xref: /openbmc/linux/drivers/net/ethernet/sfc/nic.c (revision d07df8ec)
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 /* Depth of RX flush request fifo */
53 #define EFX_RX_FLUSH_COUNT 4
54 
55 /* Driver generated events */
56 #define _EFX_CHANNEL_MAGIC_TEST		0x000101
57 #define _EFX_CHANNEL_MAGIC_FILL		0x000102
58 #define _EFX_CHANNEL_MAGIC_RX_DRAIN	0x000103
59 #define _EFX_CHANNEL_MAGIC_TX_DRAIN	0x000104
60 
61 #define _EFX_CHANNEL_MAGIC(_code, _data)	((_code) << 8 | (_data))
62 #define _EFX_CHANNEL_MAGIC_CODE(_magic)		((_magic) >> 8)
63 
64 #define EFX_CHANNEL_MAGIC_TEST(_channel)				\
65 	_EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_TEST, (_channel)->channel)
66 #define EFX_CHANNEL_MAGIC_FILL(_rx_queue)				\
67 	_EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_FILL,			\
68 			   efx_rx_queue_index(_rx_queue))
69 #define EFX_CHANNEL_MAGIC_RX_DRAIN(_rx_queue)				\
70 	_EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_RX_DRAIN,			\
71 			   efx_rx_queue_index(_rx_queue))
72 #define EFX_CHANNEL_MAGIC_TX_DRAIN(_tx_queue)				\
73 	_EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_TX_DRAIN,			\
74 			   (_tx_queue)->queue)
75 
76 static void efx_magic_event(struct efx_channel *channel, u32 magic);
77 
78 /**************************************************************************
79  *
80  * Solarstorm hardware access
81  *
82  **************************************************************************/
83 
84 static inline void efx_write_buf_tbl(struct efx_nic *efx, efx_qword_t *value,
85 				     unsigned int index)
86 {
87 	efx_sram_writeq(efx, efx->membase + efx->type->buf_tbl_base,
88 			value, index);
89 }
90 
91 /* Read the current event from the event queue */
92 static inline efx_qword_t *efx_event(struct efx_channel *channel,
93 				     unsigned int index)
94 {
95 	return ((efx_qword_t *) (channel->eventq.addr)) +
96 		(index & channel->eventq_mask);
97 }
98 
99 /* See if an event is present
100  *
101  * We check both the high and low dword of the event for all ones.  We
102  * wrote all ones when we cleared the event, and no valid event can
103  * have all ones in either its high or low dwords.  This approach is
104  * robust against reordering.
105  *
106  * Note that using a single 64-bit comparison is incorrect; even
107  * though the CPU read will be atomic, the DMA write may not be.
108  */
109 static inline int efx_event_present(efx_qword_t *event)
110 {
111 	return !(EFX_DWORD_IS_ALL_ONES(event->dword[0]) |
112 		  EFX_DWORD_IS_ALL_ONES(event->dword[1]));
113 }
114 
115 static bool efx_masked_compare_oword(const efx_oword_t *a, const efx_oword_t *b,
116 				     const efx_oword_t *mask)
117 {
118 	return ((a->u64[0] ^ b->u64[0]) & mask->u64[0]) ||
119 		((a->u64[1] ^ b->u64[1]) & mask->u64[1]);
120 }
121 
122 int efx_nic_test_registers(struct efx_nic *efx,
123 			   const struct efx_nic_register_test *regs,
124 			   size_t n_regs)
125 {
126 	unsigned address = 0, i, j;
127 	efx_oword_t mask, imask, original, reg, buf;
128 
129 	for (i = 0; i < n_regs; ++i) {
130 		address = regs[i].address;
131 		mask = imask = regs[i].mask;
132 		EFX_INVERT_OWORD(imask);
133 
134 		efx_reado(efx, &original, address);
135 
136 		/* bit sweep on and off */
137 		for (j = 0; j < 128; j++) {
138 			if (!EFX_EXTRACT_OWORD32(mask, j, j))
139 				continue;
140 
141 			/* Test this testable bit can be set in isolation */
142 			EFX_AND_OWORD(reg, original, mask);
143 			EFX_SET_OWORD32(reg, j, j, 1);
144 
145 			efx_writeo(efx, &reg, address);
146 			efx_reado(efx, &buf, address);
147 
148 			if (efx_masked_compare_oword(&reg, &buf, &mask))
149 				goto fail;
150 
151 			/* Test this testable bit can be cleared in isolation */
152 			EFX_OR_OWORD(reg, original, mask);
153 			EFX_SET_OWORD32(reg, j, j, 0);
154 
155 			efx_writeo(efx, &reg, address);
156 			efx_reado(efx, &buf, address);
157 
158 			if (efx_masked_compare_oword(&reg, &buf, &mask))
159 				goto fail;
160 		}
161 
162 		efx_writeo(efx, &original, address);
163 	}
164 
165 	return 0;
166 
167 fail:
168 	netif_err(efx, hw, efx->net_dev,
169 		  "wrote "EFX_OWORD_FMT" read "EFX_OWORD_FMT
170 		  " at address 0x%x mask "EFX_OWORD_FMT"\n", EFX_OWORD_VAL(reg),
171 		  EFX_OWORD_VAL(buf), address, EFX_OWORD_VAL(mask));
172 	return -EIO;
173 }
174 
175 /**************************************************************************
176  *
177  * Special buffer handling
178  * Special buffers are used for event queues and the TX and RX
179  * descriptor rings.
180  *
181  *************************************************************************/
182 
183 /*
184  * Initialise a special buffer
185  *
186  * This will define a buffer (previously allocated via
187  * efx_alloc_special_buffer()) in the buffer table, allowing
188  * it to be used for event queues, descriptor rings etc.
189  */
190 static void
191 efx_init_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
192 {
193 	efx_qword_t buf_desc;
194 	unsigned int index;
195 	dma_addr_t dma_addr;
196 	int i;
197 
198 	EFX_BUG_ON_PARANOID(!buffer->addr);
199 
200 	/* Write buffer descriptors to NIC */
201 	for (i = 0; i < buffer->entries; i++) {
202 		index = buffer->index + i;
203 		dma_addr = buffer->dma_addr + (i * EFX_BUF_SIZE);
204 		netif_dbg(efx, probe, efx->net_dev,
205 			  "mapping special buffer %d at %llx\n",
206 			  index, (unsigned long long)dma_addr);
207 		EFX_POPULATE_QWORD_3(buf_desc,
208 				     FRF_AZ_BUF_ADR_REGION, 0,
209 				     FRF_AZ_BUF_ADR_FBUF, dma_addr >> 12,
210 				     FRF_AZ_BUF_OWNER_ID_FBUF, 0);
211 		efx_write_buf_tbl(efx, &buf_desc, index);
212 	}
213 }
214 
215 /* Unmaps a buffer and clears the buffer table entries */
216 static void
217 efx_fini_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
218 {
219 	efx_oword_t buf_tbl_upd;
220 	unsigned int start = buffer->index;
221 	unsigned int end = (buffer->index + buffer->entries - 1);
222 
223 	if (!buffer->entries)
224 		return;
225 
226 	netif_dbg(efx, hw, efx->net_dev, "unmapping special buffers %d-%d\n",
227 		  buffer->index, buffer->index + buffer->entries - 1);
228 
229 	EFX_POPULATE_OWORD_4(buf_tbl_upd,
230 			     FRF_AZ_BUF_UPD_CMD, 0,
231 			     FRF_AZ_BUF_CLR_CMD, 1,
232 			     FRF_AZ_BUF_CLR_END_ID, end,
233 			     FRF_AZ_BUF_CLR_START_ID, start);
234 	efx_writeo(efx, &buf_tbl_upd, FR_AZ_BUF_TBL_UPD);
235 }
236 
237 /*
238  * Allocate a new special buffer
239  *
240  * This allocates memory for a new buffer, clears it and allocates a
241  * new buffer ID range.  It does not write into the buffer table.
242  *
243  * This call will allocate 4KB buffers, since 8KB buffers can't be
244  * used for event queues and descriptor rings.
245  */
246 static int efx_alloc_special_buffer(struct efx_nic *efx,
247 				    struct efx_special_buffer *buffer,
248 				    unsigned int len)
249 {
250 	len = ALIGN(len, EFX_BUF_SIZE);
251 
252 	buffer->addr = dma_alloc_coherent(&efx->pci_dev->dev, len,
253 					  &buffer->dma_addr, GFP_KERNEL);
254 	if (!buffer->addr)
255 		return -ENOMEM;
256 	buffer->len = len;
257 	buffer->entries = len / EFX_BUF_SIZE;
258 	BUG_ON(buffer->dma_addr & (EFX_BUF_SIZE - 1));
259 
260 	/* Select new buffer ID */
261 	buffer->index = efx->next_buffer_table;
262 	efx->next_buffer_table += buffer->entries;
263 #ifdef CONFIG_SFC_SRIOV
264 	BUG_ON(efx_sriov_enabled(efx) &&
265 	       efx->vf_buftbl_base < efx->next_buffer_table);
266 #endif
267 
268 	netif_dbg(efx, probe, efx->net_dev,
269 		  "allocating special buffers %d-%d at %llx+%x "
270 		  "(virt %p phys %llx)\n", buffer->index,
271 		  buffer->index + buffer->entries - 1,
272 		  (u64)buffer->dma_addr, len,
273 		  buffer->addr, (u64)virt_to_phys(buffer->addr));
274 
275 	return 0;
276 }
277 
278 static void
279 efx_free_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
280 {
281 	if (!buffer->addr)
282 		return;
283 
284 	netif_dbg(efx, hw, efx->net_dev,
285 		  "deallocating special buffers %d-%d at %llx+%x "
286 		  "(virt %p phys %llx)\n", buffer->index,
287 		  buffer->index + buffer->entries - 1,
288 		  (u64)buffer->dma_addr, buffer->len,
289 		  buffer->addr, (u64)virt_to_phys(buffer->addr));
290 
291 	dma_free_coherent(&efx->pci_dev->dev, buffer->len, buffer->addr,
292 			  buffer->dma_addr);
293 	buffer->addr = NULL;
294 	buffer->entries = 0;
295 }
296 
297 /**************************************************************************
298  *
299  * Generic buffer handling
300  * These buffers are used for interrupt status, MAC stats, etc.
301  *
302  **************************************************************************/
303 
304 int efx_nic_alloc_buffer(struct efx_nic *efx, struct efx_buffer *buffer,
305 			 unsigned int len)
306 {
307 	buffer->addr = dma_alloc_coherent(&efx->pci_dev->dev, len,
308 					  &buffer->dma_addr,
309 					  GFP_ATOMIC | __GFP_ZERO);
310 	if (!buffer->addr)
311 		return -ENOMEM;
312 	buffer->len = len;
313 	return 0;
314 }
315 
316 void efx_nic_free_buffer(struct efx_nic *efx, struct efx_buffer *buffer)
317 {
318 	if (buffer->addr) {
319 		dma_free_coherent(&efx->pci_dev->dev, buffer->len,
320 				  buffer->addr, buffer->dma_addr);
321 		buffer->addr = NULL;
322 	}
323 }
324 
325 /**************************************************************************
326  *
327  * TX path
328  *
329  **************************************************************************/
330 
331 /* Returns a pointer to the specified transmit descriptor in the TX
332  * descriptor queue belonging to the specified channel.
333  */
334 static inline efx_qword_t *
335 efx_tx_desc(struct efx_tx_queue *tx_queue, unsigned int index)
336 {
337 	return ((efx_qword_t *) (tx_queue->txd.addr)) + index;
338 }
339 
340 /* This writes to the TX_DESC_WPTR; write pointer for TX descriptor ring */
341 static inline void efx_notify_tx_desc(struct efx_tx_queue *tx_queue)
342 {
343 	unsigned write_ptr;
344 	efx_dword_t reg;
345 
346 	write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
347 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_TX_DESC_WPTR_DWORD, write_ptr);
348 	efx_writed_page(tx_queue->efx, &reg,
349 			FR_AZ_TX_DESC_UPD_DWORD_P0, tx_queue->queue);
350 }
351 
352 /* Write pointer and first descriptor for TX descriptor ring */
353 static inline void efx_push_tx_desc(struct efx_tx_queue *tx_queue,
354 				    const efx_qword_t *txd)
355 {
356 	unsigned write_ptr;
357 	efx_oword_t reg;
358 
359 	BUILD_BUG_ON(FRF_AZ_TX_DESC_LBN != 0);
360 	BUILD_BUG_ON(FR_AA_TX_DESC_UPD_KER != FR_BZ_TX_DESC_UPD_P0);
361 
362 	write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
363 	EFX_POPULATE_OWORD_2(reg, FRF_AZ_TX_DESC_PUSH_CMD, true,
364 			     FRF_AZ_TX_DESC_WPTR, write_ptr);
365 	reg.qword[0] = *txd;
366 	efx_writeo_page(tx_queue->efx, &reg,
367 			FR_BZ_TX_DESC_UPD_P0, tx_queue->queue);
368 }
369 
370 static inline bool
371 efx_may_push_tx_desc(struct efx_tx_queue *tx_queue, unsigned int write_count)
372 {
373 	unsigned empty_read_count = ACCESS_ONCE(tx_queue->empty_read_count);
374 
375 	if (empty_read_count == 0)
376 		return false;
377 
378 	tx_queue->empty_read_count = 0;
379 	return ((empty_read_count ^ write_count) & ~EFX_EMPTY_COUNT_VALID) == 0
380 		&& tx_queue->write_count - write_count == 1;
381 }
382 
383 /* For each entry inserted into the software descriptor ring, create a
384  * descriptor in the hardware TX descriptor ring (in host memory), and
385  * write a doorbell.
386  */
387 void efx_nic_push_buffers(struct efx_tx_queue *tx_queue)
388 {
389 
390 	struct efx_tx_buffer *buffer;
391 	efx_qword_t *txd;
392 	unsigned write_ptr;
393 	unsigned old_write_count = tx_queue->write_count;
394 
395 	BUG_ON(tx_queue->write_count == tx_queue->insert_count);
396 
397 	do {
398 		write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
399 		buffer = &tx_queue->buffer[write_ptr];
400 		txd = efx_tx_desc(tx_queue, write_ptr);
401 		++tx_queue->write_count;
402 
403 		/* Create TX descriptor ring entry */
404 		BUILD_BUG_ON(EFX_TX_BUF_CONT != 1);
405 		EFX_POPULATE_QWORD_4(*txd,
406 				     FSF_AZ_TX_KER_CONT,
407 				     buffer->flags & EFX_TX_BUF_CONT,
408 				     FSF_AZ_TX_KER_BYTE_COUNT, buffer->len,
409 				     FSF_AZ_TX_KER_BUF_REGION, 0,
410 				     FSF_AZ_TX_KER_BUF_ADDR, buffer->dma_addr);
411 	} while (tx_queue->write_count != tx_queue->insert_count);
412 
413 	wmb(); /* Ensure descriptors are written before they are fetched */
414 
415 	if (efx_may_push_tx_desc(tx_queue, old_write_count)) {
416 		txd = efx_tx_desc(tx_queue,
417 				  old_write_count & tx_queue->ptr_mask);
418 		efx_push_tx_desc(tx_queue, txd);
419 		++tx_queue->pushes;
420 	} else {
421 		efx_notify_tx_desc(tx_queue);
422 	}
423 }
424 
425 /* Allocate hardware resources for a TX queue */
426 int efx_nic_probe_tx(struct efx_tx_queue *tx_queue)
427 {
428 	struct efx_nic *efx = tx_queue->efx;
429 	unsigned entries;
430 
431 	entries = tx_queue->ptr_mask + 1;
432 	return efx_alloc_special_buffer(efx, &tx_queue->txd,
433 					entries * sizeof(efx_qword_t));
434 }
435 
436 void efx_nic_init_tx(struct efx_tx_queue *tx_queue)
437 {
438 	struct efx_nic *efx = tx_queue->efx;
439 	efx_oword_t reg;
440 
441 	/* Pin TX descriptor ring */
442 	efx_init_special_buffer(efx, &tx_queue->txd);
443 
444 	/* Push TX descriptor ring to card */
445 	EFX_POPULATE_OWORD_10(reg,
446 			      FRF_AZ_TX_DESCQ_EN, 1,
447 			      FRF_AZ_TX_ISCSI_DDIG_EN, 0,
448 			      FRF_AZ_TX_ISCSI_HDIG_EN, 0,
449 			      FRF_AZ_TX_DESCQ_BUF_BASE_ID, tx_queue->txd.index,
450 			      FRF_AZ_TX_DESCQ_EVQ_ID,
451 			      tx_queue->channel->channel,
452 			      FRF_AZ_TX_DESCQ_OWNER_ID, 0,
453 			      FRF_AZ_TX_DESCQ_LABEL, tx_queue->queue,
454 			      FRF_AZ_TX_DESCQ_SIZE,
455 			      __ffs(tx_queue->txd.entries),
456 			      FRF_AZ_TX_DESCQ_TYPE, 0,
457 			      FRF_BZ_TX_NON_IP_DROP_DIS, 1);
458 
459 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
460 		int csum = tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD;
461 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_IP_CHKSM_DIS, !csum);
462 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_TCP_CHKSM_DIS,
463 				    !csum);
464 	}
465 
466 	efx_writeo_table(efx, &reg, efx->type->txd_ptr_tbl_base,
467 			 tx_queue->queue);
468 
469 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0) {
470 		/* Only 128 bits in this register */
471 		BUILD_BUG_ON(EFX_MAX_TX_QUEUES > 128);
472 
473 		efx_reado(efx, &reg, FR_AA_TX_CHKSM_CFG);
474 		if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD)
475 			__clear_bit_le(tx_queue->queue, &reg);
476 		else
477 			__set_bit_le(tx_queue->queue, &reg);
478 		efx_writeo(efx, &reg, FR_AA_TX_CHKSM_CFG);
479 	}
480 
481 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
482 		EFX_POPULATE_OWORD_1(reg,
483 				     FRF_BZ_TX_PACE,
484 				     (tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
485 				     FFE_BZ_TX_PACE_OFF :
486 				     FFE_BZ_TX_PACE_RESERVED);
487 		efx_writeo_table(efx, &reg, FR_BZ_TX_PACE_TBL,
488 				 tx_queue->queue);
489 	}
490 }
491 
492 static void efx_flush_tx_queue(struct efx_tx_queue *tx_queue)
493 {
494 	struct efx_nic *efx = tx_queue->efx;
495 	efx_oword_t tx_flush_descq;
496 
497 	WARN_ON(atomic_read(&tx_queue->flush_outstanding));
498 	atomic_set(&tx_queue->flush_outstanding, 1);
499 
500 	EFX_POPULATE_OWORD_2(tx_flush_descq,
501 			     FRF_AZ_TX_FLUSH_DESCQ_CMD, 1,
502 			     FRF_AZ_TX_FLUSH_DESCQ, tx_queue->queue);
503 	efx_writeo(efx, &tx_flush_descq, FR_AZ_TX_FLUSH_DESCQ);
504 }
505 
506 void efx_nic_fini_tx(struct efx_tx_queue *tx_queue)
507 {
508 	struct efx_nic *efx = tx_queue->efx;
509 	efx_oword_t tx_desc_ptr;
510 
511 	/* Remove TX descriptor ring from card */
512 	EFX_ZERO_OWORD(tx_desc_ptr);
513 	efx_writeo_table(efx, &tx_desc_ptr, efx->type->txd_ptr_tbl_base,
514 			 tx_queue->queue);
515 
516 	/* Unpin TX descriptor ring */
517 	efx_fini_special_buffer(efx, &tx_queue->txd);
518 }
519 
520 /* Free buffers backing TX queue */
521 void efx_nic_remove_tx(struct efx_tx_queue *tx_queue)
522 {
523 	efx_free_special_buffer(tx_queue->efx, &tx_queue->txd);
524 }
525 
526 /**************************************************************************
527  *
528  * RX path
529  *
530  **************************************************************************/
531 
532 /* Returns a pointer to the specified descriptor in the RX descriptor queue */
533 static inline efx_qword_t *
534 efx_rx_desc(struct efx_rx_queue *rx_queue, unsigned int index)
535 {
536 	return ((efx_qword_t *) (rx_queue->rxd.addr)) + index;
537 }
538 
539 /* This creates an entry in the RX descriptor queue */
540 static inline void
541 efx_build_rx_desc(struct efx_rx_queue *rx_queue, unsigned index)
542 {
543 	struct efx_rx_buffer *rx_buf;
544 	efx_qword_t *rxd;
545 
546 	rxd = efx_rx_desc(rx_queue, index);
547 	rx_buf = efx_rx_buffer(rx_queue, index);
548 	EFX_POPULATE_QWORD_3(*rxd,
549 			     FSF_AZ_RX_KER_BUF_SIZE,
550 			     rx_buf->len -
551 			     rx_queue->efx->type->rx_buffer_padding,
552 			     FSF_AZ_RX_KER_BUF_REGION, 0,
553 			     FSF_AZ_RX_KER_BUF_ADDR, rx_buf->dma_addr);
554 }
555 
556 /* This writes to the RX_DESC_WPTR register for the specified receive
557  * descriptor ring.
558  */
559 void efx_nic_notify_rx_desc(struct efx_rx_queue *rx_queue)
560 {
561 	struct efx_nic *efx = rx_queue->efx;
562 	efx_dword_t reg;
563 	unsigned write_ptr;
564 
565 	while (rx_queue->notified_count != rx_queue->added_count) {
566 		efx_build_rx_desc(
567 			rx_queue,
568 			rx_queue->notified_count & rx_queue->ptr_mask);
569 		++rx_queue->notified_count;
570 	}
571 
572 	wmb();
573 	write_ptr = rx_queue->added_count & rx_queue->ptr_mask;
574 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_RX_DESC_WPTR_DWORD, write_ptr);
575 	efx_writed_page(efx, &reg, FR_AZ_RX_DESC_UPD_DWORD_P0,
576 			efx_rx_queue_index(rx_queue));
577 }
578 
579 int efx_nic_probe_rx(struct efx_rx_queue *rx_queue)
580 {
581 	struct efx_nic *efx = rx_queue->efx;
582 	unsigned entries;
583 
584 	entries = rx_queue->ptr_mask + 1;
585 	return efx_alloc_special_buffer(efx, &rx_queue->rxd,
586 					entries * sizeof(efx_qword_t));
587 }
588 
589 void efx_nic_init_rx(struct efx_rx_queue *rx_queue)
590 {
591 	efx_oword_t rx_desc_ptr;
592 	struct efx_nic *efx = rx_queue->efx;
593 	bool is_b0 = efx_nic_rev(efx) >= EFX_REV_FALCON_B0;
594 	bool iscsi_digest_en = is_b0;
595 	bool jumbo_en;
596 
597 	/* For kernel-mode queues in Falcon A1, the JUMBO flag enables
598 	 * DMA to continue after a PCIe page boundary (and scattering
599 	 * is not possible).  In Falcon B0 and Siena, it enables
600 	 * scatter.
601 	 */
602 	jumbo_en = !is_b0 || efx->rx_scatter;
603 
604 	netif_dbg(efx, hw, efx->net_dev,
605 		  "RX queue %d ring in special buffers %d-%d\n",
606 		  efx_rx_queue_index(rx_queue), rx_queue->rxd.index,
607 		  rx_queue->rxd.index + rx_queue->rxd.entries - 1);
608 
609 	rx_queue->scatter_n = 0;
610 
611 	/* Pin RX descriptor ring */
612 	efx_init_special_buffer(efx, &rx_queue->rxd);
613 
614 	/* Push RX descriptor ring to card */
615 	EFX_POPULATE_OWORD_10(rx_desc_ptr,
616 			      FRF_AZ_RX_ISCSI_DDIG_EN, iscsi_digest_en,
617 			      FRF_AZ_RX_ISCSI_HDIG_EN, iscsi_digest_en,
618 			      FRF_AZ_RX_DESCQ_BUF_BASE_ID, rx_queue->rxd.index,
619 			      FRF_AZ_RX_DESCQ_EVQ_ID,
620 			      efx_rx_queue_channel(rx_queue)->channel,
621 			      FRF_AZ_RX_DESCQ_OWNER_ID, 0,
622 			      FRF_AZ_RX_DESCQ_LABEL,
623 			      efx_rx_queue_index(rx_queue),
624 			      FRF_AZ_RX_DESCQ_SIZE,
625 			      __ffs(rx_queue->rxd.entries),
626 			      FRF_AZ_RX_DESCQ_TYPE, 0 /* kernel queue */ ,
627 			      FRF_AZ_RX_DESCQ_JUMBO, jumbo_en,
628 			      FRF_AZ_RX_DESCQ_EN, 1);
629 	efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
630 			 efx_rx_queue_index(rx_queue));
631 }
632 
633 static void efx_flush_rx_queue(struct efx_rx_queue *rx_queue)
634 {
635 	struct efx_nic *efx = rx_queue->efx;
636 	efx_oword_t rx_flush_descq;
637 
638 	EFX_POPULATE_OWORD_2(rx_flush_descq,
639 			     FRF_AZ_RX_FLUSH_DESCQ_CMD, 1,
640 			     FRF_AZ_RX_FLUSH_DESCQ,
641 			     efx_rx_queue_index(rx_queue));
642 	efx_writeo(efx, &rx_flush_descq, FR_AZ_RX_FLUSH_DESCQ);
643 }
644 
645 void efx_nic_fini_rx(struct efx_rx_queue *rx_queue)
646 {
647 	efx_oword_t rx_desc_ptr;
648 	struct efx_nic *efx = rx_queue->efx;
649 
650 	/* Remove RX descriptor ring from card */
651 	EFX_ZERO_OWORD(rx_desc_ptr);
652 	efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
653 			 efx_rx_queue_index(rx_queue));
654 
655 	/* Unpin RX descriptor ring */
656 	efx_fini_special_buffer(efx, &rx_queue->rxd);
657 }
658 
659 /* Free buffers backing RX queue */
660 void efx_nic_remove_rx(struct efx_rx_queue *rx_queue)
661 {
662 	efx_free_special_buffer(rx_queue->efx, &rx_queue->rxd);
663 }
664 
665 /**************************************************************************
666  *
667  * Flush handling
668  *
669  **************************************************************************/
670 
671 /* efx_nic_flush_queues() must be woken up when all flushes are completed,
672  * or more RX flushes can be kicked off.
673  */
674 static bool efx_flush_wake(struct efx_nic *efx)
675 {
676 	/* Ensure that all updates are visible to efx_nic_flush_queues() */
677 	smp_mb();
678 
679 	return (atomic_read(&efx->drain_pending) == 0 ||
680 		(atomic_read(&efx->rxq_flush_outstanding) < EFX_RX_FLUSH_COUNT
681 		 && atomic_read(&efx->rxq_flush_pending) > 0));
682 }
683 
684 static bool efx_check_tx_flush_complete(struct efx_nic *efx)
685 {
686 	bool i = true;
687 	efx_oword_t txd_ptr_tbl;
688 	struct efx_channel *channel;
689 	struct efx_tx_queue *tx_queue;
690 
691 	efx_for_each_channel(channel, efx) {
692 		efx_for_each_channel_tx_queue(tx_queue, channel) {
693 			efx_reado_table(efx, &txd_ptr_tbl,
694 					FR_BZ_TX_DESC_PTR_TBL, tx_queue->queue);
695 			if (EFX_OWORD_FIELD(txd_ptr_tbl,
696 					    FRF_AZ_TX_DESCQ_FLUSH) ||
697 			    EFX_OWORD_FIELD(txd_ptr_tbl,
698 					    FRF_AZ_TX_DESCQ_EN)) {
699 				netif_dbg(efx, hw, efx->net_dev,
700 					  "flush did not complete on TXQ %d\n",
701 					  tx_queue->queue);
702 				i = false;
703 			} else if (atomic_cmpxchg(&tx_queue->flush_outstanding,
704 						  1, 0)) {
705 				/* The flush is complete, but we didn't
706 				 * receive a flush completion event
707 				 */
708 				netif_dbg(efx, hw, efx->net_dev,
709 					  "flush complete on TXQ %d, so drain "
710 					  "the queue\n", tx_queue->queue);
711 				/* Don't need to increment drain_pending as it
712 				 * has already been incremented for the queues
713 				 * which did not drain
714 				 */
715 				efx_magic_event(channel,
716 						EFX_CHANNEL_MAGIC_TX_DRAIN(
717 							tx_queue));
718 			}
719 		}
720 	}
721 
722 	return i;
723 }
724 
725 /* Flush all the transmit queues, and continue flushing receive queues until
726  * they're all flushed. Wait for the DRAIN events to be recieved so that there
727  * are no more RX and TX events left on any channel. */
728 int efx_nic_flush_queues(struct efx_nic *efx)
729 {
730 	unsigned timeout = msecs_to_jiffies(5000); /* 5s for all flushes and drains */
731 	struct efx_channel *channel;
732 	struct efx_rx_queue *rx_queue;
733 	struct efx_tx_queue *tx_queue;
734 	int rc = 0;
735 
736 	efx->type->prepare_flush(efx);
737 
738 	efx_for_each_channel(channel, efx) {
739 		efx_for_each_channel_tx_queue(tx_queue, channel) {
740 			atomic_inc(&efx->drain_pending);
741 			efx_flush_tx_queue(tx_queue);
742 		}
743 		efx_for_each_channel_rx_queue(rx_queue, channel) {
744 			atomic_inc(&efx->drain_pending);
745 			rx_queue->flush_pending = true;
746 			atomic_inc(&efx->rxq_flush_pending);
747 		}
748 	}
749 
750 	while (timeout && atomic_read(&efx->drain_pending) > 0) {
751 		/* If SRIOV is enabled, then offload receive queue flushing to
752 		 * the firmware (though we will still have to poll for
753 		 * completion). If that fails, fall back to the old scheme.
754 		 */
755 		if (efx_sriov_enabled(efx)) {
756 			rc = efx_mcdi_flush_rxqs(efx);
757 			if (!rc)
758 				goto wait;
759 		}
760 
761 		/* The hardware supports four concurrent rx flushes, each of
762 		 * which may need to be retried if there is an outstanding
763 		 * descriptor fetch
764 		 */
765 		efx_for_each_channel(channel, efx) {
766 			efx_for_each_channel_rx_queue(rx_queue, channel) {
767 				if (atomic_read(&efx->rxq_flush_outstanding) >=
768 				    EFX_RX_FLUSH_COUNT)
769 					break;
770 
771 				if (rx_queue->flush_pending) {
772 					rx_queue->flush_pending = false;
773 					atomic_dec(&efx->rxq_flush_pending);
774 					atomic_inc(&efx->rxq_flush_outstanding);
775 					efx_flush_rx_queue(rx_queue);
776 				}
777 			}
778 		}
779 
780 	wait:
781 		timeout = wait_event_timeout(efx->flush_wq, efx_flush_wake(efx),
782 					     timeout);
783 	}
784 
785 	if (atomic_read(&efx->drain_pending) &&
786 	    !efx_check_tx_flush_complete(efx)) {
787 		netif_err(efx, hw, efx->net_dev, "failed to flush %d queues "
788 			  "(rx %d+%d)\n", atomic_read(&efx->drain_pending),
789 			  atomic_read(&efx->rxq_flush_outstanding),
790 			  atomic_read(&efx->rxq_flush_pending));
791 		rc = -ETIMEDOUT;
792 
793 		atomic_set(&efx->drain_pending, 0);
794 		atomic_set(&efx->rxq_flush_pending, 0);
795 		atomic_set(&efx->rxq_flush_outstanding, 0);
796 	}
797 
798 	efx->type->finish_flush(efx);
799 
800 	return rc;
801 }
802 
803 /**************************************************************************
804  *
805  * Event queue processing
806  * Event queues are processed by per-channel tasklets.
807  *
808  **************************************************************************/
809 
810 /* Update a channel's event queue's read pointer (RPTR) register
811  *
812  * This writes the EVQ_RPTR_REG register for the specified channel's
813  * event queue.
814  */
815 void efx_nic_eventq_read_ack(struct efx_channel *channel)
816 {
817 	efx_dword_t reg;
818 	struct efx_nic *efx = channel->efx;
819 
820 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_EVQ_RPTR,
821 			     channel->eventq_read_ptr & channel->eventq_mask);
822 
823 	/* For Falcon A1, EVQ_RPTR_KER is documented as having a step size
824 	 * of 4 bytes, but it is really 16 bytes just like later revisions.
825 	 */
826 	efx_writed(efx, &reg,
827 		   efx->type->evq_rptr_tbl_base +
828 		   FR_BZ_EVQ_RPTR_STEP * channel->channel);
829 }
830 
831 /* Use HW to insert a SW defined event */
832 void efx_generate_event(struct efx_nic *efx, unsigned int evq,
833 			efx_qword_t *event)
834 {
835 	efx_oword_t drv_ev_reg;
836 
837 	BUILD_BUG_ON(FRF_AZ_DRV_EV_DATA_LBN != 0 ||
838 		     FRF_AZ_DRV_EV_DATA_WIDTH != 64);
839 	drv_ev_reg.u32[0] = event->u32[0];
840 	drv_ev_reg.u32[1] = event->u32[1];
841 	drv_ev_reg.u32[2] = 0;
842 	drv_ev_reg.u32[3] = 0;
843 	EFX_SET_OWORD_FIELD(drv_ev_reg, FRF_AZ_DRV_EV_QID, evq);
844 	efx_writeo(efx, &drv_ev_reg, FR_AZ_DRV_EV);
845 }
846 
847 static void efx_magic_event(struct efx_channel *channel, u32 magic)
848 {
849 	efx_qword_t event;
850 
851 	EFX_POPULATE_QWORD_2(event, FSF_AZ_EV_CODE,
852 			     FSE_AZ_EV_CODE_DRV_GEN_EV,
853 			     FSF_AZ_DRV_GEN_EV_MAGIC, magic);
854 	efx_generate_event(channel->efx, channel->channel, &event);
855 }
856 
857 /* Handle a transmit completion event
858  *
859  * The NIC batches TX completion events; the message we receive is of
860  * the form "complete all TX events up to this index".
861  */
862 static int
863 efx_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
864 {
865 	unsigned int tx_ev_desc_ptr;
866 	unsigned int tx_ev_q_label;
867 	struct efx_tx_queue *tx_queue;
868 	struct efx_nic *efx = channel->efx;
869 	int tx_packets = 0;
870 
871 	if (unlikely(ACCESS_ONCE(efx->reset_pending)))
872 		return 0;
873 
874 	if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
875 		/* Transmit completion */
876 		tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
877 		tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
878 		tx_queue = efx_channel_get_tx_queue(
879 			channel, tx_ev_q_label % EFX_TXQ_TYPES);
880 		tx_packets = ((tx_ev_desc_ptr - tx_queue->read_count) &
881 			      tx_queue->ptr_mask);
882 		efx_xmit_done(tx_queue, tx_ev_desc_ptr);
883 	} else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
884 		/* Rewrite the FIFO write pointer */
885 		tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
886 		tx_queue = efx_channel_get_tx_queue(
887 			channel, tx_ev_q_label % EFX_TXQ_TYPES);
888 
889 		netif_tx_lock(efx->net_dev);
890 		efx_notify_tx_desc(tx_queue);
891 		netif_tx_unlock(efx->net_dev);
892 	} else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR) &&
893 		   EFX_WORKAROUND_10727(efx)) {
894 		efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
895 	} else {
896 		netif_err(efx, tx_err, efx->net_dev,
897 			  "channel %d unexpected TX event "
898 			  EFX_QWORD_FMT"\n", channel->channel,
899 			  EFX_QWORD_VAL(*event));
900 	}
901 
902 	return tx_packets;
903 }
904 
905 /* Detect errors included in the rx_evt_pkt_ok bit. */
906 static u16 efx_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
907 				const efx_qword_t *event)
908 {
909 	struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
910 	struct efx_nic *efx = rx_queue->efx;
911 	bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
912 	bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
913 	bool rx_ev_frm_trunc, rx_ev_drib_nib, rx_ev_tobe_disc;
914 	bool rx_ev_other_err, rx_ev_pause_frm;
915 	bool rx_ev_hdr_type, rx_ev_mcast_pkt;
916 	unsigned rx_ev_pkt_type;
917 
918 	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
919 	rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
920 	rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
921 	rx_ev_pkt_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_TYPE);
922 	rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
923 						 FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
924 	rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
925 						  FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
926 	rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
927 						   FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
928 	rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
929 	rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
930 	rx_ev_drib_nib = ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) ?
931 			  0 : EFX_QWORD_FIELD(*event, FSF_AA_RX_EV_DRIB_NIB));
932 	rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
933 
934 	/* Every error apart from tobe_disc and pause_frm */
935 	rx_ev_other_err = (rx_ev_drib_nib | rx_ev_tcp_udp_chksum_err |
936 			   rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
937 			   rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
938 
939 	/* Count errors that are not in MAC stats.  Ignore expected
940 	 * checksum errors during self-test. */
941 	if (rx_ev_frm_trunc)
942 		++channel->n_rx_frm_trunc;
943 	else if (rx_ev_tobe_disc)
944 		++channel->n_rx_tobe_disc;
945 	else if (!efx->loopback_selftest) {
946 		if (rx_ev_ip_hdr_chksum_err)
947 			++channel->n_rx_ip_hdr_chksum_err;
948 		else if (rx_ev_tcp_udp_chksum_err)
949 			++channel->n_rx_tcp_udp_chksum_err;
950 	}
951 
952 	/* TOBE_DISC is expected on unicast mismatches; don't print out an
953 	 * error message.  FRM_TRUNC indicates RXDP dropped the packet due
954 	 * to a FIFO overflow.
955 	 */
956 #ifdef DEBUG
957 	if (rx_ev_other_err && net_ratelimit()) {
958 		netif_dbg(efx, rx_err, efx->net_dev,
959 			  " RX queue %d unexpected RX event "
960 			  EFX_QWORD_FMT "%s%s%s%s%s%s%s%s\n",
961 			  efx_rx_queue_index(rx_queue), EFX_QWORD_VAL(*event),
962 			  rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
963 			  rx_ev_ip_hdr_chksum_err ?
964 			  " [IP_HDR_CHKSUM_ERR]" : "",
965 			  rx_ev_tcp_udp_chksum_err ?
966 			  " [TCP_UDP_CHKSUM_ERR]" : "",
967 			  rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
968 			  rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
969 			  rx_ev_drib_nib ? " [DRIB_NIB]" : "",
970 			  rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
971 			  rx_ev_pause_frm ? " [PAUSE]" : "");
972 	}
973 #endif
974 
975 	/* The frame must be discarded if any of these are true. */
976 	return (rx_ev_eth_crc_err | rx_ev_frm_trunc | rx_ev_drib_nib |
977 		rx_ev_tobe_disc | rx_ev_pause_frm) ?
978 		EFX_RX_PKT_DISCARD : 0;
979 }
980 
981 /* Handle receive events that are not in-order. Return true if this
982  * can be handled as a partial packet discard, false if it's more
983  * serious.
984  */
985 static bool
986 efx_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
987 {
988 	struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
989 	struct efx_nic *efx = rx_queue->efx;
990 	unsigned expected, dropped;
991 
992 	if (rx_queue->scatter_n &&
993 	    index == ((rx_queue->removed_count + rx_queue->scatter_n - 1) &
994 		      rx_queue->ptr_mask)) {
995 		++channel->n_rx_nodesc_trunc;
996 		return true;
997 	}
998 
999 	expected = rx_queue->removed_count & rx_queue->ptr_mask;
1000 	dropped = (index - expected) & rx_queue->ptr_mask;
1001 	netif_info(efx, rx_err, efx->net_dev,
1002 		   "dropped %d events (index=%d expected=%d)\n",
1003 		   dropped, index, expected);
1004 
1005 	efx_schedule_reset(efx, EFX_WORKAROUND_5676(efx) ?
1006 			   RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
1007 	return false;
1008 }
1009 
1010 /* Handle a packet received event
1011  *
1012  * The NIC gives a "discard" flag if it's a unicast packet with the
1013  * wrong destination address
1014  * Also "is multicast" and "matches multicast filter" flags can be used to
1015  * discard non-matching multicast packets.
1016  */
1017 static void
1018 efx_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
1019 {
1020 	unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
1021 	unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
1022 	unsigned expected_ptr;
1023 	bool rx_ev_pkt_ok, rx_ev_sop, rx_ev_cont;
1024 	u16 flags;
1025 	struct efx_rx_queue *rx_queue;
1026 	struct efx_nic *efx = channel->efx;
1027 
1028 	if (unlikely(ACCESS_ONCE(efx->reset_pending)))
1029 		return;
1030 
1031 	rx_ev_cont = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT);
1032 	rx_ev_sop = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP);
1033 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
1034 		channel->channel);
1035 
1036 	rx_queue = efx_channel_get_rx_queue(channel);
1037 
1038 	rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
1039 	expected_ptr = ((rx_queue->removed_count + rx_queue->scatter_n) &
1040 			rx_queue->ptr_mask);
1041 
1042 	/* Check for partial drops and other errors */
1043 	if (unlikely(rx_ev_desc_ptr != expected_ptr) ||
1044 	    unlikely(rx_ev_sop != (rx_queue->scatter_n == 0))) {
1045 		if (rx_ev_desc_ptr != expected_ptr &&
1046 		    !efx_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr))
1047 			return;
1048 
1049 		/* Discard all pending fragments */
1050 		if (rx_queue->scatter_n) {
1051 			efx_rx_packet(
1052 				rx_queue,
1053 				rx_queue->removed_count & rx_queue->ptr_mask,
1054 				rx_queue->scatter_n, 0, EFX_RX_PKT_DISCARD);
1055 			rx_queue->removed_count += rx_queue->scatter_n;
1056 			rx_queue->scatter_n = 0;
1057 		}
1058 
1059 		/* Return if there is no new fragment */
1060 		if (rx_ev_desc_ptr != expected_ptr)
1061 			return;
1062 
1063 		/* Discard new fragment if not SOP */
1064 		if (!rx_ev_sop) {
1065 			efx_rx_packet(
1066 				rx_queue,
1067 				rx_queue->removed_count & rx_queue->ptr_mask,
1068 				1, 0, EFX_RX_PKT_DISCARD);
1069 			++rx_queue->removed_count;
1070 			return;
1071 		}
1072 	}
1073 
1074 	++rx_queue->scatter_n;
1075 	if (rx_ev_cont)
1076 		return;
1077 
1078 	rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
1079 	rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
1080 	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
1081 
1082 	if (likely(rx_ev_pkt_ok)) {
1083 		/* If packet is marked as OK then we can rely on the
1084 		 * hardware checksum and classification.
1085 		 */
1086 		flags = 0;
1087 		switch (rx_ev_hdr_type) {
1088 		case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP:
1089 			flags |= EFX_RX_PKT_TCP;
1090 			/* fall through */
1091 		case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP:
1092 			flags |= EFX_RX_PKT_CSUMMED;
1093 			/* fall through */
1094 		case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_OTHER:
1095 		case FSE_AZ_RX_EV_HDR_TYPE_OTHER:
1096 			break;
1097 		}
1098 	} else {
1099 		flags = efx_handle_rx_not_ok(rx_queue, event);
1100 	}
1101 
1102 	/* Detect multicast packets that didn't match the filter */
1103 	rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
1104 	if (rx_ev_mcast_pkt) {
1105 		unsigned int rx_ev_mcast_hash_match =
1106 			EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
1107 
1108 		if (unlikely(!rx_ev_mcast_hash_match)) {
1109 			++channel->n_rx_mcast_mismatch;
1110 			flags |= EFX_RX_PKT_DISCARD;
1111 		}
1112 	}
1113 
1114 	channel->irq_mod_score += 2;
1115 
1116 	/* Handle received packet */
1117 	efx_rx_packet(rx_queue,
1118 		      rx_queue->removed_count & rx_queue->ptr_mask,
1119 		      rx_queue->scatter_n, rx_ev_byte_cnt, flags);
1120 	rx_queue->removed_count += rx_queue->scatter_n;
1121 	rx_queue->scatter_n = 0;
1122 }
1123 
1124 /* If this flush done event corresponds to a &struct efx_tx_queue, then
1125  * send an %EFX_CHANNEL_MAGIC_TX_DRAIN event to drain the event queue
1126  * of all transmit completions.
1127  */
1128 static void
1129 efx_handle_tx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1130 {
1131 	struct efx_tx_queue *tx_queue;
1132 	int qid;
1133 
1134 	qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1135 	if (qid < EFX_TXQ_TYPES * efx->n_tx_channels) {
1136 		tx_queue = efx_get_tx_queue(efx, qid / EFX_TXQ_TYPES,
1137 					    qid % EFX_TXQ_TYPES);
1138 		if (atomic_cmpxchg(&tx_queue->flush_outstanding, 1, 0)) {
1139 			efx_magic_event(tx_queue->channel,
1140 					EFX_CHANNEL_MAGIC_TX_DRAIN(tx_queue));
1141 		}
1142 	}
1143 }
1144 
1145 /* If this flush done event corresponds to a &struct efx_rx_queue: If the flush
1146  * was succesful then send an %EFX_CHANNEL_MAGIC_RX_DRAIN, otherwise add
1147  * the RX queue back to the mask of RX queues in need of flushing.
1148  */
1149 static void
1150 efx_handle_rx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1151 {
1152 	struct efx_channel *channel;
1153 	struct efx_rx_queue *rx_queue;
1154 	int qid;
1155 	bool failed;
1156 
1157 	qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
1158 	failed = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
1159 	if (qid >= efx->n_channels)
1160 		return;
1161 	channel = efx_get_channel(efx, qid);
1162 	if (!efx_channel_has_rx_queue(channel))
1163 		return;
1164 	rx_queue = efx_channel_get_rx_queue(channel);
1165 
1166 	if (failed) {
1167 		netif_info(efx, hw, efx->net_dev,
1168 			   "RXQ %d flush retry\n", qid);
1169 		rx_queue->flush_pending = true;
1170 		atomic_inc(&efx->rxq_flush_pending);
1171 	} else {
1172 		efx_magic_event(efx_rx_queue_channel(rx_queue),
1173 				EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue));
1174 	}
1175 	atomic_dec(&efx->rxq_flush_outstanding);
1176 	if (efx_flush_wake(efx))
1177 		wake_up(&efx->flush_wq);
1178 }
1179 
1180 static void
1181 efx_handle_drain_event(struct efx_channel *channel)
1182 {
1183 	struct efx_nic *efx = channel->efx;
1184 
1185 	WARN_ON(atomic_read(&efx->drain_pending) == 0);
1186 	atomic_dec(&efx->drain_pending);
1187 	if (efx_flush_wake(efx))
1188 		wake_up(&efx->flush_wq);
1189 }
1190 
1191 static void
1192 efx_handle_generated_event(struct efx_channel *channel, efx_qword_t *event)
1193 {
1194 	struct efx_nic *efx = channel->efx;
1195 	struct efx_rx_queue *rx_queue =
1196 		efx_channel_has_rx_queue(channel) ?
1197 		efx_channel_get_rx_queue(channel) : NULL;
1198 	unsigned magic, code;
1199 
1200 	magic = EFX_QWORD_FIELD(*event, FSF_AZ_DRV_GEN_EV_MAGIC);
1201 	code = _EFX_CHANNEL_MAGIC_CODE(magic);
1202 
1203 	if (magic == EFX_CHANNEL_MAGIC_TEST(channel)) {
1204 		channel->event_test_cpu = raw_smp_processor_id();
1205 	} else if (rx_queue && magic == EFX_CHANNEL_MAGIC_FILL(rx_queue)) {
1206 		/* The queue must be empty, so we won't receive any rx
1207 		 * events, so efx_process_channel() won't refill the
1208 		 * queue. Refill it here */
1209 		efx_fast_push_rx_descriptors(rx_queue);
1210 	} else if (rx_queue && magic == EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue)) {
1211 		rx_queue->enabled = false;
1212 		efx_handle_drain_event(channel);
1213 	} else if (code == _EFX_CHANNEL_MAGIC_TX_DRAIN) {
1214 		efx_handle_drain_event(channel);
1215 	} else {
1216 		netif_dbg(efx, hw, efx->net_dev, "channel %d received "
1217 			  "generated event "EFX_QWORD_FMT"\n",
1218 			  channel->channel, EFX_QWORD_VAL(*event));
1219 	}
1220 }
1221 
1222 static void
1223 efx_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
1224 {
1225 	struct efx_nic *efx = channel->efx;
1226 	unsigned int ev_sub_code;
1227 	unsigned int ev_sub_data;
1228 
1229 	ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
1230 	ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1231 
1232 	switch (ev_sub_code) {
1233 	case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
1234 		netif_vdbg(efx, hw, efx->net_dev, "channel %d TXQ %d flushed\n",
1235 			   channel->channel, ev_sub_data);
1236 		efx_handle_tx_flush_done(efx, event);
1237 		efx_sriov_tx_flush_done(efx, event);
1238 		break;
1239 	case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
1240 		netif_vdbg(efx, hw, efx->net_dev, "channel %d RXQ %d flushed\n",
1241 			   channel->channel, ev_sub_data);
1242 		efx_handle_rx_flush_done(efx, event);
1243 		efx_sriov_rx_flush_done(efx, event);
1244 		break;
1245 	case FSE_AZ_EVQ_INIT_DONE_EV:
1246 		netif_dbg(efx, hw, efx->net_dev,
1247 			  "channel %d EVQ %d initialised\n",
1248 			  channel->channel, ev_sub_data);
1249 		break;
1250 	case FSE_AZ_SRM_UPD_DONE_EV:
1251 		netif_vdbg(efx, hw, efx->net_dev,
1252 			   "channel %d SRAM update done\n", channel->channel);
1253 		break;
1254 	case FSE_AZ_WAKE_UP_EV:
1255 		netif_vdbg(efx, hw, efx->net_dev,
1256 			   "channel %d RXQ %d wakeup event\n",
1257 			   channel->channel, ev_sub_data);
1258 		break;
1259 	case FSE_AZ_TIMER_EV:
1260 		netif_vdbg(efx, hw, efx->net_dev,
1261 			   "channel %d RX queue %d timer expired\n",
1262 			   channel->channel, ev_sub_data);
1263 		break;
1264 	case FSE_AA_RX_RECOVER_EV:
1265 		netif_err(efx, rx_err, efx->net_dev,
1266 			  "channel %d seen DRIVER RX_RESET event. "
1267 			"Resetting.\n", channel->channel);
1268 		atomic_inc(&efx->rx_reset);
1269 		efx_schedule_reset(efx,
1270 				   EFX_WORKAROUND_6555(efx) ?
1271 				   RESET_TYPE_RX_RECOVERY :
1272 				   RESET_TYPE_DISABLE);
1273 		break;
1274 	case FSE_BZ_RX_DSC_ERROR_EV:
1275 		if (ev_sub_data < EFX_VI_BASE) {
1276 			netif_err(efx, rx_err, efx->net_dev,
1277 				  "RX DMA Q %d reports descriptor fetch error."
1278 				  " RX Q %d is disabled.\n", ev_sub_data,
1279 				  ev_sub_data);
1280 			efx_schedule_reset(efx, RESET_TYPE_RX_DESC_FETCH);
1281 		} else
1282 			efx_sriov_desc_fetch_err(efx, ev_sub_data);
1283 		break;
1284 	case FSE_BZ_TX_DSC_ERROR_EV:
1285 		if (ev_sub_data < EFX_VI_BASE) {
1286 			netif_err(efx, tx_err, efx->net_dev,
1287 				  "TX DMA Q %d reports descriptor fetch error."
1288 				  " TX Q %d is disabled.\n", ev_sub_data,
1289 				  ev_sub_data);
1290 			efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
1291 		} else
1292 			efx_sriov_desc_fetch_err(efx, ev_sub_data);
1293 		break;
1294 	default:
1295 		netif_vdbg(efx, hw, efx->net_dev,
1296 			   "channel %d unknown driver event code %d "
1297 			   "data %04x\n", channel->channel, ev_sub_code,
1298 			   ev_sub_data);
1299 		break;
1300 	}
1301 }
1302 
1303 int efx_nic_process_eventq(struct efx_channel *channel, int budget)
1304 {
1305 	struct efx_nic *efx = channel->efx;
1306 	unsigned int read_ptr;
1307 	efx_qword_t event, *p_event;
1308 	int ev_code;
1309 	int tx_packets = 0;
1310 	int spent = 0;
1311 
1312 	read_ptr = channel->eventq_read_ptr;
1313 
1314 	for (;;) {
1315 		p_event = efx_event(channel, read_ptr);
1316 		event = *p_event;
1317 
1318 		if (!efx_event_present(&event))
1319 			/* End of events */
1320 			break;
1321 
1322 		netif_vdbg(channel->efx, intr, channel->efx->net_dev,
1323 			   "channel %d event is "EFX_QWORD_FMT"\n",
1324 			   channel->channel, EFX_QWORD_VAL(event));
1325 
1326 		/* Clear this event by marking it all ones */
1327 		EFX_SET_QWORD(*p_event);
1328 
1329 		++read_ptr;
1330 
1331 		ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
1332 
1333 		switch (ev_code) {
1334 		case FSE_AZ_EV_CODE_RX_EV:
1335 			efx_handle_rx_event(channel, &event);
1336 			if (++spent == budget)
1337 				goto out;
1338 			break;
1339 		case FSE_AZ_EV_CODE_TX_EV:
1340 			tx_packets += efx_handle_tx_event(channel, &event);
1341 			if (tx_packets > efx->txq_entries) {
1342 				spent = budget;
1343 				goto out;
1344 			}
1345 			break;
1346 		case FSE_AZ_EV_CODE_DRV_GEN_EV:
1347 			efx_handle_generated_event(channel, &event);
1348 			break;
1349 		case FSE_AZ_EV_CODE_DRIVER_EV:
1350 			efx_handle_driver_event(channel, &event);
1351 			break;
1352 		case FSE_CZ_EV_CODE_USER_EV:
1353 			efx_sriov_event(channel, &event);
1354 			break;
1355 		case FSE_CZ_EV_CODE_MCDI_EV:
1356 			efx_mcdi_process_event(channel, &event);
1357 			break;
1358 		case FSE_AZ_EV_CODE_GLOBAL_EV:
1359 			if (efx->type->handle_global_event &&
1360 			    efx->type->handle_global_event(channel, &event))
1361 				break;
1362 			/* else fall through */
1363 		default:
1364 			netif_err(channel->efx, hw, channel->efx->net_dev,
1365 				  "channel %d unknown event type %d (data "
1366 				  EFX_QWORD_FMT ")\n", channel->channel,
1367 				  ev_code, EFX_QWORD_VAL(event));
1368 		}
1369 	}
1370 
1371 out:
1372 	channel->eventq_read_ptr = read_ptr;
1373 	return spent;
1374 }
1375 
1376 /* Check whether an event is present in the eventq at the current
1377  * read pointer.  Only useful for self-test.
1378  */
1379 bool efx_nic_event_present(struct efx_channel *channel)
1380 {
1381 	return efx_event_present(efx_event(channel, channel->eventq_read_ptr));
1382 }
1383 
1384 /* Allocate buffer table entries for event queue */
1385 int efx_nic_probe_eventq(struct efx_channel *channel)
1386 {
1387 	struct efx_nic *efx = channel->efx;
1388 	unsigned entries;
1389 
1390 	entries = channel->eventq_mask + 1;
1391 	return efx_alloc_special_buffer(efx, &channel->eventq,
1392 					entries * sizeof(efx_qword_t));
1393 }
1394 
1395 void efx_nic_init_eventq(struct efx_channel *channel)
1396 {
1397 	efx_oword_t reg;
1398 	struct efx_nic *efx = channel->efx;
1399 
1400 	netif_dbg(efx, hw, efx->net_dev,
1401 		  "channel %d event queue in special buffers %d-%d\n",
1402 		  channel->channel, channel->eventq.index,
1403 		  channel->eventq.index + channel->eventq.entries - 1);
1404 
1405 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1406 		EFX_POPULATE_OWORD_3(reg,
1407 				     FRF_CZ_TIMER_Q_EN, 1,
1408 				     FRF_CZ_HOST_NOTIFY_MODE, 0,
1409 				     FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
1410 		efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1411 	}
1412 
1413 	/* Pin event queue buffer */
1414 	efx_init_special_buffer(efx, &channel->eventq);
1415 
1416 	/* Fill event queue with all ones (i.e. empty events) */
1417 	memset(channel->eventq.addr, 0xff, channel->eventq.len);
1418 
1419 	/* Push event queue to card */
1420 	EFX_POPULATE_OWORD_3(reg,
1421 			     FRF_AZ_EVQ_EN, 1,
1422 			     FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
1423 			     FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
1424 	efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1425 			 channel->channel);
1426 
1427 	efx->type->push_irq_moderation(channel);
1428 }
1429 
1430 void efx_nic_fini_eventq(struct efx_channel *channel)
1431 {
1432 	efx_oword_t reg;
1433 	struct efx_nic *efx = channel->efx;
1434 
1435 	/* Remove event queue from card */
1436 	EFX_ZERO_OWORD(reg);
1437 	efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1438 			 channel->channel);
1439 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1440 		efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1441 
1442 	/* Unpin event queue */
1443 	efx_fini_special_buffer(efx, &channel->eventq);
1444 }
1445 
1446 /* Free buffers backing event queue */
1447 void efx_nic_remove_eventq(struct efx_channel *channel)
1448 {
1449 	efx_free_special_buffer(channel->efx, &channel->eventq);
1450 }
1451 
1452 
1453 void efx_nic_event_test_start(struct efx_channel *channel)
1454 {
1455 	channel->event_test_cpu = -1;
1456 	smp_wmb();
1457 	efx_magic_event(channel, EFX_CHANNEL_MAGIC_TEST(channel));
1458 }
1459 
1460 void efx_nic_generate_fill_event(struct efx_rx_queue *rx_queue)
1461 {
1462 	efx_magic_event(efx_rx_queue_channel(rx_queue),
1463 			EFX_CHANNEL_MAGIC_FILL(rx_queue));
1464 }
1465 
1466 /**************************************************************************
1467  *
1468  * Hardware interrupts
1469  * The hardware interrupt handler does very little work; all the event
1470  * queue processing is carried out by per-channel tasklets.
1471  *
1472  **************************************************************************/
1473 
1474 /* Enable/disable/generate interrupts */
1475 static inline void efx_nic_interrupts(struct efx_nic *efx,
1476 				      bool enabled, bool force)
1477 {
1478 	efx_oword_t int_en_reg_ker;
1479 
1480 	EFX_POPULATE_OWORD_3(int_en_reg_ker,
1481 			     FRF_AZ_KER_INT_LEVE_SEL, efx->irq_level,
1482 			     FRF_AZ_KER_INT_KER, force,
1483 			     FRF_AZ_DRV_INT_EN_KER, enabled);
1484 	efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
1485 }
1486 
1487 void efx_nic_enable_interrupts(struct efx_nic *efx)
1488 {
1489 	EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
1490 	wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
1491 
1492 	efx_nic_interrupts(efx, true, false);
1493 }
1494 
1495 void efx_nic_disable_interrupts(struct efx_nic *efx)
1496 {
1497 	/* Disable interrupts */
1498 	efx_nic_interrupts(efx, false, false);
1499 }
1500 
1501 /* Generate a test interrupt
1502  * Interrupt must already have been enabled, otherwise nasty things
1503  * may happen.
1504  */
1505 void efx_nic_irq_test_start(struct efx_nic *efx)
1506 {
1507 	efx->last_irq_cpu = -1;
1508 	smp_wmb();
1509 	efx_nic_interrupts(efx, true, true);
1510 }
1511 
1512 /* Process a fatal interrupt
1513  * Disable bus mastering ASAP and schedule a reset
1514  */
1515 irqreturn_t efx_nic_fatal_interrupt(struct efx_nic *efx)
1516 {
1517 	struct falcon_nic_data *nic_data = efx->nic_data;
1518 	efx_oword_t *int_ker = efx->irq_status.addr;
1519 	efx_oword_t fatal_intr;
1520 	int error, mem_perr;
1521 
1522 	efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
1523 	error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
1524 
1525 	netif_err(efx, hw, efx->net_dev, "SYSTEM ERROR "EFX_OWORD_FMT" status "
1526 		  EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
1527 		  EFX_OWORD_VAL(fatal_intr),
1528 		  error ? "disabling bus mastering" : "no recognised error");
1529 
1530 	/* If this is a memory parity error dump which blocks are offending */
1531 	mem_perr = (EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER) ||
1532 		    EFX_OWORD_FIELD(fatal_intr, FRF_AZ_SRM_PERR_INT_KER));
1533 	if (mem_perr) {
1534 		efx_oword_t reg;
1535 		efx_reado(efx, &reg, FR_AZ_MEM_STAT);
1536 		netif_err(efx, hw, efx->net_dev,
1537 			  "SYSTEM ERROR: memory parity error "EFX_OWORD_FMT"\n",
1538 			  EFX_OWORD_VAL(reg));
1539 	}
1540 
1541 	/* Disable both devices */
1542 	pci_clear_master(efx->pci_dev);
1543 	if (efx_nic_is_dual_func(efx))
1544 		pci_clear_master(nic_data->pci_dev2);
1545 	efx_nic_disable_interrupts(efx);
1546 
1547 	/* Count errors and reset or disable the NIC accordingly */
1548 	if (efx->int_error_count == 0 ||
1549 	    time_after(jiffies, efx->int_error_expire)) {
1550 		efx->int_error_count = 0;
1551 		efx->int_error_expire =
1552 			jiffies + EFX_INT_ERROR_EXPIRE * HZ;
1553 	}
1554 	if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
1555 		netif_err(efx, hw, efx->net_dev,
1556 			  "SYSTEM ERROR - reset scheduled\n");
1557 		efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
1558 	} else {
1559 		netif_err(efx, hw, efx->net_dev,
1560 			  "SYSTEM ERROR - max number of errors seen."
1561 			  "NIC will be disabled\n");
1562 		efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1563 	}
1564 
1565 	return IRQ_HANDLED;
1566 }
1567 
1568 /* Handle a legacy interrupt
1569  * Acknowledges the interrupt and schedule event queue processing.
1570  */
1571 static irqreturn_t efx_legacy_interrupt(int irq, void *dev_id)
1572 {
1573 	struct efx_nic *efx = dev_id;
1574 	efx_oword_t *int_ker = efx->irq_status.addr;
1575 	irqreturn_t result = IRQ_NONE;
1576 	struct efx_channel *channel;
1577 	efx_dword_t reg;
1578 	u32 queues;
1579 	int syserr;
1580 
1581 	/* Could this be ours?  If interrupts are disabled then the
1582 	 * channel state may not be valid.
1583 	 */
1584 	if (!efx->legacy_irq_enabled)
1585 		return result;
1586 
1587 	/* Read the ISR which also ACKs the interrupts */
1588 	efx_readd(efx, &reg, FR_BZ_INT_ISR0);
1589 	queues = EFX_EXTRACT_DWORD(reg, 0, 31);
1590 
1591 	/* Legacy interrupts are disabled too late by the EEH kernel
1592 	 * code. Disable them earlier.
1593 	 * If an EEH error occurred, the read will have returned all ones.
1594 	 */
1595 	if (EFX_DWORD_IS_ALL_ONES(reg) && efx_try_recovery(efx) &&
1596 	    !efx->eeh_disabled_legacy_irq) {
1597 		disable_irq_nosync(efx->legacy_irq);
1598 		efx->eeh_disabled_legacy_irq = true;
1599 	}
1600 
1601 	/* Handle non-event-queue sources */
1602 	if (queues & (1U << efx->irq_level)) {
1603 		syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1604 		if (unlikely(syserr))
1605 			return efx_nic_fatal_interrupt(efx);
1606 		efx->last_irq_cpu = raw_smp_processor_id();
1607 	}
1608 
1609 	if (queues != 0) {
1610 		if (EFX_WORKAROUND_15783(efx))
1611 			efx->irq_zero_count = 0;
1612 
1613 		/* Schedule processing of any interrupting queues */
1614 		efx_for_each_channel(channel, efx) {
1615 			if (queues & 1)
1616 				efx_schedule_channel_irq(channel);
1617 			queues >>= 1;
1618 		}
1619 		result = IRQ_HANDLED;
1620 
1621 	} else if (EFX_WORKAROUND_15783(efx)) {
1622 		efx_qword_t *event;
1623 
1624 		/* We can't return IRQ_HANDLED more than once on seeing ISR=0
1625 		 * because this might be a shared interrupt. */
1626 		if (efx->irq_zero_count++ == 0)
1627 			result = IRQ_HANDLED;
1628 
1629 		/* Ensure we schedule or rearm all event queues */
1630 		efx_for_each_channel(channel, efx) {
1631 			event = efx_event(channel, channel->eventq_read_ptr);
1632 			if (efx_event_present(event))
1633 				efx_schedule_channel_irq(channel);
1634 			else
1635 				efx_nic_eventq_read_ack(channel);
1636 		}
1637 	}
1638 
1639 	if (result == IRQ_HANDLED)
1640 		netif_vdbg(efx, intr, efx->net_dev,
1641 			   "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
1642 			   irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
1643 
1644 	return result;
1645 }
1646 
1647 /* Handle an MSI interrupt
1648  *
1649  * Handle an MSI hardware interrupt.  This routine schedules event
1650  * queue processing.  No interrupt acknowledgement cycle is necessary.
1651  * Also, we never need to check that the interrupt is for us, since
1652  * MSI interrupts cannot be shared.
1653  */
1654 static irqreturn_t efx_msi_interrupt(int irq, void *dev_id)
1655 {
1656 	struct efx_channel *channel = *(struct efx_channel **)dev_id;
1657 	struct efx_nic *efx = channel->efx;
1658 	efx_oword_t *int_ker = efx->irq_status.addr;
1659 	int syserr;
1660 
1661 	netif_vdbg(efx, intr, efx->net_dev,
1662 		   "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
1663 		   irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
1664 
1665 	/* Handle non-event-queue sources */
1666 	if (channel->channel == efx->irq_level) {
1667 		syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1668 		if (unlikely(syserr))
1669 			return efx_nic_fatal_interrupt(efx);
1670 		efx->last_irq_cpu = raw_smp_processor_id();
1671 	}
1672 
1673 	/* Schedule processing of the channel */
1674 	efx_schedule_channel_irq(channel);
1675 
1676 	return IRQ_HANDLED;
1677 }
1678 
1679 
1680 /* Setup RSS indirection table.
1681  * This maps from the hash value of the packet to RXQ
1682  */
1683 void efx_nic_push_rx_indir_table(struct efx_nic *efx)
1684 {
1685 	size_t i = 0;
1686 	efx_dword_t dword;
1687 
1688 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0)
1689 		return;
1690 
1691 	BUILD_BUG_ON(ARRAY_SIZE(efx->rx_indir_table) !=
1692 		     FR_BZ_RX_INDIRECTION_TBL_ROWS);
1693 
1694 	for (i = 0; i < FR_BZ_RX_INDIRECTION_TBL_ROWS; i++) {
1695 		EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
1696 				     efx->rx_indir_table[i]);
1697 		efx_writed(efx, &dword,
1698 			   FR_BZ_RX_INDIRECTION_TBL +
1699 			   FR_BZ_RX_INDIRECTION_TBL_STEP * i);
1700 	}
1701 }
1702 
1703 /* Hook interrupt handler(s)
1704  * Try MSI and then legacy interrupts.
1705  */
1706 int efx_nic_init_interrupt(struct efx_nic *efx)
1707 {
1708 	struct efx_channel *channel;
1709 	int rc;
1710 
1711 	if (!EFX_INT_MODE_USE_MSI(efx)) {
1712 		irq_handler_t handler;
1713 		if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1714 			handler = efx_legacy_interrupt;
1715 		else
1716 			handler = falcon_legacy_interrupt_a1;
1717 
1718 		rc = request_irq(efx->legacy_irq, handler, IRQF_SHARED,
1719 				 efx->name, efx);
1720 		if (rc) {
1721 			netif_err(efx, drv, efx->net_dev,
1722 				  "failed to hook legacy IRQ %d\n",
1723 				  efx->pci_dev->irq);
1724 			goto fail1;
1725 		}
1726 		return 0;
1727 	}
1728 
1729 	/* Hook MSI or MSI-X interrupt */
1730 	efx_for_each_channel(channel, efx) {
1731 		rc = request_irq(channel->irq, efx_msi_interrupt,
1732 				 IRQF_PROBE_SHARED, /* Not shared */
1733 				 efx->channel_name[channel->channel],
1734 				 &efx->channel[channel->channel]);
1735 		if (rc) {
1736 			netif_err(efx, drv, efx->net_dev,
1737 				  "failed to hook IRQ %d\n", channel->irq);
1738 			goto fail2;
1739 		}
1740 	}
1741 
1742 	return 0;
1743 
1744  fail2:
1745 	efx_for_each_channel(channel, efx)
1746 		free_irq(channel->irq, &efx->channel[channel->channel]);
1747  fail1:
1748 	return rc;
1749 }
1750 
1751 void efx_nic_fini_interrupt(struct efx_nic *efx)
1752 {
1753 	struct efx_channel *channel;
1754 	efx_oword_t reg;
1755 
1756 	/* Disable MSI/MSI-X interrupts */
1757 	efx_for_each_channel(channel, efx) {
1758 		if (channel->irq)
1759 			free_irq(channel->irq, &efx->channel[channel->channel]);
1760 	}
1761 
1762 	/* ACK legacy interrupt */
1763 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1764 		efx_reado(efx, &reg, FR_BZ_INT_ISR0);
1765 	else
1766 		falcon_irq_ack_a1(efx);
1767 
1768 	/* Disable legacy interrupt */
1769 	if (efx->legacy_irq)
1770 		free_irq(efx->legacy_irq, efx);
1771 }
1772 
1773 /* Looks at available SRAM resources and works out how many queues we
1774  * can support, and where things like descriptor caches should live.
1775  *
1776  * SRAM is split up as follows:
1777  * 0                          buftbl entries for channels
1778  * efx->vf_buftbl_base        buftbl entries for SR-IOV
1779  * efx->rx_dc_base            RX descriptor caches
1780  * efx->tx_dc_base            TX descriptor caches
1781  */
1782 void efx_nic_dimension_resources(struct efx_nic *efx, unsigned sram_lim_qw)
1783 {
1784 	unsigned vi_count, buftbl_min;
1785 
1786 	/* Account for the buffer table entries backing the datapath channels
1787 	 * and the descriptor caches for those channels.
1788 	 */
1789 	buftbl_min = ((efx->n_rx_channels * EFX_MAX_DMAQ_SIZE +
1790 		       efx->n_tx_channels * EFX_TXQ_TYPES * EFX_MAX_DMAQ_SIZE +
1791 		       efx->n_channels * EFX_MAX_EVQ_SIZE)
1792 		      * sizeof(efx_qword_t) / EFX_BUF_SIZE);
1793 	vi_count = max(efx->n_channels, efx->n_tx_channels * EFX_TXQ_TYPES);
1794 
1795 #ifdef CONFIG_SFC_SRIOV
1796 	if (efx_sriov_wanted(efx)) {
1797 		unsigned vi_dc_entries, buftbl_free, entries_per_vf, vf_limit;
1798 
1799 		efx->vf_buftbl_base = buftbl_min;
1800 
1801 		vi_dc_entries = RX_DC_ENTRIES + TX_DC_ENTRIES;
1802 		vi_count = max(vi_count, EFX_VI_BASE);
1803 		buftbl_free = (sram_lim_qw - buftbl_min -
1804 			       vi_count * vi_dc_entries);
1805 
1806 		entries_per_vf = ((vi_dc_entries + EFX_VF_BUFTBL_PER_VI) *
1807 				  efx_vf_size(efx));
1808 		vf_limit = min(buftbl_free / entries_per_vf,
1809 			       (1024U - EFX_VI_BASE) >> efx->vi_scale);
1810 
1811 		if (efx->vf_count > vf_limit) {
1812 			netif_err(efx, probe, efx->net_dev,
1813 				  "Reducing VF count from from %d to %d\n",
1814 				  efx->vf_count, vf_limit);
1815 			efx->vf_count = vf_limit;
1816 		}
1817 		vi_count += efx->vf_count * efx_vf_size(efx);
1818 	}
1819 #endif
1820 
1821 	efx->tx_dc_base = sram_lim_qw - vi_count * TX_DC_ENTRIES;
1822 	efx->rx_dc_base = efx->tx_dc_base - vi_count * RX_DC_ENTRIES;
1823 }
1824 
1825 u32 efx_nic_fpga_ver(struct efx_nic *efx)
1826 {
1827 	efx_oword_t altera_build;
1828 	efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
1829 	return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
1830 }
1831 
1832 void efx_nic_init_common(struct efx_nic *efx)
1833 {
1834 	efx_oword_t temp;
1835 
1836 	/* Set positions of descriptor caches in SRAM. */
1837 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR, efx->tx_dc_base);
1838 	efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
1839 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR, efx->rx_dc_base);
1840 	efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
1841 
1842 	/* Set TX descriptor cache size. */
1843 	BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
1844 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
1845 	efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
1846 
1847 	/* Set RX descriptor cache size.  Set low watermark to size-8, as
1848 	 * this allows most efficient prefetching.
1849 	 */
1850 	BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
1851 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
1852 	efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
1853 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
1854 	efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
1855 
1856 	/* Program INT_KER address */
1857 	EFX_POPULATE_OWORD_2(temp,
1858 			     FRF_AZ_NORM_INT_VEC_DIS_KER,
1859 			     EFX_INT_MODE_USE_MSI(efx),
1860 			     FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
1861 	efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
1862 
1863 	if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
1864 		/* Use an interrupt level unused by event queues */
1865 		efx->irq_level = 0x1f;
1866 	else
1867 		/* Use a valid MSI-X vector */
1868 		efx->irq_level = 0;
1869 
1870 	/* Enable all the genuinely fatal interrupts.  (They are still
1871 	 * masked by the overall interrupt mask, controlled by
1872 	 * falcon_interrupts()).
1873 	 *
1874 	 * Note: All other fatal interrupts are enabled
1875 	 */
1876 	EFX_POPULATE_OWORD_3(temp,
1877 			     FRF_AZ_ILL_ADR_INT_KER_EN, 1,
1878 			     FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
1879 			     FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
1880 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1881 		EFX_SET_OWORD_FIELD(temp, FRF_CZ_SRAM_PERR_INT_P_KER_EN, 1);
1882 	EFX_INVERT_OWORD(temp);
1883 	efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
1884 
1885 	efx_nic_push_rx_indir_table(efx);
1886 
1887 	/* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
1888 	 * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
1889 	 */
1890 	efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
1891 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
1892 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
1893 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
1894 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 1);
1895 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
1896 	/* Enable SW_EV to inherit in char driver - assume harmless here */
1897 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
1898 	/* Prefetch threshold 2 => fetch when descriptor cache half empty */
1899 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
1900 	/* Disable hardware watchdog which can misfire */
1901 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
1902 	/* Squash TX of packets of 16 bytes or less */
1903 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1904 		EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
1905 	efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
1906 
1907 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
1908 		EFX_POPULATE_OWORD_4(temp,
1909 				     /* Default values */
1910 				     FRF_BZ_TX_PACE_SB_NOT_AF, 0x15,
1911 				     FRF_BZ_TX_PACE_SB_AF, 0xb,
1912 				     FRF_BZ_TX_PACE_FB_BASE, 0,
1913 				     /* Allow large pace values in the
1914 				      * fast bin. */
1915 				     FRF_BZ_TX_PACE_BIN_TH,
1916 				     FFE_BZ_TX_PACE_RESERVED);
1917 		efx_writeo(efx, &temp, FR_BZ_TX_PACE);
1918 	}
1919 }
1920 
1921 /* Register dump */
1922 
1923 #define REGISTER_REVISION_A	1
1924 #define REGISTER_REVISION_B	2
1925 #define REGISTER_REVISION_C	3
1926 #define REGISTER_REVISION_Z	3	/* latest revision */
1927 
1928 struct efx_nic_reg {
1929 	u32 offset:24;
1930 	u32 min_revision:2, max_revision:2;
1931 };
1932 
1933 #define REGISTER(name, min_rev, max_rev) {				\
1934 	FR_ ## min_rev ## max_rev ## _ ## name,				\
1935 	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev	\
1936 }
1937 #define REGISTER_AA(name) REGISTER(name, A, A)
1938 #define REGISTER_AB(name) REGISTER(name, A, B)
1939 #define REGISTER_AZ(name) REGISTER(name, A, Z)
1940 #define REGISTER_BB(name) REGISTER(name, B, B)
1941 #define REGISTER_BZ(name) REGISTER(name, B, Z)
1942 #define REGISTER_CZ(name) REGISTER(name, C, Z)
1943 
1944 static const struct efx_nic_reg efx_nic_regs[] = {
1945 	REGISTER_AZ(ADR_REGION),
1946 	REGISTER_AZ(INT_EN_KER),
1947 	REGISTER_BZ(INT_EN_CHAR),
1948 	REGISTER_AZ(INT_ADR_KER),
1949 	REGISTER_BZ(INT_ADR_CHAR),
1950 	/* INT_ACK_KER is WO */
1951 	/* INT_ISR0 is RC */
1952 	REGISTER_AZ(HW_INIT),
1953 	REGISTER_CZ(USR_EV_CFG),
1954 	REGISTER_AB(EE_SPI_HCMD),
1955 	REGISTER_AB(EE_SPI_HADR),
1956 	REGISTER_AB(EE_SPI_HDATA),
1957 	REGISTER_AB(EE_BASE_PAGE),
1958 	REGISTER_AB(EE_VPD_CFG0),
1959 	/* EE_VPD_SW_CNTL and EE_VPD_SW_DATA are not used */
1960 	/* PMBX_DBG_IADDR and PBMX_DBG_IDATA are indirect */
1961 	/* PCIE_CORE_INDIRECT is indirect */
1962 	REGISTER_AB(NIC_STAT),
1963 	REGISTER_AB(GPIO_CTL),
1964 	REGISTER_AB(GLB_CTL),
1965 	/* FATAL_INTR_KER and FATAL_INTR_CHAR are partly RC */
1966 	REGISTER_BZ(DP_CTRL),
1967 	REGISTER_AZ(MEM_STAT),
1968 	REGISTER_AZ(CS_DEBUG),
1969 	REGISTER_AZ(ALTERA_BUILD),
1970 	REGISTER_AZ(CSR_SPARE),
1971 	REGISTER_AB(PCIE_SD_CTL0123),
1972 	REGISTER_AB(PCIE_SD_CTL45),
1973 	REGISTER_AB(PCIE_PCS_CTL_STAT),
1974 	/* DEBUG_DATA_OUT is not used */
1975 	/* DRV_EV is WO */
1976 	REGISTER_AZ(EVQ_CTL),
1977 	REGISTER_AZ(EVQ_CNT1),
1978 	REGISTER_AZ(EVQ_CNT2),
1979 	REGISTER_AZ(BUF_TBL_CFG),
1980 	REGISTER_AZ(SRM_RX_DC_CFG),
1981 	REGISTER_AZ(SRM_TX_DC_CFG),
1982 	REGISTER_AZ(SRM_CFG),
1983 	/* BUF_TBL_UPD is WO */
1984 	REGISTER_AZ(SRM_UPD_EVQ),
1985 	REGISTER_AZ(SRAM_PARITY),
1986 	REGISTER_AZ(RX_CFG),
1987 	REGISTER_BZ(RX_FILTER_CTL),
1988 	/* RX_FLUSH_DESCQ is WO */
1989 	REGISTER_AZ(RX_DC_CFG),
1990 	REGISTER_AZ(RX_DC_PF_WM),
1991 	REGISTER_BZ(RX_RSS_TKEY),
1992 	/* RX_NODESC_DROP is RC */
1993 	REGISTER_AA(RX_SELF_RST),
1994 	/* RX_DEBUG, RX_PUSH_DROP are not used */
1995 	REGISTER_CZ(RX_RSS_IPV6_REG1),
1996 	REGISTER_CZ(RX_RSS_IPV6_REG2),
1997 	REGISTER_CZ(RX_RSS_IPV6_REG3),
1998 	/* TX_FLUSH_DESCQ is WO */
1999 	REGISTER_AZ(TX_DC_CFG),
2000 	REGISTER_AA(TX_CHKSM_CFG),
2001 	REGISTER_AZ(TX_CFG),
2002 	/* TX_PUSH_DROP is not used */
2003 	REGISTER_AZ(TX_RESERVED),
2004 	REGISTER_BZ(TX_PACE),
2005 	/* TX_PACE_DROP_QID is RC */
2006 	REGISTER_BB(TX_VLAN),
2007 	REGISTER_BZ(TX_IPFIL_PORTEN),
2008 	REGISTER_AB(MD_TXD),
2009 	REGISTER_AB(MD_RXD),
2010 	REGISTER_AB(MD_CS),
2011 	REGISTER_AB(MD_PHY_ADR),
2012 	REGISTER_AB(MD_ID),
2013 	/* MD_STAT is RC */
2014 	REGISTER_AB(MAC_STAT_DMA),
2015 	REGISTER_AB(MAC_CTRL),
2016 	REGISTER_BB(GEN_MODE),
2017 	REGISTER_AB(MAC_MC_HASH_REG0),
2018 	REGISTER_AB(MAC_MC_HASH_REG1),
2019 	REGISTER_AB(GM_CFG1),
2020 	REGISTER_AB(GM_CFG2),
2021 	/* GM_IPG and GM_HD are not used */
2022 	REGISTER_AB(GM_MAX_FLEN),
2023 	/* GM_TEST is not used */
2024 	REGISTER_AB(GM_ADR1),
2025 	REGISTER_AB(GM_ADR2),
2026 	REGISTER_AB(GMF_CFG0),
2027 	REGISTER_AB(GMF_CFG1),
2028 	REGISTER_AB(GMF_CFG2),
2029 	REGISTER_AB(GMF_CFG3),
2030 	REGISTER_AB(GMF_CFG4),
2031 	REGISTER_AB(GMF_CFG5),
2032 	REGISTER_BB(TX_SRC_MAC_CTL),
2033 	REGISTER_AB(XM_ADR_LO),
2034 	REGISTER_AB(XM_ADR_HI),
2035 	REGISTER_AB(XM_GLB_CFG),
2036 	REGISTER_AB(XM_TX_CFG),
2037 	REGISTER_AB(XM_RX_CFG),
2038 	REGISTER_AB(XM_MGT_INT_MASK),
2039 	REGISTER_AB(XM_FC),
2040 	REGISTER_AB(XM_PAUSE_TIME),
2041 	REGISTER_AB(XM_TX_PARAM),
2042 	REGISTER_AB(XM_RX_PARAM),
2043 	/* XM_MGT_INT_MSK (note no 'A') is RC */
2044 	REGISTER_AB(XX_PWR_RST),
2045 	REGISTER_AB(XX_SD_CTL),
2046 	REGISTER_AB(XX_TXDRV_CTL),
2047 	/* XX_PRBS_CTL, XX_PRBS_CHK and XX_PRBS_ERR are not used */
2048 	/* XX_CORE_STAT is partly RC */
2049 };
2050 
2051 struct efx_nic_reg_table {
2052 	u32 offset:24;
2053 	u32 min_revision:2, max_revision:2;
2054 	u32 step:6, rows:21;
2055 };
2056 
2057 #define REGISTER_TABLE_DIMENSIONS(_, offset, min_rev, max_rev, step, rows) { \
2058 	offset,								\
2059 	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev,	\
2060 	step, rows							\
2061 }
2062 #define REGISTER_TABLE(name, min_rev, max_rev)				\
2063 	REGISTER_TABLE_DIMENSIONS(					\
2064 		name, FR_ ## min_rev ## max_rev ## _ ## name,		\
2065 		min_rev, max_rev,					\
2066 		FR_ ## min_rev ## max_rev ## _ ## name ## _STEP,	\
2067 		FR_ ## min_rev ## max_rev ## _ ## name ## _ROWS)
2068 #define REGISTER_TABLE_AA(name) REGISTER_TABLE(name, A, A)
2069 #define REGISTER_TABLE_AZ(name) REGISTER_TABLE(name, A, Z)
2070 #define REGISTER_TABLE_BB(name) REGISTER_TABLE(name, B, B)
2071 #define REGISTER_TABLE_BZ(name) REGISTER_TABLE(name, B, Z)
2072 #define REGISTER_TABLE_BB_CZ(name)					\
2073 	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, B, B,		\
2074 				  FR_BZ_ ## name ## _STEP,		\
2075 				  FR_BB_ ## name ## _ROWS),		\
2076 	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, C, Z,		\
2077 				  FR_BZ_ ## name ## _STEP,		\
2078 				  FR_CZ_ ## name ## _ROWS)
2079 #define REGISTER_TABLE_CZ(name) REGISTER_TABLE(name, C, Z)
2080 
2081 static const struct efx_nic_reg_table efx_nic_reg_tables[] = {
2082 	/* DRIVER is not used */
2083 	/* EVQ_RPTR, TIMER_COMMAND, USR_EV and {RX,TX}_DESC_UPD are WO */
2084 	REGISTER_TABLE_BB(TX_IPFIL_TBL),
2085 	REGISTER_TABLE_BB(TX_SRC_MAC_TBL),
2086 	REGISTER_TABLE_AA(RX_DESC_PTR_TBL_KER),
2087 	REGISTER_TABLE_BB_CZ(RX_DESC_PTR_TBL),
2088 	REGISTER_TABLE_AA(TX_DESC_PTR_TBL_KER),
2089 	REGISTER_TABLE_BB_CZ(TX_DESC_PTR_TBL),
2090 	REGISTER_TABLE_AA(EVQ_PTR_TBL_KER),
2091 	REGISTER_TABLE_BB_CZ(EVQ_PTR_TBL),
2092 	/* We can't reasonably read all of the buffer table (up to 8MB!).
2093 	 * However this driver will only use a few entries.  Reading
2094 	 * 1K entries allows for some expansion of queue count and
2095 	 * size before we need to change the version. */
2096 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL_KER, FR_AA_BUF_FULL_TBL_KER,
2097 				  A, A, 8, 1024),
2098 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL, FR_BZ_BUF_FULL_TBL,
2099 				  B, Z, 8, 1024),
2100 	REGISTER_TABLE_CZ(RX_MAC_FILTER_TBL0),
2101 	REGISTER_TABLE_BB_CZ(TIMER_TBL),
2102 	REGISTER_TABLE_BB_CZ(TX_PACE_TBL),
2103 	REGISTER_TABLE_BZ(RX_INDIRECTION_TBL),
2104 	/* TX_FILTER_TBL0 is huge and not used by this driver */
2105 	REGISTER_TABLE_CZ(TX_MAC_FILTER_TBL0),
2106 	REGISTER_TABLE_CZ(MC_TREG_SMEM),
2107 	/* MSIX_PBA_TABLE is not mapped */
2108 	/* SRM_DBG is not mapped (and is redundant with BUF_FLL_TBL) */
2109 	REGISTER_TABLE_BZ(RX_FILTER_TBL0),
2110 };
2111 
2112 size_t efx_nic_get_regs_len(struct efx_nic *efx)
2113 {
2114 	const struct efx_nic_reg *reg;
2115 	const struct efx_nic_reg_table *table;
2116 	size_t len = 0;
2117 
2118 	for (reg = efx_nic_regs;
2119 	     reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
2120 	     reg++)
2121 		if (efx->type->revision >= reg->min_revision &&
2122 		    efx->type->revision <= reg->max_revision)
2123 			len += sizeof(efx_oword_t);
2124 
2125 	for (table = efx_nic_reg_tables;
2126 	     table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
2127 	     table++)
2128 		if (efx->type->revision >= table->min_revision &&
2129 		    efx->type->revision <= table->max_revision)
2130 			len += table->rows * min_t(size_t, table->step, 16);
2131 
2132 	return len;
2133 }
2134 
2135 void efx_nic_get_regs(struct efx_nic *efx, void *buf)
2136 {
2137 	const struct efx_nic_reg *reg;
2138 	const struct efx_nic_reg_table *table;
2139 
2140 	for (reg = efx_nic_regs;
2141 	     reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
2142 	     reg++) {
2143 		if (efx->type->revision >= reg->min_revision &&
2144 		    efx->type->revision <= reg->max_revision) {
2145 			efx_reado(efx, (efx_oword_t *)buf, reg->offset);
2146 			buf += sizeof(efx_oword_t);
2147 		}
2148 	}
2149 
2150 	for (table = efx_nic_reg_tables;
2151 	     table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
2152 	     table++) {
2153 		size_t size, i;
2154 
2155 		if (!(efx->type->revision >= table->min_revision &&
2156 		      efx->type->revision <= table->max_revision))
2157 			continue;
2158 
2159 		size = min_t(size_t, table->step, 16);
2160 
2161 		for (i = 0; i < table->rows; i++) {
2162 			switch (table->step) {
2163 			case 4: /* 32-bit SRAM */
2164 				efx_readd(efx, buf, table->offset + 4 * i);
2165 				break;
2166 			case 8: /* 64-bit SRAM */
2167 				efx_sram_readq(efx,
2168 					       efx->membase + table->offset,
2169 					       buf, i);
2170 				break;
2171 			case 16: /* 128-bit-readable register */
2172 				efx_reado_table(efx, buf, table->offset, i);
2173 				break;
2174 			case 32: /* 128-bit register, interleaved */
2175 				efx_reado_table(efx, buf, table->offset, 2 * i);
2176 				break;
2177 			default:
2178 				WARN_ON(1);
2179 				return;
2180 			}
2181 			buf += size;
2182 		}
2183 	}
2184 }
2185