xref: /openbmc/linux/drivers/net/ethernet/sfc/nic.c (revision 4ef594eb)
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(_channel)				\
70 	_EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_FILL, (_channel)->channel)
71 
72 /**************************************************************************
73  *
74  * Solarstorm hardware access
75  *
76  **************************************************************************/
77 
78 static inline void efx_write_buf_tbl(struct efx_nic *efx, efx_qword_t *value,
79 				     unsigned int index)
80 {
81 	efx_sram_writeq(efx, efx->membase + efx->type->buf_tbl_base,
82 			value, index);
83 }
84 
85 /* Read the current event from the event queue */
86 static inline efx_qword_t *efx_event(struct efx_channel *channel,
87 				     unsigned int index)
88 {
89 	return ((efx_qword_t *) (channel->eventq.addr)) +
90 		(index & channel->eventq_mask);
91 }
92 
93 /* See if an event is present
94  *
95  * We check both the high and low dword of the event for all ones.  We
96  * wrote all ones when we cleared the event, and no valid event can
97  * have all ones in either its high or low dwords.  This approach is
98  * robust against reordering.
99  *
100  * Note that using a single 64-bit comparison is incorrect; even
101  * though the CPU read will be atomic, the DMA write may not be.
102  */
103 static inline int efx_event_present(efx_qword_t *event)
104 {
105 	return !(EFX_DWORD_IS_ALL_ONES(event->dword[0]) |
106 		  EFX_DWORD_IS_ALL_ONES(event->dword[1]));
107 }
108 
109 static bool efx_masked_compare_oword(const efx_oword_t *a, const efx_oword_t *b,
110 				     const efx_oword_t *mask)
111 {
112 	return ((a->u64[0] ^ b->u64[0]) & mask->u64[0]) ||
113 		((a->u64[1] ^ b->u64[1]) & mask->u64[1]);
114 }
115 
116 int efx_nic_test_registers(struct efx_nic *efx,
117 			   const struct efx_nic_register_test *regs,
118 			   size_t n_regs)
119 {
120 	unsigned address = 0, i, j;
121 	efx_oword_t mask, imask, original, reg, buf;
122 
123 	/* Falcon should be in loopback to isolate the XMAC from the PHY */
124 	WARN_ON(!LOOPBACK_INTERNAL(efx));
125 
126 	for (i = 0; i < n_regs; ++i) {
127 		address = regs[i].address;
128 		mask = imask = regs[i].mask;
129 		EFX_INVERT_OWORD(imask);
130 
131 		efx_reado(efx, &original, address);
132 
133 		/* bit sweep on and off */
134 		for (j = 0; j < 128; j++) {
135 			if (!EFX_EXTRACT_OWORD32(mask, j, j))
136 				continue;
137 
138 			/* Test this testable bit can be set in isolation */
139 			EFX_AND_OWORD(reg, original, mask);
140 			EFX_SET_OWORD32(reg, j, j, 1);
141 
142 			efx_writeo(efx, &reg, address);
143 			efx_reado(efx, &buf, address);
144 
145 			if (efx_masked_compare_oword(&reg, &buf, &mask))
146 				goto fail;
147 
148 			/* Test this testable bit can be cleared in isolation */
149 			EFX_OR_OWORD(reg, original, mask);
150 			EFX_SET_OWORD32(reg, j, j, 0);
151 
152 			efx_writeo(efx, &reg, address);
153 			efx_reado(efx, &buf, address);
154 
155 			if (efx_masked_compare_oword(&reg, &buf, &mask))
156 				goto fail;
157 		}
158 
159 		efx_writeo(efx, &original, address);
160 	}
161 
162 	return 0;
163 
164 fail:
165 	netif_err(efx, hw, efx->net_dev,
166 		  "wrote "EFX_OWORD_FMT" read "EFX_OWORD_FMT
167 		  " at address 0x%x mask "EFX_OWORD_FMT"\n", EFX_OWORD_VAL(reg),
168 		  EFX_OWORD_VAL(buf), address, EFX_OWORD_VAL(mask));
169 	return -EIO;
170 }
171 
172 /**************************************************************************
173  *
174  * Special buffer handling
175  * Special buffers are used for event queues and the TX and RX
176  * descriptor rings.
177  *
178  *************************************************************************/
179 
180 /*
181  * Initialise a special buffer
182  *
183  * This will define a buffer (previously allocated via
184  * efx_alloc_special_buffer()) in the buffer table, allowing
185  * it to be used for event queues, descriptor rings etc.
186  */
187 static void
188 efx_init_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
189 {
190 	efx_qword_t buf_desc;
191 	int index;
192 	dma_addr_t dma_addr;
193 	int i;
194 
195 	EFX_BUG_ON_PARANOID(!buffer->addr);
196 
197 	/* Write buffer descriptors to NIC */
198 	for (i = 0; i < buffer->entries; i++) {
199 		index = buffer->index + i;
200 		dma_addr = buffer->dma_addr + (i * EFX_BUF_SIZE);
201 		netif_dbg(efx, probe, efx->net_dev,
202 			  "mapping special buffer %d at %llx\n",
203 			  index, (unsigned long long)dma_addr);
204 		EFX_POPULATE_QWORD_3(buf_desc,
205 				     FRF_AZ_BUF_ADR_REGION, 0,
206 				     FRF_AZ_BUF_ADR_FBUF, dma_addr >> 12,
207 				     FRF_AZ_BUF_OWNER_ID_FBUF, 0);
208 		efx_write_buf_tbl(efx, &buf_desc, index);
209 	}
210 }
211 
212 /* Unmaps a buffer and clears the buffer table entries */
213 static void
214 efx_fini_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
215 {
216 	efx_oword_t buf_tbl_upd;
217 	unsigned int start = buffer->index;
218 	unsigned int end = (buffer->index + buffer->entries - 1);
219 
220 	if (!buffer->entries)
221 		return;
222 
223 	netif_dbg(efx, hw, efx->net_dev, "unmapping special buffers %d-%d\n",
224 		  buffer->index, buffer->index + buffer->entries - 1);
225 
226 	EFX_POPULATE_OWORD_4(buf_tbl_upd,
227 			     FRF_AZ_BUF_UPD_CMD, 0,
228 			     FRF_AZ_BUF_CLR_CMD, 1,
229 			     FRF_AZ_BUF_CLR_END_ID, end,
230 			     FRF_AZ_BUF_CLR_START_ID, start);
231 	efx_writeo(efx, &buf_tbl_upd, FR_AZ_BUF_TBL_UPD);
232 }
233 
234 /*
235  * Allocate a new special buffer
236  *
237  * This allocates memory for a new buffer, clears it and allocates a
238  * new buffer ID range.  It does not write into the buffer table.
239  *
240  * This call will allocate 4KB buffers, since 8KB buffers can't be
241  * used for event queues and descriptor rings.
242  */
243 static int efx_alloc_special_buffer(struct efx_nic *efx,
244 				    struct efx_special_buffer *buffer,
245 				    unsigned int len)
246 {
247 	len = ALIGN(len, EFX_BUF_SIZE);
248 
249 	buffer->addr = dma_alloc_coherent(&efx->pci_dev->dev, len,
250 					  &buffer->dma_addr, GFP_KERNEL);
251 	if (!buffer->addr)
252 		return -ENOMEM;
253 	buffer->len = len;
254 	buffer->entries = len / EFX_BUF_SIZE;
255 	BUG_ON(buffer->dma_addr & (EFX_BUF_SIZE - 1));
256 
257 	/* All zeros is a potentially valid event so memset to 0xff */
258 	memset(buffer->addr, 0xff, len);
259 
260 	/* Select new buffer ID */
261 	buffer->index = efx->next_buffer_table;
262 	efx->next_buffer_table += buffer->entries;
263 
264 	netif_dbg(efx, probe, efx->net_dev,
265 		  "allocating special buffers %d-%d at %llx+%x "
266 		  "(virt %p phys %llx)\n", buffer->index,
267 		  buffer->index + buffer->entries - 1,
268 		  (u64)buffer->dma_addr, len,
269 		  buffer->addr, (u64)virt_to_phys(buffer->addr));
270 
271 	return 0;
272 }
273 
274 static void
275 efx_free_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
276 {
277 	if (!buffer->addr)
278 		return;
279 
280 	netif_dbg(efx, hw, efx->net_dev,
281 		  "deallocating special buffers %d-%d at %llx+%x "
282 		  "(virt %p phys %llx)\n", buffer->index,
283 		  buffer->index + buffer->entries - 1,
284 		  (u64)buffer->dma_addr, buffer->len,
285 		  buffer->addr, (u64)virt_to_phys(buffer->addr));
286 
287 	dma_free_coherent(&efx->pci_dev->dev, buffer->len, buffer->addr,
288 			  buffer->dma_addr);
289 	buffer->addr = NULL;
290 	buffer->entries = 0;
291 }
292 
293 /**************************************************************************
294  *
295  * Generic buffer handling
296  * These buffers are used for interrupt status and MAC stats
297  *
298  **************************************************************************/
299 
300 int efx_nic_alloc_buffer(struct efx_nic *efx, struct efx_buffer *buffer,
301 			 unsigned int len)
302 {
303 	buffer->addr = pci_alloc_consistent(efx->pci_dev, len,
304 					    &buffer->dma_addr);
305 	if (!buffer->addr)
306 		return -ENOMEM;
307 	buffer->len = len;
308 	memset(buffer->addr, 0, len);
309 	return 0;
310 }
311 
312 void efx_nic_free_buffer(struct efx_nic *efx, struct efx_buffer *buffer)
313 {
314 	if (buffer->addr) {
315 		pci_free_consistent(efx->pci_dev, buffer->len,
316 				    buffer->addr, buffer->dma_addr);
317 		buffer->addr = NULL;
318 	}
319 }
320 
321 /**************************************************************************
322  *
323  * TX path
324  *
325  **************************************************************************/
326 
327 /* Returns a pointer to the specified transmit descriptor in the TX
328  * descriptor queue belonging to the specified channel.
329  */
330 static inline efx_qword_t *
331 efx_tx_desc(struct efx_tx_queue *tx_queue, unsigned int index)
332 {
333 	return ((efx_qword_t *) (tx_queue->txd.addr)) + index;
334 }
335 
336 /* This writes to the TX_DESC_WPTR; write pointer for TX descriptor ring */
337 static inline void efx_notify_tx_desc(struct efx_tx_queue *tx_queue)
338 {
339 	unsigned write_ptr;
340 	efx_dword_t reg;
341 
342 	write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
343 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_TX_DESC_WPTR_DWORD, write_ptr);
344 	efx_writed_page(tx_queue->efx, &reg,
345 			FR_AZ_TX_DESC_UPD_DWORD_P0, tx_queue->queue);
346 }
347 
348 /* Write pointer and first descriptor for TX descriptor ring */
349 static inline void efx_push_tx_desc(struct efx_tx_queue *tx_queue,
350 				    const efx_qword_t *txd)
351 {
352 	unsigned write_ptr;
353 	efx_oword_t reg;
354 
355 	BUILD_BUG_ON(FRF_AZ_TX_DESC_LBN != 0);
356 	BUILD_BUG_ON(FR_AA_TX_DESC_UPD_KER != FR_BZ_TX_DESC_UPD_P0);
357 
358 	write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
359 	EFX_POPULATE_OWORD_2(reg, FRF_AZ_TX_DESC_PUSH_CMD, true,
360 			     FRF_AZ_TX_DESC_WPTR, write_ptr);
361 	reg.qword[0] = *txd;
362 	efx_writeo_page(tx_queue->efx, &reg,
363 			FR_BZ_TX_DESC_UPD_P0, tx_queue->queue);
364 }
365 
366 static inline bool
367 efx_may_push_tx_desc(struct efx_tx_queue *tx_queue, unsigned int write_count)
368 {
369 	unsigned empty_read_count = ACCESS_ONCE(tx_queue->empty_read_count);
370 
371 	if (empty_read_count == 0)
372 		return false;
373 
374 	tx_queue->empty_read_count = 0;
375 	return ((empty_read_count ^ write_count) & ~EFX_EMPTY_COUNT_VALID) == 0;
376 }
377 
378 /* For each entry inserted into the software descriptor ring, create a
379  * descriptor in the hardware TX descriptor ring (in host memory), and
380  * write a doorbell.
381  */
382 void efx_nic_push_buffers(struct efx_tx_queue *tx_queue)
383 {
384 
385 	struct efx_tx_buffer *buffer;
386 	efx_qword_t *txd;
387 	unsigned write_ptr;
388 	unsigned old_write_count = tx_queue->write_count;
389 
390 	BUG_ON(tx_queue->write_count == tx_queue->insert_count);
391 
392 	do {
393 		write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
394 		buffer = &tx_queue->buffer[write_ptr];
395 		txd = efx_tx_desc(tx_queue, write_ptr);
396 		++tx_queue->write_count;
397 
398 		/* Create TX descriptor ring entry */
399 		EFX_POPULATE_QWORD_4(*txd,
400 				     FSF_AZ_TX_KER_CONT, buffer->continuation,
401 				     FSF_AZ_TX_KER_BYTE_COUNT, buffer->len,
402 				     FSF_AZ_TX_KER_BUF_REGION, 0,
403 				     FSF_AZ_TX_KER_BUF_ADDR, buffer->dma_addr);
404 	} while (tx_queue->write_count != tx_queue->insert_count);
405 
406 	wmb(); /* Ensure descriptors are written before they are fetched */
407 
408 	if (efx_may_push_tx_desc(tx_queue, old_write_count)) {
409 		txd = efx_tx_desc(tx_queue,
410 				  old_write_count & tx_queue->ptr_mask);
411 		efx_push_tx_desc(tx_queue, txd);
412 		++tx_queue->pushes;
413 	} else {
414 		efx_notify_tx_desc(tx_queue);
415 	}
416 }
417 
418 /* Allocate hardware resources for a TX queue */
419 int efx_nic_probe_tx(struct efx_tx_queue *tx_queue)
420 {
421 	struct efx_nic *efx = tx_queue->efx;
422 	unsigned entries;
423 
424 	entries = tx_queue->ptr_mask + 1;
425 	return efx_alloc_special_buffer(efx, &tx_queue->txd,
426 					entries * sizeof(efx_qword_t));
427 }
428 
429 void efx_nic_init_tx(struct efx_tx_queue *tx_queue)
430 {
431 	struct efx_nic *efx = tx_queue->efx;
432 	efx_oword_t reg;
433 
434 	tx_queue->flushed = FLUSH_NONE;
435 
436 	/* Pin TX descriptor ring */
437 	efx_init_special_buffer(efx, &tx_queue->txd);
438 
439 	/* Push TX descriptor ring to card */
440 	EFX_POPULATE_OWORD_10(reg,
441 			      FRF_AZ_TX_DESCQ_EN, 1,
442 			      FRF_AZ_TX_ISCSI_DDIG_EN, 0,
443 			      FRF_AZ_TX_ISCSI_HDIG_EN, 0,
444 			      FRF_AZ_TX_DESCQ_BUF_BASE_ID, tx_queue->txd.index,
445 			      FRF_AZ_TX_DESCQ_EVQ_ID,
446 			      tx_queue->channel->channel,
447 			      FRF_AZ_TX_DESCQ_OWNER_ID, 0,
448 			      FRF_AZ_TX_DESCQ_LABEL, tx_queue->queue,
449 			      FRF_AZ_TX_DESCQ_SIZE,
450 			      __ffs(tx_queue->txd.entries),
451 			      FRF_AZ_TX_DESCQ_TYPE, 0,
452 			      FRF_BZ_TX_NON_IP_DROP_DIS, 1);
453 
454 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
455 		int csum = tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD;
456 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_IP_CHKSM_DIS, !csum);
457 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_TCP_CHKSM_DIS,
458 				    !csum);
459 	}
460 
461 	efx_writeo_table(efx, &reg, efx->type->txd_ptr_tbl_base,
462 			 tx_queue->queue);
463 
464 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0) {
465 		/* Only 128 bits in this register */
466 		BUILD_BUG_ON(EFX_MAX_TX_QUEUES > 128);
467 
468 		efx_reado(efx, &reg, FR_AA_TX_CHKSM_CFG);
469 		if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD)
470 			clear_bit_le(tx_queue->queue, (void *)&reg);
471 		else
472 			set_bit_le(tx_queue->queue, (void *)&reg);
473 		efx_writeo(efx, &reg, FR_AA_TX_CHKSM_CFG);
474 	}
475 
476 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
477 		EFX_POPULATE_OWORD_1(reg,
478 				     FRF_BZ_TX_PACE,
479 				     (tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
480 				     FFE_BZ_TX_PACE_OFF :
481 				     FFE_BZ_TX_PACE_RESERVED);
482 		efx_writeo_table(efx, &reg, FR_BZ_TX_PACE_TBL,
483 				 tx_queue->queue);
484 	}
485 }
486 
487 static void efx_flush_tx_queue(struct efx_tx_queue *tx_queue)
488 {
489 	struct efx_nic *efx = tx_queue->efx;
490 	efx_oword_t tx_flush_descq;
491 
492 	tx_queue->flushed = FLUSH_PENDING;
493 
494 	/* Post a flush command */
495 	EFX_POPULATE_OWORD_2(tx_flush_descq,
496 			     FRF_AZ_TX_FLUSH_DESCQ_CMD, 1,
497 			     FRF_AZ_TX_FLUSH_DESCQ, tx_queue->queue);
498 	efx_writeo(efx, &tx_flush_descq, FR_AZ_TX_FLUSH_DESCQ);
499 }
500 
501 void efx_nic_fini_tx(struct efx_tx_queue *tx_queue)
502 {
503 	struct efx_nic *efx = tx_queue->efx;
504 	efx_oword_t tx_desc_ptr;
505 
506 	/* The queue should have been flushed */
507 	WARN_ON(tx_queue->flushed != FLUSH_DONE);
508 
509 	/* Remove TX descriptor ring from card */
510 	EFX_ZERO_OWORD(tx_desc_ptr);
511 	efx_writeo_table(efx, &tx_desc_ptr, efx->type->txd_ptr_tbl_base,
512 			 tx_queue->queue);
513 
514 	/* Unpin TX descriptor ring */
515 	efx_fini_special_buffer(efx, &tx_queue->txd);
516 }
517 
518 /* Free buffers backing TX queue */
519 void efx_nic_remove_tx(struct efx_tx_queue *tx_queue)
520 {
521 	efx_free_special_buffer(tx_queue->efx, &tx_queue->txd);
522 }
523 
524 /**************************************************************************
525  *
526  * RX path
527  *
528  **************************************************************************/
529 
530 /* Returns a pointer to the specified descriptor in the RX descriptor queue */
531 static inline efx_qword_t *
532 efx_rx_desc(struct efx_rx_queue *rx_queue, unsigned int index)
533 {
534 	return ((efx_qword_t *) (rx_queue->rxd.addr)) + index;
535 }
536 
537 /* This creates an entry in the RX descriptor queue */
538 static inline void
539 efx_build_rx_desc(struct efx_rx_queue *rx_queue, unsigned index)
540 {
541 	struct efx_rx_buffer *rx_buf;
542 	efx_qword_t *rxd;
543 
544 	rxd = efx_rx_desc(rx_queue, index);
545 	rx_buf = efx_rx_buffer(rx_queue, index);
546 	EFX_POPULATE_QWORD_3(*rxd,
547 			     FSF_AZ_RX_KER_BUF_SIZE,
548 			     rx_buf->len -
549 			     rx_queue->efx->type->rx_buffer_padding,
550 			     FSF_AZ_RX_KER_BUF_REGION, 0,
551 			     FSF_AZ_RX_KER_BUF_ADDR, rx_buf->dma_addr);
552 }
553 
554 /* This writes to the RX_DESC_WPTR register for the specified receive
555  * descriptor ring.
556  */
557 void efx_nic_notify_rx_desc(struct efx_rx_queue *rx_queue)
558 {
559 	struct efx_nic *efx = rx_queue->efx;
560 	efx_dword_t reg;
561 	unsigned write_ptr;
562 
563 	while (rx_queue->notified_count != rx_queue->added_count) {
564 		efx_build_rx_desc(
565 			rx_queue,
566 			rx_queue->notified_count & rx_queue->ptr_mask);
567 		++rx_queue->notified_count;
568 	}
569 
570 	wmb();
571 	write_ptr = rx_queue->added_count & rx_queue->ptr_mask;
572 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_RX_DESC_WPTR_DWORD, write_ptr);
573 	efx_writed_page(efx, &reg, FR_AZ_RX_DESC_UPD_DWORD_P0,
574 			efx_rx_queue_index(rx_queue));
575 }
576 
577 int efx_nic_probe_rx(struct efx_rx_queue *rx_queue)
578 {
579 	struct efx_nic *efx = rx_queue->efx;
580 	unsigned entries;
581 
582 	entries = rx_queue->ptr_mask + 1;
583 	return efx_alloc_special_buffer(efx, &rx_queue->rxd,
584 					entries * sizeof(efx_qword_t));
585 }
586 
587 void efx_nic_init_rx(struct efx_rx_queue *rx_queue)
588 {
589 	efx_oword_t rx_desc_ptr;
590 	struct efx_nic *efx = rx_queue->efx;
591 	bool is_b0 = efx_nic_rev(efx) >= EFX_REV_FALCON_B0;
592 	bool iscsi_digest_en = is_b0;
593 
594 	netif_dbg(efx, hw, efx->net_dev,
595 		  "RX queue %d ring in special buffers %d-%d\n",
596 		  efx_rx_queue_index(rx_queue), rx_queue->rxd.index,
597 		  rx_queue->rxd.index + rx_queue->rxd.entries - 1);
598 
599 	rx_queue->flushed = FLUSH_NONE;
600 
601 	/* Pin RX descriptor ring */
602 	efx_init_special_buffer(efx, &rx_queue->rxd);
603 
604 	/* Push RX descriptor ring to card */
605 	EFX_POPULATE_OWORD_10(rx_desc_ptr,
606 			      FRF_AZ_RX_ISCSI_DDIG_EN, iscsi_digest_en,
607 			      FRF_AZ_RX_ISCSI_HDIG_EN, iscsi_digest_en,
608 			      FRF_AZ_RX_DESCQ_BUF_BASE_ID, rx_queue->rxd.index,
609 			      FRF_AZ_RX_DESCQ_EVQ_ID,
610 			      efx_rx_queue_channel(rx_queue)->channel,
611 			      FRF_AZ_RX_DESCQ_OWNER_ID, 0,
612 			      FRF_AZ_RX_DESCQ_LABEL,
613 			      efx_rx_queue_index(rx_queue),
614 			      FRF_AZ_RX_DESCQ_SIZE,
615 			      __ffs(rx_queue->rxd.entries),
616 			      FRF_AZ_RX_DESCQ_TYPE, 0 /* kernel queue */ ,
617 			      /* For >=B0 this is scatter so disable */
618 			      FRF_AZ_RX_DESCQ_JUMBO, !is_b0,
619 			      FRF_AZ_RX_DESCQ_EN, 1);
620 	efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
621 			 efx_rx_queue_index(rx_queue));
622 }
623 
624 static void efx_flush_rx_queue(struct efx_rx_queue *rx_queue)
625 {
626 	struct efx_nic *efx = rx_queue->efx;
627 	efx_oword_t rx_flush_descq;
628 
629 	rx_queue->flushed = FLUSH_PENDING;
630 
631 	/* Post a flush command */
632 	EFX_POPULATE_OWORD_2(rx_flush_descq,
633 			     FRF_AZ_RX_FLUSH_DESCQ_CMD, 1,
634 			     FRF_AZ_RX_FLUSH_DESCQ,
635 			     efx_rx_queue_index(rx_queue));
636 	efx_writeo(efx, &rx_flush_descq, FR_AZ_RX_FLUSH_DESCQ);
637 }
638 
639 void efx_nic_fini_rx(struct efx_rx_queue *rx_queue)
640 {
641 	efx_oword_t rx_desc_ptr;
642 	struct efx_nic *efx = rx_queue->efx;
643 
644 	/* The queue should already have been flushed */
645 	WARN_ON(rx_queue->flushed != FLUSH_DONE);
646 
647 	/* Remove RX descriptor ring from card */
648 	EFX_ZERO_OWORD(rx_desc_ptr);
649 	efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
650 			 efx_rx_queue_index(rx_queue));
651 
652 	/* Unpin RX descriptor ring */
653 	efx_fini_special_buffer(efx, &rx_queue->rxd);
654 }
655 
656 /* Free buffers backing RX queue */
657 void efx_nic_remove_rx(struct efx_rx_queue *rx_queue)
658 {
659 	efx_free_special_buffer(rx_queue->efx, &rx_queue->rxd);
660 }
661 
662 /**************************************************************************
663  *
664  * Event queue processing
665  * Event queues are processed by per-channel tasklets.
666  *
667  **************************************************************************/
668 
669 /* Update a channel's event queue's read pointer (RPTR) register
670  *
671  * This writes the EVQ_RPTR_REG register for the specified channel's
672  * event queue.
673  */
674 void efx_nic_eventq_read_ack(struct efx_channel *channel)
675 {
676 	efx_dword_t reg;
677 	struct efx_nic *efx = channel->efx;
678 
679 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_EVQ_RPTR,
680 			     channel->eventq_read_ptr & channel->eventq_mask);
681 	efx_writed_table(efx, &reg, efx->type->evq_rptr_tbl_base,
682 			 channel->channel);
683 }
684 
685 /* Use HW to insert a SW defined event */
686 static void efx_generate_event(struct efx_channel *channel, efx_qword_t *event)
687 {
688 	efx_oword_t drv_ev_reg;
689 
690 	BUILD_BUG_ON(FRF_AZ_DRV_EV_DATA_LBN != 0 ||
691 		     FRF_AZ_DRV_EV_DATA_WIDTH != 64);
692 	drv_ev_reg.u32[0] = event->u32[0];
693 	drv_ev_reg.u32[1] = event->u32[1];
694 	drv_ev_reg.u32[2] = 0;
695 	drv_ev_reg.u32[3] = 0;
696 	EFX_SET_OWORD_FIELD(drv_ev_reg, FRF_AZ_DRV_EV_QID, channel->channel);
697 	efx_writeo(channel->efx, &drv_ev_reg, FR_AZ_DRV_EV);
698 }
699 
700 static void efx_magic_event(struct efx_channel *channel, u32 magic)
701 {
702 	efx_qword_t event;
703 
704 	EFX_POPULATE_QWORD_2(event, FSF_AZ_EV_CODE,
705 			     FSE_AZ_EV_CODE_DRV_GEN_EV,
706 			     FSF_AZ_DRV_GEN_EV_MAGIC, magic);
707 	efx_generate_event(channel, &event);
708 }
709 
710 /* Handle a transmit completion event
711  *
712  * The NIC batches TX completion events; the message we receive is of
713  * the form "complete all TX events up to this index".
714  */
715 static int
716 efx_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
717 {
718 	unsigned int tx_ev_desc_ptr;
719 	unsigned int tx_ev_q_label;
720 	struct efx_tx_queue *tx_queue;
721 	struct efx_nic *efx = channel->efx;
722 	int tx_packets = 0;
723 
724 	if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
725 		/* Transmit completion */
726 		tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
727 		tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
728 		tx_queue = efx_channel_get_tx_queue(
729 			channel, tx_ev_q_label % EFX_TXQ_TYPES);
730 		tx_packets = ((tx_ev_desc_ptr - tx_queue->read_count) &
731 			      tx_queue->ptr_mask);
732 		channel->irq_mod_score += tx_packets;
733 		efx_xmit_done(tx_queue, tx_ev_desc_ptr);
734 	} else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
735 		/* Rewrite the FIFO write pointer */
736 		tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
737 		tx_queue = efx_channel_get_tx_queue(
738 			channel, tx_ev_q_label % EFX_TXQ_TYPES);
739 
740 		netif_tx_lock(efx->net_dev);
741 		efx_notify_tx_desc(tx_queue);
742 		netif_tx_unlock(efx->net_dev);
743 	} else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR) &&
744 		   EFX_WORKAROUND_10727(efx)) {
745 		efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
746 	} else {
747 		netif_err(efx, tx_err, efx->net_dev,
748 			  "channel %d unexpected TX event "
749 			  EFX_QWORD_FMT"\n", channel->channel,
750 			  EFX_QWORD_VAL(*event));
751 	}
752 
753 	return tx_packets;
754 }
755 
756 /* Detect errors included in the rx_evt_pkt_ok bit. */
757 static u16 efx_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
758 				const efx_qword_t *event)
759 {
760 	struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
761 	struct efx_nic *efx = rx_queue->efx;
762 	bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
763 	bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
764 	bool rx_ev_frm_trunc, rx_ev_drib_nib, rx_ev_tobe_disc;
765 	bool rx_ev_other_err, rx_ev_pause_frm;
766 	bool rx_ev_hdr_type, rx_ev_mcast_pkt;
767 	unsigned rx_ev_pkt_type;
768 
769 	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
770 	rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
771 	rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
772 	rx_ev_pkt_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_TYPE);
773 	rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
774 						 FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
775 	rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
776 						  FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
777 	rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
778 						   FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
779 	rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
780 	rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
781 	rx_ev_drib_nib = ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) ?
782 			  0 : EFX_QWORD_FIELD(*event, FSF_AA_RX_EV_DRIB_NIB));
783 	rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
784 
785 	/* Every error apart from tobe_disc and pause_frm */
786 	rx_ev_other_err = (rx_ev_drib_nib | rx_ev_tcp_udp_chksum_err |
787 			   rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
788 			   rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
789 
790 	/* Count errors that are not in MAC stats.  Ignore expected
791 	 * checksum errors during self-test. */
792 	if (rx_ev_frm_trunc)
793 		++channel->n_rx_frm_trunc;
794 	else if (rx_ev_tobe_disc)
795 		++channel->n_rx_tobe_disc;
796 	else if (!efx->loopback_selftest) {
797 		if (rx_ev_ip_hdr_chksum_err)
798 			++channel->n_rx_ip_hdr_chksum_err;
799 		else if (rx_ev_tcp_udp_chksum_err)
800 			++channel->n_rx_tcp_udp_chksum_err;
801 	}
802 
803 	/* TOBE_DISC is expected on unicast mismatches; don't print out an
804 	 * error message.  FRM_TRUNC indicates RXDP dropped the packet due
805 	 * to a FIFO overflow.
806 	 */
807 #ifdef DEBUG
808 	if (rx_ev_other_err && net_ratelimit()) {
809 		netif_dbg(efx, rx_err, efx->net_dev,
810 			  " RX queue %d unexpected RX event "
811 			  EFX_QWORD_FMT "%s%s%s%s%s%s%s%s\n",
812 			  efx_rx_queue_index(rx_queue), EFX_QWORD_VAL(*event),
813 			  rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
814 			  rx_ev_ip_hdr_chksum_err ?
815 			  " [IP_HDR_CHKSUM_ERR]" : "",
816 			  rx_ev_tcp_udp_chksum_err ?
817 			  " [TCP_UDP_CHKSUM_ERR]" : "",
818 			  rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
819 			  rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
820 			  rx_ev_drib_nib ? " [DRIB_NIB]" : "",
821 			  rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
822 			  rx_ev_pause_frm ? " [PAUSE]" : "");
823 	}
824 #endif
825 
826 	/* The frame must be discarded if any of these are true. */
827 	return (rx_ev_eth_crc_err | rx_ev_frm_trunc | rx_ev_drib_nib |
828 		rx_ev_tobe_disc | rx_ev_pause_frm) ?
829 		EFX_RX_PKT_DISCARD : 0;
830 }
831 
832 /* Handle receive events that are not in-order. */
833 static void
834 efx_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
835 {
836 	struct efx_nic *efx = rx_queue->efx;
837 	unsigned expected, dropped;
838 
839 	expected = rx_queue->removed_count & rx_queue->ptr_mask;
840 	dropped = (index - expected) & rx_queue->ptr_mask;
841 	netif_info(efx, rx_err, efx->net_dev,
842 		   "dropped %d events (index=%d expected=%d)\n",
843 		   dropped, index, expected);
844 
845 	efx_schedule_reset(efx, EFX_WORKAROUND_5676(efx) ?
846 			   RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
847 }
848 
849 /* Handle a packet received event
850  *
851  * The NIC gives a "discard" flag if it's a unicast packet with the
852  * wrong destination address
853  * Also "is multicast" and "matches multicast filter" flags can be used to
854  * discard non-matching multicast packets.
855  */
856 static void
857 efx_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
858 {
859 	unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
860 	unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
861 	unsigned expected_ptr;
862 	bool rx_ev_pkt_ok;
863 	u16 flags;
864 	struct efx_rx_queue *rx_queue;
865 
866 	/* Basic packet information */
867 	rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
868 	rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
869 	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
870 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT));
871 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP) != 1);
872 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
873 		channel->channel);
874 
875 	rx_queue = efx_channel_get_rx_queue(channel);
876 
877 	rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
878 	expected_ptr = rx_queue->removed_count & rx_queue->ptr_mask;
879 	if (unlikely(rx_ev_desc_ptr != expected_ptr))
880 		efx_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr);
881 
882 	if (likely(rx_ev_pkt_ok)) {
883 		/* If packet is marked as OK and packet type is TCP/IP or
884 		 * UDP/IP, then we can rely on the hardware checksum.
885 		 */
886 		flags = (rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP ||
887 			 rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP) ?
888 			EFX_RX_PKT_CSUMMED : 0;
889 	} else {
890 		flags = efx_handle_rx_not_ok(rx_queue, event);
891 	}
892 
893 	/* Detect multicast packets that didn't match the filter */
894 	rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
895 	if (rx_ev_mcast_pkt) {
896 		unsigned int rx_ev_mcast_hash_match =
897 			EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
898 
899 		if (unlikely(!rx_ev_mcast_hash_match)) {
900 			++channel->n_rx_mcast_mismatch;
901 			flags |= EFX_RX_PKT_DISCARD;
902 		}
903 	}
904 
905 	channel->irq_mod_score += 2;
906 
907 	/* Handle received packet */
908 	efx_rx_packet(rx_queue, rx_ev_desc_ptr, rx_ev_byte_cnt, flags);
909 }
910 
911 static void
912 efx_handle_generated_event(struct efx_channel *channel, efx_qword_t *event)
913 {
914 	struct efx_nic *efx = channel->efx;
915 	unsigned magic;
916 
917 	magic = EFX_QWORD_FIELD(*event, FSF_AZ_DRV_GEN_EV_MAGIC);
918 
919 	if (magic == EFX_CHANNEL_MAGIC_TEST(channel))
920 		; /* ignore */
921 	else if (magic == EFX_CHANNEL_MAGIC_FILL(channel))
922 		/* The queue must be empty, so we won't receive any rx
923 		 * events, so efx_process_channel() won't refill the
924 		 * queue. Refill it here */
925 		efx_fast_push_rx_descriptors(efx_channel_get_rx_queue(channel));
926 	else
927 		netif_dbg(efx, hw, efx->net_dev, "channel %d received "
928 			  "generated event "EFX_QWORD_FMT"\n",
929 			  channel->channel, EFX_QWORD_VAL(*event));
930 }
931 
932 static void
933 efx_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
934 {
935 	struct efx_nic *efx = channel->efx;
936 	unsigned int ev_sub_code;
937 	unsigned int ev_sub_data;
938 
939 	ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
940 	ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
941 
942 	switch (ev_sub_code) {
943 	case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
944 		netif_vdbg(efx, hw, efx->net_dev, "channel %d TXQ %d flushed\n",
945 			   channel->channel, ev_sub_data);
946 		break;
947 	case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
948 		netif_vdbg(efx, hw, efx->net_dev, "channel %d RXQ %d flushed\n",
949 			   channel->channel, ev_sub_data);
950 		break;
951 	case FSE_AZ_EVQ_INIT_DONE_EV:
952 		netif_dbg(efx, hw, efx->net_dev,
953 			  "channel %d EVQ %d initialised\n",
954 			  channel->channel, ev_sub_data);
955 		break;
956 	case FSE_AZ_SRM_UPD_DONE_EV:
957 		netif_vdbg(efx, hw, efx->net_dev,
958 			   "channel %d SRAM update done\n", channel->channel);
959 		break;
960 	case FSE_AZ_WAKE_UP_EV:
961 		netif_vdbg(efx, hw, efx->net_dev,
962 			   "channel %d RXQ %d wakeup event\n",
963 			   channel->channel, ev_sub_data);
964 		break;
965 	case FSE_AZ_TIMER_EV:
966 		netif_vdbg(efx, hw, efx->net_dev,
967 			   "channel %d RX queue %d timer expired\n",
968 			   channel->channel, ev_sub_data);
969 		break;
970 	case FSE_AA_RX_RECOVER_EV:
971 		netif_err(efx, rx_err, efx->net_dev,
972 			  "channel %d seen DRIVER RX_RESET event. "
973 			"Resetting.\n", channel->channel);
974 		atomic_inc(&efx->rx_reset);
975 		efx_schedule_reset(efx,
976 				   EFX_WORKAROUND_6555(efx) ?
977 				   RESET_TYPE_RX_RECOVERY :
978 				   RESET_TYPE_DISABLE);
979 		break;
980 	case FSE_BZ_RX_DSC_ERROR_EV:
981 		netif_err(efx, rx_err, efx->net_dev,
982 			  "RX DMA Q %d reports descriptor fetch error."
983 			  " RX Q %d is disabled.\n", ev_sub_data, ev_sub_data);
984 		efx_schedule_reset(efx, RESET_TYPE_RX_DESC_FETCH);
985 		break;
986 	case FSE_BZ_TX_DSC_ERROR_EV:
987 		netif_err(efx, tx_err, efx->net_dev,
988 			  "TX DMA Q %d reports descriptor fetch error."
989 			  " TX Q %d is disabled.\n", ev_sub_data, ev_sub_data);
990 		efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
991 		break;
992 	default:
993 		netif_vdbg(efx, hw, efx->net_dev,
994 			   "channel %d unknown driver event code %d "
995 			   "data %04x\n", channel->channel, ev_sub_code,
996 			   ev_sub_data);
997 		break;
998 	}
999 }
1000 
1001 int efx_nic_process_eventq(struct efx_channel *channel, int budget)
1002 {
1003 	struct efx_nic *efx = channel->efx;
1004 	unsigned int read_ptr;
1005 	efx_qword_t event, *p_event;
1006 	int ev_code;
1007 	int tx_packets = 0;
1008 	int spent = 0;
1009 
1010 	read_ptr = channel->eventq_read_ptr;
1011 
1012 	for (;;) {
1013 		p_event = efx_event(channel, read_ptr);
1014 		event = *p_event;
1015 
1016 		if (!efx_event_present(&event))
1017 			/* End of events */
1018 			break;
1019 
1020 		netif_vdbg(channel->efx, intr, channel->efx->net_dev,
1021 			   "channel %d event is "EFX_QWORD_FMT"\n",
1022 			   channel->channel, EFX_QWORD_VAL(event));
1023 
1024 		/* Clear this event by marking it all ones */
1025 		EFX_SET_QWORD(*p_event);
1026 
1027 		++read_ptr;
1028 
1029 		ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
1030 
1031 		switch (ev_code) {
1032 		case FSE_AZ_EV_CODE_RX_EV:
1033 			efx_handle_rx_event(channel, &event);
1034 			if (++spent == budget)
1035 				goto out;
1036 			break;
1037 		case FSE_AZ_EV_CODE_TX_EV:
1038 			tx_packets += efx_handle_tx_event(channel, &event);
1039 			if (tx_packets > efx->txq_entries) {
1040 				spent = budget;
1041 				goto out;
1042 			}
1043 			break;
1044 		case FSE_AZ_EV_CODE_DRV_GEN_EV:
1045 			efx_handle_generated_event(channel, &event);
1046 			break;
1047 		case FSE_AZ_EV_CODE_DRIVER_EV:
1048 			efx_handle_driver_event(channel, &event);
1049 			break;
1050 		case FSE_CZ_EV_CODE_MCDI_EV:
1051 			efx_mcdi_process_event(channel, &event);
1052 			break;
1053 		case FSE_AZ_EV_CODE_GLOBAL_EV:
1054 			if (efx->type->handle_global_event &&
1055 			    efx->type->handle_global_event(channel, &event))
1056 				break;
1057 			/* else fall through */
1058 		default:
1059 			netif_err(channel->efx, hw, channel->efx->net_dev,
1060 				  "channel %d unknown event type %d (data "
1061 				  EFX_QWORD_FMT ")\n", channel->channel,
1062 				  ev_code, EFX_QWORD_VAL(event));
1063 		}
1064 	}
1065 
1066 out:
1067 	channel->eventq_read_ptr = read_ptr;
1068 	return spent;
1069 }
1070 
1071 /* Check whether an event is present in the eventq at the current
1072  * read pointer.  Only useful for self-test.
1073  */
1074 bool efx_nic_event_present(struct efx_channel *channel)
1075 {
1076 	return efx_event_present(efx_event(channel, channel->eventq_read_ptr));
1077 }
1078 
1079 /* Allocate buffer table entries for event queue */
1080 int efx_nic_probe_eventq(struct efx_channel *channel)
1081 {
1082 	struct efx_nic *efx = channel->efx;
1083 	unsigned entries;
1084 
1085 	entries = channel->eventq_mask + 1;
1086 	return efx_alloc_special_buffer(efx, &channel->eventq,
1087 					entries * sizeof(efx_qword_t));
1088 }
1089 
1090 void efx_nic_init_eventq(struct efx_channel *channel)
1091 {
1092 	efx_oword_t reg;
1093 	struct efx_nic *efx = channel->efx;
1094 
1095 	netif_dbg(efx, hw, efx->net_dev,
1096 		  "channel %d event queue in special buffers %d-%d\n",
1097 		  channel->channel, channel->eventq.index,
1098 		  channel->eventq.index + channel->eventq.entries - 1);
1099 
1100 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1101 		EFX_POPULATE_OWORD_3(reg,
1102 				     FRF_CZ_TIMER_Q_EN, 1,
1103 				     FRF_CZ_HOST_NOTIFY_MODE, 0,
1104 				     FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
1105 		efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1106 	}
1107 
1108 	/* Pin event queue buffer */
1109 	efx_init_special_buffer(efx, &channel->eventq);
1110 
1111 	/* Fill event queue with all ones (i.e. empty events) */
1112 	memset(channel->eventq.addr, 0xff, channel->eventq.len);
1113 
1114 	/* Push event queue to card */
1115 	EFX_POPULATE_OWORD_3(reg,
1116 			     FRF_AZ_EVQ_EN, 1,
1117 			     FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
1118 			     FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
1119 	efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1120 			 channel->channel);
1121 
1122 	efx->type->push_irq_moderation(channel);
1123 }
1124 
1125 void efx_nic_fini_eventq(struct efx_channel *channel)
1126 {
1127 	efx_oword_t reg;
1128 	struct efx_nic *efx = channel->efx;
1129 
1130 	/* Remove event queue from card */
1131 	EFX_ZERO_OWORD(reg);
1132 	efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1133 			 channel->channel);
1134 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1135 		efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1136 
1137 	/* Unpin event queue */
1138 	efx_fini_special_buffer(efx, &channel->eventq);
1139 }
1140 
1141 /* Free buffers backing event queue */
1142 void efx_nic_remove_eventq(struct efx_channel *channel)
1143 {
1144 	efx_free_special_buffer(channel->efx, &channel->eventq);
1145 }
1146 
1147 
1148 void efx_nic_generate_test_event(struct efx_channel *channel)
1149 {
1150 	efx_magic_event(channel, EFX_CHANNEL_MAGIC_TEST(channel));
1151 }
1152 
1153 void efx_nic_generate_fill_event(struct efx_channel *channel)
1154 {
1155 	efx_magic_event(channel, EFX_CHANNEL_MAGIC_FILL(channel));
1156 }
1157 
1158 /**************************************************************************
1159  *
1160  * Flush handling
1161  *
1162  **************************************************************************/
1163 
1164 
1165 static void efx_poll_flush_events(struct efx_nic *efx)
1166 {
1167 	struct efx_channel *channel = efx_get_channel(efx, 0);
1168 	struct efx_tx_queue *tx_queue;
1169 	struct efx_rx_queue *rx_queue;
1170 	unsigned int read_ptr = channel->eventq_read_ptr;
1171 	unsigned int end_ptr = read_ptr + channel->eventq_mask - 1;
1172 
1173 	do {
1174 		efx_qword_t *event = efx_event(channel, read_ptr);
1175 		int ev_code, ev_sub_code, ev_queue;
1176 		bool ev_failed;
1177 
1178 		if (!efx_event_present(event))
1179 			break;
1180 
1181 		ev_code = EFX_QWORD_FIELD(*event, FSF_AZ_EV_CODE);
1182 		ev_sub_code = EFX_QWORD_FIELD(*event,
1183 					      FSF_AZ_DRIVER_EV_SUBCODE);
1184 		if (ev_code == FSE_AZ_EV_CODE_DRIVER_EV &&
1185 		    ev_sub_code == FSE_AZ_TX_DESCQ_FLS_DONE_EV) {
1186 			ev_queue = EFX_QWORD_FIELD(*event,
1187 						   FSF_AZ_DRIVER_EV_SUBDATA);
1188 			if (ev_queue < EFX_TXQ_TYPES * efx->n_tx_channels) {
1189 				tx_queue = efx_get_tx_queue(
1190 					efx, ev_queue / EFX_TXQ_TYPES,
1191 					ev_queue % EFX_TXQ_TYPES);
1192 				tx_queue->flushed = FLUSH_DONE;
1193 			}
1194 		} else if (ev_code == FSE_AZ_EV_CODE_DRIVER_EV &&
1195 			   ev_sub_code == FSE_AZ_RX_DESCQ_FLS_DONE_EV) {
1196 			ev_queue = EFX_QWORD_FIELD(
1197 				*event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
1198 			ev_failed = EFX_QWORD_FIELD(
1199 				*event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
1200 			if (ev_queue < efx->n_rx_channels) {
1201 				rx_queue = efx_get_rx_queue(efx, ev_queue);
1202 				rx_queue->flushed =
1203 					ev_failed ? FLUSH_FAILED : FLUSH_DONE;
1204 			}
1205 		}
1206 
1207 		/* We're about to destroy the queue anyway, so
1208 		 * it's ok to throw away every non-flush event */
1209 		EFX_SET_QWORD(*event);
1210 
1211 		++read_ptr;
1212 	} while (read_ptr != end_ptr);
1213 
1214 	channel->eventq_read_ptr = read_ptr;
1215 }
1216 
1217 /* Handle tx and rx flushes at the same time, since they run in
1218  * parallel in the hardware and there's no reason for us to
1219  * serialise them */
1220 int efx_nic_flush_queues(struct efx_nic *efx)
1221 {
1222 	struct efx_channel *channel;
1223 	struct efx_rx_queue *rx_queue;
1224 	struct efx_tx_queue *tx_queue;
1225 	int i, tx_pending, rx_pending;
1226 
1227 	/* If necessary prepare the hardware for flushing */
1228 	efx->type->prepare_flush(efx);
1229 
1230 	/* Flush all tx queues in parallel */
1231 	efx_for_each_channel(channel, efx) {
1232 		efx_for_each_possible_channel_tx_queue(tx_queue, channel) {
1233 			if (tx_queue->initialised)
1234 				efx_flush_tx_queue(tx_queue);
1235 		}
1236 	}
1237 
1238 	/* The hardware supports four concurrent rx flushes, each of which may
1239 	 * need to be retried if there is an outstanding descriptor fetch */
1240 	for (i = 0; i < EFX_FLUSH_POLL_COUNT; ++i) {
1241 		rx_pending = tx_pending = 0;
1242 		efx_for_each_channel(channel, efx) {
1243 			efx_for_each_channel_rx_queue(rx_queue, channel) {
1244 				if (rx_queue->flushed == FLUSH_PENDING)
1245 					++rx_pending;
1246 			}
1247 		}
1248 		efx_for_each_channel(channel, efx) {
1249 			efx_for_each_channel_rx_queue(rx_queue, channel) {
1250 				if (rx_pending == EFX_RX_FLUSH_COUNT)
1251 					break;
1252 				if (rx_queue->flushed == FLUSH_FAILED ||
1253 				    rx_queue->flushed == FLUSH_NONE) {
1254 					efx_flush_rx_queue(rx_queue);
1255 					++rx_pending;
1256 				}
1257 			}
1258 			efx_for_each_possible_channel_tx_queue(tx_queue, channel) {
1259 				if (tx_queue->initialised &&
1260 				    tx_queue->flushed != FLUSH_DONE)
1261 					++tx_pending;
1262 			}
1263 		}
1264 
1265 		if (rx_pending == 0 && tx_pending == 0)
1266 			return 0;
1267 
1268 		msleep(EFX_FLUSH_INTERVAL);
1269 		efx_poll_flush_events(efx);
1270 	}
1271 
1272 	/* Mark the queues as all flushed. We're going to return failure
1273 	 * leading to a reset, or fake up success anyway */
1274 	efx_for_each_channel(channel, efx) {
1275 		efx_for_each_possible_channel_tx_queue(tx_queue, channel) {
1276 			if (tx_queue->initialised &&
1277 			    tx_queue->flushed != FLUSH_DONE)
1278 				netif_err(efx, hw, efx->net_dev,
1279 					  "tx queue %d flush command timed out\n",
1280 					  tx_queue->queue);
1281 			tx_queue->flushed = FLUSH_DONE;
1282 		}
1283 		efx_for_each_channel_rx_queue(rx_queue, channel) {
1284 			if (rx_queue->flushed != FLUSH_DONE)
1285 				netif_err(efx, hw, efx->net_dev,
1286 					  "rx queue %d flush command timed out\n",
1287 					  efx_rx_queue_index(rx_queue));
1288 			rx_queue->flushed = FLUSH_DONE;
1289 		}
1290 	}
1291 
1292 	return -ETIMEDOUT;
1293 }
1294 
1295 /**************************************************************************
1296  *
1297  * Hardware interrupts
1298  * The hardware interrupt handler does very little work; all the event
1299  * queue processing is carried out by per-channel tasklets.
1300  *
1301  **************************************************************************/
1302 
1303 /* Enable/disable/generate interrupts */
1304 static inline void efx_nic_interrupts(struct efx_nic *efx,
1305 				      bool enabled, bool force)
1306 {
1307 	efx_oword_t int_en_reg_ker;
1308 
1309 	EFX_POPULATE_OWORD_3(int_en_reg_ker,
1310 			     FRF_AZ_KER_INT_LEVE_SEL, efx->irq_level,
1311 			     FRF_AZ_KER_INT_KER, force,
1312 			     FRF_AZ_DRV_INT_EN_KER, enabled);
1313 	efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
1314 }
1315 
1316 void efx_nic_enable_interrupts(struct efx_nic *efx)
1317 {
1318 	struct efx_channel *channel;
1319 
1320 	EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
1321 	wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
1322 
1323 	/* Enable interrupts */
1324 	efx_nic_interrupts(efx, true, false);
1325 
1326 	/* Force processing of all the channels to get the EVQ RPTRs up to
1327 	   date */
1328 	efx_for_each_channel(channel, efx)
1329 		efx_schedule_channel(channel);
1330 }
1331 
1332 void efx_nic_disable_interrupts(struct efx_nic *efx)
1333 {
1334 	/* Disable interrupts */
1335 	efx_nic_interrupts(efx, false, false);
1336 }
1337 
1338 /* Generate a test interrupt
1339  * Interrupt must already have been enabled, otherwise nasty things
1340  * may happen.
1341  */
1342 void efx_nic_generate_interrupt(struct efx_nic *efx)
1343 {
1344 	efx_nic_interrupts(efx, true, true);
1345 }
1346 
1347 /* Process a fatal interrupt
1348  * Disable bus mastering ASAP and schedule a reset
1349  */
1350 irqreturn_t efx_nic_fatal_interrupt(struct efx_nic *efx)
1351 {
1352 	struct falcon_nic_data *nic_data = efx->nic_data;
1353 	efx_oword_t *int_ker = efx->irq_status.addr;
1354 	efx_oword_t fatal_intr;
1355 	int error, mem_perr;
1356 
1357 	efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
1358 	error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
1359 
1360 	netif_err(efx, hw, efx->net_dev, "SYSTEM ERROR "EFX_OWORD_FMT" status "
1361 		  EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
1362 		  EFX_OWORD_VAL(fatal_intr),
1363 		  error ? "disabling bus mastering" : "no recognised error");
1364 
1365 	/* If this is a memory parity error dump which blocks are offending */
1366 	mem_perr = (EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER) ||
1367 		    EFX_OWORD_FIELD(fatal_intr, FRF_AZ_SRM_PERR_INT_KER));
1368 	if (mem_perr) {
1369 		efx_oword_t reg;
1370 		efx_reado(efx, &reg, FR_AZ_MEM_STAT);
1371 		netif_err(efx, hw, efx->net_dev,
1372 			  "SYSTEM ERROR: memory parity error "EFX_OWORD_FMT"\n",
1373 			  EFX_OWORD_VAL(reg));
1374 	}
1375 
1376 	/* Disable both devices */
1377 	pci_clear_master(efx->pci_dev);
1378 	if (efx_nic_is_dual_func(efx))
1379 		pci_clear_master(nic_data->pci_dev2);
1380 	efx_nic_disable_interrupts(efx);
1381 
1382 	/* Count errors and reset or disable the NIC accordingly */
1383 	if (efx->int_error_count == 0 ||
1384 	    time_after(jiffies, efx->int_error_expire)) {
1385 		efx->int_error_count = 0;
1386 		efx->int_error_expire =
1387 			jiffies + EFX_INT_ERROR_EXPIRE * HZ;
1388 	}
1389 	if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
1390 		netif_err(efx, hw, efx->net_dev,
1391 			  "SYSTEM ERROR - reset scheduled\n");
1392 		efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
1393 	} else {
1394 		netif_err(efx, hw, efx->net_dev,
1395 			  "SYSTEM ERROR - max number of errors seen."
1396 			  "NIC will be disabled\n");
1397 		efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1398 	}
1399 
1400 	return IRQ_HANDLED;
1401 }
1402 
1403 /* Handle a legacy interrupt
1404  * Acknowledges the interrupt and schedule event queue processing.
1405  */
1406 static irqreturn_t efx_legacy_interrupt(int irq, void *dev_id)
1407 {
1408 	struct efx_nic *efx = dev_id;
1409 	efx_oword_t *int_ker = efx->irq_status.addr;
1410 	irqreturn_t result = IRQ_NONE;
1411 	struct efx_channel *channel;
1412 	efx_dword_t reg;
1413 	u32 queues;
1414 	int syserr;
1415 
1416 	/* Could this be ours?  If interrupts are disabled then the
1417 	 * channel state may not be valid.
1418 	 */
1419 	if (!efx->legacy_irq_enabled)
1420 		return result;
1421 
1422 	/* Read the ISR which also ACKs the interrupts */
1423 	efx_readd(efx, &reg, FR_BZ_INT_ISR0);
1424 	queues = EFX_EXTRACT_DWORD(reg, 0, 31);
1425 
1426 	/* Handle non-event-queue sources */
1427 	if (queues & (1U << efx->irq_level)) {
1428 		syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1429 		if (unlikely(syserr))
1430 			return efx_nic_fatal_interrupt(efx);
1431 		efx->last_irq_cpu = raw_smp_processor_id();
1432 	}
1433 
1434 	if (queues != 0) {
1435 		if (EFX_WORKAROUND_15783(efx))
1436 			efx->irq_zero_count = 0;
1437 
1438 		/* Schedule processing of any interrupting queues */
1439 		efx_for_each_channel(channel, efx) {
1440 			if (queues & 1)
1441 				efx_schedule_channel_irq(channel);
1442 			queues >>= 1;
1443 		}
1444 		result = IRQ_HANDLED;
1445 
1446 	} else if (EFX_WORKAROUND_15783(efx)) {
1447 		efx_qword_t *event;
1448 
1449 		/* We can't return IRQ_HANDLED more than once on seeing ISR=0
1450 		 * because this might be a shared interrupt. */
1451 		if (efx->irq_zero_count++ == 0)
1452 			result = IRQ_HANDLED;
1453 
1454 		/* Ensure we schedule or rearm all event queues */
1455 		efx_for_each_channel(channel, efx) {
1456 			event = efx_event(channel, channel->eventq_read_ptr);
1457 			if (efx_event_present(event))
1458 				efx_schedule_channel_irq(channel);
1459 			else
1460 				efx_nic_eventq_read_ack(channel);
1461 		}
1462 	}
1463 
1464 	if (result == IRQ_HANDLED)
1465 		netif_vdbg(efx, intr, efx->net_dev,
1466 			   "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
1467 			   irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
1468 
1469 	return result;
1470 }
1471 
1472 /* Handle an MSI interrupt
1473  *
1474  * Handle an MSI hardware interrupt.  This routine schedules event
1475  * queue processing.  No interrupt acknowledgement cycle is necessary.
1476  * Also, we never need to check that the interrupt is for us, since
1477  * MSI interrupts cannot be shared.
1478  */
1479 static irqreturn_t efx_msi_interrupt(int irq, void *dev_id)
1480 {
1481 	struct efx_channel *channel = *(struct efx_channel **)dev_id;
1482 	struct efx_nic *efx = channel->efx;
1483 	efx_oword_t *int_ker = efx->irq_status.addr;
1484 	int syserr;
1485 
1486 	netif_vdbg(efx, intr, efx->net_dev,
1487 		   "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
1488 		   irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
1489 
1490 	/* Handle non-event-queue sources */
1491 	if (channel->channel == efx->irq_level) {
1492 		syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1493 		if (unlikely(syserr))
1494 			return efx_nic_fatal_interrupt(efx);
1495 		efx->last_irq_cpu = raw_smp_processor_id();
1496 	}
1497 
1498 	/* Schedule processing of the channel */
1499 	efx_schedule_channel_irq(channel);
1500 
1501 	return IRQ_HANDLED;
1502 }
1503 
1504 
1505 /* Setup RSS indirection table.
1506  * This maps from the hash value of the packet to RXQ
1507  */
1508 void efx_nic_push_rx_indir_table(struct efx_nic *efx)
1509 {
1510 	size_t i = 0;
1511 	efx_dword_t dword;
1512 
1513 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0)
1514 		return;
1515 
1516 	BUILD_BUG_ON(ARRAY_SIZE(efx->rx_indir_table) !=
1517 		     FR_BZ_RX_INDIRECTION_TBL_ROWS);
1518 
1519 	for (i = 0; i < FR_BZ_RX_INDIRECTION_TBL_ROWS; i++) {
1520 		EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
1521 				     efx->rx_indir_table[i]);
1522 		efx_writed_table(efx, &dword, FR_BZ_RX_INDIRECTION_TBL, i);
1523 	}
1524 }
1525 
1526 /* Hook interrupt handler(s)
1527  * Try MSI and then legacy interrupts.
1528  */
1529 int efx_nic_init_interrupt(struct efx_nic *efx)
1530 {
1531 	struct efx_channel *channel;
1532 	int rc;
1533 
1534 	if (!EFX_INT_MODE_USE_MSI(efx)) {
1535 		irq_handler_t handler;
1536 		if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1537 			handler = efx_legacy_interrupt;
1538 		else
1539 			handler = falcon_legacy_interrupt_a1;
1540 
1541 		rc = request_irq(efx->legacy_irq, handler, IRQF_SHARED,
1542 				 efx->name, efx);
1543 		if (rc) {
1544 			netif_err(efx, drv, efx->net_dev,
1545 				  "failed to hook legacy IRQ %d\n",
1546 				  efx->pci_dev->irq);
1547 			goto fail1;
1548 		}
1549 		return 0;
1550 	}
1551 
1552 	/* Hook MSI or MSI-X interrupt */
1553 	efx_for_each_channel(channel, efx) {
1554 		rc = request_irq(channel->irq, efx_msi_interrupt,
1555 				 IRQF_PROBE_SHARED, /* Not shared */
1556 				 efx->channel_name[channel->channel],
1557 				 &efx->channel[channel->channel]);
1558 		if (rc) {
1559 			netif_err(efx, drv, efx->net_dev,
1560 				  "failed to hook IRQ %d\n", channel->irq);
1561 			goto fail2;
1562 		}
1563 	}
1564 
1565 	return 0;
1566 
1567  fail2:
1568 	efx_for_each_channel(channel, efx)
1569 		free_irq(channel->irq, &efx->channel[channel->channel]);
1570  fail1:
1571 	return rc;
1572 }
1573 
1574 void efx_nic_fini_interrupt(struct efx_nic *efx)
1575 {
1576 	struct efx_channel *channel;
1577 	efx_oword_t reg;
1578 
1579 	/* Disable MSI/MSI-X interrupts */
1580 	efx_for_each_channel(channel, efx) {
1581 		if (channel->irq)
1582 			free_irq(channel->irq, &efx->channel[channel->channel]);
1583 	}
1584 
1585 	/* ACK legacy interrupt */
1586 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1587 		efx_reado(efx, &reg, FR_BZ_INT_ISR0);
1588 	else
1589 		falcon_irq_ack_a1(efx);
1590 
1591 	/* Disable legacy interrupt */
1592 	if (efx->legacy_irq)
1593 		free_irq(efx->legacy_irq, efx);
1594 }
1595 
1596 u32 efx_nic_fpga_ver(struct efx_nic *efx)
1597 {
1598 	efx_oword_t altera_build;
1599 	efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
1600 	return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
1601 }
1602 
1603 void efx_nic_init_common(struct efx_nic *efx)
1604 {
1605 	efx_oword_t temp;
1606 
1607 	/* Set positions of descriptor caches in SRAM. */
1608 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR,
1609 			     efx->type->tx_dc_base / 8);
1610 	efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
1611 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR,
1612 			     efx->type->rx_dc_base / 8);
1613 	efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
1614 
1615 	/* Set TX descriptor cache size. */
1616 	BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
1617 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
1618 	efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
1619 
1620 	/* Set RX descriptor cache size.  Set low watermark to size-8, as
1621 	 * this allows most efficient prefetching.
1622 	 */
1623 	BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
1624 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
1625 	efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
1626 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
1627 	efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
1628 
1629 	/* Program INT_KER address */
1630 	EFX_POPULATE_OWORD_2(temp,
1631 			     FRF_AZ_NORM_INT_VEC_DIS_KER,
1632 			     EFX_INT_MODE_USE_MSI(efx),
1633 			     FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
1634 	efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
1635 
1636 	if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
1637 		/* Use an interrupt level unused by event queues */
1638 		efx->irq_level = 0x1f;
1639 	else
1640 		/* Use a valid MSI-X vector */
1641 		efx->irq_level = 0;
1642 
1643 	/* Enable all the genuinely fatal interrupts.  (They are still
1644 	 * masked by the overall interrupt mask, controlled by
1645 	 * falcon_interrupts()).
1646 	 *
1647 	 * Note: All other fatal interrupts are enabled
1648 	 */
1649 	EFX_POPULATE_OWORD_3(temp,
1650 			     FRF_AZ_ILL_ADR_INT_KER_EN, 1,
1651 			     FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
1652 			     FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
1653 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1654 		EFX_SET_OWORD_FIELD(temp, FRF_CZ_SRAM_PERR_INT_P_KER_EN, 1);
1655 	EFX_INVERT_OWORD(temp);
1656 	efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
1657 
1658 	efx_nic_push_rx_indir_table(efx);
1659 
1660 	/* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
1661 	 * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
1662 	 */
1663 	efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
1664 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
1665 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
1666 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
1667 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 1);
1668 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
1669 	/* Enable SW_EV to inherit in char driver - assume harmless here */
1670 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
1671 	/* Prefetch threshold 2 => fetch when descriptor cache half empty */
1672 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
1673 	/* Disable hardware watchdog which can misfire */
1674 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
1675 	/* Squash TX of packets of 16 bytes or less */
1676 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1677 		EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
1678 	efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
1679 
1680 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
1681 		EFX_POPULATE_OWORD_4(temp,
1682 				     /* Default values */
1683 				     FRF_BZ_TX_PACE_SB_NOT_AF, 0x15,
1684 				     FRF_BZ_TX_PACE_SB_AF, 0xb,
1685 				     FRF_BZ_TX_PACE_FB_BASE, 0,
1686 				     /* Allow large pace values in the
1687 				      * fast bin. */
1688 				     FRF_BZ_TX_PACE_BIN_TH,
1689 				     FFE_BZ_TX_PACE_RESERVED);
1690 		efx_writeo(efx, &temp, FR_BZ_TX_PACE);
1691 	}
1692 }
1693 
1694 /* Register dump */
1695 
1696 #define REGISTER_REVISION_A	1
1697 #define REGISTER_REVISION_B	2
1698 #define REGISTER_REVISION_C	3
1699 #define REGISTER_REVISION_Z	3	/* latest revision */
1700 
1701 struct efx_nic_reg {
1702 	u32 offset:24;
1703 	u32 min_revision:2, max_revision:2;
1704 };
1705 
1706 #define REGISTER(name, min_rev, max_rev) {				\
1707 	FR_ ## min_rev ## max_rev ## _ ## name,				\
1708 	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev	\
1709 }
1710 #define REGISTER_AA(name) REGISTER(name, A, A)
1711 #define REGISTER_AB(name) REGISTER(name, A, B)
1712 #define REGISTER_AZ(name) REGISTER(name, A, Z)
1713 #define REGISTER_BB(name) REGISTER(name, B, B)
1714 #define REGISTER_BZ(name) REGISTER(name, B, Z)
1715 #define REGISTER_CZ(name) REGISTER(name, C, Z)
1716 
1717 static const struct efx_nic_reg efx_nic_regs[] = {
1718 	REGISTER_AZ(ADR_REGION),
1719 	REGISTER_AZ(INT_EN_KER),
1720 	REGISTER_BZ(INT_EN_CHAR),
1721 	REGISTER_AZ(INT_ADR_KER),
1722 	REGISTER_BZ(INT_ADR_CHAR),
1723 	/* INT_ACK_KER is WO */
1724 	/* INT_ISR0 is RC */
1725 	REGISTER_AZ(HW_INIT),
1726 	REGISTER_CZ(USR_EV_CFG),
1727 	REGISTER_AB(EE_SPI_HCMD),
1728 	REGISTER_AB(EE_SPI_HADR),
1729 	REGISTER_AB(EE_SPI_HDATA),
1730 	REGISTER_AB(EE_BASE_PAGE),
1731 	REGISTER_AB(EE_VPD_CFG0),
1732 	/* EE_VPD_SW_CNTL and EE_VPD_SW_DATA are not used */
1733 	/* PMBX_DBG_IADDR and PBMX_DBG_IDATA are indirect */
1734 	/* PCIE_CORE_INDIRECT is indirect */
1735 	REGISTER_AB(NIC_STAT),
1736 	REGISTER_AB(GPIO_CTL),
1737 	REGISTER_AB(GLB_CTL),
1738 	/* FATAL_INTR_KER and FATAL_INTR_CHAR are partly RC */
1739 	REGISTER_BZ(DP_CTRL),
1740 	REGISTER_AZ(MEM_STAT),
1741 	REGISTER_AZ(CS_DEBUG),
1742 	REGISTER_AZ(ALTERA_BUILD),
1743 	REGISTER_AZ(CSR_SPARE),
1744 	REGISTER_AB(PCIE_SD_CTL0123),
1745 	REGISTER_AB(PCIE_SD_CTL45),
1746 	REGISTER_AB(PCIE_PCS_CTL_STAT),
1747 	/* DEBUG_DATA_OUT is not used */
1748 	/* DRV_EV is WO */
1749 	REGISTER_AZ(EVQ_CTL),
1750 	REGISTER_AZ(EVQ_CNT1),
1751 	REGISTER_AZ(EVQ_CNT2),
1752 	REGISTER_AZ(BUF_TBL_CFG),
1753 	REGISTER_AZ(SRM_RX_DC_CFG),
1754 	REGISTER_AZ(SRM_TX_DC_CFG),
1755 	REGISTER_AZ(SRM_CFG),
1756 	/* BUF_TBL_UPD is WO */
1757 	REGISTER_AZ(SRM_UPD_EVQ),
1758 	REGISTER_AZ(SRAM_PARITY),
1759 	REGISTER_AZ(RX_CFG),
1760 	REGISTER_BZ(RX_FILTER_CTL),
1761 	/* RX_FLUSH_DESCQ is WO */
1762 	REGISTER_AZ(RX_DC_CFG),
1763 	REGISTER_AZ(RX_DC_PF_WM),
1764 	REGISTER_BZ(RX_RSS_TKEY),
1765 	/* RX_NODESC_DROP is RC */
1766 	REGISTER_AA(RX_SELF_RST),
1767 	/* RX_DEBUG, RX_PUSH_DROP are not used */
1768 	REGISTER_CZ(RX_RSS_IPV6_REG1),
1769 	REGISTER_CZ(RX_RSS_IPV6_REG2),
1770 	REGISTER_CZ(RX_RSS_IPV6_REG3),
1771 	/* TX_FLUSH_DESCQ is WO */
1772 	REGISTER_AZ(TX_DC_CFG),
1773 	REGISTER_AA(TX_CHKSM_CFG),
1774 	REGISTER_AZ(TX_CFG),
1775 	/* TX_PUSH_DROP is not used */
1776 	REGISTER_AZ(TX_RESERVED),
1777 	REGISTER_BZ(TX_PACE),
1778 	/* TX_PACE_DROP_QID is RC */
1779 	REGISTER_BB(TX_VLAN),
1780 	REGISTER_BZ(TX_IPFIL_PORTEN),
1781 	REGISTER_AB(MD_TXD),
1782 	REGISTER_AB(MD_RXD),
1783 	REGISTER_AB(MD_CS),
1784 	REGISTER_AB(MD_PHY_ADR),
1785 	REGISTER_AB(MD_ID),
1786 	/* MD_STAT is RC */
1787 	REGISTER_AB(MAC_STAT_DMA),
1788 	REGISTER_AB(MAC_CTRL),
1789 	REGISTER_BB(GEN_MODE),
1790 	REGISTER_AB(MAC_MC_HASH_REG0),
1791 	REGISTER_AB(MAC_MC_HASH_REG1),
1792 	REGISTER_AB(GM_CFG1),
1793 	REGISTER_AB(GM_CFG2),
1794 	/* GM_IPG and GM_HD are not used */
1795 	REGISTER_AB(GM_MAX_FLEN),
1796 	/* GM_TEST is not used */
1797 	REGISTER_AB(GM_ADR1),
1798 	REGISTER_AB(GM_ADR2),
1799 	REGISTER_AB(GMF_CFG0),
1800 	REGISTER_AB(GMF_CFG1),
1801 	REGISTER_AB(GMF_CFG2),
1802 	REGISTER_AB(GMF_CFG3),
1803 	REGISTER_AB(GMF_CFG4),
1804 	REGISTER_AB(GMF_CFG5),
1805 	REGISTER_BB(TX_SRC_MAC_CTL),
1806 	REGISTER_AB(XM_ADR_LO),
1807 	REGISTER_AB(XM_ADR_HI),
1808 	REGISTER_AB(XM_GLB_CFG),
1809 	REGISTER_AB(XM_TX_CFG),
1810 	REGISTER_AB(XM_RX_CFG),
1811 	REGISTER_AB(XM_MGT_INT_MASK),
1812 	REGISTER_AB(XM_FC),
1813 	REGISTER_AB(XM_PAUSE_TIME),
1814 	REGISTER_AB(XM_TX_PARAM),
1815 	REGISTER_AB(XM_RX_PARAM),
1816 	/* XM_MGT_INT_MSK (note no 'A') is RC */
1817 	REGISTER_AB(XX_PWR_RST),
1818 	REGISTER_AB(XX_SD_CTL),
1819 	REGISTER_AB(XX_TXDRV_CTL),
1820 	/* XX_PRBS_CTL, XX_PRBS_CHK and XX_PRBS_ERR are not used */
1821 	/* XX_CORE_STAT is partly RC */
1822 };
1823 
1824 struct efx_nic_reg_table {
1825 	u32 offset:24;
1826 	u32 min_revision:2, max_revision:2;
1827 	u32 step:6, rows:21;
1828 };
1829 
1830 #define REGISTER_TABLE_DIMENSIONS(_, offset, min_rev, max_rev, step, rows) { \
1831 	offset,								\
1832 	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev,	\
1833 	step, rows							\
1834 }
1835 #define REGISTER_TABLE(name, min_rev, max_rev)				\
1836 	REGISTER_TABLE_DIMENSIONS(					\
1837 		name, FR_ ## min_rev ## max_rev ## _ ## name,		\
1838 		min_rev, max_rev,					\
1839 		FR_ ## min_rev ## max_rev ## _ ## name ## _STEP,	\
1840 		FR_ ## min_rev ## max_rev ## _ ## name ## _ROWS)
1841 #define REGISTER_TABLE_AA(name) REGISTER_TABLE(name, A, A)
1842 #define REGISTER_TABLE_AZ(name) REGISTER_TABLE(name, A, Z)
1843 #define REGISTER_TABLE_BB(name) REGISTER_TABLE(name, B, B)
1844 #define REGISTER_TABLE_BZ(name) REGISTER_TABLE(name, B, Z)
1845 #define REGISTER_TABLE_BB_CZ(name)					\
1846 	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, B, B,		\
1847 				  FR_BZ_ ## name ## _STEP,		\
1848 				  FR_BB_ ## name ## _ROWS),		\
1849 	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, C, Z,		\
1850 				  FR_BZ_ ## name ## _STEP,		\
1851 				  FR_CZ_ ## name ## _ROWS)
1852 #define REGISTER_TABLE_CZ(name) REGISTER_TABLE(name, C, Z)
1853 
1854 static const struct efx_nic_reg_table efx_nic_reg_tables[] = {
1855 	/* DRIVER is not used */
1856 	/* EVQ_RPTR, TIMER_COMMAND, USR_EV and {RX,TX}_DESC_UPD are WO */
1857 	REGISTER_TABLE_BB(TX_IPFIL_TBL),
1858 	REGISTER_TABLE_BB(TX_SRC_MAC_TBL),
1859 	REGISTER_TABLE_AA(RX_DESC_PTR_TBL_KER),
1860 	REGISTER_TABLE_BB_CZ(RX_DESC_PTR_TBL),
1861 	REGISTER_TABLE_AA(TX_DESC_PTR_TBL_KER),
1862 	REGISTER_TABLE_BB_CZ(TX_DESC_PTR_TBL),
1863 	REGISTER_TABLE_AA(EVQ_PTR_TBL_KER),
1864 	REGISTER_TABLE_BB_CZ(EVQ_PTR_TBL),
1865 	/* We can't reasonably read all of the buffer table (up to 8MB!).
1866 	 * However this driver will only use a few entries.  Reading
1867 	 * 1K entries allows for some expansion of queue count and
1868 	 * size before we need to change the version. */
1869 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL_KER, FR_AA_BUF_FULL_TBL_KER,
1870 				  A, A, 8, 1024),
1871 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL, FR_BZ_BUF_FULL_TBL,
1872 				  B, Z, 8, 1024),
1873 	REGISTER_TABLE_CZ(RX_MAC_FILTER_TBL0),
1874 	REGISTER_TABLE_BB_CZ(TIMER_TBL),
1875 	REGISTER_TABLE_BB_CZ(TX_PACE_TBL),
1876 	REGISTER_TABLE_BZ(RX_INDIRECTION_TBL),
1877 	/* TX_FILTER_TBL0 is huge and not used by this driver */
1878 	REGISTER_TABLE_CZ(TX_MAC_FILTER_TBL0),
1879 	REGISTER_TABLE_CZ(MC_TREG_SMEM),
1880 	/* MSIX_PBA_TABLE is not mapped */
1881 	/* SRM_DBG is not mapped (and is redundant with BUF_FLL_TBL) */
1882 	REGISTER_TABLE_BZ(RX_FILTER_TBL0),
1883 };
1884 
1885 size_t efx_nic_get_regs_len(struct efx_nic *efx)
1886 {
1887 	const struct efx_nic_reg *reg;
1888 	const struct efx_nic_reg_table *table;
1889 	size_t len = 0;
1890 
1891 	for (reg = efx_nic_regs;
1892 	     reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
1893 	     reg++)
1894 		if (efx->type->revision >= reg->min_revision &&
1895 		    efx->type->revision <= reg->max_revision)
1896 			len += sizeof(efx_oword_t);
1897 
1898 	for (table = efx_nic_reg_tables;
1899 	     table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
1900 	     table++)
1901 		if (efx->type->revision >= table->min_revision &&
1902 		    efx->type->revision <= table->max_revision)
1903 			len += table->rows * min_t(size_t, table->step, 16);
1904 
1905 	return len;
1906 }
1907 
1908 void efx_nic_get_regs(struct efx_nic *efx, void *buf)
1909 {
1910 	const struct efx_nic_reg *reg;
1911 	const struct efx_nic_reg_table *table;
1912 
1913 	for (reg = efx_nic_regs;
1914 	     reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
1915 	     reg++) {
1916 		if (efx->type->revision >= reg->min_revision &&
1917 		    efx->type->revision <= reg->max_revision) {
1918 			efx_reado(efx, (efx_oword_t *)buf, reg->offset);
1919 			buf += sizeof(efx_oword_t);
1920 		}
1921 	}
1922 
1923 	for (table = efx_nic_reg_tables;
1924 	     table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
1925 	     table++) {
1926 		size_t size, i;
1927 
1928 		if (!(efx->type->revision >= table->min_revision &&
1929 		      efx->type->revision <= table->max_revision))
1930 			continue;
1931 
1932 		size = min_t(size_t, table->step, 16);
1933 
1934 		for (i = 0; i < table->rows; i++) {
1935 			switch (table->step) {
1936 			case 4: /* 32-bit register or SRAM */
1937 				efx_readd_table(efx, buf, table->offset, i);
1938 				break;
1939 			case 8: /* 64-bit SRAM */
1940 				efx_sram_readq(efx,
1941 					       efx->membase + table->offset,
1942 					       buf, i);
1943 				break;
1944 			case 16: /* 128-bit register */
1945 				efx_reado_table(efx, buf, table->offset, i);
1946 				break;
1947 			case 32: /* 128-bit register, interleaved */
1948 				efx_reado_table(efx, buf, table->offset, 2 * i);
1949 				break;
1950 			default:
1951 				WARN_ON(1);
1952 				return;
1953 			}
1954 			buf += size;
1955 		}
1956 	}
1957 }
1958