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 "siena_sriov.h"
26 #include "workarounds.h"
28 /* Falcon-architecture (SFC4000 and SFC9000-family) support */
30 /**************************************************************************
34 **************************************************************************
37 /* This is set to 16 for a good reason. In summary, if larger than
38 * 16, the descriptor cache holds more than a default socket
39 * buffer's worth of packets (for UDP we can only have at most one
40 * socket buffer's worth outstanding). This combined with the fact
41 * that we only get 1 TX event per descriptor cache means the NIC
44 #define TX_DC_ENTRIES 16
45 #define TX_DC_ENTRIES_ORDER 1
47 #define RX_DC_ENTRIES 64
48 #define RX_DC_ENTRIES_ORDER 3
50 /* If EFX_MAX_INT_ERRORS internal errors occur within
51 * EFX_INT_ERROR_EXPIRE seconds, we consider the NIC broken and
54 #define EFX_INT_ERROR_EXPIRE 3600
55 #define EFX_MAX_INT_ERRORS 5
57 /* Depth of RX flush request fifo */
58 #define EFX_RX_FLUSH_COUNT 4
60 /* Driver generated events */
61 #define _EFX_CHANNEL_MAGIC_TEST 0x000101
62 #define _EFX_CHANNEL_MAGIC_FILL 0x000102
63 #define _EFX_CHANNEL_MAGIC_RX_DRAIN 0x000103
64 #define _EFX_CHANNEL_MAGIC_TX_DRAIN 0x000104
66 #define _EFX_CHANNEL_MAGIC(_code, _data) ((_code) << 8 | (_data))
67 #define _EFX_CHANNEL_MAGIC_CODE(_magic) ((_magic) >> 8)
69 #define EFX_CHANNEL_MAGIC_TEST(_channel) \
70 _EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_TEST, (_channel)->channel)
71 #define EFX_CHANNEL_MAGIC_FILL(_rx_queue) \
72 _EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_FILL, \
73 efx_rx_queue_index(_rx_queue))
74 #define EFX_CHANNEL_MAGIC_RX_DRAIN(_rx_queue) \
75 _EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_RX_DRAIN, \
76 efx_rx_queue_index(_rx_queue))
77 #define EFX_CHANNEL_MAGIC_TX_DRAIN(_tx_queue) \
78 _EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_TX_DRAIN, \
81 static void efx_farch_magic_event(struct efx_channel *channel, u32 magic);
83 /**************************************************************************
87 **************************************************************************/
89 static inline void efx_write_buf_tbl(struct efx_nic *efx, efx_qword_t *value,
92 efx_sram_writeq(efx, efx->membase + efx->type->buf_tbl_base,
96 static bool efx_masked_compare_oword(const efx_oword_t *a, const efx_oword_t *b,
97 const efx_oword_t *mask)
99 return ((a->u64[0] ^ b->u64[0]) & mask->u64[0]) ||
100 ((a->u64[1] ^ b->u64[1]) & mask->u64[1]);
103 int efx_farch_test_registers(struct efx_nic *efx,
104 const struct efx_farch_register_test *regs,
107 unsigned address = 0;
109 efx_oword_t mask, imask, original, reg, buf;
111 for (i = 0; i < n_regs; ++i) {
112 address = regs[i].address;
113 mask = imask = regs[i].mask;
114 EFX_INVERT_OWORD(imask);
116 efx_reado(efx, &original, address);
118 /* bit sweep on and off */
119 for (j = 0; j < 128; j++) {
120 if (!EFX_EXTRACT_OWORD32(mask, j, j))
123 /* Test this testable bit can be set in isolation */
124 EFX_AND_OWORD(reg, original, mask);
125 EFX_SET_OWORD32(reg, j, j, 1);
127 efx_writeo(efx, ®, address);
128 efx_reado(efx, &buf, address);
130 if (efx_masked_compare_oword(®, &buf, &mask))
133 /* Test this testable bit can be cleared in isolation */
134 EFX_OR_OWORD(reg, original, mask);
135 EFX_SET_OWORD32(reg, j, j, 0);
137 efx_writeo(efx, ®, address);
138 efx_reado(efx, &buf, address);
140 if (efx_masked_compare_oword(®, &buf, &mask))
144 efx_writeo(efx, &original, address);
150 netif_err(efx, hw, efx->net_dev,
151 "wrote "EFX_OWORD_FMT" read "EFX_OWORD_FMT
152 " at address 0x%x mask "EFX_OWORD_FMT"\n", EFX_OWORD_VAL(reg),
153 EFX_OWORD_VAL(buf), address, EFX_OWORD_VAL(mask));
157 /**************************************************************************
159 * Special buffer handling
160 * Special buffers are used for event queues and the TX and RX
163 *************************************************************************/
166 * Initialise a special buffer
168 * This will define a buffer (previously allocated via
169 * efx_alloc_special_buffer()) in the buffer table, allowing
170 * it to be used for event queues, descriptor rings etc.
173 efx_init_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
175 efx_qword_t buf_desc;
180 EFX_BUG_ON_PARANOID(!buffer->buf.addr);
182 /* Write buffer descriptors to NIC */
183 for (i = 0; i < buffer->entries; i++) {
184 index = buffer->index + i;
185 dma_addr = buffer->buf.dma_addr + (i * EFX_BUF_SIZE);
186 netif_dbg(efx, probe, efx->net_dev,
187 "mapping special buffer %d at %llx\n",
188 index, (unsigned long long)dma_addr);
189 EFX_POPULATE_QWORD_3(buf_desc,
190 FRF_AZ_BUF_ADR_REGION, 0,
191 FRF_AZ_BUF_ADR_FBUF, dma_addr >> 12,
192 FRF_AZ_BUF_OWNER_ID_FBUF, 0);
193 efx_write_buf_tbl(efx, &buf_desc, index);
197 /* Unmaps a buffer and clears the buffer table entries */
199 efx_fini_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
201 efx_oword_t buf_tbl_upd;
202 unsigned int start = buffer->index;
203 unsigned int end = (buffer->index + buffer->entries - 1);
205 if (!buffer->entries)
208 netif_dbg(efx, hw, efx->net_dev, "unmapping special buffers %d-%d\n",
209 buffer->index, buffer->index + buffer->entries - 1);
211 EFX_POPULATE_OWORD_4(buf_tbl_upd,
212 FRF_AZ_BUF_UPD_CMD, 0,
213 FRF_AZ_BUF_CLR_CMD, 1,
214 FRF_AZ_BUF_CLR_END_ID, end,
215 FRF_AZ_BUF_CLR_START_ID, start);
216 efx_writeo(efx, &buf_tbl_upd, FR_AZ_BUF_TBL_UPD);
220 * Allocate a new special buffer
222 * This allocates memory for a new buffer, clears it and allocates a
223 * new buffer ID range. It does not write into the buffer table.
225 * This call will allocate 4KB buffers, since 8KB buffers can't be
226 * used for event queues and descriptor rings.
228 static int efx_alloc_special_buffer(struct efx_nic *efx,
229 struct efx_special_buffer *buffer,
232 #ifdef CONFIG_SFC_SRIOV
233 struct siena_nic_data *nic_data = efx->nic_data;
235 len = ALIGN(len, EFX_BUF_SIZE);
237 if (efx_nic_alloc_buffer(efx, &buffer->buf, len, GFP_KERNEL))
239 buffer->entries = len / EFX_BUF_SIZE;
240 BUG_ON(buffer->buf.dma_addr & (EFX_BUF_SIZE - 1));
242 /* Select new buffer ID */
243 buffer->index = efx->next_buffer_table;
244 efx->next_buffer_table += buffer->entries;
245 #ifdef CONFIG_SFC_SRIOV
246 BUG_ON(efx_siena_sriov_enabled(efx) &&
247 nic_data->vf_buftbl_base < efx->next_buffer_table);
250 netif_dbg(efx, probe, efx->net_dev,
251 "allocating special buffers %d-%d at %llx+%x "
252 "(virt %p phys %llx)\n", buffer->index,
253 buffer->index + buffer->entries - 1,
254 (u64)buffer->buf.dma_addr, len,
255 buffer->buf.addr, (u64)virt_to_phys(buffer->buf.addr));
261 efx_free_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
263 if (!buffer->buf.addr)
266 netif_dbg(efx, hw, efx->net_dev,
267 "deallocating special buffers %d-%d at %llx+%x "
268 "(virt %p phys %llx)\n", buffer->index,
269 buffer->index + buffer->entries - 1,
270 (u64)buffer->buf.dma_addr, buffer->buf.len,
271 buffer->buf.addr, (u64)virt_to_phys(buffer->buf.addr));
273 efx_nic_free_buffer(efx, &buffer->buf);
277 /**************************************************************************
281 **************************************************************************/
283 /* This writes to the TX_DESC_WPTR; write pointer for TX descriptor ring */
284 static inline void efx_farch_notify_tx_desc(struct efx_tx_queue *tx_queue)
289 write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
290 EFX_POPULATE_DWORD_1(reg, FRF_AZ_TX_DESC_WPTR_DWORD, write_ptr);
291 efx_writed_page(tx_queue->efx, ®,
292 FR_AZ_TX_DESC_UPD_DWORD_P0, tx_queue->queue);
295 /* Write pointer and first descriptor for TX descriptor ring */
296 static inline void efx_farch_push_tx_desc(struct efx_tx_queue *tx_queue,
297 const efx_qword_t *txd)
302 BUILD_BUG_ON(FRF_AZ_TX_DESC_LBN != 0);
303 BUILD_BUG_ON(FR_AA_TX_DESC_UPD_KER != FR_BZ_TX_DESC_UPD_P0);
305 write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
306 EFX_POPULATE_OWORD_2(reg, FRF_AZ_TX_DESC_PUSH_CMD, true,
307 FRF_AZ_TX_DESC_WPTR, write_ptr);
309 efx_writeo_page(tx_queue->efx, ®,
310 FR_BZ_TX_DESC_UPD_P0, tx_queue->queue);
314 /* For each entry inserted into the software descriptor ring, create a
315 * descriptor in the hardware TX descriptor ring (in host memory), and
318 void efx_farch_tx_write(struct efx_tx_queue *tx_queue)
320 struct efx_tx_buffer *buffer;
323 unsigned old_write_count = tx_queue->write_count;
325 tx_queue->xmit_more_available = false;
326 if (unlikely(tx_queue->write_count == tx_queue->insert_count))
330 write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
331 buffer = &tx_queue->buffer[write_ptr];
332 txd = efx_tx_desc(tx_queue, write_ptr);
333 ++tx_queue->write_count;
335 EFX_BUG_ON_PARANOID(buffer->flags & EFX_TX_BUF_OPTION);
337 /* Create TX descriptor ring entry */
338 BUILD_BUG_ON(EFX_TX_BUF_CONT != 1);
339 EFX_POPULATE_QWORD_4(*txd,
341 buffer->flags & EFX_TX_BUF_CONT,
342 FSF_AZ_TX_KER_BYTE_COUNT, buffer->len,
343 FSF_AZ_TX_KER_BUF_REGION, 0,
344 FSF_AZ_TX_KER_BUF_ADDR, buffer->dma_addr);
345 } while (tx_queue->write_count != tx_queue->insert_count);
347 wmb(); /* Ensure descriptors are written before they are fetched */
349 if (efx_nic_may_push_tx_desc(tx_queue, old_write_count)) {
350 txd = efx_tx_desc(tx_queue,
351 old_write_count & tx_queue->ptr_mask);
352 efx_farch_push_tx_desc(tx_queue, txd);
355 efx_farch_notify_tx_desc(tx_queue);
359 /* Allocate hardware resources for a TX queue */
360 int efx_farch_tx_probe(struct efx_tx_queue *tx_queue)
362 struct efx_nic *efx = tx_queue->efx;
365 entries = tx_queue->ptr_mask + 1;
366 return efx_alloc_special_buffer(efx, &tx_queue->txd,
367 entries * sizeof(efx_qword_t));
370 void efx_farch_tx_init(struct efx_tx_queue *tx_queue)
372 struct efx_nic *efx = tx_queue->efx;
375 /* Pin TX descriptor ring */
376 efx_init_special_buffer(efx, &tx_queue->txd);
378 /* Push TX descriptor ring to card */
379 EFX_POPULATE_OWORD_10(reg,
380 FRF_AZ_TX_DESCQ_EN, 1,
381 FRF_AZ_TX_ISCSI_DDIG_EN, 0,
382 FRF_AZ_TX_ISCSI_HDIG_EN, 0,
383 FRF_AZ_TX_DESCQ_BUF_BASE_ID, tx_queue->txd.index,
384 FRF_AZ_TX_DESCQ_EVQ_ID,
385 tx_queue->channel->channel,
386 FRF_AZ_TX_DESCQ_OWNER_ID, 0,
387 FRF_AZ_TX_DESCQ_LABEL, tx_queue->queue,
388 FRF_AZ_TX_DESCQ_SIZE,
389 __ffs(tx_queue->txd.entries),
390 FRF_AZ_TX_DESCQ_TYPE, 0,
391 FRF_BZ_TX_NON_IP_DROP_DIS, 1);
393 if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
394 int csum = tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD;
395 EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_IP_CHKSM_DIS, !csum);
396 EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_TCP_CHKSM_DIS,
400 efx_writeo_table(efx, ®, efx->type->txd_ptr_tbl_base,
403 if (efx_nic_rev(efx) < EFX_REV_FALCON_B0) {
404 /* Only 128 bits in this register */
405 BUILD_BUG_ON(EFX_MAX_TX_QUEUES > 128);
407 efx_reado(efx, ®, FR_AA_TX_CHKSM_CFG);
408 if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD)
409 __clear_bit_le(tx_queue->queue, ®);
411 __set_bit_le(tx_queue->queue, ®);
412 efx_writeo(efx, ®, FR_AA_TX_CHKSM_CFG);
415 if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
416 EFX_POPULATE_OWORD_1(reg,
418 (tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
420 FFE_BZ_TX_PACE_RESERVED);
421 efx_writeo_table(efx, ®, FR_BZ_TX_PACE_TBL,
426 static void efx_farch_flush_tx_queue(struct efx_tx_queue *tx_queue)
428 struct efx_nic *efx = tx_queue->efx;
429 efx_oword_t tx_flush_descq;
431 WARN_ON(atomic_read(&tx_queue->flush_outstanding));
432 atomic_set(&tx_queue->flush_outstanding, 1);
434 EFX_POPULATE_OWORD_2(tx_flush_descq,
435 FRF_AZ_TX_FLUSH_DESCQ_CMD, 1,
436 FRF_AZ_TX_FLUSH_DESCQ, tx_queue->queue);
437 efx_writeo(efx, &tx_flush_descq, FR_AZ_TX_FLUSH_DESCQ);
440 void efx_farch_tx_fini(struct efx_tx_queue *tx_queue)
442 struct efx_nic *efx = tx_queue->efx;
443 efx_oword_t tx_desc_ptr;
445 /* Remove TX descriptor ring from card */
446 EFX_ZERO_OWORD(tx_desc_ptr);
447 efx_writeo_table(efx, &tx_desc_ptr, efx->type->txd_ptr_tbl_base,
450 /* Unpin TX descriptor ring */
451 efx_fini_special_buffer(efx, &tx_queue->txd);
454 /* Free buffers backing TX queue */
455 void efx_farch_tx_remove(struct efx_tx_queue *tx_queue)
457 efx_free_special_buffer(tx_queue->efx, &tx_queue->txd);
460 /**************************************************************************
464 **************************************************************************/
466 /* This creates an entry in the RX descriptor queue */
468 efx_farch_build_rx_desc(struct efx_rx_queue *rx_queue, unsigned index)
470 struct efx_rx_buffer *rx_buf;
473 rxd = efx_rx_desc(rx_queue, index);
474 rx_buf = efx_rx_buffer(rx_queue, index);
475 EFX_POPULATE_QWORD_3(*rxd,
476 FSF_AZ_RX_KER_BUF_SIZE,
478 rx_queue->efx->type->rx_buffer_padding,
479 FSF_AZ_RX_KER_BUF_REGION, 0,
480 FSF_AZ_RX_KER_BUF_ADDR, rx_buf->dma_addr);
483 /* This writes to the RX_DESC_WPTR register for the specified receive
486 void efx_farch_rx_write(struct efx_rx_queue *rx_queue)
488 struct efx_nic *efx = rx_queue->efx;
492 while (rx_queue->notified_count != rx_queue->added_count) {
493 efx_farch_build_rx_desc(
495 rx_queue->notified_count & rx_queue->ptr_mask);
496 ++rx_queue->notified_count;
500 write_ptr = rx_queue->added_count & rx_queue->ptr_mask;
501 EFX_POPULATE_DWORD_1(reg, FRF_AZ_RX_DESC_WPTR_DWORD, write_ptr);
502 efx_writed_page(efx, ®, FR_AZ_RX_DESC_UPD_DWORD_P0,
503 efx_rx_queue_index(rx_queue));
506 int efx_farch_rx_probe(struct efx_rx_queue *rx_queue)
508 struct efx_nic *efx = rx_queue->efx;
511 entries = rx_queue->ptr_mask + 1;
512 return efx_alloc_special_buffer(efx, &rx_queue->rxd,
513 entries * sizeof(efx_qword_t));
516 void efx_farch_rx_init(struct efx_rx_queue *rx_queue)
518 efx_oword_t rx_desc_ptr;
519 struct efx_nic *efx = rx_queue->efx;
520 bool is_b0 = efx_nic_rev(efx) >= EFX_REV_FALCON_B0;
521 bool iscsi_digest_en = is_b0;
524 /* For kernel-mode queues in Falcon A1, the JUMBO flag enables
525 * DMA to continue after a PCIe page boundary (and scattering
526 * is not possible). In Falcon B0 and Siena, it enables
529 jumbo_en = !is_b0 || efx->rx_scatter;
531 netif_dbg(efx, hw, efx->net_dev,
532 "RX queue %d ring in special buffers %d-%d\n",
533 efx_rx_queue_index(rx_queue), rx_queue->rxd.index,
534 rx_queue->rxd.index + rx_queue->rxd.entries - 1);
536 rx_queue->scatter_n = 0;
538 /* Pin RX descriptor ring */
539 efx_init_special_buffer(efx, &rx_queue->rxd);
541 /* Push RX descriptor ring to card */
542 EFX_POPULATE_OWORD_10(rx_desc_ptr,
543 FRF_AZ_RX_ISCSI_DDIG_EN, iscsi_digest_en,
544 FRF_AZ_RX_ISCSI_HDIG_EN, iscsi_digest_en,
545 FRF_AZ_RX_DESCQ_BUF_BASE_ID, rx_queue->rxd.index,
546 FRF_AZ_RX_DESCQ_EVQ_ID,
547 efx_rx_queue_channel(rx_queue)->channel,
548 FRF_AZ_RX_DESCQ_OWNER_ID, 0,
549 FRF_AZ_RX_DESCQ_LABEL,
550 efx_rx_queue_index(rx_queue),
551 FRF_AZ_RX_DESCQ_SIZE,
552 __ffs(rx_queue->rxd.entries),
553 FRF_AZ_RX_DESCQ_TYPE, 0 /* kernel queue */ ,
554 FRF_AZ_RX_DESCQ_JUMBO, jumbo_en,
555 FRF_AZ_RX_DESCQ_EN, 1);
556 efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
557 efx_rx_queue_index(rx_queue));
560 static void efx_farch_flush_rx_queue(struct efx_rx_queue *rx_queue)
562 struct efx_nic *efx = rx_queue->efx;
563 efx_oword_t rx_flush_descq;
565 EFX_POPULATE_OWORD_2(rx_flush_descq,
566 FRF_AZ_RX_FLUSH_DESCQ_CMD, 1,
567 FRF_AZ_RX_FLUSH_DESCQ,
568 efx_rx_queue_index(rx_queue));
569 efx_writeo(efx, &rx_flush_descq, FR_AZ_RX_FLUSH_DESCQ);
572 void efx_farch_rx_fini(struct efx_rx_queue *rx_queue)
574 efx_oword_t rx_desc_ptr;
575 struct efx_nic *efx = rx_queue->efx;
577 /* Remove RX descriptor ring from card */
578 EFX_ZERO_OWORD(rx_desc_ptr);
579 efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
580 efx_rx_queue_index(rx_queue));
582 /* Unpin RX descriptor ring */
583 efx_fini_special_buffer(efx, &rx_queue->rxd);
586 /* Free buffers backing RX queue */
587 void efx_farch_rx_remove(struct efx_rx_queue *rx_queue)
589 efx_free_special_buffer(rx_queue->efx, &rx_queue->rxd);
592 /**************************************************************************
596 **************************************************************************/
598 /* efx_farch_flush_queues() must be woken up when all flushes are completed,
599 * or more RX flushes can be kicked off.
601 static bool efx_farch_flush_wake(struct efx_nic *efx)
603 /* Ensure that all updates are visible to efx_farch_flush_queues() */
606 return (atomic_read(&efx->active_queues) == 0 ||
607 (atomic_read(&efx->rxq_flush_outstanding) < EFX_RX_FLUSH_COUNT
608 && atomic_read(&efx->rxq_flush_pending) > 0));
611 static bool efx_check_tx_flush_complete(struct efx_nic *efx)
614 efx_oword_t txd_ptr_tbl;
615 struct efx_channel *channel;
616 struct efx_tx_queue *tx_queue;
618 efx_for_each_channel(channel, efx) {
619 efx_for_each_channel_tx_queue(tx_queue, channel) {
620 efx_reado_table(efx, &txd_ptr_tbl,
621 FR_BZ_TX_DESC_PTR_TBL, tx_queue->queue);
622 if (EFX_OWORD_FIELD(txd_ptr_tbl,
623 FRF_AZ_TX_DESCQ_FLUSH) ||
624 EFX_OWORD_FIELD(txd_ptr_tbl,
625 FRF_AZ_TX_DESCQ_EN)) {
626 netif_dbg(efx, hw, efx->net_dev,
627 "flush did not complete on TXQ %d\n",
630 } else if (atomic_cmpxchg(&tx_queue->flush_outstanding,
632 /* The flush is complete, but we didn't
633 * receive a flush completion event
635 netif_dbg(efx, hw, efx->net_dev,
636 "flush complete on TXQ %d, so drain "
637 "the queue\n", tx_queue->queue);
638 /* Don't need to increment active_queues as it
639 * has already been incremented for the queues
640 * which did not drain
642 efx_farch_magic_event(channel,
643 EFX_CHANNEL_MAGIC_TX_DRAIN(
652 /* Flush all the transmit queues, and continue flushing receive queues until
653 * they're all flushed. Wait for the DRAIN events to be received so that there
654 * are no more RX and TX events left on any channel. */
655 static int efx_farch_do_flush(struct efx_nic *efx)
657 unsigned timeout = msecs_to_jiffies(5000); /* 5s for all flushes and drains */
658 struct efx_channel *channel;
659 struct efx_rx_queue *rx_queue;
660 struct efx_tx_queue *tx_queue;
663 efx_for_each_channel(channel, efx) {
664 efx_for_each_channel_tx_queue(tx_queue, channel) {
665 efx_farch_flush_tx_queue(tx_queue);
667 efx_for_each_channel_rx_queue(rx_queue, channel) {
668 rx_queue->flush_pending = true;
669 atomic_inc(&efx->rxq_flush_pending);
673 while (timeout && atomic_read(&efx->active_queues) > 0) {
674 /* If SRIOV is enabled, then offload receive queue flushing to
675 * the firmware (though we will still have to poll for
676 * completion). If that fails, fall back to the old scheme.
678 if (efx_siena_sriov_enabled(efx)) {
679 rc = efx_mcdi_flush_rxqs(efx);
684 /* The hardware supports four concurrent rx flushes, each of
685 * which may need to be retried if there is an outstanding
688 efx_for_each_channel(channel, efx) {
689 efx_for_each_channel_rx_queue(rx_queue, channel) {
690 if (atomic_read(&efx->rxq_flush_outstanding) >=
694 if (rx_queue->flush_pending) {
695 rx_queue->flush_pending = false;
696 atomic_dec(&efx->rxq_flush_pending);
697 atomic_inc(&efx->rxq_flush_outstanding);
698 efx_farch_flush_rx_queue(rx_queue);
704 timeout = wait_event_timeout(efx->flush_wq,
705 efx_farch_flush_wake(efx),
709 if (atomic_read(&efx->active_queues) &&
710 !efx_check_tx_flush_complete(efx)) {
711 netif_err(efx, hw, efx->net_dev, "failed to flush %d queues "
712 "(rx %d+%d)\n", atomic_read(&efx->active_queues),
713 atomic_read(&efx->rxq_flush_outstanding),
714 atomic_read(&efx->rxq_flush_pending));
717 atomic_set(&efx->active_queues, 0);
718 atomic_set(&efx->rxq_flush_pending, 0);
719 atomic_set(&efx->rxq_flush_outstanding, 0);
725 int efx_farch_fini_dmaq(struct efx_nic *efx)
727 struct efx_channel *channel;
728 struct efx_tx_queue *tx_queue;
729 struct efx_rx_queue *rx_queue;
732 /* Do not attempt to write to the NIC during EEH recovery */
733 if (efx->state != STATE_RECOVERY) {
734 /* Only perform flush if DMA is enabled */
735 if (efx->pci_dev->is_busmaster) {
736 efx->type->prepare_flush(efx);
737 rc = efx_farch_do_flush(efx);
738 efx->type->finish_flush(efx);
741 efx_for_each_channel(channel, efx) {
742 efx_for_each_channel_rx_queue(rx_queue, channel)
743 efx_farch_rx_fini(rx_queue);
744 efx_for_each_channel_tx_queue(tx_queue, channel)
745 efx_farch_tx_fini(tx_queue);
752 /* Reset queue and flush accounting after FLR
754 * One possible cause of FLR recovery is that DMA may be failing (eg. if bus
755 * mastering was disabled), in which case we don't receive (RXQ) flush
756 * completion events. This means that efx->rxq_flush_outstanding remained at 4
757 * after the FLR; also, efx->active_queues was non-zero (as no flush completion
758 * events were received, and we didn't go through efx_check_tx_flush_complete())
759 * If we don't fix this up, on the next call to efx_realloc_channels() we won't
760 * flush any RX queues because efx->rxq_flush_outstanding is at the limit of 4
761 * for batched flush requests; and the efx->active_queues gets messed up because
762 * we keep incrementing for the newly initialised queues, but it never went to
763 * zero previously. Then we get a timeout every time we try to restart the
764 * queues, as it doesn't go back to zero when we should be flushing the queues.
766 void efx_farch_finish_flr(struct efx_nic *efx)
768 atomic_set(&efx->rxq_flush_pending, 0);
769 atomic_set(&efx->rxq_flush_outstanding, 0);
770 atomic_set(&efx->active_queues, 0);
774 /**************************************************************************
776 * Event queue processing
777 * Event queues are processed by per-channel tasklets.
779 **************************************************************************/
781 /* Update a channel's event queue's read pointer (RPTR) register
783 * This writes the EVQ_RPTR_REG register for the specified channel's
786 void efx_farch_ev_read_ack(struct efx_channel *channel)
789 struct efx_nic *efx = channel->efx;
791 EFX_POPULATE_DWORD_1(reg, FRF_AZ_EVQ_RPTR,
792 channel->eventq_read_ptr & channel->eventq_mask);
794 /* For Falcon A1, EVQ_RPTR_KER is documented as having a step size
795 * of 4 bytes, but it is really 16 bytes just like later revisions.
797 efx_writed(efx, ®,
798 efx->type->evq_rptr_tbl_base +
799 FR_BZ_EVQ_RPTR_STEP * channel->channel);
802 /* Use HW to insert a SW defined event */
803 void efx_farch_generate_event(struct efx_nic *efx, unsigned int evq,
806 efx_oword_t drv_ev_reg;
808 BUILD_BUG_ON(FRF_AZ_DRV_EV_DATA_LBN != 0 ||
809 FRF_AZ_DRV_EV_DATA_WIDTH != 64);
810 drv_ev_reg.u32[0] = event->u32[0];
811 drv_ev_reg.u32[1] = event->u32[1];
812 drv_ev_reg.u32[2] = 0;
813 drv_ev_reg.u32[3] = 0;
814 EFX_SET_OWORD_FIELD(drv_ev_reg, FRF_AZ_DRV_EV_QID, evq);
815 efx_writeo(efx, &drv_ev_reg, FR_AZ_DRV_EV);
818 static void efx_farch_magic_event(struct efx_channel *channel, u32 magic)
822 EFX_POPULATE_QWORD_2(event, FSF_AZ_EV_CODE,
823 FSE_AZ_EV_CODE_DRV_GEN_EV,
824 FSF_AZ_DRV_GEN_EV_MAGIC, magic);
825 efx_farch_generate_event(channel->efx, channel->channel, &event);
828 /* Handle a transmit completion event
830 * The NIC batches TX completion events; the message we receive is of
831 * the form "complete all TX events up to this index".
834 efx_farch_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
836 unsigned int tx_ev_desc_ptr;
837 unsigned int tx_ev_q_label;
838 struct efx_tx_queue *tx_queue;
839 struct efx_nic *efx = channel->efx;
842 if (unlikely(ACCESS_ONCE(efx->reset_pending)))
845 if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
846 /* Transmit completion */
847 tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
848 tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
849 tx_queue = efx_channel_get_tx_queue(
850 channel, tx_ev_q_label % EFX_TXQ_TYPES);
851 tx_packets = ((tx_ev_desc_ptr - tx_queue->read_count) &
853 efx_xmit_done(tx_queue, tx_ev_desc_ptr);
854 } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
855 /* Rewrite the FIFO write pointer */
856 tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
857 tx_queue = efx_channel_get_tx_queue(
858 channel, tx_ev_q_label % EFX_TXQ_TYPES);
860 netif_tx_lock(efx->net_dev);
861 efx_farch_notify_tx_desc(tx_queue);
862 netif_tx_unlock(efx->net_dev);
863 } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR)) {
864 efx_schedule_reset(efx, RESET_TYPE_DMA_ERROR);
866 netif_err(efx, tx_err, efx->net_dev,
867 "channel %d unexpected TX event "
868 EFX_QWORD_FMT"\n", channel->channel,
869 EFX_QWORD_VAL(*event));
875 /* Detect errors included in the rx_evt_pkt_ok bit. */
876 static u16 efx_farch_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
877 const efx_qword_t *event)
879 struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
880 struct efx_nic *efx = rx_queue->efx;
881 bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
882 bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
883 bool rx_ev_frm_trunc, rx_ev_drib_nib, rx_ev_tobe_disc;
884 bool rx_ev_other_err, rx_ev_pause_frm;
885 bool rx_ev_hdr_type, rx_ev_mcast_pkt;
886 unsigned rx_ev_pkt_type;
888 rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
889 rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
890 rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
891 rx_ev_pkt_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_TYPE);
892 rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
893 FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
894 rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
895 FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
896 rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
897 FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
898 rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
899 rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
900 rx_ev_drib_nib = ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) ?
901 0 : EFX_QWORD_FIELD(*event, FSF_AA_RX_EV_DRIB_NIB));
902 rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
904 /* Every error apart from tobe_disc and pause_frm */
905 rx_ev_other_err = (rx_ev_drib_nib | rx_ev_tcp_udp_chksum_err |
906 rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
907 rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
909 /* Count errors that are not in MAC stats. Ignore expected
910 * checksum errors during self-test. */
912 ++channel->n_rx_frm_trunc;
913 else if (rx_ev_tobe_disc)
914 ++channel->n_rx_tobe_disc;
915 else if (!efx->loopback_selftest) {
916 if (rx_ev_ip_hdr_chksum_err)
917 ++channel->n_rx_ip_hdr_chksum_err;
918 else if (rx_ev_tcp_udp_chksum_err)
919 ++channel->n_rx_tcp_udp_chksum_err;
922 /* TOBE_DISC is expected on unicast mismatches; don't print out an
923 * error message. FRM_TRUNC indicates RXDP dropped the packet due
924 * to a FIFO overflow.
927 if (rx_ev_other_err && net_ratelimit()) {
928 netif_dbg(efx, rx_err, efx->net_dev,
929 " RX queue %d unexpected RX event "
930 EFX_QWORD_FMT "%s%s%s%s%s%s%s%s\n",
931 efx_rx_queue_index(rx_queue), EFX_QWORD_VAL(*event),
932 rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
933 rx_ev_ip_hdr_chksum_err ?
934 " [IP_HDR_CHKSUM_ERR]" : "",
935 rx_ev_tcp_udp_chksum_err ?
936 " [TCP_UDP_CHKSUM_ERR]" : "",
937 rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
938 rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
939 rx_ev_drib_nib ? " [DRIB_NIB]" : "",
940 rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
941 rx_ev_pause_frm ? " [PAUSE]" : "");
945 /* The frame must be discarded if any of these are true. */
946 return (rx_ev_eth_crc_err | rx_ev_frm_trunc | rx_ev_drib_nib |
947 rx_ev_tobe_disc | rx_ev_pause_frm) ?
948 EFX_RX_PKT_DISCARD : 0;
951 /* Handle receive events that are not in-order. Return true if this
952 * can be handled as a partial packet discard, false if it's more
956 efx_farch_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
958 struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
959 struct efx_nic *efx = rx_queue->efx;
960 unsigned expected, dropped;
962 if (rx_queue->scatter_n &&
963 index == ((rx_queue->removed_count + rx_queue->scatter_n - 1) &
964 rx_queue->ptr_mask)) {
965 ++channel->n_rx_nodesc_trunc;
969 expected = rx_queue->removed_count & rx_queue->ptr_mask;
970 dropped = (index - expected) & rx_queue->ptr_mask;
971 netif_info(efx, rx_err, efx->net_dev,
972 "dropped %d events (index=%d expected=%d)\n",
973 dropped, index, expected);
975 efx_schedule_reset(efx, EFX_WORKAROUND_5676(efx) ?
976 RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
980 /* Handle a packet received event
982 * The NIC gives a "discard" flag if it's a unicast packet with the
983 * wrong destination address
984 * Also "is multicast" and "matches multicast filter" flags can be used to
985 * discard non-matching multicast packets.
988 efx_farch_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
990 unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
991 unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
992 unsigned expected_ptr;
993 bool rx_ev_pkt_ok, rx_ev_sop, rx_ev_cont;
995 struct efx_rx_queue *rx_queue;
996 struct efx_nic *efx = channel->efx;
998 if (unlikely(ACCESS_ONCE(efx->reset_pending)))
1001 rx_ev_cont = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT);
1002 rx_ev_sop = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP);
1003 WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
1006 rx_queue = efx_channel_get_rx_queue(channel);
1008 rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
1009 expected_ptr = ((rx_queue->removed_count + rx_queue->scatter_n) &
1010 rx_queue->ptr_mask);
1012 /* Check for partial drops and other errors */
1013 if (unlikely(rx_ev_desc_ptr != expected_ptr) ||
1014 unlikely(rx_ev_sop != (rx_queue->scatter_n == 0))) {
1015 if (rx_ev_desc_ptr != expected_ptr &&
1016 !efx_farch_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr))
1019 /* Discard all pending fragments */
1020 if (rx_queue->scatter_n) {
1023 rx_queue->removed_count & rx_queue->ptr_mask,
1024 rx_queue->scatter_n, 0, EFX_RX_PKT_DISCARD);
1025 rx_queue->removed_count += rx_queue->scatter_n;
1026 rx_queue->scatter_n = 0;
1029 /* Return if there is no new fragment */
1030 if (rx_ev_desc_ptr != expected_ptr)
1033 /* Discard new fragment if not SOP */
1037 rx_queue->removed_count & rx_queue->ptr_mask,
1038 1, 0, EFX_RX_PKT_DISCARD);
1039 ++rx_queue->removed_count;
1044 ++rx_queue->scatter_n;
1048 rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
1049 rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
1050 rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
1052 if (likely(rx_ev_pkt_ok)) {
1053 /* If packet is marked as OK then we can rely on the
1054 * hardware checksum and classification.
1057 switch (rx_ev_hdr_type) {
1058 case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP:
1059 flags |= EFX_RX_PKT_TCP;
1061 case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP:
1062 flags |= EFX_RX_PKT_CSUMMED;
1064 case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_OTHER:
1065 case FSE_AZ_RX_EV_HDR_TYPE_OTHER:
1069 flags = efx_farch_handle_rx_not_ok(rx_queue, event);
1072 /* Detect multicast packets that didn't match the filter */
1073 rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
1074 if (rx_ev_mcast_pkt) {
1075 unsigned int rx_ev_mcast_hash_match =
1076 EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
1078 if (unlikely(!rx_ev_mcast_hash_match)) {
1079 ++channel->n_rx_mcast_mismatch;
1080 flags |= EFX_RX_PKT_DISCARD;
1084 channel->irq_mod_score += 2;
1086 /* Handle received packet */
1087 efx_rx_packet(rx_queue,
1088 rx_queue->removed_count & rx_queue->ptr_mask,
1089 rx_queue->scatter_n, rx_ev_byte_cnt, flags);
1090 rx_queue->removed_count += rx_queue->scatter_n;
1091 rx_queue->scatter_n = 0;
1094 /* If this flush done event corresponds to a &struct efx_tx_queue, then
1095 * send an %EFX_CHANNEL_MAGIC_TX_DRAIN event to drain the event queue
1096 * of all transmit completions.
1099 efx_farch_handle_tx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1101 struct efx_tx_queue *tx_queue;
1104 qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1105 if (qid < EFX_TXQ_TYPES * efx->n_tx_channels) {
1106 tx_queue = efx_get_tx_queue(efx, qid / EFX_TXQ_TYPES,
1107 qid % EFX_TXQ_TYPES);
1108 if (atomic_cmpxchg(&tx_queue->flush_outstanding, 1, 0)) {
1109 efx_farch_magic_event(tx_queue->channel,
1110 EFX_CHANNEL_MAGIC_TX_DRAIN(tx_queue));
1115 /* If this flush done event corresponds to a &struct efx_rx_queue: If the flush
1116 * was successful then send an %EFX_CHANNEL_MAGIC_RX_DRAIN, otherwise add
1117 * the RX queue back to the mask of RX queues in need of flushing.
1120 efx_farch_handle_rx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1122 struct efx_channel *channel;
1123 struct efx_rx_queue *rx_queue;
1127 qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
1128 failed = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
1129 if (qid >= efx->n_channels)
1131 channel = efx_get_channel(efx, qid);
1132 if (!efx_channel_has_rx_queue(channel))
1134 rx_queue = efx_channel_get_rx_queue(channel);
1137 netif_info(efx, hw, efx->net_dev,
1138 "RXQ %d flush retry\n", qid);
1139 rx_queue->flush_pending = true;
1140 atomic_inc(&efx->rxq_flush_pending);
1142 efx_farch_magic_event(efx_rx_queue_channel(rx_queue),
1143 EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue));
1145 atomic_dec(&efx->rxq_flush_outstanding);
1146 if (efx_farch_flush_wake(efx))
1147 wake_up(&efx->flush_wq);
1151 efx_farch_handle_drain_event(struct efx_channel *channel)
1153 struct efx_nic *efx = channel->efx;
1155 WARN_ON(atomic_read(&efx->active_queues) == 0);
1156 atomic_dec(&efx->active_queues);
1157 if (efx_farch_flush_wake(efx))
1158 wake_up(&efx->flush_wq);
1161 static void efx_farch_handle_generated_event(struct efx_channel *channel,
1164 struct efx_nic *efx = channel->efx;
1165 struct efx_rx_queue *rx_queue =
1166 efx_channel_has_rx_queue(channel) ?
1167 efx_channel_get_rx_queue(channel) : NULL;
1168 unsigned magic, code;
1170 magic = EFX_QWORD_FIELD(*event, FSF_AZ_DRV_GEN_EV_MAGIC);
1171 code = _EFX_CHANNEL_MAGIC_CODE(magic);
1173 if (magic == EFX_CHANNEL_MAGIC_TEST(channel)) {
1174 channel->event_test_cpu = raw_smp_processor_id();
1175 } else if (rx_queue && magic == EFX_CHANNEL_MAGIC_FILL(rx_queue)) {
1176 /* The queue must be empty, so we won't receive any rx
1177 * events, so efx_process_channel() won't refill the
1178 * queue. Refill it here */
1179 efx_fast_push_rx_descriptors(rx_queue, true);
1180 } else if (rx_queue && magic == EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue)) {
1181 efx_farch_handle_drain_event(channel);
1182 } else if (code == _EFX_CHANNEL_MAGIC_TX_DRAIN) {
1183 efx_farch_handle_drain_event(channel);
1185 netif_dbg(efx, hw, efx->net_dev, "channel %d received "
1186 "generated event "EFX_QWORD_FMT"\n",
1187 channel->channel, EFX_QWORD_VAL(*event));
1192 efx_farch_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
1194 struct efx_nic *efx = channel->efx;
1195 unsigned int ev_sub_code;
1196 unsigned int ev_sub_data;
1198 ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
1199 ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1201 switch (ev_sub_code) {
1202 case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
1203 netif_vdbg(efx, hw, efx->net_dev, "channel %d TXQ %d flushed\n",
1204 channel->channel, ev_sub_data);
1205 efx_farch_handle_tx_flush_done(efx, event);
1206 #ifdef CONFIG_SFC_SRIOV
1207 efx_siena_sriov_tx_flush_done(efx, event);
1210 case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
1211 netif_vdbg(efx, hw, efx->net_dev, "channel %d RXQ %d flushed\n",
1212 channel->channel, ev_sub_data);
1213 efx_farch_handle_rx_flush_done(efx, event);
1214 #ifdef CONFIG_SFC_SRIOV
1215 efx_siena_sriov_rx_flush_done(efx, event);
1218 case FSE_AZ_EVQ_INIT_DONE_EV:
1219 netif_dbg(efx, hw, efx->net_dev,
1220 "channel %d EVQ %d initialised\n",
1221 channel->channel, ev_sub_data);
1223 case FSE_AZ_SRM_UPD_DONE_EV:
1224 netif_vdbg(efx, hw, efx->net_dev,
1225 "channel %d SRAM update done\n", channel->channel);
1227 case FSE_AZ_WAKE_UP_EV:
1228 netif_vdbg(efx, hw, efx->net_dev,
1229 "channel %d RXQ %d wakeup event\n",
1230 channel->channel, ev_sub_data);
1232 case FSE_AZ_TIMER_EV:
1233 netif_vdbg(efx, hw, efx->net_dev,
1234 "channel %d RX queue %d timer expired\n",
1235 channel->channel, ev_sub_data);
1237 case FSE_AA_RX_RECOVER_EV:
1238 netif_err(efx, rx_err, efx->net_dev,
1239 "channel %d seen DRIVER RX_RESET event. "
1240 "Resetting.\n", channel->channel);
1241 atomic_inc(&efx->rx_reset);
1242 efx_schedule_reset(efx,
1243 EFX_WORKAROUND_6555(efx) ?
1244 RESET_TYPE_RX_RECOVERY :
1245 RESET_TYPE_DISABLE);
1247 case FSE_BZ_RX_DSC_ERROR_EV:
1248 if (ev_sub_data < EFX_VI_BASE) {
1249 netif_err(efx, rx_err, efx->net_dev,
1250 "RX DMA Q %d reports descriptor fetch error."
1251 " RX Q %d is disabled.\n", ev_sub_data,
1253 efx_schedule_reset(efx, RESET_TYPE_DMA_ERROR);
1255 #ifdef CONFIG_SFC_SRIOV
1257 efx_siena_sriov_desc_fetch_err(efx, ev_sub_data);
1260 case FSE_BZ_TX_DSC_ERROR_EV:
1261 if (ev_sub_data < EFX_VI_BASE) {
1262 netif_err(efx, tx_err, efx->net_dev,
1263 "TX DMA Q %d reports descriptor fetch error."
1264 " TX Q %d is disabled.\n", ev_sub_data,
1266 efx_schedule_reset(efx, RESET_TYPE_DMA_ERROR);
1268 #ifdef CONFIG_SFC_SRIOV
1270 efx_siena_sriov_desc_fetch_err(efx, ev_sub_data);
1274 netif_vdbg(efx, hw, efx->net_dev,
1275 "channel %d unknown driver event code %d "
1276 "data %04x\n", channel->channel, ev_sub_code,
1282 int efx_farch_ev_process(struct efx_channel *channel, int budget)
1284 struct efx_nic *efx = channel->efx;
1285 unsigned int read_ptr;
1286 efx_qword_t event, *p_event;
1294 read_ptr = channel->eventq_read_ptr;
1297 p_event = efx_event(channel, read_ptr);
1300 if (!efx_event_present(&event))
1304 netif_vdbg(channel->efx, intr, channel->efx->net_dev,
1305 "channel %d event is "EFX_QWORD_FMT"\n",
1306 channel->channel, EFX_QWORD_VAL(event));
1308 /* Clear this event by marking it all ones */
1309 EFX_SET_QWORD(*p_event);
1313 ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
1316 case FSE_AZ_EV_CODE_RX_EV:
1317 efx_farch_handle_rx_event(channel, &event);
1318 if (++spent == budget)
1321 case FSE_AZ_EV_CODE_TX_EV:
1322 tx_packets += efx_farch_handle_tx_event(channel,
1324 if (tx_packets > efx->txq_entries) {
1329 case FSE_AZ_EV_CODE_DRV_GEN_EV:
1330 efx_farch_handle_generated_event(channel, &event);
1332 case FSE_AZ_EV_CODE_DRIVER_EV:
1333 efx_farch_handle_driver_event(channel, &event);
1335 #ifdef CONFIG_SFC_SRIOV
1336 case FSE_CZ_EV_CODE_USER_EV:
1337 efx_siena_sriov_event(channel, &event);
1340 case FSE_CZ_EV_CODE_MCDI_EV:
1341 efx_mcdi_process_event(channel, &event);
1343 case FSE_AZ_EV_CODE_GLOBAL_EV:
1344 if (efx->type->handle_global_event &&
1345 efx->type->handle_global_event(channel, &event))
1347 /* else fall through */
1349 netif_err(channel->efx, hw, channel->efx->net_dev,
1350 "channel %d unknown event type %d (data "
1351 EFX_QWORD_FMT ")\n", channel->channel,
1352 ev_code, EFX_QWORD_VAL(event));
1357 channel->eventq_read_ptr = read_ptr;
1361 /* Allocate buffer table entries for event queue */
1362 int efx_farch_ev_probe(struct efx_channel *channel)
1364 struct efx_nic *efx = channel->efx;
1367 entries = channel->eventq_mask + 1;
1368 return efx_alloc_special_buffer(efx, &channel->eventq,
1369 entries * sizeof(efx_qword_t));
1372 int efx_farch_ev_init(struct efx_channel *channel)
1375 struct efx_nic *efx = channel->efx;
1377 netif_dbg(efx, hw, efx->net_dev,
1378 "channel %d event queue in special buffers %d-%d\n",
1379 channel->channel, channel->eventq.index,
1380 channel->eventq.index + channel->eventq.entries - 1);
1382 if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1383 EFX_POPULATE_OWORD_3(reg,
1384 FRF_CZ_TIMER_Q_EN, 1,
1385 FRF_CZ_HOST_NOTIFY_MODE, 0,
1386 FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
1387 efx_writeo_table(efx, ®, FR_BZ_TIMER_TBL, channel->channel);
1390 /* Pin event queue buffer */
1391 efx_init_special_buffer(efx, &channel->eventq);
1393 /* Fill event queue with all ones (i.e. empty events) */
1394 memset(channel->eventq.buf.addr, 0xff, channel->eventq.buf.len);
1396 /* Push event queue to card */
1397 EFX_POPULATE_OWORD_3(reg,
1399 FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
1400 FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
1401 efx_writeo_table(efx, ®, efx->type->evq_ptr_tbl_base,
1407 void efx_farch_ev_fini(struct efx_channel *channel)
1410 struct efx_nic *efx = channel->efx;
1412 /* Remove event queue from card */
1413 EFX_ZERO_OWORD(reg);
1414 efx_writeo_table(efx, ®, efx->type->evq_ptr_tbl_base,
1416 if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1417 efx_writeo_table(efx, ®, FR_BZ_TIMER_TBL, channel->channel);
1419 /* Unpin event queue */
1420 efx_fini_special_buffer(efx, &channel->eventq);
1423 /* Free buffers backing event queue */
1424 void efx_farch_ev_remove(struct efx_channel *channel)
1426 efx_free_special_buffer(channel->efx, &channel->eventq);
1430 void efx_farch_ev_test_generate(struct efx_channel *channel)
1432 efx_farch_magic_event(channel, EFX_CHANNEL_MAGIC_TEST(channel));
1435 void efx_farch_rx_defer_refill(struct efx_rx_queue *rx_queue)
1437 efx_farch_magic_event(efx_rx_queue_channel(rx_queue),
1438 EFX_CHANNEL_MAGIC_FILL(rx_queue));
1441 /**************************************************************************
1443 * Hardware interrupts
1444 * The hardware interrupt handler does very little work; all the event
1445 * queue processing is carried out by per-channel tasklets.
1447 **************************************************************************/
1449 /* Enable/disable/generate interrupts */
1450 static inline void efx_farch_interrupts(struct efx_nic *efx,
1451 bool enabled, bool force)
1453 efx_oword_t int_en_reg_ker;
1455 EFX_POPULATE_OWORD_3(int_en_reg_ker,
1456 FRF_AZ_KER_INT_LEVE_SEL, efx->irq_level,
1457 FRF_AZ_KER_INT_KER, force,
1458 FRF_AZ_DRV_INT_EN_KER, enabled);
1459 efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
1462 void efx_farch_irq_enable_master(struct efx_nic *efx)
1464 EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
1465 wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
1467 efx_farch_interrupts(efx, true, false);
1470 void efx_farch_irq_disable_master(struct efx_nic *efx)
1472 /* Disable interrupts */
1473 efx_farch_interrupts(efx, false, false);
1476 /* Generate a test interrupt
1477 * Interrupt must already have been enabled, otherwise nasty things
1480 void efx_farch_irq_test_generate(struct efx_nic *efx)
1482 efx_farch_interrupts(efx, true, true);
1485 /* Process a fatal interrupt
1486 * Disable bus mastering ASAP and schedule a reset
1488 irqreturn_t efx_farch_fatal_interrupt(struct efx_nic *efx)
1490 struct falcon_nic_data *nic_data = efx->nic_data;
1491 efx_oword_t *int_ker = efx->irq_status.addr;
1492 efx_oword_t fatal_intr;
1493 int error, mem_perr;
1495 efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
1496 error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
1498 netif_err(efx, hw, efx->net_dev, "SYSTEM ERROR "EFX_OWORD_FMT" status "
1499 EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
1500 EFX_OWORD_VAL(fatal_intr),
1501 error ? "disabling bus mastering" : "no recognised error");
1503 /* If this is a memory parity error dump which blocks are offending */
1504 mem_perr = (EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER) ||
1505 EFX_OWORD_FIELD(fatal_intr, FRF_AZ_SRM_PERR_INT_KER));
1508 efx_reado(efx, ®, FR_AZ_MEM_STAT);
1509 netif_err(efx, hw, efx->net_dev,
1510 "SYSTEM ERROR: memory parity error "EFX_OWORD_FMT"\n",
1511 EFX_OWORD_VAL(reg));
1514 /* Disable both devices */
1515 pci_clear_master(efx->pci_dev);
1516 if (efx_nic_is_dual_func(efx))
1517 pci_clear_master(nic_data->pci_dev2);
1518 efx_farch_irq_disable_master(efx);
1520 /* Count errors and reset or disable the NIC accordingly */
1521 if (efx->int_error_count == 0 ||
1522 time_after(jiffies, efx->int_error_expire)) {
1523 efx->int_error_count = 0;
1524 efx->int_error_expire =
1525 jiffies + EFX_INT_ERROR_EXPIRE * HZ;
1527 if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
1528 netif_err(efx, hw, efx->net_dev,
1529 "SYSTEM ERROR - reset scheduled\n");
1530 efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
1532 netif_err(efx, hw, efx->net_dev,
1533 "SYSTEM ERROR - max number of errors seen."
1534 "NIC will be disabled\n");
1535 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1541 /* Handle a legacy interrupt
1542 * Acknowledges the interrupt and schedule event queue processing.
1544 irqreturn_t efx_farch_legacy_interrupt(int irq, void *dev_id)
1546 struct efx_nic *efx = dev_id;
1547 bool soft_enabled = ACCESS_ONCE(efx->irq_soft_enabled);
1548 efx_oword_t *int_ker = efx->irq_status.addr;
1549 irqreturn_t result = IRQ_NONE;
1550 struct efx_channel *channel;
1555 /* Read the ISR which also ACKs the interrupts */
1556 efx_readd(efx, ®, FR_BZ_INT_ISR0);
1557 queues = EFX_EXTRACT_DWORD(reg, 0, 31);
1559 /* Legacy interrupts are disabled too late by the EEH kernel
1560 * code. Disable them earlier.
1561 * If an EEH error occurred, the read will have returned all ones.
1563 if (EFX_DWORD_IS_ALL_ONES(reg) && efx_try_recovery(efx) &&
1564 !efx->eeh_disabled_legacy_irq) {
1565 disable_irq_nosync(efx->legacy_irq);
1566 efx->eeh_disabled_legacy_irq = true;
1569 /* Handle non-event-queue sources */
1570 if (queues & (1U << efx->irq_level) && soft_enabled) {
1571 syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1572 if (unlikely(syserr))
1573 return efx_farch_fatal_interrupt(efx);
1574 efx->last_irq_cpu = raw_smp_processor_id();
1578 efx->irq_zero_count = 0;
1580 /* Schedule processing of any interrupting queues */
1581 if (likely(soft_enabled)) {
1582 efx_for_each_channel(channel, efx) {
1584 efx_schedule_channel_irq(channel);
1588 result = IRQ_HANDLED;
1593 /* Legacy ISR read can return zero once (SF bug 15783) */
1595 /* We can't return IRQ_HANDLED more than once on seeing ISR=0
1596 * because this might be a shared interrupt. */
1597 if (efx->irq_zero_count++ == 0)
1598 result = IRQ_HANDLED;
1600 /* Ensure we schedule or rearm all event queues */
1601 if (likely(soft_enabled)) {
1602 efx_for_each_channel(channel, efx) {
1603 event = efx_event(channel,
1604 channel->eventq_read_ptr);
1605 if (efx_event_present(event))
1606 efx_schedule_channel_irq(channel);
1608 efx_farch_ev_read_ack(channel);
1613 if (result == IRQ_HANDLED)
1614 netif_vdbg(efx, intr, efx->net_dev,
1615 "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
1616 irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
1621 /* Handle an MSI interrupt
1623 * Handle an MSI hardware interrupt. This routine schedules event
1624 * queue processing. No interrupt acknowledgement cycle is necessary.
1625 * Also, we never need to check that the interrupt is for us, since
1626 * MSI interrupts cannot be shared.
1628 irqreturn_t efx_farch_msi_interrupt(int irq, void *dev_id)
1630 struct efx_msi_context *context = dev_id;
1631 struct efx_nic *efx = context->efx;
1632 efx_oword_t *int_ker = efx->irq_status.addr;
1635 netif_vdbg(efx, intr, efx->net_dev,
1636 "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
1637 irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
1639 if (!likely(ACCESS_ONCE(efx->irq_soft_enabled)))
1642 /* Handle non-event-queue sources */
1643 if (context->index == efx->irq_level) {
1644 syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1645 if (unlikely(syserr))
1646 return efx_farch_fatal_interrupt(efx);
1647 efx->last_irq_cpu = raw_smp_processor_id();
1650 /* Schedule processing of the channel */
1651 efx_schedule_channel_irq(efx->channel[context->index]);
1656 /* Setup RSS indirection table.
1657 * This maps from the hash value of the packet to RXQ
1659 void efx_farch_rx_push_indir_table(struct efx_nic *efx)
1664 BUG_ON(efx_nic_rev(efx) < EFX_REV_FALCON_B0);
1666 BUILD_BUG_ON(ARRAY_SIZE(efx->rx_indir_table) !=
1667 FR_BZ_RX_INDIRECTION_TBL_ROWS);
1669 for (i = 0; i < FR_BZ_RX_INDIRECTION_TBL_ROWS; i++) {
1670 EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
1671 efx->rx_indir_table[i]);
1672 efx_writed(efx, &dword,
1673 FR_BZ_RX_INDIRECTION_TBL +
1674 FR_BZ_RX_INDIRECTION_TBL_STEP * i);
1678 /* Looks at available SRAM resources and works out how many queues we
1679 * can support, and where things like descriptor caches should live.
1681 * SRAM is split up as follows:
1682 * 0 buftbl entries for channels
1683 * efx->vf_buftbl_base buftbl entries for SR-IOV
1684 * efx->rx_dc_base RX descriptor caches
1685 * efx->tx_dc_base TX descriptor caches
1687 void efx_farch_dimension_resources(struct efx_nic *efx, unsigned sram_lim_qw)
1689 unsigned vi_count, buftbl_min;
1691 #ifdef CONFIG_SFC_SRIOV
1692 struct siena_nic_data *nic_data = efx->nic_data;
1695 /* Account for the buffer table entries backing the datapath channels
1696 * and the descriptor caches for those channels.
1698 buftbl_min = ((efx->n_rx_channels * EFX_MAX_DMAQ_SIZE +
1699 efx->n_tx_channels * EFX_TXQ_TYPES * EFX_MAX_DMAQ_SIZE +
1700 efx->n_channels * EFX_MAX_EVQ_SIZE)
1701 * sizeof(efx_qword_t) / EFX_BUF_SIZE);
1702 vi_count = max(efx->n_channels, efx->n_tx_channels * EFX_TXQ_TYPES);
1704 #ifdef CONFIG_SFC_SRIOV
1705 if (efx->type->sriov_wanted) {
1706 if (efx->type->sriov_wanted(efx)) {
1707 unsigned vi_dc_entries, buftbl_free;
1708 unsigned entries_per_vf, vf_limit;
1710 nic_data->vf_buftbl_base = buftbl_min;
1712 vi_dc_entries = RX_DC_ENTRIES + TX_DC_ENTRIES;
1713 vi_count = max(vi_count, EFX_VI_BASE);
1714 buftbl_free = (sram_lim_qw - buftbl_min -
1715 vi_count * vi_dc_entries);
1717 entries_per_vf = ((vi_dc_entries +
1718 EFX_VF_BUFTBL_PER_VI) *
1720 vf_limit = min(buftbl_free / entries_per_vf,
1721 (1024U - EFX_VI_BASE) >> efx->vi_scale);
1723 if (efx->vf_count > vf_limit) {
1724 netif_err(efx, probe, efx->net_dev,
1725 "Reducing VF count from from %d to %d\n",
1726 efx->vf_count, vf_limit);
1727 efx->vf_count = vf_limit;
1729 vi_count += efx->vf_count * efx_vf_size(efx);
1734 efx->tx_dc_base = sram_lim_qw - vi_count * TX_DC_ENTRIES;
1735 efx->rx_dc_base = efx->tx_dc_base - vi_count * RX_DC_ENTRIES;
1738 u32 efx_farch_fpga_ver(struct efx_nic *efx)
1740 efx_oword_t altera_build;
1741 efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
1742 return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
1745 void efx_farch_init_common(struct efx_nic *efx)
1749 /* Set positions of descriptor caches in SRAM. */
1750 EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR, efx->tx_dc_base);
1751 efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
1752 EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR, efx->rx_dc_base);
1753 efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
1755 /* Set TX descriptor cache size. */
1756 BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
1757 EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
1758 efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
1760 /* Set RX descriptor cache size. Set low watermark to size-8, as
1761 * this allows most efficient prefetching.
1763 BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
1764 EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
1765 efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
1766 EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
1767 efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
1769 /* Program INT_KER address */
1770 EFX_POPULATE_OWORD_2(temp,
1771 FRF_AZ_NORM_INT_VEC_DIS_KER,
1772 EFX_INT_MODE_USE_MSI(efx),
1773 FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
1774 efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
1776 if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
1777 /* Use an interrupt level unused by event queues */
1778 efx->irq_level = 0x1f;
1780 /* Use a valid MSI-X vector */
1783 /* Enable all the genuinely fatal interrupts. (They are still
1784 * masked by the overall interrupt mask, controlled by
1785 * falcon_interrupts()).
1787 * Note: All other fatal interrupts are enabled
1789 EFX_POPULATE_OWORD_3(temp,
1790 FRF_AZ_ILL_ADR_INT_KER_EN, 1,
1791 FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
1792 FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
1793 if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1794 EFX_SET_OWORD_FIELD(temp, FRF_CZ_SRAM_PERR_INT_P_KER_EN, 1);
1795 EFX_INVERT_OWORD(temp);
1796 efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
1798 /* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
1799 * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
1801 efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
1802 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
1803 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
1804 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
1805 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 1);
1806 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
1807 /* Enable SW_EV to inherit in char driver - assume harmless here */
1808 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
1809 /* Prefetch threshold 2 => fetch when descriptor cache half empty */
1810 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
1811 /* Disable hardware watchdog which can misfire */
1812 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
1813 /* Squash TX of packets of 16 bytes or less */
1814 if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1815 EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
1816 efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
1818 if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
1819 EFX_POPULATE_OWORD_4(temp,
1820 /* Default values */
1821 FRF_BZ_TX_PACE_SB_NOT_AF, 0x15,
1822 FRF_BZ_TX_PACE_SB_AF, 0xb,
1823 FRF_BZ_TX_PACE_FB_BASE, 0,
1824 /* Allow large pace values in the
1826 FRF_BZ_TX_PACE_BIN_TH,
1827 FFE_BZ_TX_PACE_RESERVED);
1828 efx_writeo(efx, &temp, FR_BZ_TX_PACE);
1832 /**************************************************************************
1836 **************************************************************************
1839 /* "Fudge factors" - difference between programmed value and actual depth.
1840 * Due to pipelined implementation we need to program H/W with a value that
1841 * is larger than the hop limit we want.
1843 #define EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD 3
1844 #define EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL 1
1846 /* Hard maximum search limit. Hardware will time-out beyond 200-something.
1847 * We also need to avoid infinite loops in efx_farch_filter_search() when the
1850 #define EFX_FARCH_FILTER_CTL_SRCH_MAX 200
1852 /* Don't try very hard to find space for performance hints, as this is
1853 * counter-productive. */
1854 #define EFX_FARCH_FILTER_CTL_SRCH_HINT_MAX 5
1856 enum efx_farch_filter_type {
1857 EFX_FARCH_FILTER_TCP_FULL = 0,
1858 EFX_FARCH_FILTER_TCP_WILD,
1859 EFX_FARCH_FILTER_UDP_FULL,
1860 EFX_FARCH_FILTER_UDP_WILD,
1861 EFX_FARCH_FILTER_MAC_FULL = 4,
1862 EFX_FARCH_FILTER_MAC_WILD,
1863 EFX_FARCH_FILTER_UC_DEF = 8,
1864 EFX_FARCH_FILTER_MC_DEF,
1865 EFX_FARCH_FILTER_TYPE_COUNT, /* number of specific types */
1868 enum efx_farch_filter_table_id {
1869 EFX_FARCH_FILTER_TABLE_RX_IP = 0,
1870 EFX_FARCH_FILTER_TABLE_RX_MAC,
1871 EFX_FARCH_FILTER_TABLE_RX_DEF,
1872 EFX_FARCH_FILTER_TABLE_TX_MAC,
1873 EFX_FARCH_FILTER_TABLE_COUNT,
1876 enum efx_farch_filter_index {
1877 EFX_FARCH_FILTER_INDEX_UC_DEF,
1878 EFX_FARCH_FILTER_INDEX_MC_DEF,
1879 EFX_FARCH_FILTER_SIZE_RX_DEF,
1882 struct efx_farch_filter_spec {
1890 struct efx_farch_filter_table {
1891 enum efx_farch_filter_table_id id;
1892 u32 offset; /* address of table relative to BAR */
1893 unsigned size; /* number of entries */
1894 unsigned step; /* step between entries */
1895 unsigned used; /* number currently used */
1896 unsigned long *used_bitmap;
1897 struct efx_farch_filter_spec *spec;
1898 unsigned search_limit[EFX_FARCH_FILTER_TYPE_COUNT];
1901 struct efx_farch_filter_state {
1902 struct efx_farch_filter_table table[EFX_FARCH_FILTER_TABLE_COUNT];
1906 efx_farch_filter_table_clear_entry(struct efx_nic *efx,
1907 struct efx_farch_filter_table *table,
1908 unsigned int filter_idx);
1910 /* The filter hash function is LFSR polynomial x^16 + x^3 + 1 of a 32-bit
1911 * key derived from the n-tuple. The initial LFSR state is 0xffff. */
1912 static u16 efx_farch_filter_hash(u32 key)
1916 /* First 16 rounds */
1917 tmp = 0x1fff ^ key >> 16;
1918 tmp = tmp ^ tmp >> 3 ^ tmp >> 6;
1919 tmp = tmp ^ tmp >> 9;
1920 /* Last 16 rounds */
1921 tmp = tmp ^ tmp << 13 ^ key;
1922 tmp = tmp ^ tmp >> 3 ^ tmp >> 6;
1923 return tmp ^ tmp >> 9;
1926 /* To allow for hash collisions, filter search continues at these
1927 * increments from the first possible entry selected by the hash. */
1928 static u16 efx_farch_filter_increment(u32 key)
1933 static enum efx_farch_filter_table_id
1934 efx_farch_filter_spec_table_id(const struct efx_farch_filter_spec *spec)
1936 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1937 (EFX_FARCH_FILTER_TCP_FULL >> 2));
1938 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1939 (EFX_FARCH_FILTER_TCP_WILD >> 2));
1940 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1941 (EFX_FARCH_FILTER_UDP_FULL >> 2));
1942 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1943 (EFX_FARCH_FILTER_UDP_WILD >> 2));
1944 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_MAC !=
1945 (EFX_FARCH_FILTER_MAC_FULL >> 2));
1946 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_MAC !=
1947 (EFX_FARCH_FILTER_MAC_WILD >> 2));
1948 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_TX_MAC !=
1949 EFX_FARCH_FILTER_TABLE_RX_MAC + 2);
1950 return (spec->type >> 2) + ((spec->flags & EFX_FILTER_FLAG_TX) ? 2 : 0);
1953 static void efx_farch_filter_push_rx_config(struct efx_nic *efx)
1955 struct efx_farch_filter_state *state = efx->filter_state;
1956 struct efx_farch_filter_table *table;
1957 efx_oword_t filter_ctl;
1959 efx_reado(efx, &filter_ctl, FR_BZ_RX_FILTER_CTL);
1961 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_IP];
1962 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_TCP_FULL_SRCH_LIMIT,
1963 table->search_limit[EFX_FARCH_FILTER_TCP_FULL] +
1964 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
1965 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_TCP_WILD_SRCH_LIMIT,
1966 table->search_limit[EFX_FARCH_FILTER_TCP_WILD] +
1967 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
1968 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_UDP_FULL_SRCH_LIMIT,
1969 table->search_limit[EFX_FARCH_FILTER_UDP_FULL] +
1970 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
1971 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_UDP_WILD_SRCH_LIMIT,
1972 table->search_limit[EFX_FARCH_FILTER_UDP_WILD] +
1973 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
1975 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_MAC];
1977 EFX_SET_OWORD_FIELD(
1978 filter_ctl, FRF_CZ_ETHERNET_FULL_SEARCH_LIMIT,
1979 table->search_limit[EFX_FARCH_FILTER_MAC_FULL] +
1980 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
1981 EFX_SET_OWORD_FIELD(
1982 filter_ctl, FRF_CZ_ETHERNET_WILDCARD_SEARCH_LIMIT,
1983 table->search_limit[EFX_FARCH_FILTER_MAC_WILD] +
1984 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
1987 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_DEF];
1989 EFX_SET_OWORD_FIELD(
1990 filter_ctl, FRF_CZ_UNICAST_NOMATCH_Q_ID,
1991 table->spec[EFX_FARCH_FILTER_INDEX_UC_DEF].dmaq_id);
1992 EFX_SET_OWORD_FIELD(
1993 filter_ctl, FRF_CZ_UNICAST_NOMATCH_RSS_ENABLED,
1994 !!(table->spec[EFX_FARCH_FILTER_INDEX_UC_DEF].flags &
1995 EFX_FILTER_FLAG_RX_RSS));
1996 EFX_SET_OWORD_FIELD(
1997 filter_ctl, FRF_CZ_MULTICAST_NOMATCH_Q_ID,
1998 table->spec[EFX_FARCH_FILTER_INDEX_MC_DEF].dmaq_id);
1999 EFX_SET_OWORD_FIELD(
2000 filter_ctl, FRF_CZ_MULTICAST_NOMATCH_RSS_ENABLED,
2001 !!(table->spec[EFX_FARCH_FILTER_INDEX_MC_DEF].flags &
2002 EFX_FILTER_FLAG_RX_RSS));
2004 /* There is a single bit to enable RX scatter for all
2005 * unmatched packets. Only set it if scatter is
2006 * enabled in both filter specs.
2008 EFX_SET_OWORD_FIELD(
2009 filter_ctl, FRF_BZ_SCATTER_ENBL_NO_MATCH_Q,
2010 !!(table->spec[EFX_FARCH_FILTER_INDEX_UC_DEF].flags &
2011 table->spec[EFX_FARCH_FILTER_INDEX_MC_DEF].flags &
2012 EFX_FILTER_FLAG_RX_SCATTER));
2013 } else if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
2014 /* We don't expose 'default' filters because unmatched
2015 * packets always go to the queue number found in the
2016 * RSS table. But we still need to set the RX scatter
2019 EFX_SET_OWORD_FIELD(
2020 filter_ctl, FRF_BZ_SCATTER_ENBL_NO_MATCH_Q,
2024 efx_writeo(efx, &filter_ctl, FR_BZ_RX_FILTER_CTL);
2027 static void efx_farch_filter_push_tx_limits(struct efx_nic *efx)
2029 struct efx_farch_filter_state *state = efx->filter_state;
2030 struct efx_farch_filter_table *table;
2033 efx_reado(efx, &tx_cfg, FR_AZ_TX_CFG);
2035 table = &state->table[EFX_FARCH_FILTER_TABLE_TX_MAC];
2037 EFX_SET_OWORD_FIELD(
2038 tx_cfg, FRF_CZ_TX_ETH_FILTER_FULL_SEARCH_RANGE,
2039 table->search_limit[EFX_FARCH_FILTER_MAC_FULL] +
2040 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
2041 EFX_SET_OWORD_FIELD(
2042 tx_cfg, FRF_CZ_TX_ETH_FILTER_WILD_SEARCH_RANGE,
2043 table->search_limit[EFX_FARCH_FILTER_MAC_WILD] +
2044 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
2047 efx_writeo(efx, &tx_cfg, FR_AZ_TX_CFG);
2051 efx_farch_filter_from_gen_spec(struct efx_farch_filter_spec *spec,
2052 const struct efx_filter_spec *gen_spec)
2054 bool is_full = false;
2056 if ((gen_spec->flags & EFX_FILTER_FLAG_RX_RSS) &&
2057 gen_spec->rss_context != EFX_FILTER_RSS_CONTEXT_DEFAULT)
2060 spec->priority = gen_spec->priority;
2061 spec->flags = gen_spec->flags;
2062 spec->dmaq_id = gen_spec->dmaq_id;
2064 switch (gen_spec->match_flags) {
2065 case (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
2066 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
2067 EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT):
2070 case (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
2071 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT): {
2072 __be32 rhost, host1, host2;
2073 __be16 rport, port1, port2;
2075 EFX_BUG_ON_PARANOID(!(gen_spec->flags & EFX_FILTER_FLAG_RX));
2077 if (gen_spec->ether_type != htons(ETH_P_IP))
2078 return -EPROTONOSUPPORT;
2079 if (gen_spec->loc_port == 0 ||
2080 (is_full && gen_spec->rem_port == 0))
2081 return -EADDRNOTAVAIL;
2082 switch (gen_spec->ip_proto) {
2084 spec->type = (is_full ? EFX_FARCH_FILTER_TCP_FULL :
2085 EFX_FARCH_FILTER_TCP_WILD);
2088 spec->type = (is_full ? EFX_FARCH_FILTER_UDP_FULL :
2089 EFX_FARCH_FILTER_UDP_WILD);
2092 return -EPROTONOSUPPORT;
2095 /* Filter is constructed in terms of source and destination,
2096 * with the odd wrinkle that the ports are swapped in a UDP
2097 * wildcard filter. We need to convert from local and remote
2098 * (= zero for wildcard) addresses.
2100 rhost = is_full ? gen_spec->rem_host[0] : 0;
2101 rport = is_full ? gen_spec->rem_port : 0;
2103 host2 = gen_spec->loc_host[0];
2104 if (!is_full && gen_spec->ip_proto == IPPROTO_UDP) {
2105 port1 = gen_spec->loc_port;
2109 port2 = gen_spec->loc_port;
2111 spec->data[0] = ntohl(host1) << 16 | ntohs(port1);
2112 spec->data[1] = ntohs(port2) << 16 | ntohl(host1) >> 16;
2113 spec->data[2] = ntohl(host2);
2118 case EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_OUTER_VID:
2121 case EFX_FILTER_MATCH_LOC_MAC:
2122 spec->type = (is_full ? EFX_FARCH_FILTER_MAC_FULL :
2123 EFX_FARCH_FILTER_MAC_WILD);
2124 spec->data[0] = is_full ? ntohs(gen_spec->outer_vid) : 0;
2125 spec->data[1] = (gen_spec->loc_mac[2] << 24 |
2126 gen_spec->loc_mac[3] << 16 |
2127 gen_spec->loc_mac[4] << 8 |
2128 gen_spec->loc_mac[5]);
2129 spec->data[2] = (gen_spec->loc_mac[0] << 8 |
2130 gen_spec->loc_mac[1]);
2133 case EFX_FILTER_MATCH_LOC_MAC_IG:
2134 spec->type = (is_multicast_ether_addr(gen_spec->loc_mac) ?
2135 EFX_FARCH_FILTER_MC_DEF :
2136 EFX_FARCH_FILTER_UC_DEF);
2137 memset(spec->data, 0, sizeof(spec->data)); /* ensure equality */
2141 return -EPROTONOSUPPORT;
2148 efx_farch_filter_to_gen_spec(struct efx_filter_spec *gen_spec,
2149 const struct efx_farch_filter_spec *spec)
2151 bool is_full = false;
2153 /* *gen_spec should be completely initialised, to be consistent
2154 * with efx_filter_init_{rx,tx}() and in case we want to copy
2155 * it back to userland.
2157 memset(gen_spec, 0, sizeof(*gen_spec));
2159 gen_spec->priority = spec->priority;
2160 gen_spec->flags = spec->flags;
2161 gen_spec->dmaq_id = spec->dmaq_id;
2163 switch (spec->type) {
2164 case EFX_FARCH_FILTER_TCP_FULL:
2165 case EFX_FARCH_FILTER_UDP_FULL:
2168 case EFX_FARCH_FILTER_TCP_WILD:
2169 case EFX_FARCH_FILTER_UDP_WILD: {
2170 __be32 host1, host2;
2171 __be16 port1, port2;
2173 gen_spec->match_flags =
2174 EFX_FILTER_MATCH_ETHER_TYPE |
2175 EFX_FILTER_MATCH_IP_PROTO |
2176 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT;
2178 gen_spec->match_flags |= (EFX_FILTER_MATCH_REM_HOST |
2179 EFX_FILTER_MATCH_REM_PORT);
2180 gen_spec->ether_type = htons(ETH_P_IP);
2181 gen_spec->ip_proto =
2182 (spec->type == EFX_FARCH_FILTER_TCP_FULL ||
2183 spec->type == EFX_FARCH_FILTER_TCP_WILD) ?
2184 IPPROTO_TCP : IPPROTO_UDP;
2186 host1 = htonl(spec->data[0] >> 16 | spec->data[1] << 16);
2187 port1 = htons(spec->data[0]);
2188 host2 = htonl(spec->data[2]);
2189 port2 = htons(spec->data[1] >> 16);
2190 if (spec->flags & EFX_FILTER_FLAG_TX) {
2191 gen_spec->loc_host[0] = host1;
2192 gen_spec->rem_host[0] = host2;
2194 gen_spec->loc_host[0] = host2;
2195 gen_spec->rem_host[0] = host1;
2197 if (!!(gen_spec->flags & EFX_FILTER_FLAG_TX) ^
2198 (!is_full && gen_spec->ip_proto == IPPROTO_UDP)) {
2199 gen_spec->loc_port = port1;
2200 gen_spec->rem_port = port2;
2202 gen_spec->loc_port = port2;
2203 gen_spec->rem_port = port1;
2209 case EFX_FARCH_FILTER_MAC_FULL:
2212 case EFX_FARCH_FILTER_MAC_WILD:
2213 gen_spec->match_flags = EFX_FILTER_MATCH_LOC_MAC;
2215 gen_spec->match_flags |= EFX_FILTER_MATCH_OUTER_VID;
2216 gen_spec->loc_mac[0] = spec->data[2] >> 8;
2217 gen_spec->loc_mac[1] = spec->data[2];
2218 gen_spec->loc_mac[2] = spec->data[1] >> 24;
2219 gen_spec->loc_mac[3] = spec->data[1] >> 16;
2220 gen_spec->loc_mac[4] = spec->data[1] >> 8;
2221 gen_spec->loc_mac[5] = spec->data[1];
2222 gen_spec->outer_vid = htons(spec->data[0]);
2225 case EFX_FARCH_FILTER_UC_DEF:
2226 case EFX_FARCH_FILTER_MC_DEF:
2227 gen_spec->match_flags = EFX_FILTER_MATCH_LOC_MAC_IG;
2228 gen_spec->loc_mac[0] = spec->type == EFX_FARCH_FILTER_MC_DEF;
2238 efx_farch_filter_init_rx_auto(struct efx_nic *efx,
2239 struct efx_farch_filter_spec *spec)
2241 /* If there's only one channel then disable RSS for non VF
2242 * traffic, thereby allowing VFs to use RSS when the PF can't.
2244 spec->priority = EFX_FILTER_PRI_AUTO;
2245 spec->flags = (EFX_FILTER_FLAG_RX |
2246 (efx_rss_enabled(efx) ? EFX_FILTER_FLAG_RX_RSS : 0) |
2247 (efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0));
2251 /* Build a filter entry and return its n-tuple key. */
2252 static u32 efx_farch_filter_build(efx_oword_t *filter,
2253 struct efx_farch_filter_spec *spec)
2257 switch (efx_farch_filter_spec_table_id(spec)) {
2258 case EFX_FARCH_FILTER_TABLE_RX_IP: {
2259 bool is_udp = (spec->type == EFX_FARCH_FILTER_UDP_FULL ||
2260 spec->type == EFX_FARCH_FILTER_UDP_WILD);
2261 EFX_POPULATE_OWORD_7(
2264 !!(spec->flags & EFX_FILTER_FLAG_RX_RSS),
2266 !!(spec->flags & EFX_FILTER_FLAG_RX_SCATTER),
2267 FRF_BZ_TCP_UDP, is_udp,
2268 FRF_BZ_RXQ_ID, spec->dmaq_id,
2269 EFX_DWORD_2, spec->data[2],
2270 EFX_DWORD_1, spec->data[1],
2271 EFX_DWORD_0, spec->data[0]);
2276 case EFX_FARCH_FILTER_TABLE_RX_MAC: {
2277 bool is_wild = spec->type == EFX_FARCH_FILTER_MAC_WILD;
2278 EFX_POPULATE_OWORD_7(
2281 !!(spec->flags & EFX_FILTER_FLAG_RX_RSS),
2282 FRF_CZ_RMFT_SCATTER_EN,
2283 !!(spec->flags & EFX_FILTER_FLAG_RX_SCATTER),
2284 FRF_CZ_RMFT_RXQ_ID, spec->dmaq_id,
2285 FRF_CZ_RMFT_WILDCARD_MATCH, is_wild,
2286 FRF_CZ_RMFT_DEST_MAC_HI, spec->data[2],
2287 FRF_CZ_RMFT_DEST_MAC_LO, spec->data[1],
2288 FRF_CZ_RMFT_VLAN_ID, spec->data[0]);
2293 case EFX_FARCH_FILTER_TABLE_TX_MAC: {
2294 bool is_wild = spec->type == EFX_FARCH_FILTER_MAC_WILD;
2295 EFX_POPULATE_OWORD_5(*filter,
2296 FRF_CZ_TMFT_TXQ_ID, spec->dmaq_id,
2297 FRF_CZ_TMFT_WILDCARD_MATCH, is_wild,
2298 FRF_CZ_TMFT_SRC_MAC_HI, spec->data[2],
2299 FRF_CZ_TMFT_SRC_MAC_LO, spec->data[1],
2300 FRF_CZ_TMFT_VLAN_ID, spec->data[0]);
2301 data3 = is_wild | spec->dmaq_id << 1;
2309 return spec->data[0] ^ spec->data[1] ^ spec->data[2] ^ data3;
2312 static bool efx_farch_filter_equal(const struct efx_farch_filter_spec *left,
2313 const struct efx_farch_filter_spec *right)
2315 if (left->type != right->type ||
2316 memcmp(left->data, right->data, sizeof(left->data)))
2319 if (left->flags & EFX_FILTER_FLAG_TX &&
2320 left->dmaq_id != right->dmaq_id)
2327 * Construct/deconstruct external filter IDs. At least the RX filter
2328 * IDs must be ordered by matching priority, for RX NFC semantics.
2330 * Deconstruction needs to be robust against invalid IDs so that
2331 * efx_filter_remove_id_safe() and efx_filter_get_filter_safe() can
2332 * accept user-provided IDs.
2335 #define EFX_FARCH_FILTER_MATCH_PRI_COUNT 5
2337 static const u8 efx_farch_filter_type_match_pri[EFX_FARCH_FILTER_TYPE_COUNT] = {
2338 [EFX_FARCH_FILTER_TCP_FULL] = 0,
2339 [EFX_FARCH_FILTER_UDP_FULL] = 0,
2340 [EFX_FARCH_FILTER_TCP_WILD] = 1,
2341 [EFX_FARCH_FILTER_UDP_WILD] = 1,
2342 [EFX_FARCH_FILTER_MAC_FULL] = 2,
2343 [EFX_FARCH_FILTER_MAC_WILD] = 3,
2344 [EFX_FARCH_FILTER_UC_DEF] = 4,
2345 [EFX_FARCH_FILTER_MC_DEF] = 4,
2348 static const enum efx_farch_filter_table_id efx_farch_filter_range_table[] = {
2349 EFX_FARCH_FILTER_TABLE_RX_IP, /* RX match pri 0 */
2350 EFX_FARCH_FILTER_TABLE_RX_IP,
2351 EFX_FARCH_FILTER_TABLE_RX_MAC,
2352 EFX_FARCH_FILTER_TABLE_RX_MAC,
2353 EFX_FARCH_FILTER_TABLE_RX_DEF, /* RX match pri 4 */
2354 EFX_FARCH_FILTER_TABLE_TX_MAC, /* TX match pri 0 */
2355 EFX_FARCH_FILTER_TABLE_TX_MAC, /* TX match pri 1 */
2358 #define EFX_FARCH_FILTER_INDEX_WIDTH 13
2359 #define EFX_FARCH_FILTER_INDEX_MASK ((1 << EFX_FARCH_FILTER_INDEX_WIDTH) - 1)
2362 efx_farch_filter_make_id(const struct efx_farch_filter_spec *spec,
2367 range = efx_farch_filter_type_match_pri[spec->type];
2368 if (!(spec->flags & EFX_FILTER_FLAG_RX))
2369 range += EFX_FARCH_FILTER_MATCH_PRI_COUNT;
2371 return range << EFX_FARCH_FILTER_INDEX_WIDTH | index;
2374 static inline enum efx_farch_filter_table_id
2375 efx_farch_filter_id_table_id(u32 id)
2377 unsigned int range = id >> EFX_FARCH_FILTER_INDEX_WIDTH;
2379 if (range < ARRAY_SIZE(efx_farch_filter_range_table))
2380 return efx_farch_filter_range_table[range];
2382 return EFX_FARCH_FILTER_TABLE_COUNT; /* invalid */
2385 static inline unsigned int efx_farch_filter_id_index(u32 id)
2387 return id & EFX_FARCH_FILTER_INDEX_MASK;
2390 u32 efx_farch_filter_get_rx_id_limit(struct efx_nic *efx)
2392 struct efx_farch_filter_state *state = efx->filter_state;
2393 unsigned int range = EFX_FARCH_FILTER_MATCH_PRI_COUNT - 1;
2394 enum efx_farch_filter_table_id table_id;
2397 table_id = efx_farch_filter_range_table[range];
2398 if (state->table[table_id].size != 0)
2399 return range << EFX_FARCH_FILTER_INDEX_WIDTH |
2400 state->table[table_id].size;
2406 s32 efx_farch_filter_insert(struct efx_nic *efx,
2407 struct efx_filter_spec *gen_spec,
2410 struct efx_farch_filter_state *state = efx->filter_state;
2411 struct efx_farch_filter_table *table;
2412 struct efx_farch_filter_spec spec;
2414 int rep_index, ins_index;
2415 unsigned int depth = 0;
2418 rc = efx_farch_filter_from_gen_spec(&spec, gen_spec);
2422 table = &state->table[efx_farch_filter_spec_table_id(&spec)];
2423 if (table->size == 0)
2426 netif_vdbg(efx, hw, efx->net_dev,
2427 "%s: type %d search_limit=%d", __func__, spec.type,
2428 table->search_limit[spec.type]);
2430 if (table->id == EFX_FARCH_FILTER_TABLE_RX_DEF) {
2431 /* One filter spec per type */
2432 BUILD_BUG_ON(EFX_FARCH_FILTER_INDEX_UC_DEF != 0);
2433 BUILD_BUG_ON(EFX_FARCH_FILTER_INDEX_MC_DEF !=
2434 EFX_FARCH_FILTER_MC_DEF - EFX_FARCH_FILTER_UC_DEF);
2435 rep_index = spec.type - EFX_FARCH_FILTER_UC_DEF;
2436 ins_index = rep_index;
2438 spin_lock_bh(&efx->filter_lock);
2440 /* Search concurrently for
2441 * (1) a filter to be replaced (rep_index): any filter
2442 * with the same match values, up to the current
2443 * search depth for this type, and
2444 * (2) the insertion point (ins_index): (1) or any
2445 * free slot before it or up to the maximum search
2446 * depth for this priority
2447 * We fail if we cannot find (2).
2449 * We can stop once either
2450 * (a) we find (1), in which case we have definitely
2451 * found (2) as well; or
2452 * (b) we have searched exhaustively for (1), and have
2453 * either found (2) or searched exhaustively for it
2455 u32 key = efx_farch_filter_build(&filter, &spec);
2456 unsigned int hash = efx_farch_filter_hash(key);
2457 unsigned int incr = efx_farch_filter_increment(key);
2458 unsigned int max_rep_depth = table->search_limit[spec.type];
2459 unsigned int max_ins_depth =
2460 spec.priority <= EFX_FILTER_PRI_HINT ?
2461 EFX_FARCH_FILTER_CTL_SRCH_HINT_MAX :
2462 EFX_FARCH_FILTER_CTL_SRCH_MAX;
2463 unsigned int i = hash & (table->size - 1);
2468 spin_lock_bh(&efx->filter_lock);
2471 if (!test_bit(i, table->used_bitmap)) {
2474 } else if (efx_farch_filter_equal(&spec,
2483 if (depth >= max_rep_depth &&
2484 (ins_index >= 0 || depth >= max_ins_depth)) {
2486 if (ins_index < 0) {
2494 i = (i + incr) & (table->size - 1);
2499 /* If we found a filter to be replaced, check whether we
2502 if (rep_index >= 0) {
2503 struct efx_farch_filter_spec *saved_spec =
2504 &table->spec[rep_index];
2506 if (spec.priority == saved_spec->priority && !replace_equal) {
2510 if (spec.priority < saved_spec->priority) {
2514 if (saved_spec->priority == EFX_FILTER_PRI_AUTO ||
2515 saved_spec->flags & EFX_FILTER_FLAG_RX_OVER_AUTO)
2516 spec.flags |= EFX_FILTER_FLAG_RX_OVER_AUTO;
2519 /* Insert the filter */
2520 if (ins_index != rep_index) {
2521 __set_bit(ins_index, table->used_bitmap);
2524 table->spec[ins_index] = spec;
2526 if (table->id == EFX_FARCH_FILTER_TABLE_RX_DEF) {
2527 efx_farch_filter_push_rx_config(efx);
2529 if (table->search_limit[spec.type] < depth) {
2530 table->search_limit[spec.type] = depth;
2531 if (spec.flags & EFX_FILTER_FLAG_TX)
2532 efx_farch_filter_push_tx_limits(efx);
2534 efx_farch_filter_push_rx_config(efx);
2537 efx_writeo(efx, &filter,
2538 table->offset + table->step * ins_index);
2540 /* If we were able to replace a filter by inserting
2541 * at a lower depth, clear the replaced filter
2543 if (ins_index != rep_index && rep_index >= 0)
2544 efx_farch_filter_table_clear_entry(efx, table,
2548 netif_vdbg(efx, hw, efx->net_dev,
2549 "%s: filter type %d index %d rxq %u set",
2550 __func__, spec.type, ins_index, spec.dmaq_id);
2551 rc = efx_farch_filter_make_id(&spec, ins_index);
2554 spin_unlock_bh(&efx->filter_lock);
2559 efx_farch_filter_table_clear_entry(struct efx_nic *efx,
2560 struct efx_farch_filter_table *table,
2561 unsigned int filter_idx)
2563 static efx_oword_t filter;
2565 EFX_WARN_ON_PARANOID(!test_bit(filter_idx, table->used_bitmap));
2566 BUG_ON(table->offset == 0); /* can't clear MAC default filters */
2568 __clear_bit(filter_idx, table->used_bitmap);
2570 memset(&table->spec[filter_idx], 0, sizeof(table->spec[0]));
2572 efx_writeo(efx, &filter, table->offset + table->step * filter_idx);
2574 /* If this filter required a greater search depth than
2575 * any other, the search limit for its type can now be
2576 * decreased. However, it is hard to determine that
2577 * unless the table has become completely empty - in
2578 * which case, all its search limits can be set to 0.
2580 if (unlikely(table->used == 0)) {
2581 memset(table->search_limit, 0, sizeof(table->search_limit));
2582 if (table->id == EFX_FARCH_FILTER_TABLE_TX_MAC)
2583 efx_farch_filter_push_tx_limits(efx);
2585 efx_farch_filter_push_rx_config(efx);
2589 static int efx_farch_filter_remove(struct efx_nic *efx,
2590 struct efx_farch_filter_table *table,
2591 unsigned int filter_idx,
2592 enum efx_filter_priority priority)
2594 struct efx_farch_filter_spec *spec = &table->spec[filter_idx];
2596 if (!test_bit(filter_idx, table->used_bitmap) ||
2597 spec->priority != priority)
2600 if (spec->flags & EFX_FILTER_FLAG_RX_OVER_AUTO) {
2601 efx_farch_filter_init_rx_auto(efx, spec);
2602 efx_farch_filter_push_rx_config(efx);
2604 efx_farch_filter_table_clear_entry(efx, table, filter_idx);
2610 int efx_farch_filter_remove_safe(struct efx_nic *efx,
2611 enum efx_filter_priority priority,
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 unsigned int filter_idx;
2618 struct efx_farch_filter_spec *spec;
2621 table_id = efx_farch_filter_id_table_id(filter_id);
2622 if ((unsigned int)table_id >= EFX_FARCH_FILTER_TABLE_COUNT)
2624 table = &state->table[table_id];
2626 filter_idx = efx_farch_filter_id_index(filter_id);
2627 if (filter_idx >= table->size)
2629 spec = &table->spec[filter_idx];
2631 spin_lock_bh(&efx->filter_lock);
2632 rc = efx_farch_filter_remove(efx, table, filter_idx, priority);
2633 spin_unlock_bh(&efx->filter_lock);
2638 int efx_farch_filter_get_safe(struct efx_nic *efx,
2639 enum efx_filter_priority priority,
2640 u32 filter_id, struct efx_filter_spec *spec_buf)
2642 struct efx_farch_filter_state *state = efx->filter_state;
2643 enum efx_farch_filter_table_id table_id;
2644 struct efx_farch_filter_table *table;
2645 struct efx_farch_filter_spec *spec;
2646 unsigned int filter_idx;
2649 table_id = efx_farch_filter_id_table_id(filter_id);
2650 if ((unsigned int)table_id >= EFX_FARCH_FILTER_TABLE_COUNT)
2652 table = &state->table[table_id];
2654 filter_idx = efx_farch_filter_id_index(filter_id);
2655 if (filter_idx >= table->size)
2657 spec = &table->spec[filter_idx];
2659 spin_lock_bh(&efx->filter_lock);
2661 if (test_bit(filter_idx, table->used_bitmap) &&
2662 spec->priority == priority) {
2663 efx_farch_filter_to_gen_spec(spec_buf, spec);
2669 spin_unlock_bh(&efx->filter_lock);
2675 efx_farch_filter_table_clear(struct efx_nic *efx,
2676 enum efx_farch_filter_table_id table_id,
2677 enum efx_filter_priority priority)
2679 struct efx_farch_filter_state *state = efx->filter_state;
2680 struct efx_farch_filter_table *table = &state->table[table_id];
2681 unsigned int filter_idx;
2683 spin_lock_bh(&efx->filter_lock);
2684 for (filter_idx = 0; filter_idx < table->size; ++filter_idx) {
2685 if (table->spec[filter_idx].priority != EFX_FILTER_PRI_AUTO)
2686 efx_farch_filter_remove(efx, table,
2687 filter_idx, priority);
2689 spin_unlock_bh(&efx->filter_lock);
2692 int efx_farch_filter_clear_rx(struct efx_nic *efx,
2693 enum efx_filter_priority priority)
2695 efx_farch_filter_table_clear(efx, EFX_FARCH_FILTER_TABLE_RX_IP,
2697 efx_farch_filter_table_clear(efx, EFX_FARCH_FILTER_TABLE_RX_MAC,
2699 efx_farch_filter_table_clear(efx, EFX_FARCH_FILTER_TABLE_RX_DEF,
2704 u32 efx_farch_filter_count_rx_used(struct efx_nic *efx,
2705 enum efx_filter_priority priority)
2707 struct efx_farch_filter_state *state = efx->filter_state;
2708 enum efx_farch_filter_table_id table_id;
2709 struct efx_farch_filter_table *table;
2710 unsigned int filter_idx;
2713 spin_lock_bh(&efx->filter_lock);
2715 for (table_id = EFX_FARCH_FILTER_TABLE_RX_IP;
2716 table_id <= EFX_FARCH_FILTER_TABLE_RX_DEF;
2718 table = &state->table[table_id];
2719 for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
2720 if (test_bit(filter_idx, table->used_bitmap) &&
2721 table->spec[filter_idx].priority == priority)
2726 spin_unlock_bh(&efx->filter_lock);
2731 s32 efx_farch_filter_get_rx_ids(struct efx_nic *efx,
2732 enum efx_filter_priority priority,
2735 struct efx_farch_filter_state *state = efx->filter_state;
2736 enum efx_farch_filter_table_id table_id;
2737 struct efx_farch_filter_table *table;
2738 unsigned int filter_idx;
2741 spin_lock_bh(&efx->filter_lock);
2743 for (table_id = EFX_FARCH_FILTER_TABLE_RX_IP;
2744 table_id <= EFX_FARCH_FILTER_TABLE_RX_DEF;
2746 table = &state->table[table_id];
2747 for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
2748 if (test_bit(filter_idx, table->used_bitmap) &&
2749 table->spec[filter_idx].priority == priority) {
2750 if (count == size) {
2754 buf[count++] = efx_farch_filter_make_id(
2755 &table->spec[filter_idx], filter_idx);
2760 spin_unlock_bh(&efx->filter_lock);
2765 /* Restore filter stater after reset */
2766 void efx_farch_filter_table_restore(struct efx_nic *efx)
2768 struct efx_farch_filter_state *state = efx->filter_state;
2769 enum efx_farch_filter_table_id table_id;
2770 struct efx_farch_filter_table *table;
2772 unsigned int filter_idx;
2774 spin_lock_bh(&efx->filter_lock);
2776 for (table_id = 0; table_id < EFX_FARCH_FILTER_TABLE_COUNT; table_id++) {
2777 table = &state->table[table_id];
2779 /* Check whether this is a regular register table */
2780 if (table->step == 0)
2783 for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
2784 if (!test_bit(filter_idx, table->used_bitmap))
2786 efx_farch_filter_build(&filter, &table->spec[filter_idx]);
2787 efx_writeo(efx, &filter,
2788 table->offset + table->step * filter_idx);
2792 efx_farch_filter_push_rx_config(efx);
2793 efx_farch_filter_push_tx_limits(efx);
2795 spin_unlock_bh(&efx->filter_lock);
2798 void efx_farch_filter_table_remove(struct efx_nic *efx)
2800 struct efx_farch_filter_state *state = efx->filter_state;
2801 enum efx_farch_filter_table_id table_id;
2803 for (table_id = 0; table_id < EFX_FARCH_FILTER_TABLE_COUNT; table_id++) {
2804 kfree(state->table[table_id].used_bitmap);
2805 vfree(state->table[table_id].spec);
2810 int efx_farch_filter_table_probe(struct efx_nic *efx)
2812 struct efx_farch_filter_state *state;
2813 struct efx_farch_filter_table *table;
2816 state = kzalloc(sizeof(struct efx_farch_filter_state), GFP_KERNEL);
2819 efx->filter_state = state;
2821 if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
2822 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_IP];
2823 table->id = EFX_FARCH_FILTER_TABLE_RX_IP;
2824 table->offset = FR_BZ_RX_FILTER_TBL0;
2825 table->size = FR_BZ_RX_FILTER_TBL0_ROWS;
2826 table->step = FR_BZ_RX_FILTER_TBL0_STEP;
2829 if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
2830 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_MAC];
2831 table->id = EFX_FARCH_FILTER_TABLE_RX_MAC;
2832 table->offset = FR_CZ_RX_MAC_FILTER_TBL0;
2833 table->size = FR_CZ_RX_MAC_FILTER_TBL0_ROWS;
2834 table->step = FR_CZ_RX_MAC_FILTER_TBL0_STEP;
2836 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_DEF];
2837 table->id = EFX_FARCH_FILTER_TABLE_RX_DEF;
2838 table->size = EFX_FARCH_FILTER_SIZE_RX_DEF;
2840 table = &state->table[EFX_FARCH_FILTER_TABLE_TX_MAC];
2841 table->id = EFX_FARCH_FILTER_TABLE_TX_MAC;
2842 table->offset = FR_CZ_TX_MAC_FILTER_TBL0;
2843 table->size = FR_CZ_TX_MAC_FILTER_TBL0_ROWS;
2844 table->step = FR_CZ_TX_MAC_FILTER_TBL0_STEP;
2847 for (table_id = 0; table_id < EFX_FARCH_FILTER_TABLE_COUNT; table_id++) {
2848 table = &state->table[table_id];
2849 if (table->size == 0)
2851 table->used_bitmap = kcalloc(BITS_TO_LONGS(table->size),
2852 sizeof(unsigned long),
2854 if (!table->used_bitmap)
2856 table->spec = vzalloc(table->size * sizeof(*table->spec));
2861 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_DEF];
2863 /* RX default filters must always exist */
2864 struct efx_farch_filter_spec *spec;
2867 for (i = 0; i < EFX_FARCH_FILTER_SIZE_RX_DEF; i++) {
2868 spec = &table->spec[i];
2869 spec->type = EFX_FARCH_FILTER_UC_DEF + i;
2870 efx_farch_filter_init_rx_auto(efx, spec);
2871 __set_bit(i, table->used_bitmap);
2875 efx_farch_filter_push_rx_config(efx);
2880 efx_farch_filter_table_remove(efx);
2884 /* Update scatter enable flags for filters pointing to our own RX queues */
2885 void efx_farch_filter_update_rx_scatter(struct efx_nic *efx)
2887 struct efx_farch_filter_state *state = efx->filter_state;
2888 enum efx_farch_filter_table_id table_id;
2889 struct efx_farch_filter_table *table;
2891 unsigned int filter_idx;
2893 spin_lock_bh(&efx->filter_lock);
2895 for (table_id = EFX_FARCH_FILTER_TABLE_RX_IP;
2896 table_id <= EFX_FARCH_FILTER_TABLE_RX_DEF;
2898 table = &state->table[table_id];
2900 for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
2901 if (!test_bit(filter_idx, table->used_bitmap) ||
2902 table->spec[filter_idx].dmaq_id >=
2906 if (efx->rx_scatter)
2907 table->spec[filter_idx].flags |=
2908 EFX_FILTER_FLAG_RX_SCATTER;
2910 table->spec[filter_idx].flags &=
2911 ~EFX_FILTER_FLAG_RX_SCATTER;
2913 if (table_id == EFX_FARCH_FILTER_TABLE_RX_DEF)
2914 /* Pushed by efx_farch_filter_push_rx_config() */
2917 efx_farch_filter_build(&filter, &table->spec[filter_idx]);
2918 efx_writeo(efx, &filter,
2919 table->offset + table->step * filter_idx);
2923 efx_farch_filter_push_rx_config(efx);
2925 spin_unlock_bh(&efx->filter_lock);
2928 #ifdef CONFIG_RFS_ACCEL
2930 s32 efx_farch_filter_rfs_insert(struct efx_nic *efx,
2931 struct efx_filter_spec *gen_spec)
2933 return efx_farch_filter_insert(efx, gen_spec, true);
2936 bool efx_farch_filter_rfs_expire_one(struct efx_nic *efx, u32 flow_id,
2939 struct efx_farch_filter_state *state = efx->filter_state;
2940 struct efx_farch_filter_table *table =
2941 &state->table[EFX_FARCH_FILTER_TABLE_RX_IP];
2943 if (test_bit(index, table->used_bitmap) &&
2944 table->spec[index].priority == EFX_FILTER_PRI_HINT &&
2945 rps_may_expire_flow(efx->net_dev, table->spec[index].dmaq_id,
2947 efx_farch_filter_table_clear_entry(efx, table, index);
2954 #endif /* CONFIG_RFS_ACCEL */
2956 void efx_farch_filter_sync_rx_mode(struct efx_nic *efx)
2958 struct net_device *net_dev = efx->net_dev;
2959 struct netdev_hw_addr *ha;
2960 union efx_multicast_hash *mc_hash = &efx->multicast_hash;
2964 if (!efx_dev_registered(efx))
2967 netif_addr_lock_bh(net_dev);
2969 efx->unicast_filter = !(net_dev->flags & IFF_PROMISC);
2971 /* Build multicast hash table */
2972 if (net_dev->flags & (IFF_PROMISC | IFF_ALLMULTI)) {
2973 memset(mc_hash, 0xff, sizeof(*mc_hash));
2975 memset(mc_hash, 0x00, sizeof(*mc_hash));
2976 netdev_for_each_mc_addr(ha, net_dev) {
2977 crc = ether_crc_le(ETH_ALEN, ha->addr);
2978 bit = crc & (EFX_MCAST_HASH_ENTRIES - 1);
2979 __set_bit_le(bit, mc_hash);
2982 /* Broadcast packets go through the multicast hash filter.
2983 * ether_crc_le() of the broadcast address is 0xbe2612ff
2984 * so we always add bit 0xff to the mask.
2986 __set_bit_le(0xff, mc_hash);
2989 netif_addr_unlock_bh(net_dev);