1 /*
2 * Copyright (c) 2005, 2006, 2007, 2008 Mellanox Technologies. All rights reserved.
3 * Copyright (c) 2005, 2006, 2007 Cisco Systems, Inc. All rights reserved.
4 *
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
10 *
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
13 * conditions are met:
14 *
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
17 * disclaimer.
18 *
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
23 *
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31 * SOFTWARE.
32 */
33
34 #include <linux/init.h>
35 #include <linux/interrupt.h>
36 #include <linux/slab.h>
37 #include <linux/export.h>
38 #include <linux/mm.h>
39 #include <linux/dma-mapping.h>
40
41 #include <linux/mlx4/cmd.h>
42
43 #include "mlx4.h"
44 #include "fw.h"
45
46 enum {
47 MLX4_IRQNAME_SIZE = 32
48 };
49
50 enum {
51 MLX4_NUM_ASYNC_EQE = 0x100,
52 MLX4_NUM_SPARE_EQE = 0x80,
53 MLX4_EQ_ENTRY_SIZE = 0x20
54 };
55
56 #define MLX4_EQ_STATUS_OK ( 0 << 28)
57 #define MLX4_EQ_STATUS_WRITE_FAIL (10 << 28)
58 #define MLX4_EQ_OWNER_SW ( 0 << 24)
59 #define MLX4_EQ_OWNER_HW ( 1 << 24)
60 #define MLX4_EQ_FLAG_EC ( 1 << 18)
61 #define MLX4_EQ_FLAG_OI ( 1 << 17)
62 #define MLX4_EQ_STATE_ARMED ( 9 << 8)
63 #define MLX4_EQ_STATE_FIRED (10 << 8)
64 #define MLX4_EQ_STATE_ALWAYS_ARMED (11 << 8)
65
66 #define MLX4_ASYNC_EVENT_MASK ((1ull << MLX4_EVENT_TYPE_PATH_MIG) | \
67 (1ull << MLX4_EVENT_TYPE_COMM_EST) | \
68 (1ull << MLX4_EVENT_TYPE_SQ_DRAINED) | \
69 (1ull << MLX4_EVENT_TYPE_CQ_ERROR) | \
70 (1ull << MLX4_EVENT_TYPE_WQ_CATAS_ERROR) | \
71 (1ull << MLX4_EVENT_TYPE_EEC_CATAS_ERROR) | \
72 (1ull << MLX4_EVENT_TYPE_PATH_MIG_FAILED) | \
73 (1ull << MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR) | \
74 (1ull << MLX4_EVENT_TYPE_WQ_ACCESS_ERROR) | \
75 (1ull << MLX4_EVENT_TYPE_PORT_CHANGE) | \
76 (1ull << MLX4_EVENT_TYPE_ECC_DETECT) | \
77 (1ull << MLX4_EVENT_TYPE_SRQ_CATAS_ERROR) | \
78 (1ull << MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE) | \
79 (1ull << MLX4_EVENT_TYPE_SRQ_LIMIT) | \
80 (1ull << MLX4_EVENT_TYPE_CMD) | \
81 (1ull << MLX4_EVENT_TYPE_COMM_CHANNEL) | \
82 (1ull << MLX4_EVENT_TYPE_FLR_EVENT))
83
eq_set_ci(struct mlx4_eq * eq,int req_not)84 static void eq_set_ci(struct mlx4_eq *eq, int req_not)
85 {
86 __raw_writel((__force u32) cpu_to_be32((eq->cons_index & 0xffffff) |
87 req_not << 31),
88 eq->doorbell);
89 /* We still want ordering, just not swabbing, so add a barrier */
90 mb();
91 }
92
get_eqe(struct mlx4_eq * eq,u32 entry)93 static struct mlx4_eqe *get_eqe(struct mlx4_eq *eq, u32 entry)
94 {
95 unsigned long off = (entry & (eq->nent - 1)) * MLX4_EQ_ENTRY_SIZE;
96 return eq->page_list[off / PAGE_SIZE].buf + off % PAGE_SIZE;
97 }
98
next_eqe_sw(struct mlx4_eq * eq)99 static struct mlx4_eqe *next_eqe_sw(struct mlx4_eq *eq)
100 {
101 struct mlx4_eqe *eqe = get_eqe(eq, eq->cons_index);
102 return !!(eqe->owner & 0x80) ^ !!(eq->cons_index & eq->nent) ? NULL : eqe;
103 }
104
next_slave_event_eqe(struct mlx4_slave_event_eq * slave_eq)105 static struct mlx4_eqe *next_slave_event_eqe(struct mlx4_slave_event_eq *slave_eq)
106 {
107 struct mlx4_eqe *eqe =
108 &slave_eq->event_eqe[slave_eq->cons & (SLAVE_EVENT_EQ_SIZE - 1)];
109 return (!!(eqe->owner & 0x80) ^
110 !!(slave_eq->cons & SLAVE_EVENT_EQ_SIZE)) ?
111 eqe : NULL;
112 }
113
mlx4_gen_slave_eqe(struct work_struct * work)114 void mlx4_gen_slave_eqe(struct work_struct *work)
115 {
116 struct mlx4_mfunc_master_ctx *master =
117 container_of(work, struct mlx4_mfunc_master_ctx,
118 slave_event_work);
119 struct mlx4_mfunc *mfunc =
120 container_of(master, struct mlx4_mfunc, master);
121 struct mlx4_priv *priv = container_of(mfunc, struct mlx4_priv, mfunc);
122 struct mlx4_dev *dev = &priv->dev;
123 struct mlx4_slave_event_eq *slave_eq = &mfunc->master.slave_eq;
124 struct mlx4_eqe *eqe;
125 u8 slave;
126 int i;
127
128 for (eqe = next_slave_event_eqe(slave_eq); eqe;
129 eqe = next_slave_event_eqe(slave_eq)) {
130 slave = eqe->slave_id;
131
132 /* All active slaves need to receive the event */
133 if (slave == ALL_SLAVES) {
134 for (i = 0; i < dev->num_slaves; i++) {
135 if (i != dev->caps.function &&
136 master->slave_state[i].active)
137 if (mlx4_GEN_EQE(dev, i, eqe))
138 mlx4_warn(dev, "Failed to "
139 " generate event "
140 "for slave %d\n", i);
141 }
142 } else {
143 if (mlx4_GEN_EQE(dev, slave, eqe))
144 mlx4_warn(dev, "Failed to generate event "
145 "for slave %d\n", slave);
146 }
147 ++slave_eq->cons;
148 }
149 }
150
151
slave_event(struct mlx4_dev * dev,u8 slave,struct mlx4_eqe * eqe)152 static void slave_event(struct mlx4_dev *dev, u8 slave, struct mlx4_eqe *eqe)
153 {
154 struct mlx4_priv *priv = mlx4_priv(dev);
155 struct mlx4_slave_event_eq *slave_eq = &priv->mfunc.master.slave_eq;
156 struct mlx4_eqe *s_eqe =
157 &slave_eq->event_eqe[slave_eq->prod & (SLAVE_EVENT_EQ_SIZE - 1)];
158
159 if ((!!(s_eqe->owner & 0x80)) ^
160 (!!(slave_eq->prod & SLAVE_EVENT_EQ_SIZE))) {
161 mlx4_warn(dev, "Master failed to generate an EQE for slave: %d. "
162 "No free EQE on slave events queue\n", slave);
163 return;
164 }
165
166 memcpy(s_eqe, eqe, sizeof(struct mlx4_eqe) - 1);
167 s_eqe->slave_id = slave;
168 /* ensure all information is written before setting the ownersip bit */
169 wmb();
170 s_eqe->owner = !!(slave_eq->prod & SLAVE_EVENT_EQ_SIZE) ? 0x0 : 0x80;
171 ++slave_eq->prod;
172
173 queue_work(priv->mfunc.master.comm_wq,
174 &priv->mfunc.master.slave_event_work);
175 }
176
mlx4_slave_event(struct mlx4_dev * dev,int slave,struct mlx4_eqe * eqe)177 static void mlx4_slave_event(struct mlx4_dev *dev, int slave,
178 struct mlx4_eqe *eqe)
179 {
180 struct mlx4_priv *priv = mlx4_priv(dev);
181 struct mlx4_slave_state *s_slave =
182 &priv->mfunc.master.slave_state[slave];
183
184 if (!s_slave->active) {
185 /*mlx4_warn(dev, "Trying to pass event to inactive slave\n");*/
186 return;
187 }
188
189 slave_event(dev, slave, eqe);
190 }
191
mlx4_master_handle_slave_flr(struct work_struct * work)192 void mlx4_master_handle_slave_flr(struct work_struct *work)
193 {
194 struct mlx4_mfunc_master_ctx *master =
195 container_of(work, struct mlx4_mfunc_master_ctx,
196 slave_flr_event_work);
197 struct mlx4_mfunc *mfunc =
198 container_of(master, struct mlx4_mfunc, master);
199 struct mlx4_priv *priv =
200 container_of(mfunc, struct mlx4_priv, mfunc);
201 struct mlx4_dev *dev = &priv->dev;
202 struct mlx4_slave_state *slave_state = priv->mfunc.master.slave_state;
203 int i;
204 int err;
205
206 mlx4_dbg(dev, "mlx4_handle_slave_flr\n");
207
208 for (i = 0 ; i < dev->num_slaves; i++) {
209
210 if (MLX4_COMM_CMD_FLR == slave_state[i].last_cmd) {
211 mlx4_dbg(dev, "mlx4_handle_slave_flr: "
212 "clean slave: %d\n", i);
213
214 mlx4_delete_all_resources_for_slave(dev, i);
215 /*return the slave to running mode*/
216 spin_lock(&priv->mfunc.master.slave_state_lock);
217 slave_state[i].last_cmd = MLX4_COMM_CMD_RESET;
218 slave_state[i].is_slave_going_down = 0;
219 spin_unlock(&priv->mfunc.master.slave_state_lock);
220 /*notify the FW:*/
221 err = mlx4_cmd(dev, 0, i, 0, MLX4_CMD_INFORM_FLR_DONE,
222 MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED);
223 if (err)
224 mlx4_warn(dev, "Failed to notify FW on "
225 "FLR done (slave:%d)\n", i);
226 }
227 }
228 }
229
mlx4_eq_int(struct mlx4_dev * dev,struct mlx4_eq * eq)230 static int mlx4_eq_int(struct mlx4_dev *dev, struct mlx4_eq *eq)
231 {
232 struct mlx4_priv *priv = mlx4_priv(dev);
233 struct mlx4_eqe *eqe;
234 int cqn;
235 int eqes_found = 0;
236 int set_ci = 0;
237 int port;
238 int slave = 0;
239 int ret;
240 u32 flr_slave;
241 u8 update_slave_state;
242 int i;
243
244 while ((eqe = next_eqe_sw(eq))) {
245 /*
246 * Make sure we read EQ entry contents after we've
247 * checked the ownership bit.
248 */
249 rmb();
250
251 switch (eqe->type) {
252 case MLX4_EVENT_TYPE_COMP:
253 cqn = be32_to_cpu(eqe->event.comp.cqn) & 0xffffff;
254 mlx4_cq_completion(dev, cqn);
255 break;
256
257 case MLX4_EVENT_TYPE_PATH_MIG:
258 case MLX4_EVENT_TYPE_COMM_EST:
259 case MLX4_EVENT_TYPE_SQ_DRAINED:
260 case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
261 case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
262 case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
263 case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
264 case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
265 mlx4_dbg(dev, "event %d arrived\n", eqe->type);
266 if (mlx4_is_master(dev)) {
267 /* forward only to slave owning the QP */
268 ret = mlx4_get_slave_from_resource_id(dev,
269 RES_QP,
270 be32_to_cpu(eqe->event.qp.qpn)
271 & 0xffffff, &slave);
272 if (ret && ret != -ENOENT) {
273 mlx4_dbg(dev, "QP event %02x(%02x) on "
274 "EQ %d at index %u: could "
275 "not get slave id (%d)\n",
276 eqe->type, eqe->subtype,
277 eq->eqn, eq->cons_index, ret);
278 break;
279 }
280
281 if (!ret && slave != dev->caps.function) {
282 mlx4_slave_event(dev, slave, eqe);
283 break;
284 }
285
286 }
287 mlx4_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) &
288 0xffffff, eqe->type);
289 break;
290
291 case MLX4_EVENT_TYPE_SRQ_LIMIT:
292 mlx4_warn(dev, "%s: MLX4_EVENT_TYPE_SRQ_LIMIT\n",
293 __func__);
294 case MLX4_EVENT_TYPE_SRQ_CATAS_ERROR:
295 if (mlx4_is_master(dev)) {
296 /* forward only to slave owning the SRQ */
297 ret = mlx4_get_slave_from_resource_id(dev,
298 RES_SRQ,
299 be32_to_cpu(eqe->event.srq.srqn)
300 & 0xffffff,
301 &slave);
302 if (ret && ret != -ENOENT) {
303 mlx4_warn(dev, "SRQ event %02x(%02x) "
304 "on EQ %d at index %u: could"
305 " not get slave id (%d)\n",
306 eqe->type, eqe->subtype,
307 eq->eqn, eq->cons_index, ret);
308 break;
309 }
310 mlx4_warn(dev, "%s: slave:%d, srq_no:0x%x,"
311 " event: %02x(%02x)\n", __func__,
312 slave,
313 be32_to_cpu(eqe->event.srq.srqn),
314 eqe->type, eqe->subtype);
315
316 if (!ret && slave != dev->caps.function) {
317 mlx4_warn(dev, "%s: sending event "
318 "%02x(%02x) to slave:%d\n",
319 __func__, eqe->type,
320 eqe->subtype, slave);
321 mlx4_slave_event(dev, slave, eqe);
322 break;
323 }
324 }
325 mlx4_srq_event(dev, be32_to_cpu(eqe->event.srq.srqn) &
326 0xffffff, eqe->type);
327 break;
328
329 case MLX4_EVENT_TYPE_CMD:
330 mlx4_cmd_event(dev,
331 be16_to_cpu(eqe->event.cmd.token),
332 eqe->event.cmd.status,
333 be64_to_cpu(eqe->event.cmd.out_param));
334 break;
335
336 case MLX4_EVENT_TYPE_PORT_CHANGE:
337 port = be32_to_cpu(eqe->event.port_change.port) >> 28;
338 if (eqe->subtype == MLX4_PORT_CHANGE_SUBTYPE_DOWN) {
339 mlx4_dispatch_event(dev,
340 MLX4_DEV_EVENT_PORT_DOWN,
341 port);
342 mlx4_priv(dev)->sense.do_sense_port[port] = 1;
343 if (mlx4_is_master(dev))
344 /*change the state of all slave's port
345 * to down:*/
346 for (i = 0; i < dev->num_slaves; i++) {
347 mlx4_dbg(dev, "%s: Sending "
348 "MLX4_PORT_CHANGE_SUBTYPE_DOWN"
349 " to slave: %d, port:%d\n",
350 __func__, i, port);
351 if (i == dev->caps.function)
352 continue;
353 mlx4_slave_event(dev, i, eqe);
354 }
355 } else {
356 mlx4_dispatch_event(dev,
357 MLX4_DEV_EVENT_PORT_UP,
358 port);
359 mlx4_priv(dev)->sense.do_sense_port[port] = 0;
360
361 if (mlx4_is_master(dev)) {
362 for (i = 0; i < dev->num_slaves; i++) {
363 if (i == dev->caps.function)
364 continue;
365 mlx4_slave_event(dev, i, eqe);
366 }
367 }
368 }
369 break;
370
371 case MLX4_EVENT_TYPE_CQ_ERROR:
372 mlx4_warn(dev, "CQ %s on CQN %06x\n",
373 eqe->event.cq_err.syndrome == 1 ?
374 "overrun" : "access violation",
375 be32_to_cpu(eqe->event.cq_err.cqn) & 0xffffff);
376 if (mlx4_is_master(dev)) {
377 ret = mlx4_get_slave_from_resource_id(dev,
378 RES_CQ,
379 be32_to_cpu(eqe->event.cq_err.cqn)
380 & 0xffffff, &slave);
381 if (ret && ret != -ENOENT) {
382 mlx4_dbg(dev, "CQ event %02x(%02x) on "
383 "EQ %d at index %u: could "
384 "not get slave id (%d)\n",
385 eqe->type, eqe->subtype,
386 eq->eqn, eq->cons_index, ret);
387 break;
388 }
389
390 if (!ret && slave != dev->caps.function) {
391 mlx4_slave_event(dev, slave, eqe);
392 break;
393 }
394 }
395 mlx4_cq_event(dev,
396 be32_to_cpu(eqe->event.cq_err.cqn)
397 & 0xffffff,
398 eqe->type);
399 break;
400
401 case MLX4_EVENT_TYPE_EQ_OVERFLOW:
402 mlx4_warn(dev, "EQ overrun on EQN %d\n", eq->eqn);
403 break;
404
405 case MLX4_EVENT_TYPE_COMM_CHANNEL:
406 if (!mlx4_is_master(dev)) {
407 mlx4_warn(dev, "Received comm channel event "
408 "for non master device\n");
409 break;
410 }
411 memcpy(&priv->mfunc.master.comm_arm_bit_vector,
412 eqe->event.comm_channel_arm.bit_vec,
413 sizeof eqe->event.comm_channel_arm.bit_vec);
414 queue_work(priv->mfunc.master.comm_wq,
415 &priv->mfunc.master.comm_work);
416 break;
417
418 case MLX4_EVENT_TYPE_FLR_EVENT:
419 flr_slave = be32_to_cpu(eqe->event.flr_event.slave_id);
420 if (!mlx4_is_master(dev)) {
421 mlx4_warn(dev, "Non-master function received"
422 "FLR event\n");
423 break;
424 }
425
426 mlx4_dbg(dev, "FLR event for slave: %d\n", flr_slave);
427
428 if (flr_slave > dev->num_slaves) {
429 mlx4_warn(dev,
430 "Got FLR for unknown function: %d\n",
431 flr_slave);
432 update_slave_state = 0;
433 } else
434 update_slave_state = 1;
435
436 spin_lock(&priv->mfunc.master.slave_state_lock);
437 if (update_slave_state) {
438 priv->mfunc.master.slave_state[flr_slave].active = false;
439 priv->mfunc.master.slave_state[flr_slave].last_cmd = MLX4_COMM_CMD_FLR;
440 priv->mfunc.master.slave_state[flr_slave].is_slave_going_down = 1;
441 }
442 spin_unlock(&priv->mfunc.master.slave_state_lock);
443 queue_work(priv->mfunc.master.comm_wq,
444 &priv->mfunc.master.slave_flr_event_work);
445 break;
446 case MLX4_EVENT_TYPE_EEC_CATAS_ERROR:
447 case MLX4_EVENT_TYPE_ECC_DETECT:
448 default:
449 mlx4_warn(dev, "Unhandled event %02x(%02x) on EQ %d at "
450 "index %u. owner=%x, nent=0x%x, slave=%x, "
451 "ownership=%s\n",
452 eqe->type, eqe->subtype, eq->eqn,
453 eq->cons_index, eqe->owner, eq->nent,
454 eqe->slave_id,
455 !!(eqe->owner & 0x80) ^
456 !!(eq->cons_index & eq->nent) ? "HW" : "SW");
457 break;
458 };
459
460 ++eq->cons_index;
461 eqes_found = 1;
462 ++set_ci;
463
464 /*
465 * The HCA will think the queue has overflowed if we
466 * don't tell it we've been processing events. We
467 * create our EQs with MLX4_NUM_SPARE_EQE extra
468 * entries, so we must update our consumer index at
469 * least that often.
470 */
471 if (unlikely(set_ci >= MLX4_NUM_SPARE_EQE)) {
472 eq_set_ci(eq, 0);
473 set_ci = 0;
474 }
475 }
476
477 eq_set_ci(eq, 1);
478
479 return eqes_found;
480 }
481
mlx4_interrupt(int irq,void * dev_ptr)482 static irqreturn_t mlx4_interrupt(int irq, void *dev_ptr)
483 {
484 struct mlx4_dev *dev = dev_ptr;
485 struct mlx4_priv *priv = mlx4_priv(dev);
486 int work = 0;
487 int i;
488
489 writel(priv->eq_table.clr_mask, priv->eq_table.clr_int);
490
491 for (i = 0; i < dev->caps.num_comp_vectors + 1; ++i)
492 work |= mlx4_eq_int(dev, &priv->eq_table.eq[i]);
493
494 return IRQ_RETVAL(work);
495 }
496
mlx4_msi_x_interrupt(int irq,void * eq_ptr)497 static irqreturn_t mlx4_msi_x_interrupt(int irq, void *eq_ptr)
498 {
499 struct mlx4_eq *eq = eq_ptr;
500 struct mlx4_dev *dev = eq->dev;
501
502 mlx4_eq_int(dev, eq);
503
504 /* MSI-X vectors always belong to us */
505 return IRQ_HANDLED;
506 }
507
mlx4_MAP_EQ_wrapper(struct mlx4_dev * dev,int slave,struct mlx4_vhcr * vhcr,struct mlx4_cmd_mailbox * inbox,struct mlx4_cmd_mailbox * outbox,struct mlx4_cmd_info * cmd)508 int mlx4_MAP_EQ_wrapper(struct mlx4_dev *dev, int slave,
509 struct mlx4_vhcr *vhcr,
510 struct mlx4_cmd_mailbox *inbox,
511 struct mlx4_cmd_mailbox *outbox,
512 struct mlx4_cmd_info *cmd)
513 {
514 struct mlx4_priv *priv = mlx4_priv(dev);
515 struct mlx4_slave_event_eq_info *event_eq =
516 priv->mfunc.master.slave_state[slave].event_eq;
517 u32 in_modifier = vhcr->in_modifier;
518 u32 eqn = in_modifier & 0x1FF;
519 u64 in_param = vhcr->in_param;
520 int err = 0;
521 int i;
522
523 if (slave == dev->caps.function)
524 err = mlx4_cmd(dev, in_param, (in_modifier & 0x80000000) | eqn,
525 0, MLX4_CMD_MAP_EQ, MLX4_CMD_TIME_CLASS_B,
526 MLX4_CMD_NATIVE);
527 if (!err)
528 for (i = 0; i < MLX4_EVENT_TYPES_NUM; ++i)
529 if (in_param & (1LL << i))
530 event_eq[i].eqn = in_modifier >> 31 ? -1 : eqn;
531
532 return err;
533 }
534
mlx4_MAP_EQ(struct mlx4_dev * dev,u64 event_mask,int unmap,int eq_num)535 static int mlx4_MAP_EQ(struct mlx4_dev *dev, u64 event_mask, int unmap,
536 int eq_num)
537 {
538 return mlx4_cmd(dev, event_mask, (unmap << 31) | eq_num,
539 0, MLX4_CMD_MAP_EQ, MLX4_CMD_TIME_CLASS_B,
540 MLX4_CMD_WRAPPED);
541 }
542
mlx4_SW2HW_EQ(struct mlx4_dev * dev,struct mlx4_cmd_mailbox * mailbox,int eq_num)543 static int mlx4_SW2HW_EQ(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
544 int eq_num)
545 {
546 return mlx4_cmd(dev, mailbox->dma, eq_num, 0,
547 MLX4_CMD_SW2HW_EQ, MLX4_CMD_TIME_CLASS_A,
548 MLX4_CMD_WRAPPED);
549 }
550
mlx4_HW2SW_EQ(struct mlx4_dev * dev,struct mlx4_cmd_mailbox * mailbox,int eq_num)551 static int mlx4_HW2SW_EQ(struct mlx4_dev *dev, struct mlx4_cmd_mailbox *mailbox,
552 int eq_num)
553 {
554 return mlx4_cmd_box(dev, 0, mailbox->dma, eq_num,
555 0, MLX4_CMD_HW2SW_EQ, MLX4_CMD_TIME_CLASS_A,
556 MLX4_CMD_WRAPPED);
557 }
558
mlx4_num_eq_uar(struct mlx4_dev * dev)559 static int mlx4_num_eq_uar(struct mlx4_dev *dev)
560 {
561 /*
562 * Each UAR holds 4 EQ doorbells. To figure out how many UARs
563 * we need to map, take the difference of highest index and
564 * the lowest index we'll use and add 1.
565 */
566 return (dev->caps.num_comp_vectors + 1 + dev->caps.reserved_eqs +
567 dev->caps.comp_pool)/4 - dev->caps.reserved_eqs/4 + 1;
568 }
569
mlx4_get_eq_uar(struct mlx4_dev * dev,struct mlx4_eq * eq)570 static void __iomem *mlx4_get_eq_uar(struct mlx4_dev *dev, struct mlx4_eq *eq)
571 {
572 struct mlx4_priv *priv = mlx4_priv(dev);
573 int index;
574
575 index = eq->eqn / 4 - dev->caps.reserved_eqs / 4;
576
577 if (!priv->eq_table.uar_map[index]) {
578 priv->eq_table.uar_map[index] =
579 ioremap(pci_resource_start(dev->pdev, 2) +
580 ((eq->eqn / 4) << PAGE_SHIFT),
581 PAGE_SIZE);
582 if (!priv->eq_table.uar_map[index]) {
583 mlx4_err(dev, "Couldn't map EQ doorbell for EQN 0x%06x\n",
584 eq->eqn);
585 return NULL;
586 }
587 }
588
589 return priv->eq_table.uar_map[index] + 0x800 + 8 * (eq->eqn % 4);
590 }
591
mlx4_create_eq(struct mlx4_dev * dev,int nent,u8 intr,struct mlx4_eq * eq)592 static int mlx4_create_eq(struct mlx4_dev *dev, int nent,
593 u8 intr, struct mlx4_eq *eq)
594 {
595 struct mlx4_priv *priv = mlx4_priv(dev);
596 struct mlx4_cmd_mailbox *mailbox;
597 struct mlx4_eq_context *eq_context;
598 int npages;
599 u64 *dma_list = NULL;
600 dma_addr_t t;
601 u64 mtt_addr;
602 int err = -ENOMEM;
603 int i;
604
605 eq->dev = dev;
606 eq->nent = roundup_pow_of_two(max(nent, 2));
607 npages = PAGE_ALIGN(eq->nent * MLX4_EQ_ENTRY_SIZE) / PAGE_SIZE;
608
609 eq->page_list = kmalloc(npages * sizeof *eq->page_list,
610 GFP_KERNEL);
611 if (!eq->page_list)
612 goto err_out;
613
614 for (i = 0; i < npages; ++i)
615 eq->page_list[i].buf = NULL;
616
617 dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
618 if (!dma_list)
619 goto err_out_free;
620
621 mailbox = mlx4_alloc_cmd_mailbox(dev);
622 if (IS_ERR(mailbox))
623 goto err_out_free;
624 eq_context = mailbox->buf;
625
626 for (i = 0; i < npages; ++i) {
627 eq->page_list[i].buf = dma_alloc_coherent(&dev->pdev->dev,
628 PAGE_SIZE, &t, GFP_KERNEL);
629 if (!eq->page_list[i].buf)
630 goto err_out_free_pages;
631
632 dma_list[i] = t;
633 eq->page_list[i].map = t;
634
635 memset(eq->page_list[i].buf, 0, PAGE_SIZE);
636 }
637
638 eq->eqn = mlx4_bitmap_alloc(&priv->eq_table.bitmap);
639 if (eq->eqn == -1)
640 goto err_out_free_pages;
641
642 eq->doorbell = mlx4_get_eq_uar(dev, eq);
643 if (!eq->doorbell) {
644 err = -ENOMEM;
645 goto err_out_free_eq;
646 }
647
648 err = mlx4_mtt_init(dev, npages, PAGE_SHIFT, &eq->mtt);
649 if (err)
650 goto err_out_free_eq;
651
652 err = mlx4_write_mtt(dev, &eq->mtt, 0, npages, dma_list);
653 if (err)
654 goto err_out_free_mtt;
655
656 memset(eq_context, 0, sizeof *eq_context);
657 eq_context->flags = cpu_to_be32(MLX4_EQ_STATUS_OK |
658 MLX4_EQ_STATE_ARMED);
659 eq_context->log_eq_size = ilog2(eq->nent);
660 eq_context->intr = intr;
661 eq_context->log_page_size = PAGE_SHIFT - MLX4_ICM_PAGE_SHIFT;
662
663 mtt_addr = mlx4_mtt_addr(dev, &eq->mtt);
664 eq_context->mtt_base_addr_h = mtt_addr >> 32;
665 eq_context->mtt_base_addr_l = cpu_to_be32(mtt_addr & 0xffffffff);
666
667 err = mlx4_SW2HW_EQ(dev, mailbox, eq->eqn);
668 if (err) {
669 mlx4_warn(dev, "SW2HW_EQ failed (%d)\n", err);
670 goto err_out_free_mtt;
671 }
672
673 kfree(dma_list);
674 mlx4_free_cmd_mailbox(dev, mailbox);
675
676 eq->cons_index = 0;
677
678 return err;
679
680 err_out_free_mtt:
681 mlx4_mtt_cleanup(dev, &eq->mtt);
682
683 err_out_free_eq:
684 mlx4_bitmap_free(&priv->eq_table.bitmap, eq->eqn);
685
686 err_out_free_pages:
687 for (i = 0; i < npages; ++i)
688 if (eq->page_list[i].buf)
689 dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
690 eq->page_list[i].buf,
691 eq->page_list[i].map);
692
693 mlx4_free_cmd_mailbox(dev, mailbox);
694
695 err_out_free:
696 kfree(eq->page_list);
697 kfree(dma_list);
698
699 err_out:
700 return err;
701 }
702
mlx4_free_eq(struct mlx4_dev * dev,struct mlx4_eq * eq)703 static void mlx4_free_eq(struct mlx4_dev *dev,
704 struct mlx4_eq *eq)
705 {
706 struct mlx4_priv *priv = mlx4_priv(dev);
707 struct mlx4_cmd_mailbox *mailbox;
708 int err;
709 int npages = PAGE_ALIGN(MLX4_EQ_ENTRY_SIZE * eq->nent) / PAGE_SIZE;
710 int i;
711
712 mailbox = mlx4_alloc_cmd_mailbox(dev);
713 if (IS_ERR(mailbox))
714 return;
715
716 err = mlx4_HW2SW_EQ(dev, mailbox, eq->eqn);
717 if (err)
718 mlx4_warn(dev, "HW2SW_EQ failed (%d)\n", err);
719
720 if (0) {
721 mlx4_dbg(dev, "Dumping EQ context %02x:\n", eq->eqn);
722 for (i = 0; i < sizeof (struct mlx4_eq_context) / 4; ++i) {
723 if (i % 4 == 0)
724 pr_cont("[%02x] ", i * 4);
725 pr_cont(" %08x", be32_to_cpup(mailbox->buf + i * 4));
726 if ((i + 1) % 4 == 0)
727 pr_cont("\n");
728 }
729 }
730
731 mlx4_mtt_cleanup(dev, &eq->mtt);
732 for (i = 0; i < npages; ++i)
733 dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
734 eq->page_list[i].buf,
735 eq->page_list[i].map);
736
737 kfree(eq->page_list);
738 mlx4_bitmap_free(&priv->eq_table.bitmap, eq->eqn);
739 mlx4_free_cmd_mailbox(dev, mailbox);
740 }
741
mlx4_free_irqs(struct mlx4_dev * dev)742 static void mlx4_free_irqs(struct mlx4_dev *dev)
743 {
744 struct mlx4_eq_table *eq_table = &mlx4_priv(dev)->eq_table;
745 struct mlx4_priv *priv = mlx4_priv(dev);
746 int i, vec;
747
748 if (eq_table->have_irq)
749 free_irq(dev->pdev->irq, dev);
750
751 for (i = 0; i < dev->caps.num_comp_vectors + 1; ++i)
752 if (eq_table->eq[i].have_irq) {
753 free_irq(eq_table->eq[i].irq, eq_table->eq + i);
754 eq_table->eq[i].have_irq = 0;
755 }
756
757 for (i = 0; i < dev->caps.comp_pool; i++) {
758 /*
759 * Freeing the assigned irq's
760 * all bits should be 0, but we need to validate
761 */
762 if (priv->msix_ctl.pool_bm & 1ULL << i) {
763 /* NO need protecting*/
764 vec = dev->caps.num_comp_vectors + 1 + i;
765 free_irq(priv->eq_table.eq[vec].irq,
766 &priv->eq_table.eq[vec]);
767 }
768 }
769
770
771 kfree(eq_table->irq_names);
772 }
773
mlx4_map_clr_int(struct mlx4_dev * dev)774 static int mlx4_map_clr_int(struct mlx4_dev *dev)
775 {
776 struct mlx4_priv *priv = mlx4_priv(dev);
777
778 priv->clr_base = ioremap(pci_resource_start(dev->pdev, priv->fw.clr_int_bar) +
779 priv->fw.clr_int_base, MLX4_CLR_INT_SIZE);
780 if (!priv->clr_base) {
781 mlx4_err(dev, "Couldn't map interrupt clear register, aborting.\n");
782 return -ENOMEM;
783 }
784
785 return 0;
786 }
787
mlx4_unmap_clr_int(struct mlx4_dev * dev)788 static void mlx4_unmap_clr_int(struct mlx4_dev *dev)
789 {
790 struct mlx4_priv *priv = mlx4_priv(dev);
791
792 iounmap(priv->clr_base);
793 }
794
mlx4_alloc_eq_table(struct mlx4_dev * dev)795 int mlx4_alloc_eq_table(struct mlx4_dev *dev)
796 {
797 struct mlx4_priv *priv = mlx4_priv(dev);
798
799 priv->eq_table.eq = kcalloc(dev->caps.num_eqs - dev->caps.reserved_eqs,
800 sizeof *priv->eq_table.eq, GFP_KERNEL);
801 if (!priv->eq_table.eq)
802 return -ENOMEM;
803
804 return 0;
805 }
806
mlx4_free_eq_table(struct mlx4_dev * dev)807 void mlx4_free_eq_table(struct mlx4_dev *dev)
808 {
809 kfree(mlx4_priv(dev)->eq_table.eq);
810 }
811
mlx4_init_eq_table(struct mlx4_dev * dev)812 int mlx4_init_eq_table(struct mlx4_dev *dev)
813 {
814 struct mlx4_priv *priv = mlx4_priv(dev);
815 int err;
816 int i;
817
818 priv->eq_table.uar_map = kcalloc(mlx4_num_eq_uar(dev),
819 sizeof *priv->eq_table.uar_map,
820 GFP_KERNEL);
821 if (!priv->eq_table.uar_map) {
822 err = -ENOMEM;
823 goto err_out_free;
824 }
825
826 err = mlx4_bitmap_init(&priv->eq_table.bitmap, dev->caps.num_eqs,
827 dev->caps.num_eqs - 1, dev->caps.reserved_eqs, 0);
828 if (err)
829 goto err_out_free;
830
831 for (i = 0; i < mlx4_num_eq_uar(dev); ++i)
832 priv->eq_table.uar_map[i] = NULL;
833
834 if (!mlx4_is_slave(dev)) {
835 err = mlx4_map_clr_int(dev);
836 if (err)
837 goto err_out_bitmap;
838
839 priv->eq_table.clr_mask =
840 swab32(1 << (priv->eq_table.inta_pin & 31));
841 priv->eq_table.clr_int = priv->clr_base +
842 (priv->eq_table.inta_pin < 32 ? 4 : 0);
843 }
844
845 priv->eq_table.irq_names =
846 kmalloc(MLX4_IRQNAME_SIZE * (dev->caps.num_comp_vectors + 1 +
847 dev->caps.comp_pool),
848 GFP_KERNEL);
849 if (!priv->eq_table.irq_names) {
850 err = -ENOMEM;
851 goto err_out_bitmap;
852 }
853
854 for (i = 0; i < dev->caps.num_comp_vectors; ++i) {
855 err = mlx4_create_eq(dev, dev->caps.num_cqs -
856 dev->caps.reserved_cqs +
857 MLX4_NUM_SPARE_EQE,
858 (dev->flags & MLX4_FLAG_MSI_X) ? i : 0,
859 &priv->eq_table.eq[i]);
860 if (err) {
861 --i;
862 goto err_out_unmap;
863 }
864 }
865
866 err = mlx4_create_eq(dev, MLX4_NUM_ASYNC_EQE + MLX4_NUM_SPARE_EQE,
867 (dev->flags & MLX4_FLAG_MSI_X) ? dev->caps.num_comp_vectors : 0,
868 &priv->eq_table.eq[dev->caps.num_comp_vectors]);
869 if (err)
870 goto err_out_comp;
871
872 /*if additional completion vectors poolsize is 0 this loop will not run*/
873 for (i = dev->caps.num_comp_vectors + 1;
874 i < dev->caps.num_comp_vectors + dev->caps.comp_pool + 1; ++i) {
875
876 err = mlx4_create_eq(dev, dev->caps.num_cqs -
877 dev->caps.reserved_cqs +
878 MLX4_NUM_SPARE_EQE,
879 (dev->flags & MLX4_FLAG_MSI_X) ? i : 0,
880 &priv->eq_table.eq[i]);
881 if (err) {
882 --i;
883 goto err_out_unmap;
884 }
885 }
886
887
888 if (dev->flags & MLX4_FLAG_MSI_X) {
889 const char *eq_name;
890
891 for (i = 0; i < dev->caps.num_comp_vectors + 1; ++i) {
892 if (i < dev->caps.num_comp_vectors) {
893 snprintf(priv->eq_table.irq_names +
894 i * MLX4_IRQNAME_SIZE,
895 MLX4_IRQNAME_SIZE,
896 "mlx4-comp-%d@pci:%s", i,
897 pci_name(dev->pdev));
898 } else {
899 snprintf(priv->eq_table.irq_names +
900 i * MLX4_IRQNAME_SIZE,
901 MLX4_IRQNAME_SIZE,
902 "mlx4-async@pci:%s",
903 pci_name(dev->pdev));
904 }
905
906 eq_name = priv->eq_table.irq_names +
907 i * MLX4_IRQNAME_SIZE;
908 err = request_irq(priv->eq_table.eq[i].irq,
909 mlx4_msi_x_interrupt, 0, eq_name,
910 priv->eq_table.eq + i);
911 if (err)
912 goto err_out_async;
913
914 priv->eq_table.eq[i].have_irq = 1;
915 }
916 } else {
917 snprintf(priv->eq_table.irq_names,
918 MLX4_IRQNAME_SIZE,
919 DRV_NAME "@pci:%s",
920 pci_name(dev->pdev));
921 err = request_irq(dev->pdev->irq, mlx4_interrupt,
922 IRQF_SHARED, priv->eq_table.irq_names, dev);
923 if (err)
924 goto err_out_async;
925
926 priv->eq_table.have_irq = 1;
927 }
928
929 err = mlx4_MAP_EQ(dev, MLX4_ASYNC_EVENT_MASK, 0,
930 priv->eq_table.eq[dev->caps.num_comp_vectors].eqn);
931 if (err)
932 mlx4_warn(dev, "MAP_EQ for async EQ %d failed (%d)\n",
933 priv->eq_table.eq[dev->caps.num_comp_vectors].eqn, err);
934
935 for (i = 0; i < dev->caps.num_comp_vectors + 1; ++i)
936 eq_set_ci(&priv->eq_table.eq[i], 1);
937
938 return 0;
939
940 err_out_async:
941 mlx4_free_eq(dev, &priv->eq_table.eq[dev->caps.num_comp_vectors]);
942
943 err_out_comp:
944 i = dev->caps.num_comp_vectors - 1;
945
946 err_out_unmap:
947 while (i >= 0) {
948 mlx4_free_eq(dev, &priv->eq_table.eq[i]);
949 --i;
950 }
951 if (!mlx4_is_slave(dev))
952 mlx4_unmap_clr_int(dev);
953 mlx4_free_irqs(dev);
954
955 err_out_bitmap:
956 mlx4_bitmap_cleanup(&priv->eq_table.bitmap);
957
958 err_out_free:
959 kfree(priv->eq_table.uar_map);
960
961 return err;
962 }
963
mlx4_cleanup_eq_table(struct mlx4_dev * dev)964 void mlx4_cleanup_eq_table(struct mlx4_dev *dev)
965 {
966 struct mlx4_priv *priv = mlx4_priv(dev);
967 int i;
968
969 mlx4_MAP_EQ(dev, MLX4_ASYNC_EVENT_MASK, 1,
970 priv->eq_table.eq[dev->caps.num_comp_vectors].eqn);
971
972 mlx4_free_irqs(dev);
973
974 for (i = 0; i < dev->caps.num_comp_vectors + dev->caps.comp_pool + 1; ++i)
975 mlx4_free_eq(dev, &priv->eq_table.eq[i]);
976
977 if (!mlx4_is_slave(dev))
978 mlx4_unmap_clr_int(dev);
979
980 for (i = 0; i < mlx4_num_eq_uar(dev); ++i)
981 if (priv->eq_table.uar_map[i])
982 iounmap(priv->eq_table.uar_map[i]);
983
984 mlx4_bitmap_cleanup(&priv->eq_table.bitmap);
985
986 kfree(priv->eq_table.uar_map);
987 }
988
989 /* A test that verifies that we can accept interrupts on all
990 * the irq vectors of the device.
991 * Interrupts are checked using the NOP command.
992 */
mlx4_test_interrupts(struct mlx4_dev * dev)993 int mlx4_test_interrupts(struct mlx4_dev *dev)
994 {
995 struct mlx4_priv *priv = mlx4_priv(dev);
996 int i;
997 int err;
998
999 err = mlx4_NOP(dev);
1000 /* When not in MSI_X, there is only one irq to check */
1001 if (!(dev->flags & MLX4_FLAG_MSI_X) || mlx4_is_slave(dev))
1002 return err;
1003
1004 /* A loop over all completion vectors, for each vector we will check
1005 * whether it works by mapping command completions to that vector
1006 * and performing a NOP command
1007 */
1008 for(i = 0; !err && (i < dev->caps.num_comp_vectors); ++i) {
1009 /* Temporary use polling for command completions */
1010 mlx4_cmd_use_polling(dev);
1011
1012 /* Map the new eq to handle all asyncronous events */
1013 err = mlx4_MAP_EQ(dev, MLX4_ASYNC_EVENT_MASK, 0,
1014 priv->eq_table.eq[i].eqn);
1015 if (err) {
1016 mlx4_warn(dev, "Failed mapping eq for interrupt test\n");
1017 mlx4_cmd_use_events(dev);
1018 break;
1019 }
1020
1021 /* Go back to using events */
1022 mlx4_cmd_use_events(dev);
1023 err = mlx4_NOP(dev);
1024 }
1025
1026 /* Return to default */
1027 mlx4_MAP_EQ(dev, MLX4_ASYNC_EVENT_MASK, 0,
1028 priv->eq_table.eq[dev->caps.num_comp_vectors].eqn);
1029 return err;
1030 }
1031 EXPORT_SYMBOL(mlx4_test_interrupts);
1032
mlx4_assign_eq(struct mlx4_dev * dev,char * name,int * vector)1033 int mlx4_assign_eq(struct mlx4_dev *dev, char* name, int * vector)
1034 {
1035
1036 struct mlx4_priv *priv = mlx4_priv(dev);
1037 int vec = 0, err = 0, i;
1038
1039 mutex_lock(&priv->msix_ctl.pool_lock);
1040 for (i = 0; !vec && i < dev->caps.comp_pool; i++) {
1041 if (~priv->msix_ctl.pool_bm & 1ULL << i) {
1042 priv->msix_ctl.pool_bm |= 1ULL << i;
1043 vec = dev->caps.num_comp_vectors + 1 + i;
1044 snprintf(priv->eq_table.irq_names +
1045 vec * MLX4_IRQNAME_SIZE,
1046 MLX4_IRQNAME_SIZE, "%s", name);
1047 err = request_irq(priv->eq_table.eq[vec].irq,
1048 mlx4_msi_x_interrupt, 0,
1049 &priv->eq_table.irq_names[vec<<5],
1050 priv->eq_table.eq + vec);
1051 if (err) {
1052 /*zero out bit by fliping it*/
1053 priv->msix_ctl.pool_bm ^= 1 << i;
1054 vec = 0;
1055 continue;
1056 /*we dont want to break here*/
1057 }
1058 eq_set_ci(&priv->eq_table.eq[vec], 1);
1059 }
1060 }
1061 mutex_unlock(&priv->msix_ctl.pool_lock);
1062
1063 if (vec) {
1064 *vector = vec;
1065 } else {
1066 *vector = 0;
1067 err = (i == dev->caps.comp_pool) ? -ENOSPC : err;
1068 }
1069 return err;
1070 }
1071 EXPORT_SYMBOL(mlx4_assign_eq);
1072
mlx4_release_eq(struct mlx4_dev * dev,int vec)1073 void mlx4_release_eq(struct mlx4_dev *dev, int vec)
1074 {
1075 struct mlx4_priv *priv = mlx4_priv(dev);
1076 /*bm index*/
1077 int i = vec - dev->caps.num_comp_vectors - 1;
1078
1079 if (likely(i >= 0)) {
1080 /*sanity check , making sure were not trying to free irq's
1081 Belonging to a legacy EQ*/
1082 mutex_lock(&priv->msix_ctl.pool_lock);
1083 if (priv->msix_ctl.pool_bm & 1ULL << i) {
1084 free_irq(priv->eq_table.eq[vec].irq,
1085 &priv->eq_table.eq[vec]);
1086 priv->msix_ctl.pool_bm &= ~(1ULL << i);
1087 }
1088 mutex_unlock(&priv->msix_ctl.pool_lock);
1089 }
1090
1091 }
1092 EXPORT_SYMBOL(mlx4_release_eq);
1093
1094