1 // SPDX-License-Identifier: (GPL-2.0 OR MIT)
2 /* Google virtual Ethernet (gve) driver
3  *
4  * Copyright (C) 2015-2019 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_adminq.h"
18 #include "gve_register.h"
19 
20 #define GVE_DEFAULT_RX_COPYBREAK	(256)
21 
22 #define DEFAULT_MSG_LEVEL	(NETIF_MSG_DRV | NETIF_MSG_LINK)
23 #define GVE_VERSION		"1.0.0"
24 #define GVE_VERSION_PREFIX	"GVE-"
25 
26 const char gve_version_str[] = GVE_VERSION;
27 static const char gve_version_prefix[] = GVE_VERSION_PREFIX;
28 
29 static void gve_get_stats(struct net_device *dev, struct rtnl_link_stats64 *s)
30 {
31 	struct gve_priv *priv = netdev_priv(dev);
32 	unsigned int start;
33 	int ring;
34 
35 	if (priv->rx) {
36 		for (ring = 0; ring < priv->rx_cfg.num_queues; ring++) {
37 			do {
38 				start =
39 				  u64_stats_fetch_begin(&priv->rx[ring].statss);
40 				s->rx_packets += priv->rx[ring].rpackets;
41 				s->rx_bytes += priv->rx[ring].rbytes;
42 			} while (u64_stats_fetch_retry(&priv->rx[ring].statss,
43 						       start));
44 		}
45 	}
46 	if (priv->tx) {
47 		for (ring = 0; ring < priv->tx_cfg.num_queues; ring++) {
48 			do {
49 				start =
50 				  u64_stats_fetch_begin(&priv->tx[ring].statss);
51 				s->tx_packets += priv->tx[ring].pkt_done;
52 				s->tx_bytes += priv->tx[ring].bytes_done;
53 			} while (u64_stats_fetch_retry(&priv->rx[ring].statss,
54 						       start));
55 		}
56 	}
57 }
58 
59 static int gve_alloc_counter_array(struct gve_priv *priv)
60 {
61 	priv->counter_array =
62 		dma_alloc_coherent(&priv->pdev->dev,
63 				   priv->num_event_counters *
64 				   sizeof(*priv->counter_array),
65 				   &priv->counter_array_bus, GFP_KERNEL);
66 	if (!priv->counter_array)
67 		return -ENOMEM;
68 
69 	return 0;
70 }
71 
72 static void gve_free_counter_array(struct gve_priv *priv)
73 {
74 	dma_free_coherent(&priv->pdev->dev,
75 			  priv->num_event_counters *
76 			  sizeof(*priv->counter_array),
77 			  priv->counter_array, priv->counter_array_bus);
78 	priv->counter_array = NULL;
79 }
80 
81 static irqreturn_t gve_mgmnt_intr(int irq, void *arg)
82 {
83 	struct gve_priv *priv = arg;
84 
85 	queue_work(priv->gve_wq, &priv->service_task);
86 	return IRQ_HANDLED;
87 }
88 
89 static irqreturn_t gve_intr(int irq, void *arg)
90 {
91 	struct gve_notify_block *block = arg;
92 	struct gve_priv *priv = block->priv;
93 
94 	iowrite32be(GVE_IRQ_MASK, gve_irq_doorbell(priv, block));
95 	napi_schedule_irqoff(&block->napi);
96 	return IRQ_HANDLED;
97 }
98 
99 static int gve_napi_poll(struct napi_struct *napi, int budget)
100 {
101 	struct gve_notify_block *block;
102 	__be32 __iomem *irq_doorbell;
103 	bool reschedule = false;
104 	struct gve_priv *priv;
105 
106 	block = container_of(napi, struct gve_notify_block, napi);
107 	priv = block->priv;
108 
109 	if (block->tx)
110 		reschedule |= gve_tx_poll(block, budget);
111 	if (block->rx)
112 		reschedule |= gve_rx_poll(block, budget);
113 
114 	if (reschedule)
115 		return budget;
116 
117 	napi_complete(napi);
118 	irq_doorbell = gve_irq_doorbell(priv, block);
119 	iowrite32be(GVE_IRQ_ACK | GVE_IRQ_EVENT, irq_doorbell);
120 
121 	/* Double check we have no extra work.
122 	 * Ensure unmask synchronizes with checking for work.
123 	 */
124 	dma_rmb();
125 	if (block->tx)
126 		reschedule |= gve_tx_poll(block, -1);
127 	if (block->rx)
128 		reschedule |= gve_rx_poll(block, -1);
129 	if (reschedule && napi_reschedule(napi))
130 		iowrite32be(GVE_IRQ_MASK, irq_doorbell);
131 
132 	return 0;
133 }
134 
135 static int gve_alloc_notify_blocks(struct gve_priv *priv)
136 {
137 	int num_vecs_requested = priv->num_ntfy_blks + 1;
138 	char *name = priv->dev->name;
139 	unsigned int active_cpus;
140 	int vecs_enabled;
141 	int i, j;
142 	int err;
143 
144 	priv->msix_vectors = kvzalloc(num_vecs_requested *
145 				      sizeof(*priv->msix_vectors), GFP_KERNEL);
146 	if (!priv->msix_vectors)
147 		return -ENOMEM;
148 	for (i = 0; i < num_vecs_requested; i++)
149 		priv->msix_vectors[i].entry = i;
150 	vecs_enabled = pci_enable_msix_range(priv->pdev, priv->msix_vectors,
151 					     GVE_MIN_MSIX, num_vecs_requested);
152 	if (vecs_enabled < 0) {
153 		dev_err(&priv->pdev->dev, "Could not enable min msix %d/%d\n",
154 			GVE_MIN_MSIX, vecs_enabled);
155 		err = vecs_enabled;
156 		goto abort_with_msix_vectors;
157 	}
158 	if (vecs_enabled != num_vecs_requested) {
159 		int new_num_ntfy_blks = (vecs_enabled - 1) & ~0x1;
160 		int vecs_per_type = new_num_ntfy_blks / 2;
161 		int vecs_left = new_num_ntfy_blks % 2;
162 
163 		priv->num_ntfy_blks = new_num_ntfy_blks;
164 		priv->tx_cfg.max_queues = min_t(int, priv->tx_cfg.max_queues,
165 						vecs_per_type);
166 		priv->rx_cfg.max_queues = min_t(int, priv->rx_cfg.max_queues,
167 						vecs_per_type + vecs_left);
168 		dev_err(&priv->pdev->dev,
169 			"Could not enable desired msix, only enabled %d, adjusting tx max queues to %d, and rx max queues to %d\n",
170 			vecs_enabled, priv->tx_cfg.max_queues,
171 			priv->rx_cfg.max_queues);
172 		if (priv->tx_cfg.num_queues > priv->tx_cfg.max_queues)
173 			priv->tx_cfg.num_queues = priv->tx_cfg.max_queues;
174 		if (priv->rx_cfg.num_queues > priv->rx_cfg.max_queues)
175 			priv->rx_cfg.num_queues = priv->rx_cfg.max_queues;
176 	}
177 	/* Half the notification blocks go to TX and half to RX */
178 	active_cpus = min_t(int, priv->num_ntfy_blks / 2, num_online_cpus());
179 
180 	/* Setup Management Vector  - the last vector */
181 	snprintf(priv->mgmt_msix_name, sizeof(priv->mgmt_msix_name), "%s-mgmnt",
182 		 name);
183 	err = request_irq(priv->msix_vectors[priv->mgmt_msix_idx].vector,
184 			  gve_mgmnt_intr, 0, priv->mgmt_msix_name, priv);
185 	if (err) {
186 		dev_err(&priv->pdev->dev, "Did not receive management vector.\n");
187 		goto abort_with_msix_enabled;
188 	}
189 	priv->ntfy_blocks =
190 		dma_alloc_coherent(&priv->pdev->dev,
191 				   priv->num_ntfy_blks *
192 				   sizeof(*priv->ntfy_blocks),
193 				   &priv->ntfy_block_bus, GFP_KERNEL);
194 	if (!priv->ntfy_blocks) {
195 		err = -ENOMEM;
196 		goto abort_with_mgmt_vector;
197 	}
198 	/* Setup the other blocks - the first n-1 vectors */
199 	for (i = 0; i < priv->num_ntfy_blks; i++) {
200 		struct gve_notify_block *block = &priv->ntfy_blocks[i];
201 		int msix_idx = i;
202 
203 		snprintf(block->name, sizeof(block->name), "%s-ntfy-block.%d",
204 			 name, i);
205 		block->priv = priv;
206 		err = request_irq(priv->msix_vectors[msix_idx].vector,
207 				  gve_intr, 0, block->name, block);
208 		if (err) {
209 			dev_err(&priv->pdev->dev,
210 				"Failed to receive msix vector %d\n", i);
211 			goto abort_with_some_ntfy_blocks;
212 		}
213 		irq_set_affinity_hint(priv->msix_vectors[msix_idx].vector,
214 				      get_cpu_mask(i % active_cpus));
215 	}
216 	return 0;
217 abort_with_some_ntfy_blocks:
218 	for (j = 0; j < i; j++) {
219 		struct gve_notify_block *block = &priv->ntfy_blocks[j];
220 		int msix_idx = j;
221 
222 		irq_set_affinity_hint(priv->msix_vectors[msix_idx].vector,
223 				      NULL);
224 		free_irq(priv->msix_vectors[msix_idx].vector, block);
225 	}
226 	dma_free_coherent(&priv->pdev->dev, priv->num_ntfy_blks *
227 			  sizeof(*priv->ntfy_blocks),
228 			  priv->ntfy_blocks, priv->ntfy_block_bus);
229 	priv->ntfy_blocks = NULL;
230 abort_with_mgmt_vector:
231 	free_irq(priv->msix_vectors[priv->mgmt_msix_idx].vector, priv);
232 abort_with_msix_enabled:
233 	pci_disable_msix(priv->pdev);
234 abort_with_msix_vectors:
235 	kvfree(priv->msix_vectors);
236 	priv->msix_vectors = NULL;
237 	return err;
238 }
239 
240 static void gve_free_notify_blocks(struct gve_priv *priv)
241 {
242 	int i;
243 
244 	/* Free the irqs */
245 	for (i = 0; i < priv->num_ntfy_blks; i++) {
246 		struct gve_notify_block *block = &priv->ntfy_blocks[i];
247 		int msix_idx = i;
248 
249 		irq_set_affinity_hint(priv->msix_vectors[msix_idx].vector,
250 				      NULL);
251 		free_irq(priv->msix_vectors[msix_idx].vector, block);
252 	}
253 	dma_free_coherent(&priv->pdev->dev,
254 			  priv->num_ntfy_blks * sizeof(*priv->ntfy_blocks),
255 			  priv->ntfy_blocks, priv->ntfy_block_bus);
256 	priv->ntfy_blocks = NULL;
257 	free_irq(priv->msix_vectors[priv->mgmt_msix_idx].vector, priv);
258 	pci_disable_msix(priv->pdev);
259 	kvfree(priv->msix_vectors);
260 	priv->msix_vectors = NULL;
261 }
262 
263 static int gve_setup_device_resources(struct gve_priv *priv)
264 {
265 	int err;
266 
267 	err = gve_alloc_counter_array(priv);
268 	if (err)
269 		return err;
270 	err = gve_alloc_notify_blocks(priv);
271 	if (err)
272 		goto abort_with_counter;
273 	err = gve_adminq_configure_device_resources(priv,
274 						    priv->counter_array_bus,
275 						    priv->num_event_counters,
276 						    priv->ntfy_block_bus,
277 						    priv->num_ntfy_blks);
278 	if (unlikely(err)) {
279 		dev_err(&priv->pdev->dev,
280 			"could not setup device_resources: err=%d\n", err);
281 		err = -ENXIO;
282 		goto abort_with_ntfy_blocks;
283 	}
284 	gve_set_device_resources_ok(priv);
285 	return 0;
286 abort_with_ntfy_blocks:
287 	gve_free_notify_blocks(priv);
288 abort_with_counter:
289 	gve_free_counter_array(priv);
290 	return err;
291 }
292 
293 static void gve_trigger_reset(struct gve_priv *priv);
294 
295 static void gve_teardown_device_resources(struct gve_priv *priv)
296 {
297 	int err;
298 
299 	/* Tell device its resources are being freed */
300 	if (gve_get_device_resources_ok(priv)) {
301 		err = gve_adminq_deconfigure_device_resources(priv);
302 		if (err) {
303 			dev_err(&priv->pdev->dev,
304 				"Could not deconfigure device resources: err=%d\n",
305 				err);
306 			gve_trigger_reset(priv);
307 		}
308 	}
309 	gve_free_counter_array(priv);
310 	gve_free_notify_blocks(priv);
311 	gve_clear_device_resources_ok(priv);
312 }
313 
314 static void gve_add_napi(struct gve_priv *priv, int ntfy_idx)
315 {
316 	struct gve_notify_block *block = &priv->ntfy_blocks[ntfy_idx];
317 
318 	netif_napi_add(priv->dev, &block->napi, gve_napi_poll,
319 		       NAPI_POLL_WEIGHT);
320 }
321 
322 static void gve_remove_napi(struct gve_priv *priv, int ntfy_idx)
323 {
324 	struct gve_notify_block *block = &priv->ntfy_blocks[ntfy_idx];
325 
326 	netif_napi_del(&block->napi);
327 }
328 
329 static int gve_register_qpls(struct gve_priv *priv)
330 {
331 	int num_qpls = gve_num_tx_qpls(priv) + gve_num_rx_qpls(priv);
332 	int err;
333 	int i;
334 
335 	for (i = 0; i < num_qpls; i++) {
336 		err = gve_adminq_register_page_list(priv, &priv->qpls[i]);
337 		if (err) {
338 			netif_err(priv, drv, priv->dev,
339 				  "failed to register queue page list %d\n",
340 				  priv->qpls[i].id);
341 			/* This failure will trigger a reset - no need to clean
342 			 * up
343 			 */
344 			return err;
345 		}
346 	}
347 	return 0;
348 }
349 
350 static int gve_unregister_qpls(struct gve_priv *priv)
351 {
352 	int num_qpls = gve_num_tx_qpls(priv) + gve_num_rx_qpls(priv);
353 	int err;
354 	int i;
355 
356 	for (i = 0; i < num_qpls; i++) {
357 		err = gve_adminq_unregister_page_list(priv, priv->qpls[i].id);
358 		/* This failure will trigger a reset - no need to clean up */
359 		if (err) {
360 			netif_err(priv, drv, priv->dev,
361 				  "Failed to unregister queue page list %d\n",
362 				  priv->qpls[i].id);
363 			return err;
364 		}
365 	}
366 	return 0;
367 }
368 
369 static int gve_create_rings(struct gve_priv *priv)
370 {
371 	int err;
372 	int i;
373 
374 	for (i = 0; i < priv->tx_cfg.num_queues; i++) {
375 		err = gve_adminq_create_tx_queue(priv, i);
376 		if (err) {
377 			netif_err(priv, drv, priv->dev, "failed to create tx queue %d\n",
378 				  i);
379 			/* This failure will trigger a reset - no need to clean
380 			 * up
381 			 */
382 			return err;
383 		}
384 		netif_dbg(priv, drv, priv->dev, "created tx queue %d\n", i);
385 	}
386 	for (i = 0; i < priv->rx_cfg.num_queues; i++) {
387 		err = gve_adminq_create_rx_queue(priv, i);
388 		if (err) {
389 			netif_err(priv, drv, priv->dev, "failed to create rx queue %d\n",
390 				  i);
391 			/* This failure will trigger a reset - no need to clean
392 			 * up
393 			 */
394 			return err;
395 		}
396 		/* Rx data ring has been prefilled with packet buffers at
397 		 * queue allocation time.
398 		 * Write the doorbell to provide descriptor slots and packet
399 		 * buffers to the NIC.
400 		 */
401 		gve_rx_write_doorbell(priv, &priv->rx[i]);
402 		netif_dbg(priv, drv, priv->dev, "created rx queue %d\n", i);
403 	}
404 
405 	return 0;
406 }
407 
408 static int gve_alloc_rings(struct gve_priv *priv)
409 {
410 	int ntfy_idx;
411 	int err;
412 	int i;
413 
414 	/* Setup tx rings */
415 	priv->tx = kvzalloc(priv->tx_cfg.num_queues * sizeof(*priv->tx),
416 			    GFP_KERNEL);
417 	if (!priv->tx)
418 		return -ENOMEM;
419 	err = gve_tx_alloc_rings(priv);
420 	if (err)
421 		goto free_tx;
422 	/* Setup rx rings */
423 	priv->rx = kvzalloc(priv->rx_cfg.num_queues * sizeof(*priv->rx),
424 			    GFP_KERNEL);
425 	if (!priv->rx) {
426 		err = -ENOMEM;
427 		goto free_tx_queue;
428 	}
429 	err = gve_rx_alloc_rings(priv);
430 	if (err)
431 		goto free_rx;
432 	/* Add tx napi & init sync stats*/
433 	for (i = 0; i < priv->tx_cfg.num_queues; i++) {
434 		u64_stats_init(&priv->tx[i].statss);
435 		ntfy_idx = gve_tx_idx_to_ntfy(priv, i);
436 		gve_add_napi(priv, ntfy_idx);
437 	}
438 	/* Add rx napi  & init sync stats*/
439 	for (i = 0; i < priv->rx_cfg.num_queues; i++) {
440 		u64_stats_init(&priv->rx[i].statss);
441 		ntfy_idx = gve_rx_idx_to_ntfy(priv, i);
442 		gve_add_napi(priv, ntfy_idx);
443 	}
444 
445 	return 0;
446 
447 free_rx:
448 	kvfree(priv->rx);
449 	priv->rx = NULL;
450 free_tx_queue:
451 	gve_tx_free_rings(priv);
452 free_tx:
453 	kvfree(priv->tx);
454 	priv->tx = NULL;
455 	return err;
456 }
457 
458 static int gve_destroy_rings(struct gve_priv *priv)
459 {
460 	int err;
461 	int i;
462 
463 	for (i = 0; i < priv->tx_cfg.num_queues; i++) {
464 		err = gve_adminq_destroy_tx_queue(priv, i);
465 		if (err) {
466 			netif_err(priv, drv, priv->dev,
467 				  "failed to destroy tx queue %d\n",
468 				  i);
469 			/* This failure will trigger a reset - no need to clean
470 			 * up
471 			 */
472 			return err;
473 		}
474 		netif_dbg(priv, drv, priv->dev, "destroyed tx queue %d\n", i);
475 	}
476 	for (i = 0; i < priv->rx_cfg.num_queues; i++) {
477 		err = gve_adminq_destroy_rx_queue(priv, i);
478 		if (err) {
479 			netif_err(priv, drv, priv->dev,
480 				  "failed to destroy rx queue %d\n",
481 				  i);
482 			/* This failure will trigger a reset - no need to clean
483 			 * up
484 			 */
485 			return err;
486 		}
487 		netif_dbg(priv, drv, priv->dev, "destroyed rx queue %d\n", i);
488 	}
489 	return 0;
490 }
491 
492 static void gve_free_rings(struct gve_priv *priv)
493 {
494 	int ntfy_idx;
495 	int i;
496 
497 	if (priv->tx) {
498 		for (i = 0; i < priv->tx_cfg.num_queues; i++) {
499 			ntfy_idx = gve_tx_idx_to_ntfy(priv, i);
500 			gve_remove_napi(priv, ntfy_idx);
501 		}
502 		gve_tx_free_rings(priv);
503 		kvfree(priv->tx);
504 		priv->tx = NULL;
505 	}
506 	if (priv->rx) {
507 		for (i = 0; i < priv->rx_cfg.num_queues; i++) {
508 			ntfy_idx = gve_rx_idx_to_ntfy(priv, i);
509 			gve_remove_napi(priv, ntfy_idx);
510 		}
511 		gve_rx_free_rings(priv);
512 		kvfree(priv->rx);
513 		priv->rx = NULL;
514 	}
515 }
516 
517 int gve_alloc_page(struct device *dev, struct page **page, dma_addr_t *dma,
518 		   enum dma_data_direction dir)
519 {
520 	*page = alloc_page(GFP_KERNEL);
521 	if (!*page)
522 		return -ENOMEM;
523 	*dma = dma_map_page(dev, *page, 0, PAGE_SIZE, dir);
524 	if (dma_mapping_error(dev, *dma)) {
525 		put_page(*page);
526 		return -ENOMEM;
527 	}
528 	return 0;
529 }
530 
531 static int gve_alloc_queue_page_list(struct gve_priv *priv, u32 id,
532 				     int pages)
533 {
534 	struct gve_queue_page_list *qpl = &priv->qpls[id];
535 	int err;
536 	int i;
537 
538 	if (pages + priv->num_registered_pages > priv->max_registered_pages) {
539 		netif_err(priv, drv, priv->dev,
540 			  "Reached max number of registered pages %llu > %llu\n",
541 			  pages + priv->num_registered_pages,
542 			  priv->max_registered_pages);
543 		return -EINVAL;
544 	}
545 
546 	qpl->id = id;
547 	qpl->num_entries = pages;
548 	qpl->pages = kvzalloc(pages * sizeof(*qpl->pages), GFP_KERNEL);
549 	/* caller handles clean up */
550 	if (!qpl->pages)
551 		return -ENOMEM;
552 	qpl->page_buses = kvzalloc(pages * sizeof(*qpl->page_buses),
553 				   GFP_KERNEL);
554 	/* caller handles clean up */
555 	if (!qpl->page_buses)
556 		return -ENOMEM;
557 
558 	for (i = 0; i < pages; i++) {
559 		err = gve_alloc_page(&priv->pdev->dev, &qpl->pages[i],
560 				     &qpl->page_buses[i],
561 				     gve_qpl_dma_dir(priv, id));
562 		/* caller handles clean up */
563 		if (err)
564 			return -ENOMEM;
565 	}
566 	priv->num_registered_pages += pages;
567 
568 	return 0;
569 }
570 
571 void gve_free_page(struct device *dev, struct page *page, dma_addr_t dma,
572 		   enum dma_data_direction dir)
573 {
574 	if (!dma_mapping_error(dev, dma))
575 		dma_unmap_page(dev, dma, PAGE_SIZE, dir);
576 	if (page)
577 		put_page(page);
578 }
579 
580 static void gve_free_queue_page_list(struct gve_priv *priv,
581 				     int id)
582 {
583 	struct gve_queue_page_list *qpl = &priv->qpls[id];
584 	int i;
585 
586 	if (!qpl->pages)
587 		return;
588 	if (!qpl->page_buses)
589 		goto free_pages;
590 
591 	for (i = 0; i < qpl->num_entries; i++)
592 		gve_free_page(&priv->pdev->dev, qpl->pages[i],
593 			      qpl->page_buses[i], gve_qpl_dma_dir(priv, id));
594 
595 	kvfree(qpl->page_buses);
596 free_pages:
597 	kvfree(qpl->pages);
598 	priv->num_registered_pages -= qpl->num_entries;
599 }
600 
601 static int gve_alloc_qpls(struct gve_priv *priv)
602 {
603 	int num_qpls = gve_num_tx_qpls(priv) + gve_num_rx_qpls(priv);
604 	int i, j;
605 	int err;
606 
607 	priv->qpls = kvzalloc(num_qpls * sizeof(*priv->qpls), GFP_KERNEL);
608 	if (!priv->qpls)
609 		return -ENOMEM;
610 
611 	for (i = 0; i < gve_num_tx_qpls(priv); i++) {
612 		err = gve_alloc_queue_page_list(priv, i,
613 						priv->tx_pages_per_qpl);
614 		if (err)
615 			goto free_qpls;
616 	}
617 	for (; i < num_qpls; i++) {
618 		err = gve_alloc_queue_page_list(priv, i,
619 						priv->rx_pages_per_qpl);
620 		if (err)
621 			goto free_qpls;
622 	}
623 
624 	priv->qpl_cfg.qpl_map_size = BITS_TO_LONGS(num_qpls) *
625 				     sizeof(unsigned long) * BITS_PER_BYTE;
626 	priv->qpl_cfg.qpl_id_map = kvzalloc(BITS_TO_LONGS(num_qpls) *
627 					    sizeof(unsigned long), GFP_KERNEL);
628 	if (!priv->qpl_cfg.qpl_id_map) {
629 		err = -ENOMEM;
630 		goto free_qpls;
631 	}
632 
633 	return 0;
634 
635 free_qpls:
636 	for (j = 0; j <= i; j++)
637 		gve_free_queue_page_list(priv, j);
638 	kvfree(priv->qpls);
639 	return err;
640 }
641 
642 static void gve_free_qpls(struct gve_priv *priv)
643 {
644 	int num_qpls = gve_num_tx_qpls(priv) + gve_num_rx_qpls(priv);
645 	int i;
646 
647 	kvfree(priv->qpl_cfg.qpl_id_map);
648 
649 	for (i = 0; i < num_qpls; i++)
650 		gve_free_queue_page_list(priv, i);
651 
652 	kvfree(priv->qpls);
653 }
654 
655 /* Use this to schedule a reset when the device is capable of continuing
656  * to handle other requests in its current state. If it is not, do a reset
657  * in thread instead.
658  */
659 void gve_schedule_reset(struct gve_priv *priv)
660 {
661 	gve_set_do_reset(priv);
662 	queue_work(priv->gve_wq, &priv->service_task);
663 }
664 
665 static void gve_reset_and_teardown(struct gve_priv *priv, bool was_up);
666 static int gve_reset_recovery(struct gve_priv *priv, bool was_up);
667 static void gve_turndown(struct gve_priv *priv);
668 static void gve_turnup(struct gve_priv *priv);
669 
670 static int gve_open(struct net_device *dev)
671 {
672 	struct gve_priv *priv = netdev_priv(dev);
673 	int err;
674 
675 	err = gve_alloc_qpls(priv);
676 	if (err)
677 		return err;
678 	err = gve_alloc_rings(priv);
679 	if (err)
680 		goto free_qpls;
681 
682 	err = netif_set_real_num_tx_queues(dev, priv->tx_cfg.num_queues);
683 	if (err)
684 		goto free_rings;
685 	err = netif_set_real_num_rx_queues(dev, priv->rx_cfg.num_queues);
686 	if (err)
687 		goto free_rings;
688 
689 	err = gve_register_qpls(priv);
690 	if (err)
691 		goto reset;
692 	err = gve_create_rings(priv);
693 	if (err)
694 		goto reset;
695 	gve_set_device_rings_ok(priv);
696 
697 	gve_turnup(priv);
698 	netif_carrier_on(dev);
699 	return 0;
700 
701 free_rings:
702 	gve_free_rings(priv);
703 free_qpls:
704 	gve_free_qpls(priv);
705 	return err;
706 
707 reset:
708 	/* This must have been called from a reset due to the rtnl lock
709 	 * so just return at this point.
710 	 */
711 	if (gve_get_reset_in_progress(priv))
712 		return err;
713 	/* Otherwise reset before returning */
714 	gve_reset_and_teardown(priv, true);
715 	/* if this fails there is nothing we can do so just ignore the return */
716 	gve_reset_recovery(priv, false);
717 	/* return the original error */
718 	return err;
719 }
720 
721 static int gve_close(struct net_device *dev)
722 {
723 	struct gve_priv *priv = netdev_priv(dev);
724 	int err;
725 
726 	netif_carrier_off(dev);
727 	if (gve_get_device_rings_ok(priv)) {
728 		gve_turndown(priv);
729 		err = gve_destroy_rings(priv);
730 		if (err)
731 			goto err;
732 		err = gve_unregister_qpls(priv);
733 		if (err)
734 			goto err;
735 		gve_clear_device_rings_ok(priv);
736 	}
737 
738 	gve_free_rings(priv);
739 	gve_free_qpls(priv);
740 	return 0;
741 
742 err:
743 	/* This must have been called from a reset due to the rtnl lock
744 	 * so just return at this point.
745 	 */
746 	if (gve_get_reset_in_progress(priv))
747 		return err;
748 	/* Otherwise reset before returning */
749 	gve_reset_and_teardown(priv, true);
750 	return gve_reset_recovery(priv, false);
751 }
752 
753 int gve_adjust_queues(struct gve_priv *priv,
754 		      struct gve_queue_config new_rx_config,
755 		      struct gve_queue_config new_tx_config)
756 {
757 	int err;
758 
759 	if (netif_carrier_ok(priv->dev)) {
760 		/* To make this process as simple as possible we teardown the
761 		 * device, set the new configuration, and then bring the device
762 		 * up again.
763 		 */
764 		err = gve_close(priv->dev);
765 		/* we have already tried to reset in close,
766 		 * just fail at this point
767 		 */
768 		if (err)
769 			return err;
770 		priv->tx_cfg = new_tx_config;
771 		priv->rx_cfg = new_rx_config;
772 
773 		err = gve_open(priv->dev);
774 		if (err)
775 			goto err;
776 
777 		return 0;
778 	}
779 	/* Set the config for the next up. */
780 	priv->tx_cfg = new_tx_config;
781 	priv->rx_cfg = new_rx_config;
782 
783 	return 0;
784 err:
785 	netif_err(priv, drv, priv->dev,
786 		  "Adjust queues failed! !!! DISABLING ALL QUEUES !!!\n");
787 	gve_turndown(priv);
788 	return err;
789 }
790 
791 static void gve_turndown(struct gve_priv *priv)
792 {
793 	int idx;
794 
795 	if (netif_carrier_ok(priv->dev))
796 		netif_carrier_off(priv->dev);
797 
798 	if (!gve_get_napi_enabled(priv))
799 		return;
800 
801 	/* Disable napi to prevent more work from coming in */
802 	for (idx = 0; idx < priv->tx_cfg.num_queues; idx++) {
803 		int ntfy_idx = gve_tx_idx_to_ntfy(priv, idx);
804 		struct gve_notify_block *block = &priv->ntfy_blocks[ntfy_idx];
805 
806 		napi_disable(&block->napi);
807 	}
808 	for (idx = 0; idx < priv->rx_cfg.num_queues; idx++) {
809 		int ntfy_idx = gve_rx_idx_to_ntfy(priv, idx);
810 		struct gve_notify_block *block = &priv->ntfy_blocks[ntfy_idx];
811 
812 		napi_disable(&block->napi);
813 	}
814 
815 	/* Stop tx queues */
816 	netif_tx_disable(priv->dev);
817 
818 	gve_clear_napi_enabled(priv);
819 }
820 
821 static void gve_turnup(struct gve_priv *priv)
822 {
823 	int idx;
824 
825 	/* Start the tx queues */
826 	netif_tx_start_all_queues(priv->dev);
827 
828 	/* Enable napi and unmask interrupts for all queues */
829 	for (idx = 0; idx < priv->tx_cfg.num_queues; idx++) {
830 		int ntfy_idx = gve_tx_idx_to_ntfy(priv, idx);
831 		struct gve_notify_block *block = &priv->ntfy_blocks[ntfy_idx];
832 
833 		napi_enable(&block->napi);
834 		iowrite32be(0, gve_irq_doorbell(priv, block));
835 	}
836 	for (idx = 0; idx < priv->rx_cfg.num_queues; idx++) {
837 		int ntfy_idx = gve_rx_idx_to_ntfy(priv, idx);
838 		struct gve_notify_block *block = &priv->ntfy_blocks[ntfy_idx];
839 
840 		napi_enable(&block->napi);
841 		iowrite32be(0, gve_irq_doorbell(priv, block));
842 	}
843 
844 	gve_set_napi_enabled(priv);
845 }
846 
847 static void gve_tx_timeout(struct net_device *dev)
848 {
849 	struct gve_priv *priv = netdev_priv(dev);
850 
851 	gve_schedule_reset(priv);
852 	priv->tx_timeo_cnt++;
853 }
854 
855 static const struct net_device_ops gve_netdev_ops = {
856 	.ndo_start_xmit		=	gve_tx,
857 	.ndo_open		=	gve_open,
858 	.ndo_stop		=	gve_close,
859 	.ndo_get_stats64	=	gve_get_stats,
860 	.ndo_tx_timeout         =       gve_tx_timeout,
861 };
862 
863 static void gve_handle_status(struct gve_priv *priv, u32 status)
864 {
865 	if (GVE_DEVICE_STATUS_RESET_MASK & status) {
866 		dev_info(&priv->pdev->dev, "Device requested reset.\n");
867 		gve_set_do_reset(priv);
868 	}
869 }
870 
871 static void gve_handle_reset(struct gve_priv *priv)
872 {
873 	/* A service task will be scheduled at the end of probe to catch any
874 	 * resets that need to happen, and we don't want to reset until
875 	 * probe is done.
876 	 */
877 	if (gve_get_probe_in_progress(priv))
878 		return;
879 
880 	if (gve_get_do_reset(priv)) {
881 		rtnl_lock();
882 		gve_reset(priv, false);
883 		rtnl_unlock();
884 	}
885 }
886 
887 /* Handle NIC status register changes and reset requests */
888 static void gve_service_task(struct work_struct *work)
889 {
890 	struct gve_priv *priv = container_of(work, struct gve_priv,
891 					     service_task);
892 
893 	gve_handle_status(priv,
894 			  ioread32be(&priv->reg_bar0->device_status));
895 
896 	gve_handle_reset(priv);
897 }
898 
899 static int gve_init_priv(struct gve_priv *priv, bool skip_describe_device)
900 {
901 	int num_ntfy;
902 	int err;
903 
904 	/* Set up the adminq */
905 	err = gve_adminq_alloc(&priv->pdev->dev, priv);
906 	if (err) {
907 		dev_err(&priv->pdev->dev,
908 			"Failed to alloc admin queue: err=%d\n", err);
909 		return err;
910 	}
911 
912 	if (skip_describe_device)
913 		goto setup_device;
914 
915 	/* Get the initial information we need from the device */
916 	err = gve_adminq_describe_device(priv);
917 	if (err) {
918 		dev_err(&priv->pdev->dev,
919 			"Could not get device information: err=%d\n", err);
920 		goto err;
921 	}
922 	if (priv->dev->max_mtu > PAGE_SIZE) {
923 		priv->dev->max_mtu = PAGE_SIZE;
924 		err = gve_adminq_set_mtu(priv, priv->dev->mtu);
925 		if (err) {
926 			netif_err(priv, drv, priv->dev, "Could not set mtu");
927 			goto err;
928 		}
929 	}
930 	priv->dev->mtu = priv->dev->max_mtu;
931 	num_ntfy = pci_msix_vec_count(priv->pdev);
932 	if (num_ntfy <= 0) {
933 		dev_err(&priv->pdev->dev,
934 			"could not count MSI-x vectors: err=%d\n", num_ntfy);
935 		err = num_ntfy;
936 		goto err;
937 	} else if (num_ntfy < GVE_MIN_MSIX) {
938 		dev_err(&priv->pdev->dev, "gve needs at least %d MSI-x vectors, but only has %d\n",
939 			GVE_MIN_MSIX, num_ntfy);
940 		err = -EINVAL;
941 		goto err;
942 	}
943 
944 	priv->num_registered_pages = 0;
945 	priv->rx_copybreak = GVE_DEFAULT_RX_COPYBREAK;
946 	/* gvnic has one Notification Block per MSI-x vector, except for the
947 	 * management vector
948 	 */
949 	priv->num_ntfy_blks = (num_ntfy - 1) & ~0x1;
950 	priv->mgmt_msix_idx = priv->num_ntfy_blks;
951 
952 	priv->tx_cfg.max_queues =
953 		min_t(int, priv->tx_cfg.max_queues, priv->num_ntfy_blks / 2);
954 	priv->rx_cfg.max_queues =
955 		min_t(int, priv->rx_cfg.max_queues, priv->num_ntfy_blks / 2);
956 
957 	priv->tx_cfg.num_queues = priv->tx_cfg.max_queues;
958 	priv->rx_cfg.num_queues = priv->rx_cfg.max_queues;
959 	if (priv->default_num_queues > 0) {
960 		priv->tx_cfg.num_queues = min_t(int, priv->default_num_queues,
961 						priv->tx_cfg.num_queues);
962 		priv->rx_cfg.num_queues = min_t(int, priv->default_num_queues,
963 						priv->rx_cfg.num_queues);
964 	}
965 
966 	netif_info(priv, drv, priv->dev, "TX queues %d, RX queues %d\n",
967 		   priv->tx_cfg.num_queues, priv->rx_cfg.num_queues);
968 	netif_info(priv, drv, priv->dev, "Max TX queues %d, Max RX queues %d\n",
969 		   priv->tx_cfg.max_queues, priv->rx_cfg.max_queues);
970 
971 setup_device:
972 	err = gve_setup_device_resources(priv);
973 	if (!err)
974 		return 0;
975 err:
976 	gve_adminq_free(&priv->pdev->dev, priv);
977 	return err;
978 }
979 
980 static void gve_teardown_priv_resources(struct gve_priv *priv)
981 {
982 	gve_teardown_device_resources(priv);
983 	gve_adminq_free(&priv->pdev->dev, priv);
984 }
985 
986 static void gve_trigger_reset(struct gve_priv *priv)
987 {
988 	/* Reset the device by releasing the AQ */
989 	gve_adminq_release(priv);
990 }
991 
992 static void gve_reset_and_teardown(struct gve_priv *priv, bool was_up)
993 {
994 	gve_trigger_reset(priv);
995 	/* With the reset having already happened, close cannot fail */
996 	if (was_up)
997 		gve_close(priv->dev);
998 	gve_teardown_priv_resources(priv);
999 }
1000 
1001 static int gve_reset_recovery(struct gve_priv *priv, bool was_up)
1002 {
1003 	int err;
1004 
1005 	err = gve_init_priv(priv, true);
1006 	if (err)
1007 		goto err;
1008 	if (was_up) {
1009 		err = gve_open(priv->dev);
1010 		if (err)
1011 			goto err;
1012 	}
1013 	return 0;
1014 err:
1015 	dev_err(&priv->pdev->dev, "Reset failed! !!! DISABLING ALL QUEUES !!!\n");
1016 	gve_turndown(priv);
1017 	return err;
1018 }
1019 
1020 int gve_reset(struct gve_priv *priv, bool attempt_teardown)
1021 {
1022 	bool was_up = netif_carrier_ok(priv->dev);
1023 	int err;
1024 
1025 	dev_info(&priv->pdev->dev, "Performing reset\n");
1026 	gve_clear_do_reset(priv);
1027 	gve_set_reset_in_progress(priv);
1028 	/* If we aren't attempting to teardown normally, just go turndown and
1029 	 * reset right away.
1030 	 */
1031 	if (!attempt_teardown) {
1032 		gve_turndown(priv);
1033 		gve_reset_and_teardown(priv, was_up);
1034 	} else {
1035 		/* Otherwise attempt to close normally */
1036 		if (was_up) {
1037 			err = gve_close(priv->dev);
1038 			/* If that fails reset as we did above */
1039 			if (err)
1040 				gve_reset_and_teardown(priv, was_up);
1041 		}
1042 		/* Clean up any remaining resources */
1043 		gve_teardown_priv_resources(priv);
1044 	}
1045 
1046 	/* Set it all back up */
1047 	err = gve_reset_recovery(priv, was_up);
1048 	gve_clear_reset_in_progress(priv);
1049 	return err;
1050 }
1051 
1052 static void gve_write_version(u8 __iomem *driver_version_register)
1053 {
1054 	const char *c = gve_version_prefix;
1055 
1056 	while (*c) {
1057 		writeb(*c, driver_version_register);
1058 		c++;
1059 	}
1060 
1061 	c = gve_version_str;
1062 	while (*c) {
1063 		writeb(*c, driver_version_register);
1064 		c++;
1065 	}
1066 	writeb('\n', driver_version_register);
1067 }
1068 
1069 static int gve_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
1070 {
1071 	int max_tx_queues, max_rx_queues;
1072 	struct net_device *dev;
1073 	__be32 __iomem *db_bar;
1074 	struct gve_registers __iomem *reg_bar;
1075 	struct gve_priv *priv;
1076 	int err;
1077 
1078 	err = pci_enable_device(pdev);
1079 	if (err)
1080 		return -ENXIO;
1081 
1082 	err = pci_request_regions(pdev, "gvnic-cfg");
1083 	if (err)
1084 		goto abort_with_enabled;
1085 
1086 	pci_set_master(pdev);
1087 
1088 	err = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
1089 	if (err) {
1090 		dev_err(&pdev->dev, "Failed to set dma mask: err=%d\n", err);
1091 		goto abort_with_pci_region;
1092 	}
1093 
1094 	err = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
1095 	if (err) {
1096 		dev_err(&pdev->dev,
1097 			"Failed to set consistent dma mask: err=%d\n", err);
1098 		goto abort_with_pci_region;
1099 	}
1100 
1101 	reg_bar = pci_iomap(pdev, GVE_REGISTER_BAR, 0);
1102 	if (!reg_bar) {
1103 		dev_err(&pdev->dev, "Failed to map pci bar!\n");
1104 		err = -ENOMEM;
1105 		goto abort_with_pci_region;
1106 	}
1107 
1108 	db_bar = pci_iomap(pdev, GVE_DOORBELL_BAR, 0);
1109 	if (!db_bar) {
1110 		dev_err(&pdev->dev, "Failed to map doorbell bar!\n");
1111 		err = -ENOMEM;
1112 		goto abort_with_reg_bar;
1113 	}
1114 
1115 	gve_write_version(&reg_bar->driver_version);
1116 	/* Get max queues to alloc etherdev */
1117 	max_rx_queues = ioread32be(&reg_bar->max_tx_queues);
1118 	max_tx_queues = ioread32be(&reg_bar->max_rx_queues);
1119 	/* Alloc and setup the netdev and priv */
1120 	dev = alloc_etherdev_mqs(sizeof(*priv), max_tx_queues, max_rx_queues);
1121 	if (!dev) {
1122 		dev_err(&pdev->dev, "could not allocate netdev\n");
1123 		goto abort_with_db_bar;
1124 	}
1125 	SET_NETDEV_DEV(dev, &pdev->dev);
1126 	pci_set_drvdata(pdev, dev);
1127 	dev->ethtool_ops = &gve_ethtool_ops;
1128 	dev->netdev_ops = &gve_netdev_ops;
1129 	/* advertise features */
1130 	dev->hw_features = NETIF_F_HIGHDMA;
1131 	dev->hw_features |= NETIF_F_SG;
1132 	dev->hw_features |= NETIF_F_HW_CSUM;
1133 	dev->hw_features |= NETIF_F_TSO;
1134 	dev->hw_features |= NETIF_F_TSO6;
1135 	dev->hw_features |= NETIF_F_TSO_ECN;
1136 	dev->hw_features |= NETIF_F_RXCSUM;
1137 	dev->hw_features |= NETIF_F_RXHASH;
1138 	dev->features = dev->hw_features;
1139 	dev->watchdog_timeo = 5 * HZ;
1140 	dev->min_mtu = ETH_MIN_MTU;
1141 	netif_carrier_off(dev);
1142 
1143 	priv = netdev_priv(dev);
1144 	priv->dev = dev;
1145 	priv->pdev = pdev;
1146 	priv->msg_enable = DEFAULT_MSG_LEVEL;
1147 	priv->reg_bar0 = reg_bar;
1148 	priv->db_bar2 = db_bar;
1149 	priv->service_task_flags = 0x0;
1150 	priv->state_flags = 0x0;
1151 
1152 	gve_set_probe_in_progress(priv);
1153 	priv->gve_wq = alloc_ordered_workqueue("gve", 0);
1154 	if (!priv->gve_wq) {
1155 		dev_err(&pdev->dev, "Could not allocate workqueue");
1156 		err = -ENOMEM;
1157 		goto abort_with_netdev;
1158 	}
1159 	INIT_WORK(&priv->service_task, gve_service_task);
1160 	priv->tx_cfg.max_queues = max_tx_queues;
1161 	priv->rx_cfg.max_queues = max_rx_queues;
1162 
1163 	err = gve_init_priv(priv, false);
1164 	if (err)
1165 		goto abort_with_wq;
1166 
1167 	err = register_netdev(dev);
1168 	if (err)
1169 		goto abort_with_wq;
1170 
1171 	dev_info(&pdev->dev, "GVE version %s\n", gve_version_str);
1172 	gve_clear_probe_in_progress(priv);
1173 	queue_work(priv->gve_wq, &priv->service_task);
1174 	return 0;
1175 
1176 abort_with_wq:
1177 	destroy_workqueue(priv->gve_wq);
1178 
1179 abort_with_netdev:
1180 	free_netdev(dev);
1181 
1182 abort_with_db_bar:
1183 	pci_iounmap(pdev, db_bar);
1184 
1185 abort_with_reg_bar:
1186 	pci_iounmap(pdev, reg_bar);
1187 
1188 abort_with_pci_region:
1189 	pci_release_regions(pdev);
1190 
1191 abort_with_enabled:
1192 	pci_disable_device(pdev);
1193 	return -ENXIO;
1194 }
1195 
1196 static void gve_remove(struct pci_dev *pdev)
1197 {
1198 	struct net_device *netdev = pci_get_drvdata(pdev);
1199 	struct gve_priv *priv = netdev_priv(netdev);
1200 	__be32 __iomem *db_bar = priv->db_bar2;
1201 	void __iomem *reg_bar = priv->reg_bar0;
1202 
1203 	unregister_netdev(netdev);
1204 	gve_teardown_priv_resources(priv);
1205 	destroy_workqueue(priv->gve_wq);
1206 	free_netdev(netdev);
1207 	pci_iounmap(pdev, db_bar);
1208 	pci_iounmap(pdev, reg_bar);
1209 	pci_release_regions(pdev);
1210 	pci_disable_device(pdev);
1211 }
1212 
1213 static const struct pci_device_id gve_id_table[] = {
1214 	{ PCI_DEVICE(PCI_VENDOR_ID_GOOGLE, PCI_DEV_ID_GVNIC) },
1215 	{ }
1216 };
1217 
1218 static struct pci_driver gvnic_driver = {
1219 	.name		= "gvnic",
1220 	.id_table	= gve_id_table,
1221 	.probe		= gve_probe,
1222 	.remove		= gve_remove,
1223 };
1224 
1225 module_pci_driver(gvnic_driver);
1226 
1227 MODULE_DEVICE_TABLE(pci, gve_id_table);
1228 MODULE_AUTHOR("Google, Inc.");
1229 MODULE_DESCRIPTION("gVNIC Driver");
1230 MODULE_LICENSE("Dual MIT/GPL");
1231 MODULE_VERSION(GVE_VERSION);
1232