xref: /linux/drivers/infiniband/sw/rdmavt/qp.c (revision 32786fdc9506aeba98278c1844d4bfb766863832)
1 /*
2  * Copyright(c) 2016 Intel Corporation.
3  *
4  * This file is provided under a dual BSD/GPLv2 license.  When using or
5  * redistributing this file, you may do so under either license.
6  *
7  * GPL LICENSE SUMMARY
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of version 2 of the GNU General Public License as
11  * published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful, but
14  * WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * General Public License for more details.
17  *
18  * BSD LICENSE
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  *
24  *  - Redistributions of source code must retain the above copyright
25  *    notice, this list of conditions and the following disclaimer.
26  *  - Redistributions in binary form must reproduce the above copyright
27  *    notice, this list of conditions and the following disclaimer in
28  *    the documentation and/or other materials provided with the
29  *    distribution.
30  *  - Neither the name of Intel Corporation nor the names of its
31  *    contributors may be used to endorse or promote products derived
32  *    from this software without specific prior written permission.
33  *
34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45  *
46  */
47 
48 #include <linux/hash.h>
49 #include <linux/bitops.h>
50 #include <linux/lockdep.h>
51 #include <linux/vmalloc.h>
52 #include <linux/slab.h>
53 #include <rdma/ib_verbs.h>
54 #include "qp.h"
55 #include "vt.h"
56 #include "trace.h"
57 
58 /*
59  * Note that it is OK to post send work requests in the SQE and ERR
60  * states; rvt_do_send() will process them and generate error
61  * completions as per IB 1.2 C10-96.
62  */
63 const int ib_rvt_state_ops[IB_QPS_ERR + 1] = {
64 	[IB_QPS_RESET] = 0,
65 	[IB_QPS_INIT] = RVT_POST_RECV_OK,
66 	[IB_QPS_RTR] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK,
67 	[IB_QPS_RTS] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK |
68 	    RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK |
69 	    RVT_PROCESS_NEXT_SEND_OK,
70 	[IB_QPS_SQD] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK |
71 	    RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK,
72 	[IB_QPS_SQE] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK |
73 	    RVT_POST_SEND_OK | RVT_FLUSH_SEND,
74 	[IB_QPS_ERR] = RVT_POST_RECV_OK | RVT_FLUSH_RECV |
75 	    RVT_POST_SEND_OK | RVT_FLUSH_SEND,
76 };
77 EXPORT_SYMBOL(ib_rvt_state_ops);
78 
79 /*
80  * Translate ib_wr_opcode into ib_wc_opcode.
81  */
82 const enum ib_wc_opcode ib_rvt_wc_opcode[] = {
83 	[IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
84 	[IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
85 	[IB_WR_SEND] = IB_WC_SEND,
86 	[IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
87 	[IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
88 	[IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
89 	[IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD,
90 	[IB_WR_SEND_WITH_INV] = IB_WC_SEND,
91 	[IB_WR_LOCAL_INV] = IB_WC_LOCAL_INV,
92 	[IB_WR_REG_MR] = IB_WC_REG_MR
93 };
94 EXPORT_SYMBOL(ib_rvt_wc_opcode);
95 
96 static void get_map_page(struct rvt_qpn_table *qpt,
97 			 struct rvt_qpn_map *map,
98 			 gfp_t gfp)
99 {
100 	unsigned long page = get_zeroed_page(gfp);
101 
102 	/*
103 	 * Free the page if someone raced with us installing it.
104 	 */
105 
106 	spin_lock(&qpt->lock);
107 	if (map->page)
108 		free_page(page);
109 	else
110 		map->page = (void *)page;
111 	spin_unlock(&qpt->lock);
112 }
113 
114 /**
115  * init_qpn_table - initialize the QP number table for a device
116  * @qpt: the QPN table
117  */
118 static int init_qpn_table(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt)
119 {
120 	u32 offset, i;
121 	struct rvt_qpn_map *map;
122 	int ret = 0;
123 
124 	if (!(rdi->dparms.qpn_res_end >= rdi->dparms.qpn_res_start))
125 		return -EINVAL;
126 
127 	spin_lock_init(&qpt->lock);
128 
129 	qpt->last = rdi->dparms.qpn_start;
130 	qpt->incr = rdi->dparms.qpn_inc << rdi->dparms.qos_shift;
131 
132 	/*
133 	 * Drivers may want some QPs beyond what we need for verbs let them use
134 	 * our qpn table. No need for two. Lets go ahead and mark the bitmaps
135 	 * for those. The reserved range must be *after* the range which verbs
136 	 * will pick from.
137 	 */
138 
139 	/* Figure out number of bit maps needed before reserved range */
140 	qpt->nmaps = rdi->dparms.qpn_res_start / RVT_BITS_PER_PAGE;
141 
142 	/* This should always be zero */
143 	offset = rdi->dparms.qpn_res_start & RVT_BITS_PER_PAGE_MASK;
144 
145 	/* Starting with the first reserved bit map */
146 	map = &qpt->map[qpt->nmaps];
147 
148 	rvt_pr_info(rdi, "Reserving QPNs from 0x%x to 0x%x for non-verbs use\n",
149 		    rdi->dparms.qpn_res_start, rdi->dparms.qpn_res_end);
150 	for (i = rdi->dparms.qpn_res_start; i <= rdi->dparms.qpn_res_end; i++) {
151 		if (!map->page) {
152 			get_map_page(qpt, map, GFP_KERNEL);
153 			if (!map->page) {
154 				ret = -ENOMEM;
155 				break;
156 			}
157 		}
158 		set_bit(offset, map->page);
159 		offset++;
160 		if (offset == RVT_BITS_PER_PAGE) {
161 			/* next page */
162 			qpt->nmaps++;
163 			map++;
164 			offset = 0;
165 		}
166 	}
167 	return ret;
168 }
169 
170 /**
171  * free_qpn_table - free the QP number table for a device
172  * @qpt: the QPN table
173  */
174 static void free_qpn_table(struct rvt_qpn_table *qpt)
175 {
176 	int i;
177 
178 	for (i = 0; i < ARRAY_SIZE(qpt->map); i++)
179 		free_page((unsigned long)qpt->map[i].page);
180 }
181 
182 /**
183  * rvt_driver_qp_init - Init driver qp resources
184  * @rdi: rvt dev strucutre
185  *
186  * Return: 0 on success
187  */
188 int rvt_driver_qp_init(struct rvt_dev_info *rdi)
189 {
190 	int i;
191 	int ret = -ENOMEM;
192 
193 	if (!rdi->dparms.qp_table_size)
194 		return -EINVAL;
195 
196 	/*
197 	 * If driver is not doing any QP allocation then make sure it is
198 	 * providing the necessary QP functions.
199 	 */
200 	if (!rdi->driver_f.free_all_qps ||
201 	    !rdi->driver_f.qp_priv_alloc ||
202 	    !rdi->driver_f.qp_priv_free ||
203 	    !rdi->driver_f.notify_qp_reset)
204 		return -EINVAL;
205 
206 	/* allocate parent object */
207 	rdi->qp_dev = kzalloc_node(sizeof(*rdi->qp_dev), GFP_KERNEL,
208 				   rdi->dparms.node);
209 	if (!rdi->qp_dev)
210 		return -ENOMEM;
211 
212 	/* allocate hash table */
213 	rdi->qp_dev->qp_table_size = rdi->dparms.qp_table_size;
214 	rdi->qp_dev->qp_table_bits = ilog2(rdi->dparms.qp_table_size);
215 	rdi->qp_dev->qp_table =
216 		kmalloc_node(rdi->qp_dev->qp_table_size *
217 			     sizeof(*rdi->qp_dev->qp_table),
218 			     GFP_KERNEL, rdi->dparms.node);
219 	if (!rdi->qp_dev->qp_table)
220 		goto no_qp_table;
221 
222 	for (i = 0; i < rdi->qp_dev->qp_table_size; i++)
223 		RCU_INIT_POINTER(rdi->qp_dev->qp_table[i], NULL);
224 
225 	spin_lock_init(&rdi->qp_dev->qpt_lock);
226 
227 	/* initialize qpn map */
228 	if (init_qpn_table(rdi, &rdi->qp_dev->qpn_table))
229 		goto fail_table;
230 
231 	spin_lock_init(&rdi->n_qps_lock);
232 
233 	return 0;
234 
235 fail_table:
236 	kfree(rdi->qp_dev->qp_table);
237 	free_qpn_table(&rdi->qp_dev->qpn_table);
238 
239 no_qp_table:
240 	kfree(rdi->qp_dev);
241 
242 	return ret;
243 }
244 
245 /**
246  * free_all_qps - check for QPs still in use
247  * @qpt: the QP table to empty
248  *
249  * There should not be any QPs still in use.
250  * Free memory for table.
251  */
252 static unsigned rvt_free_all_qps(struct rvt_dev_info *rdi)
253 {
254 	unsigned long flags;
255 	struct rvt_qp *qp;
256 	unsigned n, qp_inuse = 0;
257 	spinlock_t *ql; /* work around too long line below */
258 
259 	if (rdi->driver_f.free_all_qps)
260 		qp_inuse = rdi->driver_f.free_all_qps(rdi);
261 
262 	qp_inuse += rvt_mcast_tree_empty(rdi);
263 
264 	if (!rdi->qp_dev)
265 		return qp_inuse;
266 
267 	ql = &rdi->qp_dev->qpt_lock;
268 	spin_lock_irqsave(ql, flags);
269 	for (n = 0; n < rdi->qp_dev->qp_table_size; n++) {
270 		qp = rcu_dereference_protected(rdi->qp_dev->qp_table[n],
271 					       lockdep_is_held(ql));
272 		RCU_INIT_POINTER(rdi->qp_dev->qp_table[n], NULL);
273 
274 		for (; qp; qp = rcu_dereference_protected(qp->next,
275 							  lockdep_is_held(ql)))
276 			qp_inuse++;
277 	}
278 	spin_unlock_irqrestore(ql, flags);
279 	synchronize_rcu();
280 	return qp_inuse;
281 }
282 
283 /**
284  * rvt_qp_exit - clean up qps on device exit
285  * @rdi: rvt dev structure
286  *
287  * Check for qp leaks and free resources.
288  */
289 void rvt_qp_exit(struct rvt_dev_info *rdi)
290 {
291 	u32 qps_inuse = rvt_free_all_qps(rdi);
292 
293 	if (qps_inuse)
294 		rvt_pr_err(rdi, "QP memory leak! %u still in use\n",
295 			   qps_inuse);
296 	if (!rdi->qp_dev)
297 		return;
298 
299 	kfree(rdi->qp_dev->qp_table);
300 	free_qpn_table(&rdi->qp_dev->qpn_table);
301 	kfree(rdi->qp_dev);
302 }
303 
304 static inline unsigned mk_qpn(struct rvt_qpn_table *qpt,
305 			      struct rvt_qpn_map *map, unsigned off)
306 {
307 	return (map - qpt->map) * RVT_BITS_PER_PAGE + off;
308 }
309 
310 /**
311  * alloc_qpn - Allocate the next available qpn or zero/one for QP type
312  *	       IB_QPT_SMI/IB_QPT_GSI
313  *@rdi:	rvt device info structure
314  *@qpt: queue pair number table pointer
315  *@port_num: IB port number, 1 based, comes from core
316  *
317  * Return: The queue pair number
318  */
319 static int alloc_qpn(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt,
320 		     enum ib_qp_type type, u8 port_num, gfp_t gfp)
321 {
322 	u32 i, offset, max_scan, qpn;
323 	struct rvt_qpn_map *map;
324 	u32 ret;
325 
326 	if (rdi->driver_f.alloc_qpn)
327 		return rdi->driver_f.alloc_qpn(rdi, qpt, type, port_num, gfp);
328 
329 	if (type == IB_QPT_SMI || type == IB_QPT_GSI) {
330 		unsigned n;
331 
332 		ret = type == IB_QPT_GSI;
333 		n = 1 << (ret + 2 * (port_num - 1));
334 		spin_lock(&qpt->lock);
335 		if (qpt->flags & n)
336 			ret = -EINVAL;
337 		else
338 			qpt->flags |= n;
339 		spin_unlock(&qpt->lock);
340 		goto bail;
341 	}
342 
343 	qpn = qpt->last + qpt->incr;
344 	if (qpn >= RVT_QPN_MAX)
345 		qpn = qpt->incr | ((qpt->last & 1) ^ 1);
346 	/* offset carries bit 0 */
347 	offset = qpn & RVT_BITS_PER_PAGE_MASK;
348 	map = &qpt->map[qpn / RVT_BITS_PER_PAGE];
349 	max_scan = qpt->nmaps - !offset;
350 	for (i = 0;;) {
351 		if (unlikely(!map->page)) {
352 			get_map_page(qpt, map, gfp);
353 			if (unlikely(!map->page))
354 				break;
355 		}
356 		do {
357 			if (!test_and_set_bit(offset, map->page)) {
358 				qpt->last = qpn;
359 				ret = qpn;
360 				goto bail;
361 			}
362 			offset += qpt->incr;
363 			/*
364 			 * This qpn might be bogus if offset >= BITS_PER_PAGE.
365 			 * That is OK.   It gets re-assigned below
366 			 */
367 			qpn = mk_qpn(qpt, map, offset);
368 		} while (offset < RVT_BITS_PER_PAGE && qpn < RVT_QPN_MAX);
369 		/*
370 		 * In order to keep the number of pages allocated to a
371 		 * minimum, we scan the all existing pages before increasing
372 		 * the size of the bitmap table.
373 		 */
374 		if (++i > max_scan) {
375 			if (qpt->nmaps == RVT_QPNMAP_ENTRIES)
376 				break;
377 			map = &qpt->map[qpt->nmaps++];
378 			/* start at incr with current bit 0 */
379 			offset = qpt->incr | (offset & 1);
380 		} else if (map < &qpt->map[qpt->nmaps]) {
381 			++map;
382 			/* start at incr with current bit 0 */
383 			offset = qpt->incr | (offset & 1);
384 		} else {
385 			map = &qpt->map[0];
386 			/* wrap to first map page, invert bit 0 */
387 			offset = qpt->incr | ((offset & 1) ^ 1);
388 		}
389 		/* there can be no set bits in low-order QoS bits */
390 		WARN_ON(offset & (BIT(rdi->dparms.qos_shift) - 1));
391 		qpn = mk_qpn(qpt, map, offset);
392 	}
393 
394 	ret = -ENOMEM;
395 
396 bail:
397 	return ret;
398 }
399 
400 static void free_qpn(struct rvt_qpn_table *qpt, u32 qpn)
401 {
402 	struct rvt_qpn_map *map;
403 
404 	map = qpt->map + qpn / RVT_BITS_PER_PAGE;
405 	if (map->page)
406 		clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page);
407 }
408 
409 /**
410  * rvt_clear_mr_refs - Drop help mr refs
411  * @qp: rvt qp data structure
412  * @clr_sends: If shoudl clear send side or not
413  */
414 static void rvt_clear_mr_refs(struct rvt_qp *qp, int clr_sends)
415 {
416 	unsigned n;
417 	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
418 
419 	if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags))
420 		rvt_put_ss(&qp->s_rdma_read_sge);
421 
422 	rvt_put_ss(&qp->r_sge);
423 
424 	if (clr_sends) {
425 		while (qp->s_last != qp->s_head) {
426 			struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_last);
427 			unsigned i;
428 
429 			for (i = 0; i < wqe->wr.num_sge; i++) {
430 				struct rvt_sge *sge = &wqe->sg_list[i];
431 
432 				rvt_put_mr(sge->mr);
433 			}
434 			if (qp->ibqp.qp_type == IB_QPT_UD ||
435 			    qp->ibqp.qp_type == IB_QPT_SMI ||
436 			    qp->ibqp.qp_type == IB_QPT_GSI)
437 				atomic_dec(&ibah_to_rvtah(
438 						wqe->ud_wr.ah)->refcount);
439 			if (++qp->s_last >= qp->s_size)
440 				qp->s_last = 0;
441 			smp_wmb(); /* see qp_set_savail */
442 		}
443 		if (qp->s_rdma_mr) {
444 			rvt_put_mr(qp->s_rdma_mr);
445 			qp->s_rdma_mr = NULL;
446 		}
447 	}
448 
449 	if (qp->ibqp.qp_type != IB_QPT_RC)
450 		return;
451 
452 	for (n = 0; n < rvt_max_atomic(rdi); n++) {
453 		struct rvt_ack_entry *e = &qp->s_ack_queue[n];
454 
455 		if (e->rdma_sge.mr) {
456 			rvt_put_mr(e->rdma_sge.mr);
457 			e->rdma_sge.mr = NULL;
458 		}
459 	}
460 }
461 
462 /**
463  * rvt_remove_qp - remove qp form table
464  * @rdi: rvt dev struct
465  * @qp: qp to remove
466  *
467  * Remove the QP from the table so it can't be found asynchronously by
468  * the receive routine.
469  */
470 static void rvt_remove_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp)
471 {
472 	struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1];
473 	u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits);
474 	unsigned long flags;
475 	int removed = 1;
476 
477 	spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags);
478 
479 	if (rcu_dereference_protected(rvp->qp[0],
480 			lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) {
481 		RCU_INIT_POINTER(rvp->qp[0], NULL);
482 	} else if (rcu_dereference_protected(rvp->qp[1],
483 			lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) {
484 		RCU_INIT_POINTER(rvp->qp[1], NULL);
485 	} else {
486 		struct rvt_qp *q;
487 		struct rvt_qp __rcu **qpp;
488 
489 		removed = 0;
490 		qpp = &rdi->qp_dev->qp_table[n];
491 		for (; (q = rcu_dereference_protected(*qpp,
492 			lockdep_is_held(&rdi->qp_dev->qpt_lock))) != NULL;
493 			qpp = &q->next) {
494 			if (q == qp) {
495 				RCU_INIT_POINTER(*qpp,
496 				     rcu_dereference_protected(qp->next,
497 				     lockdep_is_held(&rdi->qp_dev->qpt_lock)));
498 				removed = 1;
499 				trace_rvt_qpremove(qp, n);
500 				break;
501 			}
502 		}
503 	}
504 
505 	spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags);
506 	if (removed) {
507 		synchronize_rcu();
508 		rvt_put_qp(qp);
509 	}
510 }
511 
512 /**
513  * rvt_init_qp - initialize the QP state to the reset state
514  * @qp: the QP to init or reinit
515  * @type: the QP type
516  *
517  * This function is called from both rvt_create_qp() and
518  * rvt_reset_qp().   The difference is that the reset
519  * patch the necessary locks to protect against concurent
520  * access.
521  */
522 static void rvt_init_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp,
523 			enum ib_qp_type type)
524 {
525 	qp->remote_qpn = 0;
526 	qp->qkey = 0;
527 	qp->qp_access_flags = 0;
528 	qp->s_flags &= RVT_S_SIGNAL_REQ_WR;
529 	qp->s_hdrwords = 0;
530 	qp->s_wqe = NULL;
531 	qp->s_draining = 0;
532 	qp->s_next_psn = 0;
533 	qp->s_last_psn = 0;
534 	qp->s_sending_psn = 0;
535 	qp->s_sending_hpsn = 0;
536 	qp->s_psn = 0;
537 	qp->r_psn = 0;
538 	qp->r_msn = 0;
539 	if (type == IB_QPT_RC) {
540 		qp->s_state = IB_OPCODE_RC_SEND_LAST;
541 		qp->r_state = IB_OPCODE_RC_SEND_LAST;
542 	} else {
543 		qp->s_state = IB_OPCODE_UC_SEND_LAST;
544 		qp->r_state = IB_OPCODE_UC_SEND_LAST;
545 	}
546 	qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
547 	qp->r_nak_state = 0;
548 	qp->r_aflags = 0;
549 	qp->r_flags = 0;
550 	qp->s_head = 0;
551 	qp->s_tail = 0;
552 	qp->s_cur = 0;
553 	qp->s_acked = 0;
554 	qp->s_last = 0;
555 	qp->s_ssn = 1;
556 	qp->s_lsn = 0;
557 	qp->s_mig_state = IB_MIG_MIGRATED;
558 	qp->r_head_ack_queue = 0;
559 	qp->s_tail_ack_queue = 0;
560 	qp->s_num_rd_atomic = 0;
561 	if (qp->r_rq.wq) {
562 		qp->r_rq.wq->head = 0;
563 		qp->r_rq.wq->tail = 0;
564 	}
565 	qp->r_sge.num_sge = 0;
566 	atomic_set(&qp->s_reserved_used, 0);
567 }
568 
569 /**
570  * rvt_reset_qp - initialize the QP state to the reset state
571  * @qp: the QP to reset
572  * @type: the QP type
573  *
574  * r_lock, s_hlock, and s_lock are required to be held by the caller
575  */
576 static void rvt_reset_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp,
577 			 enum ib_qp_type type)
578 	__must_hold(&qp->s_lock)
579 	__must_hold(&qp->s_hlock)
580 	__must_hold(&qp->r_lock)
581 {
582 	lockdep_assert_held(&qp->r_lock);
583 	lockdep_assert_held(&qp->s_hlock);
584 	lockdep_assert_held(&qp->s_lock);
585 	if (qp->state != IB_QPS_RESET) {
586 		qp->state = IB_QPS_RESET;
587 
588 		/* Let drivers flush their waitlist */
589 		rdi->driver_f.flush_qp_waiters(qp);
590 		qp->s_flags &= ~(RVT_S_TIMER | RVT_S_ANY_WAIT);
591 		spin_unlock(&qp->s_lock);
592 		spin_unlock(&qp->s_hlock);
593 		spin_unlock_irq(&qp->r_lock);
594 
595 		/* Stop the send queue and the retry timer */
596 		rdi->driver_f.stop_send_queue(qp);
597 
598 		/* Wait for things to stop */
599 		rdi->driver_f.quiesce_qp(qp);
600 
601 		/* take qp out the hash and wait for it to be unused */
602 		rvt_remove_qp(rdi, qp);
603 		wait_event(qp->wait, !atomic_read(&qp->refcount));
604 
605 		/* grab the lock b/c it was locked at call time */
606 		spin_lock_irq(&qp->r_lock);
607 		spin_lock(&qp->s_hlock);
608 		spin_lock(&qp->s_lock);
609 
610 		rvt_clear_mr_refs(qp, 1);
611 		/*
612 		 * Let the driver do any tear down or re-init it needs to for
613 		 * a qp that has been reset
614 		 */
615 		rdi->driver_f.notify_qp_reset(qp);
616 	}
617 	rvt_init_qp(rdi, qp, type);
618 	lockdep_assert_held(&qp->r_lock);
619 	lockdep_assert_held(&qp->s_hlock);
620 	lockdep_assert_held(&qp->s_lock);
621 }
622 
623 /**
624  * rvt_create_qp - create a queue pair for a device
625  * @ibpd: the protection domain who's device we create the queue pair for
626  * @init_attr: the attributes of the queue pair
627  * @udata: user data for libibverbs.so
628  *
629  * Queue pair creation is mostly an rvt issue. However, drivers have their own
630  * unique idea of what queue pair numbers mean. For instance there is a reserved
631  * range for PSM.
632  *
633  * Return: the queue pair on success, otherwise returns an errno.
634  *
635  * Called by the ib_create_qp() core verbs function.
636  */
637 struct ib_qp *rvt_create_qp(struct ib_pd *ibpd,
638 			    struct ib_qp_init_attr *init_attr,
639 			    struct ib_udata *udata)
640 {
641 	struct rvt_qp *qp;
642 	int err;
643 	struct rvt_swqe *swq = NULL;
644 	size_t sz;
645 	size_t sg_list_sz;
646 	struct ib_qp *ret = ERR_PTR(-ENOMEM);
647 	struct rvt_dev_info *rdi = ib_to_rvt(ibpd->device);
648 	void *priv = NULL;
649 	gfp_t gfp;
650 	size_t sqsize;
651 
652 	if (!rdi)
653 		return ERR_PTR(-EINVAL);
654 
655 	if (init_attr->cap.max_send_sge > rdi->dparms.props.max_sge ||
656 	    init_attr->cap.max_send_wr > rdi->dparms.props.max_qp_wr ||
657 	    init_attr->create_flags & ~(IB_QP_CREATE_USE_GFP_NOIO))
658 		return ERR_PTR(-EINVAL);
659 
660 	/* GFP_NOIO is applicable to RC QP's only */
661 
662 	if (init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO &&
663 	    init_attr->qp_type != IB_QPT_RC)
664 		return ERR_PTR(-EINVAL);
665 
666 	gfp = init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO ?
667 						GFP_NOIO : GFP_KERNEL;
668 
669 	/* Check receive queue parameters if no SRQ is specified. */
670 	if (!init_attr->srq) {
671 		if (init_attr->cap.max_recv_sge > rdi->dparms.props.max_sge ||
672 		    init_attr->cap.max_recv_wr > rdi->dparms.props.max_qp_wr)
673 			return ERR_PTR(-EINVAL);
674 
675 		if (init_attr->cap.max_send_sge +
676 		    init_attr->cap.max_send_wr +
677 		    init_attr->cap.max_recv_sge +
678 		    init_attr->cap.max_recv_wr == 0)
679 			return ERR_PTR(-EINVAL);
680 	}
681 	sqsize =
682 		init_attr->cap.max_send_wr + 1 +
683 		rdi->dparms.reserved_operations;
684 	switch (init_attr->qp_type) {
685 	case IB_QPT_SMI:
686 	case IB_QPT_GSI:
687 		if (init_attr->port_num == 0 ||
688 		    init_attr->port_num > ibpd->device->phys_port_cnt)
689 			return ERR_PTR(-EINVAL);
690 	case IB_QPT_UC:
691 	case IB_QPT_RC:
692 	case IB_QPT_UD:
693 		sz = sizeof(struct rvt_sge) *
694 			init_attr->cap.max_send_sge +
695 			sizeof(struct rvt_swqe);
696 		if (gfp == GFP_NOIO)
697 			swq = __vmalloc(
698 				sqsize * sz,
699 				gfp | __GFP_ZERO, PAGE_KERNEL);
700 		else
701 			swq = vzalloc_node(
702 				sqsize * sz,
703 				rdi->dparms.node);
704 		if (!swq)
705 			return ERR_PTR(-ENOMEM);
706 
707 		sz = sizeof(*qp);
708 		sg_list_sz = 0;
709 		if (init_attr->srq) {
710 			struct rvt_srq *srq = ibsrq_to_rvtsrq(init_attr->srq);
711 
712 			if (srq->rq.max_sge > 1)
713 				sg_list_sz = sizeof(*qp->r_sg_list) *
714 					(srq->rq.max_sge - 1);
715 		} else if (init_attr->cap.max_recv_sge > 1)
716 			sg_list_sz = sizeof(*qp->r_sg_list) *
717 				(init_attr->cap.max_recv_sge - 1);
718 		qp = kzalloc_node(sz + sg_list_sz, gfp, rdi->dparms.node);
719 		if (!qp)
720 			goto bail_swq;
721 
722 		RCU_INIT_POINTER(qp->next, NULL);
723 		if (init_attr->qp_type == IB_QPT_RC) {
724 			qp->s_ack_queue =
725 				kzalloc_node(
726 					sizeof(*qp->s_ack_queue) *
727 					 rvt_max_atomic(rdi),
728 					gfp,
729 					rdi->dparms.node);
730 			if (!qp->s_ack_queue)
731 				goto bail_qp;
732 		}
733 
734 		/*
735 		 * Driver needs to set up it's private QP structure and do any
736 		 * initialization that is needed.
737 		 */
738 		priv = rdi->driver_f.qp_priv_alloc(rdi, qp, gfp);
739 		if (IS_ERR(priv)) {
740 			ret = priv;
741 			goto bail_qp;
742 		}
743 		qp->priv = priv;
744 		qp->timeout_jiffies =
745 			usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
746 				1000UL);
747 		if (init_attr->srq) {
748 			sz = 0;
749 		} else {
750 			qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
751 			qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
752 			sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
753 				sizeof(struct rvt_rwqe);
754 			if (udata)
755 				qp->r_rq.wq = vmalloc_user(
756 						sizeof(struct rvt_rwq) +
757 						qp->r_rq.size * sz);
758 			else if (gfp == GFP_NOIO)
759 				qp->r_rq.wq = __vmalloc(
760 						sizeof(struct rvt_rwq) +
761 						qp->r_rq.size * sz,
762 						gfp | __GFP_ZERO, PAGE_KERNEL);
763 			else
764 				qp->r_rq.wq = vzalloc_node(
765 						sizeof(struct rvt_rwq) +
766 						qp->r_rq.size * sz,
767 						rdi->dparms.node);
768 			if (!qp->r_rq.wq)
769 				goto bail_driver_priv;
770 		}
771 
772 		/*
773 		 * ib_create_qp() will initialize qp->ibqp
774 		 * except for qp->ibqp.qp_num.
775 		 */
776 		spin_lock_init(&qp->r_lock);
777 		spin_lock_init(&qp->s_hlock);
778 		spin_lock_init(&qp->s_lock);
779 		spin_lock_init(&qp->r_rq.lock);
780 		atomic_set(&qp->refcount, 0);
781 		atomic_set(&qp->local_ops_pending, 0);
782 		init_waitqueue_head(&qp->wait);
783 		init_timer(&qp->s_timer);
784 		qp->s_timer.data = (unsigned long)qp;
785 		INIT_LIST_HEAD(&qp->rspwait);
786 		qp->state = IB_QPS_RESET;
787 		qp->s_wq = swq;
788 		qp->s_size = sqsize;
789 		qp->s_avail = init_attr->cap.max_send_wr;
790 		qp->s_max_sge = init_attr->cap.max_send_sge;
791 		if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
792 			qp->s_flags = RVT_S_SIGNAL_REQ_WR;
793 
794 		err = alloc_qpn(rdi, &rdi->qp_dev->qpn_table,
795 				init_attr->qp_type,
796 				init_attr->port_num, gfp);
797 		if (err < 0) {
798 			ret = ERR_PTR(err);
799 			goto bail_rq_wq;
800 		}
801 		qp->ibqp.qp_num = err;
802 		qp->port_num = init_attr->port_num;
803 		rvt_init_qp(rdi, qp, init_attr->qp_type);
804 		break;
805 
806 	default:
807 		/* Don't support raw QPs */
808 		return ERR_PTR(-EINVAL);
809 	}
810 
811 	init_attr->cap.max_inline_data = 0;
812 
813 	/*
814 	 * Return the address of the RWQ as the offset to mmap.
815 	 * See rvt_mmap() for details.
816 	 */
817 	if (udata && udata->outlen >= sizeof(__u64)) {
818 		if (!qp->r_rq.wq) {
819 			__u64 offset = 0;
820 
821 			err = ib_copy_to_udata(udata, &offset,
822 					       sizeof(offset));
823 			if (err) {
824 				ret = ERR_PTR(err);
825 				goto bail_qpn;
826 			}
827 		} else {
828 			u32 s = sizeof(struct rvt_rwq) + qp->r_rq.size * sz;
829 
830 			qp->ip = rvt_create_mmap_info(rdi, s,
831 						      ibpd->uobject->context,
832 						      qp->r_rq.wq);
833 			if (!qp->ip) {
834 				ret = ERR_PTR(-ENOMEM);
835 				goto bail_qpn;
836 			}
837 
838 			err = ib_copy_to_udata(udata, &qp->ip->offset,
839 					       sizeof(qp->ip->offset));
840 			if (err) {
841 				ret = ERR_PTR(err);
842 				goto bail_ip;
843 			}
844 		}
845 		qp->pid = current->pid;
846 	}
847 
848 	spin_lock(&rdi->n_qps_lock);
849 	if (rdi->n_qps_allocated == rdi->dparms.props.max_qp) {
850 		spin_unlock(&rdi->n_qps_lock);
851 		ret = ERR_PTR(-ENOMEM);
852 		goto bail_ip;
853 	}
854 
855 	rdi->n_qps_allocated++;
856 	/*
857 	 * Maintain a busy_jiffies variable that will be added to the timeout
858 	 * period in mod_retry_timer and add_retry_timer. This busy jiffies
859 	 * is scaled by the number of rc qps created for the device to reduce
860 	 * the number of timeouts occurring when there is a large number of
861 	 * qps. busy_jiffies is incremented every rc qp scaling interval.
862 	 * The scaling interval is selected based on extensive performance
863 	 * evaluation of targeted workloads.
864 	 */
865 	if (init_attr->qp_type == IB_QPT_RC) {
866 		rdi->n_rc_qps++;
867 		rdi->busy_jiffies = rdi->n_rc_qps / RC_QP_SCALING_INTERVAL;
868 	}
869 	spin_unlock(&rdi->n_qps_lock);
870 
871 	if (qp->ip) {
872 		spin_lock_irq(&rdi->pending_lock);
873 		list_add(&qp->ip->pending_mmaps, &rdi->pending_mmaps);
874 		spin_unlock_irq(&rdi->pending_lock);
875 	}
876 
877 	ret = &qp->ibqp;
878 
879 	/*
880 	 * We have our QP and its good, now keep track of what types of opcodes
881 	 * can be processed on this QP. We do this by keeping track of what the
882 	 * 3 high order bits of the opcode are.
883 	 */
884 	switch (init_attr->qp_type) {
885 	case IB_QPT_SMI:
886 	case IB_QPT_GSI:
887 	case IB_QPT_UD:
888 		qp->allowed_ops = IB_OPCODE_UD;
889 		break;
890 	case IB_QPT_RC:
891 		qp->allowed_ops = IB_OPCODE_RC;
892 		break;
893 	case IB_QPT_UC:
894 		qp->allowed_ops = IB_OPCODE_UC;
895 		break;
896 	default:
897 		ret = ERR_PTR(-EINVAL);
898 		goto bail_ip;
899 	}
900 
901 	return ret;
902 
903 bail_ip:
904 	if (qp->ip)
905 		kref_put(&qp->ip->ref, rvt_release_mmap_info);
906 
907 bail_qpn:
908 	free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num);
909 
910 bail_rq_wq:
911 	if (!qp->ip)
912 		vfree(qp->r_rq.wq);
913 
914 bail_driver_priv:
915 	rdi->driver_f.qp_priv_free(rdi, qp);
916 
917 bail_qp:
918 	kfree(qp->s_ack_queue);
919 	kfree(qp);
920 
921 bail_swq:
922 	vfree(swq);
923 
924 	return ret;
925 }
926 
927 /**
928  * rvt_error_qp - put a QP into the error state
929  * @qp: the QP to put into the error state
930  * @err: the receive completion error to signal if a RWQE is active
931  *
932  * Flushes both send and receive work queues.
933  *
934  * Return: true if last WQE event should be generated.
935  * The QP r_lock and s_lock should be held and interrupts disabled.
936  * If we are already in error state, just return.
937  */
938 int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err)
939 {
940 	struct ib_wc wc;
941 	int ret = 0;
942 	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
943 
944 	lockdep_assert_held(&qp->r_lock);
945 	lockdep_assert_held(&qp->s_lock);
946 	if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET)
947 		goto bail;
948 
949 	qp->state = IB_QPS_ERR;
950 
951 	if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) {
952 		qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR);
953 		del_timer(&qp->s_timer);
954 	}
955 
956 	if (qp->s_flags & RVT_S_ANY_WAIT_SEND)
957 		qp->s_flags &= ~RVT_S_ANY_WAIT_SEND;
958 
959 	rdi->driver_f.notify_error_qp(qp);
960 
961 	/* Schedule the sending tasklet to drain the send work queue. */
962 	if (ACCESS_ONCE(qp->s_last) != qp->s_head)
963 		rdi->driver_f.schedule_send(qp);
964 
965 	rvt_clear_mr_refs(qp, 0);
966 
967 	memset(&wc, 0, sizeof(wc));
968 	wc.qp = &qp->ibqp;
969 	wc.opcode = IB_WC_RECV;
970 
971 	if (test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) {
972 		wc.wr_id = qp->r_wr_id;
973 		wc.status = err;
974 		rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
975 	}
976 	wc.status = IB_WC_WR_FLUSH_ERR;
977 
978 	if (qp->r_rq.wq) {
979 		struct rvt_rwq *wq;
980 		u32 head;
981 		u32 tail;
982 
983 		spin_lock(&qp->r_rq.lock);
984 
985 		/* sanity check pointers before trusting them */
986 		wq = qp->r_rq.wq;
987 		head = wq->head;
988 		if (head >= qp->r_rq.size)
989 			head = 0;
990 		tail = wq->tail;
991 		if (tail >= qp->r_rq.size)
992 			tail = 0;
993 		while (tail != head) {
994 			wc.wr_id = rvt_get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
995 			if (++tail >= qp->r_rq.size)
996 				tail = 0;
997 			rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
998 		}
999 		wq->tail = tail;
1000 
1001 		spin_unlock(&qp->r_rq.lock);
1002 	} else if (qp->ibqp.event_handler) {
1003 		ret = 1;
1004 	}
1005 
1006 bail:
1007 	return ret;
1008 }
1009 EXPORT_SYMBOL(rvt_error_qp);
1010 
1011 /*
1012  * Put the QP into the hash table.
1013  * The hash table holds a reference to the QP.
1014  */
1015 static void rvt_insert_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp)
1016 {
1017 	struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1];
1018 	unsigned long flags;
1019 
1020 	rvt_get_qp(qp);
1021 	spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags);
1022 
1023 	if (qp->ibqp.qp_num <= 1) {
1024 		rcu_assign_pointer(rvp->qp[qp->ibqp.qp_num], qp);
1025 	} else {
1026 		u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits);
1027 
1028 		qp->next = rdi->qp_dev->qp_table[n];
1029 		rcu_assign_pointer(rdi->qp_dev->qp_table[n], qp);
1030 		trace_rvt_qpinsert(qp, n);
1031 	}
1032 
1033 	spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags);
1034 }
1035 
1036 /**
1037  * rvt_modify_qp - modify the attributes of a queue pair
1038  * @ibqp: the queue pair who's attributes we're modifying
1039  * @attr: the new attributes
1040  * @attr_mask: the mask of attributes to modify
1041  * @udata: user data for libibverbs.so
1042  *
1043  * Return: 0 on success, otherwise returns an errno.
1044  */
1045 int rvt_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1046 		  int attr_mask, struct ib_udata *udata)
1047 {
1048 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1049 	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1050 	enum ib_qp_state cur_state, new_state;
1051 	struct ib_event ev;
1052 	int lastwqe = 0;
1053 	int mig = 0;
1054 	int pmtu = 0; /* for gcc warning only */
1055 	enum rdma_link_layer link;
1056 
1057 	link = rdma_port_get_link_layer(ibqp->device, qp->port_num);
1058 
1059 	spin_lock_irq(&qp->r_lock);
1060 	spin_lock(&qp->s_hlock);
1061 	spin_lock(&qp->s_lock);
1062 
1063 	cur_state = attr_mask & IB_QP_CUR_STATE ?
1064 		attr->cur_qp_state : qp->state;
1065 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1066 
1067 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1068 				attr_mask, link))
1069 		goto inval;
1070 
1071 	if (rdi->driver_f.check_modify_qp &&
1072 	    rdi->driver_f.check_modify_qp(qp, attr, attr_mask, udata))
1073 		goto inval;
1074 
1075 	if (attr_mask & IB_QP_AV) {
1076 		if (attr->ah_attr.dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE))
1077 			goto inval;
1078 		if (rvt_check_ah(qp->ibqp.device, &attr->ah_attr))
1079 			goto inval;
1080 	}
1081 
1082 	if (attr_mask & IB_QP_ALT_PATH) {
1083 		if (attr->alt_ah_attr.dlid >=
1084 		    be16_to_cpu(IB_MULTICAST_LID_BASE))
1085 			goto inval;
1086 		if (rvt_check_ah(qp->ibqp.device, &attr->alt_ah_attr))
1087 			goto inval;
1088 		if (attr->alt_pkey_index >= rvt_get_npkeys(rdi))
1089 			goto inval;
1090 	}
1091 
1092 	if (attr_mask & IB_QP_PKEY_INDEX)
1093 		if (attr->pkey_index >= rvt_get_npkeys(rdi))
1094 			goto inval;
1095 
1096 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
1097 		if (attr->min_rnr_timer > 31)
1098 			goto inval;
1099 
1100 	if (attr_mask & IB_QP_PORT)
1101 		if (qp->ibqp.qp_type == IB_QPT_SMI ||
1102 		    qp->ibqp.qp_type == IB_QPT_GSI ||
1103 		    attr->port_num == 0 ||
1104 		    attr->port_num > ibqp->device->phys_port_cnt)
1105 			goto inval;
1106 
1107 	if (attr_mask & IB_QP_DEST_QPN)
1108 		if (attr->dest_qp_num > RVT_QPN_MASK)
1109 			goto inval;
1110 
1111 	if (attr_mask & IB_QP_RETRY_CNT)
1112 		if (attr->retry_cnt > 7)
1113 			goto inval;
1114 
1115 	if (attr_mask & IB_QP_RNR_RETRY)
1116 		if (attr->rnr_retry > 7)
1117 			goto inval;
1118 
1119 	/*
1120 	 * Don't allow invalid path_mtu values.  OK to set greater
1121 	 * than the active mtu (or even the max_cap, if we have tuned
1122 	 * that to a small mtu.  We'll set qp->path_mtu
1123 	 * to the lesser of requested attribute mtu and active,
1124 	 * for packetizing messages.
1125 	 * Note that the QP port has to be set in INIT and MTU in RTR.
1126 	 */
1127 	if (attr_mask & IB_QP_PATH_MTU) {
1128 		pmtu = rdi->driver_f.get_pmtu_from_attr(rdi, qp, attr);
1129 		if (pmtu < 0)
1130 			goto inval;
1131 	}
1132 
1133 	if (attr_mask & IB_QP_PATH_MIG_STATE) {
1134 		if (attr->path_mig_state == IB_MIG_REARM) {
1135 			if (qp->s_mig_state == IB_MIG_ARMED)
1136 				goto inval;
1137 			if (new_state != IB_QPS_RTS)
1138 				goto inval;
1139 		} else if (attr->path_mig_state == IB_MIG_MIGRATED) {
1140 			if (qp->s_mig_state == IB_MIG_REARM)
1141 				goto inval;
1142 			if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD)
1143 				goto inval;
1144 			if (qp->s_mig_state == IB_MIG_ARMED)
1145 				mig = 1;
1146 		} else {
1147 			goto inval;
1148 		}
1149 	}
1150 
1151 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1152 		if (attr->max_dest_rd_atomic > rdi->dparms.max_rdma_atomic)
1153 			goto inval;
1154 
1155 	switch (new_state) {
1156 	case IB_QPS_RESET:
1157 		if (qp->state != IB_QPS_RESET)
1158 			rvt_reset_qp(rdi, qp, ibqp->qp_type);
1159 		break;
1160 
1161 	case IB_QPS_RTR:
1162 		/* Allow event to re-trigger if QP set to RTR more than once */
1163 		qp->r_flags &= ~RVT_R_COMM_EST;
1164 		qp->state = new_state;
1165 		break;
1166 
1167 	case IB_QPS_SQD:
1168 		qp->s_draining = qp->s_last != qp->s_cur;
1169 		qp->state = new_state;
1170 		break;
1171 
1172 	case IB_QPS_SQE:
1173 		if (qp->ibqp.qp_type == IB_QPT_RC)
1174 			goto inval;
1175 		qp->state = new_state;
1176 		break;
1177 
1178 	case IB_QPS_ERR:
1179 		lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1180 		break;
1181 
1182 	default:
1183 		qp->state = new_state;
1184 		break;
1185 	}
1186 
1187 	if (attr_mask & IB_QP_PKEY_INDEX)
1188 		qp->s_pkey_index = attr->pkey_index;
1189 
1190 	if (attr_mask & IB_QP_PORT)
1191 		qp->port_num = attr->port_num;
1192 
1193 	if (attr_mask & IB_QP_DEST_QPN)
1194 		qp->remote_qpn = attr->dest_qp_num;
1195 
1196 	if (attr_mask & IB_QP_SQ_PSN) {
1197 		qp->s_next_psn = attr->sq_psn & rdi->dparms.psn_modify_mask;
1198 		qp->s_psn = qp->s_next_psn;
1199 		qp->s_sending_psn = qp->s_next_psn;
1200 		qp->s_last_psn = qp->s_next_psn - 1;
1201 		qp->s_sending_hpsn = qp->s_last_psn;
1202 	}
1203 
1204 	if (attr_mask & IB_QP_RQ_PSN)
1205 		qp->r_psn = attr->rq_psn & rdi->dparms.psn_modify_mask;
1206 
1207 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1208 		qp->qp_access_flags = attr->qp_access_flags;
1209 
1210 	if (attr_mask & IB_QP_AV) {
1211 		qp->remote_ah_attr = attr->ah_attr;
1212 		qp->s_srate = attr->ah_attr.static_rate;
1213 		qp->srate_mbps = ib_rate_to_mbps(qp->s_srate);
1214 	}
1215 
1216 	if (attr_mask & IB_QP_ALT_PATH) {
1217 		qp->alt_ah_attr = attr->alt_ah_attr;
1218 		qp->s_alt_pkey_index = attr->alt_pkey_index;
1219 	}
1220 
1221 	if (attr_mask & IB_QP_PATH_MIG_STATE) {
1222 		qp->s_mig_state = attr->path_mig_state;
1223 		if (mig) {
1224 			qp->remote_ah_attr = qp->alt_ah_attr;
1225 			qp->port_num = qp->alt_ah_attr.port_num;
1226 			qp->s_pkey_index = qp->s_alt_pkey_index;
1227 		}
1228 	}
1229 
1230 	if (attr_mask & IB_QP_PATH_MTU) {
1231 		qp->pmtu = rdi->driver_f.mtu_from_qp(rdi, qp, pmtu);
1232 		qp->path_mtu = rdi->driver_f.mtu_to_path_mtu(qp->pmtu);
1233 		qp->log_pmtu = ilog2(qp->pmtu);
1234 	}
1235 
1236 	if (attr_mask & IB_QP_RETRY_CNT) {
1237 		qp->s_retry_cnt = attr->retry_cnt;
1238 		qp->s_retry = attr->retry_cnt;
1239 	}
1240 
1241 	if (attr_mask & IB_QP_RNR_RETRY) {
1242 		qp->s_rnr_retry_cnt = attr->rnr_retry;
1243 		qp->s_rnr_retry = attr->rnr_retry;
1244 	}
1245 
1246 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
1247 		qp->r_min_rnr_timer = attr->min_rnr_timer;
1248 
1249 	if (attr_mask & IB_QP_TIMEOUT) {
1250 		qp->timeout = attr->timeout;
1251 		qp->timeout_jiffies =
1252 			usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
1253 				1000UL);
1254 	}
1255 
1256 	if (attr_mask & IB_QP_QKEY)
1257 		qp->qkey = attr->qkey;
1258 
1259 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1260 		qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
1261 
1262 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
1263 		qp->s_max_rd_atomic = attr->max_rd_atomic;
1264 
1265 	if (rdi->driver_f.modify_qp)
1266 		rdi->driver_f.modify_qp(qp, attr, attr_mask, udata);
1267 
1268 	spin_unlock(&qp->s_lock);
1269 	spin_unlock(&qp->s_hlock);
1270 	spin_unlock_irq(&qp->r_lock);
1271 
1272 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1273 		rvt_insert_qp(rdi, qp);
1274 
1275 	if (lastwqe) {
1276 		ev.device = qp->ibqp.device;
1277 		ev.element.qp = &qp->ibqp;
1278 		ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
1279 		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1280 	}
1281 	if (mig) {
1282 		ev.device = qp->ibqp.device;
1283 		ev.element.qp = &qp->ibqp;
1284 		ev.event = IB_EVENT_PATH_MIG;
1285 		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1286 	}
1287 	return 0;
1288 
1289 inval:
1290 	spin_unlock(&qp->s_lock);
1291 	spin_unlock(&qp->s_hlock);
1292 	spin_unlock_irq(&qp->r_lock);
1293 	return -EINVAL;
1294 }
1295 
1296 /** rvt_free_qpn - Free a qpn from the bit map
1297  * @qpt: QP table
1298  * @qpn: queue pair number to free
1299  */
1300 static void rvt_free_qpn(struct rvt_qpn_table *qpt, u32 qpn)
1301 {
1302 	struct rvt_qpn_map *map;
1303 
1304 	map = qpt->map + qpn / RVT_BITS_PER_PAGE;
1305 	if (map->page)
1306 		clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page);
1307 }
1308 
1309 /**
1310  * rvt_destroy_qp - destroy a queue pair
1311  * @ibqp: the queue pair to destroy
1312  *
1313  * Note that this can be called while the QP is actively sending or
1314  * receiving!
1315  *
1316  * Return: 0 on success.
1317  */
1318 int rvt_destroy_qp(struct ib_qp *ibqp)
1319 {
1320 	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1321 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1322 
1323 	spin_lock_irq(&qp->r_lock);
1324 	spin_lock(&qp->s_hlock);
1325 	spin_lock(&qp->s_lock);
1326 	rvt_reset_qp(rdi, qp, ibqp->qp_type);
1327 	spin_unlock(&qp->s_lock);
1328 	spin_unlock(&qp->s_hlock);
1329 	spin_unlock_irq(&qp->r_lock);
1330 
1331 	/* qpn is now available for use again */
1332 	rvt_free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num);
1333 
1334 	spin_lock(&rdi->n_qps_lock);
1335 	rdi->n_qps_allocated--;
1336 	if (qp->ibqp.qp_type == IB_QPT_RC) {
1337 		rdi->n_rc_qps--;
1338 		rdi->busy_jiffies = rdi->n_rc_qps / RC_QP_SCALING_INTERVAL;
1339 	}
1340 	spin_unlock(&rdi->n_qps_lock);
1341 
1342 	if (qp->ip)
1343 		kref_put(&qp->ip->ref, rvt_release_mmap_info);
1344 	else
1345 		vfree(qp->r_rq.wq);
1346 	vfree(qp->s_wq);
1347 	rdi->driver_f.qp_priv_free(rdi, qp);
1348 	kfree(qp->s_ack_queue);
1349 	kfree(qp);
1350 	return 0;
1351 }
1352 
1353 /**
1354  * rvt_query_qp - query an ipbq
1355  * @ibqp: IB qp to query
1356  * @attr: attr struct to fill in
1357  * @attr_mask: attr mask ignored
1358  * @init_attr: struct to fill in
1359  *
1360  * Return: always 0
1361  */
1362 int rvt_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1363 		 int attr_mask, struct ib_qp_init_attr *init_attr)
1364 {
1365 	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1366 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1367 
1368 	attr->qp_state = qp->state;
1369 	attr->cur_qp_state = attr->qp_state;
1370 	attr->path_mtu = qp->path_mtu;
1371 	attr->path_mig_state = qp->s_mig_state;
1372 	attr->qkey = qp->qkey;
1373 	attr->rq_psn = qp->r_psn & rdi->dparms.psn_mask;
1374 	attr->sq_psn = qp->s_next_psn & rdi->dparms.psn_mask;
1375 	attr->dest_qp_num = qp->remote_qpn;
1376 	attr->qp_access_flags = qp->qp_access_flags;
1377 	attr->cap.max_send_wr = qp->s_size - 1 -
1378 		rdi->dparms.reserved_operations;
1379 	attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
1380 	attr->cap.max_send_sge = qp->s_max_sge;
1381 	attr->cap.max_recv_sge = qp->r_rq.max_sge;
1382 	attr->cap.max_inline_data = 0;
1383 	attr->ah_attr = qp->remote_ah_attr;
1384 	attr->alt_ah_attr = qp->alt_ah_attr;
1385 	attr->pkey_index = qp->s_pkey_index;
1386 	attr->alt_pkey_index = qp->s_alt_pkey_index;
1387 	attr->en_sqd_async_notify = 0;
1388 	attr->sq_draining = qp->s_draining;
1389 	attr->max_rd_atomic = qp->s_max_rd_atomic;
1390 	attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
1391 	attr->min_rnr_timer = qp->r_min_rnr_timer;
1392 	attr->port_num = qp->port_num;
1393 	attr->timeout = qp->timeout;
1394 	attr->retry_cnt = qp->s_retry_cnt;
1395 	attr->rnr_retry = qp->s_rnr_retry_cnt;
1396 	attr->alt_port_num = qp->alt_ah_attr.port_num;
1397 	attr->alt_timeout = qp->alt_timeout;
1398 
1399 	init_attr->event_handler = qp->ibqp.event_handler;
1400 	init_attr->qp_context = qp->ibqp.qp_context;
1401 	init_attr->send_cq = qp->ibqp.send_cq;
1402 	init_attr->recv_cq = qp->ibqp.recv_cq;
1403 	init_attr->srq = qp->ibqp.srq;
1404 	init_attr->cap = attr->cap;
1405 	if (qp->s_flags & RVT_S_SIGNAL_REQ_WR)
1406 		init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
1407 	else
1408 		init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
1409 	init_attr->qp_type = qp->ibqp.qp_type;
1410 	init_attr->port_num = qp->port_num;
1411 	return 0;
1412 }
1413 
1414 /**
1415  * rvt_post_receive - post a receive on a QP
1416  * @ibqp: the QP to post the receive on
1417  * @wr: the WR to post
1418  * @bad_wr: the first bad WR is put here
1419  *
1420  * This may be called from interrupt context.
1421  *
1422  * Return: 0 on success otherwise errno
1423  */
1424 int rvt_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1425 		  struct ib_recv_wr **bad_wr)
1426 {
1427 	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1428 	struct rvt_rwq *wq = qp->r_rq.wq;
1429 	unsigned long flags;
1430 	int qp_err_flush = (ib_rvt_state_ops[qp->state] & RVT_FLUSH_RECV) &&
1431 				!qp->ibqp.srq;
1432 
1433 	/* Check that state is OK to post receive. */
1434 	if (!(ib_rvt_state_ops[qp->state] & RVT_POST_RECV_OK) || !wq) {
1435 		*bad_wr = wr;
1436 		return -EINVAL;
1437 	}
1438 
1439 	for (; wr; wr = wr->next) {
1440 		struct rvt_rwqe *wqe;
1441 		u32 next;
1442 		int i;
1443 
1444 		if ((unsigned)wr->num_sge > qp->r_rq.max_sge) {
1445 			*bad_wr = wr;
1446 			return -EINVAL;
1447 		}
1448 
1449 		spin_lock_irqsave(&qp->r_rq.lock, flags);
1450 		next = wq->head + 1;
1451 		if (next >= qp->r_rq.size)
1452 			next = 0;
1453 		if (next == wq->tail) {
1454 			spin_unlock_irqrestore(&qp->r_rq.lock, flags);
1455 			*bad_wr = wr;
1456 			return -ENOMEM;
1457 		}
1458 		if (unlikely(qp_err_flush)) {
1459 			struct ib_wc wc;
1460 
1461 			memset(&wc, 0, sizeof(wc));
1462 			wc.qp = &qp->ibqp;
1463 			wc.opcode = IB_WC_RECV;
1464 			wc.wr_id = wr->wr_id;
1465 			wc.status = IB_WC_WR_FLUSH_ERR;
1466 			rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
1467 		} else {
1468 			wqe = rvt_get_rwqe_ptr(&qp->r_rq, wq->head);
1469 			wqe->wr_id = wr->wr_id;
1470 			wqe->num_sge = wr->num_sge;
1471 			for (i = 0; i < wr->num_sge; i++)
1472 				wqe->sg_list[i] = wr->sg_list[i];
1473 			/*
1474 			 * Make sure queue entry is written
1475 			 * before the head index.
1476 			 */
1477 			smp_wmb();
1478 			wq->head = next;
1479 		}
1480 		spin_unlock_irqrestore(&qp->r_rq.lock, flags);
1481 	}
1482 	return 0;
1483 }
1484 
1485 /**
1486  * rvt_qp_valid_operation - validate post send wr request
1487  * @qp - the qp
1488  * @post-parms - the post send table for the driver
1489  * @wr - the work request
1490  *
1491  * The routine validates the operation based on the
1492  * validation table an returns the length of the operation
1493  * which can extend beyond the ib_send_bw.  Operation
1494  * dependent flags key atomic operation validation.
1495  *
1496  * There is an exception for UD qps that validates the pd and
1497  * overrides the length to include the additional UD specific
1498  * length.
1499  *
1500  * Returns a negative error or the length of the work request
1501  * for building the swqe.
1502  */
1503 static inline int rvt_qp_valid_operation(
1504 	struct rvt_qp *qp,
1505 	const struct rvt_operation_params *post_parms,
1506 	struct ib_send_wr *wr)
1507 {
1508 	int len;
1509 
1510 	if (wr->opcode >= RVT_OPERATION_MAX || !post_parms[wr->opcode].length)
1511 		return -EINVAL;
1512 	if (!(post_parms[wr->opcode].qpt_support & BIT(qp->ibqp.qp_type)))
1513 		return -EINVAL;
1514 	if ((post_parms[wr->opcode].flags & RVT_OPERATION_PRIV) &&
1515 	    ibpd_to_rvtpd(qp->ibqp.pd)->user)
1516 		return -EINVAL;
1517 	if (post_parms[wr->opcode].flags & RVT_OPERATION_ATOMIC_SGE &&
1518 	    (wr->num_sge == 0 ||
1519 	     wr->sg_list[0].length < sizeof(u64) ||
1520 	     wr->sg_list[0].addr & (sizeof(u64) - 1)))
1521 		return -EINVAL;
1522 	if (post_parms[wr->opcode].flags & RVT_OPERATION_ATOMIC &&
1523 	    !qp->s_max_rd_atomic)
1524 		return -EINVAL;
1525 	len = post_parms[wr->opcode].length;
1526 	/* UD specific */
1527 	if (qp->ibqp.qp_type != IB_QPT_UC &&
1528 	    qp->ibqp.qp_type != IB_QPT_RC) {
1529 		if (qp->ibqp.pd != ud_wr(wr)->ah->pd)
1530 			return -EINVAL;
1531 		len = sizeof(struct ib_ud_wr);
1532 	}
1533 	return len;
1534 }
1535 
1536 /**
1537  * rvt_qp_is_avail - determine queue capacity
1538  * @qp - the qp
1539  * @rdi - the rdmavt device
1540  * @reserved_op - is reserved operation
1541  *
1542  * This assumes the s_hlock is held but the s_last
1543  * qp variable is uncontrolled.
1544  *
1545  * For non reserved operations, the qp->s_avail
1546  * may be changed.
1547  *
1548  * The return value is zero or a -ENOMEM.
1549  */
1550 static inline int rvt_qp_is_avail(
1551 	struct rvt_qp *qp,
1552 	struct rvt_dev_info *rdi,
1553 	bool reserved_op)
1554 {
1555 	u32 slast;
1556 	u32 avail;
1557 	u32 reserved_used;
1558 
1559 	/* see rvt_qp_wqe_unreserve() */
1560 	smp_mb__before_atomic();
1561 	reserved_used = atomic_read(&qp->s_reserved_used);
1562 	if (unlikely(reserved_op)) {
1563 		/* see rvt_qp_wqe_unreserve() */
1564 		smp_mb__before_atomic();
1565 		if (reserved_used >= rdi->dparms.reserved_operations)
1566 			return -ENOMEM;
1567 		return 0;
1568 	}
1569 	/* non-reserved operations */
1570 	if (likely(qp->s_avail))
1571 		return 0;
1572 	smp_read_barrier_depends(); /* see rc.c */
1573 	slast = ACCESS_ONCE(qp->s_last);
1574 	if (qp->s_head >= slast)
1575 		avail = qp->s_size - (qp->s_head - slast);
1576 	else
1577 		avail = slast - qp->s_head;
1578 
1579 	/* see rvt_qp_wqe_unreserve() */
1580 	smp_mb__before_atomic();
1581 	reserved_used = atomic_read(&qp->s_reserved_used);
1582 	avail =  avail - 1 -
1583 		(rdi->dparms.reserved_operations - reserved_used);
1584 	/* insure we don't assign a negative s_avail */
1585 	if ((s32)avail <= 0)
1586 		return -ENOMEM;
1587 	qp->s_avail = avail;
1588 	if (WARN_ON(qp->s_avail >
1589 		    (qp->s_size - 1 - rdi->dparms.reserved_operations)))
1590 		rvt_pr_err(rdi,
1591 			   "More avail entries than QP RB size.\nQP: %u, size: %u, avail: %u\nhead: %u, tail: %u, cur: %u, acked: %u, last: %u",
1592 			   qp->ibqp.qp_num, qp->s_size, qp->s_avail,
1593 			   qp->s_head, qp->s_tail, qp->s_cur,
1594 			   qp->s_acked, qp->s_last);
1595 	return 0;
1596 }
1597 
1598 /**
1599  * rvt_post_one_wr - post one RC, UC, or UD send work request
1600  * @qp: the QP to post on
1601  * @wr: the work request to send
1602  */
1603 static int rvt_post_one_wr(struct rvt_qp *qp,
1604 			   struct ib_send_wr *wr,
1605 			   int *call_send)
1606 {
1607 	struct rvt_swqe *wqe;
1608 	u32 next;
1609 	int i;
1610 	int j;
1611 	int acc;
1612 	struct rvt_lkey_table *rkt;
1613 	struct rvt_pd *pd;
1614 	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
1615 	u8 log_pmtu;
1616 	int ret;
1617 	size_t cplen;
1618 	bool reserved_op;
1619 	int local_ops_delayed = 0;
1620 
1621 	BUILD_BUG_ON(IB_QPT_MAX >= (sizeof(u32) * BITS_PER_BYTE));
1622 
1623 	/* IB spec says that num_sge == 0 is OK. */
1624 	if (unlikely(wr->num_sge > qp->s_max_sge))
1625 		return -EINVAL;
1626 
1627 	ret = rvt_qp_valid_operation(qp, rdi->post_parms, wr);
1628 	if (ret < 0)
1629 		return ret;
1630 	cplen = ret;
1631 
1632 	/*
1633 	 * Local operations include fast register and local invalidate.
1634 	 * Fast register needs to be processed immediately because the
1635 	 * registered lkey may be used by following work requests and the
1636 	 * lkey needs to be valid at the time those requests are posted.
1637 	 * Local invalidate can be processed immediately if fencing is
1638 	 * not required and no previous local invalidate ops are pending.
1639 	 * Signaled local operations that have been processed immediately
1640 	 * need to have requests with "completion only" flags set posted
1641 	 * to the send queue in order to generate completions.
1642 	 */
1643 	if ((rdi->post_parms[wr->opcode].flags & RVT_OPERATION_LOCAL)) {
1644 		switch (wr->opcode) {
1645 		case IB_WR_REG_MR:
1646 			ret = rvt_fast_reg_mr(qp,
1647 					      reg_wr(wr)->mr,
1648 					      reg_wr(wr)->key,
1649 					      reg_wr(wr)->access);
1650 			if (ret || !(wr->send_flags & IB_SEND_SIGNALED))
1651 				return ret;
1652 			break;
1653 		case IB_WR_LOCAL_INV:
1654 			if ((wr->send_flags & IB_SEND_FENCE) ||
1655 			    atomic_read(&qp->local_ops_pending)) {
1656 				local_ops_delayed = 1;
1657 			} else {
1658 				ret = rvt_invalidate_rkey(
1659 					qp, wr->ex.invalidate_rkey);
1660 				if (ret || !(wr->send_flags & IB_SEND_SIGNALED))
1661 					return ret;
1662 			}
1663 			break;
1664 		default:
1665 			return -EINVAL;
1666 		}
1667 	}
1668 
1669 	reserved_op = rdi->post_parms[wr->opcode].flags &
1670 			RVT_OPERATION_USE_RESERVE;
1671 	/* check for avail */
1672 	ret = rvt_qp_is_avail(qp, rdi, reserved_op);
1673 	if (ret)
1674 		return ret;
1675 	next = qp->s_head + 1;
1676 	if (next >= qp->s_size)
1677 		next = 0;
1678 
1679 	rkt = &rdi->lkey_table;
1680 	pd = ibpd_to_rvtpd(qp->ibqp.pd);
1681 	wqe = rvt_get_swqe_ptr(qp, qp->s_head);
1682 
1683 	/* cplen has length from above */
1684 	memcpy(&wqe->wr, wr, cplen);
1685 
1686 	wqe->length = 0;
1687 	j = 0;
1688 	if (wr->num_sge) {
1689 		acc = wr->opcode >= IB_WR_RDMA_READ ?
1690 			IB_ACCESS_LOCAL_WRITE : 0;
1691 		for (i = 0; i < wr->num_sge; i++) {
1692 			u32 length = wr->sg_list[i].length;
1693 			int ok;
1694 
1695 			if (length == 0)
1696 				continue;
1697 			ok = rvt_lkey_ok(rkt, pd, &wqe->sg_list[j],
1698 					 &wr->sg_list[i], acc);
1699 			if (!ok) {
1700 				ret = -EINVAL;
1701 				goto bail_inval_free;
1702 			}
1703 			wqe->length += length;
1704 			j++;
1705 		}
1706 		wqe->wr.num_sge = j;
1707 	}
1708 
1709 	/* general part of wqe valid - allow for driver checks */
1710 	if (rdi->driver_f.check_send_wqe) {
1711 		ret = rdi->driver_f.check_send_wqe(qp, wqe);
1712 		if (ret < 0)
1713 			goto bail_inval_free;
1714 		if (ret)
1715 			*call_send = ret;
1716 	}
1717 
1718 	log_pmtu = qp->log_pmtu;
1719 	if (qp->ibqp.qp_type != IB_QPT_UC &&
1720 	    qp->ibqp.qp_type != IB_QPT_RC) {
1721 		struct rvt_ah *ah = ibah_to_rvtah(wqe->ud_wr.ah);
1722 
1723 		log_pmtu = ah->log_pmtu;
1724 		atomic_inc(&ibah_to_rvtah(ud_wr(wr)->ah)->refcount);
1725 	}
1726 
1727 	if (rdi->post_parms[wr->opcode].flags & RVT_OPERATION_LOCAL) {
1728 		if (local_ops_delayed)
1729 			atomic_inc(&qp->local_ops_pending);
1730 		else
1731 			wqe->wr.send_flags |= RVT_SEND_COMPLETION_ONLY;
1732 		wqe->ssn = 0;
1733 		wqe->psn = 0;
1734 		wqe->lpsn = 0;
1735 	} else {
1736 		wqe->ssn = qp->s_ssn++;
1737 		wqe->psn = qp->s_next_psn;
1738 		wqe->lpsn = wqe->psn +
1739 				(wqe->length ?
1740 					((wqe->length - 1) >> log_pmtu) :
1741 					0);
1742 		qp->s_next_psn = wqe->lpsn + 1;
1743 	}
1744 	trace_rvt_post_one_wr(qp, wqe);
1745 	if (unlikely(reserved_op))
1746 		rvt_qp_wqe_reserve(qp, wqe);
1747 	else
1748 		qp->s_avail--;
1749 	smp_wmb(); /* see request builders */
1750 	qp->s_head = next;
1751 
1752 	return 0;
1753 
1754 bail_inval_free:
1755 	/* release mr holds */
1756 	while (j) {
1757 		struct rvt_sge *sge = &wqe->sg_list[--j];
1758 
1759 		rvt_put_mr(sge->mr);
1760 	}
1761 	return ret;
1762 }
1763 
1764 /**
1765  * rvt_post_send - post a send on a QP
1766  * @ibqp: the QP to post the send on
1767  * @wr: the list of work requests to post
1768  * @bad_wr: the first bad WR is put here
1769  *
1770  * This may be called from interrupt context.
1771  *
1772  * Return: 0 on success else errno
1773  */
1774 int rvt_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1775 		  struct ib_send_wr **bad_wr)
1776 {
1777 	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1778 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1779 	unsigned long flags = 0;
1780 	int call_send;
1781 	unsigned nreq = 0;
1782 	int err = 0;
1783 
1784 	spin_lock_irqsave(&qp->s_hlock, flags);
1785 
1786 	/*
1787 	 * Ensure QP state is such that we can send. If not bail out early,
1788 	 * there is no need to do this every time we post a send.
1789 	 */
1790 	if (unlikely(!(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))) {
1791 		spin_unlock_irqrestore(&qp->s_hlock, flags);
1792 		return -EINVAL;
1793 	}
1794 
1795 	/*
1796 	 * If the send queue is empty, and we only have a single WR then just go
1797 	 * ahead and kick the send engine into gear. Otherwise we will always
1798 	 * just schedule the send to happen later.
1799 	 */
1800 	call_send = qp->s_head == ACCESS_ONCE(qp->s_last) && !wr->next;
1801 
1802 	for (; wr; wr = wr->next) {
1803 		err = rvt_post_one_wr(qp, wr, &call_send);
1804 		if (unlikely(err)) {
1805 			*bad_wr = wr;
1806 			goto bail;
1807 		}
1808 		nreq++;
1809 	}
1810 bail:
1811 	spin_unlock_irqrestore(&qp->s_hlock, flags);
1812 	if (nreq) {
1813 		if (call_send)
1814 			rdi->driver_f.do_send(qp);
1815 		else
1816 			rdi->driver_f.schedule_send_no_lock(qp);
1817 	}
1818 	return err;
1819 }
1820 
1821 /**
1822  * rvt_post_srq_receive - post a receive on a shared receive queue
1823  * @ibsrq: the SRQ to post the receive on
1824  * @wr: the list of work requests to post
1825  * @bad_wr: A pointer to the first WR to cause a problem is put here
1826  *
1827  * This may be called from interrupt context.
1828  *
1829  * Return: 0 on success else errno
1830  */
1831 int rvt_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
1832 		      struct ib_recv_wr **bad_wr)
1833 {
1834 	struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq);
1835 	struct rvt_rwq *wq;
1836 	unsigned long flags;
1837 
1838 	for (; wr; wr = wr->next) {
1839 		struct rvt_rwqe *wqe;
1840 		u32 next;
1841 		int i;
1842 
1843 		if ((unsigned)wr->num_sge > srq->rq.max_sge) {
1844 			*bad_wr = wr;
1845 			return -EINVAL;
1846 		}
1847 
1848 		spin_lock_irqsave(&srq->rq.lock, flags);
1849 		wq = srq->rq.wq;
1850 		next = wq->head + 1;
1851 		if (next >= srq->rq.size)
1852 			next = 0;
1853 		if (next == wq->tail) {
1854 			spin_unlock_irqrestore(&srq->rq.lock, flags);
1855 			*bad_wr = wr;
1856 			return -ENOMEM;
1857 		}
1858 
1859 		wqe = rvt_get_rwqe_ptr(&srq->rq, wq->head);
1860 		wqe->wr_id = wr->wr_id;
1861 		wqe->num_sge = wr->num_sge;
1862 		for (i = 0; i < wr->num_sge; i++)
1863 			wqe->sg_list[i] = wr->sg_list[i];
1864 		/* Make sure queue entry is written before the head index. */
1865 		smp_wmb();
1866 		wq->head = next;
1867 		spin_unlock_irqrestore(&srq->rq.lock, flags);
1868 	}
1869 	return 0;
1870 }
1871