Merge tag 'trace-printf-v6.13' of git://git.kernel.org/pub/scm/linux/kernel/git/trace...
[drm/drm-misc.git] / drivers / infiniband / hw / efa / efa_com.c
blobbafd210dd43e86087160e1715c50709b6006e0a9
1 // SPDX-License-Identifier: GPL-2.0 OR BSD-2-Clause
2 /*
3 * Copyright 2018-2024 Amazon.com, Inc. or its affiliates. All rights reserved.
4 */
6 #include "efa_com.h"
7 #include "efa_regs_defs.h"
9 #define ADMIN_CMD_TIMEOUT_US 30000000 /* usecs */
11 #define EFA_REG_READ_TIMEOUT_US 50000 /* usecs */
12 #define EFA_MMIO_READ_INVALID 0xffffffff
14 #define EFA_POLL_INTERVAL_MS 100 /* msecs */
16 #define EFA_ASYNC_QUEUE_DEPTH 16
17 #define EFA_ADMIN_QUEUE_DEPTH 32
19 #define EFA_CTRL_MAJOR 0
20 #define EFA_CTRL_MINOR 0
21 #define EFA_CTRL_SUB_MINOR 1
23 enum efa_cmd_status {
24 EFA_CMD_SUBMITTED,
25 EFA_CMD_COMPLETED,
28 struct efa_comp_ctx {
29 struct completion wait_event;
30 struct efa_admin_acq_entry *user_cqe;
31 u32 comp_size;
32 enum efa_cmd_status status;
33 u8 cmd_opcode;
34 u8 occupied;
37 static const char *efa_com_cmd_str(u8 cmd)
39 #define EFA_CMD_STR_CASE(_cmd) case EFA_ADMIN_##_cmd: return #_cmd
41 switch (cmd) {
42 EFA_CMD_STR_CASE(CREATE_QP);
43 EFA_CMD_STR_CASE(MODIFY_QP);
44 EFA_CMD_STR_CASE(QUERY_QP);
45 EFA_CMD_STR_CASE(DESTROY_QP);
46 EFA_CMD_STR_CASE(CREATE_AH);
47 EFA_CMD_STR_CASE(DESTROY_AH);
48 EFA_CMD_STR_CASE(REG_MR);
49 EFA_CMD_STR_CASE(DEREG_MR);
50 EFA_CMD_STR_CASE(CREATE_CQ);
51 EFA_CMD_STR_CASE(DESTROY_CQ);
52 EFA_CMD_STR_CASE(GET_FEATURE);
53 EFA_CMD_STR_CASE(SET_FEATURE);
54 EFA_CMD_STR_CASE(GET_STATS);
55 EFA_CMD_STR_CASE(ALLOC_PD);
56 EFA_CMD_STR_CASE(DEALLOC_PD);
57 EFA_CMD_STR_CASE(ALLOC_UAR);
58 EFA_CMD_STR_CASE(DEALLOC_UAR);
59 EFA_CMD_STR_CASE(CREATE_EQ);
60 EFA_CMD_STR_CASE(DESTROY_EQ);
61 default: return "unknown command opcode";
63 #undef EFA_CMD_STR_CASE
66 void efa_com_set_dma_addr(dma_addr_t addr, u32 *addr_high, u32 *addr_low)
68 *addr_low = lower_32_bits(addr);
69 *addr_high = upper_32_bits(addr);
72 static u32 efa_com_reg_read32(struct efa_com_dev *edev, u16 offset)
74 struct efa_com_mmio_read *mmio_read = &edev->mmio_read;
75 struct efa_admin_mmio_req_read_less_resp *read_resp;
76 unsigned long exp_time;
77 u32 mmio_read_reg = 0;
78 u32 err;
80 read_resp = mmio_read->read_resp;
82 spin_lock(&mmio_read->lock);
83 mmio_read->seq_num++;
85 /* trash DMA req_id to identify when hardware is done */
86 read_resp->req_id = mmio_read->seq_num + 0x9aL;
87 EFA_SET(&mmio_read_reg, EFA_REGS_MMIO_REG_READ_REG_OFF, offset);
88 EFA_SET(&mmio_read_reg, EFA_REGS_MMIO_REG_READ_REQ_ID,
89 mmio_read->seq_num);
91 writel(mmio_read_reg, edev->reg_bar + EFA_REGS_MMIO_REG_READ_OFF);
93 exp_time = jiffies + usecs_to_jiffies(mmio_read->mmio_read_timeout);
94 do {
95 if (READ_ONCE(read_resp->req_id) == mmio_read->seq_num)
96 break;
97 udelay(1);
98 } while (time_is_after_jiffies(exp_time));
100 if (read_resp->req_id != mmio_read->seq_num) {
101 ibdev_err_ratelimited(
102 edev->efa_dev,
103 "Reading register timed out. expected: req id[%u] offset[%#x] actual: req id[%u] offset[%#x]\n",
104 mmio_read->seq_num, offset, read_resp->req_id,
105 read_resp->reg_off);
106 err = EFA_MMIO_READ_INVALID;
107 goto out;
110 if (read_resp->reg_off != offset) {
111 ibdev_err_ratelimited(
112 edev->efa_dev,
113 "Reading register failed: wrong offset provided\n");
114 err = EFA_MMIO_READ_INVALID;
115 goto out;
118 err = read_resp->reg_val;
119 out:
120 spin_unlock(&mmio_read->lock);
121 return err;
124 static int efa_com_admin_init_sq(struct efa_com_dev *edev)
126 struct efa_com_admin_queue *aq = &edev->aq;
127 struct efa_com_admin_sq *sq = &aq->sq;
128 u16 size = aq->depth * sizeof(*sq->entries);
129 u32 aq_caps = 0;
130 u32 addr_high;
131 u32 addr_low;
133 sq->entries =
134 dma_alloc_coherent(aq->dmadev, size, &sq->dma_addr, GFP_KERNEL);
135 if (!sq->entries)
136 return -ENOMEM;
138 spin_lock_init(&sq->lock);
140 sq->cc = 0;
141 sq->pc = 0;
142 sq->phase = 1;
144 sq->db_addr = (u32 __iomem *)(edev->reg_bar + EFA_REGS_AQ_PROD_DB_OFF);
146 addr_high = upper_32_bits(sq->dma_addr);
147 addr_low = lower_32_bits(sq->dma_addr);
149 writel(addr_low, edev->reg_bar + EFA_REGS_AQ_BASE_LO_OFF);
150 writel(addr_high, edev->reg_bar + EFA_REGS_AQ_BASE_HI_OFF);
152 EFA_SET(&aq_caps, EFA_REGS_AQ_CAPS_AQ_DEPTH, aq->depth);
153 EFA_SET(&aq_caps, EFA_REGS_AQ_CAPS_AQ_ENTRY_SIZE,
154 sizeof(struct efa_admin_aq_entry));
156 writel(aq_caps, edev->reg_bar + EFA_REGS_AQ_CAPS_OFF);
158 return 0;
161 static int efa_com_admin_init_cq(struct efa_com_dev *edev)
163 struct efa_com_admin_queue *aq = &edev->aq;
164 struct efa_com_admin_cq *cq = &aq->cq;
165 u16 size = aq->depth * sizeof(*cq->entries);
166 u32 acq_caps = 0;
167 u32 addr_high;
168 u32 addr_low;
170 cq->entries =
171 dma_alloc_coherent(aq->dmadev, size, &cq->dma_addr, GFP_KERNEL);
172 if (!cq->entries)
173 return -ENOMEM;
175 spin_lock_init(&cq->lock);
177 cq->cc = 0;
178 cq->phase = 1;
180 addr_high = upper_32_bits(cq->dma_addr);
181 addr_low = lower_32_bits(cq->dma_addr);
183 writel(addr_low, edev->reg_bar + EFA_REGS_ACQ_BASE_LO_OFF);
184 writel(addr_high, edev->reg_bar + EFA_REGS_ACQ_BASE_HI_OFF);
186 EFA_SET(&acq_caps, EFA_REGS_ACQ_CAPS_ACQ_DEPTH, aq->depth);
187 EFA_SET(&acq_caps, EFA_REGS_ACQ_CAPS_ACQ_ENTRY_SIZE,
188 sizeof(struct efa_admin_acq_entry));
189 EFA_SET(&acq_caps, EFA_REGS_ACQ_CAPS_ACQ_MSIX_VECTOR,
190 aq->msix_vector_idx);
192 writel(acq_caps, edev->reg_bar + EFA_REGS_ACQ_CAPS_OFF);
194 return 0;
197 static int efa_com_admin_init_aenq(struct efa_com_dev *edev,
198 struct efa_aenq_handlers *aenq_handlers)
200 struct efa_com_aenq *aenq = &edev->aenq;
201 u32 addr_low, addr_high;
202 u32 aenq_caps = 0;
203 u16 size;
205 if (!aenq_handlers) {
206 ibdev_err(edev->efa_dev, "aenq handlers pointer is NULL\n");
207 return -EINVAL;
210 size = EFA_ASYNC_QUEUE_DEPTH * sizeof(*aenq->entries);
211 aenq->entries = dma_alloc_coherent(edev->dmadev, size, &aenq->dma_addr,
212 GFP_KERNEL);
213 if (!aenq->entries)
214 return -ENOMEM;
216 aenq->aenq_handlers = aenq_handlers;
217 aenq->depth = EFA_ASYNC_QUEUE_DEPTH;
218 aenq->cc = 0;
219 aenq->phase = 1;
221 addr_low = lower_32_bits(aenq->dma_addr);
222 addr_high = upper_32_bits(aenq->dma_addr);
224 writel(addr_low, edev->reg_bar + EFA_REGS_AENQ_BASE_LO_OFF);
225 writel(addr_high, edev->reg_bar + EFA_REGS_AENQ_BASE_HI_OFF);
227 EFA_SET(&aenq_caps, EFA_REGS_AENQ_CAPS_AENQ_DEPTH, aenq->depth);
228 EFA_SET(&aenq_caps, EFA_REGS_AENQ_CAPS_AENQ_ENTRY_SIZE,
229 sizeof(struct efa_admin_aenq_entry));
230 EFA_SET(&aenq_caps, EFA_REGS_AENQ_CAPS_AENQ_MSIX_VECTOR,
231 aenq->msix_vector_idx);
232 writel(aenq_caps, edev->reg_bar + EFA_REGS_AENQ_CAPS_OFF);
235 * Init cons_db to mark that all entries in the queue
236 * are initially available
238 writel(edev->aenq.cc, edev->reg_bar + EFA_REGS_AENQ_CONS_DB_OFF);
240 return 0;
243 /* ID to be used with efa_com_get_comp_ctx */
244 static u16 efa_com_alloc_ctx_id(struct efa_com_admin_queue *aq)
246 u16 ctx_id;
248 spin_lock(&aq->comp_ctx_lock);
249 ctx_id = aq->comp_ctx_pool[aq->comp_ctx_pool_next];
250 aq->comp_ctx_pool_next++;
251 spin_unlock(&aq->comp_ctx_lock);
253 return ctx_id;
256 static void efa_com_dealloc_ctx_id(struct efa_com_admin_queue *aq,
257 u16 ctx_id)
259 spin_lock(&aq->comp_ctx_lock);
260 aq->comp_ctx_pool_next--;
261 aq->comp_ctx_pool[aq->comp_ctx_pool_next] = ctx_id;
262 spin_unlock(&aq->comp_ctx_lock);
265 static inline void efa_com_put_comp_ctx(struct efa_com_admin_queue *aq,
266 struct efa_comp_ctx *comp_ctx)
268 u16 cmd_id = EFA_GET(&comp_ctx->user_cqe->acq_common_descriptor.command,
269 EFA_ADMIN_ACQ_COMMON_DESC_COMMAND_ID);
270 u16 ctx_id = cmd_id & (aq->depth - 1);
272 ibdev_dbg(aq->efa_dev, "Put completion command_id %#x\n", cmd_id);
273 comp_ctx->occupied = 0;
274 efa_com_dealloc_ctx_id(aq, ctx_id);
277 static struct efa_comp_ctx *efa_com_get_comp_ctx(struct efa_com_admin_queue *aq,
278 u16 cmd_id, bool capture)
280 u16 ctx_id = cmd_id & (aq->depth - 1);
282 if (aq->comp_ctx[ctx_id].occupied && capture) {
283 ibdev_err_ratelimited(
284 aq->efa_dev,
285 "Completion context for command_id %#x is occupied\n",
286 cmd_id);
287 return NULL;
290 if (capture) {
291 aq->comp_ctx[ctx_id].occupied = 1;
292 ibdev_dbg(aq->efa_dev,
293 "Take completion ctxt for command_id %#x\n", cmd_id);
296 return &aq->comp_ctx[ctx_id];
299 static struct efa_comp_ctx *__efa_com_submit_admin_cmd(struct efa_com_admin_queue *aq,
300 struct efa_admin_aq_entry *cmd,
301 size_t cmd_size_in_bytes,
302 struct efa_admin_acq_entry *comp,
303 size_t comp_size_in_bytes)
305 struct efa_admin_aq_entry *aqe;
306 struct efa_comp_ctx *comp_ctx;
307 u16 queue_size_mask;
308 u16 cmd_id;
309 u16 ctx_id;
310 u16 pi;
312 queue_size_mask = aq->depth - 1;
313 pi = aq->sq.pc & queue_size_mask;
315 ctx_id = efa_com_alloc_ctx_id(aq);
317 /* cmd_id LSBs are the ctx_id and MSBs are entropy bits from pc */
318 cmd_id = ctx_id & queue_size_mask;
319 cmd_id |= aq->sq.pc & ~queue_size_mask;
320 cmd_id &= EFA_ADMIN_AQ_COMMON_DESC_COMMAND_ID_MASK;
322 cmd->aq_common_descriptor.command_id = cmd_id;
323 EFA_SET(&cmd->aq_common_descriptor.flags,
324 EFA_ADMIN_AQ_COMMON_DESC_PHASE, aq->sq.phase);
326 comp_ctx = efa_com_get_comp_ctx(aq, cmd_id, true);
327 if (!comp_ctx) {
328 efa_com_dealloc_ctx_id(aq, ctx_id);
329 return ERR_PTR(-EINVAL);
332 comp_ctx->status = EFA_CMD_SUBMITTED;
333 comp_ctx->comp_size = comp_size_in_bytes;
334 comp_ctx->user_cqe = comp;
335 comp_ctx->cmd_opcode = cmd->aq_common_descriptor.opcode;
337 reinit_completion(&comp_ctx->wait_event);
339 aqe = &aq->sq.entries[pi];
340 memset(aqe, 0, sizeof(*aqe));
341 memcpy(aqe, cmd, cmd_size_in_bytes);
343 aq->sq.pc++;
344 atomic64_inc(&aq->stats.submitted_cmd);
346 if ((aq->sq.pc & queue_size_mask) == 0)
347 aq->sq.phase = !aq->sq.phase;
349 /* barrier not needed in case of writel */
350 writel(aq->sq.pc, aq->sq.db_addr);
352 return comp_ctx;
355 static inline int efa_com_init_comp_ctxt(struct efa_com_admin_queue *aq)
357 size_t pool_size = aq->depth * sizeof(*aq->comp_ctx_pool);
358 size_t size = aq->depth * sizeof(struct efa_comp_ctx);
359 struct efa_comp_ctx *comp_ctx;
360 u16 i;
362 aq->comp_ctx = devm_kzalloc(aq->dmadev, size, GFP_KERNEL);
363 aq->comp_ctx_pool = devm_kzalloc(aq->dmadev, pool_size, GFP_KERNEL);
364 if (!aq->comp_ctx || !aq->comp_ctx_pool) {
365 devm_kfree(aq->dmadev, aq->comp_ctx_pool);
366 devm_kfree(aq->dmadev, aq->comp_ctx);
367 return -ENOMEM;
370 for (i = 0; i < aq->depth; i++) {
371 comp_ctx = efa_com_get_comp_ctx(aq, i, false);
372 if (comp_ctx)
373 init_completion(&comp_ctx->wait_event);
375 aq->comp_ctx_pool[i] = i;
378 spin_lock_init(&aq->comp_ctx_lock);
380 aq->comp_ctx_pool_next = 0;
382 return 0;
385 static struct efa_comp_ctx *efa_com_submit_admin_cmd(struct efa_com_admin_queue *aq,
386 struct efa_admin_aq_entry *cmd,
387 size_t cmd_size_in_bytes,
388 struct efa_admin_acq_entry *comp,
389 size_t comp_size_in_bytes)
391 struct efa_comp_ctx *comp_ctx;
393 spin_lock(&aq->sq.lock);
394 if (!test_bit(EFA_AQ_STATE_RUNNING_BIT, &aq->state)) {
395 ibdev_err_ratelimited(aq->efa_dev, "Admin queue is closed\n");
396 spin_unlock(&aq->sq.lock);
397 return ERR_PTR(-ENODEV);
400 comp_ctx = __efa_com_submit_admin_cmd(aq, cmd, cmd_size_in_bytes, comp,
401 comp_size_in_bytes);
402 spin_unlock(&aq->sq.lock);
403 if (IS_ERR(comp_ctx))
404 clear_bit(EFA_AQ_STATE_RUNNING_BIT, &aq->state);
406 return comp_ctx;
409 static int efa_com_handle_single_admin_completion(struct efa_com_admin_queue *aq,
410 struct efa_admin_acq_entry *cqe)
412 struct efa_comp_ctx *comp_ctx;
413 u16 cmd_id;
415 cmd_id = EFA_GET(&cqe->acq_common_descriptor.command,
416 EFA_ADMIN_ACQ_COMMON_DESC_COMMAND_ID);
418 comp_ctx = efa_com_get_comp_ctx(aq, cmd_id, false);
419 if (comp_ctx->status != EFA_CMD_SUBMITTED) {
420 ibdev_err(aq->efa_dev,
421 "Received completion with unexpected command id[%d], sq producer: %d, sq consumer: %d, cq consumer: %d\n",
422 cmd_id, aq->sq.pc, aq->sq.cc, aq->cq.cc);
423 return -EINVAL;
426 comp_ctx->status = EFA_CMD_COMPLETED;
427 memcpy(comp_ctx->user_cqe, cqe, comp_ctx->comp_size);
429 if (!test_bit(EFA_AQ_STATE_POLLING_BIT, &aq->state))
430 complete(&comp_ctx->wait_event);
432 return 0;
435 static void efa_com_handle_admin_completion(struct efa_com_admin_queue *aq)
437 struct efa_admin_acq_entry *cqe;
438 u16 queue_size_mask;
439 u16 comp_cmds = 0;
440 u8 phase;
441 int err;
442 u16 ci;
444 queue_size_mask = aq->depth - 1;
446 ci = aq->cq.cc & queue_size_mask;
447 phase = aq->cq.phase;
449 cqe = &aq->cq.entries[ci];
451 /* Go over all the completions */
452 while ((READ_ONCE(cqe->acq_common_descriptor.flags) &
453 EFA_ADMIN_ACQ_COMMON_DESC_PHASE_MASK) == phase) {
455 * Do not read the rest of the completion entry before the
456 * phase bit was validated
458 dma_rmb();
459 err = efa_com_handle_single_admin_completion(aq, cqe);
460 if (!err)
461 comp_cmds++;
463 aq->cq.cc++;
464 ci++;
465 if (ci == aq->depth) {
466 ci = 0;
467 phase = !phase;
470 cqe = &aq->cq.entries[ci];
473 aq->cq.phase = phase;
474 aq->sq.cc += comp_cmds;
475 atomic64_add(comp_cmds, &aq->stats.completed_cmd);
478 static int efa_com_comp_status_to_errno(u8 comp_status)
480 switch (comp_status) {
481 case EFA_ADMIN_SUCCESS:
482 return 0;
483 case EFA_ADMIN_RESOURCE_ALLOCATION_FAILURE:
484 return -ENOMEM;
485 case EFA_ADMIN_UNSUPPORTED_OPCODE:
486 return -EOPNOTSUPP;
487 case EFA_ADMIN_BAD_OPCODE:
488 case EFA_ADMIN_MALFORMED_REQUEST:
489 case EFA_ADMIN_ILLEGAL_PARAMETER:
490 case EFA_ADMIN_UNKNOWN_ERROR:
491 return -EINVAL;
492 default:
493 return -EINVAL;
497 static int efa_com_wait_and_process_admin_cq_polling(struct efa_comp_ctx *comp_ctx,
498 struct efa_com_admin_queue *aq)
500 unsigned long timeout;
501 unsigned long flags;
502 int err;
504 timeout = jiffies + usecs_to_jiffies(aq->completion_timeout);
506 while (1) {
507 spin_lock_irqsave(&aq->cq.lock, flags);
508 efa_com_handle_admin_completion(aq);
509 spin_unlock_irqrestore(&aq->cq.lock, flags);
511 if (comp_ctx->status != EFA_CMD_SUBMITTED)
512 break;
514 if (time_is_before_jiffies(timeout)) {
515 ibdev_err_ratelimited(
516 aq->efa_dev,
517 "Wait for completion (polling) timeout\n");
518 /* EFA didn't have any completion */
519 atomic64_inc(&aq->stats.no_completion);
521 clear_bit(EFA_AQ_STATE_RUNNING_BIT, &aq->state);
522 err = -ETIME;
523 goto out;
526 msleep(aq->poll_interval);
529 err = efa_com_comp_status_to_errno(comp_ctx->user_cqe->acq_common_descriptor.status);
530 out:
531 efa_com_put_comp_ctx(aq, comp_ctx);
532 return err;
535 static int efa_com_wait_and_process_admin_cq_interrupts(struct efa_comp_ctx *comp_ctx,
536 struct efa_com_admin_queue *aq)
538 unsigned long flags;
539 int err;
541 wait_for_completion_timeout(&comp_ctx->wait_event,
542 usecs_to_jiffies(aq->completion_timeout));
545 * In case the command wasn't completed find out the root cause.
546 * There might be 2 kinds of errors
547 * 1) No completion (timeout reached)
548 * 2) There is completion but the device didn't get any msi-x interrupt.
550 if (comp_ctx->status == EFA_CMD_SUBMITTED) {
551 spin_lock_irqsave(&aq->cq.lock, flags);
552 efa_com_handle_admin_completion(aq);
553 spin_unlock_irqrestore(&aq->cq.lock, flags);
555 atomic64_inc(&aq->stats.no_completion);
557 if (comp_ctx->status == EFA_CMD_COMPLETED)
558 ibdev_err_ratelimited(
559 aq->efa_dev,
560 "The device sent a completion but the driver didn't receive any MSI-X interrupt for admin cmd %s(%d) status %d (ctx: 0x%p, sq producer: %d, sq consumer: %d, cq consumer: %d)\n",
561 efa_com_cmd_str(comp_ctx->cmd_opcode),
562 comp_ctx->cmd_opcode, comp_ctx->status,
563 comp_ctx, aq->sq.pc, aq->sq.cc, aq->cq.cc);
564 else
565 ibdev_err_ratelimited(
566 aq->efa_dev,
567 "The device didn't send any completion for admin cmd %s(%d) status %d (ctx 0x%p, sq producer: %d, sq consumer: %d, cq consumer: %d)\n",
568 efa_com_cmd_str(comp_ctx->cmd_opcode),
569 comp_ctx->cmd_opcode, comp_ctx->status,
570 comp_ctx, aq->sq.pc, aq->sq.cc, aq->cq.cc);
572 clear_bit(EFA_AQ_STATE_RUNNING_BIT, &aq->state);
573 err = -ETIME;
574 goto out;
577 err = efa_com_comp_status_to_errno(comp_ctx->user_cqe->acq_common_descriptor.status);
578 out:
579 efa_com_put_comp_ctx(aq, comp_ctx);
580 return err;
584 * There are two types to wait for completion.
585 * Polling mode - wait until the completion is available.
586 * Async mode - wait on wait queue until the completion is ready
587 * (or the timeout expired).
588 * It is expected that the IRQ called efa_com_handle_admin_completion
589 * to mark the completions.
591 static int efa_com_wait_and_process_admin_cq(struct efa_comp_ctx *comp_ctx,
592 struct efa_com_admin_queue *aq)
594 if (test_bit(EFA_AQ_STATE_POLLING_BIT, &aq->state))
595 return efa_com_wait_and_process_admin_cq_polling(comp_ctx, aq);
597 return efa_com_wait_and_process_admin_cq_interrupts(comp_ctx, aq);
601 * efa_com_cmd_exec - Execute admin command
602 * @aq: admin queue.
603 * @cmd: the admin command to execute.
604 * @cmd_size: the command size.
605 * @comp: command completion return entry.
606 * @comp_size: command completion size.
607 * Submit an admin command and then wait until the device will return a
608 * completion.
609 * The completion will be copied into comp.
611 * @return - 0 on success, negative value on failure.
613 int efa_com_cmd_exec(struct efa_com_admin_queue *aq,
614 struct efa_admin_aq_entry *cmd,
615 size_t cmd_size,
616 struct efa_admin_acq_entry *comp,
617 size_t comp_size)
619 struct efa_comp_ctx *comp_ctx;
620 int err;
622 might_sleep();
624 /* In case of queue FULL */
625 down(&aq->avail_cmds);
627 ibdev_dbg(aq->efa_dev, "%s (opcode %d)\n",
628 efa_com_cmd_str(cmd->aq_common_descriptor.opcode),
629 cmd->aq_common_descriptor.opcode);
630 comp_ctx = efa_com_submit_admin_cmd(aq, cmd, cmd_size, comp, comp_size);
631 if (IS_ERR(comp_ctx)) {
632 ibdev_err_ratelimited(
633 aq->efa_dev,
634 "Failed to submit command %s (opcode %u) err %ld\n",
635 efa_com_cmd_str(cmd->aq_common_descriptor.opcode),
636 cmd->aq_common_descriptor.opcode, PTR_ERR(comp_ctx));
638 up(&aq->avail_cmds);
639 atomic64_inc(&aq->stats.cmd_err);
640 return PTR_ERR(comp_ctx);
643 err = efa_com_wait_and_process_admin_cq(comp_ctx, aq);
644 if (err) {
645 ibdev_err_ratelimited(
646 aq->efa_dev,
647 "Failed to process command %s (opcode %u) comp_status %d err %d\n",
648 efa_com_cmd_str(cmd->aq_common_descriptor.opcode),
649 cmd->aq_common_descriptor.opcode,
650 comp_ctx->user_cqe->acq_common_descriptor.status, err);
651 atomic64_inc(&aq->stats.cmd_err);
654 up(&aq->avail_cmds);
656 return err;
660 * efa_com_admin_destroy - Destroy the admin and the async events queues.
661 * @edev: EFA communication layer struct
663 void efa_com_admin_destroy(struct efa_com_dev *edev)
665 struct efa_com_admin_queue *aq = &edev->aq;
666 struct efa_com_aenq *aenq = &edev->aenq;
667 struct efa_com_admin_cq *cq = &aq->cq;
668 struct efa_com_admin_sq *sq = &aq->sq;
669 u16 size;
671 clear_bit(EFA_AQ_STATE_RUNNING_BIT, &aq->state);
673 devm_kfree(edev->dmadev, aq->comp_ctx_pool);
674 devm_kfree(edev->dmadev, aq->comp_ctx);
676 size = aq->depth * sizeof(*sq->entries);
677 dma_free_coherent(edev->dmadev, size, sq->entries, sq->dma_addr);
679 size = aq->depth * sizeof(*cq->entries);
680 dma_free_coherent(edev->dmadev, size, cq->entries, cq->dma_addr);
682 size = aenq->depth * sizeof(*aenq->entries);
683 dma_free_coherent(edev->dmadev, size, aenq->entries, aenq->dma_addr);
687 * efa_com_set_admin_polling_mode - Set the admin completion queue polling mode
688 * @edev: EFA communication layer struct
689 * @polling: Enable/Disable polling mode
691 * Set the admin completion mode.
693 void efa_com_set_admin_polling_mode(struct efa_com_dev *edev, bool polling)
695 u32 mask_value = 0;
697 if (polling)
698 EFA_SET(&mask_value, EFA_REGS_INTR_MASK_EN, 1);
700 writel(mask_value, edev->reg_bar + EFA_REGS_INTR_MASK_OFF);
701 if (polling)
702 set_bit(EFA_AQ_STATE_POLLING_BIT, &edev->aq.state);
703 else
704 clear_bit(EFA_AQ_STATE_POLLING_BIT, &edev->aq.state);
707 static void efa_com_stats_init(struct efa_com_dev *edev)
709 atomic64_t *s = (atomic64_t *)&edev->aq.stats;
710 int i;
712 for (i = 0; i < sizeof(edev->aq.stats) / sizeof(*s); i++, s++)
713 atomic64_set(s, 0);
717 * efa_com_admin_init - Init the admin and the async queues
718 * @edev: EFA communication layer struct
719 * @aenq_handlers: Those handlers to be called upon event.
721 * Initialize the admin submission and completion queues.
722 * Initialize the asynchronous events notification queues.
724 * @return - 0 on success, negative value on failure.
726 int efa_com_admin_init(struct efa_com_dev *edev,
727 struct efa_aenq_handlers *aenq_handlers)
729 struct efa_com_admin_queue *aq = &edev->aq;
730 u32 timeout;
731 u32 dev_sts;
732 u32 cap;
733 int err;
735 dev_sts = efa_com_reg_read32(edev, EFA_REGS_DEV_STS_OFF);
736 if (!EFA_GET(&dev_sts, EFA_REGS_DEV_STS_READY)) {
737 ibdev_err(edev->efa_dev,
738 "Device isn't ready, abort com init %#x\n", dev_sts);
739 return -ENODEV;
742 aq->depth = EFA_ADMIN_QUEUE_DEPTH;
744 aq->dmadev = edev->dmadev;
745 aq->efa_dev = edev->efa_dev;
746 set_bit(EFA_AQ_STATE_POLLING_BIT, &aq->state);
748 sema_init(&aq->avail_cmds, aq->depth);
750 efa_com_stats_init(edev);
752 err = efa_com_init_comp_ctxt(aq);
753 if (err)
754 return err;
756 err = efa_com_admin_init_sq(edev);
757 if (err)
758 goto err_destroy_comp_ctxt;
760 err = efa_com_admin_init_cq(edev);
761 if (err)
762 goto err_destroy_sq;
764 efa_com_set_admin_polling_mode(edev, false);
766 err = efa_com_admin_init_aenq(edev, aenq_handlers);
767 if (err)
768 goto err_destroy_cq;
770 cap = efa_com_reg_read32(edev, EFA_REGS_CAPS_OFF);
771 timeout = EFA_GET(&cap, EFA_REGS_CAPS_ADMIN_CMD_TO);
772 if (timeout)
773 /* the resolution of timeout reg is 100ms */
774 aq->completion_timeout = timeout * 100000;
775 else
776 aq->completion_timeout = ADMIN_CMD_TIMEOUT_US;
778 aq->poll_interval = EFA_POLL_INTERVAL_MS;
780 set_bit(EFA_AQ_STATE_RUNNING_BIT, &aq->state);
782 return 0;
784 err_destroy_cq:
785 dma_free_coherent(edev->dmadev, aq->depth * sizeof(*aq->cq.entries),
786 aq->cq.entries, aq->cq.dma_addr);
787 err_destroy_sq:
788 dma_free_coherent(edev->dmadev, aq->depth * sizeof(*aq->sq.entries),
789 aq->sq.entries, aq->sq.dma_addr);
790 err_destroy_comp_ctxt:
791 devm_kfree(edev->dmadev, aq->comp_ctx);
793 return err;
797 * efa_com_admin_q_comp_intr_handler - admin queue interrupt handler
798 * @edev: EFA communication layer struct
800 * This method goes over the admin completion queue and wakes up
801 * all the pending threads that wait on the commands wait event.
803 * Note: Should be called after MSI-X interrupt.
805 void efa_com_admin_q_comp_intr_handler(struct efa_com_dev *edev)
807 unsigned long flags;
809 spin_lock_irqsave(&edev->aq.cq.lock, flags);
810 efa_com_handle_admin_completion(&edev->aq);
811 spin_unlock_irqrestore(&edev->aq.cq.lock, flags);
815 * efa_handle_specific_aenq_event:
816 * return the handler that is relevant to the specific event group
818 static efa_aenq_handler efa_com_get_specific_aenq_cb(struct efa_com_dev *edev,
819 u16 group)
821 struct efa_aenq_handlers *aenq_handlers = edev->aenq.aenq_handlers;
823 if (group < EFA_MAX_HANDLERS && aenq_handlers->handlers[group])
824 return aenq_handlers->handlers[group];
826 return aenq_handlers->unimplemented_handler;
830 * efa_com_aenq_intr_handler - AENQ interrupt handler
831 * @edev: EFA communication layer struct
832 * @data: Data of interrupt handler.
834 * Go over the async event notification queue and call the proper aenq handler.
836 void efa_com_aenq_intr_handler(struct efa_com_dev *edev, void *data)
838 struct efa_admin_aenq_common_desc *aenq_common;
839 struct efa_com_aenq *aenq = &edev->aenq;
840 struct efa_admin_aenq_entry *aenq_e;
841 efa_aenq_handler handler_cb;
842 u32 processed = 0;
843 u8 phase;
844 u32 ci;
846 ci = aenq->cc & (aenq->depth - 1);
847 phase = aenq->phase;
848 aenq_e = &aenq->entries[ci]; /* Get first entry */
849 aenq_common = &aenq_e->aenq_common_desc;
851 /* Go over all the events */
852 while ((READ_ONCE(aenq_common->flags) &
853 EFA_ADMIN_AENQ_COMMON_DESC_PHASE_MASK) == phase) {
855 * Do not read the rest of the completion entry before the
856 * phase bit was validated
858 dma_rmb();
860 /* Handle specific event*/
861 handler_cb = efa_com_get_specific_aenq_cb(edev,
862 aenq_common->group);
863 handler_cb(data, aenq_e); /* call the actual event handler*/
865 /* Get next event entry */
866 ci++;
867 processed++;
869 if (ci == aenq->depth) {
870 ci = 0;
871 phase = !phase;
873 aenq_e = &aenq->entries[ci];
874 aenq_common = &aenq_e->aenq_common_desc;
877 aenq->cc += processed;
878 aenq->phase = phase;
880 /* Don't update aenq doorbell if there weren't any processed events */
881 if (!processed)
882 return;
884 /* barrier not needed in case of writel */
885 writel(aenq->cc, edev->reg_bar + EFA_REGS_AENQ_CONS_DB_OFF);
888 static void efa_com_mmio_reg_read_resp_addr_init(struct efa_com_dev *edev)
890 struct efa_com_mmio_read *mmio_read = &edev->mmio_read;
891 u32 addr_high;
892 u32 addr_low;
894 /* dma_addr_bits is unknown at this point */
895 addr_high = (mmio_read->read_resp_dma_addr >> 32) & GENMASK(31, 0);
896 addr_low = mmio_read->read_resp_dma_addr & GENMASK(31, 0);
898 writel(addr_high, edev->reg_bar + EFA_REGS_MMIO_RESP_HI_OFF);
899 writel(addr_low, edev->reg_bar + EFA_REGS_MMIO_RESP_LO_OFF);
902 int efa_com_mmio_reg_read_init(struct efa_com_dev *edev)
904 struct efa_com_mmio_read *mmio_read = &edev->mmio_read;
906 spin_lock_init(&mmio_read->lock);
907 mmio_read->read_resp =
908 dma_alloc_coherent(edev->dmadev, sizeof(*mmio_read->read_resp),
909 &mmio_read->read_resp_dma_addr, GFP_KERNEL);
910 if (!mmio_read->read_resp)
911 return -ENOMEM;
913 efa_com_mmio_reg_read_resp_addr_init(edev);
915 mmio_read->read_resp->req_id = 0;
916 mmio_read->seq_num = 0;
917 mmio_read->mmio_read_timeout = EFA_REG_READ_TIMEOUT_US;
919 return 0;
922 void efa_com_mmio_reg_read_destroy(struct efa_com_dev *edev)
924 struct efa_com_mmio_read *mmio_read = &edev->mmio_read;
926 dma_free_coherent(edev->dmadev, sizeof(*mmio_read->read_resp),
927 mmio_read->read_resp, mmio_read->read_resp_dma_addr);
930 int efa_com_validate_version(struct efa_com_dev *edev)
932 u32 min_ctrl_ver = 0;
933 u32 ctrl_ver_masked;
934 u32 min_ver = 0;
935 u32 ctrl_ver;
936 u32 ver;
939 * Make sure the EFA version and the controller version are at least
940 * as the driver expects
942 ver = efa_com_reg_read32(edev, EFA_REGS_VERSION_OFF);
943 ctrl_ver = efa_com_reg_read32(edev,
944 EFA_REGS_CONTROLLER_VERSION_OFF);
946 ibdev_dbg(edev->efa_dev, "efa device version: %d.%d\n",
947 EFA_GET(&ver, EFA_REGS_VERSION_MAJOR_VERSION),
948 EFA_GET(&ver, EFA_REGS_VERSION_MINOR_VERSION));
950 EFA_SET(&min_ver, EFA_REGS_VERSION_MAJOR_VERSION,
951 EFA_ADMIN_API_VERSION_MAJOR);
952 EFA_SET(&min_ver, EFA_REGS_VERSION_MINOR_VERSION,
953 EFA_ADMIN_API_VERSION_MINOR);
954 if (ver < min_ver) {
955 ibdev_err(edev->efa_dev,
956 "EFA version is lower than the minimal version the driver supports\n");
957 return -EOPNOTSUPP;
960 ibdev_dbg(
961 edev->efa_dev,
962 "efa controller version: %d.%d.%d implementation version %d\n",
963 EFA_GET(&ctrl_ver, EFA_REGS_CONTROLLER_VERSION_MAJOR_VERSION),
964 EFA_GET(&ctrl_ver, EFA_REGS_CONTROLLER_VERSION_MINOR_VERSION),
965 EFA_GET(&ctrl_ver,
966 EFA_REGS_CONTROLLER_VERSION_SUBMINOR_VERSION),
967 EFA_GET(&ctrl_ver, EFA_REGS_CONTROLLER_VERSION_IMPL_ID));
969 ctrl_ver_masked =
970 EFA_GET(&ctrl_ver, EFA_REGS_CONTROLLER_VERSION_MAJOR_VERSION) |
971 EFA_GET(&ctrl_ver, EFA_REGS_CONTROLLER_VERSION_MINOR_VERSION) |
972 EFA_GET(&ctrl_ver,
973 EFA_REGS_CONTROLLER_VERSION_SUBMINOR_VERSION);
975 EFA_SET(&min_ctrl_ver, EFA_REGS_CONTROLLER_VERSION_MAJOR_VERSION,
976 EFA_CTRL_MAJOR);
977 EFA_SET(&min_ctrl_ver, EFA_REGS_CONTROLLER_VERSION_MINOR_VERSION,
978 EFA_CTRL_MINOR);
979 EFA_SET(&min_ctrl_ver, EFA_REGS_CONTROLLER_VERSION_SUBMINOR_VERSION,
980 EFA_CTRL_SUB_MINOR);
981 /* Validate the ctrl version without the implementation ID */
982 if (ctrl_ver_masked < min_ctrl_ver) {
983 ibdev_err(edev->efa_dev,
984 "EFA ctrl version is lower than the minimal ctrl version the driver supports\n");
985 return -EOPNOTSUPP;
988 return 0;
992 * efa_com_get_dma_width - Retrieve physical dma address width the device
993 * supports.
994 * @edev: EFA communication layer struct
996 * Retrieve the maximum physical address bits the device can handle.
998 * @return: > 0 on Success and negative value otherwise.
1000 int efa_com_get_dma_width(struct efa_com_dev *edev)
1002 u32 caps = efa_com_reg_read32(edev, EFA_REGS_CAPS_OFF);
1003 int width;
1005 width = EFA_GET(&caps, EFA_REGS_CAPS_DMA_ADDR_WIDTH);
1007 ibdev_dbg(edev->efa_dev, "DMA width: %d\n", width);
1009 if (width < 32 || width > 64) {
1010 ibdev_err(edev->efa_dev, "DMA width illegal value: %d\n", width);
1011 return -EINVAL;
1014 edev->dma_addr_bits = width;
1016 return width;
1019 static int wait_for_reset_state(struct efa_com_dev *edev, u32 timeout, int on)
1021 u32 val, i;
1023 for (i = 0; i < timeout; i++) {
1024 val = efa_com_reg_read32(edev, EFA_REGS_DEV_STS_OFF);
1026 if (EFA_GET(&val, EFA_REGS_DEV_STS_RESET_IN_PROGRESS) == on)
1027 return 0;
1029 ibdev_dbg(edev->efa_dev, "Reset indication val %d\n", val);
1030 msleep(EFA_POLL_INTERVAL_MS);
1033 return -ETIME;
1037 * efa_com_dev_reset - Perform device FLR to the device.
1038 * @edev: EFA communication layer struct
1039 * @reset_reason: Specify what is the trigger for the reset in case of an error.
1041 * @return - 0 on success, negative value on failure.
1043 int efa_com_dev_reset(struct efa_com_dev *edev,
1044 enum efa_regs_reset_reason_types reset_reason)
1046 u32 stat, timeout, cap;
1047 u32 reset_val = 0;
1048 int err;
1050 stat = efa_com_reg_read32(edev, EFA_REGS_DEV_STS_OFF);
1051 cap = efa_com_reg_read32(edev, EFA_REGS_CAPS_OFF);
1053 if (!EFA_GET(&stat, EFA_REGS_DEV_STS_READY)) {
1054 ibdev_err(edev->efa_dev,
1055 "Device isn't ready, can't reset device\n");
1056 return -EINVAL;
1059 timeout = EFA_GET(&cap, EFA_REGS_CAPS_RESET_TIMEOUT);
1060 if (!timeout) {
1061 ibdev_err(edev->efa_dev, "Invalid timeout value\n");
1062 return -EINVAL;
1065 /* start reset */
1066 EFA_SET(&reset_val, EFA_REGS_DEV_CTL_DEV_RESET, 1);
1067 EFA_SET(&reset_val, EFA_REGS_DEV_CTL_RESET_REASON, reset_reason);
1068 writel(reset_val, edev->reg_bar + EFA_REGS_DEV_CTL_OFF);
1070 /* reset clears the mmio readless address, restore it */
1071 efa_com_mmio_reg_read_resp_addr_init(edev);
1073 err = wait_for_reset_state(edev, timeout, 1);
1074 if (err) {
1075 ibdev_err(edev->efa_dev, "Reset indication didn't turn on\n");
1076 return err;
1079 /* reset done */
1080 writel(0, edev->reg_bar + EFA_REGS_DEV_CTL_OFF);
1081 err = wait_for_reset_state(edev, timeout, 0);
1082 if (err) {
1083 ibdev_err(edev->efa_dev, "Reset indication didn't turn off\n");
1084 return err;
1087 timeout = EFA_GET(&cap, EFA_REGS_CAPS_ADMIN_CMD_TO);
1088 if (timeout)
1089 /* the resolution of timeout reg is 100ms */
1090 edev->aq.completion_timeout = timeout * 100000;
1091 else
1092 edev->aq.completion_timeout = ADMIN_CMD_TIMEOUT_US;
1094 return 0;
1097 static int efa_com_create_eq(struct efa_com_dev *edev,
1098 struct efa_com_create_eq_params *params,
1099 struct efa_com_create_eq_result *result)
1101 struct efa_com_admin_queue *aq = &edev->aq;
1102 struct efa_admin_create_eq_resp resp = {};
1103 struct efa_admin_create_eq_cmd cmd = {};
1104 int err;
1106 cmd.aq_common_descriptor.opcode = EFA_ADMIN_CREATE_EQ;
1107 EFA_SET(&cmd.caps, EFA_ADMIN_CREATE_EQ_CMD_ENTRY_SIZE_WORDS,
1108 params->entry_size_in_bytes / 4);
1109 cmd.depth = params->depth;
1110 cmd.event_bitmask = params->event_bitmask;
1111 cmd.msix_vec = params->msix_vec;
1113 efa_com_set_dma_addr(params->dma_addr, &cmd.ba.mem_addr_high,
1114 &cmd.ba.mem_addr_low);
1116 err = efa_com_cmd_exec(aq,
1117 (struct efa_admin_aq_entry *)&cmd,
1118 sizeof(cmd),
1119 (struct efa_admin_acq_entry *)&resp,
1120 sizeof(resp));
1121 if (err) {
1122 ibdev_err_ratelimited(edev->efa_dev,
1123 "Failed to create eq[%d]\n", err);
1124 return err;
1127 result->eqn = resp.eqn;
1129 return 0;
1132 static void efa_com_destroy_eq(struct efa_com_dev *edev,
1133 struct efa_com_destroy_eq_params *params)
1135 struct efa_com_admin_queue *aq = &edev->aq;
1136 struct efa_admin_destroy_eq_resp resp = {};
1137 struct efa_admin_destroy_eq_cmd cmd = {};
1138 int err;
1140 cmd.aq_common_descriptor.opcode = EFA_ADMIN_DESTROY_EQ;
1141 cmd.eqn = params->eqn;
1143 err = efa_com_cmd_exec(aq,
1144 (struct efa_admin_aq_entry *)&cmd,
1145 sizeof(cmd),
1146 (struct efa_admin_acq_entry *)&resp,
1147 sizeof(resp));
1148 if (err)
1149 ibdev_err_ratelimited(edev->efa_dev,
1150 "Failed to destroy EQ-%u [%d]\n", cmd.eqn,
1151 err);
1154 static void efa_com_arm_eq(struct efa_com_dev *edev, struct efa_com_eq *eeq)
1156 u32 val = 0;
1158 EFA_SET(&val, EFA_REGS_EQ_DB_EQN, eeq->eqn);
1159 EFA_SET(&val, EFA_REGS_EQ_DB_ARM, 1);
1161 writel(val, edev->reg_bar + EFA_REGS_EQ_DB_OFF);
1164 void efa_com_eq_comp_intr_handler(struct efa_com_dev *edev,
1165 struct efa_com_eq *eeq)
1167 struct efa_admin_eqe *eqe;
1168 u32 processed = 0;
1169 u8 phase;
1170 u32 ci;
1172 ci = eeq->cc & (eeq->depth - 1);
1173 phase = eeq->phase;
1174 eqe = &eeq->eqes[ci];
1176 /* Go over all the events */
1177 while ((READ_ONCE(eqe->common) & EFA_ADMIN_EQE_PHASE_MASK) == phase) {
1179 * Do not read the rest of the completion entry before the
1180 * phase bit was validated
1182 dma_rmb();
1184 eeq->cb(eeq, eqe);
1186 /* Get next event entry */
1187 ci++;
1188 processed++;
1190 if (ci == eeq->depth) {
1191 ci = 0;
1192 phase = !phase;
1195 eqe = &eeq->eqes[ci];
1198 eeq->cc += processed;
1199 eeq->phase = phase;
1200 efa_com_arm_eq(eeq->edev, eeq);
1203 void efa_com_eq_destroy(struct efa_com_dev *edev, struct efa_com_eq *eeq)
1205 struct efa_com_destroy_eq_params params = {
1206 .eqn = eeq->eqn,
1209 efa_com_destroy_eq(edev, &params);
1210 dma_free_coherent(edev->dmadev, eeq->depth * sizeof(*eeq->eqes),
1211 eeq->eqes, eeq->dma_addr);
1214 int efa_com_eq_init(struct efa_com_dev *edev, struct efa_com_eq *eeq,
1215 efa_eqe_handler cb, u16 depth, u8 msix_vec)
1217 struct efa_com_create_eq_params params = {};
1218 struct efa_com_create_eq_result result = {};
1219 int err;
1221 params.depth = depth;
1222 params.entry_size_in_bytes = sizeof(*eeq->eqes);
1223 EFA_SET(&params.event_bitmask,
1224 EFA_ADMIN_CREATE_EQ_CMD_COMPLETION_EVENTS, 1);
1225 params.msix_vec = msix_vec;
1227 eeq->eqes = dma_alloc_coherent(edev->dmadev,
1228 params.depth * sizeof(*eeq->eqes),
1229 &params.dma_addr, GFP_KERNEL);
1230 if (!eeq->eqes)
1231 return -ENOMEM;
1233 err = efa_com_create_eq(edev, &params, &result);
1234 if (err)
1235 goto err_free_coherent;
1237 eeq->eqn = result.eqn;
1238 eeq->edev = edev;
1239 eeq->dma_addr = params.dma_addr;
1240 eeq->phase = 1;
1241 eeq->depth = params.depth;
1242 eeq->cb = cb;
1243 efa_com_arm_eq(edev, eeq);
1245 return 0;
1247 err_free_coherent:
1248 dma_free_coherent(edev->dmadev, params.depth * sizeof(*eeq->eqes),
1249 eeq->eqes, params.dma_addr);
1250 return err;