1 // SPDX-License-Identifier: (GPL-2.0 OR MIT)
2 /* Google virtual Ethernet (gve) driver
3  *
4  * Copyright (C) 2015-2021 Google, Inc.
5  */
6 
7 #include <linux/cpumask.h>
8 #include <linux/etherdevice.h>
9 #include <linux/interrupt.h>
10 #include <linux/module.h>
11 #include <linux/pci.h>
12 #include <linux/sched.h>
13 #include <linux/timer.h>
14 #include <linux/workqueue.h>
15 #include <net/sch_generic.h>
16 #include "gve.h"
17 #include "gve_dqo.h"
18 #include "gve_adminq.h"
19 #include "gve_register.h"
20 
21 #define GVE_DEFAULT_RX_COPYBREAK	(256)
22 
23 #define DEFAULT_MSG_LEVEL	(NETIF_MSG_DRV | NETIF_MSG_LINK)
24 #define GVE_VERSION		"1.0.0"
25 #define GVE_VERSION_PREFIX	"GVE-"
26 
27 // Minimum amount of time between queue kicks in msec (10 seconds)
28 #define MIN_TX_TIMEOUT_GAP (1000 * 10)
29 
30 const char gve_version_str[] = GVE_VERSION;
31 static const char gve_version_prefix[] = GVE_VERSION_PREFIX;
32 
33 static netdev_tx_t gve_start_xmit(struct sk_buff *skb, struct net_device *dev)
34 {
35 	struct gve_priv *priv = netdev_priv(dev);
36 
37 	if (gve_is_gqi(priv))
38 		return gve_tx(skb, dev);
39 	else
40 		return gve_tx_dqo(skb, dev);
41 }
42 
43 static void gve_get_stats(struct net_device *dev, struct rtnl_link_stats64 *s)
44 {
45 	struct gve_priv *priv = netdev_priv(dev);
46 	unsigned int start;
47 	u64 packets, bytes;
48 	int ring;
49 
50 	if (priv->rx) {
51 		for (ring = 0; ring < priv->rx_cfg.num_queues; ring++) {
52 			do {
53 				start =
54 				  u64_stats_fetch_begin(&priv->rx[ring].statss);
55 				packets = priv->rx[ring].rpackets;
56 				bytes = priv->rx[ring].rbytes;
57 			} while (u64_stats_fetch_retry(&priv->rx[ring].statss,
58 						       start));
59 			s->rx_packets += packets;
60 			s->rx_bytes += bytes;
61 		}
62 	}
63 	if (priv->tx) {
64 		for (ring = 0; ring < priv->tx_cfg.num_queues; ring++) {
65 			do {
66 				start =
67 				  u64_stats_fetch_begin(&priv->tx[ring].statss);
68 				packets = priv->tx[ring].pkt_done;
69 				bytes = priv->tx[ring].bytes_done;
70 			} while (u64_stats_fetch_retry(&priv->tx[ring].statss,
71 						       start));
72 			s->tx_packets += packets;
73 			s->tx_bytes += bytes;
74 		}
75 	}
76 }
77 
78 static int gve_alloc_counter_array(struct gve_priv *priv)
79 {
80 	priv->counter_array =
81 		dma_alloc_coherent(&priv->pdev->dev,
82 				   priv->num_event_counters *
83 				   sizeof(*priv->counter_array),
84 				   &priv->counter_array_bus, GFP_KERNEL);
85 	if (!priv->counter_array)
86 		return -ENOMEM;
87 
88 	return 0;
89 }
90 
91 static void gve_free_counter_array(struct gve_priv *priv)
92 {
93 	if (!priv->counter_array)
94 		return;
95 
96 	dma_free_coherent(&priv->pdev->dev,
97 			  priv->num_event_counters *
98 			  sizeof(*priv->counter_array),
99 			  priv->counter_array, priv->counter_array_bus);
100 	priv->counter_array = NULL;
101 }
102 
103 /* NIC requests to report stats */
104 static void gve_stats_report_task(struct work_struct *work)
105 {
106 	struct gve_priv *priv = container_of(work, struct gve_priv,
107 					     stats_report_task);
108 	if (gve_get_do_report_stats(priv)) {
109 		gve_handle_report_stats(priv);
110 		gve_clear_do_report_stats(priv);
111 	}
112 }
113 
114 static void gve_stats_report_schedule(struct gve_priv *priv)
115 {
116 	if (!gve_get_probe_in_progress(priv) &&
117 	    !gve_get_reset_in_progress(priv)) {
118 		gve_set_do_report_stats(priv);
119 		queue_work(priv->gve_wq, &priv->stats_report_task);
120 	}
121 }
122 
123 static void gve_stats_report_timer(struct timer_list *t)
124 {
125 	struct gve_priv *priv = from_timer(priv, t, stats_report_timer);
126 
127 	mod_timer(&priv->stats_report_timer,
128 		  round_jiffies(jiffies +
129 		  msecs_to_jiffies(priv->stats_report_timer_period)));
130 	gve_stats_report_schedule(priv);
131 }
132 
133 static int gve_alloc_stats_report(struct gve_priv *priv)
134 {
135 	int tx_stats_num, rx_stats_num;
136 
137 	tx_stats_num = (GVE_TX_STATS_REPORT_NUM + NIC_TX_STATS_REPORT_NUM) *
138 		       priv->tx_cfg.num_queues;
139 	rx_stats_num = (GVE_RX_STATS_REPORT_NUM + NIC_RX_STATS_REPORT_NUM) *
140 		       priv->rx_cfg.num_queues;
141 	priv->stats_report_len = struct_size(priv->stats_report, stats,
142 					     tx_stats_num + rx_stats_num);
143 	priv->stats_report =
144 		dma_alloc_coherent(&priv->pdev->dev, priv->stats_report_len,
145 				   &priv->stats_report_bus, GFP_KERNEL);
146 	if (!priv->stats_report)
147 		return -ENOMEM;
148 	/* Set up timer for the report-stats task */
149 	timer_setup(&priv->stats_report_timer, gve_stats_report_timer, 0);
150 	priv->stats_report_timer_period = GVE_STATS_REPORT_TIMER_PERIOD;
151 	return 0;
152 }
153 
154 static void gve_free_stats_report(struct gve_priv *priv)
155 {
156 	if (!priv->stats_report)
157 		return;
158 
159 	del_timer_sync(&priv->stats_report_timer);
160 	dma_free_coherent(&priv->pdev->dev, priv->stats_report_len,
161 			  priv->stats_report, priv->stats_report_bus);
162 	priv->stats_report = NULL;
163 }
164 
165 static irqreturn_t gve_mgmnt_intr(int irq, void *arg)
166 {
167 	struct gve_priv *priv = arg;
168 
169 	queue_work(priv->gve_wq, &priv->service_task);
170 	return IRQ_HANDLED;
171 }
172 
173 static irqreturn_t gve_intr(int irq, void *arg)
174 {
175 	struct gve_notify_block *block = arg;
176 	struct gve_priv *priv = block->priv;
177 
178 	iowrite32be(GVE_IRQ_MASK, gve_irq_doorbell(priv, block));
179 	napi_schedule_irqoff(&block->napi);
180 	return IRQ_HANDLED;
181 }
182 
183 static irqreturn_t gve_intr_dqo(int irq, void *arg)
184 {
185 	struct gve_notify_block *block = arg;
186 
187 	/* Interrupts are automatically masked */
188 	napi_schedule_irqoff(&block->napi);
189 	return IRQ_HANDLED;
190 }
191 
192 static int gve_napi_poll(struct napi_struct *napi, int budget)
193 {
194 	struct gve_notify_block *block;
195 	__be32 __iomem *irq_doorbell;
196 	bool reschedule = false;
197 	struct gve_priv *priv;
198 	int work_done = 0;
199 
200 	block = container_of(napi, struct gve_notify_block, napi);
201 	priv = block->priv;
202 
203 	if (block->tx)
204 		reschedule |= gve_tx_poll(block, budget);
205 	if (block->rx) {
206 		work_done = gve_rx_poll(block, budget);
207 		reschedule |= work_done == budget;
208 	}
209 
210 	if (reschedule)
211 		return budget;
212 
213        /* Complete processing - don't unmask irq if busy polling is enabled */
214 	if (likely(napi_complete_done(napi, work_done))) {
215 		irq_doorbell = gve_irq_doorbell(priv, block);
216 		iowrite32be(GVE_IRQ_ACK | GVE_IRQ_EVENT, irq_doorbell);
217 
218 		/* Ensure IRQ ACK is visible before we check pending work.
219 		 * If queue had issued updates, it would be truly visible.
220 		 */
221 		mb();
222 
223 		if (block->tx)
224 			reschedule |= gve_tx_clean_pending(priv, block->tx);
225 		if (block->rx)
226 			reschedule |= gve_rx_work_pending(block->rx);
227 
228 		if (reschedule && napi_reschedule(napi))
229 			iowrite32be(GVE_IRQ_MASK, irq_doorbell);
230 	}
231 	return work_done;
232 }
233 
234 static int gve_napi_poll_dqo(struct napi_struct *napi, int budget)
235 {
236 	struct gve_notify_block *block =
237 		container_of(napi, struct gve_notify_block, napi);
238 	struct gve_priv *priv = block->priv;
239 	bool reschedule = false;
240 	int work_done = 0;
241 
242 	/* Clear PCI MSI-X Pending Bit Array (PBA)
243 	 *
244 	 * This bit is set if an interrupt event occurs while the vector is
245 	 * masked. If this bit is set and we reenable the interrupt, it will
246 	 * fire again. Since we're just about to poll the queue state, we don't
247 	 * need it to fire again.
248 	 *
249 	 * Under high softirq load, it's possible that the interrupt condition
250 	 * is triggered twice before we got the chance to process it.
251 	 */
252 	gve_write_irq_doorbell_dqo(priv, block,
253 				   GVE_ITR_NO_UPDATE_DQO | GVE_ITR_CLEAR_PBA_BIT_DQO);
254 
255 	if (block->tx)
256 		reschedule |= gve_tx_poll_dqo(block, /*do_clean=*/true);
257 
258 	if (block->rx) {
259 		work_done = gve_rx_poll_dqo(block, budget);
260 		reschedule |= work_done == budget;
261 	}
262 
263 	if (reschedule)
264 		return budget;
265 
266 	if (likely(napi_complete_done(napi, work_done))) {
267 		/* Enable interrupts again.
268 		 *
269 		 * We don't need to repoll afterwards because HW supports the
270 		 * PCI MSI-X PBA feature.
271 		 *
272 		 * Another interrupt would be triggered if a new event came in
273 		 * since the last one.
274 		 */
275 		gve_write_irq_doorbell_dqo(priv, block,
276 					   GVE_ITR_NO_UPDATE_DQO | GVE_ITR_ENABLE_BIT_DQO);
277 	}
278 
279 	return work_done;
280 }
281 
282 static int gve_alloc_notify_blocks(struct gve_priv *priv)
283 {
284 	int num_vecs_requested = priv->num_ntfy_blks + 1;
285 	char *name = priv->dev->name;
286 	unsigned int active_cpus;
287 	int vecs_enabled;
288 	int i, j;
289 	int err;
290 
291 	priv->msix_vectors = kvcalloc(num_vecs_requested,
292 				      sizeof(*priv->msix_vectors), GFP_KERNEL);
293 	if (!priv->msix_vectors)
294 		return -ENOMEM;
295 	for (i = 0; i < num_vecs_requested; i++)
296 		priv->msix_vectors[i].entry = i;
297 	vecs_enabled = pci_enable_msix_range(priv->pdev, priv->msix_vectors,
298 					     GVE_MIN_MSIX, num_vecs_requested);
299 	if (vecs_enabled < 0) {
300 		dev_err(&priv->pdev->dev, "Could not enable min msix %d/%d\n",
301 			GVE_MIN_MSIX, vecs_enabled);
302 		err = vecs_enabled;
303 		goto abort_with_msix_vectors;
304 	}
305 	if (vecs_enabled != num_vecs_requested) {
306 		int new_num_ntfy_blks = (vecs_enabled - 1) & ~0x1;
307 		int vecs_per_type = new_num_ntfy_blks / 2;
308 		int vecs_left = new_num_ntfy_blks % 2;
309 
310 		priv->num_ntfy_blks = new_num_ntfy_blks;
311 		priv->mgmt_msix_idx = priv->num_ntfy_blks;
312 		priv->tx_cfg.max_queues = min_t(int, priv->tx_cfg.max_queues,
313 						vecs_per_type);
314 		priv->rx_cfg.max_queues = min_t(int, priv->rx_cfg.max_queues,
315 						vecs_per_type + vecs_left);
316 		dev_err(&priv->pdev->dev,
317 			"Could not enable desired msix, only enabled %d, adjusting tx max queues to %d, and rx max queues to %d\n",
318 			vecs_enabled, priv->tx_cfg.max_queues,
319 			priv->rx_cfg.max_queues);
320 		if (priv->tx_cfg.num_queues > priv->tx_cfg.max_queues)
321 			priv->tx_cfg.num_queues = priv->tx_cfg.max_queues;
322 		if (priv->rx_cfg.num_queues > priv->rx_cfg.max_queues)
323 			priv->rx_cfg.num_queues = priv->rx_cfg.max_queues;
324 	}
325 	/* Half the notification blocks go to TX and half to RX */
326 	active_cpus = min_t(int, priv->num_ntfy_blks / 2, num_online_cpus());
327 
328 	/* Setup Management Vector  - the last vector */
329 	snprintf(priv->mgmt_msix_name, sizeof(priv->mgmt_msix_name), "%s-mgmnt",
330 		 name);
331 	err = request_irq(priv->msix_vectors[priv->mgmt_msix_idx].vector,
332 			  gve_mgmnt_intr, 0, priv->mgmt_msix_name, priv);
333 	if (err) {
334 		dev_err(&priv->pdev->dev, "Did not receive management vector.\n");
335 		goto abort_with_msix_enabled;
336 	}
337 	priv->irq_db_indices =
338 		dma_alloc_coherent(&priv->pdev->dev,
339 				   priv->num_ntfy_blks *
340 				   sizeof(*priv->irq_db_indices),
341 				   &priv->irq_db_indices_bus, GFP_KERNEL);
342 	if (!priv->irq_db_indices) {
343 		err = -ENOMEM;
344 		goto abort_with_mgmt_vector;
345 	}
346 
347 	priv->ntfy_blocks = kvzalloc(priv->num_ntfy_blks *
348 				     sizeof(*priv->ntfy_blocks), GFP_KERNEL);
349 	if (!priv->ntfy_blocks) {
350 		err = -ENOMEM;
351 		goto abort_with_irq_db_indices;
352 	}
353 
354 	/* Setup the other blocks - the first n-1 vectors */
355 	for (i = 0; i < priv->num_ntfy_blks; i++) {
356 		struct gve_notify_block *block = &priv->ntfy_blocks[i];
357 		int msix_idx = i;
358 
359 		snprintf(block->name, sizeof(block->name), "%s-ntfy-block.%d",
360 			 name, i);
361 		block->priv = priv;
362 		err = request_irq(priv->msix_vectors[msix_idx].vector,
363 				  gve_is_gqi(priv) ? gve_intr : gve_intr_dqo,
364 				  0, block->name, block);
365 		if (err) {
366 			dev_err(&priv->pdev->dev,
367 				"Failed to receive msix vector %d\n", i);
368 			goto abort_with_some_ntfy_blocks;
369 		}
370 		irq_set_affinity_hint(priv->msix_vectors[msix_idx].vector,
371 				      get_cpu_mask(i % active_cpus));
372 		block->irq_db_index = &priv->irq_db_indices[i].index;
373 	}
374 	return 0;
375 abort_with_some_ntfy_blocks:
376 	for (j = 0; j < i; j++) {
377 		struct gve_notify_block *block = &priv->ntfy_blocks[j];
378 		int msix_idx = j;
379 
380 		irq_set_affinity_hint(priv->msix_vectors[msix_idx].vector,
381 				      NULL);
382 		free_irq(priv->msix_vectors[msix_idx].vector, block);
383 	}
384 	kvfree(priv->ntfy_blocks);
385 	priv->ntfy_blocks = NULL;
386 abort_with_irq_db_indices:
387 	dma_free_coherent(&priv->pdev->dev, priv->num_ntfy_blks *
388 			  sizeof(*priv->irq_db_indices),
389 			  priv->irq_db_indices, priv->irq_db_indices_bus);
390 	priv->irq_db_indices = NULL;
391 abort_with_mgmt_vector:
392 	free_irq(priv->msix_vectors[priv->mgmt_msix_idx].vector, priv);
393 abort_with_msix_enabled:
394 	pci_disable_msix(priv->pdev);
395 abort_with_msix_vectors:
396 	kvfree(priv->msix_vectors);
397 	priv->msix_vectors = NULL;
398 	return err;
399 }
400 
401 static void gve_free_notify_blocks(struct gve_priv *priv)
402 {
403 	int i;
404 
405 	if (!priv->msix_vectors)
406 		return;
407 
408 	/* Free the irqs */
409 	for (i = 0; i < priv->num_ntfy_blks; i++) {
410 		struct gve_notify_block *block = &priv->ntfy_blocks[i];
411 		int msix_idx = i;
412 
413 		irq_set_affinity_hint(priv->msix_vectors[msix_idx].vector,
414 				      NULL);
415 		free_irq(priv->msix_vectors[msix_idx].vector, block);
416 	}
417 	free_irq(priv->msix_vectors[priv->mgmt_msix_idx].vector, priv);
418 	kvfree(priv->ntfy_blocks);
419 	priv->ntfy_blocks = NULL;
420 	dma_free_coherent(&priv->pdev->dev, priv->num_ntfy_blks *
421 			  sizeof(*priv->irq_db_indices),
422 			  priv->irq_db_indices, priv->irq_db_indices_bus);
423 	priv->irq_db_indices = NULL;
424 	pci_disable_msix(priv->pdev);
425 	kvfree(priv->msix_vectors);
426 	priv->msix_vectors = NULL;
427 }
428 
429 static int gve_setup_device_resources(struct gve_priv *priv)
430 {
431 	int err;
432 
433 	err = gve_alloc_counter_array(priv);
434 	if (err)
435 		return err;
436 	err = gve_alloc_notify_blocks(priv);
437 	if (err)
438 		goto abort_with_counter;
439 	err = gve_alloc_stats_report(priv);
440 	if (err)
441 		goto abort_with_ntfy_blocks;
442 	err = gve_adminq_configure_device_resources(priv,
443 						    priv->counter_array_bus,
444 						    priv->num_event_counters,
445 						    priv->irq_db_indices_bus,
446 						    priv->num_ntfy_blks);
447 	if (unlikely(err)) {
448 		dev_err(&priv->pdev->dev,
449 			"could not setup device_resources: err=%d\n", err);
450 		err = -ENXIO;
451 		goto abort_with_stats_report;
452 	}
453 
454 	if (priv->queue_format == GVE_DQO_RDA_FORMAT) {
455 		priv->ptype_lut_dqo = kvzalloc(sizeof(*priv->ptype_lut_dqo),
456 					       GFP_KERNEL);
457 		if (!priv->ptype_lut_dqo) {
458 			err = -ENOMEM;
459 			goto abort_with_stats_report;
460 		}
461 		err = gve_adminq_get_ptype_map_dqo(priv, priv->ptype_lut_dqo);
462 		if (err) {
463 			dev_err(&priv->pdev->dev,
464 				"Failed to get ptype map: err=%d\n", err);
465 			goto abort_with_ptype_lut;
466 		}
467 	}
468 
469 	err = gve_adminq_report_stats(priv, priv->stats_report_len,
470 				      priv->stats_report_bus,
471 				      GVE_STATS_REPORT_TIMER_PERIOD);
472 	if (err)
473 		dev_err(&priv->pdev->dev,
474 			"Failed to report stats: err=%d\n", err);
475 	gve_set_device_resources_ok(priv);
476 	return 0;
477 
478 abort_with_ptype_lut:
479 	kvfree(priv->ptype_lut_dqo);
480 	priv->ptype_lut_dqo = NULL;
481 abort_with_stats_report:
482 	gve_free_stats_report(priv);
483 abort_with_ntfy_blocks:
484 	gve_free_notify_blocks(priv);
485 abort_with_counter:
486 	gve_free_counter_array(priv);
487 
488 	return err;
489 }
490 
491 static void gve_trigger_reset(struct gve_priv *priv);
492 
493 static void gve_teardown_device_resources(struct gve_priv *priv)
494 {
495 	int err;
496 
497 	/* Tell device its resources are being freed */
498 	if (gve_get_device_resources_ok(priv)) {
499 		/* detach the stats report */
500 		err = gve_adminq_report_stats(priv, 0, 0x0, GVE_STATS_REPORT_TIMER_PERIOD);
501 		if (err) {
502 			dev_err(&priv->pdev->dev,
503 				"Failed to detach stats report: err=%d\n", err);
504 			gve_trigger_reset(priv);
505 		}
506 		err = gve_adminq_deconfigure_device_resources(priv);
507 		if (err) {
508 			dev_err(&priv->pdev->dev,
509 				"Could not deconfigure device resources: err=%d\n",
510 				err);
511 			gve_trigger_reset(priv);
512 		}
513 	}
514 
515 	kvfree(priv->ptype_lut_dqo);
516 	priv->ptype_lut_dqo = NULL;
517 
518 	gve_free_counter_array(priv);
519 	gve_free_notify_blocks(priv);
520 	gve_free_stats_report(priv);
521 	gve_clear_device_resources_ok(priv);
522 }
523 
524 static void gve_add_napi(struct gve_priv *priv, int ntfy_idx,
525 			 int (*gve_poll)(struct napi_struct *, int))
526 {
527 	struct gve_notify_block *block = &priv->ntfy_blocks[ntfy_idx];
528 
529 	netif_napi_add(priv->dev, &block->napi, gve_poll,
530 		       NAPI_POLL_WEIGHT);
531 }
532 
533 static void gve_remove_napi(struct gve_priv *priv, int ntfy_idx)
534 {
535 	struct gve_notify_block *block = &priv->ntfy_blocks[ntfy_idx];
536 
537 	netif_napi_del(&block->napi);
538 }
539 
540 static int gve_register_qpls(struct gve_priv *priv)
541 {
542 	int num_qpls = gve_num_tx_qpls(priv) + gve_num_rx_qpls(priv);
543 	int err;
544 	int i;
545 
546 	for (i = 0; i < num_qpls; i++) {
547 		err = gve_adminq_register_page_list(priv, &priv->qpls[i]);
548 		if (err) {
549 			netif_err(priv, drv, priv->dev,
550 				  "failed to register queue page list %d\n",
551 				  priv->qpls[i].id);
552 			/* This failure will trigger a reset - no need to clean
553 			 * up
554 			 */
555 			return err;
556 		}
557 	}
558 	return 0;
559 }
560 
561 static int gve_unregister_qpls(struct gve_priv *priv)
562 {
563 	int num_qpls = gve_num_tx_qpls(priv) + gve_num_rx_qpls(priv);
564 	int err;
565 	int i;
566 
567 	for (i = 0; i < num_qpls; i++) {
568 		err = gve_adminq_unregister_page_list(priv, priv->qpls[i].id);
569 		/* This failure will trigger a reset - no need to clean up */
570 		if (err) {
571 			netif_err(priv, drv, priv->dev,
572 				  "Failed to unregister queue page list %d\n",
573 				  priv->qpls[i].id);
574 			return err;
575 		}
576 	}
577 	return 0;
578 }
579 
580 static int gve_create_rings(struct gve_priv *priv)
581 {
582 	int err;
583 	int i;
584 
585 	err = gve_adminq_create_tx_queues(priv, priv->tx_cfg.num_queues);
586 	if (err) {
587 		netif_err(priv, drv, priv->dev, "failed to create %d tx queues\n",
588 			  priv->tx_cfg.num_queues);
589 		/* This failure will trigger a reset - no need to clean
590 		 * up
591 		 */
592 		return err;
593 	}
594 	netif_dbg(priv, drv, priv->dev, "created %d tx queues\n",
595 		  priv->tx_cfg.num_queues);
596 
597 	err = gve_adminq_create_rx_queues(priv, priv->rx_cfg.num_queues);
598 	if (err) {
599 		netif_err(priv, drv, priv->dev, "failed to create %d rx queues\n",
600 			  priv->rx_cfg.num_queues);
601 		/* This failure will trigger a reset - no need to clean
602 		 * up
603 		 */
604 		return err;
605 	}
606 	netif_dbg(priv, drv, priv->dev, "created %d rx queues\n",
607 		  priv->rx_cfg.num_queues);
608 
609 	if (gve_is_gqi(priv)) {
610 		/* Rx data ring has been prefilled with packet buffers at queue
611 		 * allocation time.
612 		 *
613 		 * Write the doorbell to provide descriptor slots and packet
614 		 * buffers to the NIC.
615 		 */
616 		for (i = 0; i < priv->rx_cfg.num_queues; i++)
617 			gve_rx_write_doorbell(priv, &priv->rx[i]);
618 	} else {
619 		for (i = 0; i < priv->rx_cfg.num_queues; i++) {
620 			/* Post buffers and ring doorbell. */
621 			gve_rx_post_buffers_dqo(&priv->rx[i]);
622 		}
623 	}
624 
625 	return 0;
626 }
627 
628 static void add_napi_init_sync_stats(struct gve_priv *priv,
629 				     int (*napi_poll)(struct napi_struct *napi,
630 						      int budget))
631 {
632 	int i;
633 
634 	/* Add tx napi & init sync stats*/
635 	for (i = 0; i < priv->tx_cfg.num_queues; i++) {
636 		int ntfy_idx = gve_tx_idx_to_ntfy(priv, i);
637 
638 		u64_stats_init(&priv->tx[i].statss);
639 		priv->tx[i].ntfy_id = ntfy_idx;
640 		gve_add_napi(priv, ntfy_idx, napi_poll);
641 	}
642 	/* Add rx napi  & init sync stats*/
643 	for (i = 0; i < priv->rx_cfg.num_queues; i++) {
644 		int ntfy_idx = gve_rx_idx_to_ntfy(priv, i);
645 
646 		u64_stats_init(&priv->rx[i].statss);
647 		priv->rx[i].ntfy_id = ntfy_idx;
648 		gve_add_napi(priv, ntfy_idx, napi_poll);
649 	}
650 }
651 
652 static void gve_tx_free_rings(struct gve_priv *priv)
653 {
654 	if (gve_is_gqi(priv)) {
655 		gve_tx_free_rings_gqi(priv);
656 	} else {
657 		gve_tx_free_rings_dqo(priv);
658 	}
659 }
660 
661 static int gve_alloc_rings(struct gve_priv *priv)
662 {
663 	int err;
664 
665 	/* Setup tx rings */
666 	priv->tx = kvcalloc(priv->tx_cfg.num_queues, sizeof(*priv->tx),
667 			    GFP_KERNEL);
668 	if (!priv->tx)
669 		return -ENOMEM;
670 
671 	if (gve_is_gqi(priv))
672 		err = gve_tx_alloc_rings(priv);
673 	else
674 		err = gve_tx_alloc_rings_dqo(priv);
675 	if (err)
676 		goto free_tx;
677 
678 	/* Setup rx rings */
679 	priv->rx = kvcalloc(priv->rx_cfg.num_queues, sizeof(*priv->rx),
680 			    GFP_KERNEL);
681 	if (!priv->rx) {
682 		err = -ENOMEM;
683 		goto free_tx_queue;
684 	}
685 
686 	if (gve_is_gqi(priv))
687 		err = gve_rx_alloc_rings(priv);
688 	else
689 		err = gve_rx_alloc_rings_dqo(priv);
690 	if (err)
691 		goto free_rx;
692 
693 	if (gve_is_gqi(priv))
694 		add_napi_init_sync_stats(priv, gve_napi_poll);
695 	else
696 		add_napi_init_sync_stats(priv, gve_napi_poll_dqo);
697 
698 	return 0;
699 
700 free_rx:
701 	kvfree(priv->rx);
702 	priv->rx = NULL;
703 free_tx_queue:
704 	gve_tx_free_rings(priv);
705 free_tx:
706 	kvfree(priv->tx);
707 	priv->tx = NULL;
708 	return err;
709 }
710 
711 static int gve_destroy_rings(struct gve_priv *priv)
712 {
713 	int err;
714 
715 	err = gve_adminq_destroy_tx_queues(priv, priv->tx_cfg.num_queues);
716 	if (err) {
717 		netif_err(priv, drv, priv->dev,
718 			  "failed to destroy tx queues\n");
719 		/* This failure will trigger a reset - no need to clean up */
720 		return err;
721 	}
722 	netif_dbg(priv, drv, priv->dev, "destroyed tx queues\n");
723 	err = gve_adminq_destroy_rx_queues(priv, priv->rx_cfg.num_queues);
724 	if (err) {
725 		netif_err(priv, drv, priv->dev,
726 			  "failed to destroy rx queues\n");
727 		/* This failure will trigger a reset - no need to clean up */
728 		return err;
729 	}
730 	netif_dbg(priv, drv, priv->dev, "destroyed rx queues\n");
731 	return 0;
732 }
733 
734 static void gve_rx_free_rings(struct gve_priv *priv)
735 {
736 	if (gve_is_gqi(priv))
737 		gve_rx_free_rings_gqi(priv);
738 	else
739 		gve_rx_free_rings_dqo(priv);
740 }
741 
742 static void gve_free_rings(struct gve_priv *priv)
743 {
744 	int ntfy_idx;
745 	int i;
746 
747 	if (priv->tx) {
748 		for (i = 0; i < priv->tx_cfg.num_queues; i++) {
749 			ntfy_idx = gve_tx_idx_to_ntfy(priv, i);
750 			gve_remove_napi(priv, ntfy_idx);
751 		}
752 		gve_tx_free_rings(priv);
753 		kvfree(priv->tx);
754 		priv->tx = NULL;
755 	}
756 	if (priv->rx) {
757 		for (i = 0; i < priv->rx_cfg.num_queues; i++) {
758 			ntfy_idx = gve_rx_idx_to_ntfy(priv, i);
759 			gve_remove_napi(priv, ntfy_idx);
760 		}
761 		gve_rx_free_rings(priv);
762 		kvfree(priv->rx);
763 		priv->rx = NULL;
764 	}
765 }
766 
767 int gve_alloc_page(struct gve_priv *priv, struct device *dev,
768 		   struct page **page, dma_addr_t *dma,
769 		   enum dma_data_direction dir, gfp_t gfp_flags)
770 {
771 	*page = alloc_page(gfp_flags);
772 	if (!*page) {
773 		priv->page_alloc_fail++;
774 		return -ENOMEM;
775 	}
776 	*dma = dma_map_page(dev, *page, 0, PAGE_SIZE, dir);
777 	if (dma_mapping_error(dev, *dma)) {
778 		priv->dma_mapping_error++;
779 		put_page(*page);
780 		return -ENOMEM;
781 	}
782 	return 0;
783 }
784 
785 static int gve_alloc_queue_page_list(struct gve_priv *priv, u32 id,
786 				     int pages)
787 {
788 	struct gve_queue_page_list *qpl = &priv->qpls[id];
789 	int err;
790 	int i;
791 
792 	if (pages + priv->num_registered_pages > priv->max_registered_pages) {
793 		netif_err(priv, drv, priv->dev,
794 			  "Reached max number of registered pages %llu > %llu\n",
795 			  pages + priv->num_registered_pages,
796 			  priv->max_registered_pages);
797 		return -EINVAL;
798 	}
799 
800 	qpl->id = id;
801 	qpl->num_entries = 0;
802 	qpl->pages = kvcalloc(pages, sizeof(*qpl->pages), GFP_KERNEL);
803 	/* caller handles clean up */
804 	if (!qpl->pages)
805 		return -ENOMEM;
806 	qpl->page_buses = kvcalloc(pages, sizeof(*qpl->page_buses), GFP_KERNEL);
807 	/* caller handles clean up */
808 	if (!qpl->page_buses)
809 		return -ENOMEM;
810 
811 	for (i = 0; i < pages; i++) {
812 		err = gve_alloc_page(priv, &priv->pdev->dev, &qpl->pages[i],
813 				     &qpl->page_buses[i],
814 				     gve_qpl_dma_dir(priv, id), GFP_KERNEL);
815 		/* caller handles clean up */
816 		if (err)
817 			return -ENOMEM;
818 		qpl->num_entries++;
819 	}
820 	priv->num_registered_pages += pages;
821 
822 	return 0;
823 }
824 
825 void gve_free_page(struct device *dev, struct page *page, dma_addr_t dma,
826 		   enum dma_data_direction dir)
827 {
828 	if (!dma_mapping_error(dev, dma))
829 		dma_unmap_page(dev, dma, PAGE_SIZE, dir);
830 	if (page)
831 		put_page(page);
832 }
833 
834 static void gve_free_queue_page_list(struct gve_priv *priv, u32 id)
835 {
836 	struct gve_queue_page_list *qpl = &priv->qpls[id];
837 	int i;
838 
839 	if (!qpl->pages)
840 		return;
841 	if (!qpl->page_buses)
842 		goto free_pages;
843 
844 	for (i = 0; i < qpl->num_entries; i++)
845 		gve_free_page(&priv->pdev->dev, qpl->pages[i],
846 			      qpl->page_buses[i], gve_qpl_dma_dir(priv, id));
847 
848 	kvfree(qpl->page_buses);
849 free_pages:
850 	kvfree(qpl->pages);
851 	priv->num_registered_pages -= qpl->num_entries;
852 }
853 
854 static int gve_alloc_qpls(struct gve_priv *priv)
855 {
856 	int num_qpls = gve_num_tx_qpls(priv) + gve_num_rx_qpls(priv);
857 	int i, j;
858 	int err;
859 
860 	/* Raw addressing means no QPLs */
861 	if (priv->queue_format == GVE_GQI_RDA_FORMAT)
862 		return 0;
863 
864 	priv->qpls = kvcalloc(num_qpls, sizeof(*priv->qpls), GFP_KERNEL);
865 	if (!priv->qpls)
866 		return -ENOMEM;
867 
868 	for (i = 0; i < gve_num_tx_qpls(priv); i++) {
869 		err = gve_alloc_queue_page_list(priv, i,
870 						priv->tx_pages_per_qpl);
871 		if (err)
872 			goto free_qpls;
873 	}
874 	for (; i < num_qpls; i++) {
875 		err = gve_alloc_queue_page_list(priv, i,
876 						priv->rx_data_slot_cnt);
877 		if (err)
878 			goto free_qpls;
879 	}
880 
881 	priv->qpl_cfg.qpl_map_size = BITS_TO_LONGS(num_qpls) *
882 				     sizeof(unsigned long) * BITS_PER_BYTE;
883 	priv->qpl_cfg.qpl_id_map = kvcalloc(BITS_TO_LONGS(num_qpls),
884 					    sizeof(unsigned long), GFP_KERNEL);
885 	if (!priv->qpl_cfg.qpl_id_map) {
886 		err = -ENOMEM;
887 		goto free_qpls;
888 	}
889 
890 	return 0;
891 
892 free_qpls:
893 	for (j = 0; j <= i; j++)
894 		gve_free_queue_page_list(priv, j);
895 	kvfree(priv->qpls);
896 	return err;
897 }
898 
899 static void gve_free_qpls(struct gve_priv *priv)
900 {
901 	int num_qpls = gve_num_tx_qpls(priv) + gve_num_rx_qpls(priv);
902 	int i;
903 
904 	/* Raw addressing means no QPLs */
905 	if (priv->queue_format == GVE_GQI_RDA_FORMAT)
906 		return;
907 
908 	kvfree(priv->qpl_cfg.qpl_id_map);
909 
910 	for (i = 0; i < num_qpls; i++)
911 		gve_free_queue_page_list(priv, i);
912 
913 	kvfree(priv->qpls);
914 }
915 
916 /* Use this to schedule a reset when the device is capable of continuing
917  * to handle other requests in its current state. If it is not, do a reset
918  * in thread instead.
919  */
920 void gve_schedule_reset(struct gve_priv *priv)
921 {
922 	gve_set_do_reset(priv);
923 	queue_work(priv->gve_wq, &priv->service_task);
924 }
925 
926 static void gve_reset_and_teardown(struct gve_priv *priv, bool was_up);
927 static int gve_reset_recovery(struct gve_priv *priv, bool was_up);
928 static void gve_turndown(struct gve_priv *priv);
929 static void gve_turnup(struct gve_priv *priv);
930 
931 static int gve_open(struct net_device *dev)
932 {
933 	struct gve_priv *priv = netdev_priv(dev);
934 	int err;
935 
936 	err = gve_alloc_qpls(priv);
937 	if (err)
938 		return err;
939 
940 	err = gve_alloc_rings(priv);
941 	if (err)
942 		goto free_qpls;
943 
944 	err = netif_set_real_num_tx_queues(dev, priv->tx_cfg.num_queues);
945 	if (err)
946 		goto free_rings;
947 	err = netif_set_real_num_rx_queues(dev, priv->rx_cfg.num_queues);
948 	if (err)
949 		goto free_rings;
950 
951 	err = gve_register_qpls(priv);
952 	if (err)
953 		goto reset;
954 
955 	if (!gve_is_gqi(priv)) {
956 		/* Hard code this for now. This may be tuned in the future for
957 		 * performance.
958 		 */
959 		priv->data_buffer_size_dqo = GVE_RX_BUFFER_SIZE_DQO;
960 	}
961 	err = gve_create_rings(priv);
962 	if (err)
963 		goto reset;
964 
965 	gve_set_device_rings_ok(priv);
966 
967 	if (gve_get_report_stats(priv))
968 		mod_timer(&priv->stats_report_timer,
969 			  round_jiffies(jiffies +
970 				msecs_to_jiffies(priv->stats_report_timer_period)));
971 
972 	gve_turnup(priv);
973 	queue_work(priv->gve_wq, &priv->service_task);
974 	priv->interface_up_cnt++;
975 	return 0;
976 
977 free_rings:
978 	gve_free_rings(priv);
979 free_qpls:
980 	gve_free_qpls(priv);
981 	return err;
982 
983 reset:
984 	/* This must have been called from a reset due to the rtnl lock
985 	 * so just return at this point.
986 	 */
987 	if (gve_get_reset_in_progress(priv))
988 		return err;
989 	/* Otherwise reset before returning */
990 	gve_reset_and_teardown(priv, true);
991 	/* if this fails there is nothing we can do so just ignore the return */
992 	gve_reset_recovery(priv, false);
993 	/* return the original error */
994 	return err;
995 }
996 
997 static int gve_close(struct net_device *dev)
998 {
999 	struct gve_priv *priv = netdev_priv(dev);
1000 	int err;
1001 
1002 	netif_carrier_off(dev);
1003 	if (gve_get_device_rings_ok(priv)) {
1004 		gve_turndown(priv);
1005 		err = gve_destroy_rings(priv);
1006 		if (err)
1007 			goto err;
1008 		err = gve_unregister_qpls(priv);
1009 		if (err)
1010 			goto err;
1011 		gve_clear_device_rings_ok(priv);
1012 	}
1013 	del_timer_sync(&priv->stats_report_timer);
1014 
1015 	gve_free_rings(priv);
1016 	gve_free_qpls(priv);
1017 	priv->interface_down_cnt++;
1018 	return 0;
1019 
1020 err:
1021 	/* This must have been called from a reset due to the rtnl lock
1022 	 * so just return at this point.
1023 	 */
1024 	if (gve_get_reset_in_progress(priv))
1025 		return err;
1026 	/* Otherwise reset before returning */
1027 	gve_reset_and_teardown(priv, true);
1028 	return gve_reset_recovery(priv, false);
1029 }
1030 
1031 int gve_adjust_queues(struct gve_priv *priv,
1032 		      struct gve_queue_config new_rx_config,
1033 		      struct gve_queue_config new_tx_config)
1034 {
1035 	int err;
1036 
1037 	if (netif_carrier_ok(priv->dev)) {
1038 		/* To make this process as simple as possible we teardown the
1039 		 * device, set the new configuration, and then bring the device
1040 		 * up again.
1041 		 */
1042 		err = gve_close(priv->dev);
1043 		/* we have already tried to reset in close,
1044 		 * just fail at this point
1045 		 */
1046 		if (err)
1047 			return err;
1048 		priv->tx_cfg = new_tx_config;
1049 		priv->rx_cfg = new_rx_config;
1050 
1051 		err = gve_open(priv->dev);
1052 		if (err)
1053 			goto err;
1054 
1055 		return 0;
1056 	}
1057 	/* Set the config for the next up. */
1058 	priv->tx_cfg = new_tx_config;
1059 	priv->rx_cfg = new_rx_config;
1060 
1061 	return 0;
1062 err:
1063 	netif_err(priv, drv, priv->dev,
1064 		  "Adjust queues failed! !!! DISABLING ALL QUEUES !!!\n");
1065 	gve_turndown(priv);
1066 	return err;
1067 }
1068 
1069 static void gve_turndown(struct gve_priv *priv)
1070 {
1071 	int idx;
1072 
1073 	if (netif_carrier_ok(priv->dev))
1074 		netif_carrier_off(priv->dev);
1075 
1076 	if (!gve_get_napi_enabled(priv))
1077 		return;
1078 
1079 	/* Disable napi to prevent more work from coming in */
1080 	for (idx = 0; idx < priv->tx_cfg.num_queues; idx++) {
1081 		int ntfy_idx = gve_tx_idx_to_ntfy(priv, idx);
1082 		struct gve_notify_block *block = &priv->ntfy_blocks[ntfy_idx];
1083 
1084 		napi_disable(&block->napi);
1085 	}
1086 	for (idx = 0; idx < priv->rx_cfg.num_queues; idx++) {
1087 		int ntfy_idx = gve_rx_idx_to_ntfy(priv, idx);
1088 		struct gve_notify_block *block = &priv->ntfy_blocks[ntfy_idx];
1089 
1090 		napi_disable(&block->napi);
1091 	}
1092 
1093 	/* Stop tx queues */
1094 	netif_tx_disable(priv->dev);
1095 
1096 	gve_clear_napi_enabled(priv);
1097 	gve_clear_report_stats(priv);
1098 }
1099 
1100 static void gve_turnup(struct gve_priv *priv)
1101 {
1102 	int idx;
1103 
1104 	/* Start the tx queues */
1105 	netif_tx_start_all_queues(priv->dev);
1106 
1107 	/* Enable napi and unmask interrupts for all queues */
1108 	for (idx = 0; idx < priv->tx_cfg.num_queues; idx++) {
1109 		int ntfy_idx = gve_tx_idx_to_ntfy(priv, idx);
1110 		struct gve_notify_block *block = &priv->ntfy_blocks[ntfy_idx];
1111 
1112 		napi_enable(&block->napi);
1113 		if (gve_is_gqi(priv)) {
1114 			iowrite32be(0, gve_irq_doorbell(priv, block));
1115 		} else {
1116 			gve_set_itr_coalesce_usecs_dqo(priv, block,
1117 						       priv->tx_coalesce_usecs);
1118 		}
1119 	}
1120 	for (idx = 0; idx < priv->rx_cfg.num_queues; idx++) {
1121 		int ntfy_idx = gve_rx_idx_to_ntfy(priv, idx);
1122 		struct gve_notify_block *block = &priv->ntfy_blocks[ntfy_idx];
1123 
1124 		napi_enable(&block->napi);
1125 		if (gve_is_gqi(priv)) {
1126 			iowrite32be(0, gve_irq_doorbell(priv, block));
1127 		} else {
1128 			gve_set_itr_coalesce_usecs_dqo(priv, block,
1129 						       priv->rx_coalesce_usecs);
1130 		}
1131 	}
1132 
1133 	gve_set_napi_enabled(priv);
1134 }
1135 
1136 static void gve_tx_timeout(struct net_device *dev, unsigned int txqueue)
1137 {
1138 	struct gve_notify_block *block;
1139 	struct gve_tx_ring *tx = NULL;
1140 	struct gve_priv *priv;
1141 	u32 last_nic_done;
1142 	u32 current_time;
1143 	u32 ntfy_idx;
1144 
1145 	netdev_info(dev, "Timeout on tx queue, %d", txqueue);
1146 	priv = netdev_priv(dev);
1147 	if (txqueue > priv->tx_cfg.num_queues)
1148 		goto reset;
1149 
1150 	ntfy_idx = gve_tx_idx_to_ntfy(priv, txqueue);
1151 	if (ntfy_idx >= priv->num_ntfy_blks)
1152 		goto reset;
1153 
1154 	block = &priv->ntfy_blocks[ntfy_idx];
1155 	tx = block->tx;
1156 
1157 	current_time = jiffies_to_msecs(jiffies);
1158 	if (tx->last_kick_msec + MIN_TX_TIMEOUT_GAP > current_time)
1159 		goto reset;
1160 
1161 	/* Check to see if there are missed completions, which will allow us to
1162 	 * kick the queue.
1163 	 */
1164 	last_nic_done = gve_tx_load_event_counter(priv, tx);
1165 	if (last_nic_done - tx->done) {
1166 		netdev_info(dev, "Kicking queue %d", txqueue);
1167 		iowrite32be(GVE_IRQ_MASK, gve_irq_doorbell(priv, block));
1168 		napi_schedule(&block->napi);
1169 		tx->last_kick_msec = current_time;
1170 		goto out;
1171 	} // Else reset.
1172 
1173 reset:
1174 	gve_schedule_reset(priv);
1175 
1176 out:
1177 	if (tx)
1178 		tx->queue_timeout++;
1179 	priv->tx_timeo_cnt++;
1180 }
1181 
1182 static int gve_set_features(struct net_device *netdev,
1183 			    netdev_features_t features)
1184 {
1185 	const netdev_features_t orig_features = netdev->features;
1186 	struct gve_priv *priv = netdev_priv(netdev);
1187 	int err;
1188 
1189 	if ((netdev->features & NETIF_F_LRO) != (features & NETIF_F_LRO)) {
1190 		netdev->features ^= NETIF_F_LRO;
1191 		if (netif_carrier_ok(netdev)) {
1192 			/* To make this process as simple as possible we
1193 			 * teardown the device, set the new configuration,
1194 			 * and then bring the device up again.
1195 			 */
1196 			err = gve_close(netdev);
1197 			/* We have already tried to reset in close, just fail
1198 			 * at this point.
1199 			 */
1200 			if (err)
1201 				goto err;
1202 
1203 			err = gve_open(netdev);
1204 			if (err)
1205 				goto err;
1206 		}
1207 	}
1208 
1209 	return 0;
1210 err:
1211 	/* Reverts the change on error. */
1212 	netdev->features = orig_features;
1213 	netif_err(priv, drv, netdev,
1214 		  "Set features failed! !!! DISABLING ALL QUEUES !!!\n");
1215 	return err;
1216 }
1217 
1218 static const struct net_device_ops gve_netdev_ops = {
1219 	.ndo_start_xmit		=	gve_start_xmit,
1220 	.ndo_open		=	gve_open,
1221 	.ndo_stop		=	gve_close,
1222 	.ndo_get_stats64	=	gve_get_stats,
1223 	.ndo_tx_timeout         =       gve_tx_timeout,
1224 	.ndo_set_features	=	gve_set_features,
1225 };
1226 
1227 static void gve_handle_status(struct gve_priv *priv, u32 status)
1228 {
1229 	if (GVE_DEVICE_STATUS_RESET_MASK & status) {
1230 		dev_info(&priv->pdev->dev, "Device requested reset.\n");
1231 		gve_set_do_reset(priv);
1232 	}
1233 	if (GVE_DEVICE_STATUS_REPORT_STATS_MASK & status) {
1234 		priv->stats_report_trigger_cnt++;
1235 		gve_set_do_report_stats(priv);
1236 	}
1237 }
1238 
1239 static void gve_handle_reset(struct gve_priv *priv)
1240 {
1241 	/* A service task will be scheduled at the end of probe to catch any
1242 	 * resets that need to happen, and we don't want to reset until
1243 	 * probe is done.
1244 	 */
1245 	if (gve_get_probe_in_progress(priv))
1246 		return;
1247 
1248 	if (gve_get_do_reset(priv)) {
1249 		rtnl_lock();
1250 		gve_reset(priv, false);
1251 		rtnl_unlock();
1252 	}
1253 }
1254 
1255 void gve_handle_report_stats(struct gve_priv *priv)
1256 {
1257 	struct stats *stats = priv->stats_report->stats;
1258 	int idx, stats_idx = 0;
1259 	unsigned int start = 0;
1260 	u64 tx_bytes;
1261 
1262 	if (!gve_get_report_stats(priv))
1263 		return;
1264 
1265 	be64_add_cpu(&priv->stats_report->written_count, 1);
1266 	/* tx stats */
1267 	if (priv->tx) {
1268 		for (idx = 0; idx < priv->tx_cfg.num_queues; idx++) {
1269 			u32 last_completion = 0;
1270 			u32 tx_frames = 0;
1271 
1272 			/* DQO doesn't currently support these metrics. */
1273 			if (gve_is_gqi(priv)) {
1274 				last_completion = priv->tx[idx].done;
1275 				tx_frames = priv->tx[idx].req;
1276 			}
1277 
1278 			do {
1279 				start = u64_stats_fetch_begin(&priv->tx[idx].statss);
1280 				tx_bytes = priv->tx[idx].bytes_done;
1281 			} while (u64_stats_fetch_retry(&priv->tx[idx].statss, start));
1282 			stats[stats_idx++] = (struct stats) {
1283 				.stat_name = cpu_to_be32(TX_WAKE_CNT),
1284 				.value = cpu_to_be64(priv->tx[idx].wake_queue),
1285 				.queue_id = cpu_to_be32(idx),
1286 			};
1287 			stats[stats_idx++] = (struct stats) {
1288 				.stat_name = cpu_to_be32(TX_STOP_CNT),
1289 				.value = cpu_to_be64(priv->tx[idx].stop_queue),
1290 				.queue_id = cpu_to_be32(idx),
1291 			};
1292 			stats[stats_idx++] = (struct stats) {
1293 				.stat_name = cpu_to_be32(TX_FRAMES_SENT),
1294 				.value = cpu_to_be64(tx_frames),
1295 				.queue_id = cpu_to_be32(idx),
1296 			};
1297 			stats[stats_idx++] = (struct stats) {
1298 				.stat_name = cpu_to_be32(TX_BYTES_SENT),
1299 				.value = cpu_to_be64(tx_bytes),
1300 				.queue_id = cpu_to_be32(idx),
1301 			};
1302 			stats[stats_idx++] = (struct stats) {
1303 				.stat_name = cpu_to_be32(TX_LAST_COMPLETION_PROCESSED),
1304 				.value = cpu_to_be64(last_completion),
1305 				.queue_id = cpu_to_be32(idx),
1306 			};
1307 			stats[stats_idx++] = (struct stats) {
1308 				.stat_name = cpu_to_be32(TX_TIMEOUT_CNT),
1309 				.value = cpu_to_be64(priv->tx[idx].queue_timeout),
1310 				.queue_id = cpu_to_be32(idx),
1311 			};
1312 		}
1313 	}
1314 	/* rx stats */
1315 	if (priv->rx) {
1316 		for (idx = 0; idx < priv->rx_cfg.num_queues; idx++) {
1317 			stats[stats_idx++] = (struct stats) {
1318 				.stat_name = cpu_to_be32(RX_NEXT_EXPECTED_SEQUENCE),
1319 				.value = cpu_to_be64(priv->rx[idx].desc.seqno),
1320 				.queue_id = cpu_to_be32(idx),
1321 			};
1322 			stats[stats_idx++] = (struct stats) {
1323 				.stat_name = cpu_to_be32(RX_BUFFERS_POSTED),
1324 				.value = cpu_to_be64(priv->rx[0].fill_cnt),
1325 				.queue_id = cpu_to_be32(idx),
1326 			};
1327 		}
1328 	}
1329 }
1330 
1331 static void gve_handle_link_status(struct gve_priv *priv, bool link_status)
1332 {
1333 	if (!gve_get_napi_enabled(priv))
1334 		return;
1335 
1336 	if (link_status == netif_carrier_ok(priv->dev))
1337 		return;
1338 
1339 	if (link_status) {
1340 		netdev_info(priv->dev, "Device link is up.\n");
1341 		netif_carrier_on(priv->dev);
1342 	} else {
1343 		netdev_info(priv->dev, "Device link is down.\n");
1344 		netif_carrier_off(priv->dev);
1345 	}
1346 }
1347 
1348 /* Handle NIC status register changes, reset requests and report stats */
1349 static void gve_service_task(struct work_struct *work)
1350 {
1351 	struct gve_priv *priv = container_of(work, struct gve_priv,
1352 					     service_task);
1353 	u32 status = ioread32be(&priv->reg_bar0->device_status);
1354 
1355 	gve_handle_status(priv, status);
1356 
1357 	gve_handle_reset(priv);
1358 	gve_handle_link_status(priv, GVE_DEVICE_STATUS_LINK_STATUS_MASK & status);
1359 }
1360 
1361 static int gve_init_priv(struct gve_priv *priv, bool skip_describe_device)
1362 {
1363 	int num_ntfy;
1364 	int err;
1365 
1366 	/* Set up the adminq */
1367 	err = gve_adminq_alloc(&priv->pdev->dev, priv);
1368 	if (err) {
1369 		dev_err(&priv->pdev->dev,
1370 			"Failed to alloc admin queue: err=%d\n", err);
1371 		return err;
1372 	}
1373 
1374 	if (skip_describe_device)
1375 		goto setup_device;
1376 
1377 	priv->queue_format = GVE_QUEUE_FORMAT_UNSPECIFIED;
1378 	/* Get the initial information we need from the device */
1379 	err = gve_adminq_describe_device(priv);
1380 	if (err) {
1381 		dev_err(&priv->pdev->dev,
1382 			"Could not get device information: err=%d\n", err);
1383 		goto err;
1384 	}
1385 	priv->dev->mtu = priv->dev->max_mtu;
1386 	num_ntfy = pci_msix_vec_count(priv->pdev);
1387 	if (num_ntfy <= 0) {
1388 		dev_err(&priv->pdev->dev,
1389 			"could not count MSI-x vectors: err=%d\n", num_ntfy);
1390 		err = num_ntfy;
1391 		goto err;
1392 	} else if (num_ntfy < GVE_MIN_MSIX) {
1393 		dev_err(&priv->pdev->dev, "gve needs at least %d MSI-x vectors, but only has %d\n",
1394 			GVE_MIN_MSIX, num_ntfy);
1395 		err = -EINVAL;
1396 		goto err;
1397 	}
1398 
1399 	priv->num_registered_pages = 0;
1400 	priv->rx_copybreak = GVE_DEFAULT_RX_COPYBREAK;
1401 	/* gvnic has one Notification Block per MSI-x vector, except for the
1402 	 * management vector
1403 	 */
1404 	priv->num_ntfy_blks = (num_ntfy - 1) & ~0x1;
1405 	priv->mgmt_msix_idx = priv->num_ntfy_blks;
1406 
1407 	priv->tx_cfg.max_queues =
1408 		min_t(int, priv->tx_cfg.max_queues, priv->num_ntfy_blks / 2);
1409 	priv->rx_cfg.max_queues =
1410 		min_t(int, priv->rx_cfg.max_queues, priv->num_ntfy_blks / 2);
1411 
1412 	priv->tx_cfg.num_queues = priv->tx_cfg.max_queues;
1413 	priv->rx_cfg.num_queues = priv->rx_cfg.max_queues;
1414 	if (priv->default_num_queues > 0) {
1415 		priv->tx_cfg.num_queues = min_t(int, priv->default_num_queues,
1416 						priv->tx_cfg.num_queues);
1417 		priv->rx_cfg.num_queues = min_t(int, priv->default_num_queues,
1418 						priv->rx_cfg.num_queues);
1419 	}
1420 
1421 	dev_info(&priv->pdev->dev, "TX queues %d, RX queues %d\n",
1422 		 priv->tx_cfg.num_queues, priv->rx_cfg.num_queues);
1423 	dev_info(&priv->pdev->dev, "Max TX queues %d, Max RX queues %d\n",
1424 		 priv->tx_cfg.max_queues, priv->rx_cfg.max_queues);
1425 
1426 	if (!gve_is_gqi(priv)) {
1427 		priv->tx_coalesce_usecs = GVE_TX_IRQ_RATELIMIT_US_DQO;
1428 		priv->rx_coalesce_usecs = GVE_RX_IRQ_RATELIMIT_US_DQO;
1429 	}
1430 
1431 setup_device:
1432 	err = gve_setup_device_resources(priv);
1433 	if (!err)
1434 		return 0;
1435 err:
1436 	gve_adminq_free(&priv->pdev->dev, priv);
1437 	return err;
1438 }
1439 
1440 static void gve_teardown_priv_resources(struct gve_priv *priv)
1441 {
1442 	gve_teardown_device_resources(priv);
1443 	gve_adminq_free(&priv->pdev->dev, priv);
1444 }
1445 
1446 static void gve_trigger_reset(struct gve_priv *priv)
1447 {
1448 	/* Reset the device by releasing the AQ */
1449 	gve_adminq_release(priv);
1450 }
1451 
1452 static void gve_reset_and_teardown(struct gve_priv *priv, bool was_up)
1453 {
1454 	gve_trigger_reset(priv);
1455 	/* With the reset having already happened, close cannot fail */
1456 	if (was_up)
1457 		gve_close(priv->dev);
1458 	gve_teardown_priv_resources(priv);
1459 }
1460 
1461 static int gve_reset_recovery(struct gve_priv *priv, bool was_up)
1462 {
1463 	int err;
1464 
1465 	err = gve_init_priv(priv, true);
1466 	if (err)
1467 		goto err;
1468 	if (was_up) {
1469 		err = gve_open(priv->dev);
1470 		if (err)
1471 			goto err;
1472 	}
1473 	return 0;
1474 err:
1475 	dev_err(&priv->pdev->dev, "Reset failed! !!! DISABLING ALL QUEUES !!!\n");
1476 	gve_turndown(priv);
1477 	return err;
1478 }
1479 
1480 int gve_reset(struct gve_priv *priv, bool attempt_teardown)
1481 {
1482 	bool was_up = netif_carrier_ok(priv->dev);
1483 	int err;
1484 
1485 	dev_info(&priv->pdev->dev, "Performing reset\n");
1486 	gve_clear_do_reset(priv);
1487 	gve_set_reset_in_progress(priv);
1488 	/* If we aren't attempting to teardown normally, just go turndown and
1489 	 * reset right away.
1490 	 */
1491 	if (!attempt_teardown) {
1492 		gve_turndown(priv);
1493 		gve_reset_and_teardown(priv, was_up);
1494 	} else {
1495 		/* Otherwise attempt to close normally */
1496 		if (was_up) {
1497 			err = gve_close(priv->dev);
1498 			/* If that fails reset as we did above */
1499 			if (err)
1500 				gve_reset_and_teardown(priv, was_up);
1501 		}
1502 		/* Clean up any remaining resources */
1503 		gve_teardown_priv_resources(priv);
1504 	}
1505 
1506 	/* Set it all back up */
1507 	err = gve_reset_recovery(priv, was_up);
1508 	gve_clear_reset_in_progress(priv);
1509 	priv->reset_cnt++;
1510 	priv->interface_up_cnt = 0;
1511 	priv->interface_down_cnt = 0;
1512 	priv->stats_report_trigger_cnt = 0;
1513 	return err;
1514 }
1515 
1516 static void gve_write_version(u8 __iomem *driver_version_register)
1517 {
1518 	const char *c = gve_version_prefix;
1519 
1520 	while (*c) {
1521 		writeb(*c, driver_version_register);
1522 		c++;
1523 	}
1524 
1525 	c = gve_version_str;
1526 	while (*c) {
1527 		writeb(*c, driver_version_register);
1528 		c++;
1529 	}
1530 	writeb('\n', driver_version_register);
1531 }
1532 
1533 static int gve_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
1534 {
1535 	int max_tx_queues, max_rx_queues;
1536 	struct net_device *dev;
1537 	__be32 __iomem *db_bar;
1538 	struct gve_registers __iomem *reg_bar;
1539 	struct gve_priv *priv;
1540 	int err;
1541 
1542 	err = pci_enable_device(pdev);
1543 	if (err)
1544 		return err;
1545 
1546 	err = pci_request_regions(pdev, "gvnic-cfg");
1547 	if (err)
1548 		goto abort_with_enabled;
1549 
1550 	pci_set_master(pdev);
1551 
1552 	err = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
1553 	if (err) {
1554 		dev_err(&pdev->dev, "Failed to set dma mask: err=%d\n", err);
1555 		goto abort_with_pci_region;
1556 	}
1557 
1558 	reg_bar = pci_iomap(pdev, GVE_REGISTER_BAR, 0);
1559 	if (!reg_bar) {
1560 		dev_err(&pdev->dev, "Failed to map pci bar!\n");
1561 		err = -ENOMEM;
1562 		goto abort_with_pci_region;
1563 	}
1564 
1565 	db_bar = pci_iomap(pdev, GVE_DOORBELL_BAR, 0);
1566 	if (!db_bar) {
1567 		dev_err(&pdev->dev, "Failed to map doorbell bar!\n");
1568 		err = -ENOMEM;
1569 		goto abort_with_reg_bar;
1570 	}
1571 
1572 	gve_write_version(&reg_bar->driver_version);
1573 	/* Get max queues to alloc etherdev */
1574 	max_tx_queues = ioread32be(&reg_bar->max_tx_queues);
1575 	max_rx_queues = ioread32be(&reg_bar->max_rx_queues);
1576 	/* Alloc and setup the netdev and priv */
1577 	dev = alloc_etherdev_mqs(sizeof(*priv), max_tx_queues, max_rx_queues);
1578 	if (!dev) {
1579 		dev_err(&pdev->dev, "could not allocate netdev\n");
1580 		err = -ENOMEM;
1581 		goto abort_with_db_bar;
1582 	}
1583 	SET_NETDEV_DEV(dev, &pdev->dev);
1584 	pci_set_drvdata(pdev, dev);
1585 	dev->ethtool_ops = &gve_ethtool_ops;
1586 	dev->netdev_ops = &gve_netdev_ops;
1587 
1588 	/* Set default and supported features.
1589 	 *
1590 	 * Features might be set in other locations as well (such as
1591 	 * `gve_adminq_describe_device`).
1592 	 */
1593 	dev->hw_features = NETIF_F_HIGHDMA;
1594 	dev->hw_features |= NETIF_F_SG;
1595 	dev->hw_features |= NETIF_F_HW_CSUM;
1596 	dev->hw_features |= NETIF_F_TSO;
1597 	dev->hw_features |= NETIF_F_TSO6;
1598 	dev->hw_features |= NETIF_F_TSO_ECN;
1599 	dev->hw_features |= NETIF_F_RXCSUM;
1600 	dev->hw_features |= NETIF_F_RXHASH;
1601 	dev->features = dev->hw_features;
1602 	dev->watchdog_timeo = 5 * HZ;
1603 	dev->min_mtu = ETH_MIN_MTU;
1604 	netif_carrier_off(dev);
1605 
1606 	priv = netdev_priv(dev);
1607 	priv->dev = dev;
1608 	priv->pdev = pdev;
1609 	priv->msg_enable = DEFAULT_MSG_LEVEL;
1610 	priv->reg_bar0 = reg_bar;
1611 	priv->db_bar2 = db_bar;
1612 	priv->service_task_flags = 0x0;
1613 	priv->state_flags = 0x0;
1614 	priv->ethtool_flags = 0x0;
1615 
1616 	gve_set_probe_in_progress(priv);
1617 	priv->gve_wq = alloc_ordered_workqueue("gve", 0);
1618 	if (!priv->gve_wq) {
1619 		dev_err(&pdev->dev, "Could not allocate workqueue");
1620 		err = -ENOMEM;
1621 		goto abort_with_netdev;
1622 	}
1623 	INIT_WORK(&priv->service_task, gve_service_task);
1624 	INIT_WORK(&priv->stats_report_task, gve_stats_report_task);
1625 	priv->tx_cfg.max_queues = max_tx_queues;
1626 	priv->rx_cfg.max_queues = max_rx_queues;
1627 
1628 	err = gve_init_priv(priv, false);
1629 	if (err)
1630 		goto abort_with_wq;
1631 
1632 	err = register_netdev(dev);
1633 	if (err)
1634 		goto abort_with_gve_init;
1635 
1636 	dev_info(&pdev->dev, "GVE version %s\n", gve_version_str);
1637 	dev_info(&pdev->dev, "GVE queue format %d\n", (int)priv->queue_format);
1638 	gve_clear_probe_in_progress(priv);
1639 	queue_work(priv->gve_wq, &priv->service_task);
1640 	return 0;
1641 
1642 abort_with_gve_init:
1643 	gve_teardown_priv_resources(priv);
1644 
1645 abort_with_wq:
1646 	destroy_workqueue(priv->gve_wq);
1647 
1648 abort_with_netdev:
1649 	free_netdev(dev);
1650 
1651 abort_with_db_bar:
1652 	pci_iounmap(pdev, db_bar);
1653 
1654 abort_with_reg_bar:
1655 	pci_iounmap(pdev, reg_bar);
1656 
1657 abort_with_pci_region:
1658 	pci_release_regions(pdev);
1659 
1660 abort_with_enabled:
1661 	pci_disable_device(pdev);
1662 	return err;
1663 }
1664 
1665 static void gve_remove(struct pci_dev *pdev)
1666 {
1667 	struct net_device *netdev = pci_get_drvdata(pdev);
1668 	struct gve_priv *priv = netdev_priv(netdev);
1669 	__be32 __iomem *db_bar = priv->db_bar2;
1670 	void __iomem *reg_bar = priv->reg_bar0;
1671 
1672 	unregister_netdev(netdev);
1673 	gve_teardown_priv_resources(priv);
1674 	destroy_workqueue(priv->gve_wq);
1675 	free_netdev(netdev);
1676 	pci_iounmap(pdev, db_bar);
1677 	pci_iounmap(pdev, reg_bar);
1678 	pci_release_regions(pdev);
1679 	pci_disable_device(pdev);
1680 }
1681 
1682 static void gve_shutdown(struct pci_dev *pdev)
1683 {
1684 	struct net_device *netdev = pci_get_drvdata(pdev);
1685 	struct gve_priv *priv = netdev_priv(netdev);
1686 	bool was_up = netif_carrier_ok(priv->dev);
1687 
1688 	rtnl_lock();
1689 	if (was_up && gve_close(priv->dev)) {
1690 		/* If the dev was up, attempt to close, if close fails, reset */
1691 		gve_reset_and_teardown(priv, was_up);
1692 	} else {
1693 		/* If the dev wasn't up or close worked, finish tearing down */
1694 		gve_teardown_priv_resources(priv);
1695 	}
1696 	rtnl_unlock();
1697 }
1698 
1699 #ifdef CONFIG_PM
1700 static int gve_suspend(struct pci_dev *pdev, pm_message_t state)
1701 {
1702 	struct net_device *netdev = pci_get_drvdata(pdev);
1703 	struct gve_priv *priv = netdev_priv(netdev);
1704 	bool was_up = netif_carrier_ok(priv->dev);
1705 
1706 	priv->suspend_cnt++;
1707 	rtnl_lock();
1708 	if (was_up && gve_close(priv->dev)) {
1709 		/* If the dev was up, attempt to close, if close fails, reset */
1710 		gve_reset_and_teardown(priv, was_up);
1711 	} else {
1712 		/* If the dev wasn't up or close worked, finish tearing down */
1713 		gve_teardown_priv_resources(priv);
1714 	}
1715 	priv->up_before_suspend = was_up;
1716 	rtnl_unlock();
1717 	return 0;
1718 }
1719 
1720 static int gve_resume(struct pci_dev *pdev)
1721 {
1722 	struct net_device *netdev = pci_get_drvdata(pdev);
1723 	struct gve_priv *priv = netdev_priv(netdev);
1724 	int err;
1725 
1726 	priv->resume_cnt++;
1727 	rtnl_lock();
1728 	err = gve_reset_recovery(priv, priv->up_before_suspend);
1729 	rtnl_unlock();
1730 	return err;
1731 }
1732 #endif /* CONFIG_PM */
1733 
1734 static const struct pci_device_id gve_id_table[] = {
1735 	{ PCI_DEVICE(PCI_VENDOR_ID_GOOGLE, PCI_DEV_ID_GVNIC) },
1736 	{ }
1737 };
1738 
1739 static struct pci_driver gvnic_driver = {
1740 	.name		= "gvnic",
1741 	.id_table	= gve_id_table,
1742 	.probe		= gve_probe,
1743 	.remove		= gve_remove,
1744 	.shutdown	= gve_shutdown,
1745 #ifdef CONFIG_PM
1746 	.suspend        = gve_suspend,
1747 	.resume         = gve_resume,
1748 #endif
1749 };
1750 
1751 module_pci_driver(gvnic_driver);
1752 
1753 MODULE_DEVICE_TABLE(pci, gve_id_table);
1754 MODULE_AUTHOR("Google, Inc.");
1755 MODULE_DESCRIPTION("gVNIC Driver");
1756 MODULE_LICENSE("Dual MIT/GPL");
1757 MODULE_VERSION(GVE_VERSION);
1758