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 #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_OP_REQUIRED)	    | \
82 			       (1ull << MLX4_EVENT_TYPE_COMM_CHANNEL)       | \
83 			       (1ull << MLX4_EVENT_TYPE_FLR_EVENT)	    | \
84 			       (1ull << MLX4_EVENT_TYPE_FATAL_WARNING))
85 
86 static u64 get_async_ev_mask(struct mlx4_dev *dev)
87 {
88 	u64 async_ev_mask = MLX4_ASYNC_EVENT_MASK;
89 	if (dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV)
90 		async_ev_mask |= (1ull << MLX4_EVENT_TYPE_PORT_MNG_CHG_EVENT);
91 	if (dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_RECOVERABLE_ERROR_EVENT)
92 		async_ev_mask |= (1ull << MLX4_EVENT_TYPE_RECOVERABLE_ERROR_EVENT);
93 
94 	return async_ev_mask;
95 }
96 
97 static void eq_set_ci(struct mlx4_eq *eq, int req_not)
98 {
99 	__raw_writel((__force u32) cpu_to_be32((eq->cons_index & 0xffffff) |
100 					       req_not << 31),
101 		     eq->doorbell);
102 	/* We still want ordering, just not swabbing, so add a barrier */
103 	mb();
104 }
105 
106 static struct mlx4_eqe *get_eqe(struct mlx4_eq *eq, u32 entry, u8 eqe_factor,
107 				u8 eqe_size)
108 {
109 	/* (entry & (eq->nent - 1)) gives us a cyclic array */
110 	unsigned long offset = (entry & (eq->nent - 1)) * eqe_size;
111 	/* CX3 is capable of extending the EQE from 32 to 64 bytes with
112 	 * strides of 64B,128B and 256B.
113 	 * When 64B EQE is used, the first (in the lower addresses)
114 	 * 32 bytes in the 64 byte EQE are reserved and the next 32 bytes
115 	 * contain the legacy EQE information.
116 	 * In all other cases, the first 32B contains the legacy EQE info.
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, u8 size)
122 {
123 	struct mlx4_eqe *eqe = get_eqe(eq, eq->cons_index, eqe_factor, size);
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, phys_port, slave_port;
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->persist->num_vfs; i++) {
157 				phys_port = 0;
158 				if (eqe->type == MLX4_EVENT_TYPE_PORT_MNG_CHG_EVENT &&
159 				    eqe->subtype == MLX4_DEV_PMC_SUBTYPE_PORT_INFO) {
160 					phys_port  = eqe->event.port_mgmt_change.port;
161 					slave_port = mlx4_phys_to_slave_port(dev, i, phys_port);
162 					if (slave_port < 0) /* VF doesn't have this port */
163 						continue;
164 					eqe->event.port_mgmt_change.port = slave_port;
165 				}
166 				if (mlx4_GEN_EQE(dev, i, eqe))
167 					mlx4_warn(dev, "Failed to generate event for slave %d\n",
168 						  i);
169 				if (phys_port)
170 					eqe->event.port_mgmt_change.port = phys_port;
171 			}
172 		} else {
173 			if (mlx4_GEN_EQE(dev, slave, eqe))
174 				mlx4_warn(dev, "Failed to generate event for slave %d\n",
175 					  slave);
176 		}
177 		++slave_eq->cons;
178 	}
179 }
180 
181 
182 static void slave_event(struct mlx4_dev *dev, u8 slave, struct mlx4_eqe *eqe)
183 {
184 	struct mlx4_priv *priv = mlx4_priv(dev);
185 	struct mlx4_slave_event_eq *slave_eq = &priv->mfunc.master.slave_eq;
186 	struct mlx4_eqe *s_eqe;
187 	unsigned long flags;
188 
189 	spin_lock_irqsave(&slave_eq->event_lock, flags);
190 	s_eqe = &slave_eq->event_eqe[slave_eq->prod & (SLAVE_EVENT_EQ_SIZE - 1)];
191 	if ((!!(s_eqe->owner & 0x80)) ^
192 	    (!!(slave_eq->prod & SLAVE_EVENT_EQ_SIZE))) {
193 		mlx4_warn(dev, "Master failed to generate an EQE for slave: %d. No free EQE on slave events queue\n",
194 			  slave);
195 		spin_unlock_irqrestore(&slave_eq->event_lock, flags);
196 		return;
197 	}
198 
199 	memcpy(s_eqe, eqe, sizeof(struct mlx4_eqe) - 1);
200 	s_eqe->slave_id = slave;
201 	/* ensure all information is written before setting the ownersip bit */
202 	dma_wmb();
203 	s_eqe->owner = !!(slave_eq->prod & SLAVE_EVENT_EQ_SIZE) ? 0x0 : 0x80;
204 	++slave_eq->prod;
205 
206 	queue_work(priv->mfunc.master.comm_wq,
207 		   &priv->mfunc.master.slave_event_work);
208 	spin_unlock_irqrestore(&slave_eq->event_lock, flags);
209 }
210 
211 static void mlx4_slave_event(struct mlx4_dev *dev, int slave,
212 			     struct mlx4_eqe *eqe)
213 {
214 	struct mlx4_priv *priv = mlx4_priv(dev);
215 
216 	if (slave < 0 || slave > dev->persist->num_vfs ||
217 	    slave == dev->caps.function ||
218 	    !priv->mfunc.master.slave_state[slave].active)
219 		return;
220 
221 	slave_event(dev, slave, eqe);
222 }
223 
224 #if defined(CONFIG_SMP)
225 static void mlx4_set_eq_affinity_hint(struct mlx4_priv *priv, int vec)
226 {
227 	int hint_err;
228 	struct mlx4_dev *dev = &priv->dev;
229 	struct mlx4_eq *eq = &priv->eq_table.eq[vec];
230 
231 	if (!eq->affinity_mask || cpumask_empty(eq->affinity_mask))
232 		return;
233 
234 	hint_err = irq_set_affinity_hint(eq->irq, eq->affinity_mask);
235 	if (hint_err)
236 		mlx4_warn(dev, "irq_set_affinity_hint failed, err %d\n", hint_err);
237 }
238 #endif
239 
240 int mlx4_gen_pkey_eqe(struct mlx4_dev *dev, int slave, u8 port)
241 {
242 	struct mlx4_eqe eqe;
243 
244 	struct mlx4_priv *priv = mlx4_priv(dev);
245 	struct mlx4_slave_state *s_slave = &priv->mfunc.master.slave_state[slave];
246 
247 	if (!s_slave->active)
248 		return 0;
249 
250 	memset(&eqe, 0, sizeof eqe);
251 
252 	eqe.type = MLX4_EVENT_TYPE_PORT_MNG_CHG_EVENT;
253 	eqe.subtype = MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE;
254 	eqe.event.port_mgmt_change.port = mlx4_phys_to_slave_port(dev, slave, port);
255 
256 	return mlx4_GEN_EQE(dev, slave, &eqe);
257 }
258 EXPORT_SYMBOL(mlx4_gen_pkey_eqe);
259 
260 int mlx4_gen_guid_change_eqe(struct mlx4_dev *dev, int slave, u8 port)
261 {
262 	struct mlx4_eqe eqe;
263 
264 	/*don't send if we don't have the that slave */
265 	if (dev->persist->num_vfs < slave)
266 		return 0;
267 	memset(&eqe, 0, sizeof eqe);
268 
269 	eqe.type = MLX4_EVENT_TYPE_PORT_MNG_CHG_EVENT;
270 	eqe.subtype = MLX4_DEV_PMC_SUBTYPE_GUID_INFO;
271 	eqe.event.port_mgmt_change.port = mlx4_phys_to_slave_port(dev, slave, port);
272 
273 	return mlx4_GEN_EQE(dev, slave, &eqe);
274 }
275 EXPORT_SYMBOL(mlx4_gen_guid_change_eqe);
276 
277 int mlx4_gen_port_state_change_eqe(struct mlx4_dev *dev, int slave, u8 port,
278 				   u8 port_subtype_change)
279 {
280 	struct mlx4_eqe eqe;
281 	u8 slave_port = mlx4_phys_to_slave_port(dev, slave, port);
282 
283 	/*don't send if we don't have the that slave */
284 	if (dev->persist->num_vfs < slave)
285 		return 0;
286 	memset(&eqe, 0, sizeof eqe);
287 
288 	eqe.type = MLX4_EVENT_TYPE_PORT_CHANGE;
289 	eqe.subtype = port_subtype_change;
290 	eqe.event.port_change.port = cpu_to_be32(slave_port << 28);
291 
292 	mlx4_dbg(dev, "%s: sending: %d to slave: %d on port: %d\n", __func__,
293 		 port_subtype_change, slave, port);
294 	return mlx4_GEN_EQE(dev, slave, &eqe);
295 }
296 EXPORT_SYMBOL(mlx4_gen_port_state_change_eqe);
297 
298 enum slave_port_state mlx4_get_slave_port_state(struct mlx4_dev *dev, int slave, u8 port)
299 {
300 	struct mlx4_priv *priv = mlx4_priv(dev);
301 	struct mlx4_slave_state *s_state = priv->mfunc.master.slave_state;
302 	struct mlx4_active_ports actv_ports = mlx4_get_active_ports(dev, slave);
303 
304 	if (slave >= dev->num_slaves || port > dev->caps.num_ports ||
305 	    port <= 0 || !test_bit(port - 1, actv_ports.ports)) {
306 		pr_err("%s: Error: asking for slave:%d, port:%d\n",
307 		       __func__, slave, port);
308 		return SLAVE_PORT_DOWN;
309 	}
310 	return s_state[slave].port_state[port];
311 }
312 EXPORT_SYMBOL(mlx4_get_slave_port_state);
313 
314 static int mlx4_set_slave_port_state(struct mlx4_dev *dev, int slave, u8 port,
315 				     enum slave_port_state state)
316 {
317 	struct mlx4_priv *priv = mlx4_priv(dev);
318 	struct mlx4_slave_state *s_state = priv->mfunc.master.slave_state;
319 	struct mlx4_active_ports actv_ports = mlx4_get_active_ports(dev, slave);
320 
321 	if (slave >= dev->num_slaves || port > dev->caps.num_ports ||
322 	    port <= 0 || !test_bit(port - 1, actv_ports.ports)) {
323 		pr_err("%s: Error: asking for slave:%d, port:%d\n",
324 		       __func__, slave, port);
325 		return -1;
326 	}
327 	s_state[slave].port_state[port] = state;
328 
329 	return 0;
330 }
331 
332 static void set_all_slave_state(struct mlx4_dev *dev, u8 port, int event)
333 {
334 	int i;
335 	enum slave_port_gen_event gen_event;
336 	struct mlx4_slaves_pport slaves_pport = mlx4_phys_to_slaves_pport(dev,
337 									  port);
338 
339 	for (i = 0; i < dev->persist->num_vfs + 1; i++)
340 		if (test_bit(i, slaves_pport.slaves))
341 			set_and_calc_slave_port_state(dev, i, port,
342 						      event, &gen_event);
343 }
344 /**************************************************************************
345 	The function get as input the new event to that port,
346 	and according to the prev state change the slave's port state.
347 	The events are:
348 		MLX4_PORT_STATE_DEV_EVENT_PORT_DOWN,
349 		MLX4_PORT_STATE_DEV_EVENT_PORT_UP
350 		MLX4_PORT_STATE_IB_EVENT_GID_VALID
351 		MLX4_PORT_STATE_IB_EVENT_GID_INVALID
352 ***************************************************************************/
353 int set_and_calc_slave_port_state(struct mlx4_dev *dev, int slave,
354 				  u8 port, int event,
355 				  enum slave_port_gen_event *gen_event)
356 {
357 	struct mlx4_priv *priv = mlx4_priv(dev);
358 	struct mlx4_slave_state *ctx = NULL;
359 	unsigned long flags;
360 	int ret = -1;
361 	struct mlx4_active_ports actv_ports = mlx4_get_active_ports(dev, slave);
362 	enum slave_port_state cur_state =
363 		mlx4_get_slave_port_state(dev, slave, port);
364 
365 	*gen_event = SLAVE_PORT_GEN_EVENT_NONE;
366 
367 	if (slave >= dev->num_slaves || port > dev->caps.num_ports ||
368 	    port <= 0 || !test_bit(port - 1, actv_ports.ports)) {
369 		pr_err("%s: Error: asking for slave:%d, port:%d\n",
370 		       __func__, slave, port);
371 		return ret;
372 	}
373 
374 	ctx = &priv->mfunc.master.slave_state[slave];
375 	spin_lock_irqsave(&ctx->lock, flags);
376 
377 	switch (cur_state) {
378 	case SLAVE_PORT_DOWN:
379 		if (MLX4_PORT_STATE_DEV_EVENT_PORT_UP == event)
380 			mlx4_set_slave_port_state(dev, slave, port,
381 						  SLAVE_PENDING_UP);
382 		break;
383 	case SLAVE_PENDING_UP:
384 		if (MLX4_PORT_STATE_DEV_EVENT_PORT_DOWN == event)
385 			mlx4_set_slave_port_state(dev, slave, port,
386 						  SLAVE_PORT_DOWN);
387 		else if (MLX4_PORT_STATE_IB_PORT_STATE_EVENT_GID_VALID == event) {
388 			mlx4_set_slave_port_state(dev, slave, port,
389 						  SLAVE_PORT_UP);
390 			*gen_event = SLAVE_PORT_GEN_EVENT_UP;
391 		}
392 		break;
393 	case SLAVE_PORT_UP:
394 		if (MLX4_PORT_STATE_DEV_EVENT_PORT_DOWN == event) {
395 			mlx4_set_slave_port_state(dev, slave, port,
396 						  SLAVE_PORT_DOWN);
397 			*gen_event = SLAVE_PORT_GEN_EVENT_DOWN;
398 		} else if (MLX4_PORT_STATE_IB_EVENT_GID_INVALID ==
399 				event) {
400 			mlx4_set_slave_port_state(dev, slave, port,
401 						  SLAVE_PENDING_UP);
402 			*gen_event = SLAVE_PORT_GEN_EVENT_DOWN;
403 		}
404 		break;
405 	default:
406 		pr_err("%s: BUG!!! UNKNOWN state: slave:%d, port:%d\n",
407 		       __func__, slave, port);
408 		goto out;
409 	}
410 	ret = mlx4_get_slave_port_state(dev, slave, port);
411 
412 out:
413 	spin_unlock_irqrestore(&ctx->lock, flags);
414 	return ret;
415 }
416 
417 EXPORT_SYMBOL(set_and_calc_slave_port_state);
418 
419 int mlx4_gen_slaves_port_mgt_ev(struct mlx4_dev *dev, u8 port, int attr)
420 {
421 	struct mlx4_eqe eqe;
422 
423 	memset(&eqe, 0, sizeof eqe);
424 
425 	eqe.type = MLX4_EVENT_TYPE_PORT_MNG_CHG_EVENT;
426 	eqe.subtype = MLX4_DEV_PMC_SUBTYPE_PORT_INFO;
427 	eqe.event.port_mgmt_change.port = port;
428 	eqe.event.port_mgmt_change.params.port_info.changed_attr =
429 		cpu_to_be32((u32) attr);
430 
431 	slave_event(dev, ALL_SLAVES, &eqe);
432 	return 0;
433 }
434 EXPORT_SYMBOL(mlx4_gen_slaves_port_mgt_ev);
435 
436 void mlx4_master_handle_slave_flr(struct work_struct *work)
437 {
438 	struct mlx4_mfunc_master_ctx *master =
439 		container_of(work, struct mlx4_mfunc_master_ctx,
440 			     slave_flr_event_work);
441 	struct mlx4_mfunc *mfunc =
442 		container_of(master, struct mlx4_mfunc, master);
443 	struct mlx4_priv *priv =
444 		container_of(mfunc, struct mlx4_priv, mfunc);
445 	struct mlx4_dev *dev = &priv->dev;
446 	struct mlx4_slave_state *slave_state = priv->mfunc.master.slave_state;
447 	int i;
448 	int err;
449 	unsigned long flags;
450 
451 	mlx4_dbg(dev, "mlx4_handle_slave_flr\n");
452 
453 	for (i = 0 ; i < dev->num_slaves; i++) {
454 
455 		if (MLX4_COMM_CMD_FLR == slave_state[i].last_cmd) {
456 			mlx4_dbg(dev, "mlx4_handle_slave_flr: clean slave: %d\n",
457 				 i);
458 			/* In case of 'Reset flow' FLR can be generated for
459 			 * a slave before mlx4_load_one is done.
460 			 * make sure interface is up before trying to delete
461 			 * slave resources which weren't allocated yet.
462 			 */
463 			if (dev->persist->interface_state &
464 			    MLX4_INTERFACE_STATE_UP)
465 				mlx4_delete_all_resources_for_slave(dev, i);
466 			/*return the slave to running mode*/
467 			spin_lock_irqsave(&priv->mfunc.master.slave_state_lock, flags);
468 			slave_state[i].last_cmd = MLX4_COMM_CMD_RESET;
469 			slave_state[i].is_slave_going_down = 0;
470 			spin_unlock_irqrestore(&priv->mfunc.master.slave_state_lock, flags);
471 			/*notify the FW:*/
472 			err = mlx4_cmd(dev, 0, i, 0, MLX4_CMD_INFORM_FLR_DONE,
473 				       MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED);
474 			if (err)
475 				mlx4_warn(dev, "Failed to notify FW on FLR done (slave:%d)\n",
476 					  i);
477 		}
478 	}
479 }
480 
481 static int mlx4_eq_int(struct mlx4_dev *dev, struct mlx4_eq *eq)
482 {
483 	struct mlx4_priv *priv = mlx4_priv(dev);
484 	struct mlx4_eqe *eqe;
485 	int cqn = -1;
486 	int eqes_found = 0;
487 	int set_ci = 0;
488 	int port;
489 	int slave = 0;
490 	int ret;
491 	u32 flr_slave;
492 	u8 update_slave_state;
493 	int i;
494 	enum slave_port_gen_event gen_event;
495 	unsigned long flags;
496 	struct mlx4_vport_state *s_info;
497 	int eqe_size = dev->caps.eqe_size;
498 
499 	while ((eqe = next_eqe_sw(eq, dev->caps.eqe_factor, eqe_size))) {
500 		/*
501 		 * Make sure we read EQ entry contents after we've
502 		 * checked the ownership bit.
503 		 */
504 		dma_rmb();
505 
506 		switch (eqe->type) {
507 		case MLX4_EVENT_TYPE_COMP:
508 			cqn = be32_to_cpu(eqe->event.comp.cqn) & 0xffffff;
509 			mlx4_cq_completion(dev, cqn);
510 			break;
511 
512 		case MLX4_EVENT_TYPE_PATH_MIG:
513 		case MLX4_EVENT_TYPE_COMM_EST:
514 		case MLX4_EVENT_TYPE_SQ_DRAINED:
515 		case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
516 		case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
517 		case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
518 		case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
519 		case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
520 			mlx4_dbg(dev, "event %d arrived\n", eqe->type);
521 			if (mlx4_is_master(dev)) {
522 				/* forward only to slave owning the QP */
523 				ret = mlx4_get_slave_from_resource_id(dev,
524 						RES_QP,
525 						be32_to_cpu(eqe->event.qp.qpn)
526 						& 0xffffff, &slave);
527 				if (ret && ret != -ENOENT) {
528 					mlx4_dbg(dev, "QP event %02x(%02x) on EQ %d at index %u: could not get slave id (%d)\n",
529 						 eqe->type, eqe->subtype,
530 						 eq->eqn, eq->cons_index, ret);
531 					break;
532 				}
533 
534 				if (!ret && slave != dev->caps.function) {
535 					mlx4_slave_event(dev, slave, eqe);
536 					break;
537 				}
538 
539 			}
540 			mlx4_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) &
541 				      0xffffff, eqe->type);
542 			break;
543 
544 		case MLX4_EVENT_TYPE_SRQ_LIMIT:
545 			mlx4_dbg(dev, "%s: MLX4_EVENT_TYPE_SRQ_LIMIT\n",
546 				 __func__);
547 		case MLX4_EVENT_TYPE_SRQ_CATAS_ERROR:
548 			if (mlx4_is_master(dev)) {
549 				/* forward only to slave owning the SRQ */
550 				ret = mlx4_get_slave_from_resource_id(dev,
551 						RES_SRQ,
552 						be32_to_cpu(eqe->event.srq.srqn)
553 						& 0xffffff,
554 						&slave);
555 				if (ret && ret != -ENOENT) {
556 					mlx4_warn(dev, "SRQ event %02x(%02x) on EQ %d at index %u: could not get slave id (%d)\n",
557 						  eqe->type, eqe->subtype,
558 						  eq->eqn, eq->cons_index, ret);
559 					break;
560 				}
561 				mlx4_warn(dev, "%s: slave:%d, srq_no:0x%x, event: %02x(%02x)\n",
562 					  __func__, slave,
563 					  be32_to_cpu(eqe->event.srq.srqn),
564 					  eqe->type, eqe->subtype);
565 
566 				if (!ret && slave != dev->caps.function) {
567 					mlx4_warn(dev, "%s: sending event %02x(%02x) to slave:%d\n",
568 						  __func__, eqe->type,
569 						  eqe->subtype, slave);
570 					mlx4_slave_event(dev, slave, eqe);
571 					break;
572 				}
573 			}
574 			mlx4_srq_event(dev, be32_to_cpu(eqe->event.srq.srqn) &
575 				       0xffffff, eqe->type);
576 			break;
577 
578 		case MLX4_EVENT_TYPE_CMD:
579 			mlx4_cmd_event(dev,
580 				       be16_to_cpu(eqe->event.cmd.token),
581 				       eqe->event.cmd.status,
582 				       be64_to_cpu(eqe->event.cmd.out_param));
583 			break;
584 
585 		case MLX4_EVENT_TYPE_PORT_CHANGE: {
586 			struct mlx4_slaves_pport slaves_port;
587 			port = be32_to_cpu(eqe->event.port_change.port) >> 28;
588 			slaves_port = mlx4_phys_to_slaves_pport(dev, port);
589 			if (eqe->subtype == MLX4_PORT_CHANGE_SUBTYPE_DOWN) {
590 				mlx4_dispatch_event(dev, MLX4_DEV_EVENT_PORT_DOWN,
591 						    port);
592 				mlx4_priv(dev)->sense.do_sense_port[port] = 1;
593 				if (!mlx4_is_master(dev))
594 					break;
595 				for (i = 0; i < dev->persist->num_vfs + 1;
596 				     i++) {
597 					if (!test_bit(i, slaves_port.slaves))
598 						continue;
599 					if (dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH) {
600 						if (i == mlx4_master_func_num(dev))
601 							continue;
602 						mlx4_dbg(dev, "%s: Sending MLX4_PORT_CHANGE_SUBTYPE_DOWN to slave: %d, port:%d\n",
603 							 __func__, i, port);
604 						s_info = &priv->mfunc.master.vf_oper[i].vport[port].state;
605 						if (IFLA_VF_LINK_STATE_AUTO == s_info->link_state) {
606 							eqe->event.port_change.port =
607 								cpu_to_be32(
608 								(be32_to_cpu(eqe->event.port_change.port) & 0xFFFFFFF)
609 								| (mlx4_phys_to_slave_port(dev, i, port) << 28));
610 							mlx4_slave_event(dev, i, eqe);
611 						}
612 					} else {  /* IB port */
613 						set_and_calc_slave_port_state(dev, i, port,
614 									      MLX4_PORT_STATE_DEV_EVENT_PORT_DOWN,
615 									      &gen_event);
616 						/*we can be in pending state, then do not send port_down event*/
617 						if (SLAVE_PORT_GEN_EVENT_DOWN ==  gen_event) {
618 							if (i == mlx4_master_func_num(dev))
619 								continue;
620 							eqe->event.port_change.port =
621 								cpu_to_be32(
622 								(be32_to_cpu(eqe->event.port_change.port) & 0xFFFFFFF)
623 								| (mlx4_phys_to_slave_port(dev, i, port) << 28));
624 							mlx4_slave_event(dev, i, eqe);
625 						}
626 					}
627 				}
628 			} else {
629 				mlx4_dispatch_event(dev, MLX4_DEV_EVENT_PORT_UP, port);
630 
631 				mlx4_priv(dev)->sense.do_sense_port[port] = 0;
632 
633 				if (!mlx4_is_master(dev))
634 					break;
635 				if (dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH)
636 					for (i = 0;
637 					     i < dev->persist->num_vfs + 1;
638 					     i++) {
639 						if (!test_bit(i, slaves_port.slaves))
640 							continue;
641 						if (i == mlx4_master_func_num(dev))
642 							continue;
643 						s_info = &priv->mfunc.master.vf_oper[i].vport[port].state;
644 						if (IFLA_VF_LINK_STATE_AUTO == s_info->link_state) {
645 							eqe->event.port_change.port =
646 								cpu_to_be32(
647 								(be32_to_cpu(eqe->event.port_change.port) & 0xFFFFFFF)
648 								| (mlx4_phys_to_slave_port(dev, i, port) << 28));
649 							mlx4_slave_event(dev, i, eqe);
650 						}
651 					}
652 				else /* IB port */
653 					/* port-up event will be sent to a slave when the
654 					 * slave's alias-guid is set. This is done in alias_GUID.c
655 					 */
656 					set_all_slave_state(dev, port, MLX4_DEV_EVENT_PORT_UP);
657 			}
658 			break;
659 		}
660 
661 		case MLX4_EVENT_TYPE_CQ_ERROR:
662 			mlx4_warn(dev, "CQ %s on CQN %06x\n",
663 				  eqe->event.cq_err.syndrome == 1 ?
664 				  "overrun" : "access violation",
665 				  be32_to_cpu(eqe->event.cq_err.cqn) & 0xffffff);
666 			if (mlx4_is_master(dev)) {
667 				ret = mlx4_get_slave_from_resource_id(dev,
668 					RES_CQ,
669 					be32_to_cpu(eqe->event.cq_err.cqn)
670 					& 0xffffff, &slave);
671 				if (ret && ret != -ENOENT) {
672 					mlx4_dbg(dev, "CQ event %02x(%02x) on EQ %d at index %u: could not get slave id (%d)\n",
673 						 eqe->type, eqe->subtype,
674 						 eq->eqn, eq->cons_index, ret);
675 					break;
676 				}
677 
678 				if (!ret && slave != dev->caps.function) {
679 					mlx4_slave_event(dev, slave, eqe);
680 					break;
681 				}
682 			}
683 			mlx4_cq_event(dev,
684 				      be32_to_cpu(eqe->event.cq_err.cqn)
685 				      & 0xffffff,
686 				      eqe->type);
687 			break;
688 
689 		case MLX4_EVENT_TYPE_EQ_OVERFLOW:
690 			mlx4_warn(dev, "EQ overrun on EQN %d\n", eq->eqn);
691 			break;
692 
693 		case MLX4_EVENT_TYPE_OP_REQUIRED:
694 			atomic_inc(&priv->opreq_count);
695 			/* FW commands can't be executed from interrupt context
696 			 * working in deferred task
697 			 */
698 			queue_work(mlx4_wq, &priv->opreq_task);
699 			break;
700 
701 		case MLX4_EVENT_TYPE_COMM_CHANNEL:
702 			if (!mlx4_is_master(dev)) {
703 				mlx4_warn(dev, "Received comm channel event for non master device\n");
704 				break;
705 			}
706 			memcpy(&priv->mfunc.master.comm_arm_bit_vector,
707 			       eqe->event.comm_channel_arm.bit_vec,
708 			       sizeof eqe->event.comm_channel_arm.bit_vec);
709 			queue_work(priv->mfunc.master.comm_wq,
710 				   &priv->mfunc.master.comm_work);
711 			break;
712 
713 		case MLX4_EVENT_TYPE_FLR_EVENT:
714 			flr_slave = be32_to_cpu(eqe->event.flr_event.slave_id);
715 			if (!mlx4_is_master(dev)) {
716 				mlx4_warn(dev, "Non-master function received FLR event\n");
717 				break;
718 			}
719 
720 			mlx4_dbg(dev, "FLR event for slave: %d\n", flr_slave);
721 
722 			if (flr_slave >= dev->num_slaves) {
723 				mlx4_warn(dev,
724 					  "Got FLR for unknown function: %d\n",
725 					  flr_slave);
726 				update_slave_state = 0;
727 			} else
728 				update_slave_state = 1;
729 
730 			spin_lock_irqsave(&priv->mfunc.master.slave_state_lock, flags);
731 			if (update_slave_state) {
732 				priv->mfunc.master.slave_state[flr_slave].active = false;
733 				priv->mfunc.master.slave_state[flr_slave].last_cmd = MLX4_COMM_CMD_FLR;
734 				priv->mfunc.master.slave_state[flr_slave].is_slave_going_down = 1;
735 			}
736 			spin_unlock_irqrestore(&priv->mfunc.master.slave_state_lock, flags);
737 			mlx4_dispatch_event(dev, MLX4_DEV_EVENT_SLAVE_SHUTDOWN,
738 					    flr_slave);
739 			queue_work(priv->mfunc.master.comm_wq,
740 				   &priv->mfunc.master.slave_flr_event_work);
741 			break;
742 
743 		case MLX4_EVENT_TYPE_FATAL_WARNING:
744 			if (eqe->subtype == MLX4_FATAL_WARNING_SUBTYPE_WARMING) {
745 				if (mlx4_is_master(dev))
746 					for (i = 0; i < dev->num_slaves; i++) {
747 						mlx4_dbg(dev, "%s: Sending MLX4_FATAL_WARNING_SUBTYPE_WARMING to slave: %d\n",
748 							 __func__, i);
749 						if (i == dev->caps.function)
750 							continue;
751 						mlx4_slave_event(dev, i, eqe);
752 					}
753 				mlx4_err(dev, "Temperature Threshold was reached! Threshold: %d celsius degrees; Current Temperature: %d\n",
754 					 be16_to_cpu(eqe->event.warming.warning_threshold),
755 					 be16_to_cpu(eqe->event.warming.current_temperature));
756 			} else
757 				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",
758 					  eqe->type, eqe->subtype, eq->eqn,
759 					  eq->cons_index, eqe->owner, eq->nent,
760 					  eqe->slave_id,
761 					  !!(eqe->owner & 0x80) ^
762 					  !!(eq->cons_index & eq->nent) ? "HW" : "SW");
763 
764 			break;
765 
766 		case MLX4_EVENT_TYPE_PORT_MNG_CHG_EVENT:
767 			mlx4_dispatch_event(dev, MLX4_DEV_EVENT_PORT_MGMT_CHANGE,
768 					    (unsigned long) eqe);
769 			break;
770 
771 		case MLX4_EVENT_TYPE_RECOVERABLE_ERROR_EVENT:
772 			switch (eqe->subtype) {
773 			case MLX4_RECOVERABLE_ERROR_EVENT_SUBTYPE_BAD_CABLE:
774 				mlx4_warn(dev, "Bad cable detected on port %u\n",
775 					  eqe->event.bad_cable.port);
776 				break;
777 			case MLX4_RECOVERABLE_ERROR_EVENT_SUBTYPE_UNSUPPORTED_CABLE:
778 				mlx4_warn(dev, "Unsupported cable detected\n");
779 				break;
780 			default:
781 				mlx4_dbg(dev,
782 					 "Unhandled recoverable error event detected: %02x(%02x) on EQ %d at index %u. owner=%x, nent=0x%x, ownership=%s\n",
783 					 eqe->type, eqe->subtype, eq->eqn,
784 					 eq->cons_index, eqe->owner, eq->nent,
785 					 !!(eqe->owner & 0x80) ^
786 					 !!(eq->cons_index & eq->nent) ? "HW" : "SW");
787 				break;
788 			}
789 			break;
790 
791 		case MLX4_EVENT_TYPE_EEC_CATAS_ERROR:
792 		case MLX4_EVENT_TYPE_ECC_DETECT:
793 		default:
794 			mlx4_warn(dev, "Unhandled event %02x(%02x) on EQ %d at index %u. owner=%x, nent=0x%x, slave=%x, ownership=%s\n",
795 				  eqe->type, eqe->subtype, eq->eqn,
796 				  eq->cons_index, eqe->owner, eq->nent,
797 				  eqe->slave_id,
798 				  !!(eqe->owner & 0x80) ^
799 				  !!(eq->cons_index & eq->nent) ? "HW" : "SW");
800 			break;
801 		};
802 
803 		++eq->cons_index;
804 		eqes_found = 1;
805 		++set_ci;
806 
807 		/*
808 		 * The HCA will think the queue has overflowed if we
809 		 * don't tell it we've been processing events.  We
810 		 * create our EQs with MLX4_NUM_SPARE_EQE extra
811 		 * entries, so we must update our consumer index at
812 		 * least that often.
813 		 */
814 		if (unlikely(set_ci >= MLX4_NUM_SPARE_EQE)) {
815 			eq_set_ci(eq, 0);
816 			set_ci = 0;
817 		}
818 	}
819 
820 	eq_set_ci(eq, 1);
821 
822 	/* cqn is 24bit wide but is initialized such that its higher bits
823 	 * are ones too. Thus, if we got any event, cqn's high bits should be off
824 	 * and we need to schedule the tasklet.
825 	 */
826 	if (!(cqn & ~0xffffff))
827 		tasklet_schedule(&eq->tasklet_ctx.task);
828 
829 	return eqes_found;
830 }
831 
832 static irqreturn_t mlx4_interrupt(int irq, void *dev_ptr)
833 {
834 	struct mlx4_dev *dev = dev_ptr;
835 	struct mlx4_priv *priv = mlx4_priv(dev);
836 	int work = 0;
837 	int i;
838 
839 	writel(priv->eq_table.clr_mask, priv->eq_table.clr_int);
840 
841 	for (i = 0; i < dev->caps.num_comp_vectors + 1; ++i)
842 		work |= mlx4_eq_int(dev, &priv->eq_table.eq[i]);
843 
844 	return IRQ_RETVAL(work);
845 }
846 
847 static irqreturn_t mlx4_msi_x_interrupt(int irq, void *eq_ptr)
848 {
849 	struct mlx4_eq  *eq  = eq_ptr;
850 	struct mlx4_dev *dev = eq->dev;
851 
852 	mlx4_eq_int(dev, eq);
853 
854 	/* MSI-X vectors always belong to us */
855 	return IRQ_HANDLED;
856 }
857 
858 int mlx4_MAP_EQ_wrapper(struct mlx4_dev *dev, int slave,
859 			struct mlx4_vhcr *vhcr,
860 			struct mlx4_cmd_mailbox *inbox,
861 			struct mlx4_cmd_mailbox *outbox,
862 			struct mlx4_cmd_info *cmd)
863 {
864 	struct mlx4_priv *priv = mlx4_priv(dev);
865 	struct mlx4_slave_event_eq_info *event_eq =
866 		priv->mfunc.master.slave_state[slave].event_eq;
867 	u32 in_modifier = vhcr->in_modifier;
868 	u32 eqn = in_modifier & 0x3FF;
869 	u64 in_param =  vhcr->in_param;
870 	int err = 0;
871 	int i;
872 
873 	if (slave == dev->caps.function)
874 		err = mlx4_cmd(dev, in_param, (in_modifier & 0x80000000) | eqn,
875 			       0, MLX4_CMD_MAP_EQ, MLX4_CMD_TIME_CLASS_B,
876 			       MLX4_CMD_NATIVE);
877 	if (!err)
878 		for (i = 0; i < MLX4_EVENT_TYPES_NUM; ++i)
879 			if (in_param & (1LL << i))
880 				event_eq[i].eqn = in_modifier >> 31 ? -1 : eqn;
881 
882 	return err;
883 }
884 
885 static int mlx4_MAP_EQ(struct mlx4_dev *dev, u64 event_mask, int unmap,
886 			int eq_num)
887 {
888 	return mlx4_cmd(dev, event_mask, (unmap << 31) | eq_num,
889 			0, MLX4_CMD_MAP_EQ, MLX4_CMD_TIME_CLASS_B,
890 			MLX4_CMD_WRAPPED);
891 }
892 
893 static int mlx4_SW2HW_EQ(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
894 			 int eq_num)
895 {
896 	return mlx4_cmd(dev, mailbox->dma, eq_num, 0,
897 			MLX4_CMD_SW2HW_EQ, MLX4_CMD_TIME_CLASS_A,
898 			MLX4_CMD_WRAPPED);
899 }
900 
901 static int mlx4_HW2SW_EQ(struct mlx4_dev *dev,  int eq_num)
902 {
903 	return mlx4_cmd(dev, 0, eq_num, 1, MLX4_CMD_HW2SW_EQ,
904 			MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED);
905 }
906 
907 static int mlx4_num_eq_uar(struct mlx4_dev *dev)
908 {
909 	/*
910 	 * Each UAR holds 4 EQ doorbells.  To figure out how many UARs
911 	 * we need to map, take the difference of highest index and
912 	 * the lowest index we'll use and add 1.
913 	 */
914 	return (dev->caps.num_comp_vectors + 1 + dev->caps.reserved_eqs) / 4 -
915 		dev->caps.reserved_eqs / 4 + 1;
916 }
917 
918 static void __iomem *mlx4_get_eq_uar(struct mlx4_dev *dev, struct mlx4_eq *eq)
919 {
920 	struct mlx4_priv *priv = mlx4_priv(dev);
921 	int index;
922 
923 	index = eq->eqn / 4 - dev->caps.reserved_eqs / 4;
924 
925 	if (!priv->eq_table.uar_map[index]) {
926 		priv->eq_table.uar_map[index] =
927 			ioremap(pci_resource_start(dev->persist->pdev, 2) +
928 				((eq->eqn / 4) << PAGE_SHIFT),
929 				PAGE_SIZE);
930 		if (!priv->eq_table.uar_map[index]) {
931 			mlx4_err(dev, "Couldn't map EQ doorbell for EQN 0x%06x\n",
932 				 eq->eqn);
933 			return NULL;
934 		}
935 	}
936 
937 	return priv->eq_table.uar_map[index] + 0x800 + 8 * (eq->eqn % 4);
938 }
939 
940 static void mlx4_unmap_uar(struct mlx4_dev *dev)
941 {
942 	struct mlx4_priv *priv = mlx4_priv(dev);
943 	int i;
944 
945 	for (i = 0; i < mlx4_num_eq_uar(dev); ++i)
946 		if (priv->eq_table.uar_map[i]) {
947 			iounmap(priv->eq_table.uar_map[i]);
948 			priv->eq_table.uar_map[i] = NULL;
949 		}
950 }
951 
952 static int mlx4_create_eq(struct mlx4_dev *dev, int nent,
953 			  u8 intr, struct mlx4_eq *eq)
954 {
955 	struct mlx4_priv *priv = mlx4_priv(dev);
956 	struct mlx4_cmd_mailbox *mailbox;
957 	struct mlx4_eq_context *eq_context;
958 	int npages;
959 	u64 *dma_list = NULL;
960 	dma_addr_t t;
961 	u64 mtt_addr;
962 	int err = -ENOMEM;
963 	int i;
964 
965 	eq->dev   = dev;
966 	eq->nent  = roundup_pow_of_two(max(nent, 2));
967 	/* CX3 is capable of extending the CQE/EQE from 32 to 64 bytes, with
968 	 * strides of 64B,128B and 256B.
969 	 */
970 	npages = PAGE_ALIGN(eq->nent * dev->caps.eqe_size) / PAGE_SIZE;
971 
972 	eq->page_list = kmalloc(npages * sizeof *eq->page_list,
973 				GFP_KERNEL);
974 	if (!eq->page_list)
975 		goto err_out;
976 
977 	for (i = 0; i < npages; ++i)
978 		eq->page_list[i].buf = NULL;
979 
980 	dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
981 	if (!dma_list)
982 		goto err_out_free;
983 
984 	mailbox = mlx4_alloc_cmd_mailbox(dev);
985 	if (IS_ERR(mailbox))
986 		goto err_out_free;
987 	eq_context = mailbox->buf;
988 
989 	for (i = 0; i < npages; ++i) {
990 		eq->page_list[i].buf = dma_alloc_coherent(&dev->persist->
991 							  pdev->dev,
992 							  PAGE_SIZE, &t,
993 							  GFP_KERNEL);
994 		if (!eq->page_list[i].buf)
995 			goto err_out_free_pages;
996 
997 		dma_list[i] = t;
998 		eq->page_list[i].map = t;
999 
1000 		memset(eq->page_list[i].buf, 0, PAGE_SIZE);
1001 	}
1002 
1003 	eq->eqn = mlx4_bitmap_alloc(&priv->eq_table.bitmap);
1004 	if (eq->eqn == -1)
1005 		goto err_out_free_pages;
1006 
1007 	eq->doorbell = mlx4_get_eq_uar(dev, eq);
1008 	if (!eq->doorbell) {
1009 		err = -ENOMEM;
1010 		goto err_out_free_eq;
1011 	}
1012 
1013 	err = mlx4_mtt_init(dev, npages, PAGE_SHIFT, &eq->mtt);
1014 	if (err)
1015 		goto err_out_free_eq;
1016 
1017 	err = mlx4_write_mtt(dev, &eq->mtt, 0, npages, dma_list);
1018 	if (err)
1019 		goto err_out_free_mtt;
1020 
1021 	eq_context->flags	  = cpu_to_be32(MLX4_EQ_STATUS_OK   |
1022 						MLX4_EQ_STATE_ARMED);
1023 	eq_context->log_eq_size	  = ilog2(eq->nent);
1024 	eq_context->intr	  = intr;
1025 	eq_context->log_page_size = PAGE_SHIFT - MLX4_ICM_PAGE_SHIFT;
1026 
1027 	mtt_addr = mlx4_mtt_addr(dev, &eq->mtt);
1028 	eq_context->mtt_base_addr_h = mtt_addr >> 32;
1029 	eq_context->mtt_base_addr_l = cpu_to_be32(mtt_addr & 0xffffffff);
1030 
1031 	err = mlx4_SW2HW_EQ(dev, mailbox, eq->eqn);
1032 	if (err) {
1033 		mlx4_warn(dev, "SW2HW_EQ failed (%d)\n", err);
1034 		goto err_out_free_mtt;
1035 	}
1036 
1037 	kfree(dma_list);
1038 	mlx4_free_cmd_mailbox(dev, mailbox);
1039 
1040 	eq->cons_index = 0;
1041 
1042 	INIT_LIST_HEAD(&eq->tasklet_ctx.list);
1043 	INIT_LIST_HEAD(&eq->tasklet_ctx.process_list);
1044 	spin_lock_init(&eq->tasklet_ctx.lock);
1045 	tasklet_init(&eq->tasklet_ctx.task, mlx4_cq_tasklet_cb,
1046 		     (unsigned long)&eq->tasklet_ctx);
1047 
1048 	return err;
1049 
1050 err_out_free_mtt:
1051 	mlx4_mtt_cleanup(dev, &eq->mtt);
1052 
1053 err_out_free_eq:
1054 	mlx4_bitmap_free(&priv->eq_table.bitmap, eq->eqn, MLX4_USE_RR);
1055 
1056 err_out_free_pages:
1057 	for (i = 0; i < npages; ++i)
1058 		if (eq->page_list[i].buf)
1059 			dma_free_coherent(&dev->persist->pdev->dev, PAGE_SIZE,
1060 					  eq->page_list[i].buf,
1061 					  eq->page_list[i].map);
1062 
1063 	mlx4_free_cmd_mailbox(dev, mailbox);
1064 
1065 err_out_free:
1066 	kfree(eq->page_list);
1067 	kfree(dma_list);
1068 
1069 err_out:
1070 	return err;
1071 }
1072 
1073 static void mlx4_free_eq(struct mlx4_dev *dev,
1074 			 struct mlx4_eq *eq)
1075 {
1076 	struct mlx4_priv *priv = mlx4_priv(dev);
1077 	int err;
1078 	int i;
1079 	/* CX3 is capable of extending the CQE/EQE from 32 to 64 bytes, with
1080 	 * strides of 64B,128B and 256B
1081 	 */
1082 	int npages = PAGE_ALIGN(dev->caps.eqe_size  * eq->nent) / PAGE_SIZE;
1083 
1084 	err = mlx4_HW2SW_EQ(dev, eq->eqn);
1085 	if (err)
1086 		mlx4_warn(dev, "HW2SW_EQ failed (%d)\n", err);
1087 
1088 	synchronize_irq(eq->irq);
1089 	tasklet_disable(&eq->tasklet_ctx.task);
1090 
1091 	mlx4_mtt_cleanup(dev, &eq->mtt);
1092 	for (i = 0; i < npages; ++i)
1093 		dma_free_coherent(&dev->persist->pdev->dev, PAGE_SIZE,
1094 				  eq->page_list[i].buf,
1095 				  eq->page_list[i].map);
1096 
1097 	kfree(eq->page_list);
1098 	mlx4_bitmap_free(&priv->eq_table.bitmap, eq->eqn, MLX4_USE_RR);
1099 }
1100 
1101 static void mlx4_free_irqs(struct mlx4_dev *dev)
1102 {
1103 	struct mlx4_eq_table *eq_table = &mlx4_priv(dev)->eq_table;
1104 	int	i;
1105 
1106 	if (eq_table->have_irq)
1107 		free_irq(dev->persist->pdev->irq, dev);
1108 
1109 	for (i = 0; i < dev->caps.num_comp_vectors + 1; ++i)
1110 		if (eq_table->eq[i].have_irq) {
1111 			free_cpumask_var(eq_table->eq[i].affinity_mask);
1112 #if defined(CONFIG_SMP)
1113 			irq_set_affinity_hint(eq_table->eq[i].irq, NULL);
1114 #endif
1115 			free_irq(eq_table->eq[i].irq, eq_table->eq + i);
1116 			eq_table->eq[i].have_irq = 0;
1117 		}
1118 
1119 	kfree(eq_table->irq_names);
1120 }
1121 
1122 static int mlx4_map_clr_int(struct mlx4_dev *dev)
1123 {
1124 	struct mlx4_priv *priv = mlx4_priv(dev);
1125 
1126 	priv->clr_base = ioremap(pci_resource_start(dev->persist->pdev,
1127 				 priv->fw.clr_int_bar) +
1128 				 priv->fw.clr_int_base, MLX4_CLR_INT_SIZE);
1129 	if (!priv->clr_base) {
1130 		mlx4_err(dev, "Couldn't map interrupt clear register, aborting\n");
1131 		return -ENOMEM;
1132 	}
1133 
1134 	return 0;
1135 }
1136 
1137 static void mlx4_unmap_clr_int(struct mlx4_dev *dev)
1138 {
1139 	struct mlx4_priv *priv = mlx4_priv(dev);
1140 
1141 	iounmap(priv->clr_base);
1142 }
1143 
1144 int mlx4_alloc_eq_table(struct mlx4_dev *dev)
1145 {
1146 	struct mlx4_priv *priv = mlx4_priv(dev);
1147 
1148 	priv->eq_table.eq = kcalloc(dev->caps.num_eqs - dev->caps.reserved_eqs,
1149 				    sizeof *priv->eq_table.eq, GFP_KERNEL);
1150 	if (!priv->eq_table.eq)
1151 		return -ENOMEM;
1152 
1153 	return 0;
1154 }
1155 
1156 void mlx4_free_eq_table(struct mlx4_dev *dev)
1157 {
1158 	kfree(mlx4_priv(dev)->eq_table.eq);
1159 }
1160 
1161 int mlx4_init_eq_table(struct mlx4_dev *dev)
1162 {
1163 	struct mlx4_priv *priv = mlx4_priv(dev);
1164 	int err;
1165 	int i;
1166 
1167 	priv->eq_table.uar_map = kcalloc(mlx4_num_eq_uar(dev),
1168 					 sizeof *priv->eq_table.uar_map,
1169 					 GFP_KERNEL);
1170 	if (!priv->eq_table.uar_map) {
1171 		err = -ENOMEM;
1172 		goto err_out_free;
1173 	}
1174 
1175 	err = mlx4_bitmap_init(&priv->eq_table.bitmap,
1176 			       roundup_pow_of_two(dev->caps.num_eqs),
1177 			       dev->caps.num_eqs - 1,
1178 			       dev->caps.reserved_eqs,
1179 			       roundup_pow_of_two(dev->caps.num_eqs) -
1180 			       dev->caps.num_eqs);
1181 	if (err)
1182 		goto err_out_free;
1183 
1184 	for (i = 0; i < mlx4_num_eq_uar(dev); ++i)
1185 		priv->eq_table.uar_map[i] = NULL;
1186 
1187 	if (!mlx4_is_slave(dev)) {
1188 		err = mlx4_map_clr_int(dev);
1189 		if (err)
1190 			goto err_out_bitmap;
1191 
1192 		priv->eq_table.clr_mask =
1193 			swab32(1 << (priv->eq_table.inta_pin & 31));
1194 		priv->eq_table.clr_int  = priv->clr_base +
1195 			(priv->eq_table.inta_pin < 32 ? 4 : 0);
1196 	}
1197 
1198 	priv->eq_table.irq_names =
1199 		kmalloc(MLX4_IRQNAME_SIZE * (dev->caps.num_comp_vectors + 1),
1200 			GFP_KERNEL);
1201 	if (!priv->eq_table.irq_names) {
1202 		err = -ENOMEM;
1203 		goto err_out_clr_int;
1204 	}
1205 
1206 	for (i = 0; i < dev->caps.num_comp_vectors + 1; ++i) {
1207 		if (i == MLX4_EQ_ASYNC) {
1208 			err = mlx4_create_eq(dev,
1209 					     MLX4_NUM_ASYNC_EQE + MLX4_NUM_SPARE_EQE,
1210 					     0, &priv->eq_table.eq[MLX4_EQ_ASYNC]);
1211 		} else {
1212 			struct mlx4_eq	*eq = &priv->eq_table.eq[i];
1213 #ifdef CONFIG_RFS_ACCEL
1214 			int port = find_first_bit(eq->actv_ports.ports,
1215 						  dev->caps.num_ports) + 1;
1216 
1217 			if (port <= dev->caps.num_ports) {
1218 				struct mlx4_port_info *info =
1219 					&mlx4_priv(dev)->port[port];
1220 
1221 				if (!info->rmap) {
1222 					info->rmap = alloc_irq_cpu_rmap(
1223 						mlx4_get_eqs_per_port(dev, port));
1224 					if (!info->rmap) {
1225 						mlx4_warn(dev, "Failed to allocate cpu rmap\n");
1226 						err = -ENOMEM;
1227 						goto err_out_unmap;
1228 					}
1229 				}
1230 
1231 				err = irq_cpu_rmap_add(
1232 					info->rmap, eq->irq);
1233 				if (err)
1234 					mlx4_warn(dev, "Failed adding irq rmap\n");
1235 			}
1236 #endif
1237 			err = mlx4_create_eq(dev, dev->caps.num_cqs -
1238 						  dev->caps.reserved_cqs +
1239 						  MLX4_NUM_SPARE_EQE,
1240 					     (dev->flags & MLX4_FLAG_MSI_X) ?
1241 					     i + 1 - !!(i > MLX4_EQ_ASYNC) : 0,
1242 					     eq);
1243 		}
1244 		if (err)
1245 			goto err_out_unmap;
1246 	}
1247 
1248 	if (dev->flags & MLX4_FLAG_MSI_X) {
1249 		const char *eq_name;
1250 
1251 		snprintf(priv->eq_table.irq_names +
1252 			 MLX4_EQ_ASYNC * MLX4_IRQNAME_SIZE,
1253 			 MLX4_IRQNAME_SIZE,
1254 			 "mlx4-async@pci:%s",
1255 			 pci_name(dev->persist->pdev));
1256 		eq_name = priv->eq_table.irq_names +
1257 			MLX4_EQ_ASYNC * MLX4_IRQNAME_SIZE;
1258 
1259 		err = request_irq(priv->eq_table.eq[MLX4_EQ_ASYNC].irq,
1260 				  mlx4_msi_x_interrupt, 0, eq_name,
1261 				  priv->eq_table.eq + MLX4_EQ_ASYNC);
1262 		if (err)
1263 			goto err_out_unmap;
1264 
1265 		priv->eq_table.eq[MLX4_EQ_ASYNC].have_irq = 1;
1266 	} else {
1267 		snprintf(priv->eq_table.irq_names,
1268 			 MLX4_IRQNAME_SIZE,
1269 			 DRV_NAME "@pci:%s",
1270 			 pci_name(dev->persist->pdev));
1271 		err = request_irq(dev->persist->pdev->irq, mlx4_interrupt,
1272 				  IRQF_SHARED, priv->eq_table.irq_names, dev);
1273 		if (err)
1274 			goto err_out_unmap;
1275 
1276 		priv->eq_table.have_irq = 1;
1277 	}
1278 
1279 	err = mlx4_MAP_EQ(dev, get_async_ev_mask(dev), 0,
1280 			  priv->eq_table.eq[MLX4_EQ_ASYNC].eqn);
1281 	if (err)
1282 		mlx4_warn(dev, "MAP_EQ for async EQ %d failed (%d)\n",
1283 			   priv->eq_table.eq[MLX4_EQ_ASYNC].eqn, err);
1284 
1285 	/* arm ASYNC eq */
1286 	eq_set_ci(&priv->eq_table.eq[MLX4_EQ_ASYNC], 1);
1287 
1288 	return 0;
1289 
1290 err_out_unmap:
1291 	while (i >= 0)
1292 		mlx4_free_eq(dev, &priv->eq_table.eq[i--]);
1293 #ifdef CONFIG_RFS_ACCEL
1294 	for (i = 1; i <= dev->caps.num_ports; i++) {
1295 		if (mlx4_priv(dev)->port[i].rmap) {
1296 			free_irq_cpu_rmap(mlx4_priv(dev)->port[i].rmap);
1297 			mlx4_priv(dev)->port[i].rmap = NULL;
1298 		}
1299 	}
1300 #endif
1301 	mlx4_free_irqs(dev);
1302 
1303 err_out_clr_int:
1304 	if (!mlx4_is_slave(dev))
1305 		mlx4_unmap_clr_int(dev);
1306 
1307 err_out_bitmap:
1308 	mlx4_unmap_uar(dev);
1309 	mlx4_bitmap_cleanup(&priv->eq_table.bitmap);
1310 
1311 err_out_free:
1312 	kfree(priv->eq_table.uar_map);
1313 
1314 	return err;
1315 }
1316 
1317 void mlx4_cleanup_eq_table(struct mlx4_dev *dev)
1318 {
1319 	struct mlx4_priv *priv = mlx4_priv(dev);
1320 	int i;
1321 
1322 	mlx4_MAP_EQ(dev, get_async_ev_mask(dev), 1,
1323 		    priv->eq_table.eq[MLX4_EQ_ASYNC].eqn);
1324 
1325 #ifdef CONFIG_RFS_ACCEL
1326 	for (i = 1; i <= dev->caps.num_ports; i++) {
1327 		if (mlx4_priv(dev)->port[i].rmap) {
1328 			free_irq_cpu_rmap(mlx4_priv(dev)->port[i].rmap);
1329 			mlx4_priv(dev)->port[i].rmap = NULL;
1330 		}
1331 	}
1332 #endif
1333 	mlx4_free_irqs(dev);
1334 
1335 	for (i = 0; i < dev->caps.num_comp_vectors + 1; ++i)
1336 		mlx4_free_eq(dev, &priv->eq_table.eq[i]);
1337 
1338 	if (!mlx4_is_slave(dev))
1339 		mlx4_unmap_clr_int(dev);
1340 
1341 	mlx4_unmap_uar(dev);
1342 	mlx4_bitmap_cleanup(&priv->eq_table.bitmap);
1343 
1344 	kfree(priv->eq_table.uar_map);
1345 }
1346 
1347 /* A test that verifies that we can accept interrupts on all
1348  * the irq vectors of the device.
1349  * Interrupts are checked using the NOP command.
1350  */
1351 int mlx4_test_interrupts(struct mlx4_dev *dev)
1352 {
1353 	struct mlx4_priv *priv = mlx4_priv(dev);
1354 	int i;
1355 	int err;
1356 
1357 	err = mlx4_NOP(dev);
1358 	/* When not in MSI_X, there is only one irq to check */
1359 	if (!(dev->flags & MLX4_FLAG_MSI_X) || mlx4_is_slave(dev))
1360 		return err;
1361 
1362 	/* A loop over all completion vectors, for each vector we will check
1363 	 * whether it works by mapping command completions to that vector
1364 	 * and performing a NOP command
1365 	 */
1366 	for(i = 0; !err && (i < dev->caps.num_comp_vectors); ++i) {
1367 		/* Make sure request_irq was called */
1368 		if (!priv->eq_table.eq[i].have_irq)
1369 			continue;
1370 
1371 		/* Temporary use polling for command completions */
1372 		mlx4_cmd_use_polling(dev);
1373 
1374 		/* Map the new eq to handle all asynchronous events */
1375 		err = mlx4_MAP_EQ(dev, get_async_ev_mask(dev), 0,
1376 				  priv->eq_table.eq[i].eqn);
1377 		if (err) {
1378 			mlx4_warn(dev, "Failed mapping eq for interrupt test\n");
1379 			mlx4_cmd_use_events(dev);
1380 			break;
1381 		}
1382 
1383 		/* Go back to using events */
1384 		mlx4_cmd_use_events(dev);
1385 		err = mlx4_NOP(dev);
1386 	}
1387 
1388 	/* Return to default */
1389 	mlx4_MAP_EQ(dev, get_async_ev_mask(dev), 0,
1390 		    priv->eq_table.eq[MLX4_EQ_ASYNC].eqn);
1391 	return err;
1392 }
1393 EXPORT_SYMBOL(mlx4_test_interrupts);
1394 
1395 bool mlx4_is_eq_vector_valid(struct mlx4_dev *dev, u8 port, int vector)
1396 {
1397 	struct mlx4_priv *priv = mlx4_priv(dev);
1398 
1399 	vector = MLX4_CQ_TO_EQ_VECTOR(vector);
1400 	if (vector < 0 || (vector >= dev->caps.num_comp_vectors + 1) ||
1401 	    (vector == MLX4_EQ_ASYNC))
1402 		return false;
1403 
1404 	return test_bit(port - 1, priv->eq_table.eq[vector].actv_ports.ports);
1405 }
1406 EXPORT_SYMBOL(mlx4_is_eq_vector_valid);
1407 
1408 u32 mlx4_get_eqs_per_port(struct mlx4_dev *dev, u8 port)
1409 {
1410 	struct mlx4_priv *priv = mlx4_priv(dev);
1411 	unsigned int i;
1412 	unsigned int sum = 0;
1413 
1414 	for (i = 0; i < dev->caps.num_comp_vectors + 1; i++)
1415 		sum += !!test_bit(port - 1,
1416 				  priv->eq_table.eq[i].actv_ports.ports);
1417 
1418 	return sum;
1419 }
1420 EXPORT_SYMBOL(mlx4_get_eqs_per_port);
1421 
1422 int mlx4_is_eq_shared(struct mlx4_dev *dev, int vector)
1423 {
1424 	struct mlx4_priv *priv = mlx4_priv(dev);
1425 
1426 	vector = MLX4_CQ_TO_EQ_VECTOR(vector);
1427 	if (vector <= 0 || (vector >= dev->caps.num_comp_vectors + 1))
1428 		return -EINVAL;
1429 
1430 	return !!(bitmap_weight(priv->eq_table.eq[vector].actv_ports.ports,
1431 				dev->caps.num_ports) > 1);
1432 }
1433 EXPORT_SYMBOL(mlx4_is_eq_shared);
1434 
1435 struct cpu_rmap *mlx4_get_cpu_rmap(struct mlx4_dev *dev, int port)
1436 {
1437 	return mlx4_priv(dev)->port[port].rmap;
1438 }
1439 EXPORT_SYMBOL(mlx4_get_cpu_rmap);
1440 
1441 int mlx4_assign_eq(struct mlx4_dev *dev, u8 port, int *vector)
1442 {
1443 	struct mlx4_priv *priv = mlx4_priv(dev);
1444 	int err = 0, i = 0;
1445 	u32 min_ref_count_val = (u32)-1;
1446 	int requested_vector = MLX4_CQ_TO_EQ_VECTOR(*vector);
1447 	int *prequested_vector = NULL;
1448 
1449 
1450 	mutex_lock(&priv->msix_ctl.pool_lock);
1451 	if (requested_vector < (dev->caps.num_comp_vectors + 1) &&
1452 	    (requested_vector >= 0) &&
1453 	    (requested_vector != MLX4_EQ_ASYNC)) {
1454 		if (test_bit(port - 1,
1455 			     priv->eq_table.eq[requested_vector].actv_ports.ports)) {
1456 			prequested_vector = &requested_vector;
1457 		} else {
1458 			struct mlx4_eq *eq;
1459 
1460 			for (i = 1; i < port;
1461 			     requested_vector += mlx4_get_eqs_per_port(dev, i++))
1462 				;
1463 
1464 			eq = &priv->eq_table.eq[requested_vector];
1465 			if (requested_vector < dev->caps.num_comp_vectors + 1 &&
1466 			    test_bit(port - 1, eq->actv_ports.ports)) {
1467 				prequested_vector = &requested_vector;
1468 			}
1469 		}
1470 	}
1471 
1472 	if  (!prequested_vector) {
1473 		requested_vector = -1;
1474 		for (i = 0; min_ref_count_val && i < dev->caps.num_comp_vectors + 1;
1475 		     i++) {
1476 			struct mlx4_eq *eq = &priv->eq_table.eq[i];
1477 
1478 			if (min_ref_count_val > eq->ref_count &&
1479 			    test_bit(port - 1, eq->actv_ports.ports)) {
1480 				min_ref_count_val = eq->ref_count;
1481 				requested_vector = i;
1482 			}
1483 		}
1484 
1485 		if (requested_vector < 0) {
1486 			err = -ENOSPC;
1487 			goto err_unlock;
1488 		}
1489 
1490 		prequested_vector = &requested_vector;
1491 	}
1492 
1493 	if (!test_bit(*prequested_vector, priv->msix_ctl.pool_bm) &&
1494 	    dev->flags & MLX4_FLAG_MSI_X) {
1495 		set_bit(*prequested_vector, priv->msix_ctl.pool_bm);
1496 		snprintf(priv->eq_table.irq_names +
1497 			 *prequested_vector * MLX4_IRQNAME_SIZE,
1498 			 MLX4_IRQNAME_SIZE, "mlx4-%d@%s",
1499 			 *prequested_vector, dev_name(&dev->persist->pdev->dev));
1500 
1501 		err = request_irq(priv->eq_table.eq[*prequested_vector].irq,
1502 				  mlx4_msi_x_interrupt, 0,
1503 				  &priv->eq_table.irq_names[*prequested_vector << 5],
1504 				  priv->eq_table.eq + *prequested_vector);
1505 
1506 		if (err) {
1507 			clear_bit(*prequested_vector, priv->msix_ctl.pool_bm);
1508 			*prequested_vector = -1;
1509 		} else {
1510 #if defined(CONFIG_SMP)
1511 			mlx4_set_eq_affinity_hint(priv, *prequested_vector);
1512 #endif
1513 			eq_set_ci(&priv->eq_table.eq[*prequested_vector], 1);
1514 			priv->eq_table.eq[*prequested_vector].have_irq = 1;
1515 		}
1516 	}
1517 
1518 	if (!err && *prequested_vector >= 0)
1519 		priv->eq_table.eq[*prequested_vector].ref_count++;
1520 
1521 err_unlock:
1522 	mutex_unlock(&priv->msix_ctl.pool_lock);
1523 
1524 	if (!err && *prequested_vector >= 0)
1525 		*vector = MLX4_EQ_TO_CQ_VECTOR(*prequested_vector);
1526 	else
1527 		*vector = 0;
1528 
1529 	return err;
1530 }
1531 EXPORT_SYMBOL(mlx4_assign_eq);
1532 
1533 int mlx4_eq_get_irq(struct mlx4_dev *dev, int cq_vec)
1534 {
1535 	struct mlx4_priv *priv = mlx4_priv(dev);
1536 
1537 	return priv->eq_table.eq[MLX4_CQ_TO_EQ_VECTOR(cq_vec)].irq;
1538 }
1539 EXPORT_SYMBOL(mlx4_eq_get_irq);
1540 
1541 void mlx4_release_eq(struct mlx4_dev *dev, int vec)
1542 {
1543 	struct mlx4_priv *priv = mlx4_priv(dev);
1544 	int eq_vec = MLX4_CQ_TO_EQ_VECTOR(vec);
1545 
1546 	mutex_lock(&priv->msix_ctl.pool_lock);
1547 	priv->eq_table.eq[eq_vec].ref_count--;
1548 
1549 	/* once we allocated EQ, we don't release it because it might be binded
1550 	 * to cpu_rmap.
1551 	 */
1552 	mutex_unlock(&priv->msix_ctl.pool_lock);
1553 }
1554 EXPORT_SYMBOL(mlx4_release_eq);
1555 
1556