xref: /linux/drivers/net/ethernet/mellanox/mlx4/eq.c (revision b43ab901d671e3e3cad425ea5e9a3c74e266dcdd)
1 /*
2  * Copyright (c) 2005, 2006, 2007, 2008 Mellanox Technologies. All rights reserved.
3  * Copyright (c) 2005, 2006, 2007 Cisco Systems, 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/init.h>
35 #include <linux/interrupt.h>
36 #include <linux/slab.h>
37 #include <linux/export.h>
38 #include <linux/mm.h>
39 #include <linux/dma-mapping.h>
40 
41 #include <linux/mlx4/cmd.h>
42 
43 #include "mlx4.h"
44 #include "fw.h"
45 
46 enum {
47 	MLX4_IRQNAME_SIZE	= 32
48 };
49 
50 enum {
51 	MLX4_NUM_ASYNC_EQE	= 0x100,
52 	MLX4_NUM_SPARE_EQE	= 0x80,
53 	MLX4_EQ_ENTRY_SIZE	= 0x20
54 };
55 
56 #define MLX4_EQ_STATUS_OK	   ( 0 << 28)
57 #define MLX4_EQ_STATUS_WRITE_FAIL  (10 << 28)
58 #define MLX4_EQ_OWNER_SW	   ( 0 << 24)
59 #define MLX4_EQ_OWNER_HW	   ( 1 << 24)
60 #define MLX4_EQ_FLAG_EC		   ( 1 << 18)
61 #define MLX4_EQ_FLAG_OI		   ( 1 << 17)
62 #define MLX4_EQ_STATE_ARMED	   ( 9 <<  8)
63 #define MLX4_EQ_STATE_FIRED	   (10 <<  8)
64 #define MLX4_EQ_STATE_ALWAYS_ARMED (11 <<  8)
65 
66 #define MLX4_ASYNC_EVENT_MASK ((1ull << MLX4_EVENT_TYPE_PATH_MIG)	    | \
67 			       (1ull << MLX4_EVENT_TYPE_COMM_EST)	    | \
68 			       (1ull << MLX4_EVENT_TYPE_SQ_DRAINED)	    | \
69 			       (1ull << MLX4_EVENT_TYPE_CQ_ERROR)	    | \
70 			       (1ull << MLX4_EVENT_TYPE_WQ_CATAS_ERROR)	    | \
71 			       (1ull << MLX4_EVENT_TYPE_EEC_CATAS_ERROR)    | \
72 			       (1ull << MLX4_EVENT_TYPE_PATH_MIG_FAILED)    | \
73 			       (1ull << MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR) | \
74 			       (1ull << MLX4_EVENT_TYPE_WQ_ACCESS_ERROR)    | \
75 			       (1ull << MLX4_EVENT_TYPE_PORT_CHANGE)	    | \
76 			       (1ull << MLX4_EVENT_TYPE_ECC_DETECT)	    | \
77 			       (1ull << MLX4_EVENT_TYPE_SRQ_CATAS_ERROR)    | \
78 			       (1ull << MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE)    | \
79 			       (1ull << MLX4_EVENT_TYPE_SRQ_LIMIT)	    | \
80 			       (1ull << MLX4_EVENT_TYPE_CMD)		    | \
81 			       (1ull << MLX4_EVENT_TYPE_COMM_CHANNEL)       | \
82 			       (1ull << MLX4_EVENT_TYPE_FLR_EVENT))
83 
84 static void eq_set_ci(struct mlx4_eq *eq, int req_not)
85 {
86 	__raw_writel((__force u32) cpu_to_be32((eq->cons_index & 0xffffff) |
87 					       req_not << 31),
88 		     eq->doorbell);
89 	/* We still want ordering, just not swabbing, so add a barrier */
90 	mb();
91 }
92 
93 static struct mlx4_eqe *get_eqe(struct mlx4_eq *eq, u32 entry)
94 {
95 	unsigned long off = (entry & (eq->nent - 1)) * MLX4_EQ_ENTRY_SIZE;
96 	return eq->page_list[off / PAGE_SIZE].buf + off % PAGE_SIZE;
97 }
98 
99 static struct mlx4_eqe *next_eqe_sw(struct mlx4_eq *eq)
100 {
101 	struct mlx4_eqe *eqe = get_eqe(eq, eq->cons_index);
102 	return !!(eqe->owner & 0x80) ^ !!(eq->cons_index & eq->nent) ? NULL : eqe;
103 }
104 
105 static struct mlx4_eqe *next_slave_event_eqe(struct mlx4_slave_event_eq *slave_eq)
106 {
107 	struct mlx4_eqe *eqe =
108 		&slave_eq->event_eqe[slave_eq->cons & (SLAVE_EVENT_EQ_SIZE - 1)];
109 	return (!!(eqe->owner & 0x80) ^
110 		!!(slave_eq->cons & SLAVE_EVENT_EQ_SIZE)) ?
111 		eqe : NULL;
112 }
113 
114 void mlx4_gen_slave_eqe(struct work_struct *work)
115 {
116 	struct mlx4_mfunc_master_ctx *master =
117 		container_of(work, struct mlx4_mfunc_master_ctx,
118 			     slave_event_work);
119 	struct mlx4_mfunc *mfunc =
120 		container_of(master, struct mlx4_mfunc, master);
121 	struct mlx4_priv *priv = container_of(mfunc, struct mlx4_priv, mfunc);
122 	struct mlx4_dev *dev = &priv->dev;
123 	struct mlx4_slave_event_eq *slave_eq = &mfunc->master.slave_eq;
124 	struct mlx4_eqe *eqe;
125 	u8 slave;
126 	int i;
127 
128 	for (eqe = next_slave_event_eqe(slave_eq); eqe;
129 	      eqe = next_slave_event_eqe(slave_eq)) {
130 		slave = eqe->slave_id;
131 
132 		/* All active slaves need to receive the event */
133 		if (slave == ALL_SLAVES) {
134 			for (i = 0; i < dev->num_slaves; i++) {
135 				if (i != dev->caps.function &&
136 				    master->slave_state[i].active)
137 					if (mlx4_GEN_EQE(dev, i, eqe))
138 						mlx4_warn(dev, "Failed to "
139 							  " generate event "
140 							  "for slave %d\n", i);
141 			}
142 		} else {
143 			if (mlx4_GEN_EQE(dev, slave, eqe))
144 				mlx4_warn(dev, "Failed to generate event "
145 					       "for slave %d\n", slave);
146 		}
147 		++slave_eq->cons;
148 	}
149 }
150 
151 
152 static void slave_event(struct mlx4_dev *dev, u8 slave, struct mlx4_eqe *eqe)
153 {
154 	struct mlx4_priv *priv = mlx4_priv(dev);
155 	struct mlx4_slave_event_eq *slave_eq = &priv->mfunc.master.slave_eq;
156 	struct mlx4_eqe *s_eqe =
157 		&slave_eq->event_eqe[slave_eq->prod & (SLAVE_EVENT_EQ_SIZE - 1)];
158 
159 	if ((!!(s_eqe->owner & 0x80)) ^
160 	    (!!(slave_eq->prod & SLAVE_EVENT_EQ_SIZE))) {
161 		mlx4_warn(dev, "Master failed to generate an EQE for slave: %d. "
162 			  "No free EQE on slave events queue\n", slave);
163 		return;
164 	}
165 
166 	memcpy(s_eqe, eqe, sizeof(struct mlx4_eqe) - 1);
167 	s_eqe->slave_id = slave;
168 	/* ensure all information is written before setting the ownersip bit */
169 	wmb();
170 	s_eqe->owner = !!(slave_eq->prod & SLAVE_EVENT_EQ_SIZE) ? 0x0 : 0x80;
171 	++slave_eq->prod;
172 
173 	queue_work(priv->mfunc.master.comm_wq,
174 		   &priv->mfunc.master.slave_event_work);
175 }
176 
177 static void mlx4_slave_event(struct mlx4_dev *dev, int slave,
178 			     struct mlx4_eqe *eqe)
179 {
180 	struct mlx4_priv *priv = mlx4_priv(dev);
181 	struct mlx4_slave_state *s_slave =
182 		&priv->mfunc.master.slave_state[slave];
183 
184 	if (!s_slave->active) {
185 		/*mlx4_warn(dev, "Trying to pass event to inactive slave\n");*/
186 		return;
187 	}
188 
189 	slave_event(dev, slave, eqe);
190 }
191 
192 void mlx4_master_handle_slave_flr(struct work_struct *work)
193 {
194 	struct mlx4_mfunc_master_ctx *master =
195 		container_of(work, struct mlx4_mfunc_master_ctx,
196 			     slave_flr_event_work);
197 	struct mlx4_mfunc *mfunc =
198 		container_of(master, struct mlx4_mfunc, master);
199 	struct mlx4_priv *priv =
200 		container_of(mfunc, struct mlx4_priv, mfunc);
201 	struct mlx4_dev *dev = &priv->dev;
202 	struct mlx4_slave_state *slave_state = priv->mfunc.master.slave_state;
203 	int i;
204 	int err;
205 
206 	mlx4_dbg(dev, "mlx4_handle_slave_flr\n");
207 
208 	for (i = 0 ; i < dev->num_slaves; i++) {
209 
210 		if (MLX4_COMM_CMD_FLR == slave_state[i].last_cmd) {
211 			mlx4_dbg(dev, "mlx4_handle_slave_flr: "
212 				 "clean slave: %d\n", i);
213 
214 			mlx4_delete_all_resources_for_slave(dev, i);
215 			/*return the slave to running mode*/
216 			spin_lock(&priv->mfunc.master.slave_state_lock);
217 			slave_state[i].last_cmd = MLX4_COMM_CMD_RESET;
218 			slave_state[i].is_slave_going_down = 0;
219 			spin_unlock(&priv->mfunc.master.slave_state_lock);
220 			/*notify the FW:*/
221 			err = mlx4_cmd(dev, 0, i, 0, MLX4_CMD_INFORM_FLR_DONE,
222 				       MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED);
223 			if (err)
224 				mlx4_warn(dev, "Failed to notify FW on "
225 					  "FLR done (slave:%d)\n", i);
226 		}
227 	}
228 }
229 
230 static int mlx4_eq_int(struct mlx4_dev *dev, struct mlx4_eq *eq)
231 {
232 	struct mlx4_priv *priv = mlx4_priv(dev);
233 	struct mlx4_eqe *eqe;
234 	int cqn;
235 	int eqes_found = 0;
236 	int set_ci = 0;
237 	int port;
238 	int slave = 0;
239 	int ret;
240 	u32 flr_slave;
241 	u8 update_slave_state;
242 	int i;
243 
244 	while ((eqe = next_eqe_sw(eq))) {
245 		/*
246 		 * Make sure we read EQ entry contents after we've
247 		 * checked the ownership bit.
248 		 */
249 		rmb();
250 
251 		switch (eqe->type) {
252 		case MLX4_EVENT_TYPE_COMP:
253 			cqn = be32_to_cpu(eqe->event.comp.cqn) & 0xffffff;
254 			mlx4_cq_completion(dev, cqn);
255 			break;
256 
257 		case MLX4_EVENT_TYPE_PATH_MIG:
258 		case MLX4_EVENT_TYPE_COMM_EST:
259 		case MLX4_EVENT_TYPE_SQ_DRAINED:
260 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
261 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
262 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
263 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
264 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
265 			mlx4_dbg(dev, "event %d arrived\n", eqe->type);
266 			if (mlx4_is_master(dev)) {
267 				/* forward only to slave owning the QP */
268 				ret = mlx4_get_slave_from_resource_id(dev,
269 						RES_QP,
270 						be32_to_cpu(eqe->event.qp.qpn)
271 						& 0xffffff, &slave);
272 				if (ret && ret != -ENOENT) {
273 					mlx4_dbg(dev, "QP event %02x(%02x) on "
274 						 "EQ %d at index %u: could "
275 						 "not get slave id (%d)\n",
276 						 eqe->type, eqe->subtype,
277 						 eq->eqn, eq->cons_index, ret);
278 					break;
279 				}
280 
281 				if (!ret && slave != dev->caps.function) {
282 					mlx4_slave_event(dev, slave, eqe);
283 					break;
284 				}
285 
286 			}
287 			mlx4_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) &
288 				      0xffffff, eqe->type);
289 			break;
290 
291 		case MLX4_EVENT_TYPE_SRQ_LIMIT:
292 			mlx4_warn(dev, "%s: MLX4_EVENT_TYPE_SRQ_LIMIT\n",
293 				  __func__);
294 		case MLX4_EVENT_TYPE_SRQ_CATAS_ERROR:
295 			if (mlx4_is_master(dev)) {
296 				/* forward only to slave owning the SRQ */
297 				ret = mlx4_get_slave_from_resource_id(dev,
298 						RES_SRQ,
299 						be32_to_cpu(eqe->event.srq.srqn)
300 						& 0xffffff,
301 						&slave);
302 				if (ret && ret != -ENOENT) {
303 					mlx4_warn(dev, "SRQ event %02x(%02x) "
304 						  "on EQ %d at index %u: could"
305 						  " not get slave id (%d)\n",
306 						  eqe->type, eqe->subtype,
307 						  eq->eqn, eq->cons_index, ret);
308 					break;
309 				}
310 				mlx4_warn(dev, "%s: slave:%d, srq_no:0x%x,"
311 					  " event: %02x(%02x)\n", __func__,
312 					  slave,
313 					  be32_to_cpu(eqe->event.srq.srqn),
314 					  eqe->type, eqe->subtype);
315 
316 				if (!ret && slave != dev->caps.function) {
317 					mlx4_warn(dev, "%s: sending event "
318 						  "%02x(%02x) to slave:%d\n",
319 						   __func__, eqe->type,
320 						  eqe->subtype, slave);
321 					mlx4_slave_event(dev, slave, eqe);
322 					break;
323 				}
324 			}
325 			mlx4_srq_event(dev, be32_to_cpu(eqe->event.srq.srqn) &
326 				       0xffffff, eqe->type);
327 			break;
328 
329 		case MLX4_EVENT_TYPE_CMD:
330 			mlx4_cmd_event(dev,
331 				       be16_to_cpu(eqe->event.cmd.token),
332 				       eqe->event.cmd.status,
333 				       be64_to_cpu(eqe->event.cmd.out_param));
334 			break;
335 
336 		case MLX4_EVENT_TYPE_PORT_CHANGE:
337 			port = be32_to_cpu(eqe->event.port_change.port) >> 28;
338 			if (eqe->subtype == MLX4_PORT_CHANGE_SUBTYPE_DOWN) {
339 				mlx4_dispatch_event(dev,
340 						    MLX4_DEV_EVENT_PORT_DOWN,
341 						    port);
342 				mlx4_priv(dev)->sense.do_sense_port[port] = 1;
343 				if (mlx4_is_master(dev))
344 					/*change the state of all slave's port
345 					* to down:*/
346 					for (i = 0; i < dev->num_slaves; i++) {
347 						mlx4_dbg(dev, "%s: Sending "
348 							 "MLX4_PORT_CHANGE_SUBTYPE_DOWN"
349 							 " to slave: %d, port:%d\n",
350 							 __func__, i, port);
351 						if (i == dev->caps.function)
352 							continue;
353 						mlx4_slave_event(dev, i, eqe);
354 					}
355 			} else {
356 				mlx4_dispatch_event(dev,
357 						    MLX4_DEV_EVENT_PORT_UP,
358 						    port);
359 				mlx4_priv(dev)->sense.do_sense_port[port] = 0;
360 
361 				if (mlx4_is_master(dev)) {
362 					for (i = 0; i < dev->num_slaves; i++) {
363 						if (i == dev->caps.function)
364 							continue;
365 						mlx4_slave_event(dev, i, eqe);
366 					}
367 				}
368 			}
369 			break;
370 
371 		case MLX4_EVENT_TYPE_CQ_ERROR:
372 			mlx4_warn(dev, "CQ %s on CQN %06x\n",
373 				  eqe->event.cq_err.syndrome == 1 ?
374 				  "overrun" : "access violation",
375 				  be32_to_cpu(eqe->event.cq_err.cqn) & 0xffffff);
376 			if (mlx4_is_master(dev)) {
377 				ret = mlx4_get_slave_from_resource_id(dev,
378 					RES_CQ,
379 					be32_to_cpu(eqe->event.cq_err.cqn)
380 					& 0xffffff, &slave);
381 				if (ret && ret != -ENOENT) {
382 					mlx4_dbg(dev, "CQ event %02x(%02x) on "
383 						 "EQ %d at index %u: could "
384 						  "not get slave id (%d)\n",
385 						  eqe->type, eqe->subtype,
386 						  eq->eqn, eq->cons_index, ret);
387 					break;
388 				}
389 
390 				if (!ret && slave != dev->caps.function) {
391 					mlx4_slave_event(dev, slave, eqe);
392 					break;
393 				}
394 			}
395 			mlx4_cq_event(dev,
396 				      be32_to_cpu(eqe->event.cq_err.cqn)
397 				      & 0xffffff,
398 				      eqe->type);
399 			break;
400 
401 		case MLX4_EVENT_TYPE_EQ_OVERFLOW:
402 			mlx4_warn(dev, "EQ overrun on EQN %d\n", eq->eqn);
403 			break;
404 
405 		case MLX4_EVENT_TYPE_COMM_CHANNEL:
406 			if (!mlx4_is_master(dev)) {
407 				mlx4_warn(dev, "Received comm channel event "
408 					       "for non master device\n");
409 				break;
410 			}
411 			memcpy(&priv->mfunc.master.comm_arm_bit_vector,
412 			       eqe->event.comm_channel_arm.bit_vec,
413 			       sizeof eqe->event.comm_channel_arm.bit_vec);
414 			queue_work(priv->mfunc.master.comm_wq,
415 				   &priv->mfunc.master.comm_work);
416 			break;
417 
418 		case MLX4_EVENT_TYPE_FLR_EVENT:
419 			flr_slave = be32_to_cpu(eqe->event.flr_event.slave_id);
420 			if (!mlx4_is_master(dev)) {
421 				mlx4_warn(dev, "Non-master function received"
422 					       "FLR event\n");
423 				break;
424 			}
425 
426 			mlx4_dbg(dev, "FLR event for slave: %d\n", flr_slave);
427 
428 			if (flr_slave > dev->num_slaves) {
429 				mlx4_warn(dev,
430 					  "Got FLR for unknown function: %d\n",
431 					  flr_slave);
432 				update_slave_state = 0;
433 			} else
434 				update_slave_state = 1;
435 
436 			spin_lock(&priv->mfunc.master.slave_state_lock);
437 			if (update_slave_state) {
438 				priv->mfunc.master.slave_state[flr_slave].active = false;
439 				priv->mfunc.master.slave_state[flr_slave].last_cmd = MLX4_COMM_CMD_FLR;
440 				priv->mfunc.master.slave_state[flr_slave].is_slave_going_down = 1;
441 			}
442 			spin_unlock(&priv->mfunc.master.slave_state_lock);
443 			queue_work(priv->mfunc.master.comm_wq,
444 				   &priv->mfunc.master.slave_flr_event_work);
445 			break;
446 		case MLX4_EVENT_TYPE_EEC_CATAS_ERROR:
447 		case MLX4_EVENT_TYPE_ECC_DETECT:
448 		default:
449 			mlx4_warn(dev, "Unhandled event %02x(%02x) on EQ %d at "
450 				  "index %u. owner=%x, nent=0x%x, slave=%x, "
451 				  "ownership=%s\n",
452 				  eqe->type, eqe->subtype, eq->eqn,
453 				  eq->cons_index, eqe->owner, eq->nent,
454 				  eqe->slave_id,
455 				  !!(eqe->owner & 0x80) ^
456 				  !!(eq->cons_index & eq->nent) ? "HW" : "SW");
457 			break;
458 		};
459 
460 		++eq->cons_index;
461 		eqes_found = 1;
462 		++set_ci;
463 
464 		/*
465 		 * The HCA will think the queue has overflowed if we
466 		 * don't tell it we've been processing events.  We
467 		 * create our EQs with MLX4_NUM_SPARE_EQE extra
468 		 * entries, so we must update our consumer index at
469 		 * least that often.
470 		 */
471 		if (unlikely(set_ci >= MLX4_NUM_SPARE_EQE)) {
472 			eq_set_ci(eq, 0);
473 			set_ci = 0;
474 		}
475 	}
476 
477 	eq_set_ci(eq, 1);
478 
479 	return eqes_found;
480 }
481 
482 static irqreturn_t mlx4_interrupt(int irq, void *dev_ptr)
483 {
484 	struct mlx4_dev *dev = dev_ptr;
485 	struct mlx4_priv *priv = mlx4_priv(dev);
486 	int work = 0;
487 	int i;
488 
489 	writel(priv->eq_table.clr_mask, priv->eq_table.clr_int);
490 
491 	for (i = 0; i < dev->caps.num_comp_vectors + 1; ++i)
492 		work |= mlx4_eq_int(dev, &priv->eq_table.eq[i]);
493 
494 	return IRQ_RETVAL(work);
495 }
496 
497 static irqreturn_t mlx4_msi_x_interrupt(int irq, void *eq_ptr)
498 {
499 	struct mlx4_eq  *eq  = eq_ptr;
500 	struct mlx4_dev *dev = eq->dev;
501 
502 	mlx4_eq_int(dev, eq);
503 
504 	/* MSI-X vectors always belong to us */
505 	return IRQ_HANDLED;
506 }
507 
508 int mlx4_MAP_EQ_wrapper(struct mlx4_dev *dev, int slave,
509 			struct mlx4_vhcr *vhcr,
510 			struct mlx4_cmd_mailbox *inbox,
511 			struct mlx4_cmd_mailbox *outbox,
512 			struct mlx4_cmd_info *cmd)
513 {
514 	struct mlx4_priv *priv = mlx4_priv(dev);
515 	struct mlx4_slave_event_eq_info *event_eq =
516 		&priv->mfunc.master.slave_state[slave].event_eq;
517 	u32 in_modifier = vhcr->in_modifier;
518 	u32 eqn = in_modifier & 0x1FF;
519 	u64 in_param =  vhcr->in_param;
520 	int err = 0;
521 
522 	if (slave == dev->caps.function)
523 		err = mlx4_cmd(dev, in_param, (in_modifier & 0x80000000) | eqn,
524 			       0, MLX4_CMD_MAP_EQ, MLX4_CMD_TIME_CLASS_B,
525 			       MLX4_CMD_NATIVE);
526 	if (!err) {
527 		if (in_modifier >> 31) {
528 			/* unmap */
529 			event_eq->event_type &= ~in_param;
530 		} else {
531 			event_eq->eqn = eqn;
532 			event_eq->event_type = in_param;
533 		}
534 	}
535 	return err;
536 }
537 
538 static int mlx4_MAP_EQ(struct mlx4_dev *dev, u64 event_mask, int unmap,
539 			int eq_num)
540 {
541 	return mlx4_cmd(dev, event_mask, (unmap << 31) | eq_num,
542 			0, MLX4_CMD_MAP_EQ, MLX4_CMD_TIME_CLASS_B,
543 			MLX4_CMD_WRAPPED);
544 }
545 
546 static int mlx4_SW2HW_EQ(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
547 			 int eq_num)
548 {
549 	return mlx4_cmd(dev, mailbox->dma | dev->caps.function, eq_num, 0,
550 			MLX4_CMD_SW2HW_EQ, MLX4_CMD_TIME_CLASS_A,
551 			MLX4_CMD_WRAPPED);
552 }
553 
554 static int mlx4_HW2SW_EQ(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
555 			 int eq_num)
556 {
557 	return mlx4_cmd_box(dev, dev->caps.function, mailbox->dma, eq_num,
558 			    0, MLX4_CMD_HW2SW_EQ, MLX4_CMD_TIME_CLASS_A,
559 			    MLX4_CMD_WRAPPED);
560 }
561 
562 static int mlx4_num_eq_uar(struct mlx4_dev *dev)
563 {
564 	/*
565 	 * Each UAR holds 4 EQ doorbells.  To figure out how many UARs
566 	 * we need to map, take the difference of highest index and
567 	 * the lowest index we'll use and add 1.
568 	 */
569 	return (dev->caps.num_comp_vectors + 1 + dev->caps.reserved_eqs +
570 		 dev->caps.comp_pool)/4 - dev->caps.reserved_eqs/4 + 1;
571 }
572 
573 static void __iomem *mlx4_get_eq_uar(struct mlx4_dev *dev, struct mlx4_eq *eq)
574 {
575 	struct mlx4_priv *priv = mlx4_priv(dev);
576 	int index;
577 
578 	index = eq->eqn / 4 - dev->caps.reserved_eqs / 4;
579 
580 	if (!priv->eq_table.uar_map[index]) {
581 		priv->eq_table.uar_map[index] =
582 			ioremap(pci_resource_start(dev->pdev, 2) +
583 				((eq->eqn / 4) << PAGE_SHIFT),
584 				PAGE_SIZE);
585 		if (!priv->eq_table.uar_map[index]) {
586 			mlx4_err(dev, "Couldn't map EQ doorbell for EQN 0x%06x\n",
587 				 eq->eqn);
588 			return NULL;
589 		}
590 	}
591 
592 	return priv->eq_table.uar_map[index] + 0x800 + 8 * (eq->eqn % 4);
593 }
594 
595 static int mlx4_create_eq(struct mlx4_dev *dev, int nent,
596 			  u8 intr, struct mlx4_eq *eq)
597 {
598 	struct mlx4_priv *priv = mlx4_priv(dev);
599 	struct mlx4_cmd_mailbox *mailbox;
600 	struct mlx4_eq_context *eq_context;
601 	int npages;
602 	u64 *dma_list = NULL;
603 	dma_addr_t t;
604 	u64 mtt_addr;
605 	int err = -ENOMEM;
606 	int i;
607 
608 	eq->dev   = dev;
609 	eq->nent  = roundup_pow_of_two(max(nent, 2));
610 	npages = PAGE_ALIGN(eq->nent * MLX4_EQ_ENTRY_SIZE) / PAGE_SIZE;
611 
612 	eq->page_list = kmalloc(npages * sizeof *eq->page_list,
613 				GFP_KERNEL);
614 	if (!eq->page_list)
615 		goto err_out;
616 
617 	for (i = 0; i < npages; ++i)
618 		eq->page_list[i].buf = NULL;
619 
620 	dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
621 	if (!dma_list)
622 		goto err_out_free;
623 
624 	mailbox = mlx4_alloc_cmd_mailbox(dev);
625 	if (IS_ERR(mailbox))
626 		goto err_out_free;
627 	eq_context = mailbox->buf;
628 
629 	for (i = 0; i < npages; ++i) {
630 		eq->page_list[i].buf = dma_alloc_coherent(&dev->pdev->dev,
631 							  PAGE_SIZE, &t, GFP_KERNEL);
632 		if (!eq->page_list[i].buf)
633 			goto err_out_free_pages;
634 
635 		dma_list[i] = t;
636 		eq->page_list[i].map = t;
637 
638 		memset(eq->page_list[i].buf, 0, PAGE_SIZE);
639 	}
640 
641 	eq->eqn = mlx4_bitmap_alloc(&priv->eq_table.bitmap);
642 	if (eq->eqn == -1)
643 		goto err_out_free_pages;
644 
645 	eq->doorbell = mlx4_get_eq_uar(dev, eq);
646 	if (!eq->doorbell) {
647 		err = -ENOMEM;
648 		goto err_out_free_eq;
649 	}
650 
651 	err = mlx4_mtt_init(dev, npages, PAGE_SHIFT, &eq->mtt);
652 	if (err)
653 		goto err_out_free_eq;
654 
655 	err = mlx4_write_mtt(dev, &eq->mtt, 0, npages, dma_list);
656 	if (err)
657 		goto err_out_free_mtt;
658 
659 	memset(eq_context, 0, sizeof *eq_context);
660 	eq_context->flags	  = cpu_to_be32(MLX4_EQ_STATUS_OK   |
661 						MLX4_EQ_STATE_ARMED);
662 	eq_context->log_eq_size	  = ilog2(eq->nent);
663 	eq_context->intr	  = intr;
664 	eq_context->log_page_size = PAGE_SHIFT - MLX4_ICM_PAGE_SHIFT;
665 
666 	mtt_addr = mlx4_mtt_addr(dev, &eq->mtt);
667 	eq_context->mtt_base_addr_h = mtt_addr >> 32;
668 	eq_context->mtt_base_addr_l = cpu_to_be32(mtt_addr & 0xffffffff);
669 
670 	err = mlx4_SW2HW_EQ(dev, mailbox, eq->eqn);
671 	if (err) {
672 		mlx4_warn(dev, "SW2HW_EQ failed (%d)\n", err);
673 		goto err_out_free_mtt;
674 	}
675 
676 	kfree(dma_list);
677 	mlx4_free_cmd_mailbox(dev, mailbox);
678 
679 	eq->cons_index = 0;
680 
681 	return err;
682 
683 err_out_free_mtt:
684 	mlx4_mtt_cleanup(dev, &eq->mtt);
685 
686 err_out_free_eq:
687 	mlx4_bitmap_free(&priv->eq_table.bitmap, eq->eqn);
688 
689 err_out_free_pages:
690 	for (i = 0; i < npages; ++i)
691 		if (eq->page_list[i].buf)
692 			dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
693 					  eq->page_list[i].buf,
694 					  eq->page_list[i].map);
695 
696 	mlx4_free_cmd_mailbox(dev, mailbox);
697 
698 err_out_free:
699 	kfree(eq->page_list);
700 	kfree(dma_list);
701 
702 err_out:
703 	return err;
704 }
705 
706 static void mlx4_free_eq(struct mlx4_dev *dev,
707 			 struct mlx4_eq *eq)
708 {
709 	struct mlx4_priv *priv = mlx4_priv(dev);
710 	struct mlx4_cmd_mailbox *mailbox;
711 	int err;
712 	int npages = PAGE_ALIGN(MLX4_EQ_ENTRY_SIZE * eq->nent) / PAGE_SIZE;
713 	int i;
714 
715 	mailbox = mlx4_alloc_cmd_mailbox(dev);
716 	if (IS_ERR(mailbox))
717 		return;
718 
719 	err = mlx4_HW2SW_EQ(dev, mailbox, eq->eqn);
720 	if (err)
721 		mlx4_warn(dev, "HW2SW_EQ failed (%d)\n", err);
722 
723 	if (0) {
724 		mlx4_dbg(dev, "Dumping EQ context %02x:\n", eq->eqn);
725 		for (i = 0; i < sizeof (struct mlx4_eq_context) / 4; ++i) {
726 			if (i % 4 == 0)
727 				pr_cont("[%02x] ", i * 4);
728 			pr_cont(" %08x", be32_to_cpup(mailbox->buf + i * 4));
729 			if ((i + 1) % 4 == 0)
730 				pr_cont("\n");
731 		}
732 	}
733 
734 	mlx4_mtt_cleanup(dev, &eq->mtt);
735 	for (i = 0; i < npages; ++i)
736 		dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
737 				    eq->page_list[i].buf,
738 				    eq->page_list[i].map);
739 
740 	kfree(eq->page_list);
741 	mlx4_bitmap_free(&priv->eq_table.bitmap, eq->eqn);
742 	mlx4_free_cmd_mailbox(dev, mailbox);
743 }
744 
745 static void mlx4_free_irqs(struct mlx4_dev *dev)
746 {
747 	struct mlx4_eq_table *eq_table = &mlx4_priv(dev)->eq_table;
748 	struct mlx4_priv *priv = mlx4_priv(dev);
749 	int	i, vec;
750 
751 	if (eq_table->have_irq)
752 		free_irq(dev->pdev->irq, dev);
753 
754 	for (i = 0; i < dev->caps.num_comp_vectors + 1; ++i)
755 		if (eq_table->eq[i].have_irq) {
756 			free_irq(eq_table->eq[i].irq, eq_table->eq + i);
757 			eq_table->eq[i].have_irq = 0;
758 		}
759 
760 	for (i = 0; i < dev->caps.comp_pool; i++) {
761 		/*
762 		 * Freeing the assigned irq's
763 		 * all bits should be 0, but we need to validate
764 		 */
765 		if (priv->msix_ctl.pool_bm & 1ULL << i) {
766 			/* NO need protecting*/
767 			vec = dev->caps.num_comp_vectors + 1 + i;
768 			free_irq(priv->eq_table.eq[vec].irq,
769 				 &priv->eq_table.eq[vec]);
770 		}
771 	}
772 
773 
774 	kfree(eq_table->irq_names);
775 }
776 
777 static int mlx4_map_clr_int(struct mlx4_dev *dev)
778 {
779 	struct mlx4_priv *priv = mlx4_priv(dev);
780 
781 	priv->clr_base = ioremap(pci_resource_start(dev->pdev, priv->fw.clr_int_bar) +
782 				 priv->fw.clr_int_base, MLX4_CLR_INT_SIZE);
783 	if (!priv->clr_base) {
784 		mlx4_err(dev, "Couldn't map interrupt clear register, aborting.\n");
785 		return -ENOMEM;
786 	}
787 
788 	return 0;
789 }
790 
791 static void mlx4_unmap_clr_int(struct mlx4_dev *dev)
792 {
793 	struct mlx4_priv *priv = mlx4_priv(dev);
794 
795 	iounmap(priv->clr_base);
796 }
797 
798 int mlx4_alloc_eq_table(struct mlx4_dev *dev)
799 {
800 	struct mlx4_priv *priv = mlx4_priv(dev);
801 
802 	priv->eq_table.eq = kcalloc(dev->caps.num_eqs - dev->caps.reserved_eqs,
803 				    sizeof *priv->eq_table.eq, GFP_KERNEL);
804 	if (!priv->eq_table.eq)
805 		return -ENOMEM;
806 
807 	return 0;
808 }
809 
810 void mlx4_free_eq_table(struct mlx4_dev *dev)
811 {
812 	kfree(mlx4_priv(dev)->eq_table.eq);
813 }
814 
815 int mlx4_init_eq_table(struct mlx4_dev *dev)
816 {
817 	struct mlx4_priv *priv = mlx4_priv(dev);
818 	int err;
819 	int i;
820 
821 	priv->eq_table.uar_map = kcalloc(sizeof *priv->eq_table.uar_map,
822 					 mlx4_num_eq_uar(dev), GFP_KERNEL);
823 	if (!priv->eq_table.uar_map) {
824 		err = -ENOMEM;
825 		goto err_out_free;
826 	}
827 
828 	err = mlx4_bitmap_init(&priv->eq_table.bitmap, dev->caps.num_eqs,
829 			       dev->caps.num_eqs - 1, dev->caps.reserved_eqs, 0);
830 	if (err)
831 		goto err_out_free;
832 
833 	for (i = 0; i < mlx4_num_eq_uar(dev); ++i)
834 		priv->eq_table.uar_map[i] = NULL;
835 
836 	if (!mlx4_is_slave(dev)) {
837 		err = mlx4_map_clr_int(dev);
838 		if (err)
839 			goto err_out_bitmap;
840 
841 		priv->eq_table.clr_mask =
842 			swab32(1 << (priv->eq_table.inta_pin & 31));
843 		priv->eq_table.clr_int  = priv->clr_base +
844 			(priv->eq_table.inta_pin < 32 ? 4 : 0);
845 	}
846 
847 	priv->eq_table.irq_names =
848 		kmalloc(MLX4_IRQNAME_SIZE * (dev->caps.num_comp_vectors + 1 +
849 					     dev->caps.comp_pool),
850 			GFP_KERNEL);
851 	if (!priv->eq_table.irq_names) {
852 		err = -ENOMEM;
853 		goto err_out_bitmap;
854 	}
855 
856 	for (i = 0; i < dev->caps.num_comp_vectors; ++i) {
857 		err = mlx4_create_eq(dev, dev->caps.num_cqs -
858 					  dev->caps.reserved_cqs +
859 					  MLX4_NUM_SPARE_EQE,
860 				     (dev->flags & MLX4_FLAG_MSI_X) ? i : 0,
861 				     &priv->eq_table.eq[i]);
862 		if (err) {
863 			--i;
864 			goto err_out_unmap;
865 		}
866 	}
867 
868 	err = mlx4_create_eq(dev, MLX4_NUM_ASYNC_EQE + MLX4_NUM_SPARE_EQE,
869 			     (dev->flags & MLX4_FLAG_MSI_X) ? dev->caps.num_comp_vectors : 0,
870 			     &priv->eq_table.eq[dev->caps.num_comp_vectors]);
871 	if (err)
872 		goto err_out_comp;
873 
874 	/*if additional completion vectors poolsize is 0 this loop will not run*/
875 	for (i = dev->caps.num_comp_vectors + 1;
876 	      i < dev->caps.num_comp_vectors + dev->caps.comp_pool + 1; ++i) {
877 
878 		err = mlx4_create_eq(dev, dev->caps.num_cqs -
879 					  dev->caps.reserved_cqs +
880 					  MLX4_NUM_SPARE_EQE,
881 				     (dev->flags & MLX4_FLAG_MSI_X) ? i : 0,
882 				     &priv->eq_table.eq[i]);
883 		if (err) {
884 			--i;
885 			goto err_out_unmap;
886 		}
887 	}
888 
889 
890 	if (dev->flags & MLX4_FLAG_MSI_X) {
891 		const char *eq_name;
892 
893 		for (i = 0; i < dev->caps.num_comp_vectors + 1; ++i) {
894 			if (i < dev->caps.num_comp_vectors) {
895 				snprintf(priv->eq_table.irq_names +
896 					 i * MLX4_IRQNAME_SIZE,
897 					 MLX4_IRQNAME_SIZE,
898 					 "mlx4-comp-%d@pci:%s", i,
899 					 pci_name(dev->pdev));
900 			} else {
901 				snprintf(priv->eq_table.irq_names +
902 					 i * MLX4_IRQNAME_SIZE,
903 					 MLX4_IRQNAME_SIZE,
904 					 "mlx4-async@pci:%s",
905 					 pci_name(dev->pdev));
906 			}
907 
908 			eq_name = priv->eq_table.irq_names +
909 				  i * MLX4_IRQNAME_SIZE;
910 			err = request_irq(priv->eq_table.eq[i].irq,
911 					  mlx4_msi_x_interrupt, 0, eq_name,
912 					  priv->eq_table.eq + i);
913 			if (err)
914 				goto err_out_async;
915 
916 			priv->eq_table.eq[i].have_irq = 1;
917 		}
918 	} else {
919 		snprintf(priv->eq_table.irq_names,
920 			 MLX4_IRQNAME_SIZE,
921 			 DRV_NAME "@pci:%s",
922 			 pci_name(dev->pdev));
923 		err = request_irq(dev->pdev->irq, mlx4_interrupt,
924 				  IRQF_SHARED, priv->eq_table.irq_names, dev);
925 		if (err)
926 			goto err_out_async;
927 
928 		priv->eq_table.have_irq = 1;
929 	}
930 
931 	err = mlx4_MAP_EQ(dev, MLX4_ASYNC_EVENT_MASK, 0,
932 			  priv->eq_table.eq[dev->caps.num_comp_vectors].eqn);
933 	if (err)
934 		mlx4_warn(dev, "MAP_EQ for async EQ %d failed (%d)\n",
935 			   priv->eq_table.eq[dev->caps.num_comp_vectors].eqn, err);
936 
937 	for (i = 0; i < dev->caps.num_comp_vectors + 1; ++i)
938 		eq_set_ci(&priv->eq_table.eq[i], 1);
939 
940 	return 0;
941 
942 err_out_async:
943 	mlx4_free_eq(dev, &priv->eq_table.eq[dev->caps.num_comp_vectors]);
944 
945 err_out_comp:
946 	i = dev->caps.num_comp_vectors - 1;
947 
948 err_out_unmap:
949 	while (i >= 0) {
950 		mlx4_free_eq(dev, &priv->eq_table.eq[i]);
951 		--i;
952 	}
953 	if (!mlx4_is_slave(dev))
954 		mlx4_unmap_clr_int(dev);
955 	mlx4_free_irqs(dev);
956 
957 err_out_bitmap:
958 	mlx4_bitmap_cleanup(&priv->eq_table.bitmap);
959 
960 err_out_free:
961 	kfree(priv->eq_table.uar_map);
962 
963 	return err;
964 }
965 
966 void mlx4_cleanup_eq_table(struct mlx4_dev *dev)
967 {
968 	struct mlx4_priv *priv = mlx4_priv(dev);
969 	int i;
970 
971 	mlx4_MAP_EQ(dev, MLX4_ASYNC_EVENT_MASK, 1,
972 		    priv->eq_table.eq[dev->caps.num_comp_vectors].eqn);
973 
974 	mlx4_free_irqs(dev);
975 
976 	for (i = 0; i < dev->caps.num_comp_vectors + dev->caps.comp_pool + 1; ++i)
977 		mlx4_free_eq(dev, &priv->eq_table.eq[i]);
978 
979 	if (!mlx4_is_slave(dev))
980 		mlx4_unmap_clr_int(dev);
981 
982 	for (i = 0; i < mlx4_num_eq_uar(dev); ++i)
983 		if (priv->eq_table.uar_map[i])
984 			iounmap(priv->eq_table.uar_map[i]);
985 
986 	mlx4_bitmap_cleanup(&priv->eq_table.bitmap);
987 
988 	kfree(priv->eq_table.uar_map);
989 }
990 
991 /* A test that verifies that we can accept interrupts on all
992  * the irq vectors of the device.
993  * Interrupts are checked using the NOP command.
994  */
995 int mlx4_test_interrupts(struct mlx4_dev *dev)
996 {
997 	struct mlx4_priv *priv = mlx4_priv(dev);
998 	int i;
999 	int err;
1000 
1001 	err = mlx4_NOP(dev);
1002 	/* When not in MSI_X, there is only one irq to check */
1003 	if (!(dev->flags & MLX4_FLAG_MSI_X) || mlx4_is_slave(dev))
1004 		return err;
1005 
1006 	/* A loop over all completion vectors, for each vector we will check
1007 	 * whether it works by mapping command completions to that vector
1008 	 * and performing a NOP command
1009 	 */
1010 	for(i = 0; !err && (i < dev->caps.num_comp_vectors); ++i) {
1011 		/* Temporary use polling for command completions */
1012 		mlx4_cmd_use_polling(dev);
1013 
1014 		/* Map the new eq to handle all asyncronous events */
1015 		err = mlx4_MAP_EQ(dev, MLX4_ASYNC_EVENT_MASK, 0,
1016 				  priv->eq_table.eq[i].eqn);
1017 		if (err) {
1018 			mlx4_warn(dev, "Failed mapping eq for interrupt test\n");
1019 			mlx4_cmd_use_events(dev);
1020 			break;
1021 		}
1022 
1023 		/* Go back to using events */
1024 		mlx4_cmd_use_events(dev);
1025 		err = mlx4_NOP(dev);
1026 	}
1027 
1028 	/* Return to default */
1029 	mlx4_MAP_EQ(dev, MLX4_ASYNC_EVENT_MASK, 0,
1030 		    priv->eq_table.eq[dev->caps.num_comp_vectors].eqn);
1031 	return err;
1032 }
1033 EXPORT_SYMBOL(mlx4_test_interrupts);
1034 
1035 int mlx4_assign_eq(struct mlx4_dev *dev, char* name, int * vector)
1036 {
1037 
1038 	struct mlx4_priv *priv = mlx4_priv(dev);
1039 	int vec = 0, err = 0, i;
1040 
1041 	spin_lock(&priv->msix_ctl.pool_lock);
1042 	for (i = 0; !vec && i < dev->caps.comp_pool; i++) {
1043 		if (~priv->msix_ctl.pool_bm & 1ULL << i) {
1044 			priv->msix_ctl.pool_bm |= 1ULL << i;
1045 			vec = dev->caps.num_comp_vectors + 1 + i;
1046 			snprintf(priv->eq_table.irq_names +
1047 					vec * MLX4_IRQNAME_SIZE,
1048 					MLX4_IRQNAME_SIZE, "%s", name);
1049 			err = request_irq(priv->eq_table.eq[vec].irq,
1050 					  mlx4_msi_x_interrupt, 0,
1051 					  &priv->eq_table.irq_names[vec<<5],
1052 					  priv->eq_table.eq + vec);
1053 			if (err) {
1054 				/*zero out bit by fliping it*/
1055 				priv->msix_ctl.pool_bm ^= 1 << i;
1056 				vec = 0;
1057 				continue;
1058 				/*we dont want to break here*/
1059 			}
1060 			eq_set_ci(&priv->eq_table.eq[vec], 1);
1061 		}
1062 	}
1063 	spin_unlock(&priv->msix_ctl.pool_lock);
1064 
1065 	if (vec) {
1066 		*vector = vec;
1067 	} else {
1068 		*vector = 0;
1069 		err = (i == dev->caps.comp_pool) ? -ENOSPC : err;
1070 	}
1071 	return err;
1072 }
1073 EXPORT_SYMBOL(mlx4_assign_eq);
1074 
1075 void mlx4_release_eq(struct mlx4_dev *dev, int vec)
1076 {
1077 	struct mlx4_priv *priv = mlx4_priv(dev);
1078 	/*bm index*/
1079 	int i = vec - dev->caps.num_comp_vectors - 1;
1080 
1081 	if (likely(i >= 0)) {
1082 		/*sanity check , making sure were not trying to free irq's
1083 		  Belonging to a legacy EQ*/
1084 		spin_lock(&priv->msix_ctl.pool_lock);
1085 		if (priv->msix_ctl.pool_bm & 1ULL << i) {
1086 			free_irq(priv->eq_table.eq[vec].irq,
1087 				 &priv->eq_table.eq[vec]);
1088 			priv->msix_ctl.pool_bm &= ~(1ULL << i);
1089 		}
1090 		spin_unlock(&priv->msix_ctl.pool_lock);
1091 	}
1092 
1093 }
1094 EXPORT_SYMBOL(mlx4_release_eq);
1095 
1096