1/*
2 * Copyright (c) 2006, 2007, 2008 QLogic Corporation. All rights reserved.
3 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
4 *
5 * This software is available to you under a choice of one of two
6 * licenses.  You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
10 *
11 *     Redistribution and use in source and binary forms, with or
12 *     without modification, are permitted provided that the following
13 *     conditions are met:
14 *
15 *      - Redistributions of source code must retain the above
16 *        copyright notice, this list of conditions and the following
17 *        disclaimer.
18 *
19 *      - Redistributions in binary form must reproduce the above
20 *        copyright notice, this list of conditions and the following
21 *        disclaimer in the documentation and/or other materials
22 *        provided with the distribution.
23 *
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31 * SOFTWARE.
32 */
33
34#include <linux/err.h>
35#include <linux/sched.h>
36#include <linux/slab.h>
37#include <linux/vmalloc.h>
38
39#include "ipath_verbs.h"
40#include "ipath_kernel.h"
41
42#define BITS_PER_PAGE		(PAGE_SIZE*BITS_PER_BYTE)
43#define BITS_PER_PAGE_MASK	(BITS_PER_PAGE-1)
44#define mk_qpn(qpt, map, off)	(((map) - (qpt)->map) * BITS_PER_PAGE + \
45				 (off))
46#define find_next_offset(map, off) find_next_zero_bit((map)->page, \
47						      BITS_PER_PAGE, off)
48
49/*
50 * Convert the AETH credit code into the number of credits.
51 */
52static u32 credit_table[31] = {
53	0,			/* 0 */
54	1,			/* 1 */
55	2,			/* 2 */
56	3,			/* 3 */
57	4,			/* 4 */
58	6,			/* 5 */
59	8,			/* 6 */
60	12,			/* 7 */
61	16,			/* 8 */
62	24,			/* 9 */
63	32,			/* A */
64	48,			/* B */
65	64,			/* C */
66	96,			/* D */
67	128,			/* E */
68	192,			/* F */
69	256,			/* 10 */
70	384,			/* 11 */
71	512,			/* 12 */
72	768,			/* 13 */
73	1024,			/* 14 */
74	1536,			/* 15 */
75	2048,			/* 16 */
76	3072,			/* 17 */
77	4096,			/* 18 */
78	6144,			/* 19 */
79	8192,			/* 1A */
80	12288,			/* 1B */
81	16384,			/* 1C */
82	24576,			/* 1D */
83	32768			/* 1E */
84};
85
86
87static void get_map_page(struct ipath_qp_table *qpt, struct qpn_map *map)
88{
89	unsigned long page = get_zeroed_page(GFP_KERNEL);
90	unsigned long flags;
91
92	/*
93	 * Free the page if someone raced with us installing it.
94	 */
95
96	spin_lock_irqsave(&qpt->lock, flags);
97	if (map->page)
98		free_page(page);
99	else
100		map->page = (void *)page;
101	spin_unlock_irqrestore(&qpt->lock, flags);
102}
103
104
105static int alloc_qpn(struct ipath_qp_table *qpt, enum ib_qp_type type)
106{
107	u32 i, offset, max_scan, qpn;
108	struct qpn_map *map;
109	u32 ret = -1;
110
111	if (type == IB_QPT_SMI)
112		ret = 0;
113	else if (type == IB_QPT_GSI)
114		ret = 1;
115
116	if (ret != -1) {
117		map = &qpt->map[0];
118		if (unlikely(!map->page)) {
119			get_map_page(qpt, map);
120			if (unlikely(!map->page)) {
121				ret = -ENOMEM;
122				goto bail;
123			}
124		}
125		if (!test_and_set_bit(ret, map->page))
126			atomic_dec(&map->n_free);
127		else
128			ret = -EBUSY;
129		goto bail;
130	}
131
132	qpn = qpt->last + 1;
133	if (qpn >= QPN_MAX)
134		qpn = 2;
135	offset = qpn & BITS_PER_PAGE_MASK;
136	map = &qpt->map[qpn / BITS_PER_PAGE];
137	max_scan = qpt->nmaps - !offset;
138	for (i = 0;;) {
139		if (unlikely(!map->page)) {
140			get_map_page(qpt, map);
141			if (unlikely(!map->page))
142				break;
143		}
144		if (likely(atomic_read(&map->n_free))) {
145			do {
146				if (!test_and_set_bit(offset, map->page)) {
147					atomic_dec(&map->n_free);
148					qpt->last = qpn;
149					ret = qpn;
150					goto bail;
151				}
152				offset = find_next_offset(map, offset);
153				qpn = mk_qpn(qpt, map, offset);
154				/*
155				 * This test differs from alloc_pidmap().
156				 * If find_next_offset() does find a zero
157				 * bit, we don't need to check for QPN
158				 * wrapping around past our starting QPN.
159				 * We just need to be sure we don't loop
160				 * forever.
161				 */
162			} while (offset < BITS_PER_PAGE && qpn < QPN_MAX);
163		}
164		/*
165		 * In order to keep the number of pages allocated to a
166		 * minimum, we scan the all existing pages before increasing
167		 * the size of the bitmap table.
168		 */
169		if (++i > max_scan) {
170			if (qpt->nmaps == QPNMAP_ENTRIES)
171				break;
172			map = &qpt->map[qpt->nmaps++];
173			offset = 0;
174		} else if (map < &qpt->map[qpt->nmaps]) {
175			++map;
176			offset = 0;
177		} else {
178			map = &qpt->map[0];
179			offset = 2;
180		}
181		qpn = mk_qpn(qpt, map, offset);
182	}
183
184	ret = -ENOMEM;
185
186bail:
187	return ret;
188}
189
190static void free_qpn(struct ipath_qp_table *qpt, u32 qpn)
191{
192	struct qpn_map *map;
193
194	map = qpt->map + qpn / BITS_PER_PAGE;
195	if (map->page)
196		clear_bit(qpn & BITS_PER_PAGE_MASK, map->page);
197	atomic_inc(&map->n_free);
198}
199
200/**
201 * ipath_alloc_qpn - allocate a QP number
202 * @qpt: the QP table
203 * @qp: the QP
204 * @type: the QP type (IB_QPT_SMI and IB_QPT_GSI are special)
205 *
206 * Allocate the next available QPN and put the QP into the hash table.
207 * The hash table holds a reference to the QP.
208 */
209static int ipath_alloc_qpn(struct ipath_qp_table *qpt, struct ipath_qp *qp,
210			   enum ib_qp_type type)
211{
212	unsigned long flags;
213	int ret;
214
215	ret = alloc_qpn(qpt, type);
216	if (ret < 0)
217		goto bail;
218	qp->ibqp.qp_num = ret;
219
220	/* Add the QP to the hash table. */
221	spin_lock_irqsave(&qpt->lock, flags);
222
223	ret %= qpt->max;
224	qp->next = qpt->table[ret];
225	qpt->table[ret] = qp;
226	atomic_inc(&qp->refcount);
227
228	spin_unlock_irqrestore(&qpt->lock, flags);
229	ret = 0;
230
231bail:
232	return ret;
233}
234
235/**
236 * ipath_free_qp - remove a QP from the QP table
237 * @qpt: the QP table
238 * @qp: the QP to remove
239 *
240 * Remove the QP from the table so it can't be found asynchronously by
241 * the receive interrupt routine.
242 */
243static void ipath_free_qp(struct ipath_qp_table *qpt, struct ipath_qp *qp)
244{
245	struct ipath_qp *q, **qpp;
246	unsigned long flags;
247
248	spin_lock_irqsave(&qpt->lock, flags);
249
250	/* Remove QP from the hash table. */
251	qpp = &qpt->table[qp->ibqp.qp_num % qpt->max];
252	for (; (q = *qpp) != NULL; qpp = &q->next) {
253		if (q == qp) {
254			*qpp = qp->next;
255			qp->next = NULL;
256			atomic_dec(&qp->refcount);
257			break;
258		}
259	}
260
261	spin_unlock_irqrestore(&qpt->lock, flags);
262}
263
264/**
265 * ipath_free_all_qps - check for QPs still in use
266 * @qpt: the QP table to empty
267 *
268 * There should not be any QPs still in use.
269 * Free memory for table.
270 */
271unsigned ipath_free_all_qps(struct ipath_qp_table *qpt)
272{
273	unsigned long flags;
274	struct ipath_qp *qp;
275	u32 n, qp_inuse = 0;
276
277	spin_lock_irqsave(&qpt->lock, flags);
278	for (n = 0; n < qpt->max; n++) {
279		qp = qpt->table[n];
280		qpt->table[n] = NULL;
281
282		for (; qp; qp = qp->next)
283			qp_inuse++;
284	}
285	spin_unlock_irqrestore(&qpt->lock, flags);
286
287	for (n = 0; n < ARRAY_SIZE(qpt->map); n++)
288		if (qpt->map[n].page)
289			free_page((unsigned long) qpt->map[n].page);
290	return qp_inuse;
291}
292
293/**
294 * ipath_lookup_qpn - return the QP with the given QPN
295 * @qpt: the QP table
296 * @qpn: the QP number to look up
297 *
298 * The caller is responsible for decrementing the QP reference count
299 * when done.
300 */
301struct ipath_qp *ipath_lookup_qpn(struct ipath_qp_table *qpt, u32 qpn)
302{
303	unsigned long flags;
304	struct ipath_qp *qp;
305
306	spin_lock_irqsave(&qpt->lock, flags);
307
308	for (qp = qpt->table[qpn % qpt->max]; qp; qp = qp->next) {
309		if (qp->ibqp.qp_num == qpn) {
310			atomic_inc(&qp->refcount);
311			break;
312		}
313	}
314
315	spin_unlock_irqrestore(&qpt->lock, flags);
316	return qp;
317}
318
319/**
320 * ipath_reset_qp - initialize the QP state to the reset state
321 * @qp: the QP to reset
322 * @type: the QP type
323 */
324static void ipath_reset_qp(struct ipath_qp *qp, enum ib_qp_type type)
325{
326	qp->remote_qpn = 0;
327	qp->qkey = 0;
328	qp->qp_access_flags = 0;
329	atomic_set(&qp->s_dma_busy, 0);
330	qp->s_flags &= IPATH_S_SIGNAL_REQ_WR;
331	qp->s_hdrwords = 0;
332	qp->s_wqe = NULL;
333	qp->s_pkt_delay = 0;
334	qp->s_draining = 0;
335	qp->s_psn = 0;
336	qp->r_psn = 0;
337	qp->r_msn = 0;
338	if (type == IB_QPT_RC) {
339		qp->s_state = IB_OPCODE_RC_SEND_LAST;
340		qp->r_state = IB_OPCODE_RC_SEND_LAST;
341	} else {
342		qp->s_state = IB_OPCODE_UC_SEND_LAST;
343		qp->r_state = IB_OPCODE_UC_SEND_LAST;
344	}
345	qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
346	qp->r_nak_state = 0;
347	qp->r_aflags = 0;
348	qp->r_flags = 0;
349	qp->s_rnr_timeout = 0;
350	qp->s_head = 0;
351	qp->s_tail = 0;
352	qp->s_cur = 0;
353	qp->s_last = 0;
354	qp->s_ssn = 1;
355	qp->s_lsn = 0;
356	memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue));
357	qp->r_head_ack_queue = 0;
358	qp->s_tail_ack_queue = 0;
359	qp->s_num_rd_atomic = 0;
360	if (qp->r_rq.wq) {
361		qp->r_rq.wq->head = 0;
362		qp->r_rq.wq->tail = 0;
363	}
364}
365
366/**
367 * ipath_error_qp - put a QP into the error state
368 * @qp: the QP to put into the error state
369 * @err: the receive completion error to signal if a RWQE is active
370 *
371 * Flushes both send and receive work queues.
372 * Returns true if last WQE event should be generated.
373 * The QP s_lock should be held and interrupts disabled.
374 * If we are already in error state, just return.
375 */
376
377int ipath_error_qp(struct ipath_qp *qp, enum ib_wc_status err)
378{
379	struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
380	struct ib_wc wc;
381	int ret = 0;
382
383	if (qp->state == IB_QPS_ERR)
384		goto bail;
385
386	qp->state = IB_QPS_ERR;
387
388	spin_lock(&dev->pending_lock);
389	if (!list_empty(&qp->timerwait))
390		list_del_init(&qp->timerwait);
391	if (!list_empty(&qp->piowait))
392		list_del_init(&qp->piowait);
393	spin_unlock(&dev->pending_lock);
394
395	/* Schedule the sending tasklet to drain the send work queue. */
396	if (qp->s_last != qp->s_head)
397		ipath_schedule_send(qp);
398
399	memset(&wc, 0, sizeof(wc));
400	wc.qp = &qp->ibqp;
401	wc.opcode = IB_WC_RECV;
402
403	if (test_and_clear_bit(IPATH_R_WRID_VALID, &qp->r_aflags)) {
404		wc.wr_id = qp->r_wr_id;
405		wc.status = err;
406		ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
407	}
408	wc.status = IB_WC_WR_FLUSH_ERR;
409
410	if (qp->r_rq.wq) {
411		struct ipath_rwq *wq;
412		u32 head;
413		u32 tail;
414
415		spin_lock(&qp->r_rq.lock);
416
417		/* sanity check pointers before trusting them */
418		wq = qp->r_rq.wq;
419		head = wq->head;
420		if (head >= qp->r_rq.size)
421			head = 0;
422		tail = wq->tail;
423		if (tail >= qp->r_rq.size)
424			tail = 0;
425		while (tail != head) {
426			wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
427			if (++tail >= qp->r_rq.size)
428				tail = 0;
429			ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
430		}
431		wq->tail = tail;
432
433		spin_unlock(&qp->r_rq.lock);
434	} else if (qp->ibqp.event_handler)
435		ret = 1;
436
437bail:
438	return ret;
439}
440
441/**
442 * ipath_modify_qp - modify the attributes of a queue pair
443 * @ibqp: the queue pair who's attributes we're modifying
444 * @attr: the new attributes
445 * @attr_mask: the mask of attributes to modify
446 * @udata: user data for ipathverbs.so
447 *
448 * Returns 0 on success, otherwise returns an errno.
449 */
450int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
451		    int attr_mask, struct ib_udata *udata)
452{
453	struct ipath_ibdev *dev = to_idev(ibqp->device);
454	struct ipath_qp *qp = to_iqp(ibqp);
455	enum ib_qp_state cur_state, new_state;
456	int lastwqe = 0;
457	int ret;
458
459	spin_lock_irq(&qp->s_lock);
460
461	cur_state = attr_mask & IB_QP_CUR_STATE ?
462		attr->cur_qp_state : qp->state;
463	new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
464
465	if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
466				attr_mask))
467		goto inval;
468
469	if (attr_mask & IB_QP_AV) {
470		if (attr->ah_attr.dlid == 0 ||
471		    attr->ah_attr.dlid >= IPATH_MULTICAST_LID_BASE)
472			goto inval;
473
474		if ((attr->ah_attr.ah_flags & IB_AH_GRH) &&
475		    (attr->ah_attr.grh.sgid_index > 1))
476			goto inval;
477	}
478
479	if (attr_mask & IB_QP_PKEY_INDEX)
480		if (attr->pkey_index >= ipath_get_npkeys(dev->dd))
481			goto inval;
482
483	if (attr_mask & IB_QP_MIN_RNR_TIMER)
484		if (attr->min_rnr_timer > 31)
485			goto inval;
486
487	if (attr_mask & IB_QP_PORT)
488		if (attr->port_num == 0 ||
489		    attr->port_num > ibqp->device->phys_port_cnt)
490			goto inval;
491
492	/*
493	 * don't allow invalid Path MTU values or greater than 2048
494	 * unless we are configured for a 4KB MTU
495	 */
496	if ((attr_mask & IB_QP_PATH_MTU) &&
497		(ib_mtu_enum_to_int(attr->path_mtu) == -1 ||
498		(attr->path_mtu > IB_MTU_2048 && !ipath_mtu4096)))
499		goto inval;
500
501	if (attr_mask & IB_QP_PATH_MIG_STATE)
502		if (attr->path_mig_state != IB_MIG_MIGRATED &&
503		    attr->path_mig_state != IB_MIG_REARM)
504			goto inval;
505
506	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
507		if (attr->max_dest_rd_atomic > IPATH_MAX_RDMA_ATOMIC)
508			goto inval;
509
510	switch (new_state) {
511	case IB_QPS_RESET:
512		if (qp->state != IB_QPS_RESET) {
513			qp->state = IB_QPS_RESET;
514			spin_lock(&dev->pending_lock);
515			if (!list_empty(&qp->timerwait))
516				list_del_init(&qp->timerwait);
517			if (!list_empty(&qp->piowait))
518				list_del_init(&qp->piowait);
519			spin_unlock(&dev->pending_lock);
520			qp->s_flags &= ~IPATH_S_ANY_WAIT;
521			spin_unlock_irq(&qp->s_lock);
522			/* Stop the sending tasklet */
523			tasklet_kill(&qp->s_task);
524			wait_event(qp->wait_dma, !atomic_read(&qp->s_dma_busy));
525			spin_lock_irq(&qp->s_lock);
526		}
527		ipath_reset_qp(qp, ibqp->qp_type);
528		break;
529
530	case IB_QPS_SQD:
531		qp->s_draining = qp->s_last != qp->s_cur;
532		qp->state = new_state;
533		break;
534
535	case IB_QPS_SQE:
536		if (qp->ibqp.qp_type == IB_QPT_RC)
537			goto inval;
538		qp->state = new_state;
539		break;
540
541	case IB_QPS_ERR:
542		lastwqe = ipath_error_qp(qp, IB_WC_WR_FLUSH_ERR);
543		break;
544
545	default:
546		qp->state = new_state;
547		break;
548	}
549
550	if (attr_mask & IB_QP_PKEY_INDEX)
551		qp->s_pkey_index = attr->pkey_index;
552
553	if (attr_mask & IB_QP_DEST_QPN)
554		qp->remote_qpn = attr->dest_qp_num;
555
556	if (attr_mask & IB_QP_SQ_PSN) {
557		qp->s_psn = qp->s_next_psn = attr->sq_psn;
558		qp->s_last_psn = qp->s_next_psn - 1;
559	}
560
561	if (attr_mask & IB_QP_RQ_PSN)
562		qp->r_psn = attr->rq_psn;
563
564	if (attr_mask & IB_QP_ACCESS_FLAGS)
565		qp->qp_access_flags = attr->qp_access_flags;
566
567	if (attr_mask & IB_QP_AV) {
568		qp->remote_ah_attr = attr->ah_attr;
569		qp->s_dmult = ipath_ib_rate_to_mult(attr->ah_attr.static_rate);
570	}
571
572	if (attr_mask & IB_QP_PATH_MTU)
573		qp->path_mtu = attr->path_mtu;
574
575	if (attr_mask & IB_QP_RETRY_CNT)
576		qp->s_retry = qp->s_retry_cnt = attr->retry_cnt;
577
578	if (attr_mask & IB_QP_RNR_RETRY) {
579		qp->s_rnr_retry = attr->rnr_retry;
580		if (qp->s_rnr_retry > 7)
581			qp->s_rnr_retry = 7;
582		qp->s_rnr_retry_cnt = qp->s_rnr_retry;
583	}
584
585	if (attr_mask & IB_QP_MIN_RNR_TIMER)
586		qp->r_min_rnr_timer = attr->min_rnr_timer;
587
588	if (attr_mask & IB_QP_TIMEOUT)
589		qp->timeout = attr->timeout;
590
591	if (attr_mask & IB_QP_QKEY)
592		qp->qkey = attr->qkey;
593
594	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
595		qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
596
597	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
598		qp->s_max_rd_atomic = attr->max_rd_atomic;
599
600	spin_unlock_irq(&qp->s_lock);
601
602	if (lastwqe) {
603		struct ib_event ev;
604
605		ev.device = qp->ibqp.device;
606		ev.element.qp = &qp->ibqp;
607		ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
608		qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
609	}
610	ret = 0;
611	goto bail;
612
613inval:
614	spin_unlock_irq(&qp->s_lock);
615	ret = -EINVAL;
616
617bail:
618	return ret;
619}
620
621int ipath_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
622		   int attr_mask, struct ib_qp_init_attr *init_attr)
623{
624	struct ipath_qp *qp = to_iqp(ibqp);
625
626	attr->qp_state = qp->state;
627	attr->cur_qp_state = attr->qp_state;
628	attr->path_mtu = qp->path_mtu;
629	attr->path_mig_state = 0;
630	attr->qkey = qp->qkey;
631	attr->rq_psn = qp->r_psn;
632	attr->sq_psn = qp->s_next_psn;
633	attr->dest_qp_num = qp->remote_qpn;
634	attr->qp_access_flags = qp->qp_access_flags;
635	attr->cap.max_send_wr = qp->s_size - 1;
636	attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
637	attr->cap.max_send_sge = qp->s_max_sge;
638	attr->cap.max_recv_sge = qp->r_rq.max_sge;
639	attr->cap.max_inline_data = 0;
640	attr->ah_attr = qp->remote_ah_attr;
641	memset(&attr->alt_ah_attr, 0, sizeof(attr->alt_ah_attr));
642	attr->pkey_index = qp->s_pkey_index;
643	attr->alt_pkey_index = 0;
644	attr->en_sqd_async_notify = 0;
645	attr->sq_draining = qp->s_draining;
646	attr->max_rd_atomic = qp->s_max_rd_atomic;
647	attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
648	attr->min_rnr_timer = qp->r_min_rnr_timer;
649	attr->port_num = 1;
650	attr->timeout = qp->timeout;
651	attr->retry_cnt = qp->s_retry_cnt;
652	attr->rnr_retry = qp->s_rnr_retry_cnt;
653	attr->alt_port_num = 0;
654	attr->alt_timeout = 0;
655
656	init_attr->event_handler = qp->ibqp.event_handler;
657	init_attr->qp_context = qp->ibqp.qp_context;
658	init_attr->send_cq = qp->ibqp.send_cq;
659	init_attr->recv_cq = qp->ibqp.recv_cq;
660	init_attr->srq = qp->ibqp.srq;
661	init_attr->cap = attr->cap;
662	if (qp->s_flags & IPATH_S_SIGNAL_REQ_WR)
663		init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
664	else
665		init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
666	init_attr->qp_type = qp->ibqp.qp_type;
667	init_attr->port_num = 1;
668	return 0;
669}
670
671/**
672 * ipath_compute_aeth - compute the AETH (syndrome + MSN)
673 * @qp: the queue pair to compute the AETH for
674 *
675 * Returns the AETH.
676 */
677__be32 ipath_compute_aeth(struct ipath_qp *qp)
678{
679	u32 aeth = qp->r_msn & IPATH_MSN_MASK;
680
681	if (qp->ibqp.srq) {
682		/*
683		 * Shared receive queues don't generate credits.
684		 * Set the credit field to the invalid value.
685		 */
686		aeth |= IPATH_AETH_CREDIT_INVAL << IPATH_AETH_CREDIT_SHIFT;
687	} else {
688		u32 min, max, x;
689		u32 credits;
690		struct ipath_rwq *wq = qp->r_rq.wq;
691		u32 head;
692		u32 tail;
693
694		/* sanity check pointers before trusting them */
695		head = wq->head;
696		if (head >= qp->r_rq.size)
697			head = 0;
698		tail = wq->tail;
699		if (tail >= qp->r_rq.size)
700			tail = 0;
701		/*
702		 * Compute the number of credits available (RWQEs).
703		 * XXX Not holding the r_rq.lock here so there is a small
704		 * chance that the pair of reads are not atomic.
705		 */
706		credits = head - tail;
707		if ((int)credits < 0)
708			credits += qp->r_rq.size;
709		/*
710		 * Binary search the credit table to find the code to
711		 * use.
712		 */
713		min = 0;
714		max = 31;
715		for (;;) {
716			x = (min + max) / 2;
717			if (credit_table[x] == credits)
718				break;
719			if (credit_table[x] > credits)
720				max = x;
721			else if (min == x)
722				break;
723			else
724				min = x;
725		}
726		aeth |= x << IPATH_AETH_CREDIT_SHIFT;
727	}
728	return cpu_to_be32(aeth);
729}
730
731/**
732 * ipath_create_qp - create a queue pair for a device
733 * @ibpd: the protection domain who's device we create the queue pair for
734 * @init_attr: the attributes of the queue pair
735 * @udata: unused by InfiniPath
736 *
737 * Returns the queue pair on success, otherwise returns an errno.
738 *
739 * Called by the ib_create_qp() core verbs function.
740 */
741struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
742			      struct ib_qp_init_attr *init_attr,
743			      struct ib_udata *udata)
744{
745	struct ipath_qp *qp;
746	int err;
747	struct ipath_swqe *swq = NULL;
748	struct ipath_ibdev *dev;
749	size_t sz;
750	size_t sg_list_sz;
751	struct ib_qp *ret;
752
753	if (init_attr->create_flags) {
754		ret = ERR_PTR(-EINVAL);
755		goto bail;
756	}
757
758	if (init_attr->cap.max_send_sge > ib_ipath_max_sges ||
759	    init_attr->cap.max_send_wr > ib_ipath_max_qp_wrs) {
760		ret = ERR_PTR(-EINVAL);
761		goto bail;
762	}
763
764	/* Check receive queue parameters if no SRQ is specified. */
765	if (!init_attr->srq) {
766		if (init_attr->cap.max_recv_sge > ib_ipath_max_sges ||
767		    init_attr->cap.max_recv_wr > ib_ipath_max_qp_wrs) {
768			ret = ERR_PTR(-EINVAL);
769			goto bail;
770		}
771		if (init_attr->cap.max_send_sge +
772		    init_attr->cap.max_send_wr +
773		    init_attr->cap.max_recv_sge +
774		    init_attr->cap.max_recv_wr == 0) {
775			ret = ERR_PTR(-EINVAL);
776			goto bail;
777		}
778	}
779
780	switch (init_attr->qp_type) {
781	case IB_QPT_UC:
782	case IB_QPT_RC:
783	case IB_QPT_UD:
784	case IB_QPT_SMI:
785	case IB_QPT_GSI:
786		sz = sizeof(struct ipath_sge) *
787			init_attr->cap.max_send_sge +
788			sizeof(struct ipath_swqe);
789		swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz);
790		if (swq == NULL) {
791			ret = ERR_PTR(-ENOMEM);
792			goto bail;
793		}
794		sz = sizeof(*qp);
795		sg_list_sz = 0;
796		if (init_attr->srq) {
797			struct ipath_srq *srq = to_isrq(init_attr->srq);
798
799			if (srq->rq.max_sge > 1)
800				sg_list_sz = sizeof(*qp->r_sg_list) *
801					(srq->rq.max_sge - 1);
802		} else if (init_attr->cap.max_recv_sge > 1)
803			sg_list_sz = sizeof(*qp->r_sg_list) *
804				(init_attr->cap.max_recv_sge - 1);
805		qp = kmalloc(sz + sg_list_sz, GFP_KERNEL);
806		if (!qp) {
807			ret = ERR_PTR(-ENOMEM);
808			goto bail_swq;
809		}
810		if (sg_list_sz && (init_attr->qp_type == IB_QPT_UD ||
811		    init_attr->qp_type == IB_QPT_SMI ||
812		    init_attr->qp_type == IB_QPT_GSI)) {
813			qp->r_ud_sg_list = kmalloc(sg_list_sz, GFP_KERNEL);
814			if (!qp->r_ud_sg_list) {
815				ret = ERR_PTR(-ENOMEM);
816				goto bail_qp;
817			}
818		} else
819			qp->r_ud_sg_list = NULL;
820		if (init_attr->srq) {
821			sz = 0;
822			qp->r_rq.size = 0;
823			qp->r_rq.max_sge = 0;
824			qp->r_rq.wq = NULL;
825			init_attr->cap.max_recv_wr = 0;
826			init_attr->cap.max_recv_sge = 0;
827		} else {
828			qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
829			qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
830			sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
831				sizeof(struct ipath_rwqe);
832			qp->r_rq.wq = vmalloc_user(sizeof(struct ipath_rwq) +
833					      qp->r_rq.size * sz);
834			if (!qp->r_rq.wq) {
835				ret = ERR_PTR(-ENOMEM);
836				goto bail_sg_list;
837			}
838		}
839
840		/*
841		 * ib_create_qp() will initialize qp->ibqp
842		 * except for qp->ibqp.qp_num.
843		 */
844		spin_lock_init(&qp->s_lock);
845		spin_lock_init(&qp->r_rq.lock);
846		atomic_set(&qp->refcount, 0);
847		init_waitqueue_head(&qp->wait);
848		init_waitqueue_head(&qp->wait_dma);
849		tasklet_init(&qp->s_task, ipath_do_send, (unsigned long)qp);
850		INIT_LIST_HEAD(&qp->piowait);
851		INIT_LIST_HEAD(&qp->timerwait);
852		qp->state = IB_QPS_RESET;
853		qp->s_wq = swq;
854		qp->s_size = init_attr->cap.max_send_wr + 1;
855		qp->s_max_sge = init_attr->cap.max_send_sge;
856		if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
857			qp->s_flags = IPATH_S_SIGNAL_REQ_WR;
858		else
859			qp->s_flags = 0;
860		dev = to_idev(ibpd->device);
861		err = ipath_alloc_qpn(&dev->qp_table, qp,
862				      init_attr->qp_type);
863		if (err) {
864			ret = ERR_PTR(err);
865			vfree(qp->r_rq.wq);
866			goto bail_sg_list;
867		}
868		qp->ip = NULL;
869		qp->s_tx = NULL;
870		ipath_reset_qp(qp, init_attr->qp_type);
871		break;
872
873	default:
874		/* Don't support raw QPs */
875		ret = ERR_PTR(-ENOSYS);
876		goto bail;
877	}
878
879	init_attr->cap.max_inline_data = 0;
880
881	/*
882	 * Return the address of the RWQ as the offset to mmap.
883	 * See ipath_mmap() for details.
884	 */
885	if (udata && udata->outlen >= sizeof(__u64)) {
886		if (!qp->r_rq.wq) {
887			__u64 offset = 0;
888
889			err = ib_copy_to_udata(udata, &offset,
890					       sizeof(offset));
891			if (err) {
892				ret = ERR_PTR(err);
893				goto bail_ip;
894			}
895		} else {
896			u32 s = sizeof(struct ipath_rwq) +
897				qp->r_rq.size * sz;
898
899			qp->ip =
900			    ipath_create_mmap_info(dev, s,
901						   ibpd->uobject->context,
902						   qp->r_rq.wq);
903			if (!qp->ip) {
904				ret = ERR_PTR(-ENOMEM);
905				goto bail_ip;
906			}
907
908			err = ib_copy_to_udata(udata, &(qp->ip->offset),
909					       sizeof(qp->ip->offset));
910			if (err) {
911				ret = ERR_PTR(err);
912				goto bail_ip;
913			}
914		}
915	}
916
917	spin_lock(&dev->n_qps_lock);
918	if (dev->n_qps_allocated == ib_ipath_max_qps) {
919		spin_unlock(&dev->n_qps_lock);
920		ret = ERR_PTR(-ENOMEM);
921		goto bail_ip;
922	}
923
924	dev->n_qps_allocated++;
925	spin_unlock(&dev->n_qps_lock);
926
927	if (qp->ip) {
928		spin_lock_irq(&dev->pending_lock);
929		list_add(&qp->ip->pending_mmaps, &dev->pending_mmaps);
930		spin_unlock_irq(&dev->pending_lock);
931	}
932
933	ret = &qp->ibqp;
934	goto bail;
935
936bail_ip:
937	if (qp->ip)
938		kref_put(&qp->ip->ref, ipath_release_mmap_info);
939	else
940		vfree(qp->r_rq.wq);
941	ipath_free_qp(&dev->qp_table, qp);
942	free_qpn(&dev->qp_table, qp->ibqp.qp_num);
943bail_sg_list:
944	kfree(qp->r_ud_sg_list);
945bail_qp:
946	kfree(qp);
947bail_swq:
948	vfree(swq);
949bail:
950	return ret;
951}
952
953/**
954 * ipath_destroy_qp - destroy a queue pair
955 * @ibqp: the queue pair to destroy
956 *
957 * Returns 0 on success.
958 *
959 * Note that this can be called while the QP is actively sending or
960 * receiving!
961 */
962int ipath_destroy_qp(struct ib_qp *ibqp)
963{
964	struct ipath_qp *qp = to_iqp(ibqp);
965	struct ipath_ibdev *dev = to_idev(ibqp->device);
966
967	/* Make sure HW and driver activity is stopped. */
968	spin_lock_irq(&qp->s_lock);
969	if (qp->state != IB_QPS_RESET) {
970		qp->state = IB_QPS_RESET;
971		spin_lock(&dev->pending_lock);
972		if (!list_empty(&qp->timerwait))
973			list_del_init(&qp->timerwait);
974		if (!list_empty(&qp->piowait))
975			list_del_init(&qp->piowait);
976		spin_unlock(&dev->pending_lock);
977		qp->s_flags &= ~IPATH_S_ANY_WAIT;
978		spin_unlock_irq(&qp->s_lock);
979		/* Stop the sending tasklet */
980		tasklet_kill(&qp->s_task);
981		wait_event(qp->wait_dma, !atomic_read(&qp->s_dma_busy));
982	} else
983		spin_unlock_irq(&qp->s_lock);
984
985	ipath_free_qp(&dev->qp_table, qp);
986
987	if (qp->s_tx) {
988		atomic_dec(&qp->refcount);
989		if (qp->s_tx->txreq.flags & IPATH_SDMA_TXREQ_F_FREEBUF)
990			kfree(qp->s_tx->txreq.map_addr);
991		spin_lock_irq(&dev->pending_lock);
992		list_add(&qp->s_tx->txreq.list, &dev->txreq_free);
993		spin_unlock_irq(&dev->pending_lock);
994		qp->s_tx = NULL;
995	}
996
997	wait_event(qp->wait, !atomic_read(&qp->refcount));
998
999	/* all user's cleaned up, mark it available */
1000	free_qpn(&dev->qp_table, qp->ibqp.qp_num);
1001	spin_lock(&dev->n_qps_lock);
1002	dev->n_qps_allocated--;
1003	spin_unlock(&dev->n_qps_lock);
1004
1005	if (qp->ip)
1006		kref_put(&qp->ip->ref, ipath_release_mmap_info);
1007	else
1008		vfree(qp->r_rq.wq);
1009	kfree(qp->r_ud_sg_list);
1010	vfree(qp->s_wq);
1011	kfree(qp);
1012	return 0;
1013}
1014
1015/**
1016 * ipath_init_qp_table - initialize the QP table for a device
1017 * @idev: the device who's QP table we're initializing
1018 * @size: the size of the QP table
1019 *
1020 * Returns 0 on success, otherwise returns an errno.
1021 */
1022int ipath_init_qp_table(struct ipath_ibdev *idev, int size)
1023{
1024	int i;
1025	int ret;
1026
1027	idev->qp_table.last = 1;	/* QPN 0 and 1 are special. */
1028	idev->qp_table.max = size;
1029	idev->qp_table.nmaps = 1;
1030	idev->qp_table.table = kzalloc(size * sizeof(*idev->qp_table.table),
1031				       GFP_KERNEL);
1032	if (idev->qp_table.table == NULL) {
1033		ret = -ENOMEM;
1034		goto bail;
1035	}
1036
1037	for (i = 0; i < ARRAY_SIZE(idev->qp_table.map); i++) {
1038		atomic_set(&idev->qp_table.map[i].n_free, BITS_PER_PAGE);
1039		idev->qp_table.map[i].page = NULL;
1040	}
1041
1042	ret = 0;
1043
1044bail:
1045	return ret;
1046}
1047
1048/**
1049 * ipath_get_credit - flush the send work queue of a QP
1050 * @qp: the qp who's send work queue to flush
1051 * @aeth: the Acknowledge Extended Transport Header
1052 *
1053 * The QP s_lock should be held.
1054 */
1055void ipath_get_credit(struct ipath_qp *qp, u32 aeth)
1056{
1057	u32 credit = (aeth >> IPATH_AETH_CREDIT_SHIFT) & IPATH_AETH_CREDIT_MASK;
1058
1059	/*
1060	 * If the credit is invalid, we can send
1061	 * as many packets as we like.  Otherwise, we have to
1062	 * honor the credit field.
1063	 */
1064	if (credit == IPATH_AETH_CREDIT_INVAL)
1065		qp->s_lsn = (u32) -1;
1066	else if (qp->s_lsn != (u32) -1) {
1067		/* Compute new LSN (i.e., MSN + credit) */
1068		credit = (aeth + credit_table[credit]) & IPATH_MSN_MASK;
1069		if (ipath_cmp24(credit, qp->s_lsn) > 0)
1070			qp->s_lsn = credit;
1071	}
1072
1073	/* Restart sending if it was blocked due to lack of credits. */
1074	if ((qp->s_flags & IPATH_S_WAIT_SSN_CREDIT) &&
1075	    qp->s_cur != qp->s_head &&
1076	    (qp->s_lsn == (u32) -1 ||
1077	     ipath_cmp24(get_swqe_ptr(qp, qp->s_cur)->ssn,
1078			 qp->s_lsn + 1) <= 0))
1079		ipath_schedule_send(qp);
1080}
1081