1 // SPDX-License-Identifier: GPL-2.0-only
2 /****************************************************************************
3 * Driver for Solarflare network controllers and boards
4 * Copyright 2005-2006 Fen Systems Ltd.
5 * Copyright 2006-2012 Solarflare Communications Inc.
8 #include <linux/netdevice.h>
9 #include <linux/module.h>
10 #include <linux/delay.h>
11 #include <linux/kernel_stat.h>
12 #include <linux/pci.h>
13 #include <linux/ethtool.h>
16 #include <linux/udp.h>
17 #include <linux/rtnetlink.h>
18 #include <linux/slab.h>
19 #include "net_driver.h"
21 #include "efx_common.h"
22 #include "efx_channels.h"
25 #include "workarounds.h"
27 /* IRQ latency can be enormous because:
28 * - All IRQs may be disabled on a CPU for a *long* time by e.g. a
29 * slow serial console or an old IDE driver doing error recovery
30 * - The PREEMPT_RT patches mostly deal with this, but also allow a
31 * tasklet or normal task to be given higher priority than our IRQ
33 * Try to avoid blaming the hardware for this.
35 #define IRQ_TIMEOUT HZ
38 * Loopback test packet structure
40 * The self-test should stress every RSS vector, and unfortunately
41 * Falcon only performs RSS on TCP/UDP packets.
43 struct efx_loopback_payload {
51 /* Loopback test source MAC address */
52 static const u8 payload_source[ETH_ALEN] __aligned(2) = {
53 0x00, 0x0f, 0x53, 0x1b, 0x1b, 0x1b,
56 static const char payload_msg[] =
57 "Hello world! This is an Efx loopback test in progress!";
59 /* Interrupt mode names */
60 static const unsigned int efx_interrupt_mode_max = EFX_INT_MODE_MAX;
61 static const char *const efx_interrupt_mode_names[] = {
62 [EFX_INT_MODE_MSIX] = "MSI-X",
63 [EFX_INT_MODE_MSI] = "MSI",
64 [EFX_INT_MODE_LEGACY] = "legacy",
66 #define INT_MODE(efx) \
67 STRING_TABLE_LOOKUP(efx->interrupt_mode, efx_interrupt_mode)
70 * efx_loopback_state - persistent state during a loopback selftest
71 * @flush: Drop all packets in efx_loopback_rx_packet
72 * @packet_count: Number of packets being used in this test
73 * @skbs: An array of skbs transmitted
74 * @offload_csum: Checksums are being offloaded
75 * @rx_good: RX good packet count
76 * @rx_bad: RX bad packet count
77 * @payload: Payload used in tests
79 struct efx_loopback_state {
82 struct sk_buff **skbs;
86 struct efx_loopback_payload payload;
89 /* How long to wait for all the packets to arrive (in ms) */
90 #define LOOPBACK_TIMEOUT_MS 1000
92 /**************************************************************************
94 * MII, NVRAM and register tests
96 **************************************************************************/
98 static int efx_test_phy_alive(struct efx_nic *efx, struct efx_self_tests *tests)
102 if (efx->phy_op->test_alive) {
103 rc = efx->phy_op->test_alive(efx);
104 tests->phy_alive = rc ? -1 : 1;
110 static int efx_test_nvram(struct efx_nic *efx, struct efx_self_tests *tests)
114 if (efx->type->test_nvram) {
115 rc = efx->type->test_nvram(efx);
119 tests->nvram = rc ? -1 : 1;
125 /**************************************************************************
127 * Interrupt and event queue testing
129 **************************************************************************/
131 /* Test generation and receipt of interrupts */
132 static int efx_test_interrupts(struct efx_nic *efx,
133 struct efx_self_tests *tests)
135 unsigned long timeout, wait;
139 netif_dbg(efx, drv, efx->net_dev, "testing interrupts\n");
140 tests->interrupt = -1;
142 rc = efx_nic_irq_test_start(efx);
143 if (rc == -ENOTSUPP) {
144 netif_dbg(efx, drv, efx->net_dev,
145 "direct interrupt testing not supported\n");
146 tests->interrupt = 0;
150 timeout = jiffies + IRQ_TIMEOUT;
153 /* Wait for arrival of test interrupt. */
154 netif_dbg(efx, drv, efx->net_dev, "waiting for test interrupt\n");
156 schedule_timeout_uninterruptible(wait);
157 cpu = efx_nic_irq_test_irq_cpu(efx);
161 } while (time_before(jiffies, timeout));
163 netif_err(efx, drv, efx->net_dev, "timed out waiting for interrupt\n");
167 netif_dbg(efx, drv, efx->net_dev, "%s test interrupt seen on CPU%d\n",
169 tests->interrupt = 1;
173 /* Test generation and receipt of interrupting events */
174 static int efx_test_eventq_irq(struct efx_nic *efx,
175 struct efx_self_tests *tests)
177 struct efx_channel *channel;
178 unsigned int read_ptr[EFX_MAX_CHANNELS];
179 unsigned long napi_ran = 0, dma_pend = 0, int_pend = 0;
180 unsigned long timeout, wait;
182 BUILD_BUG_ON(EFX_MAX_CHANNELS > BITS_PER_LONG);
184 efx_for_each_channel(channel, efx) {
185 read_ptr[channel->channel] = channel->eventq_read_ptr;
186 set_bit(channel->channel, &dma_pend);
187 set_bit(channel->channel, &int_pend);
188 efx_nic_event_test_start(channel);
191 timeout = jiffies + IRQ_TIMEOUT;
194 /* Wait for arrival of interrupts. NAPI processing may or may
195 * not complete in time, but we can cope in any case.
198 schedule_timeout_uninterruptible(wait);
200 efx_for_each_channel(channel, efx) {
201 efx_stop_eventq(channel);
202 if (channel->eventq_read_ptr !=
203 read_ptr[channel->channel]) {
204 set_bit(channel->channel, &napi_ran);
205 clear_bit(channel->channel, &dma_pend);
206 clear_bit(channel->channel, &int_pend);
208 if (efx_nic_event_present(channel))
209 clear_bit(channel->channel, &dma_pend);
210 if (efx_nic_event_test_irq_cpu(channel) >= 0)
211 clear_bit(channel->channel, &int_pend);
213 efx_start_eventq(channel);
217 } while ((dma_pend || int_pend) && time_before(jiffies, timeout));
219 efx_for_each_channel(channel, efx) {
220 bool dma_seen = !test_bit(channel->channel, &dma_pend);
221 bool int_seen = !test_bit(channel->channel, &int_pend);
223 tests->eventq_dma[channel->channel] = dma_seen ? 1 : -1;
224 tests->eventq_int[channel->channel] = int_seen ? 1 : -1;
226 if (dma_seen && int_seen) {
227 netif_dbg(efx, drv, efx->net_dev,
228 "channel %d event queue passed (with%s NAPI)\n",
230 test_bit(channel->channel, &napi_ran) ?
233 /* Report failure and whether either interrupt or DMA
236 netif_err(efx, drv, efx->net_dev,
237 "channel %d timed out waiting for event queue\n",
240 netif_err(efx, drv, efx->net_dev,
241 "channel %d saw interrupt "
242 "during event queue test\n",
245 netif_err(efx, drv, efx->net_dev,
246 "channel %d event was generated, but "
247 "failed to trigger an interrupt\n",
252 return (dma_pend || int_pend) ? -ETIMEDOUT : 0;
255 static int efx_test_phy(struct efx_nic *efx, struct efx_self_tests *tests,
260 if (!efx->phy_op->run_tests)
263 mutex_lock(&efx->mac_lock);
264 rc = efx->phy_op->run_tests(efx, tests->phy_ext, flags);
265 mutex_unlock(&efx->mac_lock);
269 netif_info(efx, drv, efx->net_dev,
270 "%s phy selftest\n", rc ? "Failed" : "Passed");
275 /**************************************************************************
278 * NB Only one loopback test can be executing concurrently.
280 **************************************************************************/
282 /* Loopback test RX callback
283 * This is called for each received packet during loopback testing.
285 void efx_loopback_rx_packet(struct efx_nic *efx,
286 const char *buf_ptr, int pkt_len)
288 struct efx_loopback_state *state = efx->loopback_selftest;
289 struct efx_loopback_payload *received;
290 struct efx_loopback_payload *payload;
294 /* If we are just flushing, then drop the packet */
295 if ((state == NULL) || state->flush)
298 payload = &state->payload;
300 received = (struct efx_loopback_payload *) buf_ptr;
301 received->ip.saddr = payload->ip.saddr;
302 if (state->offload_csum)
303 received->ip.check = payload->ip.check;
305 /* Check that header exists */
306 if (pkt_len < sizeof(received->header)) {
307 netif_err(efx, drv, efx->net_dev,
308 "saw runt RX packet (length %d) in %s loopback "
309 "test\n", pkt_len, LOOPBACK_MODE(efx));
313 /* Check that the ethernet header exists */
314 if (memcmp(&received->header, &payload->header, ETH_HLEN) != 0) {
315 netif_err(efx, drv, efx->net_dev,
316 "saw non-loopback RX packet in %s loopback test\n",
321 /* Check packet length */
322 if (pkt_len != sizeof(*payload)) {
323 netif_err(efx, drv, efx->net_dev,
324 "saw incorrect RX packet length %d (wanted %d) in "
325 "%s loopback test\n", pkt_len, (int)sizeof(*payload),
330 /* Check that IP header matches */
331 if (memcmp(&received->ip, &payload->ip, sizeof(payload->ip)) != 0) {
332 netif_err(efx, drv, efx->net_dev,
333 "saw corrupted IP header in %s loopback test\n",
338 /* Check that msg and padding matches */
339 if (memcmp(&received->msg, &payload->msg, sizeof(received->msg)) != 0) {
340 netif_err(efx, drv, efx->net_dev,
341 "saw corrupted RX packet in %s loopback test\n",
346 /* Check that iteration matches */
347 if (received->iteration != payload->iteration) {
348 netif_err(efx, drv, efx->net_dev,
349 "saw RX packet from iteration %d (wanted %d) in "
350 "%s loopback test\n", ntohs(received->iteration),
351 ntohs(payload->iteration), LOOPBACK_MODE(efx));
355 /* Increase correct RX count */
356 netif_vdbg(efx, drv, efx->net_dev,
357 "got loopback RX in %s loopback test\n", LOOPBACK_MODE(efx));
359 atomic_inc(&state->rx_good);
364 if (atomic_read(&state->rx_bad) == 0) {
365 netif_err(efx, drv, efx->net_dev, "received packet:\n");
366 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 0x10, 1,
367 buf_ptr, pkt_len, 0);
368 netif_err(efx, drv, efx->net_dev, "expected packet:\n");
369 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 0x10, 1,
370 &state->payload, sizeof(state->payload), 0);
373 atomic_inc(&state->rx_bad);
376 /* Initialise an efx_selftest_state for a new iteration */
377 static void efx_iterate_state(struct efx_nic *efx)
379 struct efx_loopback_state *state = efx->loopback_selftest;
380 struct net_device *net_dev = efx->net_dev;
381 struct efx_loopback_payload *payload = &state->payload;
383 /* Initialise the layerII header */
384 ether_addr_copy((u8 *)&payload->header.h_dest, net_dev->dev_addr);
385 ether_addr_copy((u8 *)&payload->header.h_source, payload_source);
386 payload->header.h_proto = htons(ETH_P_IP);
388 /* saddr set later and used as incrementing count */
389 payload->ip.daddr = htonl(INADDR_LOOPBACK);
391 payload->ip.check = (__force __sum16) htons(0xdead);
392 payload->ip.tot_len = htons(sizeof(*payload) - sizeof(struct ethhdr));
393 payload->ip.version = IPVERSION;
394 payload->ip.protocol = IPPROTO_UDP;
396 /* Initialise udp header */
397 payload->udp.source = 0;
398 payload->udp.len = htons(sizeof(*payload) - sizeof(struct ethhdr) -
399 sizeof(struct iphdr));
400 payload->udp.check = 0; /* checksum ignored */
402 /* Fill out payload */
403 payload->iteration = htons(ntohs(payload->iteration) + 1);
404 memcpy(&payload->msg, payload_msg, sizeof(payload_msg));
406 /* Fill out remaining state members */
407 atomic_set(&state->rx_good, 0);
408 atomic_set(&state->rx_bad, 0);
412 static int efx_begin_loopback(struct efx_tx_queue *tx_queue)
414 struct efx_nic *efx = tx_queue->efx;
415 struct efx_loopback_state *state = efx->loopback_selftest;
416 struct efx_loopback_payload *payload;
421 /* Transmit N copies of buffer */
422 for (i = 0; i < state->packet_count; i++) {
423 /* Allocate an skb, holding an extra reference for
424 * transmit completion counting */
425 skb = alloc_skb(sizeof(state->payload), GFP_KERNEL);
428 state->skbs[i] = skb;
431 /* Copy the payload in, incrementing the source address to
432 * exercise the rss vectors */
433 payload = skb_put(skb, sizeof(state->payload));
434 memcpy(payload, &state->payload, sizeof(state->payload));
435 payload->ip.saddr = htonl(INADDR_LOOPBACK | (i << 2));
437 /* Ensure everything we've written is visible to the
438 * interrupt handler. */
441 netif_tx_lock_bh(efx->net_dev);
442 rc = efx_enqueue_skb(tx_queue, skb);
443 netif_tx_unlock_bh(efx->net_dev);
445 if (rc != NETDEV_TX_OK) {
446 netif_err(efx, drv, efx->net_dev,
447 "TX queue %d could not transmit packet %d of "
448 "%d in %s loopback test\n", tx_queue->label,
449 i + 1, state->packet_count,
452 /* Defer cleaning up the other skbs for the caller */
461 static int efx_poll_loopback(struct efx_nic *efx)
463 struct efx_loopback_state *state = efx->loopback_selftest;
465 return atomic_read(&state->rx_good) == state->packet_count;
468 static int efx_end_loopback(struct efx_tx_queue *tx_queue,
469 struct efx_loopback_self_tests *lb_tests)
471 struct efx_nic *efx = tx_queue->efx;
472 struct efx_loopback_state *state = efx->loopback_selftest;
474 int tx_done = 0, rx_good, rx_bad;
477 netif_tx_lock_bh(efx->net_dev);
479 /* Count the number of tx completions, and decrement the refcnt. Any
480 * skbs not already completed will be free'd when the queue is flushed */
481 for (i = 0; i < state->packet_count; i++) {
482 skb = state->skbs[i];
483 if (skb && !skb_shared(skb))
488 netif_tx_unlock_bh(efx->net_dev);
490 /* Check TX completion and received packet counts */
491 rx_good = atomic_read(&state->rx_good);
492 rx_bad = atomic_read(&state->rx_bad);
493 if (tx_done != state->packet_count) {
494 /* Don't free the skbs; they will be picked up on TX
495 * overflow or channel teardown.
497 netif_err(efx, drv, efx->net_dev,
498 "TX queue %d saw only %d out of an expected %d "
499 "TX completion events in %s loopback test\n",
500 tx_queue->label, tx_done, state->packet_count,
503 /* Allow to fall through so we see the RX errors as well */
506 /* We may always be up to a flush away from our desired packet total */
507 if (rx_good != state->packet_count) {
508 netif_dbg(efx, drv, efx->net_dev,
509 "TX queue %d saw only %d out of an expected %d "
510 "received packets in %s loopback test\n",
511 tx_queue->label, rx_good, state->packet_count,
517 /* Update loopback test structure */
518 lb_tests->tx_sent[tx_queue->label] += state->packet_count;
519 lb_tests->tx_done[tx_queue->label] += tx_done;
520 lb_tests->rx_good += rx_good;
521 lb_tests->rx_bad += rx_bad;
527 efx_test_loopback(struct efx_tx_queue *tx_queue,
528 struct efx_loopback_self_tests *lb_tests)
530 struct efx_nic *efx = tx_queue->efx;
531 struct efx_loopback_state *state = efx->loopback_selftest;
532 int i, begin_rc, end_rc;
534 for (i = 0; i < 3; i++) {
535 /* Determine how many packets to send */
536 state->packet_count = efx->txq_entries / 3;
537 state->packet_count = min(1 << (i << 2), state->packet_count);
538 state->skbs = kcalloc(state->packet_count,
539 sizeof(state->skbs[0]), GFP_KERNEL);
542 state->flush = false;
544 netif_dbg(efx, drv, efx->net_dev,
545 "TX queue %d (hw %d) testing %s loopback with %d packets\n",
546 tx_queue->label, tx_queue->queue, LOOPBACK_MODE(efx),
547 state->packet_count);
549 efx_iterate_state(efx);
550 begin_rc = efx_begin_loopback(tx_queue);
552 /* This will normally complete very quickly, but be
553 * prepared to wait much longer. */
555 if (!efx_poll_loopback(efx)) {
556 msleep(LOOPBACK_TIMEOUT_MS);
557 efx_poll_loopback(efx);
560 end_rc = efx_end_loopback(tx_queue, lb_tests);
563 if (begin_rc || end_rc) {
564 /* Wait a while to ensure there are no packets
565 * floating around after a failure. */
566 schedule_timeout_uninterruptible(HZ / 10);
567 return begin_rc ? begin_rc : end_rc;
571 netif_dbg(efx, drv, efx->net_dev,
572 "TX queue %d passed %s loopback test with a burst length "
573 "of %d packets\n", tx_queue->label, LOOPBACK_MODE(efx),
574 state->packet_count);
579 /* Wait for link up. On Falcon, we would prefer to rely on efx_monitor, but
580 * any contention on the mac lock (via e.g. efx_mac_mcast_work) causes it
581 * to delay and retry. Therefore, it's safer to just poll directly. Wait
582 * for link up and any faults to dissipate. */
583 static int efx_wait_for_link(struct efx_nic *efx)
585 struct efx_link_state *link_state = &efx->link_state;
586 int count, link_up_count = 0;
589 for (count = 0; count < 40; count++) {
590 schedule_timeout_uninterruptible(HZ / 10);
592 if (efx->type->monitor != NULL) {
593 mutex_lock(&efx->mac_lock);
594 efx->type->monitor(efx);
595 mutex_unlock(&efx->mac_lock);
598 mutex_lock(&efx->mac_lock);
599 link_up = link_state->up;
601 link_up = !efx->type->check_mac_fault(efx);
602 mutex_unlock(&efx->mac_lock);
605 if (++link_up_count == 2)
615 static int efx_test_loopbacks(struct efx_nic *efx, struct efx_self_tests *tests,
616 unsigned int loopback_modes)
618 enum efx_loopback_mode mode;
619 struct efx_loopback_state *state;
620 struct efx_channel *channel =
621 efx_get_channel(efx, efx->tx_channel_offset);
622 struct efx_tx_queue *tx_queue;
625 /* Set the port loopback_selftest member. From this point on
626 * all received packets will be dropped. Mark the state as
627 * "flushing" so all inflight packets are dropped */
628 state = kzalloc(sizeof(*state), GFP_KERNEL);
631 BUG_ON(efx->loopback_selftest);
633 efx->loopback_selftest = state;
635 /* Test all supported loopback modes */
636 for (mode = LOOPBACK_NONE; mode <= LOOPBACK_TEST_MAX; mode++) {
637 if (!(loopback_modes & (1 << mode)))
640 /* Move the port into the specified loopback mode. */
642 mutex_lock(&efx->mac_lock);
643 efx->loopback_mode = mode;
644 rc = __efx_reconfigure_port(efx);
645 mutex_unlock(&efx->mac_lock);
647 netif_err(efx, drv, efx->net_dev,
648 "unable to move into %s loopback\n",
653 rc = efx_wait_for_link(efx);
655 netif_err(efx, drv, efx->net_dev,
656 "loopback %s never came up\n",
661 /* Test all enabled types of TX queue */
662 efx_for_each_channel_tx_queue(tx_queue, channel) {
663 state->offload_csum = (tx_queue->label &
664 EFX_TXQ_TYPE_OFFLOAD);
665 rc = efx_test_loopback(tx_queue,
666 &tests->loopback[mode]);
673 /* Remove the flush. The caller will remove the loopback setting */
675 efx->loopback_selftest = NULL;
685 /**************************************************************************
689 *************************************************************************/
691 int efx_selftest(struct efx_nic *efx, struct efx_self_tests *tests,
694 enum efx_loopback_mode loopback_mode = efx->loopback_mode;
695 int phy_mode = efx->phy_mode;
696 int rc_test = 0, rc_reset, rc;
698 efx_selftest_async_cancel(efx);
700 /* Online (i.e. non-disruptive) testing
701 * This checks interrupt generation, event delivery and PHY presence. */
703 rc = efx_test_phy_alive(efx, tests);
707 rc = efx_test_nvram(efx, tests);
711 rc = efx_test_interrupts(efx, tests);
715 rc = efx_test_eventq_irq(efx, tests);
722 if (!(flags & ETH_TEST_FL_OFFLINE))
723 return efx_test_phy(efx, tests, flags);
725 /* Offline (i.e. disruptive) testing
726 * This checks MAC and PHY loopback on the specified port. */
728 /* Detach the device so the kernel doesn't transmit during the
729 * loopback test and the watchdog timeout doesn't fire.
731 efx_device_detach_sync(efx);
733 if (efx->type->test_chip) {
734 rc_reset = efx->type->test_chip(efx, tests);
736 netif_err(efx, hw, efx->net_dev,
737 "Unable to recover from chip test\n");
738 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
742 if ((tests->memory < 0 || tests->registers < 0) && !rc_test)
746 /* Ensure that the phy is powered and out of loopback
747 * for the bist and loopback tests */
748 mutex_lock(&efx->mac_lock);
749 efx->phy_mode &= ~PHY_MODE_LOW_POWER;
750 efx->loopback_mode = LOOPBACK_NONE;
751 __efx_reconfigure_port(efx);
752 mutex_unlock(&efx->mac_lock);
754 rc = efx_test_phy(efx, tests, flags);
758 rc = efx_test_loopbacks(efx, tests, efx->loopback_modes);
762 /* restore the PHY to the previous state */
763 mutex_lock(&efx->mac_lock);
764 efx->phy_mode = phy_mode;
765 efx->loopback_mode = loopback_mode;
766 __efx_reconfigure_port(efx);
767 mutex_unlock(&efx->mac_lock);
769 efx_device_attach_if_not_resetting(efx);
774 void efx_selftest_async_start(struct efx_nic *efx)
776 struct efx_channel *channel;
778 efx_for_each_channel(channel, efx)
779 efx_nic_event_test_start(channel);
780 schedule_delayed_work(&efx->selftest_work, IRQ_TIMEOUT);
783 void efx_selftest_async_cancel(struct efx_nic *efx)
785 cancel_delayed_work_sync(&efx->selftest_work);
788 static void efx_selftest_async_work(struct work_struct *data)
790 struct efx_nic *efx = container_of(data, struct efx_nic,
792 struct efx_channel *channel;
795 efx_for_each_channel(channel, efx) {
796 cpu = efx_nic_event_test_irq_cpu(channel);
798 netif_err(efx, ifup, efx->net_dev,
799 "channel %d failed to trigger an interrupt\n",
802 netif_dbg(efx, ifup, efx->net_dev,
803 "channel %d triggered interrupt on CPU %d\n",
804 channel->channel, cpu);
808 void efx_selftest_async_init(struct efx_nic *efx)
810 INIT_DELAYED_WORK(&efx->selftest_work, efx_selftest_async_work);