xref: /openbmc/linux/drivers/net/ethernet/sfc/nic.c (revision 2ae75dac)
1 /****************************************************************************
2  * Driver for Solarflare Solarstorm network controllers and boards
3  * Copyright 2005-2006 Fen Systems Ltd.
4  * Copyright 2006-2011 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 "net_driver.h"
18 #include "bitfield.h"
19 #include "efx.h"
20 #include "nic.h"
21 #include "regs.h"
22 #include "io.h"
23 #include "workarounds.h"
24 
25 /**************************************************************************
26  *
27  * Configurable values
28  *
29  **************************************************************************
30  */
31 
32 /* This is set to 16 for a good reason.  In summary, if larger than
33  * 16, the descriptor cache holds more than a default socket
34  * buffer's worth of packets (for UDP we can only have at most one
35  * socket buffer's worth outstanding).  This combined with the fact
36  * that we only get 1 TX event per descriptor cache means the NIC
37  * goes idle.
38  */
39 #define TX_DC_ENTRIES 16
40 #define TX_DC_ENTRIES_ORDER 1
41 
42 #define RX_DC_ENTRIES 64
43 #define RX_DC_ENTRIES_ORDER 3
44 
45 /* If EFX_MAX_INT_ERRORS internal errors occur within
46  * EFX_INT_ERROR_EXPIRE seconds, we consider the NIC broken and
47  * disable it.
48  */
49 #define EFX_INT_ERROR_EXPIRE 3600
50 #define EFX_MAX_INT_ERRORS 5
51 
52 /* We poll for events every FLUSH_INTERVAL ms, and check FLUSH_POLL_COUNT times
53  */
54 #define EFX_FLUSH_INTERVAL 10
55 #define EFX_FLUSH_POLL_COUNT 100
56 
57 /* Depth of RX flush request fifo */
58 #define EFX_RX_FLUSH_COUNT 4
59 
60 /* Driver generated events */
61 #define _EFX_CHANNEL_MAGIC_TEST		0x000101
62 #define _EFX_CHANNEL_MAGIC_FILL		0x000102
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 
73 /**************************************************************************
74  *
75  * Solarstorm hardware access
76  *
77  **************************************************************************/
78 
79 static inline void efx_write_buf_tbl(struct efx_nic *efx, efx_qword_t *value,
80 				     unsigned int index)
81 {
82 	efx_sram_writeq(efx, efx->membase + efx->type->buf_tbl_base,
83 			value, index);
84 }
85 
86 /* Read the current event from the event queue */
87 static inline efx_qword_t *efx_event(struct efx_channel *channel,
88 				     unsigned int index)
89 {
90 	return ((efx_qword_t *) (channel->eventq.addr)) +
91 		(index & channel->eventq_mask);
92 }
93 
94 /* See if an event is present
95  *
96  * We check both the high and low dword of the event for all ones.  We
97  * wrote all ones when we cleared the event, and no valid event can
98  * have all ones in either its high or low dwords.  This approach is
99  * robust against reordering.
100  *
101  * Note that using a single 64-bit comparison is incorrect; even
102  * though the CPU read will be atomic, the DMA write may not be.
103  */
104 static inline int efx_event_present(efx_qword_t *event)
105 {
106 	return !(EFX_DWORD_IS_ALL_ONES(event->dword[0]) |
107 		  EFX_DWORD_IS_ALL_ONES(event->dword[1]));
108 }
109 
110 static bool efx_masked_compare_oword(const efx_oword_t *a, const efx_oword_t *b,
111 				     const efx_oword_t *mask)
112 {
113 	return ((a->u64[0] ^ b->u64[0]) & mask->u64[0]) ||
114 		((a->u64[1] ^ b->u64[1]) & mask->u64[1]);
115 }
116 
117 int efx_nic_test_registers(struct efx_nic *efx,
118 			   const struct efx_nic_register_test *regs,
119 			   size_t n_regs)
120 {
121 	unsigned address = 0, i, j;
122 	efx_oword_t mask, imask, original, reg, buf;
123 
124 	/* Falcon should be in loopback to isolate the XMAC from the PHY */
125 	WARN_ON(!LOOPBACK_INTERNAL(efx));
126 
127 	for (i = 0; i < n_regs; ++i) {
128 		address = regs[i].address;
129 		mask = imask = regs[i].mask;
130 		EFX_INVERT_OWORD(imask);
131 
132 		efx_reado(efx, &original, address);
133 
134 		/* bit sweep on and off */
135 		for (j = 0; j < 128; j++) {
136 			if (!EFX_EXTRACT_OWORD32(mask, j, j))
137 				continue;
138 
139 			/* Test this testable bit can be set in isolation */
140 			EFX_AND_OWORD(reg, original, mask);
141 			EFX_SET_OWORD32(reg, j, j, 1);
142 
143 			efx_writeo(efx, &reg, address);
144 			efx_reado(efx, &buf, address);
145 
146 			if (efx_masked_compare_oword(&reg, &buf, &mask))
147 				goto fail;
148 
149 			/* Test this testable bit can be cleared in isolation */
150 			EFX_OR_OWORD(reg, original, mask);
151 			EFX_SET_OWORD32(reg, j, j, 0);
152 
153 			efx_writeo(efx, &reg, address);
154 			efx_reado(efx, &buf, address);
155 
156 			if (efx_masked_compare_oword(&reg, &buf, &mask))
157 				goto fail;
158 		}
159 
160 		efx_writeo(efx, &original, address);
161 	}
162 
163 	return 0;
164 
165 fail:
166 	netif_err(efx, hw, efx->net_dev,
167 		  "wrote "EFX_OWORD_FMT" read "EFX_OWORD_FMT
168 		  " at address 0x%x mask "EFX_OWORD_FMT"\n", EFX_OWORD_VAL(reg),
169 		  EFX_OWORD_VAL(buf), address, EFX_OWORD_VAL(mask));
170 	return -EIO;
171 }
172 
173 /**************************************************************************
174  *
175  * Special buffer handling
176  * Special buffers are used for event queues and the TX and RX
177  * descriptor rings.
178  *
179  *************************************************************************/
180 
181 /*
182  * Initialise a special buffer
183  *
184  * This will define a buffer (previously allocated via
185  * efx_alloc_special_buffer()) in the buffer table, allowing
186  * it to be used for event queues, descriptor rings etc.
187  */
188 static void
189 efx_init_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
190 {
191 	efx_qword_t buf_desc;
192 	int index;
193 	dma_addr_t dma_addr;
194 	int i;
195 
196 	EFX_BUG_ON_PARANOID(!buffer->addr);
197 
198 	/* Write buffer descriptors to NIC */
199 	for (i = 0; i < buffer->entries; i++) {
200 		index = buffer->index + i;
201 		dma_addr = buffer->dma_addr + (i * EFX_BUF_SIZE);
202 		netif_dbg(efx, probe, efx->net_dev,
203 			  "mapping special buffer %d at %llx\n",
204 			  index, (unsigned long long)dma_addr);
205 		EFX_POPULATE_QWORD_3(buf_desc,
206 				     FRF_AZ_BUF_ADR_REGION, 0,
207 				     FRF_AZ_BUF_ADR_FBUF, dma_addr >> 12,
208 				     FRF_AZ_BUF_OWNER_ID_FBUF, 0);
209 		efx_write_buf_tbl(efx, &buf_desc, index);
210 	}
211 }
212 
213 /* Unmaps a buffer and clears the buffer table entries */
214 static void
215 efx_fini_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
216 {
217 	efx_oword_t buf_tbl_upd;
218 	unsigned int start = buffer->index;
219 	unsigned int end = (buffer->index + buffer->entries - 1);
220 
221 	if (!buffer->entries)
222 		return;
223 
224 	netif_dbg(efx, hw, efx->net_dev, "unmapping special buffers %d-%d\n",
225 		  buffer->index, buffer->index + buffer->entries - 1);
226 
227 	EFX_POPULATE_OWORD_4(buf_tbl_upd,
228 			     FRF_AZ_BUF_UPD_CMD, 0,
229 			     FRF_AZ_BUF_CLR_CMD, 1,
230 			     FRF_AZ_BUF_CLR_END_ID, end,
231 			     FRF_AZ_BUF_CLR_START_ID, start);
232 	efx_writeo(efx, &buf_tbl_upd, FR_AZ_BUF_TBL_UPD);
233 }
234 
235 /*
236  * Allocate a new special buffer
237  *
238  * This allocates memory for a new buffer, clears it and allocates a
239  * new buffer ID range.  It does not write into the buffer table.
240  *
241  * This call will allocate 4KB buffers, since 8KB buffers can't be
242  * used for event queues and descriptor rings.
243  */
244 static int efx_alloc_special_buffer(struct efx_nic *efx,
245 				    struct efx_special_buffer *buffer,
246 				    unsigned int len)
247 {
248 	len = ALIGN(len, EFX_BUF_SIZE);
249 
250 	buffer->addr = dma_alloc_coherent(&efx->pci_dev->dev, len,
251 					  &buffer->dma_addr, GFP_KERNEL);
252 	if (!buffer->addr)
253 		return -ENOMEM;
254 	buffer->len = len;
255 	buffer->entries = len / EFX_BUF_SIZE;
256 	BUG_ON(buffer->dma_addr & (EFX_BUF_SIZE - 1));
257 
258 	/* All zeros is a potentially valid event so memset to 0xff */
259 	memset(buffer->addr, 0xff, len);
260 
261 	/* Select new buffer ID */
262 	buffer->index = efx->next_buffer_table;
263 	efx->next_buffer_table += buffer->entries;
264 
265 	netif_dbg(efx, probe, efx->net_dev,
266 		  "allocating special buffers %d-%d at %llx+%x "
267 		  "(virt %p phys %llx)\n", buffer->index,
268 		  buffer->index + buffer->entries - 1,
269 		  (u64)buffer->dma_addr, len,
270 		  buffer->addr, (u64)virt_to_phys(buffer->addr));
271 
272 	return 0;
273 }
274 
275 static void
276 efx_free_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
277 {
278 	if (!buffer->addr)
279 		return;
280 
281 	netif_dbg(efx, hw, efx->net_dev,
282 		  "deallocating special buffers %d-%d at %llx+%x "
283 		  "(virt %p phys %llx)\n", buffer->index,
284 		  buffer->index + buffer->entries - 1,
285 		  (u64)buffer->dma_addr, buffer->len,
286 		  buffer->addr, (u64)virt_to_phys(buffer->addr));
287 
288 	dma_free_coherent(&efx->pci_dev->dev, buffer->len, buffer->addr,
289 			  buffer->dma_addr);
290 	buffer->addr = NULL;
291 	buffer->entries = 0;
292 }
293 
294 /**************************************************************************
295  *
296  * Generic buffer handling
297  * These buffers are used for interrupt status and MAC stats
298  *
299  **************************************************************************/
300 
301 int efx_nic_alloc_buffer(struct efx_nic *efx, struct efx_buffer *buffer,
302 			 unsigned int len)
303 {
304 	buffer->addr = pci_alloc_consistent(efx->pci_dev, len,
305 					    &buffer->dma_addr);
306 	if (!buffer->addr)
307 		return -ENOMEM;
308 	buffer->len = len;
309 	memset(buffer->addr, 0, len);
310 	return 0;
311 }
312 
313 void efx_nic_free_buffer(struct efx_nic *efx, struct efx_buffer *buffer)
314 {
315 	if (buffer->addr) {
316 		pci_free_consistent(efx->pci_dev, buffer->len,
317 				    buffer->addr, buffer->dma_addr);
318 		buffer->addr = NULL;
319 	}
320 }
321 
322 /**************************************************************************
323  *
324  * TX path
325  *
326  **************************************************************************/
327 
328 /* Returns a pointer to the specified transmit descriptor in the TX
329  * descriptor queue belonging to the specified channel.
330  */
331 static inline efx_qword_t *
332 efx_tx_desc(struct efx_tx_queue *tx_queue, unsigned int index)
333 {
334 	return ((efx_qword_t *) (tx_queue->txd.addr)) + index;
335 }
336 
337 /* This writes to the TX_DESC_WPTR; write pointer for TX descriptor ring */
338 static inline void efx_notify_tx_desc(struct efx_tx_queue *tx_queue)
339 {
340 	unsigned write_ptr;
341 	efx_dword_t reg;
342 
343 	write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
344 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_TX_DESC_WPTR_DWORD, write_ptr);
345 	efx_writed_page(tx_queue->efx, &reg,
346 			FR_AZ_TX_DESC_UPD_DWORD_P0, tx_queue->queue);
347 }
348 
349 /* Write pointer and first descriptor for TX descriptor ring */
350 static inline void efx_push_tx_desc(struct efx_tx_queue *tx_queue,
351 				    const efx_qword_t *txd)
352 {
353 	unsigned write_ptr;
354 	efx_oword_t reg;
355 
356 	BUILD_BUG_ON(FRF_AZ_TX_DESC_LBN != 0);
357 	BUILD_BUG_ON(FR_AA_TX_DESC_UPD_KER != FR_BZ_TX_DESC_UPD_P0);
358 
359 	write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
360 	EFX_POPULATE_OWORD_2(reg, FRF_AZ_TX_DESC_PUSH_CMD, true,
361 			     FRF_AZ_TX_DESC_WPTR, write_ptr);
362 	reg.qword[0] = *txd;
363 	efx_writeo_page(tx_queue->efx, &reg,
364 			FR_BZ_TX_DESC_UPD_P0, tx_queue->queue);
365 }
366 
367 static inline bool
368 efx_may_push_tx_desc(struct efx_tx_queue *tx_queue, unsigned int write_count)
369 {
370 	unsigned empty_read_count = ACCESS_ONCE(tx_queue->empty_read_count);
371 
372 	if (empty_read_count == 0)
373 		return false;
374 
375 	tx_queue->empty_read_count = 0;
376 	return ((empty_read_count ^ write_count) & ~EFX_EMPTY_COUNT_VALID) == 0;
377 }
378 
379 /* For each entry inserted into the software descriptor ring, create a
380  * descriptor in the hardware TX descriptor ring (in host memory), and
381  * write a doorbell.
382  */
383 void efx_nic_push_buffers(struct efx_tx_queue *tx_queue)
384 {
385 
386 	struct efx_tx_buffer *buffer;
387 	efx_qword_t *txd;
388 	unsigned write_ptr;
389 	unsigned old_write_count = tx_queue->write_count;
390 
391 	BUG_ON(tx_queue->write_count == tx_queue->insert_count);
392 
393 	do {
394 		write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
395 		buffer = &tx_queue->buffer[write_ptr];
396 		txd = efx_tx_desc(tx_queue, write_ptr);
397 		++tx_queue->write_count;
398 
399 		/* Create TX descriptor ring entry */
400 		EFX_POPULATE_QWORD_4(*txd,
401 				     FSF_AZ_TX_KER_CONT, buffer->continuation,
402 				     FSF_AZ_TX_KER_BYTE_COUNT, buffer->len,
403 				     FSF_AZ_TX_KER_BUF_REGION, 0,
404 				     FSF_AZ_TX_KER_BUF_ADDR, buffer->dma_addr);
405 	} while (tx_queue->write_count != tx_queue->insert_count);
406 
407 	wmb(); /* Ensure descriptors are written before they are fetched */
408 
409 	if (efx_may_push_tx_desc(tx_queue, old_write_count)) {
410 		txd = efx_tx_desc(tx_queue,
411 				  old_write_count & tx_queue->ptr_mask);
412 		efx_push_tx_desc(tx_queue, txd);
413 		++tx_queue->pushes;
414 	} else {
415 		efx_notify_tx_desc(tx_queue);
416 	}
417 }
418 
419 /* Allocate hardware resources for a TX queue */
420 int efx_nic_probe_tx(struct efx_tx_queue *tx_queue)
421 {
422 	struct efx_nic *efx = tx_queue->efx;
423 	unsigned entries;
424 
425 	entries = tx_queue->ptr_mask + 1;
426 	return efx_alloc_special_buffer(efx, &tx_queue->txd,
427 					entries * sizeof(efx_qword_t));
428 }
429 
430 void efx_nic_init_tx(struct efx_tx_queue *tx_queue)
431 {
432 	struct efx_nic *efx = tx_queue->efx;
433 	efx_oword_t reg;
434 
435 	tx_queue->flushed = FLUSH_NONE;
436 
437 	/* Pin TX descriptor ring */
438 	efx_init_special_buffer(efx, &tx_queue->txd);
439 
440 	/* Push TX descriptor ring to card */
441 	EFX_POPULATE_OWORD_10(reg,
442 			      FRF_AZ_TX_DESCQ_EN, 1,
443 			      FRF_AZ_TX_ISCSI_DDIG_EN, 0,
444 			      FRF_AZ_TX_ISCSI_HDIG_EN, 0,
445 			      FRF_AZ_TX_DESCQ_BUF_BASE_ID, tx_queue->txd.index,
446 			      FRF_AZ_TX_DESCQ_EVQ_ID,
447 			      tx_queue->channel->channel,
448 			      FRF_AZ_TX_DESCQ_OWNER_ID, 0,
449 			      FRF_AZ_TX_DESCQ_LABEL, tx_queue->queue,
450 			      FRF_AZ_TX_DESCQ_SIZE,
451 			      __ffs(tx_queue->txd.entries),
452 			      FRF_AZ_TX_DESCQ_TYPE, 0,
453 			      FRF_BZ_TX_NON_IP_DROP_DIS, 1);
454 
455 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
456 		int csum = tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD;
457 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_IP_CHKSM_DIS, !csum);
458 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_TCP_CHKSM_DIS,
459 				    !csum);
460 	}
461 
462 	efx_writeo_table(efx, &reg, efx->type->txd_ptr_tbl_base,
463 			 tx_queue->queue);
464 
465 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0) {
466 		/* Only 128 bits in this register */
467 		BUILD_BUG_ON(EFX_MAX_TX_QUEUES > 128);
468 
469 		efx_reado(efx, &reg, FR_AA_TX_CHKSM_CFG);
470 		if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD)
471 			clear_bit_le(tx_queue->queue, (void *)&reg);
472 		else
473 			set_bit_le(tx_queue->queue, (void *)&reg);
474 		efx_writeo(efx, &reg, FR_AA_TX_CHKSM_CFG);
475 	}
476 
477 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
478 		EFX_POPULATE_OWORD_1(reg,
479 				     FRF_BZ_TX_PACE,
480 				     (tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
481 				     FFE_BZ_TX_PACE_OFF :
482 				     FFE_BZ_TX_PACE_RESERVED);
483 		efx_writeo_table(efx, &reg, FR_BZ_TX_PACE_TBL,
484 				 tx_queue->queue);
485 	}
486 }
487 
488 static void efx_flush_tx_queue(struct efx_tx_queue *tx_queue)
489 {
490 	struct efx_nic *efx = tx_queue->efx;
491 	efx_oword_t tx_flush_descq;
492 
493 	tx_queue->flushed = FLUSH_PENDING;
494 
495 	/* Post a flush command */
496 	EFX_POPULATE_OWORD_2(tx_flush_descq,
497 			     FRF_AZ_TX_FLUSH_DESCQ_CMD, 1,
498 			     FRF_AZ_TX_FLUSH_DESCQ, tx_queue->queue);
499 	efx_writeo(efx, &tx_flush_descq, FR_AZ_TX_FLUSH_DESCQ);
500 }
501 
502 void efx_nic_fini_tx(struct efx_tx_queue *tx_queue)
503 {
504 	struct efx_nic *efx = tx_queue->efx;
505 	efx_oword_t tx_desc_ptr;
506 
507 	/* The queue should have been flushed */
508 	WARN_ON(tx_queue->flushed != FLUSH_DONE);
509 
510 	/* Remove TX descriptor ring from card */
511 	EFX_ZERO_OWORD(tx_desc_ptr);
512 	efx_writeo_table(efx, &tx_desc_ptr, efx->type->txd_ptr_tbl_base,
513 			 tx_queue->queue);
514 
515 	/* Unpin TX descriptor ring */
516 	efx_fini_special_buffer(efx, &tx_queue->txd);
517 }
518 
519 /* Free buffers backing TX queue */
520 void efx_nic_remove_tx(struct efx_tx_queue *tx_queue)
521 {
522 	efx_free_special_buffer(tx_queue->efx, &tx_queue->txd);
523 }
524 
525 /**************************************************************************
526  *
527  * RX path
528  *
529  **************************************************************************/
530 
531 /* Returns a pointer to the specified descriptor in the RX descriptor queue */
532 static inline efx_qword_t *
533 efx_rx_desc(struct efx_rx_queue *rx_queue, unsigned int index)
534 {
535 	return ((efx_qword_t *) (rx_queue->rxd.addr)) + index;
536 }
537 
538 /* This creates an entry in the RX descriptor queue */
539 static inline void
540 efx_build_rx_desc(struct efx_rx_queue *rx_queue, unsigned index)
541 {
542 	struct efx_rx_buffer *rx_buf;
543 	efx_qword_t *rxd;
544 
545 	rxd = efx_rx_desc(rx_queue, index);
546 	rx_buf = efx_rx_buffer(rx_queue, index);
547 	EFX_POPULATE_QWORD_3(*rxd,
548 			     FSF_AZ_RX_KER_BUF_SIZE,
549 			     rx_buf->len -
550 			     rx_queue->efx->type->rx_buffer_padding,
551 			     FSF_AZ_RX_KER_BUF_REGION, 0,
552 			     FSF_AZ_RX_KER_BUF_ADDR, rx_buf->dma_addr);
553 }
554 
555 /* This writes to the RX_DESC_WPTR register for the specified receive
556  * descriptor ring.
557  */
558 void efx_nic_notify_rx_desc(struct efx_rx_queue *rx_queue)
559 {
560 	struct efx_nic *efx = rx_queue->efx;
561 	efx_dword_t reg;
562 	unsigned write_ptr;
563 
564 	while (rx_queue->notified_count != rx_queue->added_count) {
565 		efx_build_rx_desc(
566 			rx_queue,
567 			rx_queue->notified_count & rx_queue->ptr_mask);
568 		++rx_queue->notified_count;
569 	}
570 
571 	wmb();
572 	write_ptr = rx_queue->added_count & rx_queue->ptr_mask;
573 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_RX_DESC_WPTR_DWORD, write_ptr);
574 	efx_writed_page(efx, &reg, FR_AZ_RX_DESC_UPD_DWORD_P0,
575 			efx_rx_queue_index(rx_queue));
576 }
577 
578 int efx_nic_probe_rx(struct efx_rx_queue *rx_queue)
579 {
580 	struct efx_nic *efx = rx_queue->efx;
581 	unsigned entries;
582 
583 	entries = rx_queue->ptr_mask + 1;
584 	return efx_alloc_special_buffer(efx, &rx_queue->rxd,
585 					entries * sizeof(efx_qword_t));
586 }
587 
588 void efx_nic_init_rx(struct efx_rx_queue *rx_queue)
589 {
590 	efx_oword_t rx_desc_ptr;
591 	struct efx_nic *efx = rx_queue->efx;
592 	bool is_b0 = efx_nic_rev(efx) >= EFX_REV_FALCON_B0;
593 	bool iscsi_digest_en = is_b0;
594 
595 	netif_dbg(efx, hw, efx->net_dev,
596 		  "RX queue %d ring in special buffers %d-%d\n",
597 		  efx_rx_queue_index(rx_queue), rx_queue->rxd.index,
598 		  rx_queue->rxd.index + rx_queue->rxd.entries - 1);
599 
600 	rx_queue->flushed = FLUSH_NONE;
601 
602 	/* Pin RX descriptor ring */
603 	efx_init_special_buffer(efx, &rx_queue->rxd);
604 
605 	/* Push RX descriptor ring to card */
606 	EFX_POPULATE_OWORD_10(rx_desc_ptr,
607 			      FRF_AZ_RX_ISCSI_DDIG_EN, iscsi_digest_en,
608 			      FRF_AZ_RX_ISCSI_HDIG_EN, iscsi_digest_en,
609 			      FRF_AZ_RX_DESCQ_BUF_BASE_ID, rx_queue->rxd.index,
610 			      FRF_AZ_RX_DESCQ_EVQ_ID,
611 			      efx_rx_queue_channel(rx_queue)->channel,
612 			      FRF_AZ_RX_DESCQ_OWNER_ID, 0,
613 			      FRF_AZ_RX_DESCQ_LABEL,
614 			      efx_rx_queue_index(rx_queue),
615 			      FRF_AZ_RX_DESCQ_SIZE,
616 			      __ffs(rx_queue->rxd.entries),
617 			      FRF_AZ_RX_DESCQ_TYPE, 0 /* kernel queue */ ,
618 			      /* For >=B0 this is scatter so disable */
619 			      FRF_AZ_RX_DESCQ_JUMBO, !is_b0,
620 			      FRF_AZ_RX_DESCQ_EN, 1);
621 	efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
622 			 efx_rx_queue_index(rx_queue));
623 }
624 
625 static void efx_flush_rx_queue(struct efx_rx_queue *rx_queue)
626 {
627 	struct efx_nic *efx = rx_queue->efx;
628 	efx_oword_t rx_flush_descq;
629 
630 	rx_queue->flushed = FLUSH_PENDING;
631 
632 	/* Post a flush command */
633 	EFX_POPULATE_OWORD_2(rx_flush_descq,
634 			     FRF_AZ_RX_FLUSH_DESCQ_CMD, 1,
635 			     FRF_AZ_RX_FLUSH_DESCQ,
636 			     efx_rx_queue_index(rx_queue));
637 	efx_writeo(efx, &rx_flush_descq, FR_AZ_RX_FLUSH_DESCQ);
638 }
639 
640 void efx_nic_fini_rx(struct efx_rx_queue *rx_queue)
641 {
642 	efx_oword_t rx_desc_ptr;
643 	struct efx_nic *efx = rx_queue->efx;
644 
645 	/* The queue should already have been flushed */
646 	WARN_ON(rx_queue->flushed != FLUSH_DONE);
647 
648 	/* Remove RX descriptor ring from card */
649 	EFX_ZERO_OWORD(rx_desc_ptr);
650 	efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
651 			 efx_rx_queue_index(rx_queue));
652 
653 	/* Unpin RX descriptor ring */
654 	efx_fini_special_buffer(efx, &rx_queue->rxd);
655 }
656 
657 /* Free buffers backing RX queue */
658 void efx_nic_remove_rx(struct efx_rx_queue *rx_queue)
659 {
660 	efx_free_special_buffer(rx_queue->efx, &rx_queue->rxd);
661 }
662 
663 /**************************************************************************
664  *
665  * Event queue processing
666  * Event queues are processed by per-channel tasklets.
667  *
668  **************************************************************************/
669 
670 /* Update a channel's event queue's read pointer (RPTR) register
671  *
672  * This writes the EVQ_RPTR_REG register for the specified channel's
673  * event queue.
674  */
675 void efx_nic_eventq_read_ack(struct efx_channel *channel)
676 {
677 	efx_dword_t reg;
678 	struct efx_nic *efx = channel->efx;
679 
680 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_EVQ_RPTR,
681 			     channel->eventq_read_ptr & channel->eventq_mask);
682 	efx_writed_table(efx, &reg, efx->type->evq_rptr_tbl_base,
683 			 channel->channel);
684 }
685 
686 /* Use HW to insert a SW defined event */
687 static void efx_generate_event(struct efx_channel *channel, efx_qword_t *event)
688 {
689 	efx_oword_t drv_ev_reg;
690 
691 	BUILD_BUG_ON(FRF_AZ_DRV_EV_DATA_LBN != 0 ||
692 		     FRF_AZ_DRV_EV_DATA_WIDTH != 64);
693 	drv_ev_reg.u32[0] = event->u32[0];
694 	drv_ev_reg.u32[1] = event->u32[1];
695 	drv_ev_reg.u32[2] = 0;
696 	drv_ev_reg.u32[3] = 0;
697 	EFX_SET_OWORD_FIELD(drv_ev_reg, FRF_AZ_DRV_EV_QID, channel->channel);
698 	efx_writeo(channel->efx, &drv_ev_reg, FR_AZ_DRV_EV);
699 }
700 
701 static void efx_magic_event(struct efx_channel *channel, u32 magic)
702 {
703 	efx_qword_t event;
704 
705 	EFX_POPULATE_QWORD_2(event, FSF_AZ_EV_CODE,
706 			     FSE_AZ_EV_CODE_DRV_GEN_EV,
707 			     FSF_AZ_DRV_GEN_EV_MAGIC, magic);
708 	efx_generate_event(channel, &event);
709 }
710 
711 /* Handle a transmit completion event
712  *
713  * The NIC batches TX completion events; the message we receive is of
714  * the form "complete all TX events up to this index".
715  */
716 static int
717 efx_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
718 {
719 	unsigned int tx_ev_desc_ptr;
720 	unsigned int tx_ev_q_label;
721 	struct efx_tx_queue *tx_queue;
722 	struct efx_nic *efx = channel->efx;
723 	int tx_packets = 0;
724 
725 	if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
726 		/* Transmit completion */
727 		tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
728 		tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
729 		tx_queue = efx_channel_get_tx_queue(
730 			channel, tx_ev_q_label % EFX_TXQ_TYPES);
731 		tx_packets = ((tx_ev_desc_ptr - tx_queue->read_count) &
732 			      tx_queue->ptr_mask);
733 		channel->irq_mod_score += tx_packets;
734 		efx_xmit_done(tx_queue, tx_ev_desc_ptr);
735 	} else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
736 		/* Rewrite the FIFO write pointer */
737 		tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
738 		tx_queue = efx_channel_get_tx_queue(
739 			channel, tx_ev_q_label % EFX_TXQ_TYPES);
740 
741 		netif_tx_lock(efx->net_dev);
742 		efx_notify_tx_desc(tx_queue);
743 		netif_tx_unlock(efx->net_dev);
744 	} else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR) &&
745 		   EFX_WORKAROUND_10727(efx)) {
746 		efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
747 	} else {
748 		netif_err(efx, tx_err, efx->net_dev,
749 			  "channel %d unexpected TX event "
750 			  EFX_QWORD_FMT"\n", channel->channel,
751 			  EFX_QWORD_VAL(*event));
752 	}
753 
754 	return tx_packets;
755 }
756 
757 /* Detect errors included in the rx_evt_pkt_ok bit. */
758 static u16 efx_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
759 				const efx_qword_t *event)
760 {
761 	struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
762 	struct efx_nic *efx = rx_queue->efx;
763 	bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
764 	bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
765 	bool rx_ev_frm_trunc, rx_ev_drib_nib, rx_ev_tobe_disc;
766 	bool rx_ev_other_err, rx_ev_pause_frm;
767 	bool rx_ev_hdr_type, rx_ev_mcast_pkt;
768 	unsigned rx_ev_pkt_type;
769 
770 	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
771 	rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
772 	rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
773 	rx_ev_pkt_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_TYPE);
774 	rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
775 						 FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
776 	rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
777 						  FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
778 	rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
779 						   FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
780 	rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
781 	rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
782 	rx_ev_drib_nib = ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) ?
783 			  0 : EFX_QWORD_FIELD(*event, FSF_AA_RX_EV_DRIB_NIB));
784 	rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
785 
786 	/* Every error apart from tobe_disc and pause_frm */
787 	rx_ev_other_err = (rx_ev_drib_nib | rx_ev_tcp_udp_chksum_err |
788 			   rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
789 			   rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
790 
791 	/* Count errors that are not in MAC stats.  Ignore expected
792 	 * checksum errors during self-test. */
793 	if (rx_ev_frm_trunc)
794 		++channel->n_rx_frm_trunc;
795 	else if (rx_ev_tobe_disc)
796 		++channel->n_rx_tobe_disc;
797 	else if (!efx->loopback_selftest) {
798 		if (rx_ev_ip_hdr_chksum_err)
799 			++channel->n_rx_ip_hdr_chksum_err;
800 		else if (rx_ev_tcp_udp_chksum_err)
801 			++channel->n_rx_tcp_udp_chksum_err;
802 	}
803 
804 	/* TOBE_DISC is expected on unicast mismatches; don't print out an
805 	 * error message.  FRM_TRUNC indicates RXDP dropped the packet due
806 	 * to a FIFO overflow.
807 	 */
808 #ifdef DEBUG
809 	if (rx_ev_other_err && net_ratelimit()) {
810 		netif_dbg(efx, rx_err, efx->net_dev,
811 			  " RX queue %d unexpected RX event "
812 			  EFX_QWORD_FMT "%s%s%s%s%s%s%s%s\n",
813 			  efx_rx_queue_index(rx_queue), EFX_QWORD_VAL(*event),
814 			  rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
815 			  rx_ev_ip_hdr_chksum_err ?
816 			  " [IP_HDR_CHKSUM_ERR]" : "",
817 			  rx_ev_tcp_udp_chksum_err ?
818 			  " [TCP_UDP_CHKSUM_ERR]" : "",
819 			  rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
820 			  rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
821 			  rx_ev_drib_nib ? " [DRIB_NIB]" : "",
822 			  rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
823 			  rx_ev_pause_frm ? " [PAUSE]" : "");
824 	}
825 #endif
826 
827 	/* The frame must be discarded if any of these are true. */
828 	return (rx_ev_eth_crc_err | rx_ev_frm_trunc | rx_ev_drib_nib |
829 		rx_ev_tobe_disc | rx_ev_pause_frm) ?
830 		EFX_RX_PKT_DISCARD : 0;
831 }
832 
833 /* Handle receive events that are not in-order. */
834 static void
835 efx_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
836 {
837 	struct efx_nic *efx = rx_queue->efx;
838 	unsigned expected, dropped;
839 
840 	expected = rx_queue->removed_count & rx_queue->ptr_mask;
841 	dropped = (index - expected) & rx_queue->ptr_mask;
842 	netif_info(efx, rx_err, efx->net_dev,
843 		   "dropped %d events (index=%d expected=%d)\n",
844 		   dropped, index, expected);
845 
846 	efx_schedule_reset(efx, EFX_WORKAROUND_5676(efx) ?
847 			   RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
848 }
849 
850 /* Handle a packet received event
851  *
852  * The NIC gives a "discard" flag if it's a unicast packet with the
853  * wrong destination address
854  * Also "is multicast" and "matches multicast filter" flags can be used to
855  * discard non-matching multicast packets.
856  */
857 static void
858 efx_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
859 {
860 	unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
861 	unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
862 	unsigned expected_ptr;
863 	bool rx_ev_pkt_ok;
864 	u16 flags;
865 	struct efx_rx_queue *rx_queue;
866 
867 	/* Basic packet information */
868 	rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
869 	rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
870 	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
871 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT));
872 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP) != 1);
873 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
874 		channel->channel);
875 
876 	rx_queue = efx_channel_get_rx_queue(channel);
877 
878 	rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
879 	expected_ptr = rx_queue->removed_count & rx_queue->ptr_mask;
880 	if (unlikely(rx_ev_desc_ptr != expected_ptr))
881 		efx_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr);
882 
883 	if (likely(rx_ev_pkt_ok)) {
884 		/* If packet is marked as OK and packet type is TCP/IP or
885 		 * UDP/IP, then we can rely on the hardware checksum.
886 		 */
887 		flags = (rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP ||
888 			 rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP) ?
889 			EFX_RX_PKT_CSUMMED : 0;
890 	} else {
891 		flags = efx_handle_rx_not_ok(rx_queue, event);
892 	}
893 
894 	/* Detect multicast packets that didn't match the filter */
895 	rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
896 	if (rx_ev_mcast_pkt) {
897 		unsigned int rx_ev_mcast_hash_match =
898 			EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
899 
900 		if (unlikely(!rx_ev_mcast_hash_match)) {
901 			++channel->n_rx_mcast_mismatch;
902 			flags |= EFX_RX_PKT_DISCARD;
903 		}
904 	}
905 
906 	channel->irq_mod_score += 2;
907 
908 	/* Handle received packet */
909 	efx_rx_packet(rx_queue, rx_ev_desc_ptr, rx_ev_byte_cnt, flags);
910 }
911 
912 static void
913 efx_handle_generated_event(struct efx_channel *channel, efx_qword_t *event)
914 {
915 	struct efx_nic *efx = channel->efx;
916 	struct efx_rx_queue *rx_queue =
917 		efx_channel_has_rx_queue(channel) ?
918 		efx_channel_get_rx_queue(channel) : NULL;
919 	unsigned magic;
920 
921 	magic = EFX_QWORD_FIELD(*event, FSF_AZ_DRV_GEN_EV_MAGIC);
922 
923 	if (magic == EFX_CHANNEL_MAGIC_TEST(channel))
924 		; /* ignore */
925 	else if (rx_queue && magic == EFX_CHANNEL_MAGIC_FILL(rx_queue))
926 		/* The queue must be empty, so we won't receive any rx
927 		 * events, so efx_process_channel() won't refill the
928 		 * queue. Refill it here */
929 		efx_fast_push_rx_descriptors(rx_queue);
930 	else
931 		netif_dbg(efx, hw, efx->net_dev, "channel %d received "
932 			  "generated event "EFX_QWORD_FMT"\n",
933 			  channel->channel, EFX_QWORD_VAL(*event));
934 }
935 
936 static void
937 efx_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
938 {
939 	struct efx_nic *efx = channel->efx;
940 	unsigned int ev_sub_code;
941 	unsigned int ev_sub_data;
942 
943 	ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
944 	ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
945 
946 	switch (ev_sub_code) {
947 	case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
948 		netif_vdbg(efx, hw, efx->net_dev, "channel %d TXQ %d flushed\n",
949 			   channel->channel, ev_sub_data);
950 		break;
951 	case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
952 		netif_vdbg(efx, hw, efx->net_dev, "channel %d RXQ %d flushed\n",
953 			   channel->channel, ev_sub_data);
954 		break;
955 	case FSE_AZ_EVQ_INIT_DONE_EV:
956 		netif_dbg(efx, hw, efx->net_dev,
957 			  "channel %d EVQ %d initialised\n",
958 			  channel->channel, ev_sub_data);
959 		break;
960 	case FSE_AZ_SRM_UPD_DONE_EV:
961 		netif_vdbg(efx, hw, efx->net_dev,
962 			   "channel %d SRAM update done\n", channel->channel);
963 		break;
964 	case FSE_AZ_WAKE_UP_EV:
965 		netif_vdbg(efx, hw, efx->net_dev,
966 			   "channel %d RXQ %d wakeup event\n",
967 			   channel->channel, ev_sub_data);
968 		break;
969 	case FSE_AZ_TIMER_EV:
970 		netif_vdbg(efx, hw, efx->net_dev,
971 			   "channel %d RX queue %d timer expired\n",
972 			   channel->channel, ev_sub_data);
973 		break;
974 	case FSE_AA_RX_RECOVER_EV:
975 		netif_err(efx, rx_err, efx->net_dev,
976 			  "channel %d seen DRIVER RX_RESET event. "
977 			"Resetting.\n", channel->channel);
978 		atomic_inc(&efx->rx_reset);
979 		efx_schedule_reset(efx,
980 				   EFX_WORKAROUND_6555(efx) ?
981 				   RESET_TYPE_RX_RECOVERY :
982 				   RESET_TYPE_DISABLE);
983 		break;
984 	case FSE_BZ_RX_DSC_ERROR_EV:
985 		netif_err(efx, rx_err, efx->net_dev,
986 			  "RX DMA Q %d reports descriptor fetch error."
987 			  " RX Q %d is disabled.\n", ev_sub_data, ev_sub_data);
988 		efx_schedule_reset(efx, RESET_TYPE_RX_DESC_FETCH);
989 		break;
990 	case FSE_BZ_TX_DSC_ERROR_EV:
991 		netif_err(efx, tx_err, efx->net_dev,
992 			  "TX DMA Q %d reports descriptor fetch error."
993 			  " TX Q %d is disabled.\n", ev_sub_data, ev_sub_data);
994 		efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
995 		break;
996 	default:
997 		netif_vdbg(efx, hw, efx->net_dev,
998 			   "channel %d unknown driver event code %d "
999 			   "data %04x\n", channel->channel, ev_sub_code,
1000 			   ev_sub_data);
1001 		break;
1002 	}
1003 }
1004 
1005 int efx_nic_process_eventq(struct efx_channel *channel, int budget)
1006 {
1007 	struct efx_nic *efx = channel->efx;
1008 	unsigned int read_ptr;
1009 	efx_qword_t event, *p_event;
1010 	int ev_code;
1011 	int tx_packets = 0;
1012 	int spent = 0;
1013 
1014 	read_ptr = channel->eventq_read_ptr;
1015 
1016 	for (;;) {
1017 		p_event = efx_event(channel, read_ptr);
1018 		event = *p_event;
1019 
1020 		if (!efx_event_present(&event))
1021 			/* End of events */
1022 			break;
1023 
1024 		netif_vdbg(channel->efx, intr, channel->efx->net_dev,
1025 			   "channel %d event is "EFX_QWORD_FMT"\n",
1026 			   channel->channel, EFX_QWORD_VAL(event));
1027 
1028 		/* Clear this event by marking it all ones */
1029 		EFX_SET_QWORD(*p_event);
1030 
1031 		++read_ptr;
1032 
1033 		ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
1034 
1035 		switch (ev_code) {
1036 		case FSE_AZ_EV_CODE_RX_EV:
1037 			efx_handle_rx_event(channel, &event);
1038 			if (++spent == budget)
1039 				goto out;
1040 			break;
1041 		case FSE_AZ_EV_CODE_TX_EV:
1042 			tx_packets += efx_handle_tx_event(channel, &event);
1043 			if (tx_packets > efx->txq_entries) {
1044 				spent = budget;
1045 				goto out;
1046 			}
1047 			break;
1048 		case FSE_AZ_EV_CODE_DRV_GEN_EV:
1049 			efx_handle_generated_event(channel, &event);
1050 			break;
1051 		case FSE_AZ_EV_CODE_DRIVER_EV:
1052 			efx_handle_driver_event(channel, &event);
1053 			break;
1054 		case FSE_CZ_EV_CODE_MCDI_EV:
1055 			efx_mcdi_process_event(channel, &event);
1056 			break;
1057 		case FSE_AZ_EV_CODE_GLOBAL_EV:
1058 			if (efx->type->handle_global_event &&
1059 			    efx->type->handle_global_event(channel, &event))
1060 				break;
1061 			/* else fall through */
1062 		default:
1063 			netif_err(channel->efx, hw, channel->efx->net_dev,
1064 				  "channel %d unknown event type %d (data "
1065 				  EFX_QWORD_FMT ")\n", channel->channel,
1066 				  ev_code, EFX_QWORD_VAL(event));
1067 		}
1068 	}
1069 
1070 out:
1071 	channel->eventq_read_ptr = read_ptr;
1072 	return spent;
1073 }
1074 
1075 /* Check whether an event is present in the eventq at the current
1076  * read pointer.  Only useful for self-test.
1077  */
1078 bool efx_nic_event_present(struct efx_channel *channel)
1079 {
1080 	return efx_event_present(efx_event(channel, channel->eventq_read_ptr));
1081 }
1082 
1083 /* Allocate buffer table entries for event queue */
1084 int efx_nic_probe_eventq(struct efx_channel *channel)
1085 {
1086 	struct efx_nic *efx = channel->efx;
1087 	unsigned entries;
1088 
1089 	entries = channel->eventq_mask + 1;
1090 	return efx_alloc_special_buffer(efx, &channel->eventq,
1091 					entries * sizeof(efx_qword_t));
1092 }
1093 
1094 void efx_nic_init_eventq(struct efx_channel *channel)
1095 {
1096 	efx_oword_t reg;
1097 	struct efx_nic *efx = channel->efx;
1098 
1099 	netif_dbg(efx, hw, efx->net_dev,
1100 		  "channel %d event queue in special buffers %d-%d\n",
1101 		  channel->channel, channel->eventq.index,
1102 		  channel->eventq.index + channel->eventq.entries - 1);
1103 
1104 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1105 		EFX_POPULATE_OWORD_3(reg,
1106 				     FRF_CZ_TIMER_Q_EN, 1,
1107 				     FRF_CZ_HOST_NOTIFY_MODE, 0,
1108 				     FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
1109 		efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1110 	}
1111 
1112 	/* Pin event queue buffer */
1113 	efx_init_special_buffer(efx, &channel->eventq);
1114 
1115 	/* Fill event queue with all ones (i.e. empty events) */
1116 	memset(channel->eventq.addr, 0xff, channel->eventq.len);
1117 
1118 	/* Push event queue to card */
1119 	EFX_POPULATE_OWORD_3(reg,
1120 			     FRF_AZ_EVQ_EN, 1,
1121 			     FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
1122 			     FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
1123 	efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1124 			 channel->channel);
1125 
1126 	efx->type->push_irq_moderation(channel);
1127 }
1128 
1129 void efx_nic_fini_eventq(struct efx_channel *channel)
1130 {
1131 	efx_oword_t reg;
1132 	struct efx_nic *efx = channel->efx;
1133 
1134 	/* Remove event queue from card */
1135 	EFX_ZERO_OWORD(reg);
1136 	efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1137 			 channel->channel);
1138 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1139 		efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1140 
1141 	/* Unpin event queue */
1142 	efx_fini_special_buffer(efx, &channel->eventq);
1143 }
1144 
1145 /* Free buffers backing event queue */
1146 void efx_nic_remove_eventq(struct efx_channel *channel)
1147 {
1148 	efx_free_special_buffer(channel->efx, &channel->eventq);
1149 }
1150 
1151 
1152 void efx_nic_generate_test_event(struct efx_channel *channel)
1153 {
1154 	efx_magic_event(channel, EFX_CHANNEL_MAGIC_TEST(channel));
1155 }
1156 
1157 void efx_nic_generate_fill_event(struct efx_rx_queue *rx_queue)
1158 {
1159 	efx_magic_event(efx_rx_queue_channel(rx_queue),
1160 			EFX_CHANNEL_MAGIC_FILL(rx_queue));
1161 }
1162 
1163 /**************************************************************************
1164  *
1165  * Flush handling
1166  *
1167  **************************************************************************/
1168 
1169 
1170 static void efx_poll_flush_events(struct efx_nic *efx)
1171 {
1172 	struct efx_channel *channel = efx_get_channel(efx, 0);
1173 	struct efx_tx_queue *tx_queue;
1174 	struct efx_rx_queue *rx_queue;
1175 	unsigned int read_ptr = channel->eventq_read_ptr;
1176 	unsigned int end_ptr = read_ptr + channel->eventq_mask - 1;
1177 
1178 	do {
1179 		efx_qword_t *event = efx_event(channel, read_ptr);
1180 		int ev_code, ev_sub_code, ev_queue;
1181 		bool ev_failed;
1182 
1183 		if (!efx_event_present(event))
1184 			break;
1185 
1186 		ev_code = EFX_QWORD_FIELD(*event, FSF_AZ_EV_CODE);
1187 		ev_sub_code = EFX_QWORD_FIELD(*event,
1188 					      FSF_AZ_DRIVER_EV_SUBCODE);
1189 		if (ev_code == FSE_AZ_EV_CODE_DRIVER_EV &&
1190 		    ev_sub_code == FSE_AZ_TX_DESCQ_FLS_DONE_EV) {
1191 			ev_queue = EFX_QWORD_FIELD(*event,
1192 						   FSF_AZ_DRIVER_EV_SUBDATA);
1193 			if (ev_queue < EFX_TXQ_TYPES * efx->n_tx_channels) {
1194 				tx_queue = efx_get_tx_queue(
1195 					efx, ev_queue / EFX_TXQ_TYPES,
1196 					ev_queue % EFX_TXQ_TYPES);
1197 				tx_queue->flushed = FLUSH_DONE;
1198 			}
1199 		} else if (ev_code == FSE_AZ_EV_CODE_DRIVER_EV &&
1200 			   ev_sub_code == FSE_AZ_RX_DESCQ_FLS_DONE_EV) {
1201 			ev_queue = EFX_QWORD_FIELD(
1202 				*event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
1203 			ev_failed = EFX_QWORD_FIELD(
1204 				*event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
1205 			if (ev_queue < efx->n_rx_channels) {
1206 				rx_queue = efx_get_rx_queue(efx, ev_queue);
1207 				rx_queue->flushed =
1208 					ev_failed ? FLUSH_FAILED : FLUSH_DONE;
1209 			}
1210 		}
1211 
1212 		/* We're about to destroy the queue anyway, so
1213 		 * it's ok to throw away every non-flush event */
1214 		EFX_SET_QWORD(*event);
1215 
1216 		++read_ptr;
1217 	} while (read_ptr != end_ptr);
1218 
1219 	channel->eventq_read_ptr = read_ptr;
1220 }
1221 
1222 /* Handle tx and rx flushes at the same time, since they run in
1223  * parallel in the hardware and there's no reason for us to
1224  * serialise them */
1225 int efx_nic_flush_queues(struct efx_nic *efx)
1226 {
1227 	struct efx_channel *channel;
1228 	struct efx_rx_queue *rx_queue;
1229 	struct efx_tx_queue *tx_queue;
1230 	int i, tx_pending, rx_pending;
1231 
1232 	/* If necessary prepare the hardware for flushing */
1233 	efx->type->prepare_flush(efx);
1234 
1235 	/* Flush all tx queues in parallel */
1236 	efx_for_each_channel(channel, efx) {
1237 		efx_for_each_possible_channel_tx_queue(tx_queue, channel) {
1238 			if (tx_queue->initialised)
1239 				efx_flush_tx_queue(tx_queue);
1240 		}
1241 	}
1242 
1243 	/* The hardware supports four concurrent rx flushes, each of which may
1244 	 * need to be retried if there is an outstanding descriptor fetch */
1245 	for (i = 0; i < EFX_FLUSH_POLL_COUNT; ++i) {
1246 		rx_pending = tx_pending = 0;
1247 		efx_for_each_channel(channel, efx) {
1248 			efx_for_each_channel_rx_queue(rx_queue, channel) {
1249 				if (rx_queue->flushed == FLUSH_PENDING)
1250 					++rx_pending;
1251 			}
1252 		}
1253 		efx_for_each_channel(channel, efx) {
1254 			efx_for_each_channel_rx_queue(rx_queue, channel) {
1255 				if (rx_pending == EFX_RX_FLUSH_COUNT)
1256 					break;
1257 				if (rx_queue->flushed == FLUSH_FAILED ||
1258 				    rx_queue->flushed == FLUSH_NONE) {
1259 					efx_flush_rx_queue(rx_queue);
1260 					++rx_pending;
1261 				}
1262 			}
1263 			efx_for_each_possible_channel_tx_queue(tx_queue, channel) {
1264 				if (tx_queue->initialised &&
1265 				    tx_queue->flushed != FLUSH_DONE)
1266 					++tx_pending;
1267 			}
1268 		}
1269 
1270 		if (rx_pending == 0 && tx_pending == 0)
1271 			return 0;
1272 
1273 		msleep(EFX_FLUSH_INTERVAL);
1274 		efx_poll_flush_events(efx);
1275 	}
1276 
1277 	/* Mark the queues as all flushed. We're going to return failure
1278 	 * leading to a reset, or fake up success anyway */
1279 	efx_for_each_channel(channel, efx) {
1280 		efx_for_each_possible_channel_tx_queue(tx_queue, channel) {
1281 			if (tx_queue->initialised &&
1282 			    tx_queue->flushed != FLUSH_DONE)
1283 				netif_err(efx, hw, efx->net_dev,
1284 					  "tx queue %d flush command timed out\n",
1285 					  tx_queue->queue);
1286 			tx_queue->flushed = FLUSH_DONE;
1287 		}
1288 		efx_for_each_channel_rx_queue(rx_queue, channel) {
1289 			if (rx_queue->flushed != FLUSH_DONE)
1290 				netif_err(efx, hw, efx->net_dev,
1291 					  "rx queue %d flush command timed out\n",
1292 					  efx_rx_queue_index(rx_queue));
1293 			rx_queue->flushed = FLUSH_DONE;
1294 		}
1295 	}
1296 
1297 	return -ETIMEDOUT;
1298 }
1299 
1300 /**************************************************************************
1301  *
1302  * Hardware interrupts
1303  * The hardware interrupt handler does very little work; all the event
1304  * queue processing is carried out by per-channel tasklets.
1305  *
1306  **************************************************************************/
1307 
1308 /* Enable/disable/generate interrupts */
1309 static inline void efx_nic_interrupts(struct efx_nic *efx,
1310 				      bool enabled, bool force)
1311 {
1312 	efx_oword_t int_en_reg_ker;
1313 
1314 	EFX_POPULATE_OWORD_3(int_en_reg_ker,
1315 			     FRF_AZ_KER_INT_LEVE_SEL, efx->irq_level,
1316 			     FRF_AZ_KER_INT_KER, force,
1317 			     FRF_AZ_DRV_INT_EN_KER, enabled);
1318 	efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
1319 }
1320 
1321 void efx_nic_enable_interrupts(struct efx_nic *efx)
1322 {
1323 	struct efx_channel *channel;
1324 
1325 	EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
1326 	wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
1327 
1328 	/* Enable interrupts */
1329 	efx_nic_interrupts(efx, true, false);
1330 
1331 	/* Force processing of all the channels to get the EVQ RPTRs up to
1332 	   date */
1333 	efx_for_each_channel(channel, efx)
1334 		efx_schedule_channel(channel);
1335 }
1336 
1337 void efx_nic_disable_interrupts(struct efx_nic *efx)
1338 {
1339 	/* Disable interrupts */
1340 	efx_nic_interrupts(efx, false, false);
1341 }
1342 
1343 /* Generate a test interrupt
1344  * Interrupt must already have been enabled, otherwise nasty things
1345  * may happen.
1346  */
1347 void efx_nic_generate_interrupt(struct efx_nic *efx)
1348 {
1349 	efx_nic_interrupts(efx, true, true);
1350 }
1351 
1352 /* Process a fatal interrupt
1353  * Disable bus mastering ASAP and schedule a reset
1354  */
1355 irqreturn_t efx_nic_fatal_interrupt(struct efx_nic *efx)
1356 {
1357 	struct falcon_nic_data *nic_data = efx->nic_data;
1358 	efx_oword_t *int_ker = efx->irq_status.addr;
1359 	efx_oword_t fatal_intr;
1360 	int error, mem_perr;
1361 
1362 	efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
1363 	error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
1364 
1365 	netif_err(efx, hw, efx->net_dev, "SYSTEM ERROR "EFX_OWORD_FMT" status "
1366 		  EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
1367 		  EFX_OWORD_VAL(fatal_intr),
1368 		  error ? "disabling bus mastering" : "no recognised error");
1369 
1370 	/* If this is a memory parity error dump which blocks are offending */
1371 	mem_perr = (EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER) ||
1372 		    EFX_OWORD_FIELD(fatal_intr, FRF_AZ_SRM_PERR_INT_KER));
1373 	if (mem_perr) {
1374 		efx_oword_t reg;
1375 		efx_reado(efx, &reg, FR_AZ_MEM_STAT);
1376 		netif_err(efx, hw, efx->net_dev,
1377 			  "SYSTEM ERROR: memory parity error "EFX_OWORD_FMT"\n",
1378 			  EFX_OWORD_VAL(reg));
1379 	}
1380 
1381 	/* Disable both devices */
1382 	pci_clear_master(efx->pci_dev);
1383 	if (efx_nic_is_dual_func(efx))
1384 		pci_clear_master(nic_data->pci_dev2);
1385 	efx_nic_disable_interrupts(efx);
1386 
1387 	/* Count errors and reset or disable the NIC accordingly */
1388 	if (efx->int_error_count == 0 ||
1389 	    time_after(jiffies, efx->int_error_expire)) {
1390 		efx->int_error_count = 0;
1391 		efx->int_error_expire =
1392 			jiffies + EFX_INT_ERROR_EXPIRE * HZ;
1393 	}
1394 	if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
1395 		netif_err(efx, hw, efx->net_dev,
1396 			  "SYSTEM ERROR - reset scheduled\n");
1397 		efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
1398 	} else {
1399 		netif_err(efx, hw, efx->net_dev,
1400 			  "SYSTEM ERROR - max number of errors seen."
1401 			  "NIC will be disabled\n");
1402 		efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1403 	}
1404 
1405 	return IRQ_HANDLED;
1406 }
1407 
1408 /* Handle a legacy interrupt
1409  * Acknowledges the interrupt and schedule event queue processing.
1410  */
1411 static irqreturn_t efx_legacy_interrupt(int irq, void *dev_id)
1412 {
1413 	struct efx_nic *efx = dev_id;
1414 	efx_oword_t *int_ker = efx->irq_status.addr;
1415 	irqreturn_t result = IRQ_NONE;
1416 	struct efx_channel *channel;
1417 	efx_dword_t reg;
1418 	u32 queues;
1419 	int syserr;
1420 
1421 	/* Could this be ours?  If interrupts are disabled then the
1422 	 * channel state may not be valid.
1423 	 */
1424 	if (!efx->legacy_irq_enabled)
1425 		return result;
1426 
1427 	/* Read the ISR which also ACKs the interrupts */
1428 	efx_readd(efx, &reg, FR_BZ_INT_ISR0);
1429 	queues = EFX_EXTRACT_DWORD(reg, 0, 31);
1430 
1431 	/* Handle non-event-queue sources */
1432 	if (queues & (1U << efx->irq_level)) {
1433 		syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1434 		if (unlikely(syserr))
1435 			return efx_nic_fatal_interrupt(efx);
1436 		efx->last_irq_cpu = raw_smp_processor_id();
1437 	}
1438 
1439 	if (queues != 0) {
1440 		if (EFX_WORKAROUND_15783(efx))
1441 			efx->irq_zero_count = 0;
1442 
1443 		/* Schedule processing of any interrupting queues */
1444 		efx_for_each_channel(channel, efx) {
1445 			if (queues & 1)
1446 				efx_schedule_channel_irq(channel);
1447 			queues >>= 1;
1448 		}
1449 		result = IRQ_HANDLED;
1450 
1451 	} else if (EFX_WORKAROUND_15783(efx)) {
1452 		efx_qword_t *event;
1453 
1454 		/* We can't return IRQ_HANDLED more than once on seeing ISR=0
1455 		 * because this might be a shared interrupt. */
1456 		if (efx->irq_zero_count++ == 0)
1457 			result = IRQ_HANDLED;
1458 
1459 		/* Ensure we schedule or rearm all event queues */
1460 		efx_for_each_channel(channel, efx) {
1461 			event = efx_event(channel, channel->eventq_read_ptr);
1462 			if (efx_event_present(event))
1463 				efx_schedule_channel_irq(channel);
1464 			else
1465 				efx_nic_eventq_read_ack(channel);
1466 		}
1467 	}
1468 
1469 	if (result == IRQ_HANDLED)
1470 		netif_vdbg(efx, intr, efx->net_dev,
1471 			   "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
1472 			   irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
1473 
1474 	return result;
1475 }
1476 
1477 /* Handle an MSI interrupt
1478  *
1479  * Handle an MSI hardware interrupt.  This routine schedules event
1480  * queue processing.  No interrupt acknowledgement cycle is necessary.
1481  * Also, we never need to check that the interrupt is for us, since
1482  * MSI interrupts cannot be shared.
1483  */
1484 static irqreturn_t efx_msi_interrupt(int irq, void *dev_id)
1485 {
1486 	struct efx_channel *channel = *(struct efx_channel **)dev_id;
1487 	struct efx_nic *efx = channel->efx;
1488 	efx_oword_t *int_ker = efx->irq_status.addr;
1489 	int syserr;
1490 
1491 	netif_vdbg(efx, intr, efx->net_dev,
1492 		   "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
1493 		   irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
1494 
1495 	/* Handle non-event-queue sources */
1496 	if (channel->channel == efx->irq_level) {
1497 		syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1498 		if (unlikely(syserr))
1499 			return efx_nic_fatal_interrupt(efx);
1500 		efx->last_irq_cpu = raw_smp_processor_id();
1501 	}
1502 
1503 	/* Schedule processing of the channel */
1504 	efx_schedule_channel_irq(channel);
1505 
1506 	return IRQ_HANDLED;
1507 }
1508 
1509 
1510 /* Setup RSS indirection table.
1511  * This maps from the hash value of the packet to RXQ
1512  */
1513 void efx_nic_push_rx_indir_table(struct efx_nic *efx)
1514 {
1515 	size_t i = 0;
1516 	efx_dword_t dword;
1517 
1518 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0)
1519 		return;
1520 
1521 	BUILD_BUG_ON(ARRAY_SIZE(efx->rx_indir_table) !=
1522 		     FR_BZ_RX_INDIRECTION_TBL_ROWS);
1523 
1524 	for (i = 0; i < FR_BZ_RX_INDIRECTION_TBL_ROWS; i++) {
1525 		EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
1526 				     efx->rx_indir_table[i]);
1527 		efx_writed_table(efx, &dword, FR_BZ_RX_INDIRECTION_TBL, i);
1528 	}
1529 }
1530 
1531 /* Hook interrupt handler(s)
1532  * Try MSI and then legacy interrupts.
1533  */
1534 int efx_nic_init_interrupt(struct efx_nic *efx)
1535 {
1536 	struct efx_channel *channel;
1537 	int rc;
1538 
1539 	if (!EFX_INT_MODE_USE_MSI(efx)) {
1540 		irq_handler_t handler;
1541 		if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1542 			handler = efx_legacy_interrupt;
1543 		else
1544 			handler = falcon_legacy_interrupt_a1;
1545 
1546 		rc = request_irq(efx->legacy_irq, handler, IRQF_SHARED,
1547 				 efx->name, efx);
1548 		if (rc) {
1549 			netif_err(efx, drv, efx->net_dev,
1550 				  "failed to hook legacy IRQ %d\n",
1551 				  efx->pci_dev->irq);
1552 			goto fail1;
1553 		}
1554 		return 0;
1555 	}
1556 
1557 	/* Hook MSI or MSI-X interrupt */
1558 	efx_for_each_channel(channel, efx) {
1559 		rc = request_irq(channel->irq, efx_msi_interrupt,
1560 				 IRQF_PROBE_SHARED, /* Not shared */
1561 				 efx->channel_name[channel->channel],
1562 				 &efx->channel[channel->channel]);
1563 		if (rc) {
1564 			netif_err(efx, drv, efx->net_dev,
1565 				  "failed to hook IRQ %d\n", channel->irq);
1566 			goto fail2;
1567 		}
1568 	}
1569 
1570 	return 0;
1571 
1572  fail2:
1573 	efx_for_each_channel(channel, efx)
1574 		free_irq(channel->irq, &efx->channel[channel->channel]);
1575  fail1:
1576 	return rc;
1577 }
1578 
1579 void efx_nic_fini_interrupt(struct efx_nic *efx)
1580 {
1581 	struct efx_channel *channel;
1582 	efx_oword_t reg;
1583 
1584 	/* Disable MSI/MSI-X interrupts */
1585 	efx_for_each_channel(channel, efx) {
1586 		if (channel->irq)
1587 			free_irq(channel->irq, &efx->channel[channel->channel]);
1588 	}
1589 
1590 	/* ACK legacy interrupt */
1591 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1592 		efx_reado(efx, &reg, FR_BZ_INT_ISR0);
1593 	else
1594 		falcon_irq_ack_a1(efx);
1595 
1596 	/* Disable legacy interrupt */
1597 	if (efx->legacy_irq)
1598 		free_irq(efx->legacy_irq, efx);
1599 }
1600 
1601 u32 efx_nic_fpga_ver(struct efx_nic *efx)
1602 {
1603 	efx_oword_t altera_build;
1604 	efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
1605 	return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
1606 }
1607 
1608 void efx_nic_init_common(struct efx_nic *efx)
1609 {
1610 	efx_oword_t temp;
1611 
1612 	/* Set positions of descriptor caches in SRAM. */
1613 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR,
1614 			     efx->type->tx_dc_base / 8);
1615 	efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
1616 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR,
1617 			     efx->type->rx_dc_base / 8);
1618 	efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
1619 
1620 	/* Set TX descriptor cache size. */
1621 	BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
1622 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
1623 	efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
1624 
1625 	/* Set RX descriptor cache size.  Set low watermark to size-8, as
1626 	 * this allows most efficient prefetching.
1627 	 */
1628 	BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
1629 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
1630 	efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
1631 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
1632 	efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
1633 
1634 	/* Program INT_KER address */
1635 	EFX_POPULATE_OWORD_2(temp,
1636 			     FRF_AZ_NORM_INT_VEC_DIS_KER,
1637 			     EFX_INT_MODE_USE_MSI(efx),
1638 			     FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
1639 	efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
1640 
1641 	if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
1642 		/* Use an interrupt level unused by event queues */
1643 		efx->irq_level = 0x1f;
1644 	else
1645 		/* Use a valid MSI-X vector */
1646 		efx->irq_level = 0;
1647 
1648 	/* Enable all the genuinely fatal interrupts.  (They are still
1649 	 * masked by the overall interrupt mask, controlled by
1650 	 * falcon_interrupts()).
1651 	 *
1652 	 * Note: All other fatal interrupts are enabled
1653 	 */
1654 	EFX_POPULATE_OWORD_3(temp,
1655 			     FRF_AZ_ILL_ADR_INT_KER_EN, 1,
1656 			     FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
1657 			     FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
1658 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1659 		EFX_SET_OWORD_FIELD(temp, FRF_CZ_SRAM_PERR_INT_P_KER_EN, 1);
1660 	EFX_INVERT_OWORD(temp);
1661 	efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
1662 
1663 	efx_nic_push_rx_indir_table(efx);
1664 
1665 	/* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
1666 	 * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
1667 	 */
1668 	efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
1669 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
1670 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
1671 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
1672 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 1);
1673 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
1674 	/* Enable SW_EV to inherit in char driver - assume harmless here */
1675 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
1676 	/* Prefetch threshold 2 => fetch when descriptor cache half empty */
1677 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
1678 	/* Disable hardware watchdog which can misfire */
1679 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
1680 	/* Squash TX of packets of 16 bytes or less */
1681 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1682 		EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
1683 	efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
1684 
1685 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
1686 		EFX_POPULATE_OWORD_4(temp,
1687 				     /* Default values */
1688 				     FRF_BZ_TX_PACE_SB_NOT_AF, 0x15,
1689 				     FRF_BZ_TX_PACE_SB_AF, 0xb,
1690 				     FRF_BZ_TX_PACE_FB_BASE, 0,
1691 				     /* Allow large pace values in the
1692 				      * fast bin. */
1693 				     FRF_BZ_TX_PACE_BIN_TH,
1694 				     FFE_BZ_TX_PACE_RESERVED);
1695 		efx_writeo(efx, &temp, FR_BZ_TX_PACE);
1696 	}
1697 }
1698 
1699 /* Register dump */
1700 
1701 #define REGISTER_REVISION_A	1
1702 #define REGISTER_REVISION_B	2
1703 #define REGISTER_REVISION_C	3
1704 #define REGISTER_REVISION_Z	3	/* latest revision */
1705 
1706 struct efx_nic_reg {
1707 	u32 offset:24;
1708 	u32 min_revision:2, max_revision:2;
1709 };
1710 
1711 #define REGISTER(name, min_rev, max_rev) {				\
1712 	FR_ ## min_rev ## max_rev ## _ ## name,				\
1713 	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev	\
1714 }
1715 #define REGISTER_AA(name) REGISTER(name, A, A)
1716 #define REGISTER_AB(name) REGISTER(name, A, B)
1717 #define REGISTER_AZ(name) REGISTER(name, A, Z)
1718 #define REGISTER_BB(name) REGISTER(name, B, B)
1719 #define REGISTER_BZ(name) REGISTER(name, B, Z)
1720 #define REGISTER_CZ(name) REGISTER(name, C, Z)
1721 
1722 static const struct efx_nic_reg efx_nic_regs[] = {
1723 	REGISTER_AZ(ADR_REGION),
1724 	REGISTER_AZ(INT_EN_KER),
1725 	REGISTER_BZ(INT_EN_CHAR),
1726 	REGISTER_AZ(INT_ADR_KER),
1727 	REGISTER_BZ(INT_ADR_CHAR),
1728 	/* INT_ACK_KER is WO */
1729 	/* INT_ISR0 is RC */
1730 	REGISTER_AZ(HW_INIT),
1731 	REGISTER_CZ(USR_EV_CFG),
1732 	REGISTER_AB(EE_SPI_HCMD),
1733 	REGISTER_AB(EE_SPI_HADR),
1734 	REGISTER_AB(EE_SPI_HDATA),
1735 	REGISTER_AB(EE_BASE_PAGE),
1736 	REGISTER_AB(EE_VPD_CFG0),
1737 	/* EE_VPD_SW_CNTL and EE_VPD_SW_DATA are not used */
1738 	/* PMBX_DBG_IADDR and PBMX_DBG_IDATA are indirect */
1739 	/* PCIE_CORE_INDIRECT is indirect */
1740 	REGISTER_AB(NIC_STAT),
1741 	REGISTER_AB(GPIO_CTL),
1742 	REGISTER_AB(GLB_CTL),
1743 	/* FATAL_INTR_KER and FATAL_INTR_CHAR are partly RC */
1744 	REGISTER_BZ(DP_CTRL),
1745 	REGISTER_AZ(MEM_STAT),
1746 	REGISTER_AZ(CS_DEBUG),
1747 	REGISTER_AZ(ALTERA_BUILD),
1748 	REGISTER_AZ(CSR_SPARE),
1749 	REGISTER_AB(PCIE_SD_CTL0123),
1750 	REGISTER_AB(PCIE_SD_CTL45),
1751 	REGISTER_AB(PCIE_PCS_CTL_STAT),
1752 	/* DEBUG_DATA_OUT is not used */
1753 	/* DRV_EV is WO */
1754 	REGISTER_AZ(EVQ_CTL),
1755 	REGISTER_AZ(EVQ_CNT1),
1756 	REGISTER_AZ(EVQ_CNT2),
1757 	REGISTER_AZ(BUF_TBL_CFG),
1758 	REGISTER_AZ(SRM_RX_DC_CFG),
1759 	REGISTER_AZ(SRM_TX_DC_CFG),
1760 	REGISTER_AZ(SRM_CFG),
1761 	/* BUF_TBL_UPD is WO */
1762 	REGISTER_AZ(SRM_UPD_EVQ),
1763 	REGISTER_AZ(SRAM_PARITY),
1764 	REGISTER_AZ(RX_CFG),
1765 	REGISTER_BZ(RX_FILTER_CTL),
1766 	/* RX_FLUSH_DESCQ is WO */
1767 	REGISTER_AZ(RX_DC_CFG),
1768 	REGISTER_AZ(RX_DC_PF_WM),
1769 	REGISTER_BZ(RX_RSS_TKEY),
1770 	/* RX_NODESC_DROP is RC */
1771 	REGISTER_AA(RX_SELF_RST),
1772 	/* RX_DEBUG, RX_PUSH_DROP are not used */
1773 	REGISTER_CZ(RX_RSS_IPV6_REG1),
1774 	REGISTER_CZ(RX_RSS_IPV6_REG2),
1775 	REGISTER_CZ(RX_RSS_IPV6_REG3),
1776 	/* TX_FLUSH_DESCQ is WO */
1777 	REGISTER_AZ(TX_DC_CFG),
1778 	REGISTER_AA(TX_CHKSM_CFG),
1779 	REGISTER_AZ(TX_CFG),
1780 	/* TX_PUSH_DROP is not used */
1781 	REGISTER_AZ(TX_RESERVED),
1782 	REGISTER_BZ(TX_PACE),
1783 	/* TX_PACE_DROP_QID is RC */
1784 	REGISTER_BB(TX_VLAN),
1785 	REGISTER_BZ(TX_IPFIL_PORTEN),
1786 	REGISTER_AB(MD_TXD),
1787 	REGISTER_AB(MD_RXD),
1788 	REGISTER_AB(MD_CS),
1789 	REGISTER_AB(MD_PHY_ADR),
1790 	REGISTER_AB(MD_ID),
1791 	/* MD_STAT is RC */
1792 	REGISTER_AB(MAC_STAT_DMA),
1793 	REGISTER_AB(MAC_CTRL),
1794 	REGISTER_BB(GEN_MODE),
1795 	REGISTER_AB(MAC_MC_HASH_REG0),
1796 	REGISTER_AB(MAC_MC_HASH_REG1),
1797 	REGISTER_AB(GM_CFG1),
1798 	REGISTER_AB(GM_CFG2),
1799 	/* GM_IPG and GM_HD are not used */
1800 	REGISTER_AB(GM_MAX_FLEN),
1801 	/* GM_TEST is not used */
1802 	REGISTER_AB(GM_ADR1),
1803 	REGISTER_AB(GM_ADR2),
1804 	REGISTER_AB(GMF_CFG0),
1805 	REGISTER_AB(GMF_CFG1),
1806 	REGISTER_AB(GMF_CFG2),
1807 	REGISTER_AB(GMF_CFG3),
1808 	REGISTER_AB(GMF_CFG4),
1809 	REGISTER_AB(GMF_CFG5),
1810 	REGISTER_BB(TX_SRC_MAC_CTL),
1811 	REGISTER_AB(XM_ADR_LO),
1812 	REGISTER_AB(XM_ADR_HI),
1813 	REGISTER_AB(XM_GLB_CFG),
1814 	REGISTER_AB(XM_TX_CFG),
1815 	REGISTER_AB(XM_RX_CFG),
1816 	REGISTER_AB(XM_MGT_INT_MASK),
1817 	REGISTER_AB(XM_FC),
1818 	REGISTER_AB(XM_PAUSE_TIME),
1819 	REGISTER_AB(XM_TX_PARAM),
1820 	REGISTER_AB(XM_RX_PARAM),
1821 	/* XM_MGT_INT_MSK (note no 'A') is RC */
1822 	REGISTER_AB(XX_PWR_RST),
1823 	REGISTER_AB(XX_SD_CTL),
1824 	REGISTER_AB(XX_TXDRV_CTL),
1825 	/* XX_PRBS_CTL, XX_PRBS_CHK and XX_PRBS_ERR are not used */
1826 	/* XX_CORE_STAT is partly RC */
1827 };
1828 
1829 struct efx_nic_reg_table {
1830 	u32 offset:24;
1831 	u32 min_revision:2, max_revision:2;
1832 	u32 step:6, rows:21;
1833 };
1834 
1835 #define REGISTER_TABLE_DIMENSIONS(_, offset, min_rev, max_rev, step, rows) { \
1836 	offset,								\
1837 	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev,	\
1838 	step, rows							\
1839 }
1840 #define REGISTER_TABLE(name, min_rev, max_rev)				\
1841 	REGISTER_TABLE_DIMENSIONS(					\
1842 		name, FR_ ## min_rev ## max_rev ## _ ## name,		\
1843 		min_rev, max_rev,					\
1844 		FR_ ## min_rev ## max_rev ## _ ## name ## _STEP,	\
1845 		FR_ ## min_rev ## max_rev ## _ ## name ## _ROWS)
1846 #define REGISTER_TABLE_AA(name) REGISTER_TABLE(name, A, A)
1847 #define REGISTER_TABLE_AZ(name) REGISTER_TABLE(name, A, Z)
1848 #define REGISTER_TABLE_BB(name) REGISTER_TABLE(name, B, B)
1849 #define REGISTER_TABLE_BZ(name) REGISTER_TABLE(name, B, Z)
1850 #define REGISTER_TABLE_BB_CZ(name)					\
1851 	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, B, B,		\
1852 				  FR_BZ_ ## name ## _STEP,		\
1853 				  FR_BB_ ## name ## _ROWS),		\
1854 	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, C, Z,		\
1855 				  FR_BZ_ ## name ## _STEP,		\
1856 				  FR_CZ_ ## name ## _ROWS)
1857 #define REGISTER_TABLE_CZ(name) REGISTER_TABLE(name, C, Z)
1858 
1859 static const struct efx_nic_reg_table efx_nic_reg_tables[] = {
1860 	/* DRIVER is not used */
1861 	/* EVQ_RPTR, TIMER_COMMAND, USR_EV and {RX,TX}_DESC_UPD are WO */
1862 	REGISTER_TABLE_BB(TX_IPFIL_TBL),
1863 	REGISTER_TABLE_BB(TX_SRC_MAC_TBL),
1864 	REGISTER_TABLE_AA(RX_DESC_PTR_TBL_KER),
1865 	REGISTER_TABLE_BB_CZ(RX_DESC_PTR_TBL),
1866 	REGISTER_TABLE_AA(TX_DESC_PTR_TBL_KER),
1867 	REGISTER_TABLE_BB_CZ(TX_DESC_PTR_TBL),
1868 	REGISTER_TABLE_AA(EVQ_PTR_TBL_KER),
1869 	REGISTER_TABLE_BB_CZ(EVQ_PTR_TBL),
1870 	/* We can't reasonably read all of the buffer table (up to 8MB!).
1871 	 * However this driver will only use a few entries.  Reading
1872 	 * 1K entries allows for some expansion of queue count and
1873 	 * size before we need to change the version. */
1874 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL_KER, FR_AA_BUF_FULL_TBL_KER,
1875 				  A, A, 8, 1024),
1876 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL, FR_BZ_BUF_FULL_TBL,
1877 				  B, Z, 8, 1024),
1878 	REGISTER_TABLE_CZ(RX_MAC_FILTER_TBL0),
1879 	REGISTER_TABLE_BB_CZ(TIMER_TBL),
1880 	REGISTER_TABLE_BB_CZ(TX_PACE_TBL),
1881 	REGISTER_TABLE_BZ(RX_INDIRECTION_TBL),
1882 	/* TX_FILTER_TBL0 is huge and not used by this driver */
1883 	REGISTER_TABLE_CZ(TX_MAC_FILTER_TBL0),
1884 	REGISTER_TABLE_CZ(MC_TREG_SMEM),
1885 	/* MSIX_PBA_TABLE is not mapped */
1886 	/* SRM_DBG is not mapped (and is redundant with BUF_FLL_TBL) */
1887 	REGISTER_TABLE_BZ(RX_FILTER_TBL0),
1888 };
1889 
1890 size_t efx_nic_get_regs_len(struct efx_nic *efx)
1891 {
1892 	const struct efx_nic_reg *reg;
1893 	const struct efx_nic_reg_table *table;
1894 	size_t len = 0;
1895 
1896 	for (reg = efx_nic_regs;
1897 	     reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
1898 	     reg++)
1899 		if (efx->type->revision >= reg->min_revision &&
1900 		    efx->type->revision <= reg->max_revision)
1901 			len += sizeof(efx_oword_t);
1902 
1903 	for (table = efx_nic_reg_tables;
1904 	     table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
1905 	     table++)
1906 		if (efx->type->revision >= table->min_revision &&
1907 		    efx->type->revision <= table->max_revision)
1908 			len += table->rows * min_t(size_t, table->step, 16);
1909 
1910 	return len;
1911 }
1912 
1913 void efx_nic_get_regs(struct efx_nic *efx, void *buf)
1914 {
1915 	const struct efx_nic_reg *reg;
1916 	const struct efx_nic_reg_table *table;
1917 
1918 	for (reg = efx_nic_regs;
1919 	     reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
1920 	     reg++) {
1921 		if (efx->type->revision >= reg->min_revision &&
1922 		    efx->type->revision <= reg->max_revision) {
1923 			efx_reado(efx, (efx_oword_t *)buf, reg->offset);
1924 			buf += sizeof(efx_oword_t);
1925 		}
1926 	}
1927 
1928 	for (table = efx_nic_reg_tables;
1929 	     table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
1930 	     table++) {
1931 		size_t size, i;
1932 
1933 		if (!(efx->type->revision >= table->min_revision &&
1934 		      efx->type->revision <= table->max_revision))
1935 			continue;
1936 
1937 		size = min_t(size_t, table->step, 16);
1938 
1939 		for (i = 0; i < table->rows; i++) {
1940 			switch (table->step) {
1941 			case 4: /* 32-bit register or SRAM */
1942 				efx_readd_table(efx, buf, table->offset, i);
1943 				break;
1944 			case 8: /* 64-bit SRAM */
1945 				efx_sram_readq(efx,
1946 					       efx->membase + table->offset,
1947 					       buf, i);
1948 				break;
1949 			case 16: /* 128-bit register */
1950 				efx_reado_table(efx, buf, table->offset, i);
1951 				break;
1952 			case 32: /* 128-bit register, interleaved */
1953 				efx_reado_table(efx, buf, table->offset, 2 * i);
1954 				break;
1955 			default:
1956 				WARN_ON(1);
1957 				return;
1958 			}
1959 			buf += size;
1960 		}
1961 	}
1962 }
1963