xref: /openbmc/linux/drivers/infiniband/hw/qib/qib_qp.c (revision d2999e1b)
1 /*
2  * Copyright (c) 2012, 2013 Intel Corporation.  All rights reserved.
3  * Copyright (c) 2006 - 2012 QLogic Corporation.  * All rights reserved.
4  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34 
35 #include <linux/err.h>
36 #include <linux/vmalloc.h>
37 #include <linux/jhash.h>
38 #ifdef CONFIG_DEBUG_FS
39 #include <linux/seq_file.h>
40 #endif
41 
42 #include "qib.h"
43 
44 #define BITS_PER_PAGE           (PAGE_SIZE*BITS_PER_BYTE)
45 #define BITS_PER_PAGE_MASK      (BITS_PER_PAGE-1)
46 
47 static inline unsigned mk_qpn(struct qib_qpn_table *qpt,
48 			      struct qpn_map *map, unsigned off)
49 {
50 	return (map - qpt->map) * BITS_PER_PAGE + off;
51 }
52 
53 static inline unsigned find_next_offset(struct qib_qpn_table *qpt,
54 					struct qpn_map *map, unsigned off,
55 					unsigned n)
56 {
57 	if (qpt->mask) {
58 		off++;
59 		if (((off & qpt->mask) >> 1) >= n)
60 			off = (off | qpt->mask) + 2;
61 	} else
62 		off = find_next_zero_bit(map->page, BITS_PER_PAGE, off);
63 	return off;
64 }
65 
66 /*
67  * Convert the AETH credit code into the number of credits.
68  */
69 static u32 credit_table[31] = {
70 	0,                      /* 0 */
71 	1,                      /* 1 */
72 	2,                      /* 2 */
73 	3,                      /* 3 */
74 	4,                      /* 4 */
75 	6,                      /* 5 */
76 	8,                      /* 6 */
77 	12,                     /* 7 */
78 	16,                     /* 8 */
79 	24,                     /* 9 */
80 	32,                     /* A */
81 	48,                     /* B */
82 	64,                     /* C */
83 	96,                     /* D */
84 	128,                    /* E */
85 	192,                    /* F */
86 	256,                    /* 10 */
87 	384,                    /* 11 */
88 	512,                    /* 12 */
89 	768,                    /* 13 */
90 	1024,                   /* 14 */
91 	1536,                   /* 15 */
92 	2048,                   /* 16 */
93 	3072,                   /* 17 */
94 	4096,                   /* 18 */
95 	6144,                   /* 19 */
96 	8192,                   /* 1A */
97 	12288,                  /* 1B */
98 	16384,                  /* 1C */
99 	24576,                  /* 1D */
100 	32768                   /* 1E */
101 };
102 
103 static void get_map_page(struct qib_qpn_table *qpt, struct qpn_map *map)
104 {
105 	unsigned long page = get_zeroed_page(GFP_KERNEL);
106 
107 	/*
108 	 * Free the page if someone raced with us installing it.
109 	 */
110 
111 	spin_lock(&qpt->lock);
112 	if (map->page)
113 		free_page(page);
114 	else
115 		map->page = (void *)page;
116 	spin_unlock(&qpt->lock);
117 }
118 
119 /*
120  * Allocate the next available QPN or
121  * zero/one for QP type IB_QPT_SMI/IB_QPT_GSI.
122  */
123 static int alloc_qpn(struct qib_devdata *dd, struct qib_qpn_table *qpt,
124 		     enum ib_qp_type type, u8 port)
125 {
126 	u32 i, offset, max_scan, qpn;
127 	struct qpn_map *map;
128 	u32 ret;
129 
130 	if (type == IB_QPT_SMI || type == IB_QPT_GSI) {
131 		unsigned n;
132 
133 		ret = type == IB_QPT_GSI;
134 		n = 1 << (ret + 2 * (port - 1));
135 		spin_lock(&qpt->lock);
136 		if (qpt->flags & n)
137 			ret = -EINVAL;
138 		else
139 			qpt->flags |= n;
140 		spin_unlock(&qpt->lock);
141 		goto bail;
142 	}
143 
144 	qpn = qpt->last + 2;
145 	if (qpn >= QPN_MAX)
146 		qpn = 2;
147 	if (qpt->mask && ((qpn & qpt->mask) >> 1) >= dd->n_krcv_queues)
148 		qpn = (qpn | qpt->mask) + 2;
149 	offset = qpn & BITS_PER_PAGE_MASK;
150 	map = &qpt->map[qpn / BITS_PER_PAGE];
151 	max_scan = qpt->nmaps - !offset;
152 	for (i = 0;;) {
153 		if (unlikely(!map->page)) {
154 			get_map_page(qpt, map);
155 			if (unlikely(!map->page))
156 				break;
157 		}
158 		do {
159 			if (!test_and_set_bit(offset, map->page)) {
160 				qpt->last = qpn;
161 				ret = qpn;
162 				goto bail;
163 			}
164 			offset = find_next_offset(qpt, map, offset,
165 				dd->n_krcv_queues);
166 			qpn = mk_qpn(qpt, map, offset);
167 			/*
168 			 * This test differs from alloc_pidmap().
169 			 * If find_next_offset() does find a zero
170 			 * bit, we don't need to check for QPN
171 			 * wrapping around past our starting QPN.
172 			 * We just need to be sure we don't loop
173 			 * forever.
174 			 */
175 		} while (offset < BITS_PER_PAGE && qpn < QPN_MAX);
176 		/*
177 		 * In order to keep the number of pages allocated to a
178 		 * minimum, we scan the all existing pages before increasing
179 		 * the size of the bitmap table.
180 		 */
181 		if (++i > max_scan) {
182 			if (qpt->nmaps == QPNMAP_ENTRIES)
183 				break;
184 			map = &qpt->map[qpt->nmaps++];
185 			offset = 0;
186 		} else if (map < &qpt->map[qpt->nmaps]) {
187 			++map;
188 			offset = 0;
189 		} else {
190 			map = &qpt->map[0];
191 			offset = 2;
192 		}
193 		qpn = mk_qpn(qpt, map, offset);
194 	}
195 
196 	ret = -ENOMEM;
197 
198 bail:
199 	return ret;
200 }
201 
202 static void free_qpn(struct qib_qpn_table *qpt, u32 qpn)
203 {
204 	struct qpn_map *map;
205 
206 	map = qpt->map + qpn / BITS_PER_PAGE;
207 	if (map->page)
208 		clear_bit(qpn & BITS_PER_PAGE_MASK, map->page);
209 }
210 
211 static inline unsigned qpn_hash(struct qib_ibdev *dev, u32 qpn)
212 {
213 	return jhash_1word(qpn, dev->qp_rnd) &
214 		(dev->qp_table_size - 1);
215 }
216 
217 
218 /*
219  * Put the QP into the hash table.
220  * The hash table holds a reference to the QP.
221  */
222 static void insert_qp(struct qib_ibdev *dev, struct qib_qp *qp)
223 {
224 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
225 	unsigned long flags;
226 	unsigned n = qpn_hash(dev, qp->ibqp.qp_num);
227 
228 	atomic_inc(&qp->refcount);
229 	spin_lock_irqsave(&dev->qpt_lock, flags);
230 
231 	if (qp->ibqp.qp_num == 0)
232 		rcu_assign_pointer(ibp->qp0, qp);
233 	else if (qp->ibqp.qp_num == 1)
234 		rcu_assign_pointer(ibp->qp1, qp);
235 	else {
236 		qp->next = dev->qp_table[n];
237 		rcu_assign_pointer(dev->qp_table[n], qp);
238 	}
239 
240 	spin_unlock_irqrestore(&dev->qpt_lock, flags);
241 }
242 
243 /*
244  * Remove the QP from the table so it can't be found asynchronously by
245  * the receive interrupt routine.
246  */
247 static void remove_qp(struct qib_ibdev *dev, struct qib_qp *qp)
248 {
249 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
250 	unsigned n = qpn_hash(dev, qp->ibqp.qp_num);
251 	unsigned long flags;
252 	int removed = 1;
253 
254 	spin_lock_irqsave(&dev->qpt_lock, flags);
255 
256 	if (rcu_dereference_protected(ibp->qp0,
257 			lockdep_is_held(&dev->qpt_lock)) == qp) {
258 		rcu_assign_pointer(ibp->qp0, NULL);
259 	} else if (rcu_dereference_protected(ibp->qp1,
260 			lockdep_is_held(&dev->qpt_lock)) == qp) {
261 		rcu_assign_pointer(ibp->qp1, NULL);
262 	} else {
263 		struct qib_qp *q;
264 		struct qib_qp __rcu **qpp;
265 
266 		removed = 0;
267 		qpp = &dev->qp_table[n];
268 		for (; (q = rcu_dereference_protected(*qpp,
269 				lockdep_is_held(&dev->qpt_lock))) != NULL;
270 				qpp = &q->next)
271 			if (q == qp) {
272 				rcu_assign_pointer(*qpp,
273 					rcu_dereference_protected(qp->next,
274 					 lockdep_is_held(&dev->qpt_lock)));
275 				removed = 1;
276 				break;
277 			}
278 	}
279 
280 	spin_unlock_irqrestore(&dev->qpt_lock, flags);
281 	if (removed) {
282 		synchronize_rcu();
283 		atomic_dec(&qp->refcount);
284 	}
285 }
286 
287 /**
288  * qib_free_all_qps - check for QPs still in use
289  * @qpt: the QP table to empty
290  *
291  * There should not be any QPs still in use.
292  * Free memory for table.
293  */
294 unsigned qib_free_all_qps(struct qib_devdata *dd)
295 {
296 	struct qib_ibdev *dev = &dd->verbs_dev;
297 	unsigned long flags;
298 	struct qib_qp *qp;
299 	unsigned n, qp_inuse = 0;
300 
301 	for (n = 0; n < dd->num_pports; n++) {
302 		struct qib_ibport *ibp = &dd->pport[n].ibport_data;
303 
304 		if (!qib_mcast_tree_empty(ibp))
305 			qp_inuse++;
306 		rcu_read_lock();
307 		if (rcu_dereference(ibp->qp0))
308 			qp_inuse++;
309 		if (rcu_dereference(ibp->qp1))
310 			qp_inuse++;
311 		rcu_read_unlock();
312 	}
313 
314 	spin_lock_irqsave(&dev->qpt_lock, flags);
315 	for (n = 0; n < dev->qp_table_size; n++) {
316 		qp = rcu_dereference_protected(dev->qp_table[n],
317 			lockdep_is_held(&dev->qpt_lock));
318 		rcu_assign_pointer(dev->qp_table[n], NULL);
319 
320 		for (; qp; qp = rcu_dereference_protected(qp->next,
321 					lockdep_is_held(&dev->qpt_lock)))
322 			qp_inuse++;
323 	}
324 	spin_unlock_irqrestore(&dev->qpt_lock, flags);
325 	synchronize_rcu();
326 
327 	return qp_inuse;
328 }
329 
330 /**
331  * qib_lookup_qpn - return the QP with the given QPN
332  * @qpt: the QP table
333  * @qpn: the QP number to look up
334  *
335  * The caller is responsible for decrementing the QP reference count
336  * when done.
337  */
338 struct qib_qp *qib_lookup_qpn(struct qib_ibport *ibp, u32 qpn)
339 {
340 	struct qib_qp *qp = NULL;
341 
342 	rcu_read_lock();
343 	if (unlikely(qpn <= 1)) {
344 		if (qpn == 0)
345 			qp = rcu_dereference(ibp->qp0);
346 		else
347 			qp = rcu_dereference(ibp->qp1);
348 		if (qp)
349 			atomic_inc(&qp->refcount);
350 	} else {
351 		struct qib_ibdev *dev = &ppd_from_ibp(ibp)->dd->verbs_dev;
352 		unsigned n = qpn_hash(dev, qpn);
353 
354 		for (qp = rcu_dereference(dev->qp_table[n]); qp;
355 			qp = rcu_dereference(qp->next))
356 			if (qp->ibqp.qp_num == qpn) {
357 				atomic_inc(&qp->refcount);
358 				break;
359 			}
360 	}
361 	rcu_read_unlock();
362 	return qp;
363 }
364 
365 /**
366  * qib_reset_qp - initialize the QP state to the reset state
367  * @qp: the QP to reset
368  * @type: the QP type
369  */
370 static void qib_reset_qp(struct qib_qp *qp, enum ib_qp_type type)
371 {
372 	qp->remote_qpn = 0;
373 	qp->qkey = 0;
374 	qp->qp_access_flags = 0;
375 	atomic_set(&qp->s_dma_busy, 0);
376 	qp->s_flags &= QIB_S_SIGNAL_REQ_WR;
377 	qp->s_hdrwords = 0;
378 	qp->s_wqe = NULL;
379 	qp->s_draining = 0;
380 	qp->s_next_psn = 0;
381 	qp->s_last_psn = 0;
382 	qp->s_sending_psn = 0;
383 	qp->s_sending_hpsn = 0;
384 	qp->s_psn = 0;
385 	qp->r_psn = 0;
386 	qp->r_msn = 0;
387 	if (type == IB_QPT_RC) {
388 		qp->s_state = IB_OPCODE_RC_SEND_LAST;
389 		qp->r_state = IB_OPCODE_RC_SEND_LAST;
390 	} else {
391 		qp->s_state = IB_OPCODE_UC_SEND_LAST;
392 		qp->r_state = IB_OPCODE_UC_SEND_LAST;
393 	}
394 	qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
395 	qp->r_nak_state = 0;
396 	qp->r_aflags = 0;
397 	qp->r_flags = 0;
398 	qp->s_head = 0;
399 	qp->s_tail = 0;
400 	qp->s_cur = 0;
401 	qp->s_acked = 0;
402 	qp->s_last = 0;
403 	qp->s_ssn = 1;
404 	qp->s_lsn = 0;
405 	qp->s_mig_state = IB_MIG_MIGRATED;
406 	memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue));
407 	qp->r_head_ack_queue = 0;
408 	qp->s_tail_ack_queue = 0;
409 	qp->s_num_rd_atomic = 0;
410 	if (qp->r_rq.wq) {
411 		qp->r_rq.wq->head = 0;
412 		qp->r_rq.wq->tail = 0;
413 	}
414 	qp->r_sge.num_sge = 0;
415 }
416 
417 static void clear_mr_refs(struct qib_qp *qp, int clr_sends)
418 {
419 	unsigned n;
420 
421 	if (test_and_clear_bit(QIB_R_REWIND_SGE, &qp->r_aflags))
422 		qib_put_ss(&qp->s_rdma_read_sge);
423 
424 	qib_put_ss(&qp->r_sge);
425 
426 	if (clr_sends) {
427 		while (qp->s_last != qp->s_head) {
428 			struct qib_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
429 			unsigned i;
430 
431 			for (i = 0; i < wqe->wr.num_sge; i++) {
432 				struct qib_sge *sge = &wqe->sg_list[i];
433 
434 				qib_put_mr(sge->mr);
435 			}
436 			if (qp->ibqp.qp_type == IB_QPT_UD ||
437 			    qp->ibqp.qp_type == IB_QPT_SMI ||
438 			    qp->ibqp.qp_type == IB_QPT_GSI)
439 				atomic_dec(&to_iah(wqe->wr.wr.ud.ah)->refcount);
440 			if (++qp->s_last >= qp->s_size)
441 				qp->s_last = 0;
442 		}
443 		if (qp->s_rdma_mr) {
444 			qib_put_mr(qp->s_rdma_mr);
445 			qp->s_rdma_mr = NULL;
446 		}
447 	}
448 
449 	if (qp->ibqp.qp_type != IB_QPT_RC)
450 		return;
451 
452 	for (n = 0; n < ARRAY_SIZE(qp->s_ack_queue); n++) {
453 		struct qib_ack_entry *e = &qp->s_ack_queue[n];
454 
455 		if (e->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST &&
456 		    e->rdma_sge.mr) {
457 			qib_put_mr(e->rdma_sge.mr);
458 			e->rdma_sge.mr = NULL;
459 		}
460 	}
461 }
462 
463 /**
464  * qib_error_qp - put a QP into the error state
465  * @qp: the QP to put into the error state
466  * @err: the receive completion error to signal if a RWQE is active
467  *
468  * Flushes both send and receive work queues.
469  * Returns true if last WQE event should be generated.
470  * The QP r_lock and s_lock should be held and interrupts disabled.
471  * If we are already in error state, just return.
472  */
473 int qib_error_qp(struct qib_qp *qp, enum ib_wc_status err)
474 {
475 	struct qib_ibdev *dev = to_idev(qp->ibqp.device);
476 	struct ib_wc wc;
477 	int ret = 0;
478 
479 	if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET)
480 		goto bail;
481 
482 	qp->state = IB_QPS_ERR;
483 
484 	if (qp->s_flags & (QIB_S_TIMER | QIB_S_WAIT_RNR)) {
485 		qp->s_flags &= ~(QIB_S_TIMER | QIB_S_WAIT_RNR);
486 		del_timer(&qp->s_timer);
487 	}
488 
489 	if (qp->s_flags & QIB_S_ANY_WAIT_SEND)
490 		qp->s_flags &= ~QIB_S_ANY_WAIT_SEND;
491 
492 	spin_lock(&dev->pending_lock);
493 	if (!list_empty(&qp->iowait) && !(qp->s_flags & QIB_S_BUSY)) {
494 		qp->s_flags &= ~QIB_S_ANY_WAIT_IO;
495 		list_del_init(&qp->iowait);
496 	}
497 	spin_unlock(&dev->pending_lock);
498 
499 	if (!(qp->s_flags & QIB_S_BUSY)) {
500 		qp->s_hdrwords = 0;
501 		if (qp->s_rdma_mr) {
502 			qib_put_mr(qp->s_rdma_mr);
503 			qp->s_rdma_mr = NULL;
504 		}
505 		if (qp->s_tx) {
506 			qib_put_txreq(qp->s_tx);
507 			qp->s_tx = NULL;
508 		}
509 	}
510 
511 	/* Schedule the sending tasklet to drain the send work queue. */
512 	if (qp->s_last != qp->s_head)
513 		qib_schedule_send(qp);
514 
515 	clear_mr_refs(qp, 0);
516 
517 	memset(&wc, 0, sizeof(wc));
518 	wc.qp = &qp->ibqp;
519 	wc.opcode = IB_WC_RECV;
520 
521 	if (test_and_clear_bit(QIB_R_WRID_VALID, &qp->r_aflags)) {
522 		wc.wr_id = qp->r_wr_id;
523 		wc.status = err;
524 		qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
525 	}
526 	wc.status = IB_WC_WR_FLUSH_ERR;
527 
528 	if (qp->r_rq.wq) {
529 		struct qib_rwq *wq;
530 		u32 head;
531 		u32 tail;
532 
533 		spin_lock(&qp->r_rq.lock);
534 
535 		/* sanity check pointers before trusting them */
536 		wq = qp->r_rq.wq;
537 		head = wq->head;
538 		if (head >= qp->r_rq.size)
539 			head = 0;
540 		tail = wq->tail;
541 		if (tail >= qp->r_rq.size)
542 			tail = 0;
543 		while (tail != head) {
544 			wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
545 			if (++tail >= qp->r_rq.size)
546 				tail = 0;
547 			qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
548 		}
549 		wq->tail = tail;
550 
551 		spin_unlock(&qp->r_rq.lock);
552 	} else if (qp->ibqp.event_handler)
553 		ret = 1;
554 
555 bail:
556 	return ret;
557 }
558 
559 /**
560  * qib_modify_qp - modify the attributes of a queue pair
561  * @ibqp: the queue pair who's attributes we're modifying
562  * @attr: the new attributes
563  * @attr_mask: the mask of attributes to modify
564  * @udata: user data for libibverbs.so
565  *
566  * Returns 0 on success, otherwise returns an errno.
567  */
568 int qib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
569 		  int attr_mask, struct ib_udata *udata)
570 {
571 	struct qib_ibdev *dev = to_idev(ibqp->device);
572 	struct qib_qp *qp = to_iqp(ibqp);
573 	enum ib_qp_state cur_state, new_state;
574 	struct ib_event ev;
575 	int lastwqe = 0;
576 	int mig = 0;
577 	int ret;
578 	u32 pmtu = 0; /* for gcc warning only */
579 
580 	spin_lock_irq(&qp->r_lock);
581 	spin_lock(&qp->s_lock);
582 
583 	cur_state = attr_mask & IB_QP_CUR_STATE ?
584 		attr->cur_qp_state : qp->state;
585 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
586 
587 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
588 				attr_mask, IB_LINK_LAYER_UNSPECIFIED))
589 		goto inval;
590 
591 	if (attr_mask & IB_QP_AV) {
592 		if (attr->ah_attr.dlid >= QIB_MULTICAST_LID_BASE)
593 			goto inval;
594 		if (qib_check_ah(qp->ibqp.device, &attr->ah_attr))
595 			goto inval;
596 	}
597 
598 	if (attr_mask & IB_QP_ALT_PATH) {
599 		if (attr->alt_ah_attr.dlid >= QIB_MULTICAST_LID_BASE)
600 			goto inval;
601 		if (qib_check_ah(qp->ibqp.device, &attr->alt_ah_attr))
602 			goto inval;
603 		if (attr->alt_pkey_index >= qib_get_npkeys(dd_from_dev(dev)))
604 			goto inval;
605 	}
606 
607 	if (attr_mask & IB_QP_PKEY_INDEX)
608 		if (attr->pkey_index >= qib_get_npkeys(dd_from_dev(dev)))
609 			goto inval;
610 
611 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
612 		if (attr->min_rnr_timer > 31)
613 			goto inval;
614 
615 	if (attr_mask & IB_QP_PORT)
616 		if (qp->ibqp.qp_type == IB_QPT_SMI ||
617 		    qp->ibqp.qp_type == IB_QPT_GSI ||
618 		    attr->port_num == 0 ||
619 		    attr->port_num > ibqp->device->phys_port_cnt)
620 			goto inval;
621 
622 	if (attr_mask & IB_QP_DEST_QPN)
623 		if (attr->dest_qp_num > QIB_QPN_MASK)
624 			goto inval;
625 
626 	if (attr_mask & IB_QP_RETRY_CNT)
627 		if (attr->retry_cnt > 7)
628 			goto inval;
629 
630 	if (attr_mask & IB_QP_RNR_RETRY)
631 		if (attr->rnr_retry > 7)
632 			goto inval;
633 
634 	/*
635 	 * Don't allow invalid path_mtu values.  OK to set greater
636 	 * than the active mtu (or even the max_cap, if we have tuned
637 	 * that to a small mtu.  We'll set qp->path_mtu
638 	 * to the lesser of requested attribute mtu and active,
639 	 * for packetizing messages.
640 	 * Note that the QP port has to be set in INIT and MTU in RTR.
641 	 */
642 	if (attr_mask & IB_QP_PATH_MTU) {
643 		struct qib_devdata *dd = dd_from_dev(dev);
644 		int mtu, pidx = qp->port_num - 1;
645 
646 		mtu = ib_mtu_enum_to_int(attr->path_mtu);
647 		if (mtu == -1)
648 			goto inval;
649 		if (mtu > dd->pport[pidx].ibmtu) {
650 			switch (dd->pport[pidx].ibmtu) {
651 			case 4096:
652 				pmtu = IB_MTU_4096;
653 				break;
654 			case 2048:
655 				pmtu = IB_MTU_2048;
656 				break;
657 			case 1024:
658 				pmtu = IB_MTU_1024;
659 				break;
660 			case 512:
661 				pmtu = IB_MTU_512;
662 				break;
663 			case 256:
664 				pmtu = IB_MTU_256;
665 				break;
666 			default:
667 				pmtu = IB_MTU_2048;
668 			}
669 		} else
670 			pmtu = attr->path_mtu;
671 	}
672 
673 	if (attr_mask & IB_QP_PATH_MIG_STATE) {
674 		if (attr->path_mig_state == IB_MIG_REARM) {
675 			if (qp->s_mig_state == IB_MIG_ARMED)
676 				goto inval;
677 			if (new_state != IB_QPS_RTS)
678 				goto inval;
679 		} else if (attr->path_mig_state == IB_MIG_MIGRATED) {
680 			if (qp->s_mig_state == IB_MIG_REARM)
681 				goto inval;
682 			if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD)
683 				goto inval;
684 			if (qp->s_mig_state == IB_MIG_ARMED)
685 				mig = 1;
686 		} else
687 			goto inval;
688 	}
689 
690 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
691 		if (attr->max_dest_rd_atomic > QIB_MAX_RDMA_ATOMIC)
692 			goto inval;
693 
694 	switch (new_state) {
695 	case IB_QPS_RESET:
696 		if (qp->state != IB_QPS_RESET) {
697 			qp->state = IB_QPS_RESET;
698 			spin_lock(&dev->pending_lock);
699 			if (!list_empty(&qp->iowait))
700 				list_del_init(&qp->iowait);
701 			spin_unlock(&dev->pending_lock);
702 			qp->s_flags &= ~(QIB_S_TIMER | QIB_S_ANY_WAIT);
703 			spin_unlock(&qp->s_lock);
704 			spin_unlock_irq(&qp->r_lock);
705 			/* Stop the sending work queue and retry timer */
706 			cancel_work_sync(&qp->s_work);
707 			del_timer_sync(&qp->s_timer);
708 			wait_event(qp->wait_dma, !atomic_read(&qp->s_dma_busy));
709 			if (qp->s_tx) {
710 				qib_put_txreq(qp->s_tx);
711 				qp->s_tx = NULL;
712 			}
713 			remove_qp(dev, qp);
714 			wait_event(qp->wait, !atomic_read(&qp->refcount));
715 			spin_lock_irq(&qp->r_lock);
716 			spin_lock(&qp->s_lock);
717 			clear_mr_refs(qp, 1);
718 			qib_reset_qp(qp, ibqp->qp_type);
719 		}
720 		break;
721 
722 	case IB_QPS_RTR:
723 		/* Allow event to retrigger if QP set to RTR more than once */
724 		qp->r_flags &= ~QIB_R_COMM_EST;
725 		qp->state = new_state;
726 		break;
727 
728 	case IB_QPS_SQD:
729 		qp->s_draining = qp->s_last != qp->s_cur;
730 		qp->state = new_state;
731 		break;
732 
733 	case IB_QPS_SQE:
734 		if (qp->ibqp.qp_type == IB_QPT_RC)
735 			goto inval;
736 		qp->state = new_state;
737 		break;
738 
739 	case IB_QPS_ERR:
740 		lastwqe = qib_error_qp(qp, IB_WC_WR_FLUSH_ERR);
741 		break;
742 
743 	default:
744 		qp->state = new_state;
745 		break;
746 	}
747 
748 	if (attr_mask & IB_QP_PKEY_INDEX)
749 		qp->s_pkey_index = attr->pkey_index;
750 
751 	if (attr_mask & IB_QP_PORT)
752 		qp->port_num = attr->port_num;
753 
754 	if (attr_mask & IB_QP_DEST_QPN)
755 		qp->remote_qpn = attr->dest_qp_num;
756 
757 	if (attr_mask & IB_QP_SQ_PSN) {
758 		qp->s_next_psn = attr->sq_psn & QIB_PSN_MASK;
759 		qp->s_psn = qp->s_next_psn;
760 		qp->s_sending_psn = qp->s_next_psn;
761 		qp->s_last_psn = qp->s_next_psn - 1;
762 		qp->s_sending_hpsn = qp->s_last_psn;
763 	}
764 
765 	if (attr_mask & IB_QP_RQ_PSN)
766 		qp->r_psn = attr->rq_psn & QIB_PSN_MASK;
767 
768 	if (attr_mask & IB_QP_ACCESS_FLAGS)
769 		qp->qp_access_flags = attr->qp_access_flags;
770 
771 	if (attr_mask & IB_QP_AV) {
772 		qp->remote_ah_attr = attr->ah_attr;
773 		qp->s_srate = attr->ah_attr.static_rate;
774 	}
775 
776 	if (attr_mask & IB_QP_ALT_PATH) {
777 		qp->alt_ah_attr = attr->alt_ah_attr;
778 		qp->s_alt_pkey_index = attr->alt_pkey_index;
779 	}
780 
781 	if (attr_mask & IB_QP_PATH_MIG_STATE) {
782 		qp->s_mig_state = attr->path_mig_state;
783 		if (mig) {
784 			qp->remote_ah_attr = qp->alt_ah_attr;
785 			qp->port_num = qp->alt_ah_attr.port_num;
786 			qp->s_pkey_index = qp->s_alt_pkey_index;
787 		}
788 	}
789 
790 	if (attr_mask & IB_QP_PATH_MTU) {
791 		qp->path_mtu = pmtu;
792 		qp->pmtu = ib_mtu_enum_to_int(pmtu);
793 	}
794 
795 	if (attr_mask & IB_QP_RETRY_CNT) {
796 		qp->s_retry_cnt = attr->retry_cnt;
797 		qp->s_retry = attr->retry_cnt;
798 	}
799 
800 	if (attr_mask & IB_QP_RNR_RETRY) {
801 		qp->s_rnr_retry_cnt = attr->rnr_retry;
802 		qp->s_rnr_retry = attr->rnr_retry;
803 	}
804 
805 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
806 		qp->r_min_rnr_timer = attr->min_rnr_timer;
807 
808 	if (attr_mask & IB_QP_TIMEOUT) {
809 		qp->timeout = attr->timeout;
810 		qp->timeout_jiffies =
811 			usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
812 				1000UL);
813 	}
814 
815 	if (attr_mask & IB_QP_QKEY)
816 		qp->qkey = attr->qkey;
817 
818 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
819 		qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
820 
821 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
822 		qp->s_max_rd_atomic = attr->max_rd_atomic;
823 
824 	spin_unlock(&qp->s_lock);
825 	spin_unlock_irq(&qp->r_lock);
826 
827 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
828 		insert_qp(dev, qp);
829 
830 	if (lastwqe) {
831 		ev.device = qp->ibqp.device;
832 		ev.element.qp = &qp->ibqp;
833 		ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
834 		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
835 	}
836 	if (mig) {
837 		ev.device = qp->ibqp.device;
838 		ev.element.qp = &qp->ibqp;
839 		ev.event = IB_EVENT_PATH_MIG;
840 		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
841 	}
842 	ret = 0;
843 	goto bail;
844 
845 inval:
846 	spin_unlock(&qp->s_lock);
847 	spin_unlock_irq(&qp->r_lock);
848 	ret = -EINVAL;
849 
850 bail:
851 	return ret;
852 }
853 
854 int qib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
855 		 int attr_mask, struct ib_qp_init_attr *init_attr)
856 {
857 	struct qib_qp *qp = to_iqp(ibqp);
858 
859 	attr->qp_state = qp->state;
860 	attr->cur_qp_state = attr->qp_state;
861 	attr->path_mtu = qp->path_mtu;
862 	attr->path_mig_state = qp->s_mig_state;
863 	attr->qkey = qp->qkey;
864 	attr->rq_psn = qp->r_psn & QIB_PSN_MASK;
865 	attr->sq_psn = qp->s_next_psn & QIB_PSN_MASK;
866 	attr->dest_qp_num = qp->remote_qpn;
867 	attr->qp_access_flags = qp->qp_access_flags;
868 	attr->cap.max_send_wr = qp->s_size - 1;
869 	attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
870 	attr->cap.max_send_sge = qp->s_max_sge;
871 	attr->cap.max_recv_sge = qp->r_rq.max_sge;
872 	attr->cap.max_inline_data = 0;
873 	attr->ah_attr = qp->remote_ah_attr;
874 	attr->alt_ah_attr = qp->alt_ah_attr;
875 	attr->pkey_index = qp->s_pkey_index;
876 	attr->alt_pkey_index = qp->s_alt_pkey_index;
877 	attr->en_sqd_async_notify = 0;
878 	attr->sq_draining = qp->s_draining;
879 	attr->max_rd_atomic = qp->s_max_rd_atomic;
880 	attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
881 	attr->min_rnr_timer = qp->r_min_rnr_timer;
882 	attr->port_num = qp->port_num;
883 	attr->timeout = qp->timeout;
884 	attr->retry_cnt = qp->s_retry_cnt;
885 	attr->rnr_retry = qp->s_rnr_retry_cnt;
886 	attr->alt_port_num = qp->alt_ah_attr.port_num;
887 	attr->alt_timeout = qp->alt_timeout;
888 
889 	init_attr->event_handler = qp->ibqp.event_handler;
890 	init_attr->qp_context = qp->ibqp.qp_context;
891 	init_attr->send_cq = qp->ibqp.send_cq;
892 	init_attr->recv_cq = qp->ibqp.recv_cq;
893 	init_attr->srq = qp->ibqp.srq;
894 	init_attr->cap = attr->cap;
895 	if (qp->s_flags & QIB_S_SIGNAL_REQ_WR)
896 		init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
897 	else
898 		init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
899 	init_attr->qp_type = qp->ibqp.qp_type;
900 	init_attr->port_num = qp->port_num;
901 	return 0;
902 }
903 
904 /**
905  * qib_compute_aeth - compute the AETH (syndrome + MSN)
906  * @qp: the queue pair to compute the AETH for
907  *
908  * Returns the AETH.
909  */
910 __be32 qib_compute_aeth(struct qib_qp *qp)
911 {
912 	u32 aeth = qp->r_msn & QIB_MSN_MASK;
913 
914 	if (qp->ibqp.srq) {
915 		/*
916 		 * Shared receive queues don't generate credits.
917 		 * Set the credit field to the invalid value.
918 		 */
919 		aeth |= QIB_AETH_CREDIT_INVAL << QIB_AETH_CREDIT_SHIFT;
920 	} else {
921 		u32 min, max, x;
922 		u32 credits;
923 		struct qib_rwq *wq = qp->r_rq.wq;
924 		u32 head;
925 		u32 tail;
926 
927 		/* sanity check pointers before trusting them */
928 		head = wq->head;
929 		if (head >= qp->r_rq.size)
930 			head = 0;
931 		tail = wq->tail;
932 		if (tail >= qp->r_rq.size)
933 			tail = 0;
934 		/*
935 		 * Compute the number of credits available (RWQEs).
936 		 * XXX Not holding the r_rq.lock here so there is a small
937 		 * chance that the pair of reads are not atomic.
938 		 */
939 		credits = head - tail;
940 		if ((int)credits < 0)
941 			credits += qp->r_rq.size;
942 		/*
943 		 * Binary search the credit table to find the code to
944 		 * use.
945 		 */
946 		min = 0;
947 		max = 31;
948 		for (;;) {
949 			x = (min + max) / 2;
950 			if (credit_table[x] == credits)
951 				break;
952 			if (credit_table[x] > credits)
953 				max = x;
954 			else if (min == x)
955 				break;
956 			else
957 				min = x;
958 		}
959 		aeth |= x << QIB_AETH_CREDIT_SHIFT;
960 	}
961 	return cpu_to_be32(aeth);
962 }
963 
964 /**
965  * qib_create_qp - create a queue pair for a device
966  * @ibpd: the protection domain who's device we create the queue pair for
967  * @init_attr: the attributes of the queue pair
968  * @udata: user data for libibverbs.so
969  *
970  * Returns the queue pair on success, otherwise returns an errno.
971  *
972  * Called by the ib_create_qp() core verbs function.
973  */
974 struct ib_qp *qib_create_qp(struct ib_pd *ibpd,
975 			    struct ib_qp_init_attr *init_attr,
976 			    struct ib_udata *udata)
977 {
978 	struct qib_qp *qp;
979 	int err;
980 	struct qib_swqe *swq = NULL;
981 	struct qib_ibdev *dev;
982 	struct qib_devdata *dd;
983 	size_t sz;
984 	size_t sg_list_sz;
985 	struct ib_qp *ret;
986 
987 	if (init_attr->cap.max_send_sge > ib_qib_max_sges ||
988 	    init_attr->cap.max_send_wr > ib_qib_max_qp_wrs ||
989 	    init_attr->create_flags) {
990 		ret = ERR_PTR(-EINVAL);
991 		goto bail;
992 	}
993 
994 	/* Check receive queue parameters if no SRQ is specified. */
995 	if (!init_attr->srq) {
996 		if (init_attr->cap.max_recv_sge > ib_qib_max_sges ||
997 		    init_attr->cap.max_recv_wr > ib_qib_max_qp_wrs) {
998 			ret = ERR_PTR(-EINVAL);
999 			goto bail;
1000 		}
1001 		if (init_attr->cap.max_send_sge +
1002 		    init_attr->cap.max_send_wr +
1003 		    init_attr->cap.max_recv_sge +
1004 		    init_attr->cap.max_recv_wr == 0) {
1005 			ret = ERR_PTR(-EINVAL);
1006 			goto bail;
1007 		}
1008 	}
1009 
1010 	switch (init_attr->qp_type) {
1011 	case IB_QPT_SMI:
1012 	case IB_QPT_GSI:
1013 		if (init_attr->port_num == 0 ||
1014 		    init_attr->port_num > ibpd->device->phys_port_cnt) {
1015 			ret = ERR_PTR(-EINVAL);
1016 			goto bail;
1017 		}
1018 	case IB_QPT_UC:
1019 	case IB_QPT_RC:
1020 	case IB_QPT_UD:
1021 		sz = sizeof(struct qib_sge) *
1022 			init_attr->cap.max_send_sge +
1023 			sizeof(struct qib_swqe);
1024 		swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz);
1025 		if (swq == NULL) {
1026 			ret = ERR_PTR(-ENOMEM);
1027 			goto bail;
1028 		}
1029 		sz = sizeof(*qp);
1030 		sg_list_sz = 0;
1031 		if (init_attr->srq) {
1032 			struct qib_srq *srq = to_isrq(init_attr->srq);
1033 
1034 			if (srq->rq.max_sge > 1)
1035 				sg_list_sz = sizeof(*qp->r_sg_list) *
1036 					(srq->rq.max_sge - 1);
1037 		} else if (init_attr->cap.max_recv_sge > 1)
1038 			sg_list_sz = sizeof(*qp->r_sg_list) *
1039 				(init_attr->cap.max_recv_sge - 1);
1040 		qp = kzalloc(sz + sg_list_sz, GFP_KERNEL);
1041 		if (!qp) {
1042 			ret = ERR_PTR(-ENOMEM);
1043 			goto bail_swq;
1044 		}
1045 		RCU_INIT_POINTER(qp->next, NULL);
1046 		qp->s_hdr = kzalloc(sizeof(*qp->s_hdr), GFP_KERNEL);
1047 		if (!qp->s_hdr) {
1048 			ret = ERR_PTR(-ENOMEM);
1049 			goto bail_qp;
1050 		}
1051 		qp->timeout_jiffies =
1052 			usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
1053 				1000UL);
1054 		if (init_attr->srq)
1055 			sz = 0;
1056 		else {
1057 			qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
1058 			qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
1059 			sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
1060 				sizeof(struct qib_rwqe);
1061 			qp->r_rq.wq = vmalloc_user(sizeof(struct qib_rwq) +
1062 						   qp->r_rq.size * sz);
1063 			if (!qp->r_rq.wq) {
1064 				ret = ERR_PTR(-ENOMEM);
1065 				goto bail_qp;
1066 			}
1067 		}
1068 
1069 		/*
1070 		 * ib_create_qp() will initialize qp->ibqp
1071 		 * except for qp->ibqp.qp_num.
1072 		 */
1073 		spin_lock_init(&qp->r_lock);
1074 		spin_lock_init(&qp->s_lock);
1075 		spin_lock_init(&qp->r_rq.lock);
1076 		atomic_set(&qp->refcount, 0);
1077 		init_waitqueue_head(&qp->wait);
1078 		init_waitqueue_head(&qp->wait_dma);
1079 		init_timer(&qp->s_timer);
1080 		qp->s_timer.data = (unsigned long)qp;
1081 		INIT_WORK(&qp->s_work, qib_do_send);
1082 		INIT_LIST_HEAD(&qp->iowait);
1083 		INIT_LIST_HEAD(&qp->rspwait);
1084 		qp->state = IB_QPS_RESET;
1085 		qp->s_wq = swq;
1086 		qp->s_size = init_attr->cap.max_send_wr + 1;
1087 		qp->s_max_sge = init_attr->cap.max_send_sge;
1088 		if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
1089 			qp->s_flags = QIB_S_SIGNAL_REQ_WR;
1090 		dev = to_idev(ibpd->device);
1091 		dd = dd_from_dev(dev);
1092 		err = alloc_qpn(dd, &dev->qpn_table, init_attr->qp_type,
1093 				init_attr->port_num);
1094 		if (err < 0) {
1095 			ret = ERR_PTR(err);
1096 			vfree(qp->r_rq.wq);
1097 			goto bail_qp;
1098 		}
1099 		qp->ibqp.qp_num = err;
1100 		qp->port_num = init_attr->port_num;
1101 		qib_reset_qp(qp, init_attr->qp_type);
1102 		break;
1103 
1104 	default:
1105 		/* Don't support raw QPs */
1106 		ret = ERR_PTR(-ENOSYS);
1107 		goto bail;
1108 	}
1109 
1110 	init_attr->cap.max_inline_data = 0;
1111 
1112 	/*
1113 	 * Return the address of the RWQ as the offset to mmap.
1114 	 * See qib_mmap() for details.
1115 	 */
1116 	if (udata && udata->outlen >= sizeof(__u64)) {
1117 		if (!qp->r_rq.wq) {
1118 			__u64 offset = 0;
1119 
1120 			err = ib_copy_to_udata(udata, &offset,
1121 					       sizeof(offset));
1122 			if (err) {
1123 				ret = ERR_PTR(err);
1124 				goto bail_ip;
1125 			}
1126 		} else {
1127 			u32 s = sizeof(struct qib_rwq) + qp->r_rq.size * sz;
1128 
1129 			qp->ip = qib_create_mmap_info(dev, s,
1130 						      ibpd->uobject->context,
1131 						      qp->r_rq.wq);
1132 			if (!qp->ip) {
1133 				ret = ERR_PTR(-ENOMEM);
1134 				goto bail_ip;
1135 			}
1136 
1137 			err = ib_copy_to_udata(udata, &(qp->ip->offset),
1138 					       sizeof(qp->ip->offset));
1139 			if (err) {
1140 				ret = ERR_PTR(err);
1141 				goto bail_ip;
1142 			}
1143 		}
1144 	}
1145 
1146 	spin_lock(&dev->n_qps_lock);
1147 	if (dev->n_qps_allocated == ib_qib_max_qps) {
1148 		spin_unlock(&dev->n_qps_lock);
1149 		ret = ERR_PTR(-ENOMEM);
1150 		goto bail_ip;
1151 	}
1152 
1153 	dev->n_qps_allocated++;
1154 	spin_unlock(&dev->n_qps_lock);
1155 
1156 	if (qp->ip) {
1157 		spin_lock_irq(&dev->pending_lock);
1158 		list_add(&qp->ip->pending_mmaps, &dev->pending_mmaps);
1159 		spin_unlock_irq(&dev->pending_lock);
1160 	}
1161 
1162 	ret = &qp->ibqp;
1163 	goto bail;
1164 
1165 bail_ip:
1166 	if (qp->ip)
1167 		kref_put(&qp->ip->ref, qib_release_mmap_info);
1168 	else
1169 		vfree(qp->r_rq.wq);
1170 	free_qpn(&dev->qpn_table, qp->ibqp.qp_num);
1171 bail_qp:
1172 	kfree(qp->s_hdr);
1173 	kfree(qp);
1174 bail_swq:
1175 	vfree(swq);
1176 bail:
1177 	return ret;
1178 }
1179 
1180 /**
1181  * qib_destroy_qp - destroy a queue pair
1182  * @ibqp: the queue pair to destroy
1183  *
1184  * Returns 0 on success.
1185  *
1186  * Note that this can be called while the QP is actively sending or
1187  * receiving!
1188  */
1189 int qib_destroy_qp(struct ib_qp *ibqp)
1190 {
1191 	struct qib_qp *qp = to_iqp(ibqp);
1192 	struct qib_ibdev *dev = to_idev(ibqp->device);
1193 
1194 	/* Make sure HW and driver activity is stopped. */
1195 	spin_lock_irq(&qp->s_lock);
1196 	if (qp->state != IB_QPS_RESET) {
1197 		qp->state = IB_QPS_RESET;
1198 		spin_lock(&dev->pending_lock);
1199 		if (!list_empty(&qp->iowait))
1200 			list_del_init(&qp->iowait);
1201 		spin_unlock(&dev->pending_lock);
1202 		qp->s_flags &= ~(QIB_S_TIMER | QIB_S_ANY_WAIT);
1203 		spin_unlock_irq(&qp->s_lock);
1204 		cancel_work_sync(&qp->s_work);
1205 		del_timer_sync(&qp->s_timer);
1206 		wait_event(qp->wait_dma, !atomic_read(&qp->s_dma_busy));
1207 		if (qp->s_tx) {
1208 			qib_put_txreq(qp->s_tx);
1209 			qp->s_tx = NULL;
1210 		}
1211 		remove_qp(dev, qp);
1212 		wait_event(qp->wait, !atomic_read(&qp->refcount));
1213 		clear_mr_refs(qp, 1);
1214 	} else
1215 		spin_unlock_irq(&qp->s_lock);
1216 
1217 	/* all user's cleaned up, mark it available */
1218 	free_qpn(&dev->qpn_table, qp->ibqp.qp_num);
1219 	spin_lock(&dev->n_qps_lock);
1220 	dev->n_qps_allocated--;
1221 	spin_unlock(&dev->n_qps_lock);
1222 
1223 	if (qp->ip)
1224 		kref_put(&qp->ip->ref, qib_release_mmap_info);
1225 	else
1226 		vfree(qp->r_rq.wq);
1227 	vfree(qp->s_wq);
1228 	kfree(qp->s_hdr);
1229 	kfree(qp);
1230 	return 0;
1231 }
1232 
1233 /**
1234  * qib_init_qpn_table - initialize the QP number table for a device
1235  * @qpt: the QPN table
1236  */
1237 void qib_init_qpn_table(struct qib_devdata *dd, struct qib_qpn_table *qpt)
1238 {
1239 	spin_lock_init(&qpt->lock);
1240 	qpt->last = 1;          /* start with QPN 2 */
1241 	qpt->nmaps = 1;
1242 	qpt->mask = dd->qpn_mask;
1243 }
1244 
1245 /**
1246  * qib_free_qpn_table - free the QP number table for a device
1247  * @qpt: the QPN table
1248  */
1249 void qib_free_qpn_table(struct qib_qpn_table *qpt)
1250 {
1251 	int i;
1252 
1253 	for (i = 0; i < ARRAY_SIZE(qpt->map); i++)
1254 		if (qpt->map[i].page)
1255 			free_page((unsigned long) qpt->map[i].page);
1256 }
1257 
1258 /**
1259  * qib_get_credit - flush the send work queue of a QP
1260  * @qp: the qp who's send work queue to flush
1261  * @aeth: the Acknowledge Extended Transport Header
1262  *
1263  * The QP s_lock should be held.
1264  */
1265 void qib_get_credit(struct qib_qp *qp, u32 aeth)
1266 {
1267 	u32 credit = (aeth >> QIB_AETH_CREDIT_SHIFT) & QIB_AETH_CREDIT_MASK;
1268 
1269 	/*
1270 	 * If the credit is invalid, we can send
1271 	 * as many packets as we like.  Otherwise, we have to
1272 	 * honor the credit field.
1273 	 */
1274 	if (credit == QIB_AETH_CREDIT_INVAL) {
1275 		if (!(qp->s_flags & QIB_S_UNLIMITED_CREDIT)) {
1276 			qp->s_flags |= QIB_S_UNLIMITED_CREDIT;
1277 			if (qp->s_flags & QIB_S_WAIT_SSN_CREDIT) {
1278 				qp->s_flags &= ~QIB_S_WAIT_SSN_CREDIT;
1279 				qib_schedule_send(qp);
1280 			}
1281 		}
1282 	} else if (!(qp->s_flags & QIB_S_UNLIMITED_CREDIT)) {
1283 		/* Compute new LSN (i.e., MSN + credit) */
1284 		credit = (aeth + credit_table[credit]) & QIB_MSN_MASK;
1285 		if (qib_cmp24(credit, qp->s_lsn) > 0) {
1286 			qp->s_lsn = credit;
1287 			if (qp->s_flags & QIB_S_WAIT_SSN_CREDIT) {
1288 				qp->s_flags &= ~QIB_S_WAIT_SSN_CREDIT;
1289 				qib_schedule_send(qp);
1290 			}
1291 		}
1292 	}
1293 }
1294 
1295 #ifdef CONFIG_DEBUG_FS
1296 
1297 struct qib_qp_iter {
1298 	struct qib_ibdev *dev;
1299 	struct qib_qp *qp;
1300 	int n;
1301 };
1302 
1303 struct qib_qp_iter *qib_qp_iter_init(struct qib_ibdev *dev)
1304 {
1305 	struct qib_qp_iter *iter;
1306 
1307 	iter = kzalloc(sizeof(*iter), GFP_KERNEL);
1308 	if (!iter)
1309 		return NULL;
1310 
1311 	iter->dev = dev;
1312 	if (qib_qp_iter_next(iter)) {
1313 		kfree(iter);
1314 		return NULL;
1315 	}
1316 
1317 	return iter;
1318 }
1319 
1320 int qib_qp_iter_next(struct qib_qp_iter *iter)
1321 {
1322 	struct qib_ibdev *dev = iter->dev;
1323 	int n = iter->n;
1324 	int ret = 1;
1325 	struct qib_qp *pqp = iter->qp;
1326 	struct qib_qp *qp;
1327 
1328 	rcu_read_lock();
1329 	for (; n < dev->qp_table_size; n++) {
1330 		if (pqp)
1331 			qp = rcu_dereference(pqp->next);
1332 		else
1333 			qp = rcu_dereference(dev->qp_table[n]);
1334 		pqp = qp;
1335 		if (qp) {
1336 			if (iter->qp)
1337 				atomic_dec(&iter->qp->refcount);
1338 			atomic_inc(&qp->refcount);
1339 			rcu_read_unlock();
1340 			iter->qp = qp;
1341 			iter->n = n;
1342 			return 0;
1343 		}
1344 	}
1345 	rcu_read_unlock();
1346 	if (iter->qp)
1347 		atomic_dec(&iter->qp->refcount);
1348 	return ret;
1349 }
1350 
1351 static const char * const qp_type_str[] = {
1352 	"SMI", "GSI", "RC", "UC", "UD",
1353 };
1354 
1355 void qib_qp_iter_print(struct seq_file *s, struct qib_qp_iter *iter)
1356 {
1357 	struct qib_swqe *wqe;
1358 	struct qib_qp *qp = iter->qp;
1359 
1360 	wqe = get_swqe_ptr(qp, qp->s_last);
1361 	seq_printf(s,
1362 		   "N %d QP%u %s %u %u %u f=%x %u %u %u %u %u PSN %x %x %x %x %x (%u %u %u %u %u %u) QP%u LID %x\n",
1363 		   iter->n,
1364 		   qp->ibqp.qp_num,
1365 		   qp_type_str[qp->ibqp.qp_type],
1366 		   qp->state,
1367 		   wqe->wr.opcode,
1368 		   qp->s_hdrwords,
1369 		   qp->s_flags,
1370 		   atomic_read(&qp->s_dma_busy),
1371 		   !list_empty(&qp->iowait),
1372 		   qp->timeout,
1373 		   wqe->ssn,
1374 		   qp->s_lsn,
1375 		   qp->s_last_psn,
1376 		   qp->s_psn, qp->s_next_psn,
1377 		   qp->s_sending_psn, qp->s_sending_hpsn,
1378 		   qp->s_last, qp->s_acked, qp->s_cur,
1379 		   qp->s_tail, qp->s_head, qp->s_size,
1380 		   qp->remote_qpn,
1381 		   qp->remote_ah_attr.dlid);
1382 }
1383 
1384 #endif
1385