xref: /openbmc/linux/drivers/infiniband/sw/rdmavt/qp.c (revision 7587eb18)
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 static void get_map_page(struct rvt_qpn_table *qpt,
80 			 struct rvt_qpn_map *map,
81 			 gfp_t gfp)
82 {
83 	unsigned long page = get_zeroed_page(gfp);
84 
85 	/*
86 	 * Free the page if someone raced with us installing it.
87 	 */
88 
89 	spin_lock(&qpt->lock);
90 	if (map->page)
91 		free_page(page);
92 	else
93 		map->page = (void *)page;
94 	spin_unlock(&qpt->lock);
95 }
96 
97 /**
98  * init_qpn_table - initialize the QP number table for a device
99  * @qpt: the QPN table
100  */
101 static int init_qpn_table(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt)
102 {
103 	u32 offset, i;
104 	struct rvt_qpn_map *map;
105 	int ret = 0;
106 
107 	if (!(rdi->dparms.qpn_res_end >= rdi->dparms.qpn_res_start))
108 		return -EINVAL;
109 
110 	spin_lock_init(&qpt->lock);
111 
112 	qpt->last = rdi->dparms.qpn_start;
113 	qpt->incr = rdi->dparms.qpn_inc << rdi->dparms.qos_shift;
114 
115 	/*
116 	 * Drivers may want some QPs beyond what we need for verbs let them use
117 	 * our qpn table. No need for two. Lets go ahead and mark the bitmaps
118 	 * for those. The reserved range must be *after* the range which verbs
119 	 * will pick from.
120 	 */
121 
122 	/* Figure out number of bit maps needed before reserved range */
123 	qpt->nmaps = rdi->dparms.qpn_res_start / RVT_BITS_PER_PAGE;
124 
125 	/* This should always be zero */
126 	offset = rdi->dparms.qpn_res_start & RVT_BITS_PER_PAGE_MASK;
127 
128 	/* Starting with the first reserved bit map */
129 	map = &qpt->map[qpt->nmaps];
130 
131 	rvt_pr_info(rdi, "Reserving QPNs from 0x%x to 0x%x for non-verbs use\n",
132 		    rdi->dparms.qpn_res_start, rdi->dparms.qpn_res_end);
133 	for (i = rdi->dparms.qpn_res_start; i <= rdi->dparms.qpn_res_end; i++) {
134 		if (!map->page) {
135 			get_map_page(qpt, map, GFP_KERNEL);
136 			if (!map->page) {
137 				ret = -ENOMEM;
138 				break;
139 			}
140 		}
141 		set_bit(offset, map->page);
142 		offset++;
143 		if (offset == RVT_BITS_PER_PAGE) {
144 			/* next page */
145 			qpt->nmaps++;
146 			map++;
147 			offset = 0;
148 		}
149 	}
150 	return ret;
151 }
152 
153 /**
154  * free_qpn_table - free the QP number table for a device
155  * @qpt: the QPN table
156  */
157 static void free_qpn_table(struct rvt_qpn_table *qpt)
158 {
159 	int i;
160 
161 	for (i = 0; i < ARRAY_SIZE(qpt->map); i++)
162 		free_page((unsigned long)qpt->map[i].page);
163 }
164 
165 /**
166  * rvt_driver_qp_init - Init driver qp resources
167  * @rdi: rvt dev strucutre
168  *
169  * Return: 0 on success
170  */
171 int rvt_driver_qp_init(struct rvt_dev_info *rdi)
172 {
173 	int i;
174 	int ret = -ENOMEM;
175 
176 	if (!rdi->dparms.qp_table_size)
177 		return -EINVAL;
178 
179 	/*
180 	 * If driver is not doing any QP allocation then make sure it is
181 	 * providing the necessary QP functions.
182 	 */
183 	if (!rdi->driver_f.free_all_qps ||
184 	    !rdi->driver_f.qp_priv_alloc ||
185 	    !rdi->driver_f.qp_priv_free ||
186 	    !rdi->driver_f.notify_qp_reset)
187 		return -EINVAL;
188 
189 	/* allocate parent object */
190 	rdi->qp_dev = kzalloc_node(sizeof(*rdi->qp_dev), GFP_KERNEL,
191 				   rdi->dparms.node);
192 	if (!rdi->qp_dev)
193 		return -ENOMEM;
194 
195 	/* allocate hash table */
196 	rdi->qp_dev->qp_table_size = rdi->dparms.qp_table_size;
197 	rdi->qp_dev->qp_table_bits = ilog2(rdi->dparms.qp_table_size);
198 	rdi->qp_dev->qp_table =
199 		kmalloc_node(rdi->qp_dev->qp_table_size *
200 			     sizeof(*rdi->qp_dev->qp_table),
201 			     GFP_KERNEL, rdi->dparms.node);
202 	if (!rdi->qp_dev->qp_table)
203 		goto no_qp_table;
204 
205 	for (i = 0; i < rdi->qp_dev->qp_table_size; i++)
206 		RCU_INIT_POINTER(rdi->qp_dev->qp_table[i], NULL);
207 
208 	spin_lock_init(&rdi->qp_dev->qpt_lock);
209 
210 	/* initialize qpn map */
211 	if (init_qpn_table(rdi, &rdi->qp_dev->qpn_table))
212 		goto fail_table;
213 
214 	spin_lock_init(&rdi->n_qps_lock);
215 
216 	return 0;
217 
218 fail_table:
219 	kfree(rdi->qp_dev->qp_table);
220 	free_qpn_table(&rdi->qp_dev->qpn_table);
221 
222 no_qp_table:
223 	kfree(rdi->qp_dev);
224 
225 	return ret;
226 }
227 
228 /**
229  * free_all_qps - check for QPs still in use
230  * @qpt: the QP table to empty
231  *
232  * There should not be any QPs still in use.
233  * Free memory for table.
234  */
235 static unsigned rvt_free_all_qps(struct rvt_dev_info *rdi)
236 {
237 	unsigned long flags;
238 	struct rvt_qp *qp;
239 	unsigned n, qp_inuse = 0;
240 	spinlock_t *ql; /* work around too long line below */
241 
242 	if (rdi->driver_f.free_all_qps)
243 		qp_inuse = rdi->driver_f.free_all_qps(rdi);
244 
245 	qp_inuse += rvt_mcast_tree_empty(rdi);
246 
247 	if (!rdi->qp_dev)
248 		return qp_inuse;
249 
250 	ql = &rdi->qp_dev->qpt_lock;
251 	spin_lock_irqsave(ql, flags);
252 	for (n = 0; n < rdi->qp_dev->qp_table_size; n++) {
253 		qp = rcu_dereference_protected(rdi->qp_dev->qp_table[n],
254 					       lockdep_is_held(ql));
255 		RCU_INIT_POINTER(rdi->qp_dev->qp_table[n], NULL);
256 
257 		for (; qp; qp = rcu_dereference_protected(qp->next,
258 							  lockdep_is_held(ql)))
259 			qp_inuse++;
260 	}
261 	spin_unlock_irqrestore(ql, flags);
262 	synchronize_rcu();
263 	return qp_inuse;
264 }
265 
266 /**
267  * rvt_qp_exit - clean up qps on device exit
268  * @rdi: rvt dev structure
269  *
270  * Check for qp leaks and free resources.
271  */
272 void rvt_qp_exit(struct rvt_dev_info *rdi)
273 {
274 	u32 qps_inuse = rvt_free_all_qps(rdi);
275 
276 	if (qps_inuse)
277 		rvt_pr_err(rdi, "QP memory leak! %u still in use\n",
278 			   qps_inuse);
279 	if (!rdi->qp_dev)
280 		return;
281 
282 	kfree(rdi->qp_dev->qp_table);
283 	free_qpn_table(&rdi->qp_dev->qpn_table);
284 	kfree(rdi->qp_dev);
285 }
286 
287 static inline unsigned mk_qpn(struct rvt_qpn_table *qpt,
288 			      struct rvt_qpn_map *map, unsigned off)
289 {
290 	return (map - qpt->map) * RVT_BITS_PER_PAGE + off;
291 }
292 
293 /**
294  * alloc_qpn - Allocate the next available qpn or zero/one for QP type
295  *	       IB_QPT_SMI/IB_QPT_GSI
296  *@rdi:	rvt device info structure
297  *@qpt: queue pair number table pointer
298  *@port_num: IB port number, 1 based, comes from core
299  *
300  * Return: The queue pair number
301  */
302 static int alloc_qpn(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt,
303 		     enum ib_qp_type type, u8 port_num, gfp_t gfp)
304 {
305 	u32 i, offset, max_scan, qpn;
306 	struct rvt_qpn_map *map;
307 	u32 ret;
308 
309 	if (rdi->driver_f.alloc_qpn)
310 		return rdi->driver_f.alloc_qpn(rdi, qpt, type, port_num, gfp);
311 
312 	if (type == IB_QPT_SMI || type == IB_QPT_GSI) {
313 		unsigned n;
314 
315 		ret = type == IB_QPT_GSI;
316 		n = 1 << (ret + 2 * (port_num - 1));
317 		spin_lock(&qpt->lock);
318 		if (qpt->flags & n)
319 			ret = -EINVAL;
320 		else
321 			qpt->flags |= n;
322 		spin_unlock(&qpt->lock);
323 		goto bail;
324 	}
325 
326 	qpn = qpt->last + qpt->incr;
327 	if (qpn >= RVT_QPN_MAX)
328 		qpn = qpt->incr | ((qpt->last & 1) ^ 1);
329 	/* offset carries bit 0 */
330 	offset = qpn & RVT_BITS_PER_PAGE_MASK;
331 	map = &qpt->map[qpn / RVT_BITS_PER_PAGE];
332 	max_scan = qpt->nmaps - !offset;
333 	for (i = 0;;) {
334 		if (unlikely(!map->page)) {
335 			get_map_page(qpt, map, gfp);
336 			if (unlikely(!map->page))
337 				break;
338 		}
339 		do {
340 			if (!test_and_set_bit(offset, map->page)) {
341 				qpt->last = qpn;
342 				ret = qpn;
343 				goto bail;
344 			}
345 			offset += qpt->incr;
346 			/*
347 			 * This qpn might be bogus if offset >= BITS_PER_PAGE.
348 			 * That is OK.   It gets re-assigned below
349 			 */
350 			qpn = mk_qpn(qpt, map, offset);
351 		} while (offset < RVT_BITS_PER_PAGE && qpn < RVT_QPN_MAX);
352 		/*
353 		 * In order to keep the number of pages allocated to a
354 		 * minimum, we scan the all existing pages before increasing
355 		 * the size of the bitmap table.
356 		 */
357 		if (++i > max_scan) {
358 			if (qpt->nmaps == RVT_QPNMAP_ENTRIES)
359 				break;
360 			map = &qpt->map[qpt->nmaps++];
361 			/* start at incr with current bit 0 */
362 			offset = qpt->incr | (offset & 1);
363 		} else if (map < &qpt->map[qpt->nmaps]) {
364 			++map;
365 			/* start at incr with current bit 0 */
366 			offset = qpt->incr | (offset & 1);
367 		} else {
368 			map = &qpt->map[0];
369 			/* wrap to first map page, invert bit 0 */
370 			offset = qpt->incr | ((offset & 1) ^ 1);
371 		}
372 		/* there can be no bits at shift and below */
373 		WARN_ON(offset & (rdi->dparms.qos_shift - 1));
374 		qpn = mk_qpn(qpt, map, offset);
375 	}
376 
377 	ret = -ENOMEM;
378 
379 bail:
380 	return ret;
381 }
382 
383 static void free_qpn(struct rvt_qpn_table *qpt, u32 qpn)
384 {
385 	struct rvt_qpn_map *map;
386 
387 	map = qpt->map + qpn / RVT_BITS_PER_PAGE;
388 	if (map->page)
389 		clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page);
390 }
391 
392 /**
393  * rvt_clear_mr_refs - Drop help mr refs
394  * @qp: rvt qp data structure
395  * @clr_sends: If shoudl clear send side or not
396  */
397 static void rvt_clear_mr_refs(struct rvt_qp *qp, int clr_sends)
398 {
399 	unsigned n;
400 	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
401 
402 	if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags))
403 		rvt_put_ss(&qp->s_rdma_read_sge);
404 
405 	rvt_put_ss(&qp->r_sge);
406 
407 	if (clr_sends) {
408 		while (qp->s_last != qp->s_head) {
409 			struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_last);
410 			unsigned i;
411 
412 			for (i = 0; i < wqe->wr.num_sge; i++) {
413 				struct rvt_sge *sge = &wqe->sg_list[i];
414 
415 				rvt_put_mr(sge->mr);
416 			}
417 			if (qp->ibqp.qp_type == IB_QPT_UD ||
418 			    qp->ibqp.qp_type == IB_QPT_SMI ||
419 			    qp->ibqp.qp_type == IB_QPT_GSI)
420 				atomic_dec(&ibah_to_rvtah(
421 						wqe->ud_wr.ah)->refcount);
422 			if (++qp->s_last >= qp->s_size)
423 				qp->s_last = 0;
424 			smp_wmb(); /* see qp_set_savail */
425 		}
426 		if (qp->s_rdma_mr) {
427 			rvt_put_mr(qp->s_rdma_mr);
428 			qp->s_rdma_mr = NULL;
429 		}
430 	}
431 
432 	if (qp->ibqp.qp_type != IB_QPT_RC)
433 		return;
434 
435 	for (n = 0; n < rvt_max_atomic(rdi); n++) {
436 		struct rvt_ack_entry *e = &qp->s_ack_queue[n];
437 
438 		if (e->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST &&
439 		    e->rdma_sge.mr) {
440 			rvt_put_mr(e->rdma_sge.mr);
441 			e->rdma_sge.mr = NULL;
442 		}
443 	}
444 }
445 
446 /**
447  * rvt_remove_qp - remove qp form table
448  * @rdi: rvt dev struct
449  * @qp: qp to remove
450  *
451  * Remove the QP from the table so it can't be found asynchronously by
452  * the receive routine.
453  */
454 static void rvt_remove_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp)
455 {
456 	struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1];
457 	u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits);
458 	unsigned long flags;
459 	int removed = 1;
460 
461 	spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags);
462 
463 	if (rcu_dereference_protected(rvp->qp[0],
464 			lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) {
465 		RCU_INIT_POINTER(rvp->qp[0], NULL);
466 	} else if (rcu_dereference_protected(rvp->qp[1],
467 			lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) {
468 		RCU_INIT_POINTER(rvp->qp[1], NULL);
469 	} else {
470 		struct rvt_qp *q;
471 		struct rvt_qp __rcu **qpp;
472 
473 		removed = 0;
474 		qpp = &rdi->qp_dev->qp_table[n];
475 		for (; (q = rcu_dereference_protected(*qpp,
476 			lockdep_is_held(&rdi->qp_dev->qpt_lock))) != NULL;
477 			qpp = &q->next) {
478 			if (q == qp) {
479 				RCU_INIT_POINTER(*qpp,
480 				     rcu_dereference_protected(qp->next,
481 				     lockdep_is_held(&rdi->qp_dev->qpt_lock)));
482 				removed = 1;
483 				trace_rvt_qpremove(qp, n);
484 				break;
485 			}
486 		}
487 	}
488 
489 	spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags);
490 	if (removed) {
491 		synchronize_rcu();
492 		if (atomic_dec_and_test(&qp->refcount))
493 			wake_up(&qp->wait);
494 	}
495 }
496 
497 /**
498  * reset_qp - initialize the QP state to the reset state
499  * @qp: the QP to reset
500  * @type: the QP type
501  * r and s lock are required to be held by the caller
502  */
503 static void rvt_reset_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp,
504 		  enum ib_qp_type type)
505 	__releases(&qp->s_lock)
506 	__releases(&qp->s_hlock)
507 	__releases(&qp->r_lock)
508 	__acquires(&qp->r_lock)
509 	__acquires(&qp->s_hlock)
510 	__acquires(&qp->s_lock)
511 {
512 	if (qp->state != IB_QPS_RESET) {
513 		qp->state = IB_QPS_RESET;
514 
515 		/* Let drivers flush their waitlist */
516 		rdi->driver_f.flush_qp_waiters(qp);
517 		qp->s_flags &= ~(RVT_S_TIMER | RVT_S_ANY_WAIT);
518 		spin_unlock(&qp->s_lock);
519 		spin_unlock(&qp->s_hlock);
520 		spin_unlock_irq(&qp->r_lock);
521 
522 		/* Stop the send queue and the retry timer */
523 		rdi->driver_f.stop_send_queue(qp);
524 
525 		/* Wait for things to stop */
526 		rdi->driver_f.quiesce_qp(qp);
527 
528 		/* take qp out the hash and wait for it to be unused */
529 		rvt_remove_qp(rdi, qp);
530 		wait_event(qp->wait, !atomic_read(&qp->refcount));
531 
532 		/* grab the lock b/c it was locked at call time */
533 		spin_lock_irq(&qp->r_lock);
534 		spin_lock(&qp->s_hlock);
535 		spin_lock(&qp->s_lock);
536 
537 		rvt_clear_mr_refs(qp, 1);
538 	}
539 
540 	/*
541 	 * Let the driver do any tear down it needs to for a qp
542 	 * that has been reset
543 	 */
544 	rdi->driver_f.notify_qp_reset(qp);
545 
546 	qp->remote_qpn = 0;
547 	qp->qkey = 0;
548 	qp->qp_access_flags = 0;
549 	qp->s_flags &= RVT_S_SIGNAL_REQ_WR;
550 	qp->s_hdrwords = 0;
551 	qp->s_wqe = NULL;
552 	qp->s_draining = 0;
553 	qp->s_next_psn = 0;
554 	qp->s_last_psn = 0;
555 	qp->s_sending_psn = 0;
556 	qp->s_sending_hpsn = 0;
557 	qp->s_psn = 0;
558 	qp->r_psn = 0;
559 	qp->r_msn = 0;
560 	if (type == IB_QPT_RC) {
561 		qp->s_state = IB_OPCODE_RC_SEND_LAST;
562 		qp->r_state = IB_OPCODE_RC_SEND_LAST;
563 	} else {
564 		qp->s_state = IB_OPCODE_UC_SEND_LAST;
565 		qp->r_state = IB_OPCODE_UC_SEND_LAST;
566 	}
567 	qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
568 	qp->r_nak_state = 0;
569 	qp->r_aflags = 0;
570 	qp->r_flags = 0;
571 	qp->s_head = 0;
572 	qp->s_tail = 0;
573 	qp->s_cur = 0;
574 	qp->s_acked = 0;
575 	qp->s_last = 0;
576 	qp->s_ssn = 1;
577 	qp->s_lsn = 0;
578 	qp->s_mig_state = IB_MIG_MIGRATED;
579 	if (qp->s_ack_queue)
580 		memset(
581 			qp->s_ack_queue,
582 			0,
583 			rvt_max_atomic(rdi) *
584 				sizeof(*qp->s_ack_queue));
585 	qp->r_head_ack_queue = 0;
586 	qp->s_tail_ack_queue = 0;
587 	qp->s_num_rd_atomic = 0;
588 	if (qp->r_rq.wq) {
589 		qp->r_rq.wq->head = 0;
590 		qp->r_rq.wq->tail = 0;
591 	}
592 	qp->r_sge.num_sge = 0;
593 }
594 
595 /**
596  * rvt_create_qp - create a queue pair for a device
597  * @ibpd: the protection domain who's device we create the queue pair for
598  * @init_attr: the attributes of the queue pair
599  * @udata: user data for libibverbs.so
600  *
601  * Queue pair creation is mostly an rvt issue. However, drivers have their own
602  * unique idea of what queue pair numbers mean. For instance there is a reserved
603  * range for PSM.
604  *
605  * Return: the queue pair on success, otherwise returns an errno.
606  *
607  * Called by the ib_create_qp() core verbs function.
608  */
609 struct ib_qp *rvt_create_qp(struct ib_pd *ibpd,
610 			    struct ib_qp_init_attr *init_attr,
611 			    struct ib_udata *udata)
612 {
613 	struct rvt_qp *qp;
614 	int err;
615 	struct rvt_swqe *swq = NULL;
616 	size_t sz;
617 	size_t sg_list_sz;
618 	struct ib_qp *ret = ERR_PTR(-ENOMEM);
619 	struct rvt_dev_info *rdi = ib_to_rvt(ibpd->device);
620 	void *priv = NULL;
621 	gfp_t gfp;
622 
623 	if (!rdi)
624 		return ERR_PTR(-EINVAL);
625 
626 	if (init_attr->cap.max_send_sge > rdi->dparms.props.max_sge ||
627 	    init_attr->cap.max_send_wr > rdi->dparms.props.max_qp_wr ||
628 	    init_attr->create_flags & ~(IB_QP_CREATE_USE_GFP_NOIO))
629 		return ERR_PTR(-EINVAL);
630 
631 	/* GFP_NOIO is applicable to RC QP's only */
632 
633 	if (init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO &&
634 	    init_attr->qp_type != IB_QPT_RC)
635 		return ERR_PTR(-EINVAL);
636 
637 	gfp = init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO ?
638 						GFP_NOIO : GFP_KERNEL;
639 
640 	/* Check receive queue parameters if no SRQ is specified. */
641 	if (!init_attr->srq) {
642 		if (init_attr->cap.max_recv_sge > rdi->dparms.props.max_sge ||
643 		    init_attr->cap.max_recv_wr > rdi->dparms.props.max_qp_wr)
644 			return ERR_PTR(-EINVAL);
645 
646 		if (init_attr->cap.max_send_sge +
647 		    init_attr->cap.max_send_wr +
648 		    init_attr->cap.max_recv_sge +
649 		    init_attr->cap.max_recv_wr == 0)
650 			return ERR_PTR(-EINVAL);
651 	}
652 
653 	switch (init_attr->qp_type) {
654 	case IB_QPT_SMI:
655 	case IB_QPT_GSI:
656 		if (init_attr->port_num == 0 ||
657 		    init_attr->port_num > ibpd->device->phys_port_cnt)
658 			return ERR_PTR(-EINVAL);
659 	case IB_QPT_UC:
660 	case IB_QPT_RC:
661 	case IB_QPT_UD:
662 		sz = sizeof(struct rvt_sge) *
663 			init_attr->cap.max_send_sge +
664 			sizeof(struct rvt_swqe);
665 		if (gfp == GFP_NOIO)
666 			swq = __vmalloc(
667 				(init_attr->cap.max_send_wr + 1) * sz,
668 				gfp | __GFP_ZERO, PAGE_KERNEL);
669 		else
670 			swq = vzalloc_node(
671 				(init_attr->cap.max_send_wr + 1) * sz,
672 				rdi->dparms.node);
673 		if (!swq)
674 			return ERR_PTR(-ENOMEM);
675 
676 		sz = sizeof(*qp);
677 		sg_list_sz = 0;
678 		if (init_attr->srq) {
679 			struct rvt_srq *srq = ibsrq_to_rvtsrq(init_attr->srq);
680 
681 			if (srq->rq.max_sge > 1)
682 				sg_list_sz = sizeof(*qp->r_sg_list) *
683 					(srq->rq.max_sge - 1);
684 		} else if (init_attr->cap.max_recv_sge > 1)
685 			sg_list_sz = sizeof(*qp->r_sg_list) *
686 				(init_attr->cap.max_recv_sge - 1);
687 		qp = kzalloc_node(sz + sg_list_sz, gfp, rdi->dparms.node);
688 		if (!qp)
689 			goto bail_swq;
690 
691 		RCU_INIT_POINTER(qp->next, NULL);
692 		if (init_attr->qp_type == IB_QPT_RC) {
693 			qp->s_ack_queue =
694 				kzalloc_node(
695 					sizeof(*qp->s_ack_queue) *
696 					 rvt_max_atomic(rdi),
697 					gfp,
698 					rdi->dparms.node);
699 			if (!qp->s_ack_queue)
700 				goto bail_qp;
701 		}
702 
703 		/*
704 		 * Driver needs to set up it's private QP structure and do any
705 		 * initialization that is needed.
706 		 */
707 		priv = rdi->driver_f.qp_priv_alloc(rdi, qp, gfp);
708 		if (!priv)
709 			goto bail_qp;
710 		qp->priv = priv;
711 		qp->timeout_jiffies =
712 			usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
713 				1000UL);
714 		if (init_attr->srq) {
715 			sz = 0;
716 		} else {
717 			qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
718 			qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
719 			sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
720 				sizeof(struct rvt_rwqe);
721 			if (udata)
722 				qp->r_rq.wq = vmalloc_user(
723 						sizeof(struct rvt_rwq) +
724 						qp->r_rq.size * sz);
725 			else if (gfp == GFP_NOIO)
726 				qp->r_rq.wq = __vmalloc(
727 						sizeof(struct rvt_rwq) +
728 						qp->r_rq.size * sz,
729 						gfp | __GFP_ZERO, PAGE_KERNEL);
730 			else
731 				qp->r_rq.wq = vzalloc_node(
732 						sizeof(struct rvt_rwq) +
733 						qp->r_rq.size * sz,
734 						rdi->dparms.node);
735 			if (!qp->r_rq.wq)
736 				goto bail_driver_priv;
737 		}
738 
739 		/*
740 		 * ib_create_qp() will initialize qp->ibqp
741 		 * except for qp->ibqp.qp_num.
742 		 */
743 		spin_lock_init(&qp->r_lock);
744 		spin_lock_init(&qp->s_hlock);
745 		spin_lock_init(&qp->s_lock);
746 		spin_lock_init(&qp->r_rq.lock);
747 		atomic_set(&qp->refcount, 0);
748 		init_waitqueue_head(&qp->wait);
749 		init_timer(&qp->s_timer);
750 		qp->s_timer.data = (unsigned long)qp;
751 		INIT_LIST_HEAD(&qp->rspwait);
752 		qp->state = IB_QPS_RESET;
753 		qp->s_wq = swq;
754 		qp->s_size = init_attr->cap.max_send_wr + 1;
755 		qp->s_avail = init_attr->cap.max_send_wr;
756 		qp->s_max_sge = init_attr->cap.max_send_sge;
757 		if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
758 			qp->s_flags = RVT_S_SIGNAL_REQ_WR;
759 
760 		err = alloc_qpn(rdi, &rdi->qp_dev->qpn_table,
761 				init_attr->qp_type,
762 				init_attr->port_num, gfp);
763 		if (err < 0) {
764 			ret = ERR_PTR(err);
765 			goto bail_rq_wq;
766 		}
767 		qp->ibqp.qp_num = err;
768 		qp->port_num = init_attr->port_num;
769 		rvt_reset_qp(rdi, qp, init_attr->qp_type);
770 		break;
771 
772 	default:
773 		/* Don't support raw QPs */
774 		return ERR_PTR(-EINVAL);
775 	}
776 
777 	init_attr->cap.max_inline_data = 0;
778 
779 	/*
780 	 * Return the address of the RWQ as the offset to mmap.
781 	 * See rvt_mmap() for details.
782 	 */
783 	if (udata && udata->outlen >= sizeof(__u64)) {
784 		if (!qp->r_rq.wq) {
785 			__u64 offset = 0;
786 
787 			err = ib_copy_to_udata(udata, &offset,
788 					       sizeof(offset));
789 			if (err) {
790 				ret = ERR_PTR(err);
791 				goto bail_qpn;
792 			}
793 		} else {
794 			u32 s = sizeof(struct rvt_rwq) + qp->r_rq.size * sz;
795 
796 			qp->ip = rvt_create_mmap_info(rdi, s,
797 						      ibpd->uobject->context,
798 						      qp->r_rq.wq);
799 			if (!qp->ip) {
800 				ret = ERR_PTR(-ENOMEM);
801 				goto bail_qpn;
802 			}
803 
804 			err = ib_copy_to_udata(udata, &qp->ip->offset,
805 					       sizeof(qp->ip->offset));
806 			if (err) {
807 				ret = ERR_PTR(err);
808 				goto bail_ip;
809 			}
810 		}
811 		qp->pid = current->pid;
812 	}
813 
814 	spin_lock(&rdi->n_qps_lock);
815 	if (rdi->n_qps_allocated == rdi->dparms.props.max_qp) {
816 		spin_unlock(&rdi->n_qps_lock);
817 		ret = ERR_PTR(-ENOMEM);
818 		goto bail_ip;
819 	}
820 
821 	rdi->n_qps_allocated++;
822 	/*
823 	 * Maintain a busy_jiffies variable that will be added to the timeout
824 	 * period in mod_retry_timer and add_retry_timer. This busy jiffies
825 	 * is scaled by the number of rc qps created for the device to reduce
826 	 * the number of timeouts occurring when there is a large number of
827 	 * qps. busy_jiffies is incremented every rc qp scaling interval.
828 	 * The scaling interval is selected based on extensive performance
829 	 * evaluation of targeted workloads.
830 	 */
831 	if (init_attr->qp_type == IB_QPT_RC) {
832 		rdi->n_rc_qps++;
833 		rdi->busy_jiffies = rdi->n_rc_qps / RC_QP_SCALING_INTERVAL;
834 	}
835 	spin_unlock(&rdi->n_qps_lock);
836 
837 	if (qp->ip) {
838 		spin_lock_irq(&rdi->pending_lock);
839 		list_add(&qp->ip->pending_mmaps, &rdi->pending_mmaps);
840 		spin_unlock_irq(&rdi->pending_lock);
841 	}
842 
843 	ret = &qp->ibqp;
844 
845 	/*
846 	 * We have our QP and its good, now keep track of what types of opcodes
847 	 * can be processed on this QP. We do this by keeping track of what the
848 	 * 3 high order bits of the opcode are.
849 	 */
850 	switch (init_attr->qp_type) {
851 	case IB_QPT_SMI:
852 	case IB_QPT_GSI:
853 	case IB_QPT_UD:
854 		qp->allowed_ops = IB_OPCODE_UD;
855 		break;
856 	case IB_QPT_RC:
857 		qp->allowed_ops = IB_OPCODE_RC;
858 		break;
859 	case IB_QPT_UC:
860 		qp->allowed_ops = IB_OPCODE_UC;
861 		break;
862 	default:
863 		ret = ERR_PTR(-EINVAL);
864 		goto bail_ip;
865 	}
866 
867 	return ret;
868 
869 bail_ip:
870 	kref_put(&qp->ip->ref, rvt_release_mmap_info);
871 
872 bail_qpn:
873 	free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num);
874 
875 bail_rq_wq:
876 	vfree(qp->r_rq.wq);
877 
878 bail_driver_priv:
879 	rdi->driver_f.qp_priv_free(rdi, qp);
880 
881 bail_qp:
882 	kfree(qp->s_ack_queue);
883 	kfree(qp);
884 
885 bail_swq:
886 	vfree(swq);
887 
888 	return ret;
889 }
890 
891 /**
892  * rvt_error_qp - put a QP into the error state
893  * @qp: the QP to put into the error state
894  * @err: the receive completion error to signal if a RWQE is active
895  *
896  * Flushes both send and receive work queues.
897  *
898  * Return: true if last WQE event should be generated.
899  * The QP r_lock and s_lock should be held and interrupts disabled.
900  * If we are already in error state, just return.
901  */
902 int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err)
903 {
904 	struct ib_wc wc;
905 	int ret = 0;
906 	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
907 
908 	if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET)
909 		goto bail;
910 
911 	qp->state = IB_QPS_ERR;
912 
913 	if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) {
914 		qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR);
915 		del_timer(&qp->s_timer);
916 	}
917 
918 	if (qp->s_flags & RVT_S_ANY_WAIT_SEND)
919 		qp->s_flags &= ~RVT_S_ANY_WAIT_SEND;
920 
921 	rdi->driver_f.notify_error_qp(qp);
922 
923 	/* Schedule the sending tasklet to drain the send work queue. */
924 	if (ACCESS_ONCE(qp->s_last) != qp->s_head)
925 		rdi->driver_f.schedule_send(qp);
926 
927 	rvt_clear_mr_refs(qp, 0);
928 
929 	memset(&wc, 0, sizeof(wc));
930 	wc.qp = &qp->ibqp;
931 	wc.opcode = IB_WC_RECV;
932 
933 	if (test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) {
934 		wc.wr_id = qp->r_wr_id;
935 		wc.status = err;
936 		rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
937 	}
938 	wc.status = IB_WC_WR_FLUSH_ERR;
939 
940 	if (qp->r_rq.wq) {
941 		struct rvt_rwq *wq;
942 		u32 head;
943 		u32 tail;
944 
945 		spin_lock(&qp->r_rq.lock);
946 
947 		/* sanity check pointers before trusting them */
948 		wq = qp->r_rq.wq;
949 		head = wq->head;
950 		if (head >= qp->r_rq.size)
951 			head = 0;
952 		tail = wq->tail;
953 		if (tail >= qp->r_rq.size)
954 			tail = 0;
955 		while (tail != head) {
956 			wc.wr_id = rvt_get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
957 			if (++tail >= qp->r_rq.size)
958 				tail = 0;
959 			rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
960 		}
961 		wq->tail = tail;
962 
963 		spin_unlock(&qp->r_rq.lock);
964 	} else if (qp->ibqp.event_handler) {
965 		ret = 1;
966 	}
967 
968 bail:
969 	return ret;
970 }
971 EXPORT_SYMBOL(rvt_error_qp);
972 
973 /*
974  * Put the QP into the hash table.
975  * The hash table holds a reference to the QP.
976  */
977 static void rvt_insert_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp)
978 {
979 	struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1];
980 	unsigned long flags;
981 
982 	atomic_inc(&qp->refcount);
983 	spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags);
984 
985 	if (qp->ibqp.qp_num <= 1) {
986 		rcu_assign_pointer(rvp->qp[qp->ibqp.qp_num], qp);
987 	} else {
988 		u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits);
989 
990 		qp->next = rdi->qp_dev->qp_table[n];
991 		rcu_assign_pointer(rdi->qp_dev->qp_table[n], qp);
992 		trace_rvt_qpinsert(qp, n);
993 	}
994 
995 	spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags);
996 }
997 
998 /**
999  * qib_modify_qp - modify the attributes of a queue pair
1000  * @ibqp: the queue pair who's attributes we're modifying
1001  * @attr: the new attributes
1002  * @attr_mask: the mask of attributes to modify
1003  * @udata: user data for libibverbs.so
1004  *
1005  * Return: 0 on success, otherwise returns an errno.
1006  */
1007 int rvt_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1008 		  int attr_mask, struct ib_udata *udata)
1009 {
1010 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1011 	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1012 	enum ib_qp_state cur_state, new_state;
1013 	struct ib_event ev;
1014 	int lastwqe = 0;
1015 	int mig = 0;
1016 	int pmtu = 0; /* for gcc warning only */
1017 	enum rdma_link_layer link;
1018 
1019 	link = rdma_port_get_link_layer(ibqp->device, qp->port_num);
1020 
1021 	spin_lock_irq(&qp->r_lock);
1022 	spin_lock(&qp->s_hlock);
1023 	spin_lock(&qp->s_lock);
1024 
1025 	cur_state = attr_mask & IB_QP_CUR_STATE ?
1026 		attr->cur_qp_state : qp->state;
1027 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1028 
1029 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1030 				attr_mask, link))
1031 		goto inval;
1032 
1033 	if (rdi->driver_f.check_modify_qp &&
1034 	    rdi->driver_f.check_modify_qp(qp, attr, attr_mask, udata))
1035 		goto inval;
1036 
1037 	if (attr_mask & IB_QP_AV) {
1038 		if (attr->ah_attr.dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE))
1039 			goto inval;
1040 		if (rvt_check_ah(qp->ibqp.device, &attr->ah_attr))
1041 			goto inval;
1042 	}
1043 
1044 	if (attr_mask & IB_QP_ALT_PATH) {
1045 		if (attr->alt_ah_attr.dlid >=
1046 		    be16_to_cpu(IB_MULTICAST_LID_BASE))
1047 			goto inval;
1048 		if (rvt_check_ah(qp->ibqp.device, &attr->alt_ah_attr))
1049 			goto inval;
1050 		if (attr->alt_pkey_index >= rvt_get_npkeys(rdi))
1051 			goto inval;
1052 	}
1053 
1054 	if (attr_mask & IB_QP_PKEY_INDEX)
1055 		if (attr->pkey_index >= rvt_get_npkeys(rdi))
1056 			goto inval;
1057 
1058 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
1059 		if (attr->min_rnr_timer > 31)
1060 			goto inval;
1061 
1062 	if (attr_mask & IB_QP_PORT)
1063 		if (qp->ibqp.qp_type == IB_QPT_SMI ||
1064 		    qp->ibqp.qp_type == IB_QPT_GSI ||
1065 		    attr->port_num == 0 ||
1066 		    attr->port_num > ibqp->device->phys_port_cnt)
1067 			goto inval;
1068 
1069 	if (attr_mask & IB_QP_DEST_QPN)
1070 		if (attr->dest_qp_num > RVT_QPN_MASK)
1071 			goto inval;
1072 
1073 	if (attr_mask & IB_QP_RETRY_CNT)
1074 		if (attr->retry_cnt > 7)
1075 			goto inval;
1076 
1077 	if (attr_mask & IB_QP_RNR_RETRY)
1078 		if (attr->rnr_retry > 7)
1079 			goto inval;
1080 
1081 	/*
1082 	 * Don't allow invalid path_mtu values.  OK to set greater
1083 	 * than the active mtu (or even the max_cap, if we have tuned
1084 	 * that to a small mtu.  We'll set qp->path_mtu
1085 	 * to the lesser of requested attribute mtu and active,
1086 	 * for packetizing messages.
1087 	 * Note that the QP port has to be set in INIT and MTU in RTR.
1088 	 */
1089 	if (attr_mask & IB_QP_PATH_MTU) {
1090 		pmtu = rdi->driver_f.get_pmtu_from_attr(rdi, qp, attr);
1091 		if (pmtu < 0)
1092 			goto inval;
1093 	}
1094 
1095 	if (attr_mask & IB_QP_PATH_MIG_STATE) {
1096 		if (attr->path_mig_state == IB_MIG_REARM) {
1097 			if (qp->s_mig_state == IB_MIG_ARMED)
1098 				goto inval;
1099 			if (new_state != IB_QPS_RTS)
1100 				goto inval;
1101 		} else if (attr->path_mig_state == IB_MIG_MIGRATED) {
1102 			if (qp->s_mig_state == IB_MIG_REARM)
1103 				goto inval;
1104 			if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD)
1105 				goto inval;
1106 			if (qp->s_mig_state == IB_MIG_ARMED)
1107 				mig = 1;
1108 		} else {
1109 			goto inval;
1110 		}
1111 	}
1112 
1113 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1114 		if (attr->max_dest_rd_atomic > rdi->dparms.max_rdma_atomic)
1115 			goto inval;
1116 
1117 	switch (new_state) {
1118 	case IB_QPS_RESET:
1119 		if (qp->state != IB_QPS_RESET)
1120 			rvt_reset_qp(rdi, qp, ibqp->qp_type);
1121 		break;
1122 
1123 	case IB_QPS_RTR:
1124 		/* Allow event to re-trigger if QP set to RTR more than once */
1125 		qp->r_flags &= ~RVT_R_COMM_EST;
1126 		qp->state = new_state;
1127 		break;
1128 
1129 	case IB_QPS_SQD:
1130 		qp->s_draining = qp->s_last != qp->s_cur;
1131 		qp->state = new_state;
1132 		break;
1133 
1134 	case IB_QPS_SQE:
1135 		if (qp->ibqp.qp_type == IB_QPT_RC)
1136 			goto inval;
1137 		qp->state = new_state;
1138 		break;
1139 
1140 	case IB_QPS_ERR:
1141 		lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1142 		break;
1143 
1144 	default:
1145 		qp->state = new_state;
1146 		break;
1147 	}
1148 
1149 	if (attr_mask & IB_QP_PKEY_INDEX)
1150 		qp->s_pkey_index = attr->pkey_index;
1151 
1152 	if (attr_mask & IB_QP_PORT)
1153 		qp->port_num = attr->port_num;
1154 
1155 	if (attr_mask & IB_QP_DEST_QPN)
1156 		qp->remote_qpn = attr->dest_qp_num;
1157 
1158 	if (attr_mask & IB_QP_SQ_PSN) {
1159 		qp->s_next_psn = attr->sq_psn & rdi->dparms.psn_modify_mask;
1160 		qp->s_psn = qp->s_next_psn;
1161 		qp->s_sending_psn = qp->s_next_psn;
1162 		qp->s_last_psn = qp->s_next_psn - 1;
1163 		qp->s_sending_hpsn = qp->s_last_psn;
1164 	}
1165 
1166 	if (attr_mask & IB_QP_RQ_PSN)
1167 		qp->r_psn = attr->rq_psn & rdi->dparms.psn_modify_mask;
1168 
1169 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1170 		qp->qp_access_flags = attr->qp_access_flags;
1171 
1172 	if (attr_mask & IB_QP_AV) {
1173 		qp->remote_ah_attr = attr->ah_attr;
1174 		qp->s_srate = attr->ah_attr.static_rate;
1175 		qp->srate_mbps = ib_rate_to_mbps(qp->s_srate);
1176 	}
1177 
1178 	if (attr_mask & IB_QP_ALT_PATH) {
1179 		qp->alt_ah_attr = attr->alt_ah_attr;
1180 		qp->s_alt_pkey_index = attr->alt_pkey_index;
1181 	}
1182 
1183 	if (attr_mask & IB_QP_PATH_MIG_STATE) {
1184 		qp->s_mig_state = attr->path_mig_state;
1185 		if (mig) {
1186 			qp->remote_ah_attr = qp->alt_ah_attr;
1187 			qp->port_num = qp->alt_ah_attr.port_num;
1188 			qp->s_pkey_index = qp->s_alt_pkey_index;
1189 		}
1190 	}
1191 
1192 	if (attr_mask & IB_QP_PATH_MTU) {
1193 		qp->pmtu = rdi->driver_f.mtu_from_qp(rdi, qp, pmtu);
1194 		qp->path_mtu = rdi->driver_f.mtu_to_path_mtu(qp->pmtu);
1195 		qp->log_pmtu = ilog2(qp->pmtu);
1196 	}
1197 
1198 	if (attr_mask & IB_QP_RETRY_CNT) {
1199 		qp->s_retry_cnt = attr->retry_cnt;
1200 		qp->s_retry = attr->retry_cnt;
1201 	}
1202 
1203 	if (attr_mask & IB_QP_RNR_RETRY) {
1204 		qp->s_rnr_retry_cnt = attr->rnr_retry;
1205 		qp->s_rnr_retry = attr->rnr_retry;
1206 	}
1207 
1208 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
1209 		qp->r_min_rnr_timer = attr->min_rnr_timer;
1210 
1211 	if (attr_mask & IB_QP_TIMEOUT) {
1212 		qp->timeout = attr->timeout;
1213 		qp->timeout_jiffies =
1214 			usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
1215 				1000UL);
1216 	}
1217 
1218 	if (attr_mask & IB_QP_QKEY)
1219 		qp->qkey = attr->qkey;
1220 
1221 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1222 		qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
1223 
1224 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
1225 		qp->s_max_rd_atomic = attr->max_rd_atomic;
1226 
1227 	if (rdi->driver_f.modify_qp)
1228 		rdi->driver_f.modify_qp(qp, attr, attr_mask, udata);
1229 
1230 	spin_unlock(&qp->s_lock);
1231 	spin_unlock(&qp->s_hlock);
1232 	spin_unlock_irq(&qp->r_lock);
1233 
1234 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1235 		rvt_insert_qp(rdi, qp);
1236 
1237 	if (lastwqe) {
1238 		ev.device = qp->ibqp.device;
1239 		ev.element.qp = &qp->ibqp;
1240 		ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
1241 		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1242 	}
1243 	if (mig) {
1244 		ev.device = qp->ibqp.device;
1245 		ev.element.qp = &qp->ibqp;
1246 		ev.event = IB_EVENT_PATH_MIG;
1247 		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1248 	}
1249 	return 0;
1250 
1251 inval:
1252 	spin_unlock(&qp->s_lock);
1253 	spin_unlock(&qp->s_hlock);
1254 	spin_unlock_irq(&qp->r_lock);
1255 	return -EINVAL;
1256 }
1257 
1258 /** rvt_free_qpn - Free a qpn from the bit map
1259  * @qpt: QP table
1260  * @qpn: queue pair number to free
1261  */
1262 static void rvt_free_qpn(struct rvt_qpn_table *qpt, u32 qpn)
1263 {
1264 	struct rvt_qpn_map *map;
1265 
1266 	map = qpt->map + qpn / RVT_BITS_PER_PAGE;
1267 	if (map->page)
1268 		clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page);
1269 }
1270 
1271 /**
1272  * rvt_destroy_qp - destroy a queue pair
1273  * @ibqp: the queue pair to destroy
1274  *
1275  * Note that this can be called while the QP is actively sending or
1276  * receiving!
1277  *
1278  * Return: 0 on success.
1279  */
1280 int rvt_destroy_qp(struct ib_qp *ibqp)
1281 {
1282 	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1283 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1284 
1285 	spin_lock_irq(&qp->r_lock);
1286 	spin_lock(&qp->s_hlock);
1287 	spin_lock(&qp->s_lock);
1288 	rvt_reset_qp(rdi, qp, ibqp->qp_type);
1289 	spin_unlock(&qp->s_lock);
1290 	spin_unlock(&qp->s_hlock);
1291 	spin_unlock_irq(&qp->r_lock);
1292 
1293 	/* qpn is now available for use again */
1294 	rvt_free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num);
1295 
1296 	spin_lock(&rdi->n_qps_lock);
1297 	rdi->n_qps_allocated--;
1298 	if (qp->ibqp.qp_type == IB_QPT_RC) {
1299 		rdi->n_rc_qps--;
1300 		rdi->busy_jiffies = rdi->n_rc_qps / RC_QP_SCALING_INTERVAL;
1301 	}
1302 	spin_unlock(&rdi->n_qps_lock);
1303 
1304 	if (qp->ip)
1305 		kref_put(&qp->ip->ref, rvt_release_mmap_info);
1306 	else
1307 		vfree(qp->r_rq.wq);
1308 	vfree(qp->s_wq);
1309 	rdi->driver_f.qp_priv_free(rdi, qp);
1310 	kfree(qp->s_ack_queue);
1311 	kfree(qp);
1312 	return 0;
1313 }
1314 
1315 /**
1316  * rvt_query_qp - query an ipbq
1317  * @ibqp: IB qp to query
1318  * @attr: attr struct to fill in
1319  * @attr_mask: attr mask ignored
1320  * @init_attr: struct to fill in
1321  *
1322  * Return: always 0
1323  */
1324 int rvt_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1325 		 int attr_mask, struct ib_qp_init_attr *init_attr)
1326 {
1327 	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1328 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1329 
1330 	attr->qp_state = qp->state;
1331 	attr->cur_qp_state = attr->qp_state;
1332 	attr->path_mtu = qp->path_mtu;
1333 	attr->path_mig_state = qp->s_mig_state;
1334 	attr->qkey = qp->qkey;
1335 	attr->rq_psn = qp->r_psn & rdi->dparms.psn_mask;
1336 	attr->sq_psn = qp->s_next_psn & rdi->dparms.psn_mask;
1337 	attr->dest_qp_num = qp->remote_qpn;
1338 	attr->qp_access_flags = qp->qp_access_flags;
1339 	attr->cap.max_send_wr = qp->s_size - 1;
1340 	attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
1341 	attr->cap.max_send_sge = qp->s_max_sge;
1342 	attr->cap.max_recv_sge = qp->r_rq.max_sge;
1343 	attr->cap.max_inline_data = 0;
1344 	attr->ah_attr = qp->remote_ah_attr;
1345 	attr->alt_ah_attr = qp->alt_ah_attr;
1346 	attr->pkey_index = qp->s_pkey_index;
1347 	attr->alt_pkey_index = qp->s_alt_pkey_index;
1348 	attr->en_sqd_async_notify = 0;
1349 	attr->sq_draining = qp->s_draining;
1350 	attr->max_rd_atomic = qp->s_max_rd_atomic;
1351 	attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
1352 	attr->min_rnr_timer = qp->r_min_rnr_timer;
1353 	attr->port_num = qp->port_num;
1354 	attr->timeout = qp->timeout;
1355 	attr->retry_cnt = qp->s_retry_cnt;
1356 	attr->rnr_retry = qp->s_rnr_retry_cnt;
1357 	attr->alt_port_num = qp->alt_ah_attr.port_num;
1358 	attr->alt_timeout = qp->alt_timeout;
1359 
1360 	init_attr->event_handler = qp->ibqp.event_handler;
1361 	init_attr->qp_context = qp->ibqp.qp_context;
1362 	init_attr->send_cq = qp->ibqp.send_cq;
1363 	init_attr->recv_cq = qp->ibqp.recv_cq;
1364 	init_attr->srq = qp->ibqp.srq;
1365 	init_attr->cap = attr->cap;
1366 	if (qp->s_flags & RVT_S_SIGNAL_REQ_WR)
1367 		init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
1368 	else
1369 		init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
1370 	init_attr->qp_type = qp->ibqp.qp_type;
1371 	init_attr->port_num = qp->port_num;
1372 	return 0;
1373 }
1374 
1375 /**
1376  * rvt_post_receive - post a receive on a QP
1377  * @ibqp: the QP to post the receive on
1378  * @wr: the WR to post
1379  * @bad_wr: the first bad WR is put here
1380  *
1381  * This may be called from interrupt context.
1382  *
1383  * Return: 0 on success otherwise errno
1384  */
1385 int rvt_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1386 		  struct ib_recv_wr **bad_wr)
1387 {
1388 	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1389 	struct rvt_rwq *wq = qp->r_rq.wq;
1390 	unsigned long flags;
1391 	int qp_err_flush = (ib_rvt_state_ops[qp->state] & RVT_FLUSH_RECV) &&
1392 				!qp->ibqp.srq;
1393 
1394 	/* Check that state is OK to post receive. */
1395 	if (!(ib_rvt_state_ops[qp->state] & RVT_POST_RECV_OK) || !wq) {
1396 		*bad_wr = wr;
1397 		return -EINVAL;
1398 	}
1399 
1400 	for (; wr; wr = wr->next) {
1401 		struct rvt_rwqe *wqe;
1402 		u32 next;
1403 		int i;
1404 
1405 		if ((unsigned)wr->num_sge > qp->r_rq.max_sge) {
1406 			*bad_wr = wr;
1407 			return -EINVAL;
1408 		}
1409 
1410 		spin_lock_irqsave(&qp->r_rq.lock, flags);
1411 		next = wq->head + 1;
1412 		if (next >= qp->r_rq.size)
1413 			next = 0;
1414 		if (next == wq->tail) {
1415 			spin_unlock_irqrestore(&qp->r_rq.lock, flags);
1416 			*bad_wr = wr;
1417 			return -ENOMEM;
1418 		}
1419 		if (unlikely(qp_err_flush)) {
1420 			struct ib_wc wc;
1421 
1422 			memset(&wc, 0, sizeof(wc));
1423 			wc.qp = &qp->ibqp;
1424 			wc.opcode = IB_WC_RECV;
1425 			wc.wr_id = wr->wr_id;
1426 			wc.status = IB_WC_WR_FLUSH_ERR;
1427 			rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
1428 		} else {
1429 			wqe = rvt_get_rwqe_ptr(&qp->r_rq, wq->head);
1430 			wqe->wr_id = wr->wr_id;
1431 			wqe->num_sge = wr->num_sge;
1432 			for (i = 0; i < wr->num_sge; i++)
1433 				wqe->sg_list[i] = wr->sg_list[i];
1434 			/*
1435 			 * Make sure queue entry is written
1436 			 * before the head index.
1437 			 */
1438 			smp_wmb();
1439 			wq->head = next;
1440 		}
1441 		spin_unlock_irqrestore(&qp->r_rq.lock, flags);
1442 	}
1443 	return 0;
1444 }
1445 
1446 /**
1447  * qp_get_savail - return number of avail send entries
1448  *
1449  * @qp - the qp
1450  *
1451  * This assumes the s_hlock is held but the s_last
1452  * qp variable is uncontrolled.
1453  */
1454 static inline u32 qp_get_savail(struct rvt_qp *qp)
1455 {
1456 	u32 slast;
1457 	u32 ret;
1458 
1459 	smp_read_barrier_depends(); /* see rc.c */
1460 	slast = ACCESS_ONCE(qp->s_last);
1461 	if (qp->s_head >= slast)
1462 		ret = qp->s_size - (qp->s_head - slast);
1463 	else
1464 		ret = slast - qp->s_head;
1465 	return ret - 1;
1466 }
1467 
1468 /**
1469  * rvt_post_one_wr - post one RC, UC, or UD send work request
1470  * @qp: the QP to post on
1471  * @wr: the work request to send
1472  */
1473 static int rvt_post_one_wr(struct rvt_qp *qp,
1474 			   struct ib_send_wr *wr,
1475 			   int *call_send)
1476 {
1477 	struct rvt_swqe *wqe;
1478 	u32 next;
1479 	int i;
1480 	int j;
1481 	int acc;
1482 	struct rvt_lkey_table *rkt;
1483 	struct rvt_pd *pd;
1484 	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
1485 	u8 log_pmtu;
1486 	int ret;
1487 
1488 	/* IB spec says that num_sge == 0 is OK. */
1489 	if (unlikely(wr->num_sge > qp->s_max_sge))
1490 		return -EINVAL;
1491 
1492 	/*
1493 	 * Don't allow RDMA reads or atomic operations on UC or
1494 	 * undefined operations.
1495 	 * Make sure buffer is large enough to hold the result for atomics.
1496 	 */
1497 	if (qp->ibqp.qp_type == IB_QPT_UC) {
1498 		if ((unsigned)wr->opcode >= IB_WR_RDMA_READ)
1499 			return -EINVAL;
1500 	} else if (qp->ibqp.qp_type != IB_QPT_RC) {
1501 		/* Check IB_QPT_SMI, IB_QPT_GSI, IB_QPT_UD opcode */
1502 		if (wr->opcode != IB_WR_SEND &&
1503 		    wr->opcode != IB_WR_SEND_WITH_IMM)
1504 			return -EINVAL;
1505 		/* Check UD destination address PD */
1506 		if (qp->ibqp.pd != ud_wr(wr)->ah->pd)
1507 			return -EINVAL;
1508 	} else if ((unsigned)wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD) {
1509 		return -EINVAL;
1510 	} else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
1511 		   (wr->num_sge == 0 ||
1512 		    wr->sg_list[0].length < sizeof(u64) ||
1513 		    wr->sg_list[0].addr & (sizeof(u64) - 1))) {
1514 		return -EINVAL;
1515 	} else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic) {
1516 		return -EINVAL;
1517 	}
1518 	/* check for avail */
1519 	if (unlikely(!qp->s_avail)) {
1520 		qp->s_avail = qp_get_savail(qp);
1521 		if (WARN_ON(qp->s_avail > (qp->s_size - 1)))
1522 			rvt_pr_err(rdi,
1523 				   "More avail entries than QP RB size.\nQP: %u, size: %u, avail: %u\nhead: %u, tail: %u, cur: %u, acked: %u, last: %u",
1524 				   qp->ibqp.qp_num, qp->s_size, qp->s_avail,
1525 				   qp->s_head, qp->s_tail, qp->s_cur,
1526 				   qp->s_acked, qp->s_last);
1527 		if (!qp->s_avail)
1528 			return -ENOMEM;
1529 	}
1530 	next = qp->s_head + 1;
1531 	if (next >= qp->s_size)
1532 		next = 0;
1533 
1534 	rkt = &rdi->lkey_table;
1535 	pd = ibpd_to_rvtpd(qp->ibqp.pd);
1536 	wqe = rvt_get_swqe_ptr(qp, qp->s_head);
1537 
1538 	if (qp->ibqp.qp_type != IB_QPT_UC &&
1539 	    qp->ibqp.qp_type != IB_QPT_RC)
1540 		memcpy(&wqe->ud_wr, ud_wr(wr), sizeof(wqe->ud_wr));
1541 	else if (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
1542 		 wr->opcode == IB_WR_RDMA_WRITE ||
1543 		 wr->opcode == IB_WR_RDMA_READ)
1544 		memcpy(&wqe->rdma_wr, rdma_wr(wr), sizeof(wqe->rdma_wr));
1545 	else if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1546 		 wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
1547 		memcpy(&wqe->atomic_wr, atomic_wr(wr), sizeof(wqe->atomic_wr));
1548 	else
1549 		memcpy(&wqe->wr, wr, sizeof(wqe->wr));
1550 
1551 	wqe->length = 0;
1552 	j = 0;
1553 	if (wr->num_sge) {
1554 		acc = wr->opcode >= IB_WR_RDMA_READ ?
1555 			IB_ACCESS_LOCAL_WRITE : 0;
1556 		for (i = 0; i < wr->num_sge; i++) {
1557 			u32 length = wr->sg_list[i].length;
1558 			int ok;
1559 
1560 			if (length == 0)
1561 				continue;
1562 			ok = rvt_lkey_ok(rkt, pd, &wqe->sg_list[j],
1563 					 &wr->sg_list[i], acc);
1564 			if (!ok) {
1565 				ret = -EINVAL;
1566 				goto bail_inval_free;
1567 			}
1568 			wqe->length += length;
1569 			j++;
1570 		}
1571 		wqe->wr.num_sge = j;
1572 	}
1573 
1574 	/* general part of wqe valid - allow for driver checks */
1575 	if (rdi->driver_f.check_send_wqe) {
1576 		ret = rdi->driver_f.check_send_wqe(qp, wqe);
1577 		if (ret < 0)
1578 			goto bail_inval_free;
1579 		if (ret)
1580 			*call_send = ret;
1581 	}
1582 
1583 	log_pmtu = qp->log_pmtu;
1584 	if (qp->ibqp.qp_type != IB_QPT_UC &&
1585 	    qp->ibqp.qp_type != IB_QPT_RC) {
1586 		struct rvt_ah *ah = ibah_to_rvtah(wqe->ud_wr.ah);
1587 
1588 		log_pmtu = ah->log_pmtu;
1589 		atomic_inc(&ibah_to_rvtah(ud_wr(wr)->ah)->refcount);
1590 	}
1591 
1592 	wqe->ssn = qp->s_ssn++;
1593 	wqe->psn = qp->s_next_psn;
1594 	wqe->lpsn = wqe->psn +
1595 			(wqe->length ? ((wqe->length - 1) >> log_pmtu) : 0);
1596 	qp->s_next_psn = wqe->lpsn + 1;
1597 	trace_rvt_post_one_wr(qp, wqe);
1598 	smp_wmb(); /* see request builders */
1599 	qp->s_avail--;
1600 	qp->s_head = next;
1601 
1602 	return 0;
1603 
1604 bail_inval_free:
1605 	/* release mr holds */
1606 	while (j) {
1607 		struct rvt_sge *sge = &wqe->sg_list[--j];
1608 
1609 		rvt_put_mr(sge->mr);
1610 	}
1611 	return ret;
1612 }
1613 
1614 /**
1615  * rvt_post_send - post a send on a QP
1616  * @ibqp: the QP to post the send on
1617  * @wr: the list of work requests to post
1618  * @bad_wr: the first bad WR is put here
1619  *
1620  * This may be called from interrupt context.
1621  *
1622  * Return: 0 on success else errno
1623  */
1624 int rvt_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1625 		  struct ib_send_wr **bad_wr)
1626 {
1627 	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1628 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1629 	unsigned long flags = 0;
1630 	int call_send;
1631 	unsigned nreq = 0;
1632 	int err = 0;
1633 
1634 	spin_lock_irqsave(&qp->s_hlock, flags);
1635 
1636 	/*
1637 	 * Ensure QP state is such that we can send. If not bail out early,
1638 	 * there is no need to do this every time we post a send.
1639 	 */
1640 	if (unlikely(!(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))) {
1641 		spin_unlock_irqrestore(&qp->s_hlock, flags);
1642 		return -EINVAL;
1643 	}
1644 
1645 	/*
1646 	 * If the send queue is empty, and we only have a single WR then just go
1647 	 * ahead and kick the send engine into gear. Otherwise we will always
1648 	 * just schedule the send to happen later.
1649 	 */
1650 	call_send = qp->s_head == ACCESS_ONCE(qp->s_last) && !wr->next;
1651 
1652 	for (; wr; wr = wr->next) {
1653 		err = rvt_post_one_wr(qp, wr, &call_send);
1654 		if (unlikely(err)) {
1655 			*bad_wr = wr;
1656 			goto bail;
1657 		}
1658 		nreq++;
1659 	}
1660 bail:
1661 	spin_unlock_irqrestore(&qp->s_hlock, flags);
1662 	if (nreq) {
1663 		if (call_send)
1664 			rdi->driver_f.do_send(qp);
1665 		else
1666 			rdi->driver_f.schedule_send_no_lock(qp);
1667 	}
1668 	return err;
1669 }
1670 
1671 /**
1672  * rvt_post_srq_receive - post a receive on a shared receive queue
1673  * @ibsrq: the SRQ to post the receive on
1674  * @wr: the list of work requests to post
1675  * @bad_wr: A pointer to the first WR to cause a problem is put here
1676  *
1677  * This may be called from interrupt context.
1678  *
1679  * Return: 0 on success else errno
1680  */
1681 int rvt_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
1682 		      struct ib_recv_wr **bad_wr)
1683 {
1684 	struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq);
1685 	struct rvt_rwq *wq;
1686 	unsigned long flags;
1687 
1688 	for (; wr; wr = wr->next) {
1689 		struct rvt_rwqe *wqe;
1690 		u32 next;
1691 		int i;
1692 
1693 		if ((unsigned)wr->num_sge > srq->rq.max_sge) {
1694 			*bad_wr = wr;
1695 			return -EINVAL;
1696 		}
1697 
1698 		spin_lock_irqsave(&srq->rq.lock, flags);
1699 		wq = srq->rq.wq;
1700 		next = wq->head + 1;
1701 		if (next >= srq->rq.size)
1702 			next = 0;
1703 		if (next == wq->tail) {
1704 			spin_unlock_irqrestore(&srq->rq.lock, flags);
1705 			*bad_wr = wr;
1706 			return -ENOMEM;
1707 		}
1708 
1709 		wqe = rvt_get_rwqe_ptr(&srq->rq, wq->head);
1710 		wqe->wr_id = wr->wr_id;
1711 		wqe->num_sge = wr->num_sge;
1712 		for (i = 0; i < wr->num_sge; i++)
1713 			wqe->sg_list[i] = wr->sg_list[i];
1714 		/* Make sure queue entry is written before the head index. */
1715 		smp_wmb();
1716 		wq->head = next;
1717 		spin_unlock_irqrestore(&srq->rq.lock, flags);
1718 	}
1719 	return 0;
1720 }
1721