xref: /linux/drivers/infiniband/hw/efa/efa_com.c (revision c7decec2f2d2ab0366567f9e30c0e1418cece43f)
1 // SPDX-License-Identifier: GPL-2.0 OR BSD-2-Clause
2 /*
3  * Copyright 2018-2025 Amazon.com, Inc. or its affiliates. All rights reserved.
4  */
5 
6 #include <linux/log2.h>
7 
8 #include "efa_com.h"
9 #include "efa_regs_defs.h"
10 
11 #define ADMIN_CMD_TIMEOUT_US 30000000 /* usecs */
12 
13 #define EFA_REG_READ_TIMEOUT_US 50000 /* usecs */
14 #define EFA_MMIO_READ_INVALID 0xffffffff
15 
16 #define EFA_POLL_INTERVAL_MS 100 /* msecs */
17 
18 #define EFA_ASYNC_QUEUE_DEPTH 16
19 #define EFA_ADMIN_QUEUE_DEPTH 32
20 
21 #define EFA_CTRL_MAJOR          0
22 #define EFA_CTRL_MINOR          0
23 #define EFA_CTRL_SUB_MINOR      1
24 
25 enum efa_cmd_status {
26 	EFA_CMD_UNUSED,
27 	EFA_CMD_ALLOCATED,
28 	EFA_CMD_SUBMITTED,
29 	EFA_CMD_COMPLETED,
30 };
31 
32 struct efa_comp_ctx {
33 	struct completion wait_event;
34 	struct efa_admin_acq_entry *user_cqe;
35 	u32 comp_size;
36 	enum efa_cmd_status status;
37 	u16 cmd_id;
38 	u8 cmd_opcode;
39 };
40 
41 static const char *efa_com_cmd_str(u8 cmd)
42 {
43 #define EFA_CMD_STR_CASE(_cmd) case EFA_ADMIN_##_cmd: return #_cmd
44 
45 	switch (cmd) {
46 	EFA_CMD_STR_CASE(CREATE_QP);
47 	EFA_CMD_STR_CASE(MODIFY_QP);
48 	EFA_CMD_STR_CASE(QUERY_QP);
49 	EFA_CMD_STR_CASE(DESTROY_QP);
50 	EFA_CMD_STR_CASE(CREATE_AH);
51 	EFA_CMD_STR_CASE(DESTROY_AH);
52 	EFA_CMD_STR_CASE(REG_MR);
53 	EFA_CMD_STR_CASE(DEREG_MR);
54 	EFA_CMD_STR_CASE(CREATE_CQ);
55 	EFA_CMD_STR_CASE(DESTROY_CQ);
56 	EFA_CMD_STR_CASE(GET_FEATURE);
57 	EFA_CMD_STR_CASE(SET_FEATURE);
58 	EFA_CMD_STR_CASE(GET_STATS);
59 	EFA_CMD_STR_CASE(ALLOC_PD);
60 	EFA_CMD_STR_CASE(DEALLOC_PD);
61 	EFA_CMD_STR_CASE(ALLOC_UAR);
62 	EFA_CMD_STR_CASE(DEALLOC_UAR);
63 	EFA_CMD_STR_CASE(CREATE_EQ);
64 	EFA_CMD_STR_CASE(DESTROY_EQ);
65 	default: return "unknown command opcode";
66 	}
67 #undef EFA_CMD_STR_CASE
68 }
69 
70 void efa_com_set_dma_addr(dma_addr_t addr, u32 *addr_high, u32 *addr_low)
71 {
72 	*addr_low = lower_32_bits(addr);
73 	*addr_high = upper_32_bits(addr);
74 }
75 
76 static u32 efa_com_reg_read32(struct efa_com_dev *edev, u16 offset)
77 {
78 	struct efa_com_mmio_read *mmio_read = &edev->mmio_read;
79 	struct efa_admin_mmio_req_read_less_resp *read_resp;
80 	unsigned long exp_time;
81 	u32 mmio_read_reg = 0;
82 	u32 err;
83 
84 	read_resp = mmio_read->read_resp;
85 
86 	spin_lock(&mmio_read->lock);
87 	mmio_read->seq_num++;
88 
89 	/* trash DMA req_id to identify when hardware is done */
90 	read_resp->req_id = mmio_read->seq_num + 0x9aL;
91 	EFA_SET(&mmio_read_reg, EFA_REGS_MMIO_REG_READ_REG_OFF, offset);
92 	EFA_SET(&mmio_read_reg, EFA_REGS_MMIO_REG_READ_REQ_ID,
93 		mmio_read->seq_num);
94 
95 	writel(mmio_read_reg, edev->reg_bar + EFA_REGS_MMIO_REG_READ_OFF);
96 
97 	exp_time = jiffies + usecs_to_jiffies(mmio_read->mmio_read_timeout);
98 	do {
99 		if (READ_ONCE(read_resp->req_id) == mmio_read->seq_num)
100 			break;
101 		udelay(1);
102 	} while (time_is_after_jiffies(exp_time));
103 
104 	if (read_resp->req_id != mmio_read->seq_num) {
105 		ibdev_err_ratelimited(
106 			edev->efa_dev,
107 			"Reading register timed out. expected: req id[%u] offset[%#x] actual: req id[%u] offset[%#x]\n",
108 			mmio_read->seq_num, offset, read_resp->req_id,
109 			read_resp->reg_off);
110 		err = EFA_MMIO_READ_INVALID;
111 		goto out;
112 	}
113 
114 	if (read_resp->reg_off != offset) {
115 		ibdev_err_ratelimited(
116 			edev->efa_dev,
117 			"Reading register failed: wrong offset provided\n");
118 		err = EFA_MMIO_READ_INVALID;
119 		goto out;
120 	}
121 
122 	err = read_resp->reg_val;
123 out:
124 	spin_unlock(&mmio_read->lock);
125 	return err;
126 }
127 
128 static int efa_com_admin_init_sq(struct efa_com_dev *edev)
129 {
130 	struct efa_com_admin_queue *aq = &edev->aq;
131 	struct efa_com_admin_sq *sq = &aq->sq;
132 	u16 size = aq->depth * sizeof(*sq->entries);
133 	u32 aq_caps = 0;
134 	u32 addr_high;
135 	u32 addr_low;
136 
137 	sq->entries =
138 		dma_alloc_coherent(aq->dmadev, size, &sq->dma_addr, GFP_KERNEL);
139 	if (!sq->entries)
140 		return -ENOMEM;
141 
142 	spin_lock_init(&sq->lock);
143 
144 	sq->cc = 0;
145 	sq->pc = 0;
146 	sq->phase = 1;
147 
148 	sq->db_addr = (u32 __iomem *)(edev->reg_bar + EFA_REGS_AQ_PROD_DB_OFF);
149 
150 	addr_high = upper_32_bits(sq->dma_addr);
151 	addr_low = lower_32_bits(sq->dma_addr);
152 
153 	writel(addr_low, edev->reg_bar + EFA_REGS_AQ_BASE_LO_OFF);
154 	writel(addr_high, edev->reg_bar + EFA_REGS_AQ_BASE_HI_OFF);
155 
156 	EFA_SET(&aq_caps, EFA_REGS_AQ_CAPS_AQ_DEPTH, aq->depth);
157 	EFA_SET(&aq_caps, EFA_REGS_AQ_CAPS_AQ_ENTRY_SIZE,
158 		sizeof(struct efa_admin_aq_entry));
159 
160 	writel(aq_caps, edev->reg_bar + EFA_REGS_AQ_CAPS_OFF);
161 
162 	return 0;
163 }
164 
165 static int efa_com_admin_init_cq(struct efa_com_dev *edev)
166 {
167 	struct efa_com_admin_queue *aq = &edev->aq;
168 	struct efa_com_admin_cq *cq = &aq->cq;
169 	u16 size = aq->depth * sizeof(*cq->entries);
170 	u32 acq_caps = 0;
171 	u32 addr_high;
172 	u32 addr_low;
173 
174 	cq->entries =
175 		dma_alloc_coherent(aq->dmadev, size, &cq->dma_addr, GFP_KERNEL);
176 	if (!cq->entries)
177 		return -ENOMEM;
178 
179 	spin_lock_init(&cq->lock);
180 
181 	cq->cc = 0;
182 	cq->phase = 1;
183 
184 	addr_high = upper_32_bits(cq->dma_addr);
185 	addr_low = lower_32_bits(cq->dma_addr);
186 
187 	writel(addr_low, edev->reg_bar + EFA_REGS_ACQ_BASE_LO_OFF);
188 	writel(addr_high, edev->reg_bar + EFA_REGS_ACQ_BASE_HI_OFF);
189 
190 	EFA_SET(&acq_caps, EFA_REGS_ACQ_CAPS_ACQ_DEPTH, aq->depth);
191 	EFA_SET(&acq_caps, EFA_REGS_ACQ_CAPS_ACQ_ENTRY_SIZE,
192 		sizeof(struct efa_admin_acq_entry));
193 	EFA_SET(&acq_caps, EFA_REGS_ACQ_CAPS_ACQ_MSIX_VECTOR,
194 		aq->msix_vector_idx);
195 
196 	writel(acq_caps, edev->reg_bar + EFA_REGS_ACQ_CAPS_OFF);
197 
198 	return 0;
199 }
200 
201 static int efa_com_admin_init_aenq(struct efa_com_dev *edev,
202 				   struct efa_aenq_handlers *aenq_handlers)
203 {
204 	struct efa_com_aenq *aenq = &edev->aenq;
205 	u32 addr_low, addr_high;
206 	u32 aenq_caps = 0;
207 	u16 size;
208 
209 	if (!aenq_handlers) {
210 		ibdev_err(edev->efa_dev, "aenq handlers pointer is NULL\n");
211 		return -EINVAL;
212 	}
213 
214 	size = EFA_ASYNC_QUEUE_DEPTH * sizeof(*aenq->entries);
215 	aenq->entries = dma_alloc_coherent(edev->dmadev, size, &aenq->dma_addr,
216 					   GFP_KERNEL);
217 	if (!aenq->entries)
218 		return -ENOMEM;
219 
220 	aenq->aenq_handlers = aenq_handlers;
221 	aenq->depth = EFA_ASYNC_QUEUE_DEPTH;
222 	aenq->cc = 0;
223 	aenq->phase = 1;
224 
225 	addr_low = lower_32_bits(aenq->dma_addr);
226 	addr_high = upper_32_bits(aenq->dma_addr);
227 
228 	writel(addr_low, edev->reg_bar + EFA_REGS_AENQ_BASE_LO_OFF);
229 	writel(addr_high, edev->reg_bar + EFA_REGS_AENQ_BASE_HI_OFF);
230 
231 	EFA_SET(&aenq_caps, EFA_REGS_AENQ_CAPS_AENQ_DEPTH, aenq->depth);
232 	EFA_SET(&aenq_caps, EFA_REGS_AENQ_CAPS_AENQ_ENTRY_SIZE,
233 		sizeof(struct efa_admin_aenq_entry));
234 	EFA_SET(&aenq_caps, EFA_REGS_AENQ_CAPS_AENQ_MSIX_VECTOR,
235 		aenq->msix_vector_idx);
236 	writel(aenq_caps, edev->reg_bar + EFA_REGS_AENQ_CAPS_OFF);
237 
238 	/*
239 	 * Init cons_db to mark that all entries in the queue
240 	 * are initially available
241 	 */
242 	writel(edev->aenq.cc, edev->reg_bar + EFA_REGS_AENQ_CONS_DB_OFF);
243 
244 	return 0;
245 }
246 
247 static u16 efa_com_alloc_ctx_id(struct efa_com_admin_queue *aq)
248 {
249 	u16 ctx_id;
250 
251 	spin_lock(&aq->comp_ctx_lock);
252 	ctx_id = aq->comp_ctx_pool[aq->comp_ctx_pool_next];
253 	aq->comp_ctx_pool_next++;
254 	spin_unlock(&aq->comp_ctx_lock);
255 
256 	return ctx_id;
257 }
258 
259 static void efa_com_dealloc_ctx_id(struct efa_com_admin_queue *aq,
260 				   u16 ctx_id)
261 {
262 	spin_lock(&aq->comp_ctx_lock);
263 	aq->comp_ctx_pool_next--;
264 	aq->comp_ctx_pool[aq->comp_ctx_pool_next] = ctx_id;
265 	spin_unlock(&aq->comp_ctx_lock);
266 }
267 
268 static struct efa_comp_ctx *efa_com_alloc_comp_ctx(struct efa_com_admin_queue *aq)
269 {
270 	struct efa_comp_ctx *comp_ctx;
271 	u16 ctx_id;
272 
273 	ctx_id = efa_com_alloc_ctx_id(aq);
274 
275 	comp_ctx = &aq->comp_ctx[ctx_id];
276 	if (comp_ctx->status != EFA_CMD_UNUSED) {
277 		efa_com_dealloc_ctx_id(aq, ctx_id);
278 		ibdev_err_ratelimited(aq->efa_dev,
279 				      "Completion context[%u] is used[%u]\n",
280 				      ctx_id, comp_ctx->status);
281 		return NULL;
282 	}
283 
284 	comp_ctx->status = EFA_CMD_ALLOCATED;
285 	ibdev_dbg(aq->efa_dev, "Take completion context[%u]\n", ctx_id);
286 	return comp_ctx;
287 }
288 
289 static inline u16 efa_com_get_comp_ctx_id(struct efa_com_admin_queue *aq,
290 					  struct efa_comp_ctx *comp_ctx)
291 {
292 	return comp_ctx - aq->comp_ctx;
293 }
294 
295 static inline void efa_com_dealloc_comp_ctx(struct efa_com_admin_queue *aq,
296 					    struct efa_comp_ctx *comp_ctx)
297 {
298 	u16 ctx_id = efa_com_get_comp_ctx_id(aq, comp_ctx);
299 
300 	ibdev_dbg(aq->efa_dev, "Put completion context[%u]\n", ctx_id);
301 	comp_ctx->status = EFA_CMD_UNUSED;
302 	efa_com_dealloc_ctx_id(aq, ctx_id);
303 }
304 
305 static inline struct efa_comp_ctx *efa_com_get_comp_ctx_by_cmd_id(struct efa_com_admin_queue *aq,
306 								  u16 cmd_id)
307 {
308 	u16 ctx_id = cmd_id & (aq->depth - 1);
309 
310 	return &aq->comp_ctx[ctx_id];
311 }
312 
313 static struct efa_comp_ctx *__efa_com_submit_admin_cmd(struct efa_com_admin_queue *aq,
314 						       struct efa_admin_aq_entry *cmd,
315 						       size_t cmd_size_in_bytes,
316 						       struct efa_admin_acq_entry *comp,
317 						       size_t comp_size_in_bytes)
318 {
319 	struct efa_admin_aq_entry *aqe;
320 	struct efa_comp_ctx *comp_ctx;
321 	u16 queue_size_mask;
322 	u16 cmd_id;
323 	u16 ctx_id;
324 	u16 pi;
325 
326 	comp_ctx = efa_com_alloc_comp_ctx(aq);
327 	if (!comp_ctx)
328 		return ERR_PTR(-EINVAL);
329 
330 	queue_size_mask = aq->depth - 1;
331 	pi = aq->sq.pc & queue_size_mask;
332 	ctx_id = efa_com_get_comp_ctx_id(aq, comp_ctx);
333 
334 	/* cmd_id LSBs are the ctx_id and MSBs are entropy bits from pc */
335 	cmd_id = ctx_id & queue_size_mask;
336 	cmd_id |= aq->sq.pc << ilog2(aq->depth);
337 	cmd_id &= EFA_ADMIN_AQ_COMMON_DESC_COMMAND_ID_MASK;
338 
339 	cmd->aq_common_descriptor.command_id = cmd_id;
340 	EFA_SET(&cmd->aq_common_descriptor.flags,
341 		EFA_ADMIN_AQ_COMMON_DESC_PHASE, aq->sq.phase);
342 
343 	comp_ctx->status = EFA_CMD_SUBMITTED;
344 	comp_ctx->comp_size = comp_size_in_bytes;
345 	comp_ctx->user_cqe = comp;
346 	comp_ctx->cmd_opcode = cmd->aq_common_descriptor.opcode;
347 	comp_ctx->cmd_id = cmd_id;
348 
349 	reinit_completion(&comp_ctx->wait_event);
350 
351 	aqe = &aq->sq.entries[pi];
352 	memset(aqe, 0, sizeof(*aqe));
353 	memcpy(aqe, cmd, cmd_size_in_bytes);
354 
355 	aq->sq.pc++;
356 	atomic64_inc(&aq->stats.submitted_cmd);
357 
358 	if ((aq->sq.pc & queue_size_mask) == 0)
359 		aq->sq.phase = !aq->sq.phase;
360 
361 	/* barrier not needed in case of writel */
362 	writel(aq->sq.pc, aq->sq.db_addr);
363 
364 	return comp_ctx;
365 }
366 
367 static inline int efa_com_init_comp_ctxt(struct efa_com_admin_queue *aq)
368 {
369 	size_t pool_size = aq->depth * sizeof(*aq->comp_ctx_pool);
370 	size_t size = aq->depth * sizeof(struct efa_comp_ctx);
371 	struct efa_comp_ctx *comp_ctx;
372 	u16 i;
373 
374 	aq->comp_ctx = devm_kzalloc(aq->dmadev, size, GFP_KERNEL);
375 	aq->comp_ctx_pool = devm_kzalloc(aq->dmadev, pool_size, GFP_KERNEL);
376 	if (!aq->comp_ctx || !aq->comp_ctx_pool) {
377 		devm_kfree(aq->dmadev, aq->comp_ctx_pool);
378 		devm_kfree(aq->dmadev, aq->comp_ctx);
379 		return -ENOMEM;
380 	}
381 
382 	for (i = 0; i < aq->depth; i++) {
383 		comp_ctx = &aq->comp_ctx[i];
384 		comp_ctx->status = EFA_CMD_UNUSED;
385 		init_completion(&comp_ctx->wait_event);
386 
387 		aq->comp_ctx_pool[i] = i;
388 	}
389 
390 	spin_lock_init(&aq->comp_ctx_lock);
391 
392 	aq->comp_ctx_pool_next = 0;
393 
394 	return 0;
395 }
396 
397 static struct efa_comp_ctx *efa_com_submit_admin_cmd(struct efa_com_admin_queue *aq,
398 						     struct efa_admin_aq_entry *cmd,
399 						     size_t cmd_size_in_bytes,
400 						     struct efa_admin_acq_entry *comp,
401 						     size_t comp_size_in_bytes)
402 {
403 	struct efa_comp_ctx *comp_ctx;
404 
405 	spin_lock(&aq->sq.lock);
406 	if (!test_bit(EFA_AQ_STATE_RUNNING_BIT, &aq->state)) {
407 		ibdev_err_ratelimited(aq->efa_dev, "Admin queue is closed\n");
408 		spin_unlock(&aq->sq.lock);
409 		return ERR_PTR(-ENODEV);
410 	}
411 
412 	comp_ctx = __efa_com_submit_admin_cmd(aq, cmd, cmd_size_in_bytes, comp,
413 					      comp_size_in_bytes);
414 	spin_unlock(&aq->sq.lock);
415 	if (IS_ERR(comp_ctx))
416 		clear_bit(EFA_AQ_STATE_RUNNING_BIT, &aq->state);
417 
418 	return comp_ctx;
419 }
420 
421 static int efa_com_handle_single_admin_completion(struct efa_com_admin_queue *aq,
422 						  struct efa_admin_acq_entry *cqe)
423 {
424 	struct efa_comp_ctx *comp_ctx;
425 	u16 cmd_id;
426 
427 	cmd_id = EFA_GET(&cqe->acq_common_descriptor.command,
428 			 EFA_ADMIN_ACQ_COMMON_DESC_COMMAND_ID);
429 
430 	comp_ctx = efa_com_get_comp_ctx_by_cmd_id(aq, cmd_id);
431 	if (comp_ctx->status != EFA_CMD_SUBMITTED || comp_ctx->cmd_id != cmd_id) {
432 		ibdev_err(aq->efa_dev,
433 			  "Received completion with unexpected command id[%x], status[%d] sq producer[%d], sq consumer[%d], cq consumer[%d]\n",
434 			  cmd_id, comp_ctx->status, aq->sq.pc, aq->sq.cc,
435 			  aq->cq.cc);
436 		return -EINVAL;
437 	}
438 
439 	comp_ctx->status = EFA_CMD_COMPLETED;
440 	memcpy(comp_ctx->user_cqe, cqe, comp_ctx->comp_size);
441 
442 	if (!test_bit(EFA_AQ_STATE_POLLING_BIT, &aq->state))
443 		complete(&comp_ctx->wait_event);
444 
445 	return 0;
446 }
447 
448 static void efa_com_handle_admin_completion(struct efa_com_admin_queue *aq)
449 {
450 	struct efa_admin_acq_entry *cqe;
451 	u16 queue_size_mask;
452 	u16 comp_cmds = 0;
453 	u8 phase;
454 	int err;
455 	u16 ci;
456 
457 	queue_size_mask = aq->depth - 1;
458 
459 	ci = aq->cq.cc & queue_size_mask;
460 	phase = aq->cq.phase;
461 
462 	cqe = &aq->cq.entries[ci];
463 
464 	/* Go over all the completions */
465 	while ((READ_ONCE(cqe->acq_common_descriptor.flags) &
466 		EFA_ADMIN_ACQ_COMMON_DESC_PHASE_MASK) == phase) {
467 		/*
468 		 * Do not read the rest of the completion entry before the
469 		 * phase bit was validated
470 		 */
471 		dma_rmb();
472 		err = efa_com_handle_single_admin_completion(aq, cqe);
473 		if (!err)
474 			comp_cmds++;
475 
476 		aq->cq.cc++;
477 		ci++;
478 		if (ci == aq->depth) {
479 			ci = 0;
480 			phase = !phase;
481 		}
482 
483 		cqe = &aq->cq.entries[ci];
484 	}
485 
486 	aq->cq.phase = phase;
487 	aq->sq.cc += comp_cmds;
488 	atomic64_add(comp_cmds, &aq->stats.completed_cmd);
489 }
490 
491 static int efa_com_comp_status_to_errno(u8 comp_status)
492 {
493 	switch (comp_status) {
494 	case EFA_ADMIN_SUCCESS:
495 		return 0;
496 	case EFA_ADMIN_RESOURCE_ALLOCATION_FAILURE:
497 		return -ENOMEM;
498 	case EFA_ADMIN_UNSUPPORTED_OPCODE:
499 		return -EOPNOTSUPP;
500 	case EFA_ADMIN_BAD_OPCODE:
501 	case EFA_ADMIN_MALFORMED_REQUEST:
502 	case EFA_ADMIN_ILLEGAL_PARAMETER:
503 	case EFA_ADMIN_UNKNOWN_ERROR:
504 		return -EINVAL;
505 	default:
506 		return -EINVAL;
507 	}
508 }
509 
510 static int efa_com_wait_and_process_admin_cq_polling(struct efa_comp_ctx *comp_ctx,
511 						     struct efa_com_admin_queue *aq)
512 {
513 	unsigned long timeout;
514 	unsigned long flags;
515 	int err;
516 
517 	timeout = jiffies + usecs_to_jiffies(aq->completion_timeout);
518 
519 	while (1) {
520 		spin_lock_irqsave(&aq->cq.lock, flags);
521 		efa_com_handle_admin_completion(aq);
522 		spin_unlock_irqrestore(&aq->cq.lock, flags);
523 
524 		if (comp_ctx->status != EFA_CMD_SUBMITTED)
525 			break;
526 
527 		if (time_is_before_jiffies(timeout)) {
528 			ibdev_err_ratelimited(
529 				aq->efa_dev,
530 				"Wait for completion (polling) timeout\n");
531 			/* EFA didn't have any completion */
532 			atomic64_inc(&aq->stats.no_completion);
533 
534 			clear_bit(EFA_AQ_STATE_RUNNING_BIT, &aq->state);
535 			err = -ETIME;
536 			goto out;
537 		}
538 
539 		msleep(aq->poll_interval);
540 	}
541 
542 	err = efa_com_comp_status_to_errno(comp_ctx->user_cqe->acq_common_descriptor.status);
543 out:
544 	efa_com_dealloc_comp_ctx(aq, comp_ctx);
545 	return err;
546 }
547 
548 static int efa_com_wait_and_process_admin_cq_interrupts(struct efa_comp_ctx *comp_ctx,
549 							struct efa_com_admin_queue *aq)
550 {
551 	unsigned long flags;
552 	int err;
553 
554 	wait_for_completion_timeout(&comp_ctx->wait_event,
555 				    usecs_to_jiffies(aq->completion_timeout));
556 
557 	/*
558 	 * In case the command wasn't completed find out the root cause.
559 	 * There might be 2 kinds of errors
560 	 * 1) No completion (timeout reached)
561 	 * 2) There is completion but the device didn't get any msi-x interrupt.
562 	 */
563 	if (comp_ctx->status == EFA_CMD_SUBMITTED) {
564 		spin_lock_irqsave(&aq->cq.lock, flags);
565 		efa_com_handle_admin_completion(aq);
566 		spin_unlock_irqrestore(&aq->cq.lock, flags);
567 
568 		atomic64_inc(&aq->stats.no_completion);
569 
570 		if (comp_ctx->status == EFA_CMD_COMPLETED)
571 			ibdev_err_ratelimited(
572 				aq->efa_dev,
573 				"The device sent a completion but the driver didn't receive any MSI-X interrupt for admin cmd %s(%d) status %d (id: %d, sq producer: %d, sq consumer: %d, cq consumer: %d)\n",
574 				efa_com_cmd_str(comp_ctx->cmd_opcode),
575 				comp_ctx->cmd_opcode, comp_ctx->status,
576 				comp_ctx->cmd_id, aq->sq.pc, aq->sq.cc,
577 				aq->cq.cc);
578 		else
579 			ibdev_err_ratelimited(
580 				aq->efa_dev,
581 				"The device didn't send any completion for admin cmd %s(%d) status %d (id: %d, sq producer: %d, sq consumer: %d, cq consumer: %d)\n",
582 				efa_com_cmd_str(comp_ctx->cmd_opcode),
583 				comp_ctx->cmd_opcode, comp_ctx->status,
584 				comp_ctx->cmd_id, aq->sq.pc, aq->sq.cc,
585 				aq->cq.cc);
586 
587 		clear_bit(EFA_AQ_STATE_RUNNING_BIT, &aq->state);
588 		err = -ETIME;
589 		goto out;
590 	}
591 
592 	err = efa_com_comp_status_to_errno(comp_ctx->user_cqe->acq_common_descriptor.status);
593 out:
594 	efa_com_dealloc_comp_ctx(aq, comp_ctx);
595 	return err;
596 }
597 
598 /*
599  * There are two types to wait for completion.
600  * Polling mode - wait until the completion is available.
601  * Async mode - wait on wait queue until the completion is ready
602  * (or the timeout expired).
603  * It is expected that the IRQ called efa_com_handle_admin_completion
604  * to mark the completions.
605  */
606 static int efa_com_wait_and_process_admin_cq(struct efa_comp_ctx *comp_ctx,
607 					     struct efa_com_admin_queue *aq)
608 {
609 	if (test_bit(EFA_AQ_STATE_POLLING_BIT, &aq->state))
610 		return efa_com_wait_and_process_admin_cq_polling(comp_ctx, aq);
611 
612 	return efa_com_wait_and_process_admin_cq_interrupts(comp_ctx, aq);
613 }
614 
615 /**
616  * efa_com_cmd_exec - Execute admin command
617  * @aq: admin queue.
618  * @cmd: the admin command to execute.
619  * @cmd_size: the command size.
620  * @comp: command completion return entry.
621  * @comp_size: command completion size.
622  * Submit an admin command and then wait until the device will return a
623  * completion.
624  * The completion will be copied into comp.
625  *
626  * @return - 0 on success, negative value on failure.
627  */
628 int efa_com_cmd_exec(struct efa_com_admin_queue *aq,
629 		     struct efa_admin_aq_entry *cmd,
630 		     size_t cmd_size,
631 		     struct efa_admin_acq_entry *comp,
632 		     size_t comp_size)
633 {
634 	struct efa_comp_ctx *comp_ctx;
635 	int err;
636 
637 	might_sleep();
638 
639 	/* In case of queue FULL */
640 	down(&aq->avail_cmds);
641 
642 	ibdev_dbg(aq->efa_dev, "%s (opcode %d)\n",
643 		  efa_com_cmd_str(cmd->aq_common_descriptor.opcode),
644 		  cmd->aq_common_descriptor.opcode);
645 	comp_ctx = efa_com_submit_admin_cmd(aq, cmd, cmd_size, comp, comp_size);
646 	if (IS_ERR(comp_ctx)) {
647 		ibdev_err_ratelimited(
648 			aq->efa_dev,
649 			"Failed to submit command %s (opcode %u) err %pe\n",
650 			efa_com_cmd_str(cmd->aq_common_descriptor.opcode),
651 			cmd->aq_common_descriptor.opcode, comp_ctx);
652 
653 		up(&aq->avail_cmds);
654 		atomic64_inc(&aq->stats.cmd_err);
655 		return PTR_ERR(comp_ctx);
656 	}
657 
658 	err = efa_com_wait_and_process_admin_cq(comp_ctx, aq);
659 	if (err) {
660 		ibdev_err_ratelimited(
661 			aq->efa_dev,
662 			"Failed to process command %s (opcode %u) comp_status %d err %d\n",
663 			efa_com_cmd_str(cmd->aq_common_descriptor.opcode),
664 			cmd->aq_common_descriptor.opcode,
665 			comp_ctx->user_cqe->acq_common_descriptor.status, err);
666 		atomic64_inc(&aq->stats.cmd_err);
667 	}
668 
669 	up(&aq->avail_cmds);
670 
671 	return err;
672 }
673 
674 /**
675  * efa_com_admin_destroy - Destroy the admin and the async events queues.
676  * @edev: EFA communication layer struct
677  */
678 void efa_com_admin_destroy(struct efa_com_dev *edev)
679 {
680 	struct efa_com_admin_queue *aq = &edev->aq;
681 	struct efa_com_aenq *aenq = &edev->aenq;
682 	struct efa_com_admin_cq *cq = &aq->cq;
683 	struct efa_com_admin_sq *sq = &aq->sq;
684 	u16 size;
685 
686 	clear_bit(EFA_AQ_STATE_RUNNING_BIT, &aq->state);
687 
688 	devm_kfree(edev->dmadev, aq->comp_ctx_pool);
689 	devm_kfree(edev->dmadev, aq->comp_ctx);
690 
691 	size = aq->depth * sizeof(*sq->entries);
692 	dma_free_coherent(edev->dmadev, size, sq->entries, sq->dma_addr);
693 
694 	size = aq->depth * sizeof(*cq->entries);
695 	dma_free_coherent(edev->dmadev, size, cq->entries, cq->dma_addr);
696 
697 	size = aenq->depth * sizeof(*aenq->entries);
698 	dma_free_coherent(edev->dmadev, size, aenq->entries, aenq->dma_addr);
699 }
700 
701 /**
702  * efa_com_set_admin_polling_mode - Set the admin completion queue polling mode
703  * @edev: EFA communication layer struct
704  * @polling: Enable/Disable polling mode
705  *
706  * Set the admin completion mode.
707  */
708 void efa_com_set_admin_polling_mode(struct efa_com_dev *edev, bool polling)
709 {
710 	u32 mask_value = 0;
711 
712 	if (polling)
713 		EFA_SET(&mask_value, EFA_REGS_INTR_MASK_EN, 1);
714 
715 	writel(mask_value, edev->reg_bar + EFA_REGS_INTR_MASK_OFF);
716 	if (polling)
717 		set_bit(EFA_AQ_STATE_POLLING_BIT, &edev->aq.state);
718 	else
719 		clear_bit(EFA_AQ_STATE_POLLING_BIT, &edev->aq.state);
720 }
721 
722 static void efa_com_stats_init(struct efa_com_dev *edev)
723 {
724 	atomic64_t *s = (atomic64_t *)&edev->aq.stats;
725 	int i;
726 
727 	for (i = 0; i < sizeof(edev->aq.stats) / sizeof(*s); i++, s++)
728 		atomic64_set(s, 0);
729 }
730 
731 /**
732  * efa_com_admin_init - Init the admin and the async queues
733  * @edev: EFA communication layer struct
734  * @aenq_handlers: Those handlers to be called upon event.
735  *
736  * Initialize the admin submission and completion queues.
737  * Initialize the asynchronous events notification queues.
738  *
739  * @return - 0 on success, negative value on failure.
740  */
741 int efa_com_admin_init(struct efa_com_dev *edev,
742 		       struct efa_aenq_handlers *aenq_handlers)
743 {
744 	struct efa_com_admin_queue *aq = &edev->aq;
745 	u32 timeout;
746 	u32 dev_sts;
747 	u32 cap;
748 	int err;
749 
750 	dev_sts = efa_com_reg_read32(edev, EFA_REGS_DEV_STS_OFF);
751 	if (!EFA_GET(&dev_sts, EFA_REGS_DEV_STS_READY)) {
752 		ibdev_err(edev->efa_dev,
753 			  "Device isn't ready, abort com init %#x\n", dev_sts);
754 		return -ENODEV;
755 	}
756 
757 	aq->depth = EFA_ADMIN_QUEUE_DEPTH;
758 
759 	aq->dmadev = edev->dmadev;
760 	aq->efa_dev = edev->efa_dev;
761 	set_bit(EFA_AQ_STATE_POLLING_BIT, &aq->state);
762 
763 	sema_init(&aq->avail_cmds, aq->depth);
764 
765 	efa_com_stats_init(edev);
766 
767 	err = efa_com_init_comp_ctxt(aq);
768 	if (err)
769 		return err;
770 
771 	err = efa_com_admin_init_sq(edev);
772 	if (err)
773 		goto err_destroy_comp_ctxt;
774 
775 	err = efa_com_admin_init_cq(edev);
776 	if (err)
777 		goto err_destroy_sq;
778 
779 	efa_com_set_admin_polling_mode(edev, false);
780 
781 	err = efa_com_admin_init_aenq(edev, aenq_handlers);
782 	if (err)
783 		goto err_destroy_cq;
784 
785 	cap = efa_com_reg_read32(edev, EFA_REGS_CAPS_OFF);
786 	timeout = EFA_GET(&cap, EFA_REGS_CAPS_ADMIN_CMD_TO);
787 	if (timeout)
788 		/* the resolution of timeout reg is 100ms */
789 		aq->completion_timeout = timeout * 100000;
790 	else
791 		aq->completion_timeout = ADMIN_CMD_TIMEOUT_US;
792 
793 	aq->poll_interval = EFA_POLL_INTERVAL_MS;
794 
795 	set_bit(EFA_AQ_STATE_RUNNING_BIT, &aq->state);
796 
797 	return 0;
798 
799 err_destroy_cq:
800 	dma_free_coherent(edev->dmadev, aq->depth * sizeof(*aq->cq.entries),
801 			  aq->cq.entries, aq->cq.dma_addr);
802 err_destroy_sq:
803 	dma_free_coherent(edev->dmadev, aq->depth * sizeof(*aq->sq.entries),
804 			  aq->sq.entries, aq->sq.dma_addr);
805 err_destroy_comp_ctxt:
806 	devm_kfree(edev->dmadev, aq->comp_ctx);
807 
808 	return err;
809 }
810 
811 /**
812  * efa_com_admin_q_comp_intr_handler - admin queue interrupt handler
813  * @edev: EFA communication layer struct
814  *
815  * This method goes over the admin completion queue and wakes up
816  * all the pending threads that wait on the commands wait event.
817  *
818  * Note: Should be called after MSI-X interrupt.
819  */
820 void efa_com_admin_q_comp_intr_handler(struct efa_com_dev *edev)
821 {
822 	unsigned long flags;
823 
824 	spin_lock_irqsave(&edev->aq.cq.lock, flags);
825 	efa_com_handle_admin_completion(&edev->aq);
826 	spin_unlock_irqrestore(&edev->aq.cq.lock, flags);
827 }
828 
829 /*
830  * efa_handle_specific_aenq_event:
831  * return the handler that is relevant to the specific event group
832  */
833 static efa_aenq_handler efa_com_get_specific_aenq_cb(struct efa_com_dev *edev,
834 						     u16 group)
835 {
836 	struct efa_aenq_handlers *aenq_handlers = edev->aenq.aenq_handlers;
837 
838 	if (group < EFA_MAX_HANDLERS && aenq_handlers->handlers[group])
839 		return aenq_handlers->handlers[group];
840 
841 	return aenq_handlers->unimplemented_handler;
842 }
843 
844 /**
845  * efa_com_aenq_intr_handler - AENQ interrupt handler
846  * @edev: EFA communication layer struct
847  * @data: Data of interrupt handler.
848  *
849  * Go over the async event notification queue and call the proper aenq handler.
850  */
851 void efa_com_aenq_intr_handler(struct efa_com_dev *edev, void *data)
852 {
853 	struct efa_admin_aenq_common_desc *aenq_common;
854 	struct efa_com_aenq *aenq = &edev->aenq;
855 	struct efa_admin_aenq_entry *aenq_e;
856 	efa_aenq_handler handler_cb;
857 	u32 processed = 0;
858 	u8 phase;
859 	u32 ci;
860 
861 	ci = aenq->cc & (aenq->depth - 1);
862 	phase = aenq->phase;
863 	aenq_e = &aenq->entries[ci]; /* Get first entry */
864 	aenq_common = &aenq_e->aenq_common_desc;
865 
866 	/* Go over all the events */
867 	while ((READ_ONCE(aenq_common->flags) &
868 		EFA_ADMIN_AENQ_COMMON_DESC_PHASE_MASK) == phase) {
869 		/*
870 		 * Do not read the rest of the completion entry before the
871 		 * phase bit was validated
872 		 */
873 		dma_rmb();
874 
875 		/* Handle specific event*/
876 		handler_cb = efa_com_get_specific_aenq_cb(edev,
877 							  aenq_common->group);
878 		handler_cb(data, aenq_e); /* call the actual event handler*/
879 
880 		/* Get next event entry */
881 		ci++;
882 		processed++;
883 
884 		if (ci == aenq->depth) {
885 			ci = 0;
886 			phase = !phase;
887 		}
888 		aenq_e = &aenq->entries[ci];
889 		aenq_common = &aenq_e->aenq_common_desc;
890 	}
891 
892 	aenq->cc += processed;
893 	aenq->phase = phase;
894 
895 	/* Don't update aenq doorbell if there weren't any processed events */
896 	if (!processed)
897 		return;
898 
899 	/* barrier not needed in case of writel */
900 	writel(aenq->cc, edev->reg_bar + EFA_REGS_AENQ_CONS_DB_OFF);
901 }
902 
903 static void efa_com_mmio_reg_read_resp_addr_init(struct efa_com_dev *edev)
904 {
905 	struct efa_com_mmio_read *mmio_read = &edev->mmio_read;
906 	u32 addr_high;
907 	u32 addr_low;
908 
909 	/* dma_addr_bits is unknown at this point */
910 	addr_high = (mmio_read->read_resp_dma_addr >> 32) & GENMASK(31, 0);
911 	addr_low = mmio_read->read_resp_dma_addr & GENMASK(31, 0);
912 
913 	writel(addr_high, edev->reg_bar + EFA_REGS_MMIO_RESP_HI_OFF);
914 	writel(addr_low, edev->reg_bar + EFA_REGS_MMIO_RESP_LO_OFF);
915 }
916 
917 int efa_com_mmio_reg_read_init(struct efa_com_dev *edev)
918 {
919 	struct efa_com_mmio_read *mmio_read = &edev->mmio_read;
920 
921 	spin_lock_init(&mmio_read->lock);
922 	mmio_read->read_resp =
923 		dma_alloc_coherent(edev->dmadev, sizeof(*mmio_read->read_resp),
924 				   &mmio_read->read_resp_dma_addr, GFP_KERNEL);
925 	if (!mmio_read->read_resp)
926 		return -ENOMEM;
927 
928 	efa_com_mmio_reg_read_resp_addr_init(edev);
929 
930 	mmio_read->read_resp->req_id = 0;
931 	mmio_read->seq_num = 0;
932 	mmio_read->mmio_read_timeout = EFA_REG_READ_TIMEOUT_US;
933 
934 	return 0;
935 }
936 
937 void efa_com_mmio_reg_read_destroy(struct efa_com_dev *edev)
938 {
939 	struct efa_com_mmio_read *mmio_read = &edev->mmio_read;
940 
941 	dma_free_coherent(edev->dmadev, sizeof(*mmio_read->read_resp),
942 			  mmio_read->read_resp, mmio_read->read_resp_dma_addr);
943 }
944 
945 int efa_com_validate_version(struct efa_com_dev *edev)
946 {
947 	u32 min_ctrl_ver = 0;
948 	u32 ctrl_ver_masked;
949 	u32 min_ver = 0;
950 	u32 ctrl_ver;
951 	u32 ver;
952 
953 	/*
954 	 * Make sure the EFA version and the controller version are at least
955 	 * as the driver expects
956 	 */
957 	ver = efa_com_reg_read32(edev, EFA_REGS_VERSION_OFF);
958 	ctrl_ver = efa_com_reg_read32(edev,
959 				      EFA_REGS_CONTROLLER_VERSION_OFF);
960 
961 	ibdev_dbg(edev->efa_dev, "efa device version: %d.%d\n",
962 		  EFA_GET(&ver, EFA_REGS_VERSION_MAJOR_VERSION),
963 		  EFA_GET(&ver, EFA_REGS_VERSION_MINOR_VERSION));
964 
965 	EFA_SET(&min_ver, EFA_REGS_VERSION_MAJOR_VERSION,
966 		EFA_ADMIN_API_VERSION_MAJOR);
967 	EFA_SET(&min_ver, EFA_REGS_VERSION_MINOR_VERSION,
968 		EFA_ADMIN_API_VERSION_MINOR);
969 	if (ver < min_ver) {
970 		ibdev_err(edev->efa_dev,
971 			  "EFA version is lower than the minimal version the driver supports\n");
972 		return -EOPNOTSUPP;
973 	}
974 
975 	ibdev_dbg(
976 		edev->efa_dev,
977 		"efa controller version: %d.%d.%d implementation version %d\n",
978 		EFA_GET(&ctrl_ver, EFA_REGS_CONTROLLER_VERSION_MAJOR_VERSION),
979 		EFA_GET(&ctrl_ver, EFA_REGS_CONTROLLER_VERSION_MINOR_VERSION),
980 		EFA_GET(&ctrl_ver,
981 			EFA_REGS_CONTROLLER_VERSION_SUBMINOR_VERSION),
982 		EFA_GET(&ctrl_ver, EFA_REGS_CONTROLLER_VERSION_IMPL_ID));
983 
984 	ctrl_ver_masked =
985 		EFA_GET(&ctrl_ver, EFA_REGS_CONTROLLER_VERSION_MAJOR_VERSION) |
986 		EFA_GET(&ctrl_ver, EFA_REGS_CONTROLLER_VERSION_MINOR_VERSION) |
987 		EFA_GET(&ctrl_ver,
988 			EFA_REGS_CONTROLLER_VERSION_SUBMINOR_VERSION);
989 
990 	EFA_SET(&min_ctrl_ver, EFA_REGS_CONTROLLER_VERSION_MAJOR_VERSION,
991 		EFA_CTRL_MAJOR);
992 	EFA_SET(&min_ctrl_ver, EFA_REGS_CONTROLLER_VERSION_MINOR_VERSION,
993 		EFA_CTRL_MINOR);
994 	EFA_SET(&min_ctrl_ver, EFA_REGS_CONTROLLER_VERSION_SUBMINOR_VERSION,
995 		EFA_CTRL_SUB_MINOR);
996 	/* Validate the ctrl version without the implementation ID */
997 	if (ctrl_ver_masked < min_ctrl_ver) {
998 		ibdev_err(edev->efa_dev,
999 			  "EFA ctrl version is lower than the minimal ctrl version the driver supports\n");
1000 		return -EOPNOTSUPP;
1001 	}
1002 
1003 	return 0;
1004 }
1005 
1006 /**
1007  * efa_com_get_dma_width - Retrieve physical dma address width the device
1008  * supports.
1009  * @edev: EFA communication layer struct
1010  *
1011  * Retrieve the maximum physical address bits the device can handle.
1012  *
1013  * @return: > 0 on Success and negative value otherwise.
1014  */
1015 int efa_com_get_dma_width(struct efa_com_dev *edev)
1016 {
1017 	u32 caps = efa_com_reg_read32(edev, EFA_REGS_CAPS_OFF);
1018 	int width;
1019 
1020 	width = EFA_GET(&caps, EFA_REGS_CAPS_DMA_ADDR_WIDTH);
1021 
1022 	ibdev_dbg(edev->efa_dev, "DMA width: %d\n", width);
1023 
1024 	if (width < 32 || width > 64) {
1025 		ibdev_err(edev->efa_dev, "DMA width illegal value: %d\n", width);
1026 		return -EINVAL;
1027 	}
1028 
1029 	edev->dma_addr_bits = width;
1030 
1031 	return width;
1032 }
1033 
1034 static int wait_for_reset_state(struct efa_com_dev *edev, u32 timeout, int on)
1035 {
1036 	u32 val, i;
1037 
1038 	for (i = 0; i < timeout; i++) {
1039 		val = efa_com_reg_read32(edev, EFA_REGS_DEV_STS_OFF);
1040 
1041 		if (EFA_GET(&val, EFA_REGS_DEV_STS_RESET_IN_PROGRESS) == on)
1042 			return 0;
1043 
1044 		ibdev_dbg(edev->efa_dev, "Reset indication val %d\n", val);
1045 		msleep(EFA_POLL_INTERVAL_MS);
1046 	}
1047 
1048 	return -ETIME;
1049 }
1050 
1051 /**
1052  * efa_com_dev_reset - Perform device FLR to the device.
1053  * @edev: EFA communication layer struct
1054  * @reset_reason: Specify what is the trigger for the reset in case of an error.
1055  *
1056  * @return - 0 on success, negative value on failure.
1057  */
1058 int efa_com_dev_reset(struct efa_com_dev *edev,
1059 		      enum efa_regs_reset_reason_types reset_reason)
1060 {
1061 	u32 stat, timeout, cap;
1062 	u32 reset_val = 0;
1063 	int err;
1064 
1065 	stat = efa_com_reg_read32(edev, EFA_REGS_DEV_STS_OFF);
1066 	cap = efa_com_reg_read32(edev, EFA_REGS_CAPS_OFF);
1067 
1068 	if (!EFA_GET(&stat, EFA_REGS_DEV_STS_READY)) {
1069 		ibdev_err(edev->efa_dev,
1070 			  "Device isn't ready, can't reset device\n");
1071 		return -EINVAL;
1072 	}
1073 
1074 	timeout = EFA_GET(&cap, EFA_REGS_CAPS_RESET_TIMEOUT);
1075 	if (!timeout) {
1076 		ibdev_err(edev->efa_dev, "Invalid timeout value\n");
1077 		return -EINVAL;
1078 	}
1079 
1080 	/* start reset */
1081 	EFA_SET(&reset_val, EFA_REGS_DEV_CTL_DEV_RESET, 1);
1082 	EFA_SET(&reset_val, EFA_REGS_DEV_CTL_RESET_REASON, reset_reason);
1083 	writel(reset_val, edev->reg_bar + EFA_REGS_DEV_CTL_OFF);
1084 
1085 	/* reset clears the mmio readless address, restore it */
1086 	efa_com_mmio_reg_read_resp_addr_init(edev);
1087 
1088 	err = wait_for_reset_state(edev, timeout, 1);
1089 	if (err) {
1090 		ibdev_err(edev->efa_dev, "Reset indication didn't turn on\n");
1091 		return err;
1092 	}
1093 
1094 	/* reset done */
1095 	writel(0, edev->reg_bar + EFA_REGS_DEV_CTL_OFF);
1096 	err = wait_for_reset_state(edev, timeout, 0);
1097 	if (err) {
1098 		ibdev_err(edev->efa_dev, "Reset indication didn't turn off\n");
1099 		return err;
1100 	}
1101 
1102 	timeout = EFA_GET(&cap, EFA_REGS_CAPS_ADMIN_CMD_TO);
1103 	if (timeout)
1104 		/* the resolution of timeout reg is 100ms */
1105 		edev->aq.completion_timeout = timeout * 100000;
1106 	else
1107 		edev->aq.completion_timeout = ADMIN_CMD_TIMEOUT_US;
1108 
1109 	return 0;
1110 }
1111 
1112 static int efa_com_create_eq(struct efa_com_dev *edev,
1113 			     struct efa_com_create_eq_params *params,
1114 			     struct efa_com_create_eq_result *result)
1115 {
1116 	struct efa_com_admin_queue *aq = &edev->aq;
1117 	struct efa_admin_create_eq_resp resp = {};
1118 	struct efa_admin_create_eq_cmd cmd = {};
1119 	int err;
1120 
1121 	cmd.aq_common_descriptor.opcode = EFA_ADMIN_CREATE_EQ;
1122 	EFA_SET(&cmd.caps, EFA_ADMIN_CREATE_EQ_CMD_ENTRY_SIZE_WORDS,
1123 		params->entry_size_in_bytes / 4);
1124 	cmd.depth = params->depth;
1125 	cmd.event_bitmask = params->event_bitmask;
1126 	cmd.msix_vec = params->msix_vec;
1127 
1128 	efa_com_set_dma_addr(params->dma_addr, &cmd.ba.mem_addr_high,
1129 			     &cmd.ba.mem_addr_low);
1130 
1131 	err = efa_com_cmd_exec(aq,
1132 			       (struct efa_admin_aq_entry *)&cmd,
1133 			       sizeof(cmd),
1134 			       (struct efa_admin_acq_entry *)&resp,
1135 			       sizeof(resp));
1136 	if (err) {
1137 		ibdev_err_ratelimited(edev->efa_dev,
1138 				      "Failed to create eq[%d]\n", err);
1139 		return err;
1140 	}
1141 
1142 	result->eqn = resp.eqn;
1143 
1144 	return 0;
1145 }
1146 
1147 static void efa_com_destroy_eq(struct efa_com_dev *edev,
1148 			       struct efa_com_destroy_eq_params *params)
1149 {
1150 	struct efa_com_admin_queue *aq = &edev->aq;
1151 	struct efa_admin_destroy_eq_resp resp = {};
1152 	struct efa_admin_destroy_eq_cmd cmd = {};
1153 	int err;
1154 
1155 	cmd.aq_common_descriptor.opcode = EFA_ADMIN_DESTROY_EQ;
1156 	cmd.eqn = params->eqn;
1157 
1158 	err = efa_com_cmd_exec(aq,
1159 			       (struct efa_admin_aq_entry *)&cmd,
1160 			       sizeof(cmd),
1161 			       (struct efa_admin_acq_entry *)&resp,
1162 			       sizeof(resp));
1163 	if (err)
1164 		ibdev_err_ratelimited(edev->efa_dev,
1165 				      "Failed to destroy EQ-%u [%d]\n", cmd.eqn,
1166 				      err);
1167 }
1168 
1169 static void efa_com_arm_eq(struct efa_com_dev *edev, struct efa_com_eq *eeq)
1170 {
1171 	u32 val = 0;
1172 
1173 	EFA_SET(&val, EFA_REGS_EQ_DB_EQN, eeq->eqn);
1174 	EFA_SET(&val, EFA_REGS_EQ_DB_ARM, 1);
1175 
1176 	writel(val, edev->reg_bar + EFA_REGS_EQ_DB_OFF);
1177 }
1178 
1179 void efa_com_eq_comp_intr_handler(struct efa_com_dev *edev,
1180 				  struct efa_com_eq *eeq)
1181 {
1182 	struct efa_admin_eqe *eqe;
1183 	u32 processed = 0;
1184 	u8 phase;
1185 	u32 ci;
1186 
1187 	ci = eeq->cc & (eeq->depth - 1);
1188 	phase = eeq->phase;
1189 	eqe = &eeq->eqes[ci];
1190 
1191 	/* Go over all the events */
1192 	while ((READ_ONCE(eqe->common) & EFA_ADMIN_EQE_PHASE_MASK) == phase) {
1193 		/*
1194 		 * Do not read the rest of the completion entry before the
1195 		 * phase bit was validated
1196 		 */
1197 		dma_rmb();
1198 
1199 		eeq->cb(eeq, eqe);
1200 
1201 		/* Get next event entry */
1202 		ci++;
1203 		processed++;
1204 
1205 		if (ci == eeq->depth) {
1206 			ci = 0;
1207 			phase = !phase;
1208 		}
1209 
1210 		eqe = &eeq->eqes[ci];
1211 	}
1212 
1213 	eeq->cc += processed;
1214 	eeq->phase = phase;
1215 	efa_com_arm_eq(eeq->edev, eeq);
1216 }
1217 
1218 void efa_com_eq_destroy(struct efa_com_dev *edev, struct efa_com_eq *eeq)
1219 {
1220 	struct efa_com_destroy_eq_params params = {
1221 		.eqn = eeq->eqn,
1222 	};
1223 
1224 	efa_com_destroy_eq(edev, &params);
1225 	dma_free_coherent(edev->dmadev, eeq->depth * sizeof(*eeq->eqes),
1226 			  eeq->eqes, eeq->dma_addr);
1227 }
1228 
1229 int efa_com_eq_init(struct efa_com_dev *edev, struct efa_com_eq *eeq,
1230 		    efa_eqe_handler cb, u16 depth, u8 msix_vec)
1231 {
1232 	struct efa_com_create_eq_params params = {};
1233 	struct efa_com_create_eq_result result = {};
1234 	int err;
1235 
1236 	params.depth = depth;
1237 	params.entry_size_in_bytes = sizeof(*eeq->eqes);
1238 	EFA_SET(&params.event_bitmask,
1239 		EFA_ADMIN_CREATE_EQ_CMD_COMPLETION_EVENTS, 1);
1240 	params.msix_vec = msix_vec;
1241 
1242 	eeq->eqes = dma_alloc_coherent(edev->dmadev,
1243 				       params.depth * sizeof(*eeq->eqes),
1244 				       &params.dma_addr, GFP_KERNEL);
1245 	if (!eeq->eqes)
1246 		return -ENOMEM;
1247 
1248 	err = efa_com_create_eq(edev, &params, &result);
1249 	if (err)
1250 		goto err_free_coherent;
1251 
1252 	eeq->eqn = result.eqn;
1253 	eeq->edev = edev;
1254 	eeq->dma_addr = params.dma_addr;
1255 	eeq->phase = 1;
1256 	eeq->depth = params.depth;
1257 	eeq->cb = cb;
1258 	efa_com_arm_eq(edev, eeq);
1259 
1260 	return 0;
1261 
1262 err_free_coherent:
1263 	dma_free_coherent(edev->dmadev, params.depth * sizeof(*eeq->eqes),
1264 			  eeq->eqes, params.dma_addr);
1265 	return err;
1266 }
1267