1 /****************************************************************************
2 * Driver for Solarflare network controllers and boards
3 * Copyright 2005-2006 Fen Systems Ltd.
4 * Copyright 2006-2013 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/bitops.h>
12 #include <linux/delay.h>
13 #include <linux/interrupt.h>
14 #include <linux/pci.h>
15 #include <linux/module.h>
16 #include <linux/seq_file.h>
17 #include <linux/crc32.h>
18 #include "net_driver.h"
22 #include "farch_regs.h"
24 #include "workarounds.h"
26 /* Falcon-architecture (SFC4000 and 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, i, j;
106 efx_oword_t mask, imask, original, reg, buf;
108 for (i = 0; i < n_regs; ++i) {
109 address = regs[i].address;
110 mask = imask = regs[i].mask;
111 EFX_INVERT_OWORD(imask);
113 efx_reado(efx, &original, address);
115 /* bit sweep on and off */
116 for (j = 0; j < 128; j++) {
117 if (!EFX_EXTRACT_OWORD32(mask, j, j))
120 /* Test this testable bit can be set in isolation */
121 EFX_AND_OWORD(reg, original, mask);
122 EFX_SET_OWORD32(reg, j, j, 1);
124 efx_writeo(efx, ®, address);
125 efx_reado(efx, &buf, address);
127 if (efx_masked_compare_oword(®, &buf, &mask))
130 /* Test this testable bit can be cleared in isolation */
131 EFX_OR_OWORD(reg, original, mask);
132 EFX_SET_OWORD32(reg, j, j, 0);
134 efx_writeo(efx, ®, address);
135 efx_reado(efx, &buf, address);
137 if (efx_masked_compare_oword(®, &buf, &mask))
141 efx_writeo(efx, &original, address);
147 netif_err(efx, hw, efx->net_dev,
148 "wrote "EFX_OWORD_FMT" read "EFX_OWORD_FMT
149 " at address 0x%x mask "EFX_OWORD_FMT"\n", EFX_OWORD_VAL(reg),
150 EFX_OWORD_VAL(buf), address, EFX_OWORD_VAL(mask));
154 /**************************************************************************
156 * Special buffer handling
157 * Special buffers are used for event queues and the TX and RX
160 *************************************************************************/
163 * Initialise a special buffer
165 * This will define a buffer (previously allocated via
166 * efx_alloc_special_buffer()) in the buffer table, allowing
167 * it to be used for event queues, descriptor rings etc.
170 efx_init_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
172 efx_qword_t buf_desc;
177 EFX_BUG_ON_PARANOID(!buffer->buf.addr);
179 /* Write buffer descriptors to NIC */
180 for (i = 0; i < buffer->entries; i++) {
181 index = buffer->index + i;
182 dma_addr = buffer->buf.dma_addr + (i * EFX_BUF_SIZE);
183 netif_dbg(efx, probe, efx->net_dev,
184 "mapping special buffer %d at %llx\n",
185 index, (unsigned long long)dma_addr);
186 EFX_POPULATE_QWORD_3(buf_desc,
187 FRF_AZ_BUF_ADR_REGION, 0,
188 FRF_AZ_BUF_ADR_FBUF, dma_addr >> 12,
189 FRF_AZ_BUF_OWNER_ID_FBUF, 0);
190 efx_write_buf_tbl(efx, &buf_desc, index);
194 /* Unmaps a buffer and clears the buffer table entries */
196 efx_fini_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
198 efx_oword_t buf_tbl_upd;
199 unsigned int start = buffer->index;
200 unsigned int end = (buffer->index + buffer->entries - 1);
202 if (!buffer->entries)
205 netif_dbg(efx, hw, efx->net_dev, "unmapping special buffers %d-%d\n",
206 buffer->index, buffer->index + buffer->entries - 1);
208 EFX_POPULATE_OWORD_4(buf_tbl_upd,
209 FRF_AZ_BUF_UPD_CMD, 0,
210 FRF_AZ_BUF_CLR_CMD, 1,
211 FRF_AZ_BUF_CLR_END_ID, end,
212 FRF_AZ_BUF_CLR_START_ID, start);
213 efx_writeo(efx, &buf_tbl_upd, FR_AZ_BUF_TBL_UPD);
217 * Allocate a new special buffer
219 * This allocates memory for a new buffer, clears it and allocates a
220 * new buffer ID range. It does not write into the buffer table.
222 * This call will allocate 4KB buffers, since 8KB buffers can't be
223 * used for event queues and descriptor rings.
225 static int efx_alloc_special_buffer(struct efx_nic *efx,
226 struct efx_special_buffer *buffer,
229 len = ALIGN(len, EFX_BUF_SIZE);
231 if (efx_nic_alloc_buffer(efx, &buffer->buf, len, GFP_KERNEL))
233 buffer->entries = len / EFX_BUF_SIZE;
234 BUG_ON(buffer->buf.dma_addr & (EFX_BUF_SIZE - 1));
236 /* Select new buffer ID */
237 buffer->index = efx->next_buffer_table;
238 efx->next_buffer_table += buffer->entries;
239 #ifdef CONFIG_SFC_SRIOV
240 BUG_ON(efx_sriov_enabled(efx) &&
241 efx->vf_buftbl_base < efx->next_buffer_table);
244 netif_dbg(efx, probe, efx->net_dev,
245 "allocating special buffers %d-%d at %llx+%x "
246 "(virt %p phys %llx)\n", buffer->index,
247 buffer->index + buffer->entries - 1,
248 (u64)buffer->buf.dma_addr, len,
249 buffer->buf.addr, (u64)virt_to_phys(buffer->buf.addr));
255 efx_free_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
257 if (!buffer->buf.addr)
260 netif_dbg(efx, hw, efx->net_dev,
261 "deallocating special buffers %d-%d at %llx+%x "
262 "(virt %p phys %llx)\n", buffer->index,
263 buffer->index + buffer->entries - 1,
264 (u64)buffer->buf.dma_addr, buffer->buf.len,
265 buffer->buf.addr, (u64)virt_to_phys(buffer->buf.addr));
267 efx_nic_free_buffer(efx, &buffer->buf);
271 /**************************************************************************
275 **************************************************************************/
277 /* This writes to the TX_DESC_WPTR; write pointer for TX descriptor ring */
278 static inline void efx_farch_notify_tx_desc(struct efx_tx_queue *tx_queue)
283 write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
284 EFX_POPULATE_DWORD_1(reg, FRF_AZ_TX_DESC_WPTR_DWORD, write_ptr);
285 efx_writed_page(tx_queue->efx, ®,
286 FR_AZ_TX_DESC_UPD_DWORD_P0, tx_queue->queue);
289 /* Write pointer and first descriptor for TX descriptor ring */
290 static inline void efx_farch_push_tx_desc(struct efx_tx_queue *tx_queue,
291 const efx_qword_t *txd)
296 BUILD_BUG_ON(FRF_AZ_TX_DESC_LBN != 0);
297 BUILD_BUG_ON(FR_AA_TX_DESC_UPD_KER != FR_BZ_TX_DESC_UPD_P0);
299 write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
300 EFX_POPULATE_OWORD_2(reg, FRF_AZ_TX_DESC_PUSH_CMD, true,
301 FRF_AZ_TX_DESC_WPTR, write_ptr);
303 efx_writeo_page(tx_queue->efx, ®,
304 FR_BZ_TX_DESC_UPD_P0, tx_queue->queue);
308 /* For each entry inserted into the software descriptor ring, create a
309 * descriptor in the hardware TX descriptor ring (in host memory), and
312 void efx_farch_tx_write(struct efx_tx_queue *tx_queue)
314 struct efx_tx_buffer *buffer;
317 unsigned old_write_count = tx_queue->write_count;
319 BUG_ON(tx_queue->write_count == tx_queue->insert_count);
322 write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
323 buffer = &tx_queue->buffer[write_ptr];
324 txd = efx_tx_desc(tx_queue, write_ptr);
325 ++tx_queue->write_count;
327 EFX_BUG_ON_PARANOID(buffer->flags & EFX_TX_BUF_OPTION);
329 /* Create TX descriptor ring entry */
330 BUILD_BUG_ON(EFX_TX_BUF_CONT != 1);
331 EFX_POPULATE_QWORD_4(*txd,
333 buffer->flags & EFX_TX_BUF_CONT,
334 FSF_AZ_TX_KER_BYTE_COUNT, buffer->len,
335 FSF_AZ_TX_KER_BUF_REGION, 0,
336 FSF_AZ_TX_KER_BUF_ADDR, buffer->dma_addr);
337 } while (tx_queue->write_count != tx_queue->insert_count);
339 wmb(); /* Ensure descriptors are written before they are fetched */
341 if (efx_nic_may_push_tx_desc(tx_queue, old_write_count)) {
342 txd = efx_tx_desc(tx_queue,
343 old_write_count & tx_queue->ptr_mask);
344 efx_farch_push_tx_desc(tx_queue, txd);
347 efx_farch_notify_tx_desc(tx_queue);
351 /* Allocate hardware resources for a TX queue */
352 int efx_farch_tx_probe(struct efx_tx_queue *tx_queue)
354 struct efx_nic *efx = tx_queue->efx;
357 entries = tx_queue->ptr_mask + 1;
358 return efx_alloc_special_buffer(efx, &tx_queue->txd,
359 entries * sizeof(efx_qword_t));
362 void efx_farch_tx_init(struct efx_tx_queue *tx_queue)
364 struct efx_nic *efx = tx_queue->efx;
367 /* Pin TX descriptor ring */
368 efx_init_special_buffer(efx, &tx_queue->txd);
370 /* Push TX descriptor ring to card */
371 EFX_POPULATE_OWORD_10(reg,
372 FRF_AZ_TX_DESCQ_EN, 1,
373 FRF_AZ_TX_ISCSI_DDIG_EN, 0,
374 FRF_AZ_TX_ISCSI_HDIG_EN, 0,
375 FRF_AZ_TX_DESCQ_BUF_BASE_ID, tx_queue->txd.index,
376 FRF_AZ_TX_DESCQ_EVQ_ID,
377 tx_queue->channel->channel,
378 FRF_AZ_TX_DESCQ_OWNER_ID, 0,
379 FRF_AZ_TX_DESCQ_LABEL, tx_queue->queue,
380 FRF_AZ_TX_DESCQ_SIZE,
381 __ffs(tx_queue->txd.entries),
382 FRF_AZ_TX_DESCQ_TYPE, 0,
383 FRF_BZ_TX_NON_IP_DROP_DIS, 1);
385 if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
386 int csum = tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD;
387 EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_IP_CHKSM_DIS, !csum);
388 EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_TCP_CHKSM_DIS,
392 efx_writeo_table(efx, ®, efx->type->txd_ptr_tbl_base,
395 if (efx_nic_rev(efx) < EFX_REV_FALCON_B0) {
396 /* Only 128 bits in this register */
397 BUILD_BUG_ON(EFX_MAX_TX_QUEUES > 128);
399 efx_reado(efx, ®, FR_AA_TX_CHKSM_CFG);
400 if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD)
401 __clear_bit_le(tx_queue->queue, ®);
403 __set_bit_le(tx_queue->queue, ®);
404 efx_writeo(efx, ®, FR_AA_TX_CHKSM_CFG);
407 if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
408 EFX_POPULATE_OWORD_1(reg,
410 (tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
412 FFE_BZ_TX_PACE_RESERVED);
413 efx_writeo_table(efx, ®, FR_BZ_TX_PACE_TBL,
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;
512 bool is_b0 = efx_nic_rev(efx) >= EFX_REV_FALCON_B0;
513 bool iscsi_digest_en = is_b0;
516 /* For kernel-mode queues in Falcon A1, the JUMBO flag enables
517 * DMA to continue after a PCIe page boundary (and scattering
518 * is not possible). In Falcon B0 and Siena, it enables
521 jumbo_en = !is_b0 || efx->rx_scatter;
523 netif_dbg(efx, hw, efx->net_dev,
524 "RX queue %d ring in special buffers %d-%d\n",
525 efx_rx_queue_index(rx_queue), rx_queue->rxd.index,
526 rx_queue->rxd.index + rx_queue->rxd.entries - 1);
528 rx_queue->scatter_n = 0;
530 /* Pin RX descriptor ring */
531 efx_init_special_buffer(efx, &rx_queue->rxd);
533 /* Push RX descriptor ring to card */
534 EFX_POPULATE_OWORD_10(rx_desc_ptr,
535 FRF_AZ_RX_ISCSI_DDIG_EN, iscsi_digest_en,
536 FRF_AZ_RX_ISCSI_HDIG_EN, iscsi_digest_en,
537 FRF_AZ_RX_DESCQ_BUF_BASE_ID, rx_queue->rxd.index,
538 FRF_AZ_RX_DESCQ_EVQ_ID,
539 efx_rx_queue_channel(rx_queue)->channel,
540 FRF_AZ_RX_DESCQ_OWNER_ID, 0,
541 FRF_AZ_RX_DESCQ_LABEL,
542 efx_rx_queue_index(rx_queue),
543 FRF_AZ_RX_DESCQ_SIZE,
544 __ffs(rx_queue->rxd.entries),
545 FRF_AZ_RX_DESCQ_TYPE, 0 /* kernel queue */ ,
546 FRF_AZ_RX_DESCQ_JUMBO, jumbo_en,
547 FRF_AZ_RX_DESCQ_EN, 1);
548 efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
549 efx_rx_queue_index(rx_queue));
552 static void efx_farch_flush_rx_queue(struct efx_rx_queue *rx_queue)
554 struct efx_nic *efx = rx_queue->efx;
555 efx_oword_t rx_flush_descq;
557 EFX_POPULATE_OWORD_2(rx_flush_descq,
558 FRF_AZ_RX_FLUSH_DESCQ_CMD, 1,
559 FRF_AZ_RX_FLUSH_DESCQ,
560 efx_rx_queue_index(rx_queue));
561 efx_writeo(efx, &rx_flush_descq, FR_AZ_RX_FLUSH_DESCQ);
564 void efx_farch_rx_fini(struct efx_rx_queue *rx_queue)
566 efx_oword_t rx_desc_ptr;
567 struct efx_nic *efx = rx_queue->efx;
569 /* Remove RX descriptor ring from card */
570 EFX_ZERO_OWORD(rx_desc_ptr);
571 efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
572 efx_rx_queue_index(rx_queue));
574 /* Unpin RX descriptor ring */
575 efx_fini_special_buffer(efx, &rx_queue->rxd);
578 /* Free buffers backing RX queue */
579 void efx_farch_rx_remove(struct efx_rx_queue *rx_queue)
581 efx_free_special_buffer(rx_queue->efx, &rx_queue->rxd);
584 /**************************************************************************
588 **************************************************************************/
590 /* efx_farch_flush_queues() must be woken up when all flushes are completed,
591 * or more RX flushes can be kicked off.
593 static bool efx_farch_flush_wake(struct efx_nic *efx)
595 /* Ensure that all updates are visible to efx_farch_flush_queues() */
598 return (atomic_read(&efx->active_queues) == 0 ||
599 (atomic_read(&efx->rxq_flush_outstanding) < EFX_RX_FLUSH_COUNT
600 && atomic_read(&efx->rxq_flush_pending) > 0));
603 static bool efx_check_tx_flush_complete(struct efx_nic *efx)
606 efx_oword_t txd_ptr_tbl;
607 struct efx_channel *channel;
608 struct efx_tx_queue *tx_queue;
610 efx_for_each_channel(channel, efx) {
611 efx_for_each_channel_tx_queue(tx_queue, channel) {
612 efx_reado_table(efx, &txd_ptr_tbl,
613 FR_BZ_TX_DESC_PTR_TBL, tx_queue->queue);
614 if (EFX_OWORD_FIELD(txd_ptr_tbl,
615 FRF_AZ_TX_DESCQ_FLUSH) ||
616 EFX_OWORD_FIELD(txd_ptr_tbl,
617 FRF_AZ_TX_DESCQ_EN)) {
618 netif_dbg(efx, hw, efx->net_dev,
619 "flush did not complete on TXQ %d\n",
622 } else if (atomic_cmpxchg(&tx_queue->flush_outstanding,
624 /* The flush is complete, but we didn't
625 * receive a flush completion event
627 netif_dbg(efx, hw, efx->net_dev,
628 "flush complete on TXQ %d, so drain "
629 "the queue\n", tx_queue->queue);
630 /* Don't need to increment active_queues as it
631 * has already been incremented for the queues
632 * which did not drain
634 efx_farch_magic_event(channel,
635 EFX_CHANNEL_MAGIC_TX_DRAIN(
644 /* Flush all the transmit queues, and continue flushing receive queues until
645 * they're all flushed. Wait for the DRAIN events to be recieved so that there
646 * are no more RX and TX events left on any channel. */
647 static int efx_farch_do_flush(struct efx_nic *efx)
649 unsigned timeout = msecs_to_jiffies(5000); /* 5s for all flushes and drains */
650 struct efx_channel *channel;
651 struct efx_rx_queue *rx_queue;
652 struct efx_tx_queue *tx_queue;
655 efx_for_each_channel(channel, efx) {
656 efx_for_each_channel_tx_queue(tx_queue, channel) {
657 efx_farch_flush_tx_queue(tx_queue);
659 efx_for_each_channel_rx_queue(rx_queue, channel) {
660 rx_queue->flush_pending = true;
661 atomic_inc(&efx->rxq_flush_pending);
665 while (timeout && atomic_read(&efx->active_queues) > 0) {
666 /* If SRIOV is enabled, then offload receive queue flushing to
667 * the firmware (though we will still have to poll for
668 * completion). If that fails, fall back to the old scheme.
670 if (efx_sriov_enabled(efx)) {
671 rc = efx_mcdi_flush_rxqs(efx);
676 /* The hardware supports four concurrent rx flushes, each of
677 * which may need to be retried if there is an outstanding
680 efx_for_each_channel(channel, efx) {
681 efx_for_each_channel_rx_queue(rx_queue, channel) {
682 if (atomic_read(&efx->rxq_flush_outstanding) >=
686 if (rx_queue->flush_pending) {
687 rx_queue->flush_pending = false;
688 atomic_dec(&efx->rxq_flush_pending);
689 atomic_inc(&efx->rxq_flush_outstanding);
690 efx_farch_flush_rx_queue(rx_queue);
696 timeout = wait_event_timeout(efx->flush_wq,
697 efx_farch_flush_wake(efx),
701 if (atomic_read(&efx->active_queues) &&
702 !efx_check_tx_flush_complete(efx)) {
703 netif_err(efx, hw, efx->net_dev, "failed to flush %d queues "
704 "(rx %d+%d)\n", atomic_read(&efx->active_queues),
705 atomic_read(&efx->rxq_flush_outstanding),
706 atomic_read(&efx->rxq_flush_pending));
709 atomic_set(&efx->active_queues, 0);
710 atomic_set(&efx->rxq_flush_pending, 0);
711 atomic_set(&efx->rxq_flush_outstanding, 0);
717 int efx_farch_fini_dmaq(struct efx_nic *efx)
719 struct efx_channel *channel;
720 struct efx_tx_queue *tx_queue;
721 struct efx_rx_queue *rx_queue;
724 /* Do not attempt to write to the NIC during EEH recovery */
725 if (efx->state != STATE_RECOVERY) {
726 /* Only perform flush if DMA is enabled */
727 if (efx->pci_dev->is_busmaster) {
728 efx->type->prepare_flush(efx);
729 rc = efx_farch_do_flush(efx);
730 efx->type->finish_flush(efx);
733 efx_for_each_channel(channel, efx) {
734 efx_for_each_channel_rx_queue(rx_queue, channel)
735 efx_farch_rx_fini(rx_queue);
736 efx_for_each_channel_tx_queue(tx_queue, channel)
737 efx_farch_tx_fini(tx_queue);
744 /**************************************************************************
746 * Event queue processing
747 * Event queues are processed by per-channel tasklets.
749 **************************************************************************/
751 /* Update a channel's event queue's read pointer (RPTR) register
753 * This writes the EVQ_RPTR_REG register for the specified channel's
756 void efx_farch_ev_read_ack(struct efx_channel *channel)
759 struct efx_nic *efx = channel->efx;
761 EFX_POPULATE_DWORD_1(reg, FRF_AZ_EVQ_RPTR,
762 channel->eventq_read_ptr & channel->eventq_mask);
764 /* For Falcon A1, EVQ_RPTR_KER is documented as having a step size
765 * of 4 bytes, but it is really 16 bytes just like later revisions.
767 efx_writed(efx, ®,
768 efx->type->evq_rptr_tbl_base +
769 FR_BZ_EVQ_RPTR_STEP * channel->channel);
772 /* Use HW to insert a SW defined event */
773 void efx_farch_generate_event(struct efx_nic *efx, unsigned int evq,
776 efx_oword_t drv_ev_reg;
778 BUILD_BUG_ON(FRF_AZ_DRV_EV_DATA_LBN != 0 ||
779 FRF_AZ_DRV_EV_DATA_WIDTH != 64);
780 drv_ev_reg.u32[0] = event->u32[0];
781 drv_ev_reg.u32[1] = event->u32[1];
782 drv_ev_reg.u32[2] = 0;
783 drv_ev_reg.u32[3] = 0;
784 EFX_SET_OWORD_FIELD(drv_ev_reg, FRF_AZ_DRV_EV_QID, evq);
785 efx_writeo(efx, &drv_ev_reg, FR_AZ_DRV_EV);
788 static void efx_farch_magic_event(struct efx_channel *channel, u32 magic)
792 EFX_POPULATE_QWORD_2(event, FSF_AZ_EV_CODE,
793 FSE_AZ_EV_CODE_DRV_GEN_EV,
794 FSF_AZ_DRV_GEN_EV_MAGIC, magic);
795 efx_farch_generate_event(channel->efx, channel->channel, &event);
798 /* Handle a transmit completion event
800 * The NIC batches TX completion events; the message we receive is of
801 * the form "complete all TX events up to this index".
804 efx_farch_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
806 unsigned int tx_ev_desc_ptr;
807 unsigned int tx_ev_q_label;
808 struct efx_tx_queue *tx_queue;
809 struct efx_nic *efx = channel->efx;
812 if (unlikely(ACCESS_ONCE(efx->reset_pending)))
815 if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
816 /* Transmit completion */
817 tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
818 tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
819 tx_queue = efx_channel_get_tx_queue(
820 channel, tx_ev_q_label % EFX_TXQ_TYPES);
821 tx_packets = ((tx_ev_desc_ptr - tx_queue->read_count) &
823 efx_xmit_done(tx_queue, tx_ev_desc_ptr);
824 } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
825 /* Rewrite the FIFO write pointer */
826 tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
827 tx_queue = efx_channel_get_tx_queue(
828 channel, tx_ev_q_label % EFX_TXQ_TYPES);
830 netif_tx_lock(efx->net_dev);
831 efx_farch_notify_tx_desc(tx_queue);
832 netif_tx_unlock(efx->net_dev);
833 } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR)) {
834 efx_schedule_reset(efx, RESET_TYPE_DMA_ERROR);
836 netif_err(efx, tx_err, efx->net_dev,
837 "channel %d unexpected TX event "
838 EFX_QWORD_FMT"\n", channel->channel,
839 EFX_QWORD_VAL(*event));
845 /* Detect errors included in the rx_evt_pkt_ok bit. */
846 static u16 efx_farch_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
847 const efx_qword_t *event)
849 struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
850 struct efx_nic *efx = rx_queue->efx;
851 bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
852 bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
853 bool rx_ev_frm_trunc, rx_ev_drib_nib, rx_ev_tobe_disc;
854 bool rx_ev_other_err, rx_ev_pause_frm;
855 bool rx_ev_hdr_type, rx_ev_mcast_pkt;
856 unsigned rx_ev_pkt_type;
858 rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
859 rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
860 rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
861 rx_ev_pkt_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_TYPE);
862 rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
863 FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
864 rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
865 FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
866 rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
867 FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
868 rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
869 rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
870 rx_ev_drib_nib = ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) ?
871 0 : EFX_QWORD_FIELD(*event, FSF_AA_RX_EV_DRIB_NIB));
872 rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
874 /* Every error apart from tobe_disc and pause_frm */
875 rx_ev_other_err = (rx_ev_drib_nib | rx_ev_tcp_udp_chksum_err |
876 rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
877 rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
879 /* Count errors that are not in MAC stats. Ignore expected
880 * checksum errors during self-test. */
882 ++channel->n_rx_frm_trunc;
883 else if (rx_ev_tobe_disc)
884 ++channel->n_rx_tobe_disc;
885 else if (!efx->loopback_selftest) {
886 if (rx_ev_ip_hdr_chksum_err)
887 ++channel->n_rx_ip_hdr_chksum_err;
888 else if (rx_ev_tcp_udp_chksum_err)
889 ++channel->n_rx_tcp_udp_chksum_err;
892 /* TOBE_DISC is expected on unicast mismatches; don't print out an
893 * error message. FRM_TRUNC indicates RXDP dropped the packet due
894 * to a FIFO overflow.
897 if (rx_ev_other_err && net_ratelimit()) {
898 netif_dbg(efx, rx_err, efx->net_dev,
899 " RX queue %d unexpected RX event "
900 EFX_QWORD_FMT "%s%s%s%s%s%s%s%s\n",
901 efx_rx_queue_index(rx_queue), EFX_QWORD_VAL(*event),
902 rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
903 rx_ev_ip_hdr_chksum_err ?
904 " [IP_HDR_CHKSUM_ERR]" : "",
905 rx_ev_tcp_udp_chksum_err ?
906 " [TCP_UDP_CHKSUM_ERR]" : "",
907 rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
908 rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
909 rx_ev_drib_nib ? " [DRIB_NIB]" : "",
910 rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
911 rx_ev_pause_frm ? " [PAUSE]" : "");
915 /* The frame must be discarded if any of these are true. */
916 return (rx_ev_eth_crc_err | rx_ev_frm_trunc | rx_ev_drib_nib |
917 rx_ev_tobe_disc | rx_ev_pause_frm) ?
918 EFX_RX_PKT_DISCARD : 0;
921 /* Handle receive events that are not in-order. Return true if this
922 * can be handled as a partial packet discard, false if it's more
926 efx_farch_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
928 struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
929 struct efx_nic *efx = rx_queue->efx;
930 unsigned expected, dropped;
932 if (rx_queue->scatter_n &&
933 index == ((rx_queue->removed_count + rx_queue->scatter_n - 1) &
934 rx_queue->ptr_mask)) {
935 ++channel->n_rx_nodesc_trunc;
939 expected = rx_queue->removed_count & rx_queue->ptr_mask;
940 dropped = (index - expected) & rx_queue->ptr_mask;
941 netif_info(efx, rx_err, efx->net_dev,
942 "dropped %d events (index=%d expected=%d)\n",
943 dropped, index, expected);
945 efx_schedule_reset(efx, EFX_WORKAROUND_5676(efx) ?
946 RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
950 /* Handle a packet received event
952 * The NIC gives a "discard" flag if it's a unicast packet with the
953 * wrong destination address
954 * Also "is multicast" and "matches multicast filter" flags can be used to
955 * discard non-matching multicast packets.
958 efx_farch_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
960 unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
961 unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
962 unsigned expected_ptr;
963 bool rx_ev_pkt_ok, rx_ev_sop, rx_ev_cont;
965 struct efx_rx_queue *rx_queue;
966 struct efx_nic *efx = channel->efx;
968 if (unlikely(ACCESS_ONCE(efx->reset_pending)))
971 rx_ev_cont = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT);
972 rx_ev_sop = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP);
973 WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
976 rx_queue = efx_channel_get_rx_queue(channel);
978 rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
979 expected_ptr = ((rx_queue->removed_count + rx_queue->scatter_n) &
982 /* Check for partial drops and other errors */
983 if (unlikely(rx_ev_desc_ptr != expected_ptr) ||
984 unlikely(rx_ev_sop != (rx_queue->scatter_n == 0))) {
985 if (rx_ev_desc_ptr != expected_ptr &&
986 !efx_farch_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr))
989 /* Discard all pending fragments */
990 if (rx_queue->scatter_n) {
993 rx_queue->removed_count & rx_queue->ptr_mask,
994 rx_queue->scatter_n, 0, EFX_RX_PKT_DISCARD);
995 rx_queue->removed_count += rx_queue->scatter_n;
996 rx_queue->scatter_n = 0;
999 /* Return if there is no new fragment */
1000 if (rx_ev_desc_ptr != expected_ptr)
1003 /* Discard new fragment if not SOP */
1007 rx_queue->removed_count & rx_queue->ptr_mask,
1008 1, 0, EFX_RX_PKT_DISCARD);
1009 ++rx_queue->removed_count;
1014 ++rx_queue->scatter_n;
1018 rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
1019 rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
1020 rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
1022 if (likely(rx_ev_pkt_ok)) {
1023 /* If packet is marked as OK then we can rely on the
1024 * hardware checksum and classification.
1027 switch (rx_ev_hdr_type) {
1028 case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP:
1029 flags |= EFX_RX_PKT_TCP;
1031 case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP:
1032 flags |= EFX_RX_PKT_CSUMMED;
1034 case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_OTHER:
1035 case FSE_AZ_RX_EV_HDR_TYPE_OTHER:
1039 flags = efx_farch_handle_rx_not_ok(rx_queue, event);
1042 /* Detect multicast packets that didn't match the filter */
1043 rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
1044 if (rx_ev_mcast_pkt) {
1045 unsigned int rx_ev_mcast_hash_match =
1046 EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
1048 if (unlikely(!rx_ev_mcast_hash_match)) {
1049 ++channel->n_rx_mcast_mismatch;
1050 flags |= EFX_RX_PKT_DISCARD;
1054 channel->irq_mod_score += 2;
1056 /* Handle received packet */
1057 efx_rx_packet(rx_queue,
1058 rx_queue->removed_count & rx_queue->ptr_mask,
1059 rx_queue->scatter_n, rx_ev_byte_cnt, flags);
1060 rx_queue->removed_count += rx_queue->scatter_n;
1061 rx_queue->scatter_n = 0;
1064 /* If this flush done event corresponds to a &struct efx_tx_queue, then
1065 * send an %EFX_CHANNEL_MAGIC_TX_DRAIN event to drain the event queue
1066 * of all transmit completions.
1069 efx_farch_handle_tx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1071 struct efx_tx_queue *tx_queue;
1074 qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1075 if (qid < EFX_TXQ_TYPES * efx->n_tx_channels) {
1076 tx_queue = efx_get_tx_queue(efx, qid / EFX_TXQ_TYPES,
1077 qid % EFX_TXQ_TYPES);
1078 if (atomic_cmpxchg(&tx_queue->flush_outstanding, 1, 0)) {
1079 efx_farch_magic_event(tx_queue->channel,
1080 EFX_CHANNEL_MAGIC_TX_DRAIN(tx_queue));
1085 /* If this flush done event corresponds to a &struct efx_rx_queue: If the flush
1086 * was succesful then send an %EFX_CHANNEL_MAGIC_RX_DRAIN, otherwise add
1087 * the RX queue back to the mask of RX queues in need of flushing.
1090 efx_farch_handle_rx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1092 struct efx_channel *channel;
1093 struct efx_rx_queue *rx_queue;
1097 qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
1098 failed = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
1099 if (qid >= efx->n_channels)
1101 channel = efx_get_channel(efx, qid);
1102 if (!efx_channel_has_rx_queue(channel))
1104 rx_queue = efx_channel_get_rx_queue(channel);
1107 netif_info(efx, hw, efx->net_dev,
1108 "RXQ %d flush retry\n", qid);
1109 rx_queue->flush_pending = true;
1110 atomic_inc(&efx->rxq_flush_pending);
1112 efx_farch_magic_event(efx_rx_queue_channel(rx_queue),
1113 EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue));
1115 atomic_dec(&efx->rxq_flush_outstanding);
1116 if (efx_farch_flush_wake(efx))
1117 wake_up(&efx->flush_wq);
1121 efx_farch_handle_drain_event(struct efx_channel *channel)
1123 struct efx_nic *efx = channel->efx;
1125 WARN_ON(atomic_read(&efx->active_queues) == 0);
1126 atomic_dec(&efx->active_queues);
1127 if (efx_farch_flush_wake(efx))
1128 wake_up(&efx->flush_wq);
1131 static void efx_farch_handle_generated_event(struct efx_channel *channel,
1134 struct efx_nic *efx = channel->efx;
1135 struct efx_rx_queue *rx_queue =
1136 efx_channel_has_rx_queue(channel) ?
1137 efx_channel_get_rx_queue(channel) : NULL;
1138 unsigned magic, code;
1140 magic = EFX_QWORD_FIELD(*event, FSF_AZ_DRV_GEN_EV_MAGIC);
1141 code = _EFX_CHANNEL_MAGIC_CODE(magic);
1143 if (magic == EFX_CHANNEL_MAGIC_TEST(channel)) {
1144 channel->event_test_cpu = raw_smp_processor_id();
1145 } else if (rx_queue && magic == EFX_CHANNEL_MAGIC_FILL(rx_queue)) {
1146 /* The queue must be empty, so we won't receive any rx
1147 * events, so efx_process_channel() won't refill the
1148 * queue. Refill it here */
1149 efx_fast_push_rx_descriptors(rx_queue, true);
1150 } else if (rx_queue && magic == EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue)) {
1151 efx_farch_handle_drain_event(channel);
1152 } else if (code == _EFX_CHANNEL_MAGIC_TX_DRAIN) {
1153 efx_farch_handle_drain_event(channel);
1155 netif_dbg(efx, hw, efx->net_dev, "channel %d received "
1156 "generated event "EFX_QWORD_FMT"\n",
1157 channel->channel, EFX_QWORD_VAL(*event));
1162 efx_farch_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
1164 struct efx_nic *efx = channel->efx;
1165 unsigned int ev_sub_code;
1166 unsigned int ev_sub_data;
1168 ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
1169 ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1171 switch (ev_sub_code) {
1172 case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
1173 netif_vdbg(efx, hw, efx->net_dev, "channel %d TXQ %d flushed\n",
1174 channel->channel, ev_sub_data);
1175 efx_farch_handle_tx_flush_done(efx, event);
1176 efx_sriov_tx_flush_done(efx, event);
1178 case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
1179 netif_vdbg(efx, hw, efx->net_dev, "channel %d RXQ %d flushed\n",
1180 channel->channel, ev_sub_data);
1181 efx_farch_handle_rx_flush_done(efx, event);
1182 efx_sriov_rx_flush_done(efx, event);
1184 case FSE_AZ_EVQ_INIT_DONE_EV:
1185 netif_dbg(efx, hw, efx->net_dev,
1186 "channel %d EVQ %d initialised\n",
1187 channel->channel, ev_sub_data);
1189 case FSE_AZ_SRM_UPD_DONE_EV:
1190 netif_vdbg(efx, hw, efx->net_dev,
1191 "channel %d SRAM update done\n", channel->channel);
1193 case FSE_AZ_WAKE_UP_EV:
1194 netif_vdbg(efx, hw, efx->net_dev,
1195 "channel %d RXQ %d wakeup event\n",
1196 channel->channel, ev_sub_data);
1198 case FSE_AZ_TIMER_EV:
1199 netif_vdbg(efx, hw, efx->net_dev,
1200 "channel %d RX queue %d timer expired\n",
1201 channel->channel, ev_sub_data);
1203 case FSE_AA_RX_RECOVER_EV:
1204 netif_err(efx, rx_err, efx->net_dev,
1205 "channel %d seen DRIVER RX_RESET event. "
1206 "Resetting.\n", channel->channel);
1207 atomic_inc(&efx->rx_reset);
1208 efx_schedule_reset(efx,
1209 EFX_WORKAROUND_6555(efx) ?
1210 RESET_TYPE_RX_RECOVERY :
1211 RESET_TYPE_DISABLE);
1213 case FSE_BZ_RX_DSC_ERROR_EV:
1214 if (ev_sub_data < EFX_VI_BASE) {
1215 netif_err(efx, rx_err, efx->net_dev,
1216 "RX DMA Q %d reports descriptor fetch error."
1217 " RX Q %d is disabled.\n", ev_sub_data,
1219 efx_schedule_reset(efx, RESET_TYPE_DMA_ERROR);
1221 efx_sriov_desc_fetch_err(efx, ev_sub_data);
1223 case FSE_BZ_TX_DSC_ERROR_EV:
1224 if (ev_sub_data < EFX_VI_BASE) {
1225 netif_err(efx, tx_err, efx->net_dev,
1226 "TX DMA Q %d reports descriptor fetch error."
1227 " TX Q %d is disabled.\n", ev_sub_data,
1229 efx_schedule_reset(efx, RESET_TYPE_DMA_ERROR);
1231 efx_sriov_desc_fetch_err(efx, ev_sub_data);
1234 netif_vdbg(efx, hw, efx->net_dev,
1235 "channel %d unknown driver event code %d "
1236 "data %04x\n", channel->channel, ev_sub_code,
1242 int efx_farch_ev_process(struct efx_channel *channel, int budget)
1244 struct efx_nic *efx = channel->efx;
1245 unsigned int read_ptr;
1246 efx_qword_t event, *p_event;
1254 read_ptr = channel->eventq_read_ptr;
1257 p_event = efx_event(channel, read_ptr);
1260 if (!efx_event_present(&event))
1264 netif_vdbg(channel->efx, intr, channel->efx->net_dev,
1265 "channel %d event is "EFX_QWORD_FMT"\n",
1266 channel->channel, EFX_QWORD_VAL(event));
1268 /* Clear this event by marking it all ones */
1269 EFX_SET_QWORD(*p_event);
1273 ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
1276 case FSE_AZ_EV_CODE_RX_EV:
1277 efx_farch_handle_rx_event(channel, &event);
1278 if (++spent == budget)
1281 case FSE_AZ_EV_CODE_TX_EV:
1282 tx_packets += efx_farch_handle_tx_event(channel,
1284 if (tx_packets > efx->txq_entries) {
1289 case FSE_AZ_EV_CODE_DRV_GEN_EV:
1290 efx_farch_handle_generated_event(channel, &event);
1292 case FSE_AZ_EV_CODE_DRIVER_EV:
1293 efx_farch_handle_driver_event(channel, &event);
1295 case FSE_CZ_EV_CODE_USER_EV:
1296 efx_sriov_event(channel, &event);
1298 case FSE_CZ_EV_CODE_MCDI_EV:
1299 efx_mcdi_process_event(channel, &event);
1301 case FSE_AZ_EV_CODE_GLOBAL_EV:
1302 if (efx->type->handle_global_event &&
1303 efx->type->handle_global_event(channel, &event))
1305 /* else fall through */
1307 netif_err(channel->efx, hw, channel->efx->net_dev,
1308 "channel %d unknown event type %d (data "
1309 EFX_QWORD_FMT ")\n", channel->channel,
1310 ev_code, EFX_QWORD_VAL(event));
1315 channel->eventq_read_ptr = read_ptr;
1319 /* Allocate buffer table entries for event queue */
1320 int efx_farch_ev_probe(struct efx_channel *channel)
1322 struct efx_nic *efx = channel->efx;
1325 entries = channel->eventq_mask + 1;
1326 return efx_alloc_special_buffer(efx, &channel->eventq,
1327 entries * sizeof(efx_qword_t));
1330 int efx_farch_ev_init(struct efx_channel *channel)
1333 struct efx_nic *efx = channel->efx;
1335 netif_dbg(efx, hw, efx->net_dev,
1336 "channel %d event queue in special buffers %d-%d\n",
1337 channel->channel, channel->eventq.index,
1338 channel->eventq.index + channel->eventq.entries - 1);
1340 if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1341 EFX_POPULATE_OWORD_3(reg,
1342 FRF_CZ_TIMER_Q_EN, 1,
1343 FRF_CZ_HOST_NOTIFY_MODE, 0,
1344 FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
1345 efx_writeo_table(efx, ®, FR_BZ_TIMER_TBL, channel->channel);
1348 /* Pin event queue buffer */
1349 efx_init_special_buffer(efx, &channel->eventq);
1351 /* Fill event queue with all ones (i.e. empty events) */
1352 memset(channel->eventq.buf.addr, 0xff, channel->eventq.buf.len);
1354 /* Push event queue to card */
1355 EFX_POPULATE_OWORD_3(reg,
1357 FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
1358 FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
1359 efx_writeo_table(efx, ®, efx->type->evq_ptr_tbl_base,
1365 void efx_farch_ev_fini(struct efx_channel *channel)
1368 struct efx_nic *efx = channel->efx;
1370 /* Remove event queue from card */
1371 EFX_ZERO_OWORD(reg);
1372 efx_writeo_table(efx, ®, efx->type->evq_ptr_tbl_base,
1374 if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1375 efx_writeo_table(efx, ®, FR_BZ_TIMER_TBL, channel->channel);
1377 /* Unpin event queue */
1378 efx_fini_special_buffer(efx, &channel->eventq);
1381 /* Free buffers backing event queue */
1382 void efx_farch_ev_remove(struct efx_channel *channel)
1384 efx_free_special_buffer(channel->efx, &channel->eventq);
1388 void efx_farch_ev_test_generate(struct efx_channel *channel)
1390 efx_farch_magic_event(channel, EFX_CHANNEL_MAGIC_TEST(channel));
1393 void efx_farch_rx_defer_refill(struct efx_rx_queue *rx_queue)
1395 efx_farch_magic_event(efx_rx_queue_channel(rx_queue),
1396 EFX_CHANNEL_MAGIC_FILL(rx_queue));
1399 /**************************************************************************
1401 * Hardware interrupts
1402 * The hardware interrupt handler does very little work; all the event
1403 * queue processing is carried out by per-channel tasklets.
1405 **************************************************************************/
1407 /* Enable/disable/generate interrupts */
1408 static inline void efx_farch_interrupts(struct efx_nic *efx,
1409 bool enabled, bool force)
1411 efx_oword_t int_en_reg_ker;
1413 EFX_POPULATE_OWORD_3(int_en_reg_ker,
1414 FRF_AZ_KER_INT_LEVE_SEL, efx->irq_level,
1415 FRF_AZ_KER_INT_KER, force,
1416 FRF_AZ_DRV_INT_EN_KER, enabled);
1417 efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
1420 void efx_farch_irq_enable_master(struct efx_nic *efx)
1422 EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
1423 wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
1425 efx_farch_interrupts(efx, true, false);
1428 void efx_farch_irq_disable_master(struct efx_nic *efx)
1430 /* Disable interrupts */
1431 efx_farch_interrupts(efx, false, false);
1434 /* Generate a test interrupt
1435 * Interrupt must already have been enabled, otherwise nasty things
1438 void efx_farch_irq_test_generate(struct efx_nic *efx)
1440 efx_farch_interrupts(efx, true, true);
1443 /* Process a fatal interrupt
1444 * Disable bus mastering ASAP and schedule a reset
1446 irqreturn_t efx_farch_fatal_interrupt(struct efx_nic *efx)
1448 struct falcon_nic_data *nic_data = efx->nic_data;
1449 efx_oword_t *int_ker = efx->irq_status.addr;
1450 efx_oword_t fatal_intr;
1451 int error, mem_perr;
1453 efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
1454 error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
1456 netif_err(efx, hw, efx->net_dev, "SYSTEM ERROR "EFX_OWORD_FMT" status "
1457 EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
1458 EFX_OWORD_VAL(fatal_intr),
1459 error ? "disabling bus mastering" : "no recognised error");
1461 /* If this is a memory parity error dump which blocks are offending */
1462 mem_perr = (EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER) ||
1463 EFX_OWORD_FIELD(fatal_intr, FRF_AZ_SRM_PERR_INT_KER));
1466 efx_reado(efx, ®, FR_AZ_MEM_STAT);
1467 netif_err(efx, hw, efx->net_dev,
1468 "SYSTEM ERROR: memory parity error "EFX_OWORD_FMT"\n",
1469 EFX_OWORD_VAL(reg));
1472 /* Disable both devices */
1473 pci_clear_master(efx->pci_dev);
1474 if (efx_nic_is_dual_func(efx))
1475 pci_clear_master(nic_data->pci_dev2);
1476 efx_farch_irq_disable_master(efx);
1478 /* Count errors and reset or disable the NIC accordingly */
1479 if (efx->int_error_count == 0 ||
1480 time_after(jiffies, efx->int_error_expire)) {
1481 efx->int_error_count = 0;
1482 efx->int_error_expire =
1483 jiffies + EFX_INT_ERROR_EXPIRE * HZ;
1485 if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
1486 netif_err(efx, hw, efx->net_dev,
1487 "SYSTEM ERROR - reset scheduled\n");
1488 efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
1490 netif_err(efx, hw, efx->net_dev,
1491 "SYSTEM ERROR - max number of errors seen."
1492 "NIC will be disabled\n");
1493 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1499 /* Handle a legacy interrupt
1500 * Acknowledges the interrupt and schedule event queue processing.
1502 irqreturn_t efx_farch_legacy_interrupt(int irq, void *dev_id)
1504 struct efx_nic *efx = dev_id;
1505 bool soft_enabled = ACCESS_ONCE(efx->irq_soft_enabled);
1506 efx_oword_t *int_ker = efx->irq_status.addr;
1507 irqreturn_t result = IRQ_NONE;
1508 struct efx_channel *channel;
1513 /* Read the ISR which also ACKs the interrupts */
1514 efx_readd(efx, ®, FR_BZ_INT_ISR0);
1515 queues = EFX_EXTRACT_DWORD(reg, 0, 31);
1517 /* Legacy interrupts are disabled too late by the EEH kernel
1518 * code. Disable them earlier.
1519 * If an EEH error occurred, the read will have returned all ones.
1521 if (EFX_DWORD_IS_ALL_ONES(reg) && efx_try_recovery(efx) &&
1522 !efx->eeh_disabled_legacy_irq) {
1523 disable_irq_nosync(efx->legacy_irq);
1524 efx->eeh_disabled_legacy_irq = true;
1527 /* Handle non-event-queue sources */
1528 if (queues & (1U << efx->irq_level) && soft_enabled) {
1529 syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1530 if (unlikely(syserr))
1531 return efx_farch_fatal_interrupt(efx);
1532 efx->last_irq_cpu = raw_smp_processor_id();
1536 efx->irq_zero_count = 0;
1538 /* Schedule processing of any interrupting queues */
1539 if (likely(soft_enabled)) {
1540 efx_for_each_channel(channel, efx) {
1542 efx_schedule_channel_irq(channel);
1546 result = IRQ_HANDLED;
1551 /* Legacy ISR read can return zero once (SF bug 15783) */
1553 /* We can't return IRQ_HANDLED more than once on seeing ISR=0
1554 * because this might be a shared interrupt. */
1555 if (efx->irq_zero_count++ == 0)
1556 result = IRQ_HANDLED;
1558 /* Ensure we schedule or rearm all event queues */
1559 if (likely(soft_enabled)) {
1560 efx_for_each_channel(channel, efx) {
1561 event = efx_event(channel,
1562 channel->eventq_read_ptr);
1563 if (efx_event_present(event))
1564 efx_schedule_channel_irq(channel);
1566 efx_farch_ev_read_ack(channel);
1571 if (result == IRQ_HANDLED)
1572 netif_vdbg(efx, intr, efx->net_dev,
1573 "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
1574 irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
1579 /* Handle an MSI interrupt
1581 * Handle an MSI hardware interrupt. This routine schedules event
1582 * queue processing. No interrupt acknowledgement cycle is necessary.
1583 * Also, we never need to check that the interrupt is for us, since
1584 * MSI interrupts cannot be shared.
1586 irqreturn_t efx_farch_msi_interrupt(int irq, void *dev_id)
1588 struct efx_msi_context *context = dev_id;
1589 struct efx_nic *efx = context->efx;
1590 efx_oword_t *int_ker = efx->irq_status.addr;
1593 netif_vdbg(efx, intr, efx->net_dev,
1594 "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
1595 irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
1597 if (!likely(ACCESS_ONCE(efx->irq_soft_enabled)))
1600 /* Handle non-event-queue sources */
1601 if (context->index == efx->irq_level) {
1602 syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1603 if (unlikely(syserr))
1604 return efx_farch_fatal_interrupt(efx);
1605 efx->last_irq_cpu = raw_smp_processor_id();
1608 /* Schedule processing of the channel */
1609 efx_schedule_channel_irq(efx->channel[context->index]);
1614 /* Setup RSS indirection table.
1615 * This maps from the hash value of the packet to RXQ
1617 void efx_farch_rx_push_indir_table(struct efx_nic *efx)
1622 BUG_ON(efx_nic_rev(efx) < EFX_REV_FALCON_B0);
1624 BUILD_BUG_ON(ARRAY_SIZE(efx->rx_indir_table) !=
1625 FR_BZ_RX_INDIRECTION_TBL_ROWS);
1627 for (i = 0; i < FR_BZ_RX_INDIRECTION_TBL_ROWS; i++) {
1628 EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
1629 efx->rx_indir_table[i]);
1630 efx_writed(efx, &dword,
1631 FR_BZ_RX_INDIRECTION_TBL +
1632 FR_BZ_RX_INDIRECTION_TBL_STEP * i);
1636 /* Looks at available SRAM resources and works out how many queues we
1637 * can support, and where things like descriptor caches should live.
1639 * SRAM is split up as follows:
1640 * 0 buftbl entries for channels
1641 * efx->vf_buftbl_base buftbl entries for SR-IOV
1642 * efx->rx_dc_base RX descriptor caches
1643 * efx->tx_dc_base TX descriptor caches
1645 void efx_farch_dimension_resources(struct efx_nic *efx, unsigned sram_lim_qw)
1647 unsigned vi_count, buftbl_min;
1649 /* Account for the buffer table entries backing the datapath channels
1650 * and the descriptor caches for those channels.
1652 buftbl_min = ((efx->n_rx_channels * EFX_MAX_DMAQ_SIZE +
1653 efx->n_tx_channels * EFX_TXQ_TYPES * EFX_MAX_DMAQ_SIZE +
1654 efx->n_channels * EFX_MAX_EVQ_SIZE)
1655 * sizeof(efx_qword_t) / EFX_BUF_SIZE);
1656 vi_count = max(efx->n_channels, efx->n_tx_channels * EFX_TXQ_TYPES);
1658 #ifdef CONFIG_SFC_SRIOV
1659 if (efx_sriov_wanted(efx)) {
1660 unsigned vi_dc_entries, buftbl_free, entries_per_vf, vf_limit;
1662 efx->vf_buftbl_base = buftbl_min;
1664 vi_dc_entries = RX_DC_ENTRIES + TX_DC_ENTRIES;
1665 vi_count = max(vi_count, EFX_VI_BASE);
1666 buftbl_free = (sram_lim_qw - buftbl_min -
1667 vi_count * vi_dc_entries);
1669 entries_per_vf = ((vi_dc_entries + EFX_VF_BUFTBL_PER_VI) *
1671 vf_limit = min(buftbl_free / entries_per_vf,
1672 (1024U - EFX_VI_BASE) >> efx->vi_scale);
1674 if (efx->vf_count > vf_limit) {
1675 netif_err(efx, probe, efx->net_dev,
1676 "Reducing VF count from from %d to %d\n",
1677 efx->vf_count, vf_limit);
1678 efx->vf_count = vf_limit;
1680 vi_count += efx->vf_count * efx_vf_size(efx);
1684 efx->tx_dc_base = sram_lim_qw - vi_count * TX_DC_ENTRIES;
1685 efx->rx_dc_base = efx->tx_dc_base - vi_count * RX_DC_ENTRIES;
1688 u32 efx_farch_fpga_ver(struct efx_nic *efx)
1690 efx_oword_t altera_build;
1691 efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
1692 return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
1695 void efx_farch_init_common(struct efx_nic *efx)
1699 /* Set positions of descriptor caches in SRAM. */
1700 EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR, efx->tx_dc_base);
1701 efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
1702 EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR, efx->rx_dc_base);
1703 efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
1705 /* Set TX descriptor cache size. */
1706 BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
1707 EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
1708 efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
1710 /* Set RX descriptor cache size. Set low watermark to size-8, as
1711 * this allows most efficient prefetching.
1713 BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
1714 EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
1715 efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
1716 EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
1717 efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
1719 /* Program INT_KER address */
1720 EFX_POPULATE_OWORD_2(temp,
1721 FRF_AZ_NORM_INT_VEC_DIS_KER,
1722 EFX_INT_MODE_USE_MSI(efx),
1723 FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
1724 efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
1726 if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
1727 /* Use an interrupt level unused by event queues */
1728 efx->irq_level = 0x1f;
1730 /* Use a valid MSI-X vector */
1733 /* Enable all the genuinely fatal interrupts. (They are still
1734 * masked by the overall interrupt mask, controlled by
1735 * falcon_interrupts()).
1737 * Note: All other fatal interrupts are enabled
1739 EFX_POPULATE_OWORD_3(temp,
1740 FRF_AZ_ILL_ADR_INT_KER_EN, 1,
1741 FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
1742 FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
1743 if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1744 EFX_SET_OWORD_FIELD(temp, FRF_CZ_SRAM_PERR_INT_P_KER_EN, 1);
1745 EFX_INVERT_OWORD(temp);
1746 efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
1748 /* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
1749 * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
1751 efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
1752 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
1753 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
1754 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
1755 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 1);
1756 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
1757 /* Enable SW_EV to inherit in char driver - assume harmless here */
1758 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
1759 /* Prefetch threshold 2 => fetch when descriptor cache half empty */
1760 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
1761 /* Disable hardware watchdog which can misfire */
1762 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
1763 /* Squash TX of packets of 16 bytes or less */
1764 if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1765 EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
1766 efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
1768 if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
1769 EFX_POPULATE_OWORD_4(temp,
1770 /* Default values */
1771 FRF_BZ_TX_PACE_SB_NOT_AF, 0x15,
1772 FRF_BZ_TX_PACE_SB_AF, 0xb,
1773 FRF_BZ_TX_PACE_FB_BASE, 0,
1774 /* Allow large pace values in the
1776 FRF_BZ_TX_PACE_BIN_TH,
1777 FFE_BZ_TX_PACE_RESERVED);
1778 efx_writeo(efx, &temp, FR_BZ_TX_PACE);
1782 /**************************************************************************
1786 **************************************************************************
1789 /* "Fudge factors" - difference between programmed value and actual depth.
1790 * Due to pipelined implementation we need to program H/W with a value that
1791 * is larger than the hop limit we want.
1793 #define EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD 3
1794 #define EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL 1
1796 /* Hard maximum search limit. Hardware will time-out beyond 200-something.
1797 * We also need to avoid infinite loops in efx_farch_filter_search() when the
1800 #define EFX_FARCH_FILTER_CTL_SRCH_MAX 200
1802 /* Don't try very hard to find space for performance hints, as this is
1803 * counter-productive. */
1804 #define EFX_FARCH_FILTER_CTL_SRCH_HINT_MAX 5
1806 enum efx_farch_filter_type {
1807 EFX_FARCH_FILTER_TCP_FULL = 0,
1808 EFX_FARCH_FILTER_TCP_WILD,
1809 EFX_FARCH_FILTER_UDP_FULL,
1810 EFX_FARCH_FILTER_UDP_WILD,
1811 EFX_FARCH_FILTER_MAC_FULL = 4,
1812 EFX_FARCH_FILTER_MAC_WILD,
1813 EFX_FARCH_FILTER_UC_DEF = 8,
1814 EFX_FARCH_FILTER_MC_DEF,
1815 EFX_FARCH_FILTER_TYPE_COUNT, /* number of specific types */
1818 enum efx_farch_filter_table_id {
1819 EFX_FARCH_FILTER_TABLE_RX_IP = 0,
1820 EFX_FARCH_FILTER_TABLE_RX_MAC,
1821 EFX_FARCH_FILTER_TABLE_RX_DEF,
1822 EFX_FARCH_FILTER_TABLE_TX_MAC,
1823 EFX_FARCH_FILTER_TABLE_COUNT,
1826 enum efx_farch_filter_index {
1827 EFX_FARCH_FILTER_INDEX_UC_DEF,
1828 EFX_FARCH_FILTER_INDEX_MC_DEF,
1829 EFX_FARCH_FILTER_SIZE_RX_DEF,
1832 struct efx_farch_filter_spec {
1840 struct efx_farch_filter_table {
1841 enum efx_farch_filter_table_id id;
1842 u32 offset; /* address of table relative to BAR */
1843 unsigned size; /* number of entries */
1844 unsigned step; /* step between entries */
1845 unsigned used; /* number currently used */
1846 unsigned long *used_bitmap;
1847 struct efx_farch_filter_spec *spec;
1848 unsigned search_limit[EFX_FARCH_FILTER_TYPE_COUNT];
1851 struct efx_farch_filter_state {
1852 struct efx_farch_filter_table table[EFX_FARCH_FILTER_TABLE_COUNT];
1856 efx_farch_filter_table_clear_entry(struct efx_nic *efx,
1857 struct efx_farch_filter_table *table,
1858 unsigned int filter_idx);
1860 /* The filter hash function is LFSR polynomial x^16 + x^3 + 1 of a 32-bit
1861 * key derived from the n-tuple. The initial LFSR state is 0xffff. */
1862 static u16 efx_farch_filter_hash(u32 key)
1866 /* First 16 rounds */
1867 tmp = 0x1fff ^ key >> 16;
1868 tmp = tmp ^ tmp >> 3 ^ tmp >> 6;
1869 tmp = tmp ^ tmp >> 9;
1870 /* Last 16 rounds */
1871 tmp = tmp ^ tmp << 13 ^ key;
1872 tmp = tmp ^ tmp >> 3 ^ tmp >> 6;
1873 return tmp ^ tmp >> 9;
1876 /* To allow for hash collisions, filter search continues at these
1877 * increments from the first possible entry selected by the hash. */
1878 static u16 efx_farch_filter_increment(u32 key)
1883 static enum efx_farch_filter_table_id
1884 efx_farch_filter_spec_table_id(const struct efx_farch_filter_spec *spec)
1886 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1887 (EFX_FARCH_FILTER_TCP_FULL >> 2));
1888 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1889 (EFX_FARCH_FILTER_TCP_WILD >> 2));
1890 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1891 (EFX_FARCH_FILTER_UDP_FULL >> 2));
1892 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1893 (EFX_FARCH_FILTER_UDP_WILD >> 2));
1894 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_MAC !=
1895 (EFX_FARCH_FILTER_MAC_FULL >> 2));
1896 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_MAC !=
1897 (EFX_FARCH_FILTER_MAC_WILD >> 2));
1898 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_TX_MAC !=
1899 EFX_FARCH_FILTER_TABLE_RX_MAC + 2);
1900 return (spec->type >> 2) + ((spec->flags & EFX_FILTER_FLAG_TX) ? 2 : 0);
1903 static void efx_farch_filter_push_rx_config(struct efx_nic *efx)
1905 struct efx_farch_filter_state *state = efx->filter_state;
1906 struct efx_farch_filter_table *table;
1907 efx_oword_t filter_ctl;
1909 efx_reado(efx, &filter_ctl, FR_BZ_RX_FILTER_CTL);
1911 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_IP];
1912 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_TCP_FULL_SRCH_LIMIT,
1913 table->search_limit[EFX_FARCH_FILTER_TCP_FULL] +
1914 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
1915 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_TCP_WILD_SRCH_LIMIT,
1916 table->search_limit[EFX_FARCH_FILTER_TCP_WILD] +
1917 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
1918 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_UDP_FULL_SRCH_LIMIT,
1919 table->search_limit[EFX_FARCH_FILTER_UDP_FULL] +
1920 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
1921 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_UDP_WILD_SRCH_LIMIT,
1922 table->search_limit[EFX_FARCH_FILTER_UDP_WILD] +
1923 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
1925 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_MAC];
1927 EFX_SET_OWORD_FIELD(
1928 filter_ctl, FRF_CZ_ETHERNET_FULL_SEARCH_LIMIT,
1929 table->search_limit[EFX_FARCH_FILTER_MAC_FULL] +
1930 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
1931 EFX_SET_OWORD_FIELD(
1932 filter_ctl, FRF_CZ_ETHERNET_WILDCARD_SEARCH_LIMIT,
1933 table->search_limit[EFX_FARCH_FILTER_MAC_WILD] +
1934 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
1937 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_DEF];
1939 EFX_SET_OWORD_FIELD(
1940 filter_ctl, FRF_CZ_UNICAST_NOMATCH_Q_ID,
1941 table->spec[EFX_FARCH_FILTER_INDEX_UC_DEF].dmaq_id);
1942 EFX_SET_OWORD_FIELD(
1943 filter_ctl, FRF_CZ_UNICAST_NOMATCH_RSS_ENABLED,
1944 !!(table->spec[EFX_FARCH_FILTER_INDEX_UC_DEF].flags &
1945 EFX_FILTER_FLAG_RX_RSS));
1946 EFX_SET_OWORD_FIELD(
1947 filter_ctl, FRF_CZ_MULTICAST_NOMATCH_Q_ID,
1948 table->spec[EFX_FARCH_FILTER_INDEX_MC_DEF].dmaq_id);
1949 EFX_SET_OWORD_FIELD(
1950 filter_ctl, FRF_CZ_MULTICAST_NOMATCH_RSS_ENABLED,
1951 !!(table->spec[EFX_FARCH_FILTER_INDEX_MC_DEF].flags &
1952 EFX_FILTER_FLAG_RX_RSS));
1954 /* There is a single bit to enable RX scatter for all
1955 * unmatched packets. Only set it if scatter is
1956 * enabled in both filter specs.
1958 EFX_SET_OWORD_FIELD(
1959 filter_ctl, FRF_BZ_SCATTER_ENBL_NO_MATCH_Q,
1960 !!(table->spec[EFX_FARCH_FILTER_INDEX_UC_DEF].flags &
1961 table->spec[EFX_FARCH_FILTER_INDEX_MC_DEF].flags &
1962 EFX_FILTER_FLAG_RX_SCATTER));
1963 } else if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
1964 /* We don't expose 'default' filters because unmatched
1965 * packets always go to the queue number found in the
1966 * RSS table. But we still need to set the RX scatter
1969 EFX_SET_OWORD_FIELD(
1970 filter_ctl, FRF_BZ_SCATTER_ENBL_NO_MATCH_Q,
1974 efx_writeo(efx, &filter_ctl, FR_BZ_RX_FILTER_CTL);
1977 static void efx_farch_filter_push_tx_limits(struct efx_nic *efx)
1979 struct efx_farch_filter_state *state = efx->filter_state;
1980 struct efx_farch_filter_table *table;
1983 efx_reado(efx, &tx_cfg, FR_AZ_TX_CFG);
1985 table = &state->table[EFX_FARCH_FILTER_TABLE_TX_MAC];
1987 EFX_SET_OWORD_FIELD(
1988 tx_cfg, FRF_CZ_TX_ETH_FILTER_FULL_SEARCH_RANGE,
1989 table->search_limit[EFX_FARCH_FILTER_MAC_FULL] +
1990 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
1991 EFX_SET_OWORD_FIELD(
1992 tx_cfg, FRF_CZ_TX_ETH_FILTER_WILD_SEARCH_RANGE,
1993 table->search_limit[EFX_FARCH_FILTER_MAC_WILD] +
1994 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
1997 efx_writeo(efx, &tx_cfg, FR_AZ_TX_CFG);
2001 efx_farch_filter_from_gen_spec(struct efx_farch_filter_spec *spec,
2002 const struct efx_filter_spec *gen_spec)
2004 bool is_full = false;
2006 if ((gen_spec->flags & EFX_FILTER_FLAG_RX_RSS) &&
2007 gen_spec->rss_context != EFX_FILTER_RSS_CONTEXT_DEFAULT)
2010 spec->priority = gen_spec->priority;
2011 spec->flags = gen_spec->flags;
2012 spec->dmaq_id = gen_spec->dmaq_id;
2014 switch (gen_spec->match_flags) {
2015 case (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
2016 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
2017 EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT):
2020 case (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
2021 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT): {
2022 __be32 rhost, host1, host2;
2023 __be16 rport, port1, port2;
2025 EFX_BUG_ON_PARANOID(!(gen_spec->flags & EFX_FILTER_FLAG_RX));
2027 if (gen_spec->ether_type != htons(ETH_P_IP))
2028 return -EPROTONOSUPPORT;
2029 if (gen_spec->loc_port == 0 ||
2030 (is_full && gen_spec->rem_port == 0))
2031 return -EADDRNOTAVAIL;
2032 switch (gen_spec->ip_proto) {
2034 spec->type = (is_full ? EFX_FARCH_FILTER_TCP_FULL :
2035 EFX_FARCH_FILTER_TCP_WILD);
2038 spec->type = (is_full ? EFX_FARCH_FILTER_UDP_FULL :
2039 EFX_FARCH_FILTER_UDP_WILD);
2042 return -EPROTONOSUPPORT;
2045 /* Filter is constructed in terms of source and destination,
2046 * with the odd wrinkle that the ports are swapped in a UDP
2047 * wildcard filter. We need to convert from local and remote
2048 * (= zero for wildcard) addresses.
2050 rhost = is_full ? gen_spec->rem_host[0] : 0;
2051 rport = is_full ? gen_spec->rem_port : 0;
2053 host2 = gen_spec->loc_host[0];
2054 if (!is_full && gen_spec->ip_proto == IPPROTO_UDP) {
2055 port1 = gen_spec->loc_port;
2059 port2 = gen_spec->loc_port;
2061 spec->data[0] = ntohl(host1) << 16 | ntohs(port1);
2062 spec->data[1] = ntohs(port2) << 16 | ntohl(host1) >> 16;
2063 spec->data[2] = ntohl(host2);
2068 case EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_OUTER_VID:
2071 case EFX_FILTER_MATCH_LOC_MAC:
2072 spec->type = (is_full ? EFX_FARCH_FILTER_MAC_FULL :
2073 EFX_FARCH_FILTER_MAC_WILD);
2074 spec->data[0] = is_full ? ntohs(gen_spec->outer_vid) : 0;
2075 spec->data[1] = (gen_spec->loc_mac[2] << 24 |
2076 gen_spec->loc_mac[3] << 16 |
2077 gen_spec->loc_mac[4] << 8 |
2078 gen_spec->loc_mac[5]);
2079 spec->data[2] = (gen_spec->loc_mac[0] << 8 |
2080 gen_spec->loc_mac[1]);
2083 case EFX_FILTER_MATCH_LOC_MAC_IG:
2084 spec->type = (is_multicast_ether_addr(gen_spec->loc_mac) ?
2085 EFX_FARCH_FILTER_MC_DEF :
2086 EFX_FARCH_FILTER_UC_DEF);
2087 memset(spec->data, 0, sizeof(spec->data)); /* ensure equality */
2091 return -EPROTONOSUPPORT;
2098 efx_farch_filter_to_gen_spec(struct efx_filter_spec *gen_spec,
2099 const struct efx_farch_filter_spec *spec)
2101 bool is_full = false;
2103 /* *gen_spec should be completely initialised, to be consistent
2104 * with efx_filter_init_{rx,tx}() and in case we want to copy
2105 * it back to userland.
2107 memset(gen_spec, 0, sizeof(*gen_spec));
2109 gen_spec->priority = spec->priority;
2110 gen_spec->flags = spec->flags;
2111 gen_spec->dmaq_id = spec->dmaq_id;
2113 switch (spec->type) {
2114 case EFX_FARCH_FILTER_TCP_FULL:
2115 case EFX_FARCH_FILTER_UDP_FULL:
2118 case EFX_FARCH_FILTER_TCP_WILD:
2119 case EFX_FARCH_FILTER_UDP_WILD: {
2120 __be32 host1, host2;
2121 __be16 port1, port2;
2123 gen_spec->match_flags =
2124 EFX_FILTER_MATCH_ETHER_TYPE |
2125 EFX_FILTER_MATCH_IP_PROTO |
2126 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT;
2128 gen_spec->match_flags |= (EFX_FILTER_MATCH_REM_HOST |
2129 EFX_FILTER_MATCH_REM_PORT);
2130 gen_spec->ether_type = htons(ETH_P_IP);
2131 gen_spec->ip_proto =
2132 (spec->type == EFX_FARCH_FILTER_TCP_FULL ||
2133 spec->type == EFX_FARCH_FILTER_TCP_WILD) ?
2134 IPPROTO_TCP : IPPROTO_UDP;
2136 host1 = htonl(spec->data[0] >> 16 | spec->data[1] << 16);
2137 port1 = htons(spec->data[0]);
2138 host2 = htonl(spec->data[2]);
2139 port2 = htons(spec->data[1] >> 16);
2140 if (spec->flags & EFX_FILTER_FLAG_TX) {
2141 gen_spec->loc_host[0] = host1;
2142 gen_spec->rem_host[0] = host2;
2144 gen_spec->loc_host[0] = host2;
2145 gen_spec->rem_host[0] = host1;
2147 if (!!(gen_spec->flags & EFX_FILTER_FLAG_TX) ^
2148 (!is_full && gen_spec->ip_proto == IPPROTO_UDP)) {
2149 gen_spec->loc_port = port1;
2150 gen_spec->rem_port = port2;
2152 gen_spec->loc_port = port2;
2153 gen_spec->rem_port = port1;
2159 case EFX_FARCH_FILTER_MAC_FULL:
2162 case EFX_FARCH_FILTER_MAC_WILD:
2163 gen_spec->match_flags = EFX_FILTER_MATCH_LOC_MAC;
2165 gen_spec->match_flags |= EFX_FILTER_MATCH_OUTER_VID;
2166 gen_spec->loc_mac[0] = spec->data[2] >> 8;
2167 gen_spec->loc_mac[1] = spec->data[2];
2168 gen_spec->loc_mac[2] = spec->data[1] >> 24;
2169 gen_spec->loc_mac[3] = spec->data[1] >> 16;
2170 gen_spec->loc_mac[4] = spec->data[1] >> 8;
2171 gen_spec->loc_mac[5] = spec->data[1];
2172 gen_spec->outer_vid = htons(spec->data[0]);
2175 case EFX_FARCH_FILTER_UC_DEF:
2176 case EFX_FARCH_FILTER_MC_DEF:
2177 gen_spec->match_flags = EFX_FILTER_MATCH_LOC_MAC_IG;
2178 gen_spec->loc_mac[0] = spec->type == EFX_FARCH_FILTER_MC_DEF;
2188 efx_farch_filter_init_rx_auto(struct efx_nic *efx,
2189 struct efx_farch_filter_spec *spec)
2191 /* If there's only one channel then disable RSS for non VF
2192 * traffic, thereby allowing VFs to use RSS when the PF can't.
2194 spec->priority = EFX_FILTER_PRI_AUTO;
2195 spec->flags = (EFX_FILTER_FLAG_RX |
2196 (efx->n_rx_channels > 1 ? EFX_FILTER_FLAG_RX_RSS : 0) |
2197 (efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0));
2201 /* Build a filter entry and return its n-tuple key. */
2202 static u32 efx_farch_filter_build(efx_oword_t *filter,
2203 struct efx_farch_filter_spec *spec)
2207 switch (efx_farch_filter_spec_table_id(spec)) {
2208 case EFX_FARCH_FILTER_TABLE_RX_IP: {
2209 bool is_udp = (spec->type == EFX_FARCH_FILTER_UDP_FULL ||
2210 spec->type == EFX_FARCH_FILTER_UDP_WILD);
2211 EFX_POPULATE_OWORD_7(
2214 !!(spec->flags & EFX_FILTER_FLAG_RX_RSS),
2216 !!(spec->flags & EFX_FILTER_FLAG_RX_SCATTER),
2217 FRF_BZ_TCP_UDP, is_udp,
2218 FRF_BZ_RXQ_ID, spec->dmaq_id,
2219 EFX_DWORD_2, spec->data[2],
2220 EFX_DWORD_1, spec->data[1],
2221 EFX_DWORD_0, spec->data[0]);
2226 case EFX_FARCH_FILTER_TABLE_RX_MAC: {
2227 bool is_wild = spec->type == EFX_FARCH_FILTER_MAC_WILD;
2228 EFX_POPULATE_OWORD_7(
2231 !!(spec->flags & EFX_FILTER_FLAG_RX_RSS),
2232 FRF_CZ_RMFT_SCATTER_EN,
2233 !!(spec->flags & EFX_FILTER_FLAG_RX_SCATTER),
2234 FRF_CZ_RMFT_RXQ_ID, spec->dmaq_id,
2235 FRF_CZ_RMFT_WILDCARD_MATCH, is_wild,
2236 FRF_CZ_RMFT_DEST_MAC_HI, spec->data[2],
2237 FRF_CZ_RMFT_DEST_MAC_LO, spec->data[1],
2238 FRF_CZ_RMFT_VLAN_ID, spec->data[0]);
2243 case EFX_FARCH_FILTER_TABLE_TX_MAC: {
2244 bool is_wild = spec->type == EFX_FARCH_FILTER_MAC_WILD;
2245 EFX_POPULATE_OWORD_5(*filter,
2246 FRF_CZ_TMFT_TXQ_ID, spec->dmaq_id,
2247 FRF_CZ_TMFT_WILDCARD_MATCH, is_wild,
2248 FRF_CZ_TMFT_SRC_MAC_HI, spec->data[2],
2249 FRF_CZ_TMFT_SRC_MAC_LO, spec->data[1],
2250 FRF_CZ_TMFT_VLAN_ID, spec->data[0]);
2251 data3 = is_wild | spec->dmaq_id << 1;
2259 return spec->data[0] ^ spec->data[1] ^ spec->data[2] ^ data3;
2262 static bool efx_farch_filter_equal(const struct efx_farch_filter_spec *left,
2263 const struct efx_farch_filter_spec *right)
2265 if (left->type != right->type ||
2266 memcmp(left->data, right->data, sizeof(left->data)))
2269 if (left->flags & EFX_FILTER_FLAG_TX &&
2270 left->dmaq_id != right->dmaq_id)
2277 * Construct/deconstruct external filter IDs. At least the RX filter
2278 * IDs must be ordered by matching priority, for RX NFC semantics.
2280 * Deconstruction needs to be robust against invalid IDs so that
2281 * efx_filter_remove_id_safe() and efx_filter_get_filter_safe() can
2282 * accept user-provided IDs.
2285 #define EFX_FARCH_FILTER_MATCH_PRI_COUNT 5
2287 static const u8 efx_farch_filter_type_match_pri[EFX_FARCH_FILTER_TYPE_COUNT] = {
2288 [EFX_FARCH_FILTER_TCP_FULL] = 0,
2289 [EFX_FARCH_FILTER_UDP_FULL] = 0,
2290 [EFX_FARCH_FILTER_TCP_WILD] = 1,
2291 [EFX_FARCH_FILTER_UDP_WILD] = 1,
2292 [EFX_FARCH_FILTER_MAC_FULL] = 2,
2293 [EFX_FARCH_FILTER_MAC_WILD] = 3,
2294 [EFX_FARCH_FILTER_UC_DEF] = 4,
2295 [EFX_FARCH_FILTER_MC_DEF] = 4,
2298 static const enum efx_farch_filter_table_id efx_farch_filter_range_table[] = {
2299 EFX_FARCH_FILTER_TABLE_RX_IP, /* RX match pri 0 */
2300 EFX_FARCH_FILTER_TABLE_RX_IP,
2301 EFX_FARCH_FILTER_TABLE_RX_MAC,
2302 EFX_FARCH_FILTER_TABLE_RX_MAC,
2303 EFX_FARCH_FILTER_TABLE_RX_DEF, /* RX match pri 4 */
2304 EFX_FARCH_FILTER_TABLE_TX_MAC, /* TX match pri 0 */
2305 EFX_FARCH_FILTER_TABLE_TX_MAC, /* TX match pri 1 */
2308 #define EFX_FARCH_FILTER_INDEX_WIDTH 13
2309 #define EFX_FARCH_FILTER_INDEX_MASK ((1 << EFX_FARCH_FILTER_INDEX_WIDTH) - 1)
2312 efx_farch_filter_make_id(const struct efx_farch_filter_spec *spec,
2317 range = efx_farch_filter_type_match_pri[spec->type];
2318 if (!(spec->flags & EFX_FILTER_FLAG_RX))
2319 range += EFX_FARCH_FILTER_MATCH_PRI_COUNT;
2321 return range << EFX_FARCH_FILTER_INDEX_WIDTH | index;
2324 static inline enum efx_farch_filter_table_id
2325 efx_farch_filter_id_table_id(u32 id)
2327 unsigned int range = id >> EFX_FARCH_FILTER_INDEX_WIDTH;
2329 if (range < ARRAY_SIZE(efx_farch_filter_range_table))
2330 return efx_farch_filter_range_table[range];
2332 return EFX_FARCH_FILTER_TABLE_COUNT; /* invalid */
2335 static inline unsigned int efx_farch_filter_id_index(u32 id)
2337 return id & EFX_FARCH_FILTER_INDEX_MASK;
2340 u32 efx_farch_filter_get_rx_id_limit(struct efx_nic *efx)
2342 struct efx_farch_filter_state *state = efx->filter_state;
2343 unsigned int range = EFX_FARCH_FILTER_MATCH_PRI_COUNT - 1;
2344 enum efx_farch_filter_table_id table_id;
2347 table_id = efx_farch_filter_range_table[range];
2348 if (state->table[table_id].size != 0)
2349 return range << EFX_FARCH_FILTER_INDEX_WIDTH |
2350 state->table[table_id].size;
2356 s32 efx_farch_filter_insert(struct efx_nic *efx,
2357 struct efx_filter_spec *gen_spec,
2360 struct efx_farch_filter_state *state = efx->filter_state;
2361 struct efx_farch_filter_table *table;
2362 struct efx_farch_filter_spec spec;
2364 int rep_index, ins_index;
2365 unsigned int depth = 0;
2368 rc = efx_farch_filter_from_gen_spec(&spec, gen_spec);
2372 table = &state->table[efx_farch_filter_spec_table_id(&spec)];
2373 if (table->size == 0)
2376 netif_vdbg(efx, hw, efx->net_dev,
2377 "%s: type %d search_limit=%d", __func__, spec.type,
2378 table->search_limit[spec.type]);
2380 if (table->id == EFX_FARCH_FILTER_TABLE_RX_DEF) {
2381 /* One filter spec per type */
2382 BUILD_BUG_ON(EFX_FARCH_FILTER_INDEX_UC_DEF != 0);
2383 BUILD_BUG_ON(EFX_FARCH_FILTER_INDEX_MC_DEF !=
2384 EFX_FARCH_FILTER_MC_DEF - EFX_FARCH_FILTER_UC_DEF);
2385 rep_index = spec.type - EFX_FARCH_FILTER_UC_DEF;
2386 ins_index = rep_index;
2388 spin_lock_bh(&efx->filter_lock);
2390 /* Search concurrently for
2391 * (1) a filter to be replaced (rep_index): any filter
2392 * with the same match values, up to the current
2393 * search depth for this type, and
2394 * (2) the insertion point (ins_index): (1) or any
2395 * free slot before it or up to the maximum search
2396 * depth for this priority
2397 * We fail if we cannot find (2).
2399 * We can stop once either
2400 * (a) we find (1), in which case we have definitely
2401 * found (2) as well; or
2402 * (b) we have searched exhaustively for (1), and have
2403 * either found (2) or searched exhaustively for it
2405 u32 key = efx_farch_filter_build(&filter, &spec);
2406 unsigned int hash = efx_farch_filter_hash(key);
2407 unsigned int incr = efx_farch_filter_increment(key);
2408 unsigned int max_rep_depth = table->search_limit[spec.type];
2409 unsigned int max_ins_depth =
2410 spec.priority <= EFX_FILTER_PRI_HINT ?
2411 EFX_FARCH_FILTER_CTL_SRCH_HINT_MAX :
2412 EFX_FARCH_FILTER_CTL_SRCH_MAX;
2413 unsigned int i = hash & (table->size - 1);
2418 spin_lock_bh(&efx->filter_lock);
2421 if (!test_bit(i, table->used_bitmap)) {
2424 } else if (efx_farch_filter_equal(&spec,
2433 if (depth >= max_rep_depth &&
2434 (ins_index >= 0 || depth >= max_ins_depth)) {
2436 if (ins_index < 0) {
2444 i = (i + incr) & (table->size - 1);
2449 /* If we found a filter to be replaced, check whether we
2452 if (rep_index >= 0) {
2453 struct efx_farch_filter_spec *saved_spec =
2454 &table->spec[rep_index];
2456 if (spec.priority == saved_spec->priority && !replace_equal) {
2460 if (spec.priority < saved_spec->priority) {
2464 if (saved_spec->priority == EFX_FILTER_PRI_AUTO ||
2465 saved_spec->flags & EFX_FILTER_FLAG_RX_OVER_AUTO)
2466 spec.flags |= EFX_FILTER_FLAG_RX_OVER_AUTO;
2469 /* Insert the filter */
2470 if (ins_index != rep_index) {
2471 __set_bit(ins_index, table->used_bitmap);
2474 table->spec[ins_index] = spec;
2476 if (table->id == EFX_FARCH_FILTER_TABLE_RX_DEF) {
2477 efx_farch_filter_push_rx_config(efx);
2479 if (table->search_limit[spec.type] < depth) {
2480 table->search_limit[spec.type] = depth;
2481 if (spec.flags & EFX_FILTER_FLAG_TX)
2482 efx_farch_filter_push_tx_limits(efx);
2484 efx_farch_filter_push_rx_config(efx);
2487 efx_writeo(efx, &filter,
2488 table->offset + table->step * ins_index);
2490 /* If we were able to replace a filter by inserting
2491 * at a lower depth, clear the replaced filter
2493 if (ins_index != rep_index && rep_index >= 0)
2494 efx_farch_filter_table_clear_entry(efx, table,
2498 netif_vdbg(efx, hw, efx->net_dev,
2499 "%s: filter type %d index %d rxq %u set",
2500 __func__, spec.type, ins_index, spec.dmaq_id);
2501 rc = efx_farch_filter_make_id(&spec, ins_index);
2504 spin_unlock_bh(&efx->filter_lock);
2509 efx_farch_filter_table_clear_entry(struct efx_nic *efx,
2510 struct efx_farch_filter_table *table,
2511 unsigned int filter_idx)
2513 static efx_oword_t filter;
2515 EFX_WARN_ON_PARANOID(!test_bit(filter_idx, table->used_bitmap));
2516 BUG_ON(table->offset == 0); /* can't clear MAC default filters */
2518 __clear_bit(filter_idx, table->used_bitmap);
2520 memset(&table->spec[filter_idx], 0, sizeof(table->spec[0]));
2522 efx_writeo(efx, &filter, table->offset + table->step * filter_idx);
2524 /* If this filter required a greater search depth than
2525 * any other, the search limit for its type can now be
2526 * decreased. However, it is hard to determine that
2527 * unless the table has become completely empty - in
2528 * which case, all its search limits can be set to 0.
2530 if (unlikely(table->used == 0)) {
2531 memset(table->search_limit, 0, sizeof(table->search_limit));
2532 if (table->id == EFX_FARCH_FILTER_TABLE_TX_MAC)
2533 efx_farch_filter_push_tx_limits(efx);
2535 efx_farch_filter_push_rx_config(efx);
2539 static int efx_farch_filter_remove(struct efx_nic *efx,
2540 struct efx_farch_filter_table *table,
2541 unsigned int filter_idx,
2542 enum efx_filter_priority priority)
2544 struct efx_farch_filter_spec *spec = &table->spec[filter_idx];
2546 if (!test_bit(filter_idx, table->used_bitmap) ||
2547 spec->priority != priority)
2550 if (spec->flags & EFX_FILTER_FLAG_RX_OVER_AUTO) {
2551 efx_farch_filter_init_rx_auto(efx, spec);
2552 efx_farch_filter_push_rx_config(efx);
2554 efx_farch_filter_table_clear_entry(efx, table, filter_idx);
2560 int efx_farch_filter_remove_safe(struct efx_nic *efx,
2561 enum efx_filter_priority priority,
2564 struct efx_farch_filter_state *state = efx->filter_state;
2565 enum efx_farch_filter_table_id table_id;
2566 struct efx_farch_filter_table *table;
2567 unsigned int filter_idx;
2568 struct efx_farch_filter_spec *spec;
2571 table_id = efx_farch_filter_id_table_id(filter_id);
2572 if ((unsigned int)table_id >= EFX_FARCH_FILTER_TABLE_COUNT)
2574 table = &state->table[table_id];
2576 filter_idx = efx_farch_filter_id_index(filter_id);
2577 if (filter_idx >= table->size)
2579 spec = &table->spec[filter_idx];
2581 spin_lock_bh(&efx->filter_lock);
2582 rc = efx_farch_filter_remove(efx, table, filter_idx, priority);
2583 spin_unlock_bh(&efx->filter_lock);
2588 int efx_farch_filter_get_safe(struct efx_nic *efx,
2589 enum efx_filter_priority priority,
2590 u32 filter_id, struct efx_filter_spec *spec_buf)
2592 struct efx_farch_filter_state *state = efx->filter_state;
2593 enum efx_farch_filter_table_id table_id;
2594 struct efx_farch_filter_table *table;
2595 struct efx_farch_filter_spec *spec;
2596 unsigned int filter_idx;
2599 table_id = efx_farch_filter_id_table_id(filter_id);
2600 if ((unsigned int)table_id >= EFX_FARCH_FILTER_TABLE_COUNT)
2602 table = &state->table[table_id];
2604 filter_idx = efx_farch_filter_id_index(filter_id);
2605 if (filter_idx >= table->size)
2607 spec = &table->spec[filter_idx];
2609 spin_lock_bh(&efx->filter_lock);
2611 if (test_bit(filter_idx, table->used_bitmap) &&
2612 spec->priority == priority) {
2613 efx_farch_filter_to_gen_spec(spec_buf, spec);
2619 spin_unlock_bh(&efx->filter_lock);
2625 efx_farch_filter_table_clear(struct efx_nic *efx,
2626 enum efx_farch_filter_table_id table_id,
2627 enum efx_filter_priority priority)
2629 struct efx_farch_filter_state *state = efx->filter_state;
2630 struct efx_farch_filter_table *table = &state->table[table_id];
2631 unsigned int filter_idx;
2633 spin_lock_bh(&efx->filter_lock);
2634 for (filter_idx = 0; filter_idx < table->size; ++filter_idx) {
2635 if (table->spec[filter_idx].priority != EFX_FILTER_PRI_AUTO)
2636 efx_farch_filter_remove(efx, table,
2637 filter_idx, priority);
2639 spin_unlock_bh(&efx->filter_lock);
2642 int efx_farch_filter_clear_rx(struct efx_nic *efx,
2643 enum efx_filter_priority priority)
2645 efx_farch_filter_table_clear(efx, EFX_FARCH_FILTER_TABLE_RX_IP,
2647 efx_farch_filter_table_clear(efx, EFX_FARCH_FILTER_TABLE_RX_MAC,
2649 efx_farch_filter_table_clear(efx, EFX_FARCH_FILTER_TABLE_RX_DEF,
2654 u32 efx_farch_filter_count_rx_used(struct efx_nic *efx,
2655 enum efx_filter_priority priority)
2657 struct efx_farch_filter_state *state = efx->filter_state;
2658 enum efx_farch_filter_table_id table_id;
2659 struct efx_farch_filter_table *table;
2660 unsigned int filter_idx;
2663 spin_lock_bh(&efx->filter_lock);
2665 for (table_id = EFX_FARCH_FILTER_TABLE_RX_IP;
2666 table_id <= EFX_FARCH_FILTER_TABLE_RX_DEF;
2668 table = &state->table[table_id];
2669 for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
2670 if (test_bit(filter_idx, table->used_bitmap) &&
2671 table->spec[filter_idx].priority == priority)
2676 spin_unlock_bh(&efx->filter_lock);
2681 s32 efx_farch_filter_get_rx_ids(struct efx_nic *efx,
2682 enum efx_filter_priority priority,
2685 struct efx_farch_filter_state *state = efx->filter_state;
2686 enum efx_farch_filter_table_id table_id;
2687 struct efx_farch_filter_table *table;
2688 unsigned int filter_idx;
2691 spin_lock_bh(&efx->filter_lock);
2693 for (table_id = EFX_FARCH_FILTER_TABLE_RX_IP;
2694 table_id <= EFX_FARCH_FILTER_TABLE_RX_DEF;
2696 table = &state->table[table_id];
2697 for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
2698 if (test_bit(filter_idx, table->used_bitmap) &&
2699 table->spec[filter_idx].priority == priority) {
2700 if (count == size) {
2704 buf[count++] = efx_farch_filter_make_id(
2705 &table->spec[filter_idx], filter_idx);
2710 spin_unlock_bh(&efx->filter_lock);
2715 /* Restore filter stater after reset */
2716 void efx_farch_filter_table_restore(struct efx_nic *efx)
2718 struct efx_farch_filter_state *state = efx->filter_state;
2719 enum efx_farch_filter_table_id table_id;
2720 struct efx_farch_filter_table *table;
2722 unsigned int filter_idx;
2724 spin_lock_bh(&efx->filter_lock);
2726 for (table_id = 0; table_id < EFX_FARCH_FILTER_TABLE_COUNT; table_id++) {
2727 table = &state->table[table_id];
2729 /* Check whether this is a regular register table */
2730 if (table->step == 0)
2733 for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
2734 if (!test_bit(filter_idx, table->used_bitmap))
2736 efx_farch_filter_build(&filter, &table->spec[filter_idx]);
2737 efx_writeo(efx, &filter,
2738 table->offset + table->step * filter_idx);
2742 efx_farch_filter_push_rx_config(efx);
2743 efx_farch_filter_push_tx_limits(efx);
2745 spin_unlock_bh(&efx->filter_lock);
2748 void efx_farch_filter_table_remove(struct efx_nic *efx)
2750 struct efx_farch_filter_state *state = efx->filter_state;
2751 enum efx_farch_filter_table_id table_id;
2753 for (table_id = 0; table_id < EFX_FARCH_FILTER_TABLE_COUNT; table_id++) {
2754 kfree(state->table[table_id].used_bitmap);
2755 vfree(state->table[table_id].spec);
2760 int efx_farch_filter_table_probe(struct efx_nic *efx)
2762 struct efx_farch_filter_state *state;
2763 struct efx_farch_filter_table *table;
2766 state = kzalloc(sizeof(struct efx_farch_filter_state), GFP_KERNEL);
2769 efx->filter_state = state;
2771 if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
2772 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_IP];
2773 table->id = EFX_FARCH_FILTER_TABLE_RX_IP;
2774 table->offset = FR_BZ_RX_FILTER_TBL0;
2775 table->size = FR_BZ_RX_FILTER_TBL0_ROWS;
2776 table->step = FR_BZ_RX_FILTER_TBL0_STEP;
2779 if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
2780 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_MAC];
2781 table->id = EFX_FARCH_FILTER_TABLE_RX_MAC;
2782 table->offset = FR_CZ_RX_MAC_FILTER_TBL0;
2783 table->size = FR_CZ_RX_MAC_FILTER_TBL0_ROWS;
2784 table->step = FR_CZ_RX_MAC_FILTER_TBL0_STEP;
2786 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_DEF];
2787 table->id = EFX_FARCH_FILTER_TABLE_RX_DEF;
2788 table->size = EFX_FARCH_FILTER_SIZE_RX_DEF;
2790 table = &state->table[EFX_FARCH_FILTER_TABLE_TX_MAC];
2791 table->id = EFX_FARCH_FILTER_TABLE_TX_MAC;
2792 table->offset = FR_CZ_TX_MAC_FILTER_TBL0;
2793 table->size = FR_CZ_TX_MAC_FILTER_TBL0_ROWS;
2794 table->step = FR_CZ_TX_MAC_FILTER_TBL0_STEP;
2797 for (table_id = 0; table_id < EFX_FARCH_FILTER_TABLE_COUNT; table_id++) {
2798 table = &state->table[table_id];
2799 if (table->size == 0)
2801 table->used_bitmap = kcalloc(BITS_TO_LONGS(table->size),
2802 sizeof(unsigned long),
2804 if (!table->used_bitmap)
2806 table->spec = vzalloc(table->size * sizeof(*table->spec));
2811 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_DEF];
2813 /* RX default filters must always exist */
2814 struct efx_farch_filter_spec *spec;
2817 for (i = 0; i < EFX_FARCH_FILTER_SIZE_RX_DEF; i++) {
2818 spec = &table->spec[i];
2819 spec->type = EFX_FARCH_FILTER_UC_DEF + i;
2820 efx_farch_filter_init_rx_auto(efx, spec);
2821 __set_bit(i, table->used_bitmap);
2825 efx_farch_filter_push_rx_config(efx);
2830 efx_farch_filter_table_remove(efx);
2834 /* Update scatter enable flags for filters pointing to our own RX queues */
2835 void efx_farch_filter_update_rx_scatter(struct efx_nic *efx)
2837 struct efx_farch_filter_state *state = efx->filter_state;
2838 enum efx_farch_filter_table_id table_id;
2839 struct efx_farch_filter_table *table;
2841 unsigned int filter_idx;
2843 spin_lock_bh(&efx->filter_lock);
2845 for (table_id = EFX_FARCH_FILTER_TABLE_RX_IP;
2846 table_id <= EFX_FARCH_FILTER_TABLE_RX_DEF;
2848 table = &state->table[table_id];
2850 for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
2851 if (!test_bit(filter_idx, table->used_bitmap) ||
2852 table->spec[filter_idx].dmaq_id >=
2856 if (efx->rx_scatter)
2857 table->spec[filter_idx].flags |=
2858 EFX_FILTER_FLAG_RX_SCATTER;
2860 table->spec[filter_idx].flags &=
2861 ~EFX_FILTER_FLAG_RX_SCATTER;
2863 if (table_id == EFX_FARCH_FILTER_TABLE_RX_DEF)
2864 /* Pushed by efx_farch_filter_push_rx_config() */
2867 efx_farch_filter_build(&filter, &table->spec[filter_idx]);
2868 efx_writeo(efx, &filter,
2869 table->offset + table->step * filter_idx);
2873 efx_farch_filter_push_rx_config(efx);
2875 spin_unlock_bh(&efx->filter_lock);
2878 #ifdef CONFIG_RFS_ACCEL
2880 s32 efx_farch_filter_rfs_insert(struct efx_nic *efx,
2881 struct efx_filter_spec *gen_spec)
2883 return efx_farch_filter_insert(efx, gen_spec, true);
2886 bool efx_farch_filter_rfs_expire_one(struct efx_nic *efx, u32 flow_id,
2889 struct efx_farch_filter_state *state = efx->filter_state;
2890 struct efx_farch_filter_table *table =
2891 &state->table[EFX_FARCH_FILTER_TABLE_RX_IP];
2893 if (test_bit(index, table->used_bitmap) &&
2894 table->spec[index].priority == EFX_FILTER_PRI_HINT &&
2895 rps_may_expire_flow(efx->net_dev, table->spec[index].dmaq_id,
2897 efx_farch_filter_table_clear_entry(efx, table, index);
2904 #endif /* CONFIG_RFS_ACCEL */
2906 void efx_farch_filter_sync_rx_mode(struct efx_nic *efx)
2908 struct net_device *net_dev = efx->net_dev;
2909 struct netdev_hw_addr *ha;
2910 union efx_multicast_hash *mc_hash = &efx->multicast_hash;
2914 netif_addr_lock_bh(net_dev);
2916 efx->unicast_filter = !(net_dev->flags & IFF_PROMISC);
2918 /* Build multicast hash table */
2919 if (net_dev->flags & (IFF_PROMISC | IFF_ALLMULTI)) {
2920 memset(mc_hash, 0xff, sizeof(*mc_hash));
2922 memset(mc_hash, 0x00, sizeof(*mc_hash));
2923 netdev_for_each_mc_addr(ha, net_dev) {
2924 crc = ether_crc_le(ETH_ALEN, ha->addr);
2925 bit = crc & (EFX_MCAST_HASH_ENTRIES - 1);
2926 __set_bit_le(bit, mc_hash);
2929 /* Broadcast packets go through the multicast hash filter.
2930 * ether_crc_le() of the broadcast address is 0xbe2612ff
2931 * so we always add bit 0xff to the mask.
2933 __set_bit_le(0xff, mc_hash);
2936 netif_addr_unlock_bh(net_dev);