xref: /openbmc/linux/drivers/infiniband/hw/mthca/mthca_eq.c (revision 1da177e4c3f41524e886b7f1b8a0c1fc7321cac2)
1 /*
2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  *
32  * $Id: mthca_eq.c 1382 2004-12-24 02:21:02Z roland $
33  */
34 
35 #include <linux/init.h>
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_EQ_ENTRY_SIZE = 0x20
48 };
49 
50 /*
51  * Must be packed because start is 64 bits but only aligned to 32 bits.
52  */
53 struct mthca_eq_context {
54 	u32 flags;
55 	u64 start;
56 	u32 logsize_usrpage;
57 	u32 tavor_pd;		/* reserved for Arbel */
58 	u8  reserved1[3];
59 	u8  intr;
60 	u32 arbel_pd;		/* lost_count for Tavor */
61 	u32 lkey;
62 	u32 reserved2[2];
63 	u32 consumer_index;
64 	u32 producer_index;
65 	u32 reserved3[4];
66 } __attribute__((packed));
67 
68 #define MTHCA_EQ_STATUS_OK          ( 0 << 28)
69 #define MTHCA_EQ_STATUS_OVERFLOW    ( 9 << 28)
70 #define MTHCA_EQ_STATUS_WRITE_FAIL  (10 << 28)
71 #define MTHCA_EQ_OWNER_SW           ( 0 << 24)
72 #define MTHCA_EQ_OWNER_HW           ( 1 << 24)
73 #define MTHCA_EQ_FLAG_TR            ( 1 << 18)
74 #define MTHCA_EQ_FLAG_OI            ( 1 << 17)
75 #define MTHCA_EQ_STATE_ARMED        ( 1 <<  8)
76 #define MTHCA_EQ_STATE_FIRED        ( 2 <<  8)
77 #define MTHCA_EQ_STATE_ALWAYS_ARMED ( 3 <<  8)
78 #define MTHCA_EQ_STATE_ARBEL        ( 8 <<  8)
79 
80 enum {
81 	MTHCA_EVENT_TYPE_COMP       	    = 0x00,
82 	MTHCA_EVENT_TYPE_PATH_MIG   	    = 0x01,
83 	MTHCA_EVENT_TYPE_COMM_EST   	    = 0x02,
84 	MTHCA_EVENT_TYPE_SQ_DRAINED 	    = 0x03,
85 	MTHCA_EVENT_TYPE_SRQ_LAST_WQE       = 0x13,
86 	MTHCA_EVENT_TYPE_CQ_ERROR   	    = 0x04,
87 	MTHCA_EVENT_TYPE_WQ_CATAS_ERROR     = 0x05,
88 	MTHCA_EVENT_TYPE_EEC_CATAS_ERROR    = 0x06,
89 	MTHCA_EVENT_TYPE_PATH_MIG_FAILED    = 0x07,
90 	MTHCA_EVENT_TYPE_WQ_INVAL_REQ_ERROR = 0x10,
91 	MTHCA_EVENT_TYPE_WQ_ACCESS_ERROR    = 0x11,
92 	MTHCA_EVENT_TYPE_SRQ_CATAS_ERROR    = 0x12,
93 	MTHCA_EVENT_TYPE_LOCAL_CATAS_ERROR  = 0x08,
94 	MTHCA_EVENT_TYPE_PORT_CHANGE        = 0x09,
95 	MTHCA_EVENT_TYPE_EQ_OVERFLOW        = 0x0f,
96 	MTHCA_EVENT_TYPE_ECC_DETECT         = 0x0e,
97 	MTHCA_EVENT_TYPE_CMD                = 0x0a
98 };
99 
100 #define MTHCA_ASYNC_EVENT_MASK ((1ULL << MTHCA_EVENT_TYPE_PATH_MIG)           | \
101 				(1ULL << MTHCA_EVENT_TYPE_COMM_EST)           | \
102 				(1ULL << MTHCA_EVENT_TYPE_SQ_DRAINED)         | \
103 				(1ULL << MTHCA_EVENT_TYPE_CQ_ERROR)           | \
104 				(1ULL << MTHCA_EVENT_TYPE_WQ_CATAS_ERROR)     | \
105 				(1ULL << MTHCA_EVENT_TYPE_EEC_CATAS_ERROR)    | \
106 				(1ULL << MTHCA_EVENT_TYPE_PATH_MIG_FAILED)    | \
107 				(1ULL << MTHCA_EVENT_TYPE_WQ_INVAL_REQ_ERROR) | \
108 				(1ULL << MTHCA_EVENT_TYPE_WQ_ACCESS_ERROR)    | \
109 				(1ULL << MTHCA_EVENT_TYPE_LOCAL_CATAS_ERROR)  | \
110 				(1ULL << MTHCA_EVENT_TYPE_PORT_CHANGE)        | \
111 				(1ULL << MTHCA_EVENT_TYPE_ECC_DETECT))
112 #define MTHCA_SRQ_EVENT_MASK    (1ULL << MTHCA_EVENT_TYPE_SRQ_CATAS_ERROR)    | \
113 				(1ULL << MTHCA_EVENT_TYPE_SRQ_LAST_WQE)
114 #define MTHCA_CMD_EVENT_MASK    (1ULL << MTHCA_EVENT_TYPE_CMD)
115 
116 #define MTHCA_EQ_DB_INC_CI     (1 << 24)
117 #define MTHCA_EQ_DB_REQ_NOT    (2 << 24)
118 #define MTHCA_EQ_DB_DISARM_CQ  (3 << 24)
119 #define MTHCA_EQ_DB_SET_CI     (4 << 24)
120 #define MTHCA_EQ_DB_ALWAYS_ARM (5 << 24)
121 
122 struct mthca_eqe {
123 	u8 reserved1;
124 	u8 type;
125 	u8 reserved2;
126 	u8 subtype;
127 	union {
128 		u32 raw[6];
129 		struct {
130 			u32 cqn;
131 		} __attribute__((packed)) comp;
132 		struct {
133 			u16 reserved1;
134 			u16 token;
135 			u32 reserved2;
136 			u8  reserved3[3];
137 			u8  status;
138 			u64 out_param;
139 		} __attribute__((packed)) cmd;
140 		struct {
141 			u32 qpn;
142 		} __attribute__((packed)) qp;
143 		struct {
144 			u32 cqn;
145 			u32 reserved1;
146 			u8  reserved2[3];
147 			u8  syndrome;
148 		} __attribute__((packed)) cq_err;
149 		struct {
150 			u32 reserved1[2];
151 			u32 port;
152 		} __attribute__((packed)) port_change;
153 	} event;
154 	u8 reserved3[3];
155 	u8 owner;
156 } __attribute__((packed));
157 
158 #define  MTHCA_EQ_ENTRY_OWNER_SW      (0 << 7)
159 #define  MTHCA_EQ_ENTRY_OWNER_HW      (1 << 7)
160 
161 static inline u64 async_mask(struct mthca_dev *dev)
162 {
163 	return dev->mthca_flags & MTHCA_FLAG_SRQ ?
164 		MTHCA_ASYNC_EVENT_MASK | MTHCA_SRQ_EVENT_MASK :
165 		MTHCA_ASYNC_EVENT_MASK;
166 }
167 
168 static inline void tavor_set_eq_ci(struct mthca_dev *dev, struct mthca_eq *eq, u32 ci)
169 {
170 	u32 doorbell[2];
171 
172 	doorbell[0] = cpu_to_be32(MTHCA_EQ_DB_SET_CI | eq->eqn);
173 	doorbell[1] = cpu_to_be32(ci & (eq->nent - 1));
174 
175 	/*
176 	 * This barrier makes sure that all updates to ownership bits
177 	 * done by set_eqe_hw() hit memory before the consumer index
178 	 * is updated.  set_eq_ci() allows the HCA to possibly write
179 	 * more EQ entries, and we want to avoid the exceedingly
180 	 * unlikely possibility of the HCA writing an entry and then
181 	 * having set_eqe_hw() overwrite the owner field.
182 	 */
183 	wmb();
184 	mthca_write64(doorbell,
185 		      dev->kar + MTHCA_EQ_DOORBELL,
186 		      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
187 }
188 
189 static inline void arbel_set_eq_ci(struct mthca_dev *dev, struct mthca_eq *eq, u32 ci)
190 {
191 	/* See comment in tavor_set_eq_ci() above. */
192 	wmb();
193 	__raw_writel(cpu_to_be32(ci), dev->eq_regs.arbel.eq_set_ci_base +
194 		     eq->eqn * 8);
195 	/* We still want ordering, just not swabbing, so add a barrier */
196 	mb();
197 }
198 
199 static inline void set_eq_ci(struct mthca_dev *dev, struct mthca_eq *eq, u32 ci)
200 {
201 	if (dev->hca_type == ARBEL_NATIVE)
202 		arbel_set_eq_ci(dev, eq, ci);
203 	else
204 		tavor_set_eq_ci(dev, eq, ci);
205 }
206 
207 static inline void tavor_eq_req_not(struct mthca_dev *dev, int eqn)
208 {
209 	u32 doorbell[2];
210 
211 	doorbell[0] = cpu_to_be32(MTHCA_EQ_DB_REQ_NOT | eqn);
212 	doorbell[1] = 0;
213 
214 	mthca_write64(doorbell,
215 		      dev->kar + MTHCA_EQ_DOORBELL,
216 		      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
217 }
218 
219 static inline void arbel_eq_req_not(struct mthca_dev *dev, u32 eqn_mask)
220 {
221 	writel(eqn_mask, dev->eq_regs.arbel.eq_arm);
222 }
223 
224 static inline void disarm_cq(struct mthca_dev *dev, int eqn, int cqn)
225 {
226 	if (dev->hca_type != ARBEL_NATIVE) {
227 		u32 doorbell[2];
228 
229 		doorbell[0] = cpu_to_be32(MTHCA_EQ_DB_DISARM_CQ | eqn);
230 		doorbell[1] = cpu_to_be32(cqn);
231 
232 		mthca_write64(doorbell,
233 			      dev->kar + MTHCA_EQ_DOORBELL,
234 			      MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
235 	}
236 }
237 
238 static inline struct mthca_eqe *get_eqe(struct mthca_eq *eq, u32 entry)
239 {
240 	unsigned long off = (entry & (eq->nent - 1)) * MTHCA_EQ_ENTRY_SIZE;
241 	return eq->page_list[off / PAGE_SIZE].buf + off % PAGE_SIZE;
242 }
243 
244 static inline struct mthca_eqe* next_eqe_sw(struct mthca_eq *eq)
245 {
246 	struct mthca_eqe* eqe;
247 	eqe = get_eqe(eq, eq->cons_index);
248 	return (MTHCA_EQ_ENTRY_OWNER_HW & eqe->owner) ? NULL : eqe;
249 }
250 
251 static inline void set_eqe_hw(struct mthca_eqe *eqe)
252 {
253 	eqe->owner =  MTHCA_EQ_ENTRY_OWNER_HW;
254 }
255 
256 static void port_change(struct mthca_dev *dev, int port, int active)
257 {
258 	struct ib_event record;
259 
260 	mthca_dbg(dev, "Port change to %s for port %d\n",
261 		  active ? "active" : "down", port);
262 
263 	record.device = &dev->ib_dev;
264 	record.event  = active ? IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
265 	record.element.port_num = port;
266 
267 	ib_dispatch_event(&record);
268 }
269 
270 static int mthca_eq_int(struct mthca_dev *dev, struct mthca_eq *eq)
271 {
272 	struct mthca_eqe *eqe;
273 	int disarm_cqn;
274 	int  eqes_found = 0;
275 
276 	while ((eqe = next_eqe_sw(eq))) {
277 		int set_ci = 0;
278 
279 		/*
280 		 * Make sure we read EQ entry contents after we've
281 		 * checked the ownership bit.
282 		 */
283 		rmb();
284 
285 		switch (eqe->type) {
286 		case MTHCA_EVENT_TYPE_COMP:
287 			disarm_cqn = be32_to_cpu(eqe->event.comp.cqn) & 0xffffff;
288 			disarm_cq(dev, eq->eqn, disarm_cqn);
289 			mthca_cq_event(dev, disarm_cqn);
290 			break;
291 
292 		case MTHCA_EVENT_TYPE_PATH_MIG:
293 			mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
294 				       IB_EVENT_PATH_MIG);
295 			break;
296 
297 		case MTHCA_EVENT_TYPE_COMM_EST:
298 			mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
299 				       IB_EVENT_COMM_EST);
300 			break;
301 
302 		case MTHCA_EVENT_TYPE_SQ_DRAINED:
303 			mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
304 				       IB_EVENT_SQ_DRAINED);
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 			/*
333 			 * cmd_event() may add more commands.
334 			 * The card will think the queue has overflowed if
335 			 * we don't tell it we've been processing events.
336 			 */
337 			set_ci = 1;
338 			break;
339 
340 		case MTHCA_EVENT_TYPE_PORT_CHANGE:
341 			port_change(dev,
342 				    (be32_to_cpu(eqe->event.port_change.port) >> 28) & 3,
343 				    eqe->subtype == 0x4);
344 			break;
345 
346 		case MTHCA_EVENT_TYPE_CQ_ERROR:
347 			mthca_warn(dev, "CQ %s on CQN %08x\n",
348 				   eqe->event.cq_err.syndrome == 1 ?
349 				   "overrun" : "access violation",
350 				   be32_to_cpu(eqe->event.cq_err.cqn));
351 			break;
352 
353 		case MTHCA_EVENT_TYPE_EQ_OVERFLOW:
354 			mthca_warn(dev, "EQ overrun on EQN %d\n", eq->eqn);
355 			break;
356 
357 		case MTHCA_EVENT_TYPE_EEC_CATAS_ERROR:
358 		case MTHCA_EVENT_TYPE_SRQ_CATAS_ERROR:
359 		case MTHCA_EVENT_TYPE_LOCAL_CATAS_ERROR:
360 		case MTHCA_EVENT_TYPE_ECC_DETECT:
361 		default:
362 			mthca_warn(dev, "Unhandled event %02x(%02x) on EQ %d\n",
363 				   eqe->type, eqe->subtype, eq->eqn);
364 			break;
365 		};
366 
367 		set_eqe_hw(eqe);
368 		++eq->cons_index;
369 		eqes_found = 1;
370 
371 		if (unlikely(set_ci)) {
372 			/*
373 			 * Conditional on hca_type is OK here because
374 			 * this is a rare case, not the fast path.
375 			 */
376 			set_eq_ci(dev, eq, eq->cons_index);
377 			set_ci = 0;
378 		}
379 	}
380 
381 	/*
382 	 * Rely on caller to set consumer index so that we don't have
383 	 * to test hca_type in our interrupt handling fast path.
384 	 */
385 	return eqes_found;
386 }
387 
388 static irqreturn_t mthca_tavor_interrupt(int irq, void *dev_ptr, struct pt_regs *regs)
389 {
390 	struct mthca_dev *dev = dev_ptr;
391 	u32 ecr;
392 	int i;
393 
394 	if (dev->eq_table.clr_mask)
395 		writel(dev->eq_table.clr_mask, dev->eq_table.clr_int);
396 
397 	ecr = readl(dev->eq_regs.tavor.ecr_base + 4);
398 	if (ecr) {
399 		writel(ecr, dev->eq_regs.tavor.ecr_base +
400 		       MTHCA_ECR_CLR_BASE - MTHCA_ECR_BASE + 4);
401 
402 		for (i = 0; i < MTHCA_NUM_EQ; ++i)
403 			if (ecr & dev->eq_table.eq[i].eqn_mask &&
404 			    mthca_eq_int(dev, &dev->eq_table.eq[i])) {
405 				tavor_set_eq_ci(dev, &dev->eq_table.eq[i],
406 						dev->eq_table.eq[i].cons_index);
407 				tavor_eq_req_not(dev, dev->eq_table.eq[i].eqn);
408 			}
409 	}
410 
411 	return IRQ_RETVAL(ecr);
412 }
413 
414 static irqreturn_t mthca_tavor_msi_x_interrupt(int irq, void *eq_ptr,
415 					 struct pt_regs *regs)
416 {
417 	struct mthca_eq  *eq  = eq_ptr;
418 	struct mthca_dev *dev = eq->dev;
419 
420 	mthca_eq_int(dev, eq);
421 	tavor_set_eq_ci(dev, eq, eq->cons_index);
422 	tavor_eq_req_not(dev, eq->eqn);
423 
424 	/* MSI-X vectors always belong to us */
425 	return IRQ_HANDLED;
426 }
427 
428 static irqreturn_t mthca_arbel_interrupt(int irq, void *dev_ptr, struct pt_regs *regs)
429 {
430 	struct mthca_dev *dev = dev_ptr;
431 	int work = 0;
432 	int i;
433 
434 	if (dev->eq_table.clr_mask)
435 		writel(dev->eq_table.clr_mask, dev->eq_table.clr_int);
436 
437 	for (i = 0; i < MTHCA_NUM_EQ; ++i)
438 		if (mthca_eq_int(dev, &dev->eq_table.eq[i])) {
439 			work = 1;
440 			arbel_set_eq_ci(dev, &dev->eq_table.eq[i],
441 					dev->eq_table.eq[i].cons_index);
442 		}
443 
444 	arbel_eq_req_not(dev, dev->eq_table.arm_mask);
445 
446 	return IRQ_RETVAL(work);
447 }
448 
449 static irqreturn_t mthca_arbel_msi_x_interrupt(int irq, void *eq_ptr,
450 					       struct pt_regs *regs)
451 {
452 	struct mthca_eq  *eq  = eq_ptr;
453 	struct mthca_dev *dev = eq->dev;
454 
455 	mthca_eq_int(dev, eq);
456 	arbel_set_eq_ci(dev, eq, eq->cons_index);
457 	arbel_eq_req_not(dev, eq->eqn_mask);
458 
459 	/* MSI-X vectors always belong to us */
460 	return IRQ_HANDLED;
461 }
462 
463 static int __devinit mthca_create_eq(struct mthca_dev *dev,
464 				     int nent,
465 				     u8 intr,
466 				     struct mthca_eq *eq)
467 {
468 	int npages = (nent * MTHCA_EQ_ENTRY_SIZE + PAGE_SIZE - 1) /
469 		PAGE_SIZE;
470 	u64 *dma_list = NULL;
471 	dma_addr_t t;
472 	void *mailbox = NULL;
473 	struct mthca_eq_context *eq_context;
474 	int err = -ENOMEM;
475 	int i;
476 	u8 status;
477 
478 	/* Make sure EQ size is aligned to a power of 2 size. */
479 	for (i = 1; i < nent; i <<= 1)
480 		; /* nothing */
481 	nent = i;
482 
483 	eq->dev = dev;
484 
485 	eq->page_list = kmalloc(npages * sizeof *eq->page_list,
486 				GFP_KERNEL);
487 	if (!eq->page_list)
488 		goto err_out;
489 
490 	for (i = 0; i < npages; ++i)
491 		eq->page_list[i].buf = NULL;
492 
493 	dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
494 	if (!dma_list)
495 		goto err_out_free;
496 
497 	mailbox = kmalloc(sizeof *eq_context + MTHCA_CMD_MAILBOX_EXTRA,
498 			  GFP_KERNEL);
499 	if (!mailbox)
500 		goto err_out_free;
501 	eq_context = MAILBOX_ALIGN(mailbox);
502 
503 	for (i = 0; i < npages; ++i) {
504 		eq->page_list[i].buf = pci_alloc_consistent(dev->pdev,
505 							    PAGE_SIZE, &t);
506 		if (!eq->page_list[i].buf)
507 			goto err_out_free;
508 
509 		dma_list[i] = t;
510 		pci_unmap_addr_set(&eq->page_list[i], mapping, t);
511 
512 		memset(eq->page_list[i].buf, 0, PAGE_SIZE);
513 	}
514 
515 	for (i = 0; i < nent; ++i)
516 		set_eqe_hw(get_eqe(eq, i));
517 
518 	eq->eqn = mthca_alloc(&dev->eq_table.alloc);
519 	if (eq->eqn == -1)
520 		goto err_out_free;
521 
522 	err = mthca_mr_alloc_phys(dev, dev->driver_pd.pd_num,
523 				  dma_list, PAGE_SHIFT, npages,
524 				  0, npages * PAGE_SIZE,
525 				  MTHCA_MPT_FLAG_LOCAL_WRITE |
526 				  MTHCA_MPT_FLAG_LOCAL_READ,
527 				  &eq->mr);
528 	if (err)
529 		goto err_out_free_eq;
530 
531 	eq->nent = nent;
532 
533 	memset(eq_context, 0, sizeof *eq_context);
534 	eq_context->flags           = cpu_to_be32(MTHCA_EQ_STATUS_OK   |
535 						  MTHCA_EQ_OWNER_HW    |
536 						  MTHCA_EQ_STATE_ARMED |
537 						  MTHCA_EQ_FLAG_TR);
538 	if (dev->hca_type == ARBEL_NATIVE)
539 		eq_context->flags  |= cpu_to_be32(MTHCA_EQ_STATE_ARBEL);
540 
541 	eq_context->logsize_usrpage = cpu_to_be32((ffs(nent) - 1) << 24);
542 	if (dev->hca_type == ARBEL_NATIVE) {
543 		eq_context->arbel_pd = cpu_to_be32(dev->driver_pd.pd_num);
544 	} else {
545 		eq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
546 		eq_context->tavor_pd         = cpu_to_be32(dev->driver_pd.pd_num);
547 	}
548 	eq_context->intr            = intr;
549 	eq_context->lkey            = cpu_to_be32(eq->mr.ibmr.lkey);
550 
551 	err = mthca_SW2HW_EQ(dev, eq_context, eq->eqn, &status);
552 	if (err) {
553 		mthca_warn(dev, "SW2HW_EQ failed (%d)\n", err);
554 		goto err_out_free_mr;
555 	}
556 	if (status) {
557 		mthca_warn(dev, "SW2HW_EQ returned status 0x%02x\n",
558 			   status);
559 		err = -EINVAL;
560 		goto err_out_free_mr;
561 	}
562 
563 	kfree(dma_list);
564 	kfree(mailbox);
565 
566 	eq->eqn_mask   = swab32(1 << eq->eqn);
567 	eq->cons_index = 0;
568 
569 	dev->eq_table.arm_mask |= eq->eqn_mask;
570 
571 	mthca_dbg(dev, "Allocated EQ %d with %d entries\n",
572 		  eq->eqn, nent);
573 
574 	return err;
575 
576  err_out_free_mr:
577 	mthca_free_mr(dev, &eq->mr);
578 
579  err_out_free_eq:
580 	mthca_free(&dev->eq_table.alloc, eq->eqn);
581 
582  err_out_free:
583 	for (i = 0; i < npages; ++i)
584 		if (eq->page_list[i].buf)
585 			pci_free_consistent(dev->pdev, PAGE_SIZE,
586 					    eq->page_list[i].buf,
587 					    pci_unmap_addr(&eq->page_list[i],
588 							   mapping));
589 
590 	kfree(eq->page_list);
591 	kfree(dma_list);
592 	kfree(mailbox);
593 
594  err_out:
595 	return err;
596 }
597 
598 static void mthca_free_eq(struct mthca_dev *dev,
599 			  struct mthca_eq *eq)
600 {
601 	void *mailbox = NULL;
602 	int err;
603 	u8 status;
604 	int npages = (eq->nent * MTHCA_EQ_ENTRY_SIZE + PAGE_SIZE - 1) /
605 		PAGE_SIZE;
606 	int i;
607 
608 	mailbox = kmalloc(sizeof (struct mthca_eq_context) + MTHCA_CMD_MAILBOX_EXTRA,
609 			  GFP_KERNEL);
610 	if (!mailbox)
611 		return;
612 
613 	err = mthca_HW2SW_EQ(dev, MAILBOX_ALIGN(mailbox),
614 			     eq->eqn, &status);
615 	if (err)
616 		mthca_warn(dev, "HW2SW_EQ failed (%d)\n", err);
617 	if (status)
618 		mthca_warn(dev, "HW2SW_EQ returned status 0x%02x\n",
619 			   status);
620 
621 	dev->eq_table.arm_mask &= ~eq->eqn_mask;
622 
623 	if (0) {
624 		mthca_dbg(dev, "Dumping EQ context %02x:\n", eq->eqn);
625 		for (i = 0; i < sizeof (struct mthca_eq_context) / 4; ++i) {
626 			if (i % 4 == 0)
627 				printk("[%02x] ", i * 4);
628 			printk(" %08x", be32_to_cpup(MAILBOX_ALIGN(mailbox) + i * 4));
629 			if ((i + 1) % 4 == 0)
630 				printk("\n");
631 		}
632 	}
633 
634 	mthca_free_mr(dev, &eq->mr);
635 	for (i = 0; i < npages; ++i)
636 		pci_free_consistent(dev->pdev, PAGE_SIZE,
637 				    eq->page_list[i].buf,
638 				    pci_unmap_addr(&eq->page_list[i], mapping));
639 
640 	kfree(eq->page_list);
641 	kfree(mailbox);
642 }
643 
644 static void mthca_free_irqs(struct mthca_dev *dev)
645 {
646 	int i;
647 
648 	if (dev->eq_table.have_irq)
649 		free_irq(dev->pdev->irq, dev);
650 	for (i = 0; i < MTHCA_NUM_EQ; ++i)
651 		if (dev->eq_table.eq[i].have_irq)
652 			free_irq(dev->eq_table.eq[i].msi_x_vector,
653 				 dev->eq_table.eq + i);
654 }
655 
656 static int __devinit mthca_map_reg(struct mthca_dev *dev,
657 				   unsigned long offset, unsigned long size,
658 				   void __iomem **map)
659 {
660 	unsigned long base = pci_resource_start(dev->pdev, 0);
661 
662 	if (!request_mem_region(base + offset, size, DRV_NAME))
663 		return -EBUSY;
664 
665 	*map = ioremap(base + offset, size);
666 	if (!*map) {
667 		release_mem_region(base + offset, size);
668 		return -ENOMEM;
669 	}
670 
671 	return 0;
672 }
673 
674 static void mthca_unmap_reg(struct mthca_dev *dev, unsigned long offset,
675 			    unsigned long size, void __iomem *map)
676 {
677 	unsigned long base = pci_resource_start(dev->pdev, 0);
678 
679 	release_mem_region(base + offset, size);
680 	iounmap(map);
681 }
682 
683 static int __devinit mthca_map_eq_regs(struct mthca_dev *dev)
684 {
685 	unsigned long mthca_base;
686 
687 	mthca_base = pci_resource_start(dev->pdev, 0);
688 
689 	if (dev->hca_type == ARBEL_NATIVE) {
690 		/*
691 		 * We assume that the EQ arm and EQ set CI registers
692 		 * fall within the first BAR.  We can't trust the
693 		 * values firmware gives us, since those addresses are
694 		 * valid on the HCA's side of the PCI bus but not
695 		 * necessarily the host side.
696 		 */
697 		if (mthca_map_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
698 				  dev->fw.arbel.clr_int_base, MTHCA_CLR_INT_SIZE,
699 				  &dev->clr_base)) {
700 			mthca_err(dev, "Couldn't map interrupt clear register, "
701 				  "aborting.\n");
702 			return -ENOMEM;
703 		}
704 
705 		/*
706 		 * Add 4 because we limit ourselves to EQs 0 ... 31,
707 		 * so we only need the low word of the register.
708 		 */
709 		if (mthca_map_reg(dev, ((pci_resource_len(dev->pdev, 0) - 1) &
710 					dev->fw.arbel.eq_arm_base) + 4, 4,
711 				  &dev->eq_regs.arbel.eq_arm)) {
712 			mthca_err(dev, "Couldn't map interrupt clear register, "
713 				  "aborting.\n");
714 			mthca_unmap_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
715 					dev->fw.arbel.clr_int_base, MTHCA_CLR_INT_SIZE,
716 					dev->clr_base);
717 			return -ENOMEM;
718 		}
719 
720 		if (mthca_map_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
721 				  dev->fw.arbel.eq_set_ci_base,
722 				  MTHCA_EQ_SET_CI_SIZE,
723 				  &dev->eq_regs.arbel.eq_set_ci_base)) {
724 			mthca_err(dev, "Couldn't map interrupt clear register, "
725 				  "aborting.\n");
726 			mthca_unmap_reg(dev, ((pci_resource_len(dev->pdev, 0) - 1) &
727 					      dev->fw.arbel.eq_arm_base) + 4, 4,
728 					dev->eq_regs.arbel.eq_arm);
729 			mthca_unmap_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
730 					dev->fw.arbel.clr_int_base, MTHCA_CLR_INT_SIZE,
731 					dev->clr_base);
732 			return -ENOMEM;
733 		}
734 	} else {
735 		if (mthca_map_reg(dev, MTHCA_CLR_INT_BASE, MTHCA_CLR_INT_SIZE,
736 				  &dev->clr_base)) {
737 			mthca_err(dev, "Couldn't map interrupt clear register, "
738 				  "aborting.\n");
739 			return -ENOMEM;
740 		}
741 
742 		if (mthca_map_reg(dev, MTHCA_ECR_BASE,
743 				  MTHCA_ECR_SIZE + MTHCA_ECR_CLR_SIZE,
744 				  &dev->eq_regs.tavor.ecr_base)) {
745 			mthca_err(dev, "Couldn't map ecr register, "
746 				  "aborting.\n");
747 			mthca_unmap_reg(dev, MTHCA_CLR_INT_BASE, MTHCA_CLR_INT_SIZE,
748 					dev->clr_base);
749 			return -ENOMEM;
750 		}
751 	}
752 
753 	return 0;
754 
755 }
756 
757 static void __devexit mthca_unmap_eq_regs(struct mthca_dev *dev)
758 {
759 	if (dev->hca_type == ARBEL_NATIVE) {
760 		mthca_unmap_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
761 				dev->fw.arbel.eq_set_ci_base,
762 				MTHCA_EQ_SET_CI_SIZE,
763 				dev->eq_regs.arbel.eq_set_ci_base);
764 		mthca_unmap_reg(dev, ((pci_resource_len(dev->pdev, 0) - 1) &
765 				      dev->fw.arbel.eq_arm_base) + 4, 4,
766 				dev->eq_regs.arbel.eq_arm);
767 		mthca_unmap_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
768 				dev->fw.arbel.clr_int_base, MTHCA_CLR_INT_SIZE,
769 				dev->clr_base);
770 	} else {
771 		mthca_unmap_reg(dev, MTHCA_ECR_BASE,
772 				MTHCA_ECR_SIZE + MTHCA_ECR_CLR_SIZE,
773 				dev->eq_regs.tavor.ecr_base);
774 		mthca_unmap_reg(dev, MTHCA_CLR_INT_BASE, MTHCA_CLR_INT_SIZE,
775 				dev->clr_base);
776 	}
777 }
778 
779 int __devinit mthca_map_eq_icm(struct mthca_dev *dev, u64 icm_virt)
780 {
781 	int ret;
782 	u8 status;
783 
784 	/*
785 	 * We assume that mapping one page is enough for the whole EQ
786 	 * context table.  This is fine with all current HCAs, because
787 	 * we only use 32 EQs and each EQ uses 32 bytes of context
788 	 * memory, or 1 KB total.
789 	 */
790 	dev->eq_table.icm_virt = icm_virt;
791 	dev->eq_table.icm_page = alloc_page(GFP_HIGHUSER);
792 	if (!dev->eq_table.icm_page)
793 		return -ENOMEM;
794 	dev->eq_table.icm_dma  = pci_map_page(dev->pdev, dev->eq_table.icm_page, 0,
795 					      PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
796 	if (pci_dma_mapping_error(dev->eq_table.icm_dma)) {
797 		__free_page(dev->eq_table.icm_page);
798 		return -ENOMEM;
799 	}
800 
801 	ret = mthca_MAP_ICM_page(dev, dev->eq_table.icm_dma, icm_virt, &status);
802 	if (!ret && status)
803 		ret = -EINVAL;
804 	if (ret) {
805 		pci_unmap_page(dev->pdev, dev->eq_table.icm_dma, PAGE_SIZE,
806 			       PCI_DMA_BIDIRECTIONAL);
807 		__free_page(dev->eq_table.icm_page);
808 	}
809 
810 	return ret;
811 }
812 
813 void __devexit mthca_unmap_eq_icm(struct mthca_dev *dev)
814 {
815 	u8 status;
816 
817 	mthca_UNMAP_ICM(dev, dev->eq_table.icm_virt, PAGE_SIZE / 4096, &status);
818 	pci_unmap_page(dev->pdev, dev->eq_table.icm_dma, PAGE_SIZE,
819 		       PCI_DMA_BIDIRECTIONAL);
820 	__free_page(dev->eq_table.icm_page);
821 }
822 
823 int __devinit mthca_init_eq_table(struct mthca_dev *dev)
824 {
825 	int err;
826 	u8 status;
827 	u8 intr;
828 	int i;
829 
830 	err = mthca_alloc_init(&dev->eq_table.alloc,
831 			       dev->limits.num_eqs,
832 			       dev->limits.num_eqs - 1,
833 			       dev->limits.reserved_eqs);
834 	if (err)
835 		return err;
836 
837 	err = mthca_map_eq_regs(dev);
838 	if (err)
839 		goto err_out_free;
840 
841 	if (dev->mthca_flags & MTHCA_FLAG_MSI ||
842 	    dev->mthca_flags & MTHCA_FLAG_MSI_X) {
843 		dev->eq_table.clr_mask = 0;
844 	} else {
845 		dev->eq_table.clr_mask =
846 			swab32(1 << (dev->eq_table.inta_pin & 31));
847 		dev->eq_table.clr_int  = dev->clr_base +
848 			(dev->eq_table.inta_pin < 31 ? 4 : 0);
849 	}
850 
851 	dev->eq_table.arm_mask = 0;
852 
853 	intr = (dev->mthca_flags & MTHCA_FLAG_MSI) ?
854 		128 : dev->eq_table.inta_pin;
855 
856 	err = mthca_create_eq(dev, dev->limits.num_cqs,
857 			      (dev->mthca_flags & MTHCA_FLAG_MSI_X) ? 128 : intr,
858 			      &dev->eq_table.eq[MTHCA_EQ_COMP]);
859 	if (err)
860 		goto err_out_unmap;
861 
862 	err = mthca_create_eq(dev, MTHCA_NUM_ASYNC_EQE,
863 			      (dev->mthca_flags & MTHCA_FLAG_MSI_X) ? 129 : intr,
864 			      &dev->eq_table.eq[MTHCA_EQ_ASYNC]);
865 	if (err)
866 		goto err_out_comp;
867 
868 	err = mthca_create_eq(dev, MTHCA_NUM_CMD_EQE,
869 			      (dev->mthca_flags & MTHCA_FLAG_MSI_X) ? 130 : intr,
870 			      &dev->eq_table.eq[MTHCA_EQ_CMD]);
871 	if (err)
872 		goto err_out_async;
873 
874 	if (dev->mthca_flags & MTHCA_FLAG_MSI_X) {
875 		static const char *eq_name[] = {
876 			[MTHCA_EQ_COMP]  = DRV_NAME " (comp)",
877 			[MTHCA_EQ_ASYNC] = DRV_NAME " (async)",
878 			[MTHCA_EQ_CMD]   = DRV_NAME " (cmd)"
879 		};
880 
881 		for (i = 0; i < MTHCA_NUM_EQ; ++i) {
882 			err = request_irq(dev->eq_table.eq[i].msi_x_vector,
883 					  dev->hca_type == ARBEL_NATIVE ?
884 					  mthca_arbel_msi_x_interrupt :
885 					  mthca_tavor_msi_x_interrupt,
886 					  0, eq_name[i], dev->eq_table.eq + i);
887 			if (err)
888 				goto err_out_cmd;
889 			dev->eq_table.eq[i].have_irq = 1;
890 		}
891 	} else {
892 		err = request_irq(dev->pdev->irq,
893 				  dev->hca_type == ARBEL_NATIVE ?
894 				  mthca_arbel_interrupt :
895 				  mthca_tavor_interrupt,
896 				  SA_SHIRQ, DRV_NAME, dev);
897 		if (err)
898 			goto err_out_cmd;
899 		dev->eq_table.have_irq = 1;
900 	}
901 
902 	err = mthca_MAP_EQ(dev, async_mask(dev),
903 			   0, dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn, &status);
904 	if (err)
905 		mthca_warn(dev, "MAP_EQ for async EQ %d failed (%d)\n",
906 			   dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn, err);
907 	if (status)
908 		mthca_warn(dev, "MAP_EQ for async EQ %d returned status 0x%02x\n",
909 			   dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn, status);
910 
911 	err = mthca_MAP_EQ(dev, MTHCA_CMD_EVENT_MASK,
912 			   0, dev->eq_table.eq[MTHCA_EQ_CMD].eqn, &status);
913 	if (err)
914 		mthca_warn(dev, "MAP_EQ for cmd EQ %d failed (%d)\n",
915 			   dev->eq_table.eq[MTHCA_EQ_CMD].eqn, err);
916 	if (status)
917 		mthca_warn(dev, "MAP_EQ for cmd EQ %d returned status 0x%02x\n",
918 			   dev->eq_table.eq[MTHCA_EQ_CMD].eqn, status);
919 
920 	for (i = 0; i < MTHCA_EQ_CMD; ++i)
921 		if (dev->hca_type == ARBEL_NATIVE)
922 			arbel_eq_req_not(dev, dev->eq_table.eq[i].eqn_mask);
923 		else
924 			tavor_eq_req_not(dev, dev->eq_table.eq[i].eqn);
925 
926 	return 0;
927 
928 err_out_cmd:
929 	mthca_free_irqs(dev);
930 	mthca_free_eq(dev, &dev->eq_table.eq[MTHCA_EQ_CMD]);
931 
932 err_out_async:
933 	mthca_free_eq(dev, &dev->eq_table.eq[MTHCA_EQ_ASYNC]);
934 
935 err_out_comp:
936 	mthca_free_eq(dev, &dev->eq_table.eq[MTHCA_EQ_COMP]);
937 
938 err_out_unmap:
939 	mthca_unmap_eq_regs(dev);
940 
941 err_out_free:
942 	mthca_alloc_cleanup(&dev->eq_table.alloc);
943 	return err;
944 }
945 
946 void __devexit mthca_cleanup_eq_table(struct mthca_dev *dev)
947 {
948 	u8 status;
949 	int i;
950 
951 	mthca_free_irqs(dev);
952 
953 	mthca_MAP_EQ(dev, async_mask(dev),
954 		     1, dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn, &status);
955 	mthca_MAP_EQ(dev, MTHCA_CMD_EVENT_MASK,
956 		     1, dev->eq_table.eq[MTHCA_EQ_CMD].eqn, &status);
957 
958 	for (i = 0; i < MTHCA_NUM_EQ; ++i)
959 		mthca_free_eq(dev, &dev->eq_table.eq[i]);
960 
961 	mthca_unmap_eq_regs(dev);
962 
963 	mthca_alloc_cleanup(&dev->eq_table.alloc);
964 }
965