1 /*
2 * Copyright (c) 2004, 2005 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Mellanox Technologies. 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/errno.h>
35 #include <linux/interrupt.h>
36 #include <linux/pci.h>
37 #include <linux/slab.h>
38
39 #include "mthca_dev.h"
40 #include "mthca_cmd.h"
41 #include "mthca_config_reg.h"
42
43 enum {
44 MTHCA_NUM_ASYNC_EQE = 0x80,
45 MTHCA_NUM_CMD_EQE = 0x80,
46 MTHCA_NUM_SPARE_EQE = 0x80,
47 MTHCA_EQ_ENTRY_SIZE = 0x20
48 };
49
50 /*
51 * Must be packed because start is 64 bits but only aligned to 32 bits.
52 */
53 struct mthca_eq_context {
54 __be32 flags;
55 __be64 start;
56 __be32 logsize_usrpage;
57 __be32 tavor_pd; /* reserved for Arbel */
58 u8 reserved1[3];
59 u8 intr;
60 __be32 arbel_pd; /* lost_count for Tavor */
61 __be32 lkey;
62 u32 reserved2[2];
63 __be32 consumer_index;
64 __be32 producer_index;
65 u32 reserved3[4];
66 } __attribute__((packed));
67
68 #define MTHCA_EQ_STATUS_OK ( 0 << 28)
69 #define MTHCA_EQ_STATUS_OVERFLOW ( 9 << 28)
70 #define MTHCA_EQ_STATUS_WRITE_FAIL (10 << 28)
71 #define MTHCA_EQ_OWNER_SW ( 0 << 24)
72 #define MTHCA_EQ_OWNER_HW ( 1 << 24)
73 #define MTHCA_EQ_FLAG_TR ( 1 << 18)
74 #define MTHCA_EQ_FLAG_OI ( 1 << 17)
75 #define MTHCA_EQ_STATE_ARMED ( 1 << 8)
76 #define MTHCA_EQ_STATE_FIRED ( 2 << 8)
77 #define MTHCA_EQ_STATE_ALWAYS_ARMED ( 3 << 8)
78 #define MTHCA_EQ_STATE_ARBEL ( 8 << 8)
79
80 enum {
81 MTHCA_EVENT_TYPE_COMP = 0x00,
82 MTHCA_EVENT_TYPE_PATH_MIG = 0x01,
83 MTHCA_EVENT_TYPE_COMM_EST = 0x02,
84 MTHCA_EVENT_TYPE_SQ_DRAINED = 0x03,
85 MTHCA_EVENT_TYPE_SRQ_QP_LAST_WQE = 0x13,
86 MTHCA_EVENT_TYPE_SRQ_LIMIT = 0x14,
87 MTHCA_EVENT_TYPE_CQ_ERROR = 0x04,
88 MTHCA_EVENT_TYPE_WQ_CATAS_ERROR = 0x05,
89 MTHCA_EVENT_TYPE_EEC_CATAS_ERROR = 0x06,
90 MTHCA_EVENT_TYPE_PATH_MIG_FAILED = 0x07,
91 MTHCA_EVENT_TYPE_WQ_INVAL_REQ_ERROR = 0x10,
92 MTHCA_EVENT_TYPE_WQ_ACCESS_ERROR = 0x11,
93 MTHCA_EVENT_TYPE_SRQ_CATAS_ERROR = 0x12,
94 MTHCA_EVENT_TYPE_LOCAL_CATAS_ERROR = 0x08,
95 MTHCA_EVENT_TYPE_PORT_CHANGE = 0x09,
96 MTHCA_EVENT_TYPE_EQ_OVERFLOW = 0x0f,
97 MTHCA_EVENT_TYPE_ECC_DETECT = 0x0e,
98 MTHCA_EVENT_TYPE_CMD = 0x0a
99 };
100
101 #define MTHCA_ASYNC_EVENT_MASK ((1ULL << MTHCA_EVENT_TYPE_PATH_MIG) | \
102 (1ULL << MTHCA_EVENT_TYPE_COMM_EST) | \
103 (1ULL << MTHCA_EVENT_TYPE_SQ_DRAINED) | \
104 (1ULL << MTHCA_EVENT_TYPE_CQ_ERROR) | \
105 (1ULL << MTHCA_EVENT_TYPE_WQ_CATAS_ERROR) | \
106 (1ULL << MTHCA_EVENT_TYPE_EEC_CATAS_ERROR) | \
107 (1ULL << MTHCA_EVENT_TYPE_PATH_MIG_FAILED) | \
108 (1ULL << MTHCA_EVENT_TYPE_WQ_INVAL_REQ_ERROR) | \
109 (1ULL << MTHCA_EVENT_TYPE_WQ_ACCESS_ERROR) | \
110 (1ULL << MTHCA_EVENT_TYPE_LOCAL_CATAS_ERROR) | \
111 (1ULL << MTHCA_EVENT_TYPE_PORT_CHANGE) | \
112 (1ULL << MTHCA_EVENT_TYPE_ECC_DETECT))
113 #define MTHCA_SRQ_EVENT_MASK ((1ULL << MTHCA_EVENT_TYPE_SRQ_CATAS_ERROR) | \
114 (1ULL << MTHCA_EVENT_TYPE_SRQ_QP_LAST_WQE) | \
115 (1ULL << MTHCA_EVENT_TYPE_SRQ_LIMIT))
116 #define MTHCA_CMD_EVENT_MASK (1ULL << MTHCA_EVENT_TYPE_CMD)
117
118 #define MTHCA_EQ_DB_INC_CI (1 << 24)
119 #define MTHCA_EQ_DB_REQ_NOT (2 << 24)
120 #define MTHCA_EQ_DB_DISARM_CQ (3 << 24)
121 #define MTHCA_EQ_DB_SET_CI (4 << 24)
122 #define MTHCA_EQ_DB_ALWAYS_ARM (5 << 24)
123
124 struct mthca_eqe {
125 u8 reserved1;
126 u8 type;
127 u8 reserved2;
128 u8 subtype;
129 union {
130 u32 raw[6];
131 struct {
132 __be32 cqn;
133 } __attribute__((packed)) comp;
134 struct {
135 u16 reserved1;
136 __be16 token;
137 u32 reserved2;
138 u8 reserved3[3];
139 u8 status;
140 __be64 out_param;
141 } __attribute__((packed)) cmd;
142 struct {
143 __be32 qpn;
144 } __attribute__((packed)) qp;
145 struct {
146 __be32 srqn;
147 } __attribute__((packed)) srq;
148 struct {
149 __be32 cqn;
150 u32 reserved1;
151 u8 reserved2[3];
152 u8 syndrome;
153 } __attribute__((packed)) cq_err;
154 struct {
155 u32 reserved1[2];
156 __be32 port;
157 } __attribute__((packed)) port_change;
158 } event;
159 u8 reserved3[3];
160 u8 owner;
161 } __attribute__((packed));
162
163 #define MTHCA_EQ_ENTRY_OWNER_SW (0 << 7)
164 #define MTHCA_EQ_ENTRY_OWNER_HW (1 << 7)
165
async_mask(struct mthca_dev * dev)166 static inline u64 async_mask(struct mthca_dev *dev)
167 {
168 return dev->mthca_flags & MTHCA_FLAG_SRQ ?
169 MTHCA_ASYNC_EVENT_MASK | MTHCA_SRQ_EVENT_MASK :
170 MTHCA_ASYNC_EVENT_MASK;
171 }
172
tavor_set_eq_ci(struct mthca_dev * dev,struct mthca_eq * eq,u32 ci)173 static inline void tavor_set_eq_ci(struct mthca_dev *dev, struct mthca_eq *eq, u32 ci)
174 {
175 /*
176 * This barrier makes sure that all updates to ownership bits
177 * done by set_eqe_hw() hit memory before the consumer index
178 * is updated. set_eq_ci() allows the HCA to possibly write
179 * more EQ entries, and we want to avoid the exceedingly
180 * unlikely possibility of the HCA writing an entry and then
181 * having set_eqe_hw() overwrite the owner field.
182 */
183 wmb();
184 mthca_write64(MTHCA_EQ_DB_SET_CI | eq->eqn, ci & (eq->nent - 1),
185 dev->kar + MTHCA_EQ_DOORBELL,
186 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
187 }
188
arbel_set_eq_ci(struct mthca_dev * dev,struct mthca_eq * eq,u32 ci)189 static inline void arbel_set_eq_ci(struct mthca_dev *dev, struct mthca_eq *eq, u32 ci)
190 {
191 /* See comment in tavor_set_eq_ci() above. */
192 wmb();
193 __raw_writel((__force u32) cpu_to_be32(ci),
194 dev->eq_regs.arbel.eq_set_ci_base + eq->eqn * 8);
195 /* We still want ordering, just not swabbing, so add a barrier */
196 mb();
197 }
198
set_eq_ci(struct mthca_dev * dev,struct mthca_eq * eq,u32 ci)199 static inline void set_eq_ci(struct mthca_dev *dev, struct mthca_eq *eq, u32 ci)
200 {
201 if (mthca_is_memfree(dev))
202 arbel_set_eq_ci(dev, eq, ci);
203 else
204 tavor_set_eq_ci(dev, eq, ci);
205 }
206
tavor_eq_req_not(struct mthca_dev * dev,int eqn)207 static inline void tavor_eq_req_not(struct mthca_dev *dev, int eqn)
208 {
209 mthca_write64(MTHCA_EQ_DB_REQ_NOT | eqn, 0,
210 dev->kar + MTHCA_EQ_DOORBELL,
211 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
212 }
213
arbel_eq_req_not(struct mthca_dev * dev,u32 eqn_mask)214 static inline void arbel_eq_req_not(struct mthca_dev *dev, u32 eqn_mask)
215 {
216 writel(eqn_mask, dev->eq_regs.arbel.eq_arm);
217 }
218
disarm_cq(struct mthca_dev * dev,int eqn,int cqn)219 static inline void disarm_cq(struct mthca_dev *dev, int eqn, int cqn)
220 {
221 if (!mthca_is_memfree(dev)) {
222 mthca_write64(MTHCA_EQ_DB_DISARM_CQ | eqn, cqn,
223 dev->kar + MTHCA_EQ_DOORBELL,
224 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
225 }
226 }
227
get_eqe(struct mthca_eq * eq,u32 entry)228 static inline struct mthca_eqe *get_eqe(struct mthca_eq *eq, u32 entry)
229 {
230 unsigned long off = (entry & (eq->nent - 1)) * MTHCA_EQ_ENTRY_SIZE;
231 return eq->page_list[off / PAGE_SIZE].buf + off % PAGE_SIZE;
232 }
233
next_eqe_sw(struct mthca_eq * eq)234 static inline struct mthca_eqe *next_eqe_sw(struct mthca_eq *eq)
235 {
236 struct mthca_eqe *eqe;
237 eqe = get_eqe(eq, eq->cons_index);
238 return (MTHCA_EQ_ENTRY_OWNER_HW & eqe->owner) ? NULL : eqe;
239 }
240
set_eqe_hw(struct mthca_eqe * eqe)241 static inline void set_eqe_hw(struct mthca_eqe *eqe)
242 {
243 eqe->owner = MTHCA_EQ_ENTRY_OWNER_HW;
244 }
245
port_change(struct mthca_dev * dev,int port,int active)246 static void port_change(struct mthca_dev *dev, int port, int active)
247 {
248 struct ib_event record;
249
250 mthca_dbg(dev, "Port change to %s for port %d\n",
251 active ? "active" : "down", port);
252
253 record.device = &dev->ib_dev;
254 record.event = active ? IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
255 record.element.port_num = port;
256
257 ib_dispatch_event(&record);
258 }
259
mthca_eq_int(struct mthca_dev * dev,struct mthca_eq * eq)260 static int mthca_eq_int(struct mthca_dev *dev, struct mthca_eq *eq)
261 {
262 struct mthca_eqe *eqe;
263 int disarm_cqn;
264 int eqes_found = 0;
265 int set_ci = 0;
266
267 while ((eqe = next_eqe_sw(eq))) {
268 /*
269 * Make sure we read EQ entry contents after we've
270 * checked the ownership bit.
271 */
272 rmb();
273
274 switch (eqe->type) {
275 case MTHCA_EVENT_TYPE_COMP:
276 disarm_cqn = be32_to_cpu(eqe->event.comp.cqn) & 0xffffff;
277 disarm_cq(dev, eq->eqn, disarm_cqn);
278 mthca_cq_completion(dev, disarm_cqn);
279 break;
280
281 case MTHCA_EVENT_TYPE_PATH_MIG:
282 mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
283 IB_EVENT_PATH_MIG);
284 break;
285
286 case MTHCA_EVENT_TYPE_COMM_EST:
287 mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
288 IB_EVENT_COMM_EST);
289 break;
290
291 case MTHCA_EVENT_TYPE_SQ_DRAINED:
292 mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
293 IB_EVENT_SQ_DRAINED);
294 break;
295
296 case MTHCA_EVENT_TYPE_SRQ_QP_LAST_WQE:
297 mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
298 IB_EVENT_QP_LAST_WQE_REACHED);
299 break;
300
301 case MTHCA_EVENT_TYPE_SRQ_LIMIT:
302 mthca_srq_event(dev, be32_to_cpu(eqe->event.srq.srqn) & 0xffffff,
303 IB_EVENT_SRQ_LIMIT_REACHED);
304 break;
305
306 case MTHCA_EVENT_TYPE_WQ_CATAS_ERROR:
307 mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
308 IB_EVENT_QP_FATAL);
309 break;
310
311 case MTHCA_EVENT_TYPE_PATH_MIG_FAILED:
312 mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
313 IB_EVENT_PATH_MIG_ERR);
314 break;
315
316 case MTHCA_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
317 mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
318 IB_EVENT_QP_REQ_ERR);
319 break;
320
321 case MTHCA_EVENT_TYPE_WQ_ACCESS_ERROR:
322 mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
323 IB_EVENT_QP_ACCESS_ERR);
324 break;
325
326 case MTHCA_EVENT_TYPE_CMD:
327 mthca_cmd_event(dev,
328 be16_to_cpu(eqe->event.cmd.token),
329 eqe->event.cmd.status,
330 be64_to_cpu(eqe->event.cmd.out_param));
331 break;
332
333 case MTHCA_EVENT_TYPE_PORT_CHANGE:
334 port_change(dev,
335 (be32_to_cpu(eqe->event.port_change.port) >> 28) & 3,
336 eqe->subtype == 0x4);
337 break;
338
339 case MTHCA_EVENT_TYPE_CQ_ERROR:
340 mthca_warn(dev, "CQ %s on CQN %06x\n",
341 eqe->event.cq_err.syndrome == 1 ?
342 "overrun" : "access violation",
343 be32_to_cpu(eqe->event.cq_err.cqn) & 0xffffff);
344 mthca_cq_event(dev, be32_to_cpu(eqe->event.cq_err.cqn),
345 IB_EVENT_CQ_ERR);
346 break;
347
348 case MTHCA_EVENT_TYPE_EQ_OVERFLOW:
349 mthca_warn(dev, "EQ overrun on EQN %d\n", eq->eqn);
350 break;
351
352 case MTHCA_EVENT_TYPE_EEC_CATAS_ERROR:
353 case MTHCA_EVENT_TYPE_SRQ_CATAS_ERROR:
354 case MTHCA_EVENT_TYPE_LOCAL_CATAS_ERROR:
355 case MTHCA_EVENT_TYPE_ECC_DETECT:
356 default:
357 mthca_warn(dev, "Unhandled event %02x(%02x) on EQ %d\n",
358 eqe->type, eqe->subtype, eq->eqn);
359 break;
360 };
361
362 set_eqe_hw(eqe);
363 ++eq->cons_index;
364 eqes_found = 1;
365 ++set_ci;
366
367 /*
368 * The HCA will think the queue has overflowed if we
369 * don't tell it we've been processing events. We
370 * create our EQs with MTHCA_NUM_SPARE_EQE extra
371 * entries, so we must update our consumer index at
372 * least that often.
373 */
374 if (unlikely(set_ci >= MTHCA_NUM_SPARE_EQE)) {
375 /*
376 * Conditional on hca_type is OK here because
377 * this is a rare case, not the fast path.
378 */
379 set_eq_ci(dev, eq, eq->cons_index);
380 set_ci = 0;
381 }
382 }
383
384 /*
385 * Rely on caller to set consumer index so that we don't have
386 * to test hca_type in our interrupt handling fast path.
387 */
388 return eqes_found;
389 }
390
mthca_tavor_interrupt(int irq,void * dev_ptr)391 static irqreturn_t mthca_tavor_interrupt(int irq, void *dev_ptr)
392 {
393 struct mthca_dev *dev = dev_ptr;
394 u32 ecr;
395 int i;
396
397 if (dev->eq_table.clr_mask)
398 writel(dev->eq_table.clr_mask, dev->eq_table.clr_int);
399
400 ecr = readl(dev->eq_regs.tavor.ecr_base + 4);
401 if (!ecr)
402 return IRQ_NONE;
403
404 writel(ecr, dev->eq_regs.tavor.ecr_base +
405 MTHCA_ECR_CLR_BASE - MTHCA_ECR_BASE + 4);
406
407 for (i = 0; i < MTHCA_NUM_EQ; ++i)
408 if (ecr & dev->eq_table.eq[i].eqn_mask) {
409 if (mthca_eq_int(dev, &dev->eq_table.eq[i]))
410 tavor_set_eq_ci(dev, &dev->eq_table.eq[i],
411 dev->eq_table.eq[i].cons_index);
412 tavor_eq_req_not(dev, dev->eq_table.eq[i].eqn);
413 }
414
415 return IRQ_HANDLED;
416 }
417
mthca_tavor_msi_x_interrupt(int irq,void * eq_ptr)418 static irqreturn_t mthca_tavor_msi_x_interrupt(int irq, void *eq_ptr)
419 {
420 struct mthca_eq *eq = eq_ptr;
421 struct mthca_dev *dev = eq->dev;
422
423 mthca_eq_int(dev, eq);
424 tavor_set_eq_ci(dev, eq, eq->cons_index);
425 tavor_eq_req_not(dev, eq->eqn);
426
427 /* MSI-X vectors always belong to us */
428 return IRQ_HANDLED;
429 }
430
mthca_arbel_interrupt(int irq,void * dev_ptr)431 static irqreturn_t mthca_arbel_interrupt(int irq, void *dev_ptr)
432 {
433 struct mthca_dev *dev = dev_ptr;
434 int work = 0;
435 int i;
436
437 if (dev->eq_table.clr_mask)
438 writel(dev->eq_table.clr_mask, dev->eq_table.clr_int);
439
440 for (i = 0; i < MTHCA_NUM_EQ; ++i)
441 if (mthca_eq_int(dev, &dev->eq_table.eq[i])) {
442 work = 1;
443 arbel_set_eq_ci(dev, &dev->eq_table.eq[i],
444 dev->eq_table.eq[i].cons_index);
445 }
446
447 arbel_eq_req_not(dev, dev->eq_table.arm_mask);
448
449 return IRQ_RETVAL(work);
450 }
451
mthca_arbel_msi_x_interrupt(int irq,void * eq_ptr)452 static irqreturn_t mthca_arbel_msi_x_interrupt(int irq, void *eq_ptr)
453 {
454 struct mthca_eq *eq = eq_ptr;
455 struct mthca_dev *dev = eq->dev;
456
457 mthca_eq_int(dev, eq);
458 arbel_set_eq_ci(dev, eq, eq->cons_index);
459 arbel_eq_req_not(dev, eq->eqn_mask);
460
461 /* MSI-X vectors always belong to us */
462 return IRQ_HANDLED;
463 }
464
mthca_create_eq(struct mthca_dev * dev,int nent,u8 intr,struct mthca_eq * eq)465 static int mthca_create_eq(struct mthca_dev *dev,
466 int nent,
467 u8 intr,
468 struct mthca_eq *eq)
469 {
470 int npages;
471 u64 *dma_list = NULL;
472 dma_addr_t t;
473 struct mthca_mailbox *mailbox;
474 struct mthca_eq_context *eq_context;
475 int err = -ENOMEM;
476 int i;
477
478 eq->dev = dev;
479 eq->nent = roundup_pow_of_two(max(nent, 2));
480 npages = ALIGN(eq->nent * MTHCA_EQ_ENTRY_SIZE, PAGE_SIZE) / PAGE_SIZE;
481
482 eq->page_list = kmalloc(npages * sizeof *eq->page_list,
483 GFP_KERNEL);
484 if (!eq->page_list)
485 goto err_out;
486
487 for (i = 0; i < npages; ++i)
488 eq->page_list[i].buf = NULL;
489
490 dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
491 if (!dma_list)
492 goto err_out_free;
493
494 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
495 if (IS_ERR(mailbox))
496 goto err_out_free;
497 eq_context = mailbox->buf;
498
499 for (i = 0; i < npages; ++i) {
500 eq->page_list[i].buf = dma_alloc_coherent(&dev->pdev->dev,
501 PAGE_SIZE, &t, GFP_KERNEL);
502 if (!eq->page_list[i].buf)
503 goto err_out_free_pages;
504
505 dma_list[i] = t;
506 dma_unmap_addr_set(&eq->page_list[i], mapping, t);
507
508 clear_page(eq->page_list[i].buf);
509 }
510
511 for (i = 0; i < eq->nent; ++i)
512 set_eqe_hw(get_eqe(eq, i));
513
514 eq->eqn = mthca_alloc(&dev->eq_table.alloc);
515 if (eq->eqn == -1)
516 goto err_out_free_pages;
517
518 err = mthca_mr_alloc_phys(dev, dev->driver_pd.pd_num,
519 dma_list, PAGE_SHIFT, npages,
520 0, npages * PAGE_SIZE,
521 MTHCA_MPT_FLAG_LOCAL_WRITE |
522 MTHCA_MPT_FLAG_LOCAL_READ,
523 &eq->mr);
524 if (err)
525 goto err_out_free_eq;
526
527 memset(eq_context, 0, sizeof *eq_context);
528 eq_context->flags = cpu_to_be32(MTHCA_EQ_STATUS_OK |
529 MTHCA_EQ_OWNER_HW |
530 MTHCA_EQ_STATE_ARMED |
531 MTHCA_EQ_FLAG_TR);
532 if (mthca_is_memfree(dev))
533 eq_context->flags |= cpu_to_be32(MTHCA_EQ_STATE_ARBEL);
534
535 eq_context->logsize_usrpage = cpu_to_be32((ffs(eq->nent) - 1) << 24);
536 if (mthca_is_memfree(dev)) {
537 eq_context->arbel_pd = cpu_to_be32(dev->driver_pd.pd_num);
538 } else {
539 eq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
540 eq_context->tavor_pd = cpu_to_be32(dev->driver_pd.pd_num);
541 }
542 eq_context->intr = intr;
543 eq_context->lkey = cpu_to_be32(eq->mr.ibmr.lkey);
544
545 err = mthca_SW2HW_EQ(dev, mailbox, eq->eqn);
546 if (err) {
547 mthca_warn(dev, "SW2HW_EQ returned %d\n", err);
548 goto err_out_free_mr;
549 }
550
551 kfree(dma_list);
552 mthca_free_mailbox(dev, mailbox);
553
554 eq->eqn_mask = swab32(1 << eq->eqn);
555 eq->cons_index = 0;
556
557 dev->eq_table.arm_mask |= eq->eqn_mask;
558
559 mthca_dbg(dev, "Allocated EQ %d with %d entries\n",
560 eq->eqn, eq->nent);
561
562 return err;
563
564 err_out_free_mr:
565 mthca_free_mr(dev, &eq->mr);
566
567 err_out_free_eq:
568 mthca_free(&dev->eq_table.alloc, eq->eqn);
569
570 err_out_free_pages:
571 for (i = 0; i < npages; ++i)
572 if (eq->page_list[i].buf)
573 dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
574 eq->page_list[i].buf,
575 dma_unmap_addr(&eq->page_list[i],
576 mapping));
577
578 mthca_free_mailbox(dev, mailbox);
579
580 err_out_free:
581 kfree(eq->page_list);
582 kfree(dma_list);
583
584 err_out:
585 return err;
586 }
587
mthca_free_eq(struct mthca_dev * dev,struct mthca_eq * eq)588 static void mthca_free_eq(struct mthca_dev *dev,
589 struct mthca_eq *eq)
590 {
591 struct mthca_mailbox *mailbox;
592 int err;
593 int npages = (eq->nent * MTHCA_EQ_ENTRY_SIZE + PAGE_SIZE - 1) /
594 PAGE_SIZE;
595 int i;
596
597 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
598 if (IS_ERR(mailbox))
599 return;
600
601 err = mthca_HW2SW_EQ(dev, mailbox, eq->eqn);
602 if (err)
603 mthca_warn(dev, "HW2SW_EQ returned %d\n", err);
604
605 dev->eq_table.arm_mask &= ~eq->eqn_mask;
606
607 if (0) {
608 mthca_dbg(dev, "Dumping EQ context %02x:\n", eq->eqn);
609 for (i = 0; i < sizeof (struct mthca_eq_context) / 4; ++i) {
610 if (i % 4 == 0)
611 printk("[%02x] ", i * 4);
612 printk(" %08x", be32_to_cpup(mailbox->buf + i * 4));
613 if ((i + 1) % 4 == 0)
614 printk("\n");
615 }
616 }
617
618 mthca_free_mr(dev, &eq->mr);
619 for (i = 0; i < npages; ++i)
620 pci_free_consistent(dev->pdev, PAGE_SIZE,
621 eq->page_list[i].buf,
622 dma_unmap_addr(&eq->page_list[i], mapping));
623
624 kfree(eq->page_list);
625 mthca_free_mailbox(dev, mailbox);
626 }
627
mthca_free_irqs(struct mthca_dev * dev)628 static void mthca_free_irqs(struct mthca_dev *dev)
629 {
630 int i;
631
632 if (dev->eq_table.have_irq)
633 free_irq(dev->pdev->irq, dev);
634 for (i = 0; i < MTHCA_NUM_EQ; ++i)
635 if (dev->eq_table.eq[i].have_irq) {
636 free_irq(dev->eq_table.eq[i].msi_x_vector,
637 dev->eq_table.eq + i);
638 dev->eq_table.eq[i].have_irq = 0;
639 }
640 }
641
mthca_map_reg(struct mthca_dev * dev,unsigned long offset,unsigned long size,void __iomem ** map)642 static int mthca_map_reg(struct mthca_dev *dev,
643 unsigned long offset, unsigned long size,
644 void __iomem **map)
645 {
646 phys_addr_t base = pci_resource_start(dev->pdev, 0);
647
648 *map = ioremap(base + offset, size);
649 if (!*map)
650 return -ENOMEM;
651
652 return 0;
653 }
654
mthca_map_eq_regs(struct mthca_dev * dev)655 static int mthca_map_eq_regs(struct mthca_dev *dev)
656 {
657 if (mthca_is_memfree(dev)) {
658 /*
659 * We assume that the EQ arm and EQ set CI registers
660 * fall within the first BAR. We can't trust the
661 * values firmware gives us, since those addresses are
662 * valid on the HCA's side of the PCI bus but not
663 * necessarily the host side.
664 */
665 if (mthca_map_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
666 dev->fw.arbel.clr_int_base, MTHCA_CLR_INT_SIZE,
667 &dev->clr_base)) {
668 mthca_err(dev, "Couldn't map interrupt clear register, "
669 "aborting.\n");
670 return -ENOMEM;
671 }
672
673 /*
674 * Add 4 because we limit ourselves to EQs 0 ... 31,
675 * so we only need the low word of the register.
676 */
677 if (mthca_map_reg(dev, ((pci_resource_len(dev->pdev, 0) - 1) &
678 dev->fw.arbel.eq_arm_base) + 4, 4,
679 &dev->eq_regs.arbel.eq_arm)) {
680 mthca_err(dev, "Couldn't map EQ arm register, aborting.\n");
681 iounmap(dev->clr_base);
682 return -ENOMEM;
683 }
684
685 if (mthca_map_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
686 dev->fw.arbel.eq_set_ci_base,
687 MTHCA_EQ_SET_CI_SIZE,
688 &dev->eq_regs.arbel.eq_set_ci_base)) {
689 mthca_err(dev, "Couldn't map EQ CI register, aborting.\n");
690 iounmap(dev->eq_regs.arbel.eq_arm);
691 iounmap(dev->clr_base);
692 return -ENOMEM;
693 }
694 } else {
695 if (mthca_map_reg(dev, MTHCA_CLR_INT_BASE, MTHCA_CLR_INT_SIZE,
696 &dev->clr_base)) {
697 mthca_err(dev, "Couldn't map interrupt clear register, "
698 "aborting.\n");
699 return -ENOMEM;
700 }
701
702 if (mthca_map_reg(dev, MTHCA_ECR_BASE,
703 MTHCA_ECR_SIZE + MTHCA_ECR_CLR_SIZE,
704 &dev->eq_regs.tavor.ecr_base)) {
705 mthca_err(dev, "Couldn't map ecr register, "
706 "aborting.\n");
707 iounmap(dev->clr_base);
708 return -ENOMEM;
709 }
710 }
711
712 return 0;
713
714 }
715
mthca_unmap_eq_regs(struct mthca_dev * dev)716 static void mthca_unmap_eq_regs(struct mthca_dev *dev)
717 {
718 if (mthca_is_memfree(dev)) {
719 iounmap(dev->eq_regs.arbel.eq_set_ci_base);
720 iounmap(dev->eq_regs.arbel.eq_arm);
721 iounmap(dev->clr_base);
722 } else {
723 iounmap(dev->eq_regs.tavor.ecr_base);
724 iounmap(dev->clr_base);
725 }
726 }
727
mthca_map_eq_icm(struct mthca_dev * dev,u64 icm_virt)728 int mthca_map_eq_icm(struct mthca_dev *dev, u64 icm_virt)
729 {
730 int ret;
731
732 /*
733 * We assume that mapping one page is enough for the whole EQ
734 * context table. This is fine with all current HCAs, because
735 * we only use 32 EQs and each EQ uses 32 bytes of context
736 * memory, or 1 KB total.
737 */
738 dev->eq_table.icm_virt = icm_virt;
739 dev->eq_table.icm_page = alloc_page(GFP_HIGHUSER);
740 if (!dev->eq_table.icm_page)
741 return -ENOMEM;
742 dev->eq_table.icm_dma = pci_map_page(dev->pdev, dev->eq_table.icm_page, 0,
743 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
744 if (pci_dma_mapping_error(dev->pdev, dev->eq_table.icm_dma)) {
745 __free_page(dev->eq_table.icm_page);
746 return -ENOMEM;
747 }
748
749 ret = mthca_MAP_ICM_page(dev, dev->eq_table.icm_dma, icm_virt);
750 if (ret) {
751 pci_unmap_page(dev->pdev, dev->eq_table.icm_dma, PAGE_SIZE,
752 PCI_DMA_BIDIRECTIONAL);
753 __free_page(dev->eq_table.icm_page);
754 }
755
756 return ret;
757 }
758
mthca_unmap_eq_icm(struct mthca_dev * dev)759 void mthca_unmap_eq_icm(struct mthca_dev *dev)
760 {
761 mthca_UNMAP_ICM(dev, dev->eq_table.icm_virt, 1);
762 pci_unmap_page(dev->pdev, dev->eq_table.icm_dma, PAGE_SIZE,
763 PCI_DMA_BIDIRECTIONAL);
764 __free_page(dev->eq_table.icm_page);
765 }
766
mthca_init_eq_table(struct mthca_dev * dev)767 int mthca_init_eq_table(struct mthca_dev *dev)
768 {
769 int err;
770 u8 intr;
771 int i;
772
773 err = mthca_alloc_init(&dev->eq_table.alloc,
774 dev->limits.num_eqs,
775 dev->limits.num_eqs - 1,
776 dev->limits.reserved_eqs);
777 if (err)
778 return err;
779
780 err = mthca_map_eq_regs(dev);
781 if (err)
782 goto err_out_free;
783
784 if (dev->mthca_flags & MTHCA_FLAG_MSI_X) {
785 dev->eq_table.clr_mask = 0;
786 } else {
787 dev->eq_table.clr_mask =
788 swab32(1 << (dev->eq_table.inta_pin & 31));
789 dev->eq_table.clr_int = dev->clr_base +
790 (dev->eq_table.inta_pin < 32 ? 4 : 0);
791 }
792
793 dev->eq_table.arm_mask = 0;
794
795 intr = dev->eq_table.inta_pin;
796
797 err = mthca_create_eq(dev, dev->limits.num_cqs + MTHCA_NUM_SPARE_EQE,
798 (dev->mthca_flags & MTHCA_FLAG_MSI_X) ? 128 : intr,
799 &dev->eq_table.eq[MTHCA_EQ_COMP]);
800 if (err)
801 goto err_out_unmap;
802
803 err = mthca_create_eq(dev, MTHCA_NUM_ASYNC_EQE + MTHCA_NUM_SPARE_EQE,
804 (dev->mthca_flags & MTHCA_FLAG_MSI_X) ? 129 : intr,
805 &dev->eq_table.eq[MTHCA_EQ_ASYNC]);
806 if (err)
807 goto err_out_comp;
808
809 err = mthca_create_eq(dev, MTHCA_NUM_CMD_EQE + MTHCA_NUM_SPARE_EQE,
810 (dev->mthca_flags & MTHCA_FLAG_MSI_X) ? 130 : intr,
811 &dev->eq_table.eq[MTHCA_EQ_CMD]);
812 if (err)
813 goto err_out_async;
814
815 if (dev->mthca_flags & MTHCA_FLAG_MSI_X) {
816 static const char *eq_name[] = {
817 [MTHCA_EQ_COMP] = DRV_NAME "-comp",
818 [MTHCA_EQ_ASYNC] = DRV_NAME "-async",
819 [MTHCA_EQ_CMD] = DRV_NAME "-cmd"
820 };
821
822 for (i = 0; i < MTHCA_NUM_EQ; ++i) {
823 snprintf(dev->eq_table.eq[i].irq_name,
824 IB_DEVICE_NAME_MAX,
825 "%s@pci:%s", eq_name[i],
826 pci_name(dev->pdev));
827 err = request_irq(dev->eq_table.eq[i].msi_x_vector,
828 mthca_is_memfree(dev) ?
829 mthca_arbel_msi_x_interrupt :
830 mthca_tavor_msi_x_interrupt,
831 0, dev->eq_table.eq[i].irq_name,
832 dev->eq_table.eq + i);
833 if (err)
834 goto err_out_cmd;
835 dev->eq_table.eq[i].have_irq = 1;
836 }
837 } else {
838 snprintf(dev->eq_table.eq[0].irq_name, IB_DEVICE_NAME_MAX,
839 DRV_NAME "@pci:%s", pci_name(dev->pdev));
840 err = request_irq(dev->pdev->irq,
841 mthca_is_memfree(dev) ?
842 mthca_arbel_interrupt :
843 mthca_tavor_interrupt,
844 IRQF_SHARED, dev->eq_table.eq[0].irq_name, dev);
845 if (err)
846 goto err_out_cmd;
847 dev->eq_table.have_irq = 1;
848 }
849
850 err = mthca_MAP_EQ(dev, async_mask(dev),
851 0, dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
852 if (err)
853 mthca_warn(dev, "MAP_EQ for async EQ %d failed (%d)\n",
854 dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn, err);
855
856 err = mthca_MAP_EQ(dev, MTHCA_CMD_EVENT_MASK,
857 0, dev->eq_table.eq[MTHCA_EQ_CMD].eqn);
858 if (err)
859 mthca_warn(dev, "MAP_EQ for cmd EQ %d failed (%d)\n",
860 dev->eq_table.eq[MTHCA_EQ_CMD].eqn, err);
861
862 for (i = 0; i < MTHCA_NUM_EQ; ++i)
863 if (mthca_is_memfree(dev))
864 arbel_eq_req_not(dev, dev->eq_table.eq[i].eqn_mask);
865 else
866 tavor_eq_req_not(dev, dev->eq_table.eq[i].eqn);
867
868 return 0;
869
870 err_out_cmd:
871 mthca_free_irqs(dev);
872 mthca_free_eq(dev, &dev->eq_table.eq[MTHCA_EQ_CMD]);
873
874 err_out_async:
875 mthca_free_eq(dev, &dev->eq_table.eq[MTHCA_EQ_ASYNC]);
876
877 err_out_comp:
878 mthca_free_eq(dev, &dev->eq_table.eq[MTHCA_EQ_COMP]);
879
880 err_out_unmap:
881 mthca_unmap_eq_regs(dev);
882
883 err_out_free:
884 mthca_alloc_cleanup(&dev->eq_table.alloc);
885 return err;
886 }
887
mthca_cleanup_eq_table(struct mthca_dev * dev)888 void mthca_cleanup_eq_table(struct mthca_dev *dev)
889 {
890 int i;
891
892 mthca_free_irqs(dev);
893
894 mthca_MAP_EQ(dev, async_mask(dev),
895 1, dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn);
896 mthca_MAP_EQ(dev, MTHCA_CMD_EVENT_MASK,
897 1, dev->eq_table.eq[MTHCA_EQ_CMD].eqn);
898
899 for (i = 0; i < MTHCA_NUM_EQ; ++i)
900 mthca_free_eq(dev, &dev->eq_table.eq[i]);
901
902 mthca_unmap_eq_regs(dev);
903
904 mthca_alloc_cleanup(&dev->eq_table.alloc);
905 }
906