1 // SPDX-License-Identifier: (GPL-2.0 OR MIT)
2 /* Google virtual Ethernet (gve) driver
4 * Copyright (C) 2015-2019 Google, Inc.
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>
17 #include "gve_adminq.h"
18 #include "gve_register.h"
20 #define GVE_DEFAULT_RX_COPYBREAK (256)
22 #define DEFAULT_MSG_LEVEL (NETIF_MSG_DRV | NETIF_MSG_LINK)
23 #define GVE_VERSION "1.0.0"
24 #define GVE_VERSION_PREFIX "GVE-"
26 const char gve_version_str[] = GVE_VERSION;
27 static const char gve_version_prefix[] = GVE_VERSION_PREFIX;
29 static void gve_get_stats(struct net_device *dev, struct rtnl_link_stats64 *s)
31 struct gve_priv *priv = netdev_priv(dev);
36 for (ring = 0; ring < priv->rx_cfg.num_queues; ring++) {
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,
47 for (ring = 0; ring < priv->tx_cfg.num_queues; ring++) {
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->tx[ring].statss,
59 static int gve_alloc_counter_array(struct gve_priv *priv)
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)
72 static void gve_free_counter_array(struct gve_priv *priv)
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;
81 /* NIC requests to report stats */
82 static void gve_stats_report_task(struct work_struct *work)
84 struct gve_priv *priv = container_of(work, struct gve_priv,
86 if (gve_get_do_report_stats(priv)) {
87 gve_handle_report_stats(priv);
88 gve_clear_do_report_stats(priv);
92 static void gve_stats_report_schedule(struct gve_priv *priv)
94 if (!gve_get_probe_in_progress(priv) &&
95 !gve_get_reset_in_progress(priv)) {
96 gve_set_do_report_stats(priv);
97 queue_work(priv->gve_wq, &priv->stats_report_task);
101 static void gve_stats_report_timer(struct timer_list *t)
103 struct gve_priv *priv = from_timer(priv, t, stats_report_timer);
105 mod_timer(&priv->stats_report_timer,
106 round_jiffies(jiffies +
107 msecs_to_jiffies(priv->stats_report_timer_period)));
108 gve_stats_report_schedule(priv);
111 static int gve_alloc_stats_report(struct gve_priv *priv)
113 int tx_stats_num, rx_stats_num;
115 tx_stats_num = (GVE_TX_STATS_REPORT_NUM + NIC_TX_STATS_REPORT_NUM) *
116 priv->tx_cfg.num_queues;
117 rx_stats_num = (GVE_RX_STATS_REPORT_NUM + NIC_RX_STATS_REPORT_NUM) *
118 priv->rx_cfg.num_queues;
119 priv->stats_report_len = struct_size(priv->stats_report, stats,
120 tx_stats_num + rx_stats_num);
122 dma_alloc_coherent(&priv->pdev->dev, priv->stats_report_len,
123 &priv->stats_report_bus, GFP_KERNEL);
124 if (!priv->stats_report)
126 /* Set up timer for the report-stats task */
127 timer_setup(&priv->stats_report_timer, gve_stats_report_timer, 0);
128 priv->stats_report_timer_period = GVE_STATS_REPORT_TIMER_PERIOD;
132 static void gve_free_stats_report(struct gve_priv *priv)
134 del_timer_sync(&priv->stats_report_timer);
135 dma_free_coherent(&priv->pdev->dev, priv->stats_report_len,
136 priv->stats_report, priv->stats_report_bus);
137 priv->stats_report = NULL;
140 static irqreturn_t gve_mgmnt_intr(int irq, void *arg)
142 struct gve_priv *priv = arg;
144 queue_work(priv->gve_wq, &priv->service_task);
148 static irqreturn_t gve_intr(int irq, void *arg)
150 struct gve_notify_block *block = arg;
151 struct gve_priv *priv = block->priv;
153 iowrite32be(GVE_IRQ_MASK, gve_irq_doorbell(priv, block));
154 napi_schedule_irqoff(&block->napi);
158 static int gve_napi_poll(struct napi_struct *napi, int budget)
160 struct gve_notify_block *block;
161 __be32 __iomem *irq_doorbell;
162 bool reschedule = false;
163 struct gve_priv *priv;
165 block = container_of(napi, struct gve_notify_block, napi);
169 reschedule |= gve_tx_poll(block, budget);
171 reschedule |= gve_rx_poll(block, budget);
177 irq_doorbell = gve_irq_doorbell(priv, block);
178 iowrite32be(GVE_IRQ_ACK | GVE_IRQ_EVENT, irq_doorbell);
180 /* Double check we have no extra work.
181 * Ensure unmask synchronizes with checking for work.
185 reschedule |= gve_tx_poll(block, -1);
187 reschedule |= gve_rx_poll(block, -1);
188 if (reschedule && napi_reschedule(napi))
189 iowrite32be(GVE_IRQ_MASK, irq_doorbell);
194 static int gve_alloc_notify_blocks(struct gve_priv *priv)
196 int num_vecs_requested = priv->num_ntfy_blks + 1;
197 char *name = priv->dev->name;
198 unsigned int active_cpus;
203 priv->msix_vectors = kvzalloc(num_vecs_requested *
204 sizeof(*priv->msix_vectors), GFP_KERNEL);
205 if (!priv->msix_vectors)
207 for (i = 0; i < num_vecs_requested; i++)
208 priv->msix_vectors[i].entry = i;
209 vecs_enabled = pci_enable_msix_range(priv->pdev, priv->msix_vectors,
210 GVE_MIN_MSIX, num_vecs_requested);
211 if (vecs_enabled < 0) {
212 dev_err(&priv->pdev->dev, "Could not enable min msix %d/%d\n",
213 GVE_MIN_MSIX, vecs_enabled);
215 goto abort_with_msix_vectors;
217 if (vecs_enabled != num_vecs_requested) {
218 int new_num_ntfy_blks = (vecs_enabled - 1) & ~0x1;
219 int vecs_per_type = new_num_ntfy_blks / 2;
220 int vecs_left = new_num_ntfy_blks % 2;
222 priv->num_ntfy_blks = new_num_ntfy_blks;
223 priv->mgmt_msix_idx = priv->num_ntfy_blks;
224 priv->tx_cfg.max_queues = min_t(int, priv->tx_cfg.max_queues,
226 priv->rx_cfg.max_queues = min_t(int, priv->rx_cfg.max_queues,
227 vecs_per_type + vecs_left);
228 dev_err(&priv->pdev->dev,
229 "Could not enable desired msix, only enabled %d, adjusting tx max queues to %d, and rx max queues to %d\n",
230 vecs_enabled, priv->tx_cfg.max_queues,
231 priv->rx_cfg.max_queues);
232 if (priv->tx_cfg.num_queues > priv->tx_cfg.max_queues)
233 priv->tx_cfg.num_queues = priv->tx_cfg.max_queues;
234 if (priv->rx_cfg.num_queues > priv->rx_cfg.max_queues)
235 priv->rx_cfg.num_queues = priv->rx_cfg.max_queues;
237 /* Half the notification blocks go to TX and half to RX */
238 active_cpus = min_t(int, priv->num_ntfy_blks / 2, num_online_cpus());
240 /* Setup Management Vector - the last vector */
241 snprintf(priv->mgmt_msix_name, sizeof(priv->mgmt_msix_name), "%s-mgmnt",
243 err = request_irq(priv->msix_vectors[priv->mgmt_msix_idx].vector,
244 gve_mgmnt_intr, 0, priv->mgmt_msix_name, priv);
246 dev_err(&priv->pdev->dev, "Did not receive management vector.\n");
247 goto abort_with_msix_enabled;
250 dma_alloc_coherent(&priv->pdev->dev,
251 priv->num_ntfy_blks *
252 sizeof(*priv->ntfy_blocks),
253 &priv->ntfy_block_bus, GFP_KERNEL);
254 if (!priv->ntfy_blocks) {
256 goto abort_with_mgmt_vector;
258 /* Setup the other blocks - the first n-1 vectors */
259 for (i = 0; i < priv->num_ntfy_blks; i++) {
260 struct gve_notify_block *block = &priv->ntfy_blocks[i];
263 snprintf(block->name, sizeof(block->name), "%s-ntfy-block.%d",
266 err = request_irq(priv->msix_vectors[msix_idx].vector,
267 gve_intr, 0, block->name, block);
269 dev_err(&priv->pdev->dev,
270 "Failed to receive msix vector %d\n", i);
271 goto abort_with_some_ntfy_blocks;
273 irq_set_affinity_hint(priv->msix_vectors[msix_idx].vector,
274 get_cpu_mask(i % active_cpus));
277 abort_with_some_ntfy_blocks:
278 for (j = 0; j < i; j++) {
279 struct gve_notify_block *block = &priv->ntfy_blocks[j];
282 irq_set_affinity_hint(priv->msix_vectors[msix_idx].vector,
284 free_irq(priv->msix_vectors[msix_idx].vector, block);
286 dma_free_coherent(&priv->pdev->dev, priv->num_ntfy_blks *
287 sizeof(*priv->ntfy_blocks),
288 priv->ntfy_blocks, priv->ntfy_block_bus);
289 priv->ntfy_blocks = NULL;
290 abort_with_mgmt_vector:
291 free_irq(priv->msix_vectors[priv->mgmt_msix_idx].vector, priv);
292 abort_with_msix_enabled:
293 pci_disable_msix(priv->pdev);
294 abort_with_msix_vectors:
295 kvfree(priv->msix_vectors);
296 priv->msix_vectors = NULL;
300 static void gve_free_notify_blocks(struct gve_priv *priv)
305 for (i = 0; i < priv->num_ntfy_blks; i++) {
306 struct gve_notify_block *block = &priv->ntfy_blocks[i];
309 irq_set_affinity_hint(priv->msix_vectors[msix_idx].vector,
311 free_irq(priv->msix_vectors[msix_idx].vector, block);
313 dma_free_coherent(&priv->pdev->dev,
314 priv->num_ntfy_blks * sizeof(*priv->ntfy_blocks),
315 priv->ntfy_blocks, priv->ntfy_block_bus);
316 priv->ntfy_blocks = NULL;
317 free_irq(priv->msix_vectors[priv->mgmt_msix_idx].vector, priv);
318 pci_disable_msix(priv->pdev);
319 kvfree(priv->msix_vectors);
320 priv->msix_vectors = NULL;
323 static int gve_setup_device_resources(struct gve_priv *priv)
327 err = gve_alloc_counter_array(priv);
330 err = gve_alloc_notify_blocks(priv);
332 goto abort_with_counter;
333 err = gve_alloc_stats_report(priv);
335 goto abort_with_ntfy_blocks;
336 err = gve_adminq_configure_device_resources(priv,
337 priv->counter_array_bus,
338 priv->num_event_counters,
339 priv->ntfy_block_bus,
340 priv->num_ntfy_blks);
342 dev_err(&priv->pdev->dev,
343 "could not setup device_resources: err=%d\n", err);
345 goto abort_with_stats_report;
347 err = gve_adminq_report_stats(priv, priv->stats_report_len,
348 priv->stats_report_bus,
349 GVE_STATS_REPORT_TIMER_PERIOD);
351 dev_err(&priv->pdev->dev,
352 "Failed to report stats: err=%d\n", err);
353 gve_set_device_resources_ok(priv);
355 abort_with_stats_report:
356 gve_free_stats_report(priv);
357 abort_with_ntfy_blocks:
358 gve_free_notify_blocks(priv);
360 gve_free_counter_array(priv);
364 static void gve_trigger_reset(struct gve_priv *priv);
366 static void gve_teardown_device_resources(struct gve_priv *priv)
370 /* Tell device its resources are being freed */
371 if (gve_get_device_resources_ok(priv)) {
372 /* detach the stats report */
373 err = gve_adminq_report_stats(priv, 0, 0x0, GVE_STATS_REPORT_TIMER_PERIOD);
375 dev_err(&priv->pdev->dev,
376 "Failed to detach stats report: err=%d\n", err);
377 gve_trigger_reset(priv);
379 err = gve_adminq_deconfigure_device_resources(priv);
381 dev_err(&priv->pdev->dev,
382 "Could not deconfigure device resources: err=%d\n",
384 gve_trigger_reset(priv);
387 gve_free_counter_array(priv);
388 gve_free_notify_blocks(priv);
389 gve_free_stats_report(priv);
390 gve_clear_device_resources_ok(priv);
393 static void gve_add_napi(struct gve_priv *priv, int ntfy_idx)
395 struct gve_notify_block *block = &priv->ntfy_blocks[ntfy_idx];
397 netif_napi_add(priv->dev, &block->napi, gve_napi_poll,
401 static void gve_remove_napi(struct gve_priv *priv, int ntfy_idx)
403 struct gve_notify_block *block = &priv->ntfy_blocks[ntfy_idx];
405 netif_napi_del(&block->napi);
408 static int gve_register_qpls(struct gve_priv *priv)
410 int num_qpls = gve_num_tx_qpls(priv) + gve_num_rx_qpls(priv);
414 for (i = 0; i < num_qpls; i++) {
415 err = gve_adminq_register_page_list(priv, &priv->qpls[i]);
417 netif_err(priv, drv, priv->dev,
418 "failed to register queue page list %d\n",
420 /* This failure will trigger a reset - no need to clean
429 static int gve_unregister_qpls(struct gve_priv *priv)
431 int num_qpls = gve_num_tx_qpls(priv) + gve_num_rx_qpls(priv);
435 for (i = 0; i < num_qpls; i++) {
436 err = gve_adminq_unregister_page_list(priv, priv->qpls[i].id);
437 /* This failure will trigger a reset - no need to clean up */
439 netif_err(priv, drv, priv->dev,
440 "Failed to unregister queue page list %d\n",
448 static int gve_create_rings(struct gve_priv *priv)
453 err = gve_adminq_create_tx_queues(priv, priv->tx_cfg.num_queues);
455 netif_err(priv, drv, priv->dev, "failed to create %d tx queues\n",
456 priv->tx_cfg.num_queues);
457 /* This failure will trigger a reset - no need to clean
462 netif_dbg(priv, drv, priv->dev, "created %d tx queues\n",
463 priv->tx_cfg.num_queues);
465 err = gve_adminq_create_rx_queues(priv, priv->rx_cfg.num_queues);
467 netif_err(priv, drv, priv->dev, "failed to create %d rx queues\n",
468 priv->rx_cfg.num_queues);
469 /* This failure will trigger a reset - no need to clean
474 netif_dbg(priv, drv, priv->dev, "created %d rx queues\n",
475 priv->rx_cfg.num_queues);
477 /* Rx data ring has been prefilled with packet buffers at queue
479 * Write the doorbell to provide descriptor slots and packet buffers
482 for (i = 0; i < priv->rx_cfg.num_queues; i++)
483 gve_rx_write_doorbell(priv, &priv->rx[i]);
488 static int gve_alloc_rings(struct gve_priv *priv)
495 priv->tx = kvzalloc(priv->tx_cfg.num_queues * sizeof(*priv->tx),
499 err = gve_tx_alloc_rings(priv);
503 priv->rx = kvzalloc(priv->rx_cfg.num_queues * sizeof(*priv->rx),
509 err = gve_rx_alloc_rings(priv);
512 /* Add tx napi & init sync stats*/
513 for (i = 0; i < priv->tx_cfg.num_queues; i++) {
514 u64_stats_init(&priv->tx[i].statss);
515 ntfy_idx = gve_tx_idx_to_ntfy(priv, i);
516 gve_add_napi(priv, ntfy_idx);
518 /* Add rx napi & init sync stats*/
519 for (i = 0; i < priv->rx_cfg.num_queues; i++) {
520 u64_stats_init(&priv->rx[i].statss);
521 ntfy_idx = gve_rx_idx_to_ntfy(priv, i);
522 gve_add_napi(priv, ntfy_idx);
531 gve_tx_free_rings(priv);
538 static int gve_destroy_rings(struct gve_priv *priv)
542 err = gve_adminq_destroy_tx_queues(priv, priv->tx_cfg.num_queues);
544 netif_err(priv, drv, priv->dev,
545 "failed to destroy tx queues\n");
546 /* This failure will trigger a reset - no need to clean up */
549 netif_dbg(priv, drv, priv->dev, "destroyed tx queues\n");
550 err = gve_adminq_destroy_rx_queues(priv, priv->rx_cfg.num_queues);
552 netif_err(priv, drv, priv->dev,
553 "failed to destroy rx queues\n");
554 /* This failure will trigger a reset - no need to clean up */
557 netif_dbg(priv, drv, priv->dev, "destroyed rx queues\n");
561 static void gve_free_rings(struct gve_priv *priv)
567 for (i = 0; i < priv->tx_cfg.num_queues; i++) {
568 ntfy_idx = gve_tx_idx_to_ntfy(priv, i);
569 gve_remove_napi(priv, ntfy_idx);
571 gve_tx_free_rings(priv);
576 for (i = 0; i < priv->rx_cfg.num_queues; i++) {
577 ntfy_idx = gve_rx_idx_to_ntfy(priv, i);
578 gve_remove_napi(priv, ntfy_idx);
580 gve_rx_free_rings(priv);
586 int gve_alloc_page(struct gve_priv *priv, struct device *dev,
587 struct page **page, dma_addr_t *dma,
588 enum dma_data_direction dir)
590 *page = alloc_page(GFP_KERNEL);
592 priv->page_alloc_fail++;
595 *dma = dma_map_page(dev, *page, 0, PAGE_SIZE, dir);
596 if (dma_mapping_error(dev, *dma)) {
597 priv->dma_mapping_error++;
604 static int gve_alloc_queue_page_list(struct gve_priv *priv, u32 id,
607 struct gve_queue_page_list *qpl = &priv->qpls[id];
611 if (pages + priv->num_registered_pages > priv->max_registered_pages) {
612 netif_err(priv, drv, priv->dev,
613 "Reached max number of registered pages %llu > %llu\n",
614 pages + priv->num_registered_pages,
615 priv->max_registered_pages);
620 qpl->num_entries = 0;
621 qpl->pages = kvzalloc(pages * sizeof(*qpl->pages), GFP_KERNEL);
622 /* caller handles clean up */
625 qpl->page_buses = kvzalloc(pages * sizeof(*qpl->page_buses),
627 /* caller handles clean up */
628 if (!qpl->page_buses)
631 for (i = 0; i < pages; i++) {
632 err = gve_alloc_page(priv, &priv->pdev->dev, &qpl->pages[i],
634 gve_qpl_dma_dir(priv, id));
635 /* caller handles clean up */
640 priv->num_registered_pages += pages;
645 void gve_free_page(struct device *dev, struct page *page, dma_addr_t dma,
646 enum dma_data_direction dir)
648 if (!dma_mapping_error(dev, dma))
649 dma_unmap_page(dev, dma, PAGE_SIZE, dir);
654 static void gve_free_queue_page_list(struct gve_priv *priv,
657 struct gve_queue_page_list *qpl = &priv->qpls[id];
662 if (!qpl->page_buses)
665 for (i = 0; i < qpl->num_entries; i++)
666 gve_free_page(&priv->pdev->dev, qpl->pages[i],
667 qpl->page_buses[i], gve_qpl_dma_dir(priv, id));
669 kvfree(qpl->page_buses);
672 priv->num_registered_pages -= qpl->num_entries;
675 static int gve_alloc_qpls(struct gve_priv *priv)
677 int num_qpls = gve_num_tx_qpls(priv) + gve_num_rx_qpls(priv);
681 /* Raw addressing means no QPLs */
682 if (priv->raw_addressing)
685 priv->qpls = kvzalloc(num_qpls * sizeof(*priv->qpls), GFP_KERNEL);
689 for (i = 0; i < gve_num_tx_qpls(priv); i++) {
690 err = gve_alloc_queue_page_list(priv, i,
691 priv->tx_pages_per_qpl);
695 for (; i < num_qpls; i++) {
696 err = gve_alloc_queue_page_list(priv, i,
697 priv->rx_data_slot_cnt);
702 priv->qpl_cfg.qpl_map_size = BITS_TO_LONGS(num_qpls) *
703 sizeof(unsigned long) * BITS_PER_BYTE;
704 priv->qpl_cfg.qpl_id_map = kvzalloc(BITS_TO_LONGS(num_qpls) *
705 sizeof(unsigned long), GFP_KERNEL);
706 if (!priv->qpl_cfg.qpl_id_map) {
714 for (j = 0; j <= i; j++)
715 gve_free_queue_page_list(priv, j);
720 static void gve_free_qpls(struct gve_priv *priv)
722 int num_qpls = gve_num_tx_qpls(priv) + gve_num_rx_qpls(priv);
725 /* Raw addressing means no QPLs */
726 if (priv->raw_addressing)
729 kvfree(priv->qpl_cfg.qpl_id_map);
731 for (i = 0; i < num_qpls; i++)
732 gve_free_queue_page_list(priv, i);
737 /* Use this to schedule a reset when the device is capable of continuing
738 * to handle other requests in its current state. If it is not, do a reset
741 void gve_schedule_reset(struct gve_priv *priv)
743 gve_set_do_reset(priv);
744 queue_work(priv->gve_wq, &priv->service_task);
747 static void gve_reset_and_teardown(struct gve_priv *priv, bool was_up);
748 static int gve_reset_recovery(struct gve_priv *priv, bool was_up);
749 static void gve_turndown(struct gve_priv *priv);
750 static void gve_turnup(struct gve_priv *priv);
752 static int gve_open(struct net_device *dev)
754 struct gve_priv *priv = netdev_priv(dev);
757 err = gve_alloc_qpls(priv);
760 err = gve_alloc_rings(priv);
764 err = netif_set_real_num_tx_queues(dev, priv->tx_cfg.num_queues);
767 err = netif_set_real_num_rx_queues(dev, priv->rx_cfg.num_queues);
771 err = gve_register_qpls(priv);
774 err = gve_create_rings(priv);
777 gve_set_device_rings_ok(priv);
779 if (gve_get_report_stats(priv))
780 mod_timer(&priv->stats_report_timer,
781 round_jiffies(jiffies +
782 msecs_to_jiffies(priv->stats_report_timer_period)));
785 queue_work(priv->gve_wq, &priv->service_task);
786 priv->interface_up_cnt++;
790 gve_free_rings(priv);
796 /* This must have been called from a reset due to the rtnl lock
797 * so just return at this point.
799 if (gve_get_reset_in_progress(priv))
801 /* Otherwise reset before returning */
802 gve_reset_and_teardown(priv, true);
803 /* if this fails there is nothing we can do so just ignore the return */
804 gve_reset_recovery(priv, false);
805 /* return the original error */
809 static int gve_close(struct net_device *dev)
811 struct gve_priv *priv = netdev_priv(dev);
814 netif_carrier_off(dev);
815 if (gve_get_device_rings_ok(priv)) {
817 err = gve_destroy_rings(priv);
820 err = gve_unregister_qpls(priv);
823 gve_clear_device_rings_ok(priv);
825 del_timer_sync(&priv->stats_report_timer);
827 gve_free_rings(priv);
829 priv->interface_down_cnt++;
833 /* This must have been called from a reset due to the rtnl lock
834 * so just return at this point.
836 if (gve_get_reset_in_progress(priv))
838 /* Otherwise reset before returning */
839 gve_reset_and_teardown(priv, true);
840 return gve_reset_recovery(priv, false);
843 int gve_adjust_queues(struct gve_priv *priv,
844 struct gve_queue_config new_rx_config,
845 struct gve_queue_config new_tx_config)
849 if (netif_carrier_ok(priv->dev)) {
850 /* To make this process as simple as possible we teardown the
851 * device, set the new configuration, and then bring the device
854 err = gve_close(priv->dev);
855 /* we have already tried to reset in close,
856 * just fail at this point
860 priv->tx_cfg = new_tx_config;
861 priv->rx_cfg = new_rx_config;
863 err = gve_open(priv->dev);
869 /* Set the config for the next up. */
870 priv->tx_cfg = new_tx_config;
871 priv->rx_cfg = new_rx_config;
875 netif_err(priv, drv, priv->dev,
876 "Adjust queues failed! !!! DISABLING ALL QUEUES !!!\n");
881 static void gve_turndown(struct gve_priv *priv)
885 if (netif_carrier_ok(priv->dev))
886 netif_carrier_off(priv->dev);
888 if (!gve_get_napi_enabled(priv))
891 /* Disable napi to prevent more work from coming in */
892 for (idx = 0; idx < priv->tx_cfg.num_queues; idx++) {
893 int ntfy_idx = gve_tx_idx_to_ntfy(priv, idx);
894 struct gve_notify_block *block = &priv->ntfy_blocks[ntfy_idx];
896 napi_disable(&block->napi);
898 for (idx = 0; idx < priv->rx_cfg.num_queues; idx++) {
899 int ntfy_idx = gve_rx_idx_to_ntfy(priv, idx);
900 struct gve_notify_block *block = &priv->ntfy_blocks[ntfy_idx];
902 napi_disable(&block->napi);
906 netif_tx_disable(priv->dev);
908 gve_clear_napi_enabled(priv);
909 gve_clear_report_stats(priv);
912 static void gve_turnup(struct gve_priv *priv)
916 /* Start the tx queues */
917 netif_tx_start_all_queues(priv->dev);
919 /* Enable napi and unmask interrupts for all queues */
920 for (idx = 0; idx < priv->tx_cfg.num_queues; idx++) {
921 int ntfy_idx = gve_tx_idx_to_ntfy(priv, idx);
922 struct gve_notify_block *block = &priv->ntfy_blocks[ntfy_idx];
924 napi_enable(&block->napi);
925 iowrite32be(0, gve_irq_doorbell(priv, block));
927 for (idx = 0; idx < priv->rx_cfg.num_queues; idx++) {
928 int ntfy_idx = gve_rx_idx_to_ntfy(priv, idx);
929 struct gve_notify_block *block = &priv->ntfy_blocks[ntfy_idx];
931 napi_enable(&block->napi);
932 iowrite32be(0, gve_irq_doorbell(priv, block));
935 gve_set_napi_enabled(priv);
938 static void gve_tx_timeout(struct net_device *dev, unsigned int txqueue)
940 struct gve_priv *priv = netdev_priv(dev);
942 gve_schedule_reset(priv);
943 priv->tx_timeo_cnt++;
946 static const struct net_device_ops gve_netdev_ops = {
947 .ndo_start_xmit = gve_tx,
948 .ndo_open = gve_open,
949 .ndo_stop = gve_close,
950 .ndo_get_stats64 = gve_get_stats,
951 .ndo_tx_timeout = gve_tx_timeout,
954 static void gve_handle_status(struct gve_priv *priv, u32 status)
956 if (GVE_DEVICE_STATUS_RESET_MASK & status) {
957 dev_info(&priv->pdev->dev, "Device requested reset.\n");
958 gve_set_do_reset(priv);
960 if (GVE_DEVICE_STATUS_REPORT_STATS_MASK & status) {
961 priv->stats_report_trigger_cnt++;
962 gve_set_do_report_stats(priv);
966 static void gve_handle_reset(struct gve_priv *priv)
968 /* A service task will be scheduled at the end of probe to catch any
969 * resets that need to happen, and we don't want to reset until
972 if (gve_get_probe_in_progress(priv))
975 if (gve_get_do_reset(priv)) {
977 gve_reset(priv, false);
982 void gve_handle_report_stats(struct gve_priv *priv)
984 int idx, stats_idx = 0, tx_bytes;
985 unsigned int start = 0;
986 struct stats *stats = priv->stats_report->stats;
988 if (!gve_get_report_stats(priv))
991 be64_add_cpu(&priv->stats_report->written_count, 1);
994 for (idx = 0; idx < priv->tx_cfg.num_queues; idx++) {
996 start = u64_stats_fetch_begin(&priv->tx[idx].statss);
997 tx_bytes = priv->tx[idx].bytes_done;
998 } while (u64_stats_fetch_retry(&priv->tx[idx].statss, start));
999 stats[stats_idx++] = (struct stats) {
1000 .stat_name = cpu_to_be32(TX_WAKE_CNT),
1001 .value = cpu_to_be64(priv->tx[idx].wake_queue),
1002 .queue_id = cpu_to_be32(idx),
1004 stats[stats_idx++] = (struct stats) {
1005 .stat_name = cpu_to_be32(TX_STOP_CNT),
1006 .value = cpu_to_be64(priv->tx[idx].stop_queue),
1007 .queue_id = cpu_to_be32(idx),
1009 stats[stats_idx++] = (struct stats) {
1010 .stat_name = cpu_to_be32(TX_FRAMES_SENT),
1011 .value = cpu_to_be64(priv->tx[idx].req),
1012 .queue_id = cpu_to_be32(idx),
1014 stats[stats_idx++] = (struct stats) {
1015 .stat_name = cpu_to_be32(TX_BYTES_SENT),
1016 .value = cpu_to_be64(tx_bytes),
1017 .queue_id = cpu_to_be32(idx),
1019 stats[stats_idx++] = (struct stats) {
1020 .stat_name = cpu_to_be32(TX_LAST_COMPLETION_PROCESSED),
1021 .value = cpu_to_be64(priv->tx[idx].done),
1022 .queue_id = cpu_to_be32(idx),
1028 for (idx = 0; idx < priv->rx_cfg.num_queues; idx++) {
1029 stats[stats_idx++] = (struct stats) {
1030 .stat_name = cpu_to_be32(RX_NEXT_EXPECTED_SEQUENCE),
1031 .value = cpu_to_be64(priv->rx[idx].desc.seqno),
1032 .queue_id = cpu_to_be32(idx),
1034 stats[stats_idx++] = (struct stats) {
1035 .stat_name = cpu_to_be32(RX_BUFFERS_POSTED),
1036 .value = cpu_to_be64(priv->rx[0].fill_cnt),
1037 .queue_id = cpu_to_be32(idx),
1043 static void gve_handle_link_status(struct gve_priv *priv, bool link_status)
1045 if (!gve_get_napi_enabled(priv))
1048 if (link_status == netif_carrier_ok(priv->dev))
1052 netdev_info(priv->dev, "Device link is up.\n");
1053 netif_carrier_on(priv->dev);
1055 netdev_info(priv->dev, "Device link is down.\n");
1056 netif_carrier_off(priv->dev);
1060 /* Handle NIC status register changes, reset requests and report stats */
1061 static void gve_service_task(struct work_struct *work)
1063 struct gve_priv *priv = container_of(work, struct gve_priv,
1065 u32 status = ioread32be(&priv->reg_bar0->device_status);
1067 gve_handle_status(priv, status);
1069 gve_handle_reset(priv);
1070 gve_handle_link_status(priv, GVE_DEVICE_STATUS_LINK_STATUS_MASK & status);
1073 static int gve_init_priv(struct gve_priv *priv, bool skip_describe_device)
1078 /* Set up the adminq */
1079 err = gve_adminq_alloc(&priv->pdev->dev, priv);
1081 dev_err(&priv->pdev->dev,
1082 "Failed to alloc admin queue: err=%d\n", err);
1086 if (skip_describe_device)
1089 priv->raw_addressing = false;
1090 /* Get the initial information we need from the device */
1091 err = gve_adminq_describe_device(priv);
1093 dev_err(&priv->pdev->dev,
1094 "Could not get device information: err=%d\n", err);
1097 if (priv->dev->max_mtu > PAGE_SIZE) {
1098 priv->dev->max_mtu = PAGE_SIZE;
1099 err = gve_adminq_set_mtu(priv, priv->dev->mtu);
1101 dev_err(&priv->pdev->dev, "Could not set mtu");
1105 priv->dev->mtu = priv->dev->max_mtu;
1106 num_ntfy = pci_msix_vec_count(priv->pdev);
1107 if (num_ntfy <= 0) {
1108 dev_err(&priv->pdev->dev,
1109 "could not count MSI-x vectors: err=%d\n", num_ntfy);
1112 } else if (num_ntfy < GVE_MIN_MSIX) {
1113 dev_err(&priv->pdev->dev, "gve needs at least %d MSI-x vectors, but only has %d\n",
1114 GVE_MIN_MSIX, num_ntfy);
1119 priv->num_registered_pages = 0;
1120 priv->rx_copybreak = GVE_DEFAULT_RX_COPYBREAK;
1121 /* gvnic has one Notification Block per MSI-x vector, except for the
1124 priv->num_ntfy_blks = (num_ntfy - 1) & ~0x1;
1125 priv->mgmt_msix_idx = priv->num_ntfy_blks;
1127 priv->tx_cfg.max_queues =
1128 min_t(int, priv->tx_cfg.max_queues, priv->num_ntfy_blks / 2);
1129 priv->rx_cfg.max_queues =
1130 min_t(int, priv->rx_cfg.max_queues, priv->num_ntfy_blks / 2);
1132 priv->tx_cfg.num_queues = priv->tx_cfg.max_queues;
1133 priv->rx_cfg.num_queues = priv->rx_cfg.max_queues;
1134 if (priv->default_num_queues > 0) {
1135 priv->tx_cfg.num_queues = min_t(int, priv->default_num_queues,
1136 priv->tx_cfg.num_queues);
1137 priv->rx_cfg.num_queues = min_t(int, priv->default_num_queues,
1138 priv->rx_cfg.num_queues);
1141 dev_info(&priv->pdev->dev, "TX queues %d, RX queues %d\n",
1142 priv->tx_cfg.num_queues, priv->rx_cfg.num_queues);
1143 dev_info(&priv->pdev->dev, "Max TX queues %d, Max RX queues %d\n",
1144 priv->tx_cfg.max_queues, priv->rx_cfg.max_queues);
1147 err = gve_setup_device_resources(priv);
1151 gve_adminq_free(&priv->pdev->dev, priv);
1155 static void gve_teardown_priv_resources(struct gve_priv *priv)
1157 gve_teardown_device_resources(priv);
1158 gve_adminq_free(&priv->pdev->dev, priv);
1161 static void gve_trigger_reset(struct gve_priv *priv)
1163 /* Reset the device by releasing the AQ */
1164 gve_adminq_release(priv);
1167 static void gve_reset_and_teardown(struct gve_priv *priv, bool was_up)
1169 gve_trigger_reset(priv);
1170 /* With the reset having already happened, close cannot fail */
1172 gve_close(priv->dev);
1173 gve_teardown_priv_resources(priv);
1176 static int gve_reset_recovery(struct gve_priv *priv, bool was_up)
1180 err = gve_init_priv(priv, true);
1184 err = gve_open(priv->dev);
1190 dev_err(&priv->pdev->dev, "Reset failed! !!! DISABLING ALL QUEUES !!!\n");
1195 int gve_reset(struct gve_priv *priv, bool attempt_teardown)
1197 bool was_up = netif_carrier_ok(priv->dev);
1200 dev_info(&priv->pdev->dev, "Performing reset\n");
1201 gve_clear_do_reset(priv);
1202 gve_set_reset_in_progress(priv);
1203 /* If we aren't attempting to teardown normally, just go turndown and
1206 if (!attempt_teardown) {
1208 gve_reset_and_teardown(priv, was_up);
1210 /* Otherwise attempt to close normally */
1212 err = gve_close(priv->dev);
1213 /* If that fails reset as we did above */
1215 gve_reset_and_teardown(priv, was_up);
1217 /* Clean up any remaining resources */
1218 gve_teardown_priv_resources(priv);
1221 /* Set it all back up */
1222 err = gve_reset_recovery(priv, was_up);
1223 gve_clear_reset_in_progress(priv);
1225 priv->interface_up_cnt = 0;
1226 priv->interface_down_cnt = 0;
1227 priv->stats_report_trigger_cnt = 0;
1231 static void gve_write_version(u8 __iomem *driver_version_register)
1233 const char *c = gve_version_prefix;
1236 writeb(*c, driver_version_register);
1240 c = gve_version_str;
1242 writeb(*c, driver_version_register);
1245 writeb('\n', driver_version_register);
1248 static int gve_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
1250 int max_tx_queues, max_rx_queues;
1251 struct net_device *dev;
1252 __be32 __iomem *db_bar;
1253 struct gve_registers __iomem *reg_bar;
1254 struct gve_priv *priv;
1257 err = pci_enable_device(pdev);
1261 err = pci_request_regions(pdev, "gvnic-cfg");
1263 goto abort_with_enabled;
1265 pci_set_master(pdev);
1267 err = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
1269 dev_err(&pdev->dev, "Failed to set dma mask: err=%d\n", err);
1270 goto abort_with_pci_region;
1273 err = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
1276 "Failed to set consistent dma mask: err=%d\n", err);
1277 goto abort_with_pci_region;
1280 reg_bar = pci_iomap(pdev, GVE_REGISTER_BAR, 0);
1282 dev_err(&pdev->dev, "Failed to map pci bar!\n");
1284 goto abort_with_pci_region;
1287 db_bar = pci_iomap(pdev, GVE_DOORBELL_BAR, 0);
1289 dev_err(&pdev->dev, "Failed to map doorbell bar!\n");
1291 goto abort_with_reg_bar;
1294 gve_write_version(®_bar->driver_version);
1295 /* Get max queues to alloc etherdev */
1296 max_rx_queues = ioread32be(®_bar->max_tx_queues);
1297 max_tx_queues = ioread32be(®_bar->max_rx_queues);
1298 /* Alloc and setup the netdev and priv */
1299 dev = alloc_etherdev_mqs(sizeof(*priv), max_tx_queues, max_rx_queues);
1301 dev_err(&pdev->dev, "could not allocate netdev\n");
1302 goto abort_with_db_bar;
1304 SET_NETDEV_DEV(dev, &pdev->dev);
1305 pci_set_drvdata(pdev, dev);
1306 dev->ethtool_ops = &gve_ethtool_ops;
1307 dev->netdev_ops = &gve_netdev_ops;
1308 /* advertise features */
1309 dev->hw_features = NETIF_F_HIGHDMA;
1310 dev->hw_features |= NETIF_F_SG;
1311 dev->hw_features |= NETIF_F_HW_CSUM;
1312 dev->hw_features |= NETIF_F_TSO;
1313 dev->hw_features |= NETIF_F_TSO6;
1314 dev->hw_features |= NETIF_F_TSO_ECN;
1315 dev->hw_features |= NETIF_F_RXCSUM;
1316 dev->hw_features |= NETIF_F_RXHASH;
1317 dev->features = dev->hw_features;
1318 dev->watchdog_timeo = 5 * HZ;
1319 dev->min_mtu = ETH_MIN_MTU;
1320 netif_carrier_off(dev);
1322 priv = netdev_priv(dev);
1325 priv->msg_enable = DEFAULT_MSG_LEVEL;
1326 priv->reg_bar0 = reg_bar;
1327 priv->db_bar2 = db_bar;
1328 priv->service_task_flags = 0x0;
1329 priv->state_flags = 0x0;
1330 priv->ethtool_flags = 0x0;
1332 gve_set_probe_in_progress(priv);
1333 priv->gve_wq = alloc_ordered_workqueue("gve", 0);
1334 if (!priv->gve_wq) {
1335 dev_err(&pdev->dev, "Could not allocate workqueue");
1337 goto abort_with_netdev;
1339 INIT_WORK(&priv->service_task, gve_service_task);
1340 INIT_WORK(&priv->stats_report_task, gve_stats_report_task);
1341 priv->tx_cfg.max_queues = max_tx_queues;
1342 priv->rx_cfg.max_queues = max_rx_queues;
1344 err = gve_init_priv(priv, false);
1348 err = register_netdev(dev);
1352 dev_info(&pdev->dev, "GVE version %s\n", gve_version_str);
1353 gve_clear_probe_in_progress(priv);
1354 queue_work(priv->gve_wq, &priv->service_task);
1358 destroy_workqueue(priv->gve_wq);
1364 pci_iounmap(pdev, db_bar);
1367 pci_iounmap(pdev, reg_bar);
1369 abort_with_pci_region:
1370 pci_release_regions(pdev);
1373 pci_disable_device(pdev);
1377 static void gve_remove(struct pci_dev *pdev)
1379 struct net_device *netdev = pci_get_drvdata(pdev);
1380 struct gve_priv *priv = netdev_priv(netdev);
1381 __be32 __iomem *db_bar = priv->db_bar2;
1382 void __iomem *reg_bar = priv->reg_bar0;
1384 unregister_netdev(netdev);
1385 gve_teardown_priv_resources(priv);
1386 destroy_workqueue(priv->gve_wq);
1387 free_netdev(netdev);
1388 pci_iounmap(pdev, db_bar);
1389 pci_iounmap(pdev, reg_bar);
1390 pci_release_regions(pdev);
1391 pci_disable_device(pdev);
1394 static const struct pci_device_id gve_id_table[] = {
1395 { PCI_DEVICE(PCI_VENDOR_ID_GOOGLE, PCI_DEV_ID_GVNIC) },
1399 static struct pci_driver gvnic_driver = {
1401 .id_table = gve_id_table,
1403 .remove = gve_remove,
1406 module_pci_driver(gvnic_driver);
1408 MODULE_DEVICE_TABLE(pci, gve_id_table);
1409 MODULE_AUTHOR("Google, Inc.");
1410 MODULE_DESCRIPTION("gVNIC Driver");
1411 MODULE_LICENSE("Dual MIT/GPL");
1412 MODULE_VERSION(GVE_VERSION);