Loading...
1/****************************************************************************
2 * Driver for Solarflare network controllers and boards
3 * Copyright 2005-2006 Fen Systems Ltd.
4 * Copyright 2006-2013 Solarflare Communications Inc.
5 *
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.
9 */
10
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"
19#include "bitfield.h"
20#include "efx.h"
21#include "nic.h"
22#include "farch_regs.h"
23#include "io.h"
24#include "workarounds.h"
25
26/* Falcon-architecture (SFC4000 and SFC9000-family) support */
27
28/**************************************************************************
29 *
30 * Configurable values
31 *
32 **************************************************************************
33 */
34
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
40 * goes idle.
41 */
42#define TX_DC_ENTRIES 16
43#define TX_DC_ENTRIES_ORDER 1
44
45#define RX_DC_ENTRIES 64
46#define RX_DC_ENTRIES_ORDER 3
47
48/* If EFX_MAX_INT_ERRORS internal errors occur within
49 * EFX_INT_ERROR_EXPIRE seconds, we consider the NIC broken and
50 * disable it.
51 */
52#define EFX_INT_ERROR_EXPIRE 3600
53#define EFX_MAX_INT_ERRORS 5
54
55/* Depth of RX flush request fifo */
56#define EFX_RX_FLUSH_COUNT 4
57
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
63
64#define _EFX_CHANNEL_MAGIC(_code, _data) ((_code) << 8 | (_data))
65#define _EFX_CHANNEL_MAGIC_CODE(_magic) ((_magic) >> 8)
66
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, \
77 (_tx_queue)->queue)
78
79static void efx_farch_magic_event(struct efx_channel *channel, u32 magic);
80
81/**************************************************************************
82 *
83 * Hardware access
84 *
85 **************************************************************************/
86
87static inline void efx_write_buf_tbl(struct efx_nic *efx, efx_qword_t *value,
88 unsigned int index)
89{
90 efx_sram_writeq(efx, efx->membase + efx->type->buf_tbl_base,
91 value, index);
92}
93
94static bool efx_masked_compare_oword(const efx_oword_t *a, const efx_oword_t *b,
95 const efx_oword_t *mask)
96{
97 return ((a->u64[0] ^ b->u64[0]) & mask->u64[0]) ||
98 ((a->u64[1] ^ b->u64[1]) & mask->u64[1]);
99}
100
101int efx_farch_test_registers(struct efx_nic *efx,
102 const struct efx_farch_register_test *regs,
103 size_t n_regs)
104{
105 unsigned address = 0, i, j;
106 efx_oword_t mask, imask, original, reg, buf;
107
108 for (i = 0; i < n_regs; ++i) {
109 address = regs[i].address;
110 mask = imask = regs[i].mask;
111 EFX_INVERT_OWORD(imask);
112
113 efx_reado(efx, &original, address);
114
115 /* bit sweep on and off */
116 for (j = 0; j < 128; j++) {
117 if (!EFX_EXTRACT_OWORD32(mask, j, j))
118 continue;
119
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);
123
124 efx_writeo(efx, ®, address);
125 efx_reado(efx, &buf, address);
126
127 if (efx_masked_compare_oword(®, &buf, &mask))
128 goto fail;
129
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);
133
134 efx_writeo(efx, ®, address);
135 efx_reado(efx, &buf, address);
136
137 if (efx_masked_compare_oword(®, &buf, &mask))
138 goto fail;
139 }
140
141 efx_writeo(efx, &original, address);
142 }
143
144 return 0;
145
146fail:
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));
151 return -EIO;
152}
153
154/**************************************************************************
155 *
156 * Special buffer handling
157 * Special buffers are used for event queues and the TX and RX
158 * descriptor rings.
159 *
160 *************************************************************************/
161
162/*
163 * Initialise a special buffer
164 *
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.
168 */
169static void
170efx_init_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
171{
172 efx_qword_t buf_desc;
173 unsigned int index;
174 dma_addr_t dma_addr;
175 int i;
176
177 EFX_BUG_ON_PARANOID(!buffer->buf.addr);
178
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);
191 }
192}
193
194/* Unmaps a buffer and clears the buffer table entries */
195static void
196efx_fini_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
197{
198 efx_oword_t buf_tbl_upd;
199 unsigned int start = buffer->index;
200 unsigned int end = (buffer->index + buffer->entries - 1);
201
202 if (!buffer->entries)
203 return;
204
205 netif_dbg(efx, hw, efx->net_dev, "unmapping special buffers %d-%d\n",
206 buffer->index, buffer->index + buffer->entries - 1);
207
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);
214}
215
216/*
217 * Allocate a new special buffer
218 *
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.
221 *
222 * This call will allocate 4KB buffers, since 8KB buffers can't be
223 * used for event queues and descriptor rings.
224 */
225static int efx_alloc_special_buffer(struct efx_nic *efx,
226 struct efx_special_buffer *buffer,
227 unsigned int len)
228{
229 len = ALIGN(len, EFX_BUF_SIZE);
230
231 if (efx_nic_alloc_buffer(efx, &buffer->buf, len, GFP_KERNEL))
232 return -ENOMEM;
233 buffer->entries = len / EFX_BUF_SIZE;
234 BUG_ON(buffer->buf.dma_addr & (EFX_BUF_SIZE - 1));
235
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);
242#endif
243
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));
250
251 return 0;
252}
253
254static void
255efx_free_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
256{
257 if (!buffer->buf.addr)
258 return;
259
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));
266
267 efx_nic_free_buffer(efx, &buffer->buf);
268 buffer->entries = 0;
269}
270
271/**************************************************************************
272 *
273 * TX path
274 *
275 **************************************************************************/
276
277/* This writes to the TX_DESC_WPTR; write pointer for TX descriptor ring */
278static inline void efx_farch_notify_tx_desc(struct efx_tx_queue *tx_queue)
279{
280 unsigned write_ptr;
281 efx_dword_t reg;
282
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);
287}
288
289/* Write pointer and first descriptor for TX descriptor ring */
290static inline void efx_farch_push_tx_desc(struct efx_tx_queue *tx_queue,
291 const efx_qword_t *txd)
292{
293 unsigned write_ptr;
294 efx_oword_t reg;
295
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);
298
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);
302 reg.qword[0] = *txd;
303 efx_writeo_page(tx_queue->efx, ®,
304 FR_BZ_TX_DESC_UPD_P0, tx_queue->queue);
305}
306
307
308/* For each entry inserted into the software descriptor ring, create a
309 * descriptor in the hardware TX descriptor ring (in host memory), and
310 * write a doorbell.
311 */
312void efx_farch_tx_write(struct efx_tx_queue *tx_queue)
313{
314 struct efx_tx_buffer *buffer;
315 efx_qword_t *txd;
316 unsigned write_ptr;
317 unsigned old_write_count = tx_queue->write_count;
318
319 BUG_ON(tx_queue->write_count == tx_queue->insert_count);
320
321 do {
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;
326
327 EFX_BUG_ON_PARANOID(buffer->flags & EFX_TX_BUF_OPTION);
328
329 /* Create TX descriptor ring entry */
330 BUILD_BUG_ON(EFX_TX_BUF_CONT != 1);
331 EFX_POPULATE_QWORD_4(*txd,
332 FSF_AZ_TX_KER_CONT,
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);
338
339 wmb(); /* Ensure descriptors are written before they are fetched */
340
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);
345 ++tx_queue->pushes;
346 } else {
347 efx_farch_notify_tx_desc(tx_queue);
348 }
349}
350
351/* Allocate hardware resources for a TX queue */
352int efx_farch_tx_probe(struct efx_tx_queue *tx_queue)
353{
354 struct efx_nic *efx = tx_queue->efx;
355 unsigned entries;
356
357 entries = tx_queue->ptr_mask + 1;
358 return efx_alloc_special_buffer(efx, &tx_queue->txd,
359 entries * sizeof(efx_qword_t));
360}
361
362void efx_farch_tx_init(struct efx_tx_queue *tx_queue)
363{
364 struct efx_nic *efx = tx_queue->efx;
365 efx_oword_t reg;
366
367 /* Pin TX descriptor ring */
368 efx_init_special_buffer(efx, &tx_queue->txd);
369
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);
384
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,
389 !csum);
390 }
391
392 efx_writeo_table(efx, ®, efx->type->txd_ptr_tbl_base,
393 tx_queue->queue);
394
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);
398
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, ®);
402 else
403 __set_bit_le(tx_queue->queue, ®);
404 efx_writeo(efx, ®, FR_AA_TX_CHKSM_CFG);
405 }
406
407 if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
408 EFX_POPULATE_OWORD_1(reg,
409 FRF_BZ_TX_PACE,
410 (tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
411 FFE_BZ_TX_PACE_OFF :
412 FFE_BZ_TX_PACE_RESERVED);
413 efx_writeo_table(efx, ®, FR_BZ_TX_PACE_TBL,
414 tx_queue->queue);
415 }
416}
417
418static void efx_farch_flush_tx_queue(struct efx_tx_queue *tx_queue)
419{
420 struct efx_nic *efx = tx_queue->efx;
421 efx_oword_t tx_flush_descq;
422
423 WARN_ON(atomic_read(&tx_queue->flush_outstanding));
424 atomic_set(&tx_queue->flush_outstanding, 1);
425
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);
430}
431
432void efx_farch_tx_fini(struct efx_tx_queue *tx_queue)
433{
434 struct efx_nic *efx = tx_queue->efx;
435 efx_oword_t tx_desc_ptr;
436
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,
440 tx_queue->queue);
441
442 /* Unpin TX descriptor ring */
443 efx_fini_special_buffer(efx, &tx_queue->txd);
444}
445
446/* Free buffers backing TX queue */
447void efx_farch_tx_remove(struct efx_tx_queue *tx_queue)
448{
449 efx_free_special_buffer(tx_queue->efx, &tx_queue->txd);
450}
451
452/**************************************************************************
453 *
454 * RX path
455 *
456 **************************************************************************/
457
458/* This creates an entry in the RX descriptor queue */
459static inline void
460efx_farch_build_rx_desc(struct efx_rx_queue *rx_queue, unsigned index)
461{
462 struct efx_rx_buffer *rx_buf;
463 efx_qword_t *rxd;
464
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,
469 rx_buf->len -
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);
473}
474
475/* This writes to the RX_DESC_WPTR register for the specified receive
476 * descriptor ring.
477 */
478void efx_farch_rx_write(struct efx_rx_queue *rx_queue)
479{
480 struct efx_nic *efx = rx_queue->efx;
481 efx_dword_t reg;
482 unsigned write_ptr;
483
484 while (rx_queue->notified_count != rx_queue->added_count) {
485 efx_farch_build_rx_desc(
486 rx_queue,
487 rx_queue->notified_count & rx_queue->ptr_mask);
488 ++rx_queue->notified_count;
489 }
490
491 wmb();
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));
496}
497
498int efx_farch_rx_probe(struct efx_rx_queue *rx_queue)
499{
500 struct efx_nic *efx = rx_queue->efx;
501 unsigned entries;
502
503 entries = rx_queue->ptr_mask + 1;
504 return efx_alloc_special_buffer(efx, &rx_queue->rxd,
505 entries * sizeof(efx_qword_t));
506}
507
508void efx_farch_rx_init(struct efx_rx_queue *rx_queue)
509{
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;
514 bool jumbo_en;
515
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
519 * scatter.
520 */
521 jumbo_en = !is_b0 || efx->rx_scatter;
522
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);
527
528 rx_queue->scatter_n = 0;
529
530 /* Pin RX descriptor ring */
531 efx_init_special_buffer(efx, &rx_queue->rxd);
532
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));
550}
551
552static void efx_farch_flush_rx_queue(struct efx_rx_queue *rx_queue)
553{
554 struct efx_nic *efx = rx_queue->efx;
555 efx_oword_t rx_flush_descq;
556
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);
562}
563
564void efx_farch_rx_fini(struct efx_rx_queue *rx_queue)
565{
566 efx_oword_t rx_desc_ptr;
567 struct efx_nic *efx = rx_queue->efx;
568
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));
573
574 /* Unpin RX descriptor ring */
575 efx_fini_special_buffer(efx, &rx_queue->rxd);
576}
577
578/* Free buffers backing RX queue */
579void efx_farch_rx_remove(struct efx_rx_queue *rx_queue)
580{
581 efx_free_special_buffer(rx_queue->efx, &rx_queue->rxd);
582}
583
584/**************************************************************************
585 *
586 * Flush handling
587 *
588 **************************************************************************/
589
590/* efx_farch_flush_queues() must be woken up when all flushes are completed,
591 * or more RX flushes can be kicked off.
592 */
593static bool efx_farch_flush_wake(struct efx_nic *efx)
594{
595 /* Ensure that all updates are visible to efx_farch_flush_queues() */
596 smp_mb();
597
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));
601}
602
603static bool efx_check_tx_flush_complete(struct efx_nic *efx)
604{
605 bool i = true;
606 efx_oword_t txd_ptr_tbl;
607 struct efx_channel *channel;
608 struct efx_tx_queue *tx_queue;
609
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",
620 tx_queue->queue);
621 i = false;
622 } else if (atomic_cmpxchg(&tx_queue->flush_outstanding,
623 1, 0)) {
624 /* The flush is complete, but we didn't
625 * receive a flush completion event
626 */
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
633 */
634 efx_farch_magic_event(channel,
635 EFX_CHANNEL_MAGIC_TX_DRAIN(
636 tx_queue));
637 }
638 }
639 }
640
641 return i;
642}
643
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. */
647static int efx_farch_do_flush(struct efx_nic *efx)
648{
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;
653 int rc = 0;
654
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);
658 }
659 efx_for_each_channel_rx_queue(rx_queue, channel) {
660 rx_queue->flush_pending = true;
661 atomic_inc(&efx->rxq_flush_pending);
662 }
663 }
664
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.
669 */
670 if (efx_sriov_enabled(efx)) {
671 rc = efx_mcdi_flush_rxqs(efx);
672 if (!rc)
673 goto wait;
674 }
675
676 /* The hardware supports four concurrent rx flushes, each of
677 * which may need to be retried if there is an outstanding
678 * descriptor fetch
679 */
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) >=
683 EFX_RX_FLUSH_COUNT)
684 break;
685
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);
691 }
692 }
693 }
694
695 wait:
696 timeout = wait_event_timeout(efx->flush_wq,
697 efx_farch_flush_wake(efx),
698 timeout);
699 }
700
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));
707 rc = -ETIMEDOUT;
708
709 atomic_set(&efx->active_queues, 0);
710 atomic_set(&efx->rxq_flush_pending, 0);
711 atomic_set(&efx->rxq_flush_outstanding, 0);
712 }
713
714 return rc;
715}
716
717int efx_farch_fini_dmaq(struct efx_nic *efx)
718{
719 struct efx_channel *channel;
720 struct efx_tx_queue *tx_queue;
721 struct efx_rx_queue *rx_queue;
722 int rc = 0;
723
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);
731 }
732
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);
738 }
739 }
740
741 return rc;
742}
743
744/* Reset queue and flush accounting after FLR
745 *
746 * One possible cause of FLR recovery is that DMA may be failing (eg. if bus
747 * mastering was disabled), in which case we don't receive (RXQ) flush
748 * completion events. This means that efx->rxq_flush_outstanding remained at 4
749 * after the FLR; also, efx->active_queues was non-zero (as no flush completion
750 * events were received, and we didn't go through efx_check_tx_flush_complete())
751 * If we don't fix this up, on the next call to efx_realloc_channels() we won't
752 * flush any RX queues because efx->rxq_flush_outstanding is at the limit of 4
753 * for batched flush requests; and the efx->active_queues gets messed up because
754 * we keep incrementing for the newly initialised queues, but it never went to
755 * zero previously. Then we get a timeout every time we try to restart the
756 * queues, as it doesn't go back to zero when we should be flushing the queues.
757 */
758void efx_farch_finish_flr(struct efx_nic *efx)
759{
760 atomic_set(&efx->rxq_flush_pending, 0);
761 atomic_set(&efx->rxq_flush_outstanding, 0);
762 atomic_set(&efx->active_queues, 0);
763}
764
765
766/**************************************************************************
767 *
768 * Event queue processing
769 * Event queues are processed by per-channel tasklets.
770 *
771 **************************************************************************/
772
773/* Update a channel's event queue's read pointer (RPTR) register
774 *
775 * This writes the EVQ_RPTR_REG register for the specified channel's
776 * event queue.
777 */
778void efx_farch_ev_read_ack(struct efx_channel *channel)
779{
780 efx_dword_t reg;
781 struct efx_nic *efx = channel->efx;
782
783 EFX_POPULATE_DWORD_1(reg, FRF_AZ_EVQ_RPTR,
784 channel->eventq_read_ptr & channel->eventq_mask);
785
786 /* For Falcon A1, EVQ_RPTR_KER is documented as having a step size
787 * of 4 bytes, but it is really 16 bytes just like later revisions.
788 */
789 efx_writed(efx, ®,
790 efx->type->evq_rptr_tbl_base +
791 FR_BZ_EVQ_RPTR_STEP * channel->channel);
792}
793
794/* Use HW to insert a SW defined event */
795void efx_farch_generate_event(struct efx_nic *efx, unsigned int evq,
796 efx_qword_t *event)
797{
798 efx_oword_t drv_ev_reg;
799
800 BUILD_BUG_ON(FRF_AZ_DRV_EV_DATA_LBN != 0 ||
801 FRF_AZ_DRV_EV_DATA_WIDTH != 64);
802 drv_ev_reg.u32[0] = event->u32[0];
803 drv_ev_reg.u32[1] = event->u32[1];
804 drv_ev_reg.u32[2] = 0;
805 drv_ev_reg.u32[3] = 0;
806 EFX_SET_OWORD_FIELD(drv_ev_reg, FRF_AZ_DRV_EV_QID, evq);
807 efx_writeo(efx, &drv_ev_reg, FR_AZ_DRV_EV);
808}
809
810static void efx_farch_magic_event(struct efx_channel *channel, u32 magic)
811{
812 efx_qword_t event;
813
814 EFX_POPULATE_QWORD_2(event, FSF_AZ_EV_CODE,
815 FSE_AZ_EV_CODE_DRV_GEN_EV,
816 FSF_AZ_DRV_GEN_EV_MAGIC, magic);
817 efx_farch_generate_event(channel->efx, channel->channel, &event);
818}
819
820/* Handle a transmit completion event
821 *
822 * The NIC batches TX completion events; the message we receive is of
823 * the form "complete all TX events up to this index".
824 */
825static int
826efx_farch_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
827{
828 unsigned int tx_ev_desc_ptr;
829 unsigned int tx_ev_q_label;
830 struct efx_tx_queue *tx_queue;
831 struct efx_nic *efx = channel->efx;
832 int tx_packets = 0;
833
834 if (unlikely(ACCESS_ONCE(efx->reset_pending)))
835 return 0;
836
837 if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
838 /* Transmit completion */
839 tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
840 tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
841 tx_queue = efx_channel_get_tx_queue(
842 channel, tx_ev_q_label % EFX_TXQ_TYPES);
843 tx_packets = ((tx_ev_desc_ptr - tx_queue->read_count) &
844 tx_queue->ptr_mask);
845 efx_xmit_done(tx_queue, tx_ev_desc_ptr);
846 } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
847 /* Rewrite the FIFO write pointer */
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
852 netif_tx_lock(efx->net_dev);
853 efx_farch_notify_tx_desc(tx_queue);
854 netif_tx_unlock(efx->net_dev);
855 } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR)) {
856 efx_schedule_reset(efx, RESET_TYPE_DMA_ERROR);
857 } else {
858 netif_err(efx, tx_err, efx->net_dev,
859 "channel %d unexpected TX event "
860 EFX_QWORD_FMT"\n", channel->channel,
861 EFX_QWORD_VAL(*event));
862 }
863
864 return tx_packets;
865}
866
867/* Detect errors included in the rx_evt_pkt_ok bit. */
868static u16 efx_farch_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
869 const efx_qword_t *event)
870{
871 struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
872 struct efx_nic *efx = rx_queue->efx;
873 bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
874 bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
875 bool rx_ev_frm_trunc, rx_ev_drib_nib, rx_ev_tobe_disc;
876 bool rx_ev_other_err, rx_ev_pause_frm;
877 bool rx_ev_hdr_type, rx_ev_mcast_pkt;
878 unsigned rx_ev_pkt_type;
879
880 rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
881 rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
882 rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
883 rx_ev_pkt_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_TYPE);
884 rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
885 FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
886 rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
887 FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
888 rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
889 FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
890 rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
891 rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
892 rx_ev_drib_nib = ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) ?
893 0 : EFX_QWORD_FIELD(*event, FSF_AA_RX_EV_DRIB_NIB));
894 rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
895
896 /* Every error apart from tobe_disc and pause_frm */
897 rx_ev_other_err = (rx_ev_drib_nib | rx_ev_tcp_udp_chksum_err |
898 rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
899 rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
900
901 /* Count errors that are not in MAC stats. Ignore expected
902 * checksum errors during self-test. */
903 if (rx_ev_frm_trunc)
904 ++channel->n_rx_frm_trunc;
905 else if (rx_ev_tobe_disc)
906 ++channel->n_rx_tobe_disc;
907 else if (!efx->loopback_selftest) {
908 if (rx_ev_ip_hdr_chksum_err)
909 ++channel->n_rx_ip_hdr_chksum_err;
910 else if (rx_ev_tcp_udp_chksum_err)
911 ++channel->n_rx_tcp_udp_chksum_err;
912 }
913
914 /* TOBE_DISC is expected on unicast mismatches; don't print out an
915 * error message. FRM_TRUNC indicates RXDP dropped the packet due
916 * to a FIFO overflow.
917 */
918#ifdef DEBUG
919 if (rx_ev_other_err && net_ratelimit()) {
920 netif_dbg(efx, rx_err, efx->net_dev,
921 " RX queue %d unexpected RX event "
922 EFX_QWORD_FMT "%s%s%s%s%s%s%s%s\n",
923 efx_rx_queue_index(rx_queue), EFX_QWORD_VAL(*event),
924 rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
925 rx_ev_ip_hdr_chksum_err ?
926 " [IP_HDR_CHKSUM_ERR]" : "",
927 rx_ev_tcp_udp_chksum_err ?
928 " [TCP_UDP_CHKSUM_ERR]" : "",
929 rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
930 rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
931 rx_ev_drib_nib ? " [DRIB_NIB]" : "",
932 rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
933 rx_ev_pause_frm ? " [PAUSE]" : "");
934 }
935#endif
936
937 /* The frame must be discarded if any of these are true. */
938 return (rx_ev_eth_crc_err | rx_ev_frm_trunc | rx_ev_drib_nib |
939 rx_ev_tobe_disc | rx_ev_pause_frm) ?
940 EFX_RX_PKT_DISCARD : 0;
941}
942
943/* Handle receive events that are not in-order. Return true if this
944 * can be handled as a partial packet discard, false if it's more
945 * serious.
946 */
947static bool
948efx_farch_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
949{
950 struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
951 struct efx_nic *efx = rx_queue->efx;
952 unsigned expected, dropped;
953
954 if (rx_queue->scatter_n &&
955 index == ((rx_queue->removed_count + rx_queue->scatter_n - 1) &
956 rx_queue->ptr_mask)) {
957 ++channel->n_rx_nodesc_trunc;
958 return true;
959 }
960
961 expected = rx_queue->removed_count & rx_queue->ptr_mask;
962 dropped = (index - expected) & rx_queue->ptr_mask;
963 netif_info(efx, rx_err, efx->net_dev,
964 "dropped %d events (index=%d expected=%d)\n",
965 dropped, index, expected);
966
967 efx_schedule_reset(efx, EFX_WORKAROUND_5676(efx) ?
968 RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
969 return false;
970}
971
972/* Handle a packet received event
973 *
974 * The NIC gives a "discard" flag if it's a unicast packet with the
975 * wrong destination address
976 * Also "is multicast" and "matches multicast filter" flags can be used to
977 * discard non-matching multicast packets.
978 */
979static void
980efx_farch_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
981{
982 unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
983 unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
984 unsigned expected_ptr;
985 bool rx_ev_pkt_ok, rx_ev_sop, rx_ev_cont;
986 u16 flags;
987 struct efx_rx_queue *rx_queue;
988 struct efx_nic *efx = channel->efx;
989
990 if (unlikely(ACCESS_ONCE(efx->reset_pending)))
991 return;
992
993 rx_ev_cont = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT);
994 rx_ev_sop = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP);
995 WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
996 channel->channel);
997
998 rx_queue = efx_channel_get_rx_queue(channel);
999
1000 rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
1001 expected_ptr = ((rx_queue->removed_count + rx_queue->scatter_n) &
1002 rx_queue->ptr_mask);
1003
1004 /* Check for partial drops and other errors */
1005 if (unlikely(rx_ev_desc_ptr != expected_ptr) ||
1006 unlikely(rx_ev_sop != (rx_queue->scatter_n == 0))) {
1007 if (rx_ev_desc_ptr != expected_ptr &&
1008 !efx_farch_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr))
1009 return;
1010
1011 /* Discard all pending fragments */
1012 if (rx_queue->scatter_n) {
1013 efx_rx_packet(
1014 rx_queue,
1015 rx_queue->removed_count & rx_queue->ptr_mask,
1016 rx_queue->scatter_n, 0, EFX_RX_PKT_DISCARD);
1017 rx_queue->removed_count += rx_queue->scatter_n;
1018 rx_queue->scatter_n = 0;
1019 }
1020
1021 /* Return if there is no new fragment */
1022 if (rx_ev_desc_ptr != expected_ptr)
1023 return;
1024
1025 /* Discard new fragment if not SOP */
1026 if (!rx_ev_sop) {
1027 efx_rx_packet(
1028 rx_queue,
1029 rx_queue->removed_count & rx_queue->ptr_mask,
1030 1, 0, EFX_RX_PKT_DISCARD);
1031 ++rx_queue->removed_count;
1032 return;
1033 }
1034 }
1035
1036 ++rx_queue->scatter_n;
1037 if (rx_ev_cont)
1038 return;
1039
1040 rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
1041 rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
1042 rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
1043
1044 if (likely(rx_ev_pkt_ok)) {
1045 /* If packet is marked as OK then we can rely on the
1046 * hardware checksum and classification.
1047 */
1048 flags = 0;
1049 switch (rx_ev_hdr_type) {
1050 case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP:
1051 flags |= EFX_RX_PKT_TCP;
1052 /* fall through */
1053 case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP:
1054 flags |= EFX_RX_PKT_CSUMMED;
1055 /* fall through */
1056 case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_OTHER:
1057 case FSE_AZ_RX_EV_HDR_TYPE_OTHER:
1058 break;
1059 }
1060 } else {
1061 flags = efx_farch_handle_rx_not_ok(rx_queue, event);
1062 }
1063
1064 /* Detect multicast packets that didn't match the filter */
1065 rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
1066 if (rx_ev_mcast_pkt) {
1067 unsigned int rx_ev_mcast_hash_match =
1068 EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
1069
1070 if (unlikely(!rx_ev_mcast_hash_match)) {
1071 ++channel->n_rx_mcast_mismatch;
1072 flags |= EFX_RX_PKT_DISCARD;
1073 }
1074 }
1075
1076 channel->irq_mod_score += 2;
1077
1078 /* Handle received packet */
1079 efx_rx_packet(rx_queue,
1080 rx_queue->removed_count & rx_queue->ptr_mask,
1081 rx_queue->scatter_n, rx_ev_byte_cnt, flags);
1082 rx_queue->removed_count += rx_queue->scatter_n;
1083 rx_queue->scatter_n = 0;
1084}
1085
1086/* If this flush done event corresponds to a &struct efx_tx_queue, then
1087 * send an %EFX_CHANNEL_MAGIC_TX_DRAIN event to drain the event queue
1088 * of all transmit completions.
1089 */
1090static void
1091efx_farch_handle_tx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1092{
1093 struct efx_tx_queue *tx_queue;
1094 int qid;
1095
1096 qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1097 if (qid < EFX_TXQ_TYPES * efx->n_tx_channels) {
1098 tx_queue = efx_get_tx_queue(efx, qid / EFX_TXQ_TYPES,
1099 qid % EFX_TXQ_TYPES);
1100 if (atomic_cmpxchg(&tx_queue->flush_outstanding, 1, 0)) {
1101 efx_farch_magic_event(tx_queue->channel,
1102 EFX_CHANNEL_MAGIC_TX_DRAIN(tx_queue));
1103 }
1104 }
1105}
1106
1107/* If this flush done event corresponds to a &struct efx_rx_queue: If the flush
1108 * was succesful then send an %EFX_CHANNEL_MAGIC_RX_DRAIN, otherwise add
1109 * the RX queue back to the mask of RX queues in need of flushing.
1110 */
1111static void
1112efx_farch_handle_rx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1113{
1114 struct efx_channel *channel;
1115 struct efx_rx_queue *rx_queue;
1116 int qid;
1117 bool failed;
1118
1119 qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
1120 failed = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
1121 if (qid >= efx->n_channels)
1122 return;
1123 channel = efx_get_channel(efx, qid);
1124 if (!efx_channel_has_rx_queue(channel))
1125 return;
1126 rx_queue = efx_channel_get_rx_queue(channel);
1127
1128 if (failed) {
1129 netif_info(efx, hw, efx->net_dev,
1130 "RXQ %d flush retry\n", qid);
1131 rx_queue->flush_pending = true;
1132 atomic_inc(&efx->rxq_flush_pending);
1133 } else {
1134 efx_farch_magic_event(efx_rx_queue_channel(rx_queue),
1135 EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue));
1136 }
1137 atomic_dec(&efx->rxq_flush_outstanding);
1138 if (efx_farch_flush_wake(efx))
1139 wake_up(&efx->flush_wq);
1140}
1141
1142static void
1143efx_farch_handle_drain_event(struct efx_channel *channel)
1144{
1145 struct efx_nic *efx = channel->efx;
1146
1147 WARN_ON(atomic_read(&efx->active_queues) == 0);
1148 atomic_dec(&efx->active_queues);
1149 if (efx_farch_flush_wake(efx))
1150 wake_up(&efx->flush_wq);
1151}
1152
1153static void efx_farch_handle_generated_event(struct efx_channel *channel,
1154 efx_qword_t *event)
1155{
1156 struct efx_nic *efx = channel->efx;
1157 struct efx_rx_queue *rx_queue =
1158 efx_channel_has_rx_queue(channel) ?
1159 efx_channel_get_rx_queue(channel) : NULL;
1160 unsigned magic, code;
1161
1162 magic = EFX_QWORD_FIELD(*event, FSF_AZ_DRV_GEN_EV_MAGIC);
1163 code = _EFX_CHANNEL_MAGIC_CODE(magic);
1164
1165 if (magic == EFX_CHANNEL_MAGIC_TEST(channel)) {
1166 channel->event_test_cpu = raw_smp_processor_id();
1167 } else if (rx_queue && magic == EFX_CHANNEL_MAGIC_FILL(rx_queue)) {
1168 /* The queue must be empty, so we won't receive any rx
1169 * events, so efx_process_channel() won't refill the
1170 * queue. Refill it here */
1171 efx_fast_push_rx_descriptors(rx_queue, true);
1172 } else if (rx_queue && magic == EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue)) {
1173 efx_farch_handle_drain_event(channel);
1174 } else if (code == _EFX_CHANNEL_MAGIC_TX_DRAIN) {
1175 efx_farch_handle_drain_event(channel);
1176 } else {
1177 netif_dbg(efx, hw, efx->net_dev, "channel %d received "
1178 "generated event "EFX_QWORD_FMT"\n",
1179 channel->channel, EFX_QWORD_VAL(*event));
1180 }
1181}
1182
1183static void
1184efx_farch_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
1185{
1186 struct efx_nic *efx = channel->efx;
1187 unsigned int ev_sub_code;
1188 unsigned int ev_sub_data;
1189
1190 ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
1191 ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1192
1193 switch (ev_sub_code) {
1194 case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
1195 netif_vdbg(efx, hw, efx->net_dev, "channel %d TXQ %d flushed\n",
1196 channel->channel, ev_sub_data);
1197 efx_farch_handle_tx_flush_done(efx, event);
1198 efx_sriov_tx_flush_done(efx, event);
1199 break;
1200 case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
1201 netif_vdbg(efx, hw, efx->net_dev, "channel %d RXQ %d flushed\n",
1202 channel->channel, ev_sub_data);
1203 efx_farch_handle_rx_flush_done(efx, event);
1204 efx_sriov_rx_flush_done(efx, event);
1205 break;
1206 case FSE_AZ_EVQ_INIT_DONE_EV:
1207 netif_dbg(efx, hw, efx->net_dev,
1208 "channel %d EVQ %d initialised\n",
1209 channel->channel, ev_sub_data);
1210 break;
1211 case FSE_AZ_SRM_UPD_DONE_EV:
1212 netif_vdbg(efx, hw, efx->net_dev,
1213 "channel %d SRAM update done\n", channel->channel);
1214 break;
1215 case FSE_AZ_WAKE_UP_EV:
1216 netif_vdbg(efx, hw, efx->net_dev,
1217 "channel %d RXQ %d wakeup event\n",
1218 channel->channel, ev_sub_data);
1219 break;
1220 case FSE_AZ_TIMER_EV:
1221 netif_vdbg(efx, hw, efx->net_dev,
1222 "channel %d RX queue %d timer expired\n",
1223 channel->channel, ev_sub_data);
1224 break;
1225 case FSE_AA_RX_RECOVER_EV:
1226 netif_err(efx, rx_err, efx->net_dev,
1227 "channel %d seen DRIVER RX_RESET event. "
1228 "Resetting.\n", channel->channel);
1229 atomic_inc(&efx->rx_reset);
1230 efx_schedule_reset(efx,
1231 EFX_WORKAROUND_6555(efx) ?
1232 RESET_TYPE_RX_RECOVERY :
1233 RESET_TYPE_DISABLE);
1234 break;
1235 case FSE_BZ_RX_DSC_ERROR_EV:
1236 if (ev_sub_data < EFX_VI_BASE) {
1237 netif_err(efx, rx_err, efx->net_dev,
1238 "RX DMA Q %d reports descriptor fetch error."
1239 " RX Q %d is disabled.\n", ev_sub_data,
1240 ev_sub_data);
1241 efx_schedule_reset(efx, RESET_TYPE_DMA_ERROR);
1242 } else
1243 efx_sriov_desc_fetch_err(efx, ev_sub_data);
1244 break;
1245 case FSE_BZ_TX_DSC_ERROR_EV:
1246 if (ev_sub_data < EFX_VI_BASE) {
1247 netif_err(efx, tx_err, efx->net_dev,
1248 "TX DMA Q %d reports descriptor fetch error."
1249 " TX Q %d is disabled.\n", ev_sub_data,
1250 ev_sub_data);
1251 efx_schedule_reset(efx, RESET_TYPE_DMA_ERROR);
1252 } else
1253 efx_sriov_desc_fetch_err(efx, ev_sub_data);
1254 break;
1255 default:
1256 netif_vdbg(efx, hw, efx->net_dev,
1257 "channel %d unknown driver event code %d "
1258 "data %04x\n", channel->channel, ev_sub_code,
1259 ev_sub_data);
1260 break;
1261 }
1262}
1263
1264int efx_farch_ev_process(struct efx_channel *channel, int budget)
1265{
1266 struct efx_nic *efx = channel->efx;
1267 unsigned int read_ptr;
1268 efx_qword_t event, *p_event;
1269 int ev_code;
1270 int tx_packets = 0;
1271 int spent = 0;
1272
1273 if (budget <= 0)
1274 return spent;
1275
1276 read_ptr = channel->eventq_read_ptr;
1277
1278 for (;;) {
1279 p_event = efx_event(channel, read_ptr);
1280 event = *p_event;
1281
1282 if (!efx_event_present(&event))
1283 /* End of events */
1284 break;
1285
1286 netif_vdbg(channel->efx, intr, channel->efx->net_dev,
1287 "channel %d event is "EFX_QWORD_FMT"\n",
1288 channel->channel, EFX_QWORD_VAL(event));
1289
1290 /* Clear this event by marking it all ones */
1291 EFX_SET_QWORD(*p_event);
1292
1293 ++read_ptr;
1294
1295 ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
1296
1297 switch (ev_code) {
1298 case FSE_AZ_EV_CODE_RX_EV:
1299 efx_farch_handle_rx_event(channel, &event);
1300 if (++spent == budget)
1301 goto out;
1302 break;
1303 case FSE_AZ_EV_CODE_TX_EV:
1304 tx_packets += efx_farch_handle_tx_event(channel,
1305 &event);
1306 if (tx_packets > efx->txq_entries) {
1307 spent = budget;
1308 goto out;
1309 }
1310 break;
1311 case FSE_AZ_EV_CODE_DRV_GEN_EV:
1312 efx_farch_handle_generated_event(channel, &event);
1313 break;
1314 case FSE_AZ_EV_CODE_DRIVER_EV:
1315 efx_farch_handle_driver_event(channel, &event);
1316 break;
1317 case FSE_CZ_EV_CODE_USER_EV:
1318 efx_sriov_event(channel, &event);
1319 break;
1320 case FSE_CZ_EV_CODE_MCDI_EV:
1321 efx_mcdi_process_event(channel, &event);
1322 break;
1323 case FSE_AZ_EV_CODE_GLOBAL_EV:
1324 if (efx->type->handle_global_event &&
1325 efx->type->handle_global_event(channel, &event))
1326 break;
1327 /* else fall through */
1328 default:
1329 netif_err(channel->efx, hw, channel->efx->net_dev,
1330 "channel %d unknown event type %d (data "
1331 EFX_QWORD_FMT ")\n", channel->channel,
1332 ev_code, EFX_QWORD_VAL(event));
1333 }
1334 }
1335
1336out:
1337 channel->eventq_read_ptr = read_ptr;
1338 return spent;
1339}
1340
1341/* Allocate buffer table entries for event queue */
1342int efx_farch_ev_probe(struct efx_channel *channel)
1343{
1344 struct efx_nic *efx = channel->efx;
1345 unsigned entries;
1346
1347 entries = channel->eventq_mask + 1;
1348 return efx_alloc_special_buffer(efx, &channel->eventq,
1349 entries * sizeof(efx_qword_t));
1350}
1351
1352int efx_farch_ev_init(struct efx_channel *channel)
1353{
1354 efx_oword_t reg;
1355 struct efx_nic *efx = channel->efx;
1356
1357 netif_dbg(efx, hw, efx->net_dev,
1358 "channel %d event queue in special buffers %d-%d\n",
1359 channel->channel, channel->eventq.index,
1360 channel->eventq.index + channel->eventq.entries - 1);
1361
1362 if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1363 EFX_POPULATE_OWORD_3(reg,
1364 FRF_CZ_TIMER_Q_EN, 1,
1365 FRF_CZ_HOST_NOTIFY_MODE, 0,
1366 FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
1367 efx_writeo_table(efx, ®, FR_BZ_TIMER_TBL, channel->channel);
1368 }
1369
1370 /* Pin event queue buffer */
1371 efx_init_special_buffer(efx, &channel->eventq);
1372
1373 /* Fill event queue with all ones (i.e. empty events) */
1374 memset(channel->eventq.buf.addr, 0xff, channel->eventq.buf.len);
1375
1376 /* Push event queue to card */
1377 EFX_POPULATE_OWORD_3(reg,
1378 FRF_AZ_EVQ_EN, 1,
1379 FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
1380 FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
1381 efx_writeo_table(efx, ®, efx->type->evq_ptr_tbl_base,
1382 channel->channel);
1383
1384 return 0;
1385}
1386
1387void efx_farch_ev_fini(struct efx_channel *channel)
1388{
1389 efx_oword_t reg;
1390 struct efx_nic *efx = channel->efx;
1391
1392 /* Remove event queue from card */
1393 EFX_ZERO_OWORD(reg);
1394 efx_writeo_table(efx, ®, efx->type->evq_ptr_tbl_base,
1395 channel->channel);
1396 if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1397 efx_writeo_table(efx, ®, FR_BZ_TIMER_TBL, channel->channel);
1398
1399 /* Unpin event queue */
1400 efx_fini_special_buffer(efx, &channel->eventq);
1401}
1402
1403/* Free buffers backing event queue */
1404void efx_farch_ev_remove(struct efx_channel *channel)
1405{
1406 efx_free_special_buffer(channel->efx, &channel->eventq);
1407}
1408
1409
1410void efx_farch_ev_test_generate(struct efx_channel *channel)
1411{
1412 efx_farch_magic_event(channel, EFX_CHANNEL_MAGIC_TEST(channel));
1413}
1414
1415void efx_farch_rx_defer_refill(struct efx_rx_queue *rx_queue)
1416{
1417 efx_farch_magic_event(efx_rx_queue_channel(rx_queue),
1418 EFX_CHANNEL_MAGIC_FILL(rx_queue));
1419}
1420
1421/**************************************************************************
1422 *
1423 * Hardware interrupts
1424 * The hardware interrupt handler does very little work; all the event
1425 * queue processing is carried out by per-channel tasklets.
1426 *
1427 **************************************************************************/
1428
1429/* Enable/disable/generate interrupts */
1430static inline void efx_farch_interrupts(struct efx_nic *efx,
1431 bool enabled, bool force)
1432{
1433 efx_oword_t int_en_reg_ker;
1434
1435 EFX_POPULATE_OWORD_3(int_en_reg_ker,
1436 FRF_AZ_KER_INT_LEVE_SEL, efx->irq_level,
1437 FRF_AZ_KER_INT_KER, force,
1438 FRF_AZ_DRV_INT_EN_KER, enabled);
1439 efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
1440}
1441
1442void efx_farch_irq_enable_master(struct efx_nic *efx)
1443{
1444 EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
1445 wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
1446
1447 efx_farch_interrupts(efx, true, false);
1448}
1449
1450void efx_farch_irq_disable_master(struct efx_nic *efx)
1451{
1452 /* Disable interrupts */
1453 efx_farch_interrupts(efx, false, false);
1454}
1455
1456/* Generate a test interrupt
1457 * Interrupt must already have been enabled, otherwise nasty things
1458 * may happen.
1459 */
1460void efx_farch_irq_test_generate(struct efx_nic *efx)
1461{
1462 efx_farch_interrupts(efx, true, true);
1463}
1464
1465/* Process a fatal interrupt
1466 * Disable bus mastering ASAP and schedule a reset
1467 */
1468irqreturn_t efx_farch_fatal_interrupt(struct efx_nic *efx)
1469{
1470 struct falcon_nic_data *nic_data = efx->nic_data;
1471 efx_oword_t *int_ker = efx->irq_status.addr;
1472 efx_oword_t fatal_intr;
1473 int error, mem_perr;
1474
1475 efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
1476 error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
1477
1478 netif_err(efx, hw, efx->net_dev, "SYSTEM ERROR "EFX_OWORD_FMT" status "
1479 EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
1480 EFX_OWORD_VAL(fatal_intr),
1481 error ? "disabling bus mastering" : "no recognised error");
1482
1483 /* If this is a memory parity error dump which blocks are offending */
1484 mem_perr = (EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER) ||
1485 EFX_OWORD_FIELD(fatal_intr, FRF_AZ_SRM_PERR_INT_KER));
1486 if (mem_perr) {
1487 efx_oword_t reg;
1488 efx_reado(efx, ®, FR_AZ_MEM_STAT);
1489 netif_err(efx, hw, efx->net_dev,
1490 "SYSTEM ERROR: memory parity error "EFX_OWORD_FMT"\n",
1491 EFX_OWORD_VAL(reg));
1492 }
1493
1494 /* Disable both devices */
1495 pci_clear_master(efx->pci_dev);
1496 if (efx_nic_is_dual_func(efx))
1497 pci_clear_master(nic_data->pci_dev2);
1498 efx_farch_irq_disable_master(efx);
1499
1500 /* Count errors and reset or disable the NIC accordingly */
1501 if (efx->int_error_count == 0 ||
1502 time_after(jiffies, efx->int_error_expire)) {
1503 efx->int_error_count = 0;
1504 efx->int_error_expire =
1505 jiffies + EFX_INT_ERROR_EXPIRE * HZ;
1506 }
1507 if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
1508 netif_err(efx, hw, efx->net_dev,
1509 "SYSTEM ERROR - reset scheduled\n");
1510 efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
1511 } else {
1512 netif_err(efx, hw, efx->net_dev,
1513 "SYSTEM ERROR - max number of errors seen."
1514 "NIC will be disabled\n");
1515 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1516 }
1517
1518 return IRQ_HANDLED;
1519}
1520
1521/* Handle a legacy interrupt
1522 * Acknowledges the interrupt and schedule event queue processing.
1523 */
1524irqreturn_t efx_farch_legacy_interrupt(int irq, void *dev_id)
1525{
1526 struct efx_nic *efx = dev_id;
1527 bool soft_enabled = ACCESS_ONCE(efx->irq_soft_enabled);
1528 efx_oword_t *int_ker = efx->irq_status.addr;
1529 irqreturn_t result = IRQ_NONE;
1530 struct efx_channel *channel;
1531 efx_dword_t reg;
1532 u32 queues;
1533 int syserr;
1534
1535 /* Read the ISR which also ACKs the interrupts */
1536 efx_readd(efx, ®, FR_BZ_INT_ISR0);
1537 queues = EFX_EXTRACT_DWORD(reg, 0, 31);
1538
1539 /* Legacy interrupts are disabled too late by the EEH kernel
1540 * code. Disable them earlier.
1541 * If an EEH error occurred, the read will have returned all ones.
1542 */
1543 if (EFX_DWORD_IS_ALL_ONES(reg) && efx_try_recovery(efx) &&
1544 !efx->eeh_disabled_legacy_irq) {
1545 disable_irq_nosync(efx->legacy_irq);
1546 efx->eeh_disabled_legacy_irq = true;
1547 }
1548
1549 /* Handle non-event-queue sources */
1550 if (queues & (1U << efx->irq_level) && soft_enabled) {
1551 syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1552 if (unlikely(syserr))
1553 return efx_farch_fatal_interrupt(efx);
1554 efx->last_irq_cpu = raw_smp_processor_id();
1555 }
1556
1557 if (queues != 0) {
1558 efx->irq_zero_count = 0;
1559
1560 /* Schedule processing of any interrupting queues */
1561 if (likely(soft_enabled)) {
1562 efx_for_each_channel(channel, efx) {
1563 if (queues & 1)
1564 efx_schedule_channel_irq(channel);
1565 queues >>= 1;
1566 }
1567 }
1568 result = IRQ_HANDLED;
1569
1570 } else {
1571 efx_qword_t *event;
1572
1573 /* Legacy ISR read can return zero once (SF bug 15783) */
1574
1575 /* We can't return IRQ_HANDLED more than once on seeing ISR=0
1576 * because this might be a shared interrupt. */
1577 if (efx->irq_zero_count++ == 0)
1578 result = IRQ_HANDLED;
1579
1580 /* Ensure we schedule or rearm all event queues */
1581 if (likely(soft_enabled)) {
1582 efx_for_each_channel(channel, efx) {
1583 event = efx_event(channel,
1584 channel->eventq_read_ptr);
1585 if (efx_event_present(event))
1586 efx_schedule_channel_irq(channel);
1587 else
1588 efx_farch_ev_read_ack(channel);
1589 }
1590 }
1591 }
1592
1593 if (result == IRQ_HANDLED)
1594 netif_vdbg(efx, intr, efx->net_dev,
1595 "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
1596 irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
1597
1598 return result;
1599}
1600
1601/* Handle an MSI interrupt
1602 *
1603 * Handle an MSI hardware interrupt. This routine schedules event
1604 * queue processing. No interrupt acknowledgement cycle is necessary.
1605 * Also, we never need to check that the interrupt is for us, since
1606 * MSI interrupts cannot be shared.
1607 */
1608irqreturn_t efx_farch_msi_interrupt(int irq, void *dev_id)
1609{
1610 struct efx_msi_context *context = dev_id;
1611 struct efx_nic *efx = context->efx;
1612 efx_oword_t *int_ker = efx->irq_status.addr;
1613 int syserr;
1614
1615 netif_vdbg(efx, intr, efx->net_dev,
1616 "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
1617 irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
1618
1619 if (!likely(ACCESS_ONCE(efx->irq_soft_enabled)))
1620 return IRQ_HANDLED;
1621
1622 /* Handle non-event-queue sources */
1623 if (context->index == efx->irq_level) {
1624 syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1625 if (unlikely(syserr))
1626 return efx_farch_fatal_interrupt(efx);
1627 efx->last_irq_cpu = raw_smp_processor_id();
1628 }
1629
1630 /* Schedule processing of the channel */
1631 efx_schedule_channel_irq(efx->channel[context->index]);
1632
1633 return IRQ_HANDLED;
1634}
1635
1636/* Setup RSS indirection table.
1637 * This maps from the hash value of the packet to RXQ
1638 */
1639void efx_farch_rx_push_indir_table(struct efx_nic *efx)
1640{
1641 size_t i = 0;
1642 efx_dword_t dword;
1643
1644 BUG_ON(efx_nic_rev(efx) < EFX_REV_FALCON_B0);
1645
1646 BUILD_BUG_ON(ARRAY_SIZE(efx->rx_indir_table) !=
1647 FR_BZ_RX_INDIRECTION_TBL_ROWS);
1648
1649 for (i = 0; i < FR_BZ_RX_INDIRECTION_TBL_ROWS; i++) {
1650 EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
1651 efx->rx_indir_table[i]);
1652 efx_writed(efx, &dword,
1653 FR_BZ_RX_INDIRECTION_TBL +
1654 FR_BZ_RX_INDIRECTION_TBL_STEP * i);
1655 }
1656}
1657
1658/* Looks at available SRAM resources and works out how many queues we
1659 * can support, and where things like descriptor caches should live.
1660 *
1661 * SRAM is split up as follows:
1662 * 0 buftbl entries for channels
1663 * efx->vf_buftbl_base buftbl entries for SR-IOV
1664 * efx->rx_dc_base RX descriptor caches
1665 * efx->tx_dc_base TX descriptor caches
1666 */
1667void efx_farch_dimension_resources(struct efx_nic *efx, unsigned sram_lim_qw)
1668{
1669 unsigned vi_count, buftbl_min;
1670
1671 /* Account for the buffer table entries backing the datapath channels
1672 * and the descriptor caches for those channels.
1673 */
1674 buftbl_min = ((efx->n_rx_channels * EFX_MAX_DMAQ_SIZE +
1675 efx->n_tx_channels * EFX_TXQ_TYPES * EFX_MAX_DMAQ_SIZE +
1676 efx->n_channels * EFX_MAX_EVQ_SIZE)
1677 * sizeof(efx_qword_t) / EFX_BUF_SIZE);
1678 vi_count = max(efx->n_channels, efx->n_tx_channels * EFX_TXQ_TYPES);
1679
1680#ifdef CONFIG_SFC_SRIOV
1681 if (efx_sriov_wanted(efx)) {
1682 unsigned vi_dc_entries, buftbl_free, entries_per_vf, vf_limit;
1683
1684 efx->vf_buftbl_base = buftbl_min;
1685
1686 vi_dc_entries = RX_DC_ENTRIES + TX_DC_ENTRIES;
1687 vi_count = max(vi_count, EFX_VI_BASE);
1688 buftbl_free = (sram_lim_qw - buftbl_min -
1689 vi_count * vi_dc_entries);
1690
1691 entries_per_vf = ((vi_dc_entries + EFX_VF_BUFTBL_PER_VI) *
1692 efx_vf_size(efx));
1693 vf_limit = min(buftbl_free / entries_per_vf,
1694 (1024U - EFX_VI_BASE) >> efx->vi_scale);
1695
1696 if (efx->vf_count > vf_limit) {
1697 netif_err(efx, probe, efx->net_dev,
1698 "Reducing VF count from from %d to %d\n",
1699 efx->vf_count, vf_limit);
1700 efx->vf_count = vf_limit;
1701 }
1702 vi_count += efx->vf_count * efx_vf_size(efx);
1703 }
1704#endif
1705
1706 efx->tx_dc_base = sram_lim_qw - vi_count * TX_DC_ENTRIES;
1707 efx->rx_dc_base = efx->tx_dc_base - vi_count * RX_DC_ENTRIES;
1708}
1709
1710u32 efx_farch_fpga_ver(struct efx_nic *efx)
1711{
1712 efx_oword_t altera_build;
1713 efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
1714 return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
1715}
1716
1717void efx_farch_init_common(struct efx_nic *efx)
1718{
1719 efx_oword_t temp;
1720
1721 /* Set positions of descriptor caches in SRAM. */
1722 EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR, efx->tx_dc_base);
1723 efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
1724 EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR, efx->rx_dc_base);
1725 efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
1726
1727 /* Set TX descriptor cache size. */
1728 BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
1729 EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
1730 efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
1731
1732 /* Set RX descriptor cache size. Set low watermark to size-8, as
1733 * this allows most efficient prefetching.
1734 */
1735 BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
1736 EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
1737 efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
1738 EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
1739 efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
1740
1741 /* Program INT_KER address */
1742 EFX_POPULATE_OWORD_2(temp,
1743 FRF_AZ_NORM_INT_VEC_DIS_KER,
1744 EFX_INT_MODE_USE_MSI(efx),
1745 FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
1746 efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
1747
1748 if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
1749 /* Use an interrupt level unused by event queues */
1750 efx->irq_level = 0x1f;
1751 else
1752 /* Use a valid MSI-X vector */
1753 efx->irq_level = 0;
1754
1755 /* Enable all the genuinely fatal interrupts. (They are still
1756 * masked by the overall interrupt mask, controlled by
1757 * falcon_interrupts()).
1758 *
1759 * Note: All other fatal interrupts are enabled
1760 */
1761 EFX_POPULATE_OWORD_3(temp,
1762 FRF_AZ_ILL_ADR_INT_KER_EN, 1,
1763 FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
1764 FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
1765 if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1766 EFX_SET_OWORD_FIELD(temp, FRF_CZ_SRAM_PERR_INT_P_KER_EN, 1);
1767 EFX_INVERT_OWORD(temp);
1768 efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
1769
1770 /* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
1771 * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
1772 */
1773 efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
1774 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
1775 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
1776 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
1777 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 1);
1778 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
1779 /* Enable SW_EV to inherit in char driver - assume harmless here */
1780 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
1781 /* Prefetch threshold 2 => fetch when descriptor cache half empty */
1782 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
1783 /* Disable hardware watchdog which can misfire */
1784 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
1785 /* Squash TX of packets of 16 bytes or less */
1786 if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1787 EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
1788 efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
1789
1790 if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
1791 EFX_POPULATE_OWORD_4(temp,
1792 /* Default values */
1793 FRF_BZ_TX_PACE_SB_NOT_AF, 0x15,
1794 FRF_BZ_TX_PACE_SB_AF, 0xb,
1795 FRF_BZ_TX_PACE_FB_BASE, 0,
1796 /* Allow large pace values in the
1797 * fast bin. */
1798 FRF_BZ_TX_PACE_BIN_TH,
1799 FFE_BZ_TX_PACE_RESERVED);
1800 efx_writeo(efx, &temp, FR_BZ_TX_PACE);
1801 }
1802}
1803
1804/**************************************************************************
1805 *
1806 * Filter tables
1807 *
1808 **************************************************************************
1809 */
1810
1811/* "Fudge factors" - difference between programmed value and actual depth.
1812 * Due to pipelined implementation we need to program H/W with a value that
1813 * is larger than the hop limit we want.
1814 */
1815#define EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD 3
1816#define EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL 1
1817
1818/* Hard maximum search limit. Hardware will time-out beyond 200-something.
1819 * We also need to avoid infinite loops in efx_farch_filter_search() when the
1820 * table is full.
1821 */
1822#define EFX_FARCH_FILTER_CTL_SRCH_MAX 200
1823
1824/* Don't try very hard to find space for performance hints, as this is
1825 * counter-productive. */
1826#define EFX_FARCH_FILTER_CTL_SRCH_HINT_MAX 5
1827
1828enum efx_farch_filter_type {
1829 EFX_FARCH_FILTER_TCP_FULL = 0,
1830 EFX_FARCH_FILTER_TCP_WILD,
1831 EFX_FARCH_FILTER_UDP_FULL,
1832 EFX_FARCH_FILTER_UDP_WILD,
1833 EFX_FARCH_FILTER_MAC_FULL = 4,
1834 EFX_FARCH_FILTER_MAC_WILD,
1835 EFX_FARCH_FILTER_UC_DEF = 8,
1836 EFX_FARCH_FILTER_MC_DEF,
1837 EFX_FARCH_FILTER_TYPE_COUNT, /* number of specific types */
1838};
1839
1840enum efx_farch_filter_table_id {
1841 EFX_FARCH_FILTER_TABLE_RX_IP = 0,
1842 EFX_FARCH_FILTER_TABLE_RX_MAC,
1843 EFX_FARCH_FILTER_TABLE_RX_DEF,
1844 EFX_FARCH_FILTER_TABLE_TX_MAC,
1845 EFX_FARCH_FILTER_TABLE_COUNT,
1846};
1847
1848enum efx_farch_filter_index {
1849 EFX_FARCH_FILTER_INDEX_UC_DEF,
1850 EFX_FARCH_FILTER_INDEX_MC_DEF,
1851 EFX_FARCH_FILTER_SIZE_RX_DEF,
1852};
1853
1854struct efx_farch_filter_spec {
1855 u8 type:4;
1856 u8 priority:4;
1857 u8 flags;
1858 u16 dmaq_id;
1859 u32 data[3];
1860};
1861
1862struct efx_farch_filter_table {
1863 enum efx_farch_filter_table_id id;
1864 u32 offset; /* address of table relative to BAR */
1865 unsigned size; /* number of entries */
1866 unsigned step; /* step between entries */
1867 unsigned used; /* number currently used */
1868 unsigned long *used_bitmap;
1869 struct efx_farch_filter_spec *spec;
1870 unsigned search_limit[EFX_FARCH_FILTER_TYPE_COUNT];
1871};
1872
1873struct efx_farch_filter_state {
1874 struct efx_farch_filter_table table[EFX_FARCH_FILTER_TABLE_COUNT];
1875};
1876
1877static void
1878efx_farch_filter_table_clear_entry(struct efx_nic *efx,
1879 struct efx_farch_filter_table *table,
1880 unsigned int filter_idx);
1881
1882/* The filter hash function is LFSR polynomial x^16 + x^3 + 1 of a 32-bit
1883 * key derived from the n-tuple. The initial LFSR state is 0xffff. */
1884static u16 efx_farch_filter_hash(u32 key)
1885{
1886 u16 tmp;
1887
1888 /* First 16 rounds */
1889 tmp = 0x1fff ^ key >> 16;
1890 tmp = tmp ^ tmp >> 3 ^ tmp >> 6;
1891 tmp = tmp ^ tmp >> 9;
1892 /* Last 16 rounds */
1893 tmp = tmp ^ tmp << 13 ^ key;
1894 tmp = tmp ^ tmp >> 3 ^ tmp >> 6;
1895 return tmp ^ tmp >> 9;
1896}
1897
1898/* To allow for hash collisions, filter search continues at these
1899 * increments from the first possible entry selected by the hash. */
1900static u16 efx_farch_filter_increment(u32 key)
1901{
1902 return key * 2 - 1;
1903}
1904
1905static enum efx_farch_filter_table_id
1906efx_farch_filter_spec_table_id(const struct efx_farch_filter_spec *spec)
1907{
1908 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1909 (EFX_FARCH_FILTER_TCP_FULL >> 2));
1910 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1911 (EFX_FARCH_FILTER_TCP_WILD >> 2));
1912 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1913 (EFX_FARCH_FILTER_UDP_FULL >> 2));
1914 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1915 (EFX_FARCH_FILTER_UDP_WILD >> 2));
1916 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_MAC !=
1917 (EFX_FARCH_FILTER_MAC_FULL >> 2));
1918 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_MAC !=
1919 (EFX_FARCH_FILTER_MAC_WILD >> 2));
1920 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_TX_MAC !=
1921 EFX_FARCH_FILTER_TABLE_RX_MAC + 2);
1922 return (spec->type >> 2) + ((spec->flags & EFX_FILTER_FLAG_TX) ? 2 : 0);
1923}
1924
1925static void efx_farch_filter_push_rx_config(struct efx_nic *efx)
1926{
1927 struct efx_farch_filter_state *state = efx->filter_state;
1928 struct efx_farch_filter_table *table;
1929 efx_oword_t filter_ctl;
1930
1931 efx_reado(efx, &filter_ctl, FR_BZ_RX_FILTER_CTL);
1932
1933 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_IP];
1934 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_TCP_FULL_SRCH_LIMIT,
1935 table->search_limit[EFX_FARCH_FILTER_TCP_FULL] +
1936 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
1937 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_TCP_WILD_SRCH_LIMIT,
1938 table->search_limit[EFX_FARCH_FILTER_TCP_WILD] +
1939 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
1940 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_UDP_FULL_SRCH_LIMIT,
1941 table->search_limit[EFX_FARCH_FILTER_UDP_FULL] +
1942 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
1943 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_UDP_WILD_SRCH_LIMIT,
1944 table->search_limit[EFX_FARCH_FILTER_UDP_WILD] +
1945 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
1946
1947 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_MAC];
1948 if (table->size) {
1949 EFX_SET_OWORD_FIELD(
1950 filter_ctl, FRF_CZ_ETHERNET_FULL_SEARCH_LIMIT,
1951 table->search_limit[EFX_FARCH_FILTER_MAC_FULL] +
1952 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
1953 EFX_SET_OWORD_FIELD(
1954 filter_ctl, FRF_CZ_ETHERNET_WILDCARD_SEARCH_LIMIT,
1955 table->search_limit[EFX_FARCH_FILTER_MAC_WILD] +
1956 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
1957 }
1958
1959 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_DEF];
1960 if (table->size) {
1961 EFX_SET_OWORD_FIELD(
1962 filter_ctl, FRF_CZ_UNICAST_NOMATCH_Q_ID,
1963 table->spec[EFX_FARCH_FILTER_INDEX_UC_DEF].dmaq_id);
1964 EFX_SET_OWORD_FIELD(
1965 filter_ctl, FRF_CZ_UNICAST_NOMATCH_RSS_ENABLED,
1966 !!(table->spec[EFX_FARCH_FILTER_INDEX_UC_DEF].flags &
1967 EFX_FILTER_FLAG_RX_RSS));
1968 EFX_SET_OWORD_FIELD(
1969 filter_ctl, FRF_CZ_MULTICAST_NOMATCH_Q_ID,
1970 table->spec[EFX_FARCH_FILTER_INDEX_MC_DEF].dmaq_id);
1971 EFX_SET_OWORD_FIELD(
1972 filter_ctl, FRF_CZ_MULTICAST_NOMATCH_RSS_ENABLED,
1973 !!(table->spec[EFX_FARCH_FILTER_INDEX_MC_DEF].flags &
1974 EFX_FILTER_FLAG_RX_RSS));
1975
1976 /* There is a single bit to enable RX scatter for all
1977 * unmatched packets. Only set it if scatter is
1978 * enabled in both filter specs.
1979 */
1980 EFX_SET_OWORD_FIELD(
1981 filter_ctl, FRF_BZ_SCATTER_ENBL_NO_MATCH_Q,
1982 !!(table->spec[EFX_FARCH_FILTER_INDEX_UC_DEF].flags &
1983 table->spec[EFX_FARCH_FILTER_INDEX_MC_DEF].flags &
1984 EFX_FILTER_FLAG_RX_SCATTER));
1985 } else if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
1986 /* We don't expose 'default' filters because unmatched
1987 * packets always go to the queue number found in the
1988 * RSS table. But we still need to set the RX scatter
1989 * bit here.
1990 */
1991 EFX_SET_OWORD_FIELD(
1992 filter_ctl, FRF_BZ_SCATTER_ENBL_NO_MATCH_Q,
1993 efx->rx_scatter);
1994 }
1995
1996 efx_writeo(efx, &filter_ctl, FR_BZ_RX_FILTER_CTL);
1997}
1998
1999static void efx_farch_filter_push_tx_limits(struct efx_nic *efx)
2000{
2001 struct efx_farch_filter_state *state = efx->filter_state;
2002 struct efx_farch_filter_table *table;
2003 efx_oword_t tx_cfg;
2004
2005 efx_reado(efx, &tx_cfg, FR_AZ_TX_CFG);
2006
2007 table = &state->table[EFX_FARCH_FILTER_TABLE_TX_MAC];
2008 if (table->size) {
2009 EFX_SET_OWORD_FIELD(
2010 tx_cfg, FRF_CZ_TX_ETH_FILTER_FULL_SEARCH_RANGE,
2011 table->search_limit[EFX_FARCH_FILTER_MAC_FULL] +
2012 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
2013 EFX_SET_OWORD_FIELD(
2014 tx_cfg, FRF_CZ_TX_ETH_FILTER_WILD_SEARCH_RANGE,
2015 table->search_limit[EFX_FARCH_FILTER_MAC_WILD] +
2016 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
2017 }
2018
2019 efx_writeo(efx, &tx_cfg, FR_AZ_TX_CFG);
2020}
2021
2022static int
2023efx_farch_filter_from_gen_spec(struct efx_farch_filter_spec *spec,
2024 const struct efx_filter_spec *gen_spec)
2025{
2026 bool is_full = false;
2027
2028 if ((gen_spec->flags & EFX_FILTER_FLAG_RX_RSS) &&
2029 gen_spec->rss_context != EFX_FILTER_RSS_CONTEXT_DEFAULT)
2030 return -EINVAL;
2031
2032 spec->priority = gen_spec->priority;
2033 spec->flags = gen_spec->flags;
2034 spec->dmaq_id = gen_spec->dmaq_id;
2035
2036 switch (gen_spec->match_flags) {
2037 case (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
2038 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
2039 EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT):
2040 is_full = true;
2041 /* fall through */
2042 case (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
2043 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT): {
2044 __be32 rhost, host1, host2;
2045 __be16 rport, port1, port2;
2046
2047 EFX_BUG_ON_PARANOID(!(gen_spec->flags & EFX_FILTER_FLAG_RX));
2048
2049 if (gen_spec->ether_type != htons(ETH_P_IP))
2050 return -EPROTONOSUPPORT;
2051 if (gen_spec->loc_port == 0 ||
2052 (is_full && gen_spec->rem_port == 0))
2053 return -EADDRNOTAVAIL;
2054 switch (gen_spec->ip_proto) {
2055 case IPPROTO_TCP:
2056 spec->type = (is_full ? EFX_FARCH_FILTER_TCP_FULL :
2057 EFX_FARCH_FILTER_TCP_WILD);
2058 break;
2059 case IPPROTO_UDP:
2060 spec->type = (is_full ? EFX_FARCH_FILTER_UDP_FULL :
2061 EFX_FARCH_FILTER_UDP_WILD);
2062 break;
2063 default:
2064 return -EPROTONOSUPPORT;
2065 }
2066
2067 /* Filter is constructed in terms of source and destination,
2068 * with the odd wrinkle that the ports are swapped in a UDP
2069 * wildcard filter. We need to convert from local and remote
2070 * (= zero for wildcard) addresses.
2071 */
2072 rhost = is_full ? gen_spec->rem_host[0] : 0;
2073 rport = is_full ? gen_spec->rem_port : 0;
2074 host1 = rhost;
2075 host2 = gen_spec->loc_host[0];
2076 if (!is_full && gen_spec->ip_proto == IPPROTO_UDP) {
2077 port1 = gen_spec->loc_port;
2078 port2 = rport;
2079 } else {
2080 port1 = rport;
2081 port2 = gen_spec->loc_port;
2082 }
2083 spec->data[0] = ntohl(host1) << 16 | ntohs(port1);
2084 spec->data[1] = ntohs(port2) << 16 | ntohl(host1) >> 16;
2085 spec->data[2] = ntohl(host2);
2086
2087 break;
2088 }
2089
2090 case EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_OUTER_VID:
2091 is_full = true;
2092 /* fall through */
2093 case EFX_FILTER_MATCH_LOC_MAC:
2094 spec->type = (is_full ? EFX_FARCH_FILTER_MAC_FULL :
2095 EFX_FARCH_FILTER_MAC_WILD);
2096 spec->data[0] = is_full ? ntohs(gen_spec->outer_vid) : 0;
2097 spec->data[1] = (gen_spec->loc_mac[2] << 24 |
2098 gen_spec->loc_mac[3] << 16 |
2099 gen_spec->loc_mac[4] << 8 |
2100 gen_spec->loc_mac[5]);
2101 spec->data[2] = (gen_spec->loc_mac[0] << 8 |
2102 gen_spec->loc_mac[1]);
2103 break;
2104
2105 case EFX_FILTER_MATCH_LOC_MAC_IG:
2106 spec->type = (is_multicast_ether_addr(gen_spec->loc_mac) ?
2107 EFX_FARCH_FILTER_MC_DEF :
2108 EFX_FARCH_FILTER_UC_DEF);
2109 memset(spec->data, 0, sizeof(spec->data)); /* ensure equality */
2110 break;
2111
2112 default:
2113 return -EPROTONOSUPPORT;
2114 }
2115
2116 return 0;
2117}
2118
2119static void
2120efx_farch_filter_to_gen_spec(struct efx_filter_spec *gen_spec,
2121 const struct efx_farch_filter_spec *spec)
2122{
2123 bool is_full = false;
2124
2125 /* *gen_spec should be completely initialised, to be consistent
2126 * with efx_filter_init_{rx,tx}() and in case we want to copy
2127 * it back to userland.
2128 */
2129 memset(gen_spec, 0, sizeof(*gen_spec));
2130
2131 gen_spec->priority = spec->priority;
2132 gen_spec->flags = spec->flags;
2133 gen_spec->dmaq_id = spec->dmaq_id;
2134
2135 switch (spec->type) {
2136 case EFX_FARCH_FILTER_TCP_FULL:
2137 case EFX_FARCH_FILTER_UDP_FULL:
2138 is_full = true;
2139 /* fall through */
2140 case EFX_FARCH_FILTER_TCP_WILD:
2141 case EFX_FARCH_FILTER_UDP_WILD: {
2142 __be32 host1, host2;
2143 __be16 port1, port2;
2144
2145 gen_spec->match_flags =
2146 EFX_FILTER_MATCH_ETHER_TYPE |
2147 EFX_FILTER_MATCH_IP_PROTO |
2148 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT;
2149 if (is_full)
2150 gen_spec->match_flags |= (EFX_FILTER_MATCH_REM_HOST |
2151 EFX_FILTER_MATCH_REM_PORT);
2152 gen_spec->ether_type = htons(ETH_P_IP);
2153 gen_spec->ip_proto =
2154 (spec->type == EFX_FARCH_FILTER_TCP_FULL ||
2155 spec->type == EFX_FARCH_FILTER_TCP_WILD) ?
2156 IPPROTO_TCP : IPPROTO_UDP;
2157
2158 host1 = htonl(spec->data[0] >> 16 | spec->data[1] << 16);
2159 port1 = htons(spec->data[0]);
2160 host2 = htonl(spec->data[2]);
2161 port2 = htons(spec->data[1] >> 16);
2162 if (spec->flags & EFX_FILTER_FLAG_TX) {
2163 gen_spec->loc_host[0] = host1;
2164 gen_spec->rem_host[0] = host2;
2165 } else {
2166 gen_spec->loc_host[0] = host2;
2167 gen_spec->rem_host[0] = host1;
2168 }
2169 if (!!(gen_spec->flags & EFX_FILTER_FLAG_TX) ^
2170 (!is_full && gen_spec->ip_proto == IPPROTO_UDP)) {
2171 gen_spec->loc_port = port1;
2172 gen_spec->rem_port = port2;
2173 } else {
2174 gen_spec->loc_port = port2;
2175 gen_spec->rem_port = port1;
2176 }
2177
2178 break;
2179 }
2180
2181 case EFX_FARCH_FILTER_MAC_FULL:
2182 is_full = true;
2183 /* fall through */
2184 case EFX_FARCH_FILTER_MAC_WILD:
2185 gen_spec->match_flags = EFX_FILTER_MATCH_LOC_MAC;
2186 if (is_full)
2187 gen_spec->match_flags |= EFX_FILTER_MATCH_OUTER_VID;
2188 gen_spec->loc_mac[0] = spec->data[2] >> 8;
2189 gen_spec->loc_mac[1] = spec->data[2];
2190 gen_spec->loc_mac[2] = spec->data[1] >> 24;
2191 gen_spec->loc_mac[3] = spec->data[1] >> 16;
2192 gen_spec->loc_mac[4] = spec->data[1] >> 8;
2193 gen_spec->loc_mac[5] = spec->data[1];
2194 gen_spec->outer_vid = htons(spec->data[0]);
2195 break;
2196
2197 case EFX_FARCH_FILTER_UC_DEF:
2198 case EFX_FARCH_FILTER_MC_DEF:
2199 gen_spec->match_flags = EFX_FILTER_MATCH_LOC_MAC_IG;
2200 gen_spec->loc_mac[0] = spec->type == EFX_FARCH_FILTER_MC_DEF;
2201 break;
2202
2203 default:
2204 WARN_ON(1);
2205 break;
2206 }
2207}
2208
2209static void
2210efx_farch_filter_init_rx_auto(struct efx_nic *efx,
2211 struct efx_farch_filter_spec *spec)
2212{
2213 /* If there's only one channel then disable RSS for non VF
2214 * traffic, thereby allowing VFs to use RSS when the PF can't.
2215 */
2216 spec->priority = EFX_FILTER_PRI_AUTO;
2217 spec->flags = (EFX_FILTER_FLAG_RX |
2218 (efx->n_rx_channels > 1 ? EFX_FILTER_FLAG_RX_RSS : 0) |
2219 (efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0));
2220 spec->dmaq_id = 0;
2221}
2222
2223/* Build a filter entry and return its n-tuple key. */
2224static u32 efx_farch_filter_build(efx_oword_t *filter,
2225 struct efx_farch_filter_spec *spec)
2226{
2227 u32 data3;
2228
2229 switch (efx_farch_filter_spec_table_id(spec)) {
2230 case EFX_FARCH_FILTER_TABLE_RX_IP: {
2231 bool is_udp = (spec->type == EFX_FARCH_FILTER_UDP_FULL ||
2232 spec->type == EFX_FARCH_FILTER_UDP_WILD);
2233 EFX_POPULATE_OWORD_7(
2234 *filter,
2235 FRF_BZ_RSS_EN,
2236 !!(spec->flags & EFX_FILTER_FLAG_RX_RSS),
2237 FRF_BZ_SCATTER_EN,
2238 !!(spec->flags & EFX_FILTER_FLAG_RX_SCATTER),
2239 FRF_BZ_TCP_UDP, is_udp,
2240 FRF_BZ_RXQ_ID, spec->dmaq_id,
2241 EFX_DWORD_2, spec->data[2],
2242 EFX_DWORD_1, spec->data[1],
2243 EFX_DWORD_0, spec->data[0]);
2244 data3 = is_udp;
2245 break;
2246 }
2247
2248 case EFX_FARCH_FILTER_TABLE_RX_MAC: {
2249 bool is_wild = spec->type == EFX_FARCH_FILTER_MAC_WILD;
2250 EFX_POPULATE_OWORD_7(
2251 *filter,
2252 FRF_CZ_RMFT_RSS_EN,
2253 !!(spec->flags & EFX_FILTER_FLAG_RX_RSS),
2254 FRF_CZ_RMFT_SCATTER_EN,
2255 !!(spec->flags & EFX_FILTER_FLAG_RX_SCATTER),
2256 FRF_CZ_RMFT_RXQ_ID, spec->dmaq_id,
2257 FRF_CZ_RMFT_WILDCARD_MATCH, is_wild,
2258 FRF_CZ_RMFT_DEST_MAC_HI, spec->data[2],
2259 FRF_CZ_RMFT_DEST_MAC_LO, spec->data[1],
2260 FRF_CZ_RMFT_VLAN_ID, spec->data[0]);
2261 data3 = is_wild;
2262 break;
2263 }
2264
2265 case EFX_FARCH_FILTER_TABLE_TX_MAC: {
2266 bool is_wild = spec->type == EFX_FARCH_FILTER_MAC_WILD;
2267 EFX_POPULATE_OWORD_5(*filter,
2268 FRF_CZ_TMFT_TXQ_ID, spec->dmaq_id,
2269 FRF_CZ_TMFT_WILDCARD_MATCH, is_wild,
2270 FRF_CZ_TMFT_SRC_MAC_HI, spec->data[2],
2271 FRF_CZ_TMFT_SRC_MAC_LO, spec->data[1],
2272 FRF_CZ_TMFT_VLAN_ID, spec->data[0]);
2273 data3 = is_wild | spec->dmaq_id << 1;
2274 break;
2275 }
2276
2277 default:
2278 BUG();
2279 }
2280
2281 return spec->data[0] ^ spec->data[1] ^ spec->data[2] ^ data3;
2282}
2283
2284static bool efx_farch_filter_equal(const struct efx_farch_filter_spec *left,
2285 const struct efx_farch_filter_spec *right)
2286{
2287 if (left->type != right->type ||
2288 memcmp(left->data, right->data, sizeof(left->data)))
2289 return false;
2290
2291 if (left->flags & EFX_FILTER_FLAG_TX &&
2292 left->dmaq_id != right->dmaq_id)
2293 return false;
2294
2295 return true;
2296}
2297
2298/*
2299 * Construct/deconstruct external filter IDs. At least the RX filter
2300 * IDs must be ordered by matching priority, for RX NFC semantics.
2301 *
2302 * Deconstruction needs to be robust against invalid IDs so that
2303 * efx_filter_remove_id_safe() and efx_filter_get_filter_safe() can
2304 * accept user-provided IDs.
2305 */
2306
2307#define EFX_FARCH_FILTER_MATCH_PRI_COUNT 5
2308
2309static const u8 efx_farch_filter_type_match_pri[EFX_FARCH_FILTER_TYPE_COUNT] = {
2310 [EFX_FARCH_FILTER_TCP_FULL] = 0,
2311 [EFX_FARCH_FILTER_UDP_FULL] = 0,
2312 [EFX_FARCH_FILTER_TCP_WILD] = 1,
2313 [EFX_FARCH_FILTER_UDP_WILD] = 1,
2314 [EFX_FARCH_FILTER_MAC_FULL] = 2,
2315 [EFX_FARCH_FILTER_MAC_WILD] = 3,
2316 [EFX_FARCH_FILTER_UC_DEF] = 4,
2317 [EFX_FARCH_FILTER_MC_DEF] = 4,
2318};
2319
2320static const enum efx_farch_filter_table_id efx_farch_filter_range_table[] = {
2321 EFX_FARCH_FILTER_TABLE_RX_IP, /* RX match pri 0 */
2322 EFX_FARCH_FILTER_TABLE_RX_IP,
2323 EFX_FARCH_FILTER_TABLE_RX_MAC,
2324 EFX_FARCH_FILTER_TABLE_RX_MAC,
2325 EFX_FARCH_FILTER_TABLE_RX_DEF, /* RX match pri 4 */
2326 EFX_FARCH_FILTER_TABLE_TX_MAC, /* TX match pri 0 */
2327 EFX_FARCH_FILTER_TABLE_TX_MAC, /* TX match pri 1 */
2328};
2329
2330#define EFX_FARCH_FILTER_INDEX_WIDTH 13
2331#define EFX_FARCH_FILTER_INDEX_MASK ((1 << EFX_FARCH_FILTER_INDEX_WIDTH) - 1)
2332
2333static inline u32
2334efx_farch_filter_make_id(const struct efx_farch_filter_spec *spec,
2335 unsigned int index)
2336{
2337 unsigned int range;
2338
2339 range = efx_farch_filter_type_match_pri[spec->type];
2340 if (!(spec->flags & EFX_FILTER_FLAG_RX))
2341 range += EFX_FARCH_FILTER_MATCH_PRI_COUNT;
2342
2343 return range << EFX_FARCH_FILTER_INDEX_WIDTH | index;
2344}
2345
2346static inline enum efx_farch_filter_table_id
2347efx_farch_filter_id_table_id(u32 id)
2348{
2349 unsigned int range = id >> EFX_FARCH_FILTER_INDEX_WIDTH;
2350
2351 if (range < ARRAY_SIZE(efx_farch_filter_range_table))
2352 return efx_farch_filter_range_table[range];
2353 else
2354 return EFX_FARCH_FILTER_TABLE_COUNT; /* invalid */
2355}
2356
2357static inline unsigned int efx_farch_filter_id_index(u32 id)
2358{
2359 return id & EFX_FARCH_FILTER_INDEX_MASK;
2360}
2361
2362u32 efx_farch_filter_get_rx_id_limit(struct efx_nic *efx)
2363{
2364 struct efx_farch_filter_state *state = efx->filter_state;
2365 unsigned int range = EFX_FARCH_FILTER_MATCH_PRI_COUNT - 1;
2366 enum efx_farch_filter_table_id table_id;
2367
2368 do {
2369 table_id = efx_farch_filter_range_table[range];
2370 if (state->table[table_id].size != 0)
2371 return range << EFX_FARCH_FILTER_INDEX_WIDTH |
2372 state->table[table_id].size;
2373 } while (range--);
2374
2375 return 0;
2376}
2377
2378s32 efx_farch_filter_insert(struct efx_nic *efx,
2379 struct efx_filter_spec *gen_spec,
2380 bool replace_equal)
2381{
2382 struct efx_farch_filter_state *state = efx->filter_state;
2383 struct efx_farch_filter_table *table;
2384 struct efx_farch_filter_spec spec;
2385 efx_oword_t filter;
2386 int rep_index, ins_index;
2387 unsigned int depth = 0;
2388 int rc;
2389
2390 rc = efx_farch_filter_from_gen_spec(&spec, gen_spec);
2391 if (rc)
2392 return rc;
2393
2394 table = &state->table[efx_farch_filter_spec_table_id(&spec)];
2395 if (table->size == 0)
2396 return -EINVAL;
2397
2398 netif_vdbg(efx, hw, efx->net_dev,
2399 "%s: type %d search_limit=%d", __func__, spec.type,
2400 table->search_limit[spec.type]);
2401
2402 if (table->id == EFX_FARCH_FILTER_TABLE_RX_DEF) {
2403 /* One filter spec per type */
2404 BUILD_BUG_ON(EFX_FARCH_FILTER_INDEX_UC_DEF != 0);
2405 BUILD_BUG_ON(EFX_FARCH_FILTER_INDEX_MC_DEF !=
2406 EFX_FARCH_FILTER_MC_DEF - EFX_FARCH_FILTER_UC_DEF);
2407 rep_index = spec.type - EFX_FARCH_FILTER_UC_DEF;
2408 ins_index = rep_index;
2409
2410 spin_lock_bh(&efx->filter_lock);
2411 } else {
2412 /* Search concurrently for
2413 * (1) a filter to be replaced (rep_index): any filter
2414 * with the same match values, up to the current
2415 * search depth for this type, and
2416 * (2) the insertion point (ins_index): (1) or any
2417 * free slot before it or up to the maximum search
2418 * depth for this priority
2419 * We fail if we cannot find (2).
2420 *
2421 * We can stop once either
2422 * (a) we find (1), in which case we have definitely
2423 * found (2) as well; or
2424 * (b) we have searched exhaustively for (1), and have
2425 * either found (2) or searched exhaustively for it
2426 */
2427 u32 key = efx_farch_filter_build(&filter, &spec);
2428 unsigned int hash = efx_farch_filter_hash(key);
2429 unsigned int incr = efx_farch_filter_increment(key);
2430 unsigned int max_rep_depth = table->search_limit[spec.type];
2431 unsigned int max_ins_depth =
2432 spec.priority <= EFX_FILTER_PRI_HINT ?
2433 EFX_FARCH_FILTER_CTL_SRCH_HINT_MAX :
2434 EFX_FARCH_FILTER_CTL_SRCH_MAX;
2435 unsigned int i = hash & (table->size - 1);
2436
2437 ins_index = -1;
2438 depth = 1;
2439
2440 spin_lock_bh(&efx->filter_lock);
2441
2442 for (;;) {
2443 if (!test_bit(i, table->used_bitmap)) {
2444 if (ins_index < 0)
2445 ins_index = i;
2446 } else if (efx_farch_filter_equal(&spec,
2447 &table->spec[i])) {
2448 /* Case (a) */
2449 if (ins_index < 0)
2450 ins_index = i;
2451 rep_index = i;
2452 break;
2453 }
2454
2455 if (depth >= max_rep_depth &&
2456 (ins_index >= 0 || depth >= max_ins_depth)) {
2457 /* Case (b) */
2458 if (ins_index < 0) {
2459 rc = -EBUSY;
2460 goto out;
2461 }
2462 rep_index = -1;
2463 break;
2464 }
2465
2466 i = (i + incr) & (table->size - 1);
2467 ++depth;
2468 }
2469 }
2470
2471 /* If we found a filter to be replaced, check whether we
2472 * should do so
2473 */
2474 if (rep_index >= 0) {
2475 struct efx_farch_filter_spec *saved_spec =
2476 &table->spec[rep_index];
2477
2478 if (spec.priority == saved_spec->priority && !replace_equal) {
2479 rc = -EEXIST;
2480 goto out;
2481 }
2482 if (spec.priority < saved_spec->priority) {
2483 rc = -EPERM;
2484 goto out;
2485 }
2486 if (saved_spec->priority == EFX_FILTER_PRI_AUTO ||
2487 saved_spec->flags & EFX_FILTER_FLAG_RX_OVER_AUTO)
2488 spec.flags |= EFX_FILTER_FLAG_RX_OVER_AUTO;
2489 }
2490
2491 /* Insert the filter */
2492 if (ins_index != rep_index) {
2493 __set_bit(ins_index, table->used_bitmap);
2494 ++table->used;
2495 }
2496 table->spec[ins_index] = spec;
2497
2498 if (table->id == EFX_FARCH_FILTER_TABLE_RX_DEF) {
2499 efx_farch_filter_push_rx_config(efx);
2500 } else {
2501 if (table->search_limit[spec.type] < depth) {
2502 table->search_limit[spec.type] = depth;
2503 if (spec.flags & EFX_FILTER_FLAG_TX)
2504 efx_farch_filter_push_tx_limits(efx);
2505 else
2506 efx_farch_filter_push_rx_config(efx);
2507 }
2508
2509 efx_writeo(efx, &filter,
2510 table->offset + table->step * ins_index);
2511
2512 /* If we were able to replace a filter by inserting
2513 * at a lower depth, clear the replaced filter
2514 */
2515 if (ins_index != rep_index && rep_index >= 0)
2516 efx_farch_filter_table_clear_entry(efx, table,
2517 rep_index);
2518 }
2519
2520 netif_vdbg(efx, hw, efx->net_dev,
2521 "%s: filter type %d index %d rxq %u set",
2522 __func__, spec.type, ins_index, spec.dmaq_id);
2523 rc = efx_farch_filter_make_id(&spec, ins_index);
2524
2525out:
2526 spin_unlock_bh(&efx->filter_lock);
2527 return rc;
2528}
2529
2530static void
2531efx_farch_filter_table_clear_entry(struct efx_nic *efx,
2532 struct efx_farch_filter_table *table,
2533 unsigned int filter_idx)
2534{
2535 static efx_oword_t filter;
2536
2537 EFX_WARN_ON_PARANOID(!test_bit(filter_idx, table->used_bitmap));
2538 BUG_ON(table->offset == 0); /* can't clear MAC default filters */
2539
2540 __clear_bit(filter_idx, table->used_bitmap);
2541 --table->used;
2542 memset(&table->spec[filter_idx], 0, sizeof(table->spec[0]));
2543
2544 efx_writeo(efx, &filter, table->offset + table->step * filter_idx);
2545
2546 /* If this filter required a greater search depth than
2547 * any other, the search limit for its type can now be
2548 * decreased. However, it is hard to determine that
2549 * unless the table has become completely empty - in
2550 * which case, all its search limits can be set to 0.
2551 */
2552 if (unlikely(table->used == 0)) {
2553 memset(table->search_limit, 0, sizeof(table->search_limit));
2554 if (table->id == EFX_FARCH_FILTER_TABLE_TX_MAC)
2555 efx_farch_filter_push_tx_limits(efx);
2556 else
2557 efx_farch_filter_push_rx_config(efx);
2558 }
2559}
2560
2561static int efx_farch_filter_remove(struct efx_nic *efx,
2562 struct efx_farch_filter_table *table,
2563 unsigned int filter_idx,
2564 enum efx_filter_priority priority)
2565{
2566 struct efx_farch_filter_spec *spec = &table->spec[filter_idx];
2567
2568 if (!test_bit(filter_idx, table->used_bitmap) ||
2569 spec->priority != priority)
2570 return -ENOENT;
2571
2572 if (spec->flags & EFX_FILTER_FLAG_RX_OVER_AUTO) {
2573 efx_farch_filter_init_rx_auto(efx, spec);
2574 efx_farch_filter_push_rx_config(efx);
2575 } else {
2576 efx_farch_filter_table_clear_entry(efx, table, filter_idx);
2577 }
2578
2579 return 0;
2580}
2581
2582int efx_farch_filter_remove_safe(struct efx_nic *efx,
2583 enum efx_filter_priority priority,
2584 u32 filter_id)
2585{
2586 struct efx_farch_filter_state *state = efx->filter_state;
2587 enum efx_farch_filter_table_id table_id;
2588 struct efx_farch_filter_table *table;
2589 unsigned int filter_idx;
2590 struct efx_farch_filter_spec *spec;
2591 int rc;
2592
2593 table_id = efx_farch_filter_id_table_id(filter_id);
2594 if ((unsigned int)table_id >= EFX_FARCH_FILTER_TABLE_COUNT)
2595 return -ENOENT;
2596 table = &state->table[table_id];
2597
2598 filter_idx = efx_farch_filter_id_index(filter_id);
2599 if (filter_idx >= table->size)
2600 return -ENOENT;
2601 spec = &table->spec[filter_idx];
2602
2603 spin_lock_bh(&efx->filter_lock);
2604 rc = efx_farch_filter_remove(efx, table, filter_idx, priority);
2605 spin_unlock_bh(&efx->filter_lock);
2606
2607 return rc;
2608}
2609
2610int efx_farch_filter_get_safe(struct efx_nic *efx,
2611 enum efx_filter_priority priority,
2612 u32 filter_id, struct efx_filter_spec *spec_buf)
2613{
2614 struct efx_farch_filter_state *state = efx->filter_state;
2615 enum efx_farch_filter_table_id table_id;
2616 struct efx_farch_filter_table *table;
2617 struct efx_farch_filter_spec *spec;
2618 unsigned int filter_idx;
2619 int rc;
2620
2621 table_id = efx_farch_filter_id_table_id(filter_id);
2622 if ((unsigned int)table_id >= EFX_FARCH_FILTER_TABLE_COUNT)
2623 return -ENOENT;
2624 table = &state->table[table_id];
2625
2626 filter_idx = efx_farch_filter_id_index(filter_id);
2627 if (filter_idx >= table->size)
2628 return -ENOENT;
2629 spec = &table->spec[filter_idx];
2630
2631 spin_lock_bh(&efx->filter_lock);
2632
2633 if (test_bit(filter_idx, table->used_bitmap) &&
2634 spec->priority == priority) {
2635 efx_farch_filter_to_gen_spec(spec_buf, spec);
2636 rc = 0;
2637 } else {
2638 rc = -ENOENT;
2639 }
2640
2641 spin_unlock_bh(&efx->filter_lock);
2642
2643 return rc;
2644}
2645
2646static void
2647efx_farch_filter_table_clear(struct efx_nic *efx,
2648 enum efx_farch_filter_table_id table_id,
2649 enum efx_filter_priority priority)
2650{
2651 struct efx_farch_filter_state *state = efx->filter_state;
2652 struct efx_farch_filter_table *table = &state->table[table_id];
2653 unsigned int filter_idx;
2654
2655 spin_lock_bh(&efx->filter_lock);
2656 for (filter_idx = 0; filter_idx < table->size; ++filter_idx) {
2657 if (table->spec[filter_idx].priority != EFX_FILTER_PRI_AUTO)
2658 efx_farch_filter_remove(efx, table,
2659 filter_idx, priority);
2660 }
2661 spin_unlock_bh(&efx->filter_lock);
2662}
2663
2664int efx_farch_filter_clear_rx(struct efx_nic *efx,
2665 enum efx_filter_priority priority)
2666{
2667 efx_farch_filter_table_clear(efx, EFX_FARCH_FILTER_TABLE_RX_IP,
2668 priority);
2669 efx_farch_filter_table_clear(efx, EFX_FARCH_FILTER_TABLE_RX_MAC,
2670 priority);
2671 efx_farch_filter_table_clear(efx, EFX_FARCH_FILTER_TABLE_RX_DEF,
2672 priority);
2673 return 0;
2674}
2675
2676u32 efx_farch_filter_count_rx_used(struct efx_nic *efx,
2677 enum efx_filter_priority priority)
2678{
2679 struct efx_farch_filter_state *state = efx->filter_state;
2680 enum efx_farch_filter_table_id table_id;
2681 struct efx_farch_filter_table *table;
2682 unsigned int filter_idx;
2683 u32 count = 0;
2684
2685 spin_lock_bh(&efx->filter_lock);
2686
2687 for (table_id = EFX_FARCH_FILTER_TABLE_RX_IP;
2688 table_id <= EFX_FARCH_FILTER_TABLE_RX_DEF;
2689 table_id++) {
2690 table = &state->table[table_id];
2691 for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
2692 if (test_bit(filter_idx, table->used_bitmap) &&
2693 table->spec[filter_idx].priority == priority)
2694 ++count;
2695 }
2696 }
2697
2698 spin_unlock_bh(&efx->filter_lock);
2699
2700 return count;
2701}
2702
2703s32 efx_farch_filter_get_rx_ids(struct efx_nic *efx,
2704 enum efx_filter_priority priority,
2705 u32 *buf, u32 size)
2706{
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;
2711 s32 count = 0;
2712
2713 spin_lock_bh(&efx->filter_lock);
2714
2715 for (table_id = EFX_FARCH_FILTER_TABLE_RX_IP;
2716 table_id <= EFX_FARCH_FILTER_TABLE_RX_DEF;
2717 table_id++) {
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) {
2722 if (count == size) {
2723 count = -EMSGSIZE;
2724 goto out;
2725 }
2726 buf[count++] = efx_farch_filter_make_id(
2727 &table->spec[filter_idx], filter_idx);
2728 }
2729 }
2730 }
2731out:
2732 spin_unlock_bh(&efx->filter_lock);
2733
2734 return count;
2735}
2736
2737/* Restore filter stater after reset */
2738void efx_farch_filter_table_restore(struct efx_nic *efx)
2739{
2740 struct efx_farch_filter_state *state = efx->filter_state;
2741 enum efx_farch_filter_table_id table_id;
2742 struct efx_farch_filter_table *table;
2743 efx_oword_t filter;
2744 unsigned int filter_idx;
2745
2746 spin_lock_bh(&efx->filter_lock);
2747
2748 for (table_id = 0; table_id < EFX_FARCH_FILTER_TABLE_COUNT; table_id++) {
2749 table = &state->table[table_id];
2750
2751 /* Check whether this is a regular register table */
2752 if (table->step == 0)
2753 continue;
2754
2755 for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
2756 if (!test_bit(filter_idx, table->used_bitmap))
2757 continue;
2758 efx_farch_filter_build(&filter, &table->spec[filter_idx]);
2759 efx_writeo(efx, &filter,
2760 table->offset + table->step * filter_idx);
2761 }
2762 }
2763
2764 efx_farch_filter_push_rx_config(efx);
2765 efx_farch_filter_push_tx_limits(efx);
2766
2767 spin_unlock_bh(&efx->filter_lock);
2768}
2769
2770void efx_farch_filter_table_remove(struct efx_nic *efx)
2771{
2772 struct efx_farch_filter_state *state = efx->filter_state;
2773 enum efx_farch_filter_table_id table_id;
2774
2775 for (table_id = 0; table_id < EFX_FARCH_FILTER_TABLE_COUNT; table_id++) {
2776 kfree(state->table[table_id].used_bitmap);
2777 vfree(state->table[table_id].spec);
2778 }
2779 kfree(state);
2780}
2781
2782int efx_farch_filter_table_probe(struct efx_nic *efx)
2783{
2784 struct efx_farch_filter_state *state;
2785 struct efx_farch_filter_table *table;
2786 unsigned table_id;
2787
2788 state = kzalloc(sizeof(struct efx_farch_filter_state), GFP_KERNEL);
2789 if (!state)
2790 return -ENOMEM;
2791 efx->filter_state = state;
2792
2793 if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
2794 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_IP];
2795 table->id = EFX_FARCH_FILTER_TABLE_RX_IP;
2796 table->offset = FR_BZ_RX_FILTER_TBL0;
2797 table->size = FR_BZ_RX_FILTER_TBL0_ROWS;
2798 table->step = FR_BZ_RX_FILTER_TBL0_STEP;
2799 }
2800
2801 if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
2802 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_MAC];
2803 table->id = EFX_FARCH_FILTER_TABLE_RX_MAC;
2804 table->offset = FR_CZ_RX_MAC_FILTER_TBL0;
2805 table->size = FR_CZ_RX_MAC_FILTER_TBL0_ROWS;
2806 table->step = FR_CZ_RX_MAC_FILTER_TBL0_STEP;
2807
2808 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_DEF];
2809 table->id = EFX_FARCH_FILTER_TABLE_RX_DEF;
2810 table->size = EFX_FARCH_FILTER_SIZE_RX_DEF;
2811
2812 table = &state->table[EFX_FARCH_FILTER_TABLE_TX_MAC];
2813 table->id = EFX_FARCH_FILTER_TABLE_TX_MAC;
2814 table->offset = FR_CZ_TX_MAC_FILTER_TBL0;
2815 table->size = FR_CZ_TX_MAC_FILTER_TBL0_ROWS;
2816 table->step = FR_CZ_TX_MAC_FILTER_TBL0_STEP;
2817 }
2818
2819 for (table_id = 0; table_id < EFX_FARCH_FILTER_TABLE_COUNT; table_id++) {
2820 table = &state->table[table_id];
2821 if (table->size == 0)
2822 continue;
2823 table->used_bitmap = kcalloc(BITS_TO_LONGS(table->size),
2824 sizeof(unsigned long),
2825 GFP_KERNEL);
2826 if (!table->used_bitmap)
2827 goto fail;
2828 table->spec = vzalloc(table->size * sizeof(*table->spec));
2829 if (!table->spec)
2830 goto fail;
2831 }
2832
2833 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_DEF];
2834 if (table->size) {
2835 /* RX default filters must always exist */
2836 struct efx_farch_filter_spec *spec;
2837 unsigned i;
2838
2839 for (i = 0; i < EFX_FARCH_FILTER_SIZE_RX_DEF; i++) {
2840 spec = &table->spec[i];
2841 spec->type = EFX_FARCH_FILTER_UC_DEF + i;
2842 efx_farch_filter_init_rx_auto(efx, spec);
2843 __set_bit(i, table->used_bitmap);
2844 }
2845 }
2846
2847 efx_farch_filter_push_rx_config(efx);
2848
2849 return 0;
2850
2851fail:
2852 efx_farch_filter_table_remove(efx);
2853 return -ENOMEM;
2854}
2855
2856/* Update scatter enable flags for filters pointing to our own RX queues */
2857void efx_farch_filter_update_rx_scatter(struct efx_nic *efx)
2858{
2859 struct efx_farch_filter_state *state = efx->filter_state;
2860 enum efx_farch_filter_table_id table_id;
2861 struct efx_farch_filter_table *table;
2862 efx_oword_t filter;
2863 unsigned int filter_idx;
2864
2865 spin_lock_bh(&efx->filter_lock);
2866
2867 for (table_id = EFX_FARCH_FILTER_TABLE_RX_IP;
2868 table_id <= EFX_FARCH_FILTER_TABLE_RX_DEF;
2869 table_id++) {
2870 table = &state->table[table_id];
2871
2872 for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
2873 if (!test_bit(filter_idx, table->used_bitmap) ||
2874 table->spec[filter_idx].dmaq_id >=
2875 efx->n_rx_channels)
2876 continue;
2877
2878 if (efx->rx_scatter)
2879 table->spec[filter_idx].flags |=
2880 EFX_FILTER_FLAG_RX_SCATTER;
2881 else
2882 table->spec[filter_idx].flags &=
2883 ~EFX_FILTER_FLAG_RX_SCATTER;
2884
2885 if (table_id == EFX_FARCH_FILTER_TABLE_RX_DEF)
2886 /* Pushed by efx_farch_filter_push_rx_config() */
2887 continue;
2888
2889 efx_farch_filter_build(&filter, &table->spec[filter_idx]);
2890 efx_writeo(efx, &filter,
2891 table->offset + table->step * filter_idx);
2892 }
2893 }
2894
2895 efx_farch_filter_push_rx_config(efx);
2896
2897 spin_unlock_bh(&efx->filter_lock);
2898}
2899
2900#ifdef CONFIG_RFS_ACCEL
2901
2902s32 efx_farch_filter_rfs_insert(struct efx_nic *efx,
2903 struct efx_filter_spec *gen_spec)
2904{
2905 return efx_farch_filter_insert(efx, gen_spec, true);
2906}
2907
2908bool efx_farch_filter_rfs_expire_one(struct efx_nic *efx, u32 flow_id,
2909 unsigned int index)
2910{
2911 struct efx_farch_filter_state *state = efx->filter_state;
2912 struct efx_farch_filter_table *table =
2913 &state->table[EFX_FARCH_FILTER_TABLE_RX_IP];
2914
2915 if (test_bit(index, table->used_bitmap) &&
2916 table->spec[index].priority == EFX_FILTER_PRI_HINT &&
2917 rps_may_expire_flow(efx->net_dev, table->spec[index].dmaq_id,
2918 flow_id, index)) {
2919 efx_farch_filter_table_clear_entry(efx, table, index);
2920 return true;
2921 }
2922
2923 return false;
2924}
2925
2926#endif /* CONFIG_RFS_ACCEL */
2927
2928void efx_farch_filter_sync_rx_mode(struct efx_nic *efx)
2929{
2930 struct net_device *net_dev = efx->net_dev;
2931 struct netdev_hw_addr *ha;
2932 union efx_multicast_hash *mc_hash = &efx->multicast_hash;
2933 u32 crc;
2934 int bit;
2935
2936 netif_addr_lock_bh(net_dev);
2937
2938 efx->unicast_filter = !(net_dev->flags & IFF_PROMISC);
2939
2940 /* Build multicast hash table */
2941 if (net_dev->flags & (IFF_PROMISC | IFF_ALLMULTI)) {
2942 memset(mc_hash, 0xff, sizeof(*mc_hash));
2943 } else {
2944 memset(mc_hash, 0x00, sizeof(*mc_hash));
2945 netdev_for_each_mc_addr(ha, net_dev) {
2946 crc = ether_crc_le(ETH_ALEN, ha->addr);
2947 bit = crc & (EFX_MCAST_HASH_ENTRIES - 1);
2948 __set_bit_le(bit, mc_hash);
2949 }
2950
2951 /* Broadcast packets go through the multicast hash filter.
2952 * ether_crc_le() of the broadcast address is 0xbe2612ff
2953 * so we always add bit 0xff to the mask.
2954 */
2955 __set_bit_le(0xff, mc_hash);
2956 }
2957
2958 netif_addr_unlock_bh(net_dev);
2959}
1// SPDX-License-Identifier: GPL-2.0-only
2/****************************************************************************
3 * Driver for Solarflare network controllers and boards
4 * Copyright 2005-2006 Fen Systems Ltd.
5 * Copyright 2006-2013 Solarflare Communications Inc.
6 */
7
8#include <linux/bitops.h>
9#include <linux/delay.h>
10#include <linux/interrupt.h>
11#include <linux/pci.h>
12#include <linux/module.h>
13#include <linux/seq_file.h>
14#include <linux/crc32.h>
15#include "net_driver.h"
16#include "bitfield.h"
17#include "efx.h"
18#include "rx_common.h"
19#include "nic.h"
20#include "farch_regs.h"
21#include "sriov.h"
22#include "siena_sriov.h"
23#include "io.h"
24#include "workarounds.h"
25
26/* Falcon-architecture (SFC9000-family) support */
27
28/**************************************************************************
29 *
30 * Configurable values
31 *
32 **************************************************************************
33 */
34
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
40 * goes idle.
41 */
42#define TX_DC_ENTRIES 16
43#define TX_DC_ENTRIES_ORDER 1
44
45#define RX_DC_ENTRIES 64
46#define RX_DC_ENTRIES_ORDER 3
47
48/* If EFX_MAX_INT_ERRORS internal errors occur within
49 * EFX_INT_ERROR_EXPIRE seconds, we consider the NIC broken and
50 * disable it.
51 */
52#define EFX_INT_ERROR_EXPIRE 3600
53#define EFX_MAX_INT_ERRORS 5
54
55/* Depth of RX flush request fifo */
56#define EFX_RX_FLUSH_COUNT 4
57
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
63
64#define _EFX_CHANNEL_MAGIC(_code, _data) ((_code) << 8 | (_data))
65#define _EFX_CHANNEL_MAGIC_CODE(_magic) ((_magic) >> 8)
66
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, \
77 (_tx_queue)->queue)
78
79static void efx_farch_magic_event(struct efx_channel *channel, u32 magic);
80
81/**************************************************************************
82 *
83 * Hardware access
84 *
85 **************************************************************************/
86
87static inline void efx_write_buf_tbl(struct efx_nic *efx, efx_qword_t *value,
88 unsigned int index)
89{
90 efx_sram_writeq(efx, efx->membase + efx->type->buf_tbl_base,
91 value, index);
92}
93
94static bool efx_masked_compare_oword(const efx_oword_t *a, const efx_oword_t *b,
95 const efx_oword_t *mask)
96{
97 return ((a->u64[0] ^ b->u64[0]) & mask->u64[0]) ||
98 ((a->u64[1] ^ b->u64[1]) & mask->u64[1]);
99}
100
101int efx_farch_test_registers(struct efx_nic *efx,
102 const struct efx_farch_register_test *regs,
103 size_t n_regs)
104{
105 unsigned address = 0;
106 int i, j;
107 efx_oword_t mask, imask, original, reg, buf;
108
109 for (i = 0; i < n_regs; ++i) {
110 address = regs[i].address;
111 mask = imask = regs[i].mask;
112 EFX_INVERT_OWORD(imask);
113
114 efx_reado(efx, &original, address);
115
116 /* bit sweep on and off */
117 for (j = 0; j < 128; j++) {
118 if (!EFX_EXTRACT_OWORD32(mask, j, j))
119 continue;
120
121 /* Test this testable bit can be set in isolation */
122 EFX_AND_OWORD(reg, original, mask);
123 EFX_SET_OWORD32(reg, j, j, 1);
124
125 efx_writeo(efx, ®, address);
126 efx_reado(efx, &buf, address);
127
128 if (efx_masked_compare_oword(®, &buf, &mask))
129 goto fail;
130
131 /* Test this testable bit can be cleared in isolation */
132 EFX_OR_OWORD(reg, original, mask);
133 EFX_SET_OWORD32(reg, j, j, 0);
134
135 efx_writeo(efx, ®, address);
136 efx_reado(efx, &buf, address);
137
138 if (efx_masked_compare_oword(®, &buf, &mask))
139 goto fail;
140 }
141
142 efx_writeo(efx, &original, address);
143 }
144
145 return 0;
146
147fail:
148 netif_err(efx, hw, efx->net_dev,
149 "wrote "EFX_OWORD_FMT" read "EFX_OWORD_FMT
150 " at address 0x%x mask "EFX_OWORD_FMT"\n", EFX_OWORD_VAL(reg),
151 EFX_OWORD_VAL(buf), address, EFX_OWORD_VAL(mask));
152 return -EIO;
153}
154
155/**************************************************************************
156 *
157 * Special buffer handling
158 * Special buffers are used for event queues and the TX and RX
159 * descriptor rings.
160 *
161 *************************************************************************/
162
163/*
164 * Initialise a special buffer
165 *
166 * This will define a buffer (previously allocated via
167 * efx_alloc_special_buffer()) in the buffer table, allowing
168 * it to be used for event queues, descriptor rings etc.
169 */
170static void
171efx_init_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
172{
173 efx_qword_t buf_desc;
174 unsigned int index;
175 dma_addr_t dma_addr;
176 int i;
177
178 EFX_WARN_ON_PARANOID(!buffer->buf.addr);
179
180 /* Write buffer descriptors to NIC */
181 for (i = 0; i < buffer->entries; i++) {
182 index = buffer->index + i;
183 dma_addr = buffer->buf.dma_addr + (i * EFX_BUF_SIZE);
184 netif_dbg(efx, probe, efx->net_dev,
185 "mapping special buffer %d at %llx\n",
186 index, (unsigned long long)dma_addr);
187 EFX_POPULATE_QWORD_3(buf_desc,
188 FRF_AZ_BUF_ADR_REGION, 0,
189 FRF_AZ_BUF_ADR_FBUF, dma_addr >> 12,
190 FRF_AZ_BUF_OWNER_ID_FBUF, 0);
191 efx_write_buf_tbl(efx, &buf_desc, index);
192 }
193}
194
195/* Unmaps a buffer and clears the buffer table entries */
196static void
197efx_fini_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
198{
199 efx_oword_t buf_tbl_upd;
200 unsigned int start = buffer->index;
201 unsigned int end = (buffer->index + buffer->entries - 1);
202
203 if (!buffer->entries)
204 return;
205
206 netif_dbg(efx, hw, efx->net_dev, "unmapping special buffers %d-%d\n",
207 buffer->index, buffer->index + buffer->entries - 1);
208
209 EFX_POPULATE_OWORD_4(buf_tbl_upd,
210 FRF_AZ_BUF_UPD_CMD, 0,
211 FRF_AZ_BUF_CLR_CMD, 1,
212 FRF_AZ_BUF_CLR_END_ID, end,
213 FRF_AZ_BUF_CLR_START_ID, start);
214 efx_writeo(efx, &buf_tbl_upd, FR_AZ_BUF_TBL_UPD);
215}
216
217/*
218 * Allocate a new special buffer
219 *
220 * This allocates memory for a new buffer, clears it and allocates a
221 * new buffer ID range. It does not write into the buffer table.
222 *
223 * This call will allocate 4KB buffers, since 8KB buffers can't be
224 * used for event queues and descriptor rings.
225 */
226static int efx_alloc_special_buffer(struct efx_nic *efx,
227 struct efx_special_buffer *buffer,
228 unsigned int len)
229{
230#ifdef CONFIG_SFC_SRIOV
231 struct siena_nic_data *nic_data = efx->nic_data;
232#endif
233 len = ALIGN(len, EFX_BUF_SIZE);
234
235 if (efx_nic_alloc_buffer(efx, &buffer->buf, len, GFP_KERNEL))
236 return -ENOMEM;
237 buffer->entries = len / EFX_BUF_SIZE;
238 BUG_ON(buffer->buf.dma_addr & (EFX_BUF_SIZE - 1));
239
240 /* Select new buffer ID */
241 buffer->index = efx->next_buffer_table;
242 efx->next_buffer_table += buffer->entries;
243#ifdef CONFIG_SFC_SRIOV
244 BUG_ON(efx_siena_sriov_enabled(efx) &&
245 nic_data->vf_buftbl_base < efx->next_buffer_table);
246#endif
247
248 netif_dbg(efx, probe, efx->net_dev,
249 "allocating special buffers %d-%d at %llx+%x "
250 "(virt %p phys %llx)\n", buffer->index,
251 buffer->index + buffer->entries - 1,
252 (u64)buffer->buf.dma_addr, len,
253 buffer->buf.addr, (u64)virt_to_phys(buffer->buf.addr));
254
255 return 0;
256}
257
258static void
259efx_free_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
260{
261 if (!buffer->buf.addr)
262 return;
263
264 netif_dbg(efx, hw, efx->net_dev,
265 "deallocating special buffers %d-%d at %llx+%x "
266 "(virt %p phys %llx)\n", buffer->index,
267 buffer->index + buffer->entries - 1,
268 (u64)buffer->buf.dma_addr, buffer->buf.len,
269 buffer->buf.addr, (u64)virt_to_phys(buffer->buf.addr));
270
271 efx_nic_free_buffer(efx, &buffer->buf);
272 buffer->entries = 0;
273}
274
275/**************************************************************************
276 *
277 * TX path
278 *
279 **************************************************************************/
280
281/* This writes to the TX_DESC_WPTR; write pointer for TX descriptor ring */
282static inline void efx_farch_notify_tx_desc(struct efx_tx_queue *tx_queue)
283{
284 unsigned write_ptr;
285 efx_dword_t reg;
286
287 write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
288 EFX_POPULATE_DWORD_1(reg, FRF_AZ_TX_DESC_WPTR_DWORD, write_ptr);
289 efx_writed_page(tx_queue->efx, ®,
290 FR_AZ_TX_DESC_UPD_DWORD_P0, tx_queue->queue);
291}
292
293/* Write pointer and first descriptor for TX descriptor ring */
294static inline void efx_farch_push_tx_desc(struct efx_tx_queue *tx_queue,
295 const efx_qword_t *txd)
296{
297 unsigned write_ptr;
298 efx_oword_t reg;
299
300 BUILD_BUG_ON(FRF_AZ_TX_DESC_LBN != 0);
301 BUILD_BUG_ON(FR_AA_TX_DESC_UPD_KER != FR_BZ_TX_DESC_UPD_P0);
302
303 write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
304 EFX_POPULATE_OWORD_2(reg, FRF_AZ_TX_DESC_PUSH_CMD, true,
305 FRF_AZ_TX_DESC_WPTR, write_ptr);
306 reg.qword[0] = *txd;
307 efx_writeo_page(tx_queue->efx, ®,
308 FR_BZ_TX_DESC_UPD_P0, tx_queue->queue);
309}
310
311
312/* For each entry inserted into the software descriptor ring, create a
313 * descriptor in the hardware TX descriptor ring (in host memory), and
314 * write a doorbell.
315 */
316void efx_farch_tx_write(struct efx_tx_queue *tx_queue)
317{
318 struct efx_tx_buffer *buffer;
319 efx_qword_t *txd;
320 unsigned write_ptr;
321 unsigned old_write_count = tx_queue->write_count;
322
323 tx_queue->xmit_more_available = false;
324 if (unlikely(tx_queue->write_count == tx_queue->insert_count))
325 return;
326
327 do {
328 write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
329 buffer = &tx_queue->buffer[write_ptr];
330 txd = efx_tx_desc(tx_queue, write_ptr);
331 ++tx_queue->write_count;
332
333 EFX_WARN_ON_ONCE_PARANOID(buffer->flags & EFX_TX_BUF_OPTION);
334
335 /* Create TX descriptor ring entry */
336 BUILD_BUG_ON(EFX_TX_BUF_CONT != 1);
337 EFX_POPULATE_QWORD_4(*txd,
338 FSF_AZ_TX_KER_CONT,
339 buffer->flags & EFX_TX_BUF_CONT,
340 FSF_AZ_TX_KER_BYTE_COUNT, buffer->len,
341 FSF_AZ_TX_KER_BUF_REGION, 0,
342 FSF_AZ_TX_KER_BUF_ADDR, buffer->dma_addr);
343 } while (tx_queue->write_count != tx_queue->insert_count);
344
345 wmb(); /* Ensure descriptors are written before they are fetched */
346
347 if (efx_nic_may_push_tx_desc(tx_queue, old_write_count)) {
348 txd = efx_tx_desc(tx_queue,
349 old_write_count & tx_queue->ptr_mask);
350 efx_farch_push_tx_desc(tx_queue, txd);
351 ++tx_queue->pushes;
352 } else {
353 efx_farch_notify_tx_desc(tx_queue);
354 }
355}
356
357unsigned int efx_farch_tx_limit_len(struct efx_tx_queue *tx_queue,
358 dma_addr_t dma_addr, unsigned int len)
359{
360 /* Don't cross 4K boundaries with descriptors. */
361 unsigned int limit = (~dma_addr & (EFX_PAGE_SIZE - 1)) + 1;
362
363 len = min(limit, len);
364
365 return len;
366}
367
368
369/* Allocate hardware resources for a TX queue */
370int efx_farch_tx_probe(struct efx_tx_queue *tx_queue)
371{
372 struct efx_nic *efx = tx_queue->efx;
373 unsigned entries;
374
375 entries = tx_queue->ptr_mask + 1;
376 return efx_alloc_special_buffer(efx, &tx_queue->txd,
377 entries * sizeof(efx_qword_t));
378}
379
380void efx_farch_tx_init(struct efx_tx_queue *tx_queue)
381{
382 int csum = tx_queue->label & EFX_TXQ_TYPE_OFFLOAD;
383 struct efx_nic *efx = tx_queue->efx;
384 efx_oword_t reg;
385
386 /* Pin TX descriptor ring */
387 efx_init_special_buffer(efx, &tx_queue->txd);
388
389 /* Push TX descriptor ring to card */
390 EFX_POPULATE_OWORD_10(reg,
391 FRF_AZ_TX_DESCQ_EN, 1,
392 FRF_AZ_TX_ISCSI_DDIG_EN, 0,
393 FRF_AZ_TX_ISCSI_HDIG_EN, 0,
394 FRF_AZ_TX_DESCQ_BUF_BASE_ID, tx_queue->txd.index,
395 FRF_AZ_TX_DESCQ_EVQ_ID,
396 tx_queue->channel->channel,
397 FRF_AZ_TX_DESCQ_OWNER_ID, 0,
398 FRF_AZ_TX_DESCQ_LABEL, tx_queue->label,
399 FRF_AZ_TX_DESCQ_SIZE,
400 __ffs(tx_queue->txd.entries),
401 FRF_AZ_TX_DESCQ_TYPE, 0,
402 FRF_BZ_TX_NON_IP_DROP_DIS, 1);
403
404 EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_IP_CHKSM_DIS, !csum);
405 EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_TCP_CHKSM_DIS, !csum);
406
407 efx_writeo_table(efx, ®, efx->type->txd_ptr_tbl_base,
408 tx_queue->queue);
409
410 EFX_POPULATE_OWORD_1(reg,
411 FRF_BZ_TX_PACE,
412 (tx_queue->label & EFX_TXQ_TYPE_HIGHPRI) ?
413 FFE_BZ_TX_PACE_OFF :
414 FFE_BZ_TX_PACE_RESERVED);
415 efx_writeo_table(efx, ®, FR_BZ_TX_PACE_TBL, tx_queue->queue);
416}
417
418static void efx_farch_flush_tx_queue(struct efx_tx_queue *tx_queue)
419{
420 struct efx_nic *efx = tx_queue->efx;
421 efx_oword_t tx_flush_descq;
422
423 WARN_ON(atomic_read(&tx_queue->flush_outstanding));
424 atomic_set(&tx_queue->flush_outstanding, 1);
425
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);
430}
431
432void efx_farch_tx_fini(struct efx_tx_queue *tx_queue)
433{
434 struct efx_nic *efx = tx_queue->efx;
435 efx_oword_t tx_desc_ptr;
436
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,
440 tx_queue->queue);
441
442 /* Unpin TX descriptor ring */
443 efx_fini_special_buffer(efx, &tx_queue->txd);
444}
445
446/* Free buffers backing TX queue */
447void efx_farch_tx_remove(struct efx_tx_queue *tx_queue)
448{
449 efx_free_special_buffer(tx_queue->efx, &tx_queue->txd);
450}
451
452/**************************************************************************
453 *
454 * RX path
455 *
456 **************************************************************************/
457
458/* This creates an entry in the RX descriptor queue */
459static inline void
460efx_farch_build_rx_desc(struct efx_rx_queue *rx_queue, unsigned index)
461{
462 struct efx_rx_buffer *rx_buf;
463 efx_qword_t *rxd;
464
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,
469 rx_buf->len -
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);
473}
474
475/* This writes to the RX_DESC_WPTR register for the specified receive
476 * descriptor ring.
477 */
478void efx_farch_rx_write(struct efx_rx_queue *rx_queue)
479{
480 struct efx_nic *efx = rx_queue->efx;
481 efx_dword_t reg;
482 unsigned write_ptr;
483
484 while (rx_queue->notified_count != rx_queue->added_count) {
485 efx_farch_build_rx_desc(
486 rx_queue,
487 rx_queue->notified_count & rx_queue->ptr_mask);
488 ++rx_queue->notified_count;
489 }
490
491 wmb();
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));
496}
497
498int efx_farch_rx_probe(struct efx_rx_queue *rx_queue)
499{
500 struct efx_nic *efx = rx_queue->efx;
501 unsigned entries;
502
503 entries = rx_queue->ptr_mask + 1;
504 return efx_alloc_special_buffer(efx, &rx_queue->rxd,
505 entries * sizeof(efx_qword_t));
506}
507
508void efx_farch_rx_init(struct efx_rx_queue *rx_queue)
509{
510 efx_oword_t rx_desc_ptr;
511 struct efx_nic *efx = rx_queue->efx;
512 bool jumbo_en;
513
514 /* For kernel-mode queues in Siena, the JUMBO flag enables scatter. */
515 jumbo_en = efx->rx_scatter;
516
517 netif_dbg(efx, hw, efx->net_dev,
518 "RX queue %d ring in special buffers %d-%d\n",
519 efx_rx_queue_index(rx_queue), rx_queue->rxd.index,
520 rx_queue->rxd.index + rx_queue->rxd.entries - 1);
521
522 rx_queue->scatter_n = 0;
523
524 /* Pin RX descriptor ring */
525 efx_init_special_buffer(efx, &rx_queue->rxd);
526
527 /* Push RX descriptor ring to card */
528 EFX_POPULATE_OWORD_10(rx_desc_ptr,
529 FRF_AZ_RX_ISCSI_DDIG_EN, true,
530 FRF_AZ_RX_ISCSI_HDIG_EN, true,
531 FRF_AZ_RX_DESCQ_BUF_BASE_ID, rx_queue->rxd.index,
532 FRF_AZ_RX_DESCQ_EVQ_ID,
533 efx_rx_queue_channel(rx_queue)->channel,
534 FRF_AZ_RX_DESCQ_OWNER_ID, 0,
535 FRF_AZ_RX_DESCQ_LABEL,
536 efx_rx_queue_index(rx_queue),
537 FRF_AZ_RX_DESCQ_SIZE,
538 __ffs(rx_queue->rxd.entries),
539 FRF_AZ_RX_DESCQ_TYPE, 0 /* kernel queue */ ,
540 FRF_AZ_RX_DESCQ_JUMBO, jumbo_en,
541 FRF_AZ_RX_DESCQ_EN, 1);
542 efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
543 efx_rx_queue_index(rx_queue));
544}
545
546static void efx_farch_flush_rx_queue(struct efx_rx_queue *rx_queue)
547{
548 struct efx_nic *efx = rx_queue->efx;
549 efx_oword_t rx_flush_descq;
550
551 EFX_POPULATE_OWORD_2(rx_flush_descq,
552 FRF_AZ_RX_FLUSH_DESCQ_CMD, 1,
553 FRF_AZ_RX_FLUSH_DESCQ,
554 efx_rx_queue_index(rx_queue));
555 efx_writeo(efx, &rx_flush_descq, FR_AZ_RX_FLUSH_DESCQ);
556}
557
558void efx_farch_rx_fini(struct efx_rx_queue *rx_queue)
559{
560 efx_oword_t rx_desc_ptr;
561 struct efx_nic *efx = rx_queue->efx;
562
563 /* Remove RX descriptor ring from card */
564 EFX_ZERO_OWORD(rx_desc_ptr);
565 efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
566 efx_rx_queue_index(rx_queue));
567
568 /* Unpin RX descriptor ring */
569 efx_fini_special_buffer(efx, &rx_queue->rxd);
570}
571
572/* Free buffers backing RX queue */
573void efx_farch_rx_remove(struct efx_rx_queue *rx_queue)
574{
575 efx_free_special_buffer(rx_queue->efx, &rx_queue->rxd);
576}
577
578/**************************************************************************
579 *
580 * Flush handling
581 *
582 **************************************************************************/
583
584/* efx_farch_flush_queues() must be woken up when all flushes are completed,
585 * or more RX flushes can be kicked off.
586 */
587static bool efx_farch_flush_wake(struct efx_nic *efx)
588{
589 /* Ensure that all updates are visible to efx_farch_flush_queues() */
590 smp_mb();
591
592 return (atomic_read(&efx->active_queues) == 0 ||
593 (atomic_read(&efx->rxq_flush_outstanding) < EFX_RX_FLUSH_COUNT
594 && atomic_read(&efx->rxq_flush_pending) > 0));
595}
596
597static bool efx_check_tx_flush_complete(struct efx_nic *efx)
598{
599 bool i = true;
600 efx_oword_t txd_ptr_tbl;
601 struct efx_channel *channel;
602 struct efx_tx_queue *tx_queue;
603
604 efx_for_each_channel(channel, efx) {
605 efx_for_each_channel_tx_queue(tx_queue, channel) {
606 efx_reado_table(efx, &txd_ptr_tbl,
607 FR_BZ_TX_DESC_PTR_TBL, tx_queue->queue);
608 if (EFX_OWORD_FIELD(txd_ptr_tbl,
609 FRF_AZ_TX_DESCQ_FLUSH) ||
610 EFX_OWORD_FIELD(txd_ptr_tbl,
611 FRF_AZ_TX_DESCQ_EN)) {
612 netif_dbg(efx, hw, efx->net_dev,
613 "flush did not complete on TXQ %d\n",
614 tx_queue->queue);
615 i = false;
616 } else if (atomic_cmpxchg(&tx_queue->flush_outstanding,
617 1, 0)) {
618 /* The flush is complete, but we didn't
619 * receive a flush completion event
620 */
621 netif_dbg(efx, hw, efx->net_dev,
622 "flush complete on TXQ %d, so drain "
623 "the queue\n", tx_queue->queue);
624 /* Don't need to increment active_queues as it
625 * has already been incremented for the queues
626 * which did not drain
627 */
628 efx_farch_magic_event(channel,
629 EFX_CHANNEL_MAGIC_TX_DRAIN(
630 tx_queue));
631 }
632 }
633 }
634
635 return i;
636}
637
638/* Flush all the transmit queues, and continue flushing receive queues until
639 * they're all flushed. Wait for the DRAIN events to be received so that there
640 * are no more RX and TX events left on any channel. */
641static int efx_farch_do_flush(struct efx_nic *efx)
642{
643 unsigned timeout = msecs_to_jiffies(5000); /* 5s for all flushes and drains */
644 struct efx_channel *channel;
645 struct efx_rx_queue *rx_queue;
646 struct efx_tx_queue *tx_queue;
647 int rc = 0;
648
649 efx_for_each_channel(channel, efx) {
650 efx_for_each_channel_tx_queue(tx_queue, channel) {
651 efx_farch_flush_tx_queue(tx_queue);
652 }
653 efx_for_each_channel_rx_queue(rx_queue, channel) {
654 rx_queue->flush_pending = true;
655 atomic_inc(&efx->rxq_flush_pending);
656 }
657 }
658
659 while (timeout && atomic_read(&efx->active_queues) > 0) {
660 /* If SRIOV is enabled, then offload receive queue flushing to
661 * the firmware (though we will still have to poll for
662 * completion). If that fails, fall back to the old scheme.
663 */
664 if (efx_siena_sriov_enabled(efx)) {
665 rc = efx_mcdi_flush_rxqs(efx);
666 if (!rc)
667 goto wait;
668 }
669
670 /* The hardware supports four concurrent rx flushes, each of
671 * which may need to be retried if there is an outstanding
672 * descriptor fetch
673 */
674 efx_for_each_channel(channel, efx) {
675 efx_for_each_channel_rx_queue(rx_queue, channel) {
676 if (atomic_read(&efx->rxq_flush_outstanding) >=
677 EFX_RX_FLUSH_COUNT)
678 break;
679
680 if (rx_queue->flush_pending) {
681 rx_queue->flush_pending = false;
682 atomic_dec(&efx->rxq_flush_pending);
683 atomic_inc(&efx->rxq_flush_outstanding);
684 efx_farch_flush_rx_queue(rx_queue);
685 }
686 }
687 }
688
689 wait:
690 timeout = wait_event_timeout(efx->flush_wq,
691 efx_farch_flush_wake(efx),
692 timeout);
693 }
694
695 if (atomic_read(&efx->active_queues) &&
696 !efx_check_tx_flush_complete(efx)) {
697 netif_err(efx, hw, efx->net_dev, "failed to flush %d queues "
698 "(rx %d+%d)\n", atomic_read(&efx->active_queues),
699 atomic_read(&efx->rxq_flush_outstanding),
700 atomic_read(&efx->rxq_flush_pending));
701 rc = -ETIMEDOUT;
702
703 atomic_set(&efx->active_queues, 0);
704 atomic_set(&efx->rxq_flush_pending, 0);
705 atomic_set(&efx->rxq_flush_outstanding, 0);
706 }
707
708 return rc;
709}
710
711int efx_farch_fini_dmaq(struct efx_nic *efx)
712{
713 struct efx_channel *channel;
714 struct efx_tx_queue *tx_queue;
715 struct efx_rx_queue *rx_queue;
716 int rc = 0;
717
718 /* Do not attempt to write to the NIC during EEH recovery */
719 if (efx->state != STATE_RECOVERY) {
720 /* Only perform flush if DMA is enabled */
721 if (efx->pci_dev->is_busmaster) {
722 efx->type->prepare_flush(efx);
723 rc = efx_farch_do_flush(efx);
724 efx->type->finish_flush(efx);
725 }
726
727 efx_for_each_channel(channel, efx) {
728 efx_for_each_channel_rx_queue(rx_queue, channel)
729 efx_farch_rx_fini(rx_queue);
730 efx_for_each_channel_tx_queue(tx_queue, channel)
731 efx_farch_tx_fini(tx_queue);
732 }
733 }
734
735 return rc;
736}
737
738/* Reset queue and flush accounting after FLR
739 *
740 * One possible cause of FLR recovery is that DMA may be failing (eg. if bus
741 * mastering was disabled), in which case we don't receive (RXQ) flush
742 * completion events. This means that efx->rxq_flush_outstanding remained at 4
743 * after the FLR; also, efx->active_queues was non-zero (as no flush completion
744 * events were received, and we didn't go through efx_check_tx_flush_complete())
745 * If we don't fix this up, on the next call to efx_realloc_channels() we won't
746 * flush any RX queues because efx->rxq_flush_outstanding is at the limit of 4
747 * for batched flush requests; and the efx->active_queues gets messed up because
748 * we keep incrementing for the newly initialised queues, but it never went to
749 * zero previously. Then we get a timeout every time we try to restart the
750 * queues, as it doesn't go back to zero when we should be flushing the queues.
751 */
752void efx_farch_finish_flr(struct efx_nic *efx)
753{
754 atomic_set(&efx->rxq_flush_pending, 0);
755 atomic_set(&efx->rxq_flush_outstanding, 0);
756 atomic_set(&efx->active_queues, 0);
757}
758
759
760/**************************************************************************
761 *
762 * Event queue processing
763 * Event queues are processed by per-channel tasklets.
764 *
765 **************************************************************************/
766
767/* Update a channel's event queue's read pointer (RPTR) register
768 *
769 * This writes the EVQ_RPTR_REG register for the specified channel's
770 * event queue.
771 */
772void efx_farch_ev_read_ack(struct efx_channel *channel)
773{
774 efx_dword_t reg;
775 struct efx_nic *efx = channel->efx;
776
777 EFX_POPULATE_DWORD_1(reg, FRF_AZ_EVQ_RPTR,
778 channel->eventq_read_ptr & channel->eventq_mask);
779
780 /* For Falcon A1, EVQ_RPTR_KER is documented as having a step size
781 * of 4 bytes, but it is really 16 bytes just like later revisions.
782 */
783 efx_writed(efx, ®,
784 efx->type->evq_rptr_tbl_base +
785 FR_BZ_EVQ_RPTR_STEP * channel->channel);
786}
787
788/* Use HW to insert a SW defined event */
789void efx_farch_generate_event(struct efx_nic *efx, unsigned int evq,
790 efx_qword_t *event)
791{
792 efx_oword_t drv_ev_reg;
793
794 BUILD_BUG_ON(FRF_AZ_DRV_EV_DATA_LBN != 0 ||
795 FRF_AZ_DRV_EV_DATA_WIDTH != 64);
796 drv_ev_reg.u32[0] = event->u32[0];
797 drv_ev_reg.u32[1] = event->u32[1];
798 drv_ev_reg.u32[2] = 0;
799 drv_ev_reg.u32[3] = 0;
800 EFX_SET_OWORD_FIELD(drv_ev_reg, FRF_AZ_DRV_EV_QID, evq);
801 efx_writeo(efx, &drv_ev_reg, FR_AZ_DRV_EV);
802}
803
804static void efx_farch_magic_event(struct efx_channel *channel, u32 magic)
805{
806 efx_qword_t event;
807
808 EFX_POPULATE_QWORD_2(event, FSF_AZ_EV_CODE,
809 FSE_AZ_EV_CODE_DRV_GEN_EV,
810 FSF_AZ_DRV_GEN_EV_MAGIC, magic);
811 efx_farch_generate_event(channel->efx, channel->channel, &event);
812}
813
814/* Handle a transmit completion event
815 *
816 * The NIC batches TX completion events; the message we receive is of
817 * the form "complete all TX events up to this index".
818 */
819static void
820efx_farch_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
821{
822 unsigned int tx_ev_desc_ptr;
823 unsigned int tx_ev_q_label;
824 struct efx_tx_queue *tx_queue;
825 struct efx_nic *efx = channel->efx;
826
827 if (unlikely(READ_ONCE(efx->reset_pending)))
828 return;
829
830 if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
831 /* Transmit completion */
832 tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
833 tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
834 tx_queue = efx_channel_get_tx_queue(
835 channel, tx_ev_q_label % EFX_TXQ_TYPES);
836 efx_xmit_done(tx_queue, tx_ev_desc_ptr);
837 } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
838 /* Rewrite the FIFO write pointer */
839 tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
840 tx_queue = efx_channel_get_tx_queue(
841 channel, tx_ev_q_label % EFX_TXQ_TYPES);
842
843 netif_tx_lock(efx->net_dev);
844 efx_farch_notify_tx_desc(tx_queue);
845 netif_tx_unlock(efx->net_dev);
846 } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR)) {
847 efx_schedule_reset(efx, RESET_TYPE_DMA_ERROR);
848 } else {
849 netif_err(efx, tx_err, efx->net_dev,
850 "channel %d unexpected TX event "
851 EFX_QWORD_FMT"\n", channel->channel,
852 EFX_QWORD_VAL(*event));
853 }
854}
855
856/* Detect errors included in the rx_evt_pkt_ok bit. */
857static u16 efx_farch_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
858 const efx_qword_t *event)
859{
860 struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
861 struct efx_nic *efx = rx_queue->efx;
862 bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
863 bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
864 bool rx_ev_frm_trunc, rx_ev_tobe_disc;
865 bool rx_ev_other_err, rx_ev_pause_frm;
866 bool rx_ev_hdr_type, rx_ev_mcast_pkt;
867 unsigned rx_ev_pkt_type;
868
869 rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
870 rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
871 rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
872 rx_ev_pkt_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_TYPE);
873 rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
874 FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
875 rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
876 FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
877 rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
878 FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
879 rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
880 rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
881 rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
882
883 /* Every error apart from tobe_disc and pause_frm */
884 rx_ev_other_err = (rx_ev_tcp_udp_chksum_err |
885 rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
886 rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
887
888 /* Count errors that are not in MAC stats. Ignore expected
889 * checksum errors during self-test. */
890 if (rx_ev_frm_trunc)
891 ++channel->n_rx_frm_trunc;
892 else if (rx_ev_tobe_disc)
893 ++channel->n_rx_tobe_disc;
894 else if (!efx->loopback_selftest) {
895 if (rx_ev_ip_hdr_chksum_err)
896 ++channel->n_rx_ip_hdr_chksum_err;
897 else if (rx_ev_tcp_udp_chksum_err)
898 ++channel->n_rx_tcp_udp_chksum_err;
899 }
900
901 /* TOBE_DISC is expected on unicast mismatches; don't print out an
902 * error message. FRM_TRUNC indicates RXDP dropped the packet due
903 * to a FIFO overflow.
904 */
905#ifdef DEBUG
906 if (rx_ev_other_err && net_ratelimit()) {
907 netif_dbg(efx, rx_err, efx->net_dev,
908 " RX queue %d unexpected RX event "
909 EFX_QWORD_FMT "%s%s%s%s%s%s%s\n",
910 efx_rx_queue_index(rx_queue), EFX_QWORD_VAL(*event),
911 rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
912 rx_ev_ip_hdr_chksum_err ?
913 " [IP_HDR_CHKSUM_ERR]" : "",
914 rx_ev_tcp_udp_chksum_err ?
915 " [TCP_UDP_CHKSUM_ERR]" : "",
916 rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
917 rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
918 rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
919 rx_ev_pause_frm ? " [PAUSE]" : "");
920 }
921#endif
922
923 if (efx->net_dev->features & NETIF_F_RXALL)
924 /* don't discard frame for CRC error */
925 rx_ev_eth_crc_err = false;
926
927 /* The frame must be discarded if any of these are true. */
928 return (rx_ev_eth_crc_err | rx_ev_frm_trunc |
929 rx_ev_tobe_disc | rx_ev_pause_frm) ?
930 EFX_RX_PKT_DISCARD : 0;
931}
932
933/* Handle receive events that are not in-order. Return true if this
934 * can be handled as a partial packet discard, false if it's more
935 * serious.
936 */
937static bool
938efx_farch_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
939{
940 struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
941 struct efx_nic *efx = rx_queue->efx;
942 unsigned expected, dropped;
943
944 if (rx_queue->scatter_n &&
945 index == ((rx_queue->removed_count + rx_queue->scatter_n - 1) &
946 rx_queue->ptr_mask)) {
947 ++channel->n_rx_nodesc_trunc;
948 return true;
949 }
950
951 expected = rx_queue->removed_count & rx_queue->ptr_mask;
952 dropped = (index - expected) & rx_queue->ptr_mask;
953 netif_info(efx, rx_err, efx->net_dev,
954 "dropped %d events (index=%d expected=%d)\n",
955 dropped, index, expected);
956
957 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
958 return false;
959}
960
961/* Handle a packet received event
962 *
963 * The NIC gives a "discard" flag if it's a unicast packet with the
964 * wrong destination address
965 * Also "is multicast" and "matches multicast filter" flags can be used to
966 * discard non-matching multicast packets.
967 */
968static void
969efx_farch_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
970{
971 unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
972 unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
973 unsigned expected_ptr;
974 bool rx_ev_pkt_ok, rx_ev_sop, rx_ev_cont;
975 u16 flags;
976 struct efx_rx_queue *rx_queue;
977 struct efx_nic *efx = channel->efx;
978
979 if (unlikely(READ_ONCE(efx->reset_pending)))
980 return;
981
982 rx_ev_cont = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT);
983 rx_ev_sop = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP);
984 WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
985 channel->channel);
986
987 rx_queue = efx_channel_get_rx_queue(channel);
988
989 rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
990 expected_ptr = ((rx_queue->removed_count + rx_queue->scatter_n) &
991 rx_queue->ptr_mask);
992
993 /* Check for partial drops and other errors */
994 if (unlikely(rx_ev_desc_ptr != expected_ptr) ||
995 unlikely(rx_ev_sop != (rx_queue->scatter_n == 0))) {
996 if (rx_ev_desc_ptr != expected_ptr &&
997 !efx_farch_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr))
998 return;
999
1000 /* Discard all pending fragments */
1001 if (rx_queue->scatter_n) {
1002 efx_rx_packet(
1003 rx_queue,
1004 rx_queue->removed_count & rx_queue->ptr_mask,
1005 rx_queue->scatter_n, 0, EFX_RX_PKT_DISCARD);
1006 rx_queue->removed_count += rx_queue->scatter_n;
1007 rx_queue->scatter_n = 0;
1008 }
1009
1010 /* Return if there is no new fragment */
1011 if (rx_ev_desc_ptr != expected_ptr)
1012 return;
1013
1014 /* Discard new fragment if not SOP */
1015 if (!rx_ev_sop) {
1016 efx_rx_packet(
1017 rx_queue,
1018 rx_queue->removed_count & rx_queue->ptr_mask,
1019 1, 0, EFX_RX_PKT_DISCARD);
1020 ++rx_queue->removed_count;
1021 return;
1022 }
1023 }
1024
1025 ++rx_queue->scatter_n;
1026 if (rx_ev_cont)
1027 return;
1028
1029 rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
1030 rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
1031 rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
1032
1033 if (likely(rx_ev_pkt_ok)) {
1034 /* If packet is marked as OK then we can rely on the
1035 * hardware checksum and classification.
1036 */
1037 flags = 0;
1038 switch (rx_ev_hdr_type) {
1039 case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP:
1040 flags |= EFX_RX_PKT_TCP;
1041 fallthrough;
1042 case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP:
1043 flags |= EFX_RX_PKT_CSUMMED;
1044 fallthrough;
1045 case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_OTHER:
1046 case FSE_AZ_RX_EV_HDR_TYPE_OTHER:
1047 break;
1048 }
1049 } else {
1050 flags = efx_farch_handle_rx_not_ok(rx_queue, event);
1051 }
1052
1053 /* Detect multicast packets that didn't match the filter */
1054 rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
1055 if (rx_ev_mcast_pkt) {
1056 unsigned int rx_ev_mcast_hash_match =
1057 EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
1058
1059 if (unlikely(!rx_ev_mcast_hash_match)) {
1060 ++channel->n_rx_mcast_mismatch;
1061 flags |= EFX_RX_PKT_DISCARD;
1062 }
1063 }
1064
1065 channel->irq_mod_score += 2;
1066
1067 /* Handle received packet */
1068 efx_rx_packet(rx_queue,
1069 rx_queue->removed_count & rx_queue->ptr_mask,
1070 rx_queue->scatter_n, rx_ev_byte_cnt, flags);
1071 rx_queue->removed_count += rx_queue->scatter_n;
1072 rx_queue->scatter_n = 0;
1073}
1074
1075/* If this flush done event corresponds to a &struct efx_tx_queue, then
1076 * send an %EFX_CHANNEL_MAGIC_TX_DRAIN event to drain the event queue
1077 * of all transmit completions.
1078 */
1079static void
1080efx_farch_handle_tx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1081{
1082 struct efx_tx_queue *tx_queue;
1083 int qid;
1084
1085 qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1086 if (qid < EFX_TXQ_TYPES * (efx->n_tx_channels + efx->n_extra_tx_channels)) {
1087 tx_queue = efx_get_tx_queue(efx, qid / EFX_TXQ_TYPES,
1088 qid % EFX_TXQ_TYPES);
1089 if (atomic_cmpxchg(&tx_queue->flush_outstanding, 1, 0)) {
1090 efx_farch_magic_event(tx_queue->channel,
1091 EFX_CHANNEL_MAGIC_TX_DRAIN(tx_queue));
1092 }
1093 }
1094}
1095
1096/* If this flush done event corresponds to a &struct efx_rx_queue: If the flush
1097 * was successful then send an %EFX_CHANNEL_MAGIC_RX_DRAIN, otherwise add
1098 * the RX queue back to the mask of RX queues in need of flushing.
1099 */
1100static void
1101efx_farch_handle_rx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1102{
1103 struct efx_channel *channel;
1104 struct efx_rx_queue *rx_queue;
1105 int qid;
1106 bool failed;
1107
1108 qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
1109 failed = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
1110 if (qid >= efx->n_channels)
1111 return;
1112 channel = efx_get_channel(efx, qid);
1113 if (!efx_channel_has_rx_queue(channel))
1114 return;
1115 rx_queue = efx_channel_get_rx_queue(channel);
1116
1117 if (failed) {
1118 netif_info(efx, hw, efx->net_dev,
1119 "RXQ %d flush retry\n", qid);
1120 rx_queue->flush_pending = true;
1121 atomic_inc(&efx->rxq_flush_pending);
1122 } else {
1123 efx_farch_magic_event(efx_rx_queue_channel(rx_queue),
1124 EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue));
1125 }
1126 atomic_dec(&efx->rxq_flush_outstanding);
1127 if (efx_farch_flush_wake(efx))
1128 wake_up(&efx->flush_wq);
1129}
1130
1131static void
1132efx_farch_handle_drain_event(struct efx_channel *channel)
1133{
1134 struct efx_nic *efx = channel->efx;
1135
1136 WARN_ON(atomic_read(&efx->active_queues) == 0);
1137 atomic_dec(&efx->active_queues);
1138 if (efx_farch_flush_wake(efx))
1139 wake_up(&efx->flush_wq);
1140}
1141
1142static void efx_farch_handle_generated_event(struct efx_channel *channel,
1143 efx_qword_t *event)
1144{
1145 struct efx_nic *efx = channel->efx;
1146 struct efx_rx_queue *rx_queue =
1147 efx_channel_has_rx_queue(channel) ?
1148 efx_channel_get_rx_queue(channel) : NULL;
1149 unsigned magic, code;
1150
1151 magic = EFX_QWORD_FIELD(*event, FSF_AZ_DRV_GEN_EV_MAGIC);
1152 code = _EFX_CHANNEL_MAGIC_CODE(magic);
1153
1154 if (magic == EFX_CHANNEL_MAGIC_TEST(channel)) {
1155 channel->event_test_cpu = raw_smp_processor_id();
1156 } else if (rx_queue && magic == EFX_CHANNEL_MAGIC_FILL(rx_queue)) {
1157 /* The queue must be empty, so we won't receive any rx
1158 * events, so efx_process_channel() won't refill the
1159 * queue. Refill it here */
1160 efx_fast_push_rx_descriptors(rx_queue, true);
1161 } else if (rx_queue && magic == EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue)) {
1162 efx_farch_handle_drain_event(channel);
1163 } else if (code == _EFX_CHANNEL_MAGIC_TX_DRAIN) {
1164 efx_farch_handle_drain_event(channel);
1165 } else {
1166 netif_dbg(efx, hw, efx->net_dev, "channel %d received "
1167 "generated event "EFX_QWORD_FMT"\n",
1168 channel->channel, EFX_QWORD_VAL(*event));
1169 }
1170}
1171
1172static void
1173efx_farch_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
1174{
1175 struct efx_nic *efx = channel->efx;
1176 unsigned int ev_sub_code;
1177 unsigned int ev_sub_data;
1178
1179 ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
1180 ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1181
1182 switch (ev_sub_code) {
1183 case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
1184 netif_vdbg(efx, hw, efx->net_dev, "channel %d TXQ %d flushed\n",
1185 channel->channel, ev_sub_data);
1186 efx_farch_handle_tx_flush_done(efx, event);
1187#ifdef CONFIG_SFC_SRIOV
1188 efx_siena_sriov_tx_flush_done(efx, event);
1189#endif
1190 break;
1191 case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
1192 netif_vdbg(efx, hw, efx->net_dev, "channel %d RXQ %d flushed\n",
1193 channel->channel, ev_sub_data);
1194 efx_farch_handle_rx_flush_done(efx, event);
1195#ifdef CONFIG_SFC_SRIOV
1196 efx_siena_sriov_rx_flush_done(efx, event);
1197#endif
1198 break;
1199 case FSE_AZ_EVQ_INIT_DONE_EV:
1200 netif_dbg(efx, hw, efx->net_dev,
1201 "channel %d EVQ %d initialised\n",
1202 channel->channel, ev_sub_data);
1203 break;
1204 case FSE_AZ_SRM_UPD_DONE_EV:
1205 netif_vdbg(efx, hw, efx->net_dev,
1206 "channel %d SRAM update done\n", channel->channel);
1207 break;
1208 case FSE_AZ_WAKE_UP_EV:
1209 netif_vdbg(efx, hw, efx->net_dev,
1210 "channel %d RXQ %d wakeup event\n",
1211 channel->channel, ev_sub_data);
1212 break;
1213 case FSE_AZ_TIMER_EV:
1214 netif_vdbg(efx, hw, efx->net_dev,
1215 "channel %d RX queue %d timer expired\n",
1216 channel->channel, ev_sub_data);
1217 break;
1218 case FSE_AA_RX_RECOVER_EV:
1219 netif_err(efx, rx_err, efx->net_dev,
1220 "channel %d seen DRIVER RX_RESET event. "
1221 "Resetting.\n", channel->channel);
1222 atomic_inc(&efx->rx_reset);
1223 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1224 break;
1225 case FSE_BZ_RX_DSC_ERROR_EV:
1226 if (ev_sub_data < EFX_VI_BASE) {
1227 netif_err(efx, rx_err, efx->net_dev,
1228 "RX DMA Q %d reports descriptor fetch error."
1229 " RX Q %d is disabled.\n", ev_sub_data,
1230 ev_sub_data);
1231 efx_schedule_reset(efx, RESET_TYPE_DMA_ERROR);
1232 }
1233#ifdef CONFIG_SFC_SRIOV
1234 else
1235 efx_siena_sriov_desc_fetch_err(efx, ev_sub_data);
1236#endif
1237 break;
1238 case FSE_BZ_TX_DSC_ERROR_EV:
1239 if (ev_sub_data < EFX_VI_BASE) {
1240 netif_err(efx, tx_err, efx->net_dev,
1241 "TX DMA Q %d reports descriptor fetch error."
1242 " TX Q %d is disabled.\n", ev_sub_data,
1243 ev_sub_data);
1244 efx_schedule_reset(efx, RESET_TYPE_DMA_ERROR);
1245 }
1246#ifdef CONFIG_SFC_SRIOV
1247 else
1248 efx_siena_sriov_desc_fetch_err(efx, ev_sub_data);
1249#endif
1250 break;
1251 default:
1252 netif_vdbg(efx, hw, efx->net_dev,
1253 "channel %d unknown driver event code %d "
1254 "data %04x\n", channel->channel, ev_sub_code,
1255 ev_sub_data);
1256 break;
1257 }
1258}
1259
1260int efx_farch_ev_process(struct efx_channel *channel, int budget)
1261{
1262 struct efx_nic *efx = channel->efx;
1263 unsigned int read_ptr;
1264 efx_qword_t event, *p_event;
1265 int ev_code;
1266 int spent = 0;
1267
1268 if (budget <= 0)
1269 return spent;
1270
1271 read_ptr = channel->eventq_read_ptr;
1272
1273 for (;;) {
1274 p_event = efx_event(channel, read_ptr);
1275 event = *p_event;
1276
1277 if (!efx_event_present(&event))
1278 /* End of events */
1279 break;
1280
1281 netif_vdbg(channel->efx, intr, channel->efx->net_dev,
1282 "channel %d event is "EFX_QWORD_FMT"\n",
1283 channel->channel, EFX_QWORD_VAL(event));
1284
1285 /* Clear this event by marking it all ones */
1286 EFX_SET_QWORD(*p_event);
1287
1288 ++read_ptr;
1289
1290 ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
1291
1292 switch (ev_code) {
1293 case FSE_AZ_EV_CODE_RX_EV:
1294 efx_farch_handle_rx_event(channel, &event);
1295 if (++spent == budget)
1296 goto out;
1297 break;
1298 case FSE_AZ_EV_CODE_TX_EV:
1299 efx_farch_handle_tx_event(channel, &event);
1300 break;
1301 case FSE_AZ_EV_CODE_DRV_GEN_EV:
1302 efx_farch_handle_generated_event(channel, &event);
1303 break;
1304 case FSE_AZ_EV_CODE_DRIVER_EV:
1305 efx_farch_handle_driver_event(channel, &event);
1306 break;
1307#ifdef CONFIG_SFC_SRIOV
1308 case FSE_CZ_EV_CODE_USER_EV:
1309 efx_siena_sriov_event(channel, &event);
1310 break;
1311#endif
1312 case FSE_CZ_EV_CODE_MCDI_EV:
1313 efx_mcdi_process_event(channel, &event);
1314 break;
1315 case FSE_AZ_EV_CODE_GLOBAL_EV:
1316 if (efx->type->handle_global_event &&
1317 efx->type->handle_global_event(channel, &event))
1318 break;
1319 fallthrough;
1320 default:
1321 netif_err(channel->efx, hw, channel->efx->net_dev,
1322 "channel %d unknown event type %d (data "
1323 EFX_QWORD_FMT ")\n", channel->channel,
1324 ev_code, EFX_QWORD_VAL(event));
1325 }
1326 }
1327
1328out:
1329 channel->eventq_read_ptr = read_ptr;
1330 return spent;
1331}
1332
1333/* Allocate buffer table entries for event queue */
1334int efx_farch_ev_probe(struct efx_channel *channel)
1335{
1336 struct efx_nic *efx = channel->efx;
1337 unsigned entries;
1338
1339 entries = channel->eventq_mask + 1;
1340 return efx_alloc_special_buffer(efx, &channel->eventq,
1341 entries * sizeof(efx_qword_t));
1342}
1343
1344int efx_farch_ev_init(struct efx_channel *channel)
1345{
1346 efx_oword_t reg;
1347 struct efx_nic *efx = channel->efx;
1348
1349 netif_dbg(efx, hw, efx->net_dev,
1350 "channel %d event queue in special buffers %d-%d\n",
1351 channel->channel, channel->eventq.index,
1352 channel->eventq.index + channel->eventq.entries - 1);
1353
1354 EFX_POPULATE_OWORD_3(reg,
1355 FRF_CZ_TIMER_Q_EN, 1,
1356 FRF_CZ_HOST_NOTIFY_MODE, 0,
1357 FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
1358 efx_writeo_table(efx, ®, FR_BZ_TIMER_TBL, channel->channel);
1359
1360 /* Pin event queue buffer */
1361 efx_init_special_buffer(efx, &channel->eventq);
1362
1363 /* Fill event queue with all ones (i.e. empty events) */
1364 memset(channel->eventq.buf.addr, 0xff, channel->eventq.buf.len);
1365
1366 /* Push event queue to card */
1367 EFX_POPULATE_OWORD_3(reg,
1368 FRF_AZ_EVQ_EN, 1,
1369 FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
1370 FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
1371 efx_writeo_table(efx, ®, efx->type->evq_ptr_tbl_base,
1372 channel->channel);
1373
1374 return 0;
1375}
1376
1377void efx_farch_ev_fini(struct efx_channel *channel)
1378{
1379 efx_oword_t reg;
1380 struct efx_nic *efx = channel->efx;
1381
1382 /* Remove event queue from card */
1383 EFX_ZERO_OWORD(reg);
1384 efx_writeo_table(efx, ®, efx->type->evq_ptr_tbl_base,
1385 channel->channel);
1386 efx_writeo_table(efx, ®, FR_BZ_TIMER_TBL, channel->channel);
1387
1388 /* Unpin event queue */
1389 efx_fini_special_buffer(efx, &channel->eventq);
1390}
1391
1392/* Free buffers backing event queue */
1393void efx_farch_ev_remove(struct efx_channel *channel)
1394{
1395 efx_free_special_buffer(channel->efx, &channel->eventq);
1396}
1397
1398
1399void efx_farch_ev_test_generate(struct efx_channel *channel)
1400{
1401 efx_farch_magic_event(channel, EFX_CHANNEL_MAGIC_TEST(channel));
1402}
1403
1404void efx_farch_rx_defer_refill(struct efx_rx_queue *rx_queue)
1405{
1406 efx_farch_magic_event(efx_rx_queue_channel(rx_queue),
1407 EFX_CHANNEL_MAGIC_FILL(rx_queue));
1408}
1409
1410/**************************************************************************
1411 *
1412 * Hardware interrupts
1413 * The hardware interrupt handler does very little work; all the event
1414 * queue processing is carried out by per-channel tasklets.
1415 *
1416 **************************************************************************/
1417
1418/* Enable/disable/generate interrupts */
1419static inline void efx_farch_interrupts(struct efx_nic *efx,
1420 bool enabled, bool force)
1421{
1422 efx_oword_t int_en_reg_ker;
1423
1424 EFX_POPULATE_OWORD_3(int_en_reg_ker,
1425 FRF_AZ_KER_INT_LEVE_SEL, efx->irq_level,
1426 FRF_AZ_KER_INT_KER, force,
1427 FRF_AZ_DRV_INT_EN_KER, enabled);
1428 efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
1429}
1430
1431void efx_farch_irq_enable_master(struct efx_nic *efx)
1432{
1433 EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
1434 wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
1435
1436 efx_farch_interrupts(efx, true, false);
1437}
1438
1439void efx_farch_irq_disable_master(struct efx_nic *efx)
1440{
1441 /* Disable interrupts */
1442 efx_farch_interrupts(efx, false, false);
1443}
1444
1445/* Generate a test interrupt
1446 * Interrupt must already have been enabled, otherwise nasty things
1447 * may happen.
1448 */
1449int efx_farch_irq_test_generate(struct efx_nic *efx)
1450{
1451 efx_farch_interrupts(efx, true, true);
1452 return 0;
1453}
1454
1455/* Process a fatal interrupt
1456 * Disable bus mastering ASAP and schedule a reset
1457 */
1458irqreturn_t efx_farch_fatal_interrupt(struct efx_nic *efx)
1459{
1460 efx_oword_t *int_ker = efx->irq_status.addr;
1461 efx_oword_t fatal_intr;
1462 int error, mem_perr;
1463
1464 efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
1465 error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
1466
1467 netif_err(efx, hw, efx->net_dev, "SYSTEM ERROR "EFX_OWORD_FMT" status "
1468 EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
1469 EFX_OWORD_VAL(fatal_intr),
1470 error ? "disabling bus mastering" : "no recognised error");
1471
1472 /* If this is a memory parity error dump which blocks are offending */
1473 mem_perr = (EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER) ||
1474 EFX_OWORD_FIELD(fatal_intr, FRF_AZ_SRM_PERR_INT_KER));
1475 if (mem_perr) {
1476 efx_oword_t reg;
1477 efx_reado(efx, ®, FR_AZ_MEM_STAT);
1478 netif_err(efx, hw, efx->net_dev,
1479 "SYSTEM ERROR: memory parity error "EFX_OWORD_FMT"\n",
1480 EFX_OWORD_VAL(reg));
1481 }
1482
1483 /* Disable both devices */
1484 pci_clear_master(efx->pci_dev);
1485 efx_farch_irq_disable_master(efx);
1486
1487 /* Count errors and reset or disable the NIC accordingly */
1488 if (efx->int_error_count == 0 ||
1489 time_after(jiffies, efx->int_error_expire)) {
1490 efx->int_error_count = 0;
1491 efx->int_error_expire =
1492 jiffies + EFX_INT_ERROR_EXPIRE * HZ;
1493 }
1494 if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
1495 netif_err(efx, hw, efx->net_dev,
1496 "SYSTEM ERROR - reset scheduled\n");
1497 efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
1498 } else {
1499 netif_err(efx, hw, efx->net_dev,
1500 "SYSTEM ERROR - max number of errors seen."
1501 "NIC will be disabled\n");
1502 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1503 }
1504
1505 return IRQ_HANDLED;
1506}
1507
1508/* Handle a legacy interrupt
1509 * Acknowledges the interrupt and schedule event queue processing.
1510 */
1511irqreturn_t efx_farch_legacy_interrupt(int irq, void *dev_id)
1512{
1513 struct efx_nic *efx = dev_id;
1514 bool soft_enabled = READ_ONCE(efx->irq_soft_enabled);
1515 efx_oword_t *int_ker = efx->irq_status.addr;
1516 irqreturn_t result = IRQ_NONE;
1517 struct efx_channel *channel;
1518 efx_dword_t reg;
1519 u32 queues;
1520 int syserr;
1521
1522 /* Read the ISR which also ACKs the interrupts */
1523 efx_readd(efx, ®, FR_BZ_INT_ISR0);
1524 queues = EFX_EXTRACT_DWORD(reg, 0, 31);
1525
1526 /* Legacy interrupts are disabled too late by the EEH kernel
1527 * code. Disable them earlier.
1528 * If an EEH error occurred, the read will have returned all ones.
1529 */
1530 if (EFX_DWORD_IS_ALL_ONES(reg) && efx_try_recovery(efx) &&
1531 !efx->eeh_disabled_legacy_irq) {
1532 disable_irq_nosync(efx->legacy_irq);
1533 efx->eeh_disabled_legacy_irq = true;
1534 }
1535
1536 /* Handle non-event-queue sources */
1537 if (queues & (1U << efx->irq_level) && soft_enabled) {
1538 syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1539 if (unlikely(syserr))
1540 return efx_farch_fatal_interrupt(efx);
1541 efx->last_irq_cpu = raw_smp_processor_id();
1542 }
1543
1544 if (queues != 0) {
1545 efx->irq_zero_count = 0;
1546
1547 /* Schedule processing of any interrupting queues */
1548 if (likely(soft_enabled)) {
1549 efx_for_each_channel(channel, efx) {
1550 if (queues & 1)
1551 efx_schedule_channel_irq(channel);
1552 queues >>= 1;
1553 }
1554 }
1555 result = IRQ_HANDLED;
1556
1557 } else {
1558 efx_qword_t *event;
1559
1560 /* Legacy ISR read can return zero once (SF bug 15783) */
1561
1562 /* We can't return IRQ_HANDLED more than once on seeing ISR=0
1563 * because this might be a shared interrupt. */
1564 if (efx->irq_zero_count++ == 0)
1565 result = IRQ_HANDLED;
1566
1567 /* Ensure we schedule or rearm all event queues */
1568 if (likely(soft_enabled)) {
1569 efx_for_each_channel(channel, efx) {
1570 event = efx_event(channel,
1571 channel->eventq_read_ptr);
1572 if (efx_event_present(event))
1573 efx_schedule_channel_irq(channel);
1574 else
1575 efx_farch_ev_read_ack(channel);
1576 }
1577 }
1578 }
1579
1580 if (result == IRQ_HANDLED)
1581 netif_vdbg(efx, intr, efx->net_dev,
1582 "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
1583 irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
1584
1585 return result;
1586}
1587
1588/* Handle an MSI interrupt
1589 *
1590 * Handle an MSI hardware interrupt. This routine schedules event
1591 * queue processing. No interrupt acknowledgement cycle is necessary.
1592 * Also, we never need to check that the interrupt is for us, since
1593 * MSI interrupts cannot be shared.
1594 */
1595irqreturn_t efx_farch_msi_interrupt(int irq, void *dev_id)
1596{
1597 struct efx_msi_context *context = dev_id;
1598 struct efx_nic *efx = context->efx;
1599 efx_oword_t *int_ker = efx->irq_status.addr;
1600 int syserr;
1601
1602 netif_vdbg(efx, intr, efx->net_dev,
1603 "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
1604 irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
1605
1606 if (!likely(READ_ONCE(efx->irq_soft_enabled)))
1607 return IRQ_HANDLED;
1608
1609 /* Handle non-event-queue sources */
1610 if (context->index == efx->irq_level) {
1611 syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1612 if (unlikely(syserr))
1613 return efx_farch_fatal_interrupt(efx);
1614 efx->last_irq_cpu = raw_smp_processor_id();
1615 }
1616
1617 /* Schedule processing of the channel */
1618 efx_schedule_channel_irq(efx->channel[context->index]);
1619
1620 return IRQ_HANDLED;
1621}
1622
1623/* Setup RSS indirection table.
1624 * This maps from the hash value of the packet to RXQ
1625 */
1626void efx_farch_rx_push_indir_table(struct efx_nic *efx)
1627{
1628 size_t i = 0;
1629 efx_dword_t dword;
1630
1631 BUILD_BUG_ON(ARRAY_SIZE(efx->rss_context.rx_indir_table) !=
1632 FR_BZ_RX_INDIRECTION_TBL_ROWS);
1633
1634 for (i = 0; i < FR_BZ_RX_INDIRECTION_TBL_ROWS; i++) {
1635 EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
1636 efx->rss_context.rx_indir_table[i]);
1637 efx_writed(efx, &dword,
1638 FR_BZ_RX_INDIRECTION_TBL +
1639 FR_BZ_RX_INDIRECTION_TBL_STEP * i);
1640 }
1641}
1642
1643void efx_farch_rx_pull_indir_table(struct efx_nic *efx)
1644{
1645 size_t i = 0;
1646 efx_dword_t dword;
1647
1648 BUILD_BUG_ON(ARRAY_SIZE(efx->rss_context.rx_indir_table) !=
1649 FR_BZ_RX_INDIRECTION_TBL_ROWS);
1650
1651 for (i = 0; i < FR_BZ_RX_INDIRECTION_TBL_ROWS; i++) {
1652 efx_readd(efx, &dword,
1653 FR_BZ_RX_INDIRECTION_TBL +
1654 FR_BZ_RX_INDIRECTION_TBL_STEP * i);
1655 efx->rss_context.rx_indir_table[i] = EFX_DWORD_FIELD(dword, FRF_BZ_IT_QUEUE);
1656 }
1657}
1658
1659/* Looks at available SRAM resources and works out how many queues we
1660 * can support, and where things like descriptor caches should live.
1661 *
1662 * SRAM is split up as follows:
1663 * 0 buftbl entries for channels
1664 * efx->vf_buftbl_base buftbl entries for SR-IOV
1665 * efx->rx_dc_base RX descriptor caches
1666 * efx->tx_dc_base TX descriptor caches
1667 */
1668void efx_farch_dimension_resources(struct efx_nic *efx, unsigned sram_lim_qw)
1669{
1670 unsigned vi_count, buftbl_min, total_tx_channels;
1671
1672#ifdef CONFIG_SFC_SRIOV
1673 struct siena_nic_data *nic_data = efx->nic_data;
1674#endif
1675
1676 total_tx_channels = efx->n_tx_channels + efx->n_extra_tx_channels;
1677 /* Account for the buffer table entries backing the datapath channels
1678 * and the descriptor caches for those channels.
1679 */
1680 buftbl_min = ((efx->n_rx_channels * EFX_MAX_DMAQ_SIZE +
1681 total_tx_channels * EFX_TXQ_TYPES * EFX_MAX_DMAQ_SIZE +
1682 efx->n_channels * EFX_MAX_EVQ_SIZE)
1683 * sizeof(efx_qword_t) / EFX_BUF_SIZE);
1684 vi_count = max(efx->n_channels, total_tx_channels * EFX_TXQ_TYPES);
1685
1686#ifdef CONFIG_SFC_SRIOV
1687 if (efx->type->sriov_wanted) {
1688 if (efx->type->sriov_wanted(efx)) {
1689 unsigned vi_dc_entries, buftbl_free;
1690 unsigned entries_per_vf, vf_limit;
1691
1692 nic_data->vf_buftbl_base = buftbl_min;
1693
1694 vi_dc_entries = RX_DC_ENTRIES + TX_DC_ENTRIES;
1695 vi_count = max(vi_count, EFX_VI_BASE);
1696 buftbl_free = (sram_lim_qw - buftbl_min -
1697 vi_count * vi_dc_entries);
1698
1699 entries_per_vf = ((vi_dc_entries +
1700 EFX_VF_BUFTBL_PER_VI) *
1701 efx_vf_size(efx));
1702 vf_limit = min(buftbl_free / entries_per_vf,
1703 (1024U - EFX_VI_BASE) >> efx->vi_scale);
1704
1705 if (efx->vf_count > vf_limit) {
1706 netif_err(efx, probe, efx->net_dev,
1707 "Reducing VF count from from %d to %d\n",
1708 efx->vf_count, vf_limit);
1709 efx->vf_count = vf_limit;
1710 }
1711 vi_count += efx->vf_count * efx_vf_size(efx);
1712 }
1713 }
1714#endif
1715
1716 efx->tx_dc_base = sram_lim_qw - vi_count * TX_DC_ENTRIES;
1717 efx->rx_dc_base = efx->tx_dc_base - vi_count * RX_DC_ENTRIES;
1718}
1719
1720u32 efx_farch_fpga_ver(struct efx_nic *efx)
1721{
1722 efx_oword_t altera_build;
1723 efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
1724 return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
1725}
1726
1727void efx_farch_init_common(struct efx_nic *efx)
1728{
1729 efx_oword_t temp;
1730
1731 /* Set positions of descriptor caches in SRAM. */
1732 EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR, efx->tx_dc_base);
1733 efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
1734 EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR, efx->rx_dc_base);
1735 efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
1736
1737 /* Set TX descriptor cache size. */
1738 BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
1739 EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
1740 efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
1741
1742 /* Set RX descriptor cache size. Set low watermark to size-8, as
1743 * this allows most efficient prefetching.
1744 */
1745 BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
1746 EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
1747 efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
1748 EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
1749 efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
1750
1751 /* Program INT_KER address */
1752 EFX_POPULATE_OWORD_2(temp,
1753 FRF_AZ_NORM_INT_VEC_DIS_KER,
1754 EFX_INT_MODE_USE_MSI(efx),
1755 FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
1756 efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
1757
1758 if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
1759 /* Use an interrupt level unused by event queues */
1760 efx->irq_level = 0x1f;
1761 else
1762 /* Use a valid MSI-X vector */
1763 efx->irq_level = 0;
1764
1765 /* Enable all the genuinely fatal interrupts. (They are still
1766 * masked by the overall interrupt mask, controlled by
1767 * falcon_interrupts()).
1768 *
1769 * Note: All other fatal interrupts are enabled
1770 */
1771 EFX_POPULATE_OWORD_3(temp,
1772 FRF_AZ_ILL_ADR_INT_KER_EN, 1,
1773 FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
1774 FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
1775 EFX_SET_OWORD_FIELD(temp, FRF_CZ_SRAM_PERR_INT_P_KER_EN, 1);
1776 EFX_INVERT_OWORD(temp);
1777 efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
1778
1779 /* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
1780 * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
1781 */
1782 efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
1783 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
1784 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
1785 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
1786 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 1);
1787 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
1788 /* Enable SW_EV to inherit in char driver - assume harmless here */
1789 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
1790 /* Prefetch threshold 2 => fetch when descriptor cache half empty */
1791 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
1792 /* Disable hardware watchdog which can misfire */
1793 EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
1794 /* Squash TX of packets of 16 bytes or less */
1795 EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
1796 efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
1797
1798 EFX_POPULATE_OWORD_4(temp,
1799 /* Default values */
1800 FRF_BZ_TX_PACE_SB_NOT_AF, 0x15,
1801 FRF_BZ_TX_PACE_SB_AF, 0xb,
1802 FRF_BZ_TX_PACE_FB_BASE, 0,
1803 /* Allow large pace values in the fast bin. */
1804 FRF_BZ_TX_PACE_BIN_TH,
1805 FFE_BZ_TX_PACE_RESERVED);
1806 efx_writeo(efx, &temp, FR_BZ_TX_PACE);
1807}
1808
1809/**************************************************************************
1810 *
1811 * Filter tables
1812 *
1813 **************************************************************************
1814 */
1815
1816/* "Fudge factors" - difference between programmed value and actual depth.
1817 * Due to pipelined implementation we need to program H/W with a value that
1818 * is larger than the hop limit we want.
1819 */
1820#define EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD 3
1821#define EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL 1
1822
1823/* Hard maximum search limit. Hardware will time-out beyond 200-something.
1824 * We also need to avoid infinite loops in efx_farch_filter_search() when the
1825 * table is full.
1826 */
1827#define EFX_FARCH_FILTER_CTL_SRCH_MAX 200
1828
1829/* Don't try very hard to find space for performance hints, as this is
1830 * counter-productive. */
1831#define EFX_FARCH_FILTER_CTL_SRCH_HINT_MAX 5
1832
1833enum efx_farch_filter_type {
1834 EFX_FARCH_FILTER_TCP_FULL = 0,
1835 EFX_FARCH_FILTER_TCP_WILD,
1836 EFX_FARCH_FILTER_UDP_FULL,
1837 EFX_FARCH_FILTER_UDP_WILD,
1838 EFX_FARCH_FILTER_MAC_FULL = 4,
1839 EFX_FARCH_FILTER_MAC_WILD,
1840 EFX_FARCH_FILTER_UC_DEF = 8,
1841 EFX_FARCH_FILTER_MC_DEF,
1842 EFX_FARCH_FILTER_TYPE_COUNT, /* number of specific types */
1843};
1844
1845enum efx_farch_filter_table_id {
1846 EFX_FARCH_FILTER_TABLE_RX_IP = 0,
1847 EFX_FARCH_FILTER_TABLE_RX_MAC,
1848 EFX_FARCH_FILTER_TABLE_RX_DEF,
1849 EFX_FARCH_FILTER_TABLE_TX_MAC,
1850 EFX_FARCH_FILTER_TABLE_COUNT,
1851};
1852
1853enum efx_farch_filter_index {
1854 EFX_FARCH_FILTER_INDEX_UC_DEF,
1855 EFX_FARCH_FILTER_INDEX_MC_DEF,
1856 EFX_FARCH_FILTER_SIZE_RX_DEF,
1857};
1858
1859struct efx_farch_filter_spec {
1860 u8 type:4;
1861 u8 priority:4;
1862 u8 flags;
1863 u16 dmaq_id;
1864 u32 data[3];
1865};
1866
1867struct efx_farch_filter_table {
1868 enum efx_farch_filter_table_id id;
1869 u32 offset; /* address of table relative to BAR */
1870 unsigned size; /* number of entries */
1871 unsigned step; /* step between entries */
1872 unsigned used; /* number currently used */
1873 unsigned long *used_bitmap;
1874 struct efx_farch_filter_spec *spec;
1875 unsigned search_limit[EFX_FARCH_FILTER_TYPE_COUNT];
1876};
1877
1878struct efx_farch_filter_state {
1879 struct rw_semaphore lock; /* Protects table contents */
1880 struct efx_farch_filter_table table[EFX_FARCH_FILTER_TABLE_COUNT];
1881};
1882
1883static void
1884efx_farch_filter_table_clear_entry(struct efx_nic *efx,
1885 struct efx_farch_filter_table *table,
1886 unsigned int filter_idx);
1887
1888/* The filter hash function is LFSR polynomial x^16 + x^3 + 1 of a 32-bit
1889 * key derived from the n-tuple. The initial LFSR state is 0xffff. */
1890static u16 efx_farch_filter_hash(u32 key)
1891{
1892 u16 tmp;
1893
1894 /* First 16 rounds */
1895 tmp = 0x1fff ^ key >> 16;
1896 tmp = tmp ^ tmp >> 3 ^ tmp >> 6;
1897 tmp = tmp ^ tmp >> 9;
1898 /* Last 16 rounds */
1899 tmp = tmp ^ tmp << 13 ^ key;
1900 tmp = tmp ^ tmp >> 3 ^ tmp >> 6;
1901 return tmp ^ tmp >> 9;
1902}
1903
1904/* To allow for hash collisions, filter search continues at these
1905 * increments from the first possible entry selected by the hash. */
1906static u16 efx_farch_filter_increment(u32 key)
1907{
1908 return key * 2 - 1;
1909}
1910
1911static enum efx_farch_filter_table_id
1912efx_farch_filter_spec_table_id(const struct efx_farch_filter_spec *spec)
1913{
1914 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1915 (EFX_FARCH_FILTER_TCP_FULL >> 2));
1916 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1917 (EFX_FARCH_FILTER_TCP_WILD >> 2));
1918 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1919 (EFX_FARCH_FILTER_UDP_FULL >> 2));
1920 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_IP !=
1921 (EFX_FARCH_FILTER_UDP_WILD >> 2));
1922 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_MAC !=
1923 (EFX_FARCH_FILTER_MAC_FULL >> 2));
1924 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_RX_MAC !=
1925 (EFX_FARCH_FILTER_MAC_WILD >> 2));
1926 BUILD_BUG_ON(EFX_FARCH_FILTER_TABLE_TX_MAC !=
1927 EFX_FARCH_FILTER_TABLE_RX_MAC + 2);
1928 return (spec->type >> 2) + ((spec->flags & EFX_FILTER_FLAG_TX) ? 2 : 0);
1929}
1930
1931static void efx_farch_filter_push_rx_config(struct efx_nic *efx)
1932{
1933 struct efx_farch_filter_state *state = efx->filter_state;
1934 struct efx_farch_filter_table *table;
1935 efx_oword_t filter_ctl;
1936
1937 efx_reado(efx, &filter_ctl, FR_BZ_RX_FILTER_CTL);
1938
1939 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_IP];
1940 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_TCP_FULL_SRCH_LIMIT,
1941 table->search_limit[EFX_FARCH_FILTER_TCP_FULL] +
1942 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
1943 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_TCP_WILD_SRCH_LIMIT,
1944 table->search_limit[EFX_FARCH_FILTER_TCP_WILD] +
1945 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
1946 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_UDP_FULL_SRCH_LIMIT,
1947 table->search_limit[EFX_FARCH_FILTER_UDP_FULL] +
1948 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
1949 EFX_SET_OWORD_FIELD(filter_ctl, FRF_BZ_UDP_WILD_SRCH_LIMIT,
1950 table->search_limit[EFX_FARCH_FILTER_UDP_WILD] +
1951 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
1952
1953 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_MAC];
1954 if (table->size) {
1955 EFX_SET_OWORD_FIELD(
1956 filter_ctl, FRF_CZ_ETHERNET_FULL_SEARCH_LIMIT,
1957 table->search_limit[EFX_FARCH_FILTER_MAC_FULL] +
1958 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
1959 EFX_SET_OWORD_FIELD(
1960 filter_ctl, FRF_CZ_ETHERNET_WILDCARD_SEARCH_LIMIT,
1961 table->search_limit[EFX_FARCH_FILTER_MAC_WILD] +
1962 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
1963 }
1964
1965 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_DEF];
1966 if (table->size) {
1967 EFX_SET_OWORD_FIELD(
1968 filter_ctl, FRF_CZ_UNICAST_NOMATCH_Q_ID,
1969 table->spec[EFX_FARCH_FILTER_INDEX_UC_DEF].dmaq_id);
1970 EFX_SET_OWORD_FIELD(
1971 filter_ctl, FRF_CZ_UNICAST_NOMATCH_RSS_ENABLED,
1972 !!(table->spec[EFX_FARCH_FILTER_INDEX_UC_DEF].flags &
1973 EFX_FILTER_FLAG_RX_RSS));
1974 EFX_SET_OWORD_FIELD(
1975 filter_ctl, FRF_CZ_MULTICAST_NOMATCH_Q_ID,
1976 table->spec[EFX_FARCH_FILTER_INDEX_MC_DEF].dmaq_id);
1977 EFX_SET_OWORD_FIELD(
1978 filter_ctl, FRF_CZ_MULTICAST_NOMATCH_RSS_ENABLED,
1979 !!(table->spec[EFX_FARCH_FILTER_INDEX_MC_DEF].flags &
1980 EFX_FILTER_FLAG_RX_RSS));
1981
1982 /* There is a single bit to enable RX scatter for all
1983 * unmatched packets. Only set it if scatter is
1984 * enabled in both filter specs.
1985 */
1986 EFX_SET_OWORD_FIELD(
1987 filter_ctl, FRF_BZ_SCATTER_ENBL_NO_MATCH_Q,
1988 !!(table->spec[EFX_FARCH_FILTER_INDEX_UC_DEF].flags &
1989 table->spec[EFX_FARCH_FILTER_INDEX_MC_DEF].flags &
1990 EFX_FILTER_FLAG_RX_SCATTER));
1991 } else {
1992 /* We don't expose 'default' filters because unmatched
1993 * packets always go to the queue number found in the
1994 * RSS table. But we still need to set the RX scatter
1995 * bit here.
1996 */
1997 EFX_SET_OWORD_FIELD(
1998 filter_ctl, FRF_BZ_SCATTER_ENBL_NO_MATCH_Q,
1999 efx->rx_scatter);
2000 }
2001
2002 efx_writeo(efx, &filter_ctl, FR_BZ_RX_FILTER_CTL);
2003}
2004
2005static void efx_farch_filter_push_tx_limits(struct efx_nic *efx)
2006{
2007 struct efx_farch_filter_state *state = efx->filter_state;
2008 struct efx_farch_filter_table *table;
2009 efx_oword_t tx_cfg;
2010
2011 efx_reado(efx, &tx_cfg, FR_AZ_TX_CFG);
2012
2013 table = &state->table[EFX_FARCH_FILTER_TABLE_TX_MAC];
2014 if (table->size) {
2015 EFX_SET_OWORD_FIELD(
2016 tx_cfg, FRF_CZ_TX_ETH_FILTER_FULL_SEARCH_RANGE,
2017 table->search_limit[EFX_FARCH_FILTER_MAC_FULL] +
2018 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_FULL);
2019 EFX_SET_OWORD_FIELD(
2020 tx_cfg, FRF_CZ_TX_ETH_FILTER_WILD_SEARCH_RANGE,
2021 table->search_limit[EFX_FARCH_FILTER_MAC_WILD] +
2022 EFX_FARCH_FILTER_CTL_SRCH_FUDGE_WILD);
2023 }
2024
2025 efx_writeo(efx, &tx_cfg, FR_AZ_TX_CFG);
2026}
2027
2028static int
2029efx_farch_filter_from_gen_spec(struct efx_farch_filter_spec *spec,
2030 const struct efx_filter_spec *gen_spec)
2031{
2032 bool is_full = false;
2033
2034 if ((gen_spec->flags & EFX_FILTER_FLAG_RX_RSS) && gen_spec->rss_context)
2035 return -EINVAL;
2036
2037 spec->priority = gen_spec->priority;
2038 spec->flags = gen_spec->flags;
2039 spec->dmaq_id = gen_spec->dmaq_id;
2040
2041 switch (gen_spec->match_flags) {
2042 case (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
2043 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
2044 EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT):
2045 is_full = true;
2046 fallthrough;
2047 case (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
2048 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT): {
2049 __be32 rhost, host1, host2;
2050 __be16 rport, port1, port2;
2051
2052 EFX_WARN_ON_PARANOID(!(gen_spec->flags & EFX_FILTER_FLAG_RX));
2053
2054 if (gen_spec->ether_type != htons(ETH_P_IP))
2055 return -EPROTONOSUPPORT;
2056 if (gen_spec->loc_port == 0 ||
2057 (is_full && gen_spec->rem_port == 0))
2058 return -EADDRNOTAVAIL;
2059 switch (gen_spec->ip_proto) {
2060 case IPPROTO_TCP:
2061 spec->type = (is_full ? EFX_FARCH_FILTER_TCP_FULL :
2062 EFX_FARCH_FILTER_TCP_WILD);
2063 break;
2064 case IPPROTO_UDP:
2065 spec->type = (is_full ? EFX_FARCH_FILTER_UDP_FULL :
2066 EFX_FARCH_FILTER_UDP_WILD);
2067 break;
2068 default:
2069 return -EPROTONOSUPPORT;
2070 }
2071
2072 /* Filter is constructed in terms of source and destination,
2073 * with the odd wrinkle that the ports are swapped in a UDP
2074 * wildcard filter. We need to convert from local and remote
2075 * (= zero for wildcard) addresses.
2076 */
2077 rhost = is_full ? gen_spec->rem_host[0] : 0;
2078 rport = is_full ? gen_spec->rem_port : 0;
2079 host1 = rhost;
2080 host2 = gen_spec->loc_host[0];
2081 if (!is_full && gen_spec->ip_proto == IPPROTO_UDP) {
2082 port1 = gen_spec->loc_port;
2083 port2 = rport;
2084 } else {
2085 port1 = rport;
2086 port2 = gen_spec->loc_port;
2087 }
2088 spec->data[0] = ntohl(host1) << 16 | ntohs(port1);
2089 spec->data[1] = ntohs(port2) << 16 | ntohl(host1) >> 16;
2090 spec->data[2] = ntohl(host2);
2091
2092 break;
2093 }
2094
2095 case EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_OUTER_VID:
2096 is_full = true;
2097 fallthrough;
2098 case EFX_FILTER_MATCH_LOC_MAC:
2099 spec->type = (is_full ? EFX_FARCH_FILTER_MAC_FULL :
2100 EFX_FARCH_FILTER_MAC_WILD);
2101 spec->data[0] = is_full ? ntohs(gen_spec->outer_vid) : 0;
2102 spec->data[1] = (gen_spec->loc_mac[2] << 24 |
2103 gen_spec->loc_mac[3] << 16 |
2104 gen_spec->loc_mac[4] << 8 |
2105 gen_spec->loc_mac[5]);
2106 spec->data[2] = (gen_spec->loc_mac[0] << 8 |
2107 gen_spec->loc_mac[1]);
2108 break;
2109
2110 case EFX_FILTER_MATCH_LOC_MAC_IG:
2111 spec->type = (is_multicast_ether_addr(gen_spec->loc_mac) ?
2112 EFX_FARCH_FILTER_MC_DEF :
2113 EFX_FARCH_FILTER_UC_DEF);
2114 memset(spec->data, 0, sizeof(spec->data)); /* ensure equality */
2115 break;
2116
2117 default:
2118 return -EPROTONOSUPPORT;
2119 }
2120
2121 return 0;
2122}
2123
2124static void
2125efx_farch_filter_to_gen_spec(struct efx_filter_spec *gen_spec,
2126 const struct efx_farch_filter_spec *spec)
2127{
2128 bool is_full = false;
2129
2130 /* *gen_spec should be completely initialised, to be consistent
2131 * with efx_filter_init_{rx,tx}() and in case we want to copy
2132 * it back to userland.
2133 */
2134 memset(gen_spec, 0, sizeof(*gen_spec));
2135
2136 gen_spec->priority = spec->priority;
2137 gen_spec->flags = spec->flags;
2138 gen_spec->dmaq_id = spec->dmaq_id;
2139
2140 switch (spec->type) {
2141 case EFX_FARCH_FILTER_TCP_FULL:
2142 case EFX_FARCH_FILTER_UDP_FULL:
2143 is_full = true;
2144 fallthrough;
2145 case EFX_FARCH_FILTER_TCP_WILD:
2146 case EFX_FARCH_FILTER_UDP_WILD: {
2147 __be32 host1, host2;
2148 __be16 port1, port2;
2149
2150 gen_spec->match_flags =
2151 EFX_FILTER_MATCH_ETHER_TYPE |
2152 EFX_FILTER_MATCH_IP_PROTO |
2153 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT;
2154 if (is_full)
2155 gen_spec->match_flags |= (EFX_FILTER_MATCH_REM_HOST |
2156 EFX_FILTER_MATCH_REM_PORT);
2157 gen_spec->ether_type = htons(ETH_P_IP);
2158 gen_spec->ip_proto =
2159 (spec->type == EFX_FARCH_FILTER_TCP_FULL ||
2160 spec->type == EFX_FARCH_FILTER_TCP_WILD) ?
2161 IPPROTO_TCP : IPPROTO_UDP;
2162
2163 host1 = htonl(spec->data[0] >> 16 | spec->data[1] << 16);
2164 port1 = htons(spec->data[0]);
2165 host2 = htonl(spec->data[2]);
2166 port2 = htons(spec->data[1] >> 16);
2167 if (spec->flags & EFX_FILTER_FLAG_TX) {
2168 gen_spec->loc_host[0] = host1;
2169 gen_spec->rem_host[0] = host2;
2170 } else {
2171 gen_spec->loc_host[0] = host2;
2172 gen_spec->rem_host[0] = host1;
2173 }
2174 if (!!(gen_spec->flags & EFX_FILTER_FLAG_TX) ^
2175 (!is_full && gen_spec->ip_proto == IPPROTO_UDP)) {
2176 gen_spec->loc_port = port1;
2177 gen_spec->rem_port = port2;
2178 } else {
2179 gen_spec->loc_port = port2;
2180 gen_spec->rem_port = port1;
2181 }
2182
2183 break;
2184 }
2185
2186 case EFX_FARCH_FILTER_MAC_FULL:
2187 is_full = true;
2188 fallthrough;
2189 case EFX_FARCH_FILTER_MAC_WILD:
2190 gen_spec->match_flags = EFX_FILTER_MATCH_LOC_MAC;
2191 if (is_full)
2192 gen_spec->match_flags |= EFX_FILTER_MATCH_OUTER_VID;
2193 gen_spec->loc_mac[0] = spec->data[2] >> 8;
2194 gen_spec->loc_mac[1] = spec->data[2];
2195 gen_spec->loc_mac[2] = spec->data[1] >> 24;
2196 gen_spec->loc_mac[3] = spec->data[1] >> 16;
2197 gen_spec->loc_mac[4] = spec->data[1] >> 8;
2198 gen_spec->loc_mac[5] = spec->data[1];
2199 gen_spec->outer_vid = htons(spec->data[0]);
2200 break;
2201
2202 case EFX_FARCH_FILTER_UC_DEF:
2203 case EFX_FARCH_FILTER_MC_DEF:
2204 gen_spec->match_flags = EFX_FILTER_MATCH_LOC_MAC_IG;
2205 gen_spec->loc_mac[0] = spec->type == EFX_FARCH_FILTER_MC_DEF;
2206 break;
2207
2208 default:
2209 WARN_ON(1);
2210 break;
2211 }
2212}
2213
2214static void
2215efx_farch_filter_init_rx_auto(struct efx_nic *efx,
2216 struct efx_farch_filter_spec *spec)
2217{
2218 /* If there's only one channel then disable RSS for non VF
2219 * traffic, thereby allowing VFs to use RSS when the PF can't.
2220 */
2221 spec->priority = EFX_FILTER_PRI_AUTO;
2222 spec->flags = (EFX_FILTER_FLAG_RX |
2223 (efx_rss_enabled(efx) ? EFX_FILTER_FLAG_RX_RSS : 0) |
2224 (efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0));
2225 spec->dmaq_id = 0;
2226}
2227
2228/* Build a filter entry and return its n-tuple key. */
2229static u32 efx_farch_filter_build(efx_oword_t *filter,
2230 struct efx_farch_filter_spec *spec)
2231{
2232 u32 data3;
2233
2234 switch (efx_farch_filter_spec_table_id(spec)) {
2235 case EFX_FARCH_FILTER_TABLE_RX_IP: {
2236 bool is_udp = (spec->type == EFX_FARCH_FILTER_UDP_FULL ||
2237 spec->type == EFX_FARCH_FILTER_UDP_WILD);
2238 EFX_POPULATE_OWORD_7(
2239 *filter,
2240 FRF_BZ_RSS_EN,
2241 !!(spec->flags & EFX_FILTER_FLAG_RX_RSS),
2242 FRF_BZ_SCATTER_EN,
2243 !!(spec->flags & EFX_FILTER_FLAG_RX_SCATTER),
2244 FRF_BZ_TCP_UDP, is_udp,
2245 FRF_BZ_RXQ_ID, spec->dmaq_id,
2246 EFX_DWORD_2, spec->data[2],
2247 EFX_DWORD_1, spec->data[1],
2248 EFX_DWORD_0, spec->data[0]);
2249 data3 = is_udp;
2250 break;
2251 }
2252
2253 case EFX_FARCH_FILTER_TABLE_RX_MAC: {
2254 bool is_wild = spec->type == EFX_FARCH_FILTER_MAC_WILD;
2255 EFX_POPULATE_OWORD_7(
2256 *filter,
2257 FRF_CZ_RMFT_RSS_EN,
2258 !!(spec->flags & EFX_FILTER_FLAG_RX_RSS),
2259 FRF_CZ_RMFT_SCATTER_EN,
2260 !!(spec->flags & EFX_FILTER_FLAG_RX_SCATTER),
2261 FRF_CZ_RMFT_RXQ_ID, spec->dmaq_id,
2262 FRF_CZ_RMFT_WILDCARD_MATCH, is_wild,
2263 FRF_CZ_RMFT_DEST_MAC_HI, spec->data[2],
2264 FRF_CZ_RMFT_DEST_MAC_LO, spec->data[1],
2265 FRF_CZ_RMFT_VLAN_ID, spec->data[0]);
2266 data3 = is_wild;
2267 break;
2268 }
2269
2270 case EFX_FARCH_FILTER_TABLE_TX_MAC: {
2271 bool is_wild = spec->type == EFX_FARCH_FILTER_MAC_WILD;
2272 EFX_POPULATE_OWORD_5(*filter,
2273 FRF_CZ_TMFT_TXQ_ID, spec->dmaq_id,
2274 FRF_CZ_TMFT_WILDCARD_MATCH, is_wild,
2275 FRF_CZ_TMFT_SRC_MAC_HI, spec->data[2],
2276 FRF_CZ_TMFT_SRC_MAC_LO, spec->data[1],
2277 FRF_CZ_TMFT_VLAN_ID, spec->data[0]);
2278 data3 = is_wild | spec->dmaq_id << 1;
2279 break;
2280 }
2281
2282 default:
2283 BUG();
2284 }
2285
2286 return spec->data[0] ^ spec->data[1] ^ spec->data[2] ^ data3;
2287}
2288
2289static bool efx_farch_filter_equal(const struct efx_farch_filter_spec *left,
2290 const struct efx_farch_filter_spec *right)
2291{
2292 if (left->type != right->type ||
2293 memcmp(left->data, right->data, sizeof(left->data)))
2294 return false;
2295
2296 if (left->flags & EFX_FILTER_FLAG_TX &&
2297 left->dmaq_id != right->dmaq_id)
2298 return false;
2299
2300 return true;
2301}
2302
2303/*
2304 * Construct/deconstruct external filter IDs. At least the RX filter
2305 * IDs must be ordered by matching priority, for RX NFC semantics.
2306 *
2307 * Deconstruction needs to be robust against invalid IDs so that
2308 * efx_filter_remove_id_safe() and efx_filter_get_filter_safe() can
2309 * accept user-provided IDs.
2310 */
2311
2312#define EFX_FARCH_FILTER_MATCH_PRI_COUNT 5
2313
2314static const u8 efx_farch_filter_type_match_pri[EFX_FARCH_FILTER_TYPE_COUNT] = {
2315 [EFX_FARCH_FILTER_TCP_FULL] = 0,
2316 [EFX_FARCH_FILTER_UDP_FULL] = 0,
2317 [EFX_FARCH_FILTER_TCP_WILD] = 1,
2318 [EFX_FARCH_FILTER_UDP_WILD] = 1,
2319 [EFX_FARCH_FILTER_MAC_FULL] = 2,
2320 [EFX_FARCH_FILTER_MAC_WILD] = 3,
2321 [EFX_FARCH_FILTER_UC_DEF] = 4,
2322 [EFX_FARCH_FILTER_MC_DEF] = 4,
2323};
2324
2325static const enum efx_farch_filter_table_id efx_farch_filter_range_table[] = {
2326 EFX_FARCH_FILTER_TABLE_RX_IP, /* RX match pri 0 */
2327 EFX_FARCH_FILTER_TABLE_RX_IP,
2328 EFX_FARCH_FILTER_TABLE_RX_MAC,
2329 EFX_FARCH_FILTER_TABLE_RX_MAC,
2330 EFX_FARCH_FILTER_TABLE_RX_DEF, /* RX match pri 4 */
2331 EFX_FARCH_FILTER_TABLE_TX_MAC, /* TX match pri 0 */
2332 EFX_FARCH_FILTER_TABLE_TX_MAC, /* TX match pri 1 */
2333};
2334
2335#define EFX_FARCH_FILTER_INDEX_WIDTH 13
2336#define EFX_FARCH_FILTER_INDEX_MASK ((1 << EFX_FARCH_FILTER_INDEX_WIDTH) - 1)
2337
2338static inline u32
2339efx_farch_filter_make_id(const struct efx_farch_filter_spec *spec,
2340 unsigned int index)
2341{
2342 unsigned int range;
2343
2344 range = efx_farch_filter_type_match_pri[spec->type];
2345 if (!(spec->flags & EFX_FILTER_FLAG_RX))
2346 range += EFX_FARCH_FILTER_MATCH_PRI_COUNT;
2347
2348 return range << EFX_FARCH_FILTER_INDEX_WIDTH | index;
2349}
2350
2351static inline enum efx_farch_filter_table_id
2352efx_farch_filter_id_table_id(u32 id)
2353{
2354 unsigned int range = id >> EFX_FARCH_FILTER_INDEX_WIDTH;
2355
2356 if (range < ARRAY_SIZE(efx_farch_filter_range_table))
2357 return efx_farch_filter_range_table[range];
2358 else
2359 return EFX_FARCH_FILTER_TABLE_COUNT; /* invalid */
2360}
2361
2362static inline unsigned int efx_farch_filter_id_index(u32 id)
2363{
2364 return id & EFX_FARCH_FILTER_INDEX_MASK;
2365}
2366
2367u32 efx_farch_filter_get_rx_id_limit(struct efx_nic *efx)
2368{
2369 struct efx_farch_filter_state *state = efx->filter_state;
2370 unsigned int range = EFX_FARCH_FILTER_MATCH_PRI_COUNT - 1;
2371 enum efx_farch_filter_table_id table_id;
2372
2373 do {
2374 table_id = efx_farch_filter_range_table[range];
2375 if (state->table[table_id].size != 0)
2376 return range << EFX_FARCH_FILTER_INDEX_WIDTH |
2377 state->table[table_id].size;
2378 } while (range--);
2379
2380 return 0;
2381}
2382
2383s32 efx_farch_filter_insert(struct efx_nic *efx,
2384 struct efx_filter_spec *gen_spec,
2385 bool replace_equal)
2386{
2387 struct efx_farch_filter_state *state = efx->filter_state;
2388 struct efx_farch_filter_table *table;
2389 struct efx_farch_filter_spec spec;
2390 efx_oword_t filter;
2391 int rep_index, ins_index;
2392 unsigned int depth = 0;
2393 int rc;
2394
2395 rc = efx_farch_filter_from_gen_spec(&spec, gen_spec);
2396 if (rc)
2397 return rc;
2398
2399 down_write(&state->lock);
2400
2401 table = &state->table[efx_farch_filter_spec_table_id(&spec)];
2402 if (table->size == 0) {
2403 rc = -EINVAL;
2404 goto out_unlock;
2405 }
2406
2407 netif_vdbg(efx, hw, efx->net_dev,
2408 "%s: type %d search_limit=%d", __func__, spec.type,
2409 table->search_limit[spec.type]);
2410
2411 if (table->id == EFX_FARCH_FILTER_TABLE_RX_DEF) {
2412 /* One filter spec per type */
2413 BUILD_BUG_ON(EFX_FARCH_FILTER_INDEX_UC_DEF != 0);
2414 BUILD_BUG_ON(EFX_FARCH_FILTER_INDEX_MC_DEF !=
2415 EFX_FARCH_FILTER_MC_DEF - EFX_FARCH_FILTER_UC_DEF);
2416 rep_index = spec.type - EFX_FARCH_FILTER_UC_DEF;
2417 ins_index = rep_index;
2418 } else {
2419 /* Search concurrently for
2420 * (1) a filter to be replaced (rep_index): any filter
2421 * with the same match values, up to the current
2422 * search depth for this type, and
2423 * (2) the insertion point (ins_index): (1) or any
2424 * free slot before it or up to the maximum search
2425 * depth for this priority
2426 * We fail if we cannot find (2).
2427 *
2428 * We can stop once either
2429 * (a) we find (1), in which case we have definitely
2430 * found (2) as well; or
2431 * (b) we have searched exhaustively for (1), and have
2432 * either found (2) or searched exhaustively for it
2433 */
2434 u32 key = efx_farch_filter_build(&filter, &spec);
2435 unsigned int hash = efx_farch_filter_hash(key);
2436 unsigned int incr = efx_farch_filter_increment(key);
2437 unsigned int max_rep_depth = table->search_limit[spec.type];
2438 unsigned int max_ins_depth =
2439 spec.priority <= EFX_FILTER_PRI_HINT ?
2440 EFX_FARCH_FILTER_CTL_SRCH_HINT_MAX :
2441 EFX_FARCH_FILTER_CTL_SRCH_MAX;
2442 unsigned int i = hash & (table->size - 1);
2443
2444 ins_index = -1;
2445 depth = 1;
2446
2447 for (;;) {
2448 if (!test_bit(i, table->used_bitmap)) {
2449 if (ins_index < 0)
2450 ins_index = i;
2451 } else if (efx_farch_filter_equal(&spec,
2452 &table->spec[i])) {
2453 /* Case (a) */
2454 if (ins_index < 0)
2455 ins_index = i;
2456 rep_index = i;
2457 break;
2458 }
2459
2460 if (depth >= max_rep_depth &&
2461 (ins_index >= 0 || depth >= max_ins_depth)) {
2462 /* Case (b) */
2463 if (ins_index < 0) {
2464 rc = -EBUSY;
2465 goto out_unlock;
2466 }
2467 rep_index = -1;
2468 break;
2469 }
2470
2471 i = (i + incr) & (table->size - 1);
2472 ++depth;
2473 }
2474 }
2475
2476 /* If we found a filter to be replaced, check whether we
2477 * should do so
2478 */
2479 if (rep_index >= 0) {
2480 struct efx_farch_filter_spec *saved_spec =
2481 &table->spec[rep_index];
2482
2483 if (spec.priority == saved_spec->priority && !replace_equal) {
2484 rc = -EEXIST;
2485 goto out_unlock;
2486 }
2487 if (spec.priority < saved_spec->priority) {
2488 rc = -EPERM;
2489 goto out_unlock;
2490 }
2491 if (saved_spec->priority == EFX_FILTER_PRI_AUTO ||
2492 saved_spec->flags & EFX_FILTER_FLAG_RX_OVER_AUTO)
2493 spec.flags |= EFX_FILTER_FLAG_RX_OVER_AUTO;
2494 }
2495
2496 /* Insert the filter */
2497 if (ins_index != rep_index) {
2498 __set_bit(ins_index, table->used_bitmap);
2499 ++table->used;
2500 }
2501 table->spec[ins_index] = spec;
2502
2503 if (table->id == EFX_FARCH_FILTER_TABLE_RX_DEF) {
2504 efx_farch_filter_push_rx_config(efx);
2505 } else {
2506 if (table->search_limit[spec.type] < depth) {
2507 table->search_limit[spec.type] = depth;
2508 if (spec.flags & EFX_FILTER_FLAG_TX)
2509 efx_farch_filter_push_tx_limits(efx);
2510 else
2511 efx_farch_filter_push_rx_config(efx);
2512 }
2513
2514 efx_writeo(efx, &filter,
2515 table->offset + table->step * ins_index);
2516
2517 /* If we were able to replace a filter by inserting
2518 * at a lower depth, clear the replaced filter
2519 */
2520 if (ins_index != rep_index && rep_index >= 0)
2521 efx_farch_filter_table_clear_entry(efx, table,
2522 rep_index);
2523 }
2524
2525 netif_vdbg(efx, hw, efx->net_dev,
2526 "%s: filter type %d index %d rxq %u set",
2527 __func__, spec.type, ins_index, spec.dmaq_id);
2528 rc = efx_farch_filter_make_id(&spec, ins_index);
2529
2530out_unlock:
2531 up_write(&state->lock);
2532 return rc;
2533}
2534
2535static void
2536efx_farch_filter_table_clear_entry(struct efx_nic *efx,
2537 struct efx_farch_filter_table *table,
2538 unsigned int filter_idx)
2539{
2540 static efx_oword_t filter;
2541
2542 EFX_WARN_ON_PARANOID(!test_bit(filter_idx, table->used_bitmap));
2543 BUG_ON(table->offset == 0); /* can't clear MAC default filters */
2544
2545 __clear_bit(filter_idx, table->used_bitmap);
2546 --table->used;
2547 memset(&table->spec[filter_idx], 0, sizeof(table->spec[0]));
2548
2549 efx_writeo(efx, &filter, table->offset + table->step * filter_idx);
2550
2551 /* If this filter required a greater search depth than
2552 * any other, the search limit for its type can now be
2553 * decreased. However, it is hard to determine that
2554 * unless the table has become completely empty - in
2555 * which case, all its search limits can be set to 0.
2556 */
2557 if (unlikely(table->used == 0)) {
2558 memset(table->search_limit, 0, sizeof(table->search_limit));
2559 if (table->id == EFX_FARCH_FILTER_TABLE_TX_MAC)
2560 efx_farch_filter_push_tx_limits(efx);
2561 else
2562 efx_farch_filter_push_rx_config(efx);
2563 }
2564}
2565
2566static int efx_farch_filter_remove(struct efx_nic *efx,
2567 struct efx_farch_filter_table *table,
2568 unsigned int filter_idx,
2569 enum efx_filter_priority priority)
2570{
2571 struct efx_farch_filter_spec *spec = &table->spec[filter_idx];
2572
2573 if (!test_bit(filter_idx, table->used_bitmap) ||
2574 spec->priority != priority)
2575 return -ENOENT;
2576
2577 if (spec->flags & EFX_FILTER_FLAG_RX_OVER_AUTO) {
2578 efx_farch_filter_init_rx_auto(efx, spec);
2579 efx_farch_filter_push_rx_config(efx);
2580 } else {
2581 efx_farch_filter_table_clear_entry(efx, table, filter_idx);
2582 }
2583
2584 return 0;
2585}
2586
2587int efx_farch_filter_remove_safe(struct efx_nic *efx,
2588 enum efx_filter_priority priority,
2589 u32 filter_id)
2590{
2591 struct efx_farch_filter_state *state = efx->filter_state;
2592 enum efx_farch_filter_table_id table_id;
2593 struct efx_farch_filter_table *table;
2594 unsigned int filter_idx;
2595 struct efx_farch_filter_spec *spec;
2596 int rc;
2597
2598 table_id = efx_farch_filter_id_table_id(filter_id);
2599 if ((unsigned int)table_id >= EFX_FARCH_FILTER_TABLE_COUNT)
2600 return -ENOENT;
2601 table = &state->table[table_id];
2602
2603 filter_idx = efx_farch_filter_id_index(filter_id);
2604 if (filter_idx >= table->size)
2605 return -ENOENT;
2606 down_write(&state->lock);
2607 spec = &table->spec[filter_idx];
2608
2609 rc = efx_farch_filter_remove(efx, table, filter_idx, priority);
2610 up_write(&state->lock);
2611
2612 return rc;
2613}
2614
2615int efx_farch_filter_get_safe(struct efx_nic *efx,
2616 enum efx_filter_priority priority,
2617 u32 filter_id, struct efx_filter_spec *spec_buf)
2618{
2619 struct efx_farch_filter_state *state = efx->filter_state;
2620 enum efx_farch_filter_table_id table_id;
2621 struct efx_farch_filter_table *table;
2622 struct efx_farch_filter_spec *spec;
2623 unsigned int filter_idx;
2624 int rc = -ENOENT;
2625
2626 down_read(&state->lock);
2627
2628 table_id = efx_farch_filter_id_table_id(filter_id);
2629 if ((unsigned int)table_id >= EFX_FARCH_FILTER_TABLE_COUNT)
2630 goto out_unlock;
2631 table = &state->table[table_id];
2632
2633 filter_idx = efx_farch_filter_id_index(filter_id);
2634 if (filter_idx >= table->size)
2635 goto out_unlock;
2636 spec = &table->spec[filter_idx];
2637
2638 if (test_bit(filter_idx, table->used_bitmap) &&
2639 spec->priority == priority) {
2640 efx_farch_filter_to_gen_spec(spec_buf, spec);
2641 rc = 0;
2642 }
2643
2644out_unlock:
2645 up_read(&state->lock);
2646 return rc;
2647}
2648
2649static void
2650efx_farch_filter_table_clear(struct efx_nic *efx,
2651 enum efx_farch_filter_table_id table_id,
2652 enum efx_filter_priority priority)
2653{
2654 struct efx_farch_filter_state *state = efx->filter_state;
2655 struct efx_farch_filter_table *table = &state->table[table_id];
2656 unsigned int filter_idx;
2657
2658 down_write(&state->lock);
2659 for (filter_idx = 0; filter_idx < table->size; ++filter_idx) {
2660 if (table->spec[filter_idx].priority != EFX_FILTER_PRI_AUTO)
2661 efx_farch_filter_remove(efx, table,
2662 filter_idx, priority);
2663 }
2664 up_write(&state->lock);
2665}
2666
2667int efx_farch_filter_clear_rx(struct efx_nic *efx,
2668 enum efx_filter_priority priority)
2669{
2670 efx_farch_filter_table_clear(efx, EFX_FARCH_FILTER_TABLE_RX_IP,
2671 priority);
2672 efx_farch_filter_table_clear(efx, EFX_FARCH_FILTER_TABLE_RX_MAC,
2673 priority);
2674 efx_farch_filter_table_clear(efx, EFX_FARCH_FILTER_TABLE_RX_DEF,
2675 priority);
2676 return 0;
2677}
2678
2679u32 efx_farch_filter_count_rx_used(struct efx_nic *efx,
2680 enum efx_filter_priority priority)
2681{
2682 struct efx_farch_filter_state *state = efx->filter_state;
2683 enum efx_farch_filter_table_id table_id;
2684 struct efx_farch_filter_table *table;
2685 unsigned int filter_idx;
2686 u32 count = 0;
2687
2688 down_read(&state->lock);
2689
2690 for (table_id = EFX_FARCH_FILTER_TABLE_RX_IP;
2691 table_id <= EFX_FARCH_FILTER_TABLE_RX_DEF;
2692 table_id++) {
2693 table = &state->table[table_id];
2694 for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
2695 if (test_bit(filter_idx, table->used_bitmap) &&
2696 table->spec[filter_idx].priority == priority)
2697 ++count;
2698 }
2699 }
2700
2701 up_read(&state->lock);
2702
2703 return count;
2704}
2705
2706s32 efx_farch_filter_get_rx_ids(struct efx_nic *efx,
2707 enum efx_filter_priority priority,
2708 u32 *buf, u32 size)
2709{
2710 struct efx_farch_filter_state *state = efx->filter_state;
2711 enum efx_farch_filter_table_id table_id;
2712 struct efx_farch_filter_table *table;
2713 unsigned int filter_idx;
2714 s32 count = 0;
2715
2716 down_read(&state->lock);
2717
2718 for (table_id = EFX_FARCH_FILTER_TABLE_RX_IP;
2719 table_id <= EFX_FARCH_FILTER_TABLE_RX_DEF;
2720 table_id++) {
2721 table = &state->table[table_id];
2722 for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
2723 if (test_bit(filter_idx, table->used_bitmap) &&
2724 table->spec[filter_idx].priority == priority) {
2725 if (count == size) {
2726 count = -EMSGSIZE;
2727 goto out;
2728 }
2729 buf[count++] = efx_farch_filter_make_id(
2730 &table->spec[filter_idx], filter_idx);
2731 }
2732 }
2733 }
2734out:
2735 up_read(&state->lock);
2736
2737 return count;
2738}
2739
2740/* Restore filter stater after reset */
2741void efx_farch_filter_table_restore(struct efx_nic *efx)
2742{
2743 struct efx_farch_filter_state *state = efx->filter_state;
2744 enum efx_farch_filter_table_id table_id;
2745 struct efx_farch_filter_table *table;
2746 efx_oword_t filter;
2747 unsigned int filter_idx;
2748
2749 down_write(&state->lock);
2750
2751 for (table_id = 0; table_id < EFX_FARCH_FILTER_TABLE_COUNT; table_id++) {
2752 table = &state->table[table_id];
2753
2754 /* Check whether this is a regular register table */
2755 if (table->step == 0)
2756 continue;
2757
2758 for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
2759 if (!test_bit(filter_idx, table->used_bitmap))
2760 continue;
2761 efx_farch_filter_build(&filter, &table->spec[filter_idx]);
2762 efx_writeo(efx, &filter,
2763 table->offset + table->step * filter_idx);
2764 }
2765 }
2766
2767 efx_farch_filter_push_rx_config(efx);
2768 efx_farch_filter_push_tx_limits(efx);
2769
2770 up_write(&state->lock);
2771}
2772
2773void efx_farch_filter_table_remove(struct efx_nic *efx)
2774{
2775 struct efx_farch_filter_state *state = efx->filter_state;
2776 enum efx_farch_filter_table_id table_id;
2777
2778 for (table_id = 0; table_id < EFX_FARCH_FILTER_TABLE_COUNT; table_id++) {
2779 kfree(state->table[table_id].used_bitmap);
2780 vfree(state->table[table_id].spec);
2781 }
2782 kfree(state);
2783}
2784
2785int efx_farch_filter_table_probe(struct efx_nic *efx)
2786{
2787 struct efx_farch_filter_state *state;
2788 struct efx_farch_filter_table *table;
2789 unsigned table_id;
2790
2791 state = kzalloc(sizeof(struct efx_farch_filter_state), GFP_KERNEL);
2792 if (!state)
2793 return -ENOMEM;
2794 efx->filter_state = state;
2795 init_rwsem(&state->lock);
2796
2797 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_IP];
2798 table->id = EFX_FARCH_FILTER_TABLE_RX_IP;
2799 table->offset = FR_BZ_RX_FILTER_TBL0;
2800 table->size = FR_BZ_RX_FILTER_TBL0_ROWS;
2801 table->step = FR_BZ_RX_FILTER_TBL0_STEP;
2802
2803 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_MAC];
2804 table->id = EFX_FARCH_FILTER_TABLE_RX_MAC;
2805 table->offset = FR_CZ_RX_MAC_FILTER_TBL0;
2806 table->size = FR_CZ_RX_MAC_FILTER_TBL0_ROWS;
2807 table->step = FR_CZ_RX_MAC_FILTER_TBL0_STEP;
2808
2809 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_DEF];
2810 table->id = EFX_FARCH_FILTER_TABLE_RX_DEF;
2811 table->size = EFX_FARCH_FILTER_SIZE_RX_DEF;
2812
2813 table = &state->table[EFX_FARCH_FILTER_TABLE_TX_MAC];
2814 table->id = EFX_FARCH_FILTER_TABLE_TX_MAC;
2815 table->offset = FR_CZ_TX_MAC_FILTER_TBL0;
2816 table->size = FR_CZ_TX_MAC_FILTER_TBL0_ROWS;
2817 table->step = FR_CZ_TX_MAC_FILTER_TBL0_STEP;
2818
2819 for (table_id = 0; table_id < EFX_FARCH_FILTER_TABLE_COUNT; table_id++) {
2820 table = &state->table[table_id];
2821 if (table->size == 0)
2822 continue;
2823 table->used_bitmap = kcalloc(BITS_TO_LONGS(table->size),
2824 sizeof(unsigned long),
2825 GFP_KERNEL);
2826 if (!table->used_bitmap)
2827 goto fail;
2828 table->spec = vzalloc(array_size(sizeof(*table->spec),
2829 table->size));
2830 if (!table->spec)
2831 goto fail;
2832 }
2833
2834 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_DEF];
2835 if (table->size) {
2836 /* RX default filters must always exist */
2837 struct efx_farch_filter_spec *spec;
2838 unsigned i;
2839
2840 for (i = 0; i < EFX_FARCH_FILTER_SIZE_RX_DEF; i++) {
2841 spec = &table->spec[i];
2842 spec->type = EFX_FARCH_FILTER_UC_DEF + i;
2843 efx_farch_filter_init_rx_auto(efx, spec);
2844 __set_bit(i, table->used_bitmap);
2845 }
2846 }
2847
2848 efx_farch_filter_push_rx_config(efx);
2849
2850 return 0;
2851
2852fail:
2853 efx_farch_filter_table_remove(efx);
2854 return -ENOMEM;
2855}
2856
2857/* Update scatter enable flags for filters pointing to our own RX queues */
2858void efx_farch_filter_update_rx_scatter(struct efx_nic *efx)
2859{
2860 struct efx_farch_filter_state *state = efx->filter_state;
2861 enum efx_farch_filter_table_id table_id;
2862 struct efx_farch_filter_table *table;
2863 efx_oword_t filter;
2864 unsigned int filter_idx;
2865
2866 down_write(&state->lock);
2867
2868 for (table_id = EFX_FARCH_FILTER_TABLE_RX_IP;
2869 table_id <= EFX_FARCH_FILTER_TABLE_RX_DEF;
2870 table_id++) {
2871 table = &state->table[table_id];
2872
2873 for (filter_idx = 0; filter_idx < table->size; filter_idx++) {
2874 if (!test_bit(filter_idx, table->used_bitmap) ||
2875 table->spec[filter_idx].dmaq_id >=
2876 efx->n_rx_channels)
2877 continue;
2878
2879 if (efx->rx_scatter)
2880 table->spec[filter_idx].flags |=
2881 EFX_FILTER_FLAG_RX_SCATTER;
2882 else
2883 table->spec[filter_idx].flags &=
2884 ~EFX_FILTER_FLAG_RX_SCATTER;
2885
2886 if (table_id == EFX_FARCH_FILTER_TABLE_RX_DEF)
2887 /* Pushed by efx_farch_filter_push_rx_config() */
2888 continue;
2889
2890 efx_farch_filter_build(&filter, &table->spec[filter_idx]);
2891 efx_writeo(efx, &filter,
2892 table->offset + table->step * filter_idx);
2893 }
2894 }
2895
2896 efx_farch_filter_push_rx_config(efx);
2897
2898 up_write(&state->lock);
2899}
2900
2901#ifdef CONFIG_RFS_ACCEL
2902
2903bool efx_farch_filter_rfs_expire_one(struct efx_nic *efx, u32 flow_id,
2904 unsigned int index)
2905{
2906 struct efx_farch_filter_state *state = efx->filter_state;
2907 struct efx_farch_filter_table *table;
2908 bool ret = false, force = false;
2909 u16 arfs_id;
2910
2911 down_write(&state->lock);
2912 spin_lock_bh(&efx->rps_hash_lock);
2913 table = &state->table[EFX_FARCH_FILTER_TABLE_RX_IP];
2914 if (test_bit(index, table->used_bitmap) &&
2915 table->spec[index].priority == EFX_FILTER_PRI_HINT) {
2916 struct efx_arfs_rule *rule = NULL;
2917 struct efx_filter_spec spec;
2918
2919 efx_farch_filter_to_gen_spec(&spec, &table->spec[index]);
2920 if (!efx->rps_hash_table) {
2921 /* In the absence of the table, we always returned 0 to
2922 * ARFS, so use the same to query it.
2923 */
2924 arfs_id = 0;
2925 } else {
2926 rule = efx_rps_hash_find(efx, &spec);
2927 if (!rule) {
2928 /* ARFS table doesn't know of this filter, remove it */
2929 force = true;
2930 } else {
2931 arfs_id = rule->arfs_id;
2932 if (!efx_rps_check_rule(rule, index, &force))
2933 goto out_unlock;
2934 }
2935 }
2936 if (force || rps_may_expire_flow(efx->net_dev, spec.dmaq_id,
2937 flow_id, arfs_id)) {
2938 if (rule)
2939 rule->filter_id = EFX_ARFS_FILTER_ID_REMOVING;
2940 efx_rps_hash_del(efx, &spec);
2941 efx_farch_filter_table_clear_entry(efx, table, index);
2942 ret = true;
2943 }
2944 }
2945out_unlock:
2946 spin_unlock_bh(&efx->rps_hash_lock);
2947 up_write(&state->lock);
2948 return ret;
2949}
2950
2951#endif /* CONFIG_RFS_ACCEL */
2952
2953void efx_farch_filter_sync_rx_mode(struct efx_nic *efx)
2954{
2955 struct net_device *net_dev = efx->net_dev;
2956 struct netdev_hw_addr *ha;
2957 union efx_multicast_hash *mc_hash = &efx->multicast_hash;
2958 u32 crc;
2959 int bit;
2960
2961 if (!efx_dev_registered(efx))
2962 return;
2963
2964 netif_addr_lock_bh(net_dev);
2965
2966 efx->unicast_filter = !(net_dev->flags & IFF_PROMISC);
2967
2968 /* Build multicast hash table */
2969 if (net_dev->flags & (IFF_PROMISC | IFF_ALLMULTI)) {
2970 memset(mc_hash, 0xff, sizeof(*mc_hash));
2971 } else {
2972 memset(mc_hash, 0x00, sizeof(*mc_hash));
2973 netdev_for_each_mc_addr(ha, net_dev) {
2974 crc = ether_crc_le(ETH_ALEN, ha->addr);
2975 bit = crc & (EFX_MCAST_HASH_ENTRIES - 1);
2976 __set_bit_le(bit, mc_hash);
2977 }
2978
2979 /* Broadcast packets go through the multicast hash filter.
2980 * ether_crc_le() of the broadcast address is 0xbe2612ff
2981 * so we always add bit 0xff to the mask.
2982 */
2983 __set_bit_le(0xff, mc_hash);
2984 }
2985
2986 netif_addr_unlock_bh(net_dev);
2987}