1 /*
2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Mellanox Technologies. 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_eq.c 1382 2004-12-24 02:21:02Z roland $
34  */
35 
36 #include <linux/errno.h>
37 #include <linux/interrupt.h>
38 #include <linux/pci.h>
39 
40 #include "mthca_dev.h"
41 #include "mthca_cmd.h"
42 #include "mthca_config_reg.h"
43 
44 enum {
45 	MTHCA_NUM_ASYNC_EQE = 0x80,
46 	MTHCA_NUM_CMD_EQE   = 0x80,
47 	MTHCA_NUM_SPARE_EQE = 0x80,
48 	MTHCA_EQ_ENTRY_SIZE = 0x20
49 };
50 
51 /*
52  * Must be packed because start is 64 bits but only aligned to 32 bits.
53  */
54 struct mthca_eq_context {
55 	__be32 flags;
56 	__be64 start;
57 	__be32 logsize_usrpage;
58 	__be32 tavor_pd;	/* reserved for Arbel */
59 	u8     reserved1[3];
60 	u8     intr;
61 	__be32 arbel_pd;	/* lost_count for Tavor */
62 	__be32 lkey;
63 	u32    reserved2[2];
64 	__be32 consumer_index;
65 	__be32 producer_index;
66 	u32    reserved3[4];
67 } __attribute__((packed));
68 
69 #define MTHCA_EQ_STATUS_OK          ( 0 << 28)
70 #define MTHCA_EQ_STATUS_OVERFLOW    ( 9 << 28)
71 #define MTHCA_EQ_STATUS_WRITE_FAIL  (10 << 28)
72 #define MTHCA_EQ_OWNER_SW           ( 0 << 24)
73 #define MTHCA_EQ_OWNER_HW           ( 1 << 24)
74 #define MTHCA_EQ_FLAG_TR            ( 1 << 18)
75 #define MTHCA_EQ_FLAG_OI            ( 1 << 17)
76 #define MTHCA_EQ_STATE_ARMED        ( 1 <<  8)
77 #define MTHCA_EQ_STATE_FIRED        ( 2 <<  8)
78 #define MTHCA_EQ_STATE_ALWAYS_ARMED ( 3 <<  8)
79 #define MTHCA_EQ_STATE_ARBEL        ( 8 <<  8)
80 
81 enum {
82 	MTHCA_EVENT_TYPE_COMP       	    = 0x00,
83 	MTHCA_EVENT_TYPE_PATH_MIG   	    = 0x01,
84 	MTHCA_EVENT_TYPE_COMM_EST   	    = 0x02,
85 	MTHCA_EVENT_TYPE_SQ_DRAINED 	    = 0x03,
86 	MTHCA_EVENT_TYPE_SRQ_QP_LAST_WQE    = 0x13,
87 	MTHCA_EVENT_TYPE_SRQ_LIMIT	    = 0x14,
88 	MTHCA_EVENT_TYPE_CQ_ERROR   	    = 0x04,
89 	MTHCA_EVENT_TYPE_WQ_CATAS_ERROR     = 0x05,
90 	MTHCA_EVENT_TYPE_EEC_CATAS_ERROR    = 0x06,
91 	MTHCA_EVENT_TYPE_PATH_MIG_FAILED    = 0x07,
92 	MTHCA_EVENT_TYPE_WQ_INVAL_REQ_ERROR = 0x10,
93 	MTHCA_EVENT_TYPE_WQ_ACCESS_ERROR    = 0x11,
94 	MTHCA_EVENT_TYPE_SRQ_CATAS_ERROR    = 0x12,
95 	MTHCA_EVENT_TYPE_LOCAL_CATAS_ERROR  = 0x08,
96 	MTHCA_EVENT_TYPE_PORT_CHANGE        = 0x09,
97 	MTHCA_EVENT_TYPE_EQ_OVERFLOW        = 0x0f,
98 	MTHCA_EVENT_TYPE_ECC_DETECT         = 0x0e,
99 	MTHCA_EVENT_TYPE_CMD                = 0x0a
100 };
101 
102 #define MTHCA_ASYNC_EVENT_MASK ((1ULL << MTHCA_EVENT_TYPE_PATH_MIG)           | \
103 				(1ULL << MTHCA_EVENT_TYPE_COMM_EST)           | \
104 				(1ULL << MTHCA_EVENT_TYPE_SQ_DRAINED)         | \
105 				(1ULL << MTHCA_EVENT_TYPE_CQ_ERROR)           | \
106 				(1ULL << MTHCA_EVENT_TYPE_WQ_CATAS_ERROR)     | \
107 				(1ULL << MTHCA_EVENT_TYPE_EEC_CATAS_ERROR)    | \
108 				(1ULL << MTHCA_EVENT_TYPE_PATH_MIG_FAILED)    | \
109 				(1ULL << MTHCA_EVENT_TYPE_WQ_INVAL_REQ_ERROR) | \
110 				(1ULL << MTHCA_EVENT_TYPE_WQ_ACCESS_ERROR)    | \
111 				(1ULL << MTHCA_EVENT_TYPE_LOCAL_CATAS_ERROR)  | \
112 				(1ULL << MTHCA_EVENT_TYPE_PORT_CHANGE)        | \
113 				(1ULL << MTHCA_EVENT_TYPE_ECC_DETECT))
114 #define MTHCA_SRQ_EVENT_MASK   ((1ULL << MTHCA_EVENT_TYPE_SRQ_CATAS_ERROR)    | \
115 				(1ULL << MTHCA_EVENT_TYPE_SRQ_QP_LAST_WQE)    | \
116 				(1ULL << MTHCA_EVENT_TYPE_SRQ_LIMIT))
117 #define MTHCA_CMD_EVENT_MASK    (1ULL << MTHCA_EVENT_TYPE_CMD)
118 
119 #define MTHCA_EQ_DB_INC_CI     (1 << 24)
120 #define MTHCA_EQ_DB_REQ_NOT    (2 << 24)
121 #define MTHCA_EQ_DB_DISARM_CQ  (3 << 24)
122 #define MTHCA_EQ_DB_SET_CI     (4 << 24)
123 #define MTHCA_EQ_DB_ALWAYS_ARM (5 << 24)
124 
125 struct mthca_eqe {
126 	u8 reserved1;
127 	u8 type;
128 	u8 reserved2;
129 	u8 subtype;
130 	union {
131 		u32 raw[6];
132 		struct {
133 			__be32 cqn;
134 		} __attribute__((packed)) comp;
135 		struct {
136 			u16    reserved1;
137 			__be16 token;
138 			u32    reserved2;
139 			u8     reserved3[3];
140 			u8     status;
141 			__be64 out_param;
142 		} __attribute__((packed)) cmd;
143 		struct {
144 			__be32 qpn;
145 		} __attribute__((packed)) qp;
146 		struct {
147 			__be32 srqn;
148 		} __attribute__((packed)) srq;
149 		struct {
150 			__be32 cqn;
151 			u32    reserved1;
152 			u8     reserved2[3];
153 			u8     syndrome;
154 		} __attribute__((packed)) cq_err;
155 		struct {
156 			u32    reserved1[2];
157 			__be32 port;
158 		} __attribute__((packed)) port_change;
159 	} event;
160 	u8 reserved3[3];
161 	u8 owner;
162 } __attribute__((packed));
163 
164 #define  MTHCA_EQ_ENTRY_OWNER_SW      (0 << 7)
165 #define  MTHCA_EQ_ENTRY_OWNER_HW      (1 << 7)
166 
167 static inline u64 async_mask(struct mthca_dev *dev)
168 {
169 	return dev->mthca_flags & MTHCA_FLAG_SRQ ?
170 		MTHCA_ASYNC_EVENT_MASK | MTHCA_SRQ_EVENT_MASK :
171 		MTHCA_ASYNC_EVENT_MASK;
172 }
173 
174 static inline void tavor_set_eq_ci(struct mthca_dev *dev, struct mthca_eq *eq, u32 ci)
175 {
176 	/*
177 	 * This barrier makes sure that all updates to ownership bits
178 	 * done by set_eqe_hw() hit memory before the consumer index
179 	 * is updated.  set_eq_ci() allows the HCA to possibly write
180 	 * more EQ entries, and we want to avoid the exceedingly
181 	 * unlikely possibility of the HCA writing an entry and then
182 	 * having set_eqe_hw() overwrite the owner field.
183 	 */
184 	wmb();
185 	mthca_write64(MTHCA_EQ_DB_SET_CI | eq->eqn, ci & (eq->nent - 1),
186 		      dev->kar + MTHCA_EQ_DOORBELL,
187 		      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
188 }
189 
190 static inline void arbel_set_eq_ci(struct mthca_dev *dev, struct mthca_eq *eq, u32 ci)
191 {
192 	/* See comment in tavor_set_eq_ci() above. */
193 	wmb();
194 	__raw_writel((__force u32) cpu_to_be32(ci),
195 		     dev->eq_regs.arbel.eq_set_ci_base + eq->eqn * 8);
196 	/* We still want ordering, just not swabbing, so add a barrier */
197 	mb();
198 }
199 
200 static inline void set_eq_ci(struct mthca_dev *dev, struct mthca_eq *eq, u32 ci)
201 {
202 	if (mthca_is_memfree(dev))
203 		arbel_set_eq_ci(dev, eq, ci);
204 	else
205 		tavor_set_eq_ci(dev, eq, ci);
206 }
207 
208 static inline void tavor_eq_req_not(struct mthca_dev *dev, int eqn)
209 {
210 	mthca_write64(MTHCA_EQ_DB_REQ_NOT | eqn, 0,
211 		      dev->kar + MTHCA_EQ_DOORBELL,
212 		      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
213 }
214 
215 static inline void arbel_eq_req_not(struct mthca_dev *dev, u32 eqn_mask)
216 {
217 	writel(eqn_mask, dev->eq_regs.arbel.eq_arm);
218 }
219 
220 static inline void disarm_cq(struct mthca_dev *dev, int eqn, int cqn)
221 {
222 	if (!mthca_is_memfree(dev)) {
223 		mthca_write64(MTHCA_EQ_DB_DISARM_CQ | eqn, cqn,
224 			      dev->kar + MTHCA_EQ_DOORBELL,
225 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
226 	}
227 }
228 
229 static inline struct mthca_eqe *get_eqe(struct mthca_eq *eq, u32 entry)
230 {
231 	unsigned long off = (entry & (eq->nent - 1)) * MTHCA_EQ_ENTRY_SIZE;
232 	return eq->page_list[off / PAGE_SIZE].buf + off % PAGE_SIZE;
233 }
234 
235 static inline struct mthca_eqe *next_eqe_sw(struct mthca_eq *eq)
236 {
237 	struct mthca_eqe *eqe;
238 	eqe = get_eqe(eq, eq->cons_index);
239 	return (MTHCA_EQ_ENTRY_OWNER_HW & eqe->owner) ? NULL : eqe;
240 }
241 
242 static inline void set_eqe_hw(struct mthca_eqe *eqe)
243 {
244 	eqe->owner =  MTHCA_EQ_ENTRY_OWNER_HW;
245 }
246 
247 static void port_change(struct mthca_dev *dev, int port, int active)
248 {
249 	struct ib_event record;
250 
251 	mthca_dbg(dev, "Port change to %s for port %d\n",
252 		  active ? "active" : "down", port);
253 
254 	record.device = &dev->ib_dev;
255 	record.event  = active ? IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
256 	record.element.port_num = port;
257 
258 	ib_dispatch_event(&record);
259 }
260 
261 static int mthca_eq_int(struct mthca_dev *dev, struct mthca_eq *eq)
262 {
263 	struct mthca_eqe *eqe;
264 	int disarm_cqn;
265 	int eqes_found = 0;
266 	int set_ci = 0;
267 
268 	while ((eqe = next_eqe_sw(eq))) {
269 		/*
270 		 * Make sure we read EQ entry contents after we've
271 		 * checked the ownership bit.
272 		 */
273 		rmb();
274 
275 		switch (eqe->type) {
276 		case MTHCA_EVENT_TYPE_COMP:
277 			disarm_cqn = be32_to_cpu(eqe->event.comp.cqn) & 0xffffff;
278 			disarm_cq(dev, eq->eqn, disarm_cqn);
279 			mthca_cq_completion(dev, disarm_cqn);
280 			break;
281 
282 		case MTHCA_EVENT_TYPE_PATH_MIG:
283 			mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
284 				       IB_EVENT_PATH_MIG);
285 			break;
286 
287 		case MTHCA_EVENT_TYPE_COMM_EST:
288 			mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
289 				       IB_EVENT_COMM_EST);
290 			break;
291 
292 		case MTHCA_EVENT_TYPE_SQ_DRAINED:
293 			mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
294 				       IB_EVENT_SQ_DRAINED);
295 			break;
296 
297 		case MTHCA_EVENT_TYPE_SRQ_QP_LAST_WQE:
298 			mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
299 				       IB_EVENT_QP_LAST_WQE_REACHED);
300 			break;
301 
302 		case MTHCA_EVENT_TYPE_SRQ_LIMIT:
303 			mthca_srq_event(dev, be32_to_cpu(eqe->event.srq.srqn) & 0xffffff,
304 					IB_EVENT_SRQ_LIMIT_REACHED);
305 			break;
306 
307 		case MTHCA_EVENT_TYPE_WQ_CATAS_ERROR:
308 			mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
309 				       IB_EVENT_QP_FATAL);
310 			break;
311 
312 		case MTHCA_EVENT_TYPE_PATH_MIG_FAILED:
313 			mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
314 				       IB_EVENT_PATH_MIG_ERR);
315 			break;
316 
317 		case MTHCA_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
318 			mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
319 				       IB_EVENT_QP_REQ_ERR);
320 			break;
321 
322 		case MTHCA_EVENT_TYPE_WQ_ACCESS_ERROR:
323 			mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
324 				       IB_EVENT_QP_ACCESS_ERR);
325 			break;
326 
327 		case MTHCA_EVENT_TYPE_CMD:
328 			mthca_cmd_event(dev,
329 					be16_to_cpu(eqe->event.cmd.token),
330 					eqe->event.cmd.status,
331 					be64_to_cpu(eqe->event.cmd.out_param));
332 			break;
333 
334 		case MTHCA_EVENT_TYPE_PORT_CHANGE:
335 			port_change(dev,
336 				    (be32_to_cpu(eqe->event.port_change.port) >> 28) & 3,
337 				    eqe->subtype == 0x4);
338 			break;
339 
340 		case MTHCA_EVENT_TYPE_CQ_ERROR:
341 			mthca_warn(dev, "CQ %s on CQN %06x\n",
342 				   eqe->event.cq_err.syndrome == 1 ?
343 				   "overrun" : "access violation",
344 				   be32_to_cpu(eqe->event.cq_err.cqn) & 0xffffff);
345 			mthca_cq_event(dev, be32_to_cpu(eqe->event.cq_err.cqn),
346 				       IB_EVENT_CQ_ERR);
347 			break;
348 
349 		case MTHCA_EVENT_TYPE_EQ_OVERFLOW:
350 			mthca_warn(dev, "EQ overrun on EQN %d\n", eq->eqn);
351 			break;
352 
353 		case MTHCA_EVENT_TYPE_EEC_CATAS_ERROR:
354 		case MTHCA_EVENT_TYPE_SRQ_CATAS_ERROR:
355 		case MTHCA_EVENT_TYPE_LOCAL_CATAS_ERROR:
356 		case MTHCA_EVENT_TYPE_ECC_DETECT:
357 		default:
358 			mthca_warn(dev, "Unhandled event %02x(%02x) on EQ %d\n",
359 				   eqe->type, eqe->subtype, eq->eqn);
360 			break;
361 		};
362 
363 		set_eqe_hw(eqe);
364 		++eq->cons_index;
365 		eqes_found = 1;
366 		++set_ci;
367 
368 		/*
369 		 * The HCA will think the queue has overflowed if we
370 		 * don't tell it we've been processing events.  We
371 		 * create our EQs with MTHCA_NUM_SPARE_EQE extra
372 		 * entries, so we must update our consumer index at
373 		 * least that often.
374 		 */
375 		if (unlikely(set_ci >= MTHCA_NUM_SPARE_EQE)) {
376 			/*
377 			 * Conditional on hca_type is OK here because
378 			 * this is a rare case, not the fast path.
379 			 */
380 			set_eq_ci(dev, eq, eq->cons_index);
381 			set_ci = 0;
382 		}
383 	}
384 
385 	/*
386 	 * Rely on caller to set consumer index so that we don't have
387 	 * to test hca_type in our interrupt handling fast path.
388 	 */
389 	return eqes_found;
390 }
391 
392 static irqreturn_t mthca_tavor_interrupt(int irq, void *dev_ptr)
393 {
394 	struct mthca_dev *dev = dev_ptr;
395 	u32 ecr;
396 	int i;
397 
398 	if (dev->eq_table.clr_mask)
399 		writel(dev->eq_table.clr_mask, dev->eq_table.clr_int);
400 
401 	ecr = readl(dev->eq_regs.tavor.ecr_base + 4);
402 	if (!ecr)
403 		return IRQ_NONE;
404 
405 	writel(ecr, dev->eq_regs.tavor.ecr_base +
406 	       MTHCA_ECR_CLR_BASE - MTHCA_ECR_BASE + 4);
407 
408 	for (i = 0; i < MTHCA_NUM_EQ; ++i)
409 		if (ecr & dev->eq_table.eq[i].eqn_mask) {
410 			if (mthca_eq_int(dev, &dev->eq_table.eq[i]))
411 				tavor_set_eq_ci(dev, &dev->eq_table.eq[i],
412 						dev->eq_table.eq[i].cons_index);
413 			tavor_eq_req_not(dev, dev->eq_table.eq[i].eqn);
414 		}
415 
416 	return IRQ_HANDLED;
417 }
418 
419 static irqreturn_t mthca_tavor_msi_x_interrupt(int irq, void *eq_ptr)
420 {
421 	struct mthca_eq  *eq  = eq_ptr;
422 	struct mthca_dev *dev = eq->dev;
423 
424 	mthca_eq_int(dev, eq);
425 	tavor_set_eq_ci(dev, eq, eq->cons_index);
426 	tavor_eq_req_not(dev, eq->eqn);
427 
428 	/* MSI-X vectors always belong to us */
429 	return IRQ_HANDLED;
430 }
431 
432 static irqreturn_t mthca_arbel_interrupt(int irq, void *dev_ptr)
433 {
434 	struct mthca_dev *dev = dev_ptr;
435 	int work = 0;
436 	int i;
437 
438 	if (dev->eq_table.clr_mask)
439 		writel(dev->eq_table.clr_mask, dev->eq_table.clr_int);
440 
441 	for (i = 0; i < MTHCA_NUM_EQ; ++i)
442 		if (mthca_eq_int(dev, &dev->eq_table.eq[i])) {
443 			work = 1;
444 			arbel_set_eq_ci(dev, &dev->eq_table.eq[i],
445 					dev->eq_table.eq[i].cons_index);
446 		}
447 
448 	arbel_eq_req_not(dev, dev->eq_table.arm_mask);
449 
450 	return IRQ_RETVAL(work);
451 }
452 
453 static irqreturn_t mthca_arbel_msi_x_interrupt(int irq, void *eq_ptr)
454 {
455 	struct mthca_eq  *eq  = eq_ptr;
456 	struct mthca_dev *dev = eq->dev;
457 
458 	mthca_eq_int(dev, eq);
459 	arbel_set_eq_ci(dev, eq, eq->cons_index);
460 	arbel_eq_req_not(dev, eq->eqn_mask);
461 
462 	/* MSI-X vectors always belong to us */
463 	return IRQ_HANDLED;
464 }
465 
466 static int mthca_create_eq(struct mthca_dev *dev,
467 			   int nent,
468 			   u8 intr,
469 			   struct mthca_eq *eq)
470 {
471 	int npages;
472 	u64 *dma_list = NULL;
473 	dma_addr_t t;
474 	struct mthca_mailbox *mailbox;
475 	struct mthca_eq_context *eq_context;
476 	int err = -ENOMEM;
477 	int i;
478 	u8 status;
479 
480 	eq->dev  = dev;
481 	eq->nent = roundup_pow_of_two(max(nent, 2));
482 	npages = ALIGN(eq->nent * MTHCA_EQ_ENTRY_SIZE, PAGE_SIZE) / PAGE_SIZE;
483 
484 	eq->page_list = kmalloc(npages * sizeof *eq->page_list,
485 				GFP_KERNEL);
486 	if (!eq->page_list)
487 		goto err_out;
488 
489 	for (i = 0; i < npages; ++i)
490 		eq->page_list[i].buf = NULL;
491 
492 	dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
493 	if (!dma_list)
494 		goto err_out_free;
495 
496 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
497 	if (IS_ERR(mailbox))
498 		goto err_out_free;
499 	eq_context = mailbox->buf;
500 
501 	for (i = 0; i < npages; ++i) {
502 		eq->page_list[i].buf = dma_alloc_coherent(&dev->pdev->dev,
503 							  PAGE_SIZE, &t, GFP_KERNEL);
504 		if (!eq->page_list[i].buf)
505 			goto err_out_free_pages;
506 
507 		dma_list[i] = t;
508 		pci_unmap_addr_set(&eq->page_list[i], mapping, t);
509 
510 		clear_page(eq->page_list[i].buf);
511 	}
512 
513 	for (i = 0; i < eq->nent; ++i)
514 		set_eqe_hw(get_eqe(eq, i));
515 
516 	eq->eqn = mthca_alloc(&dev->eq_table.alloc);
517 	if (eq->eqn == -1)
518 		goto err_out_free_pages;
519 
520 	err = mthca_mr_alloc_phys(dev, dev->driver_pd.pd_num,
521 				  dma_list, PAGE_SHIFT, npages,
522 				  0, npages * PAGE_SIZE,
523 				  MTHCA_MPT_FLAG_LOCAL_WRITE |
524 				  MTHCA_MPT_FLAG_LOCAL_READ,
525 				  &eq->mr);
526 	if (err)
527 		goto err_out_free_eq;
528 
529 	memset(eq_context, 0, sizeof *eq_context);
530 	eq_context->flags           = cpu_to_be32(MTHCA_EQ_STATUS_OK   |
531 						  MTHCA_EQ_OWNER_HW    |
532 						  MTHCA_EQ_STATE_ARMED |
533 						  MTHCA_EQ_FLAG_TR);
534 	if (mthca_is_memfree(dev))
535 		eq_context->flags  |= cpu_to_be32(MTHCA_EQ_STATE_ARBEL);
536 
537 	eq_context->logsize_usrpage = cpu_to_be32((ffs(eq->nent) - 1) << 24);
538 	if (mthca_is_memfree(dev)) {
539 		eq_context->arbel_pd = cpu_to_be32(dev->driver_pd.pd_num);
540 	} else {
541 		eq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
542 		eq_context->tavor_pd         = cpu_to_be32(dev->driver_pd.pd_num);
543 	}
544 	eq_context->intr            = intr;
545 	eq_context->lkey            = cpu_to_be32(eq->mr.ibmr.lkey);
546 
547 	err = mthca_SW2HW_EQ(dev, mailbox, eq->eqn, &status);
548 	if (err) {
549 		mthca_warn(dev, "SW2HW_EQ failed (%d)\n", err);
550 		goto err_out_free_mr;
551 	}
552 	if (status) {
553 		mthca_warn(dev, "SW2HW_EQ returned status 0x%02x\n",
554 			   status);
555 		err = -EINVAL;
556 		goto err_out_free_mr;
557 	}
558 
559 	kfree(dma_list);
560 	mthca_free_mailbox(dev, mailbox);
561 
562 	eq->eqn_mask   = swab32(1 << eq->eqn);
563 	eq->cons_index = 0;
564 
565 	dev->eq_table.arm_mask |= eq->eqn_mask;
566 
567 	mthca_dbg(dev, "Allocated EQ %d with %d entries\n",
568 		  eq->eqn, eq->nent);
569 
570 	return err;
571 
572  err_out_free_mr:
573 	mthca_free_mr(dev, &eq->mr);
574 
575  err_out_free_eq:
576 	mthca_free(&dev->eq_table.alloc, eq->eqn);
577 
578  err_out_free_pages:
579 	for (i = 0; i < npages; ++i)
580 		if (eq->page_list[i].buf)
581 			dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
582 					  eq->page_list[i].buf,
583 					  pci_unmap_addr(&eq->page_list[i],
584 							 mapping));
585 
586 	mthca_free_mailbox(dev, mailbox);
587 
588  err_out_free:
589 	kfree(eq->page_list);
590 	kfree(dma_list);
591 
592  err_out:
593 	return err;
594 }
595 
596 static void mthca_free_eq(struct mthca_dev *dev,
597 			  struct mthca_eq *eq)
598 {
599 	struct mthca_mailbox *mailbox;
600 	int err;
601 	u8 status;
602 	int npages = (eq->nent * MTHCA_EQ_ENTRY_SIZE + PAGE_SIZE - 1) /
603 		PAGE_SIZE;
604 	int i;
605 
606 	mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
607 	if (IS_ERR(mailbox))
608 		return;
609 
610 	err = mthca_HW2SW_EQ(dev, mailbox, eq->eqn, &status);
611 	if (err)
612 		mthca_warn(dev, "HW2SW_EQ failed (%d)\n", err);
613 	if (status)
614 		mthca_warn(dev, "HW2SW_EQ returned status 0x%02x\n", status);
615 
616 	dev->eq_table.arm_mask &= ~eq->eqn_mask;
617 
618 	if (0) {
619 		mthca_dbg(dev, "Dumping EQ context %02x:\n", eq->eqn);
620 		for (i = 0; i < sizeof (struct mthca_eq_context) / 4; ++i) {
621 			if (i % 4 == 0)
622 				printk("[%02x] ", i * 4);
623 			printk(" %08x", be32_to_cpup(mailbox->buf + i * 4));
624 			if ((i + 1) % 4 == 0)
625 				printk("\n");
626 		}
627 	}
628 
629 	mthca_free_mr(dev, &eq->mr);
630 	for (i = 0; i < npages; ++i)
631 		pci_free_consistent(dev->pdev, PAGE_SIZE,
632 				    eq->page_list[i].buf,
633 				    pci_unmap_addr(&eq->page_list[i], mapping));
634 
635 	kfree(eq->page_list);
636 	mthca_free_mailbox(dev, mailbox);
637 }
638 
639 static void mthca_free_irqs(struct mthca_dev *dev)
640 {
641 	int i;
642 
643 	if (dev->eq_table.have_irq)
644 		free_irq(dev->pdev->irq, dev);
645 	for (i = 0; i < MTHCA_NUM_EQ; ++i)
646 		if (dev->eq_table.eq[i].have_irq)
647 			free_irq(dev->eq_table.eq[i].msi_x_vector,
648 				 dev->eq_table.eq + i);
649 }
650 
651 static int mthca_map_reg(struct mthca_dev *dev,
652 			 unsigned long offset, unsigned long size,
653 			 void __iomem **map)
654 {
655 	unsigned long base = pci_resource_start(dev->pdev, 0);
656 
657 	if (!request_mem_region(base + offset, size, DRV_NAME))
658 		return -EBUSY;
659 
660 	*map = ioremap(base + offset, size);
661 	if (!*map) {
662 		release_mem_region(base + offset, size);
663 		return -ENOMEM;
664 	}
665 
666 	return 0;
667 }
668 
669 static void mthca_unmap_reg(struct mthca_dev *dev, unsigned long offset,
670 			    unsigned long size, void __iomem *map)
671 {
672 	unsigned long base = pci_resource_start(dev->pdev, 0);
673 
674 	release_mem_region(base + offset, size);
675 	iounmap(map);
676 }
677 
678 static int mthca_map_eq_regs(struct mthca_dev *dev)
679 {
680 	if (mthca_is_memfree(dev)) {
681 		/*
682 		 * We assume that the EQ arm and EQ set CI registers
683 		 * fall within the first BAR.  We can't trust the
684 		 * values firmware gives us, since those addresses are
685 		 * valid on the HCA's side of the PCI bus but not
686 		 * necessarily the host side.
687 		 */
688 		if (mthca_map_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
689 				  dev->fw.arbel.clr_int_base, MTHCA_CLR_INT_SIZE,
690 				  &dev->clr_base)) {
691 			mthca_err(dev, "Couldn't map interrupt clear register, "
692 				  "aborting.\n");
693 			return -ENOMEM;
694 		}
695 
696 		/*
697 		 * Add 4 because we limit ourselves to EQs 0 ... 31,
698 		 * so we only need the low word of the register.
699 		 */
700 		if (mthca_map_reg(dev, ((pci_resource_len(dev->pdev, 0) - 1) &
701 					dev->fw.arbel.eq_arm_base) + 4, 4,
702 				  &dev->eq_regs.arbel.eq_arm)) {
703 			mthca_err(dev, "Couldn't map EQ arm register, aborting.\n");
704 			mthca_unmap_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
705 					dev->fw.arbel.clr_int_base, MTHCA_CLR_INT_SIZE,
706 					dev->clr_base);
707 			return -ENOMEM;
708 		}
709 
710 		if (mthca_map_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
711 				  dev->fw.arbel.eq_set_ci_base,
712 				  MTHCA_EQ_SET_CI_SIZE,
713 				  &dev->eq_regs.arbel.eq_set_ci_base)) {
714 			mthca_err(dev, "Couldn't map EQ CI register, aborting.\n");
715 			mthca_unmap_reg(dev, ((pci_resource_len(dev->pdev, 0) - 1) &
716 					      dev->fw.arbel.eq_arm_base) + 4, 4,
717 					dev->eq_regs.arbel.eq_arm);
718 			mthca_unmap_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
719 					dev->fw.arbel.clr_int_base, MTHCA_CLR_INT_SIZE,
720 					dev->clr_base);
721 			return -ENOMEM;
722 		}
723 	} else {
724 		if (mthca_map_reg(dev, MTHCA_CLR_INT_BASE, MTHCA_CLR_INT_SIZE,
725 				  &dev->clr_base)) {
726 			mthca_err(dev, "Couldn't map interrupt clear register, "
727 				  "aborting.\n");
728 			return -ENOMEM;
729 		}
730 
731 		if (mthca_map_reg(dev, MTHCA_ECR_BASE,
732 				  MTHCA_ECR_SIZE + MTHCA_ECR_CLR_SIZE,
733 				  &dev->eq_regs.tavor.ecr_base)) {
734 			mthca_err(dev, "Couldn't map ecr register, "
735 				  "aborting.\n");
736 			mthca_unmap_reg(dev, MTHCA_CLR_INT_BASE, MTHCA_CLR_INT_SIZE,
737 					dev->clr_base);
738 			return -ENOMEM;
739 		}
740 	}
741 
742 	return 0;
743 
744 }
745 
746 static void mthca_unmap_eq_regs(struct mthca_dev *dev)
747 {
748 	if (mthca_is_memfree(dev)) {
749 		mthca_unmap_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
750 				dev->fw.arbel.eq_set_ci_base,
751 				MTHCA_EQ_SET_CI_SIZE,
752 				dev->eq_regs.arbel.eq_set_ci_base);
753 		mthca_unmap_reg(dev, ((pci_resource_len(dev->pdev, 0) - 1) &
754 				      dev->fw.arbel.eq_arm_base) + 4, 4,
755 				dev->eq_regs.arbel.eq_arm);
756 		mthca_unmap_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
757 				dev->fw.arbel.clr_int_base, MTHCA_CLR_INT_SIZE,
758 				dev->clr_base);
759 	} else {
760 		mthca_unmap_reg(dev, MTHCA_ECR_BASE,
761 				MTHCA_ECR_SIZE + MTHCA_ECR_CLR_SIZE,
762 				dev->eq_regs.tavor.ecr_base);
763 		mthca_unmap_reg(dev, MTHCA_CLR_INT_BASE, MTHCA_CLR_INT_SIZE,
764 				dev->clr_base);
765 	}
766 }
767 
768 int mthca_map_eq_icm(struct mthca_dev *dev, u64 icm_virt)
769 {
770 	int ret;
771 	u8 status;
772 
773 	/*
774 	 * We assume that mapping one page is enough for the whole EQ
775 	 * context table.  This is fine with all current HCAs, because
776 	 * we only use 32 EQs and each EQ uses 32 bytes of context
777 	 * memory, or 1 KB total.
778 	 */
779 	dev->eq_table.icm_virt = icm_virt;
780 	dev->eq_table.icm_page = alloc_page(GFP_HIGHUSER);
781 	if (!dev->eq_table.icm_page)
782 		return -ENOMEM;
783 	dev->eq_table.icm_dma  = pci_map_page(dev->pdev, dev->eq_table.icm_page, 0,
784 					      PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
785 	if (pci_dma_mapping_error(dev->eq_table.icm_dma)) {
786 		__free_page(dev->eq_table.icm_page);
787 		return -ENOMEM;
788 	}
789 
790 	ret = mthca_MAP_ICM_page(dev, dev->eq_table.icm_dma, icm_virt, &status);
791 	if (!ret && status)
792 		ret = -EINVAL;
793 	if (ret) {
794 		pci_unmap_page(dev->pdev, dev->eq_table.icm_dma, PAGE_SIZE,
795 			       PCI_DMA_BIDIRECTIONAL);
796 		__free_page(dev->eq_table.icm_page);
797 	}
798 
799 	return ret;
800 }
801 
802 void mthca_unmap_eq_icm(struct mthca_dev *dev)
803 {
804 	u8 status;
805 
806 	mthca_UNMAP_ICM(dev, dev->eq_table.icm_virt, 1, &status);
807 	pci_unmap_page(dev->pdev, dev->eq_table.icm_dma, PAGE_SIZE,
808 		       PCI_DMA_BIDIRECTIONAL);
809 	__free_page(dev->eq_table.icm_page);
810 }
811 
812 int mthca_init_eq_table(struct mthca_dev *dev)
813 {
814 	int err;
815 	u8 status;
816 	u8 intr;
817 	int i;
818 
819 	err = mthca_alloc_init(&dev->eq_table.alloc,
820 			       dev->limits.num_eqs,
821 			       dev->limits.num_eqs - 1,
822 			       dev->limits.reserved_eqs);
823 	if (err)
824 		return err;
825 
826 	err = mthca_map_eq_regs(dev);
827 	if (err)
828 		goto err_out_free;
829 
830 	if (dev->mthca_flags & MTHCA_FLAG_MSI_X) {
831 		dev->eq_table.clr_mask = 0;
832 	} else {
833 		dev->eq_table.clr_mask =
834 			swab32(1 << (dev->eq_table.inta_pin & 31));
835 		dev->eq_table.clr_int  = dev->clr_base +
836 			(dev->eq_table.inta_pin < 32 ? 4 : 0);
837 	}
838 
839 	dev->eq_table.arm_mask = 0;
840 
841 	intr = dev->eq_table.inta_pin;
842 
843 	err = mthca_create_eq(dev, dev->limits.num_cqs + MTHCA_NUM_SPARE_EQE,
844 			      (dev->mthca_flags & MTHCA_FLAG_MSI_X) ? 128 : intr,
845 			      &dev->eq_table.eq[MTHCA_EQ_COMP]);
846 	if (err)
847 		goto err_out_unmap;
848 
849 	err = mthca_create_eq(dev, MTHCA_NUM_ASYNC_EQE + MTHCA_NUM_SPARE_EQE,
850 			      (dev->mthca_flags & MTHCA_FLAG_MSI_X) ? 129 : intr,
851 			      &dev->eq_table.eq[MTHCA_EQ_ASYNC]);
852 	if (err)
853 		goto err_out_comp;
854 
855 	err = mthca_create_eq(dev, MTHCA_NUM_CMD_EQE + MTHCA_NUM_SPARE_EQE,
856 			      (dev->mthca_flags & MTHCA_FLAG_MSI_X) ? 130 : intr,
857 			      &dev->eq_table.eq[MTHCA_EQ_CMD]);
858 	if (err)
859 		goto err_out_async;
860 
861 	if (dev->mthca_flags & MTHCA_FLAG_MSI_X) {
862 		static const char *eq_name[] = {
863 			[MTHCA_EQ_COMP]  = DRV_NAME " (comp)",
864 			[MTHCA_EQ_ASYNC] = DRV_NAME " (async)",
865 			[MTHCA_EQ_CMD]   = DRV_NAME " (cmd)"
866 		};
867 
868 		for (i = 0; i < MTHCA_NUM_EQ; ++i) {
869 			err = request_irq(dev->eq_table.eq[i].msi_x_vector,
870 					  mthca_is_memfree(dev) ?
871 					  mthca_arbel_msi_x_interrupt :
872 					  mthca_tavor_msi_x_interrupt,
873 					  0, eq_name[i], dev->eq_table.eq + i);
874 			if (err)
875 				goto err_out_cmd;
876 			dev->eq_table.eq[i].have_irq = 1;
877 		}
878 	} else {
879 		err = request_irq(dev->pdev->irq,
880 				  mthca_is_memfree(dev) ?
881 				  mthca_arbel_interrupt :
882 				  mthca_tavor_interrupt,
883 				  IRQF_SHARED, DRV_NAME, dev);
884 		if (err)
885 			goto err_out_cmd;
886 		dev->eq_table.have_irq = 1;
887 	}
888 
889 	err = mthca_MAP_EQ(dev, async_mask(dev),
890 			   0, dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn, &status);
891 	if (err)
892 		mthca_warn(dev, "MAP_EQ for async EQ %d failed (%d)\n",
893 			   dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn, err);
894 	if (status)
895 		mthca_warn(dev, "MAP_EQ for async EQ %d returned status 0x%02x\n",
896 			   dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn, status);
897 
898 	err = mthca_MAP_EQ(dev, MTHCA_CMD_EVENT_MASK,
899 			   0, dev->eq_table.eq[MTHCA_EQ_CMD].eqn, &status);
900 	if (err)
901 		mthca_warn(dev, "MAP_EQ for cmd EQ %d failed (%d)\n",
902 			   dev->eq_table.eq[MTHCA_EQ_CMD].eqn, err);
903 	if (status)
904 		mthca_warn(dev, "MAP_EQ for cmd EQ %d returned status 0x%02x\n",
905 			   dev->eq_table.eq[MTHCA_EQ_CMD].eqn, status);
906 
907 	for (i = 0; i < MTHCA_NUM_EQ; ++i)
908 		if (mthca_is_memfree(dev))
909 			arbel_eq_req_not(dev, dev->eq_table.eq[i].eqn_mask);
910 		else
911 			tavor_eq_req_not(dev, dev->eq_table.eq[i].eqn);
912 
913 	return 0;
914 
915 err_out_cmd:
916 	mthca_free_irqs(dev);
917 	mthca_free_eq(dev, &dev->eq_table.eq[MTHCA_EQ_CMD]);
918 
919 err_out_async:
920 	mthca_free_eq(dev, &dev->eq_table.eq[MTHCA_EQ_ASYNC]);
921 
922 err_out_comp:
923 	mthca_free_eq(dev, &dev->eq_table.eq[MTHCA_EQ_COMP]);
924 
925 err_out_unmap:
926 	mthca_unmap_eq_regs(dev);
927 
928 err_out_free:
929 	mthca_alloc_cleanup(&dev->eq_table.alloc);
930 	return err;
931 }
932 
933 void mthca_cleanup_eq_table(struct mthca_dev *dev)
934 {
935 	u8 status;
936 	int i;
937 
938 	mthca_free_irqs(dev);
939 
940 	mthca_MAP_EQ(dev, async_mask(dev),
941 		     1, dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn, &status);
942 	mthca_MAP_EQ(dev, MTHCA_CMD_EVENT_MASK,
943 		     1, dev->eq_table.eq[MTHCA_EQ_CMD].eqn, &status);
944 
945 	for (i = 0; i < MTHCA_NUM_EQ; ++i)
946 		mthca_free_eq(dev, &dev->eq_table.eq[i]);
947 
948 	mthca_unmap_eq_regs(dev);
949 
950 	mthca_alloc_cleanup(&dev->eq_table.alloc);
951 }
952