xref: /openbmc/linux/drivers/net/ethernet/sfc/nic.c (revision 85740cdf)
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, GFP_ATOMIC);
309 	if (!buffer->addr)
310 		return -ENOMEM;
311 	buffer->len = len;
312 	memset(buffer->addr, 0, 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 }
381 
382 /* For each entry inserted into the software descriptor ring, create a
383  * descriptor in the hardware TX descriptor ring (in host memory), and
384  * write a doorbell.
385  */
386 void efx_nic_push_buffers(struct efx_tx_queue *tx_queue)
387 {
388 
389 	struct efx_tx_buffer *buffer;
390 	efx_qword_t *txd;
391 	unsigned write_ptr;
392 	unsigned old_write_count = tx_queue->write_count;
393 
394 	BUG_ON(tx_queue->write_count == tx_queue->insert_count);
395 
396 	do {
397 		write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
398 		buffer = &tx_queue->buffer[write_ptr];
399 		txd = efx_tx_desc(tx_queue, write_ptr);
400 		++tx_queue->write_count;
401 
402 		/* Create TX descriptor ring entry */
403 		BUILD_BUG_ON(EFX_TX_BUF_CONT != 1);
404 		EFX_POPULATE_QWORD_4(*txd,
405 				     FSF_AZ_TX_KER_CONT,
406 				     buffer->flags & EFX_TX_BUF_CONT,
407 				     FSF_AZ_TX_KER_BYTE_COUNT, buffer->len,
408 				     FSF_AZ_TX_KER_BUF_REGION, 0,
409 				     FSF_AZ_TX_KER_BUF_ADDR, buffer->dma_addr);
410 	} while (tx_queue->write_count != tx_queue->insert_count);
411 
412 	wmb(); /* Ensure descriptors are written before they are fetched */
413 
414 	if (efx_may_push_tx_desc(tx_queue, old_write_count)) {
415 		txd = efx_tx_desc(tx_queue,
416 				  old_write_count & tx_queue->ptr_mask);
417 		efx_push_tx_desc(tx_queue, txd);
418 		++tx_queue->pushes;
419 	} else {
420 		efx_notify_tx_desc(tx_queue);
421 	}
422 }
423 
424 /* Allocate hardware resources for a TX queue */
425 int efx_nic_probe_tx(struct efx_tx_queue *tx_queue)
426 {
427 	struct efx_nic *efx = tx_queue->efx;
428 	unsigned entries;
429 
430 	entries = tx_queue->ptr_mask + 1;
431 	return efx_alloc_special_buffer(efx, &tx_queue->txd,
432 					entries * sizeof(efx_qword_t));
433 }
434 
435 void efx_nic_init_tx(struct efx_tx_queue *tx_queue)
436 {
437 	struct efx_nic *efx = tx_queue->efx;
438 	efx_oword_t reg;
439 
440 	/* Pin TX descriptor ring */
441 	efx_init_special_buffer(efx, &tx_queue->txd);
442 
443 	/* Push TX descriptor ring to card */
444 	EFX_POPULATE_OWORD_10(reg,
445 			      FRF_AZ_TX_DESCQ_EN, 1,
446 			      FRF_AZ_TX_ISCSI_DDIG_EN, 0,
447 			      FRF_AZ_TX_ISCSI_HDIG_EN, 0,
448 			      FRF_AZ_TX_DESCQ_BUF_BASE_ID, tx_queue->txd.index,
449 			      FRF_AZ_TX_DESCQ_EVQ_ID,
450 			      tx_queue->channel->channel,
451 			      FRF_AZ_TX_DESCQ_OWNER_ID, 0,
452 			      FRF_AZ_TX_DESCQ_LABEL, tx_queue->queue,
453 			      FRF_AZ_TX_DESCQ_SIZE,
454 			      __ffs(tx_queue->txd.entries),
455 			      FRF_AZ_TX_DESCQ_TYPE, 0,
456 			      FRF_BZ_TX_NON_IP_DROP_DIS, 1);
457 
458 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
459 		int csum = tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD;
460 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_IP_CHKSM_DIS, !csum);
461 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_TCP_CHKSM_DIS,
462 				    !csum);
463 	}
464 
465 	efx_writeo_table(efx, &reg, efx->type->txd_ptr_tbl_base,
466 			 tx_queue->queue);
467 
468 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0) {
469 		/* Only 128 bits in this register */
470 		BUILD_BUG_ON(EFX_MAX_TX_QUEUES > 128);
471 
472 		efx_reado(efx, &reg, FR_AA_TX_CHKSM_CFG);
473 		if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD)
474 			__clear_bit_le(tx_queue->queue, &reg);
475 		else
476 			__set_bit_le(tx_queue->queue, &reg);
477 		efx_writeo(efx, &reg, FR_AA_TX_CHKSM_CFG);
478 	}
479 
480 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
481 		EFX_POPULATE_OWORD_1(reg,
482 				     FRF_BZ_TX_PACE,
483 				     (tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
484 				     FFE_BZ_TX_PACE_OFF :
485 				     FFE_BZ_TX_PACE_RESERVED);
486 		efx_writeo_table(efx, &reg, FR_BZ_TX_PACE_TBL,
487 				 tx_queue->queue);
488 	}
489 }
490 
491 static void efx_flush_tx_queue(struct efx_tx_queue *tx_queue)
492 {
493 	struct efx_nic *efx = tx_queue->efx;
494 	efx_oword_t tx_flush_descq;
495 
496 	WARN_ON(atomic_read(&tx_queue->flush_outstanding));
497 	atomic_set(&tx_queue->flush_outstanding, 1);
498 
499 	EFX_POPULATE_OWORD_2(tx_flush_descq,
500 			     FRF_AZ_TX_FLUSH_DESCQ_CMD, 1,
501 			     FRF_AZ_TX_FLUSH_DESCQ, tx_queue->queue);
502 	efx_writeo(efx, &tx_flush_descq, FR_AZ_TX_FLUSH_DESCQ);
503 }
504 
505 void efx_nic_fini_tx(struct efx_tx_queue *tx_queue)
506 {
507 	struct efx_nic *efx = tx_queue->efx;
508 	efx_oword_t tx_desc_ptr;
509 
510 	/* Remove TX descriptor ring from card */
511 	EFX_ZERO_OWORD(tx_desc_ptr);
512 	efx_writeo_table(efx, &tx_desc_ptr, efx->type->txd_ptr_tbl_base,
513 			 tx_queue->queue);
514 
515 	/* Unpin TX descriptor ring */
516 	efx_fini_special_buffer(efx, &tx_queue->txd);
517 }
518 
519 /* Free buffers backing TX queue */
520 void efx_nic_remove_tx(struct efx_tx_queue *tx_queue)
521 {
522 	efx_free_special_buffer(tx_queue->efx, &tx_queue->txd);
523 }
524 
525 /**************************************************************************
526  *
527  * RX path
528  *
529  **************************************************************************/
530 
531 /* Returns a pointer to the specified descriptor in the RX descriptor queue */
532 static inline efx_qword_t *
533 efx_rx_desc(struct efx_rx_queue *rx_queue, unsigned int index)
534 {
535 	return ((efx_qword_t *) (rx_queue->rxd.addr)) + index;
536 }
537 
538 /* This creates an entry in the RX descriptor queue */
539 static inline void
540 efx_build_rx_desc(struct efx_rx_queue *rx_queue, unsigned index)
541 {
542 	struct efx_rx_buffer *rx_buf;
543 	efx_qword_t *rxd;
544 
545 	rxd = efx_rx_desc(rx_queue, index);
546 	rx_buf = efx_rx_buffer(rx_queue, index);
547 	EFX_POPULATE_QWORD_3(*rxd,
548 			     FSF_AZ_RX_KER_BUF_SIZE,
549 			     rx_buf->len -
550 			     rx_queue->efx->type->rx_buffer_padding,
551 			     FSF_AZ_RX_KER_BUF_REGION, 0,
552 			     FSF_AZ_RX_KER_BUF_ADDR, rx_buf->dma_addr);
553 }
554 
555 /* This writes to the RX_DESC_WPTR register for the specified receive
556  * descriptor ring.
557  */
558 void efx_nic_notify_rx_desc(struct efx_rx_queue *rx_queue)
559 {
560 	struct efx_nic *efx = rx_queue->efx;
561 	efx_dword_t reg;
562 	unsigned write_ptr;
563 
564 	while (rx_queue->notified_count != rx_queue->added_count) {
565 		efx_build_rx_desc(
566 			rx_queue,
567 			rx_queue->notified_count & rx_queue->ptr_mask);
568 		++rx_queue->notified_count;
569 	}
570 
571 	wmb();
572 	write_ptr = rx_queue->added_count & rx_queue->ptr_mask;
573 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_RX_DESC_WPTR_DWORD, write_ptr);
574 	efx_writed_page(efx, &reg, FR_AZ_RX_DESC_UPD_DWORD_P0,
575 			efx_rx_queue_index(rx_queue));
576 }
577 
578 int efx_nic_probe_rx(struct efx_rx_queue *rx_queue)
579 {
580 	struct efx_nic *efx = rx_queue->efx;
581 	unsigned entries;
582 
583 	entries = rx_queue->ptr_mask + 1;
584 	return efx_alloc_special_buffer(efx, &rx_queue->rxd,
585 					entries * sizeof(efx_qword_t));
586 }
587 
588 void efx_nic_init_rx(struct efx_rx_queue *rx_queue)
589 {
590 	efx_oword_t rx_desc_ptr;
591 	struct efx_nic *efx = rx_queue->efx;
592 	bool is_b0 = efx_nic_rev(efx) >= EFX_REV_FALCON_B0;
593 	bool iscsi_digest_en = is_b0;
594 	bool jumbo_en;
595 
596 	/* For kernel-mode queues in Falcon A1, the JUMBO flag enables
597 	 * DMA to continue after a PCIe page boundary (and scattering
598 	 * is not possible).  In Falcon B0 and Siena, it enables
599 	 * scatter.
600 	 */
601 	jumbo_en = !is_b0 || efx->rx_scatter;
602 
603 	netif_dbg(efx, hw, efx->net_dev,
604 		  "RX queue %d ring in special buffers %d-%d\n",
605 		  efx_rx_queue_index(rx_queue), rx_queue->rxd.index,
606 		  rx_queue->rxd.index + rx_queue->rxd.entries - 1);
607 
608 	rx_queue->scatter_n = 0;
609 
610 	/* Pin RX descriptor ring */
611 	efx_init_special_buffer(efx, &rx_queue->rxd);
612 
613 	/* Push RX descriptor ring to card */
614 	EFX_POPULATE_OWORD_10(rx_desc_ptr,
615 			      FRF_AZ_RX_ISCSI_DDIG_EN, iscsi_digest_en,
616 			      FRF_AZ_RX_ISCSI_HDIG_EN, iscsi_digest_en,
617 			      FRF_AZ_RX_DESCQ_BUF_BASE_ID, rx_queue->rxd.index,
618 			      FRF_AZ_RX_DESCQ_EVQ_ID,
619 			      efx_rx_queue_channel(rx_queue)->channel,
620 			      FRF_AZ_RX_DESCQ_OWNER_ID, 0,
621 			      FRF_AZ_RX_DESCQ_LABEL,
622 			      efx_rx_queue_index(rx_queue),
623 			      FRF_AZ_RX_DESCQ_SIZE,
624 			      __ffs(rx_queue->rxd.entries),
625 			      FRF_AZ_RX_DESCQ_TYPE, 0 /* kernel queue */ ,
626 			      FRF_AZ_RX_DESCQ_JUMBO, jumbo_en,
627 			      FRF_AZ_RX_DESCQ_EN, 1);
628 	efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
629 			 efx_rx_queue_index(rx_queue));
630 }
631 
632 static void efx_flush_rx_queue(struct efx_rx_queue *rx_queue)
633 {
634 	struct efx_nic *efx = rx_queue->efx;
635 	efx_oword_t rx_flush_descq;
636 
637 	EFX_POPULATE_OWORD_2(rx_flush_descq,
638 			     FRF_AZ_RX_FLUSH_DESCQ_CMD, 1,
639 			     FRF_AZ_RX_FLUSH_DESCQ,
640 			     efx_rx_queue_index(rx_queue));
641 	efx_writeo(efx, &rx_flush_descq, FR_AZ_RX_FLUSH_DESCQ);
642 }
643 
644 void efx_nic_fini_rx(struct efx_rx_queue *rx_queue)
645 {
646 	efx_oword_t rx_desc_ptr;
647 	struct efx_nic *efx = rx_queue->efx;
648 
649 	/* Remove RX descriptor ring from card */
650 	EFX_ZERO_OWORD(rx_desc_ptr);
651 	efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
652 			 efx_rx_queue_index(rx_queue));
653 
654 	/* Unpin RX descriptor ring */
655 	efx_fini_special_buffer(efx, &rx_queue->rxd);
656 }
657 
658 /* Free buffers backing RX queue */
659 void efx_nic_remove_rx(struct efx_rx_queue *rx_queue)
660 {
661 	efx_free_special_buffer(rx_queue->efx, &rx_queue->rxd);
662 }
663 
664 /**************************************************************************
665  *
666  * Flush handling
667  *
668  **************************************************************************/
669 
670 /* efx_nic_flush_queues() must be woken up when all flushes are completed,
671  * or more RX flushes can be kicked off.
672  */
673 static bool efx_flush_wake(struct efx_nic *efx)
674 {
675 	/* Ensure that all updates are visible to efx_nic_flush_queues() */
676 	smp_mb();
677 
678 	return (atomic_read(&efx->drain_pending) == 0 ||
679 		(atomic_read(&efx->rxq_flush_outstanding) < EFX_RX_FLUSH_COUNT
680 		 && atomic_read(&efx->rxq_flush_pending) > 0));
681 }
682 
683 static bool efx_check_tx_flush_complete(struct efx_nic *efx)
684 {
685 	bool i = true;
686 	efx_oword_t txd_ptr_tbl;
687 	struct efx_channel *channel;
688 	struct efx_tx_queue *tx_queue;
689 
690 	efx_for_each_channel(channel, efx) {
691 		efx_for_each_channel_tx_queue(tx_queue, channel) {
692 			efx_reado_table(efx, &txd_ptr_tbl,
693 					FR_BZ_TX_DESC_PTR_TBL, tx_queue->queue);
694 			if (EFX_OWORD_FIELD(txd_ptr_tbl,
695 					    FRF_AZ_TX_DESCQ_FLUSH) ||
696 			    EFX_OWORD_FIELD(txd_ptr_tbl,
697 					    FRF_AZ_TX_DESCQ_EN)) {
698 				netif_dbg(efx, hw, efx->net_dev,
699 					  "flush did not complete on TXQ %d\n",
700 					  tx_queue->queue);
701 				i = false;
702 			} else if (atomic_cmpxchg(&tx_queue->flush_outstanding,
703 						  1, 0)) {
704 				/* The flush is complete, but we didn't
705 				 * receive a flush completion event
706 				 */
707 				netif_dbg(efx, hw, efx->net_dev,
708 					  "flush complete on TXQ %d, so drain "
709 					  "the queue\n", tx_queue->queue);
710 				/* Don't need to increment drain_pending as it
711 				 * has already been incremented for the queues
712 				 * which did not drain
713 				 */
714 				efx_magic_event(channel,
715 						EFX_CHANNEL_MAGIC_TX_DRAIN(
716 							tx_queue));
717 			}
718 		}
719 	}
720 
721 	return i;
722 }
723 
724 /* Flush all the transmit queues, and continue flushing receive queues until
725  * they're all flushed. Wait for the DRAIN events to be recieved so that there
726  * are no more RX and TX events left on any channel. */
727 int efx_nic_flush_queues(struct efx_nic *efx)
728 {
729 	unsigned timeout = msecs_to_jiffies(5000); /* 5s for all flushes and drains */
730 	struct efx_channel *channel;
731 	struct efx_rx_queue *rx_queue;
732 	struct efx_tx_queue *tx_queue;
733 	int rc = 0;
734 
735 	efx->type->prepare_flush(efx);
736 
737 	efx_for_each_channel(channel, efx) {
738 		efx_for_each_channel_tx_queue(tx_queue, channel) {
739 			atomic_inc(&efx->drain_pending);
740 			efx_flush_tx_queue(tx_queue);
741 		}
742 		efx_for_each_channel_rx_queue(rx_queue, channel) {
743 			atomic_inc(&efx->drain_pending);
744 			rx_queue->flush_pending = true;
745 			atomic_inc(&efx->rxq_flush_pending);
746 		}
747 	}
748 
749 	while (timeout && atomic_read(&efx->drain_pending) > 0) {
750 		/* If SRIOV is enabled, then offload receive queue flushing to
751 		 * the firmware (though we will still have to poll for
752 		 * completion). If that fails, fall back to the old scheme.
753 		 */
754 		if (efx_sriov_enabled(efx)) {
755 			rc = efx_mcdi_flush_rxqs(efx);
756 			if (!rc)
757 				goto wait;
758 		}
759 
760 		/* The hardware supports four concurrent rx flushes, each of
761 		 * which may need to be retried if there is an outstanding
762 		 * descriptor fetch
763 		 */
764 		efx_for_each_channel(channel, efx) {
765 			efx_for_each_channel_rx_queue(rx_queue, channel) {
766 				if (atomic_read(&efx->rxq_flush_outstanding) >=
767 				    EFX_RX_FLUSH_COUNT)
768 					break;
769 
770 				if (rx_queue->flush_pending) {
771 					rx_queue->flush_pending = false;
772 					atomic_dec(&efx->rxq_flush_pending);
773 					atomic_inc(&efx->rxq_flush_outstanding);
774 					efx_flush_rx_queue(rx_queue);
775 				}
776 			}
777 		}
778 
779 	wait:
780 		timeout = wait_event_timeout(efx->flush_wq, efx_flush_wake(efx),
781 					     timeout);
782 	}
783 
784 	if (atomic_read(&efx->drain_pending) &&
785 	    !efx_check_tx_flush_complete(efx)) {
786 		netif_err(efx, hw, efx->net_dev, "failed to flush %d queues "
787 			  "(rx %d+%d)\n", atomic_read(&efx->drain_pending),
788 			  atomic_read(&efx->rxq_flush_outstanding),
789 			  atomic_read(&efx->rxq_flush_pending));
790 		rc = -ETIMEDOUT;
791 
792 		atomic_set(&efx->drain_pending, 0);
793 		atomic_set(&efx->rxq_flush_pending, 0);
794 		atomic_set(&efx->rxq_flush_outstanding, 0);
795 	}
796 
797 	efx->type->finish_flush(efx);
798 
799 	return rc;
800 }
801 
802 /**************************************************************************
803  *
804  * Event queue processing
805  * Event queues are processed by per-channel tasklets.
806  *
807  **************************************************************************/
808 
809 /* Update a channel's event queue's read pointer (RPTR) register
810  *
811  * This writes the EVQ_RPTR_REG register for the specified channel's
812  * event queue.
813  */
814 void efx_nic_eventq_read_ack(struct efx_channel *channel)
815 {
816 	efx_dword_t reg;
817 	struct efx_nic *efx = channel->efx;
818 
819 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_EVQ_RPTR,
820 			     channel->eventq_read_ptr & channel->eventq_mask);
821 
822 	/* For Falcon A1, EVQ_RPTR_KER is documented as having a step size
823 	 * of 4 bytes, but it is really 16 bytes just like later revisions.
824 	 */
825 	efx_writed(efx, &reg,
826 		   efx->type->evq_rptr_tbl_base +
827 		   FR_BZ_EVQ_RPTR_STEP * channel->channel);
828 }
829 
830 /* Use HW to insert a SW defined event */
831 void efx_generate_event(struct efx_nic *efx, unsigned int evq,
832 			efx_qword_t *event)
833 {
834 	efx_oword_t drv_ev_reg;
835 
836 	BUILD_BUG_ON(FRF_AZ_DRV_EV_DATA_LBN != 0 ||
837 		     FRF_AZ_DRV_EV_DATA_WIDTH != 64);
838 	drv_ev_reg.u32[0] = event->u32[0];
839 	drv_ev_reg.u32[1] = event->u32[1];
840 	drv_ev_reg.u32[2] = 0;
841 	drv_ev_reg.u32[3] = 0;
842 	EFX_SET_OWORD_FIELD(drv_ev_reg, FRF_AZ_DRV_EV_QID, evq);
843 	efx_writeo(efx, &drv_ev_reg, FR_AZ_DRV_EV);
844 }
845 
846 static void efx_magic_event(struct efx_channel *channel, u32 magic)
847 {
848 	efx_qword_t event;
849 
850 	EFX_POPULATE_QWORD_2(event, FSF_AZ_EV_CODE,
851 			     FSE_AZ_EV_CODE_DRV_GEN_EV,
852 			     FSF_AZ_DRV_GEN_EV_MAGIC, magic);
853 	efx_generate_event(channel->efx, channel->channel, &event);
854 }
855 
856 /* Handle a transmit completion event
857  *
858  * The NIC batches TX completion events; the message we receive is of
859  * the form "complete all TX events up to this index".
860  */
861 static int
862 efx_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
863 {
864 	unsigned int tx_ev_desc_ptr;
865 	unsigned int tx_ev_q_label;
866 	struct efx_tx_queue *tx_queue;
867 	struct efx_nic *efx = channel->efx;
868 	int tx_packets = 0;
869 
870 	if (unlikely(ACCESS_ONCE(efx->reset_pending)))
871 		return 0;
872 
873 	if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
874 		/* Transmit completion */
875 		tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
876 		tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
877 		tx_queue = efx_channel_get_tx_queue(
878 			channel, tx_ev_q_label % EFX_TXQ_TYPES);
879 		tx_packets = ((tx_ev_desc_ptr - tx_queue->read_count) &
880 			      tx_queue->ptr_mask);
881 		efx_xmit_done(tx_queue, tx_ev_desc_ptr);
882 	} else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
883 		/* Rewrite the FIFO write pointer */
884 		tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
885 		tx_queue = efx_channel_get_tx_queue(
886 			channel, tx_ev_q_label % EFX_TXQ_TYPES);
887 
888 		netif_tx_lock(efx->net_dev);
889 		efx_notify_tx_desc(tx_queue);
890 		netif_tx_unlock(efx->net_dev);
891 	} else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR) &&
892 		   EFX_WORKAROUND_10727(efx)) {
893 		efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
894 	} else {
895 		netif_err(efx, tx_err, efx->net_dev,
896 			  "channel %d unexpected TX event "
897 			  EFX_QWORD_FMT"\n", channel->channel,
898 			  EFX_QWORD_VAL(*event));
899 	}
900 
901 	return tx_packets;
902 }
903 
904 /* Detect errors included in the rx_evt_pkt_ok bit. */
905 static u16 efx_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
906 				const efx_qword_t *event)
907 {
908 	struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
909 	struct efx_nic *efx = rx_queue->efx;
910 	bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
911 	bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
912 	bool rx_ev_frm_trunc, rx_ev_drib_nib, rx_ev_tobe_disc;
913 	bool rx_ev_other_err, rx_ev_pause_frm;
914 	bool rx_ev_hdr_type, rx_ev_mcast_pkt;
915 	unsigned rx_ev_pkt_type;
916 
917 	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
918 	rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
919 	rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
920 	rx_ev_pkt_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_TYPE);
921 	rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
922 						 FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
923 	rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
924 						  FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
925 	rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
926 						   FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
927 	rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
928 	rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
929 	rx_ev_drib_nib = ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) ?
930 			  0 : EFX_QWORD_FIELD(*event, FSF_AA_RX_EV_DRIB_NIB));
931 	rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
932 
933 	/* Every error apart from tobe_disc and pause_frm */
934 	rx_ev_other_err = (rx_ev_drib_nib | rx_ev_tcp_udp_chksum_err |
935 			   rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
936 			   rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
937 
938 	/* Count errors that are not in MAC stats.  Ignore expected
939 	 * checksum errors during self-test. */
940 	if (rx_ev_frm_trunc)
941 		++channel->n_rx_frm_trunc;
942 	else if (rx_ev_tobe_disc)
943 		++channel->n_rx_tobe_disc;
944 	else if (!efx->loopback_selftest) {
945 		if (rx_ev_ip_hdr_chksum_err)
946 			++channel->n_rx_ip_hdr_chksum_err;
947 		else if (rx_ev_tcp_udp_chksum_err)
948 			++channel->n_rx_tcp_udp_chksum_err;
949 	}
950 
951 	/* TOBE_DISC is expected on unicast mismatches; don't print out an
952 	 * error message.  FRM_TRUNC indicates RXDP dropped the packet due
953 	 * to a FIFO overflow.
954 	 */
955 #ifdef DEBUG
956 	if (rx_ev_other_err && net_ratelimit()) {
957 		netif_dbg(efx, rx_err, efx->net_dev,
958 			  " RX queue %d unexpected RX event "
959 			  EFX_QWORD_FMT "%s%s%s%s%s%s%s%s\n",
960 			  efx_rx_queue_index(rx_queue), EFX_QWORD_VAL(*event),
961 			  rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
962 			  rx_ev_ip_hdr_chksum_err ?
963 			  " [IP_HDR_CHKSUM_ERR]" : "",
964 			  rx_ev_tcp_udp_chksum_err ?
965 			  " [TCP_UDP_CHKSUM_ERR]" : "",
966 			  rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
967 			  rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
968 			  rx_ev_drib_nib ? " [DRIB_NIB]" : "",
969 			  rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
970 			  rx_ev_pause_frm ? " [PAUSE]" : "");
971 	}
972 #endif
973 
974 	/* The frame must be discarded if any of these are true. */
975 	return (rx_ev_eth_crc_err | rx_ev_frm_trunc | rx_ev_drib_nib |
976 		rx_ev_tobe_disc | rx_ev_pause_frm) ?
977 		EFX_RX_PKT_DISCARD : 0;
978 }
979 
980 /* Handle receive events that are not in-order. Return true if this
981  * can be handled as a partial packet discard, false if it's more
982  * serious.
983  */
984 static bool
985 efx_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
986 {
987 	struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
988 	struct efx_nic *efx = rx_queue->efx;
989 	unsigned expected, dropped;
990 
991 	if (rx_queue->scatter_n &&
992 	    index == ((rx_queue->removed_count + rx_queue->scatter_n - 1) &
993 		      rx_queue->ptr_mask)) {
994 		++channel->n_rx_nodesc_trunc;
995 		return true;
996 	}
997 
998 	expected = rx_queue->removed_count & rx_queue->ptr_mask;
999 	dropped = (index - expected) & rx_queue->ptr_mask;
1000 	netif_info(efx, rx_err, efx->net_dev,
1001 		   "dropped %d events (index=%d expected=%d)\n",
1002 		   dropped, index, expected);
1003 
1004 	efx_schedule_reset(efx, EFX_WORKAROUND_5676(efx) ?
1005 			   RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
1006 	return false;
1007 }
1008 
1009 /* Handle a packet received event
1010  *
1011  * The NIC gives a "discard" flag if it's a unicast packet with the
1012  * wrong destination address
1013  * Also "is multicast" and "matches multicast filter" flags can be used to
1014  * discard non-matching multicast packets.
1015  */
1016 static void
1017 efx_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
1018 {
1019 	unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
1020 	unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
1021 	unsigned expected_ptr;
1022 	bool rx_ev_pkt_ok, rx_ev_sop, rx_ev_cont;
1023 	u16 flags;
1024 	struct efx_rx_queue *rx_queue;
1025 	struct efx_nic *efx = channel->efx;
1026 
1027 	if (unlikely(ACCESS_ONCE(efx->reset_pending)))
1028 		return;
1029 
1030 	rx_ev_cont = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT);
1031 	rx_ev_sop = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP);
1032 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
1033 		channel->channel);
1034 
1035 	rx_queue = efx_channel_get_rx_queue(channel);
1036 
1037 	rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
1038 	expected_ptr = ((rx_queue->removed_count + rx_queue->scatter_n) &
1039 			rx_queue->ptr_mask);
1040 
1041 	/* Check for partial drops and other errors */
1042 	if (unlikely(rx_ev_desc_ptr != expected_ptr) ||
1043 	    unlikely(rx_ev_sop != (rx_queue->scatter_n == 0))) {
1044 		if (rx_ev_desc_ptr != expected_ptr &&
1045 		    !efx_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr))
1046 			return;
1047 
1048 		/* Discard all pending fragments */
1049 		if (rx_queue->scatter_n) {
1050 			efx_rx_packet(
1051 				rx_queue,
1052 				rx_queue->removed_count & rx_queue->ptr_mask,
1053 				rx_queue->scatter_n, 0, EFX_RX_PKT_DISCARD);
1054 			rx_queue->removed_count += rx_queue->scatter_n;
1055 			rx_queue->scatter_n = 0;
1056 		}
1057 
1058 		/* Return if there is no new fragment */
1059 		if (rx_ev_desc_ptr != expected_ptr)
1060 			return;
1061 
1062 		/* Discard new fragment if not SOP */
1063 		if (!rx_ev_sop) {
1064 			efx_rx_packet(
1065 				rx_queue,
1066 				rx_queue->removed_count & rx_queue->ptr_mask,
1067 				1, 0, EFX_RX_PKT_DISCARD);
1068 			++rx_queue->removed_count;
1069 			return;
1070 		}
1071 	}
1072 
1073 	++rx_queue->scatter_n;
1074 	if (rx_ev_cont)
1075 		return;
1076 
1077 	rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
1078 	rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
1079 	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
1080 
1081 	if (likely(rx_ev_pkt_ok)) {
1082 		/* If packet is marked as OK and packet type is TCP/IP or
1083 		 * UDP/IP, then we can rely on the hardware checksum.
1084 		 */
1085 		flags = (rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP ||
1086 			 rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP) ?
1087 			EFX_RX_PKT_CSUMMED : 0;
1088 	} else {
1089 		flags = efx_handle_rx_not_ok(rx_queue, event);
1090 	}
1091 
1092 	/* Detect multicast packets that didn't match the filter */
1093 	rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
1094 	if (rx_ev_mcast_pkt) {
1095 		unsigned int rx_ev_mcast_hash_match =
1096 			EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
1097 
1098 		if (unlikely(!rx_ev_mcast_hash_match)) {
1099 			++channel->n_rx_mcast_mismatch;
1100 			flags |= EFX_RX_PKT_DISCARD;
1101 		}
1102 	}
1103 
1104 	channel->irq_mod_score += 2;
1105 
1106 	/* Handle received packet */
1107 	efx_rx_packet(rx_queue,
1108 		      rx_queue->removed_count & rx_queue->ptr_mask,
1109 		      rx_queue->scatter_n, rx_ev_byte_cnt, flags);
1110 	rx_queue->removed_count += rx_queue->scatter_n;
1111 	rx_queue->scatter_n = 0;
1112 }
1113 
1114 /* If this flush done event corresponds to a &struct efx_tx_queue, then
1115  * send an %EFX_CHANNEL_MAGIC_TX_DRAIN event to drain the event queue
1116  * of all transmit completions.
1117  */
1118 static void
1119 efx_handle_tx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1120 {
1121 	struct efx_tx_queue *tx_queue;
1122 	int qid;
1123 
1124 	qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1125 	if (qid < EFX_TXQ_TYPES * efx->n_tx_channels) {
1126 		tx_queue = efx_get_tx_queue(efx, qid / EFX_TXQ_TYPES,
1127 					    qid % EFX_TXQ_TYPES);
1128 		if (atomic_cmpxchg(&tx_queue->flush_outstanding, 1, 0)) {
1129 			efx_magic_event(tx_queue->channel,
1130 					EFX_CHANNEL_MAGIC_TX_DRAIN(tx_queue));
1131 		}
1132 	}
1133 }
1134 
1135 /* If this flush done event corresponds to a &struct efx_rx_queue: If the flush
1136  * was succesful then send an %EFX_CHANNEL_MAGIC_RX_DRAIN, otherwise add
1137  * the RX queue back to the mask of RX queues in need of flushing.
1138  */
1139 static void
1140 efx_handle_rx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1141 {
1142 	struct efx_channel *channel;
1143 	struct efx_rx_queue *rx_queue;
1144 	int qid;
1145 	bool failed;
1146 
1147 	qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
1148 	failed = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
1149 	if (qid >= efx->n_channels)
1150 		return;
1151 	channel = efx_get_channel(efx, qid);
1152 	if (!efx_channel_has_rx_queue(channel))
1153 		return;
1154 	rx_queue = efx_channel_get_rx_queue(channel);
1155 
1156 	if (failed) {
1157 		netif_info(efx, hw, efx->net_dev,
1158 			   "RXQ %d flush retry\n", qid);
1159 		rx_queue->flush_pending = true;
1160 		atomic_inc(&efx->rxq_flush_pending);
1161 	} else {
1162 		efx_magic_event(efx_rx_queue_channel(rx_queue),
1163 				EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue));
1164 	}
1165 	atomic_dec(&efx->rxq_flush_outstanding);
1166 	if (efx_flush_wake(efx))
1167 		wake_up(&efx->flush_wq);
1168 }
1169 
1170 static void
1171 efx_handle_drain_event(struct efx_channel *channel)
1172 {
1173 	struct efx_nic *efx = channel->efx;
1174 
1175 	WARN_ON(atomic_read(&efx->drain_pending) == 0);
1176 	atomic_dec(&efx->drain_pending);
1177 	if (efx_flush_wake(efx))
1178 		wake_up(&efx->flush_wq);
1179 }
1180 
1181 static void
1182 efx_handle_generated_event(struct efx_channel *channel, efx_qword_t *event)
1183 {
1184 	struct efx_nic *efx = channel->efx;
1185 	struct efx_rx_queue *rx_queue =
1186 		efx_channel_has_rx_queue(channel) ?
1187 		efx_channel_get_rx_queue(channel) : NULL;
1188 	unsigned magic, code;
1189 
1190 	magic = EFX_QWORD_FIELD(*event, FSF_AZ_DRV_GEN_EV_MAGIC);
1191 	code = _EFX_CHANNEL_MAGIC_CODE(magic);
1192 
1193 	if (magic == EFX_CHANNEL_MAGIC_TEST(channel)) {
1194 		channel->event_test_cpu = raw_smp_processor_id();
1195 	} else if (rx_queue && magic == EFX_CHANNEL_MAGIC_FILL(rx_queue)) {
1196 		/* The queue must be empty, so we won't receive any rx
1197 		 * events, so efx_process_channel() won't refill the
1198 		 * queue. Refill it here */
1199 		efx_fast_push_rx_descriptors(rx_queue);
1200 	} else if (rx_queue && magic == EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue)) {
1201 		rx_queue->enabled = false;
1202 		efx_handle_drain_event(channel);
1203 	} else if (code == _EFX_CHANNEL_MAGIC_TX_DRAIN) {
1204 		efx_handle_drain_event(channel);
1205 	} else {
1206 		netif_dbg(efx, hw, efx->net_dev, "channel %d received "
1207 			  "generated event "EFX_QWORD_FMT"\n",
1208 			  channel->channel, EFX_QWORD_VAL(*event));
1209 	}
1210 }
1211 
1212 static void
1213 efx_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
1214 {
1215 	struct efx_nic *efx = channel->efx;
1216 	unsigned int ev_sub_code;
1217 	unsigned int ev_sub_data;
1218 
1219 	ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
1220 	ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1221 
1222 	switch (ev_sub_code) {
1223 	case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
1224 		netif_vdbg(efx, hw, efx->net_dev, "channel %d TXQ %d flushed\n",
1225 			   channel->channel, ev_sub_data);
1226 		efx_handle_tx_flush_done(efx, event);
1227 		efx_sriov_tx_flush_done(efx, event);
1228 		break;
1229 	case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
1230 		netif_vdbg(efx, hw, efx->net_dev, "channel %d RXQ %d flushed\n",
1231 			   channel->channel, ev_sub_data);
1232 		efx_handle_rx_flush_done(efx, event);
1233 		efx_sriov_rx_flush_done(efx, event);
1234 		break;
1235 	case FSE_AZ_EVQ_INIT_DONE_EV:
1236 		netif_dbg(efx, hw, efx->net_dev,
1237 			  "channel %d EVQ %d initialised\n",
1238 			  channel->channel, ev_sub_data);
1239 		break;
1240 	case FSE_AZ_SRM_UPD_DONE_EV:
1241 		netif_vdbg(efx, hw, efx->net_dev,
1242 			   "channel %d SRAM update done\n", channel->channel);
1243 		break;
1244 	case FSE_AZ_WAKE_UP_EV:
1245 		netif_vdbg(efx, hw, efx->net_dev,
1246 			   "channel %d RXQ %d wakeup event\n",
1247 			   channel->channel, ev_sub_data);
1248 		break;
1249 	case FSE_AZ_TIMER_EV:
1250 		netif_vdbg(efx, hw, efx->net_dev,
1251 			   "channel %d RX queue %d timer expired\n",
1252 			   channel->channel, ev_sub_data);
1253 		break;
1254 	case FSE_AA_RX_RECOVER_EV:
1255 		netif_err(efx, rx_err, efx->net_dev,
1256 			  "channel %d seen DRIVER RX_RESET event. "
1257 			"Resetting.\n", channel->channel);
1258 		atomic_inc(&efx->rx_reset);
1259 		efx_schedule_reset(efx,
1260 				   EFX_WORKAROUND_6555(efx) ?
1261 				   RESET_TYPE_RX_RECOVERY :
1262 				   RESET_TYPE_DISABLE);
1263 		break;
1264 	case FSE_BZ_RX_DSC_ERROR_EV:
1265 		if (ev_sub_data < EFX_VI_BASE) {
1266 			netif_err(efx, rx_err, efx->net_dev,
1267 				  "RX DMA Q %d reports descriptor fetch error."
1268 				  " RX Q %d is disabled.\n", ev_sub_data,
1269 				  ev_sub_data);
1270 			efx_schedule_reset(efx, RESET_TYPE_RX_DESC_FETCH);
1271 		} else
1272 			efx_sriov_desc_fetch_err(efx, ev_sub_data);
1273 		break;
1274 	case FSE_BZ_TX_DSC_ERROR_EV:
1275 		if (ev_sub_data < EFX_VI_BASE) {
1276 			netif_err(efx, tx_err, efx->net_dev,
1277 				  "TX DMA Q %d reports descriptor fetch error."
1278 				  " TX Q %d is disabled.\n", ev_sub_data,
1279 				  ev_sub_data);
1280 			efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
1281 		} else
1282 			efx_sriov_desc_fetch_err(efx, ev_sub_data);
1283 		break;
1284 	default:
1285 		netif_vdbg(efx, hw, efx->net_dev,
1286 			   "channel %d unknown driver event code %d "
1287 			   "data %04x\n", channel->channel, ev_sub_code,
1288 			   ev_sub_data);
1289 		break;
1290 	}
1291 }
1292 
1293 int efx_nic_process_eventq(struct efx_channel *channel, int budget)
1294 {
1295 	struct efx_nic *efx = channel->efx;
1296 	unsigned int read_ptr;
1297 	efx_qword_t event, *p_event;
1298 	int ev_code;
1299 	int tx_packets = 0;
1300 	int spent = 0;
1301 
1302 	read_ptr = channel->eventq_read_ptr;
1303 
1304 	for (;;) {
1305 		p_event = efx_event(channel, read_ptr);
1306 		event = *p_event;
1307 
1308 		if (!efx_event_present(&event))
1309 			/* End of events */
1310 			break;
1311 
1312 		netif_vdbg(channel->efx, intr, channel->efx->net_dev,
1313 			   "channel %d event is "EFX_QWORD_FMT"\n",
1314 			   channel->channel, EFX_QWORD_VAL(event));
1315 
1316 		/* Clear this event by marking it all ones */
1317 		EFX_SET_QWORD(*p_event);
1318 
1319 		++read_ptr;
1320 
1321 		ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
1322 
1323 		switch (ev_code) {
1324 		case FSE_AZ_EV_CODE_RX_EV:
1325 			efx_handle_rx_event(channel, &event);
1326 			if (++spent == budget)
1327 				goto out;
1328 			break;
1329 		case FSE_AZ_EV_CODE_TX_EV:
1330 			tx_packets += efx_handle_tx_event(channel, &event);
1331 			if (tx_packets > efx->txq_entries) {
1332 				spent = budget;
1333 				goto out;
1334 			}
1335 			break;
1336 		case FSE_AZ_EV_CODE_DRV_GEN_EV:
1337 			efx_handle_generated_event(channel, &event);
1338 			break;
1339 		case FSE_AZ_EV_CODE_DRIVER_EV:
1340 			efx_handle_driver_event(channel, &event);
1341 			break;
1342 		case FSE_CZ_EV_CODE_USER_EV:
1343 			efx_sriov_event(channel, &event);
1344 			break;
1345 		case FSE_CZ_EV_CODE_MCDI_EV:
1346 			efx_mcdi_process_event(channel, &event);
1347 			break;
1348 		case FSE_AZ_EV_CODE_GLOBAL_EV:
1349 			if (efx->type->handle_global_event &&
1350 			    efx->type->handle_global_event(channel, &event))
1351 				break;
1352 			/* else fall through */
1353 		default:
1354 			netif_err(channel->efx, hw, channel->efx->net_dev,
1355 				  "channel %d unknown event type %d (data "
1356 				  EFX_QWORD_FMT ")\n", channel->channel,
1357 				  ev_code, EFX_QWORD_VAL(event));
1358 		}
1359 	}
1360 
1361 out:
1362 	channel->eventq_read_ptr = read_ptr;
1363 	return spent;
1364 }
1365 
1366 /* Check whether an event is present in the eventq at the current
1367  * read pointer.  Only useful for self-test.
1368  */
1369 bool efx_nic_event_present(struct efx_channel *channel)
1370 {
1371 	return efx_event_present(efx_event(channel, channel->eventq_read_ptr));
1372 }
1373 
1374 /* Allocate buffer table entries for event queue */
1375 int efx_nic_probe_eventq(struct efx_channel *channel)
1376 {
1377 	struct efx_nic *efx = channel->efx;
1378 	unsigned entries;
1379 
1380 	entries = channel->eventq_mask + 1;
1381 	return efx_alloc_special_buffer(efx, &channel->eventq,
1382 					entries * sizeof(efx_qword_t));
1383 }
1384 
1385 void efx_nic_init_eventq(struct efx_channel *channel)
1386 {
1387 	efx_oword_t reg;
1388 	struct efx_nic *efx = channel->efx;
1389 
1390 	netif_dbg(efx, hw, efx->net_dev,
1391 		  "channel %d event queue in special buffers %d-%d\n",
1392 		  channel->channel, channel->eventq.index,
1393 		  channel->eventq.index + channel->eventq.entries - 1);
1394 
1395 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1396 		EFX_POPULATE_OWORD_3(reg,
1397 				     FRF_CZ_TIMER_Q_EN, 1,
1398 				     FRF_CZ_HOST_NOTIFY_MODE, 0,
1399 				     FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
1400 		efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1401 	}
1402 
1403 	/* Pin event queue buffer */
1404 	efx_init_special_buffer(efx, &channel->eventq);
1405 
1406 	/* Fill event queue with all ones (i.e. empty events) */
1407 	memset(channel->eventq.addr, 0xff, channel->eventq.len);
1408 
1409 	/* Push event queue to card */
1410 	EFX_POPULATE_OWORD_3(reg,
1411 			     FRF_AZ_EVQ_EN, 1,
1412 			     FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
1413 			     FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
1414 	efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1415 			 channel->channel);
1416 
1417 	efx->type->push_irq_moderation(channel);
1418 }
1419 
1420 void efx_nic_fini_eventq(struct efx_channel *channel)
1421 {
1422 	efx_oword_t reg;
1423 	struct efx_nic *efx = channel->efx;
1424 
1425 	/* Remove event queue from card */
1426 	EFX_ZERO_OWORD(reg);
1427 	efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1428 			 channel->channel);
1429 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1430 		efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1431 
1432 	/* Unpin event queue */
1433 	efx_fini_special_buffer(efx, &channel->eventq);
1434 }
1435 
1436 /* Free buffers backing event queue */
1437 void efx_nic_remove_eventq(struct efx_channel *channel)
1438 {
1439 	efx_free_special_buffer(channel->efx, &channel->eventq);
1440 }
1441 
1442 
1443 void efx_nic_event_test_start(struct efx_channel *channel)
1444 {
1445 	channel->event_test_cpu = -1;
1446 	smp_wmb();
1447 	efx_magic_event(channel, EFX_CHANNEL_MAGIC_TEST(channel));
1448 }
1449 
1450 void efx_nic_generate_fill_event(struct efx_rx_queue *rx_queue)
1451 {
1452 	efx_magic_event(efx_rx_queue_channel(rx_queue),
1453 			EFX_CHANNEL_MAGIC_FILL(rx_queue));
1454 }
1455 
1456 /**************************************************************************
1457  *
1458  * Hardware interrupts
1459  * The hardware interrupt handler does very little work; all the event
1460  * queue processing is carried out by per-channel tasklets.
1461  *
1462  **************************************************************************/
1463 
1464 /* Enable/disable/generate interrupts */
1465 static inline void efx_nic_interrupts(struct efx_nic *efx,
1466 				      bool enabled, bool force)
1467 {
1468 	efx_oword_t int_en_reg_ker;
1469 
1470 	EFX_POPULATE_OWORD_3(int_en_reg_ker,
1471 			     FRF_AZ_KER_INT_LEVE_SEL, efx->irq_level,
1472 			     FRF_AZ_KER_INT_KER, force,
1473 			     FRF_AZ_DRV_INT_EN_KER, enabled);
1474 	efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
1475 }
1476 
1477 void efx_nic_enable_interrupts(struct efx_nic *efx)
1478 {
1479 	EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
1480 	wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
1481 
1482 	efx_nic_interrupts(efx, true, false);
1483 }
1484 
1485 void efx_nic_disable_interrupts(struct efx_nic *efx)
1486 {
1487 	/* Disable interrupts */
1488 	efx_nic_interrupts(efx, false, false);
1489 }
1490 
1491 /* Generate a test interrupt
1492  * Interrupt must already have been enabled, otherwise nasty things
1493  * may happen.
1494  */
1495 void efx_nic_irq_test_start(struct efx_nic *efx)
1496 {
1497 	efx->last_irq_cpu = -1;
1498 	smp_wmb();
1499 	efx_nic_interrupts(efx, true, true);
1500 }
1501 
1502 /* Process a fatal interrupt
1503  * Disable bus mastering ASAP and schedule a reset
1504  */
1505 irqreturn_t efx_nic_fatal_interrupt(struct efx_nic *efx)
1506 {
1507 	struct falcon_nic_data *nic_data = efx->nic_data;
1508 	efx_oword_t *int_ker = efx->irq_status.addr;
1509 	efx_oword_t fatal_intr;
1510 	int error, mem_perr;
1511 
1512 	efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
1513 	error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
1514 
1515 	netif_err(efx, hw, efx->net_dev, "SYSTEM ERROR "EFX_OWORD_FMT" status "
1516 		  EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
1517 		  EFX_OWORD_VAL(fatal_intr),
1518 		  error ? "disabling bus mastering" : "no recognised error");
1519 
1520 	/* If this is a memory parity error dump which blocks are offending */
1521 	mem_perr = (EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER) ||
1522 		    EFX_OWORD_FIELD(fatal_intr, FRF_AZ_SRM_PERR_INT_KER));
1523 	if (mem_perr) {
1524 		efx_oword_t reg;
1525 		efx_reado(efx, &reg, FR_AZ_MEM_STAT);
1526 		netif_err(efx, hw, efx->net_dev,
1527 			  "SYSTEM ERROR: memory parity error "EFX_OWORD_FMT"\n",
1528 			  EFX_OWORD_VAL(reg));
1529 	}
1530 
1531 	/* Disable both devices */
1532 	pci_clear_master(efx->pci_dev);
1533 	if (efx_nic_is_dual_func(efx))
1534 		pci_clear_master(nic_data->pci_dev2);
1535 	efx_nic_disable_interrupts(efx);
1536 
1537 	/* Count errors and reset or disable the NIC accordingly */
1538 	if (efx->int_error_count == 0 ||
1539 	    time_after(jiffies, efx->int_error_expire)) {
1540 		efx->int_error_count = 0;
1541 		efx->int_error_expire =
1542 			jiffies + EFX_INT_ERROR_EXPIRE * HZ;
1543 	}
1544 	if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
1545 		netif_err(efx, hw, efx->net_dev,
1546 			  "SYSTEM ERROR - reset scheduled\n");
1547 		efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
1548 	} else {
1549 		netif_err(efx, hw, efx->net_dev,
1550 			  "SYSTEM ERROR - max number of errors seen."
1551 			  "NIC will be disabled\n");
1552 		efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1553 	}
1554 
1555 	return IRQ_HANDLED;
1556 }
1557 
1558 /* Handle a legacy interrupt
1559  * Acknowledges the interrupt and schedule event queue processing.
1560  */
1561 static irqreturn_t efx_legacy_interrupt(int irq, void *dev_id)
1562 {
1563 	struct efx_nic *efx = dev_id;
1564 	efx_oword_t *int_ker = efx->irq_status.addr;
1565 	irqreturn_t result = IRQ_NONE;
1566 	struct efx_channel *channel;
1567 	efx_dword_t reg;
1568 	u32 queues;
1569 	int syserr;
1570 
1571 	/* Could this be ours?  If interrupts are disabled then the
1572 	 * channel state may not be valid.
1573 	 */
1574 	if (!efx->legacy_irq_enabled)
1575 		return result;
1576 
1577 	/* Read the ISR which also ACKs the interrupts */
1578 	efx_readd(efx, &reg, FR_BZ_INT_ISR0);
1579 	queues = EFX_EXTRACT_DWORD(reg, 0, 31);
1580 
1581 	/* Handle non-event-queue sources */
1582 	if (queues & (1U << efx->irq_level)) {
1583 		syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1584 		if (unlikely(syserr))
1585 			return efx_nic_fatal_interrupt(efx);
1586 		efx->last_irq_cpu = raw_smp_processor_id();
1587 	}
1588 
1589 	if (queues != 0) {
1590 		if (EFX_WORKAROUND_15783(efx))
1591 			efx->irq_zero_count = 0;
1592 
1593 		/* Schedule processing of any interrupting queues */
1594 		efx_for_each_channel(channel, efx) {
1595 			if (queues & 1)
1596 				efx_schedule_channel_irq(channel);
1597 			queues >>= 1;
1598 		}
1599 		result = IRQ_HANDLED;
1600 
1601 	} else if (EFX_WORKAROUND_15783(efx)) {
1602 		efx_qword_t *event;
1603 
1604 		/* We can't return IRQ_HANDLED more than once on seeing ISR=0
1605 		 * because this might be a shared interrupt. */
1606 		if (efx->irq_zero_count++ == 0)
1607 			result = IRQ_HANDLED;
1608 
1609 		/* Ensure we schedule or rearm all event queues */
1610 		efx_for_each_channel(channel, efx) {
1611 			event = efx_event(channel, channel->eventq_read_ptr);
1612 			if (efx_event_present(event))
1613 				efx_schedule_channel_irq(channel);
1614 			else
1615 				efx_nic_eventq_read_ack(channel);
1616 		}
1617 	}
1618 
1619 	if (result == IRQ_HANDLED)
1620 		netif_vdbg(efx, intr, efx->net_dev,
1621 			   "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
1622 			   irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
1623 
1624 	return result;
1625 }
1626 
1627 /* Handle an MSI interrupt
1628  *
1629  * Handle an MSI hardware interrupt.  This routine schedules event
1630  * queue processing.  No interrupt acknowledgement cycle is necessary.
1631  * Also, we never need to check that the interrupt is for us, since
1632  * MSI interrupts cannot be shared.
1633  */
1634 static irqreturn_t efx_msi_interrupt(int irq, void *dev_id)
1635 {
1636 	struct efx_channel *channel = *(struct efx_channel **)dev_id;
1637 	struct efx_nic *efx = channel->efx;
1638 	efx_oword_t *int_ker = efx->irq_status.addr;
1639 	int syserr;
1640 
1641 	netif_vdbg(efx, intr, efx->net_dev,
1642 		   "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
1643 		   irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
1644 
1645 	/* Handle non-event-queue sources */
1646 	if (channel->channel == efx->irq_level) {
1647 		syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1648 		if (unlikely(syserr))
1649 			return efx_nic_fatal_interrupt(efx);
1650 		efx->last_irq_cpu = raw_smp_processor_id();
1651 	}
1652 
1653 	/* Schedule processing of the channel */
1654 	efx_schedule_channel_irq(channel);
1655 
1656 	return IRQ_HANDLED;
1657 }
1658 
1659 
1660 /* Setup RSS indirection table.
1661  * This maps from the hash value of the packet to RXQ
1662  */
1663 void efx_nic_push_rx_indir_table(struct efx_nic *efx)
1664 {
1665 	size_t i = 0;
1666 	efx_dword_t dword;
1667 
1668 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0)
1669 		return;
1670 
1671 	BUILD_BUG_ON(ARRAY_SIZE(efx->rx_indir_table) !=
1672 		     FR_BZ_RX_INDIRECTION_TBL_ROWS);
1673 
1674 	for (i = 0; i < FR_BZ_RX_INDIRECTION_TBL_ROWS; i++) {
1675 		EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
1676 				     efx->rx_indir_table[i]);
1677 		efx_writed(efx, &dword,
1678 			   FR_BZ_RX_INDIRECTION_TBL +
1679 			   FR_BZ_RX_INDIRECTION_TBL_STEP * i);
1680 	}
1681 }
1682 
1683 /* Hook interrupt handler(s)
1684  * Try MSI and then legacy interrupts.
1685  */
1686 int efx_nic_init_interrupt(struct efx_nic *efx)
1687 {
1688 	struct efx_channel *channel;
1689 	int rc;
1690 
1691 	if (!EFX_INT_MODE_USE_MSI(efx)) {
1692 		irq_handler_t handler;
1693 		if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1694 			handler = efx_legacy_interrupt;
1695 		else
1696 			handler = falcon_legacy_interrupt_a1;
1697 
1698 		rc = request_irq(efx->legacy_irq, handler, IRQF_SHARED,
1699 				 efx->name, efx);
1700 		if (rc) {
1701 			netif_err(efx, drv, efx->net_dev,
1702 				  "failed to hook legacy IRQ %d\n",
1703 				  efx->pci_dev->irq);
1704 			goto fail1;
1705 		}
1706 		return 0;
1707 	}
1708 
1709 	/* Hook MSI or MSI-X interrupt */
1710 	efx_for_each_channel(channel, efx) {
1711 		rc = request_irq(channel->irq, efx_msi_interrupt,
1712 				 IRQF_PROBE_SHARED, /* Not shared */
1713 				 efx->channel_name[channel->channel],
1714 				 &efx->channel[channel->channel]);
1715 		if (rc) {
1716 			netif_err(efx, drv, efx->net_dev,
1717 				  "failed to hook IRQ %d\n", channel->irq);
1718 			goto fail2;
1719 		}
1720 	}
1721 
1722 	return 0;
1723 
1724  fail2:
1725 	efx_for_each_channel(channel, efx)
1726 		free_irq(channel->irq, &efx->channel[channel->channel]);
1727  fail1:
1728 	return rc;
1729 }
1730 
1731 void efx_nic_fini_interrupt(struct efx_nic *efx)
1732 {
1733 	struct efx_channel *channel;
1734 	efx_oword_t reg;
1735 
1736 	/* Disable MSI/MSI-X interrupts */
1737 	efx_for_each_channel(channel, efx) {
1738 		if (channel->irq)
1739 			free_irq(channel->irq, &efx->channel[channel->channel]);
1740 	}
1741 
1742 	/* ACK legacy interrupt */
1743 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1744 		efx_reado(efx, &reg, FR_BZ_INT_ISR0);
1745 	else
1746 		falcon_irq_ack_a1(efx);
1747 
1748 	/* Disable legacy interrupt */
1749 	if (efx->legacy_irq)
1750 		free_irq(efx->legacy_irq, efx);
1751 }
1752 
1753 /* Looks at available SRAM resources and works out how many queues we
1754  * can support, and where things like descriptor caches should live.
1755  *
1756  * SRAM is split up as follows:
1757  * 0                          buftbl entries for channels
1758  * efx->vf_buftbl_base        buftbl entries for SR-IOV
1759  * efx->rx_dc_base            RX descriptor caches
1760  * efx->tx_dc_base            TX descriptor caches
1761  */
1762 void efx_nic_dimension_resources(struct efx_nic *efx, unsigned sram_lim_qw)
1763 {
1764 	unsigned vi_count, buftbl_min;
1765 
1766 	/* Account for the buffer table entries backing the datapath channels
1767 	 * and the descriptor caches for those channels.
1768 	 */
1769 	buftbl_min = ((efx->n_rx_channels * EFX_MAX_DMAQ_SIZE +
1770 		       efx->n_tx_channels * EFX_TXQ_TYPES * EFX_MAX_DMAQ_SIZE +
1771 		       efx->n_channels * EFX_MAX_EVQ_SIZE)
1772 		      * sizeof(efx_qword_t) / EFX_BUF_SIZE);
1773 	vi_count = max(efx->n_channels, efx->n_tx_channels * EFX_TXQ_TYPES);
1774 
1775 #ifdef CONFIG_SFC_SRIOV
1776 	if (efx_sriov_wanted(efx)) {
1777 		unsigned vi_dc_entries, buftbl_free, entries_per_vf, vf_limit;
1778 
1779 		efx->vf_buftbl_base = buftbl_min;
1780 
1781 		vi_dc_entries = RX_DC_ENTRIES + TX_DC_ENTRIES;
1782 		vi_count = max(vi_count, EFX_VI_BASE);
1783 		buftbl_free = (sram_lim_qw - buftbl_min -
1784 			       vi_count * vi_dc_entries);
1785 
1786 		entries_per_vf = ((vi_dc_entries + EFX_VF_BUFTBL_PER_VI) *
1787 				  efx_vf_size(efx));
1788 		vf_limit = min(buftbl_free / entries_per_vf,
1789 			       (1024U - EFX_VI_BASE) >> efx->vi_scale);
1790 
1791 		if (efx->vf_count > vf_limit) {
1792 			netif_err(efx, probe, efx->net_dev,
1793 				  "Reducing VF count from from %d to %d\n",
1794 				  efx->vf_count, vf_limit);
1795 			efx->vf_count = vf_limit;
1796 		}
1797 		vi_count += efx->vf_count * efx_vf_size(efx);
1798 	}
1799 #endif
1800 
1801 	efx->tx_dc_base = sram_lim_qw - vi_count * TX_DC_ENTRIES;
1802 	efx->rx_dc_base = efx->tx_dc_base - vi_count * RX_DC_ENTRIES;
1803 }
1804 
1805 u32 efx_nic_fpga_ver(struct efx_nic *efx)
1806 {
1807 	efx_oword_t altera_build;
1808 	efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
1809 	return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
1810 }
1811 
1812 void efx_nic_init_common(struct efx_nic *efx)
1813 {
1814 	efx_oword_t temp;
1815 
1816 	/* Set positions of descriptor caches in SRAM. */
1817 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR, efx->tx_dc_base);
1818 	efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
1819 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR, efx->rx_dc_base);
1820 	efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
1821 
1822 	/* Set TX descriptor cache size. */
1823 	BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
1824 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
1825 	efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
1826 
1827 	/* Set RX descriptor cache size.  Set low watermark to size-8, as
1828 	 * this allows most efficient prefetching.
1829 	 */
1830 	BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
1831 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
1832 	efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
1833 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
1834 	efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
1835 
1836 	/* Program INT_KER address */
1837 	EFX_POPULATE_OWORD_2(temp,
1838 			     FRF_AZ_NORM_INT_VEC_DIS_KER,
1839 			     EFX_INT_MODE_USE_MSI(efx),
1840 			     FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
1841 	efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
1842 
1843 	if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
1844 		/* Use an interrupt level unused by event queues */
1845 		efx->irq_level = 0x1f;
1846 	else
1847 		/* Use a valid MSI-X vector */
1848 		efx->irq_level = 0;
1849 
1850 	/* Enable all the genuinely fatal interrupts.  (They are still
1851 	 * masked by the overall interrupt mask, controlled by
1852 	 * falcon_interrupts()).
1853 	 *
1854 	 * Note: All other fatal interrupts are enabled
1855 	 */
1856 	EFX_POPULATE_OWORD_3(temp,
1857 			     FRF_AZ_ILL_ADR_INT_KER_EN, 1,
1858 			     FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
1859 			     FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
1860 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1861 		EFX_SET_OWORD_FIELD(temp, FRF_CZ_SRAM_PERR_INT_P_KER_EN, 1);
1862 	EFX_INVERT_OWORD(temp);
1863 	efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
1864 
1865 	efx_nic_push_rx_indir_table(efx);
1866 
1867 	/* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
1868 	 * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
1869 	 */
1870 	efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
1871 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
1872 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
1873 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
1874 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 1);
1875 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
1876 	/* Enable SW_EV to inherit in char driver - assume harmless here */
1877 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
1878 	/* Prefetch threshold 2 => fetch when descriptor cache half empty */
1879 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
1880 	/* Disable hardware watchdog which can misfire */
1881 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
1882 	/* Squash TX of packets of 16 bytes or less */
1883 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1884 		EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
1885 	efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
1886 
1887 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
1888 		EFX_POPULATE_OWORD_4(temp,
1889 				     /* Default values */
1890 				     FRF_BZ_TX_PACE_SB_NOT_AF, 0x15,
1891 				     FRF_BZ_TX_PACE_SB_AF, 0xb,
1892 				     FRF_BZ_TX_PACE_FB_BASE, 0,
1893 				     /* Allow large pace values in the
1894 				      * fast bin. */
1895 				     FRF_BZ_TX_PACE_BIN_TH,
1896 				     FFE_BZ_TX_PACE_RESERVED);
1897 		efx_writeo(efx, &temp, FR_BZ_TX_PACE);
1898 	}
1899 }
1900 
1901 /* Register dump */
1902 
1903 #define REGISTER_REVISION_A	1
1904 #define REGISTER_REVISION_B	2
1905 #define REGISTER_REVISION_C	3
1906 #define REGISTER_REVISION_Z	3	/* latest revision */
1907 
1908 struct efx_nic_reg {
1909 	u32 offset:24;
1910 	u32 min_revision:2, max_revision:2;
1911 };
1912 
1913 #define REGISTER(name, min_rev, max_rev) {				\
1914 	FR_ ## min_rev ## max_rev ## _ ## name,				\
1915 	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev	\
1916 }
1917 #define REGISTER_AA(name) REGISTER(name, A, A)
1918 #define REGISTER_AB(name) REGISTER(name, A, B)
1919 #define REGISTER_AZ(name) REGISTER(name, A, Z)
1920 #define REGISTER_BB(name) REGISTER(name, B, B)
1921 #define REGISTER_BZ(name) REGISTER(name, B, Z)
1922 #define REGISTER_CZ(name) REGISTER(name, C, Z)
1923 
1924 static const struct efx_nic_reg efx_nic_regs[] = {
1925 	REGISTER_AZ(ADR_REGION),
1926 	REGISTER_AZ(INT_EN_KER),
1927 	REGISTER_BZ(INT_EN_CHAR),
1928 	REGISTER_AZ(INT_ADR_KER),
1929 	REGISTER_BZ(INT_ADR_CHAR),
1930 	/* INT_ACK_KER is WO */
1931 	/* INT_ISR0 is RC */
1932 	REGISTER_AZ(HW_INIT),
1933 	REGISTER_CZ(USR_EV_CFG),
1934 	REGISTER_AB(EE_SPI_HCMD),
1935 	REGISTER_AB(EE_SPI_HADR),
1936 	REGISTER_AB(EE_SPI_HDATA),
1937 	REGISTER_AB(EE_BASE_PAGE),
1938 	REGISTER_AB(EE_VPD_CFG0),
1939 	/* EE_VPD_SW_CNTL and EE_VPD_SW_DATA are not used */
1940 	/* PMBX_DBG_IADDR and PBMX_DBG_IDATA are indirect */
1941 	/* PCIE_CORE_INDIRECT is indirect */
1942 	REGISTER_AB(NIC_STAT),
1943 	REGISTER_AB(GPIO_CTL),
1944 	REGISTER_AB(GLB_CTL),
1945 	/* FATAL_INTR_KER and FATAL_INTR_CHAR are partly RC */
1946 	REGISTER_BZ(DP_CTRL),
1947 	REGISTER_AZ(MEM_STAT),
1948 	REGISTER_AZ(CS_DEBUG),
1949 	REGISTER_AZ(ALTERA_BUILD),
1950 	REGISTER_AZ(CSR_SPARE),
1951 	REGISTER_AB(PCIE_SD_CTL0123),
1952 	REGISTER_AB(PCIE_SD_CTL45),
1953 	REGISTER_AB(PCIE_PCS_CTL_STAT),
1954 	/* DEBUG_DATA_OUT is not used */
1955 	/* DRV_EV is WO */
1956 	REGISTER_AZ(EVQ_CTL),
1957 	REGISTER_AZ(EVQ_CNT1),
1958 	REGISTER_AZ(EVQ_CNT2),
1959 	REGISTER_AZ(BUF_TBL_CFG),
1960 	REGISTER_AZ(SRM_RX_DC_CFG),
1961 	REGISTER_AZ(SRM_TX_DC_CFG),
1962 	REGISTER_AZ(SRM_CFG),
1963 	/* BUF_TBL_UPD is WO */
1964 	REGISTER_AZ(SRM_UPD_EVQ),
1965 	REGISTER_AZ(SRAM_PARITY),
1966 	REGISTER_AZ(RX_CFG),
1967 	REGISTER_BZ(RX_FILTER_CTL),
1968 	/* RX_FLUSH_DESCQ is WO */
1969 	REGISTER_AZ(RX_DC_CFG),
1970 	REGISTER_AZ(RX_DC_PF_WM),
1971 	REGISTER_BZ(RX_RSS_TKEY),
1972 	/* RX_NODESC_DROP is RC */
1973 	REGISTER_AA(RX_SELF_RST),
1974 	/* RX_DEBUG, RX_PUSH_DROP are not used */
1975 	REGISTER_CZ(RX_RSS_IPV6_REG1),
1976 	REGISTER_CZ(RX_RSS_IPV6_REG2),
1977 	REGISTER_CZ(RX_RSS_IPV6_REG3),
1978 	/* TX_FLUSH_DESCQ is WO */
1979 	REGISTER_AZ(TX_DC_CFG),
1980 	REGISTER_AA(TX_CHKSM_CFG),
1981 	REGISTER_AZ(TX_CFG),
1982 	/* TX_PUSH_DROP is not used */
1983 	REGISTER_AZ(TX_RESERVED),
1984 	REGISTER_BZ(TX_PACE),
1985 	/* TX_PACE_DROP_QID is RC */
1986 	REGISTER_BB(TX_VLAN),
1987 	REGISTER_BZ(TX_IPFIL_PORTEN),
1988 	REGISTER_AB(MD_TXD),
1989 	REGISTER_AB(MD_RXD),
1990 	REGISTER_AB(MD_CS),
1991 	REGISTER_AB(MD_PHY_ADR),
1992 	REGISTER_AB(MD_ID),
1993 	/* MD_STAT is RC */
1994 	REGISTER_AB(MAC_STAT_DMA),
1995 	REGISTER_AB(MAC_CTRL),
1996 	REGISTER_BB(GEN_MODE),
1997 	REGISTER_AB(MAC_MC_HASH_REG0),
1998 	REGISTER_AB(MAC_MC_HASH_REG1),
1999 	REGISTER_AB(GM_CFG1),
2000 	REGISTER_AB(GM_CFG2),
2001 	/* GM_IPG and GM_HD are not used */
2002 	REGISTER_AB(GM_MAX_FLEN),
2003 	/* GM_TEST is not used */
2004 	REGISTER_AB(GM_ADR1),
2005 	REGISTER_AB(GM_ADR2),
2006 	REGISTER_AB(GMF_CFG0),
2007 	REGISTER_AB(GMF_CFG1),
2008 	REGISTER_AB(GMF_CFG2),
2009 	REGISTER_AB(GMF_CFG3),
2010 	REGISTER_AB(GMF_CFG4),
2011 	REGISTER_AB(GMF_CFG5),
2012 	REGISTER_BB(TX_SRC_MAC_CTL),
2013 	REGISTER_AB(XM_ADR_LO),
2014 	REGISTER_AB(XM_ADR_HI),
2015 	REGISTER_AB(XM_GLB_CFG),
2016 	REGISTER_AB(XM_TX_CFG),
2017 	REGISTER_AB(XM_RX_CFG),
2018 	REGISTER_AB(XM_MGT_INT_MASK),
2019 	REGISTER_AB(XM_FC),
2020 	REGISTER_AB(XM_PAUSE_TIME),
2021 	REGISTER_AB(XM_TX_PARAM),
2022 	REGISTER_AB(XM_RX_PARAM),
2023 	/* XM_MGT_INT_MSK (note no 'A') is RC */
2024 	REGISTER_AB(XX_PWR_RST),
2025 	REGISTER_AB(XX_SD_CTL),
2026 	REGISTER_AB(XX_TXDRV_CTL),
2027 	/* XX_PRBS_CTL, XX_PRBS_CHK and XX_PRBS_ERR are not used */
2028 	/* XX_CORE_STAT is partly RC */
2029 };
2030 
2031 struct efx_nic_reg_table {
2032 	u32 offset:24;
2033 	u32 min_revision:2, max_revision:2;
2034 	u32 step:6, rows:21;
2035 };
2036 
2037 #define REGISTER_TABLE_DIMENSIONS(_, offset, min_rev, max_rev, step, rows) { \
2038 	offset,								\
2039 	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev,	\
2040 	step, rows							\
2041 }
2042 #define REGISTER_TABLE(name, min_rev, max_rev)				\
2043 	REGISTER_TABLE_DIMENSIONS(					\
2044 		name, FR_ ## min_rev ## max_rev ## _ ## name,		\
2045 		min_rev, max_rev,					\
2046 		FR_ ## min_rev ## max_rev ## _ ## name ## _STEP,	\
2047 		FR_ ## min_rev ## max_rev ## _ ## name ## _ROWS)
2048 #define REGISTER_TABLE_AA(name) REGISTER_TABLE(name, A, A)
2049 #define REGISTER_TABLE_AZ(name) REGISTER_TABLE(name, A, Z)
2050 #define REGISTER_TABLE_BB(name) REGISTER_TABLE(name, B, B)
2051 #define REGISTER_TABLE_BZ(name) REGISTER_TABLE(name, B, Z)
2052 #define REGISTER_TABLE_BB_CZ(name)					\
2053 	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, B, B,		\
2054 				  FR_BZ_ ## name ## _STEP,		\
2055 				  FR_BB_ ## name ## _ROWS),		\
2056 	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, C, Z,		\
2057 				  FR_BZ_ ## name ## _STEP,		\
2058 				  FR_CZ_ ## name ## _ROWS)
2059 #define REGISTER_TABLE_CZ(name) REGISTER_TABLE(name, C, Z)
2060 
2061 static const struct efx_nic_reg_table efx_nic_reg_tables[] = {
2062 	/* DRIVER is not used */
2063 	/* EVQ_RPTR, TIMER_COMMAND, USR_EV and {RX,TX}_DESC_UPD are WO */
2064 	REGISTER_TABLE_BB(TX_IPFIL_TBL),
2065 	REGISTER_TABLE_BB(TX_SRC_MAC_TBL),
2066 	REGISTER_TABLE_AA(RX_DESC_PTR_TBL_KER),
2067 	REGISTER_TABLE_BB_CZ(RX_DESC_PTR_TBL),
2068 	REGISTER_TABLE_AA(TX_DESC_PTR_TBL_KER),
2069 	REGISTER_TABLE_BB_CZ(TX_DESC_PTR_TBL),
2070 	REGISTER_TABLE_AA(EVQ_PTR_TBL_KER),
2071 	REGISTER_TABLE_BB_CZ(EVQ_PTR_TBL),
2072 	/* We can't reasonably read all of the buffer table (up to 8MB!).
2073 	 * However this driver will only use a few entries.  Reading
2074 	 * 1K entries allows for some expansion of queue count and
2075 	 * size before we need to change the version. */
2076 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL_KER, FR_AA_BUF_FULL_TBL_KER,
2077 				  A, A, 8, 1024),
2078 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL, FR_BZ_BUF_FULL_TBL,
2079 				  B, Z, 8, 1024),
2080 	REGISTER_TABLE_CZ(RX_MAC_FILTER_TBL0),
2081 	REGISTER_TABLE_BB_CZ(TIMER_TBL),
2082 	REGISTER_TABLE_BB_CZ(TX_PACE_TBL),
2083 	REGISTER_TABLE_BZ(RX_INDIRECTION_TBL),
2084 	/* TX_FILTER_TBL0 is huge and not used by this driver */
2085 	REGISTER_TABLE_CZ(TX_MAC_FILTER_TBL0),
2086 	REGISTER_TABLE_CZ(MC_TREG_SMEM),
2087 	/* MSIX_PBA_TABLE is not mapped */
2088 	/* SRM_DBG is not mapped (and is redundant with BUF_FLL_TBL) */
2089 	REGISTER_TABLE_BZ(RX_FILTER_TBL0),
2090 };
2091 
2092 size_t efx_nic_get_regs_len(struct efx_nic *efx)
2093 {
2094 	const struct efx_nic_reg *reg;
2095 	const struct efx_nic_reg_table *table;
2096 	size_t len = 0;
2097 
2098 	for (reg = efx_nic_regs;
2099 	     reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
2100 	     reg++)
2101 		if (efx->type->revision >= reg->min_revision &&
2102 		    efx->type->revision <= reg->max_revision)
2103 			len += sizeof(efx_oword_t);
2104 
2105 	for (table = efx_nic_reg_tables;
2106 	     table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
2107 	     table++)
2108 		if (efx->type->revision >= table->min_revision &&
2109 		    efx->type->revision <= table->max_revision)
2110 			len += table->rows * min_t(size_t, table->step, 16);
2111 
2112 	return len;
2113 }
2114 
2115 void efx_nic_get_regs(struct efx_nic *efx, void *buf)
2116 {
2117 	const struct efx_nic_reg *reg;
2118 	const struct efx_nic_reg_table *table;
2119 
2120 	for (reg = efx_nic_regs;
2121 	     reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
2122 	     reg++) {
2123 		if (efx->type->revision >= reg->min_revision &&
2124 		    efx->type->revision <= reg->max_revision) {
2125 			efx_reado(efx, (efx_oword_t *)buf, reg->offset);
2126 			buf += sizeof(efx_oword_t);
2127 		}
2128 	}
2129 
2130 	for (table = efx_nic_reg_tables;
2131 	     table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
2132 	     table++) {
2133 		size_t size, i;
2134 
2135 		if (!(efx->type->revision >= table->min_revision &&
2136 		      efx->type->revision <= table->max_revision))
2137 			continue;
2138 
2139 		size = min_t(size_t, table->step, 16);
2140 
2141 		for (i = 0; i < table->rows; i++) {
2142 			switch (table->step) {
2143 			case 4: /* 32-bit SRAM */
2144 				efx_readd(efx, buf, table->offset + 4 * i);
2145 				break;
2146 			case 8: /* 64-bit SRAM */
2147 				efx_sram_readq(efx,
2148 					       efx->membase + table->offset,
2149 					       buf, i);
2150 				break;
2151 			case 16: /* 128-bit-readable register */
2152 				efx_reado_table(efx, buf, table->offset, i);
2153 				break;
2154 			case 32: /* 128-bit register, interleaved */
2155 				efx_reado_table(efx, buf, table->offset, 2 * i);
2156 				break;
2157 			default:
2158 				WARN_ON(1);
2159 				return;
2160 			}
2161 			buf += size;
2162 		}
2163 	}
2164 }
2165