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 2006-2013 Solarflare Communications Inc.
8 #include <linux/bitops.h>
9 #include <linux/delay.h>
10 #include <linux/interrupt.h>
11 #include <linux/pci.h>
12 #include <linux/module.h>
13 #include <linux/seq_file.h>
14 #include <linux/crc32.h>
15 #include "net_driver.h"
18 #include "rx_common.h"
20 #include "farch_regs.h"
22 #include "siena_sriov.h"
24 #include "workarounds.h"
26 /* Falcon-architecture (SFC9000-family) support */
28 /**************************************************************************
32 **************************************************************************
35 /* This is set to 16 for a good reason. In summary, if larger than
36 * 16, the descriptor cache holds more than a default socket
37 * buffer's worth of packets (for UDP we can only have at most one
38 * socket buffer's worth outstanding). This combined with the fact
39 * that we only get 1 TX event per descriptor cache means the NIC
42 #define TX_DC_ENTRIES 16
43 #define TX_DC_ENTRIES_ORDER 1
45 #define RX_DC_ENTRIES 64
46 #define RX_DC_ENTRIES_ORDER 3
48 /* If EFX_MAX_INT_ERRORS internal errors occur within
49 * EFX_INT_ERROR_EXPIRE seconds, we consider the NIC broken and
52 #define EFX_INT_ERROR_EXPIRE 3600
53 #define EFX_MAX_INT_ERRORS 5
55 /* Depth of RX flush request fifo */
56 #define EFX_RX_FLUSH_COUNT 4
58 /* Driver generated events */
59 #define _EFX_CHANNEL_MAGIC_TEST 0x000101
60 #define _EFX_CHANNEL_MAGIC_FILL 0x000102
61 #define _EFX_CHANNEL_MAGIC_RX_DRAIN 0x000103
62 #define _EFX_CHANNEL_MAGIC_TX_DRAIN 0x000104
64 #define _EFX_CHANNEL_MAGIC(_code, _data) ((_code) << 8 | (_data))
65 #define _EFX_CHANNEL_MAGIC_CODE(_magic) ((_magic) >> 8)
67 #define EFX_CHANNEL_MAGIC_TEST(_channel) \
68 _EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_TEST, (_channel)->channel)
69 #define EFX_CHANNEL_MAGIC_FILL(_rx_queue) \
70 _EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_FILL, \
71 efx_rx_queue_index(_rx_queue))
72 #define EFX_CHANNEL_MAGIC_RX_DRAIN(_rx_queue) \
73 _EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_RX_DRAIN, \
74 efx_rx_queue_index(_rx_queue))
75 #define EFX_CHANNEL_MAGIC_TX_DRAIN(_tx_queue) \
76 _EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_TX_DRAIN, \
79 static void efx_farch_magic_event(struct efx_channel *channel, u32 magic);
81 /**************************************************************************
85 **************************************************************************/
87 static inline void efx_write_buf_tbl(struct efx_nic *efx, efx_qword_t *value,
90 efx_sram_writeq(efx, efx->membase + efx->type->buf_tbl_base,
94 static bool efx_masked_compare_oword(const efx_oword_t *a, const efx_oword_t *b,
95 const efx_oword_t *mask)
97 return ((a->u64[0] ^ b->u64[0]) & mask->u64[0]) ||
98 ((a->u64[1] ^ b->u64[1]) & mask->u64[1]);
101 int efx_farch_test_registers(struct efx_nic *efx,
102 const struct efx_farch_register_test *regs,
105 unsigned address = 0;
107 efx_oword_t mask, imask, original, reg, buf;
109 for (i = 0; i < n_regs; ++i) {
110 address = regs[i].address;
111 mask = imask = regs[i].mask;
112 EFX_INVERT_OWORD(imask);
114 efx_reado(efx, &original, address);
116 /* bit sweep on and off */
117 for (j = 0; j < 128; j++) {
118 if (!EFX_EXTRACT_OWORD32(mask, j, j))
121 /* Test this testable bit can be set in isolation */
122 EFX_AND_OWORD(reg, original, mask);
123 EFX_SET_OWORD32(reg, j, j, 1);
125 efx_writeo(efx, ®, address);
126 efx_reado(efx, &buf, address);
128 if (efx_masked_compare_oword(®, &buf, &mask))
131 /* Test this testable bit can be cleared in isolation */
132 EFX_OR_OWORD(reg, original, mask);
133 EFX_SET_OWORD32(reg, j, j, 0);
135 efx_writeo(efx, ®, address);
136 efx_reado(efx, &buf, address);
138 if (efx_masked_compare_oword(®, &buf, &mask))
142 efx_writeo(efx, &original, address);
148 netif_err(efx, hw, efx->net_dev,
149 "wrote "EFX_OWORD_FMT" read "EFX_OWORD_FMT
150 " at address 0x%x mask "EFX_OWORD_FMT"\n", EFX_OWORD_VAL(reg),
151 EFX_OWORD_VAL(buf), address, EFX_OWORD_VAL(mask));
155 /**************************************************************************
157 * Special buffer handling
158 * Special buffers are used for event queues and the TX and RX
161 *************************************************************************/
164 * Initialise a special buffer
166 * This will define a buffer (previously allocated via
167 * efx_alloc_special_buffer()) in the buffer table, allowing
168 * it to be used for event queues, descriptor rings etc.
171 efx_init_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
173 efx_qword_t buf_desc;
178 EFX_WARN_ON_PARANOID(!buffer->buf.addr);
180 /* Write buffer descriptors to NIC */
181 for (i = 0; i < buffer->entries; i++) {
182 index = buffer->index + i;
183 dma_addr = buffer->buf.dma_addr + (i * EFX_BUF_SIZE);
184 netif_dbg(efx, probe, efx->net_dev,
185 "mapping special buffer %d at %llx\n",
186 index, (unsigned long long)dma_addr);
187 EFX_POPULATE_QWORD_3(buf_desc,
188 FRF_AZ_BUF_ADR_REGION, 0,
189 FRF_AZ_BUF_ADR_FBUF, dma_addr >> 12,
190 FRF_AZ_BUF_OWNER_ID_FBUF, 0);
191 efx_write_buf_tbl(efx, &buf_desc, index);
195 /* Unmaps a buffer and clears the buffer table entries */
197 efx_fini_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
199 efx_oword_t buf_tbl_upd;
200 unsigned int start = buffer->index;
201 unsigned int end = (buffer->index + buffer->entries - 1);
203 if (!buffer->entries)
206 netif_dbg(efx, hw, efx->net_dev, "unmapping special buffers %d-%d\n",
207 buffer->index, buffer->index + buffer->entries - 1);
209 EFX_POPULATE_OWORD_4(buf_tbl_upd,
210 FRF_AZ_BUF_UPD_CMD, 0,
211 FRF_AZ_BUF_CLR_CMD, 1,
212 FRF_AZ_BUF_CLR_END_ID, end,
213 FRF_AZ_BUF_CLR_START_ID, start);
214 efx_writeo(efx, &buf_tbl_upd, FR_AZ_BUF_TBL_UPD);
218 * Allocate a new special buffer
220 * This allocates memory for a new buffer, clears it and allocates a
221 * new buffer ID range. It does not write into the buffer table.
223 * This call will allocate 4KB buffers, since 8KB buffers can't be
224 * used for event queues and descriptor rings.
226 static int efx_alloc_special_buffer(struct efx_nic *efx,
227 struct efx_special_buffer *buffer,
230 #ifdef CONFIG_SFC_SRIOV
231 struct siena_nic_data *nic_data = efx->nic_data;
233 len = ALIGN(len, EFX_BUF_SIZE);
235 if (efx_nic_alloc_buffer(efx, &buffer->buf, len, GFP_KERNEL))
237 buffer->entries = len / EFX_BUF_SIZE;
238 BUG_ON(buffer->buf.dma_addr & (EFX_BUF_SIZE - 1));
240 /* Select new buffer ID */
241 buffer->index = efx->next_buffer_table;
242 efx->next_buffer_table += buffer->entries;
243 #ifdef CONFIG_SFC_SRIOV
244 BUG_ON(efx_siena_sriov_enabled(efx) &&
245 nic_data->vf_buftbl_base < efx->next_buffer_table);
248 netif_dbg(efx, probe, efx->net_dev,
249 "allocating special buffers %d-%d at %llx+%x "
250 "(virt %p phys %llx)\n", buffer->index,
251 buffer->index + buffer->entries - 1,
252 (u64)buffer->buf.dma_addr, len,
253 buffer->buf.addr, (u64)virt_to_phys(buffer->buf.addr));
259 efx_free_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
261 if (!buffer->buf.addr)
264 netif_dbg(efx, hw, efx->net_dev,
265 "deallocating special buffers %d-%d at %llx+%x "
266 "(virt %p phys %llx)\n", buffer->index,
267 buffer->index + buffer->entries - 1,
268 (u64)buffer->buf.dma_addr, buffer->buf.len,
269 buffer->buf.addr, (u64)virt_to_phys(buffer->buf.addr));
271 efx_nic_free_buffer(efx, &buffer->buf);
275 /**************************************************************************
279 **************************************************************************/
281 /* This writes to the TX_DESC_WPTR; write pointer for TX descriptor ring */
282 static inline void efx_farch_notify_tx_desc(struct efx_tx_queue *tx_queue)
287 write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
288 EFX_POPULATE_DWORD_1(reg, FRF_AZ_TX_DESC_WPTR_DWORD, write_ptr);
289 efx_writed_page(tx_queue->efx, ®,
290 FR_AZ_TX_DESC_UPD_DWORD_P0, tx_queue->queue);
293 /* Write pointer and first descriptor for TX descriptor ring */
294 static inline void efx_farch_push_tx_desc(struct efx_tx_queue *tx_queue,
295 const efx_qword_t *txd)
300 BUILD_BUG_ON(FRF_AZ_TX_DESC_LBN != 0);
301 BUILD_BUG_ON(FR_AA_TX_DESC_UPD_KER != FR_BZ_TX_DESC_UPD_P0);
303 write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
304 EFX_POPULATE_OWORD_2(reg, FRF_AZ_TX_DESC_PUSH_CMD, true,
305 FRF_AZ_TX_DESC_WPTR, write_ptr);
307 efx_writeo_page(tx_queue->efx, ®,
308 FR_BZ_TX_DESC_UPD_P0, tx_queue->queue);
312 /* For each entry inserted into the software descriptor ring, create a
313 * descriptor in the hardware TX descriptor ring (in host memory), and
316 void efx_farch_tx_write(struct efx_tx_queue *tx_queue)
318 struct efx_tx_buffer *buffer;
321 unsigned old_write_count = tx_queue->write_count;
323 tx_queue->xmit_pending = false;
324 if (unlikely(tx_queue->write_count == tx_queue->insert_count))
328 write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
329 buffer = &tx_queue->buffer[write_ptr];
330 txd = efx_tx_desc(tx_queue, write_ptr);
331 ++tx_queue->write_count;
333 EFX_WARN_ON_ONCE_PARANOID(buffer->flags & EFX_TX_BUF_OPTION);
335 /* Create TX descriptor ring entry */
336 BUILD_BUG_ON(EFX_TX_BUF_CONT != 1);
337 EFX_POPULATE_QWORD_4(*txd,
339 buffer->flags & EFX_TX_BUF_CONT,
340 FSF_AZ_TX_KER_BYTE_COUNT, buffer->len,
341 FSF_AZ_TX_KER_BUF_REGION, 0,
342 FSF_AZ_TX_KER_BUF_ADDR, buffer->dma_addr);
343 } while (tx_queue->write_count != tx_queue->insert_count);
345 wmb(); /* Ensure descriptors are written before they are fetched */
347 if (efx_nic_may_push_tx_desc(tx_queue, old_write_count)) {
348 txd = efx_tx_desc(tx_queue,
349 old_write_count & tx_queue->ptr_mask);
350 efx_farch_push_tx_desc(tx_queue, txd);
353 efx_farch_notify_tx_desc(tx_queue);
357 unsigned int efx_farch_tx_limit_len(struct efx_tx_queue *tx_queue,
358 dma_addr_t dma_addr, unsigned int len)
360 /* Don't cross 4K boundaries with descriptors. */
361 unsigned int limit = (~dma_addr & (EFX_PAGE_SIZE - 1)) + 1;
363 len = min(limit, len);
369 /* Allocate hardware resources for a TX queue */
370 int efx_farch_tx_probe(struct efx_tx_queue *tx_queue)
372 struct efx_nic *efx = tx_queue->efx;
375 entries = tx_queue->ptr_mask + 1;
376 return efx_alloc_special_buffer(efx, &tx_queue->txd,
377 entries * sizeof(efx_qword_t));
380 void efx_farch_tx_init(struct efx_tx_queue *tx_queue)
382 int csum = tx_queue->label & EFX_TXQ_TYPE_OFFLOAD;
383 struct efx_nic *efx = tx_queue->efx;
386 /* Pin TX descriptor ring */
387 efx_init_special_buffer(efx, &tx_queue->txd);
389 /* Push TX descriptor ring to card */
390 EFX_POPULATE_OWORD_10(reg,
391 FRF_AZ_TX_DESCQ_EN, 1,
392 FRF_AZ_TX_ISCSI_DDIG_EN, 0,
393 FRF_AZ_TX_ISCSI_HDIG_EN, 0,
394 FRF_AZ_TX_DESCQ_BUF_BASE_ID, tx_queue->txd.index,
395 FRF_AZ_TX_DESCQ_EVQ_ID,
396 tx_queue->channel->channel,
397 FRF_AZ_TX_DESCQ_OWNER_ID, 0,
398 FRF_AZ_TX_DESCQ_LABEL, tx_queue->label,
399 FRF_AZ_TX_DESCQ_SIZE,
400 __ffs(tx_queue->txd.entries),
401 FRF_AZ_TX_DESCQ_TYPE, 0,
402 FRF_BZ_TX_NON_IP_DROP_DIS, 1);
404 EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_IP_CHKSM_DIS, !csum);
405 EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_TCP_CHKSM_DIS, !csum);
407 efx_writeo_table(efx, ®, efx->type->txd_ptr_tbl_base,
410 EFX_POPULATE_OWORD_1(reg,
412 (tx_queue->label & EFX_TXQ_TYPE_HIGHPRI) ?
414 FFE_BZ_TX_PACE_RESERVED);
415 efx_writeo_table(efx, ®, FR_BZ_TX_PACE_TBL, tx_queue->queue);
418 static void efx_farch_flush_tx_queue(struct efx_tx_queue *tx_queue)
420 struct efx_nic *efx = tx_queue->efx;
421 efx_oword_t tx_flush_descq;
423 WARN_ON(atomic_read(&tx_queue->flush_outstanding));
424 atomic_set(&tx_queue->flush_outstanding, 1);
426 EFX_POPULATE_OWORD_2(tx_flush_descq,
427 FRF_AZ_TX_FLUSH_DESCQ_CMD, 1,
428 FRF_AZ_TX_FLUSH_DESCQ, tx_queue->queue);
429 efx_writeo(efx, &tx_flush_descq, FR_AZ_TX_FLUSH_DESCQ);
432 void efx_farch_tx_fini(struct efx_tx_queue *tx_queue)
434 struct efx_nic *efx = tx_queue->efx;
435 efx_oword_t tx_desc_ptr;
437 /* Remove TX descriptor ring from card */
438 EFX_ZERO_OWORD(tx_desc_ptr);
439 efx_writeo_table(efx, &tx_desc_ptr, efx->type->txd_ptr_tbl_base,
442 /* Unpin TX descriptor ring */
443 efx_fini_special_buffer(efx, &tx_queue->txd);
446 /* Free buffers backing TX queue */
447 void efx_farch_tx_remove(struct efx_tx_queue *tx_queue)
449 efx_free_special_buffer(tx_queue->efx, &tx_queue->txd);
452 /**************************************************************************
456 **************************************************************************/
458 /* This creates an entry in the RX descriptor queue */
460 efx_farch_build_rx_desc(struct efx_rx_queue *rx_queue, unsigned index)
462 struct efx_rx_buffer *rx_buf;
465 rxd = efx_rx_desc(rx_queue, index);
466 rx_buf = efx_rx_buffer(rx_queue, index);
467 EFX_POPULATE_QWORD_3(*rxd,
468 FSF_AZ_RX_KER_BUF_SIZE,
470 rx_queue->efx->type->rx_buffer_padding,
471 FSF_AZ_RX_KER_BUF_REGION, 0,
472 FSF_AZ_RX_KER_BUF_ADDR, rx_buf->dma_addr);
475 /* This writes to the RX_DESC_WPTR register for the specified receive
478 void efx_farch_rx_write(struct efx_rx_queue *rx_queue)
480 struct efx_nic *efx = rx_queue->efx;
484 while (rx_queue->notified_count != rx_queue->added_count) {
485 efx_farch_build_rx_desc(
487 rx_queue->notified_count & rx_queue->ptr_mask);
488 ++rx_queue->notified_count;
492 write_ptr = rx_queue->added_count & rx_queue->ptr_mask;
493 EFX_POPULATE_DWORD_1(reg, FRF_AZ_RX_DESC_WPTR_DWORD, write_ptr);
494 efx_writed_page(efx, ®, FR_AZ_RX_DESC_UPD_DWORD_P0,
495 efx_rx_queue_index(rx_queue));
498 int efx_farch_rx_probe(struct efx_rx_queue *rx_queue)
500 struct efx_nic *efx = rx_queue->efx;
503 entries = rx_queue->ptr_mask + 1;
504 return efx_alloc_special_buffer(efx, &rx_queue->rxd,
505 entries * sizeof(efx_qword_t));
508 void efx_farch_rx_init(struct efx_rx_queue *rx_queue)
510 efx_oword_t rx_desc_ptr;
511 struct efx_nic *efx = rx_queue->efx;
514 /* For kernel-mode queues in Siena, the JUMBO flag enables scatter. */
515 jumbo_en = efx->rx_scatter;
517 netif_dbg(efx, hw, efx->net_dev,
518 "RX queue %d ring in special buffers %d-%d\n",
519 efx_rx_queue_index(rx_queue), rx_queue->rxd.index,
520 rx_queue->rxd.index + rx_queue->rxd.entries - 1);
522 rx_queue->scatter_n = 0;
524 /* Pin RX descriptor ring */
525 efx_init_special_buffer(efx, &rx_queue->rxd);
527 /* Push RX descriptor ring to card */
528 EFX_POPULATE_OWORD_10(rx_desc_ptr,
529 FRF_AZ_RX_ISCSI_DDIG_EN, true,
530 FRF_AZ_RX_ISCSI_HDIG_EN, true,
531 FRF_AZ_RX_DESCQ_BUF_BASE_ID, rx_queue->rxd.index,
532 FRF_AZ_RX_DESCQ_EVQ_ID,
533 efx_rx_queue_channel(rx_queue)->channel,
534 FRF_AZ_RX_DESCQ_OWNER_ID, 0,
535 FRF_AZ_RX_DESCQ_LABEL,
536 efx_rx_queue_index(rx_queue),
537 FRF_AZ_RX_DESCQ_SIZE,
538 __ffs(rx_queue->rxd.entries),
539 FRF_AZ_RX_DESCQ_TYPE, 0 /* kernel queue */ ,
540 FRF_AZ_RX_DESCQ_JUMBO, jumbo_en,
541 FRF_AZ_RX_DESCQ_EN, 1);
542 efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
543 efx_rx_queue_index(rx_queue));
546 static void efx_farch_flush_rx_queue(struct efx_rx_queue *rx_queue)
548 struct efx_nic *efx = rx_queue->efx;
549 efx_oword_t rx_flush_descq;
551 EFX_POPULATE_OWORD_2(rx_flush_descq,
552 FRF_AZ_RX_FLUSH_DESCQ_CMD, 1,
553 FRF_AZ_RX_FLUSH_DESCQ,
554 efx_rx_queue_index(rx_queue));
555 efx_writeo(efx, &rx_flush_descq, FR_AZ_RX_FLUSH_DESCQ);
558 void efx_farch_rx_fini(struct efx_rx_queue *rx_queue)
560 efx_oword_t rx_desc_ptr;
561 struct efx_nic *efx = rx_queue->efx;
563 /* Remove RX descriptor ring from card */
564 EFX_ZERO_OWORD(rx_desc_ptr);
565 efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
566 efx_rx_queue_index(rx_queue));
568 /* Unpin RX descriptor ring */
569 efx_fini_special_buffer(efx, &rx_queue->rxd);
572 /* Free buffers backing RX queue */
573 void efx_farch_rx_remove(struct efx_rx_queue *rx_queue)
575 efx_free_special_buffer(rx_queue->efx, &rx_queue->rxd);
578 /**************************************************************************
582 **************************************************************************/
584 /* efx_farch_flush_queues() must be woken up when all flushes are completed,
585 * or more RX flushes can be kicked off.
587 static bool efx_farch_flush_wake(struct efx_nic *efx)
589 /* Ensure that all updates are visible to efx_farch_flush_queues() */
592 return (atomic_read(&efx->active_queues) == 0 ||
593 (atomic_read(&efx->rxq_flush_outstanding) < EFX_RX_FLUSH_COUNT
594 && atomic_read(&efx->rxq_flush_pending) > 0));
597 static bool efx_check_tx_flush_complete(struct efx_nic *efx)
600 efx_oword_t txd_ptr_tbl;
601 struct efx_channel *channel;
602 struct efx_tx_queue *tx_queue;
604 efx_for_each_channel(channel, efx) {
605 efx_for_each_channel_tx_queue(tx_queue, channel) {
606 efx_reado_table(efx, &txd_ptr_tbl,
607 FR_BZ_TX_DESC_PTR_TBL, tx_queue->queue);
608 if (EFX_OWORD_FIELD(txd_ptr_tbl,
609 FRF_AZ_TX_DESCQ_FLUSH) ||
610 EFX_OWORD_FIELD(txd_ptr_tbl,
611 FRF_AZ_TX_DESCQ_EN)) {
612 netif_dbg(efx, hw, efx->net_dev,
613 "flush did not complete on TXQ %d\n",
616 } else if (atomic_cmpxchg(&tx_queue->flush_outstanding,
618 /* The flush is complete, but we didn't
619 * receive a flush completion event
621 netif_dbg(efx, hw, efx->net_dev,
622 "flush complete on TXQ %d, so drain "
623 "the queue\n", tx_queue->queue);
624 /* Don't need to increment active_queues as it
625 * has already been incremented for the queues
626 * which did not drain
628 efx_farch_magic_event(channel,
629 EFX_CHANNEL_MAGIC_TX_DRAIN(
638 /* Flush all the transmit queues, and continue flushing receive queues until
639 * they're all flushed. Wait for the DRAIN events to be received so that there
640 * are no more RX and TX events left on any channel. */
641 static int efx_farch_do_flush(struct efx_nic *efx)
643 unsigned timeout = msecs_to_jiffies(5000); /* 5s for all flushes and drains */
644 struct efx_channel *channel;
645 struct efx_rx_queue *rx_queue;
646 struct efx_tx_queue *tx_queue;
649 efx_for_each_channel(channel, efx) {
650 efx_for_each_channel_tx_queue(tx_queue, channel) {
651 efx_farch_flush_tx_queue(tx_queue);
653 efx_for_each_channel_rx_queue(rx_queue, channel) {
654 rx_queue->flush_pending = true;
655 atomic_inc(&efx->rxq_flush_pending);
659 while (timeout && atomic_read(&efx->active_queues) > 0) {
660 /* If SRIOV is enabled, then offload receive queue flushing to
661 * the firmware (though we will still have to poll for
662 * completion). If that fails, fall back to the old scheme.
664 if (efx_siena_sriov_enabled(efx)) {
665 rc = efx_mcdi_flush_rxqs(efx);
670 /* The hardware supports four concurrent rx flushes, each of
671 * which may need to be retried if there is an outstanding
674 efx_for_each_channel(channel, efx) {
675 efx_for_each_channel_rx_queue(rx_queue, channel) {
676 if (atomic_read(&efx->rxq_flush_outstanding) >=
680 if (rx_queue->flush_pending) {
681 rx_queue->flush_pending = false;
682 atomic_dec(&efx->rxq_flush_pending);
683 atomic_inc(&efx->rxq_flush_outstanding);
684 efx_farch_flush_rx_queue(rx_queue);
690 timeout = wait_event_timeout(efx->flush_wq,
691 efx_farch_flush_wake(efx),
695 if (atomic_read(&efx->active_queues) &&
696 !efx_check_tx_flush_complete(efx)) {
697 netif_err(efx, hw, efx->net_dev, "failed to flush %d queues "
698 "(rx %d+%d)\n", atomic_read(&efx->active_queues),
699 atomic_read(&efx->rxq_flush_outstanding),
700 atomic_read(&efx->rxq_flush_pending));
703 atomic_set(&efx->active_queues, 0);
704 atomic_set(&efx->rxq_flush_pending, 0);
705 atomic_set(&efx->rxq_flush_outstanding, 0);
711 int efx_farch_fini_dmaq(struct efx_nic *efx)
713 struct efx_channel *channel;
714 struct efx_tx_queue *tx_queue;
715 struct efx_rx_queue *rx_queue;
718 /* Do not attempt to write to the NIC during EEH recovery */
719 if (efx->state != STATE_RECOVERY) {
720 /* Only perform flush if DMA is enabled */
721 if (efx->pci_dev->is_busmaster) {
722 efx->type->prepare_flush(efx);
723 rc = efx_farch_do_flush(efx);
724 efx->type->finish_flush(efx);
727 efx_for_each_channel(channel, efx) {
728 efx_for_each_channel_rx_queue(rx_queue, channel)
729 efx_farch_rx_fini(rx_queue);
730 efx_for_each_channel_tx_queue(tx_queue, channel)
731 efx_farch_tx_fini(tx_queue);
738 /* Reset queue and flush accounting after FLR
740 * One possible cause of FLR recovery is that DMA may be failing (eg. if bus
741 * mastering was disabled), in which case we don't receive (RXQ) flush
742 * completion events. This means that efx->rxq_flush_outstanding remained at 4
743 * after the FLR; also, efx->active_queues was non-zero (as no flush completion
744 * events were received, and we didn't go through efx_check_tx_flush_complete())
745 * If we don't fix this up, on the next call to efx_realloc_channels() we won't
746 * flush any RX queues because efx->rxq_flush_outstanding is at the limit of 4
747 * for batched flush requests; and the efx->active_queues gets messed up because
748 * we keep incrementing for the newly initialised queues, but it never went to
749 * zero previously. Then we get a timeout every time we try to restart the
750 * queues, as it doesn't go back to zero when we should be flushing the queues.
752 void efx_farch_finish_flr(struct efx_nic *efx)
754 atomic_set(&efx->rxq_flush_pending, 0);
755 atomic_set(&efx->rxq_flush_outstanding, 0);
756 atomic_set(&efx->active_queues, 0);
760 /**************************************************************************
762 * Event queue processing
763 * Event queues are processed by per-channel tasklets.
765 **************************************************************************/
767 /* Update a channel's event queue's read pointer (RPTR) register
769 * This writes the EVQ_RPTR_REG register for the specified channel's
772 void efx_farch_ev_read_ack(struct efx_channel *channel)
775 struct efx_nic *efx = channel->efx;
777 EFX_POPULATE_DWORD_1(reg, FRF_AZ_EVQ_RPTR,
778 channel->eventq_read_ptr & channel->eventq_mask);
780 /* For Falcon A1, EVQ_RPTR_KER is documented as having a step size
781 * of 4 bytes, but it is really 16 bytes just like later revisions.
783 efx_writed(efx, ®,
784 efx->type->evq_rptr_tbl_base +
785 FR_BZ_EVQ_RPTR_STEP * channel->channel);
788 /* Use HW to insert a SW defined event */
789 void efx_farch_generate_event(struct efx_nic *efx, unsigned int evq,
792 efx_oword_t drv_ev_reg;
794 BUILD_BUG_ON(FRF_AZ_DRV_EV_DATA_LBN != 0 ||
795 FRF_AZ_DRV_EV_DATA_WIDTH != 64);
796 drv_ev_reg.u32[0] = event->u32[0];
797 drv_ev_reg.u32[1] = event->u32[1];
798 drv_ev_reg.u32[2] = 0;
799 drv_ev_reg.u32[3] = 0;
800 EFX_SET_OWORD_FIELD(drv_ev_reg, FRF_AZ_DRV_EV_QID, evq);
801 efx_writeo(efx, &drv_ev_reg, FR_AZ_DRV_EV);
804 static void efx_farch_magic_event(struct efx_channel *channel, u32 magic)
808 EFX_POPULATE_QWORD_2(event, FSF_AZ_EV_CODE,
809 FSE_AZ_EV_CODE_DRV_GEN_EV,
810 FSF_AZ_DRV_GEN_EV_MAGIC, magic);
811 efx_farch_generate_event(channel->efx, channel->channel, &event);
814 /* Handle a transmit completion event
816 * The NIC batches TX completion events; the message we receive is of
817 * the form "complete all TX events up to this index".
820 efx_farch_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
822 unsigned int tx_ev_desc_ptr;
823 unsigned int tx_ev_q_label;
824 struct efx_tx_queue *tx_queue;
825 struct efx_nic *efx = channel->efx;
827 if (unlikely(READ_ONCE(efx->reset_pending)))
830 if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
831 /* Transmit completion */
832 tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
833 tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
834 tx_queue = efx_channel_get_tx_queue(
835 channel, tx_ev_q_label % EFX_TXQ_TYPES);
836 efx_xmit_done(tx_queue, tx_ev_desc_ptr);
837 } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
838 /* Rewrite the FIFO write pointer */
839 tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
840 tx_queue = efx_channel_get_tx_queue(
841 channel, tx_ev_q_label % EFX_TXQ_TYPES);
843 netif_tx_lock(efx->net_dev);
844 efx_farch_notify_tx_desc(tx_queue);
845 netif_tx_unlock(efx->net_dev);
846 } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR)) {
847 efx_schedule_reset(efx, RESET_TYPE_DMA_ERROR);
849 netif_err(efx, tx_err, efx->net_dev,
850 "channel %d unexpected TX event "
851 EFX_QWORD_FMT"\n", channel->channel,
852 EFX_QWORD_VAL(*event));
856 /* Detect errors included in the rx_evt_pkt_ok bit. */
857 static u16 efx_farch_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
858 const efx_qword_t *event)
860 struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
861 struct efx_nic *efx = rx_queue->efx;
862 bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
863 bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
864 bool rx_ev_frm_trunc, rx_ev_tobe_disc;
865 bool rx_ev_other_err, rx_ev_pause_frm;
867 rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
868 rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
869 FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
870 rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
871 FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
872 rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
873 FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
874 rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
875 rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
876 rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
878 /* Every error apart from tobe_disc and pause_frm */
879 rx_ev_other_err = (rx_ev_tcp_udp_chksum_err |
880 rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
881 rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
883 /* Count errors that are not in MAC stats. Ignore expected
884 * checksum errors during self-test. */
886 ++channel->n_rx_frm_trunc;
887 else if (rx_ev_tobe_disc)
888 ++channel->n_rx_tobe_disc;
889 else if (!efx->loopback_selftest) {
890 if (rx_ev_ip_hdr_chksum_err)
891 ++channel->n_rx_ip_hdr_chksum_err;
892 else if (rx_ev_tcp_udp_chksum_err)
893 ++channel->n_rx_tcp_udp_chksum_err;
896 /* TOBE_DISC is expected on unicast mismatches; don't print out an
897 * error message. FRM_TRUNC indicates RXDP dropped the packet due
898 * to a FIFO overflow.
901 if (rx_ev_other_err && net_ratelimit()) {
902 netif_dbg(efx, rx_err, efx->net_dev,
903 " RX queue %d unexpected RX event "
904 EFX_QWORD_FMT "%s%s%s%s%s%s%s\n",
905 efx_rx_queue_index(rx_queue), EFX_QWORD_VAL(*event),
906 rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
907 rx_ev_ip_hdr_chksum_err ?
908 " [IP_HDR_CHKSUM_ERR]" : "",
909 rx_ev_tcp_udp_chksum_err ?
910 " [TCP_UDP_CHKSUM_ERR]" : "",
911 rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
912 rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
913 rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
914 rx_ev_pause_frm ? " [PAUSE]" : "");
917 (void) rx_ev_other_err;
920 if (efx->net_dev->features & NETIF_F_RXALL)
921 /* don't discard frame for CRC error */
922 rx_ev_eth_crc_err = false;
924 /* The frame must be discarded if any of these are true. */
925 return (rx_ev_eth_crc_err | rx_ev_frm_trunc |
926 rx_ev_tobe_disc | rx_ev_pause_frm) ?
927 EFX_RX_PKT_DISCARD : 0;
930 /* Handle receive events that are not in-order. Return true if this
931 * can be handled as a partial packet discard, false if it's more
935 efx_farch_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
937 struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
938 struct efx_nic *efx = rx_queue->efx;
939 unsigned expected, dropped;
941 if (rx_queue->scatter_n &&
942 index == ((rx_queue->removed_count + rx_queue->scatter_n - 1) &
943 rx_queue->ptr_mask)) {
944 ++channel->n_rx_nodesc_trunc;
948 expected = rx_queue->removed_count & rx_queue->ptr_mask;
949 dropped = (index - expected) & rx_queue->ptr_mask;
950 netif_info(efx, rx_err, efx->net_dev,
951 "dropped %d events (index=%d expected=%d)\n",
952 dropped, index, expected);
954 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
958 /* Handle a packet received event
960 * The NIC gives a "discard" flag if it's a unicast packet with the
961 * wrong destination address
962 * Also "is multicast" and "matches multicast filter" flags can be used to
963 * discard non-matching multicast packets.
966 efx_farch_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
968 unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
969 unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
970 unsigned expected_ptr;
971 bool rx_ev_pkt_ok, rx_ev_sop, rx_ev_cont;
973 struct efx_rx_queue *rx_queue;
974 struct efx_nic *efx = channel->efx;
976 if (unlikely(READ_ONCE(efx->reset_pending)))
979 rx_ev_cont = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT);
980 rx_ev_sop = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP);
981 WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
984 rx_queue = efx_channel_get_rx_queue(channel);
986 rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
987 expected_ptr = ((rx_queue->removed_count + rx_queue->scatter_n) &
990 /* Check for partial drops and other errors */
991 if (unlikely(rx_ev_desc_ptr != expected_ptr) ||
992 unlikely(rx_ev_sop != (rx_queue->scatter_n == 0))) {
993 if (rx_ev_desc_ptr != expected_ptr &&
994 !efx_farch_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr))
997 /* Discard all pending fragments */
998 if (rx_queue->scatter_n) {
1001 rx_queue->removed_count & rx_queue->ptr_mask,
1002 rx_queue->scatter_n, 0, EFX_RX_PKT_DISCARD);
1003 rx_queue->removed_count += rx_queue->scatter_n;
1004 rx_queue->scatter_n = 0;
1007 /* Return if there is no new fragment */
1008 if (rx_ev_desc_ptr != expected_ptr)
1011 /* Discard new fragment if not SOP */
1015 rx_queue->removed_count & rx_queue->ptr_mask,
1016 1, 0, EFX_RX_PKT_DISCARD);
1017 ++rx_queue->removed_count;
1022 ++rx_queue->scatter_n;
1026 rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
1027 rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
1028 rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
1030 if (likely(rx_ev_pkt_ok)) {
1031 /* If packet is marked as OK then we can rely on the
1032 * hardware checksum and classification.
1035 switch (rx_ev_hdr_type) {
1036 case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP:
1037 flags |= EFX_RX_PKT_TCP;
1039 case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP:
1040 flags |= EFX_RX_PKT_CSUMMED;
1042 case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_OTHER:
1043 case FSE_AZ_RX_EV_HDR_TYPE_OTHER:
1047 flags = efx_farch_handle_rx_not_ok(rx_queue, event);
1050 /* Detect multicast packets that didn't match the filter */
1051 rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
1052 if (rx_ev_mcast_pkt) {
1053 unsigned int rx_ev_mcast_hash_match =
1054 EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
1056 if (unlikely(!rx_ev_mcast_hash_match)) {
1057 ++channel->n_rx_mcast_mismatch;
1058 flags |= EFX_RX_PKT_DISCARD;
1062 channel->irq_mod_score += 2;
1064 /* Handle received packet */
1065 efx_rx_packet(rx_queue,
1066 rx_queue->removed_count & rx_queue->ptr_mask,
1067 rx_queue->scatter_n, rx_ev_byte_cnt, flags);
1068 rx_queue->removed_count += rx_queue->scatter_n;
1069 rx_queue->scatter_n = 0;
1072 /* If this flush done event corresponds to a &struct efx_tx_queue, then
1073 * send an %EFX_CHANNEL_MAGIC_TX_DRAIN event to drain the event queue
1074 * of all transmit completions.
1077 efx_farch_handle_tx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1079 struct efx_tx_queue *tx_queue;
1082 qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1083 if (qid < EFX_TXQ_TYPES * (efx->n_tx_channels + efx->n_extra_tx_channels)) {
1084 tx_queue = efx_get_tx_queue(efx, qid / EFX_TXQ_TYPES,
1085 qid % EFX_TXQ_TYPES);
1086 if (atomic_cmpxchg(&tx_queue->flush_outstanding, 1, 0)) {
1087 efx_farch_magic_event(tx_queue->channel,
1088 EFX_CHANNEL_MAGIC_TX_DRAIN(tx_queue));
1093 /* If this flush done event corresponds to a &struct efx_rx_queue: If the flush
1094 * was successful then send an %EFX_CHANNEL_MAGIC_RX_DRAIN, otherwise add
1095 * the RX queue back to the mask of RX queues in need of flushing.
1098 efx_farch_handle_rx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1100 struct efx_channel *channel;
1101 struct efx_rx_queue *rx_queue;
1105 qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
1106 failed = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
1107 if (qid >= efx->n_channels)
1109 channel = efx_get_channel(efx, qid);
1110 if (!efx_channel_has_rx_queue(channel))
1112 rx_queue = efx_channel_get_rx_queue(channel);
1115 netif_info(efx, hw, efx->net_dev,
1116 "RXQ %d flush retry\n", qid);
1117 rx_queue->flush_pending = true;
1118 atomic_inc(&efx->rxq_flush_pending);
1120 efx_farch_magic_event(efx_rx_queue_channel(rx_queue),
1121 EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue));
1123 atomic_dec(&efx->rxq_flush_outstanding);
1124 if (efx_farch_flush_wake(efx))
1125 wake_up(&efx->flush_wq);
1129 efx_farch_handle_drain_event(struct efx_channel *channel)
1131 struct efx_nic *efx = channel->efx;
1133 WARN_ON(atomic_read(&efx->active_queues) == 0);
1134 atomic_dec(&efx->active_queues);
1135 if (efx_farch_flush_wake(efx))
1136 wake_up(&efx->flush_wq);
1139 static void efx_farch_handle_generated_event(struct efx_channel *channel,
1142 struct efx_nic *efx = channel->efx;
1143 struct efx_rx_queue *rx_queue =
1144 efx_channel_has_rx_queue(channel) ?
1145 efx_channel_get_rx_queue(channel) : NULL;
1146 unsigned magic, code;
1148 magic = EFX_QWORD_FIELD(*event, FSF_AZ_DRV_GEN_EV_MAGIC);
1149 code = _EFX_CHANNEL_MAGIC_CODE(magic);
1151 if (magic == EFX_CHANNEL_MAGIC_TEST(channel)) {
1152 channel->event_test_cpu = raw_smp_processor_id();
1153 } else if (rx_queue && magic == EFX_CHANNEL_MAGIC_FILL(rx_queue)) {
1154 /* The queue must be empty, so we won't receive any rx
1155 * events, so efx_process_channel() won't refill the
1156 * queue. Refill it here */
1157 efx_fast_push_rx_descriptors(rx_queue, true);
1158 } else if (rx_queue && magic == EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue)) {
1159 efx_farch_handle_drain_event(channel);
1160 } else if (code == _EFX_CHANNEL_MAGIC_TX_DRAIN) {
1161 efx_farch_handle_drain_event(channel);
1163 netif_dbg(efx, hw, efx->net_dev, "channel %d received "
1164 "generated event "EFX_QWORD_FMT"\n",
1165 channel->channel, EFX_QWORD_VAL(*event));
1170 efx_farch_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
1172 struct efx_nic *efx = channel->efx;
1173 unsigned int ev_sub_code;
1174 unsigned int ev_sub_data;
1176 ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
1177 ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1179 switch (ev_sub_code) {
1180 case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
1181 netif_vdbg(efx, hw, efx->net_dev, "channel %d TXQ %d flushed\n",
1182 channel->channel, ev_sub_data);
1183 efx_farch_handle_tx_flush_done(efx, event);
1184 #ifdef CONFIG_SFC_SRIOV
1185 efx_siena_sriov_tx_flush_done(efx, event);
1188 case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
1189 netif_vdbg(efx, hw, efx->net_dev, "channel %d RXQ %d flushed\n",
1190 channel->channel, ev_sub_data);
1191 efx_farch_handle_rx_flush_done(efx, event);
1192 #ifdef CONFIG_SFC_SRIOV
1193 efx_siena_sriov_rx_flush_done(efx, event);
1196 case FSE_AZ_EVQ_INIT_DONE_EV:
1197 netif_dbg(efx, hw, efx->net_dev,
1198 "channel %d EVQ %d initialised\n",
1199 channel->channel, ev_sub_data);
1201 case FSE_AZ_SRM_UPD_DONE_EV:
1202 netif_vdbg(efx, hw, efx->net_dev,
1203 "channel %d SRAM update done\n", channel->channel);
1205 case FSE_AZ_WAKE_UP_EV:
1206 netif_vdbg(efx, hw, efx->net_dev,
1207 "channel %d RXQ %d wakeup event\n",
1208 channel->channel, ev_sub_data);
1210 case FSE_AZ_TIMER_EV:
1211 netif_vdbg(efx, hw, efx->net_dev,
1212 "channel %d RX queue %d timer expired\n",
1213 channel->channel, ev_sub_data);
1215 case FSE_AA_RX_RECOVER_EV:
1216 netif_err(efx, rx_err, efx->net_dev,
1217 "channel %d seen DRIVER RX_RESET event. "
1218 "Resetting.\n", channel->channel);
1219 atomic_inc(&efx->rx_reset);
1220 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1222 case FSE_BZ_RX_DSC_ERROR_EV:
1223 if (ev_sub_data < EFX_VI_BASE) {
1224 netif_err(efx, rx_err, efx->net_dev,
1225 "RX DMA Q %d reports descriptor fetch error."
1226 " RX Q %d is disabled.\n", ev_sub_data,
1228 efx_schedule_reset(efx, RESET_TYPE_DMA_ERROR);
1230 #ifdef CONFIG_SFC_SRIOV
1232 efx_siena_sriov_desc_fetch_err(efx, ev_sub_data);
1235 case FSE_BZ_TX_DSC_ERROR_EV:
1236 if (ev_sub_data < EFX_VI_BASE) {
1237 netif_err(efx, tx_err, efx->net_dev,
1238 "TX DMA Q %d reports descriptor fetch error."
1239 " TX Q %d is disabled.\n", ev_sub_data,
1241 efx_schedule_reset(efx, RESET_TYPE_DMA_ERROR);
1243 #ifdef CONFIG_SFC_SRIOV
1245 efx_siena_sriov_desc_fetch_err(efx, ev_sub_data);
1249 netif_vdbg(efx, hw, efx->net_dev,
1250 "channel %d unknown driver event code %d "
1251 "data %04x\n", channel->channel, ev_sub_code,
1257 int efx_farch_ev_process(struct efx_channel *channel, int budget)
1259 struct efx_nic *efx = channel->efx;
1260 unsigned int read_ptr;
1261 efx_qword_t event, *p_event;
1268 read_ptr = channel->eventq_read_ptr;
1271 p_event = efx_event(channel, read_ptr);
1274 if (!efx_event_present(&event))
1278 netif_vdbg(channel->efx, intr, channel->efx->net_dev,
1279 "channel %d event is "EFX_QWORD_FMT"\n",
1280 channel->channel, EFX_QWORD_VAL(event));
1282 /* Clear this event by marking it all ones */
1283 EFX_SET_QWORD(*p_event);
1287 ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
1290 case FSE_AZ_EV_CODE_RX_EV:
1291 efx_farch_handle_rx_event(channel, &event);
1292 if (++spent == budget)
1295 case FSE_AZ_EV_CODE_TX_EV:
1296 efx_farch_handle_tx_event(channel, &event);
1298 case FSE_AZ_EV_CODE_DRV_GEN_EV:
1299 efx_farch_handle_generated_event(channel, &event);
1301 case FSE_AZ_EV_CODE_DRIVER_EV:
1302 efx_farch_handle_driver_event(channel, &event);
1304 #ifdef CONFIG_SFC_SRIOV
1305 case FSE_CZ_EV_CODE_USER_EV:
1306 efx_siena_sriov_event(channel, &event);
1309 case FSE_CZ_EV_CODE_MCDI_EV:
1310 efx_mcdi_process_event(channel, &event);
1312 case FSE_AZ_EV_CODE_GLOBAL_EV:
1313 if (efx->type->handle_global_event &&
1314 efx->type->handle_global_event(channel, &event))
1318 netif_err(channel->efx, hw, channel->efx->net_dev,
1319 "channel %d unknown event type %d (data "
1320 EFX_QWORD_FMT ")\n", channel->channel,
1321 ev_code, EFX_QWORD_VAL(event));
1326 channel->eventq_read_ptr = read_ptr;
1330 /* Allocate buffer table entries for event queue */
1331 int efx_farch_ev_probe(struct efx_channel *channel)
1333 struct efx_nic *efx = channel->efx;
1336 entries = channel->eventq_mask + 1;
1337 return efx_alloc_special_buffer(efx, &channel->eventq,
1338 entries * sizeof(efx_qword_t));
1341 int efx_farch_ev_init(struct efx_channel *channel)
1344 struct efx_nic *efx = channel->efx;
1346 netif_dbg(efx, hw, efx->net_dev,
1347 "channel %d event queue in special buffers %d-%d\n",
1348 channel->channel, channel->eventq.index,
1349 channel->eventq.index + channel->eventq.entries - 1);
1351 EFX_POPULATE_OWORD_3(reg,
1352 FRF_CZ_TIMER_Q_EN, 1,
1353 FRF_CZ_HOST_NOTIFY_MODE, 0,
1354 FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
1355 efx_writeo_table(efx, ®, FR_BZ_TIMER_TBL, channel->channel);
1357 /* Pin event queue buffer */
1358 efx_init_special_buffer(efx, &channel->eventq);
1360 /* Fill event queue with all ones (i.e. empty events) */
1361 memset(channel->eventq.buf.addr, 0xff, channel->eventq.buf.len);
1363 /* Push event queue to card */
1364 EFX_POPULATE_OWORD_3(reg,
1366 FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
1367 FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
1368 efx_writeo_table(efx, ®, efx->type->evq_ptr_tbl_base,
1374 void efx_farch_ev_fini(struct efx_channel *channel)
1377 struct efx_nic *efx = channel->efx;
1379 /* Remove event queue from card */
1380 EFX_ZERO_OWORD(reg);
1381 efx_writeo_table(efx, ®, efx->type->evq_ptr_tbl_base,
1383 efx_writeo_table(efx, ®, FR_BZ_TIMER_TBL, channel->channel);
1385 /* Unpin event queue */
1386 efx_fini_special_buffer(efx, &channel->eventq);
1389 /* Free buffers backing event queue */
1390 void efx_farch_ev_remove(struct efx_channel *channel)
1392 efx_free_special_buffer(channel->efx, &channel->eventq);
1396 void efx_farch_ev_test_generate(struct efx_channel *channel)
1398 efx_farch_magic_event(channel, EFX_CHANNEL_MAGIC_TEST(channel));
1401 void efx_farch_rx_defer_refill(struct efx_rx_queue *rx_queue)
1403 efx_farch_magic_event(efx_rx_queue_channel(rx_queue),
1404 EFX_CHANNEL_MAGIC_FILL(rx_queue));
1407 /**************************************************************************
1409 * Hardware interrupts
1410 * The hardware interrupt handler does very little work; all the event
1411 * queue processing is carried out by per-channel tasklets.
1413 **************************************************************************/
1415 /* Enable/disable/generate interrupts */
1416 static inline void efx_farch_interrupts(struct efx_nic *efx,
1417 bool enabled, bool force)
1419 efx_oword_t int_en_reg_ker;
1421 EFX_POPULATE_OWORD_3(int_en_reg_ker,
1422 FRF_AZ_KER_INT_LEVE_SEL, efx->irq_level,
1423 FRF_AZ_KER_INT_KER, force,
1424 FRF_AZ_DRV_INT_EN_KER, enabled);
1425 efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
1428 void efx_farch_irq_enable_master(struct efx_nic *efx)
1430 EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
1431 wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
1433 efx_farch_interrupts(efx, true, false);
1436 void efx_farch_irq_disable_master(struct efx_nic *efx)
1438 /* Disable interrupts */
1439 efx_farch_interrupts(efx, false, false);
1442 /* Generate a test interrupt
1443 * Interrupt must already have been enabled, otherwise nasty things
1446 int efx_farch_irq_test_generate(struct efx_nic *efx)
1448 efx_farch_interrupts(efx, true, true);
1452 /* Process a fatal interrupt
1453 * Disable bus mastering ASAP and schedule a reset
1455 irqreturn_t efx_farch_fatal_interrupt(struct efx_nic *efx)
1457 efx_oword_t *int_ker = efx->irq_status.addr;
1458 efx_oword_t fatal_intr;
1459 int error, mem_perr;
1461 efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
1462 error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
1464 netif_err(efx, hw, efx->net_dev, "SYSTEM ERROR "EFX_OWORD_FMT" status "
1465 EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
1466 EFX_OWORD_VAL(fatal_intr),
1467 error ? "disabling bus mastering" : "no recognised error");
1469 /* If this is a memory parity error dump which blocks are offending */
1470 mem_perr = (EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER) ||
1471 EFX_OWORD_FIELD(fatal_intr, FRF_AZ_SRM_PERR_INT_KER));
1474 efx_reado(efx, ®, FR_AZ_MEM_STAT);
1475 netif_err(efx, hw, efx->net_dev,
1476 "SYSTEM ERROR: memory parity error "EFX_OWORD_FMT"\n",
1477 EFX_OWORD_VAL(reg));
1480 /* Disable both devices */
1481 pci_clear_master(efx->pci_dev);
1482 efx_farch_irq_disable_master(efx);
1484 /* Count errors and reset or disable the NIC accordingly */
1485 if (efx->int_error_count == 0 ||
1486 time_after(jiffies, efx->int_error_expire)) {
1487 efx->int_error_count = 0;
1488 efx->int_error_expire =
1489 jiffies + EFX_INT_ERROR_EXPIRE * HZ;
1491 if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
1492 netif_err(efx, hw, efx->net_dev,
1493 "SYSTEM ERROR - reset scheduled\n");
1494 efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
1496 netif_err(efx, hw, efx->net_dev,
1497 "SYSTEM ERROR - max number of errors seen."
1498 "NIC will be disabled\n");
1499 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1505 /* Handle a legacy interrupt
1506 * Acknowledges the interrupt and schedule event queue processing.
1508 irqreturn_t efx_farch_legacy_interrupt(int irq, void *dev_id)
1510 struct efx_nic *efx = dev_id;
1511 bool soft_enabled = READ_ONCE(efx->irq_soft_enabled);
1512 efx_oword_t *int_ker = efx->irq_status.addr;
1513 irqreturn_t result = IRQ_NONE;
1514 struct efx_channel *channel;
1519 /* Read the ISR which also ACKs the interrupts */
1520 efx_readd(efx, ®, FR_BZ_INT_ISR0);
1521 queues = EFX_EXTRACT_DWORD(reg, 0, 31);
1523 /* Legacy interrupts are disabled too late by the EEH kernel
1524 * code. Disable them earlier.
1525 * If an EEH error occurred, the read will have returned all ones.
1527 if (EFX_DWORD_IS_ALL_ONES(reg) && efx_try_recovery(efx) &&
1528 !efx->eeh_disabled_legacy_irq) {
1529 disable_irq_nosync(efx->legacy_irq);
1530 efx->eeh_disabled_legacy_irq = true;
1533 /* Handle non-event-queue sources */
1534 if (queues & (1U << efx->irq_level) && soft_enabled) {
1535 syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1536 if (unlikely(syserr))
1537 return efx_farch_fatal_interrupt(efx);
1538 efx->last_irq_cpu = raw_smp_processor_id();
1542 efx->irq_zero_count = 0;
1544 /* Schedule processing of any interrupting queues */
1545 if (likely(soft_enabled)) {
1546 efx_for_each_channel(channel, efx) {
1548 efx_schedule_channel_irq(channel);
1552 result = IRQ_HANDLED;
1557 /* Legacy ISR read can return zero once (SF bug 15783) */
1559 /* We can't return IRQ_HANDLED more than once on seeing ISR=0
1560 * because this might be a shared interrupt. */
1561 if (efx->irq_zero_count++ == 0)
1562 result = IRQ_HANDLED;
1564 /* Ensure we schedule or rearm all event queues */
1565 if (likely(soft_enabled)) {
1566 efx_for_each_channel(channel, efx) {
1567 event = efx_event(channel,
1568 channel->eventq_read_ptr);
1569 if (efx_event_present(event))
1570 efx_schedule_channel_irq(channel);
1572 efx_farch_ev_read_ack(channel);
1577 if (result == IRQ_HANDLED)
1578 netif_vdbg(efx, intr, efx->net_dev,
1579 "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
1580 irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
1585 /* Handle an MSI interrupt
1587 * Handle an MSI hardware interrupt. This routine schedules event
1588 * queue processing. No interrupt acknowledgement cycle is necessary.
1589 * Also, we never need to check that the interrupt is for us, since
1590 * MSI interrupts cannot be shared.
1592 irqreturn_t efx_farch_msi_interrupt(int irq, void *dev_id)
1594 struct efx_msi_context *context = dev_id;
1595 struct efx_nic *efx = context->efx;
1596 efx_oword_t *int_ker = efx->irq_status.addr;
1599 netif_vdbg(efx, intr, efx->net_dev,
1600 "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
1601 irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
1603 if (!likely(READ_ONCE(efx->irq_soft_enabled)))
1606 /* Handle non-event-queue sources */
1607 if (context->index == efx->irq_level) {
1608 syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1609 if (unlikely(syserr))
1610 return efx_farch_fatal_interrupt(efx);
1611 efx->last_irq_cpu = raw_smp_processor_id();
1614 /* Schedule processing of the channel */
1615 efx_schedule_channel_irq(efx->channel[context->index]);
1620 /* Setup RSS indirection table.
1621 * This maps from the hash value of the packet to RXQ
1623 void efx_farch_rx_push_indir_table(struct efx_nic *efx)
1628 BUILD_BUG_ON(ARRAY_SIZE(efx->rss_context.rx_indir_table) !=
1629 FR_BZ_RX_INDIRECTION_TBL_ROWS);
1631 for (i = 0; i < FR_BZ_RX_INDIRECTION_TBL_ROWS; i++) {
1632 EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
1633 efx->rss_context.rx_indir_table[i]);
1634 efx_writed(efx, &dword,
1635 FR_BZ_RX_INDIRECTION_TBL +
1636 FR_BZ_RX_INDIRECTION_TBL_STEP * i);
1640 void efx_farch_rx_pull_indir_table(struct efx_nic *efx)
1645 BUILD_BUG_ON(ARRAY_SIZE(efx->rss_context.rx_indir_table) !=
1646 FR_BZ_RX_INDIRECTION_TBL_ROWS);
1648 for (i = 0; i < FR_BZ_RX_INDIRECTION_TBL_ROWS; i++) {
1649 efx_readd(efx, &dword,
1650 FR_BZ_RX_INDIRECTION_TBL +
1651 FR_BZ_RX_INDIRECTION_TBL_STEP * i);
1652 efx->rss_context.rx_indir_table[i] = EFX_DWORD_FIELD(dword, FRF_BZ_IT_QUEUE);
1656 /* Looks at available SRAM resources and works out how many queues we
1657 * can support, and where things like descriptor caches should live.
1659 * SRAM is split up as follows:
1660 * 0 buftbl entries for channels
1661 * efx->vf_buftbl_base buftbl entries for SR-IOV
1662 * efx->rx_dc_base RX descriptor caches
1663 * efx->tx_dc_base TX descriptor caches
1665 void efx_farch_dimension_resources(struct efx_nic *efx, unsigned sram_lim_qw)
1667 unsigned vi_count, buftbl_min, total_tx_channels;
1669 #ifdef CONFIG_SFC_SRIOV
1670 struct siena_nic_data *nic_data = efx->nic_data;
1673 total_tx_channels = efx->n_tx_channels + efx->n_extra_tx_channels;
1674 /* Account for the buffer table entries backing the datapath channels
1675 * and the descriptor caches for those channels.
1677 buftbl_min = ((efx->n_rx_channels * EFX_MAX_DMAQ_SIZE +
1678 total_tx_channels * EFX_TXQ_TYPES * EFX_MAX_DMAQ_SIZE +
1679 efx->n_channels * EFX_MAX_EVQ_SIZE)
1680 * sizeof(efx_qword_t) / EFX_BUF_SIZE);
1681 vi_count = max(efx->n_channels, total_tx_channels * EFX_TXQ_TYPES);
1683 #ifdef CONFIG_SFC_SRIOV
1684 if (efx->type->sriov_wanted) {
1685 if (efx->type->sriov_wanted(efx)) {
1686 unsigned vi_dc_entries, buftbl_free;
1687 unsigned entries_per_vf, vf_limit;
1689 nic_data->vf_buftbl_base = buftbl_min;
1691 vi_dc_entries = RX_DC_ENTRIES + TX_DC_ENTRIES;
1692 vi_count = max(vi_count, EFX_VI_BASE);
1693 buftbl_free = (sram_lim_qw - buftbl_min -
1694 vi_count * vi_dc_entries);
1696 entries_per_vf = ((vi_dc_entries +
1697 EFX_VF_BUFTBL_PER_VI) *
1699 vf_limit = min(buftbl_free / entries_per_vf,
1700 (1024U - EFX_VI_BASE) >> efx->vi_scale);
1702 if (efx->vf_count > vf_limit) {
1703 netif_err(efx, probe, efx->net_dev,
1704 "Reducing VF count from from %d to %d\n",
1705 efx->vf_count, vf_limit);
1706 efx->vf_count = vf_limit;
1708 vi_count += efx->vf_count * efx_vf_size(efx);
1713 efx->tx_dc_base = sram_lim_qw - vi_count * TX_DC_ENTRIES;
1714 efx->rx_dc_base = efx->tx_dc_base - vi_count * RX_DC_ENTRIES;
1717 u32 efx_farch_fpga_ver(struct efx_nic *efx)
1719 efx_oword_t altera_build;
1720 efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
1721 return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
1724 void efx_farch_init_common(struct efx_nic *efx)
1728 /* Set positions of descriptor caches in SRAM. */
1729 EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR, efx->tx_dc_base);
1730 efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
1731 EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR, efx->rx_dc_base);
1732 efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
1734 /* Set TX descriptor cache size. */
1735 BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
1736 EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
1737 efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
1739 /* Set RX descriptor cache size. Set low watermark to size-8, as
1740 * this allows most efficient prefetching.
1742 BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
1743 EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
1744 efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
1745 EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
1746 efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
1748 /* Program INT_KER address */
1749 EFX_POPULATE_OWORD_2(temp,
1750 FRF_AZ_NORM_INT_VEC_DIS_KER,
1751 EFX_INT_MODE_USE_MSI(efx),
1752 FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
1753 efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
1755 if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
1756 /* Use an interrupt level unused by event queues */
1757 efx->irq_level = 0x1f;
1759 /* Use a valid MSI-X vector */
1762 /* Enable all the genuinely fatal interrupts. (They are still
1763 * masked by the overall interrupt mask, controlled by
1764 * falcon_interrupts()).
1766 * Note: All other fatal interrupts are enabled
1768 EFX_POPULATE_OWORD_3(temp,
1769 FRF_AZ_ILL_ADR_INT_KER_EN, 1,
1770 FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
1771 FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
1772 EFX_SET_OWORD_FIELD(temp, FRF_CZ_SRAM_PERR_INT_P_KER_EN, 1);
1773 EFX_INVERT_OWORD(temp);
1774 efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
1776 /* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
1777 * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
1779 efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
1780 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
1781 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
1782 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
1783 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 1);
1784 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
1785 /* Enable SW_EV to inherit in char driver - assume harmless here */
1786 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
1787 /* Prefetch threshold 2 => fetch when descriptor cache half empty */
1788 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
1789 /* Disable hardware watchdog which can misfire */
1790 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
1791 /* Squash TX of packets of 16 bytes or less */
1792 EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
1793 efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
1795 EFX_POPULATE_OWORD_4(temp,
1796 /* Default values */
1797 FRF_BZ_TX_PACE_SB_NOT_AF, 0x15,
1798 FRF_BZ_TX_PACE_SB_AF, 0xb,
1799 FRF_BZ_TX_PACE_FB_BASE, 0,
1800 /* Allow large pace values in the fast bin. */
1801 FRF_BZ_TX_PACE_BIN_TH,
1802 FFE_BZ_TX_PACE_RESERVED);
1803 efx_writeo(efx, &temp, FR_BZ_TX_PACE);
1806 /**************************************************************************
1810 **************************************************************************
1813 /* "Fudge factors" - difference between programmed value and actual depth.
1814 * Due to pipelined implementation we need to program H/W with a value that
1815 * is larger than the hop limit we want.
1817 #define EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD 3
1818 #define EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL 1
1820 /* Hard maximum search limit. Hardware will time-out beyond 200-something.
1821 * We also need to avoid infinite loops in efx_farch_filter_search() when the
1824 #define EFX_FARCH_FILTER_CTL_SRCH_MAX 200
1826 /* Don't try very hard to find space for performance hints, as this is
1827 * counter-productive. */
1828 #define EFX_FARCH_FILTER_CTL_SRCH_HINT_MAX 5
1830 enum efx_farch_filter_type {
1831 EFX_FARCH_FILTER_TCP_FULL = 0,
1832 EFX_FARCH_FILTER_TCP_WILD,
1833 EFX_FARCH_FILTER_UDP_FULL,
1834 EFX_FARCH_FILTER_UDP_WILD,
1835 EFX_FARCH_FILTER_MAC_FULL = 4,
1836 EFX_FARCH_FILTER_MAC_WILD,
1837 EFX_FARCH_FILTER_UC_DEF = 8,
1838 EFX_FARCH_FILTER_MC_DEF,
1839 EFX_FARCH_FILTER_TYPE_COUNT, /* number of specific types */
1842 enum efx_farch_filter_table_id {
1843 EFX_FARCH_FILTER_TABLE_RX_IP = 0,
1844 EFX_FARCH_FILTER_TABLE_RX_MAC,
1845 EFX_FARCH_FILTER_TABLE_RX_DEF,
1846 EFX_FARCH_FILTER_TABLE_TX_MAC,
1847 EFX_FARCH_FILTER_TABLE_COUNT,
1850 enum efx_farch_filter_index {
1851 EFX_FARCH_FILTER_INDEX_UC_DEF,
1852 EFX_FARCH_FILTER_INDEX_MC_DEF,
1853 EFX_FARCH_FILTER_SIZE_RX_DEF,
1856 struct efx_farch_filter_spec {
1864 struct efx_farch_filter_table {
1865 enum efx_farch_filter_table_id id;
1866 u32 offset; /* address of table relative to BAR */
1867 unsigned size; /* number of entries */
1868 unsigned step; /* step between entries */
1869 unsigned used; /* number currently used */
1870 unsigned long *used_bitmap;
1871 struct efx_farch_filter_spec *spec;
1872 unsigned search_limit[EFX_FARCH_FILTER_TYPE_COUNT];
1875 struct efx_farch_filter_state {
1876 struct rw_semaphore lock; /* Protects table contents */
1877 struct efx_farch_filter_table table[EFX_FARCH_FILTER_TABLE_COUNT];
1881 efx_farch_filter_table_clear_entry(struct efx_nic *efx,
1882 struct efx_farch_filter_table *table,
1883 unsigned int filter_idx);
1885 /* The filter hash function is LFSR polynomial x^16 + x^3 + 1 of a 32-bit
1886 * key derived from the n-tuple. The initial LFSR state is 0xffff. */
1887 static u16 efx_farch_filter_hash(u32 key)
1891 /* First 16 rounds */
1892 tmp = 0x1fff ^ key >> 16;
1893 tmp = tmp ^ tmp >> 3 ^ tmp >> 6;
1894 tmp = tmp ^ tmp >> 9;
1895 /* Last 16 rounds */
1896 tmp = tmp ^ tmp << 13 ^ key;
1897 tmp = tmp ^ tmp >> 3 ^ tmp >> 6;
1898 return tmp ^ tmp >> 9;
1901 /* To allow for hash collisions, filter search continues at these
1902 * increments from the first possible entry selected by the hash. */
1903 static u16 efx_farch_filter_increment(u32 key)
1908 static enum efx_farch_filter_table_id
1909 efx_farch_filter_spec_table_id(const struct efx_farch_filter_spec *spec)
1911 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1912 (EFX_FARCH_FILTER_TCP_FULL >> 2));
1913 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1914 (EFX_FARCH_FILTER_TCP_WILD >> 2));
1915 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1916 (EFX_FARCH_FILTER_UDP_FULL >> 2));
1917 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1918 (EFX_FARCH_FILTER_UDP_WILD >> 2));
1919 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_MAC !=
1920 (EFX_FARCH_FILTER_MAC_FULL >> 2));
1921 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_MAC !=
1922 (EFX_FARCH_FILTER_MAC_WILD >> 2));
1923 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_TX_MAC !=
1924 EFX_FARCH_FILTER_TABLE_RX_MAC + 2);
1925 return (spec->type >> 2) + ((spec->flags & EFX_FILTER_FLAG_TX) ? 2 : 0);
1928 static void efx_farch_filter_push_rx_config(struct efx_nic *efx)
1930 struct efx_farch_filter_state *state = efx->filter_state;
1931 struct efx_farch_filter_table *table;
1932 efx_oword_t filter_ctl;
1934 efx_reado(efx, &filter_ctl, FR_BZ_RX_FILTER_CTL);
1936 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_IP];
1937 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_TCP_FULL_SRCH_LIMIT,
1938 table->search_limit[EFX_FARCH_FILTER_TCP_FULL] +
1939 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
1940 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_TCP_WILD_SRCH_LIMIT,
1941 table->search_limit[EFX_FARCH_FILTER_TCP_WILD] +
1942 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
1943 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_UDP_FULL_SRCH_LIMIT,
1944 table->search_limit[EFX_FARCH_FILTER_UDP_FULL] +
1945 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
1946 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_UDP_WILD_SRCH_LIMIT,
1947 table->search_limit[EFX_FARCH_FILTER_UDP_WILD] +
1948 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
1950 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_MAC];
1952 EFX_SET_OWORD_FIELD(
1953 filter_ctl, FRF_CZ_ETHERNET_FULL_SEARCH_LIMIT,
1954 table->search_limit[EFX_FARCH_FILTER_MAC_FULL] +
1955 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
1956 EFX_SET_OWORD_FIELD(
1957 filter_ctl, FRF_CZ_ETHERNET_WILDCARD_SEARCH_LIMIT,
1958 table->search_limit[EFX_FARCH_FILTER_MAC_WILD] +
1959 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
1962 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_DEF];
1964 EFX_SET_OWORD_FIELD(
1965 filter_ctl, FRF_CZ_UNICAST_NOMATCH_Q_ID,
1966 table->spec[EFX_FARCH_FILTER_INDEX_UC_DEF].dmaq_id);
1967 EFX_SET_OWORD_FIELD(
1968 filter_ctl, FRF_CZ_UNICAST_NOMATCH_RSS_ENABLED,
1969 !!(table->spec[EFX_FARCH_FILTER_INDEX_UC_DEF].flags &
1970 EFX_FILTER_FLAG_RX_RSS));
1971 EFX_SET_OWORD_FIELD(
1972 filter_ctl, FRF_CZ_MULTICAST_NOMATCH_Q_ID,
1973 table->spec[EFX_FARCH_FILTER_INDEX_MC_DEF].dmaq_id);
1974 EFX_SET_OWORD_FIELD(
1975 filter_ctl, FRF_CZ_MULTICAST_NOMATCH_RSS_ENABLED,
1976 !!(table->spec[EFX_FARCH_FILTER_INDEX_MC_DEF].flags &
1977 EFX_FILTER_FLAG_RX_RSS));
1979 /* There is a single bit to enable RX scatter for all
1980 * unmatched packets. Only set it if scatter is
1981 * enabled in both filter specs.
1983 EFX_SET_OWORD_FIELD(
1984 filter_ctl, FRF_BZ_SCATTER_ENBL_NO_MATCH_Q,
1985 !!(table->spec[EFX_FARCH_FILTER_INDEX_UC_DEF].flags &
1986 table->spec[EFX_FARCH_FILTER_INDEX_MC_DEF].flags &
1987 EFX_FILTER_FLAG_RX_SCATTER));
1989 /* We don't expose 'default' filters because unmatched
1990 * packets always go to the queue number found in the
1991 * RSS table. But we still need to set the RX scatter
1994 EFX_SET_OWORD_FIELD(
1995 filter_ctl, FRF_BZ_SCATTER_ENBL_NO_MATCH_Q,
1999 efx_writeo(efx, &filter_ctl, FR_BZ_RX_FILTER_CTL);
2002 static void efx_farch_filter_push_tx_limits(struct efx_nic *efx)
2004 struct efx_farch_filter_state *state = efx->filter_state;
2005 struct efx_farch_filter_table *table;
2008 efx_reado(efx, &tx_cfg, FR_AZ_TX_CFG);
2010 table = &state->table[EFX_FARCH_FILTER_TABLE_TX_MAC];
2012 EFX_SET_OWORD_FIELD(
2013 tx_cfg, FRF_CZ_TX_ETH_FILTER_FULL_SEARCH_RANGE,
2014 table->search_limit[EFX_FARCH_FILTER_MAC_FULL] +
2015 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
2016 EFX_SET_OWORD_FIELD(
2017 tx_cfg, FRF_CZ_TX_ETH_FILTER_WILD_SEARCH_RANGE,
2018 table->search_limit[EFX_FARCH_FILTER_MAC_WILD] +
2019 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
2022 efx_writeo(efx, &tx_cfg, FR_AZ_TX_CFG);
2026 efx_farch_filter_from_gen_spec(struct efx_farch_filter_spec *spec,
2027 const struct efx_filter_spec *gen_spec)
2029 bool is_full = false;
2031 if ((gen_spec->flags & EFX_FILTER_FLAG_RX_RSS) && gen_spec->rss_context)
2034 spec->priority = gen_spec->priority;
2035 spec->flags = gen_spec->flags;
2036 spec->dmaq_id = gen_spec->dmaq_id;
2038 switch (gen_spec->match_flags) {
2039 case (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
2040 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
2041 EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT):
2044 case (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
2045 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT): {
2046 __be32 rhost, host1, host2;
2047 __be16 rport, port1, port2;
2049 EFX_WARN_ON_PARANOID(!(gen_spec->flags & EFX_FILTER_FLAG_RX));
2051 if (gen_spec->ether_type != htons(ETH_P_IP))
2052 return -EPROTONOSUPPORT;
2053 if (gen_spec->loc_port == 0 ||
2054 (is_full && gen_spec->rem_port == 0))
2055 return -EADDRNOTAVAIL;
2056 switch (gen_spec->ip_proto) {
2058 spec->type = (is_full ? EFX_FARCH_FILTER_TCP_FULL :
2059 EFX_FARCH_FILTER_TCP_WILD);
2062 spec->type = (is_full ? EFX_FARCH_FILTER_UDP_FULL :
2063 EFX_FARCH_FILTER_UDP_WILD);
2066 return -EPROTONOSUPPORT;
2069 /* Filter is constructed in terms of source and destination,
2070 * with the odd wrinkle that the ports are swapped in a UDP
2071 * wildcard filter. We need to convert from local and remote
2072 * (= zero for wildcard) addresses.
2074 rhost = is_full ? gen_spec->rem_host[0] : 0;
2075 rport = is_full ? gen_spec->rem_port : 0;
2077 host2 = gen_spec->loc_host[0];
2078 if (!is_full && gen_spec->ip_proto == IPPROTO_UDP) {
2079 port1 = gen_spec->loc_port;
2083 port2 = gen_spec->loc_port;
2085 spec->data[0] = ntohl(host1) << 16 | ntohs(port1);
2086 spec->data[1] = ntohs(port2) << 16 | ntohl(host1) >> 16;
2087 spec->data[2] = ntohl(host2);
2092 case EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_OUTER_VID:
2095 case EFX_FILTER_MATCH_LOC_MAC:
2096 spec->type = (is_full ? EFX_FARCH_FILTER_MAC_FULL :
2097 EFX_FARCH_FILTER_MAC_WILD);
2098 spec->data[0] = is_full ? ntohs(gen_spec->outer_vid) : 0;
2099 spec->data[1] = (gen_spec->loc_mac[2] << 24 |
2100 gen_spec->loc_mac[3] << 16 |
2101 gen_spec->loc_mac[4] << 8 |
2102 gen_spec->loc_mac[5]);
2103 spec->data[2] = (gen_spec->loc_mac[0] << 8 |
2104 gen_spec->loc_mac[1]);
2107 case EFX_FILTER_MATCH_LOC_MAC_IG:
2108 spec->type = (is_multicast_ether_addr(gen_spec->loc_mac) ?
2109 EFX_FARCH_FILTER_MC_DEF :
2110 EFX_FARCH_FILTER_UC_DEF);
2111 memset(spec->data, 0, sizeof(spec->data)); /* ensure equality */
2115 return -EPROTONOSUPPORT;
2122 efx_farch_filter_to_gen_spec(struct efx_filter_spec *gen_spec,
2123 const struct efx_farch_filter_spec *spec)
2125 bool is_full = false;
2127 /* *gen_spec should be completely initialised, to be consistent
2128 * with efx_filter_init_{rx,tx}() and in case we want to copy
2129 * it back to userland.
2131 memset(gen_spec, 0, sizeof(*gen_spec));
2133 gen_spec->priority = spec->priority;
2134 gen_spec->flags = spec->flags;
2135 gen_spec->dmaq_id = spec->dmaq_id;
2137 switch (spec->type) {
2138 case EFX_FARCH_FILTER_TCP_FULL:
2139 case EFX_FARCH_FILTER_UDP_FULL:
2142 case EFX_FARCH_FILTER_TCP_WILD:
2143 case EFX_FARCH_FILTER_UDP_WILD: {
2144 __be32 host1, host2;
2145 __be16 port1, port2;
2147 gen_spec->match_flags =
2148 EFX_FILTER_MATCH_ETHER_TYPE |
2149 EFX_FILTER_MATCH_IP_PROTO |
2150 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT;
2152 gen_spec->match_flags |= (EFX_FILTER_MATCH_REM_HOST |
2153 EFX_FILTER_MATCH_REM_PORT);
2154 gen_spec->ether_type = htons(ETH_P_IP);
2155 gen_spec->ip_proto =
2156 (spec->type == EFX_FARCH_FILTER_TCP_FULL ||
2157 spec->type == EFX_FARCH_FILTER_TCP_WILD) ?
2158 IPPROTO_TCP : IPPROTO_UDP;
2160 host1 = htonl(spec->data[0] >> 16 | spec->data[1] << 16);
2161 port1 = htons(spec->data[0]);
2162 host2 = htonl(spec->data[2]);
2163 port2 = htons(spec->data[1] >> 16);
2164 if (spec->flags & EFX_FILTER_FLAG_TX) {
2165 gen_spec->loc_host[0] = host1;
2166 gen_spec->rem_host[0] = host2;
2168 gen_spec->loc_host[0] = host2;
2169 gen_spec->rem_host[0] = host1;
2171 if (!!(gen_spec->flags & EFX_FILTER_FLAG_TX) ^
2172 (!is_full && gen_spec->ip_proto == IPPROTO_UDP)) {
2173 gen_spec->loc_port = port1;
2174 gen_spec->rem_port = port2;
2176 gen_spec->loc_port = port2;
2177 gen_spec->rem_port = port1;
2183 case EFX_FARCH_FILTER_MAC_FULL:
2186 case EFX_FARCH_FILTER_MAC_WILD:
2187 gen_spec->match_flags = EFX_FILTER_MATCH_LOC_MAC;
2189 gen_spec->match_flags |= EFX_FILTER_MATCH_OUTER_VID;
2190 gen_spec->loc_mac[0] = spec->data[2] >> 8;
2191 gen_spec->loc_mac[1] = spec->data[2];
2192 gen_spec->loc_mac[2] = spec->data[1] >> 24;
2193 gen_spec->loc_mac[3] = spec->data[1] >> 16;
2194 gen_spec->loc_mac[4] = spec->data[1] >> 8;
2195 gen_spec->loc_mac[5] = spec->data[1];
2196 gen_spec->outer_vid = htons(spec->data[0]);
2199 case EFX_FARCH_FILTER_UC_DEF:
2200 case EFX_FARCH_FILTER_MC_DEF:
2201 gen_spec->match_flags = EFX_FILTER_MATCH_LOC_MAC_IG;
2202 gen_spec->loc_mac[0] = spec->type == EFX_FARCH_FILTER_MC_DEF;
2212 efx_farch_filter_init_rx_auto(struct efx_nic *efx,
2213 struct efx_farch_filter_spec *spec)
2215 /* If there's only one channel then disable RSS for non VF
2216 * traffic, thereby allowing VFs to use RSS when the PF can't.
2218 spec->priority = EFX_FILTER_PRI_AUTO;
2219 spec->flags = (EFX_FILTER_FLAG_RX |
2220 (efx_rss_enabled(efx) ? EFX_FILTER_FLAG_RX_RSS : 0) |
2221 (efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0));
2225 /* Build a filter entry and return its n-tuple key. */
2226 static u32 efx_farch_filter_build(efx_oword_t *filter,
2227 struct efx_farch_filter_spec *spec)
2231 switch (efx_farch_filter_spec_table_id(spec)) {
2232 case EFX_FARCH_FILTER_TABLE_RX_IP: {
2233 bool is_udp = (spec->type == EFX_FARCH_FILTER_UDP_FULL ||
2234 spec->type == EFX_FARCH_FILTER_UDP_WILD);
2235 EFX_POPULATE_OWORD_7(
2238 !!(spec->flags & EFX_FILTER_FLAG_RX_RSS),
2240 !!(spec->flags & EFX_FILTER_FLAG_RX_SCATTER),
2241 FRF_BZ_TCP_UDP, is_udp,
2242 FRF_BZ_RXQ_ID, spec->dmaq_id,
2243 EFX_DWORD_2, spec->data[2],
2244 EFX_DWORD_1, spec->data[1],
2245 EFX_DWORD_0, spec->data[0]);
2250 case EFX_FARCH_FILTER_TABLE_RX_MAC: {
2251 bool is_wild = spec->type == EFX_FARCH_FILTER_MAC_WILD;
2252 EFX_POPULATE_OWORD_7(
2255 !!(spec->flags & EFX_FILTER_FLAG_RX_RSS),
2256 FRF_CZ_RMFT_SCATTER_EN,
2257 !!(spec->flags & EFX_FILTER_FLAG_RX_SCATTER),
2258 FRF_CZ_RMFT_RXQ_ID, spec->dmaq_id,
2259 FRF_CZ_RMFT_WILDCARD_MATCH, is_wild,
2260 FRF_CZ_RMFT_DEST_MAC_HI, spec->data[2],
2261 FRF_CZ_RMFT_DEST_MAC_LO, spec->data[1],
2262 FRF_CZ_RMFT_VLAN_ID, spec->data[0]);
2267 case EFX_FARCH_FILTER_TABLE_TX_MAC: {
2268 bool is_wild = spec->type == EFX_FARCH_FILTER_MAC_WILD;
2269 EFX_POPULATE_OWORD_5(*filter,
2270 FRF_CZ_TMFT_TXQ_ID, spec->dmaq_id,
2271 FRF_CZ_TMFT_WILDCARD_MATCH, is_wild,
2272 FRF_CZ_TMFT_SRC_MAC_HI, spec->data[2],
2273 FRF_CZ_TMFT_SRC_MAC_LO, spec->data[1],
2274 FRF_CZ_TMFT_VLAN_ID, spec->data[0]);
2275 data3 = is_wild | spec->dmaq_id << 1;
2283 return spec->data[0] ^ spec->data[1] ^ spec->data[2] ^ data3;
2286 static bool efx_farch_filter_equal(const struct efx_farch_filter_spec *left,
2287 const struct efx_farch_filter_spec *right)
2289 if (left->type != right->type ||
2290 memcmp(left->data, right->data, sizeof(left->data)))
2293 if (left->flags & EFX_FILTER_FLAG_TX &&
2294 left->dmaq_id != right->dmaq_id)
2301 * Construct/deconstruct external filter IDs. At least the RX filter
2302 * IDs must be ordered by matching priority, for RX NFC semantics.
2304 * Deconstruction needs to be robust against invalid IDs so that
2305 * efx_filter_remove_id_safe() and efx_filter_get_filter_safe() can
2306 * accept user-provided IDs.
2309 #define EFX_FARCH_FILTER_MATCH_PRI_COUNT 5
2311 static const u8 efx_farch_filter_type_match_pri[EFX_FARCH_FILTER_TYPE_COUNT] = {
2312 [EFX_FARCH_FILTER_TCP_FULL] = 0,
2313 [EFX_FARCH_FILTER_UDP_FULL] = 0,
2314 [EFX_FARCH_FILTER_TCP_WILD] = 1,
2315 [EFX_FARCH_FILTER_UDP_WILD] = 1,
2316 [EFX_FARCH_FILTER_MAC_FULL] = 2,
2317 [EFX_FARCH_FILTER_MAC_WILD] = 3,
2318 [EFX_FARCH_FILTER_UC_DEF] = 4,
2319 [EFX_FARCH_FILTER_MC_DEF] = 4,
2322 static const enum efx_farch_filter_table_id efx_farch_filter_range_table[] = {
2323 EFX_FARCH_FILTER_TABLE_RX_IP, /* RX match pri 0 */
2324 EFX_FARCH_FILTER_TABLE_RX_IP,
2325 EFX_FARCH_FILTER_TABLE_RX_MAC,
2326 EFX_FARCH_FILTER_TABLE_RX_MAC,
2327 EFX_FARCH_FILTER_TABLE_RX_DEF, /* RX match pri 4 */
2328 EFX_FARCH_FILTER_TABLE_TX_MAC, /* TX match pri 0 */
2329 EFX_FARCH_FILTER_TABLE_TX_MAC, /* TX match pri 1 */
2332 #define EFX_FARCH_FILTER_INDEX_WIDTH 13
2333 #define EFX_FARCH_FILTER_INDEX_MASK ((1 << EFX_FARCH_FILTER_INDEX_WIDTH) - 1)
2336 efx_farch_filter_make_id(const struct efx_farch_filter_spec *spec,
2341 range = efx_farch_filter_type_match_pri[spec->type];
2342 if (!(spec->flags & EFX_FILTER_FLAG_RX))
2343 range += EFX_FARCH_FILTER_MATCH_PRI_COUNT;
2345 return range << EFX_FARCH_FILTER_INDEX_WIDTH | index;
2348 static inline enum efx_farch_filter_table_id
2349 efx_farch_filter_id_table_id(u32 id)
2351 unsigned int range = id >> EFX_FARCH_FILTER_INDEX_WIDTH;
2353 if (range < ARRAY_SIZE(efx_farch_filter_range_table))
2354 return efx_farch_filter_range_table[range];
2356 return EFX_FARCH_FILTER_TABLE_COUNT; /* invalid */
2359 static inline unsigned int efx_farch_filter_id_index(u32 id)
2361 return id & EFX_FARCH_FILTER_INDEX_MASK;
2364 u32 efx_farch_filter_get_rx_id_limit(struct efx_nic *efx)
2366 struct efx_farch_filter_state *state = efx->filter_state;
2367 unsigned int range = EFX_FARCH_FILTER_MATCH_PRI_COUNT - 1;
2368 enum efx_farch_filter_table_id table_id;
2371 table_id = efx_farch_filter_range_table[range];
2372 if (state->table[table_id].size != 0)
2373 return range << EFX_FARCH_FILTER_INDEX_WIDTH |
2374 state->table[table_id].size;
2380 s32 efx_farch_filter_insert(struct efx_nic *efx,
2381 struct efx_filter_spec *gen_spec,
2384 struct efx_farch_filter_state *state = efx->filter_state;
2385 struct efx_farch_filter_table *table;
2386 struct efx_farch_filter_spec spec;
2388 int rep_index, ins_index;
2389 unsigned int depth = 0;
2392 rc = efx_farch_filter_from_gen_spec(&spec, gen_spec);
2396 down_write(&state->lock);
2398 table = &state->table[efx_farch_filter_spec_table_id(&spec)];
2399 if (table->size == 0) {
2404 netif_vdbg(efx, hw, efx->net_dev,
2405 "%s: type %d search_limit=%d", __func__, spec.type,
2406 table->search_limit[spec.type]);
2408 if (table->id == EFX_FARCH_FILTER_TABLE_RX_DEF) {
2409 /* One filter spec per type */
2410 BUILD_BUG_ON(EFX_FARCH_FILTER_INDEX_UC_DEF != 0);
2411 BUILD_BUG_ON(EFX_FARCH_FILTER_INDEX_MC_DEF !=
2412 EFX_FARCH_FILTER_MC_DEF - EFX_FARCH_FILTER_UC_DEF);
2413 rep_index = spec.type - EFX_FARCH_FILTER_UC_DEF;
2414 ins_index = rep_index;
2416 /* Search concurrently for
2417 * (1) a filter to be replaced (rep_index): any filter
2418 * with the same match values, up to the current
2419 * search depth for this type, and
2420 * (2) the insertion point (ins_index): (1) or any
2421 * free slot before it or up to the maximum search
2422 * depth for this priority
2423 * We fail if we cannot find (2).
2425 * We can stop once either
2426 * (a) we find (1), in which case we have definitely
2427 * found (2) as well; or
2428 * (b) we have searched exhaustively for (1), and have
2429 * either found (2) or searched exhaustively for it
2431 u32 key = efx_farch_filter_build(&filter, &spec);
2432 unsigned int hash = efx_farch_filter_hash(key);
2433 unsigned int incr = efx_farch_filter_increment(key);
2434 unsigned int max_rep_depth = table->search_limit[spec.type];
2435 unsigned int max_ins_depth =
2436 spec.priority <= EFX_FILTER_PRI_HINT ?
2437 EFX_FARCH_FILTER_CTL_SRCH_HINT_MAX :
2438 EFX_FARCH_FILTER_CTL_SRCH_MAX;
2439 unsigned int i = hash & (table->size - 1);
2445 if (!test_bit(i, table->used_bitmap)) {
2448 } else if (efx_farch_filter_equal(&spec,
2457 if (depth >= max_rep_depth &&
2458 (ins_index >= 0 || depth >= max_ins_depth)) {
2460 if (ins_index < 0) {
2468 i = (i + incr) & (table->size - 1);
2473 /* If we found a filter to be replaced, check whether we
2476 if (rep_index >= 0) {
2477 struct efx_farch_filter_spec *saved_spec =
2478 &table->spec[rep_index];
2480 if (spec.priority == saved_spec->priority && !replace_equal) {
2484 if (spec.priority < saved_spec->priority) {
2488 if (saved_spec->priority == EFX_FILTER_PRI_AUTO ||
2489 saved_spec->flags & EFX_FILTER_FLAG_RX_OVER_AUTO)
2490 spec.flags |= EFX_FILTER_FLAG_RX_OVER_AUTO;
2493 /* Insert the filter */
2494 if (ins_index != rep_index) {
2495 __set_bit(ins_index, table->used_bitmap);
2498 table->spec[ins_index] = spec;
2500 if (table->id == EFX_FARCH_FILTER_TABLE_RX_DEF) {
2501 efx_farch_filter_push_rx_config(efx);
2503 if (table->search_limit[spec.type] < depth) {
2504 table->search_limit[spec.type] = depth;
2505 if (spec.flags & EFX_FILTER_FLAG_TX)
2506 efx_farch_filter_push_tx_limits(efx);
2508 efx_farch_filter_push_rx_config(efx);
2511 efx_writeo(efx, &filter,
2512 table->offset + table->step * ins_index);
2514 /* If we were able to replace a filter by inserting
2515 * at a lower depth, clear the replaced filter
2517 if (ins_index != rep_index && rep_index >= 0)
2518 efx_farch_filter_table_clear_entry(efx, table,
2522 netif_vdbg(efx, hw, efx->net_dev,
2523 "%s: filter type %d index %d rxq %u set",
2524 __func__, spec.type, ins_index, spec.dmaq_id);
2525 rc = efx_farch_filter_make_id(&spec, ins_index);
2528 up_write(&state->lock);
2533 efx_farch_filter_table_clear_entry(struct efx_nic *efx,
2534 struct efx_farch_filter_table *table,
2535 unsigned int filter_idx)
2537 static efx_oword_t filter;
2539 EFX_WARN_ON_PARANOID(!test_bit(filter_idx, table->used_bitmap));
2540 BUG_ON(table->offset == 0); /* can't clear MAC default filters */
2542 __clear_bit(filter_idx, table->used_bitmap);
2544 memset(&table->spec[filter_idx], 0, sizeof(table->spec[0]));
2546 efx_writeo(efx, &filter, table->offset + table->step * filter_idx);
2548 /* If this filter required a greater search depth than
2549 * any other, the search limit for its type can now be
2550 * decreased. However, it is hard to determine that
2551 * unless the table has become completely empty - in
2552 * which case, all its search limits can be set to 0.
2554 if (unlikely(table->used == 0)) {
2555 memset(table->search_limit, 0, sizeof(table->search_limit));
2556 if (table->id == EFX_FARCH_FILTER_TABLE_TX_MAC)
2557 efx_farch_filter_push_tx_limits(efx);
2559 efx_farch_filter_push_rx_config(efx);
2563 static int efx_farch_filter_remove(struct efx_nic *efx,
2564 struct efx_farch_filter_table *table,
2565 unsigned int filter_idx,
2566 enum efx_filter_priority priority)
2568 struct efx_farch_filter_spec *spec = &table->spec[filter_idx];
2570 if (!test_bit(filter_idx, table->used_bitmap) ||
2571 spec->priority != priority)
2574 if (spec->flags & EFX_FILTER_FLAG_RX_OVER_AUTO) {
2575 efx_farch_filter_init_rx_auto(efx, spec);
2576 efx_farch_filter_push_rx_config(efx);
2578 efx_farch_filter_table_clear_entry(efx, table, filter_idx);
2584 int efx_farch_filter_remove_safe(struct efx_nic *efx,
2585 enum efx_filter_priority priority,
2588 struct efx_farch_filter_state *state = efx->filter_state;
2589 enum efx_farch_filter_table_id table_id;
2590 struct efx_farch_filter_table *table;
2591 unsigned int filter_idx;
2594 table_id = efx_farch_filter_id_table_id(filter_id);
2595 if ((unsigned int)table_id >= EFX_FARCH_FILTER_TABLE_COUNT)
2597 table = &state->table[table_id];
2599 filter_idx = efx_farch_filter_id_index(filter_id);
2600 if (filter_idx >= table->size)
2602 down_write(&state->lock);
2604 rc = efx_farch_filter_remove(efx, table, filter_idx, priority);
2605 up_write(&state->lock);
2610 int efx_farch_filter_get_safe(struct efx_nic *efx,
2611 enum efx_filter_priority priority,
2612 u32 filter_id, struct efx_filter_spec *spec_buf)
2614 struct efx_farch_filter_state *state = efx->filter_state;
2615 enum efx_farch_filter_table_id table_id;
2616 struct efx_farch_filter_table *table;
2617 struct efx_farch_filter_spec *spec;
2618 unsigned int filter_idx;
2621 down_read(&state->lock);
2623 table_id = efx_farch_filter_id_table_id(filter_id);
2624 if ((unsigned int)table_id >= EFX_FARCH_FILTER_TABLE_COUNT)
2626 table = &state->table[table_id];
2628 filter_idx = efx_farch_filter_id_index(filter_id);
2629 if (filter_idx >= table->size)
2631 spec = &table->spec[filter_idx];
2633 if (test_bit(filter_idx, table->used_bitmap) &&
2634 spec->priority == priority) {
2635 efx_farch_filter_to_gen_spec(spec_buf, spec);
2640 up_read(&state->lock);
2645 efx_farch_filter_table_clear(struct efx_nic *efx,
2646 enum efx_farch_filter_table_id table_id,
2647 enum efx_filter_priority priority)
2649 struct efx_farch_filter_state *state = efx->filter_state;
2650 struct efx_farch_filter_table *table = &state->table[table_id];
2651 unsigned int filter_idx;
2653 down_write(&state->lock);
2654 for (filter_idx = 0; filter_idx < table->size; ++filter_idx) {
2655 if (table->spec[filter_idx].priority != EFX_FILTER_PRI_AUTO)
2656 efx_farch_filter_remove(efx, table,
2657 filter_idx, priority);
2659 up_write(&state->lock);
2662 int efx_farch_filter_clear_rx(struct efx_nic *efx,
2663 enum efx_filter_priority priority)
2665 efx_farch_filter_table_clear(efx, EFX_FARCH_FILTER_TABLE_RX_IP,
2667 efx_farch_filter_table_clear(efx, EFX_FARCH_FILTER_TABLE_RX_MAC,
2669 efx_farch_filter_table_clear(efx, EFX_FARCH_FILTER_TABLE_RX_DEF,
2674 u32 efx_farch_filter_count_rx_used(struct efx_nic *efx,
2675 enum efx_filter_priority priority)
2677 struct efx_farch_filter_state *state = efx->filter_state;
2678 enum efx_farch_filter_table_id table_id;
2679 struct efx_farch_filter_table *table;
2680 unsigned int filter_idx;
2683 down_read(&state->lock);
2685 for (table_id = EFX_FARCH_FILTER_TABLE_RX_IP;
2686 table_id <= EFX_FARCH_FILTER_TABLE_RX_DEF;
2688 table = &state->table[table_id];
2689 for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
2690 if (test_bit(filter_idx, table->used_bitmap) &&
2691 table->spec[filter_idx].priority == priority)
2696 up_read(&state->lock);
2701 s32 efx_farch_filter_get_rx_ids(struct efx_nic *efx,
2702 enum efx_filter_priority priority,
2705 struct efx_farch_filter_state *state = efx->filter_state;
2706 enum efx_farch_filter_table_id table_id;
2707 struct efx_farch_filter_table *table;
2708 unsigned int filter_idx;
2711 down_read(&state->lock);
2713 for (table_id = EFX_FARCH_FILTER_TABLE_RX_IP;
2714 table_id <= EFX_FARCH_FILTER_TABLE_RX_DEF;
2716 table = &state->table[table_id];
2717 for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
2718 if (test_bit(filter_idx, table->used_bitmap) &&
2719 table->spec[filter_idx].priority == priority) {
2720 if (count == size) {
2724 buf[count++] = efx_farch_filter_make_id(
2725 &table->spec[filter_idx], filter_idx);
2730 up_read(&state->lock);
2735 /* Restore filter stater after reset */
2736 void efx_farch_filter_table_restore(struct efx_nic *efx)
2738 struct efx_farch_filter_state *state = efx->filter_state;
2739 enum efx_farch_filter_table_id table_id;
2740 struct efx_farch_filter_table *table;
2742 unsigned int filter_idx;
2744 down_write(&state->lock);
2746 for (table_id = 0; table_id < EFX_FARCH_FILTER_TABLE_COUNT; table_id++) {
2747 table = &state->table[table_id];
2749 /* Check whether this is a regular register table */
2750 if (table->step == 0)
2753 for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
2754 if (!test_bit(filter_idx, table->used_bitmap))
2756 efx_farch_filter_build(&filter, &table->spec[filter_idx]);
2757 efx_writeo(efx, &filter,
2758 table->offset + table->step * filter_idx);
2762 efx_farch_filter_push_rx_config(efx);
2763 efx_farch_filter_push_tx_limits(efx);
2765 up_write(&state->lock);
2768 void efx_farch_filter_table_remove(struct efx_nic *efx)
2770 struct efx_farch_filter_state *state = efx->filter_state;
2771 enum efx_farch_filter_table_id table_id;
2773 for (table_id = 0; table_id < EFX_FARCH_FILTER_TABLE_COUNT; table_id++) {
2774 kfree(state->table[table_id].used_bitmap);
2775 vfree(state->table[table_id].spec);
2780 int efx_farch_filter_table_probe(struct efx_nic *efx)
2782 struct efx_farch_filter_state *state;
2783 struct efx_farch_filter_table *table;
2786 state = kzalloc(sizeof(struct efx_farch_filter_state), GFP_KERNEL);
2789 efx->filter_state = state;
2790 init_rwsem(&state->lock);
2792 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_IP];
2793 table->id = EFX_FARCH_FILTER_TABLE_RX_IP;
2794 table->offset = FR_BZ_RX_FILTER_TBL0;
2795 table->size = FR_BZ_RX_FILTER_TBL0_ROWS;
2796 table->step = FR_BZ_RX_FILTER_TBL0_STEP;
2798 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_MAC];
2799 table->id = EFX_FARCH_FILTER_TABLE_RX_MAC;
2800 table->offset = FR_CZ_RX_MAC_FILTER_TBL0;
2801 table->size = FR_CZ_RX_MAC_FILTER_TBL0_ROWS;
2802 table->step = FR_CZ_RX_MAC_FILTER_TBL0_STEP;
2804 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_DEF];
2805 table->id = EFX_FARCH_FILTER_TABLE_RX_DEF;
2806 table->size = EFX_FARCH_FILTER_SIZE_RX_DEF;
2808 table = &state->table[EFX_FARCH_FILTER_TABLE_TX_MAC];
2809 table->id = EFX_FARCH_FILTER_TABLE_TX_MAC;
2810 table->offset = FR_CZ_TX_MAC_FILTER_TBL0;
2811 table->size = FR_CZ_TX_MAC_FILTER_TBL0_ROWS;
2812 table->step = FR_CZ_TX_MAC_FILTER_TBL0_STEP;
2814 for (table_id = 0; table_id < EFX_FARCH_FILTER_TABLE_COUNT; table_id++) {
2815 table = &state->table[table_id];
2816 if (table->size == 0)
2818 table->used_bitmap = kcalloc(BITS_TO_LONGS(table->size),
2819 sizeof(unsigned long),
2821 if (!table->used_bitmap)
2823 table->spec = vzalloc(array_size(sizeof(*table->spec),
2829 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_DEF];
2831 /* RX default filters must always exist */
2832 struct efx_farch_filter_spec *spec;
2835 for (i = 0; i < EFX_FARCH_FILTER_SIZE_RX_DEF; i++) {
2836 spec = &table->spec[i];
2837 spec->type = EFX_FARCH_FILTER_UC_DEF + i;
2838 efx_farch_filter_init_rx_auto(efx, spec);
2839 __set_bit(i, table->used_bitmap);
2843 efx_farch_filter_push_rx_config(efx);
2848 efx_farch_filter_table_remove(efx);
2852 /* Update scatter enable flags for filters pointing to our own RX queues */
2853 void efx_farch_filter_update_rx_scatter(struct efx_nic *efx)
2855 struct efx_farch_filter_state *state = efx->filter_state;
2856 enum efx_farch_filter_table_id table_id;
2857 struct efx_farch_filter_table *table;
2859 unsigned int filter_idx;
2861 down_write(&state->lock);
2863 for (table_id = EFX_FARCH_FILTER_TABLE_RX_IP;
2864 table_id <= EFX_FARCH_FILTER_TABLE_RX_DEF;
2866 table = &state->table[table_id];
2868 for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
2869 if (!test_bit(filter_idx, table->used_bitmap) ||
2870 table->spec[filter_idx].dmaq_id >=
2874 if (efx->rx_scatter)
2875 table->spec[filter_idx].flags |=
2876 EFX_FILTER_FLAG_RX_SCATTER;
2878 table->spec[filter_idx].flags &=
2879 ~EFX_FILTER_FLAG_RX_SCATTER;
2881 if (table_id == EFX_FARCH_FILTER_TABLE_RX_DEF)
2882 /* Pushed by efx_farch_filter_push_rx_config() */
2885 efx_farch_filter_build(&filter, &table->spec[filter_idx]);
2886 efx_writeo(efx, &filter,
2887 table->offset + table->step * filter_idx);
2891 efx_farch_filter_push_rx_config(efx);
2893 up_write(&state->lock);
2896 #ifdef CONFIG_RFS_ACCEL
2898 bool efx_farch_filter_rfs_expire_one(struct efx_nic *efx, u32 flow_id,
2901 struct efx_farch_filter_state *state = efx->filter_state;
2902 struct efx_farch_filter_table *table;
2903 bool ret = false, force = false;
2906 down_write(&state->lock);
2907 spin_lock_bh(&efx->rps_hash_lock);
2908 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_IP];
2909 if (test_bit(index, table->used_bitmap) &&
2910 table->spec[index].priority == EFX_FILTER_PRI_HINT) {
2911 struct efx_arfs_rule *rule = NULL;
2912 struct efx_filter_spec spec;
2914 efx_farch_filter_to_gen_spec(&spec, &table->spec[index]);
2915 if (!efx->rps_hash_table) {
2916 /* In the absence of the table, we always returned 0 to
2917 * ARFS, so use the same to query it.
2921 rule = efx_rps_hash_find(efx, &spec);
2923 /* ARFS table doesn't know of this filter, remove it */
2926 arfs_id = rule->arfs_id;
2927 if (!efx_rps_check_rule(rule, index, &force))
2931 if (force || rps_may_expire_flow(efx->net_dev, spec.dmaq_id,
2932 flow_id, arfs_id)) {
2934 rule->filter_id = EFX_ARFS_FILTER_ID_REMOVING;
2935 efx_rps_hash_del(efx, &spec);
2936 efx_farch_filter_table_clear_entry(efx, table, index);
2941 spin_unlock_bh(&efx->rps_hash_lock);
2942 up_write(&state->lock);
2946 #endif /* CONFIG_RFS_ACCEL */
2948 void efx_farch_filter_sync_rx_mode(struct efx_nic *efx)
2950 struct net_device *net_dev = efx->net_dev;
2951 struct netdev_hw_addr *ha;
2952 union efx_multicast_hash *mc_hash = &efx->multicast_hash;
2956 if (!efx_dev_registered(efx))
2959 netif_addr_lock_bh(net_dev);
2961 efx->unicast_filter = !(net_dev->flags & IFF_PROMISC);
2963 /* Build multicast hash table */
2964 if (net_dev->flags & (IFF_PROMISC | IFF_ALLMULTI)) {
2965 memset(mc_hash, 0xff, sizeof(*mc_hash));
2967 memset(mc_hash, 0x00, sizeof(*mc_hash));
2968 netdev_for_each_mc_addr(ha, net_dev) {
2969 crc = ether_crc_le(ETH_ALEN, ha->addr);
2970 bit = crc & (EFX_MCAST_HASH_ENTRIES - 1);
2971 __set_bit_le(bit, mc_hash);
2974 /* Broadcast packets go through the multicast hash filter.
2975 * ether_crc_le() of the broadcast address is 0xbe2612ff
2976 * so we always add bit 0xff to the mask.
2978 __set_bit_le(0xff, mc_hash);
2981 netif_addr_unlock_bh(net_dev);