1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /* Copyright (c) 2021, Microsoft Corporation. */
3 
4 #include <linux/module.h>
5 #include <linux/pci.h>
6 #include <linux/utsname.h>
7 #include <linux/version.h>
8 
9 #include "mana.h"
10 
11 static u32 mana_gd_r32(struct gdma_context *g, u64 offset)
12 {
13 	return readl(g->bar0_va + offset);
14 }
15 
16 static u64 mana_gd_r64(struct gdma_context *g, u64 offset)
17 {
18 	return readq(g->bar0_va + offset);
19 }
20 
21 static void mana_gd_init_pf_regs(struct pci_dev *pdev)
22 {
23 	struct gdma_context *gc = pci_get_drvdata(pdev);
24 	void __iomem *sriov_base_va;
25 	u64 sriov_base_off;
26 
27 	gc->db_page_size = mana_gd_r32(gc, GDMA_PF_REG_DB_PAGE_SIZE) & 0xFFFF;
28 	gc->db_page_base = gc->bar0_va +
29 				mana_gd_r64(gc, GDMA_PF_REG_DB_PAGE_OFF);
30 
31 	sriov_base_off = mana_gd_r64(gc, GDMA_SRIOV_REG_CFG_BASE_OFF);
32 
33 	sriov_base_va = gc->bar0_va + sriov_base_off;
34 	gc->shm_base = sriov_base_va +
35 			mana_gd_r64(gc, sriov_base_off + GDMA_PF_REG_SHM_OFF);
36 }
37 
38 static void mana_gd_init_vf_regs(struct pci_dev *pdev)
39 {
40 	struct gdma_context *gc = pci_get_drvdata(pdev);
41 
42 	gc->db_page_size = mana_gd_r32(gc, GDMA_REG_DB_PAGE_SIZE) & 0xFFFF;
43 
44 	gc->db_page_base = gc->bar0_va +
45 				mana_gd_r64(gc, GDMA_REG_DB_PAGE_OFFSET);
46 
47 	gc->shm_base = gc->bar0_va + mana_gd_r64(gc, GDMA_REG_SHM_OFFSET);
48 }
49 
50 static void mana_gd_init_registers(struct pci_dev *pdev)
51 {
52 	struct gdma_context *gc = pci_get_drvdata(pdev);
53 
54 	if (gc->is_pf)
55 		mana_gd_init_pf_regs(pdev);
56 	else
57 		mana_gd_init_vf_regs(pdev);
58 }
59 
60 static int mana_gd_query_max_resources(struct pci_dev *pdev)
61 {
62 	struct gdma_context *gc = pci_get_drvdata(pdev);
63 	struct gdma_query_max_resources_resp resp = {};
64 	struct gdma_general_req req = {};
65 	int err;
66 
67 	mana_gd_init_req_hdr(&req.hdr, GDMA_QUERY_MAX_RESOURCES,
68 			     sizeof(req), sizeof(resp));
69 
70 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
71 	if (err || resp.hdr.status) {
72 		dev_err(gc->dev, "Failed to query resource info: %d, 0x%x\n",
73 			err, resp.hdr.status);
74 		return err ? err : -EPROTO;
75 	}
76 
77 	if (gc->num_msix_usable > resp.max_msix)
78 		gc->num_msix_usable = resp.max_msix;
79 
80 	if (gc->num_msix_usable <= 1)
81 		return -ENOSPC;
82 
83 	gc->max_num_queues = num_online_cpus();
84 	if (gc->max_num_queues > MANA_MAX_NUM_QUEUES)
85 		gc->max_num_queues = MANA_MAX_NUM_QUEUES;
86 
87 	if (gc->max_num_queues > resp.max_eq)
88 		gc->max_num_queues = resp.max_eq;
89 
90 	if (gc->max_num_queues > resp.max_cq)
91 		gc->max_num_queues = resp.max_cq;
92 
93 	if (gc->max_num_queues > resp.max_sq)
94 		gc->max_num_queues = resp.max_sq;
95 
96 	if (gc->max_num_queues > resp.max_rq)
97 		gc->max_num_queues = resp.max_rq;
98 
99 	/* The Hardware Channel (HWC) used 1 MSI-X */
100 	if (gc->max_num_queues > gc->num_msix_usable - 1)
101 		gc->max_num_queues = gc->num_msix_usable - 1;
102 
103 	return 0;
104 }
105 
106 static int mana_gd_detect_devices(struct pci_dev *pdev)
107 {
108 	struct gdma_context *gc = pci_get_drvdata(pdev);
109 	struct gdma_list_devices_resp resp = {};
110 	struct gdma_general_req req = {};
111 	struct gdma_dev_id dev;
112 	u32 i, max_num_devs;
113 	u16 dev_type;
114 	int err;
115 
116 	mana_gd_init_req_hdr(&req.hdr, GDMA_LIST_DEVICES, sizeof(req),
117 			     sizeof(resp));
118 
119 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
120 	if (err || resp.hdr.status) {
121 		dev_err(gc->dev, "Failed to detect devices: %d, 0x%x\n", err,
122 			resp.hdr.status);
123 		return err ? err : -EPROTO;
124 	}
125 
126 	max_num_devs = min_t(u32, MAX_NUM_GDMA_DEVICES, resp.num_of_devs);
127 
128 	for (i = 0; i < max_num_devs; i++) {
129 		dev = resp.devs[i];
130 		dev_type = dev.type;
131 
132 		/* HWC is already detected in mana_hwc_create_channel(). */
133 		if (dev_type == GDMA_DEVICE_HWC)
134 			continue;
135 
136 		if (dev_type == GDMA_DEVICE_MANA) {
137 			gc->mana.gdma_context = gc;
138 			gc->mana.dev_id = dev;
139 		}
140 	}
141 
142 	return gc->mana.dev_id.type == 0 ? -ENODEV : 0;
143 }
144 
145 int mana_gd_send_request(struct gdma_context *gc, u32 req_len, const void *req,
146 			 u32 resp_len, void *resp)
147 {
148 	struct hw_channel_context *hwc = gc->hwc.driver_data;
149 
150 	return mana_hwc_send_request(hwc, req_len, req, resp_len, resp);
151 }
152 
153 int mana_gd_alloc_memory(struct gdma_context *gc, unsigned int length,
154 			 struct gdma_mem_info *gmi)
155 {
156 	dma_addr_t dma_handle;
157 	void *buf;
158 
159 	if (length < PAGE_SIZE || !is_power_of_2(length))
160 		return -EINVAL;
161 
162 	gmi->dev = gc->dev;
163 	buf = dma_alloc_coherent(gmi->dev, length, &dma_handle, GFP_KERNEL);
164 	if (!buf)
165 		return -ENOMEM;
166 
167 	gmi->dma_handle = dma_handle;
168 	gmi->virt_addr = buf;
169 	gmi->length = length;
170 
171 	return 0;
172 }
173 
174 void mana_gd_free_memory(struct gdma_mem_info *gmi)
175 {
176 	dma_free_coherent(gmi->dev, gmi->length, gmi->virt_addr,
177 			  gmi->dma_handle);
178 }
179 
180 static int mana_gd_create_hw_eq(struct gdma_context *gc,
181 				struct gdma_queue *queue)
182 {
183 	struct gdma_create_queue_resp resp = {};
184 	struct gdma_create_queue_req req = {};
185 	int err;
186 
187 	if (queue->type != GDMA_EQ)
188 		return -EINVAL;
189 
190 	mana_gd_init_req_hdr(&req.hdr, GDMA_CREATE_QUEUE,
191 			     sizeof(req), sizeof(resp));
192 
193 	req.hdr.dev_id = queue->gdma_dev->dev_id;
194 	req.type = queue->type;
195 	req.pdid = queue->gdma_dev->pdid;
196 	req.doolbell_id = queue->gdma_dev->doorbell;
197 	req.gdma_region = queue->mem_info.gdma_region;
198 	req.queue_size = queue->queue_size;
199 	req.log2_throttle_limit = queue->eq.log2_throttle_limit;
200 	req.eq_pci_msix_index = queue->eq.msix_index;
201 
202 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
203 	if (err || resp.hdr.status) {
204 		dev_err(gc->dev, "Failed to create queue: %d, 0x%x\n", err,
205 			resp.hdr.status);
206 		return err ? err : -EPROTO;
207 	}
208 
209 	queue->id = resp.queue_index;
210 	queue->eq.disable_needed = true;
211 	queue->mem_info.gdma_region = GDMA_INVALID_DMA_REGION;
212 	return 0;
213 }
214 
215 static int mana_gd_disable_queue(struct gdma_queue *queue)
216 {
217 	struct gdma_context *gc = queue->gdma_dev->gdma_context;
218 	struct gdma_disable_queue_req req = {};
219 	struct gdma_general_resp resp = {};
220 	int err;
221 
222 	WARN_ON(queue->type != GDMA_EQ);
223 
224 	mana_gd_init_req_hdr(&req.hdr, GDMA_DISABLE_QUEUE,
225 			     sizeof(req), sizeof(resp));
226 
227 	req.hdr.dev_id = queue->gdma_dev->dev_id;
228 	req.type = queue->type;
229 	req.queue_index =  queue->id;
230 	req.alloc_res_id_on_creation = 1;
231 
232 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
233 	if (err || resp.hdr.status) {
234 		dev_err(gc->dev, "Failed to disable queue: %d, 0x%x\n", err,
235 			resp.hdr.status);
236 		return err ? err : -EPROTO;
237 	}
238 
239 	return 0;
240 }
241 
242 #define DOORBELL_OFFSET_SQ	0x0
243 #define DOORBELL_OFFSET_RQ	0x400
244 #define DOORBELL_OFFSET_CQ	0x800
245 #define DOORBELL_OFFSET_EQ	0xFF8
246 
247 static void mana_gd_ring_doorbell(struct gdma_context *gc, u32 db_index,
248 				  enum gdma_queue_type q_type, u32 qid,
249 				  u32 tail_ptr, u8 num_req)
250 {
251 	void __iomem *addr = gc->db_page_base + gc->db_page_size * db_index;
252 	union gdma_doorbell_entry e = {};
253 
254 	switch (q_type) {
255 	case GDMA_EQ:
256 		e.eq.id = qid;
257 		e.eq.tail_ptr = tail_ptr;
258 		e.eq.arm = num_req;
259 
260 		addr += DOORBELL_OFFSET_EQ;
261 		break;
262 
263 	case GDMA_CQ:
264 		e.cq.id = qid;
265 		e.cq.tail_ptr = tail_ptr;
266 		e.cq.arm = num_req;
267 
268 		addr += DOORBELL_OFFSET_CQ;
269 		break;
270 
271 	case GDMA_RQ:
272 		e.rq.id = qid;
273 		e.rq.tail_ptr = tail_ptr;
274 		e.rq.wqe_cnt = num_req;
275 
276 		addr += DOORBELL_OFFSET_RQ;
277 		break;
278 
279 	case GDMA_SQ:
280 		e.sq.id = qid;
281 		e.sq.tail_ptr = tail_ptr;
282 
283 		addr += DOORBELL_OFFSET_SQ;
284 		break;
285 
286 	default:
287 		WARN_ON(1);
288 		return;
289 	}
290 
291 	/* Ensure all writes are done before ring doorbell */
292 	wmb();
293 
294 	writeq(e.as_uint64, addr);
295 }
296 
297 void mana_gd_wq_ring_doorbell(struct gdma_context *gc, struct gdma_queue *queue)
298 {
299 	mana_gd_ring_doorbell(gc, queue->gdma_dev->doorbell, queue->type,
300 			      queue->id, queue->head * GDMA_WQE_BU_SIZE, 1);
301 }
302 
303 void mana_gd_ring_cq(struct gdma_queue *cq, u8 arm_bit)
304 {
305 	struct gdma_context *gc = cq->gdma_dev->gdma_context;
306 
307 	u32 num_cqe = cq->queue_size / GDMA_CQE_SIZE;
308 
309 	u32 head = cq->head % (num_cqe << GDMA_CQE_OWNER_BITS);
310 
311 	mana_gd_ring_doorbell(gc, cq->gdma_dev->doorbell, cq->type, cq->id,
312 			      head, arm_bit);
313 }
314 
315 static void mana_gd_process_eqe(struct gdma_queue *eq)
316 {
317 	u32 head = eq->head % (eq->queue_size / GDMA_EQE_SIZE);
318 	struct gdma_context *gc = eq->gdma_dev->gdma_context;
319 	struct gdma_eqe *eq_eqe_ptr = eq->queue_mem_ptr;
320 	union gdma_eqe_info eqe_info;
321 	enum gdma_eqe_type type;
322 	struct gdma_event event;
323 	struct gdma_queue *cq;
324 	struct gdma_eqe *eqe;
325 	u32 cq_id;
326 
327 	eqe = &eq_eqe_ptr[head];
328 	eqe_info.as_uint32 = eqe->eqe_info;
329 	type = eqe_info.type;
330 
331 	switch (type) {
332 	case GDMA_EQE_COMPLETION:
333 		cq_id = eqe->details[0] & 0xFFFFFF;
334 		if (WARN_ON_ONCE(cq_id >= gc->max_num_cqs))
335 			break;
336 
337 		cq = gc->cq_table[cq_id];
338 		if (WARN_ON_ONCE(!cq || cq->type != GDMA_CQ || cq->id != cq_id))
339 			break;
340 
341 		if (cq->cq.callback)
342 			cq->cq.callback(cq->cq.context, cq);
343 
344 		break;
345 
346 	case GDMA_EQE_TEST_EVENT:
347 		gc->test_event_eq_id = eq->id;
348 		complete(&gc->eq_test_event);
349 		break;
350 
351 	case GDMA_EQE_HWC_INIT_EQ_ID_DB:
352 	case GDMA_EQE_HWC_INIT_DATA:
353 	case GDMA_EQE_HWC_INIT_DONE:
354 		if (!eq->eq.callback)
355 			break;
356 
357 		event.type = type;
358 		memcpy(&event.details, &eqe->details, GDMA_EVENT_DATA_SIZE);
359 		eq->eq.callback(eq->eq.context, eq, &event);
360 		break;
361 
362 	default:
363 		break;
364 	}
365 }
366 
367 static void mana_gd_process_eq_events(void *arg)
368 {
369 	u32 owner_bits, new_bits, old_bits;
370 	union gdma_eqe_info eqe_info;
371 	struct gdma_eqe *eq_eqe_ptr;
372 	struct gdma_queue *eq = arg;
373 	struct gdma_context *gc;
374 	struct gdma_eqe *eqe;
375 	u32 head, num_eqe;
376 	int i;
377 
378 	gc = eq->gdma_dev->gdma_context;
379 
380 	num_eqe = eq->queue_size / GDMA_EQE_SIZE;
381 	eq_eqe_ptr = eq->queue_mem_ptr;
382 
383 	/* Process up to 5 EQEs at a time, and update the HW head. */
384 	for (i = 0; i < 5; i++) {
385 		eqe = &eq_eqe_ptr[eq->head % num_eqe];
386 		eqe_info.as_uint32 = eqe->eqe_info;
387 		owner_bits = eqe_info.owner_bits;
388 
389 		old_bits = (eq->head / num_eqe - 1) & GDMA_EQE_OWNER_MASK;
390 		/* No more entries */
391 		if (owner_bits == old_bits)
392 			break;
393 
394 		new_bits = (eq->head / num_eqe) & GDMA_EQE_OWNER_MASK;
395 		if (owner_bits != new_bits) {
396 			dev_err(gc->dev, "EQ %d: overflow detected\n", eq->id);
397 			break;
398 		}
399 
400 		mana_gd_process_eqe(eq);
401 
402 		eq->head++;
403 	}
404 
405 	head = eq->head % (num_eqe << GDMA_EQE_OWNER_BITS);
406 
407 	mana_gd_ring_doorbell(gc, eq->gdma_dev->doorbell, eq->type, eq->id,
408 			      head, SET_ARM_BIT);
409 }
410 
411 static int mana_gd_register_irq(struct gdma_queue *queue,
412 				const struct gdma_queue_spec *spec)
413 {
414 	struct gdma_dev *gd = queue->gdma_dev;
415 	struct gdma_irq_context *gic;
416 	struct gdma_context *gc;
417 	struct gdma_resource *r;
418 	unsigned int msi_index;
419 	unsigned long flags;
420 	struct device *dev;
421 	int err = 0;
422 
423 	gc = gd->gdma_context;
424 	r = &gc->msix_resource;
425 	dev = gc->dev;
426 
427 	spin_lock_irqsave(&r->lock, flags);
428 
429 	msi_index = find_first_zero_bit(r->map, r->size);
430 	if (msi_index >= r->size || msi_index >= gc->num_msix_usable) {
431 		err = -ENOSPC;
432 	} else {
433 		bitmap_set(r->map, msi_index, 1);
434 		queue->eq.msix_index = msi_index;
435 	}
436 
437 	spin_unlock_irqrestore(&r->lock, flags);
438 
439 	if (err) {
440 		dev_err(dev, "Register IRQ err:%d, msi:%u rsize:%u, nMSI:%u",
441 			err, msi_index, r->size, gc->num_msix_usable);
442 
443 		return err;
444 	}
445 
446 	gic = &gc->irq_contexts[msi_index];
447 
448 	WARN_ON(gic->handler || gic->arg);
449 
450 	gic->arg = queue;
451 
452 	gic->handler = mana_gd_process_eq_events;
453 
454 	return 0;
455 }
456 
457 static void mana_gd_deregiser_irq(struct gdma_queue *queue)
458 {
459 	struct gdma_dev *gd = queue->gdma_dev;
460 	struct gdma_irq_context *gic;
461 	struct gdma_context *gc;
462 	struct gdma_resource *r;
463 	unsigned int msix_index;
464 	unsigned long flags;
465 
466 	gc = gd->gdma_context;
467 	r = &gc->msix_resource;
468 
469 	/* At most num_online_cpus() + 1 interrupts are used. */
470 	msix_index = queue->eq.msix_index;
471 	if (WARN_ON(msix_index >= gc->num_msix_usable))
472 		return;
473 
474 	gic = &gc->irq_contexts[msix_index];
475 	gic->handler = NULL;
476 	gic->arg = NULL;
477 
478 	spin_lock_irqsave(&r->lock, flags);
479 	bitmap_clear(r->map, msix_index, 1);
480 	spin_unlock_irqrestore(&r->lock, flags);
481 
482 	queue->eq.msix_index = INVALID_PCI_MSIX_INDEX;
483 }
484 
485 int mana_gd_test_eq(struct gdma_context *gc, struct gdma_queue *eq)
486 {
487 	struct gdma_generate_test_event_req req = {};
488 	struct gdma_general_resp resp = {};
489 	struct device *dev = gc->dev;
490 	int err;
491 
492 	mutex_lock(&gc->eq_test_event_mutex);
493 
494 	init_completion(&gc->eq_test_event);
495 	gc->test_event_eq_id = INVALID_QUEUE_ID;
496 
497 	mana_gd_init_req_hdr(&req.hdr, GDMA_GENERATE_TEST_EQE,
498 			     sizeof(req), sizeof(resp));
499 
500 	req.hdr.dev_id = eq->gdma_dev->dev_id;
501 	req.queue_index = eq->id;
502 
503 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
504 	if (err) {
505 		dev_err(dev, "test_eq failed: %d\n", err);
506 		goto out;
507 	}
508 
509 	err = -EPROTO;
510 
511 	if (resp.hdr.status) {
512 		dev_err(dev, "test_eq failed: 0x%x\n", resp.hdr.status);
513 		goto out;
514 	}
515 
516 	if (!wait_for_completion_timeout(&gc->eq_test_event, 30 * HZ)) {
517 		dev_err(dev, "test_eq timed out on queue %d\n", eq->id);
518 		goto out;
519 	}
520 
521 	if (eq->id != gc->test_event_eq_id) {
522 		dev_err(dev, "test_eq got an event on wrong queue %d (%d)\n",
523 			gc->test_event_eq_id, eq->id);
524 		goto out;
525 	}
526 
527 	err = 0;
528 out:
529 	mutex_unlock(&gc->eq_test_event_mutex);
530 	return err;
531 }
532 
533 static void mana_gd_destroy_eq(struct gdma_context *gc, bool flush_evenets,
534 			       struct gdma_queue *queue)
535 {
536 	int err;
537 
538 	if (flush_evenets) {
539 		err = mana_gd_test_eq(gc, queue);
540 		if (err)
541 			dev_warn(gc->dev, "Failed to flush EQ: %d\n", err);
542 	}
543 
544 	mana_gd_deregiser_irq(queue);
545 
546 	if (queue->eq.disable_needed)
547 		mana_gd_disable_queue(queue);
548 }
549 
550 static int mana_gd_create_eq(struct gdma_dev *gd,
551 			     const struct gdma_queue_spec *spec,
552 			     bool create_hwq, struct gdma_queue *queue)
553 {
554 	struct gdma_context *gc = gd->gdma_context;
555 	struct device *dev = gc->dev;
556 	u32 log2_num_entries;
557 	int err;
558 
559 	queue->eq.msix_index = INVALID_PCI_MSIX_INDEX;
560 
561 	log2_num_entries = ilog2(queue->queue_size / GDMA_EQE_SIZE);
562 
563 	if (spec->eq.log2_throttle_limit > log2_num_entries) {
564 		dev_err(dev, "EQ throttling limit (%lu) > maximum EQE (%u)\n",
565 			spec->eq.log2_throttle_limit, log2_num_entries);
566 		return -EINVAL;
567 	}
568 
569 	err = mana_gd_register_irq(queue, spec);
570 	if (err) {
571 		dev_err(dev, "Failed to register irq: %d\n", err);
572 		return err;
573 	}
574 
575 	queue->eq.callback = spec->eq.callback;
576 	queue->eq.context = spec->eq.context;
577 	queue->head |= INITIALIZED_OWNER_BIT(log2_num_entries);
578 	queue->eq.log2_throttle_limit = spec->eq.log2_throttle_limit ?: 1;
579 
580 	if (create_hwq) {
581 		err = mana_gd_create_hw_eq(gc, queue);
582 		if (err)
583 			goto out;
584 
585 		err = mana_gd_test_eq(gc, queue);
586 		if (err)
587 			goto out;
588 	}
589 
590 	return 0;
591 out:
592 	dev_err(dev, "Failed to create EQ: %d\n", err);
593 	mana_gd_destroy_eq(gc, false, queue);
594 	return err;
595 }
596 
597 static void mana_gd_create_cq(const struct gdma_queue_spec *spec,
598 			      struct gdma_queue *queue)
599 {
600 	u32 log2_num_entries = ilog2(spec->queue_size / GDMA_CQE_SIZE);
601 
602 	queue->head |= INITIALIZED_OWNER_BIT(log2_num_entries);
603 	queue->cq.parent = spec->cq.parent_eq;
604 	queue->cq.context = spec->cq.context;
605 	queue->cq.callback = spec->cq.callback;
606 }
607 
608 static void mana_gd_destroy_cq(struct gdma_context *gc,
609 			       struct gdma_queue *queue)
610 {
611 	u32 id = queue->id;
612 
613 	if (id >= gc->max_num_cqs)
614 		return;
615 
616 	if (!gc->cq_table[id])
617 		return;
618 
619 	gc->cq_table[id] = NULL;
620 }
621 
622 int mana_gd_create_hwc_queue(struct gdma_dev *gd,
623 			     const struct gdma_queue_spec *spec,
624 			     struct gdma_queue **queue_ptr)
625 {
626 	struct gdma_context *gc = gd->gdma_context;
627 	struct gdma_mem_info *gmi;
628 	struct gdma_queue *queue;
629 	int err;
630 
631 	queue = kzalloc(sizeof(*queue), GFP_KERNEL);
632 	if (!queue)
633 		return -ENOMEM;
634 
635 	gmi = &queue->mem_info;
636 	err = mana_gd_alloc_memory(gc, spec->queue_size, gmi);
637 	if (err)
638 		goto free_q;
639 
640 	queue->head = 0;
641 	queue->tail = 0;
642 	queue->queue_mem_ptr = gmi->virt_addr;
643 	queue->queue_size = spec->queue_size;
644 	queue->monitor_avl_buf = spec->monitor_avl_buf;
645 	queue->type = spec->type;
646 	queue->gdma_dev = gd;
647 
648 	if (spec->type == GDMA_EQ)
649 		err = mana_gd_create_eq(gd, spec, false, queue);
650 	else if (spec->type == GDMA_CQ)
651 		mana_gd_create_cq(spec, queue);
652 
653 	if (err)
654 		goto out;
655 
656 	*queue_ptr = queue;
657 	return 0;
658 out:
659 	mana_gd_free_memory(gmi);
660 free_q:
661 	kfree(queue);
662 	return err;
663 }
664 
665 static void mana_gd_destroy_dma_region(struct gdma_context *gc, u64 gdma_region)
666 {
667 	struct gdma_destroy_dma_region_req req = {};
668 	struct gdma_general_resp resp = {};
669 	int err;
670 
671 	if (gdma_region == GDMA_INVALID_DMA_REGION)
672 		return;
673 
674 	mana_gd_init_req_hdr(&req.hdr, GDMA_DESTROY_DMA_REGION, sizeof(req),
675 			     sizeof(resp));
676 	req.gdma_region = gdma_region;
677 
678 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
679 	if (err || resp.hdr.status)
680 		dev_err(gc->dev, "Failed to destroy DMA region: %d, 0x%x\n",
681 			err, resp.hdr.status);
682 }
683 
684 static int mana_gd_create_dma_region(struct gdma_dev *gd,
685 				     struct gdma_mem_info *gmi)
686 {
687 	unsigned int num_page = gmi->length / PAGE_SIZE;
688 	struct gdma_create_dma_region_req *req = NULL;
689 	struct gdma_create_dma_region_resp resp = {};
690 	struct gdma_context *gc = gd->gdma_context;
691 	struct hw_channel_context *hwc;
692 	u32 length = gmi->length;
693 	size_t req_msg_size;
694 	int err;
695 	int i;
696 
697 	if (length < PAGE_SIZE || !is_power_of_2(length))
698 		return -EINVAL;
699 
700 	if (offset_in_page(gmi->virt_addr) != 0)
701 		return -EINVAL;
702 
703 	hwc = gc->hwc.driver_data;
704 	req_msg_size = struct_size(req, page_addr_list, num_page);
705 	if (req_msg_size > hwc->max_req_msg_size)
706 		return -EINVAL;
707 
708 	req = kzalloc(req_msg_size, GFP_KERNEL);
709 	if (!req)
710 		return -ENOMEM;
711 
712 	mana_gd_init_req_hdr(&req->hdr, GDMA_CREATE_DMA_REGION,
713 			     req_msg_size, sizeof(resp));
714 	req->length = length;
715 	req->offset_in_page = 0;
716 	req->gdma_page_type = GDMA_PAGE_TYPE_4K;
717 	req->page_count = num_page;
718 	req->page_addr_list_len = num_page;
719 
720 	for (i = 0; i < num_page; i++)
721 		req->page_addr_list[i] = gmi->dma_handle +  i * PAGE_SIZE;
722 
723 	err = mana_gd_send_request(gc, req_msg_size, req, sizeof(resp), &resp);
724 	if (err)
725 		goto out;
726 
727 	if (resp.hdr.status || resp.gdma_region == GDMA_INVALID_DMA_REGION) {
728 		dev_err(gc->dev, "Failed to create DMA region: 0x%x\n",
729 			resp.hdr.status);
730 		err = -EPROTO;
731 		goto out;
732 	}
733 
734 	gmi->gdma_region = resp.gdma_region;
735 out:
736 	kfree(req);
737 	return err;
738 }
739 
740 int mana_gd_create_mana_eq(struct gdma_dev *gd,
741 			   const struct gdma_queue_spec *spec,
742 			   struct gdma_queue **queue_ptr)
743 {
744 	struct gdma_context *gc = gd->gdma_context;
745 	struct gdma_mem_info *gmi;
746 	struct gdma_queue *queue;
747 	int err;
748 
749 	if (spec->type != GDMA_EQ)
750 		return -EINVAL;
751 
752 	queue = kzalloc(sizeof(*queue), GFP_KERNEL);
753 	if (!queue)
754 		return -ENOMEM;
755 
756 	gmi = &queue->mem_info;
757 	err = mana_gd_alloc_memory(gc, spec->queue_size, gmi);
758 	if (err)
759 		goto free_q;
760 
761 	err = mana_gd_create_dma_region(gd, gmi);
762 	if (err)
763 		goto out;
764 
765 	queue->head = 0;
766 	queue->tail = 0;
767 	queue->queue_mem_ptr = gmi->virt_addr;
768 	queue->queue_size = spec->queue_size;
769 	queue->monitor_avl_buf = spec->monitor_avl_buf;
770 	queue->type = spec->type;
771 	queue->gdma_dev = gd;
772 
773 	err = mana_gd_create_eq(gd, spec, true, queue);
774 	if (err)
775 		goto out;
776 
777 	*queue_ptr = queue;
778 	return 0;
779 out:
780 	mana_gd_free_memory(gmi);
781 free_q:
782 	kfree(queue);
783 	return err;
784 }
785 
786 int mana_gd_create_mana_wq_cq(struct gdma_dev *gd,
787 			      const struct gdma_queue_spec *spec,
788 			      struct gdma_queue **queue_ptr)
789 {
790 	struct gdma_context *gc = gd->gdma_context;
791 	struct gdma_mem_info *gmi;
792 	struct gdma_queue *queue;
793 	int err;
794 
795 	if (spec->type != GDMA_CQ && spec->type != GDMA_SQ &&
796 	    spec->type != GDMA_RQ)
797 		return -EINVAL;
798 
799 	queue = kzalloc(sizeof(*queue), GFP_KERNEL);
800 	if (!queue)
801 		return -ENOMEM;
802 
803 	gmi = &queue->mem_info;
804 	err = mana_gd_alloc_memory(gc, spec->queue_size, gmi);
805 	if (err)
806 		goto free_q;
807 
808 	err = mana_gd_create_dma_region(gd, gmi);
809 	if (err)
810 		goto out;
811 
812 	queue->head = 0;
813 	queue->tail = 0;
814 	queue->queue_mem_ptr = gmi->virt_addr;
815 	queue->queue_size = spec->queue_size;
816 	queue->monitor_avl_buf = spec->monitor_avl_buf;
817 	queue->type = spec->type;
818 	queue->gdma_dev = gd;
819 
820 	if (spec->type == GDMA_CQ)
821 		mana_gd_create_cq(spec, queue);
822 
823 	*queue_ptr = queue;
824 	return 0;
825 out:
826 	mana_gd_free_memory(gmi);
827 free_q:
828 	kfree(queue);
829 	return err;
830 }
831 
832 void mana_gd_destroy_queue(struct gdma_context *gc, struct gdma_queue *queue)
833 {
834 	struct gdma_mem_info *gmi = &queue->mem_info;
835 
836 	switch (queue->type) {
837 	case GDMA_EQ:
838 		mana_gd_destroy_eq(gc, queue->eq.disable_needed, queue);
839 		break;
840 
841 	case GDMA_CQ:
842 		mana_gd_destroy_cq(gc, queue);
843 		break;
844 
845 	case GDMA_RQ:
846 		break;
847 
848 	case GDMA_SQ:
849 		break;
850 
851 	default:
852 		dev_err(gc->dev, "Can't destroy unknown queue: type=%d\n",
853 			queue->type);
854 		return;
855 	}
856 
857 	mana_gd_destroy_dma_region(gc, gmi->gdma_region);
858 	mana_gd_free_memory(gmi);
859 	kfree(queue);
860 }
861 
862 int mana_gd_verify_vf_version(struct pci_dev *pdev)
863 {
864 	struct gdma_context *gc = pci_get_drvdata(pdev);
865 	struct gdma_verify_ver_resp resp = {};
866 	struct gdma_verify_ver_req req = {};
867 	int err;
868 
869 	mana_gd_init_req_hdr(&req.hdr, GDMA_VERIFY_VF_DRIVER_VERSION,
870 			     sizeof(req), sizeof(resp));
871 
872 	req.protocol_ver_min = GDMA_PROTOCOL_FIRST;
873 	req.protocol_ver_max = GDMA_PROTOCOL_LAST;
874 
875 	req.gd_drv_cap_flags1 = GDMA_DRV_CAP_FLAGS1;
876 	req.gd_drv_cap_flags2 = GDMA_DRV_CAP_FLAGS2;
877 	req.gd_drv_cap_flags3 = GDMA_DRV_CAP_FLAGS3;
878 	req.gd_drv_cap_flags4 = GDMA_DRV_CAP_FLAGS4;
879 
880 	req.drv_ver = 0;	/* Unused*/
881 	req.os_type = 0x10;	/* Linux */
882 	req.os_ver_major = LINUX_VERSION_MAJOR;
883 	req.os_ver_minor = LINUX_VERSION_PATCHLEVEL;
884 	req.os_ver_build = LINUX_VERSION_SUBLEVEL;
885 	strscpy(req.os_ver_str1, utsname()->sysname, sizeof(req.os_ver_str1));
886 	strscpy(req.os_ver_str2, utsname()->release, sizeof(req.os_ver_str2));
887 	strscpy(req.os_ver_str3, utsname()->version, sizeof(req.os_ver_str3));
888 
889 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
890 	if (err || resp.hdr.status) {
891 		dev_err(gc->dev, "VfVerifyVersionOutput: %d, status=0x%x\n",
892 			err, resp.hdr.status);
893 		return err ? err : -EPROTO;
894 	}
895 
896 	return 0;
897 }
898 
899 int mana_gd_register_device(struct gdma_dev *gd)
900 {
901 	struct gdma_context *gc = gd->gdma_context;
902 	struct gdma_register_device_resp resp = {};
903 	struct gdma_general_req req = {};
904 	int err;
905 
906 	gd->pdid = INVALID_PDID;
907 	gd->doorbell = INVALID_DOORBELL;
908 	gd->gpa_mkey = INVALID_MEM_KEY;
909 
910 	mana_gd_init_req_hdr(&req.hdr, GDMA_REGISTER_DEVICE, sizeof(req),
911 			     sizeof(resp));
912 
913 	req.hdr.dev_id = gd->dev_id;
914 
915 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
916 	if (err || resp.hdr.status) {
917 		dev_err(gc->dev, "gdma_register_device_resp failed: %d, 0x%x\n",
918 			err, resp.hdr.status);
919 		return err ? err : -EPROTO;
920 	}
921 
922 	gd->pdid = resp.pdid;
923 	gd->gpa_mkey = resp.gpa_mkey;
924 	gd->doorbell = resp.db_id;
925 
926 	return 0;
927 }
928 
929 int mana_gd_deregister_device(struct gdma_dev *gd)
930 {
931 	struct gdma_context *gc = gd->gdma_context;
932 	struct gdma_general_resp resp = {};
933 	struct gdma_general_req req = {};
934 	int err;
935 
936 	if (gd->pdid == INVALID_PDID)
937 		return -EINVAL;
938 
939 	mana_gd_init_req_hdr(&req.hdr, GDMA_DEREGISTER_DEVICE, sizeof(req),
940 			     sizeof(resp));
941 
942 	req.hdr.dev_id = gd->dev_id;
943 
944 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
945 	if (err || resp.hdr.status) {
946 		dev_err(gc->dev, "Failed to deregister device: %d, 0x%x\n",
947 			err, resp.hdr.status);
948 		if (!err)
949 			err = -EPROTO;
950 	}
951 
952 	gd->pdid = INVALID_PDID;
953 	gd->doorbell = INVALID_DOORBELL;
954 	gd->gpa_mkey = INVALID_MEM_KEY;
955 
956 	return err;
957 }
958 
959 u32 mana_gd_wq_avail_space(struct gdma_queue *wq)
960 {
961 	u32 used_space = (wq->head - wq->tail) * GDMA_WQE_BU_SIZE;
962 	u32 wq_size = wq->queue_size;
963 
964 	WARN_ON_ONCE(used_space > wq_size);
965 
966 	return wq_size - used_space;
967 }
968 
969 u8 *mana_gd_get_wqe_ptr(const struct gdma_queue *wq, u32 wqe_offset)
970 {
971 	u32 offset = (wqe_offset * GDMA_WQE_BU_SIZE) & (wq->queue_size - 1);
972 
973 	WARN_ON_ONCE((offset + GDMA_WQE_BU_SIZE) > wq->queue_size);
974 
975 	return wq->queue_mem_ptr + offset;
976 }
977 
978 static u32 mana_gd_write_client_oob(const struct gdma_wqe_request *wqe_req,
979 				    enum gdma_queue_type q_type,
980 				    u32 client_oob_size, u32 sgl_data_size,
981 				    u8 *wqe_ptr)
982 {
983 	bool oob_in_sgl = !!(wqe_req->flags & GDMA_WR_OOB_IN_SGL);
984 	bool pad_data = !!(wqe_req->flags & GDMA_WR_PAD_BY_SGE0);
985 	struct gdma_wqe *header = (struct gdma_wqe *)wqe_ptr;
986 	u8 *ptr;
987 
988 	memset(header, 0, sizeof(struct gdma_wqe));
989 	header->num_sge = wqe_req->num_sge;
990 	header->inline_oob_size_div4 = client_oob_size / sizeof(u32);
991 
992 	if (oob_in_sgl) {
993 		WARN_ON_ONCE(!pad_data || wqe_req->num_sge < 2);
994 
995 		header->client_oob_in_sgl = 1;
996 
997 		if (pad_data)
998 			header->last_vbytes = wqe_req->sgl[0].size;
999 	}
1000 
1001 	if (q_type == GDMA_SQ)
1002 		header->client_data_unit = wqe_req->client_data_unit;
1003 
1004 	/* The size of gdma_wqe + client_oob_size must be less than or equal
1005 	 * to one Basic Unit (i.e. 32 bytes), so the pointer can't go beyond
1006 	 * the queue memory buffer boundary.
1007 	 */
1008 	ptr = wqe_ptr + sizeof(header);
1009 
1010 	if (wqe_req->inline_oob_data && wqe_req->inline_oob_size > 0) {
1011 		memcpy(ptr, wqe_req->inline_oob_data, wqe_req->inline_oob_size);
1012 
1013 		if (client_oob_size > wqe_req->inline_oob_size)
1014 			memset(ptr + wqe_req->inline_oob_size, 0,
1015 			       client_oob_size - wqe_req->inline_oob_size);
1016 	}
1017 
1018 	return sizeof(header) + client_oob_size;
1019 }
1020 
1021 static void mana_gd_write_sgl(struct gdma_queue *wq, u8 *wqe_ptr,
1022 			      const struct gdma_wqe_request *wqe_req)
1023 {
1024 	u32 sgl_size = sizeof(struct gdma_sge) * wqe_req->num_sge;
1025 	const u8 *address = (u8 *)wqe_req->sgl;
1026 	u8 *base_ptr, *end_ptr;
1027 	u32 size_to_end;
1028 
1029 	base_ptr = wq->queue_mem_ptr;
1030 	end_ptr = base_ptr + wq->queue_size;
1031 	size_to_end = (u32)(end_ptr - wqe_ptr);
1032 
1033 	if (size_to_end < sgl_size) {
1034 		memcpy(wqe_ptr, address, size_to_end);
1035 
1036 		wqe_ptr = base_ptr;
1037 		address += size_to_end;
1038 		sgl_size -= size_to_end;
1039 	}
1040 
1041 	memcpy(wqe_ptr, address, sgl_size);
1042 }
1043 
1044 int mana_gd_post_work_request(struct gdma_queue *wq,
1045 			      const struct gdma_wqe_request *wqe_req,
1046 			      struct gdma_posted_wqe_info *wqe_info)
1047 {
1048 	u32 client_oob_size = wqe_req->inline_oob_size;
1049 	struct gdma_context *gc;
1050 	u32 sgl_data_size;
1051 	u32 max_wqe_size;
1052 	u32 wqe_size;
1053 	u8 *wqe_ptr;
1054 
1055 	if (wqe_req->num_sge == 0)
1056 		return -EINVAL;
1057 
1058 	if (wq->type == GDMA_RQ) {
1059 		if (client_oob_size != 0)
1060 			return -EINVAL;
1061 
1062 		client_oob_size = INLINE_OOB_SMALL_SIZE;
1063 
1064 		max_wqe_size = GDMA_MAX_RQE_SIZE;
1065 	} else {
1066 		if (client_oob_size != INLINE_OOB_SMALL_SIZE &&
1067 		    client_oob_size != INLINE_OOB_LARGE_SIZE)
1068 			return -EINVAL;
1069 
1070 		max_wqe_size = GDMA_MAX_SQE_SIZE;
1071 	}
1072 
1073 	sgl_data_size = sizeof(struct gdma_sge) * wqe_req->num_sge;
1074 	wqe_size = ALIGN(sizeof(struct gdma_wqe) + client_oob_size +
1075 			 sgl_data_size, GDMA_WQE_BU_SIZE);
1076 	if (wqe_size > max_wqe_size)
1077 		return -EINVAL;
1078 
1079 	if (wq->monitor_avl_buf && wqe_size > mana_gd_wq_avail_space(wq)) {
1080 		gc = wq->gdma_dev->gdma_context;
1081 		dev_err(gc->dev, "unsuccessful flow control!\n");
1082 		return -ENOSPC;
1083 	}
1084 
1085 	if (wqe_info)
1086 		wqe_info->wqe_size_in_bu = wqe_size / GDMA_WQE_BU_SIZE;
1087 
1088 	wqe_ptr = mana_gd_get_wqe_ptr(wq, wq->head);
1089 	wqe_ptr += mana_gd_write_client_oob(wqe_req, wq->type, client_oob_size,
1090 					    sgl_data_size, wqe_ptr);
1091 	if (wqe_ptr >= (u8 *)wq->queue_mem_ptr + wq->queue_size)
1092 		wqe_ptr -= wq->queue_size;
1093 
1094 	mana_gd_write_sgl(wq, wqe_ptr, wqe_req);
1095 
1096 	wq->head += wqe_size / GDMA_WQE_BU_SIZE;
1097 
1098 	return 0;
1099 }
1100 
1101 int mana_gd_post_and_ring(struct gdma_queue *queue,
1102 			  const struct gdma_wqe_request *wqe_req,
1103 			  struct gdma_posted_wqe_info *wqe_info)
1104 {
1105 	struct gdma_context *gc = queue->gdma_dev->gdma_context;
1106 	int err;
1107 
1108 	err = mana_gd_post_work_request(queue, wqe_req, wqe_info);
1109 	if (err)
1110 		return err;
1111 
1112 	mana_gd_wq_ring_doorbell(gc, queue);
1113 
1114 	return 0;
1115 }
1116 
1117 static int mana_gd_read_cqe(struct gdma_queue *cq, struct gdma_comp *comp)
1118 {
1119 	unsigned int num_cqe = cq->queue_size / sizeof(struct gdma_cqe);
1120 	struct gdma_cqe *cq_cqe = cq->queue_mem_ptr;
1121 	u32 owner_bits, new_bits, old_bits;
1122 	struct gdma_cqe *cqe;
1123 
1124 	cqe = &cq_cqe[cq->head % num_cqe];
1125 	owner_bits = cqe->cqe_info.owner_bits;
1126 
1127 	old_bits = (cq->head / num_cqe - 1) & GDMA_CQE_OWNER_MASK;
1128 	/* Return 0 if no more entries. */
1129 	if (owner_bits == old_bits)
1130 		return 0;
1131 
1132 	new_bits = (cq->head / num_cqe) & GDMA_CQE_OWNER_MASK;
1133 	/* Return -1 if overflow detected. */
1134 	if (WARN_ON_ONCE(owner_bits != new_bits))
1135 		return -1;
1136 
1137 	comp->wq_num = cqe->cqe_info.wq_num;
1138 	comp->is_sq = cqe->cqe_info.is_sq;
1139 	memcpy(comp->cqe_data, cqe->cqe_data, GDMA_COMP_DATA_SIZE);
1140 
1141 	return 1;
1142 }
1143 
1144 int mana_gd_poll_cq(struct gdma_queue *cq, struct gdma_comp *comp, int num_cqe)
1145 {
1146 	int cqe_idx;
1147 	int ret;
1148 
1149 	for (cqe_idx = 0; cqe_idx < num_cqe; cqe_idx++) {
1150 		ret = mana_gd_read_cqe(cq, &comp[cqe_idx]);
1151 
1152 		if (ret < 0) {
1153 			cq->head -= cqe_idx;
1154 			return ret;
1155 		}
1156 
1157 		if (ret == 0)
1158 			break;
1159 
1160 		cq->head++;
1161 	}
1162 
1163 	return cqe_idx;
1164 }
1165 
1166 static irqreturn_t mana_gd_intr(int irq, void *arg)
1167 {
1168 	struct gdma_irq_context *gic = arg;
1169 
1170 	if (gic->handler)
1171 		gic->handler(gic->arg);
1172 
1173 	return IRQ_HANDLED;
1174 }
1175 
1176 int mana_gd_alloc_res_map(u32 res_avail, struct gdma_resource *r)
1177 {
1178 	r->map = bitmap_zalloc(res_avail, GFP_KERNEL);
1179 	if (!r->map)
1180 		return -ENOMEM;
1181 
1182 	r->size = res_avail;
1183 	spin_lock_init(&r->lock);
1184 
1185 	return 0;
1186 }
1187 
1188 void mana_gd_free_res_map(struct gdma_resource *r)
1189 {
1190 	bitmap_free(r->map);
1191 	r->map = NULL;
1192 	r->size = 0;
1193 }
1194 
1195 static int mana_gd_setup_irqs(struct pci_dev *pdev)
1196 {
1197 	unsigned int max_queues_per_port = num_online_cpus();
1198 	struct gdma_context *gc = pci_get_drvdata(pdev);
1199 	struct gdma_irq_context *gic;
1200 	unsigned int max_irqs;
1201 	int nvec, irq;
1202 	int err, i, j;
1203 
1204 	if (max_queues_per_port > MANA_MAX_NUM_QUEUES)
1205 		max_queues_per_port = MANA_MAX_NUM_QUEUES;
1206 
1207 	/* Need 1 interrupt for the Hardware communication Channel (HWC) */
1208 	max_irqs = max_queues_per_port + 1;
1209 
1210 	nvec = pci_alloc_irq_vectors(pdev, 2, max_irqs, PCI_IRQ_MSIX);
1211 	if (nvec < 0)
1212 		return nvec;
1213 
1214 	gc->irq_contexts = kcalloc(nvec, sizeof(struct gdma_irq_context),
1215 				   GFP_KERNEL);
1216 	if (!gc->irq_contexts) {
1217 		err = -ENOMEM;
1218 		goto free_irq_vector;
1219 	}
1220 
1221 	for (i = 0; i < nvec; i++) {
1222 		gic = &gc->irq_contexts[i];
1223 		gic->handler = NULL;
1224 		gic->arg = NULL;
1225 
1226 		irq = pci_irq_vector(pdev, i);
1227 		if (irq < 0) {
1228 			err = irq;
1229 			goto free_irq;
1230 		}
1231 
1232 		err = request_irq(irq, mana_gd_intr, 0, "mana_intr", gic);
1233 		if (err)
1234 			goto free_irq;
1235 	}
1236 
1237 	err = mana_gd_alloc_res_map(nvec, &gc->msix_resource);
1238 	if (err)
1239 		goto free_irq;
1240 
1241 	gc->max_num_msix = nvec;
1242 	gc->num_msix_usable = nvec;
1243 
1244 	return 0;
1245 
1246 free_irq:
1247 	for (j = i - 1; j >= 0; j--) {
1248 		irq = pci_irq_vector(pdev, j);
1249 		gic = &gc->irq_contexts[j];
1250 		free_irq(irq, gic);
1251 	}
1252 
1253 	kfree(gc->irq_contexts);
1254 	gc->irq_contexts = NULL;
1255 free_irq_vector:
1256 	pci_free_irq_vectors(pdev);
1257 	return err;
1258 }
1259 
1260 static void mana_gd_remove_irqs(struct pci_dev *pdev)
1261 {
1262 	struct gdma_context *gc = pci_get_drvdata(pdev);
1263 	struct gdma_irq_context *gic;
1264 	int irq, i;
1265 
1266 	if (gc->max_num_msix < 1)
1267 		return;
1268 
1269 	mana_gd_free_res_map(&gc->msix_resource);
1270 
1271 	for (i = 0; i < gc->max_num_msix; i++) {
1272 		irq = pci_irq_vector(pdev, i);
1273 		if (irq < 0)
1274 			continue;
1275 
1276 		gic = &gc->irq_contexts[i];
1277 		free_irq(irq, gic);
1278 	}
1279 
1280 	pci_free_irq_vectors(pdev);
1281 
1282 	gc->max_num_msix = 0;
1283 	gc->num_msix_usable = 0;
1284 	kfree(gc->irq_contexts);
1285 	gc->irq_contexts = NULL;
1286 }
1287 
1288 static int mana_gd_setup(struct pci_dev *pdev)
1289 {
1290 	struct gdma_context *gc = pci_get_drvdata(pdev);
1291 	int err;
1292 
1293 	mana_gd_init_registers(pdev);
1294 	mana_smc_init(&gc->shm_channel, gc->dev, gc->shm_base);
1295 
1296 	err = mana_gd_setup_irqs(pdev);
1297 	if (err)
1298 		return err;
1299 
1300 	err = mana_hwc_create_channel(gc);
1301 	if (err)
1302 		goto remove_irq;
1303 
1304 	err = mana_gd_verify_vf_version(pdev);
1305 	if (err)
1306 		goto destroy_hwc;
1307 
1308 	err = mana_gd_query_max_resources(pdev);
1309 	if (err)
1310 		goto destroy_hwc;
1311 
1312 	err = mana_gd_detect_devices(pdev);
1313 	if (err)
1314 		goto destroy_hwc;
1315 
1316 	return 0;
1317 
1318 destroy_hwc:
1319 	mana_hwc_destroy_channel(gc);
1320 remove_irq:
1321 	mana_gd_remove_irqs(pdev);
1322 	return err;
1323 }
1324 
1325 static void mana_gd_cleanup(struct pci_dev *pdev)
1326 {
1327 	struct gdma_context *gc = pci_get_drvdata(pdev);
1328 
1329 	mana_hwc_destroy_channel(gc);
1330 
1331 	mana_gd_remove_irqs(pdev);
1332 }
1333 
1334 static bool mana_is_pf(unsigned short dev_id)
1335 {
1336 	return dev_id == MANA_PF_DEVICE_ID;
1337 }
1338 
1339 static int mana_gd_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
1340 {
1341 	struct gdma_context *gc;
1342 	void __iomem *bar0_va;
1343 	int bar = 0;
1344 	int err;
1345 
1346 	/* Each port has 2 CQs, each CQ has at most 1 EQE at a time */
1347 	BUILD_BUG_ON(2 * MAX_PORTS_IN_MANA_DEV * GDMA_EQE_SIZE > EQ_SIZE);
1348 
1349 	err = pci_enable_device(pdev);
1350 	if (err)
1351 		return -ENXIO;
1352 
1353 	pci_set_master(pdev);
1354 
1355 	err = pci_request_regions(pdev, "mana");
1356 	if (err)
1357 		goto disable_dev;
1358 
1359 	err = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
1360 	if (err)
1361 		goto release_region;
1362 
1363 	err = -ENOMEM;
1364 	gc = vzalloc(sizeof(*gc));
1365 	if (!gc)
1366 		goto release_region;
1367 
1368 	mutex_init(&gc->eq_test_event_mutex);
1369 	pci_set_drvdata(pdev, gc);
1370 
1371 	bar0_va = pci_iomap(pdev, bar, 0);
1372 	if (!bar0_va)
1373 		goto free_gc;
1374 
1375 	gc->is_pf = mana_is_pf(pdev->device);
1376 	gc->bar0_va = bar0_va;
1377 	gc->dev = &pdev->dev;
1378 
1379 	err = mana_gd_setup(pdev);
1380 	if (err)
1381 		goto unmap_bar;
1382 
1383 	err = mana_probe(&gc->mana, false);
1384 	if (err)
1385 		goto cleanup_gd;
1386 
1387 	return 0;
1388 
1389 cleanup_gd:
1390 	mana_gd_cleanup(pdev);
1391 unmap_bar:
1392 	pci_iounmap(pdev, bar0_va);
1393 free_gc:
1394 	pci_set_drvdata(pdev, NULL);
1395 	vfree(gc);
1396 release_region:
1397 	pci_release_regions(pdev);
1398 disable_dev:
1399 	pci_clear_master(pdev);
1400 	pci_disable_device(pdev);
1401 	dev_err(&pdev->dev, "gdma probe failed: err = %d\n", err);
1402 	return err;
1403 }
1404 
1405 static void mana_gd_remove(struct pci_dev *pdev)
1406 {
1407 	struct gdma_context *gc = pci_get_drvdata(pdev);
1408 
1409 	mana_remove(&gc->mana, false);
1410 
1411 	mana_gd_cleanup(pdev);
1412 
1413 	pci_iounmap(pdev, gc->bar0_va);
1414 
1415 	vfree(gc);
1416 
1417 	pci_release_regions(pdev);
1418 	pci_clear_master(pdev);
1419 	pci_disable_device(pdev);
1420 }
1421 
1422 /* The 'state' parameter is not used. */
1423 static int mana_gd_suspend(struct pci_dev *pdev, pm_message_t state)
1424 {
1425 	struct gdma_context *gc = pci_get_drvdata(pdev);
1426 
1427 	mana_remove(&gc->mana, true);
1428 
1429 	mana_gd_cleanup(pdev);
1430 
1431 	return 0;
1432 }
1433 
1434 /* In case the NIC hardware stops working, the suspend and resume callbacks will
1435  * fail -- if this happens, it's safer to just report an error than try to undo
1436  * what has been done.
1437  */
1438 static int mana_gd_resume(struct pci_dev *pdev)
1439 {
1440 	struct gdma_context *gc = pci_get_drvdata(pdev);
1441 	int err;
1442 
1443 	err = mana_gd_setup(pdev);
1444 	if (err)
1445 		return err;
1446 
1447 	err = mana_probe(&gc->mana, true);
1448 	if (err)
1449 		return err;
1450 
1451 	return 0;
1452 }
1453 
1454 /* Quiesce the device for kexec. This is also called upon reboot/shutdown. */
1455 static void mana_gd_shutdown(struct pci_dev *pdev)
1456 {
1457 	struct gdma_context *gc = pci_get_drvdata(pdev);
1458 
1459 	dev_info(&pdev->dev, "Shutdown was called\n");
1460 
1461 	mana_remove(&gc->mana, true);
1462 
1463 	mana_gd_cleanup(pdev);
1464 
1465 	pci_disable_device(pdev);
1466 }
1467 
1468 #ifndef PCI_VENDOR_ID_MICROSOFT
1469 #define PCI_VENDOR_ID_MICROSOFT 0x1414
1470 #endif
1471 
1472 static const struct pci_device_id mana_id_table[] = {
1473 	{ PCI_DEVICE(PCI_VENDOR_ID_MICROSOFT, MANA_PF_DEVICE_ID) },
1474 	{ PCI_DEVICE(PCI_VENDOR_ID_MICROSOFT, MANA_VF_DEVICE_ID) },
1475 	{ }
1476 };
1477 
1478 static struct pci_driver mana_driver = {
1479 	.name		= "mana",
1480 	.id_table	= mana_id_table,
1481 	.probe		= mana_gd_probe,
1482 	.remove		= mana_gd_remove,
1483 	.suspend	= mana_gd_suspend,
1484 	.resume		= mana_gd_resume,
1485 	.shutdown	= mana_gd_shutdown,
1486 };
1487 
1488 module_pci_driver(mana_driver);
1489 
1490 MODULE_DEVICE_TABLE(pci, mana_id_table);
1491 
1492 MODULE_LICENSE("Dual BSD/GPL");
1493 MODULE_DESCRIPTION("Microsoft Azure Network Adapter driver");
1494