xref: /openbmc/linux/drivers/infiniband/sw/rdmavt/qp.c (revision 4f205687)
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 set bits in low-order QoS bits */
373 		WARN_ON(offset & (BIT(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 	qp->r_head_ack_queue = 0;
580 	qp->s_tail_ack_queue = 0;
581 	qp->s_num_rd_atomic = 0;
582 	if (qp->r_rq.wq) {
583 		qp->r_rq.wq->head = 0;
584 		qp->r_rq.wq->tail = 0;
585 	}
586 	qp->r_sge.num_sge = 0;
587 }
588 
589 /**
590  * rvt_create_qp - create a queue pair for a device
591  * @ibpd: the protection domain who's device we create the queue pair for
592  * @init_attr: the attributes of the queue pair
593  * @udata: user data for libibverbs.so
594  *
595  * Queue pair creation is mostly an rvt issue. However, drivers have their own
596  * unique idea of what queue pair numbers mean. For instance there is a reserved
597  * range for PSM.
598  *
599  * Return: the queue pair on success, otherwise returns an errno.
600  *
601  * Called by the ib_create_qp() core verbs function.
602  */
603 struct ib_qp *rvt_create_qp(struct ib_pd *ibpd,
604 			    struct ib_qp_init_attr *init_attr,
605 			    struct ib_udata *udata)
606 {
607 	struct rvt_qp *qp;
608 	int err;
609 	struct rvt_swqe *swq = NULL;
610 	size_t sz;
611 	size_t sg_list_sz;
612 	struct ib_qp *ret = ERR_PTR(-ENOMEM);
613 	struct rvt_dev_info *rdi = ib_to_rvt(ibpd->device);
614 	void *priv = NULL;
615 	gfp_t gfp;
616 
617 	if (!rdi)
618 		return ERR_PTR(-EINVAL);
619 
620 	if (init_attr->cap.max_send_sge > rdi->dparms.props.max_sge ||
621 	    init_attr->cap.max_send_wr > rdi->dparms.props.max_qp_wr ||
622 	    init_attr->create_flags & ~(IB_QP_CREATE_USE_GFP_NOIO))
623 		return ERR_PTR(-EINVAL);
624 
625 	/* GFP_NOIO is applicable to RC QP's only */
626 
627 	if (init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO &&
628 	    init_attr->qp_type != IB_QPT_RC)
629 		return ERR_PTR(-EINVAL);
630 
631 	gfp = init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO ?
632 						GFP_NOIO : GFP_KERNEL;
633 
634 	/* Check receive queue parameters if no SRQ is specified. */
635 	if (!init_attr->srq) {
636 		if (init_attr->cap.max_recv_sge > rdi->dparms.props.max_sge ||
637 		    init_attr->cap.max_recv_wr > rdi->dparms.props.max_qp_wr)
638 			return ERR_PTR(-EINVAL);
639 
640 		if (init_attr->cap.max_send_sge +
641 		    init_attr->cap.max_send_wr +
642 		    init_attr->cap.max_recv_sge +
643 		    init_attr->cap.max_recv_wr == 0)
644 			return ERR_PTR(-EINVAL);
645 	}
646 
647 	switch (init_attr->qp_type) {
648 	case IB_QPT_SMI:
649 	case IB_QPT_GSI:
650 		if (init_attr->port_num == 0 ||
651 		    init_attr->port_num > ibpd->device->phys_port_cnt)
652 			return ERR_PTR(-EINVAL);
653 	case IB_QPT_UC:
654 	case IB_QPT_RC:
655 	case IB_QPT_UD:
656 		sz = sizeof(struct rvt_sge) *
657 			init_attr->cap.max_send_sge +
658 			sizeof(struct rvt_swqe);
659 		if (gfp == GFP_NOIO)
660 			swq = __vmalloc(
661 				(init_attr->cap.max_send_wr + 1) * sz,
662 				gfp | __GFP_ZERO, PAGE_KERNEL);
663 		else
664 			swq = vzalloc_node(
665 				(init_attr->cap.max_send_wr + 1) * sz,
666 				rdi->dparms.node);
667 		if (!swq)
668 			return ERR_PTR(-ENOMEM);
669 
670 		sz = sizeof(*qp);
671 		sg_list_sz = 0;
672 		if (init_attr->srq) {
673 			struct rvt_srq *srq = ibsrq_to_rvtsrq(init_attr->srq);
674 
675 			if (srq->rq.max_sge > 1)
676 				sg_list_sz = sizeof(*qp->r_sg_list) *
677 					(srq->rq.max_sge - 1);
678 		} else if (init_attr->cap.max_recv_sge > 1)
679 			sg_list_sz = sizeof(*qp->r_sg_list) *
680 				(init_attr->cap.max_recv_sge - 1);
681 		qp = kzalloc_node(sz + sg_list_sz, gfp, rdi->dparms.node);
682 		if (!qp)
683 			goto bail_swq;
684 
685 		RCU_INIT_POINTER(qp->next, NULL);
686 		if (init_attr->qp_type == IB_QPT_RC) {
687 			qp->s_ack_queue =
688 				kzalloc_node(
689 					sizeof(*qp->s_ack_queue) *
690 					 rvt_max_atomic(rdi),
691 					gfp,
692 					rdi->dparms.node);
693 			if (!qp->s_ack_queue)
694 				goto bail_qp;
695 		}
696 
697 		/*
698 		 * Driver needs to set up it's private QP structure and do any
699 		 * initialization that is needed.
700 		 */
701 		priv = rdi->driver_f.qp_priv_alloc(rdi, qp, gfp);
702 		if (IS_ERR(priv)) {
703 			ret = priv;
704 			goto bail_qp;
705 		}
706 		qp->priv = priv;
707 		qp->timeout_jiffies =
708 			usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
709 				1000UL);
710 		if (init_attr->srq) {
711 			sz = 0;
712 		} else {
713 			qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
714 			qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
715 			sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
716 				sizeof(struct rvt_rwqe);
717 			if (udata)
718 				qp->r_rq.wq = vmalloc_user(
719 						sizeof(struct rvt_rwq) +
720 						qp->r_rq.size * sz);
721 			else if (gfp == GFP_NOIO)
722 				qp->r_rq.wq = __vmalloc(
723 						sizeof(struct rvt_rwq) +
724 						qp->r_rq.size * sz,
725 						gfp | __GFP_ZERO, PAGE_KERNEL);
726 			else
727 				qp->r_rq.wq = vzalloc_node(
728 						sizeof(struct rvt_rwq) +
729 						qp->r_rq.size * sz,
730 						rdi->dparms.node);
731 			if (!qp->r_rq.wq)
732 				goto bail_driver_priv;
733 		}
734 
735 		/*
736 		 * ib_create_qp() will initialize qp->ibqp
737 		 * except for qp->ibqp.qp_num.
738 		 */
739 		spin_lock_init(&qp->r_lock);
740 		spin_lock_init(&qp->s_hlock);
741 		spin_lock_init(&qp->s_lock);
742 		spin_lock_init(&qp->r_rq.lock);
743 		atomic_set(&qp->refcount, 0);
744 		init_waitqueue_head(&qp->wait);
745 		init_timer(&qp->s_timer);
746 		qp->s_timer.data = (unsigned long)qp;
747 		INIT_LIST_HEAD(&qp->rspwait);
748 		qp->state = IB_QPS_RESET;
749 		qp->s_wq = swq;
750 		qp->s_size = init_attr->cap.max_send_wr + 1;
751 		qp->s_avail = init_attr->cap.max_send_wr;
752 		qp->s_max_sge = init_attr->cap.max_send_sge;
753 		if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
754 			qp->s_flags = RVT_S_SIGNAL_REQ_WR;
755 
756 		err = alloc_qpn(rdi, &rdi->qp_dev->qpn_table,
757 				init_attr->qp_type,
758 				init_attr->port_num, gfp);
759 		if (err < 0) {
760 			ret = ERR_PTR(err);
761 			goto bail_rq_wq;
762 		}
763 		qp->ibqp.qp_num = err;
764 		qp->port_num = init_attr->port_num;
765 		rvt_reset_qp(rdi, qp, init_attr->qp_type);
766 		break;
767 
768 	default:
769 		/* Don't support raw QPs */
770 		return ERR_PTR(-EINVAL);
771 	}
772 
773 	init_attr->cap.max_inline_data = 0;
774 
775 	/*
776 	 * Return the address of the RWQ as the offset to mmap.
777 	 * See rvt_mmap() for details.
778 	 */
779 	if (udata && udata->outlen >= sizeof(__u64)) {
780 		if (!qp->r_rq.wq) {
781 			__u64 offset = 0;
782 
783 			err = ib_copy_to_udata(udata, &offset,
784 					       sizeof(offset));
785 			if (err) {
786 				ret = ERR_PTR(err);
787 				goto bail_qpn;
788 			}
789 		} else {
790 			u32 s = sizeof(struct rvt_rwq) + qp->r_rq.size * sz;
791 
792 			qp->ip = rvt_create_mmap_info(rdi, s,
793 						      ibpd->uobject->context,
794 						      qp->r_rq.wq);
795 			if (!qp->ip) {
796 				ret = ERR_PTR(-ENOMEM);
797 				goto bail_qpn;
798 			}
799 
800 			err = ib_copy_to_udata(udata, &qp->ip->offset,
801 					       sizeof(qp->ip->offset));
802 			if (err) {
803 				ret = ERR_PTR(err);
804 				goto bail_ip;
805 			}
806 		}
807 		qp->pid = current->pid;
808 	}
809 
810 	spin_lock(&rdi->n_qps_lock);
811 	if (rdi->n_qps_allocated == rdi->dparms.props.max_qp) {
812 		spin_unlock(&rdi->n_qps_lock);
813 		ret = ERR_PTR(-ENOMEM);
814 		goto bail_ip;
815 	}
816 
817 	rdi->n_qps_allocated++;
818 	/*
819 	 * Maintain a busy_jiffies variable that will be added to the timeout
820 	 * period in mod_retry_timer and add_retry_timer. This busy jiffies
821 	 * is scaled by the number of rc qps created for the device to reduce
822 	 * the number of timeouts occurring when there is a large number of
823 	 * qps. busy_jiffies is incremented every rc qp scaling interval.
824 	 * The scaling interval is selected based on extensive performance
825 	 * evaluation of targeted workloads.
826 	 */
827 	if (init_attr->qp_type == IB_QPT_RC) {
828 		rdi->n_rc_qps++;
829 		rdi->busy_jiffies = rdi->n_rc_qps / RC_QP_SCALING_INTERVAL;
830 	}
831 	spin_unlock(&rdi->n_qps_lock);
832 
833 	if (qp->ip) {
834 		spin_lock_irq(&rdi->pending_lock);
835 		list_add(&qp->ip->pending_mmaps, &rdi->pending_mmaps);
836 		spin_unlock_irq(&rdi->pending_lock);
837 	}
838 
839 	ret = &qp->ibqp;
840 
841 	/*
842 	 * We have our QP and its good, now keep track of what types of opcodes
843 	 * can be processed on this QP. We do this by keeping track of what the
844 	 * 3 high order bits of the opcode are.
845 	 */
846 	switch (init_attr->qp_type) {
847 	case IB_QPT_SMI:
848 	case IB_QPT_GSI:
849 	case IB_QPT_UD:
850 		qp->allowed_ops = IB_OPCODE_UD;
851 		break;
852 	case IB_QPT_RC:
853 		qp->allowed_ops = IB_OPCODE_RC;
854 		break;
855 	case IB_QPT_UC:
856 		qp->allowed_ops = IB_OPCODE_UC;
857 		break;
858 	default:
859 		ret = ERR_PTR(-EINVAL);
860 		goto bail_ip;
861 	}
862 
863 	return ret;
864 
865 bail_ip:
866 	kref_put(&qp->ip->ref, rvt_release_mmap_info);
867 
868 bail_qpn:
869 	free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num);
870 
871 bail_rq_wq:
872 	vfree(qp->r_rq.wq);
873 
874 bail_driver_priv:
875 	rdi->driver_f.qp_priv_free(rdi, qp);
876 
877 bail_qp:
878 	kfree(qp->s_ack_queue);
879 	kfree(qp);
880 
881 bail_swq:
882 	vfree(swq);
883 
884 	return ret;
885 }
886 
887 /**
888  * rvt_error_qp - put a QP into the error state
889  * @qp: the QP to put into the error state
890  * @err: the receive completion error to signal if a RWQE is active
891  *
892  * Flushes both send and receive work queues.
893  *
894  * Return: true if last WQE event should be generated.
895  * The QP r_lock and s_lock should be held and interrupts disabled.
896  * If we are already in error state, just return.
897  */
898 int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err)
899 {
900 	struct ib_wc wc;
901 	int ret = 0;
902 	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
903 
904 	if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET)
905 		goto bail;
906 
907 	qp->state = IB_QPS_ERR;
908 
909 	if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) {
910 		qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR);
911 		del_timer(&qp->s_timer);
912 	}
913 
914 	if (qp->s_flags & RVT_S_ANY_WAIT_SEND)
915 		qp->s_flags &= ~RVT_S_ANY_WAIT_SEND;
916 
917 	rdi->driver_f.notify_error_qp(qp);
918 
919 	/* Schedule the sending tasklet to drain the send work queue. */
920 	if (ACCESS_ONCE(qp->s_last) != qp->s_head)
921 		rdi->driver_f.schedule_send(qp);
922 
923 	rvt_clear_mr_refs(qp, 0);
924 
925 	memset(&wc, 0, sizeof(wc));
926 	wc.qp = &qp->ibqp;
927 	wc.opcode = IB_WC_RECV;
928 
929 	if (test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) {
930 		wc.wr_id = qp->r_wr_id;
931 		wc.status = err;
932 		rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
933 	}
934 	wc.status = IB_WC_WR_FLUSH_ERR;
935 
936 	if (qp->r_rq.wq) {
937 		struct rvt_rwq *wq;
938 		u32 head;
939 		u32 tail;
940 
941 		spin_lock(&qp->r_rq.lock);
942 
943 		/* sanity check pointers before trusting them */
944 		wq = qp->r_rq.wq;
945 		head = wq->head;
946 		if (head >= qp->r_rq.size)
947 			head = 0;
948 		tail = wq->tail;
949 		if (tail >= qp->r_rq.size)
950 			tail = 0;
951 		while (tail != head) {
952 			wc.wr_id = rvt_get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
953 			if (++tail >= qp->r_rq.size)
954 				tail = 0;
955 			rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
956 		}
957 		wq->tail = tail;
958 
959 		spin_unlock(&qp->r_rq.lock);
960 	} else if (qp->ibqp.event_handler) {
961 		ret = 1;
962 	}
963 
964 bail:
965 	return ret;
966 }
967 EXPORT_SYMBOL(rvt_error_qp);
968 
969 /*
970  * Put the QP into the hash table.
971  * The hash table holds a reference to the QP.
972  */
973 static void rvt_insert_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp)
974 {
975 	struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1];
976 	unsigned long flags;
977 
978 	atomic_inc(&qp->refcount);
979 	spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags);
980 
981 	if (qp->ibqp.qp_num <= 1) {
982 		rcu_assign_pointer(rvp->qp[qp->ibqp.qp_num], qp);
983 	} else {
984 		u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits);
985 
986 		qp->next = rdi->qp_dev->qp_table[n];
987 		rcu_assign_pointer(rdi->qp_dev->qp_table[n], qp);
988 		trace_rvt_qpinsert(qp, n);
989 	}
990 
991 	spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags);
992 }
993 
994 /**
995  * qib_modify_qp - modify the attributes of a queue pair
996  * @ibqp: the queue pair who's attributes we're modifying
997  * @attr: the new attributes
998  * @attr_mask: the mask of attributes to modify
999  * @udata: user data for libibverbs.so
1000  *
1001  * Return: 0 on success, otherwise returns an errno.
1002  */
1003 int rvt_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1004 		  int attr_mask, struct ib_udata *udata)
1005 {
1006 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1007 	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1008 	enum ib_qp_state cur_state, new_state;
1009 	struct ib_event ev;
1010 	int lastwqe = 0;
1011 	int mig = 0;
1012 	int pmtu = 0; /* for gcc warning only */
1013 	enum rdma_link_layer link;
1014 
1015 	link = rdma_port_get_link_layer(ibqp->device, qp->port_num);
1016 
1017 	spin_lock_irq(&qp->r_lock);
1018 	spin_lock(&qp->s_hlock);
1019 	spin_lock(&qp->s_lock);
1020 
1021 	cur_state = attr_mask & IB_QP_CUR_STATE ?
1022 		attr->cur_qp_state : qp->state;
1023 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1024 
1025 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1026 				attr_mask, link))
1027 		goto inval;
1028 
1029 	if (rdi->driver_f.check_modify_qp &&
1030 	    rdi->driver_f.check_modify_qp(qp, attr, attr_mask, udata))
1031 		goto inval;
1032 
1033 	if (attr_mask & IB_QP_AV) {
1034 		if (attr->ah_attr.dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE))
1035 			goto inval;
1036 		if (rvt_check_ah(qp->ibqp.device, &attr->ah_attr))
1037 			goto inval;
1038 	}
1039 
1040 	if (attr_mask & IB_QP_ALT_PATH) {
1041 		if (attr->alt_ah_attr.dlid >=
1042 		    be16_to_cpu(IB_MULTICAST_LID_BASE))
1043 			goto inval;
1044 		if (rvt_check_ah(qp->ibqp.device, &attr->alt_ah_attr))
1045 			goto inval;
1046 		if (attr->alt_pkey_index >= rvt_get_npkeys(rdi))
1047 			goto inval;
1048 	}
1049 
1050 	if (attr_mask & IB_QP_PKEY_INDEX)
1051 		if (attr->pkey_index >= rvt_get_npkeys(rdi))
1052 			goto inval;
1053 
1054 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
1055 		if (attr->min_rnr_timer > 31)
1056 			goto inval;
1057 
1058 	if (attr_mask & IB_QP_PORT)
1059 		if (qp->ibqp.qp_type == IB_QPT_SMI ||
1060 		    qp->ibqp.qp_type == IB_QPT_GSI ||
1061 		    attr->port_num == 0 ||
1062 		    attr->port_num > ibqp->device->phys_port_cnt)
1063 			goto inval;
1064 
1065 	if (attr_mask & IB_QP_DEST_QPN)
1066 		if (attr->dest_qp_num > RVT_QPN_MASK)
1067 			goto inval;
1068 
1069 	if (attr_mask & IB_QP_RETRY_CNT)
1070 		if (attr->retry_cnt > 7)
1071 			goto inval;
1072 
1073 	if (attr_mask & IB_QP_RNR_RETRY)
1074 		if (attr->rnr_retry > 7)
1075 			goto inval;
1076 
1077 	/*
1078 	 * Don't allow invalid path_mtu values.  OK to set greater
1079 	 * than the active mtu (or even the max_cap, if we have tuned
1080 	 * that to a small mtu.  We'll set qp->path_mtu
1081 	 * to the lesser of requested attribute mtu and active,
1082 	 * for packetizing messages.
1083 	 * Note that the QP port has to be set in INIT and MTU in RTR.
1084 	 */
1085 	if (attr_mask & IB_QP_PATH_MTU) {
1086 		pmtu = rdi->driver_f.get_pmtu_from_attr(rdi, qp, attr);
1087 		if (pmtu < 0)
1088 			goto inval;
1089 	}
1090 
1091 	if (attr_mask & IB_QP_PATH_MIG_STATE) {
1092 		if (attr->path_mig_state == IB_MIG_REARM) {
1093 			if (qp->s_mig_state == IB_MIG_ARMED)
1094 				goto inval;
1095 			if (new_state != IB_QPS_RTS)
1096 				goto inval;
1097 		} else if (attr->path_mig_state == IB_MIG_MIGRATED) {
1098 			if (qp->s_mig_state == IB_MIG_REARM)
1099 				goto inval;
1100 			if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD)
1101 				goto inval;
1102 			if (qp->s_mig_state == IB_MIG_ARMED)
1103 				mig = 1;
1104 		} else {
1105 			goto inval;
1106 		}
1107 	}
1108 
1109 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1110 		if (attr->max_dest_rd_atomic > rdi->dparms.max_rdma_atomic)
1111 			goto inval;
1112 
1113 	switch (new_state) {
1114 	case IB_QPS_RESET:
1115 		if (qp->state != IB_QPS_RESET)
1116 			rvt_reset_qp(rdi, qp, ibqp->qp_type);
1117 		break;
1118 
1119 	case IB_QPS_RTR:
1120 		/* Allow event to re-trigger if QP set to RTR more than once */
1121 		qp->r_flags &= ~RVT_R_COMM_EST;
1122 		qp->state = new_state;
1123 		break;
1124 
1125 	case IB_QPS_SQD:
1126 		qp->s_draining = qp->s_last != qp->s_cur;
1127 		qp->state = new_state;
1128 		break;
1129 
1130 	case IB_QPS_SQE:
1131 		if (qp->ibqp.qp_type == IB_QPT_RC)
1132 			goto inval;
1133 		qp->state = new_state;
1134 		break;
1135 
1136 	case IB_QPS_ERR:
1137 		lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1138 		break;
1139 
1140 	default:
1141 		qp->state = new_state;
1142 		break;
1143 	}
1144 
1145 	if (attr_mask & IB_QP_PKEY_INDEX)
1146 		qp->s_pkey_index = attr->pkey_index;
1147 
1148 	if (attr_mask & IB_QP_PORT)
1149 		qp->port_num = attr->port_num;
1150 
1151 	if (attr_mask & IB_QP_DEST_QPN)
1152 		qp->remote_qpn = attr->dest_qp_num;
1153 
1154 	if (attr_mask & IB_QP_SQ_PSN) {
1155 		qp->s_next_psn = attr->sq_psn & rdi->dparms.psn_modify_mask;
1156 		qp->s_psn = qp->s_next_psn;
1157 		qp->s_sending_psn = qp->s_next_psn;
1158 		qp->s_last_psn = qp->s_next_psn - 1;
1159 		qp->s_sending_hpsn = qp->s_last_psn;
1160 	}
1161 
1162 	if (attr_mask & IB_QP_RQ_PSN)
1163 		qp->r_psn = attr->rq_psn & rdi->dparms.psn_modify_mask;
1164 
1165 	if (attr_mask & IB_QP_ACCESS_FLAGS)
1166 		qp->qp_access_flags = attr->qp_access_flags;
1167 
1168 	if (attr_mask & IB_QP_AV) {
1169 		qp->remote_ah_attr = attr->ah_attr;
1170 		qp->s_srate = attr->ah_attr.static_rate;
1171 		qp->srate_mbps = ib_rate_to_mbps(qp->s_srate);
1172 	}
1173 
1174 	if (attr_mask & IB_QP_ALT_PATH) {
1175 		qp->alt_ah_attr = attr->alt_ah_attr;
1176 		qp->s_alt_pkey_index = attr->alt_pkey_index;
1177 	}
1178 
1179 	if (attr_mask & IB_QP_PATH_MIG_STATE) {
1180 		qp->s_mig_state = attr->path_mig_state;
1181 		if (mig) {
1182 			qp->remote_ah_attr = qp->alt_ah_attr;
1183 			qp->port_num = qp->alt_ah_attr.port_num;
1184 			qp->s_pkey_index = qp->s_alt_pkey_index;
1185 		}
1186 	}
1187 
1188 	if (attr_mask & IB_QP_PATH_MTU) {
1189 		qp->pmtu = rdi->driver_f.mtu_from_qp(rdi, qp, pmtu);
1190 		qp->path_mtu = rdi->driver_f.mtu_to_path_mtu(qp->pmtu);
1191 		qp->log_pmtu = ilog2(qp->pmtu);
1192 	}
1193 
1194 	if (attr_mask & IB_QP_RETRY_CNT) {
1195 		qp->s_retry_cnt = attr->retry_cnt;
1196 		qp->s_retry = attr->retry_cnt;
1197 	}
1198 
1199 	if (attr_mask & IB_QP_RNR_RETRY) {
1200 		qp->s_rnr_retry_cnt = attr->rnr_retry;
1201 		qp->s_rnr_retry = attr->rnr_retry;
1202 	}
1203 
1204 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
1205 		qp->r_min_rnr_timer = attr->min_rnr_timer;
1206 
1207 	if (attr_mask & IB_QP_TIMEOUT) {
1208 		qp->timeout = attr->timeout;
1209 		qp->timeout_jiffies =
1210 			usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
1211 				1000UL);
1212 	}
1213 
1214 	if (attr_mask & IB_QP_QKEY)
1215 		qp->qkey = attr->qkey;
1216 
1217 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1218 		qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
1219 
1220 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
1221 		qp->s_max_rd_atomic = attr->max_rd_atomic;
1222 
1223 	if (rdi->driver_f.modify_qp)
1224 		rdi->driver_f.modify_qp(qp, attr, attr_mask, udata);
1225 
1226 	spin_unlock(&qp->s_lock);
1227 	spin_unlock(&qp->s_hlock);
1228 	spin_unlock_irq(&qp->r_lock);
1229 
1230 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1231 		rvt_insert_qp(rdi, qp);
1232 
1233 	if (lastwqe) {
1234 		ev.device = qp->ibqp.device;
1235 		ev.element.qp = &qp->ibqp;
1236 		ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
1237 		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1238 	}
1239 	if (mig) {
1240 		ev.device = qp->ibqp.device;
1241 		ev.element.qp = &qp->ibqp;
1242 		ev.event = IB_EVENT_PATH_MIG;
1243 		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1244 	}
1245 	return 0;
1246 
1247 inval:
1248 	spin_unlock(&qp->s_lock);
1249 	spin_unlock(&qp->s_hlock);
1250 	spin_unlock_irq(&qp->r_lock);
1251 	return -EINVAL;
1252 }
1253 
1254 /** rvt_free_qpn - Free a qpn from the bit map
1255  * @qpt: QP table
1256  * @qpn: queue pair number to free
1257  */
1258 static void rvt_free_qpn(struct rvt_qpn_table *qpt, u32 qpn)
1259 {
1260 	struct rvt_qpn_map *map;
1261 
1262 	map = qpt->map + qpn / RVT_BITS_PER_PAGE;
1263 	if (map->page)
1264 		clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page);
1265 }
1266 
1267 /**
1268  * rvt_destroy_qp - destroy a queue pair
1269  * @ibqp: the queue pair to destroy
1270  *
1271  * Note that this can be called while the QP is actively sending or
1272  * receiving!
1273  *
1274  * Return: 0 on success.
1275  */
1276 int rvt_destroy_qp(struct ib_qp *ibqp)
1277 {
1278 	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1279 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1280 
1281 	spin_lock_irq(&qp->r_lock);
1282 	spin_lock(&qp->s_hlock);
1283 	spin_lock(&qp->s_lock);
1284 	rvt_reset_qp(rdi, qp, ibqp->qp_type);
1285 	spin_unlock(&qp->s_lock);
1286 	spin_unlock(&qp->s_hlock);
1287 	spin_unlock_irq(&qp->r_lock);
1288 
1289 	/* qpn is now available for use again */
1290 	rvt_free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num);
1291 
1292 	spin_lock(&rdi->n_qps_lock);
1293 	rdi->n_qps_allocated--;
1294 	if (qp->ibqp.qp_type == IB_QPT_RC) {
1295 		rdi->n_rc_qps--;
1296 		rdi->busy_jiffies = rdi->n_rc_qps / RC_QP_SCALING_INTERVAL;
1297 	}
1298 	spin_unlock(&rdi->n_qps_lock);
1299 
1300 	if (qp->ip)
1301 		kref_put(&qp->ip->ref, rvt_release_mmap_info);
1302 	else
1303 		vfree(qp->r_rq.wq);
1304 	vfree(qp->s_wq);
1305 	rdi->driver_f.qp_priv_free(rdi, qp);
1306 	kfree(qp->s_ack_queue);
1307 	kfree(qp);
1308 	return 0;
1309 }
1310 
1311 /**
1312  * rvt_query_qp - query an ipbq
1313  * @ibqp: IB qp to query
1314  * @attr: attr struct to fill in
1315  * @attr_mask: attr mask ignored
1316  * @init_attr: struct to fill in
1317  *
1318  * Return: always 0
1319  */
1320 int rvt_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1321 		 int attr_mask, struct ib_qp_init_attr *init_attr)
1322 {
1323 	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1324 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1325 
1326 	attr->qp_state = qp->state;
1327 	attr->cur_qp_state = attr->qp_state;
1328 	attr->path_mtu = qp->path_mtu;
1329 	attr->path_mig_state = qp->s_mig_state;
1330 	attr->qkey = qp->qkey;
1331 	attr->rq_psn = qp->r_psn & rdi->dparms.psn_mask;
1332 	attr->sq_psn = qp->s_next_psn & rdi->dparms.psn_mask;
1333 	attr->dest_qp_num = qp->remote_qpn;
1334 	attr->qp_access_flags = qp->qp_access_flags;
1335 	attr->cap.max_send_wr = qp->s_size - 1;
1336 	attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
1337 	attr->cap.max_send_sge = qp->s_max_sge;
1338 	attr->cap.max_recv_sge = qp->r_rq.max_sge;
1339 	attr->cap.max_inline_data = 0;
1340 	attr->ah_attr = qp->remote_ah_attr;
1341 	attr->alt_ah_attr = qp->alt_ah_attr;
1342 	attr->pkey_index = qp->s_pkey_index;
1343 	attr->alt_pkey_index = qp->s_alt_pkey_index;
1344 	attr->en_sqd_async_notify = 0;
1345 	attr->sq_draining = qp->s_draining;
1346 	attr->max_rd_atomic = qp->s_max_rd_atomic;
1347 	attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
1348 	attr->min_rnr_timer = qp->r_min_rnr_timer;
1349 	attr->port_num = qp->port_num;
1350 	attr->timeout = qp->timeout;
1351 	attr->retry_cnt = qp->s_retry_cnt;
1352 	attr->rnr_retry = qp->s_rnr_retry_cnt;
1353 	attr->alt_port_num = qp->alt_ah_attr.port_num;
1354 	attr->alt_timeout = qp->alt_timeout;
1355 
1356 	init_attr->event_handler = qp->ibqp.event_handler;
1357 	init_attr->qp_context = qp->ibqp.qp_context;
1358 	init_attr->send_cq = qp->ibqp.send_cq;
1359 	init_attr->recv_cq = qp->ibqp.recv_cq;
1360 	init_attr->srq = qp->ibqp.srq;
1361 	init_attr->cap = attr->cap;
1362 	if (qp->s_flags & RVT_S_SIGNAL_REQ_WR)
1363 		init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
1364 	else
1365 		init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
1366 	init_attr->qp_type = qp->ibqp.qp_type;
1367 	init_attr->port_num = qp->port_num;
1368 	return 0;
1369 }
1370 
1371 /**
1372  * rvt_post_receive - post a receive on a QP
1373  * @ibqp: the QP to post the receive on
1374  * @wr: the WR to post
1375  * @bad_wr: the first bad WR is put here
1376  *
1377  * This may be called from interrupt context.
1378  *
1379  * Return: 0 on success otherwise errno
1380  */
1381 int rvt_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1382 		  struct ib_recv_wr **bad_wr)
1383 {
1384 	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1385 	struct rvt_rwq *wq = qp->r_rq.wq;
1386 	unsigned long flags;
1387 	int qp_err_flush = (ib_rvt_state_ops[qp->state] & RVT_FLUSH_RECV) &&
1388 				!qp->ibqp.srq;
1389 
1390 	/* Check that state is OK to post receive. */
1391 	if (!(ib_rvt_state_ops[qp->state] & RVT_POST_RECV_OK) || !wq) {
1392 		*bad_wr = wr;
1393 		return -EINVAL;
1394 	}
1395 
1396 	for (; wr; wr = wr->next) {
1397 		struct rvt_rwqe *wqe;
1398 		u32 next;
1399 		int i;
1400 
1401 		if ((unsigned)wr->num_sge > qp->r_rq.max_sge) {
1402 			*bad_wr = wr;
1403 			return -EINVAL;
1404 		}
1405 
1406 		spin_lock_irqsave(&qp->r_rq.lock, flags);
1407 		next = wq->head + 1;
1408 		if (next >= qp->r_rq.size)
1409 			next = 0;
1410 		if (next == wq->tail) {
1411 			spin_unlock_irqrestore(&qp->r_rq.lock, flags);
1412 			*bad_wr = wr;
1413 			return -ENOMEM;
1414 		}
1415 		if (unlikely(qp_err_flush)) {
1416 			struct ib_wc wc;
1417 
1418 			memset(&wc, 0, sizeof(wc));
1419 			wc.qp = &qp->ibqp;
1420 			wc.opcode = IB_WC_RECV;
1421 			wc.wr_id = wr->wr_id;
1422 			wc.status = IB_WC_WR_FLUSH_ERR;
1423 			rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
1424 		} else {
1425 			wqe = rvt_get_rwqe_ptr(&qp->r_rq, wq->head);
1426 			wqe->wr_id = wr->wr_id;
1427 			wqe->num_sge = wr->num_sge;
1428 			for (i = 0; i < wr->num_sge; i++)
1429 				wqe->sg_list[i] = wr->sg_list[i];
1430 			/*
1431 			 * Make sure queue entry is written
1432 			 * before the head index.
1433 			 */
1434 			smp_wmb();
1435 			wq->head = next;
1436 		}
1437 		spin_unlock_irqrestore(&qp->r_rq.lock, flags);
1438 	}
1439 	return 0;
1440 }
1441 
1442 /**
1443  * qp_get_savail - return number of avail send entries
1444  *
1445  * @qp - the qp
1446  *
1447  * This assumes the s_hlock is held but the s_last
1448  * qp variable is uncontrolled.
1449  */
1450 static inline u32 qp_get_savail(struct rvt_qp *qp)
1451 {
1452 	u32 slast;
1453 	u32 ret;
1454 
1455 	smp_read_barrier_depends(); /* see rc.c */
1456 	slast = ACCESS_ONCE(qp->s_last);
1457 	if (qp->s_head >= slast)
1458 		ret = qp->s_size - (qp->s_head - slast);
1459 	else
1460 		ret = slast - qp->s_head;
1461 	return ret - 1;
1462 }
1463 
1464 /**
1465  * rvt_post_one_wr - post one RC, UC, or UD send work request
1466  * @qp: the QP to post on
1467  * @wr: the work request to send
1468  */
1469 static int rvt_post_one_wr(struct rvt_qp *qp,
1470 			   struct ib_send_wr *wr,
1471 			   int *call_send)
1472 {
1473 	struct rvt_swqe *wqe;
1474 	u32 next;
1475 	int i;
1476 	int j;
1477 	int acc;
1478 	struct rvt_lkey_table *rkt;
1479 	struct rvt_pd *pd;
1480 	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
1481 	u8 log_pmtu;
1482 	int ret;
1483 
1484 	/* IB spec says that num_sge == 0 is OK. */
1485 	if (unlikely(wr->num_sge > qp->s_max_sge))
1486 		return -EINVAL;
1487 
1488 	/*
1489 	 * Don't allow RDMA reads or atomic operations on UC or
1490 	 * undefined operations.
1491 	 * Make sure buffer is large enough to hold the result for atomics.
1492 	 */
1493 	if (qp->ibqp.qp_type == IB_QPT_UC) {
1494 		if ((unsigned)wr->opcode >= IB_WR_RDMA_READ)
1495 			return -EINVAL;
1496 	} else if (qp->ibqp.qp_type != IB_QPT_RC) {
1497 		/* Check IB_QPT_SMI, IB_QPT_GSI, IB_QPT_UD opcode */
1498 		if (wr->opcode != IB_WR_SEND &&
1499 		    wr->opcode != IB_WR_SEND_WITH_IMM)
1500 			return -EINVAL;
1501 		/* Check UD destination address PD */
1502 		if (qp->ibqp.pd != ud_wr(wr)->ah->pd)
1503 			return -EINVAL;
1504 	} else if ((unsigned)wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD) {
1505 		return -EINVAL;
1506 	} else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
1507 		   (wr->num_sge == 0 ||
1508 		    wr->sg_list[0].length < sizeof(u64) ||
1509 		    wr->sg_list[0].addr & (sizeof(u64) - 1))) {
1510 		return -EINVAL;
1511 	} else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic) {
1512 		return -EINVAL;
1513 	}
1514 	/* check for avail */
1515 	if (unlikely(!qp->s_avail)) {
1516 		qp->s_avail = qp_get_savail(qp);
1517 		if (WARN_ON(qp->s_avail > (qp->s_size - 1)))
1518 			rvt_pr_err(rdi,
1519 				   "More avail entries than QP RB size.\nQP: %u, size: %u, avail: %u\nhead: %u, tail: %u, cur: %u, acked: %u, last: %u",
1520 				   qp->ibqp.qp_num, qp->s_size, qp->s_avail,
1521 				   qp->s_head, qp->s_tail, qp->s_cur,
1522 				   qp->s_acked, qp->s_last);
1523 		if (!qp->s_avail)
1524 			return -ENOMEM;
1525 	}
1526 	next = qp->s_head + 1;
1527 	if (next >= qp->s_size)
1528 		next = 0;
1529 
1530 	rkt = &rdi->lkey_table;
1531 	pd = ibpd_to_rvtpd(qp->ibqp.pd);
1532 	wqe = rvt_get_swqe_ptr(qp, qp->s_head);
1533 
1534 	if (qp->ibqp.qp_type != IB_QPT_UC &&
1535 	    qp->ibqp.qp_type != IB_QPT_RC)
1536 		memcpy(&wqe->ud_wr, ud_wr(wr), sizeof(wqe->ud_wr));
1537 	else if (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
1538 		 wr->opcode == IB_WR_RDMA_WRITE ||
1539 		 wr->opcode == IB_WR_RDMA_READ)
1540 		memcpy(&wqe->rdma_wr, rdma_wr(wr), sizeof(wqe->rdma_wr));
1541 	else if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1542 		 wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
1543 		memcpy(&wqe->atomic_wr, atomic_wr(wr), sizeof(wqe->atomic_wr));
1544 	else
1545 		memcpy(&wqe->wr, wr, sizeof(wqe->wr));
1546 
1547 	wqe->length = 0;
1548 	j = 0;
1549 	if (wr->num_sge) {
1550 		acc = wr->opcode >= IB_WR_RDMA_READ ?
1551 			IB_ACCESS_LOCAL_WRITE : 0;
1552 		for (i = 0; i < wr->num_sge; i++) {
1553 			u32 length = wr->sg_list[i].length;
1554 			int ok;
1555 
1556 			if (length == 0)
1557 				continue;
1558 			ok = rvt_lkey_ok(rkt, pd, &wqe->sg_list[j],
1559 					 &wr->sg_list[i], acc);
1560 			if (!ok) {
1561 				ret = -EINVAL;
1562 				goto bail_inval_free;
1563 			}
1564 			wqe->length += length;
1565 			j++;
1566 		}
1567 		wqe->wr.num_sge = j;
1568 	}
1569 
1570 	/* general part of wqe valid - allow for driver checks */
1571 	if (rdi->driver_f.check_send_wqe) {
1572 		ret = rdi->driver_f.check_send_wqe(qp, wqe);
1573 		if (ret < 0)
1574 			goto bail_inval_free;
1575 		if (ret)
1576 			*call_send = ret;
1577 	}
1578 
1579 	log_pmtu = qp->log_pmtu;
1580 	if (qp->ibqp.qp_type != IB_QPT_UC &&
1581 	    qp->ibqp.qp_type != IB_QPT_RC) {
1582 		struct rvt_ah *ah = ibah_to_rvtah(wqe->ud_wr.ah);
1583 
1584 		log_pmtu = ah->log_pmtu;
1585 		atomic_inc(&ibah_to_rvtah(ud_wr(wr)->ah)->refcount);
1586 	}
1587 
1588 	wqe->ssn = qp->s_ssn++;
1589 	wqe->psn = qp->s_next_psn;
1590 	wqe->lpsn = wqe->psn +
1591 			(wqe->length ? ((wqe->length - 1) >> log_pmtu) : 0);
1592 	qp->s_next_psn = wqe->lpsn + 1;
1593 	trace_rvt_post_one_wr(qp, wqe);
1594 	smp_wmb(); /* see request builders */
1595 	qp->s_avail--;
1596 	qp->s_head = next;
1597 
1598 	return 0;
1599 
1600 bail_inval_free:
1601 	/* release mr holds */
1602 	while (j) {
1603 		struct rvt_sge *sge = &wqe->sg_list[--j];
1604 
1605 		rvt_put_mr(sge->mr);
1606 	}
1607 	return ret;
1608 }
1609 
1610 /**
1611  * rvt_post_send - post a send on a QP
1612  * @ibqp: the QP to post the send on
1613  * @wr: the list of work requests to post
1614  * @bad_wr: the first bad WR is put here
1615  *
1616  * This may be called from interrupt context.
1617  *
1618  * Return: 0 on success else errno
1619  */
1620 int rvt_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1621 		  struct ib_send_wr **bad_wr)
1622 {
1623 	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1624 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1625 	unsigned long flags = 0;
1626 	int call_send;
1627 	unsigned nreq = 0;
1628 	int err = 0;
1629 
1630 	spin_lock_irqsave(&qp->s_hlock, flags);
1631 
1632 	/*
1633 	 * Ensure QP state is such that we can send. If not bail out early,
1634 	 * there is no need to do this every time we post a send.
1635 	 */
1636 	if (unlikely(!(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))) {
1637 		spin_unlock_irqrestore(&qp->s_hlock, flags);
1638 		return -EINVAL;
1639 	}
1640 
1641 	/*
1642 	 * If the send queue is empty, and we only have a single WR then just go
1643 	 * ahead and kick the send engine into gear. Otherwise we will always
1644 	 * just schedule the send to happen later.
1645 	 */
1646 	call_send = qp->s_head == ACCESS_ONCE(qp->s_last) && !wr->next;
1647 
1648 	for (; wr; wr = wr->next) {
1649 		err = rvt_post_one_wr(qp, wr, &call_send);
1650 		if (unlikely(err)) {
1651 			*bad_wr = wr;
1652 			goto bail;
1653 		}
1654 		nreq++;
1655 	}
1656 bail:
1657 	spin_unlock_irqrestore(&qp->s_hlock, flags);
1658 	if (nreq) {
1659 		if (call_send)
1660 			rdi->driver_f.do_send(qp);
1661 		else
1662 			rdi->driver_f.schedule_send_no_lock(qp);
1663 	}
1664 	return err;
1665 }
1666 
1667 /**
1668  * rvt_post_srq_receive - post a receive on a shared receive queue
1669  * @ibsrq: the SRQ to post the receive on
1670  * @wr: the list of work requests to post
1671  * @bad_wr: A pointer to the first WR to cause a problem is put here
1672  *
1673  * This may be called from interrupt context.
1674  *
1675  * Return: 0 on success else errno
1676  */
1677 int rvt_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
1678 		      struct ib_recv_wr **bad_wr)
1679 {
1680 	struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq);
1681 	struct rvt_rwq *wq;
1682 	unsigned long flags;
1683 
1684 	for (; wr; wr = wr->next) {
1685 		struct rvt_rwqe *wqe;
1686 		u32 next;
1687 		int i;
1688 
1689 		if ((unsigned)wr->num_sge > srq->rq.max_sge) {
1690 			*bad_wr = wr;
1691 			return -EINVAL;
1692 		}
1693 
1694 		spin_lock_irqsave(&srq->rq.lock, flags);
1695 		wq = srq->rq.wq;
1696 		next = wq->head + 1;
1697 		if (next >= srq->rq.size)
1698 			next = 0;
1699 		if (next == wq->tail) {
1700 			spin_unlock_irqrestore(&srq->rq.lock, flags);
1701 			*bad_wr = wr;
1702 			return -ENOMEM;
1703 		}
1704 
1705 		wqe = rvt_get_rwqe_ptr(&srq->rq, wq->head);
1706 		wqe->wr_id = wr->wr_id;
1707 		wqe->num_sge = wr->num_sge;
1708 		for (i = 0; i < wr->num_sge; i++)
1709 			wqe->sg_list[i] = wr->sg_list[i];
1710 		/* Make sure queue entry is written before the head index. */
1711 		smp_wmb();
1712 		wq->head = next;
1713 		spin_unlock_irqrestore(&srq->rq.lock, flags);
1714 	}
1715 	return 0;
1716 }
1717