1775 		rqe->num_sge = wr->num_sge;
1776 		siw_copy_sgl(wr->sg_list, rqe->sge, wr->num_sge);
1777 
1778 		/* Make sure S-RQE is completely written before valid */
1779 		smp_wmb();
1780 
1781 		rqe->flags = SIW_WQE_VALID;
1782 
1783 		srq->rq_put++;
1784 		wr = wr->next;
1785 	}
1786 	spin_unlock_irqrestore(&srq->lock, flags);
1787 out:
1788 	if (unlikely(rv < 0)) {
1789 		siw_dbg_pd(base_srq->pd, "[SRQ]: error %d\n", rv);
1790 		*bad_wr = wr;
1791 	}
1792 	return rv;
1793 }
1794 
1795 void siw_qp_event(struct siw_qp *qp, enum ib_event_type etype)
1796 {
1797 	struct ib_event event;
1798 	struct ib_qp *base_qp = &qp->base_qp;
1799 
1800 	/*
1801 	 * Do not report asynchronous errors on QP which gets
1802 	 * destroyed via verbs interface (siw_destroy_qp())
1803 	 */
1804 	if (qp->attrs.flags & SIW_QP_IN_DESTROY)
1805 		return;
1806 
1807 	event.event = etype;
1808 	event.device = base_qp->device;
1809 	event.element.qp = base_qp;
1810 
1811 	if (base_qp->event_handler) {
1812 		siw_dbg_qp(qp, "reporting event %d\n", etype);
1813 		base_qp->event_handler(&event, base_qp->qp_context);
1814 	}
1815 }
1816 
1817 void siw_cq_event(struct siw_cq *cq, enum ib_event_type etype)
1818 {
1819 	struct ib_event event;
1820 	struct ib_cq *base_cq = &cq->base_cq;
1821 
1822 	event.event = etype;
1823 	event.device = base_cq->device;
1824 	event.element.cq = base_cq;
1825 
1826 	if (base_cq->event_handler) {
1827 		siw_dbg_cq(cq, "reporting CQ event %d\n", etype);
1828 		base_cq->event_handler(&event, base_cq->cq_context);
1829 	}
1830 }
1831 
1832 void siw_srq_event(struct siw_srq *srq, enum ib_event_type etype)
1833 {
1834 	struct ib_event event;
1835 	struct ib_srq *base_srq = &srq->base_srq;
1836 
1837 	event.event = etype;
1838 	event.device = base_srq->device;
1839 	event.element.srq = base_srq;
1840 
1841 	if (base_srq->event_handler) {
1842 		siw_dbg_pd(srq->base_srq.pd,
1843 			   "reporting SRQ event %d\n", etype);
1844 		base_srq->event_handler(&event, base_srq->srq_context);
1845 	}
1846 }
1847 
1848 void siw_port_event(struct siw_device *sdev, u32 port, enum ib_event_type etype)
1849 {
1850 	struct ib_event event;
1851 
1852 	event.event = etype;
1853 	event.device = &sdev->base_dev;
1854 	event.element.port_num = port;
1855 
1856 	siw_dbg(&sdev->base_dev, "reporting port event %d\n", etype);
1857 
1858 	ib_dispatch_event(&event);
1859 }
1860