1/*
2 *
3 * This file is provided under a dual BSD/GPLv2 license.  When using or
4 * redistributing this file, you may do so under either license.
5 *
6 * GPL LICENSE SUMMARY
7 *
8 * Copyright(c) 2015 Intel Corporation.
9 *
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of version 2 of the GNU General Public License as
12 * published by the Free Software Foundation.
13 *
14 * This program is distributed in the hope that it will be useful, but
15 * WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
17 * General Public License for more details.
18 *
19 * BSD LICENSE
20 *
21 * Copyright(c) 2015 Intel Corporation.
22 *
23 * Redistribution and use in source and binary forms, with or without
24 * modification, are permitted provided that the following conditions
25 * are met:
26 *
27 *  - Redistributions of source code must retain the above copyright
28 *    notice, this list of conditions and the following disclaimer.
29 *  - Redistributions in binary form must reproduce the above copyright
30 *    notice, this list of conditions and the following disclaimer in
31 *    the documentation and/or other materials provided with the
32 *    distribution.
33 *  - Neither the name of Intel Corporation nor the names of its
34 *    contributors may be used to endorse or promote products derived
35 *    from this software without specific prior written permission.
36 *
37 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
38 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
39 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
40 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
41 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
42 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
43 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
44 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
45 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
46 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
47 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
48 *
49 */
50
51#include <linux/err.h>
52#include <linux/vmalloc.h>
53#include <linux/hash.h>
54#include <linux/module.h>
55#include <linux/random.h>
56#include <linux/seq_file.h>
57
58#include "hfi.h"
59#include "qp.h"
60#include "trace.h"
61#include "sdma.h"
62
63#define BITS_PER_PAGE           (PAGE_SIZE*BITS_PER_BYTE)
64#define BITS_PER_PAGE_MASK      (BITS_PER_PAGE-1)
65
66static unsigned int hfi1_qp_table_size = 256;
67module_param_named(qp_table_size, hfi1_qp_table_size, uint, S_IRUGO);
68MODULE_PARM_DESC(qp_table_size, "QP table size");
69
70static void flush_tx_list(struct hfi1_qp *qp);
71static int iowait_sleep(
72	struct sdma_engine *sde,
73	struct iowait *wait,
74	struct sdma_txreq *stx,
75	unsigned seq);
76static void iowait_wakeup(struct iowait *wait, int reason);
77
78static inline unsigned mk_qpn(struct hfi1_qpn_table *qpt,
79			      struct qpn_map *map, unsigned off)
80{
81	return (map - qpt->map) * BITS_PER_PAGE + off;
82}
83
84/*
85 * Convert the AETH credit code into the number of credits.
86 */
87static const u16 credit_table[31] = {
88	0,                      /* 0 */
89	1,                      /* 1 */
90	2,                      /* 2 */
91	3,                      /* 3 */
92	4,                      /* 4 */
93	6,                      /* 5 */
94	8,                      /* 6 */
95	12,                     /* 7 */
96	16,                     /* 8 */
97	24,                     /* 9 */
98	32,                     /* A */
99	48,                     /* B */
100	64,                     /* C */
101	96,                     /* D */
102	128,                    /* E */
103	192,                    /* F */
104	256,                    /* 10 */
105	384,                    /* 11 */
106	512,                    /* 12 */
107	768,                    /* 13 */
108	1024,                   /* 14 */
109	1536,                   /* 15 */
110	2048,                   /* 16 */
111	3072,                   /* 17 */
112	4096,                   /* 18 */
113	6144,                   /* 19 */
114	8192,                   /* 1A */
115	12288,                  /* 1B */
116	16384,                  /* 1C */
117	24576,                  /* 1D */
118	32768                   /* 1E */
119};
120
121static void get_map_page(struct hfi1_qpn_table *qpt, struct qpn_map *map)
122{
123	unsigned long page = get_zeroed_page(GFP_KERNEL);
124
125	/*
126	 * Free the page if someone raced with us installing it.
127	 */
128
129	spin_lock(&qpt->lock);
130	if (map->page)
131		free_page(page);
132	else
133		map->page = (void *)page;
134	spin_unlock(&qpt->lock);
135}
136
137/*
138 * Allocate the next available QPN or
139 * zero/one for QP type IB_QPT_SMI/IB_QPT_GSI.
140 */
141static int alloc_qpn(struct hfi1_devdata *dd, struct hfi1_qpn_table *qpt,
142		     enum ib_qp_type type, u8 port)
143{
144	u32 i, offset, max_scan, qpn;
145	struct qpn_map *map;
146	u32 ret;
147
148	if (type == IB_QPT_SMI || type == IB_QPT_GSI) {
149		unsigned n;
150
151		ret = type == IB_QPT_GSI;
152		n = 1 << (ret + 2 * (port - 1));
153		spin_lock(&qpt->lock);
154		if (qpt->flags & n)
155			ret = -EINVAL;
156		else
157			qpt->flags |= n;
158		spin_unlock(&qpt->lock);
159		goto bail;
160	}
161
162	qpn = qpt->last + qpt->incr;
163	if (qpn >= QPN_MAX)
164		qpn = qpt->incr | ((qpt->last & 1) ^ 1);
165	/* offset carries bit 0 */
166	offset = qpn & BITS_PER_PAGE_MASK;
167	map = &qpt->map[qpn / BITS_PER_PAGE];
168	max_scan = qpt->nmaps - !offset;
169	for (i = 0;;) {
170		if (unlikely(!map->page)) {
171			get_map_page(qpt, map);
172			if (unlikely(!map->page))
173				break;
174		}
175		do {
176			if (!test_and_set_bit(offset, map->page)) {
177				qpt->last = qpn;
178				ret = qpn;
179				goto bail;
180			}
181			offset += qpt->incr;
182			/*
183			 * This qpn might be bogus if offset >= BITS_PER_PAGE.
184			 * That is OK.   It gets re-assigned below
185			 */
186			qpn = mk_qpn(qpt, map, offset);
187		} while (offset < BITS_PER_PAGE && qpn < QPN_MAX);
188		/*
189		 * In order to keep the number of pages allocated to a
190		 * minimum, we scan the all existing pages before increasing
191		 * the size of the bitmap table.
192		 */
193		if (++i > max_scan) {
194			if (qpt->nmaps == QPNMAP_ENTRIES)
195				break;
196			map = &qpt->map[qpt->nmaps++];
197			/* start at incr with current bit 0 */
198			offset = qpt->incr | (offset & 1);
199		} else if (map < &qpt->map[qpt->nmaps]) {
200			++map;
201			/* start at incr with current bit 0 */
202			offset = qpt->incr | (offset & 1);
203		} else {
204			map = &qpt->map[0];
205			/* wrap to first map page, invert bit 0 */
206			offset = qpt->incr | ((offset & 1) ^ 1);
207		}
208		/* there can be no bits at shift and below */
209		WARN_ON(offset & (dd->qos_shift - 1));
210		qpn = mk_qpn(qpt, map, offset);
211	}
212
213	ret = -ENOMEM;
214
215bail:
216	return ret;
217}
218
219static void free_qpn(struct hfi1_qpn_table *qpt, u32 qpn)
220{
221	struct qpn_map *map;
222
223	map = qpt->map + qpn / BITS_PER_PAGE;
224	if (map->page)
225		clear_bit(qpn & BITS_PER_PAGE_MASK, map->page);
226}
227
228/*
229 * Put the QP into the hash table.
230 * The hash table holds a reference to the QP.
231 */
232static void insert_qp(struct hfi1_ibdev *dev, struct hfi1_qp *qp)
233{
234	struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
235	unsigned long flags;
236
237	atomic_inc(&qp->refcount);
238	spin_lock_irqsave(&dev->qp_dev->qpt_lock, flags);
239
240	if (qp->ibqp.qp_num <= 1) {
241		rcu_assign_pointer(ibp->qp[qp->ibqp.qp_num], qp);
242	} else {
243		u32 n = qpn_hash(dev->qp_dev, qp->ibqp.qp_num);
244
245		qp->next = dev->qp_dev->qp_table[n];
246		rcu_assign_pointer(dev->qp_dev->qp_table[n], qp);
247		trace_hfi1_qpinsert(qp, n);
248	}
249
250	spin_unlock_irqrestore(&dev->qp_dev->qpt_lock, flags);
251}
252
253/*
254 * Remove the QP from the table so it can't be found asynchronously by
255 * the receive interrupt routine.
256 */
257static void remove_qp(struct hfi1_ibdev *dev, struct hfi1_qp *qp)
258{
259	struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
260	u32 n = qpn_hash(dev->qp_dev, qp->ibqp.qp_num);
261	unsigned long flags;
262	int removed = 1;
263
264	spin_lock_irqsave(&dev->qp_dev->qpt_lock, flags);
265
266	if (rcu_dereference_protected(ibp->qp[0],
267			lockdep_is_held(&dev->qp_dev->qpt_lock)) == qp) {
268		RCU_INIT_POINTER(ibp->qp[0], NULL);
269	} else if (rcu_dereference_protected(ibp->qp[1],
270			lockdep_is_held(&dev->qp_dev->qpt_lock)) == qp) {
271		RCU_INIT_POINTER(ibp->qp[1], NULL);
272	} else {
273		struct hfi1_qp *q;
274		struct hfi1_qp __rcu **qpp;
275
276		removed = 0;
277		qpp = &dev->qp_dev->qp_table[n];
278		for (; (q = rcu_dereference_protected(*qpp,
279				lockdep_is_held(&dev->qp_dev->qpt_lock)))
280					!= NULL;
281				qpp = &q->next)
282			if (q == qp) {
283				RCU_INIT_POINTER(*qpp,
284				 rcu_dereference_protected(qp->next,
285				 lockdep_is_held(&dev->qp_dev->qpt_lock)));
286				removed = 1;
287				trace_hfi1_qpremove(qp, n);
288				break;
289			}
290	}
291
292	spin_unlock_irqrestore(&dev->qp_dev->qpt_lock, flags);
293	if (removed) {
294		synchronize_rcu();
295		if (atomic_dec_and_test(&qp->refcount))
296			wake_up(&qp->wait);
297	}
298}
299
300/**
301 * free_all_qps - check for QPs still in use
302 * @qpt: the QP table to empty
303 *
304 * There should not be any QPs still in use.
305 * Free memory for table.
306 */
307static unsigned free_all_qps(struct hfi1_devdata *dd)
308{
309	struct hfi1_ibdev *dev = &dd->verbs_dev;
310	unsigned long flags;
311	struct hfi1_qp *qp;
312	unsigned n, qp_inuse = 0;
313
314	for (n = 0; n < dd->num_pports; n++) {
315		struct hfi1_ibport *ibp = &dd->pport[n].ibport_data;
316
317		if (!hfi1_mcast_tree_empty(ibp))
318			qp_inuse++;
319		rcu_read_lock();
320		if (rcu_dereference(ibp->qp[0]))
321			qp_inuse++;
322		if (rcu_dereference(ibp->qp[1]))
323			qp_inuse++;
324		rcu_read_unlock();
325	}
326
327	if (!dev->qp_dev)
328		goto bail;
329	spin_lock_irqsave(&dev->qp_dev->qpt_lock, flags);
330	for (n = 0; n < dev->qp_dev->qp_table_size; n++) {
331		qp = rcu_dereference_protected(dev->qp_dev->qp_table[n],
332			lockdep_is_held(&dev->qp_dev->qpt_lock));
333		RCU_INIT_POINTER(dev->qp_dev->qp_table[n], NULL);
334
335		for (; qp; qp = rcu_dereference_protected(qp->next,
336				lockdep_is_held(&dev->qp_dev->qpt_lock)))
337			qp_inuse++;
338	}
339	spin_unlock_irqrestore(&dev->qp_dev->qpt_lock, flags);
340	synchronize_rcu();
341bail:
342	return qp_inuse;
343}
344
345/**
346 * reset_qp - initialize the QP state to the reset state
347 * @qp: the QP to reset
348 * @type: the QP type
349 */
350static void reset_qp(struct hfi1_qp *qp, enum ib_qp_type type)
351{
352	qp->remote_qpn = 0;
353	qp->qkey = 0;
354	qp->qp_access_flags = 0;
355	iowait_init(
356		&qp->s_iowait,
357		1,
358		hfi1_do_send,
359		iowait_sleep,
360		iowait_wakeup);
361	qp->s_flags &= HFI1_S_SIGNAL_REQ_WR;
362	qp->s_hdrwords = 0;
363	qp->s_wqe = NULL;
364	qp->s_draining = 0;
365	qp->s_next_psn = 0;
366	qp->s_last_psn = 0;
367	qp->s_sending_psn = 0;
368	qp->s_sending_hpsn = 0;
369	qp->s_psn = 0;
370	qp->r_psn = 0;
371	qp->r_msn = 0;
372	if (type == IB_QPT_RC) {
373		qp->s_state = IB_OPCODE_RC_SEND_LAST;
374		qp->r_state = IB_OPCODE_RC_SEND_LAST;
375	} else {
376		qp->s_state = IB_OPCODE_UC_SEND_LAST;
377		qp->r_state = IB_OPCODE_UC_SEND_LAST;
378	}
379	qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
380	qp->r_nak_state = 0;
381	qp->r_aflags = 0;
382	qp->r_flags = 0;
383	qp->s_head = 0;
384	qp->s_tail = 0;
385	qp->s_cur = 0;
386	qp->s_acked = 0;
387	qp->s_last = 0;
388	qp->s_ssn = 1;
389	qp->s_lsn = 0;
390	clear_ahg(qp);
391	qp->s_mig_state = IB_MIG_MIGRATED;
392	memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue));
393	qp->r_head_ack_queue = 0;
394	qp->s_tail_ack_queue = 0;
395	qp->s_num_rd_atomic = 0;
396	if (qp->r_rq.wq) {
397		qp->r_rq.wq->head = 0;
398		qp->r_rq.wq->tail = 0;
399	}
400	qp->r_sge.num_sge = 0;
401}
402
403static void clear_mr_refs(struct hfi1_qp *qp, int clr_sends)
404{
405	unsigned n;
406
407	if (test_and_clear_bit(HFI1_R_REWIND_SGE, &qp->r_aflags))
408		hfi1_put_ss(&qp->s_rdma_read_sge);
409
410	hfi1_put_ss(&qp->r_sge);
411
412	if (clr_sends) {
413		while (qp->s_last != qp->s_head) {
414			struct hfi1_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
415			unsigned i;
416
417			for (i = 0; i < wqe->wr.num_sge; i++) {
418				struct hfi1_sge *sge = &wqe->sg_list[i];
419
420				hfi1_put_mr(sge->mr);
421			}
422			if (qp->ibqp.qp_type == IB_QPT_UD ||
423			    qp->ibqp.qp_type == IB_QPT_SMI ||
424			    qp->ibqp.qp_type == IB_QPT_GSI)
425				atomic_dec(&to_iah(wqe->ud_wr.ah)->refcount);
426			if (++qp->s_last >= qp->s_size)
427				qp->s_last = 0;
428		}
429		if (qp->s_rdma_mr) {
430			hfi1_put_mr(qp->s_rdma_mr);
431			qp->s_rdma_mr = NULL;
432		}
433	}
434
435	if (qp->ibqp.qp_type != IB_QPT_RC)
436		return;
437
438	for (n = 0; n < ARRAY_SIZE(qp->s_ack_queue); n++) {
439		struct hfi1_ack_entry *e = &qp->s_ack_queue[n];
440
441		if (e->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST &&
442		    e->rdma_sge.mr) {
443			hfi1_put_mr(e->rdma_sge.mr);
444			e->rdma_sge.mr = NULL;
445		}
446	}
447}
448
449/**
450 * hfi1_error_qp - put a QP into the error state
451 * @qp: the QP to put into the error state
452 * @err: the receive completion error to signal if a RWQE is active
453 *
454 * Flushes both send and receive work queues.
455 * Returns true if last WQE event should be generated.
456 * The QP r_lock and s_lock should be held and interrupts disabled.
457 * If we are already in error state, just return.
458 */
459int hfi1_error_qp(struct hfi1_qp *qp, enum ib_wc_status err)
460{
461	struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
462	struct ib_wc wc;
463	int ret = 0;
464
465	if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET)
466		goto bail;
467
468	qp->state = IB_QPS_ERR;
469
470	if (qp->s_flags & (HFI1_S_TIMER | HFI1_S_WAIT_RNR)) {
471		qp->s_flags &= ~(HFI1_S_TIMER | HFI1_S_WAIT_RNR);
472		del_timer(&qp->s_timer);
473	}
474
475	if (qp->s_flags & HFI1_S_ANY_WAIT_SEND)
476		qp->s_flags &= ~HFI1_S_ANY_WAIT_SEND;
477
478	write_seqlock(&dev->iowait_lock);
479	if (!list_empty(&qp->s_iowait.list) && !(qp->s_flags & HFI1_S_BUSY)) {
480		qp->s_flags &= ~HFI1_S_ANY_WAIT_IO;
481		list_del_init(&qp->s_iowait.list);
482		if (atomic_dec_and_test(&qp->refcount))
483			wake_up(&qp->wait);
484	}
485	write_sequnlock(&dev->iowait_lock);
486
487	if (!(qp->s_flags & HFI1_S_BUSY)) {
488		qp->s_hdrwords = 0;
489		if (qp->s_rdma_mr) {
490			hfi1_put_mr(qp->s_rdma_mr);
491			qp->s_rdma_mr = NULL;
492		}
493		flush_tx_list(qp);
494	}
495
496	/* Schedule the sending tasklet to drain the send work queue. */
497	if (qp->s_last != qp->s_head)
498		hfi1_schedule_send(qp);
499
500	clear_mr_refs(qp, 0);
501
502	memset(&wc, 0, sizeof(wc));
503	wc.qp = &qp->ibqp;
504	wc.opcode = IB_WC_RECV;
505
506	if (test_and_clear_bit(HFI1_R_WRID_VALID, &qp->r_aflags)) {
507		wc.wr_id = qp->r_wr_id;
508		wc.status = err;
509		hfi1_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
510	}
511	wc.status = IB_WC_WR_FLUSH_ERR;
512
513	if (qp->r_rq.wq) {
514		struct hfi1_rwq *wq;
515		u32 head;
516		u32 tail;
517
518		spin_lock(&qp->r_rq.lock);
519
520		/* sanity check pointers before trusting them */
521		wq = qp->r_rq.wq;
522		head = wq->head;
523		if (head >= qp->r_rq.size)
524			head = 0;
525		tail = wq->tail;
526		if (tail >= qp->r_rq.size)
527			tail = 0;
528		while (tail != head) {
529			wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
530			if (++tail >= qp->r_rq.size)
531				tail = 0;
532			hfi1_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
533		}
534		wq->tail = tail;
535
536		spin_unlock(&qp->r_rq.lock);
537	} else if (qp->ibqp.event_handler)
538		ret = 1;
539
540bail:
541	return ret;
542}
543
544static void flush_tx_list(struct hfi1_qp *qp)
545{
546	while (!list_empty(&qp->s_iowait.tx_head)) {
547		struct sdma_txreq *tx;
548
549		tx = list_first_entry(
550			&qp->s_iowait.tx_head,
551			struct sdma_txreq,
552			list);
553		list_del_init(&tx->list);
554		hfi1_put_txreq(
555			container_of(tx, struct verbs_txreq, txreq));
556	}
557}
558
559static void flush_iowait(struct hfi1_qp *qp)
560{
561	struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
562	unsigned long flags;
563
564	write_seqlock_irqsave(&dev->iowait_lock, flags);
565	if (!list_empty(&qp->s_iowait.list)) {
566		list_del_init(&qp->s_iowait.list);
567		if (atomic_dec_and_test(&qp->refcount))
568			wake_up(&qp->wait);
569	}
570	write_sequnlock_irqrestore(&dev->iowait_lock, flags);
571}
572
573static inline int opa_mtu_enum_to_int(int mtu)
574{
575	switch (mtu) {
576	case OPA_MTU_8192:  return 8192;
577	case OPA_MTU_10240: return 10240;
578	default:            return -1;
579	}
580}
581
582/**
583 * This function is what we would push to the core layer if we wanted to be a
584 * "first class citizen".  Instead we hide this here and rely on Verbs ULPs
585 * to blindly pass the MTU enum value from the PathRecord to us.
586 *
587 * The actual flag used to determine "8k MTU" will change and is currently
588 * unknown.
589 */
590static inline int verbs_mtu_enum_to_int(struct ib_device *dev, enum ib_mtu mtu)
591{
592	int val = opa_mtu_enum_to_int((int)mtu);
593
594	if (val > 0)
595		return val;
596	return ib_mtu_enum_to_int(mtu);
597}
598
599
600/**
601 * hfi1_modify_qp - modify the attributes of a queue pair
602 * @ibqp: the queue pair who's attributes we're modifying
603 * @attr: the new attributes
604 * @attr_mask: the mask of attributes to modify
605 * @udata: user data for libibverbs.so
606 *
607 * Returns 0 on success, otherwise returns an errno.
608 */
609int hfi1_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
610		   int attr_mask, struct ib_udata *udata)
611{
612	struct hfi1_ibdev *dev = to_idev(ibqp->device);
613	struct hfi1_qp *qp = to_iqp(ibqp);
614	enum ib_qp_state cur_state, new_state;
615	struct ib_event ev;
616	int lastwqe = 0;
617	int mig = 0;
618	int ret;
619	u32 pmtu = 0; /* for gcc warning only */
620	struct hfi1_devdata *dd;
621
622	spin_lock_irq(&qp->r_lock);
623	spin_lock(&qp->s_lock);
624
625	cur_state = attr_mask & IB_QP_CUR_STATE ?
626		attr->cur_qp_state : qp->state;
627	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
628
629	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
630				attr_mask, IB_LINK_LAYER_UNSPECIFIED))
631		goto inval;
632
633	if (attr_mask & IB_QP_AV) {
634		if (attr->ah_attr.dlid >= HFI1_MULTICAST_LID_BASE)
635			goto inval;
636		if (hfi1_check_ah(qp->ibqp.device, &attr->ah_attr))
637			goto inval;
638	}
639
640	if (attr_mask & IB_QP_ALT_PATH) {
641		if (attr->alt_ah_attr.dlid >= HFI1_MULTICAST_LID_BASE)
642			goto inval;
643		if (hfi1_check_ah(qp->ibqp.device, &attr->alt_ah_attr))
644			goto inval;
645		if (attr->alt_pkey_index >= hfi1_get_npkeys(dd_from_dev(dev)))
646			goto inval;
647	}
648
649	if (attr_mask & IB_QP_PKEY_INDEX)
650		if (attr->pkey_index >= hfi1_get_npkeys(dd_from_dev(dev)))
651			goto inval;
652
653	if (attr_mask & IB_QP_MIN_RNR_TIMER)
654		if (attr->min_rnr_timer > 31)
655			goto inval;
656
657	if (attr_mask & IB_QP_PORT)
658		if (qp->ibqp.qp_type == IB_QPT_SMI ||
659		    qp->ibqp.qp_type == IB_QPT_GSI ||
660		    attr->port_num == 0 ||
661		    attr->port_num > ibqp->device->phys_port_cnt)
662			goto inval;
663
664	if (attr_mask & IB_QP_DEST_QPN)
665		if (attr->dest_qp_num > HFI1_QPN_MASK)
666			goto inval;
667
668	if (attr_mask & IB_QP_RETRY_CNT)
669		if (attr->retry_cnt > 7)
670			goto inval;
671
672	if (attr_mask & IB_QP_RNR_RETRY)
673		if (attr->rnr_retry > 7)
674			goto inval;
675
676	/*
677	 * Don't allow invalid path_mtu values.  OK to set greater
678	 * than the active mtu (or even the max_cap, if we have tuned
679	 * that to a small mtu.  We'll set qp->path_mtu
680	 * to the lesser of requested attribute mtu and active,
681	 * for packetizing messages.
682	 * Note that the QP port has to be set in INIT and MTU in RTR.
683	 */
684	if (attr_mask & IB_QP_PATH_MTU) {
685		int mtu, pidx = qp->port_num - 1;
686
687		dd = dd_from_dev(dev);
688		mtu = verbs_mtu_enum_to_int(ibqp->device, attr->path_mtu);
689		if (mtu == -1)
690			goto inval;
691
692		if (mtu > dd->pport[pidx].ibmtu)
693			pmtu = mtu_to_enum(dd->pport[pidx].ibmtu, IB_MTU_2048);
694		else
695			pmtu = attr->path_mtu;
696	}
697
698	if (attr_mask & IB_QP_PATH_MIG_STATE) {
699		if (attr->path_mig_state == IB_MIG_REARM) {
700			if (qp->s_mig_state == IB_MIG_ARMED)
701				goto inval;
702			if (new_state != IB_QPS_RTS)
703				goto inval;
704		} else if (attr->path_mig_state == IB_MIG_MIGRATED) {
705			if (qp->s_mig_state == IB_MIG_REARM)
706				goto inval;
707			if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD)
708				goto inval;
709			if (qp->s_mig_state == IB_MIG_ARMED)
710				mig = 1;
711		} else
712			goto inval;
713	}
714
715	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
716		if (attr->max_dest_rd_atomic > HFI1_MAX_RDMA_ATOMIC)
717			goto inval;
718
719	switch (new_state) {
720	case IB_QPS_RESET:
721		if (qp->state != IB_QPS_RESET) {
722			qp->state = IB_QPS_RESET;
723			flush_iowait(qp);
724			qp->s_flags &= ~(HFI1_S_TIMER | HFI1_S_ANY_WAIT);
725			spin_unlock(&qp->s_lock);
726			spin_unlock_irq(&qp->r_lock);
727			/* Stop the sending work queue and retry timer */
728			cancel_work_sync(&qp->s_iowait.iowork);
729			del_timer_sync(&qp->s_timer);
730			iowait_sdma_drain(&qp->s_iowait);
731			flush_tx_list(qp);
732			remove_qp(dev, qp);
733			wait_event(qp->wait, !atomic_read(&qp->refcount));
734			spin_lock_irq(&qp->r_lock);
735			spin_lock(&qp->s_lock);
736			clear_mr_refs(qp, 1);
737			clear_ahg(qp);
738			reset_qp(qp, ibqp->qp_type);
739		}
740		break;
741
742	case IB_QPS_RTR:
743		/* Allow event to re-trigger if QP set to RTR more than once */
744		qp->r_flags &= ~HFI1_R_COMM_EST;
745		qp->state = new_state;
746		break;
747
748	case IB_QPS_SQD:
749		qp->s_draining = qp->s_last != qp->s_cur;
750		qp->state = new_state;
751		break;
752
753	case IB_QPS_SQE:
754		if (qp->ibqp.qp_type == IB_QPT_RC)
755			goto inval;
756		qp->state = new_state;
757		break;
758
759	case IB_QPS_ERR:
760		lastwqe = hfi1_error_qp(qp, IB_WC_WR_FLUSH_ERR);
761		break;
762
763	default:
764		qp->state = new_state;
765		break;
766	}
767
768	if (attr_mask & IB_QP_PKEY_INDEX)
769		qp->s_pkey_index = attr->pkey_index;
770
771	if (attr_mask & IB_QP_PORT)
772		qp->port_num = attr->port_num;
773
774	if (attr_mask & IB_QP_DEST_QPN)
775		qp->remote_qpn = attr->dest_qp_num;
776
777	if (attr_mask & IB_QP_SQ_PSN) {
778		qp->s_next_psn = attr->sq_psn & PSN_MODIFY_MASK;
779		qp->s_psn = qp->s_next_psn;
780		qp->s_sending_psn = qp->s_next_psn;
781		qp->s_last_psn = qp->s_next_psn - 1;
782		qp->s_sending_hpsn = qp->s_last_psn;
783	}
784
785	if (attr_mask & IB_QP_RQ_PSN)
786		qp->r_psn = attr->rq_psn & PSN_MODIFY_MASK;
787
788	if (attr_mask & IB_QP_ACCESS_FLAGS)
789		qp->qp_access_flags = attr->qp_access_flags;
790
791	if (attr_mask & IB_QP_AV) {
792		qp->remote_ah_attr = attr->ah_attr;
793		qp->s_srate = attr->ah_attr.static_rate;
794		qp->srate_mbps = ib_rate_to_mbps(qp->s_srate);
795	}
796
797	if (attr_mask & IB_QP_ALT_PATH) {
798		qp->alt_ah_attr = attr->alt_ah_attr;
799		qp->s_alt_pkey_index = attr->alt_pkey_index;
800	}
801
802	if (attr_mask & IB_QP_PATH_MIG_STATE) {
803		qp->s_mig_state = attr->path_mig_state;
804		if (mig) {
805			qp->remote_ah_attr = qp->alt_ah_attr;
806			qp->port_num = qp->alt_ah_attr.port_num;
807			qp->s_pkey_index = qp->s_alt_pkey_index;
808			qp->s_flags |= HFI1_S_AHG_CLEAR;
809		}
810	}
811
812	if (attr_mask & IB_QP_PATH_MTU) {
813		struct hfi1_ibport *ibp;
814		u8 sc, vl;
815		u32 mtu;
816
817		dd = dd_from_dev(dev);
818		ibp = &dd->pport[qp->port_num - 1].ibport_data;
819
820		sc = ibp->sl_to_sc[qp->remote_ah_attr.sl];
821		vl = sc_to_vlt(dd, sc);
822
823		mtu = verbs_mtu_enum_to_int(ibqp->device, pmtu);
824		if (vl < PER_VL_SEND_CONTEXTS)
825			mtu = min_t(u32, mtu, dd->vld[vl].mtu);
826		pmtu = mtu_to_enum(mtu, OPA_MTU_8192);
827
828		qp->path_mtu = pmtu;
829		qp->pmtu = mtu;
830	}
831
832	if (attr_mask & IB_QP_RETRY_CNT) {
833		qp->s_retry_cnt = attr->retry_cnt;
834		qp->s_retry = attr->retry_cnt;
835	}
836
837	if (attr_mask & IB_QP_RNR_RETRY) {
838		qp->s_rnr_retry_cnt = attr->rnr_retry;
839		qp->s_rnr_retry = attr->rnr_retry;
840	}
841
842	if (attr_mask & IB_QP_MIN_RNR_TIMER)
843		qp->r_min_rnr_timer = attr->min_rnr_timer;
844
845	if (attr_mask & IB_QP_TIMEOUT) {
846		qp->timeout = attr->timeout;
847		qp->timeout_jiffies =
848			usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
849				1000UL);
850	}
851
852	if (attr_mask & IB_QP_QKEY)
853		qp->qkey = attr->qkey;
854
855	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
856		qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
857
858	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
859		qp->s_max_rd_atomic = attr->max_rd_atomic;
860
861	spin_unlock(&qp->s_lock);
862	spin_unlock_irq(&qp->r_lock);
863
864	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
865		insert_qp(dev, qp);
866
867	if (lastwqe) {
868		ev.device = qp->ibqp.device;
869		ev.element.qp = &qp->ibqp;
870		ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
871		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
872	}
873	if (mig) {
874		ev.device = qp->ibqp.device;
875		ev.element.qp = &qp->ibqp;
876		ev.event = IB_EVENT_PATH_MIG;
877		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
878	}
879	ret = 0;
880	goto bail;
881
882inval:
883	spin_unlock(&qp->s_lock);
884	spin_unlock_irq(&qp->r_lock);
885	ret = -EINVAL;
886
887bail:
888	return ret;
889}
890
891int hfi1_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
892		  int attr_mask, struct ib_qp_init_attr *init_attr)
893{
894	struct hfi1_qp *qp = to_iqp(ibqp);
895
896	attr->qp_state = qp->state;
897	attr->cur_qp_state = attr->qp_state;
898	attr->path_mtu = qp->path_mtu;
899	attr->path_mig_state = qp->s_mig_state;
900	attr->qkey = qp->qkey;
901	attr->rq_psn = mask_psn(qp->r_psn);
902	attr->sq_psn = mask_psn(qp->s_next_psn);
903	attr->dest_qp_num = qp->remote_qpn;
904	attr->qp_access_flags = qp->qp_access_flags;
905	attr->cap.max_send_wr = qp->s_size - 1;
906	attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
907	attr->cap.max_send_sge = qp->s_max_sge;
908	attr->cap.max_recv_sge = qp->r_rq.max_sge;
909	attr->cap.max_inline_data = 0;
910	attr->ah_attr = qp->remote_ah_attr;
911	attr->alt_ah_attr = qp->alt_ah_attr;
912	attr->pkey_index = qp->s_pkey_index;
913	attr->alt_pkey_index = qp->s_alt_pkey_index;
914	attr->en_sqd_async_notify = 0;
915	attr->sq_draining = qp->s_draining;
916	attr->max_rd_atomic = qp->s_max_rd_atomic;
917	attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
918	attr->min_rnr_timer = qp->r_min_rnr_timer;
919	attr->port_num = qp->port_num;
920	attr->timeout = qp->timeout;
921	attr->retry_cnt = qp->s_retry_cnt;
922	attr->rnr_retry = qp->s_rnr_retry_cnt;
923	attr->alt_port_num = qp->alt_ah_attr.port_num;
924	attr->alt_timeout = qp->alt_timeout;
925
926	init_attr->event_handler = qp->ibqp.event_handler;
927	init_attr->qp_context = qp->ibqp.qp_context;
928	init_attr->send_cq = qp->ibqp.send_cq;
929	init_attr->recv_cq = qp->ibqp.recv_cq;
930	init_attr->srq = qp->ibqp.srq;
931	init_attr->cap = attr->cap;
932	if (qp->s_flags & HFI1_S_SIGNAL_REQ_WR)
933		init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
934	else
935		init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
936	init_attr->qp_type = qp->ibqp.qp_type;
937	init_attr->port_num = qp->port_num;
938	return 0;
939}
940
941/**
942 * hfi1_compute_aeth - compute the AETH (syndrome + MSN)
943 * @qp: the queue pair to compute the AETH for
944 *
945 * Returns the AETH.
946 */
947__be32 hfi1_compute_aeth(struct hfi1_qp *qp)
948{
949	u32 aeth = qp->r_msn & HFI1_MSN_MASK;
950
951	if (qp->ibqp.srq) {
952		/*
953		 * Shared receive queues don't generate credits.
954		 * Set the credit field to the invalid value.
955		 */
956		aeth |= HFI1_AETH_CREDIT_INVAL << HFI1_AETH_CREDIT_SHIFT;
957	} else {
958		u32 min, max, x;
959		u32 credits;
960		struct hfi1_rwq *wq = qp->r_rq.wq;
961		u32 head;
962		u32 tail;
963
964		/* sanity check pointers before trusting them */
965		head = wq->head;
966		if (head >= qp->r_rq.size)
967			head = 0;
968		tail = wq->tail;
969		if (tail >= qp->r_rq.size)
970			tail = 0;
971		/*
972		 * Compute the number of credits available (RWQEs).
973		 * There is a small chance that the pair of reads are
974		 * not atomic, which is OK, since the fuzziness is
975		 * resolved as further ACKs go out.
976		 */
977		credits = head - tail;
978		if ((int)credits < 0)
979			credits += qp->r_rq.size;
980		/*
981		 * Binary search the credit table to find the code to
982		 * use.
983		 */
984		min = 0;
985		max = 31;
986		for (;;) {
987			x = (min + max) / 2;
988			if (credit_table[x] == credits)
989				break;
990			if (credit_table[x] > credits)
991				max = x;
992			else if (min == x)
993				break;
994			else
995				min = x;
996		}
997		aeth |= x << HFI1_AETH_CREDIT_SHIFT;
998	}
999	return cpu_to_be32(aeth);
1000}
1001
1002/**
1003 * hfi1_create_qp - create a queue pair for a device
1004 * @ibpd: the protection domain who's device we create the queue pair for
1005 * @init_attr: the attributes of the queue pair
1006 * @udata: user data for libibverbs.so
1007 *
1008 * Returns the queue pair on success, otherwise returns an errno.
1009 *
1010 * Called by the ib_create_qp() core verbs function.
1011 */
1012struct ib_qp *hfi1_create_qp(struct ib_pd *ibpd,
1013			     struct ib_qp_init_attr *init_attr,
1014			     struct ib_udata *udata)
1015{
1016	struct hfi1_qp *qp;
1017	int err;
1018	struct hfi1_swqe *swq = NULL;
1019	struct hfi1_ibdev *dev;
1020	struct hfi1_devdata *dd;
1021	size_t sz;
1022	size_t sg_list_sz;
1023	struct ib_qp *ret;
1024
1025	if (init_attr->cap.max_send_sge > hfi1_max_sges ||
1026	    init_attr->cap.max_send_wr > hfi1_max_qp_wrs ||
1027	    init_attr->create_flags) {
1028		ret = ERR_PTR(-EINVAL);
1029		goto bail;
1030	}
1031
1032	/* Check receive queue parameters if no SRQ is specified. */
1033	if (!init_attr->srq) {
1034		if (init_attr->cap.max_recv_sge > hfi1_max_sges ||
1035		    init_attr->cap.max_recv_wr > hfi1_max_qp_wrs) {
1036			ret = ERR_PTR(-EINVAL);
1037			goto bail;
1038		}
1039		if (init_attr->cap.max_send_sge +
1040		    init_attr->cap.max_send_wr +
1041		    init_attr->cap.max_recv_sge +
1042		    init_attr->cap.max_recv_wr == 0) {
1043			ret = ERR_PTR(-EINVAL);
1044			goto bail;
1045		}
1046	}
1047
1048	switch (init_attr->qp_type) {
1049	case IB_QPT_SMI:
1050	case IB_QPT_GSI:
1051		if (init_attr->port_num == 0 ||
1052		    init_attr->port_num > ibpd->device->phys_port_cnt) {
1053			ret = ERR_PTR(-EINVAL);
1054			goto bail;
1055		}
1056	case IB_QPT_UC:
1057	case IB_QPT_RC:
1058	case IB_QPT_UD:
1059		sz = sizeof(struct hfi1_sge) *
1060			init_attr->cap.max_send_sge +
1061			sizeof(struct hfi1_swqe);
1062		swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz);
1063		if (swq == NULL) {
1064			ret = ERR_PTR(-ENOMEM);
1065			goto bail;
1066		}
1067		sz = sizeof(*qp);
1068		sg_list_sz = 0;
1069		if (init_attr->srq) {
1070			struct hfi1_srq *srq = to_isrq(init_attr->srq);
1071
1072			if (srq->rq.max_sge > 1)
1073				sg_list_sz = sizeof(*qp->r_sg_list) *
1074					(srq->rq.max_sge - 1);
1075		} else if (init_attr->cap.max_recv_sge > 1)
1076			sg_list_sz = sizeof(*qp->r_sg_list) *
1077				(init_attr->cap.max_recv_sge - 1);
1078		qp = kzalloc(sz + sg_list_sz, GFP_KERNEL);
1079		if (!qp) {
1080			ret = ERR_PTR(-ENOMEM);
1081			goto bail_swq;
1082		}
1083		RCU_INIT_POINTER(qp->next, NULL);
1084		qp->s_hdr = kzalloc(sizeof(*qp->s_hdr), GFP_KERNEL);
1085		if (!qp->s_hdr) {
1086			ret = ERR_PTR(-ENOMEM);
1087			goto bail_qp;
1088		}
1089		qp->timeout_jiffies =
1090			usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
1091				1000UL);
1092		if (init_attr->srq)
1093			sz = 0;
1094		else {
1095			qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
1096			qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
1097			sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
1098				sizeof(struct hfi1_rwqe);
1099			qp->r_rq.wq = vmalloc_user(sizeof(struct hfi1_rwq) +
1100						   qp->r_rq.size * sz);
1101			if (!qp->r_rq.wq) {
1102				ret = ERR_PTR(-ENOMEM);
1103				goto bail_qp;
1104			}
1105		}
1106
1107		/*
1108		 * ib_create_qp() will initialize qp->ibqp
1109		 * except for qp->ibqp.qp_num.
1110		 */
1111		spin_lock_init(&qp->r_lock);
1112		spin_lock_init(&qp->s_lock);
1113		spin_lock_init(&qp->r_rq.lock);
1114		atomic_set(&qp->refcount, 0);
1115		init_waitqueue_head(&qp->wait);
1116		init_timer(&qp->s_timer);
1117		qp->s_timer.data = (unsigned long)qp;
1118		INIT_LIST_HEAD(&qp->rspwait);
1119		qp->state = IB_QPS_RESET;
1120		qp->s_wq = swq;
1121		qp->s_size = init_attr->cap.max_send_wr + 1;
1122		qp->s_max_sge = init_attr->cap.max_send_sge;
1123		if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
1124			qp->s_flags = HFI1_S_SIGNAL_REQ_WR;
1125		dev = to_idev(ibpd->device);
1126		dd = dd_from_dev(dev);
1127		err = alloc_qpn(dd, &dev->qp_dev->qpn_table, init_attr->qp_type,
1128				init_attr->port_num);
1129		if (err < 0) {
1130			ret = ERR_PTR(err);
1131			vfree(qp->r_rq.wq);
1132			goto bail_qp;
1133		}
1134		qp->ibqp.qp_num = err;
1135		qp->port_num = init_attr->port_num;
1136		reset_qp(qp, init_attr->qp_type);
1137
1138		break;
1139
1140	default:
1141		/* Don't support raw QPs */
1142		ret = ERR_PTR(-ENOSYS);
1143		goto bail;
1144	}
1145
1146	init_attr->cap.max_inline_data = 0;
1147
1148	/*
1149	 * Return the address of the RWQ as the offset to mmap.
1150	 * See hfi1_mmap() for details.
1151	 */
1152	if (udata && udata->outlen >= sizeof(__u64)) {
1153		if (!qp->r_rq.wq) {
1154			__u64 offset = 0;
1155
1156			err = ib_copy_to_udata(udata, &offset,
1157					       sizeof(offset));
1158			if (err) {
1159				ret = ERR_PTR(err);
1160				goto bail_ip;
1161			}
1162		} else {
1163			u32 s = sizeof(struct hfi1_rwq) + qp->r_rq.size * sz;
1164
1165			qp->ip = hfi1_create_mmap_info(dev, s,
1166						      ibpd->uobject->context,
1167						      qp->r_rq.wq);
1168			if (!qp->ip) {
1169				ret = ERR_PTR(-ENOMEM);
1170				goto bail_ip;
1171			}
1172
1173			err = ib_copy_to_udata(udata, &(qp->ip->offset),
1174					       sizeof(qp->ip->offset));
1175			if (err) {
1176				ret = ERR_PTR(err);
1177				goto bail_ip;
1178			}
1179		}
1180	}
1181
1182	spin_lock(&dev->n_qps_lock);
1183	if (dev->n_qps_allocated == hfi1_max_qps) {
1184		spin_unlock(&dev->n_qps_lock);
1185		ret = ERR_PTR(-ENOMEM);
1186		goto bail_ip;
1187	}
1188
1189	dev->n_qps_allocated++;
1190	spin_unlock(&dev->n_qps_lock);
1191
1192	if (qp->ip) {
1193		spin_lock_irq(&dev->pending_lock);
1194		list_add(&qp->ip->pending_mmaps, &dev->pending_mmaps);
1195		spin_unlock_irq(&dev->pending_lock);
1196	}
1197
1198	ret = &qp->ibqp;
1199
1200	/*
1201	 * We have our QP and its good, now keep track of what types of opcodes
1202	 * can be processed on this QP. We do this by keeping track of what the
1203	 * 3 high order bits of the opcode are.
1204	 */
1205	switch (init_attr->qp_type) {
1206	case IB_QPT_SMI:
1207	case IB_QPT_GSI:
1208	case IB_QPT_UD:
1209		qp->allowed_ops = IB_OPCODE_UD_SEND_ONLY & OPCODE_QP_MASK;
1210		break;
1211	case IB_QPT_RC:
1212		qp->allowed_ops = IB_OPCODE_RC_SEND_ONLY & OPCODE_QP_MASK;
1213		break;
1214	case IB_QPT_UC:
1215		qp->allowed_ops = IB_OPCODE_UC_SEND_ONLY & OPCODE_QP_MASK;
1216		break;
1217	default:
1218		ret = ERR_PTR(-EINVAL);
1219		goto bail_ip;
1220	}
1221
1222	goto bail;
1223
1224bail_ip:
1225	if (qp->ip)
1226		kref_put(&qp->ip->ref, hfi1_release_mmap_info);
1227	else
1228		vfree(qp->r_rq.wq);
1229	free_qpn(&dev->qp_dev->qpn_table, qp->ibqp.qp_num);
1230bail_qp:
1231	kfree(qp->s_hdr);
1232	kfree(qp);
1233bail_swq:
1234	vfree(swq);
1235bail:
1236	return ret;
1237}
1238
1239/**
1240 * hfi1_destroy_qp - destroy a queue pair
1241 * @ibqp: the queue pair to destroy
1242 *
1243 * Returns 0 on success.
1244 *
1245 * Note that this can be called while the QP is actively sending or
1246 * receiving!
1247 */
1248int hfi1_destroy_qp(struct ib_qp *ibqp)
1249{
1250	struct hfi1_qp *qp = to_iqp(ibqp);
1251	struct hfi1_ibdev *dev = to_idev(ibqp->device);
1252
1253	/* Make sure HW and driver activity is stopped. */
1254	spin_lock_irq(&qp->r_lock);
1255	spin_lock(&qp->s_lock);
1256	if (qp->state != IB_QPS_RESET) {
1257		qp->state = IB_QPS_RESET;
1258		flush_iowait(qp);
1259		qp->s_flags &= ~(HFI1_S_TIMER | HFI1_S_ANY_WAIT);
1260		spin_unlock(&qp->s_lock);
1261		spin_unlock_irq(&qp->r_lock);
1262		cancel_work_sync(&qp->s_iowait.iowork);
1263		del_timer_sync(&qp->s_timer);
1264		iowait_sdma_drain(&qp->s_iowait);
1265		flush_tx_list(qp);
1266		remove_qp(dev, qp);
1267		wait_event(qp->wait, !atomic_read(&qp->refcount));
1268		spin_lock_irq(&qp->r_lock);
1269		spin_lock(&qp->s_lock);
1270		clear_mr_refs(qp, 1);
1271		clear_ahg(qp);
1272	}
1273	spin_unlock(&qp->s_lock);
1274	spin_unlock_irq(&qp->r_lock);
1275
1276	/* all user's cleaned up, mark it available */
1277	free_qpn(&dev->qp_dev->qpn_table, qp->ibqp.qp_num);
1278	spin_lock(&dev->n_qps_lock);
1279	dev->n_qps_allocated--;
1280	spin_unlock(&dev->n_qps_lock);
1281
1282	if (qp->ip)
1283		kref_put(&qp->ip->ref, hfi1_release_mmap_info);
1284	else
1285		vfree(qp->r_rq.wq);
1286	vfree(qp->s_wq);
1287	kfree(qp->s_hdr);
1288	kfree(qp);
1289	return 0;
1290}
1291
1292/**
1293 * init_qpn_table - initialize the QP number table for a device
1294 * @qpt: the QPN table
1295 */
1296static int init_qpn_table(struct hfi1_devdata *dd, struct hfi1_qpn_table *qpt)
1297{
1298	u32 offset, qpn, i;
1299	struct qpn_map *map;
1300	int ret = 0;
1301
1302	spin_lock_init(&qpt->lock);
1303
1304	qpt->last = 0;
1305	qpt->incr = 1 << dd->qos_shift;
1306
1307	/* insure we don't assign QPs from KDETH 64K window */
1308	qpn = kdeth_qp << 16;
1309	qpt->nmaps = qpn / BITS_PER_PAGE;
1310	/* This should always be zero */
1311	offset = qpn & BITS_PER_PAGE_MASK;
1312	map = &qpt->map[qpt->nmaps];
1313	dd_dev_info(dd, "Reserving QPNs for KDETH window from 0x%x to 0x%x\n",
1314		qpn, qpn + 65535);
1315	for (i = 0; i < 65536; i++) {
1316		if (!map->page) {
1317			get_map_page(qpt, map);
1318			if (!map->page) {
1319				ret = -ENOMEM;
1320				break;
1321			}
1322		}
1323		set_bit(offset, map->page);
1324		offset++;
1325		if (offset == BITS_PER_PAGE) {
1326			/* next page */
1327			qpt->nmaps++;
1328			map++;
1329			offset = 0;
1330		}
1331	}
1332	return ret;
1333}
1334
1335/**
1336 * free_qpn_table - free the QP number table for a device
1337 * @qpt: the QPN table
1338 */
1339static void free_qpn_table(struct hfi1_qpn_table *qpt)
1340{
1341	int i;
1342
1343	for (i = 0; i < ARRAY_SIZE(qpt->map); i++)
1344		free_page((unsigned long) qpt->map[i].page);
1345}
1346
1347/**
1348 * hfi1_get_credit - flush the send work queue of a QP
1349 * @qp: the qp who's send work queue to flush
1350 * @aeth: the Acknowledge Extended Transport Header
1351 *
1352 * The QP s_lock should be held.
1353 */
1354void hfi1_get_credit(struct hfi1_qp *qp, u32 aeth)
1355{
1356	u32 credit = (aeth >> HFI1_AETH_CREDIT_SHIFT) & HFI1_AETH_CREDIT_MASK;
1357
1358	/*
1359	 * If the credit is invalid, we can send
1360	 * as many packets as we like.  Otherwise, we have to
1361	 * honor the credit field.
1362	 */
1363	if (credit == HFI1_AETH_CREDIT_INVAL) {
1364		if (!(qp->s_flags & HFI1_S_UNLIMITED_CREDIT)) {
1365			qp->s_flags |= HFI1_S_UNLIMITED_CREDIT;
1366			if (qp->s_flags & HFI1_S_WAIT_SSN_CREDIT) {
1367				qp->s_flags &= ~HFI1_S_WAIT_SSN_CREDIT;
1368				hfi1_schedule_send(qp);
1369			}
1370		}
1371	} else if (!(qp->s_flags & HFI1_S_UNLIMITED_CREDIT)) {
1372		/* Compute new LSN (i.e., MSN + credit) */
1373		credit = (aeth + credit_table[credit]) & HFI1_MSN_MASK;
1374		if (cmp_msn(credit, qp->s_lsn) > 0) {
1375			qp->s_lsn = credit;
1376			if (qp->s_flags & HFI1_S_WAIT_SSN_CREDIT) {
1377				qp->s_flags &= ~HFI1_S_WAIT_SSN_CREDIT;
1378				hfi1_schedule_send(qp);
1379			}
1380		}
1381	}
1382}
1383
1384void hfi1_qp_wakeup(struct hfi1_qp *qp, u32 flag)
1385{
1386	unsigned long flags;
1387
1388	spin_lock_irqsave(&qp->s_lock, flags);
1389	if (qp->s_flags & flag) {
1390		qp->s_flags &= ~flag;
1391		trace_hfi1_qpwakeup(qp, flag);
1392		hfi1_schedule_send(qp);
1393	}
1394	spin_unlock_irqrestore(&qp->s_lock, flags);
1395	/* Notify hfi1_destroy_qp() if it is waiting. */
1396	if (atomic_dec_and_test(&qp->refcount))
1397		wake_up(&qp->wait);
1398}
1399
1400static int iowait_sleep(
1401	struct sdma_engine *sde,
1402	struct iowait *wait,
1403	struct sdma_txreq *stx,
1404	unsigned seq)
1405{
1406	struct verbs_txreq *tx = container_of(stx, struct verbs_txreq, txreq);
1407	struct hfi1_qp *qp;
1408	unsigned long flags;
1409	int ret = 0;
1410	struct hfi1_ibdev *dev;
1411
1412	qp = tx->qp;
1413
1414	spin_lock_irqsave(&qp->s_lock, flags);
1415	if (ib_hfi1_state_ops[qp->state] & HFI1_PROCESS_RECV_OK) {
1416
1417		/*
1418		 * If we couldn't queue the DMA request, save the info
1419		 * and try again later rather than destroying the
1420		 * buffer and undoing the side effects of the copy.
1421		 */
1422		/* Make a common routine? */
1423		dev = &sde->dd->verbs_dev;
1424		list_add_tail(&stx->list, &wait->tx_head);
1425		write_seqlock(&dev->iowait_lock);
1426		if (sdma_progress(sde, seq, stx))
1427			goto eagain;
1428		if (list_empty(&qp->s_iowait.list)) {
1429			struct hfi1_ibport *ibp =
1430				to_iport(qp->ibqp.device, qp->port_num);
1431
1432			ibp->n_dmawait++;
1433			qp->s_flags |= HFI1_S_WAIT_DMA_DESC;
1434			list_add_tail(&qp->s_iowait.list, &sde->dmawait);
1435			trace_hfi1_qpsleep(qp, HFI1_S_WAIT_DMA_DESC);
1436			atomic_inc(&qp->refcount);
1437		}
1438		write_sequnlock(&dev->iowait_lock);
1439		qp->s_flags &= ~HFI1_S_BUSY;
1440		spin_unlock_irqrestore(&qp->s_lock, flags);
1441		ret = -EBUSY;
1442	} else {
1443		spin_unlock_irqrestore(&qp->s_lock, flags);
1444		hfi1_put_txreq(tx);
1445	}
1446	return ret;
1447eagain:
1448	write_sequnlock(&dev->iowait_lock);
1449	spin_unlock_irqrestore(&qp->s_lock, flags);
1450	list_del_init(&stx->list);
1451	return -EAGAIN;
1452}
1453
1454static void iowait_wakeup(struct iowait *wait, int reason)
1455{
1456	struct hfi1_qp *qp = container_of(wait, struct hfi1_qp, s_iowait);
1457
1458	WARN_ON(reason != SDMA_AVAIL_REASON);
1459	hfi1_qp_wakeup(qp, HFI1_S_WAIT_DMA_DESC);
1460}
1461
1462int hfi1_qp_init(struct hfi1_ibdev *dev)
1463{
1464	struct hfi1_devdata *dd = dd_from_dev(dev);
1465	int i;
1466	int ret = -ENOMEM;
1467
1468	/* allocate parent object */
1469	dev->qp_dev = kzalloc(sizeof(*dev->qp_dev), GFP_KERNEL);
1470	if (!dev->qp_dev)
1471		goto nomem;
1472	/* allocate hash table */
1473	dev->qp_dev->qp_table_size = hfi1_qp_table_size;
1474	dev->qp_dev->qp_table_bits = ilog2(hfi1_qp_table_size);
1475	dev->qp_dev->qp_table =
1476		kmalloc(dev->qp_dev->qp_table_size *
1477				sizeof(*dev->qp_dev->qp_table),
1478			GFP_KERNEL);
1479	if (!dev->qp_dev->qp_table)
1480		goto nomem;
1481	for (i = 0; i < dev->qp_dev->qp_table_size; i++)
1482		RCU_INIT_POINTER(dev->qp_dev->qp_table[i], NULL);
1483	spin_lock_init(&dev->qp_dev->qpt_lock);
1484	/* initialize qpn map */
1485	ret = init_qpn_table(dd, &dev->qp_dev->qpn_table);
1486	if (ret)
1487		goto nomem;
1488	return ret;
1489nomem:
1490	if (dev->qp_dev) {
1491		kfree(dev->qp_dev->qp_table);
1492		free_qpn_table(&dev->qp_dev->qpn_table);
1493		kfree(dev->qp_dev);
1494	}
1495	return ret;
1496}
1497
1498void hfi1_qp_exit(struct hfi1_ibdev *dev)
1499{
1500	struct hfi1_devdata *dd = dd_from_dev(dev);
1501	u32 qps_inuse;
1502
1503	qps_inuse = free_all_qps(dd);
1504	if (qps_inuse)
1505		dd_dev_err(dd, "QP memory leak! %u still in use\n",
1506			   qps_inuse);
1507	if (dev->qp_dev) {
1508		kfree(dev->qp_dev->qp_table);
1509		free_qpn_table(&dev->qp_dev->qpn_table);
1510		kfree(dev->qp_dev);
1511	}
1512}
1513
1514/**
1515 *
1516 * qp_to_sdma_engine - map a qp to a send engine
1517 * @qp: the QP
1518 * @sc5: the 5 bit sc
1519 *
1520 * Return:
1521 * A send engine for the qp or NULL for SMI type qp.
1522 */
1523struct sdma_engine *qp_to_sdma_engine(struct hfi1_qp *qp, u8 sc5)
1524{
1525	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1526	struct sdma_engine *sde;
1527
1528	if (!(dd->flags & HFI1_HAS_SEND_DMA))
1529		return NULL;
1530	switch (qp->ibqp.qp_type) {
1531	case IB_QPT_UC:
1532	case IB_QPT_RC:
1533		break;
1534	case IB_QPT_SMI:
1535		return NULL;
1536	default:
1537		break;
1538	}
1539	sde = sdma_select_engine_sc(dd, qp->ibqp.qp_num >> dd->qos_shift, sc5);
1540	return sde;
1541}
1542
1543struct qp_iter {
1544	struct hfi1_ibdev *dev;
1545	struct hfi1_qp *qp;
1546	int specials;
1547	int n;
1548};
1549
1550struct qp_iter *qp_iter_init(struct hfi1_ibdev *dev)
1551{
1552	struct qp_iter *iter;
1553
1554	iter = kzalloc(sizeof(*iter), GFP_KERNEL);
1555	if (!iter)
1556		return NULL;
1557
1558	iter->dev = dev;
1559	iter->specials = dev->ibdev.phys_port_cnt * 2;
1560	if (qp_iter_next(iter)) {
1561		kfree(iter);
1562		return NULL;
1563	}
1564
1565	return iter;
1566}
1567
1568int qp_iter_next(struct qp_iter *iter)
1569{
1570	struct hfi1_ibdev *dev = iter->dev;
1571	int n = iter->n;
1572	int ret = 1;
1573	struct hfi1_qp *pqp = iter->qp;
1574	struct hfi1_qp *qp;
1575
1576	/*
1577	 * The approach is to consider the special qps
1578	 * as an additional table entries before the
1579	 * real hash table.  Since the qp code sets
1580	 * the qp->next hash link to NULL, this works just fine.
1581	 *
1582	 * iter->specials is 2 * # ports
1583	 *
1584	 * n = 0..iter->specials is the special qp indices
1585	 *
1586	 * n = iter->specials..dev->qp_dev->qp_table_size+iter->specials are
1587	 * the potential hash bucket entries
1588	 *
1589	 */
1590	for (; n <  dev->qp_dev->qp_table_size + iter->specials; n++) {
1591		if (pqp) {
1592			qp = rcu_dereference(pqp->next);
1593		} else {
1594			if (n < iter->specials) {
1595				struct hfi1_pportdata *ppd;
1596				struct hfi1_ibport *ibp;
1597				int pidx;
1598
1599				pidx = n % dev->ibdev.phys_port_cnt;
1600				ppd = &dd_from_dev(dev)->pport[pidx];
1601				ibp = &ppd->ibport_data;
1602
1603				if (!(n & 1))
1604					qp = rcu_dereference(ibp->qp[0]);
1605				else
1606					qp = rcu_dereference(ibp->qp[1]);
1607			} else {
1608				qp = rcu_dereference(
1609					dev->qp_dev->qp_table[
1610						(n - iter->specials)]);
1611			}
1612		}
1613		pqp = qp;
1614		if (qp) {
1615			iter->qp = qp;
1616			iter->n = n;
1617			return 0;
1618		}
1619	}
1620	return ret;
1621}
1622
1623static const char * const qp_type_str[] = {
1624	"SMI", "GSI", "RC", "UC", "UD",
1625};
1626
1627static int qp_idle(struct hfi1_qp *qp)
1628{
1629	return
1630		qp->s_last == qp->s_acked &&
1631		qp->s_acked == qp->s_cur &&
1632		qp->s_cur == qp->s_tail &&
1633		qp->s_tail == qp->s_head;
1634}
1635
1636void qp_iter_print(struct seq_file *s, struct qp_iter *iter)
1637{
1638	struct hfi1_swqe *wqe;
1639	struct hfi1_qp *qp = iter->qp;
1640	struct sdma_engine *sde;
1641
1642	sde = qp_to_sdma_engine(qp, qp->s_sc);
1643	wqe = get_swqe_ptr(qp, qp->s_last);
1644	seq_printf(s,
1645		   "N %d %s QP%u R %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 SL %u MTU %d %u %u %u SDE %p,%u\n",
1646		   iter->n,
1647		   qp_idle(qp) ? "I" : "B",
1648		   qp->ibqp.qp_num,
1649		   atomic_read(&qp->refcount),
1650		   qp_type_str[qp->ibqp.qp_type],
1651		   qp->state,
1652		   wqe ? wqe->wr.opcode : 0,
1653		   qp->s_hdrwords,
1654		   qp->s_flags,
1655		   atomic_read(&qp->s_iowait.sdma_busy),
1656		   !list_empty(&qp->s_iowait.list),
1657		   qp->timeout,
1658		   wqe ? wqe->ssn : 0,
1659		   qp->s_lsn,
1660		   qp->s_last_psn,
1661		   qp->s_psn, qp->s_next_psn,
1662		   qp->s_sending_psn, qp->s_sending_hpsn,
1663		   qp->s_last, qp->s_acked, qp->s_cur,
1664		   qp->s_tail, qp->s_head, qp->s_size,
1665		   qp->remote_qpn,
1666		   qp->remote_ah_attr.dlid,
1667		   qp->remote_ah_attr.sl,
1668		   qp->pmtu,
1669		   qp->s_retry_cnt,
1670		   qp->timeout,
1671		   qp->s_rnr_retry_cnt,
1672		   sde,
1673		   sde ? sde->this_idx : 0);
1674}
1675
1676void qp_comm_est(struct hfi1_qp *qp)
1677{
1678	qp->r_flags |= HFI1_R_COMM_EST;
1679	if (qp->ibqp.event_handler) {
1680		struct ib_event ev;
1681
1682		ev.device = qp->ibqp.device;
1683		ev.element.qp = &qp->ibqp;
1684		ev.event = IB_EVENT_COMM_EST;
1685		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1686	}
1687}
1688