xref: /linux/drivers/net/ethernet/microsoft/mana/gdma_main.c (revision 1e0731c05c985deb68a97fa44c1adcd3305dda90)
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 <net/mana/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->phys_db_page_base = gc->bar0_pa +
48 				mana_gd_r64(gc, GDMA_REG_DB_PAGE_OFFSET);
49 
50 	gc->shm_base = gc->bar0_va + mana_gd_r64(gc, GDMA_REG_SHM_OFFSET);
51 }
52 
53 static void mana_gd_init_registers(struct pci_dev *pdev)
54 {
55 	struct gdma_context *gc = pci_get_drvdata(pdev);
56 
57 	if (gc->is_pf)
58 		mana_gd_init_pf_regs(pdev);
59 	else
60 		mana_gd_init_vf_regs(pdev);
61 }
62 
63 static int mana_gd_query_max_resources(struct pci_dev *pdev)
64 {
65 	struct gdma_context *gc = pci_get_drvdata(pdev);
66 	struct gdma_query_max_resources_resp resp = {};
67 	struct gdma_general_req req = {};
68 	int err;
69 
70 	mana_gd_init_req_hdr(&req.hdr, GDMA_QUERY_MAX_RESOURCES,
71 			     sizeof(req), sizeof(resp));
72 
73 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
74 	if (err || resp.hdr.status) {
75 		dev_err(gc->dev, "Failed to query resource info: %d, 0x%x\n",
76 			err, resp.hdr.status);
77 		return err ? err : -EPROTO;
78 	}
79 
80 	if (gc->num_msix_usable > resp.max_msix)
81 		gc->num_msix_usable = resp.max_msix;
82 
83 	if (gc->num_msix_usable <= 1)
84 		return -ENOSPC;
85 
86 	gc->max_num_queues = num_online_cpus();
87 	if (gc->max_num_queues > MANA_MAX_NUM_QUEUES)
88 		gc->max_num_queues = MANA_MAX_NUM_QUEUES;
89 
90 	if (gc->max_num_queues > resp.max_eq)
91 		gc->max_num_queues = resp.max_eq;
92 
93 	if (gc->max_num_queues > resp.max_cq)
94 		gc->max_num_queues = resp.max_cq;
95 
96 	if (gc->max_num_queues > resp.max_sq)
97 		gc->max_num_queues = resp.max_sq;
98 
99 	if (gc->max_num_queues > resp.max_rq)
100 		gc->max_num_queues = resp.max_rq;
101 
102 	/* The Hardware Channel (HWC) used 1 MSI-X */
103 	if (gc->max_num_queues > gc->num_msix_usable - 1)
104 		gc->max_num_queues = gc->num_msix_usable - 1;
105 
106 	return 0;
107 }
108 
109 static int mana_gd_query_hwc_timeout(struct pci_dev *pdev, u32 *timeout_val)
110 {
111 	struct gdma_context *gc = pci_get_drvdata(pdev);
112 	struct gdma_query_hwc_timeout_resp resp = {};
113 	struct gdma_query_hwc_timeout_req req = {};
114 	int err;
115 
116 	mana_gd_init_req_hdr(&req.hdr, GDMA_QUERY_HWC_TIMEOUT,
117 			     sizeof(req), sizeof(resp));
118 	req.timeout_ms = *timeout_val;
119 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
120 	if (err || resp.hdr.status)
121 		return err ? err : -EPROTO;
122 
123 	*timeout_val = resp.timeout_ms;
124 
125 	return 0;
126 }
127 
128 static int mana_gd_detect_devices(struct pci_dev *pdev)
129 {
130 	struct gdma_context *gc = pci_get_drvdata(pdev);
131 	struct gdma_list_devices_resp resp = {};
132 	struct gdma_general_req req = {};
133 	struct gdma_dev_id dev;
134 	u32 i, max_num_devs;
135 	u16 dev_type;
136 	int err;
137 
138 	mana_gd_init_req_hdr(&req.hdr, GDMA_LIST_DEVICES, sizeof(req),
139 			     sizeof(resp));
140 
141 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
142 	if (err || resp.hdr.status) {
143 		dev_err(gc->dev, "Failed to detect devices: %d, 0x%x\n", err,
144 			resp.hdr.status);
145 		return err ? err : -EPROTO;
146 	}
147 
148 	max_num_devs = min_t(u32, MAX_NUM_GDMA_DEVICES, resp.num_of_devs);
149 
150 	for (i = 0; i < max_num_devs; i++) {
151 		dev = resp.devs[i];
152 		dev_type = dev.type;
153 
154 		/* HWC is already detected in mana_hwc_create_channel(). */
155 		if (dev_type == GDMA_DEVICE_HWC)
156 			continue;
157 
158 		if (dev_type == GDMA_DEVICE_MANA) {
159 			gc->mana.gdma_context = gc;
160 			gc->mana.dev_id = dev;
161 		}
162 	}
163 
164 	return gc->mana.dev_id.type == 0 ? -ENODEV : 0;
165 }
166 
167 int mana_gd_send_request(struct gdma_context *gc, u32 req_len, const void *req,
168 			 u32 resp_len, void *resp)
169 {
170 	struct hw_channel_context *hwc = gc->hwc.driver_data;
171 
172 	return mana_hwc_send_request(hwc, req_len, req, resp_len, resp);
173 }
174 EXPORT_SYMBOL_NS(mana_gd_send_request, NET_MANA);
175 
176 int mana_gd_alloc_memory(struct gdma_context *gc, unsigned int length,
177 			 struct gdma_mem_info *gmi)
178 {
179 	dma_addr_t dma_handle;
180 	void *buf;
181 
182 	if (length < PAGE_SIZE || !is_power_of_2(length))
183 		return -EINVAL;
184 
185 	gmi->dev = gc->dev;
186 	buf = dma_alloc_coherent(gmi->dev, length, &dma_handle, GFP_KERNEL);
187 	if (!buf)
188 		return -ENOMEM;
189 
190 	gmi->dma_handle = dma_handle;
191 	gmi->virt_addr = buf;
192 	gmi->length = length;
193 
194 	return 0;
195 }
196 
197 void mana_gd_free_memory(struct gdma_mem_info *gmi)
198 {
199 	dma_free_coherent(gmi->dev, gmi->length, gmi->virt_addr,
200 			  gmi->dma_handle);
201 }
202 
203 static int mana_gd_create_hw_eq(struct gdma_context *gc,
204 				struct gdma_queue *queue)
205 {
206 	struct gdma_create_queue_resp resp = {};
207 	struct gdma_create_queue_req req = {};
208 	int err;
209 
210 	if (queue->type != GDMA_EQ)
211 		return -EINVAL;
212 
213 	mana_gd_init_req_hdr(&req.hdr, GDMA_CREATE_QUEUE,
214 			     sizeof(req), sizeof(resp));
215 
216 	req.hdr.dev_id = queue->gdma_dev->dev_id;
217 	req.type = queue->type;
218 	req.pdid = queue->gdma_dev->pdid;
219 	req.doolbell_id = queue->gdma_dev->doorbell;
220 	req.gdma_region = queue->mem_info.dma_region_handle;
221 	req.queue_size = queue->queue_size;
222 	req.log2_throttle_limit = queue->eq.log2_throttle_limit;
223 	req.eq_pci_msix_index = queue->eq.msix_index;
224 
225 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
226 	if (err || resp.hdr.status) {
227 		dev_err(gc->dev, "Failed to create queue: %d, 0x%x\n", err,
228 			resp.hdr.status);
229 		return err ? err : -EPROTO;
230 	}
231 
232 	queue->id = resp.queue_index;
233 	queue->eq.disable_needed = true;
234 	queue->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
235 	return 0;
236 }
237 
238 static int mana_gd_disable_queue(struct gdma_queue *queue)
239 {
240 	struct gdma_context *gc = queue->gdma_dev->gdma_context;
241 	struct gdma_disable_queue_req req = {};
242 	struct gdma_general_resp resp = {};
243 	int err;
244 
245 	WARN_ON(queue->type != GDMA_EQ);
246 
247 	mana_gd_init_req_hdr(&req.hdr, GDMA_DISABLE_QUEUE,
248 			     sizeof(req), sizeof(resp));
249 
250 	req.hdr.dev_id = queue->gdma_dev->dev_id;
251 	req.type = queue->type;
252 	req.queue_index =  queue->id;
253 	req.alloc_res_id_on_creation = 1;
254 
255 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
256 	if (err || resp.hdr.status) {
257 		dev_err(gc->dev, "Failed to disable queue: %d, 0x%x\n", err,
258 			resp.hdr.status);
259 		return err ? err : -EPROTO;
260 	}
261 
262 	return 0;
263 }
264 
265 #define DOORBELL_OFFSET_SQ	0x0
266 #define DOORBELL_OFFSET_RQ	0x400
267 #define DOORBELL_OFFSET_CQ	0x800
268 #define DOORBELL_OFFSET_EQ	0xFF8
269 
270 static void mana_gd_ring_doorbell(struct gdma_context *gc, u32 db_index,
271 				  enum gdma_queue_type q_type, u32 qid,
272 				  u32 tail_ptr, u8 num_req)
273 {
274 	void __iomem *addr = gc->db_page_base + gc->db_page_size * db_index;
275 	union gdma_doorbell_entry e = {};
276 
277 	switch (q_type) {
278 	case GDMA_EQ:
279 		e.eq.id = qid;
280 		e.eq.tail_ptr = tail_ptr;
281 		e.eq.arm = num_req;
282 
283 		addr += DOORBELL_OFFSET_EQ;
284 		break;
285 
286 	case GDMA_CQ:
287 		e.cq.id = qid;
288 		e.cq.tail_ptr = tail_ptr;
289 		e.cq.arm = num_req;
290 
291 		addr += DOORBELL_OFFSET_CQ;
292 		break;
293 
294 	case GDMA_RQ:
295 		e.rq.id = qid;
296 		e.rq.tail_ptr = tail_ptr;
297 		e.rq.wqe_cnt = num_req;
298 
299 		addr += DOORBELL_OFFSET_RQ;
300 		break;
301 
302 	case GDMA_SQ:
303 		e.sq.id = qid;
304 		e.sq.tail_ptr = tail_ptr;
305 
306 		addr += DOORBELL_OFFSET_SQ;
307 		break;
308 
309 	default:
310 		WARN_ON(1);
311 		return;
312 	}
313 
314 	/* Ensure all writes are done before ring doorbell */
315 	wmb();
316 
317 	writeq(e.as_uint64, addr);
318 }
319 
320 void mana_gd_wq_ring_doorbell(struct gdma_context *gc, struct gdma_queue *queue)
321 {
322 	/* Hardware Spec specifies that software client should set 0 for
323 	 * wqe_cnt for Receive Queues. This value is not used in Send Queues.
324 	 */
325 	mana_gd_ring_doorbell(gc, queue->gdma_dev->doorbell, queue->type,
326 			      queue->id, queue->head * GDMA_WQE_BU_SIZE, 0);
327 }
328 
329 void mana_gd_ring_cq(struct gdma_queue *cq, u8 arm_bit)
330 {
331 	struct gdma_context *gc = cq->gdma_dev->gdma_context;
332 
333 	u32 num_cqe = cq->queue_size / GDMA_CQE_SIZE;
334 
335 	u32 head = cq->head % (num_cqe << GDMA_CQE_OWNER_BITS);
336 
337 	mana_gd_ring_doorbell(gc, cq->gdma_dev->doorbell, cq->type, cq->id,
338 			      head, arm_bit);
339 }
340 
341 static void mana_gd_process_eqe(struct gdma_queue *eq)
342 {
343 	u32 head = eq->head % (eq->queue_size / GDMA_EQE_SIZE);
344 	struct gdma_context *gc = eq->gdma_dev->gdma_context;
345 	struct gdma_eqe *eq_eqe_ptr = eq->queue_mem_ptr;
346 	union gdma_eqe_info eqe_info;
347 	enum gdma_eqe_type type;
348 	struct gdma_event event;
349 	struct gdma_queue *cq;
350 	struct gdma_eqe *eqe;
351 	u32 cq_id;
352 
353 	eqe = &eq_eqe_ptr[head];
354 	eqe_info.as_uint32 = eqe->eqe_info;
355 	type = eqe_info.type;
356 
357 	switch (type) {
358 	case GDMA_EQE_COMPLETION:
359 		cq_id = eqe->details[0] & 0xFFFFFF;
360 		if (WARN_ON_ONCE(cq_id >= gc->max_num_cqs))
361 			break;
362 
363 		cq = gc->cq_table[cq_id];
364 		if (WARN_ON_ONCE(!cq || cq->type != GDMA_CQ || cq->id != cq_id))
365 			break;
366 
367 		if (cq->cq.callback)
368 			cq->cq.callback(cq->cq.context, cq);
369 
370 		break;
371 
372 	case GDMA_EQE_TEST_EVENT:
373 		gc->test_event_eq_id = eq->id;
374 		complete(&gc->eq_test_event);
375 		break;
376 
377 	case GDMA_EQE_HWC_INIT_EQ_ID_DB:
378 	case GDMA_EQE_HWC_INIT_DATA:
379 	case GDMA_EQE_HWC_INIT_DONE:
380 		if (!eq->eq.callback)
381 			break;
382 
383 		event.type = type;
384 		memcpy(&event.details, &eqe->details, GDMA_EVENT_DATA_SIZE);
385 		eq->eq.callback(eq->eq.context, eq, &event);
386 		break;
387 
388 	default:
389 		break;
390 	}
391 }
392 
393 static void mana_gd_process_eq_events(void *arg)
394 {
395 	u32 owner_bits, new_bits, old_bits;
396 	union gdma_eqe_info eqe_info;
397 	struct gdma_eqe *eq_eqe_ptr;
398 	struct gdma_queue *eq = arg;
399 	struct gdma_context *gc;
400 	struct gdma_eqe *eqe;
401 	u32 head, num_eqe;
402 	int i;
403 
404 	gc = eq->gdma_dev->gdma_context;
405 
406 	num_eqe = eq->queue_size / GDMA_EQE_SIZE;
407 	eq_eqe_ptr = eq->queue_mem_ptr;
408 
409 	/* Process up to 5 EQEs at a time, and update the HW head. */
410 	for (i = 0; i < 5; i++) {
411 		eqe = &eq_eqe_ptr[eq->head % num_eqe];
412 		eqe_info.as_uint32 = eqe->eqe_info;
413 		owner_bits = eqe_info.owner_bits;
414 
415 		old_bits = (eq->head / num_eqe - 1) & GDMA_EQE_OWNER_MASK;
416 		/* No more entries */
417 		if (owner_bits == old_bits)
418 			break;
419 
420 		new_bits = (eq->head / num_eqe) & GDMA_EQE_OWNER_MASK;
421 		if (owner_bits != new_bits) {
422 			dev_err(gc->dev, "EQ %d: overflow detected\n", eq->id);
423 			break;
424 		}
425 
426 		/* Per GDMA spec, rmb is necessary after checking owner_bits, before
427 		 * reading eqe.
428 		 */
429 		rmb();
430 
431 		mana_gd_process_eqe(eq);
432 
433 		eq->head++;
434 	}
435 
436 	head = eq->head % (num_eqe << GDMA_EQE_OWNER_BITS);
437 
438 	mana_gd_ring_doorbell(gc, eq->gdma_dev->doorbell, eq->type, eq->id,
439 			      head, SET_ARM_BIT);
440 }
441 
442 static int mana_gd_register_irq(struct gdma_queue *queue,
443 				const struct gdma_queue_spec *spec)
444 {
445 	struct gdma_dev *gd = queue->gdma_dev;
446 	struct gdma_irq_context *gic;
447 	struct gdma_context *gc;
448 	struct gdma_resource *r;
449 	unsigned int msi_index;
450 	unsigned long flags;
451 	struct device *dev;
452 	int err = 0;
453 
454 	gc = gd->gdma_context;
455 	r = &gc->msix_resource;
456 	dev = gc->dev;
457 
458 	spin_lock_irqsave(&r->lock, flags);
459 
460 	msi_index = find_first_zero_bit(r->map, r->size);
461 	if (msi_index >= r->size || msi_index >= gc->num_msix_usable) {
462 		err = -ENOSPC;
463 	} else {
464 		bitmap_set(r->map, msi_index, 1);
465 		queue->eq.msix_index = msi_index;
466 	}
467 
468 	spin_unlock_irqrestore(&r->lock, flags);
469 
470 	if (err) {
471 		dev_err(dev, "Register IRQ err:%d, msi:%u rsize:%u, nMSI:%u",
472 			err, msi_index, r->size, gc->num_msix_usable);
473 
474 		return err;
475 	}
476 
477 	gic = &gc->irq_contexts[msi_index];
478 
479 	WARN_ON(gic->handler || gic->arg);
480 
481 	gic->arg = queue;
482 
483 	gic->handler = mana_gd_process_eq_events;
484 
485 	return 0;
486 }
487 
488 static void mana_gd_deregiser_irq(struct gdma_queue *queue)
489 {
490 	struct gdma_dev *gd = queue->gdma_dev;
491 	struct gdma_irq_context *gic;
492 	struct gdma_context *gc;
493 	struct gdma_resource *r;
494 	unsigned int msix_index;
495 	unsigned long flags;
496 
497 	gc = gd->gdma_context;
498 	r = &gc->msix_resource;
499 
500 	/* At most num_online_cpus() + 1 interrupts are used. */
501 	msix_index = queue->eq.msix_index;
502 	if (WARN_ON(msix_index >= gc->num_msix_usable))
503 		return;
504 
505 	gic = &gc->irq_contexts[msix_index];
506 	gic->handler = NULL;
507 	gic->arg = NULL;
508 
509 	spin_lock_irqsave(&r->lock, flags);
510 	bitmap_clear(r->map, msix_index, 1);
511 	spin_unlock_irqrestore(&r->lock, flags);
512 
513 	queue->eq.msix_index = INVALID_PCI_MSIX_INDEX;
514 }
515 
516 int mana_gd_test_eq(struct gdma_context *gc, struct gdma_queue *eq)
517 {
518 	struct gdma_generate_test_event_req req = {};
519 	struct gdma_general_resp resp = {};
520 	struct device *dev = gc->dev;
521 	int err;
522 
523 	mutex_lock(&gc->eq_test_event_mutex);
524 
525 	init_completion(&gc->eq_test_event);
526 	gc->test_event_eq_id = INVALID_QUEUE_ID;
527 
528 	mana_gd_init_req_hdr(&req.hdr, GDMA_GENERATE_TEST_EQE,
529 			     sizeof(req), sizeof(resp));
530 
531 	req.hdr.dev_id = eq->gdma_dev->dev_id;
532 	req.queue_index = eq->id;
533 
534 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
535 	if (err) {
536 		dev_err(dev, "test_eq failed: %d\n", err);
537 		goto out;
538 	}
539 
540 	err = -EPROTO;
541 
542 	if (resp.hdr.status) {
543 		dev_err(dev, "test_eq failed: 0x%x\n", resp.hdr.status);
544 		goto out;
545 	}
546 
547 	if (!wait_for_completion_timeout(&gc->eq_test_event, 30 * HZ)) {
548 		dev_err(dev, "test_eq timed out on queue %d\n", eq->id);
549 		goto out;
550 	}
551 
552 	if (eq->id != gc->test_event_eq_id) {
553 		dev_err(dev, "test_eq got an event on wrong queue %d (%d)\n",
554 			gc->test_event_eq_id, eq->id);
555 		goto out;
556 	}
557 
558 	err = 0;
559 out:
560 	mutex_unlock(&gc->eq_test_event_mutex);
561 	return err;
562 }
563 
564 static void mana_gd_destroy_eq(struct gdma_context *gc, bool flush_evenets,
565 			       struct gdma_queue *queue)
566 {
567 	int err;
568 
569 	if (flush_evenets) {
570 		err = mana_gd_test_eq(gc, queue);
571 		if (err)
572 			dev_warn(gc->dev, "Failed to flush EQ: %d\n", err);
573 	}
574 
575 	mana_gd_deregiser_irq(queue);
576 
577 	if (queue->eq.disable_needed)
578 		mana_gd_disable_queue(queue);
579 }
580 
581 static int mana_gd_create_eq(struct gdma_dev *gd,
582 			     const struct gdma_queue_spec *spec,
583 			     bool create_hwq, struct gdma_queue *queue)
584 {
585 	struct gdma_context *gc = gd->gdma_context;
586 	struct device *dev = gc->dev;
587 	u32 log2_num_entries;
588 	int err;
589 
590 	queue->eq.msix_index = INVALID_PCI_MSIX_INDEX;
591 
592 	log2_num_entries = ilog2(queue->queue_size / GDMA_EQE_SIZE);
593 
594 	if (spec->eq.log2_throttle_limit > log2_num_entries) {
595 		dev_err(dev, "EQ throttling limit (%lu) > maximum EQE (%u)\n",
596 			spec->eq.log2_throttle_limit, log2_num_entries);
597 		return -EINVAL;
598 	}
599 
600 	err = mana_gd_register_irq(queue, spec);
601 	if (err) {
602 		dev_err(dev, "Failed to register irq: %d\n", err);
603 		return err;
604 	}
605 
606 	queue->eq.callback = spec->eq.callback;
607 	queue->eq.context = spec->eq.context;
608 	queue->head |= INITIALIZED_OWNER_BIT(log2_num_entries);
609 	queue->eq.log2_throttle_limit = spec->eq.log2_throttle_limit ?: 1;
610 
611 	if (create_hwq) {
612 		err = mana_gd_create_hw_eq(gc, queue);
613 		if (err)
614 			goto out;
615 
616 		err = mana_gd_test_eq(gc, queue);
617 		if (err)
618 			goto out;
619 	}
620 
621 	return 0;
622 out:
623 	dev_err(dev, "Failed to create EQ: %d\n", err);
624 	mana_gd_destroy_eq(gc, false, queue);
625 	return err;
626 }
627 
628 static void mana_gd_create_cq(const struct gdma_queue_spec *spec,
629 			      struct gdma_queue *queue)
630 {
631 	u32 log2_num_entries = ilog2(spec->queue_size / GDMA_CQE_SIZE);
632 
633 	queue->head |= INITIALIZED_OWNER_BIT(log2_num_entries);
634 	queue->cq.parent = spec->cq.parent_eq;
635 	queue->cq.context = spec->cq.context;
636 	queue->cq.callback = spec->cq.callback;
637 }
638 
639 static void mana_gd_destroy_cq(struct gdma_context *gc,
640 			       struct gdma_queue *queue)
641 {
642 	u32 id = queue->id;
643 
644 	if (id >= gc->max_num_cqs)
645 		return;
646 
647 	if (!gc->cq_table[id])
648 		return;
649 
650 	gc->cq_table[id] = NULL;
651 }
652 
653 int mana_gd_create_hwc_queue(struct gdma_dev *gd,
654 			     const struct gdma_queue_spec *spec,
655 			     struct gdma_queue **queue_ptr)
656 {
657 	struct gdma_context *gc = gd->gdma_context;
658 	struct gdma_mem_info *gmi;
659 	struct gdma_queue *queue;
660 	int err;
661 
662 	queue = kzalloc(sizeof(*queue), GFP_KERNEL);
663 	if (!queue)
664 		return -ENOMEM;
665 
666 	gmi = &queue->mem_info;
667 	err = mana_gd_alloc_memory(gc, spec->queue_size, gmi);
668 	if (err)
669 		goto free_q;
670 
671 	queue->head = 0;
672 	queue->tail = 0;
673 	queue->queue_mem_ptr = gmi->virt_addr;
674 	queue->queue_size = spec->queue_size;
675 	queue->monitor_avl_buf = spec->monitor_avl_buf;
676 	queue->type = spec->type;
677 	queue->gdma_dev = gd;
678 
679 	if (spec->type == GDMA_EQ)
680 		err = mana_gd_create_eq(gd, spec, false, queue);
681 	else if (spec->type == GDMA_CQ)
682 		mana_gd_create_cq(spec, queue);
683 
684 	if (err)
685 		goto out;
686 
687 	*queue_ptr = queue;
688 	return 0;
689 out:
690 	mana_gd_free_memory(gmi);
691 free_q:
692 	kfree(queue);
693 	return err;
694 }
695 
696 int mana_gd_destroy_dma_region(struct gdma_context *gc, u64 dma_region_handle)
697 {
698 	struct gdma_destroy_dma_region_req req = {};
699 	struct gdma_general_resp resp = {};
700 	int err;
701 
702 	if (dma_region_handle == GDMA_INVALID_DMA_REGION)
703 		return 0;
704 
705 	mana_gd_init_req_hdr(&req.hdr, GDMA_DESTROY_DMA_REGION, sizeof(req),
706 			     sizeof(resp));
707 	req.dma_region_handle = dma_region_handle;
708 
709 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
710 	if (err || resp.hdr.status) {
711 		dev_err(gc->dev, "Failed to destroy DMA region: %d, 0x%x\n",
712 			err, resp.hdr.status);
713 		return -EPROTO;
714 	}
715 
716 	return 0;
717 }
718 EXPORT_SYMBOL_NS(mana_gd_destroy_dma_region, NET_MANA);
719 
720 static int mana_gd_create_dma_region(struct gdma_dev *gd,
721 				     struct gdma_mem_info *gmi)
722 {
723 	unsigned int num_page = gmi->length / PAGE_SIZE;
724 	struct gdma_create_dma_region_req *req = NULL;
725 	struct gdma_create_dma_region_resp resp = {};
726 	struct gdma_context *gc = gd->gdma_context;
727 	struct hw_channel_context *hwc;
728 	u32 length = gmi->length;
729 	size_t req_msg_size;
730 	int err;
731 	int i;
732 
733 	if (length < PAGE_SIZE || !is_power_of_2(length))
734 		return -EINVAL;
735 
736 	if (offset_in_page(gmi->virt_addr) != 0)
737 		return -EINVAL;
738 
739 	hwc = gc->hwc.driver_data;
740 	req_msg_size = struct_size(req, page_addr_list, num_page);
741 	if (req_msg_size > hwc->max_req_msg_size)
742 		return -EINVAL;
743 
744 	req = kzalloc(req_msg_size, GFP_KERNEL);
745 	if (!req)
746 		return -ENOMEM;
747 
748 	mana_gd_init_req_hdr(&req->hdr, GDMA_CREATE_DMA_REGION,
749 			     req_msg_size, sizeof(resp));
750 	req->length = length;
751 	req->offset_in_page = 0;
752 	req->gdma_page_type = GDMA_PAGE_TYPE_4K;
753 	req->page_count = num_page;
754 	req->page_addr_list_len = num_page;
755 
756 	for (i = 0; i < num_page; i++)
757 		req->page_addr_list[i] = gmi->dma_handle +  i * PAGE_SIZE;
758 
759 	err = mana_gd_send_request(gc, req_msg_size, req, sizeof(resp), &resp);
760 	if (err)
761 		goto out;
762 
763 	if (resp.hdr.status ||
764 	    resp.dma_region_handle == GDMA_INVALID_DMA_REGION) {
765 		dev_err(gc->dev, "Failed to create DMA region: 0x%x\n",
766 			resp.hdr.status);
767 		err = -EPROTO;
768 		goto out;
769 	}
770 
771 	gmi->dma_region_handle = resp.dma_region_handle;
772 out:
773 	kfree(req);
774 	return err;
775 }
776 
777 int mana_gd_create_mana_eq(struct gdma_dev *gd,
778 			   const struct gdma_queue_spec *spec,
779 			   struct gdma_queue **queue_ptr)
780 {
781 	struct gdma_context *gc = gd->gdma_context;
782 	struct gdma_mem_info *gmi;
783 	struct gdma_queue *queue;
784 	int err;
785 
786 	if (spec->type != GDMA_EQ)
787 		return -EINVAL;
788 
789 	queue = kzalloc(sizeof(*queue), GFP_KERNEL);
790 	if (!queue)
791 		return -ENOMEM;
792 
793 	gmi = &queue->mem_info;
794 	err = mana_gd_alloc_memory(gc, spec->queue_size, gmi);
795 	if (err)
796 		goto free_q;
797 
798 	err = mana_gd_create_dma_region(gd, gmi);
799 	if (err)
800 		goto out;
801 
802 	queue->head = 0;
803 	queue->tail = 0;
804 	queue->queue_mem_ptr = gmi->virt_addr;
805 	queue->queue_size = spec->queue_size;
806 	queue->monitor_avl_buf = spec->monitor_avl_buf;
807 	queue->type = spec->type;
808 	queue->gdma_dev = gd;
809 
810 	err = mana_gd_create_eq(gd, spec, true, queue);
811 	if (err)
812 		goto out;
813 
814 	*queue_ptr = queue;
815 	return 0;
816 out:
817 	mana_gd_free_memory(gmi);
818 free_q:
819 	kfree(queue);
820 	return err;
821 }
822 
823 int mana_gd_create_mana_wq_cq(struct gdma_dev *gd,
824 			      const struct gdma_queue_spec *spec,
825 			      struct gdma_queue **queue_ptr)
826 {
827 	struct gdma_context *gc = gd->gdma_context;
828 	struct gdma_mem_info *gmi;
829 	struct gdma_queue *queue;
830 	int err;
831 
832 	if (spec->type != GDMA_CQ && spec->type != GDMA_SQ &&
833 	    spec->type != GDMA_RQ)
834 		return -EINVAL;
835 
836 	queue = kzalloc(sizeof(*queue), GFP_KERNEL);
837 	if (!queue)
838 		return -ENOMEM;
839 
840 	gmi = &queue->mem_info;
841 	err = mana_gd_alloc_memory(gc, spec->queue_size, gmi);
842 	if (err)
843 		goto free_q;
844 
845 	err = mana_gd_create_dma_region(gd, gmi);
846 	if (err)
847 		goto out;
848 
849 	queue->head = 0;
850 	queue->tail = 0;
851 	queue->queue_mem_ptr = gmi->virt_addr;
852 	queue->queue_size = spec->queue_size;
853 	queue->monitor_avl_buf = spec->monitor_avl_buf;
854 	queue->type = spec->type;
855 	queue->gdma_dev = gd;
856 
857 	if (spec->type == GDMA_CQ)
858 		mana_gd_create_cq(spec, queue);
859 
860 	*queue_ptr = queue;
861 	return 0;
862 out:
863 	mana_gd_free_memory(gmi);
864 free_q:
865 	kfree(queue);
866 	return err;
867 }
868 
869 void mana_gd_destroy_queue(struct gdma_context *gc, struct gdma_queue *queue)
870 {
871 	struct gdma_mem_info *gmi = &queue->mem_info;
872 
873 	switch (queue->type) {
874 	case GDMA_EQ:
875 		mana_gd_destroy_eq(gc, queue->eq.disable_needed, queue);
876 		break;
877 
878 	case GDMA_CQ:
879 		mana_gd_destroy_cq(gc, queue);
880 		break;
881 
882 	case GDMA_RQ:
883 		break;
884 
885 	case GDMA_SQ:
886 		break;
887 
888 	default:
889 		dev_err(gc->dev, "Can't destroy unknown queue: type=%d\n",
890 			queue->type);
891 		return;
892 	}
893 
894 	mana_gd_destroy_dma_region(gc, gmi->dma_region_handle);
895 	mana_gd_free_memory(gmi);
896 	kfree(queue);
897 }
898 
899 int mana_gd_verify_vf_version(struct pci_dev *pdev)
900 {
901 	struct gdma_context *gc = pci_get_drvdata(pdev);
902 	struct gdma_verify_ver_resp resp = {};
903 	struct gdma_verify_ver_req req = {};
904 	struct hw_channel_context *hwc;
905 	int err;
906 
907 	hwc = gc->hwc.driver_data;
908 	mana_gd_init_req_hdr(&req.hdr, GDMA_VERIFY_VF_DRIVER_VERSION,
909 			     sizeof(req), sizeof(resp));
910 
911 	req.protocol_ver_min = GDMA_PROTOCOL_FIRST;
912 	req.protocol_ver_max = GDMA_PROTOCOL_LAST;
913 
914 	req.gd_drv_cap_flags1 = GDMA_DRV_CAP_FLAGS1;
915 	req.gd_drv_cap_flags2 = GDMA_DRV_CAP_FLAGS2;
916 	req.gd_drv_cap_flags3 = GDMA_DRV_CAP_FLAGS3;
917 	req.gd_drv_cap_flags4 = GDMA_DRV_CAP_FLAGS4;
918 
919 	req.drv_ver = 0;	/* Unused*/
920 	req.os_type = 0x10;	/* Linux */
921 	req.os_ver_major = LINUX_VERSION_MAJOR;
922 	req.os_ver_minor = LINUX_VERSION_PATCHLEVEL;
923 	req.os_ver_build = LINUX_VERSION_SUBLEVEL;
924 	strscpy(req.os_ver_str1, utsname()->sysname, sizeof(req.os_ver_str1));
925 	strscpy(req.os_ver_str2, utsname()->release, sizeof(req.os_ver_str2));
926 	strscpy(req.os_ver_str3, utsname()->version, sizeof(req.os_ver_str3));
927 
928 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
929 	if (err || resp.hdr.status) {
930 		dev_err(gc->dev, "VfVerifyVersionOutput: %d, status=0x%x\n",
931 			err, resp.hdr.status);
932 		return err ? err : -EPROTO;
933 	}
934 	if (resp.pf_cap_flags1 & GDMA_DRV_CAP_FLAG_1_HWC_TIMEOUT_RECONFIG) {
935 		err = mana_gd_query_hwc_timeout(pdev, &hwc->hwc_timeout);
936 		if (err) {
937 			dev_err(gc->dev, "Failed to set the hwc timeout %d\n", err);
938 			return err;
939 		}
940 		dev_dbg(gc->dev, "set the hwc timeout to %u\n", hwc->hwc_timeout);
941 	}
942 	return 0;
943 }
944 
945 int mana_gd_register_device(struct gdma_dev *gd)
946 {
947 	struct gdma_context *gc = gd->gdma_context;
948 	struct gdma_register_device_resp resp = {};
949 	struct gdma_general_req req = {};
950 	int err;
951 
952 	gd->pdid = INVALID_PDID;
953 	gd->doorbell = INVALID_DOORBELL;
954 	gd->gpa_mkey = INVALID_MEM_KEY;
955 
956 	mana_gd_init_req_hdr(&req.hdr, GDMA_REGISTER_DEVICE, sizeof(req),
957 			     sizeof(resp));
958 
959 	req.hdr.dev_id = gd->dev_id;
960 
961 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
962 	if (err || resp.hdr.status) {
963 		dev_err(gc->dev, "gdma_register_device_resp failed: %d, 0x%x\n",
964 			err, resp.hdr.status);
965 		return err ? err : -EPROTO;
966 	}
967 
968 	gd->pdid = resp.pdid;
969 	gd->gpa_mkey = resp.gpa_mkey;
970 	gd->doorbell = resp.db_id;
971 
972 	return 0;
973 }
974 
975 int mana_gd_deregister_device(struct gdma_dev *gd)
976 {
977 	struct gdma_context *gc = gd->gdma_context;
978 	struct gdma_general_resp resp = {};
979 	struct gdma_general_req req = {};
980 	int err;
981 
982 	if (gd->pdid == INVALID_PDID)
983 		return -EINVAL;
984 
985 	mana_gd_init_req_hdr(&req.hdr, GDMA_DEREGISTER_DEVICE, sizeof(req),
986 			     sizeof(resp));
987 
988 	req.hdr.dev_id = gd->dev_id;
989 
990 	err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
991 	if (err || resp.hdr.status) {
992 		dev_err(gc->dev, "Failed to deregister device: %d, 0x%x\n",
993 			err, resp.hdr.status);
994 		if (!err)
995 			err = -EPROTO;
996 	}
997 
998 	gd->pdid = INVALID_PDID;
999 	gd->doorbell = INVALID_DOORBELL;
1000 	gd->gpa_mkey = INVALID_MEM_KEY;
1001 
1002 	return err;
1003 }
1004 
1005 u32 mana_gd_wq_avail_space(struct gdma_queue *wq)
1006 {
1007 	u32 used_space = (wq->head - wq->tail) * GDMA_WQE_BU_SIZE;
1008 	u32 wq_size = wq->queue_size;
1009 
1010 	WARN_ON_ONCE(used_space > wq_size);
1011 
1012 	return wq_size - used_space;
1013 }
1014 
1015 u8 *mana_gd_get_wqe_ptr(const struct gdma_queue *wq, u32 wqe_offset)
1016 {
1017 	u32 offset = (wqe_offset * GDMA_WQE_BU_SIZE) & (wq->queue_size - 1);
1018 
1019 	WARN_ON_ONCE((offset + GDMA_WQE_BU_SIZE) > wq->queue_size);
1020 
1021 	return wq->queue_mem_ptr + offset;
1022 }
1023 
1024 static u32 mana_gd_write_client_oob(const struct gdma_wqe_request *wqe_req,
1025 				    enum gdma_queue_type q_type,
1026 				    u32 client_oob_size, u32 sgl_data_size,
1027 				    u8 *wqe_ptr)
1028 {
1029 	bool oob_in_sgl = !!(wqe_req->flags & GDMA_WR_OOB_IN_SGL);
1030 	bool pad_data = !!(wqe_req->flags & GDMA_WR_PAD_BY_SGE0);
1031 	struct gdma_wqe *header = (struct gdma_wqe *)wqe_ptr;
1032 	u8 *ptr;
1033 
1034 	memset(header, 0, sizeof(struct gdma_wqe));
1035 	header->num_sge = wqe_req->num_sge;
1036 	header->inline_oob_size_div4 = client_oob_size / sizeof(u32);
1037 
1038 	if (oob_in_sgl) {
1039 		WARN_ON_ONCE(!pad_data || wqe_req->num_sge < 2);
1040 
1041 		header->client_oob_in_sgl = 1;
1042 
1043 		if (pad_data)
1044 			header->last_vbytes = wqe_req->sgl[0].size;
1045 	}
1046 
1047 	if (q_type == GDMA_SQ)
1048 		header->client_data_unit = wqe_req->client_data_unit;
1049 
1050 	/* The size of gdma_wqe + client_oob_size must be less than or equal
1051 	 * to one Basic Unit (i.e. 32 bytes), so the pointer can't go beyond
1052 	 * the queue memory buffer boundary.
1053 	 */
1054 	ptr = wqe_ptr + sizeof(header);
1055 
1056 	if (wqe_req->inline_oob_data && wqe_req->inline_oob_size > 0) {
1057 		memcpy(ptr, wqe_req->inline_oob_data, wqe_req->inline_oob_size);
1058 
1059 		if (client_oob_size > wqe_req->inline_oob_size)
1060 			memset(ptr + wqe_req->inline_oob_size, 0,
1061 			       client_oob_size - wqe_req->inline_oob_size);
1062 	}
1063 
1064 	return sizeof(header) + client_oob_size;
1065 }
1066 
1067 static void mana_gd_write_sgl(struct gdma_queue *wq, u8 *wqe_ptr,
1068 			      const struct gdma_wqe_request *wqe_req)
1069 {
1070 	u32 sgl_size = sizeof(struct gdma_sge) * wqe_req->num_sge;
1071 	const u8 *address = (u8 *)wqe_req->sgl;
1072 	u8 *base_ptr, *end_ptr;
1073 	u32 size_to_end;
1074 
1075 	base_ptr = wq->queue_mem_ptr;
1076 	end_ptr = base_ptr + wq->queue_size;
1077 	size_to_end = (u32)(end_ptr - wqe_ptr);
1078 
1079 	if (size_to_end < sgl_size) {
1080 		memcpy(wqe_ptr, address, size_to_end);
1081 
1082 		wqe_ptr = base_ptr;
1083 		address += size_to_end;
1084 		sgl_size -= size_to_end;
1085 	}
1086 
1087 	memcpy(wqe_ptr, address, sgl_size);
1088 }
1089 
1090 int mana_gd_post_work_request(struct gdma_queue *wq,
1091 			      const struct gdma_wqe_request *wqe_req,
1092 			      struct gdma_posted_wqe_info *wqe_info)
1093 {
1094 	u32 client_oob_size = wqe_req->inline_oob_size;
1095 	struct gdma_context *gc;
1096 	u32 sgl_data_size;
1097 	u32 max_wqe_size;
1098 	u32 wqe_size;
1099 	u8 *wqe_ptr;
1100 
1101 	if (wqe_req->num_sge == 0)
1102 		return -EINVAL;
1103 
1104 	if (wq->type == GDMA_RQ) {
1105 		if (client_oob_size != 0)
1106 			return -EINVAL;
1107 
1108 		client_oob_size = INLINE_OOB_SMALL_SIZE;
1109 
1110 		max_wqe_size = GDMA_MAX_RQE_SIZE;
1111 	} else {
1112 		if (client_oob_size != INLINE_OOB_SMALL_SIZE &&
1113 		    client_oob_size != INLINE_OOB_LARGE_SIZE)
1114 			return -EINVAL;
1115 
1116 		max_wqe_size = GDMA_MAX_SQE_SIZE;
1117 	}
1118 
1119 	sgl_data_size = sizeof(struct gdma_sge) * wqe_req->num_sge;
1120 	wqe_size = ALIGN(sizeof(struct gdma_wqe) + client_oob_size +
1121 			 sgl_data_size, GDMA_WQE_BU_SIZE);
1122 	if (wqe_size > max_wqe_size)
1123 		return -EINVAL;
1124 
1125 	if (wq->monitor_avl_buf && wqe_size > mana_gd_wq_avail_space(wq)) {
1126 		gc = wq->gdma_dev->gdma_context;
1127 		dev_err(gc->dev, "unsuccessful flow control!\n");
1128 		return -ENOSPC;
1129 	}
1130 
1131 	if (wqe_info)
1132 		wqe_info->wqe_size_in_bu = wqe_size / GDMA_WQE_BU_SIZE;
1133 
1134 	wqe_ptr = mana_gd_get_wqe_ptr(wq, wq->head);
1135 	wqe_ptr += mana_gd_write_client_oob(wqe_req, wq->type, client_oob_size,
1136 					    sgl_data_size, wqe_ptr);
1137 	if (wqe_ptr >= (u8 *)wq->queue_mem_ptr + wq->queue_size)
1138 		wqe_ptr -= wq->queue_size;
1139 
1140 	mana_gd_write_sgl(wq, wqe_ptr, wqe_req);
1141 
1142 	wq->head += wqe_size / GDMA_WQE_BU_SIZE;
1143 
1144 	return 0;
1145 }
1146 
1147 int mana_gd_post_and_ring(struct gdma_queue *queue,
1148 			  const struct gdma_wqe_request *wqe_req,
1149 			  struct gdma_posted_wqe_info *wqe_info)
1150 {
1151 	struct gdma_context *gc = queue->gdma_dev->gdma_context;
1152 	int err;
1153 
1154 	err = mana_gd_post_work_request(queue, wqe_req, wqe_info);
1155 	if (err)
1156 		return err;
1157 
1158 	mana_gd_wq_ring_doorbell(gc, queue);
1159 
1160 	return 0;
1161 }
1162 
1163 static int mana_gd_read_cqe(struct gdma_queue *cq, struct gdma_comp *comp)
1164 {
1165 	unsigned int num_cqe = cq->queue_size / sizeof(struct gdma_cqe);
1166 	struct gdma_cqe *cq_cqe = cq->queue_mem_ptr;
1167 	u32 owner_bits, new_bits, old_bits;
1168 	struct gdma_cqe *cqe;
1169 
1170 	cqe = &cq_cqe[cq->head % num_cqe];
1171 	owner_bits = cqe->cqe_info.owner_bits;
1172 
1173 	old_bits = (cq->head / num_cqe - 1) & GDMA_CQE_OWNER_MASK;
1174 	/* Return 0 if no more entries. */
1175 	if (owner_bits == old_bits)
1176 		return 0;
1177 
1178 	new_bits = (cq->head / num_cqe) & GDMA_CQE_OWNER_MASK;
1179 	/* Return -1 if overflow detected. */
1180 	if (WARN_ON_ONCE(owner_bits != new_bits))
1181 		return -1;
1182 
1183 	/* Per GDMA spec, rmb is necessary after checking owner_bits, before
1184 	 * reading completion info
1185 	 */
1186 	rmb();
1187 
1188 	comp->wq_num = cqe->cqe_info.wq_num;
1189 	comp->is_sq = cqe->cqe_info.is_sq;
1190 	memcpy(comp->cqe_data, cqe->cqe_data, GDMA_COMP_DATA_SIZE);
1191 
1192 	return 1;
1193 }
1194 
1195 int mana_gd_poll_cq(struct gdma_queue *cq, struct gdma_comp *comp, int num_cqe)
1196 {
1197 	int cqe_idx;
1198 	int ret;
1199 
1200 	for (cqe_idx = 0; cqe_idx < num_cqe; cqe_idx++) {
1201 		ret = mana_gd_read_cqe(cq, &comp[cqe_idx]);
1202 
1203 		if (ret < 0) {
1204 			cq->head -= cqe_idx;
1205 			return ret;
1206 		}
1207 
1208 		if (ret == 0)
1209 			break;
1210 
1211 		cq->head++;
1212 	}
1213 
1214 	return cqe_idx;
1215 }
1216 
1217 static irqreturn_t mana_gd_intr(int irq, void *arg)
1218 {
1219 	struct gdma_irq_context *gic = arg;
1220 
1221 	if (gic->handler)
1222 		gic->handler(gic->arg);
1223 
1224 	return IRQ_HANDLED;
1225 }
1226 
1227 int mana_gd_alloc_res_map(u32 res_avail, struct gdma_resource *r)
1228 {
1229 	r->map = bitmap_zalloc(res_avail, GFP_KERNEL);
1230 	if (!r->map)
1231 		return -ENOMEM;
1232 
1233 	r->size = res_avail;
1234 	spin_lock_init(&r->lock);
1235 
1236 	return 0;
1237 }
1238 
1239 void mana_gd_free_res_map(struct gdma_resource *r)
1240 {
1241 	bitmap_free(r->map);
1242 	r->map = NULL;
1243 	r->size = 0;
1244 }
1245 
1246 static int mana_gd_setup_irqs(struct pci_dev *pdev)
1247 {
1248 	unsigned int max_queues_per_port = num_online_cpus();
1249 	struct gdma_context *gc = pci_get_drvdata(pdev);
1250 	struct gdma_irq_context *gic;
1251 	unsigned int max_irqs, cpu;
1252 	int nvec, irq;
1253 	int err, i = 0, j;
1254 
1255 	if (max_queues_per_port > MANA_MAX_NUM_QUEUES)
1256 		max_queues_per_port = MANA_MAX_NUM_QUEUES;
1257 
1258 	/* Need 1 interrupt for the Hardware communication Channel (HWC) */
1259 	max_irqs = max_queues_per_port + 1;
1260 
1261 	nvec = pci_alloc_irq_vectors(pdev, 2, max_irqs, PCI_IRQ_MSIX);
1262 	if (nvec < 0)
1263 		return nvec;
1264 
1265 	gc->irq_contexts = kcalloc(nvec, sizeof(struct gdma_irq_context),
1266 				   GFP_KERNEL);
1267 	if (!gc->irq_contexts) {
1268 		err = -ENOMEM;
1269 		goto free_irq_vector;
1270 	}
1271 
1272 	for (i = 0; i < nvec; i++) {
1273 		gic = &gc->irq_contexts[i];
1274 		gic->handler = NULL;
1275 		gic->arg = NULL;
1276 
1277 		if (!i)
1278 			snprintf(gic->name, MANA_IRQ_NAME_SZ, "mana_hwc@pci:%s",
1279 				 pci_name(pdev));
1280 		else
1281 			snprintf(gic->name, MANA_IRQ_NAME_SZ, "mana_q%d@pci:%s",
1282 				 i - 1, pci_name(pdev));
1283 
1284 		irq = pci_irq_vector(pdev, i);
1285 		if (irq < 0) {
1286 			err = irq;
1287 			goto free_irq;
1288 		}
1289 
1290 		err = request_irq(irq, mana_gd_intr, 0, gic->name, gic);
1291 		if (err)
1292 			goto free_irq;
1293 
1294 		cpu = cpumask_local_spread(i, gc->numa_node);
1295 		irq_set_affinity_and_hint(irq, cpumask_of(cpu));
1296 	}
1297 
1298 	err = mana_gd_alloc_res_map(nvec, &gc->msix_resource);
1299 	if (err)
1300 		goto free_irq;
1301 
1302 	gc->max_num_msix = nvec;
1303 	gc->num_msix_usable = nvec;
1304 
1305 	return 0;
1306 
1307 free_irq:
1308 	for (j = i - 1; j >= 0; j--) {
1309 		irq = pci_irq_vector(pdev, j);
1310 		gic = &gc->irq_contexts[j];
1311 
1312 		irq_update_affinity_hint(irq, NULL);
1313 		free_irq(irq, gic);
1314 	}
1315 
1316 	kfree(gc->irq_contexts);
1317 	gc->irq_contexts = NULL;
1318 free_irq_vector:
1319 	pci_free_irq_vectors(pdev);
1320 	return err;
1321 }
1322 
1323 static void mana_gd_remove_irqs(struct pci_dev *pdev)
1324 {
1325 	struct gdma_context *gc = pci_get_drvdata(pdev);
1326 	struct gdma_irq_context *gic;
1327 	int irq, i;
1328 
1329 	if (gc->max_num_msix < 1)
1330 		return;
1331 
1332 	mana_gd_free_res_map(&gc->msix_resource);
1333 
1334 	for (i = 0; i < gc->max_num_msix; i++) {
1335 		irq = pci_irq_vector(pdev, i);
1336 		if (irq < 0)
1337 			continue;
1338 
1339 		gic = &gc->irq_contexts[i];
1340 
1341 		/* Need to clear the hint before free_irq */
1342 		irq_update_affinity_hint(irq, NULL);
1343 		free_irq(irq, gic);
1344 	}
1345 
1346 	pci_free_irq_vectors(pdev);
1347 
1348 	gc->max_num_msix = 0;
1349 	gc->num_msix_usable = 0;
1350 	kfree(gc->irq_contexts);
1351 	gc->irq_contexts = NULL;
1352 }
1353 
1354 static int mana_gd_setup(struct pci_dev *pdev)
1355 {
1356 	struct gdma_context *gc = pci_get_drvdata(pdev);
1357 	int err;
1358 
1359 	mana_gd_init_registers(pdev);
1360 	mana_smc_init(&gc->shm_channel, gc->dev, gc->shm_base);
1361 
1362 	err = mana_gd_setup_irqs(pdev);
1363 	if (err)
1364 		return err;
1365 
1366 	err = mana_hwc_create_channel(gc);
1367 	if (err)
1368 		goto remove_irq;
1369 
1370 	err = mana_gd_verify_vf_version(pdev);
1371 	if (err)
1372 		goto destroy_hwc;
1373 
1374 	err = mana_gd_query_max_resources(pdev);
1375 	if (err)
1376 		goto destroy_hwc;
1377 
1378 	err = mana_gd_detect_devices(pdev);
1379 	if (err)
1380 		goto destroy_hwc;
1381 
1382 	return 0;
1383 
1384 destroy_hwc:
1385 	mana_hwc_destroy_channel(gc);
1386 remove_irq:
1387 	mana_gd_remove_irqs(pdev);
1388 	return err;
1389 }
1390 
1391 static void mana_gd_cleanup(struct pci_dev *pdev)
1392 {
1393 	struct gdma_context *gc = pci_get_drvdata(pdev);
1394 
1395 	mana_hwc_destroy_channel(gc);
1396 
1397 	mana_gd_remove_irqs(pdev);
1398 }
1399 
1400 static bool mana_is_pf(unsigned short dev_id)
1401 {
1402 	return dev_id == MANA_PF_DEVICE_ID;
1403 }
1404 
1405 static int mana_gd_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
1406 {
1407 	struct gdma_context *gc;
1408 	void __iomem *bar0_va;
1409 	int bar = 0;
1410 	int err;
1411 
1412 	/* Each port has 2 CQs, each CQ has at most 1 EQE at a time */
1413 	BUILD_BUG_ON(2 * MAX_PORTS_IN_MANA_DEV * GDMA_EQE_SIZE > EQ_SIZE);
1414 
1415 	err = pci_enable_device(pdev);
1416 	if (err)
1417 		return -ENXIO;
1418 
1419 	pci_set_master(pdev);
1420 
1421 	err = pci_request_regions(pdev, "mana");
1422 	if (err)
1423 		goto disable_dev;
1424 
1425 	err = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
1426 	if (err)
1427 		goto release_region;
1428 
1429 	err = dma_set_max_seg_size(&pdev->dev, UINT_MAX);
1430 	if (err) {
1431 		dev_err(&pdev->dev, "Failed to set dma device segment size\n");
1432 		goto release_region;
1433 	}
1434 
1435 	err = -ENOMEM;
1436 	gc = vzalloc(sizeof(*gc));
1437 	if (!gc)
1438 		goto release_region;
1439 
1440 	mutex_init(&gc->eq_test_event_mutex);
1441 	pci_set_drvdata(pdev, gc);
1442 	gc->bar0_pa = pci_resource_start(pdev, 0);
1443 
1444 	bar0_va = pci_iomap(pdev, bar, 0);
1445 	if (!bar0_va)
1446 		goto free_gc;
1447 
1448 	gc->numa_node = dev_to_node(&pdev->dev);
1449 	gc->is_pf = mana_is_pf(pdev->device);
1450 	gc->bar0_va = bar0_va;
1451 	gc->dev = &pdev->dev;
1452 
1453 	err = mana_gd_setup(pdev);
1454 	if (err)
1455 		goto unmap_bar;
1456 
1457 	err = mana_probe(&gc->mana, false);
1458 	if (err)
1459 		goto cleanup_gd;
1460 
1461 	return 0;
1462 
1463 cleanup_gd:
1464 	mana_gd_cleanup(pdev);
1465 unmap_bar:
1466 	pci_iounmap(pdev, bar0_va);
1467 free_gc:
1468 	pci_set_drvdata(pdev, NULL);
1469 	vfree(gc);
1470 release_region:
1471 	pci_release_regions(pdev);
1472 disable_dev:
1473 	pci_disable_device(pdev);
1474 	dev_err(&pdev->dev, "gdma probe failed: err = %d\n", err);
1475 	return err;
1476 }
1477 
1478 static void mana_gd_remove(struct pci_dev *pdev)
1479 {
1480 	struct gdma_context *gc = pci_get_drvdata(pdev);
1481 
1482 	mana_remove(&gc->mana, false);
1483 
1484 	mana_gd_cleanup(pdev);
1485 
1486 	pci_iounmap(pdev, gc->bar0_va);
1487 
1488 	vfree(gc);
1489 
1490 	pci_release_regions(pdev);
1491 	pci_disable_device(pdev);
1492 }
1493 
1494 /* The 'state' parameter is not used. */
1495 static int mana_gd_suspend(struct pci_dev *pdev, pm_message_t state)
1496 {
1497 	struct gdma_context *gc = pci_get_drvdata(pdev);
1498 
1499 	mana_remove(&gc->mana, true);
1500 
1501 	mana_gd_cleanup(pdev);
1502 
1503 	return 0;
1504 }
1505 
1506 /* In case the NIC hardware stops working, the suspend and resume callbacks will
1507  * fail -- if this happens, it's safer to just report an error than try to undo
1508  * what has been done.
1509  */
1510 static int mana_gd_resume(struct pci_dev *pdev)
1511 {
1512 	struct gdma_context *gc = pci_get_drvdata(pdev);
1513 	int err;
1514 
1515 	err = mana_gd_setup(pdev);
1516 	if (err)
1517 		return err;
1518 
1519 	err = mana_probe(&gc->mana, true);
1520 	if (err)
1521 		return err;
1522 
1523 	return 0;
1524 }
1525 
1526 /* Quiesce the device for kexec. This is also called upon reboot/shutdown. */
1527 static void mana_gd_shutdown(struct pci_dev *pdev)
1528 {
1529 	struct gdma_context *gc = pci_get_drvdata(pdev);
1530 
1531 	dev_info(&pdev->dev, "Shutdown was called\n");
1532 
1533 	mana_remove(&gc->mana, true);
1534 
1535 	mana_gd_cleanup(pdev);
1536 
1537 	pci_disable_device(pdev);
1538 }
1539 
1540 static const struct pci_device_id mana_id_table[] = {
1541 	{ PCI_DEVICE(PCI_VENDOR_ID_MICROSOFT, MANA_PF_DEVICE_ID) },
1542 	{ PCI_DEVICE(PCI_VENDOR_ID_MICROSOFT, MANA_VF_DEVICE_ID) },
1543 	{ }
1544 };
1545 
1546 static struct pci_driver mana_driver = {
1547 	.name		= "mana",
1548 	.id_table	= mana_id_table,
1549 	.probe		= mana_gd_probe,
1550 	.remove		= mana_gd_remove,
1551 	.suspend	= mana_gd_suspend,
1552 	.resume		= mana_gd_resume,
1553 	.shutdown	= mana_gd_shutdown,
1554 };
1555 
1556 module_pci_driver(mana_driver);
1557 
1558 MODULE_DEVICE_TABLE(pci, mana_id_table);
1559 
1560 MODULE_LICENSE("Dual BSD/GPL");
1561 MODULE_DESCRIPTION("Microsoft Azure Network Adapter driver");
1562