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/interrupt.h>
35 #include <linux/slab.h>
36 #include <linux/export.h>
37 #include <linux/mm.h>
38 #include <linux/dma-mapping.h>
39 
40 #include <linux/mlx4/cmd.h>
41 #include <linux/cpu_rmap.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 struct mlx4_irq_notify {
57 	void *arg;
58 	struct irq_affinity_notify notify;
59 };
60 
61 #define MLX4_EQ_STATUS_OK	   ( 0 << 28)
62 #define MLX4_EQ_STATUS_WRITE_FAIL  (10 << 28)
63 #define MLX4_EQ_OWNER_SW	   ( 0 << 24)
64 #define MLX4_EQ_OWNER_HW	   ( 1 << 24)
65 #define MLX4_EQ_FLAG_EC		   ( 1 << 18)
66 #define MLX4_EQ_FLAG_OI		   ( 1 << 17)
67 #define MLX4_EQ_STATE_ARMED	   ( 9 <<  8)
68 #define MLX4_EQ_STATE_FIRED	   (10 <<  8)
69 #define MLX4_EQ_STATE_ALWAYS_ARMED (11 <<  8)
70 
71 #define MLX4_ASYNC_EVENT_MASK ((1ull << MLX4_EVENT_TYPE_PATH_MIG)	    | \
72 			       (1ull << MLX4_EVENT_TYPE_COMM_EST)	    | \
73 			       (1ull << MLX4_EVENT_TYPE_SQ_DRAINED)	    | \
74 			       (1ull << MLX4_EVENT_TYPE_CQ_ERROR)	    | \
75 			       (1ull << MLX4_EVENT_TYPE_WQ_CATAS_ERROR)	    | \
76 			       (1ull << MLX4_EVENT_TYPE_EEC_CATAS_ERROR)    | \
77 			       (1ull << MLX4_EVENT_TYPE_PATH_MIG_FAILED)    | \
78 			       (1ull << MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR) | \
79 			       (1ull << MLX4_EVENT_TYPE_WQ_ACCESS_ERROR)    | \
80 			       (1ull << MLX4_EVENT_TYPE_PORT_CHANGE)	    | \
81 			       (1ull << MLX4_EVENT_TYPE_ECC_DETECT)	    | \
82 			       (1ull << MLX4_EVENT_TYPE_SRQ_CATAS_ERROR)    | \
83 			       (1ull << MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE)    | \
84 			       (1ull << MLX4_EVENT_TYPE_SRQ_LIMIT)	    | \
85 			       (1ull << MLX4_EVENT_TYPE_CMD)		    | \
86 			       (1ull << MLX4_EVENT_TYPE_OP_REQUIRED)	    | \
87 			       (1ull << MLX4_EVENT_TYPE_COMM_CHANNEL)       | \
88 			       (1ull << MLX4_EVENT_TYPE_FLR_EVENT)	    | \
89 			       (1ull << MLX4_EVENT_TYPE_FATAL_WARNING))
90 
91 static u64 get_async_ev_mask(struct mlx4_dev *dev)
92 {
93 	u64 async_ev_mask = MLX4_ASYNC_EVENT_MASK;
94 	if (dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV)
95 		async_ev_mask |= (1ull << MLX4_EVENT_TYPE_PORT_MNG_CHG_EVENT);
96 
97 	return async_ev_mask;
98 }
99 
100 static void eq_set_ci(struct mlx4_eq *eq, int req_not)
101 {
102 	__raw_writel((__force u32) cpu_to_be32((eq->cons_index & 0xffffff) |
103 					       req_not << 31),
104 		     eq->doorbell);
105 	/* We still want ordering, just not swabbing, so add a barrier */
106 	mb();
107 }
108 
109 static struct mlx4_eqe *get_eqe(struct mlx4_eq *eq, u32 entry, u8 eqe_factor)
110 {
111 	/* (entry & (eq->nent - 1)) gives us a cyclic array */
112 	unsigned long offset = (entry & (eq->nent - 1)) * (MLX4_EQ_ENTRY_SIZE << eqe_factor);
113 	/* CX3 is capable of extending the EQE from 32 to 64 bytes.
114 	 * When this feature is enabled, the first (in the lower addresses)
115 	 * 32 bytes in the 64 byte EQE are reserved and the next 32 bytes
116 	 * contain the legacy EQE information.
117 	 */
118 	return eq->page_list[offset / PAGE_SIZE].buf + (offset + (eqe_factor ? MLX4_EQ_ENTRY_SIZE : 0)) % PAGE_SIZE;
119 }
120 
121 static struct mlx4_eqe *next_eqe_sw(struct mlx4_eq *eq, u8 eqe_factor)
122 {
123 	struct mlx4_eqe *eqe = get_eqe(eq, eq->cons_index, eqe_factor);
124 	return !!(eqe->owner & 0x80) ^ !!(eq->cons_index & eq->nent) ? NULL : eqe;
125 }
126 
127 static struct mlx4_eqe *next_slave_event_eqe(struct mlx4_slave_event_eq *slave_eq)
128 {
129 	struct mlx4_eqe *eqe =
130 		&slave_eq->event_eqe[slave_eq->cons & (SLAVE_EVENT_EQ_SIZE - 1)];
131 	return (!!(eqe->owner & 0x80) ^
132 		!!(slave_eq->cons & SLAVE_EVENT_EQ_SIZE)) ?
133 		eqe : NULL;
134 }
135 
136 void mlx4_gen_slave_eqe(struct work_struct *work)
137 {
138 	struct mlx4_mfunc_master_ctx *master =
139 		container_of(work, struct mlx4_mfunc_master_ctx,
140 			     slave_event_work);
141 	struct mlx4_mfunc *mfunc =
142 		container_of(master, struct mlx4_mfunc, master);
143 	struct mlx4_priv *priv = container_of(mfunc, struct mlx4_priv, mfunc);
144 	struct mlx4_dev *dev = &priv->dev;
145 	struct mlx4_slave_event_eq *slave_eq = &mfunc->master.slave_eq;
146 	struct mlx4_eqe *eqe;
147 	u8 slave;
148 	int i;
149 
150 	for (eqe = next_slave_event_eqe(slave_eq); eqe;
151 	      eqe = next_slave_event_eqe(slave_eq)) {
152 		slave = eqe->slave_id;
153 
154 		/* All active slaves need to receive the event */
155 		if (slave == ALL_SLAVES) {
156 			for (i = 0; i < dev->num_slaves; i++) {
157 				if (i != dev->caps.function &&
158 				    master->slave_state[i].active)
159 					if (mlx4_GEN_EQE(dev, i, eqe))
160 						mlx4_warn(dev, "Failed to generate event for slave %d\n",
161 							  i);
162 			}
163 		} else {
164 			if (mlx4_GEN_EQE(dev, slave, eqe))
165 				mlx4_warn(dev, "Failed to generate event for slave %d\n",
166 					  slave);
167 		}
168 		++slave_eq->cons;
169 	}
170 }
171 
172 
173 static void slave_event(struct mlx4_dev *dev, u8 slave, struct mlx4_eqe *eqe)
174 {
175 	struct mlx4_priv *priv = mlx4_priv(dev);
176 	struct mlx4_slave_event_eq *slave_eq = &priv->mfunc.master.slave_eq;
177 	struct mlx4_eqe *s_eqe;
178 	unsigned long flags;
179 
180 	spin_lock_irqsave(&slave_eq->event_lock, flags);
181 	s_eqe = &slave_eq->event_eqe[slave_eq->prod & (SLAVE_EVENT_EQ_SIZE - 1)];
182 	if ((!!(s_eqe->owner & 0x80)) ^
183 	    (!!(slave_eq->prod & SLAVE_EVENT_EQ_SIZE))) {
184 		mlx4_warn(dev, "Master failed to generate an EQE for slave: %d. No free EQE on slave events queue\n",
185 			  slave);
186 		spin_unlock_irqrestore(&slave_eq->event_lock, flags);
187 		return;
188 	}
189 
190 	memcpy(s_eqe, eqe, dev->caps.eqe_size - 1);
191 	s_eqe->slave_id = slave;
192 	/* ensure all information is written before setting the ownersip bit */
193 	wmb();
194 	s_eqe->owner = !!(slave_eq->prod & SLAVE_EVENT_EQ_SIZE) ? 0x0 : 0x80;
195 	++slave_eq->prod;
196 
197 	queue_work(priv->mfunc.master.comm_wq,
198 		   &priv->mfunc.master.slave_event_work);
199 	spin_unlock_irqrestore(&slave_eq->event_lock, flags);
200 }
201 
202 static void mlx4_slave_event(struct mlx4_dev *dev, int slave,
203 			     struct mlx4_eqe *eqe)
204 {
205 	struct mlx4_priv *priv = mlx4_priv(dev);
206 	struct mlx4_slave_state *s_slave =
207 		&priv->mfunc.master.slave_state[slave];
208 
209 	if (!s_slave->active) {
210 		/*mlx4_warn(dev, "Trying to pass event to inactive slave\n");*/
211 		return;
212 	}
213 
214 	slave_event(dev, slave, eqe);
215 }
216 
217 int mlx4_gen_pkey_eqe(struct mlx4_dev *dev, int slave, u8 port)
218 {
219 	struct mlx4_eqe eqe;
220 
221 	struct mlx4_priv *priv = mlx4_priv(dev);
222 	struct mlx4_slave_state *s_slave = &priv->mfunc.master.slave_state[slave];
223 
224 	if (!s_slave->active)
225 		return 0;
226 
227 	memset(&eqe, 0, sizeof eqe);
228 
229 	eqe.type = MLX4_EVENT_TYPE_PORT_MNG_CHG_EVENT;
230 	eqe.subtype = MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE;
231 	eqe.event.port_mgmt_change.port = port;
232 
233 	return mlx4_GEN_EQE(dev, slave, &eqe);
234 }
235 EXPORT_SYMBOL(mlx4_gen_pkey_eqe);
236 
237 int mlx4_gen_guid_change_eqe(struct mlx4_dev *dev, int slave, u8 port)
238 {
239 	struct mlx4_eqe eqe;
240 
241 	/*don't send if we don't have the that slave */
242 	if (dev->num_vfs < slave)
243 		return 0;
244 	memset(&eqe, 0, sizeof eqe);
245 
246 	eqe.type = MLX4_EVENT_TYPE_PORT_MNG_CHG_EVENT;
247 	eqe.subtype = MLX4_DEV_PMC_SUBTYPE_GUID_INFO;
248 	eqe.event.port_mgmt_change.port = port;
249 
250 	return mlx4_GEN_EQE(dev, slave, &eqe);
251 }
252 EXPORT_SYMBOL(mlx4_gen_guid_change_eqe);
253 
254 int mlx4_gen_port_state_change_eqe(struct mlx4_dev *dev, int slave, u8 port,
255 				   u8 port_subtype_change)
256 {
257 	struct mlx4_eqe eqe;
258 
259 	/*don't send if we don't have the that slave */
260 	if (dev->num_vfs < slave)
261 		return 0;
262 	memset(&eqe, 0, sizeof eqe);
263 
264 	eqe.type = MLX4_EVENT_TYPE_PORT_CHANGE;
265 	eqe.subtype = port_subtype_change;
266 	eqe.event.port_change.port = cpu_to_be32(port << 28);
267 
268 	mlx4_dbg(dev, "%s: sending: %d to slave: %d on port: %d\n", __func__,
269 		 port_subtype_change, slave, port);
270 	return mlx4_GEN_EQE(dev, slave, &eqe);
271 }
272 EXPORT_SYMBOL(mlx4_gen_port_state_change_eqe);
273 
274 enum slave_port_state mlx4_get_slave_port_state(struct mlx4_dev *dev, int slave, u8 port)
275 {
276 	struct mlx4_priv *priv = mlx4_priv(dev);
277 	struct mlx4_slave_state *s_state = priv->mfunc.master.slave_state;
278 	struct mlx4_active_ports actv_ports = mlx4_get_active_ports(dev, slave);
279 
280 	if (slave >= dev->num_slaves || port > dev->caps.num_ports ||
281 	    port <= 0 || !test_bit(port - 1, actv_ports.ports)) {
282 		pr_err("%s: Error: asking for slave:%d, port:%d\n",
283 		       __func__, slave, port);
284 		return SLAVE_PORT_DOWN;
285 	}
286 	return s_state[slave].port_state[port];
287 }
288 EXPORT_SYMBOL(mlx4_get_slave_port_state);
289 
290 static int mlx4_set_slave_port_state(struct mlx4_dev *dev, int slave, u8 port,
291 				     enum slave_port_state state)
292 {
293 	struct mlx4_priv *priv = mlx4_priv(dev);
294 	struct mlx4_slave_state *s_state = priv->mfunc.master.slave_state;
295 	struct mlx4_active_ports actv_ports = mlx4_get_active_ports(dev, slave);
296 
297 	if (slave >= dev->num_slaves || port > dev->caps.num_ports ||
298 	    port <= 0 || !test_bit(port - 1, actv_ports.ports)) {
299 		pr_err("%s: Error: asking for slave:%d, port:%d\n",
300 		       __func__, slave, port);
301 		return -1;
302 	}
303 	s_state[slave].port_state[port] = state;
304 
305 	return 0;
306 }
307 
308 static void set_all_slave_state(struct mlx4_dev *dev, u8 port, int event)
309 {
310 	int i;
311 	enum slave_port_gen_event gen_event;
312 	struct mlx4_slaves_pport slaves_pport = mlx4_phys_to_slaves_pport(dev,
313 									  port);
314 
315 	for (i = 0; i < dev->num_vfs + 1; i++)
316 		if (test_bit(i, slaves_pport.slaves))
317 			set_and_calc_slave_port_state(dev, i, port,
318 						      event, &gen_event);
319 }
320 /**************************************************************************
321 	The function get as input the new event to that port,
322 	and according to the prev state change the slave's port state.
323 	The events are:
324 		MLX4_PORT_STATE_DEV_EVENT_PORT_DOWN,
325 		MLX4_PORT_STATE_DEV_EVENT_PORT_UP
326 		MLX4_PORT_STATE_IB_EVENT_GID_VALID
327 		MLX4_PORT_STATE_IB_EVENT_GID_INVALID
328 ***************************************************************************/
329 int set_and_calc_slave_port_state(struct mlx4_dev *dev, int slave,
330 				  u8 port, int event,
331 				  enum slave_port_gen_event *gen_event)
332 {
333 	struct mlx4_priv *priv = mlx4_priv(dev);
334 	struct mlx4_slave_state *ctx = NULL;
335 	unsigned long flags;
336 	int ret = -1;
337 	struct mlx4_active_ports actv_ports = mlx4_get_active_ports(dev, slave);
338 	enum slave_port_state cur_state =
339 		mlx4_get_slave_port_state(dev, slave, port);
340 
341 	*gen_event = SLAVE_PORT_GEN_EVENT_NONE;
342 
343 	if (slave >= dev->num_slaves || port > dev->caps.num_ports ||
344 	    port <= 0 || !test_bit(port - 1, actv_ports.ports)) {
345 		pr_err("%s: Error: asking for slave:%d, port:%d\n",
346 		       __func__, slave, port);
347 		return ret;
348 	}
349 
350 	ctx = &priv->mfunc.master.slave_state[slave];
351 	spin_lock_irqsave(&ctx->lock, flags);
352 
353 	switch (cur_state) {
354 	case SLAVE_PORT_DOWN:
355 		if (MLX4_PORT_STATE_DEV_EVENT_PORT_UP == event)
356 			mlx4_set_slave_port_state(dev, slave, port,
357 						  SLAVE_PENDING_UP);
358 		break;
359 	case SLAVE_PENDING_UP:
360 		if (MLX4_PORT_STATE_DEV_EVENT_PORT_DOWN == event)
361 			mlx4_set_slave_port_state(dev, slave, port,
362 						  SLAVE_PORT_DOWN);
363 		else if (MLX4_PORT_STATE_IB_PORT_STATE_EVENT_GID_VALID == event) {
364 			mlx4_set_slave_port_state(dev, slave, port,
365 						  SLAVE_PORT_UP);
366 			*gen_event = SLAVE_PORT_GEN_EVENT_UP;
367 		}
368 		break;
369 	case SLAVE_PORT_UP:
370 		if (MLX4_PORT_STATE_DEV_EVENT_PORT_DOWN == event) {
371 			mlx4_set_slave_port_state(dev, slave, port,
372 						  SLAVE_PORT_DOWN);
373 			*gen_event = SLAVE_PORT_GEN_EVENT_DOWN;
374 		} else if (MLX4_PORT_STATE_IB_EVENT_GID_INVALID ==
375 				event) {
376 			mlx4_set_slave_port_state(dev, slave, port,
377 						  SLAVE_PENDING_UP);
378 			*gen_event = SLAVE_PORT_GEN_EVENT_DOWN;
379 		}
380 		break;
381 	default:
382 		pr_err("%s: BUG!!! UNKNOWN state: slave:%d, port:%d\n",
383 		       __func__, slave, port);
384 		goto out;
385 	}
386 	ret = mlx4_get_slave_port_state(dev, slave, port);
387 
388 out:
389 	spin_unlock_irqrestore(&ctx->lock, flags);
390 	return ret;
391 }
392 
393 EXPORT_SYMBOL(set_and_calc_slave_port_state);
394 
395 int mlx4_gen_slaves_port_mgt_ev(struct mlx4_dev *dev, u8 port, int attr)
396 {
397 	struct mlx4_eqe eqe;
398 
399 	memset(&eqe, 0, sizeof eqe);
400 
401 	eqe.type = MLX4_EVENT_TYPE_PORT_MNG_CHG_EVENT;
402 	eqe.subtype = MLX4_DEV_PMC_SUBTYPE_PORT_INFO;
403 	eqe.event.port_mgmt_change.port = port;
404 	eqe.event.port_mgmt_change.params.port_info.changed_attr =
405 		cpu_to_be32((u32) attr);
406 
407 	slave_event(dev, ALL_SLAVES, &eqe);
408 	return 0;
409 }
410 EXPORT_SYMBOL(mlx4_gen_slaves_port_mgt_ev);
411 
412 void mlx4_master_handle_slave_flr(struct work_struct *work)
413 {
414 	struct mlx4_mfunc_master_ctx *master =
415 		container_of(work, struct mlx4_mfunc_master_ctx,
416 			     slave_flr_event_work);
417 	struct mlx4_mfunc *mfunc =
418 		container_of(master, struct mlx4_mfunc, master);
419 	struct mlx4_priv *priv =
420 		container_of(mfunc, struct mlx4_priv, mfunc);
421 	struct mlx4_dev *dev = &priv->dev;
422 	struct mlx4_slave_state *slave_state = priv->mfunc.master.slave_state;
423 	int i;
424 	int err;
425 	unsigned long flags;
426 
427 	mlx4_dbg(dev, "mlx4_handle_slave_flr\n");
428 
429 	for (i = 0 ; i < dev->num_slaves; i++) {
430 
431 		if (MLX4_COMM_CMD_FLR == slave_state[i].last_cmd) {
432 			mlx4_dbg(dev, "mlx4_handle_slave_flr: clean slave: %d\n",
433 				 i);
434 
435 			mlx4_delete_all_resources_for_slave(dev, i);
436 			/*return the slave to running mode*/
437 			spin_lock_irqsave(&priv->mfunc.master.slave_state_lock, flags);
438 			slave_state[i].last_cmd = MLX4_COMM_CMD_RESET;
439 			slave_state[i].is_slave_going_down = 0;
440 			spin_unlock_irqrestore(&priv->mfunc.master.slave_state_lock, flags);
441 			/*notify the FW:*/
442 			err = mlx4_cmd(dev, 0, i, 0, MLX4_CMD_INFORM_FLR_DONE,
443 				       MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED);
444 			if (err)
445 				mlx4_warn(dev, "Failed to notify FW on FLR done (slave:%d)\n",
446 					  i);
447 		}
448 	}
449 }
450 
451 static int mlx4_eq_int(struct mlx4_dev *dev, struct mlx4_eq *eq)
452 {
453 	struct mlx4_priv *priv = mlx4_priv(dev);
454 	struct mlx4_eqe *eqe;
455 	int cqn;
456 	int eqes_found = 0;
457 	int set_ci = 0;
458 	int port;
459 	int slave = 0;
460 	int ret;
461 	u32 flr_slave;
462 	u8 update_slave_state;
463 	int i;
464 	enum slave_port_gen_event gen_event;
465 	unsigned long flags;
466 	struct mlx4_vport_state *s_info;
467 
468 	while ((eqe = next_eqe_sw(eq, dev->caps.eqe_factor))) {
469 		/*
470 		 * Make sure we read EQ entry contents after we've
471 		 * checked the ownership bit.
472 		 */
473 		rmb();
474 
475 		switch (eqe->type) {
476 		case MLX4_EVENT_TYPE_COMP:
477 			cqn = be32_to_cpu(eqe->event.comp.cqn) & 0xffffff;
478 			mlx4_cq_completion(dev, cqn);
479 			break;
480 
481 		case MLX4_EVENT_TYPE_PATH_MIG:
482 		case MLX4_EVENT_TYPE_COMM_EST:
483 		case MLX4_EVENT_TYPE_SQ_DRAINED:
484 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
485 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
486 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
487 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
488 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
489 			mlx4_dbg(dev, "event %d arrived\n", eqe->type);
490 			if (mlx4_is_master(dev)) {
491 				/* forward only to slave owning the QP */
492 				ret = mlx4_get_slave_from_resource_id(dev,
493 						RES_QP,
494 						be32_to_cpu(eqe->event.qp.qpn)
495 						& 0xffffff, &slave);
496 				if (ret && ret != -ENOENT) {
497 					mlx4_dbg(dev, "QP event %02x(%02x) on EQ %d at index %u: could not get slave id (%d)\n",
498 						 eqe->type, eqe->subtype,
499 						 eq->eqn, eq->cons_index, ret);
500 					break;
501 				}
502 
503 				if (!ret && slave != dev->caps.function) {
504 					mlx4_slave_event(dev, slave, eqe);
505 					break;
506 				}
507 
508 			}
509 			mlx4_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) &
510 				      0xffffff, eqe->type);
511 			break;
512 
513 		case MLX4_EVENT_TYPE_SRQ_LIMIT:
514 			mlx4_dbg(dev, "%s: MLX4_EVENT_TYPE_SRQ_LIMIT\n",
515 				 __func__);
516 		case MLX4_EVENT_TYPE_SRQ_CATAS_ERROR:
517 			if (mlx4_is_master(dev)) {
518 				/* forward only to slave owning the SRQ */
519 				ret = mlx4_get_slave_from_resource_id(dev,
520 						RES_SRQ,
521 						be32_to_cpu(eqe->event.srq.srqn)
522 						& 0xffffff,
523 						&slave);
524 				if (ret && ret != -ENOENT) {
525 					mlx4_warn(dev, "SRQ event %02x(%02x) on EQ %d at index %u: could not get slave id (%d)\n",
526 						  eqe->type, eqe->subtype,
527 						  eq->eqn, eq->cons_index, ret);
528 					break;
529 				}
530 				mlx4_warn(dev, "%s: slave:%d, srq_no:0x%x, event: %02x(%02x)\n",
531 					  __func__, slave,
532 					  be32_to_cpu(eqe->event.srq.srqn),
533 					  eqe->type, eqe->subtype);
534 
535 				if (!ret && slave != dev->caps.function) {
536 					mlx4_warn(dev, "%s: sending event %02x(%02x) to slave:%d\n",
537 						  __func__, eqe->type,
538 						  eqe->subtype, slave);
539 					mlx4_slave_event(dev, slave, eqe);
540 					break;
541 				}
542 			}
543 			mlx4_srq_event(dev, be32_to_cpu(eqe->event.srq.srqn) &
544 				       0xffffff, eqe->type);
545 			break;
546 
547 		case MLX4_EVENT_TYPE_CMD:
548 			mlx4_cmd_event(dev,
549 				       be16_to_cpu(eqe->event.cmd.token),
550 				       eqe->event.cmd.status,
551 				       be64_to_cpu(eqe->event.cmd.out_param));
552 			break;
553 
554 		case MLX4_EVENT_TYPE_PORT_CHANGE: {
555 			struct mlx4_slaves_pport slaves_port;
556 			port = be32_to_cpu(eqe->event.port_change.port) >> 28;
557 			slaves_port = mlx4_phys_to_slaves_pport(dev, port);
558 			if (eqe->subtype == MLX4_PORT_CHANGE_SUBTYPE_DOWN) {
559 				mlx4_dispatch_event(dev, MLX4_DEV_EVENT_PORT_DOWN,
560 						    port);
561 				mlx4_priv(dev)->sense.do_sense_port[port] = 1;
562 				if (!mlx4_is_master(dev))
563 					break;
564 				for (i = 0; i < dev->num_vfs + 1; i++) {
565 					if (!test_bit(i, slaves_port.slaves))
566 						continue;
567 					if (dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH) {
568 						if (i == mlx4_master_func_num(dev))
569 							continue;
570 						mlx4_dbg(dev, "%s: Sending MLX4_PORT_CHANGE_SUBTYPE_DOWN to slave: %d, port:%d\n",
571 							 __func__, i, port);
572 						s_info = &priv->mfunc.master.vf_oper[slave].vport[port].state;
573 						if (IFLA_VF_LINK_STATE_AUTO == s_info->link_state) {
574 							eqe->event.port_change.port =
575 								cpu_to_be32(
576 								(be32_to_cpu(eqe->event.port_change.port) & 0xFFFFFFF)
577 								| (mlx4_phys_to_slave_port(dev, i, port) << 28));
578 							mlx4_slave_event(dev, i, eqe);
579 						}
580 					} else {  /* IB port */
581 						set_and_calc_slave_port_state(dev, i, port,
582 									      MLX4_PORT_STATE_DEV_EVENT_PORT_DOWN,
583 									      &gen_event);
584 						/*we can be in pending state, then do not send port_down event*/
585 						if (SLAVE_PORT_GEN_EVENT_DOWN ==  gen_event) {
586 							if (i == mlx4_master_func_num(dev))
587 								continue;
588 							mlx4_slave_event(dev, i, eqe);
589 						}
590 					}
591 				}
592 			} else {
593 				mlx4_dispatch_event(dev, MLX4_DEV_EVENT_PORT_UP, port);
594 
595 				mlx4_priv(dev)->sense.do_sense_port[port] = 0;
596 
597 				if (!mlx4_is_master(dev))
598 					break;
599 				if (dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH)
600 					for (i = 0; i < dev->num_vfs + 1; i++) {
601 						if (!test_bit(i, slaves_port.slaves))
602 							continue;
603 						if (i == mlx4_master_func_num(dev))
604 							continue;
605 						s_info = &priv->mfunc.master.vf_oper[slave].vport[port].state;
606 						if (IFLA_VF_LINK_STATE_AUTO == s_info->link_state) {
607 							eqe->event.port_change.port =
608 								cpu_to_be32(
609 								(be32_to_cpu(eqe->event.port_change.port) & 0xFFFFFFF)
610 								| (mlx4_phys_to_slave_port(dev, i, port) << 28));
611 							mlx4_slave_event(dev, i, eqe);
612 						}
613 					}
614 				else /* IB port */
615 					/* port-up event will be sent to a slave when the
616 					 * slave's alias-guid is set. This is done in alias_GUID.c
617 					 */
618 					set_all_slave_state(dev, port, MLX4_DEV_EVENT_PORT_UP);
619 			}
620 			break;
621 		}
622 
623 		case MLX4_EVENT_TYPE_CQ_ERROR:
624 			mlx4_warn(dev, "CQ %s on CQN %06x\n",
625 				  eqe->event.cq_err.syndrome == 1 ?
626 				  "overrun" : "access violation",
627 				  be32_to_cpu(eqe->event.cq_err.cqn) & 0xffffff);
628 			if (mlx4_is_master(dev)) {
629 				ret = mlx4_get_slave_from_resource_id(dev,
630 					RES_CQ,
631 					be32_to_cpu(eqe->event.cq_err.cqn)
632 					& 0xffffff, &slave);
633 				if (ret && ret != -ENOENT) {
634 					mlx4_dbg(dev, "CQ event %02x(%02x) on EQ %d at index %u: could not get slave id (%d)\n",
635 						 eqe->type, eqe->subtype,
636 						 eq->eqn, eq->cons_index, ret);
637 					break;
638 				}
639 
640 				if (!ret && slave != dev->caps.function) {
641 					mlx4_slave_event(dev, slave, eqe);
642 					break;
643 				}
644 			}
645 			mlx4_cq_event(dev,
646 				      be32_to_cpu(eqe->event.cq_err.cqn)
647 				      & 0xffffff,
648 				      eqe->type);
649 			break;
650 
651 		case MLX4_EVENT_TYPE_EQ_OVERFLOW:
652 			mlx4_warn(dev, "EQ overrun on EQN %d\n", eq->eqn);
653 			break;
654 
655 		case MLX4_EVENT_TYPE_OP_REQUIRED:
656 			atomic_inc(&priv->opreq_count);
657 			/* FW commands can't be executed from interrupt context
658 			 * working in deferred task
659 			 */
660 			queue_work(mlx4_wq, &priv->opreq_task);
661 			break;
662 
663 		case MLX4_EVENT_TYPE_COMM_CHANNEL:
664 			if (!mlx4_is_master(dev)) {
665 				mlx4_warn(dev, "Received comm channel event for non master device\n");
666 				break;
667 			}
668 			memcpy(&priv->mfunc.master.comm_arm_bit_vector,
669 			       eqe->event.comm_channel_arm.bit_vec,
670 			       sizeof eqe->event.comm_channel_arm.bit_vec);
671 			queue_work(priv->mfunc.master.comm_wq,
672 				   &priv->mfunc.master.comm_work);
673 			break;
674 
675 		case MLX4_EVENT_TYPE_FLR_EVENT:
676 			flr_slave = be32_to_cpu(eqe->event.flr_event.slave_id);
677 			if (!mlx4_is_master(dev)) {
678 				mlx4_warn(dev, "Non-master function received FLR event\n");
679 				break;
680 			}
681 
682 			mlx4_dbg(dev, "FLR event for slave: %d\n", flr_slave);
683 
684 			if (flr_slave >= dev->num_slaves) {
685 				mlx4_warn(dev,
686 					  "Got FLR for unknown function: %d\n",
687 					  flr_slave);
688 				update_slave_state = 0;
689 			} else
690 				update_slave_state = 1;
691 
692 			spin_lock_irqsave(&priv->mfunc.master.slave_state_lock, flags);
693 			if (update_slave_state) {
694 				priv->mfunc.master.slave_state[flr_slave].active = false;
695 				priv->mfunc.master.slave_state[flr_slave].last_cmd = MLX4_COMM_CMD_FLR;
696 				priv->mfunc.master.slave_state[flr_slave].is_slave_going_down = 1;
697 			}
698 			spin_unlock_irqrestore(&priv->mfunc.master.slave_state_lock, flags);
699 			queue_work(priv->mfunc.master.comm_wq,
700 				   &priv->mfunc.master.slave_flr_event_work);
701 			break;
702 
703 		case MLX4_EVENT_TYPE_FATAL_WARNING:
704 			if (eqe->subtype == MLX4_FATAL_WARNING_SUBTYPE_WARMING) {
705 				if (mlx4_is_master(dev))
706 					for (i = 0; i < dev->num_slaves; i++) {
707 						mlx4_dbg(dev, "%s: Sending MLX4_FATAL_WARNING_SUBTYPE_WARMING to slave: %d\n",
708 							 __func__, i);
709 						if (i == dev->caps.function)
710 							continue;
711 						mlx4_slave_event(dev, i, eqe);
712 					}
713 				mlx4_err(dev, "Temperature Threshold was reached! Threshold: %d celsius degrees; Current Temperature: %d\n",
714 					 be16_to_cpu(eqe->event.warming.warning_threshold),
715 					 be16_to_cpu(eqe->event.warming.current_temperature));
716 			} else
717 				mlx4_warn(dev, "Unhandled event FATAL WARNING (%02x), subtype %02x on EQ %d at index %u. owner=%x, nent=0x%x, slave=%x, ownership=%s\n",
718 					  eqe->type, eqe->subtype, eq->eqn,
719 					  eq->cons_index, eqe->owner, eq->nent,
720 					  eqe->slave_id,
721 					  !!(eqe->owner & 0x80) ^
722 					  !!(eq->cons_index & eq->nent) ? "HW" : "SW");
723 
724 			break;
725 
726 		case MLX4_EVENT_TYPE_PORT_MNG_CHG_EVENT:
727 			mlx4_dispatch_event(dev, MLX4_DEV_EVENT_PORT_MGMT_CHANGE,
728 					    (unsigned long) eqe);
729 			break;
730 
731 		case MLX4_EVENT_TYPE_EEC_CATAS_ERROR:
732 		case MLX4_EVENT_TYPE_ECC_DETECT:
733 		default:
734 			mlx4_warn(dev, "Unhandled event %02x(%02x) on EQ %d at index %u. owner=%x, nent=0x%x, slave=%x, ownership=%s\n",
735 				  eqe->type, eqe->subtype, eq->eqn,
736 				  eq->cons_index, eqe->owner, eq->nent,
737 				  eqe->slave_id,
738 				  !!(eqe->owner & 0x80) ^
739 				  !!(eq->cons_index & eq->nent) ? "HW" : "SW");
740 			break;
741 		};
742 
743 		++eq->cons_index;
744 		eqes_found = 1;
745 		++set_ci;
746 
747 		/*
748 		 * The HCA will think the queue has overflowed if we
749 		 * don't tell it we've been processing events.  We
750 		 * create our EQs with MLX4_NUM_SPARE_EQE extra
751 		 * entries, so we must update our consumer index at
752 		 * least that often.
753 		 */
754 		if (unlikely(set_ci >= MLX4_NUM_SPARE_EQE)) {
755 			eq_set_ci(eq, 0);
756 			set_ci = 0;
757 		}
758 	}
759 
760 	eq_set_ci(eq, 1);
761 
762 	return eqes_found;
763 }
764 
765 static irqreturn_t mlx4_interrupt(int irq, void *dev_ptr)
766 {
767 	struct mlx4_dev *dev = dev_ptr;
768 	struct mlx4_priv *priv = mlx4_priv(dev);
769 	int work = 0;
770 	int i;
771 
772 	writel(priv->eq_table.clr_mask, priv->eq_table.clr_int);
773 
774 	for (i = 0; i < dev->caps.num_comp_vectors + 1; ++i)
775 		work |= mlx4_eq_int(dev, &priv->eq_table.eq[i]);
776 
777 	return IRQ_RETVAL(work);
778 }
779 
780 static irqreturn_t mlx4_msi_x_interrupt(int irq, void *eq_ptr)
781 {
782 	struct mlx4_eq  *eq  = eq_ptr;
783 	struct mlx4_dev *dev = eq->dev;
784 
785 	mlx4_eq_int(dev, eq);
786 
787 	/* MSI-X vectors always belong to us */
788 	return IRQ_HANDLED;
789 }
790 
791 int mlx4_MAP_EQ_wrapper(struct mlx4_dev *dev, int slave,
792 			struct mlx4_vhcr *vhcr,
793 			struct mlx4_cmd_mailbox *inbox,
794 			struct mlx4_cmd_mailbox *outbox,
795 			struct mlx4_cmd_info *cmd)
796 {
797 	struct mlx4_priv *priv = mlx4_priv(dev);
798 	struct mlx4_slave_event_eq_info *event_eq =
799 		priv->mfunc.master.slave_state[slave].event_eq;
800 	u32 in_modifier = vhcr->in_modifier;
801 	u32 eqn = in_modifier & 0x3FF;
802 	u64 in_param =  vhcr->in_param;
803 	int err = 0;
804 	int i;
805 
806 	if (slave == dev->caps.function)
807 		err = mlx4_cmd(dev, in_param, (in_modifier & 0x80000000) | eqn,
808 			       0, MLX4_CMD_MAP_EQ, MLX4_CMD_TIME_CLASS_B,
809 			       MLX4_CMD_NATIVE);
810 	if (!err)
811 		for (i = 0; i < MLX4_EVENT_TYPES_NUM; ++i)
812 			if (in_param & (1LL << i))
813 				event_eq[i].eqn = in_modifier >> 31 ? -1 : eqn;
814 
815 	return err;
816 }
817 
818 static int mlx4_MAP_EQ(struct mlx4_dev *dev, u64 event_mask, int unmap,
819 			int eq_num)
820 {
821 	return mlx4_cmd(dev, event_mask, (unmap << 31) | eq_num,
822 			0, MLX4_CMD_MAP_EQ, MLX4_CMD_TIME_CLASS_B,
823 			MLX4_CMD_WRAPPED);
824 }
825 
826 static int mlx4_SW2HW_EQ(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
827 			 int eq_num)
828 {
829 	return mlx4_cmd(dev, mailbox->dma, eq_num, 0,
830 			MLX4_CMD_SW2HW_EQ, MLX4_CMD_TIME_CLASS_A,
831 			MLX4_CMD_WRAPPED);
832 }
833 
834 static int mlx4_HW2SW_EQ(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
835 			 int eq_num)
836 {
837 	return mlx4_cmd_box(dev, 0, mailbox->dma, eq_num,
838 			    0, MLX4_CMD_HW2SW_EQ, MLX4_CMD_TIME_CLASS_A,
839 			    MLX4_CMD_WRAPPED);
840 }
841 
842 static int mlx4_num_eq_uar(struct mlx4_dev *dev)
843 {
844 	/*
845 	 * Each UAR holds 4 EQ doorbells.  To figure out how many UARs
846 	 * we need to map, take the difference of highest index and
847 	 * the lowest index we'll use and add 1.
848 	 */
849 	return (dev->caps.num_comp_vectors + 1 + dev->caps.reserved_eqs +
850 		 dev->caps.comp_pool)/4 - dev->caps.reserved_eqs/4 + 1;
851 }
852 
853 static void __iomem *mlx4_get_eq_uar(struct mlx4_dev *dev, struct mlx4_eq *eq)
854 {
855 	struct mlx4_priv *priv = mlx4_priv(dev);
856 	int index;
857 
858 	index = eq->eqn / 4 - dev->caps.reserved_eqs / 4;
859 
860 	if (!priv->eq_table.uar_map[index]) {
861 		priv->eq_table.uar_map[index] =
862 			ioremap(pci_resource_start(dev->pdev, 2) +
863 				((eq->eqn / 4) << PAGE_SHIFT),
864 				PAGE_SIZE);
865 		if (!priv->eq_table.uar_map[index]) {
866 			mlx4_err(dev, "Couldn't map EQ doorbell for EQN 0x%06x\n",
867 				 eq->eqn);
868 			return NULL;
869 		}
870 	}
871 
872 	return priv->eq_table.uar_map[index] + 0x800 + 8 * (eq->eqn % 4);
873 }
874 
875 static void mlx4_unmap_uar(struct mlx4_dev *dev)
876 {
877 	struct mlx4_priv *priv = mlx4_priv(dev);
878 	int i;
879 
880 	for (i = 0; i < mlx4_num_eq_uar(dev); ++i)
881 		if (priv->eq_table.uar_map[i]) {
882 			iounmap(priv->eq_table.uar_map[i]);
883 			priv->eq_table.uar_map[i] = NULL;
884 		}
885 }
886 
887 static int mlx4_create_eq(struct mlx4_dev *dev, int nent,
888 			  u8 intr, struct mlx4_eq *eq)
889 {
890 	struct mlx4_priv *priv = mlx4_priv(dev);
891 	struct mlx4_cmd_mailbox *mailbox;
892 	struct mlx4_eq_context *eq_context;
893 	int npages;
894 	u64 *dma_list = NULL;
895 	dma_addr_t t;
896 	u64 mtt_addr;
897 	int err = -ENOMEM;
898 	int i;
899 
900 	eq->dev   = dev;
901 	eq->nent  = roundup_pow_of_two(max(nent, 2));
902 	/* CX3 is capable of extending the CQE/EQE from 32 to 64 bytes */
903 	npages = PAGE_ALIGN(eq->nent * (MLX4_EQ_ENTRY_SIZE << dev->caps.eqe_factor)) / PAGE_SIZE;
904 
905 	eq->page_list = kmalloc(npages * sizeof *eq->page_list,
906 				GFP_KERNEL);
907 	if (!eq->page_list)
908 		goto err_out;
909 
910 	for (i = 0; i < npages; ++i)
911 		eq->page_list[i].buf = NULL;
912 
913 	dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
914 	if (!dma_list)
915 		goto err_out_free;
916 
917 	mailbox = mlx4_alloc_cmd_mailbox(dev);
918 	if (IS_ERR(mailbox))
919 		goto err_out_free;
920 	eq_context = mailbox->buf;
921 
922 	for (i = 0; i < npages; ++i) {
923 		eq->page_list[i].buf = dma_alloc_coherent(&dev->pdev->dev,
924 							  PAGE_SIZE, &t, GFP_KERNEL);
925 		if (!eq->page_list[i].buf)
926 			goto err_out_free_pages;
927 
928 		dma_list[i] = t;
929 		eq->page_list[i].map = t;
930 
931 		memset(eq->page_list[i].buf, 0, PAGE_SIZE);
932 	}
933 
934 	eq->eqn = mlx4_bitmap_alloc(&priv->eq_table.bitmap);
935 	if (eq->eqn == -1)
936 		goto err_out_free_pages;
937 
938 	eq->doorbell = mlx4_get_eq_uar(dev, eq);
939 	if (!eq->doorbell) {
940 		err = -ENOMEM;
941 		goto err_out_free_eq;
942 	}
943 
944 	err = mlx4_mtt_init(dev, npages, PAGE_SHIFT, &eq->mtt);
945 	if (err)
946 		goto err_out_free_eq;
947 
948 	err = mlx4_write_mtt(dev, &eq->mtt, 0, npages, dma_list);
949 	if (err)
950 		goto err_out_free_mtt;
951 
952 	eq_context->flags	  = cpu_to_be32(MLX4_EQ_STATUS_OK   |
953 						MLX4_EQ_STATE_ARMED);
954 	eq_context->log_eq_size	  = ilog2(eq->nent);
955 	eq_context->intr	  = intr;
956 	eq_context->log_page_size = PAGE_SHIFT - MLX4_ICM_PAGE_SHIFT;
957 
958 	mtt_addr = mlx4_mtt_addr(dev, &eq->mtt);
959 	eq_context->mtt_base_addr_h = mtt_addr >> 32;
960 	eq_context->mtt_base_addr_l = cpu_to_be32(mtt_addr & 0xffffffff);
961 
962 	err = mlx4_SW2HW_EQ(dev, mailbox, eq->eqn);
963 	if (err) {
964 		mlx4_warn(dev, "SW2HW_EQ failed (%d)\n", err);
965 		goto err_out_free_mtt;
966 	}
967 
968 	kfree(dma_list);
969 	mlx4_free_cmd_mailbox(dev, mailbox);
970 
971 	eq->cons_index = 0;
972 
973 	return err;
974 
975 err_out_free_mtt:
976 	mlx4_mtt_cleanup(dev, &eq->mtt);
977 
978 err_out_free_eq:
979 	mlx4_bitmap_free(&priv->eq_table.bitmap, eq->eqn, MLX4_USE_RR);
980 
981 err_out_free_pages:
982 	for (i = 0; i < npages; ++i)
983 		if (eq->page_list[i].buf)
984 			dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
985 					  eq->page_list[i].buf,
986 					  eq->page_list[i].map);
987 
988 	mlx4_free_cmd_mailbox(dev, mailbox);
989 
990 err_out_free:
991 	kfree(eq->page_list);
992 	kfree(dma_list);
993 
994 err_out:
995 	return err;
996 }
997 
998 static void mlx4_free_eq(struct mlx4_dev *dev,
999 			 struct mlx4_eq *eq)
1000 {
1001 	struct mlx4_priv *priv = mlx4_priv(dev);
1002 	struct mlx4_cmd_mailbox *mailbox;
1003 	int err;
1004 	int i;
1005 	/* CX3 is capable of extending the CQE/EQE from 32 to 64 bytes */
1006 	int npages = PAGE_ALIGN((MLX4_EQ_ENTRY_SIZE << dev->caps.eqe_factor) * eq->nent) / PAGE_SIZE;
1007 
1008 	mailbox = mlx4_alloc_cmd_mailbox(dev);
1009 	if (IS_ERR(mailbox))
1010 		return;
1011 
1012 	err = mlx4_HW2SW_EQ(dev, mailbox, eq->eqn);
1013 	if (err)
1014 		mlx4_warn(dev, "HW2SW_EQ failed (%d)\n", err);
1015 
1016 	if (0) {
1017 		mlx4_dbg(dev, "Dumping EQ context %02x:\n", eq->eqn);
1018 		for (i = 0; i < sizeof (struct mlx4_eq_context) / 4; ++i) {
1019 			if (i % 4 == 0)
1020 				pr_cont("[%02x] ", i * 4);
1021 			pr_cont(" %08x", be32_to_cpup(mailbox->buf + i * 4));
1022 			if ((i + 1) % 4 == 0)
1023 				pr_cont("\n");
1024 		}
1025 	}
1026 
1027 	mlx4_mtt_cleanup(dev, &eq->mtt);
1028 	for (i = 0; i < npages; ++i)
1029 		dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
1030 				    eq->page_list[i].buf,
1031 				    eq->page_list[i].map);
1032 
1033 	kfree(eq->page_list);
1034 	mlx4_bitmap_free(&priv->eq_table.bitmap, eq->eqn, MLX4_USE_RR);
1035 	mlx4_free_cmd_mailbox(dev, mailbox);
1036 }
1037 
1038 static void mlx4_free_irqs(struct mlx4_dev *dev)
1039 {
1040 	struct mlx4_eq_table *eq_table = &mlx4_priv(dev)->eq_table;
1041 	struct mlx4_priv *priv = mlx4_priv(dev);
1042 	int	i, vec;
1043 
1044 	if (eq_table->have_irq)
1045 		free_irq(dev->pdev->irq, dev);
1046 
1047 	for (i = 0; i < dev->caps.num_comp_vectors + 1; ++i)
1048 		if (eq_table->eq[i].have_irq) {
1049 			free_irq(eq_table->eq[i].irq, eq_table->eq + i);
1050 			eq_table->eq[i].have_irq = 0;
1051 		}
1052 
1053 	for (i = 0; i < dev->caps.comp_pool; i++) {
1054 		/*
1055 		 * Freeing the assigned irq's
1056 		 * all bits should be 0, but we need to validate
1057 		 */
1058 		if (priv->msix_ctl.pool_bm & 1ULL << i) {
1059 			/* NO need protecting*/
1060 			vec = dev->caps.num_comp_vectors + 1 + i;
1061 			free_irq(priv->eq_table.eq[vec].irq,
1062 				 &priv->eq_table.eq[vec]);
1063 		}
1064 	}
1065 
1066 
1067 	kfree(eq_table->irq_names);
1068 }
1069 
1070 static int mlx4_map_clr_int(struct mlx4_dev *dev)
1071 {
1072 	struct mlx4_priv *priv = mlx4_priv(dev);
1073 
1074 	priv->clr_base = ioremap(pci_resource_start(dev->pdev, priv->fw.clr_int_bar) +
1075 				 priv->fw.clr_int_base, MLX4_CLR_INT_SIZE);
1076 	if (!priv->clr_base) {
1077 		mlx4_err(dev, "Couldn't map interrupt clear register, aborting\n");
1078 		return -ENOMEM;
1079 	}
1080 
1081 	return 0;
1082 }
1083 
1084 static void mlx4_unmap_clr_int(struct mlx4_dev *dev)
1085 {
1086 	struct mlx4_priv *priv = mlx4_priv(dev);
1087 
1088 	iounmap(priv->clr_base);
1089 }
1090 
1091 static void mlx4_irq_notifier_notify(struct irq_affinity_notify *notify,
1092 				     const cpumask_t *mask)
1093 {
1094 	struct mlx4_irq_notify *n = container_of(notify,
1095 						 struct mlx4_irq_notify,
1096 						 notify);
1097 	struct mlx4_priv *priv = (struct mlx4_priv *)n->arg;
1098 	struct radix_tree_iter iter;
1099 	void **slot;
1100 
1101 	radix_tree_for_each_slot(slot, &priv->cq_table.tree, &iter, 0) {
1102 		struct mlx4_cq *cq = (struct mlx4_cq *)(*slot);
1103 
1104 		if (cq->irq == notify->irq)
1105 			cq->irq_affinity_change = true;
1106 	}
1107 }
1108 
1109 static void mlx4_release_irq_notifier(struct kref *ref)
1110 {
1111 	struct mlx4_irq_notify *n = container_of(ref, struct mlx4_irq_notify,
1112 						 notify.kref);
1113 	kfree(n);
1114 }
1115 
1116 static void mlx4_assign_irq_notifier(struct mlx4_priv *priv,
1117 				     struct mlx4_dev *dev, int irq)
1118 {
1119 	struct mlx4_irq_notify *irq_notifier = NULL;
1120 	int err = 0;
1121 
1122 	irq_notifier = kzalloc(sizeof(*irq_notifier), GFP_KERNEL);
1123 	if (!irq_notifier) {
1124 		mlx4_warn(dev, "Failed to allocate irq notifier. irq %d\n",
1125 			  irq);
1126 		return;
1127 	}
1128 
1129 	irq_notifier->notify.irq = irq;
1130 	irq_notifier->notify.notify = mlx4_irq_notifier_notify;
1131 	irq_notifier->notify.release = mlx4_release_irq_notifier;
1132 	irq_notifier->arg = priv;
1133 	err = irq_set_affinity_notifier(irq, &irq_notifier->notify);
1134 	if (err) {
1135 		kfree(irq_notifier);
1136 		irq_notifier = NULL;
1137 		mlx4_warn(dev, "Failed to set irq notifier. irq %d\n", irq);
1138 	}
1139 }
1140 
1141 
1142 int mlx4_alloc_eq_table(struct mlx4_dev *dev)
1143 {
1144 	struct mlx4_priv *priv = mlx4_priv(dev);
1145 
1146 	priv->eq_table.eq = kcalloc(dev->caps.num_eqs - dev->caps.reserved_eqs,
1147 				    sizeof *priv->eq_table.eq, GFP_KERNEL);
1148 	if (!priv->eq_table.eq)
1149 		return -ENOMEM;
1150 
1151 	return 0;
1152 }
1153 
1154 void mlx4_free_eq_table(struct mlx4_dev *dev)
1155 {
1156 	kfree(mlx4_priv(dev)->eq_table.eq);
1157 }
1158 
1159 int mlx4_init_eq_table(struct mlx4_dev *dev)
1160 {
1161 	struct mlx4_priv *priv = mlx4_priv(dev);
1162 	int err;
1163 	int i;
1164 
1165 	priv->eq_table.uar_map = kcalloc(mlx4_num_eq_uar(dev),
1166 					 sizeof *priv->eq_table.uar_map,
1167 					 GFP_KERNEL);
1168 	if (!priv->eq_table.uar_map) {
1169 		err = -ENOMEM;
1170 		goto err_out_free;
1171 	}
1172 
1173 	err = mlx4_bitmap_init(&priv->eq_table.bitmap, dev->caps.num_eqs,
1174 			       dev->caps.num_eqs - 1, dev->caps.reserved_eqs, 0);
1175 	if (err)
1176 		goto err_out_free;
1177 
1178 	for (i = 0; i < mlx4_num_eq_uar(dev); ++i)
1179 		priv->eq_table.uar_map[i] = NULL;
1180 
1181 	if (!mlx4_is_slave(dev)) {
1182 		err = mlx4_map_clr_int(dev);
1183 		if (err)
1184 			goto err_out_bitmap;
1185 
1186 		priv->eq_table.clr_mask =
1187 			swab32(1 << (priv->eq_table.inta_pin & 31));
1188 		priv->eq_table.clr_int  = priv->clr_base +
1189 			(priv->eq_table.inta_pin < 32 ? 4 : 0);
1190 	}
1191 
1192 	priv->eq_table.irq_names =
1193 		kmalloc(MLX4_IRQNAME_SIZE * (dev->caps.num_comp_vectors + 1 +
1194 					     dev->caps.comp_pool),
1195 			GFP_KERNEL);
1196 	if (!priv->eq_table.irq_names) {
1197 		err = -ENOMEM;
1198 		goto err_out_bitmap;
1199 	}
1200 
1201 	for (i = 0; i < dev->caps.num_comp_vectors; ++i) {
1202 		err = mlx4_create_eq(dev, dev->caps.num_cqs -
1203 					  dev->caps.reserved_cqs +
1204 					  MLX4_NUM_SPARE_EQE,
1205 				     (dev->flags & MLX4_FLAG_MSI_X) ? i : 0,
1206 				     &priv->eq_table.eq[i]);
1207 		if (err) {
1208 			--i;
1209 			goto err_out_unmap;
1210 		}
1211 	}
1212 
1213 	err = mlx4_create_eq(dev, MLX4_NUM_ASYNC_EQE + MLX4_NUM_SPARE_EQE,
1214 			     (dev->flags & MLX4_FLAG_MSI_X) ? dev->caps.num_comp_vectors : 0,
1215 			     &priv->eq_table.eq[dev->caps.num_comp_vectors]);
1216 	if (err)
1217 		goto err_out_comp;
1218 
1219 	/*if additional completion vectors poolsize is 0 this loop will not run*/
1220 	for (i = dev->caps.num_comp_vectors + 1;
1221 	      i < dev->caps.num_comp_vectors + dev->caps.comp_pool + 1; ++i) {
1222 
1223 		err = mlx4_create_eq(dev, dev->caps.num_cqs -
1224 					  dev->caps.reserved_cqs +
1225 					  MLX4_NUM_SPARE_EQE,
1226 				     (dev->flags & MLX4_FLAG_MSI_X) ? i : 0,
1227 				     &priv->eq_table.eq[i]);
1228 		if (err) {
1229 			--i;
1230 			goto err_out_unmap;
1231 		}
1232 	}
1233 
1234 
1235 	if (dev->flags & MLX4_FLAG_MSI_X) {
1236 		const char *eq_name;
1237 
1238 		for (i = 0; i < dev->caps.num_comp_vectors + 1; ++i) {
1239 			if (i < dev->caps.num_comp_vectors) {
1240 				snprintf(priv->eq_table.irq_names +
1241 					 i * MLX4_IRQNAME_SIZE,
1242 					 MLX4_IRQNAME_SIZE,
1243 					 "mlx4-comp-%d@pci:%s", i,
1244 					 pci_name(dev->pdev));
1245 			} else {
1246 				snprintf(priv->eq_table.irq_names +
1247 					 i * MLX4_IRQNAME_SIZE,
1248 					 MLX4_IRQNAME_SIZE,
1249 					 "mlx4-async@pci:%s",
1250 					 pci_name(dev->pdev));
1251 			}
1252 
1253 			eq_name = priv->eq_table.irq_names +
1254 				  i * MLX4_IRQNAME_SIZE;
1255 			err = request_irq(priv->eq_table.eq[i].irq,
1256 					  mlx4_msi_x_interrupt, 0, eq_name,
1257 					  priv->eq_table.eq + i);
1258 			if (err)
1259 				goto err_out_async;
1260 
1261 			priv->eq_table.eq[i].have_irq = 1;
1262 		}
1263 	} else {
1264 		snprintf(priv->eq_table.irq_names,
1265 			 MLX4_IRQNAME_SIZE,
1266 			 DRV_NAME "@pci:%s",
1267 			 pci_name(dev->pdev));
1268 		err = request_irq(dev->pdev->irq, mlx4_interrupt,
1269 				  IRQF_SHARED, priv->eq_table.irq_names, dev);
1270 		if (err)
1271 			goto err_out_async;
1272 
1273 		priv->eq_table.have_irq = 1;
1274 	}
1275 
1276 	err = mlx4_MAP_EQ(dev, get_async_ev_mask(dev), 0,
1277 			  priv->eq_table.eq[dev->caps.num_comp_vectors].eqn);
1278 	if (err)
1279 		mlx4_warn(dev, "MAP_EQ for async EQ %d failed (%d)\n",
1280 			   priv->eq_table.eq[dev->caps.num_comp_vectors].eqn, err);
1281 
1282 	for (i = 0; i < dev->caps.num_comp_vectors + 1; ++i)
1283 		eq_set_ci(&priv->eq_table.eq[i], 1);
1284 
1285 	return 0;
1286 
1287 err_out_async:
1288 	mlx4_free_eq(dev, &priv->eq_table.eq[dev->caps.num_comp_vectors]);
1289 
1290 err_out_comp:
1291 	i = dev->caps.num_comp_vectors - 1;
1292 
1293 err_out_unmap:
1294 	while (i >= 0) {
1295 		mlx4_free_eq(dev, &priv->eq_table.eq[i]);
1296 		--i;
1297 	}
1298 	if (!mlx4_is_slave(dev))
1299 		mlx4_unmap_clr_int(dev);
1300 	mlx4_free_irqs(dev);
1301 
1302 err_out_bitmap:
1303 	mlx4_unmap_uar(dev);
1304 	mlx4_bitmap_cleanup(&priv->eq_table.bitmap);
1305 
1306 err_out_free:
1307 	kfree(priv->eq_table.uar_map);
1308 
1309 	return err;
1310 }
1311 
1312 void mlx4_cleanup_eq_table(struct mlx4_dev *dev)
1313 {
1314 	struct mlx4_priv *priv = mlx4_priv(dev);
1315 	int i;
1316 
1317 	mlx4_MAP_EQ(dev, get_async_ev_mask(dev), 1,
1318 		    priv->eq_table.eq[dev->caps.num_comp_vectors].eqn);
1319 
1320 	mlx4_free_irqs(dev);
1321 
1322 	for (i = 0; i < dev->caps.num_comp_vectors + dev->caps.comp_pool + 1; ++i)
1323 		mlx4_free_eq(dev, &priv->eq_table.eq[i]);
1324 
1325 	if (!mlx4_is_slave(dev))
1326 		mlx4_unmap_clr_int(dev);
1327 
1328 	mlx4_unmap_uar(dev);
1329 	mlx4_bitmap_cleanup(&priv->eq_table.bitmap);
1330 
1331 	kfree(priv->eq_table.uar_map);
1332 }
1333 
1334 /* A test that verifies that we can accept interrupts on all
1335  * the irq vectors of the device.
1336  * Interrupts are checked using the NOP command.
1337  */
1338 int mlx4_test_interrupts(struct mlx4_dev *dev)
1339 {
1340 	struct mlx4_priv *priv = mlx4_priv(dev);
1341 	int i;
1342 	int err;
1343 
1344 	err = mlx4_NOP(dev);
1345 	/* When not in MSI_X, there is only one irq to check */
1346 	if (!(dev->flags & MLX4_FLAG_MSI_X) || mlx4_is_slave(dev))
1347 		return err;
1348 
1349 	/* A loop over all completion vectors, for each vector we will check
1350 	 * whether it works by mapping command completions to that vector
1351 	 * and performing a NOP command
1352 	 */
1353 	for(i = 0; !err && (i < dev->caps.num_comp_vectors); ++i) {
1354 		/* Temporary use polling for command completions */
1355 		mlx4_cmd_use_polling(dev);
1356 
1357 		/* Map the new eq to handle all asynchronous events */
1358 		err = mlx4_MAP_EQ(dev, get_async_ev_mask(dev), 0,
1359 				  priv->eq_table.eq[i].eqn);
1360 		if (err) {
1361 			mlx4_warn(dev, "Failed mapping eq for interrupt test\n");
1362 			mlx4_cmd_use_events(dev);
1363 			break;
1364 		}
1365 
1366 		/* Go back to using events */
1367 		mlx4_cmd_use_events(dev);
1368 		err = mlx4_NOP(dev);
1369 	}
1370 
1371 	/* Return to default */
1372 	mlx4_MAP_EQ(dev, get_async_ev_mask(dev), 0,
1373 		    priv->eq_table.eq[dev->caps.num_comp_vectors].eqn);
1374 	return err;
1375 }
1376 EXPORT_SYMBOL(mlx4_test_interrupts);
1377 
1378 int mlx4_assign_eq(struct mlx4_dev *dev, char *name, struct cpu_rmap *rmap,
1379 		   int *vector)
1380 {
1381 
1382 	struct mlx4_priv *priv = mlx4_priv(dev);
1383 	int vec = 0, err = 0, i;
1384 
1385 	mutex_lock(&priv->msix_ctl.pool_lock);
1386 	for (i = 0; !vec && i < dev->caps.comp_pool; i++) {
1387 		if (~priv->msix_ctl.pool_bm & 1ULL << i) {
1388 			priv->msix_ctl.pool_bm |= 1ULL << i;
1389 			vec = dev->caps.num_comp_vectors + 1 + i;
1390 			snprintf(priv->eq_table.irq_names +
1391 					vec * MLX4_IRQNAME_SIZE,
1392 					MLX4_IRQNAME_SIZE, "%s", name);
1393 #ifdef CONFIG_RFS_ACCEL
1394 			if (rmap) {
1395 				err = irq_cpu_rmap_add(rmap,
1396 						       priv->eq_table.eq[vec].irq);
1397 				if (err)
1398 					mlx4_warn(dev, "Failed adding irq rmap\n");
1399 			}
1400 #endif
1401 			err = request_irq(priv->eq_table.eq[vec].irq,
1402 					  mlx4_msi_x_interrupt, 0,
1403 					  &priv->eq_table.irq_names[vec<<5],
1404 					  priv->eq_table.eq + vec);
1405 			if (err) {
1406 				/*zero out bit by fliping it*/
1407 				priv->msix_ctl.pool_bm ^= 1 << i;
1408 				vec = 0;
1409 				continue;
1410 				/*we dont want to break here*/
1411 			}
1412 			mlx4_assign_irq_notifier(priv, dev,
1413 						 priv->eq_table.eq[vec].irq);
1414 
1415 			eq_set_ci(&priv->eq_table.eq[vec], 1);
1416 		}
1417 	}
1418 	mutex_unlock(&priv->msix_ctl.pool_lock);
1419 
1420 	if (vec) {
1421 		*vector = vec;
1422 	} else {
1423 		*vector = 0;
1424 		err = (i == dev->caps.comp_pool) ? -ENOSPC : err;
1425 	}
1426 	return err;
1427 }
1428 EXPORT_SYMBOL(mlx4_assign_eq);
1429 
1430 void mlx4_release_eq(struct mlx4_dev *dev, int vec)
1431 {
1432 	struct mlx4_priv *priv = mlx4_priv(dev);
1433 	/*bm index*/
1434 	int i = vec - dev->caps.num_comp_vectors - 1;
1435 
1436 	if (likely(i >= 0)) {
1437 		/*sanity check , making sure were not trying to free irq's
1438 		  Belonging to a legacy EQ*/
1439 		mutex_lock(&priv->msix_ctl.pool_lock);
1440 		if (priv->msix_ctl.pool_bm & 1ULL << i) {
1441 			irq_set_affinity_notifier(
1442 				priv->eq_table.eq[vec].irq,
1443 				NULL);
1444 			free_irq(priv->eq_table.eq[vec].irq,
1445 				 &priv->eq_table.eq[vec]);
1446 			priv->msix_ctl.pool_bm &= ~(1ULL << i);
1447 		}
1448 		mutex_unlock(&priv->msix_ctl.pool_lock);
1449 	}
1450 
1451 }
1452 EXPORT_SYMBOL(mlx4_release_eq);
1453 
1454