1 /****************************************************************************
2 * Driver for Solarflare Solarstorm network controllers and boards
3 * Copyright 2005-2006 Fen Systems Ltd.
4 * Copyright 2006-2010 Solarflare Communications Inc.
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation, incorporated herein by reference.
11 #include <linux/netdevice.h>
12 #include <linux/module.h>
13 #include <linux/delay.h>
14 #include <linux/kernel_stat.h>
15 #include <linux/pci.h>
16 #include <linux/ethtool.h>
19 #include <linux/udp.h>
20 #include <linux/rtnetlink.h>
21 #include <linux/slab.h>
22 #include "net_driver.h"
26 #include "workarounds.h"
28 /* IRQ latency can be enormous because:
29 * - All IRQs may be disabled on a CPU for a *long* time by e.g. a
30 * slow serial console or an old IDE driver doing error recovery
31 * - The PREEMPT_RT patches mostly deal with this, but also allow a
32 * tasklet or normal task to be given higher priority than our IRQ
34 * Try to avoid blaming the hardware for this.
36 #define IRQ_TIMEOUT HZ
39 * Loopback test packet structure
41 * The self-test should stress every RSS vector, and unfortunately
42 * Falcon only performs RSS on TCP/UDP packets.
44 struct efx_loopback_payload {
52 /* Loopback test source MAC address */
53 static const unsigned char payload_source[ETH_ALEN] = {
54 0x00, 0x0f, 0x53, 0x1b, 0x1b, 0x1b,
57 static const char payload_msg[] =
58 "Hello world! This is an Efx loopback test in progress!";
60 /* Interrupt mode names */
61 static const unsigned int efx_interrupt_mode_max = EFX_INT_MODE_MAX;
62 static const char *const efx_interrupt_mode_names[] = {
63 [EFX_INT_MODE_MSIX] = "MSI-X",
64 [EFX_INT_MODE_MSI] = "MSI",
65 [EFX_INT_MODE_LEGACY] = "legacy",
67 #define INT_MODE(efx) \
68 STRING_TABLE_LOOKUP(efx->interrupt_mode, efx_interrupt_mode)
71 * efx_loopback_state - persistent state during a loopback selftest
72 * @flush: Drop all packets in efx_loopback_rx_packet
73 * @packet_count: Number of packets being used in this test
74 * @skbs: An array of skbs transmitted
75 * @offload_csum: Checksums are being offloaded
76 * @rx_good: RX good packet count
77 * @rx_bad: RX bad packet count
78 * @payload: Payload used in tests
80 struct efx_loopback_state {
83 struct sk_buff **skbs;
87 struct efx_loopback_payload payload;
90 /* How long to wait for all the packets to arrive (in ms) */
91 #define LOOPBACK_TIMEOUT_MS 1000
93 /**************************************************************************
95 * MII, NVRAM and register tests
97 **************************************************************************/
99 static int efx_test_phy_alive(struct efx_nic *efx, struct efx_self_tests *tests)
103 if (efx->phy_op->test_alive) {
104 rc = efx->phy_op->test_alive(efx);
105 tests->phy_alive = rc ? -1 : 1;
111 static int efx_test_nvram(struct efx_nic *efx, struct efx_self_tests *tests)
115 if (efx->type->test_nvram) {
116 rc = efx->type->test_nvram(efx);
117 tests->nvram = rc ? -1 : 1;
123 static int efx_test_chip(struct efx_nic *efx, struct efx_self_tests *tests)
127 /* Test register access */
128 if (efx->type->test_registers) {
129 rc = efx->type->test_registers(efx);
130 tests->registers = rc ? -1 : 1;
136 /**************************************************************************
138 * Interrupt and event queue testing
140 **************************************************************************/
142 /* Test generation and receipt of interrupts */
143 static int efx_test_interrupts(struct efx_nic *efx,
144 struct efx_self_tests *tests)
146 unsigned long timeout, wait;
149 netif_dbg(efx, drv, efx->net_dev, "testing interrupts\n");
150 tests->interrupt = -1;
152 efx_nic_irq_test_start(efx);
153 timeout = jiffies + IRQ_TIMEOUT;
156 /* Wait for arrival of test interrupt. */
157 netif_dbg(efx, drv, efx->net_dev, "waiting for test interrupt\n");
159 schedule_timeout_uninterruptible(wait);
160 cpu = efx_nic_irq_test_irq_cpu(efx);
164 } while (time_before(jiffies, timeout));
166 netif_err(efx, drv, efx->net_dev, "timed out waiting for interrupt\n");
170 netif_dbg(efx, drv, efx->net_dev, "%s test interrupt seen on CPU%d\n",
172 tests->interrupt = 1;
176 /* Test generation and receipt of interrupting events */
177 static int efx_test_eventq_irq(struct efx_nic *efx,
178 struct efx_self_tests *tests)
180 struct efx_channel *channel;
181 unsigned int read_ptr[EFX_MAX_CHANNELS];
182 unsigned long napi_ran = 0, dma_pend = 0, int_pend = 0;
183 unsigned long timeout, wait;
185 BUILD_BUG_ON(EFX_MAX_CHANNELS > BITS_PER_LONG);
187 efx_for_each_channel(channel, efx) {
188 read_ptr[channel->channel] = channel->eventq_read_ptr;
189 set_bit(channel->channel, &dma_pend);
190 set_bit(channel->channel, &int_pend);
191 efx_nic_event_test_start(channel);
194 timeout = jiffies + IRQ_TIMEOUT;
197 /* Wait for arrival of interrupts. NAPI processing may or may
198 * not complete in time, but we can cope in any case.
201 schedule_timeout_uninterruptible(wait);
203 efx_for_each_channel(channel, efx) {
204 napi_disable(&channel->napi_str);
205 if (channel->eventq_read_ptr !=
206 read_ptr[channel->channel]) {
207 set_bit(channel->channel, &napi_ran);
208 clear_bit(channel->channel, &dma_pend);
209 clear_bit(channel->channel, &int_pend);
211 if (efx_nic_event_present(channel))
212 clear_bit(channel->channel, &dma_pend);
213 if (efx_nic_event_test_irq_cpu(channel) >= 0)
214 clear_bit(channel->channel, &int_pend);
216 napi_enable(&channel->napi_str);
217 efx_nic_eventq_read_ack(channel);
221 } while ((dma_pend || int_pend) && time_before(jiffies, timeout));
223 efx_for_each_channel(channel, efx) {
224 bool dma_seen = !test_bit(channel->channel, &dma_pend);
225 bool int_seen = !test_bit(channel->channel, &int_pend);
227 tests->eventq_dma[channel->channel] = dma_seen ? 1 : -1;
228 tests->eventq_int[channel->channel] = int_seen ? 1 : -1;
230 if (dma_seen && int_seen) {
231 netif_dbg(efx, drv, efx->net_dev,
232 "channel %d event queue passed (with%s NAPI)\n",
234 test_bit(channel->channel, &napi_ran) ?
237 /* Report failure and whether either interrupt or DMA
240 netif_err(efx, drv, efx->net_dev,
241 "channel %d timed out waiting for event queue\n",
244 netif_err(efx, drv, efx->net_dev,
245 "channel %d saw interrupt "
246 "during event queue test\n",
249 netif_err(efx, drv, efx->net_dev,
250 "channel %d event was generated, but "
251 "failed to trigger an interrupt\n",
256 return (dma_pend || int_pend) ? -ETIMEDOUT : 0;
259 static int efx_test_phy(struct efx_nic *efx, struct efx_self_tests *tests,
264 if (!efx->phy_op->run_tests)
267 mutex_lock(&efx->mac_lock);
268 rc = efx->phy_op->run_tests(efx, tests->phy_ext, flags);
269 mutex_unlock(&efx->mac_lock);
273 /**************************************************************************
276 * NB Only one loopback test can be executing concurrently.
278 **************************************************************************/
280 /* Loopback test RX callback
281 * This is called for each received packet during loopback testing.
283 void efx_loopback_rx_packet(struct efx_nic *efx,
284 const char *buf_ptr, int pkt_len)
286 struct efx_loopback_state *state = efx->loopback_selftest;
287 struct efx_loopback_payload *received;
288 struct efx_loopback_payload *payload;
292 /* If we are just flushing, then drop the packet */
293 if ((state == NULL) || state->flush)
296 payload = &state->payload;
298 received = (struct efx_loopback_payload *) buf_ptr;
299 received->ip.saddr = payload->ip.saddr;
300 if (state->offload_csum)
301 received->ip.check = payload->ip.check;
303 /* Check that header exists */
304 if (pkt_len < sizeof(received->header)) {
305 netif_err(efx, drv, efx->net_dev,
306 "saw runt RX packet (length %d) in %s loopback "
307 "test\n", pkt_len, LOOPBACK_MODE(efx));
311 /* Check that the ethernet header exists */
312 if (memcmp(&received->header, &payload->header, ETH_HLEN) != 0) {
313 netif_err(efx, drv, efx->net_dev,
314 "saw non-loopback RX packet in %s loopback test\n",
319 /* Check packet length */
320 if (pkt_len != sizeof(*payload)) {
321 netif_err(efx, drv, efx->net_dev,
322 "saw incorrect RX packet length %d (wanted %d) in "
323 "%s loopback test\n", pkt_len, (int)sizeof(*payload),
328 /* Check that IP header matches */
329 if (memcmp(&received->ip, &payload->ip, sizeof(payload->ip)) != 0) {
330 netif_err(efx, drv, efx->net_dev,
331 "saw corrupted IP header in %s loopback test\n",
336 /* Check that msg and padding matches */
337 if (memcmp(&received->msg, &payload->msg, sizeof(received->msg)) != 0) {
338 netif_err(efx, drv, efx->net_dev,
339 "saw corrupted RX packet in %s loopback test\n",
344 /* Check that iteration matches */
345 if (received->iteration != payload->iteration) {
346 netif_err(efx, drv, efx->net_dev,
347 "saw RX packet from iteration %d (wanted %d) in "
348 "%s loopback test\n", ntohs(received->iteration),
349 ntohs(payload->iteration), LOOPBACK_MODE(efx));
353 /* Increase correct RX count */
354 netif_vdbg(efx, drv, efx->net_dev,
355 "got loopback RX in %s loopback test\n", LOOPBACK_MODE(efx));
357 atomic_inc(&state->rx_good);
362 if (atomic_read(&state->rx_bad) == 0) {
363 netif_err(efx, drv, efx->net_dev, "received packet:\n");
364 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 0x10, 1,
365 buf_ptr, pkt_len, 0);
366 netif_err(efx, drv, efx->net_dev, "expected packet:\n");
367 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 0x10, 1,
368 &state->payload, sizeof(state->payload), 0);
371 atomic_inc(&state->rx_bad);
374 /* Initialise an efx_selftest_state for a new iteration */
375 static void efx_iterate_state(struct efx_nic *efx)
377 struct efx_loopback_state *state = efx->loopback_selftest;
378 struct net_device *net_dev = efx->net_dev;
379 struct efx_loopback_payload *payload = &state->payload;
381 /* Initialise the layerII header */
382 memcpy(&payload->header.h_dest, net_dev->dev_addr, ETH_ALEN);
383 memcpy(&payload->header.h_source, &payload_source, ETH_ALEN);
384 payload->header.h_proto = htons(ETH_P_IP);
386 /* saddr set later and used as incrementing count */
387 payload->ip.daddr = htonl(INADDR_LOOPBACK);
389 payload->ip.check = htons(0xdead);
390 payload->ip.tot_len = htons(sizeof(*payload) - sizeof(struct ethhdr));
391 payload->ip.version = IPVERSION;
392 payload->ip.protocol = IPPROTO_UDP;
394 /* Initialise udp header */
395 payload->udp.source = 0;
396 payload->udp.len = htons(sizeof(*payload) - sizeof(struct ethhdr) -
397 sizeof(struct iphdr));
398 payload->udp.check = 0; /* checksum ignored */
400 /* Fill out payload */
401 payload->iteration = htons(ntohs(payload->iteration) + 1);
402 memcpy(&payload->msg, payload_msg, sizeof(payload_msg));
404 /* Fill out remaining state members */
405 atomic_set(&state->rx_good, 0);
406 atomic_set(&state->rx_bad, 0);
410 static int efx_begin_loopback(struct efx_tx_queue *tx_queue)
412 struct efx_nic *efx = tx_queue->efx;
413 struct efx_loopback_state *state = efx->loopback_selftest;
414 struct efx_loopback_payload *payload;
419 /* Transmit N copies of buffer */
420 for (i = 0; i < state->packet_count; i++) {
421 /* Allocate an skb, holding an extra reference for
422 * transmit completion counting */
423 skb = alloc_skb(sizeof(state->payload), GFP_KERNEL);
426 state->skbs[i] = skb;
429 /* Copy the payload in, incrementing the source address to
430 * exercise the rss vectors */
431 payload = ((struct efx_loopback_payload *)
432 skb_put(skb, sizeof(state->payload)));
433 memcpy(payload, &state->payload, sizeof(state->payload));
434 payload->ip.saddr = htonl(INADDR_LOOPBACK | (i << 2));
436 /* Ensure everything we've written is visible to the
437 * interrupt handler. */
440 netif_tx_lock_bh(efx->net_dev);
441 rc = efx_enqueue_skb(tx_queue, skb);
442 netif_tx_unlock_bh(efx->net_dev);
444 if (rc != NETDEV_TX_OK) {
445 netif_err(efx, drv, efx->net_dev,
446 "TX queue %d could not transmit packet %d of "
447 "%d in %s loopback test\n", tx_queue->queue,
448 i + 1, state->packet_count,
451 /* Defer cleaning up the other skbs for the caller */
460 static int efx_poll_loopback(struct efx_nic *efx)
462 struct efx_loopback_state *state = efx->loopback_selftest;
463 struct efx_channel *channel;
465 /* NAPI polling is not enabled, so process channels
467 efx_for_each_channel(channel, efx) {
468 if (channel->work_pending)
469 efx_process_channel_now(channel);
471 return atomic_read(&state->rx_good) == state->packet_count;
474 static int efx_end_loopback(struct efx_tx_queue *tx_queue,
475 struct efx_loopback_self_tests *lb_tests)
477 struct efx_nic *efx = tx_queue->efx;
478 struct efx_loopback_state *state = efx->loopback_selftest;
480 int tx_done = 0, rx_good, rx_bad;
483 netif_tx_lock_bh(efx->net_dev);
485 /* Count the number of tx completions, and decrement the refcnt. Any
486 * skbs not already completed will be free'd when the queue is flushed */
487 for (i = 0; i < state->packet_count; i++) {
488 skb = state->skbs[i];
489 if (skb && !skb_shared(skb))
491 dev_kfree_skb_any(skb);
494 netif_tx_unlock_bh(efx->net_dev);
496 /* Check TX completion and received packet counts */
497 rx_good = atomic_read(&state->rx_good);
498 rx_bad = atomic_read(&state->rx_bad);
499 if (tx_done != state->packet_count) {
500 /* Don't free the skbs; they will be picked up on TX
501 * overflow or channel teardown.
503 netif_err(efx, drv, efx->net_dev,
504 "TX queue %d saw only %d out of an expected %d "
505 "TX completion events in %s loopback test\n",
506 tx_queue->queue, tx_done, state->packet_count,
509 /* Allow to fall through so we see the RX errors as well */
512 /* We may always be up to a flush away from our desired packet total */
513 if (rx_good != state->packet_count) {
514 netif_dbg(efx, drv, efx->net_dev,
515 "TX queue %d saw only %d out of an expected %d "
516 "received packets in %s loopback test\n",
517 tx_queue->queue, rx_good, state->packet_count,
523 /* Update loopback test structure */
524 lb_tests->tx_sent[tx_queue->queue] += state->packet_count;
525 lb_tests->tx_done[tx_queue->queue] += tx_done;
526 lb_tests->rx_good += rx_good;
527 lb_tests->rx_bad += rx_bad;
533 efx_test_loopback(struct efx_tx_queue *tx_queue,
534 struct efx_loopback_self_tests *lb_tests)
536 struct efx_nic *efx = tx_queue->efx;
537 struct efx_loopback_state *state = efx->loopback_selftest;
538 int i, begin_rc, end_rc;
540 for (i = 0; i < 3; i++) {
541 /* Determine how many packets to send */
542 state->packet_count = efx->txq_entries / 3;
543 state->packet_count = min(1 << (i << 2), state->packet_count);
544 state->skbs = kcalloc(state->packet_count,
545 sizeof(state->skbs[0]), GFP_KERNEL);
548 state->flush = false;
550 netif_dbg(efx, drv, efx->net_dev,
551 "TX queue %d testing %s loopback with %d packets\n",
552 tx_queue->queue, LOOPBACK_MODE(efx),
553 state->packet_count);
555 efx_iterate_state(efx);
556 begin_rc = efx_begin_loopback(tx_queue);
558 /* This will normally complete very quickly, but be
559 * prepared to wait much longer. */
561 if (!efx_poll_loopback(efx)) {
562 msleep(LOOPBACK_TIMEOUT_MS);
563 efx_poll_loopback(efx);
566 end_rc = efx_end_loopback(tx_queue, lb_tests);
569 if (begin_rc || end_rc) {
570 /* Wait a while to ensure there are no packets
571 * floating around after a failure. */
572 schedule_timeout_uninterruptible(HZ / 10);
573 return begin_rc ? begin_rc : end_rc;
577 netif_dbg(efx, drv, efx->net_dev,
578 "TX queue %d passed %s loopback test with a burst length "
579 "of %d packets\n", tx_queue->queue, LOOPBACK_MODE(efx),
580 state->packet_count);
585 /* Wait for link up. On Falcon, we would prefer to rely on efx_monitor, but
586 * any contention on the mac lock (via e.g. efx_mac_mcast_work) causes it
587 * to delay and retry. Therefore, it's safer to just poll directly. Wait
588 * for link up and any faults to dissipate. */
589 static int efx_wait_for_link(struct efx_nic *efx)
591 struct efx_link_state *link_state = &efx->link_state;
592 int count, link_up_count = 0;
595 for (count = 0; count < 40; count++) {
596 schedule_timeout_uninterruptible(HZ / 10);
598 if (efx->type->monitor != NULL) {
599 mutex_lock(&efx->mac_lock);
600 efx->type->monitor(efx);
601 mutex_unlock(&efx->mac_lock);
603 struct efx_channel *channel = efx_get_channel(efx, 0);
604 if (channel->work_pending)
605 efx_process_channel_now(channel);
608 mutex_lock(&efx->mac_lock);
609 link_up = link_state->up;
611 link_up = !efx->type->check_mac_fault(efx);
612 mutex_unlock(&efx->mac_lock);
615 if (++link_up_count == 2)
625 static int efx_test_loopbacks(struct efx_nic *efx, struct efx_self_tests *tests,
626 unsigned int loopback_modes)
628 enum efx_loopback_mode mode;
629 struct efx_loopback_state *state;
630 struct efx_channel *channel = efx_get_channel(efx, 0);
631 struct efx_tx_queue *tx_queue;
634 /* Set the port loopback_selftest member. From this point on
635 * all received packets will be dropped. Mark the state as
636 * "flushing" so all inflight packets are dropped */
637 state = kzalloc(sizeof(*state), GFP_KERNEL);
640 BUG_ON(efx->loopback_selftest);
642 efx->loopback_selftest = state;
644 /* Test all supported loopback modes */
645 for (mode = LOOPBACK_NONE; mode <= LOOPBACK_TEST_MAX; mode++) {
646 if (!(loopback_modes & (1 << mode)))
649 /* Move the port into the specified loopback mode. */
651 mutex_lock(&efx->mac_lock);
652 efx->loopback_mode = mode;
653 rc = __efx_reconfigure_port(efx);
654 mutex_unlock(&efx->mac_lock);
656 netif_err(efx, drv, efx->net_dev,
657 "unable to move into %s loopback\n",
662 rc = efx_wait_for_link(efx);
664 netif_err(efx, drv, efx->net_dev,
665 "loopback %s never came up\n",
670 /* Test all enabled types of TX queue */
671 efx_for_each_channel_tx_queue(tx_queue, channel) {
672 state->offload_csum = (tx_queue->queue &
673 EFX_TXQ_TYPE_OFFLOAD);
674 rc = efx_test_loopback(tx_queue,
675 &tests->loopback[mode]);
682 /* Remove the flush. The caller will remove the loopback setting */
684 efx->loopback_selftest = NULL;
691 /**************************************************************************
695 *************************************************************************/
697 int efx_selftest(struct efx_nic *efx, struct efx_self_tests *tests,
700 enum efx_loopback_mode loopback_mode = efx->loopback_mode;
701 int phy_mode = efx->phy_mode;
702 enum reset_type reset_method = RESET_TYPE_INVISIBLE;
703 int rc_test = 0, rc_reset = 0, rc;
705 efx_selftest_async_cancel(efx);
707 /* Online (i.e. non-disruptive) testing
708 * This checks interrupt generation, event delivery and PHY presence. */
710 rc = efx_test_phy_alive(efx, tests);
714 rc = efx_test_nvram(efx, tests);
718 rc = efx_test_interrupts(efx, tests);
722 rc = efx_test_eventq_irq(efx, tests);
729 if (!(flags & ETH_TEST_FL_OFFLINE))
730 return efx_test_phy(efx, tests, flags);
732 /* Offline (i.e. disruptive) testing
733 * This checks MAC and PHY loopback on the specified port. */
735 /* Detach the device so the kernel doesn't transmit during the
736 * loopback test and the watchdog timeout doesn't fire.
738 netif_device_detach(efx->net_dev);
740 mutex_lock(&efx->mac_lock);
741 if (efx->loopback_modes) {
742 /* We need the 312 clock from the PHY to test the XMAC
743 * registers, so move into XGMII loopback if available */
744 if (efx->loopback_modes & (1 << LOOPBACK_XGMII))
745 efx->loopback_mode = LOOPBACK_XGMII;
747 efx->loopback_mode = __ffs(efx->loopback_modes);
750 __efx_reconfigure_port(efx);
751 mutex_unlock(&efx->mac_lock);
753 /* free up all consumers of SRAM (including all the queues) */
754 efx_reset_down(efx, reset_method);
756 rc = efx_test_chip(efx, tests);
760 /* reset the chip to recover from the register test */
761 rc_reset = efx->type->reset(efx, reset_method);
763 /* Ensure that the phy is powered and out of loopback
764 * for the bist and loopback tests */
765 efx->phy_mode &= ~PHY_MODE_LOW_POWER;
766 efx->loopback_mode = LOOPBACK_NONE;
768 rc = efx_reset_up(efx, reset_method, rc_reset == 0);
773 netif_err(efx, drv, efx->net_dev,
774 "Unable to recover from chip test\n");
775 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
779 rc = efx_test_phy(efx, tests, flags);
783 rc = efx_test_loopbacks(efx, tests, efx->loopback_modes);
787 /* restore the PHY to the previous state */
788 mutex_lock(&efx->mac_lock);
789 efx->phy_mode = phy_mode;
790 efx->loopback_mode = loopback_mode;
791 __efx_reconfigure_port(efx);
792 mutex_unlock(&efx->mac_lock);
794 netif_device_attach(efx->net_dev);
799 void efx_selftest_async_start(struct efx_nic *efx)
801 struct efx_channel *channel;
803 efx_for_each_channel(channel, efx)
804 efx_nic_event_test_start(channel);
805 schedule_delayed_work(&efx->selftest_work, IRQ_TIMEOUT);
808 void efx_selftest_async_cancel(struct efx_nic *efx)
810 cancel_delayed_work_sync(&efx->selftest_work);
813 void efx_selftest_async_work(struct work_struct *data)
815 struct efx_nic *efx = container_of(data, struct efx_nic,
817 struct efx_channel *channel;
820 efx_for_each_channel(channel, efx) {
821 cpu = efx_nic_event_test_irq_cpu(channel);
823 netif_err(efx, ifup, efx->net_dev,
824 "channel %d failed to trigger an interrupt\n",
827 netif_dbg(efx, ifup, efx->net_dev,
828 "channel %d triggered interrupt on CPU %d\n",
829 channel->channel, cpu);