xref: /openbmc/linux/drivers/infiniband/hw/qib/qib_qp.c (revision 1cefc2cd)
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 <rdma/rdma_vt.h>
38 #ifdef CONFIG_DEBUG_FS
39 #include <linux/seq_file.h>
40 #endif
41 
42 #include "qib.h"
43 
44 /*
45  * mask field which was present in now deleted qib_qpn_table
46  * is not present in rvt_qpn_table. Defining the same field
47  * as qpt_mask here instead of adding the mask field to
48  * rvt_qpn_table.
49  */
50 u16 qpt_mask;
51 
52 static inline unsigned mk_qpn(struct rvt_qpn_table *qpt,
53 			      struct rvt_qpn_map *map, unsigned off)
54 {
55 	return (map - qpt->map) * RVT_BITS_PER_PAGE + off;
56 }
57 
58 static inline unsigned find_next_offset(struct rvt_qpn_table *qpt,
59 					struct rvt_qpn_map *map, unsigned off,
60 					unsigned n)
61 {
62 	if (qpt_mask) {
63 		off++;
64 		if (((off & qpt_mask) >> 1) >= n)
65 			off = (off | qpt_mask) + 2;
66 	} else {
67 		off = find_next_zero_bit(map->page, RVT_BITS_PER_PAGE, off);
68 	}
69 	return off;
70 }
71 
72 /*
73  * Convert the AETH credit code into the number of credits.
74  */
75 static u32 credit_table[31] = {
76 	0,                      /* 0 */
77 	1,                      /* 1 */
78 	2,                      /* 2 */
79 	3,                      /* 3 */
80 	4,                      /* 4 */
81 	6,                      /* 5 */
82 	8,                      /* 6 */
83 	12,                     /* 7 */
84 	16,                     /* 8 */
85 	24,                     /* 9 */
86 	32,                     /* A */
87 	48,                     /* B */
88 	64,                     /* C */
89 	96,                     /* D */
90 	128,                    /* E */
91 	192,                    /* F */
92 	256,                    /* 10 */
93 	384,                    /* 11 */
94 	512,                    /* 12 */
95 	768,                    /* 13 */
96 	1024,                   /* 14 */
97 	1536,                   /* 15 */
98 	2048,                   /* 16 */
99 	3072,                   /* 17 */
100 	4096,                   /* 18 */
101 	6144,                   /* 19 */
102 	8192,                   /* 1A */
103 	12288,                  /* 1B */
104 	16384,                  /* 1C */
105 	24576,                  /* 1D */
106 	32768                   /* 1E */
107 };
108 
109 static void get_map_page(struct rvt_qpn_table *qpt, struct rvt_qpn_map *map,
110 			 gfp_t gfp)
111 {
112 	unsigned long page = get_zeroed_page(gfp);
113 
114 	/*
115 	 * Free the page if someone raced with us installing it.
116 	 */
117 
118 	spin_lock(&qpt->lock);
119 	if (map->page)
120 		free_page(page);
121 	else
122 		map->page = (void *)page;
123 	spin_unlock(&qpt->lock);
124 }
125 
126 /*
127  * Allocate the next available QPN or
128  * zero/one for QP type IB_QPT_SMI/IB_QPT_GSI.
129  */
130 int alloc_qpn(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt,
131 	      enum ib_qp_type type, u8 port, gfp_t gfp)
132 {
133 	u32 i, offset, max_scan, qpn;
134 	struct rvt_qpn_map *map;
135 	u32 ret;
136 	struct qib_ibdev *verbs_dev = container_of(rdi, struct qib_ibdev, rdi);
137 	struct qib_devdata *dd = container_of(verbs_dev, struct qib_devdata,
138 					      verbs_dev);
139 
140 	if (type == IB_QPT_SMI || type == IB_QPT_GSI) {
141 		unsigned n;
142 
143 		ret = type == IB_QPT_GSI;
144 		n = 1 << (ret + 2 * (port - 1));
145 		spin_lock(&qpt->lock);
146 		if (qpt->flags & n)
147 			ret = -EINVAL;
148 		else
149 			qpt->flags |= n;
150 		spin_unlock(&qpt->lock);
151 		goto bail;
152 	}
153 
154 	qpn = qpt->last + 2;
155 	if (qpn >= RVT_QPN_MAX)
156 		qpn = 2;
157 	if (qpt_mask && ((qpn & qpt_mask) >> 1) >= dd->n_krcv_queues)
158 		qpn = (qpn | qpt_mask) + 2;
159 	offset = qpn & RVT_BITS_PER_PAGE_MASK;
160 	map = &qpt->map[qpn / RVT_BITS_PER_PAGE];
161 	max_scan = qpt->nmaps - !offset;
162 	for (i = 0;;) {
163 		if (unlikely(!map->page)) {
164 			get_map_page(qpt, map, gfp);
165 			if (unlikely(!map->page))
166 				break;
167 		}
168 		do {
169 			if (!test_and_set_bit(offset, map->page)) {
170 				qpt->last = qpn;
171 				ret = qpn;
172 				goto bail;
173 			}
174 			offset = find_next_offset(qpt, map, offset,
175 				dd->n_krcv_queues);
176 			qpn = mk_qpn(qpt, map, offset);
177 			/*
178 			 * This test differs from alloc_pidmap().
179 			 * If find_next_offset() does find a zero
180 			 * bit, we don't need to check for QPN
181 			 * wrapping around past our starting QPN.
182 			 * We just need to be sure we don't loop
183 			 * forever.
184 			 */
185 		} while (offset < RVT_BITS_PER_PAGE && qpn < RVT_QPN_MAX);
186 		/*
187 		 * In order to keep the number of pages allocated to a
188 		 * minimum, we scan the all existing pages before increasing
189 		 * the size of the bitmap table.
190 		 */
191 		if (++i > max_scan) {
192 			if (qpt->nmaps == RVT_QPNMAP_ENTRIES)
193 				break;
194 			map = &qpt->map[qpt->nmaps++];
195 			offset = 0;
196 		} else if (map < &qpt->map[qpt->nmaps]) {
197 			++map;
198 			offset = 0;
199 		} else {
200 			map = &qpt->map[0];
201 			offset = 2;
202 		}
203 		qpn = mk_qpn(qpt, map, offset);
204 	}
205 
206 	ret = -ENOMEM;
207 
208 bail:
209 	return ret;
210 }
211 
212 static void free_qpn(struct rvt_qpn_table *qpt, u32 qpn)
213 {
214 	struct rvt_qpn_map *map;
215 
216 	map = qpt->map + qpn / RVT_BITS_PER_PAGE;
217 	if (map->page)
218 		clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page);
219 }
220 
221 static inline unsigned qpn_hash(struct qib_ibdev *dev, u32 qpn)
222 {
223 	return hash_32(qpn, dev->rdi.qp_dev->qp_table_bits);
224 }
225 
226 
227 /*
228  * Put the QP into the hash table.
229  * The hash table holds a reference to the QP.
230  */
231 static void insert_qp(struct qib_ibdev *dev, struct rvt_qp *qp)
232 {
233 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
234 	unsigned long flags;
235 	unsigned n = qpn_hash(dev, qp->ibqp.qp_num);
236 
237 	atomic_inc(&qp->refcount);
238 	spin_lock_irqsave(&dev->rdi.qp_dev->qpt_lock, flags);
239 
240 	if (qp->ibqp.qp_num == 0)
241 		rcu_assign_pointer(ibp->rvp.qp[0], qp);
242 	else if (qp->ibqp.qp_num == 1)
243 		rcu_assign_pointer(ibp->rvp.qp[1], qp);
244 	else {
245 		qp->next = dev->rdi.qp_dev->qp_table[n];
246 		rcu_assign_pointer(dev->rdi.qp_dev->qp_table[n], qp);
247 	}
248 
249 	spin_unlock_irqrestore(&dev->rdi.qp_dev->qpt_lock, flags);
250 }
251 
252 /*
253  * Remove the QP from the table so it can't be found asynchronously by
254  * the receive interrupt routine.
255  */
256 static void remove_qp(struct qib_ibdev *dev, struct rvt_qp *qp)
257 {
258 	struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
259 	unsigned n = qpn_hash(dev, qp->ibqp.qp_num);
260 	unsigned long flags;
261 	int removed = 1;
262 	spinlock_t *qpt_lock_ptr; /* Pointer to make checkpatch happy */
263 
264 	spin_lock_irqsave(&dev->rdi.qp_dev->qpt_lock, flags);
265 
266 	qpt_lock_ptr = &dev->rdi.qp_dev->qpt_lock;
267 	if (rcu_dereference_protected(ibp->rvp.qp[0],
268 				      lockdep_is_held(qpt_lock_ptr)) == qp) {
269 		RCU_INIT_POINTER(ibp->rvp.qp[0], NULL);
270 	} else if (rcu_dereference_protected(ibp->rvp.qp[1],
271 			lockdep_is_held(&dev->rdi.qp_dev->qpt_lock)) == qp) {
272 		RCU_INIT_POINTER(ibp->rvp.qp[1], NULL);
273 	} else {
274 		struct rvt_qp *q;
275 		struct rvt_qp __rcu **qpp;
276 
277 		removed = 0;
278 		qpp = &dev->rdi.qp_dev->qp_table[n];
279 		for (; (q = rcu_dereference_protected(*qpp,
280 				lockdep_is_held(qpt_lock_ptr))) != NULL;
281 				qpp = &q->next)
282 			if (q == qp) {
283 				RCU_INIT_POINTER(*qpp,
284 					rcu_dereference_protected(qp->next,
285 					 lockdep_is_held(qpt_lock_ptr)));
286 				removed = 1;
287 				break;
288 			}
289 	}
290 
291 	spin_unlock_irqrestore(&dev->rdi.qp_dev->qpt_lock, flags);
292 	if (removed) {
293 		synchronize_rcu();
294 		if (atomic_dec_and_test(&qp->refcount))
295 			wake_up(&qp->wait);
296 	}
297 }
298 
299 /**
300  * qib_free_all_qps - check for QPs still in use
301  */
302 unsigned qib_free_all_qps(struct rvt_dev_info *rdi)
303 {
304 	struct qib_ibdev *verbs_dev = container_of(rdi, struct qib_ibdev, rdi);
305 	struct qib_devdata *dd = container_of(verbs_dev, struct qib_devdata,
306 					      verbs_dev);
307 	unsigned n, qp_inuse = 0;
308 
309 	for (n = 0; n < dd->num_pports; n++) {
310 		struct qib_ibport *ibp = &dd->pport[n].ibport_data;
311 
312 		rcu_read_lock();
313 		if (rcu_dereference(ibp->rvp.qp[0]))
314 			qp_inuse++;
315 		if (rcu_dereference(ibp->rvp.qp[1]))
316 			qp_inuse++;
317 		rcu_read_unlock();
318 	}
319 	return qp_inuse;
320 }
321 
322 void notify_qp_reset(struct rvt_qp *qp)
323 {
324 	struct qib_qp_priv *priv = qp->priv;
325 
326 	atomic_set(&priv->s_dma_busy, 0);
327 }
328 
329 static void clear_mr_refs(struct rvt_qp *qp, int clr_sends)
330 {
331 	unsigned n;
332 
333 	if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags))
334 		qib_put_ss(&qp->s_rdma_read_sge);
335 
336 	qib_put_ss(&qp->r_sge);
337 
338 	if (clr_sends) {
339 		while (qp->s_last != qp->s_head) {
340 			struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_last);
341 			unsigned i;
342 
343 			for (i = 0; i < wqe->wr.num_sge; i++) {
344 				struct rvt_sge *sge = &wqe->sg_list[i];
345 
346 				rvt_put_mr(sge->mr);
347 			}
348 			if (qp->ibqp.qp_type == IB_QPT_UD ||
349 			    qp->ibqp.qp_type == IB_QPT_SMI ||
350 			    qp->ibqp.qp_type == IB_QPT_GSI)
351 				atomic_dec(
352 				 &ibah_to_rvtah(wqe->ud_wr.ah)->refcount);
353 			if (++qp->s_last >= qp->s_size)
354 				qp->s_last = 0;
355 		}
356 		if (qp->s_rdma_mr) {
357 			rvt_put_mr(qp->s_rdma_mr);
358 			qp->s_rdma_mr = NULL;
359 		}
360 	}
361 
362 	if (qp->ibqp.qp_type != IB_QPT_RC)
363 		return;
364 
365 	for (n = 0; n < ARRAY_SIZE(qp->s_ack_queue); n++) {
366 		struct rvt_ack_entry *e = &qp->s_ack_queue[n];
367 
368 		if (e->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST &&
369 		    e->rdma_sge.mr) {
370 			rvt_put_mr(e->rdma_sge.mr);
371 			e->rdma_sge.mr = NULL;
372 		}
373 	}
374 }
375 
376 /**
377  * qib_error_qp - put a QP into the error state
378  * @qp: the QP to put into the error state
379  * @err: the receive completion error to signal if a RWQE is active
380  *
381  * Flushes both send and receive work queues.
382  * Returns true if last WQE event should be generated.
383  * The QP r_lock and s_lock should be held and interrupts disabled.
384  * If we are already in error state, just return.
385  */
386 int qib_error_qp(struct rvt_qp *qp, enum ib_wc_status err)
387 {
388 	struct qib_qp_priv *priv = qp->priv;
389 	struct qib_ibdev *dev = to_idev(qp->ibqp.device);
390 	struct ib_wc wc;
391 	int ret = 0;
392 
393 	if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET)
394 		goto bail;
395 
396 	qp->state = IB_QPS_ERR;
397 
398 	if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) {
399 		qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR);
400 		del_timer(&qp->s_timer);
401 	}
402 
403 	if (qp->s_flags & RVT_S_ANY_WAIT_SEND)
404 		qp->s_flags &= ~RVT_S_ANY_WAIT_SEND;
405 
406 	spin_lock(&dev->rdi.pending_lock);
407 	if (!list_empty(&priv->iowait) && !(qp->s_flags & RVT_S_BUSY)) {
408 		qp->s_flags &= ~RVT_S_ANY_WAIT_IO;
409 		list_del_init(&priv->iowait);
410 	}
411 	spin_unlock(&dev->rdi.pending_lock);
412 
413 	if (!(qp->s_flags & RVT_S_BUSY)) {
414 		qp->s_hdrwords = 0;
415 		if (qp->s_rdma_mr) {
416 			rvt_put_mr(qp->s_rdma_mr);
417 			qp->s_rdma_mr = NULL;
418 		}
419 		if (priv->s_tx) {
420 			qib_put_txreq(priv->s_tx);
421 			priv->s_tx = NULL;
422 		}
423 	}
424 
425 	/* Schedule the sending tasklet to drain the send work queue. */
426 	if (qp->s_last != qp->s_head)
427 		qib_schedule_send(qp);
428 
429 	clear_mr_refs(qp, 0);
430 
431 	memset(&wc, 0, sizeof(wc));
432 	wc.qp = &qp->ibqp;
433 	wc.opcode = IB_WC_RECV;
434 
435 	if (test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) {
436 		wc.wr_id = qp->r_wr_id;
437 		wc.status = err;
438 		rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
439 	}
440 	wc.status = IB_WC_WR_FLUSH_ERR;
441 
442 	if (qp->r_rq.wq) {
443 		struct rvt_rwq *wq;
444 		u32 head;
445 		u32 tail;
446 
447 		spin_lock(&qp->r_rq.lock);
448 
449 		/* sanity check pointers before trusting them */
450 		wq = qp->r_rq.wq;
451 		head = wq->head;
452 		if (head >= qp->r_rq.size)
453 			head = 0;
454 		tail = wq->tail;
455 		if (tail >= qp->r_rq.size)
456 			tail = 0;
457 		while (tail != head) {
458 			wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
459 			if (++tail >= qp->r_rq.size)
460 				tail = 0;
461 			rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
462 		}
463 		wq->tail = tail;
464 
465 		spin_unlock(&qp->r_rq.lock);
466 	} else if (qp->ibqp.event_handler)
467 		ret = 1;
468 
469 bail:
470 	return ret;
471 }
472 
473 /**
474  * qib_modify_qp - modify the attributes of a queue pair
475  * @ibqp: the queue pair who's attributes we're modifying
476  * @attr: the new attributes
477  * @attr_mask: the mask of attributes to modify
478  * @udata: user data for libibverbs.so
479  *
480  * Returns 0 on success, otherwise returns an errno.
481  */
482 int qib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
483 		  int attr_mask, struct ib_udata *udata)
484 {
485 	struct qib_ibdev *dev = to_idev(ibqp->device);
486 	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
487 	struct qib_qp_priv *priv = qp->priv;
488 	enum ib_qp_state cur_state, new_state;
489 	struct ib_event ev;
490 	int lastwqe = 0;
491 	int mig = 0;
492 	int ret;
493 	u32 pmtu = 0; /* for gcc warning only */
494 
495 	spin_lock_irq(&qp->r_lock);
496 	spin_lock(&qp->s_lock);
497 
498 	cur_state = attr_mask & IB_QP_CUR_STATE ?
499 		attr->cur_qp_state : qp->state;
500 	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
501 
502 	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
503 				attr_mask, IB_LINK_LAYER_UNSPECIFIED))
504 		goto inval;
505 
506 	if (attr_mask & IB_QP_AV) {
507 		if (attr->ah_attr.dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE))
508 			goto inval;
509 		if (rvt_check_ah(qp->ibqp.device, &attr->ah_attr))
510 			goto inval;
511 	}
512 
513 	if (attr_mask & IB_QP_ALT_PATH) {
514 		if (attr->alt_ah_attr.dlid >=
515 		    be16_to_cpu(IB_MULTICAST_LID_BASE))
516 			goto inval;
517 		if (rvt_check_ah(qp->ibqp.device, &attr->alt_ah_attr))
518 			goto inval;
519 		if (attr->alt_pkey_index >= qib_get_npkeys(dd_from_dev(dev)))
520 			goto inval;
521 	}
522 
523 	if (attr_mask & IB_QP_PKEY_INDEX)
524 		if (attr->pkey_index >= qib_get_npkeys(dd_from_dev(dev)))
525 			goto inval;
526 
527 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
528 		if (attr->min_rnr_timer > 31)
529 			goto inval;
530 
531 	if (attr_mask & IB_QP_PORT)
532 		if (qp->ibqp.qp_type == IB_QPT_SMI ||
533 		    qp->ibqp.qp_type == IB_QPT_GSI ||
534 		    attr->port_num == 0 ||
535 		    attr->port_num > ibqp->device->phys_port_cnt)
536 			goto inval;
537 
538 	if (attr_mask & IB_QP_DEST_QPN)
539 		if (attr->dest_qp_num > QIB_QPN_MASK)
540 			goto inval;
541 
542 	if (attr_mask & IB_QP_RETRY_CNT)
543 		if (attr->retry_cnt > 7)
544 			goto inval;
545 
546 	if (attr_mask & IB_QP_RNR_RETRY)
547 		if (attr->rnr_retry > 7)
548 			goto inval;
549 
550 	/*
551 	 * Don't allow invalid path_mtu values.  OK to set greater
552 	 * than the active mtu (or even the max_cap, if we have tuned
553 	 * that to a small mtu.  We'll set qp->path_mtu
554 	 * to the lesser of requested attribute mtu and active,
555 	 * for packetizing messages.
556 	 * Note that the QP port has to be set in INIT and MTU in RTR.
557 	 */
558 	if (attr_mask & IB_QP_PATH_MTU) {
559 		struct qib_devdata *dd = dd_from_dev(dev);
560 		int mtu, pidx = qp->port_num - 1;
561 
562 		mtu = ib_mtu_enum_to_int(attr->path_mtu);
563 		if (mtu == -1)
564 			goto inval;
565 		if (mtu > dd->pport[pidx].ibmtu) {
566 			switch (dd->pport[pidx].ibmtu) {
567 			case 4096:
568 				pmtu = IB_MTU_4096;
569 				break;
570 			case 2048:
571 				pmtu = IB_MTU_2048;
572 				break;
573 			case 1024:
574 				pmtu = IB_MTU_1024;
575 				break;
576 			case 512:
577 				pmtu = IB_MTU_512;
578 				break;
579 			case 256:
580 				pmtu = IB_MTU_256;
581 				break;
582 			default:
583 				pmtu = IB_MTU_2048;
584 			}
585 		} else
586 			pmtu = attr->path_mtu;
587 	}
588 
589 	if (attr_mask & IB_QP_PATH_MIG_STATE) {
590 		if (attr->path_mig_state == IB_MIG_REARM) {
591 			if (qp->s_mig_state == IB_MIG_ARMED)
592 				goto inval;
593 			if (new_state != IB_QPS_RTS)
594 				goto inval;
595 		} else if (attr->path_mig_state == IB_MIG_MIGRATED) {
596 			if (qp->s_mig_state == IB_MIG_REARM)
597 				goto inval;
598 			if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD)
599 				goto inval;
600 			if (qp->s_mig_state == IB_MIG_ARMED)
601 				mig = 1;
602 		} else
603 			goto inval;
604 	}
605 
606 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
607 		if (attr->max_dest_rd_atomic > QIB_MAX_RDMA_ATOMIC)
608 			goto inval;
609 
610 	switch (new_state) {
611 	case IB_QPS_RESET:
612 		if (qp->state != IB_QPS_RESET) {
613 			qp->state = IB_QPS_RESET;
614 			spin_lock(&dev->rdi.pending_lock);
615 			if (!list_empty(&priv->iowait))
616 				list_del_init(&priv->iowait);
617 			spin_unlock(&dev->rdi.pending_lock);
618 			qp->s_flags &= ~(RVT_S_TIMER | RVT_S_ANY_WAIT);
619 			spin_unlock(&qp->s_lock);
620 			spin_unlock_irq(&qp->r_lock);
621 			/* Stop the sending work queue and retry timer */
622 			cancel_work_sync(&priv->s_work);
623 			del_timer_sync(&qp->s_timer);
624 			wait_event(priv->wait_dma,
625 				   !atomic_read(&priv->s_dma_busy));
626 			if (priv->s_tx) {
627 				qib_put_txreq(priv->s_tx);
628 				priv->s_tx = NULL;
629 			}
630 			remove_qp(dev, qp);
631 			wait_event(qp->wait, !atomic_read(&qp->refcount));
632 			spin_lock_irq(&qp->r_lock);
633 			spin_lock(&qp->s_lock);
634 			clear_mr_refs(qp, 1);
635 			rvt_reset_qp(&dev->rdi, qp, ibqp->qp_type);
636 		}
637 		break;
638 
639 	case IB_QPS_RTR:
640 		/* Allow event to retrigger if QP set to RTR more than once */
641 		qp->r_flags &= ~RVT_R_COMM_EST;
642 		qp->state = new_state;
643 		break;
644 
645 	case IB_QPS_SQD:
646 		qp->s_draining = qp->s_last != qp->s_cur;
647 		qp->state = new_state;
648 		break;
649 
650 	case IB_QPS_SQE:
651 		if (qp->ibqp.qp_type == IB_QPT_RC)
652 			goto inval;
653 		qp->state = new_state;
654 		break;
655 
656 	case IB_QPS_ERR:
657 		lastwqe = qib_error_qp(qp, IB_WC_WR_FLUSH_ERR);
658 		break;
659 
660 	default:
661 		qp->state = new_state;
662 		break;
663 	}
664 
665 	if (attr_mask & IB_QP_PKEY_INDEX)
666 		qp->s_pkey_index = attr->pkey_index;
667 
668 	if (attr_mask & IB_QP_PORT)
669 		qp->port_num = attr->port_num;
670 
671 	if (attr_mask & IB_QP_DEST_QPN)
672 		qp->remote_qpn = attr->dest_qp_num;
673 
674 	if (attr_mask & IB_QP_SQ_PSN) {
675 		qp->s_next_psn = attr->sq_psn & QIB_PSN_MASK;
676 		qp->s_psn = qp->s_next_psn;
677 		qp->s_sending_psn = qp->s_next_psn;
678 		qp->s_last_psn = qp->s_next_psn - 1;
679 		qp->s_sending_hpsn = qp->s_last_psn;
680 	}
681 
682 	if (attr_mask & IB_QP_RQ_PSN)
683 		qp->r_psn = attr->rq_psn & QIB_PSN_MASK;
684 
685 	if (attr_mask & IB_QP_ACCESS_FLAGS)
686 		qp->qp_access_flags = attr->qp_access_flags;
687 
688 	if (attr_mask & IB_QP_AV) {
689 		qp->remote_ah_attr = attr->ah_attr;
690 		qp->s_srate = attr->ah_attr.static_rate;
691 	}
692 
693 	if (attr_mask & IB_QP_ALT_PATH) {
694 		qp->alt_ah_attr = attr->alt_ah_attr;
695 		qp->s_alt_pkey_index = attr->alt_pkey_index;
696 	}
697 
698 	if (attr_mask & IB_QP_PATH_MIG_STATE) {
699 		qp->s_mig_state = attr->path_mig_state;
700 		if (mig) {
701 			qp->remote_ah_attr = qp->alt_ah_attr;
702 			qp->port_num = qp->alt_ah_attr.port_num;
703 			qp->s_pkey_index = qp->s_alt_pkey_index;
704 		}
705 	}
706 
707 	if (attr_mask & IB_QP_PATH_MTU) {
708 		qp->path_mtu = pmtu;
709 		qp->pmtu = ib_mtu_enum_to_int(pmtu);
710 	}
711 
712 	if (attr_mask & IB_QP_RETRY_CNT) {
713 		qp->s_retry_cnt = attr->retry_cnt;
714 		qp->s_retry = attr->retry_cnt;
715 	}
716 
717 	if (attr_mask & IB_QP_RNR_RETRY) {
718 		qp->s_rnr_retry_cnt = attr->rnr_retry;
719 		qp->s_rnr_retry = attr->rnr_retry;
720 	}
721 
722 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
723 		qp->r_min_rnr_timer = attr->min_rnr_timer;
724 
725 	if (attr_mask & IB_QP_TIMEOUT) {
726 		qp->timeout = attr->timeout;
727 		qp->timeout_jiffies =
728 			usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
729 				1000UL);
730 	}
731 
732 	if (attr_mask & IB_QP_QKEY)
733 		qp->qkey = attr->qkey;
734 
735 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
736 		qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
737 
738 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
739 		qp->s_max_rd_atomic = attr->max_rd_atomic;
740 
741 	spin_unlock(&qp->s_lock);
742 	spin_unlock_irq(&qp->r_lock);
743 
744 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
745 		insert_qp(dev, qp);
746 
747 	if (lastwqe) {
748 		ev.device = qp->ibqp.device;
749 		ev.element.qp = &qp->ibqp;
750 		ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
751 		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
752 	}
753 	if (mig) {
754 		ev.device = qp->ibqp.device;
755 		ev.element.qp = &qp->ibqp;
756 		ev.event = IB_EVENT_PATH_MIG;
757 		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
758 	}
759 	ret = 0;
760 	goto bail;
761 
762 inval:
763 	spin_unlock(&qp->s_lock);
764 	spin_unlock_irq(&qp->r_lock);
765 	ret = -EINVAL;
766 
767 bail:
768 	return ret;
769 }
770 
771 /**
772  * qib_compute_aeth - compute the AETH (syndrome + MSN)
773  * @qp: the queue pair to compute the AETH for
774  *
775  * Returns the AETH.
776  */
777 __be32 qib_compute_aeth(struct rvt_qp *qp)
778 {
779 	u32 aeth = qp->r_msn & QIB_MSN_MASK;
780 
781 	if (qp->ibqp.srq) {
782 		/*
783 		 * Shared receive queues don't generate credits.
784 		 * Set the credit field to the invalid value.
785 		 */
786 		aeth |= QIB_AETH_CREDIT_INVAL << QIB_AETH_CREDIT_SHIFT;
787 	} else {
788 		u32 min, max, x;
789 		u32 credits;
790 		struct rvt_rwq *wq = qp->r_rq.wq;
791 		u32 head;
792 		u32 tail;
793 
794 		/* sanity check pointers before trusting them */
795 		head = wq->head;
796 		if (head >= qp->r_rq.size)
797 			head = 0;
798 		tail = wq->tail;
799 		if (tail >= qp->r_rq.size)
800 			tail = 0;
801 		/*
802 		 * Compute the number of credits available (RWQEs).
803 		 * XXX Not holding the r_rq.lock here so there is a small
804 		 * chance that the pair of reads are not atomic.
805 		 */
806 		credits = head - tail;
807 		if ((int)credits < 0)
808 			credits += qp->r_rq.size;
809 		/*
810 		 * Binary search the credit table to find the code to
811 		 * use.
812 		 */
813 		min = 0;
814 		max = 31;
815 		for (;;) {
816 			x = (min + max) / 2;
817 			if (credit_table[x] == credits)
818 				break;
819 			if (credit_table[x] > credits)
820 				max = x;
821 			else if (min == x)
822 				break;
823 			else
824 				min = x;
825 		}
826 		aeth |= x << QIB_AETH_CREDIT_SHIFT;
827 	}
828 	return cpu_to_be32(aeth);
829 }
830 
831 void *qp_priv_alloc(struct rvt_dev_info *rdi, struct rvt_qp *qp, gfp_t gfp)
832 {
833 	struct qib_qp_priv *priv;
834 
835 	priv = kzalloc(sizeof(*priv), gfp);
836 	if (!priv)
837 		return ERR_PTR(-ENOMEM);
838 	priv->owner = qp;
839 
840 	priv->s_hdr = kzalloc(sizeof(*priv->s_hdr), gfp);
841 	if (!priv->s_hdr) {
842 		kfree(priv);
843 		return ERR_PTR(-ENOMEM);
844 	}
845 	init_waitqueue_head(&priv->wait_dma);
846 	INIT_WORK(&priv->s_work, _qib_do_send);
847 	INIT_LIST_HEAD(&priv->iowait);
848 
849 	return priv;
850 }
851 
852 void qp_priv_free(struct rvt_dev_info *rdi, struct rvt_qp *qp)
853 {
854 	struct qib_qp_priv *priv = qp->priv;
855 
856 	kfree(priv->s_hdr);
857 	kfree(priv);
858 }
859 
860 /**
861  * qib_destroy_qp - destroy a queue pair
862  * @ibqp: the queue pair to destroy
863  *
864  * Returns 0 on success.
865  *
866  * Note that this can be called while the QP is actively sending or
867  * receiving!
868  */
869 int qib_destroy_qp(struct ib_qp *ibqp)
870 {
871 	struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
872 	struct qib_ibdev *dev = to_idev(ibqp->device);
873 	struct qib_qp_priv *priv = qp->priv;
874 
875 	/* Make sure HW and driver activity is stopped. */
876 	spin_lock_irq(&qp->s_lock);
877 	if (qp->state != IB_QPS_RESET) {
878 		qp->state = IB_QPS_RESET;
879 		spin_lock(&dev->rdi.pending_lock);
880 		if (!list_empty(&priv->iowait))
881 			list_del_init(&priv->iowait);
882 		spin_unlock(&dev->rdi.pending_lock);
883 		qp->s_flags &= ~(RVT_S_TIMER | RVT_S_ANY_WAIT);
884 		spin_unlock_irq(&qp->s_lock);
885 		cancel_work_sync(&priv->s_work);
886 		del_timer_sync(&qp->s_timer);
887 		wait_event(priv->wait_dma, !atomic_read(&priv->s_dma_busy));
888 		if (priv->s_tx) {
889 			qib_put_txreq(priv->s_tx);
890 			priv->s_tx = NULL;
891 		}
892 		remove_qp(dev, qp);
893 		wait_event(qp->wait, !atomic_read(&qp->refcount));
894 		clear_mr_refs(qp, 1);
895 	} else
896 		spin_unlock_irq(&qp->s_lock);
897 
898 	/* all user's cleaned up, mark it available */
899 	free_qpn(&dev->rdi.qp_dev->qpn_table, qp->ibqp.qp_num);
900 	spin_lock(&dev->n_qps_lock);
901 	dev->n_qps_allocated--;
902 	spin_unlock(&dev->n_qps_lock);
903 
904 	if (qp->ip)
905 		kref_put(&qp->ip->ref, rvt_release_mmap_info);
906 	else
907 		vfree(qp->r_rq.wq);
908 	vfree(qp->s_wq);
909 	kfree(priv->s_hdr);
910 	kfree(priv);
911 	kfree(qp);
912 	return 0;
913 }
914 
915 /**
916  * qib_get_credit - flush the send work queue of a QP
917  * @qp: the qp who's send work queue to flush
918  * @aeth: the Acknowledge Extended Transport Header
919  *
920  * The QP s_lock should be held.
921  */
922 void qib_get_credit(struct rvt_qp *qp, u32 aeth)
923 {
924 	u32 credit = (aeth >> QIB_AETH_CREDIT_SHIFT) & QIB_AETH_CREDIT_MASK;
925 
926 	/*
927 	 * If the credit is invalid, we can send
928 	 * as many packets as we like.  Otherwise, we have to
929 	 * honor the credit field.
930 	 */
931 	if (credit == QIB_AETH_CREDIT_INVAL) {
932 		if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) {
933 			qp->s_flags |= RVT_S_UNLIMITED_CREDIT;
934 			if (qp->s_flags & RVT_S_WAIT_SSN_CREDIT) {
935 				qp->s_flags &= ~RVT_S_WAIT_SSN_CREDIT;
936 				qib_schedule_send(qp);
937 			}
938 		}
939 	} else if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) {
940 		/* Compute new LSN (i.e., MSN + credit) */
941 		credit = (aeth + credit_table[credit]) & QIB_MSN_MASK;
942 		if (qib_cmp24(credit, qp->s_lsn) > 0) {
943 			qp->s_lsn = credit;
944 			if (qp->s_flags & RVT_S_WAIT_SSN_CREDIT) {
945 				qp->s_flags &= ~RVT_S_WAIT_SSN_CREDIT;
946 				qib_schedule_send(qp);
947 			}
948 		}
949 	}
950 }
951 
952 #ifdef CONFIG_DEBUG_FS
953 
954 struct qib_qp_iter {
955 	struct qib_ibdev *dev;
956 	struct rvt_qp *qp;
957 	int n;
958 };
959 
960 struct qib_qp_iter *qib_qp_iter_init(struct qib_ibdev *dev)
961 {
962 	struct qib_qp_iter *iter;
963 
964 	iter = kzalloc(sizeof(*iter), GFP_KERNEL);
965 	if (!iter)
966 		return NULL;
967 
968 	iter->dev = dev;
969 	if (qib_qp_iter_next(iter)) {
970 		kfree(iter);
971 		return NULL;
972 	}
973 
974 	return iter;
975 }
976 
977 int qib_qp_iter_next(struct qib_qp_iter *iter)
978 {
979 	struct qib_ibdev *dev = iter->dev;
980 	int n = iter->n;
981 	int ret = 1;
982 	struct rvt_qp *pqp = iter->qp;
983 	struct rvt_qp *qp;
984 
985 	for (; n < dev->rdi.qp_dev->qp_table_size; n++) {
986 		if (pqp)
987 			qp = rcu_dereference(pqp->next);
988 		else
989 			qp = rcu_dereference(dev->rdi.qp_dev->qp_table[n]);
990 		pqp = qp;
991 		if (qp) {
992 			iter->qp = qp;
993 			iter->n = n;
994 			return 0;
995 		}
996 	}
997 	return ret;
998 }
999 
1000 static const char * const qp_type_str[] = {
1001 	"SMI", "GSI", "RC", "UC", "UD",
1002 };
1003 
1004 void qib_qp_iter_print(struct seq_file *s, struct qib_qp_iter *iter)
1005 {
1006 	struct rvt_swqe *wqe;
1007 	struct rvt_qp *qp = iter->qp;
1008 	struct qib_qp_priv *priv = qp->priv;
1009 
1010 	wqe = rvt_get_swqe_ptr(qp, qp->s_last);
1011 	seq_printf(s,
1012 		   "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",
1013 		   iter->n,
1014 		   qp->ibqp.qp_num,
1015 		   qp_type_str[qp->ibqp.qp_type],
1016 		   qp->state,
1017 		   wqe->wr.opcode,
1018 		   qp->s_hdrwords,
1019 		   qp->s_flags,
1020 		   atomic_read(&priv->s_dma_busy),
1021 		   !list_empty(&priv->iowait),
1022 		   qp->timeout,
1023 		   wqe->ssn,
1024 		   qp->s_lsn,
1025 		   qp->s_last_psn,
1026 		   qp->s_psn, qp->s_next_psn,
1027 		   qp->s_sending_psn, qp->s_sending_hpsn,
1028 		   qp->s_last, qp->s_acked, qp->s_cur,
1029 		   qp->s_tail, qp->s_head, qp->s_size,
1030 		   qp->remote_qpn,
1031 		   qp->remote_ah_attr.dlid);
1032 }
1033 
1034 #endif
1035