mthca_cq.c revision cd4e8fb49d2326364971a56f5a4b664a976f3712
1/*
2 * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
3 * Copyright (c) 2005 Sun Microsystems, 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 * $Id: mthca_cq.c 1369 2004-12-20 16:17:07Z roland $
34 */
35
36#include <linux/init.h>
37#include <linux/hardirq.h>
38
39#include <ib_pack.h>
40
41#include "mthca_dev.h"
42#include "mthca_cmd.h"
43#include "mthca_memfree.h"
44
45enum {
46	MTHCA_MAX_DIRECT_CQ_SIZE = 4 * PAGE_SIZE
47};
48
49enum {
50	MTHCA_CQ_ENTRY_SIZE = 0x20
51};
52
53/*
54 * Must be packed because start is 64 bits but only aligned to 32 bits.
55 */
56struct mthca_cq_context {
57	u32 flags;
58	u64 start;
59	u32 logsize_usrpage;
60	u32 error_eqn;		/* Tavor only */
61	u32 comp_eqn;
62	u32 pd;
63	u32 lkey;
64	u32 last_notified_index;
65	u32 solicit_producer_index;
66	u32 consumer_index;
67	u32 producer_index;
68	u32 cqn;
69	u32 ci_db;		/* Arbel only */
70	u32 state_db;		/* Arbel only */
71	u32 reserved;
72} __attribute__((packed));
73
74#define MTHCA_CQ_STATUS_OK          ( 0 << 28)
75#define MTHCA_CQ_STATUS_OVERFLOW    ( 9 << 28)
76#define MTHCA_CQ_STATUS_WRITE_FAIL  (10 << 28)
77#define MTHCA_CQ_FLAG_TR            ( 1 << 18)
78#define MTHCA_CQ_FLAG_OI            ( 1 << 17)
79#define MTHCA_CQ_STATE_DISARMED     ( 0 <<  8)
80#define MTHCA_CQ_STATE_ARMED        ( 1 <<  8)
81#define MTHCA_CQ_STATE_ARMED_SOL    ( 4 <<  8)
82#define MTHCA_EQ_STATE_FIRED        (10 <<  8)
83
84enum {
85	MTHCA_ERROR_CQE_OPCODE_MASK = 0xfe
86};
87
88enum {
89	SYNDROME_LOCAL_LENGTH_ERR 	 = 0x01,
90	SYNDROME_LOCAL_QP_OP_ERR  	 = 0x02,
91	SYNDROME_LOCAL_EEC_OP_ERR 	 = 0x03,
92	SYNDROME_LOCAL_PROT_ERR   	 = 0x04,
93	SYNDROME_WR_FLUSH_ERR     	 = 0x05,
94	SYNDROME_MW_BIND_ERR      	 = 0x06,
95	SYNDROME_BAD_RESP_ERR     	 = 0x10,
96	SYNDROME_LOCAL_ACCESS_ERR 	 = 0x11,
97	SYNDROME_REMOTE_INVAL_REQ_ERR 	 = 0x12,
98	SYNDROME_REMOTE_ACCESS_ERR 	 = 0x13,
99	SYNDROME_REMOTE_OP_ERR     	 = 0x14,
100	SYNDROME_RETRY_EXC_ERR 		 = 0x15,
101	SYNDROME_RNR_RETRY_EXC_ERR 	 = 0x16,
102	SYNDROME_LOCAL_RDD_VIOL_ERR 	 = 0x20,
103	SYNDROME_REMOTE_INVAL_RD_REQ_ERR = 0x21,
104	SYNDROME_REMOTE_ABORTED_ERR 	 = 0x22,
105	SYNDROME_INVAL_EECN_ERR 	 = 0x23,
106	SYNDROME_INVAL_EEC_STATE_ERR 	 = 0x24
107};
108
109struct mthca_cqe {
110	u32 my_qpn;
111	u32 my_ee;
112	u32 rqpn;
113	u16 sl_g_mlpath;
114	u16 rlid;
115	u32 imm_etype_pkey_eec;
116	u32 byte_cnt;
117	u32 wqe;
118	u8  opcode;
119	u8  is_send;
120	u8  reserved;
121	u8  owner;
122};
123
124struct mthca_err_cqe {
125	u32 my_qpn;
126	u32 reserved1[3];
127	u8  syndrome;
128	u8  reserved2;
129	u16 db_cnt;
130	u32 reserved3;
131	u32 wqe;
132	u8  opcode;
133	u8  reserved4[2];
134	u8  owner;
135};
136
137#define MTHCA_CQ_ENTRY_OWNER_SW      (0 << 7)
138#define MTHCA_CQ_ENTRY_OWNER_HW      (1 << 7)
139
140#define MTHCA_TAVOR_CQ_DB_INC_CI       (1 << 24)
141#define MTHCA_TAVOR_CQ_DB_REQ_NOT      (2 << 24)
142#define MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL  (3 << 24)
143#define MTHCA_TAVOR_CQ_DB_SET_CI       (4 << 24)
144#define MTHCA_TAVOR_CQ_DB_REQ_NOT_MULT (5 << 24)
145
146#define MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL  (1 << 24)
147#define MTHCA_ARBEL_CQ_DB_REQ_NOT      (2 << 24)
148#define MTHCA_ARBEL_CQ_DB_REQ_NOT_MULT (3 << 24)
149
150static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
151{
152	if (cq->is_direct)
153		return cq->queue.direct.buf + (entry * MTHCA_CQ_ENTRY_SIZE);
154	else
155		return cq->queue.page_list[entry * MTHCA_CQ_ENTRY_SIZE / PAGE_SIZE].buf
156			+ (entry * MTHCA_CQ_ENTRY_SIZE) % PAGE_SIZE;
157}
158
159static inline struct mthca_cqe *cqe_sw(struct mthca_cq *cq, int i)
160{
161	struct mthca_cqe *cqe = get_cqe(cq, i);
162	return MTHCA_CQ_ENTRY_OWNER_HW & cqe->owner ? NULL : cqe;
163}
164
165static inline struct mthca_cqe *next_cqe_sw(struct mthca_cq *cq)
166{
167	return cqe_sw(cq, cq->cons_index & cq->ibcq.cqe);
168}
169
170static inline void set_cqe_hw(struct mthca_cqe *cqe)
171{
172	cqe->owner = MTHCA_CQ_ENTRY_OWNER_HW;
173}
174
175/*
176 * incr is ignored in native Arbel (mem-free) mode, so cq->cons_index
177 * should be correct before calling update_cons_index().
178 */
179static inline void update_cons_index(struct mthca_dev *dev, struct mthca_cq *cq,
180				     int incr)
181{
182	u32 doorbell[2];
183
184	if (mthca_is_memfree(dev)) {
185		*cq->set_ci_db = cpu_to_be32(cq->cons_index);
186		wmb();
187	} else {
188		doorbell[0] = cpu_to_be32(MTHCA_TAVOR_CQ_DB_INC_CI | cq->cqn);
189		doorbell[1] = cpu_to_be32(incr - 1);
190
191		mthca_write64(doorbell,
192			      dev->kar + MTHCA_CQ_DOORBELL,
193			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
194	}
195}
196
197void mthca_cq_event(struct mthca_dev *dev, u32 cqn)
198{
199	struct mthca_cq *cq;
200
201	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
202
203	if (!cq) {
204		mthca_warn(dev, "Completion event for bogus CQ %08x\n", cqn);
205		return;
206	}
207
208	++cq->arm_sn;
209
210	cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
211}
212
213void mthca_cq_clean(struct mthca_dev *dev, u32 cqn, u32 qpn)
214{
215	struct mthca_cq *cq;
216	struct mthca_cqe *cqe;
217	int prod_index;
218	int nfreed = 0;
219
220	spin_lock_irq(&dev->cq_table.lock);
221	cq = mthca_array_get(&dev->cq_table.cq, cqn & (dev->limits.num_cqs - 1));
222	if (cq)
223		atomic_inc(&cq->refcount);
224	spin_unlock_irq(&dev->cq_table.lock);
225
226	if (!cq)
227		return;
228
229	spin_lock_irq(&cq->lock);
230
231	/*
232	 * First we need to find the current producer index, so we
233	 * know where to start cleaning from.  It doesn't matter if HW
234	 * adds new entries after this loop -- the QP we're worried
235	 * about is already in RESET, so the new entries won't come
236	 * from our QP and therefore don't need to be checked.
237	 */
238	for (prod_index = cq->cons_index;
239	     cqe_sw(cq, prod_index & cq->ibcq.cqe);
240	     ++prod_index)
241		if (prod_index == cq->cons_index + cq->ibcq.cqe)
242			break;
243
244	if (0)
245		mthca_dbg(dev, "Cleaning QPN %06x from CQN %06x; ci %d, pi %d\n",
246			  qpn, cqn, cq->cons_index, prod_index);
247
248	/*
249	 * Now sweep backwards through the CQ, removing CQ entries
250	 * that match our QP by copying older entries on top of them.
251	 */
252	while (prod_index > cq->cons_index) {
253		cqe = get_cqe(cq, (prod_index - 1) & cq->ibcq.cqe);
254		if (cqe->my_qpn == cpu_to_be32(qpn))
255			++nfreed;
256		else if (nfreed)
257			memcpy(get_cqe(cq, (prod_index - 1 + nfreed) &
258				       cq->ibcq.cqe),
259			       cqe,
260			       MTHCA_CQ_ENTRY_SIZE);
261		--prod_index;
262	}
263
264	if (nfreed) {
265		wmb();
266		cq->cons_index += nfreed;
267		update_cons_index(dev, cq, nfreed);
268	}
269
270	spin_unlock_irq(&cq->lock);
271	if (atomic_dec_and_test(&cq->refcount))
272		wake_up(&cq->wait);
273}
274
275static int handle_error_cqe(struct mthca_dev *dev, struct mthca_cq *cq,
276			    struct mthca_qp *qp, int wqe_index, int is_send,
277			    struct mthca_err_cqe *cqe,
278			    struct ib_wc *entry, int *free_cqe)
279{
280	int err;
281	int dbd;
282	u32 new_wqe;
283
284	if (1 && cqe->syndrome != SYNDROME_WR_FLUSH_ERR) {
285		int j;
286
287		mthca_dbg(dev, "%x/%d: error CQE -> QPN %06x, WQE @ %08x\n",
288			  cq->cqn, cq->cons_index, be32_to_cpu(cqe->my_qpn),
289			  be32_to_cpu(cqe->wqe));
290
291		for (j = 0; j < 8; ++j)
292			printk(KERN_DEBUG "  [%2x] %08x\n",
293			       j * 4, be32_to_cpu(((u32 *) cqe)[j]));
294	}
295
296	/*
297	 * For completions in error, only work request ID, status (and
298	 * freed resource count for RD) have to be set.
299	 */
300	switch (cqe->syndrome) {
301	case SYNDROME_LOCAL_LENGTH_ERR:
302		entry->status = IB_WC_LOC_LEN_ERR;
303		break;
304	case SYNDROME_LOCAL_QP_OP_ERR:
305		entry->status = IB_WC_LOC_QP_OP_ERR;
306		break;
307	case SYNDROME_LOCAL_EEC_OP_ERR:
308		entry->status = IB_WC_LOC_EEC_OP_ERR;
309		break;
310	case SYNDROME_LOCAL_PROT_ERR:
311		entry->status = IB_WC_LOC_PROT_ERR;
312		break;
313	case SYNDROME_WR_FLUSH_ERR:
314		entry->status = IB_WC_WR_FLUSH_ERR;
315		break;
316	case SYNDROME_MW_BIND_ERR:
317		entry->status = IB_WC_MW_BIND_ERR;
318		break;
319	case SYNDROME_BAD_RESP_ERR:
320		entry->status = IB_WC_BAD_RESP_ERR;
321		break;
322	case SYNDROME_LOCAL_ACCESS_ERR:
323		entry->status = IB_WC_LOC_ACCESS_ERR;
324		break;
325	case SYNDROME_REMOTE_INVAL_REQ_ERR:
326		entry->status = IB_WC_REM_INV_REQ_ERR;
327		break;
328	case SYNDROME_REMOTE_ACCESS_ERR:
329		entry->status = IB_WC_REM_ACCESS_ERR;
330		break;
331	case SYNDROME_REMOTE_OP_ERR:
332		entry->status = IB_WC_REM_OP_ERR;
333		break;
334	case SYNDROME_RETRY_EXC_ERR:
335		entry->status = IB_WC_RETRY_EXC_ERR;
336		break;
337	case SYNDROME_RNR_RETRY_EXC_ERR:
338		entry->status = IB_WC_RNR_RETRY_EXC_ERR;
339		break;
340	case SYNDROME_LOCAL_RDD_VIOL_ERR:
341		entry->status = IB_WC_LOC_RDD_VIOL_ERR;
342		break;
343	case SYNDROME_REMOTE_INVAL_RD_REQ_ERR:
344		entry->status = IB_WC_REM_INV_RD_REQ_ERR;
345		break;
346	case SYNDROME_REMOTE_ABORTED_ERR:
347		entry->status = IB_WC_REM_ABORT_ERR;
348		break;
349	case SYNDROME_INVAL_EECN_ERR:
350		entry->status = IB_WC_INV_EECN_ERR;
351		break;
352	case SYNDROME_INVAL_EEC_STATE_ERR:
353		entry->status = IB_WC_INV_EEC_STATE_ERR;
354		break;
355	default:
356		entry->status = IB_WC_GENERAL_ERR;
357		break;
358	}
359
360	err = mthca_free_err_wqe(dev, qp, is_send, wqe_index, &dbd, &new_wqe);
361	if (err)
362		return err;
363
364	/*
365	 * If we're at the end of the WQE chain, or we've used up our
366	 * doorbell count, free the CQE.  Otherwise just update it for
367	 * the next poll operation.
368	 */
369	if (!(new_wqe & cpu_to_be32(0x3f)) || (!cqe->db_cnt && dbd))
370		return 0;
371
372	cqe->db_cnt   = cpu_to_be16(be16_to_cpu(cqe->db_cnt) - dbd);
373	cqe->wqe      = new_wqe;
374	cqe->syndrome = SYNDROME_WR_FLUSH_ERR;
375
376	*free_cqe = 0;
377
378	return 0;
379}
380
381static void dump_cqe(struct mthca_cqe *cqe)
382{
383	int j;
384
385	for (j = 0; j < 8; ++j)
386		printk(KERN_DEBUG "  [%2x] %08x\n",
387		       j * 4, be32_to_cpu(((u32 *) cqe)[j]));
388}
389
390static inline int mthca_poll_one(struct mthca_dev *dev,
391				 struct mthca_cq *cq,
392				 struct mthca_qp **cur_qp,
393				 int *freed,
394				 struct ib_wc *entry)
395{
396	struct mthca_wq *wq;
397	struct mthca_cqe *cqe;
398	int wqe_index;
399	int is_error;
400	int is_send;
401	int free_cqe = 1;
402	int err = 0;
403
404	cqe = next_cqe_sw(cq);
405	if (!cqe)
406		return -EAGAIN;
407
408	/*
409	 * Make sure we read CQ entry contents after we've checked the
410	 * ownership bit.
411	 */
412	rmb();
413
414	if (0) {
415		mthca_dbg(dev, "%x/%d: CQE -> QPN %06x, WQE @ %08x\n",
416			  cq->cqn, cq->cons_index, be32_to_cpu(cqe->my_qpn),
417			  be32_to_cpu(cqe->wqe));
418
419		dump_cqe(cqe);
420	}
421
422	is_error = (cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
423		MTHCA_ERROR_CQE_OPCODE_MASK;
424	is_send  = is_error ? cqe->opcode & 0x01 : cqe->is_send & 0x80;
425
426	if (!*cur_qp || be32_to_cpu(cqe->my_qpn) != (*cur_qp)->qpn) {
427		/*
428		 * We do not have to take the QP table lock here,
429		 * because CQs will be locked while QPs are removed
430		 * from the table.
431		 */
432		*cur_qp = mthca_array_get(&dev->qp_table.qp,
433					  be32_to_cpu(cqe->my_qpn) &
434					  (dev->limits.num_qps - 1));
435		if (!*cur_qp) {
436			mthca_warn(dev, "CQ entry for unknown QP %06x\n",
437				   be32_to_cpu(cqe->my_qpn) & 0xffffff);
438			err = -EINVAL;
439			goto out;
440		}
441	}
442
443	entry->qp_num = (*cur_qp)->qpn;
444
445	if (is_send) {
446		wq = &(*cur_qp)->sq;
447		wqe_index = ((be32_to_cpu(cqe->wqe) - (*cur_qp)->send_wqe_offset)
448			     >> wq->wqe_shift);
449		entry->wr_id = (*cur_qp)->wrid[wqe_index +
450					       (*cur_qp)->rq.max];
451	} else {
452		wq = &(*cur_qp)->rq;
453		wqe_index = be32_to_cpu(cqe->wqe) >> wq->wqe_shift;
454		entry->wr_id = (*cur_qp)->wrid[wqe_index];
455	}
456
457	if (wq->last_comp < wqe_index)
458		wq->tail += wqe_index - wq->last_comp;
459	else
460		wq->tail += wqe_index + wq->max - wq->last_comp;
461
462	wq->last_comp = wqe_index;
463
464	if (0)
465		mthca_dbg(dev, "%s completion for QP %06x, index %d (nr %d)\n",
466			  is_send ? "Send" : "Receive",
467			  (*cur_qp)->qpn, wqe_index, wq->max);
468
469	if (is_error) {
470		err = handle_error_cqe(dev, cq, *cur_qp, wqe_index, is_send,
471				       (struct mthca_err_cqe *) cqe,
472				       entry, &free_cqe);
473		goto out;
474	}
475
476	if (is_send) {
477		entry->wc_flags = 0;
478		switch (cqe->opcode) {
479		case MTHCA_OPCODE_RDMA_WRITE:
480			entry->opcode    = IB_WC_RDMA_WRITE;
481			break;
482		case MTHCA_OPCODE_RDMA_WRITE_IMM:
483			entry->opcode    = IB_WC_RDMA_WRITE;
484			entry->wc_flags |= IB_WC_WITH_IMM;
485			break;
486		case MTHCA_OPCODE_SEND:
487			entry->opcode    = IB_WC_SEND;
488			break;
489		case MTHCA_OPCODE_SEND_IMM:
490			entry->opcode    = IB_WC_SEND;
491			entry->wc_flags |= IB_WC_WITH_IMM;
492			break;
493		case MTHCA_OPCODE_RDMA_READ:
494			entry->opcode    = IB_WC_RDMA_READ;
495			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
496			break;
497		case MTHCA_OPCODE_ATOMIC_CS:
498			entry->opcode    = IB_WC_COMP_SWAP;
499			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
500			break;
501		case MTHCA_OPCODE_ATOMIC_FA:
502			entry->opcode    = IB_WC_FETCH_ADD;
503			entry->byte_len  = be32_to_cpu(cqe->byte_cnt);
504			break;
505		case MTHCA_OPCODE_BIND_MW:
506			entry->opcode    = IB_WC_BIND_MW;
507			break;
508		default:
509			entry->opcode    = MTHCA_OPCODE_INVALID;
510			break;
511		}
512	} else {
513		entry->byte_len = be32_to_cpu(cqe->byte_cnt);
514		switch (cqe->opcode & 0x1f) {
515		case IB_OPCODE_SEND_LAST_WITH_IMMEDIATE:
516		case IB_OPCODE_SEND_ONLY_WITH_IMMEDIATE:
517			entry->wc_flags = IB_WC_WITH_IMM;
518			entry->imm_data = cqe->imm_etype_pkey_eec;
519			entry->opcode = IB_WC_RECV;
520			break;
521		case IB_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE:
522		case IB_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE:
523			entry->wc_flags = IB_WC_WITH_IMM;
524			entry->imm_data = cqe->imm_etype_pkey_eec;
525			entry->opcode = IB_WC_RECV_RDMA_WITH_IMM;
526			break;
527		default:
528			entry->wc_flags = 0;
529			entry->opcode = IB_WC_RECV;
530			break;
531		}
532		entry->slid 	   = be16_to_cpu(cqe->rlid);
533		entry->sl   	   = be16_to_cpu(cqe->sl_g_mlpath) >> 12;
534		entry->src_qp 	   = be32_to_cpu(cqe->rqpn) & 0xffffff;
535		entry->dlid_path_bits = be16_to_cpu(cqe->sl_g_mlpath) & 0x7f;
536		entry->pkey_index  = be32_to_cpu(cqe->imm_etype_pkey_eec) >> 16;
537		entry->wc_flags   |= be16_to_cpu(cqe->sl_g_mlpath) & 0x80 ?
538					IB_WC_GRH : 0;
539	}
540
541	entry->status = IB_WC_SUCCESS;
542
543 out:
544	if (likely(free_cqe)) {
545		set_cqe_hw(cqe);
546		++(*freed);
547		++cq->cons_index;
548	}
549
550	return err;
551}
552
553int mthca_poll_cq(struct ib_cq *ibcq, int num_entries,
554		  struct ib_wc *entry)
555{
556	struct mthca_dev *dev = to_mdev(ibcq->device);
557	struct mthca_cq *cq = to_mcq(ibcq);
558	struct mthca_qp *qp = NULL;
559	unsigned long flags;
560	int err = 0;
561	int freed = 0;
562	int npolled;
563
564	spin_lock_irqsave(&cq->lock, flags);
565
566	for (npolled = 0; npolled < num_entries; ++npolled) {
567		err = mthca_poll_one(dev, cq, &qp,
568				     &freed, entry + npolled);
569		if (err)
570			break;
571	}
572
573	if (freed) {
574		wmb();
575		update_cons_index(dev, cq, freed);
576	}
577
578	spin_unlock_irqrestore(&cq->lock, flags);
579
580	return err == 0 || err == -EAGAIN ? npolled : err;
581}
582
583int mthca_tavor_arm_cq(struct ib_cq *cq, enum ib_cq_notify notify)
584{
585	u32 doorbell[2];
586
587	doorbell[0] = cpu_to_be32((notify == IB_CQ_SOLICITED ?
588				   MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL :
589				   MTHCA_TAVOR_CQ_DB_REQ_NOT)      |
590				  to_mcq(cq)->cqn);
591	doorbell[1] = 0xffffffff;
592
593	mthca_write64(doorbell,
594		      to_mdev(cq->device)->kar + MTHCA_CQ_DOORBELL,
595		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(cq->device)->doorbell_lock));
596
597	return 0;
598}
599
600int mthca_arbel_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify notify)
601{
602	struct mthca_cq *cq = to_mcq(ibcq);
603	u32 doorbell[2];
604	u32 sn;
605	u32 ci;
606
607	sn = cq->arm_sn & 3;
608	ci = cpu_to_be32(cq->cons_index);
609
610	doorbell[0] = ci;
611	doorbell[1] = cpu_to_be32((cq->cqn << 8) | (2 << 5) | (sn << 3) |
612				  (notify == IB_CQ_SOLICITED ? 1 : 2));
613
614	mthca_write_db_rec(doorbell, cq->arm_db);
615
616	/*
617	 * Make sure that the doorbell record in host memory is
618	 * written before ringing the doorbell via PCI MMIO.
619	 */
620	wmb();
621
622	doorbell[0] = cpu_to_be32((sn << 28)                       |
623				  (notify == IB_CQ_SOLICITED ?
624				   MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL :
625				   MTHCA_ARBEL_CQ_DB_REQ_NOT)      |
626				  cq->cqn);
627	doorbell[1] = ci;
628
629	mthca_write64(doorbell,
630		      to_mdev(ibcq->device)->kar + MTHCA_CQ_DOORBELL,
631		      MTHCA_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->doorbell_lock));
632
633	return 0;
634}
635
636static void mthca_free_cq_buf(struct mthca_dev *dev, struct mthca_cq *cq)
637{
638	int i;
639	int size;
640
641	if (cq->is_direct)
642		pci_free_consistent(dev->pdev,
643				    (cq->ibcq.cqe + 1) * MTHCA_CQ_ENTRY_SIZE,
644				    cq->queue.direct.buf,
645				    pci_unmap_addr(&cq->queue.direct,
646						   mapping));
647	else {
648		size = (cq->ibcq.cqe + 1) * MTHCA_CQ_ENTRY_SIZE;
649		for (i = 0; i < (size + PAGE_SIZE - 1) / PAGE_SIZE; ++i)
650			if (cq->queue.page_list[i].buf)
651				pci_free_consistent(dev->pdev, PAGE_SIZE,
652						    cq->queue.page_list[i].buf,
653						    pci_unmap_addr(&cq->queue.page_list[i],
654								   mapping));
655
656		kfree(cq->queue.page_list);
657	}
658}
659
660static int mthca_alloc_cq_buf(struct mthca_dev *dev, int size,
661			      struct mthca_cq *cq)
662{
663	int err = -ENOMEM;
664	int npages, shift;
665	u64 *dma_list = NULL;
666	dma_addr_t t;
667	int i;
668
669	if (size <= MTHCA_MAX_DIRECT_CQ_SIZE) {
670		cq->is_direct = 1;
671		npages        = 1;
672		shift         = get_order(size) + PAGE_SHIFT;
673
674		cq->queue.direct.buf = pci_alloc_consistent(dev->pdev,
675							    size, &t);
676		if (!cq->queue.direct.buf)
677			return -ENOMEM;
678
679		pci_unmap_addr_set(&cq->queue.direct, mapping, t);
680
681		memset(cq->queue.direct.buf, 0, size);
682
683		while (t & ((1 << shift) - 1)) {
684			--shift;
685			npages *= 2;
686		}
687
688		dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
689		if (!dma_list)
690			goto err_free;
691
692		for (i = 0; i < npages; ++i)
693			dma_list[i] = t + i * (1 << shift);
694	} else {
695		cq->is_direct = 0;
696		npages        = (size + PAGE_SIZE - 1) / PAGE_SIZE;
697		shift         = PAGE_SHIFT;
698
699		dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
700		if (!dma_list)
701			return -ENOMEM;
702
703		cq->queue.page_list = kmalloc(npages * sizeof *cq->queue.page_list,
704					      GFP_KERNEL);
705		if (!cq->queue.page_list)
706			goto err_out;
707
708		for (i = 0; i < npages; ++i)
709			cq->queue.page_list[i].buf = NULL;
710
711		for (i = 0; i < npages; ++i) {
712			cq->queue.page_list[i].buf =
713				pci_alloc_consistent(dev->pdev, PAGE_SIZE, &t);
714			if (!cq->queue.page_list[i].buf)
715				goto err_free;
716
717			dma_list[i] = t;
718			pci_unmap_addr_set(&cq->queue.page_list[i], mapping, t);
719
720			memset(cq->queue.page_list[i].buf, 0, PAGE_SIZE);
721		}
722	}
723
724	err = mthca_mr_alloc_phys(dev, dev->driver_pd.pd_num,
725				  dma_list, shift, npages,
726				  0, size,
727				  MTHCA_MPT_FLAG_LOCAL_WRITE |
728				  MTHCA_MPT_FLAG_LOCAL_READ,
729				  &cq->mr);
730	if (err)
731		goto err_free;
732
733	kfree(dma_list);
734
735	return 0;
736
737err_free:
738	mthca_free_cq_buf(dev, cq);
739
740err_out:
741	kfree(dma_list);
742
743	return err;
744}
745
746int mthca_init_cq(struct mthca_dev *dev, int nent,
747		  struct mthca_cq *cq)
748{
749	int size = nent * MTHCA_CQ_ENTRY_SIZE;
750	void *mailbox = NULL;
751	struct mthca_cq_context *cq_context;
752	int err = -ENOMEM;
753	u8 status;
754	int i;
755
756	might_sleep();
757
758	cq->ibcq.cqe = nent - 1;
759
760	cq->cqn = mthca_alloc(&dev->cq_table.alloc);
761	if (cq->cqn == -1)
762		return -ENOMEM;
763
764	if (mthca_is_memfree(dev)) {
765		cq->arm_sn = 1;
766
767		err = mthca_table_get(dev, dev->cq_table.table, cq->cqn);
768		if (err)
769			goto err_out;
770
771		err = -ENOMEM;
772
773		cq->set_ci_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_SET_CI,
774						     cq->cqn, &cq->set_ci_db);
775		if (cq->set_ci_db_index < 0)
776			goto err_out_icm;
777
778		cq->arm_db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_CQ_ARM,
779						  cq->cqn, &cq->arm_db);
780		if (cq->arm_db_index < 0)
781			goto err_out_ci;
782	}
783
784	mailbox = kmalloc(sizeof (struct mthca_cq_context) + MTHCA_CMD_MAILBOX_EXTRA,
785			  GFP_KERNEL);
786	if (!mailbox)
787		goto err_out_mailbox;
788
789	cq_context = MAILBOX_ALIGN(mailbox);
790
791	err = mthca_alloc_cq_buf(dev, size, cq);
792	if (err)
793		goto err_out_mailbox;
794
795	for (i = 0; i < nent; ++i)
796		set_cqe_hw(get_cqe(cq, i));
797
798	spin_lock_init(&cq->lock);
799	atomic_set(&cq->refcount, 1);
800	init_waitqueue_head(&cq->wait);
801
802	memset(cq_context, 0, sizeof *cq_context);
803	cq_context->flags           = cpu_to_be32(MTHCA_CQ_STATUS_OK      |
804						  MTHCA_CQ_STATE_DISARMED |
805						  MTHCA_CQ_FLAG_TR);
806	cq_context->start           = cpu_to_be64(0);
807	cq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24 |
808						  dev->driver_uar.index);
809	cq_context->error_eqn       = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
810	cq_context->comp_eqn        = cpu_to_be32(dev->eq_table.eq[MTHCA_EQ_COMP].eqn);
811	cq_context->pd              = cpu_to_be32(dev->driver_pd.pd_num);
812	cq_context->lkey            = cpu_to_be32(cq->mr.ibmr.lkey);
813	cq_context->cqn             = cpu_to_be32(cq->cqn);
814
815	if (mthca_is_memfree(dev)) {
816		cq_context->ci_db    = cpu_to_be32(cq->set_ci_db_index);
817		cq_context->state_db = cpu_to_be32(cq->arm_db_index);
818	}
819
820	err = mthca_SW2HW_CQ(dev, cq_context, cq->cqn, &status);
821	if (err) {
822		mthca_warn(dev, "SW2HW_CQ failed (%d)\n", err);
823		goto err_out_free_mr;
824	}
825
826	if (status) {
827		mthca_warn(dev, "SW2HW_CQ returned status 0x%02x\n",
828			   status);
829		err = -EINVAL;
830		goto err_out_free_mr;
831	}
832
833	spin_lock_irq(&dev->cq_table.lock);
834	if (mthca_array_set(&dev->cq_table.cq,
835			    cq->cqn & (dev->limits.num_cqs - 1),
836			    cq)) {
837		spin_unlock_irq(&dev->cq_table.lock);
838		goto err_out_free_mr;
839	}
840	spin_unlock_irq(&dev->cq_table.lock);
841
842	cq->cons_index = 0;
843
844	kfree(mailbox);
845
846	return 0;
847
848err_out_free_mr:
849	mthca_free_mr(dev, &cq->mr);
850	mthca_free_cq_buf(dev, cq);
851
852err_out_mailbox:
853	kfree(mailbox);
854
855	if (mthca_is_memfree(dev))
856		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM, cq->arm_db_index);
857
858err_out_ci:
859	if (mthca_is_memfree(dev))
860		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
861
862err_out_icm:
863	mthca_table_put(dev, dev->cq_table.table, cq->cqn);
864
865err_out:
866	mthca_free(&dev->cq_table.alloc, cq->cqn);
867
868	return err;
869}
870
871void mthca_free_cq(struct mthca_dev *dev,
872		   struct mthca_cq *cq)
873{
874	void *mailbox;
875	int err;
876	u8 status;
877
878	might_sleep();
879
880	mailbox = kmalloc(sizeof (struct mthca_cq_context) + MTHCA_CMD_MAILBOX_EXTRA,
881			  GFP_KERNEL);
882	if (!mailbox) {
883		mthca_warn(dev, "No memory for mailbox to free CQ.\n");
884		return;
885	}
886
887	err = mthca_HW2SW_CQ(dev, MAILBOX_ALIGN(mailbox), cq->cqn, &status);
888	if (err)
889		mthca_warn(dev, "HW2SW_CQ failed (%d)\n", err);
890	else if (status)
891		mthca_warn(dev, "HW2SW_CQ returned status 0x%02x\n",
892			   status);
893
894	if (0) {
895		u32 *ctx = MAILBOX_ALIGN(mailbox);
896		int j;
897
898		printk(KERN_ERR "context for CQN %x (cons index %x, next sw %d)\n",
899		       cq->cqn, cq->cons_index, !!next_cqe_sw(cq));
900		for (j = 0; j < 16; ++j)
901			printk(KERN_ERR "[%2x] %08x\n", j * 4, be32_to_cpu(ctx[j]));
902	}
903
904	spin_lock_irq(&dev->cq_table.lock);
905	mthca_array_clear(&dev->cq_table.cq,
906			  cq->cqn & (dev->limits.num_cqs - 1));
907	spin_unlock_irq(&dev->cq_table.lock);
908
909	if (dev->mthca_flags & MTHCA_FLAG_MSI_X)
910		synchronize_irq(dev->eq_table.eq[MTHCA_EQ_COMP].msi_x_vector);
911	else
912		synchronize_irq(dev->pdev->irq);
913
914	atomic_dec(&cq->refcount);
915	wait_event(cq->wait, !atomic_read(&cq->refcount));
916
917	mthca_free_mr(dev, &cq->mr);
918	mthca_free_cq_buf(dev, cq);
919
920	if (mthca_is_memfree(dev)) {
921		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_ARM,    cq->arm_db_index);
922		mthca_free_db(dev, MTHCA_DB_TYPE_CQ_SET_CI, cq->set_ci_db_index);
923		mthca_table_put(dev, dev->cq_table.table, cq->cqn);
924	}
925
926	mthca_free(&dev->cq_table.alloc, cq->cqn);
927	kfree(mailbox);
928}
929
930int __devinit mthca_init_cq_table(struct mthca_dev *dev)
931{
932	int err;
933
934	spin_lock_init(&dev->cq_table.lock);
935
936	err = mthca_alloc_init(&dev->cq_table.alloc,
937			       dev->limits.num_cqs,
938			       (1 << 24) - 1,
939			       dev->limits.reserved_cqs);
940	if (err)
941		return err;
942
943	err = mthca_array_init(&dev->cq_table.cq,
944			       dev->limits.num_cqs);
945	if (err)
946		mthca_alloc_cleanup(&dev->cq_table.alloc);
947
948	return err;
949}
950
951void __devexit mthca_cleanup_cq_table(struct mthca_dev *dev)
952{
953	mthca_array_cleanup(&dev->cq_table.cq, dev->limits.num_cqs);
954	mthca_alloc_cleanup(&dev->cq_table.alloc);
955}
956