xref: /openbmc/linux/drivers/net/ethernet/sfc/nic.c (revision 1899c111)
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 <linux/cpu_rmap.h>
18 #include "net_driver.h"
19 #include "bitfield.h"
20 #include "efx.h"
21 #include "nic.h"
22 #include "regs.h"
23 #include "io.h"
24 #include "workarounds.h"
25 
26 /**************************************************************************
27  *
28  * Configurable values
29  *
30  **************************************************************************
31  */
32 
33 /* This is set to 16 for a good reason.  In summary, if larger than
34  * 16, the descriptor cache holds more than a default socket
35  * buffer's worth of packets (for UDP we can only have at most one
36  * socket buffer's worth outstanding).  This combined with the fact
37  * that we only get 1 TX event per descriptor cache means the NIC
38  * goes idle.
39  */
40 #define TX_DC_ENTRIES 16
41 #define TX_DC_ENTRIES_ORDER 1
42 
43 #define RX_DC_ENTRIES 64
44 #define RX_DC_ENTRIES_ORDER 3
45 
46 /* If EFX_MAX_INT_ERRORS internal errors occur within
47  * EFX_INT_ERROR_EXPIRE seconds, we consider the NIC broken and
48  * disable it.
49  */
50 #define EFX_INT_ERROR_EXPIRE 3600
51 #define EFX_MAX_INT_ERRORS 5
52 
53 /* Depth of RX flush request fifo */
54 #define EFX_RX_FLUSH_COUNT 4
55 
56 /* Driver generated events */
57 #define _EFX_CHANNEL_MAGIC_TEST		0x000101
58 #define _EFX_CHANNEL_MAGIC_FILL		0x000102
59 #define _EFX_CHANNEL_MAGIC_RX_DRAIN	0x000103
60 #define _EFX_CHANNEL_MAGIC_TX_DRAIN	0x000104
61 
62 #define _EFX_CHANNEL_MAGIC(_code, _data)	((_code) << 8 | (_data))
63 #define _EFX_CHANNEL_MAGIC_CODE(_magic)		((_magic) >> 8)
64 
65 #define EFX_CHANNEL_MAGIC_TEST(_channel)				\
66 	_EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_TEST, (_channel)->channel)
67 #define EFX_CHANNEL_MAGIC_FILL(_rx_queue)				\
68 	_EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_FILL,			\
69 			   efx_rx_queue_index(_rx_queue))
70 #define EFX_CHANNEL_MAGIC_RX_DRAIN(_rx_queue)				\
71 	_EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_RX_DRAIN,			\
72 			   efx_rx_queue_index(_rx_queue))
73 #define EFX_CHANNEL_MAGIC_TX_DRAIN(_tx_queue)				\
74 	_EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_TX_DRAIN,			\
75 			   (_tx_queue)->queue)
76 
77 static void efx_magic_event(struct efx_channel *channel, u32 magic);
78 
79 /**************************************************************************
80  *
81  * Solarstorm hardware access
82  *
83  **************************************************************************/
84 
85 static inline void efx_write_buf_tbl(struct efx_nic *efx, efx_qword_t *value,
86 				     unsigned int index)
87 {
88 	efx_sram_writeq(efx, efx->membase + efx->type->buf_tbl_base,
89 			value, index);
90 }
91 
92 /* Read the current event from the event queue */
93 static inline efx_qword_t *efx_event(struct efx_channel *channel,
94 				     unsigned int index)
95 {
96 	return ((efx_qword_t *) (channel->eventq.addr)) +
97 		(index & channel->eventq_mask);
98 }
99 
100 /* See if an event is present
101  *
102  * We check both the high and low dword of the event for all ones.  We
103  * wrote all ones when we cleared the event, and no valid event can
104  * have all ones in either its high or low dwords.  This approach is
105  * robust against reordering.
106  *
107  * Note that using a single 64-bit comparison is incorrect; even
108  * though the CPU read will be atomic, the DMA write may not be.
109  */
110 static inline int efx_event_present(efx_qword_t *event)
111 {
112 	return !(EFX_DWORD_IS_ALL_ONES(event->dword[0]) |
113 		  EFX_DWORD_IS_ALL_ONES(event->dword[1]));
114 }
115 
116 static bool efx_masked_compare_oword(const efx_oword_t *a, const efx_oword_t *b,
117 				     const efx_oword_t *mask)
118 {
119 	return ((a->u64[0] ^ b->u64[0]) & mask->u64[0]) ||
120 		((a->u64[1] ^ b->u64[1]) & mask->u64[1]);
121 }
122 
123 int efx_nic_test_registers(struct efx_nic *efx,
124 			   const struct efx_nic_register_test *regs,
125 			   size_t n_regs)
126 {
127 	unsigned address = 0, i, j;
128 	efx_oword_t mask, imask, original, reg, buf;
129 
130 	for (i = 0; i < n_regs; ++i) {
131 		address = regs[i].address;
132 		mask = imask = regs[i].mask;
133 		EFX_INVERT_OWORD(imask);
134 
135 		efx_reado(efx, &original, address);
136 
137 		/* bit sweep on and off */
138 		for (j = 0; j < 128; j++) {
139 			if (!EFX_EXTRACT_OWORD32(mask, j, j))
140 				continue;
141 
142 			/* Test this testable bit can be set in isolation */
143 			EFX_AND_OWORD(reg, original, mask);
144 			EFX_SET_OWORD32(reg, j, j, 1);
145 
146 			efx_writeo(efx, &reg, address);
147 			efx_reado(efx, &buf, address);
148 
149 			if (efx_masked_compare_oword(&reg, &buf, &mask))
150 				goto fail;
151 
152 			/* Test this testable bit can be cleared in isolation */
153 			EFX_OR_OWORD(reg, original, mask);
154 			EFX_SET_OWORD32(reg, j, j, 0);
155 
156 			efx_writeo(efx, &reg, address);
157 			efx_reado(efx, &buf, address);
158 
159 			if (efx_masked_compare_oword(&reg, &buf, &mask))
160 				goto fail;
161 		}
162 
163 		efx_writeo(efx, &original, address);
164 	}
165 
166 	return 0;
167 
168 fail:
169 	netif_err(efx, hw, efx->net_dev,
170 		  "wrote "EFX_OWORD_FMT" read "EFX_OWORD_FMT
171 		  " at address 0x%x mask "EFX_OWORD_FMT"\n", EFX_OWORD_VAL(reg),
172 		  EFX_OWORD_VAL(buf), address, EFX_OWORD_VAL(mask));
173 	return -EIO;
174 }
175 
176 /**************************************************************************
177  *
178  * Special buffer handling
179  * Special buffers are used for event queues and the TX and RX
180  * descriptor rings.
181  *
182  *************************************************************************/
183 
184 /*
185  * Initialise a special buffer
186  *
187  * This will define a buffer (previously allocated via
188  * efx_alloc_special_buffer()) in the buffer table, allowing
189  * it to be used for event queues, descriptor rings etc.
190  */
191 static void
192 efx_init_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
193 {
194 	efx_qword_t buf_desc;
195 	unsigned int index;
196 	dma_addr_t dma_addr;
197 	int i;
198 
199 	EFX_BUG_ON_PARANOID(!buffer->addr);
200 
201 	/* Write buffer descriptors to NIC */
202 	for (i = 0; i < buffer->entries; i++) {
203 		index = buffer->index + i;
204 		dma_addr = buffer->dma_addr + (i * EFX_BUF_SIZE);
205 		netif_dbg(efx, probe, efx->net_dev,
206 			  "mapping special buffer %d at %llx\n",
207 			  index, (unsigned long long)dma_addr);
208 		EFX_POPULATE_QWORD_3(buf_desc,
209 				     FRF_AZ_BUF_ADR_REGION, 0,
210 				     FRF_AZ_BUF_ADR_FBUF, dma_addr >> 12,
211 				     FRF_AZ_BUF_OWNER_ID_FBUF, 0);
212 		efx_write_buf_tbl(efx, &buf_desc, index);
213 	}
214 }
215 
216 /* Unmaps a buffer and clears the buffer table entries */
217 static void
218 efx_fini_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
219 {
220 	efx_oword_t buf_tbl_upd;
221 	unsigned int start = buffer->index;
222 	unsigned int end = (buffer->index + buffer->entries - 1);
223 
224 	if (!buffer->entries)
225 		return;
226 
227 	netif_dbg(efx, hw, efx->net_dev, "unmapping special buffers %d-%d\n",
228 		  buffer->index, buffer->index + buffer->entries - 1);
229 
230 	EFX_POPULATE_OWORD_4(buf_tbl_upd,
231 			     FRF_AZ_BUF_UPD_CMD, 0,
232 			     FRF_AZ_BUF_CLR_CMD, 1,
233 			     FRF_AZ_BUF_CLR_END_ID, end,
234 			     FRF_AZ_BUF_CLR_START_ID, start);
235 	efx_writeo(efx, &buf_tbl_upd, FR_AZ_BUF_TBL_UPD);
236 }
237 
238 /*
239  * Allocate a new special buffer
240  *
241  * This allocates memory for a new buffer, clears it and allocates a
242  * new buffer ID range.  It does not write into the buffer table.
243  *
244  * This call will allocate 4KB buffers, since 8KB buffers can't be
245  * used for event queues and descriptor rings.
246  */
247 static int efx_alloc_special_buffer(struct efx_nic *efx,
248 				    struct efx_special_buffer *buffer,
249 				    unsigned int len)
250 {
251 	len = ALIGN(len, EFX_BUF_SIZE);
252 
253 	buffer->addr = dma_alloc_coherent(&efx->pci_dev->dev, len,
254 					  &buffer->dma_addr, GFP_KERNEL);
255 	if (!buffer->addr)
256 		return -ENOMEM;
257 	buffer->len = len;
258 	buffer->entries = len / EFX_BUF_SIZE;
259 	BUG_ON(buffer->dma_addr & (EFX_BUF_SIZE - 1));
260 
261 	/* Select new buffer ID */
262 	buffer->index = efx->next_buffer_table;
263 	efx->next_buffer_table += buffer->entries;
264 #ifdef CONFIG_SFC_SRIOV
265 	BUG_ON(efx_sriov_enabled(efx) &&
266 	       efx->vf_buftbl_base < efx->next_buffer_table);
267 #endif
268 
269 	netif_dbg(efx, probe, efx->net_dev,
270 		  "allocating special buffers %d-%d at %llx+%x "
271 		  "(virt %p phys %llx)\n", buffer->index,
272 		  buffer->index + buffer->entries - 1,
273 		  (u64)buffer->dma_addr, len,
274 		  buffer->addr, (u64)virt_to_phys(buffer->addr));
275 
276 	return 0;
277 }
278 
279 static void
280 efx_free_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
281 {
282 	if (!buffer->addr)
283 		return;
284 
285 	netif_dbg(efx, hw, efx->net_dev,
286 		  "deallocating special buffers %d-%d at %llx+%x "
287 		  "(virt %p phys %llx)\n", buffer->index,
288 		  buffer->index + buffer->entries - 1,
289 		  (u64)buffer->dma_addr, buffer->len,
290 		  buffer->addr, (u64)virt_to_phys(buffer->addr));
291 
292 	dma_free_coherent(&efx->pci_dev->dev, buffer->len, buffer->addr,
293 			  buffer->dma_addr);
294 	buffer->addr = NULL;
295 	buffer->entries = 0;
296 }
297 
298 /**************************************************************************
299  *
300  * Generic buffer handling
301  * These buffers are used for interrupt status, MAC stats, etc.
302  *
303  **************************************************************************/
304 
305 int efx_nic_alloc_buffer(struct efx_nic *efx, struct efx_buffer *buffer,
306 			 unsigned int len)
307 {
308 	buffer->addr = dma_alloc_coherent(&efx->pci_dev->dev, len,
309 					  &buffer->dma_addr,
310 					  GFP_ATOMIC | __GFP_ZERO);
311 	if (!buffer->addr)
312 		return -ENOMEM;
313 	buffer->len = len;
314 	return 0;
315 }
316 
317 void efx_nic_free_buffer(struct efx_nic *efx, struct efx_buffer *buffer)
318 {
319 	if (buffer->addr) {
320 		dma_free_coherent(&efx->pci_dev->dev, buffer->len,
321 				  buffer->addr, buffer->dma_addr);
322 		buffer->addr = NULL;
323 	}
324 }
325 
326 /**************************************************************************
327  *
328  * TX path
329  *
330  **************************************************************************/
331 
332 /* Returns a pointer to the specified transmit descriptor in the TX
333  * descriptor queue belonging to the specified channel.
334  */
335 static inline efx_qword_t *
336 efx_tx_desc(struct efx_tx_queue *tx_queue, unsigned int index)
337 {
338 	return ((efx_qword_t *) (tx_queue->txd.addr)) + index;
339 }
340 
341 /* This writes to the TX_DESC_WPTR; write pointer for TX descriptor ring */
342 static inline void efx_notify_tx_desc(struct efx_tx_queue *tx_queue)
343 {
344 	unsigned write_ptr;
345 	efx_dword_t reg;
346 
347 	write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
348 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_TX_DESC_WPTR_DWORD, write_ptr);
349 	efx_writed_page(tx_queue->efx, &reg,
350 			FR_AZ_TX_DESC_UPD_DWORD_P0, tx_queue->queue);
351 }
352 
353 /* Write pointer and first descriptor for TX descriptor ring */
354 static inline void efx_push_tx_desc(struct efx_tx_queue *tx_queue,
355 				    const efx_qword_t *txd)
356 {
357 	unsigned write_ptr;
358 	efx_oword_t reg;
359 
360 	BUILD_BUG_ON(FRF_AZ_TX_DESC_LBN != 0);
361 	BUILD_BUG_ON(FR_AA_TX_DESC_UPD_KER != FR_BZ_TX_DESC_UPD_P0);
362 
363 	write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
364 	EFX_POPULATE_OWORD_2(reg, FRF_AZ_TX_DESC_PUSH_CMD, true,
365 			     FRF_AZ_TX_DESC_WPTR, write_ptr);
366 	reg.qword[0] = *txd;
367 	efx_writeo_page(tx_queue->efx, &reg,
368 			FR_BZ_TX_DESC_UPD_P0, tx_queue->queue);
369 }
370 
371 static inline bool
372 efx_may_push_tx_desc(struct efx_tx_queue *tx_queue, unsigned int write_count)
373 {
374 	unsigned empty_read_count = ACCESS_ONCE(tx_queue->empty_read_count);
375 
376 	if (empty_read_count == 0)
377 		return false;
378 
379 	tx_queue->empty_read_count = 0;
380 	return ((empty_read_count ^ write_count) & ~EFX_EMPTY_COUNT_VALID) == 0
381 		&& tx_queue->write_count - write_count == 1;
382 }
383 
384 /* For each entry inserted into the software descriptor ring, create a
385  * descriptor in the hardware TX descriptor ring (in host memory), and
386  * write a doorbell.
387  */
388 void efx_nic_push_buffers(struct efx_tx_queue *tx_queue)
389 {
390 
391 	struct efx_tx_buffer *buffer;
392 	efx_qword_t *txd;
393 	unsigned write_ptr;
394 	unsigned old_write_count = tx_queue->write_count;
395 
396 	BUG_ON(tx_queue->write_count == tx_queue->insert_count);
397 
398 	do {
399 		write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
400 		buffer = &tx_queue->buffer[write_ptr];
401 		txd = efx_tx_desc(tx_queue, write_ptr);
402 		++tx_queue->write_count;
403 
404 		/* Create TX descriptor ring entry */
405 		BUILD_BUG_ON(EFX_TX_BUF_CONT != 1);
406 		EFX_POPULATE_QWORD_4(*txd,
407 				     FSF_AZ_TX_KER_CONT,
408 				     buffer->flags & EFX_TX_BUF_CONT,
409 				     FSF_AZ_TX_KER_BYTE_COUNT, buffer->len,
410 				     FSF_AZ_TX_KER_BUF_REGION, 0,
411 				     FSF_AZ_TX_KER_BUF_ADDR, buffer->dma_addr);
412 	} while (tx_queue->write_count != tx_queue->insert_count);
413 
414 	wmb(); /* Ensure descriptors are written before they are fetched */
415 
416 	if (efx_may_push_tx_desc(tx_queue, old_write_count)) {
417 		txd = efx_tx_desc(tx_queue,
418 				  old_write_count & tx_queue->ptr_mask);
419 		efx_push_tx_desc(tx_queue, txd);
420 		++tx_queue->pushes;
421 	} else {
422 		efx_notify_tx_desc(tx_queue);
423 	}
424 }
425 
426 /* Allocate hardware resources for a TX queue */
427 int efx_nic_probe_tx(struct efx_tx_queue *tx_queue)
428 {
429 	struct efx_nic *efx = tx_queue->efx;
430 	unsigned entries;
431 
432 	entries = tx_queue->ptr_mask + 1;
433 	return efx_alloc_special_buffer(efx, &tx_queue->txd,
434 					entries * sizeof(efx_qword_t));
435 }
436 
437 void efx_nic_init_tx(struct efx_tx_queue *tx_queue)
438 {
439 	struct efx_nic *efx = tx_queue->efx;
440 	efx_oword_t reg;
441 
442 	/* Pin TX descriptor ring */
443 	efx_init_special_buffer(efx, &tx_queue->txd);
444 
445 	/* Push TX descriptor ring to card */
446 	EFX_POPULATE_OWORD_10(reg,
447 			      FRF_AZ_TX_DESCQ_EN, 1,
448 			      FRF_AZ_TX_ISCSI_DDIG_EN, 0,
449 			      FRF_AZ_TX_ISCSI_HDIG_EN, 0,
450 			      FRF_AZ_TX_DESCQ_BUF_BASE_ID, tx_queue->txd.index,
451 			      FRF_AZ_TX_DESCQ_EVQ_ID,
452 			      tx_queue->channel->channel,
453 			      FRF_AZ_TX_DESCQ_OWNER_ID, 0,
454 			      FRF_AZ_TX_DESCQ_LABEL, tx_queue->queue,
455 			      FRF_AZ_TX_DESCQ_SIZE,
456 			      __ffs(tx_queue->txd.entries),
457 			      FRF_AZ_TX_DESCQ_TYPE, 0,
458 			      FRF_BZ_TX_NON_IP_DROP_DIS, 1);
459 
460 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
461 		int csum = tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD;
462 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_IP_CHKSM_DIS, !csum);
463 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_TCP_CHKSM_DIS,
464 				    !csum);
465 	}
466 
467 	efx_writeo_table(efx, &reg, efx->type->txd_ptr_tbl_base,
468 			 tx_queue->queue);
469 
470 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0) {
471 		/* Only 128 bits in this register */
472 		BUILD_BUG_ON(EFX_MAX_TX_QUEUES > 128);
473 
474 		efx_reado(efx, &reg, FR_AA_TX_CHKSM_CFG);
475 		if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD)
476 			__clear_bit_le(tx_queue->queue, &reg);
477 		else
478 			__set_bit_le(tx_queue->queue, &reg);
479 		efx_writeo(efx, &reg, FR_AA_TX_CHKSM_CFG);
480 	}
481 
482 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
483 		EFX_POPULATE_OWORD_1(reg,
484 				     FRF_BZ_TX_PACE,
485 				     (tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
486 				     FFE_BZ_TX_PACE_OFF :
487 				     FFE_BZ_TX_PACE_RESERVED);
488 		efx_writeo_table(efx, &reg, FR_BZ_TX_PACE_TBL,
489 				 tx_queue->queue);
490 	}
491 }
492 
493 static void efx_flush_tx_queue(struct efx_tx_queue *tx_queue)
494 {
495 	struct efx_nic *efx = tx_queue->efx;
496 	efx_oword_t tx_flush_descq;
497 
498 	WARN_ON(atomic_read(&tx_queue->flush_outstanding));
499 	atomic_set(&tx_queue->flush_outstanding, 1);
500 
501 	EFX_POPULATE_OWORD_2(tx_flush_descq,
502 			     FRF_AZ_TX_FLUSH_DESCQ_CMD, 1,
503 			     FRF_AZ_TX_FLUSH_DESCQ, tx_queue->queue);
504 	efx_writeo(efx, &tx_flush_descq, FR_AZ_TX_FLUSH_DESCQ);
505 }
506 
507 void efx_nic_fini_tx(struct efx_tx_queue *tx_queue)
508 {
509 	struct efx_nic *efx = tx_queue->efx;
510 	efx_oword_t tx_desc_ptr;
511 
512 	/* Remove TX descriptor ring from card */
513 	EFX_ZERO_OWORD(tx_desc_ptr);
514 	efx_writeo_table(efx, &tx_desc_ptr, efx->type->txd_ptr_tbl_base,
515 			 tx_queue->queue);
516 
517 	/* Unpin TX descriptor ring */
518 	efx_fini_special_buffer(efx, &tx_queue->txd);
519 }
520 
521 /* Free buffers backing TX queue */
522 void efx_nic_remove_tx(struct efx_tx_queue *tx_queue)
523 {
524 	efx_free_special_buffer(tx_queue->efx, &tx_queue->txd);
525 }
526 
527 /**************************************************************************
528  *
529  * RX path
530  *
531  **************************************************************************/
532 
533 /* Returns a pointer to the specified descriptor in the RX descriptor queue */
534 static inline efx_qword_t *
535 efx_rx_desc(struct efx_rx_queue *rx_queue, unsigned int index)
536 {
537 	return ((efx_qword_t *) (rx_queue->rxd.addr)) + index;
538 }
539 
540 /* This creates an entry in the RX descriptor queue */
541 static inline void
542 efx_build_rx_desc(struct efx_rx_queue *rx_queue, unsigned index)
543 {
544 	struct efx_rx_buffer *rx_buf;
545 	efx_qword_t *rxd;
546 
547 	rxd = efx_rx_desc(rx_queue, index);
548 	rx_buf = efx_rx_buffer(rx_queue, index);
549 	EFX_POPULATE_QWORD_3(*rxd,
550 			     FSF_AZ_RX_KER_BUF_SIZE,
551 			     rx_buf->len -
552 			     rx_queue->efx->type->rx_buffer_padding,
553 			     FSF_AZ_RX_KER_BUF_REGION, 0,
554 			     FSF_AZ_RX_KER_BUF_ADDR, rx_buf->dma_addr);
555 }
556 
557 /* This writes to the RX_DESC_WPTR register for the specified receive
558  * descriptor ring.
559  */
560 void efx_nic_notify_rx_desc(struct efx_rx_queue *rx_queue)
561 {
562 	struct efx_nic *efx = rx_queue->efx;
563 	efx_dword_t reg;
564 	unsigned write_ptr;
565 
566 	while (rx_queue->notified_count != rx_queue->added_count) {
567 		efx_build_rx_desc(
568 			rx_queue,
569 			rx_queue->notified_count & rx_queue->ptr_mask);
570 		++rx_queue->notified_count;
571 	}
572 
573 	wmb();
574 	write_ptr = rx_queue->added_count & rx_queue->ptr_mask;
575 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_RX_DESC_WPTR_DWORD, write_ptr);
576 	efx_writed_page(efx, &reg, FR_AZ_RX_DESC_UPD_DWORD_P0,
577 			efx_rx_queue_index(rx_queue));
578 }
579 
580 int efx_nic_probe_rx(struct efx_rx_queue *rx_queue)
581 {
582 	struct efx_nic *efx = rx_queue->efx;
583 	unsigned entries;
584 
585 	entries = rx_queue->ptr_mask + 1;
586 	return efx_alloc_special_buffer(efx, &rx_queue->rxd,
587 					entries * sizeof(efx_qword_t));
588 }
589 
590 void efx_nic_init_rx(struct efx_rx_queue *rx_queue)
591 {
592 	efx_oword_t rx_desc_ptr;
593 	struct efx_nic *efx = rx_queue->efx;
594 	bool is_b0 = efx_nic_rev(efx) >= EFX_REV_FALCON_B0;
595 	bool iscsi_digest_en = is_b0;
596 	bool jumbo_en;
597 
598 	/* For kernel-mode queues in Falcon A1, the JUMBO flag enables
599 	 * DMA to continue after a PCIe page boundary (and scattering
600 	 * is not possible).  In Falcon B0 and Siena, it enables
601 	 * scatter.
602 	 */
603 	jumbo_en = !is_b0 || efx->rx_scatter;
604 
605 	netif_dbg(efx, hw, efx->net_dev,
606 		  "RX queue %d ring in special buffers %d-%d\n",
607 		  efx_rx_queue_index(rx_queue), rx_queue->rxd.index,
608 		  rx_queue->rxd.index + rx_queue->rxd.entries - 1);
609 
610 	rx_queue->scatter_n = 0;
611 
612 	/* Pin RX descriptor ring */
613 	efx_init_special_buffer(efx, &rx_queue->rxd);
614 
615 	/* Push RX descriptor ring to card */
616 	EFX_POPULATE_OWORD_10(rx_desc_ptr,
617 			      FRF_AZ_RX_ISCSI_DDIG_EN, iscsi_digest_en,
618 			      FRF_AZ_RX_ISCSI_HDIG_EN, iscsi_digest_en,
619 			      FRF_AZ_RX_DESCQ_BUF_BASE_ID, rx_queue->rxd.index,
620 			      FRF_AZ_RX_DESCQ_EVQ_ID,
621 			      efx_rx_queue_channel(rx_queue)->channel,
622 			      FRF_AZ_RX_DESCQ_OWNER_ID, 0,
623 			      FRF_AZ_RX_DESCQ_LABEL,
624 			      efx_rx_queue_index(rx_queue),
625 			      FRF_AZ_RX_DESCQ_SIZE,
626 			      __ffs(rx_queue->rxd.entries),
627 			      FRF_AZ_RX_DESCQ_TYPE, 0 /* kernel queue */ ,
628 			      FRF_AZ_RX_DESCQ_JUMBO, jumbo_en,
629 			      FRF_AZ_RX_DESCQ_EN, 1);
630 	efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
631 			 efx_rx_queue_index(rx_queue));
632 }
633 
634 static void efx_flush_rx_queue(struct efx_rx_queue *rx_queue)
635 {
636 	struct efx_nic *efx = rx_queue->efx;
637 	efx_oword_t rx_flush_descq;
638 
639 	EFX_POPULATE_OWORD_2(rx_flush_descq,
640 			     FRF_AZ_RX_FLUSH_DESCQ_CMD, 1,
641 			     FRF_AZ_RX_FLUSH_DESCQ,
642 			     efx_rx_queue_index(rx_queue));
643 	efx_writeo(efx, &rx_flush_descq, FR_AZ_RX_FLUSH_DESCQ);
644 }
645 
646 void efx_nic_fini_rx(struct efx_rx_queue *rx_queue)
647 {
648 	efx_oword_t rx_desc_ptr;
649 	struct efx_nic *efx = rx_queue->efx;
650 
651 	/* Remove RX descriptor ring from card */
652 	EFX_ZERO_OWORD(rx_desc_ptr);
653 	efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
654 			 efx_rx_queue_index(rx_queue));
655 
656 	/* Unpin RX descriptor ring */
657 	efx_fini_special_buffer(efx, &rx_queue->rxd);
658 }
659 
660 /* Free buffers backing RX queue */
661 void efx_nic_remove_rx(struct efx_rx_queue *rx_queue)
662 {
663 	efx_free_special_buffer(rx_queue->efx, &rx_queue->rxd);
664 }
665 
666 /**************************************************************************
667  *
668  * Flush handling
669  *
670  **************************************************************************/
671 
672 /* efx_nic_flush_queues() must be woken up when all flushes are completed,
673  * or more RX flushes can be kicked off.
674  */
675 static bool efx_flush_wake(struct efx_nic *efx)
676 {
677 	/* Ensure that all updates are visible to efx_nic_flush_queues() */
678 	smp_mb();
679 
680 	return (atomic_read(&efx->drain_pending) == 0 ||
681 		(atomic_read(&efx->rxq_flush_outstanding) < EFX_RX_FLUSH_COUNT
682 		 && atomic_read(&efx->rxq_flush_pending) > 0));
683 }
684 
685 static bool efx_check_tx_flush_complete(struct efx_nic *efx)
686 {
687 	bool i = true;
688 	efx_oword_t txd_ptr_tbl;
689 	struct efx_channel *channel;
690 	struct efx_tx_queue *tx_queue;
691 
692 	efx_for_each_channel(channel, efx) {
693 		efx_for_each_channel_tx_queue(tx_queue, channel) {
694 			efx_reado_table(efx, &txd_ptr_tbl,
695 					FR_BZ_TX_DESC_PTR_TBL, tx_queue->queue);
696 			if (EFX_OWORD_FIELD(txd_ptr_tbl,
697 					    FRF_AZ_TX_DESCQ_FLUSH) ||
698 			    EFX_OWORD_FIELD(txd_ptr_tbl,
699 					    FRF_AZ_TX_DESCQ_EN)) {
700 				netif_dbg(efx, hw, efx->net_dev,
701 					  "flush did not complete on TXQ %d\n",
702 					  tx_queue->queue);
703 				i = false;
704 			} else if (atomic_cmpxchg(&tx_queue->flush_outstanding,
705 						  1, 0)) {
706 				/* The flush is complete, but we didn't
707 				 * receive a flush completion event
708 				 */
709 				netif_dbg(efx, hw, efx->net_dev,
710 					  "flush complete on TXQ %d, so drain "
711 					  "the queue\n", tx_queue->queue);
712 				/* Don't need to increment drain_pending as it
713 				 * has already been incremented for the queues
714 				 * which did not drain
715 				 */
716 				efx_magic_event(channel,
717 						EFX_CHANNEL_MAGIC_TX_DRAIN(
718 							tx_queue));
719 			}
720 		}
721 	}
722 
723 	return i;
724 }
725 
726 /* Flush all the transmit queues, and continue flushing receive queues until
727  * they're all flushed. Wait for the DRAIN events to be recieved so that there
728  * are no more RX and TX events left on any channel. */
729 int efx_nic_flush_queues(struct efx_nic *efx)
730 {
731 	unsigned timeout = msecs_to_jiffies(5000); /* 5s for all flushes and drains */
732 	struct efx_channel *channel;
733 	struct efx_rx_queue *rx_queue;
734 	struct efx_tx_queue *tx_queue;
735 	int rc = 0;
736 
737 	efx->type->prepare_flush(efx);
738 
739 	efx_for_each_channel(channel, efx) {
740 		efx_for_each_channel_tx_queue(tx_queue, channel) {
741 			atomic_inc(&efx->drain_pending);
742 			efx_flush_tx_queue(tx_queue);
743 		}
744 		efx_for_each_channel_rx_queue(rx_queue, channel) {
745 			atomic_inc(&efx->drain_pending);
746 			rx_queue->flush_pending = true;
747 			atomic_inc(&efx->rxq_flush_pending);
748 		}
749 	}
750 
751 	while (timeout && atomic_read(&efx->drain_pending) > 0) {
752 		/* If SRIOV is enabled, then offload receive queue flushing to
753 		 * the firmware (though we will still have to poll for
754 		 * completion). If that fails, fall back to the old scheme.
755 		 */
756 		if (efx_sriov_enabled(efx)) {
757 			rc = efx_mcdi_flush_rxqs(efx);
758 			if (!rc)
759 				goto wait;
760 		}
761 
762 		/* The hardware supports four concurrent rx flushes, each of
763 		 * which may need to be retried if there is an outstanding
764 		 * descriptor fetch
765 		 */
766 		efx_for_each_channel(channel, efx) {
767 			efx_for_each_channel_rx_queue(rx_queue, channel) {
768 				if (atomic_read(&efx->rxq_flush_outstanding) >=
769 				    EFX_RX_FLUSH_COUNT)
770 					break;
771 
772 				if (rx_queue->flush_pending) {
773 					rx_queue->flush_pending = false;
774 					atomic_dec(&efx->rxq_flush_pending);
775 					atomic_inc(&efx->rxq_flush_outstanding);
776 					efx_flush_rx_queue(rx_queue);
777 				}
778 			}
779 		}
780 
781 	wait:
782 		timeout = wait_event_timeout(efx->flush_wq, efx_flush_wake(efx),
783 					     timeout);
784 	}
785 
786 	if (atomic_read(&efx->drain_pending) &&
787 	    !efx_check_tx_flush_complete(efx)) {
788 		netif_err(efx, hw, efx->net_dev, "failed to flush %d queues "
789 			  "(rx %d+%d)\n", atomic_read(&efx->drain_pending),
790 			  atomic_read(&efx->rxq_flush_outstanding),
791 			  atomic_read(&efx->rxq_flush_pending));
792 		rc = -ETIMEDOUT;
793 
794 		atomic_set(&efx->drain_pending, 0);
795 		atomic_set(&efx->rxq_flush_pending, 0);
796 		atomic_set(&efx->rxq_flush_outstanding, 0);
797 	}
798 
799 	efx->type->finish_flush(efx);
800 
801 	return rc;
802 }
803 
804 /**************************************************************************
805  *
806  * Event queue processing
807  * Event queues are processed by per-channel tasklets.
808  *
809  **************************************************************************/
810 
811 /* Update a channel's event queue's read pointer (RPTR) register
812  *
813  * This writes the EVQ_RPTR_REG register for the specified channel's
814  * event queue.
815  */
816 void efx_nic_eventq_read_ack(struct efx_channel *channel)
817 {
818 	efx_dword_t reg;
819 	struct efx_nic *efx = channel->efx;
820 
821 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_EVQ_RPTR,
822 			     channel->eventq_read_ptr & channel->eventq_mask);
823 
824 	/* For Falcon A1, EVQ_RPTR_KER is documented as having a step size
825 	 * of 4 bytes, but it is really 16 bytes just like later revisions.
826 	 */
827 	efx_writed(efx, &reg,
828 		   efx->type->evq_rptr_tbl_base +
829 		   FR_BZ_EVQ_RPTR_STEP * channel->channel);
830 }
831 
832 /* Use HW to insert a SW defined event */
833 void efx_generate_event(struct efx_nic *efx, unsigned int evq,
834 			efx_qword_t *event)
835 {
836 	efx_oword_t drv_ev_reg;
837 
838 	BUILD_BUG_ON(FRF_AZ_DRV_EV_DATA_LBN != 0 ||
839 		     FRF_AZ_DRV_EV_DATA_WIDTH != 64);
840 	drv_ev_reg.u32[0] = event->u32[0];
841 	drv_ev_reg.u32[1] = event->u32[1];
842 	drv_ev_reg.u32[2] = 0;
843 	drv_ev_reg.u32[3] = 0;
844 	EFX_SET_OWORD_FIELD(drv_ev_reg, FRF_AZ_DRV_EV_QID, evq);
845 	efx_writeo(efx, &drv_ev_reg, FR_AZ_DRV_EV);
846 }
847 
848 static void efx_magic_event(struct efx_channel *channel, u32 magic)
849 {
850 	efx_qword_t event;
851 
852 	EFX_POPULATE_QWORD_2(event, FSF_AZ_EV_CODE,
853 			     FSE_AZ_EV_CODE_DRV_GEN_EV,
854 			     FSF_AZ_DRV_GEN_EV_MAGIC, magic);
855 	efx_generate_event(channel->efx, channel->channel, &event);
856 }
857 
858 /* Handle a transmit completion event
859  *
860  * The NIC batches TX completion events; the message we receive is of
861  * the form "complete all TX events up to this index".
862  */
863 static int
864 efx_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
865 {
866 	unsigned int tx_ev_desc_ptr;
867 	unsigned int tx_ev_q_label;
868 	struct efx_tx_queue *tx_queue;
869 	struct efx_nic *efx = channel->efx;
870 	int tx_packets = 0;
871 
872 	if (unlikely(ACCESS_ONCE(efx->reset_pending)))
873 		return 0;
874 
875 	if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
876 		/* Transmit completion */
877 		tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
878 		tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
879 		tx_queue = efx_channel_get_tx_queue(
880 			channel, tx_ev_q_label % EFX_TXQ_TYPES);
881 		tx_packets = ((tx_ev_desc_ptr - tx_queue->read_count) &
882 			      tx_queue->ptr_mask);
883 		efx_xmit_done(tx_queue, tx_ev_desc_ptr);
884 	} else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
885 		/* Rewrite the FIFO write pointer */
886 		tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
887 		tx_queue = efx_channel_get_tx_queue(
888 			channel, tx_ev_q_label % EFX_TXQ_TYPES);
889 
890 		netif_tx_lock(efx->net_dev);
891 		efx_notify_tx_desc(tx_queue);
892 		netif_tx_unlock(efx->net_dev);
893 	} else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR) &&
894 		   EFX_WORKAROUND_10727(efx)) {
895 		efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
896 	} else {
897 		netif_err(efx, tx_err, efx->net_dev,
898 			  "channel %d unexpected TX event "
899 			  EFX_QWORD_FMT"\n", channel->channel,
900 			  EFX_QWORD_VAL(*event));
901 	}
902 
903 	return tx_packets;
904 }
905 
906 /* Detect errors included in the rx_evt_pkt_ok bit. */
907 static u16 efx_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
908 				const efx_qword_t *event)
909 {
910 	struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
911 	struct efx_nic *efx = rx_queue->efx;
912 	bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
913 	bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
914 	bool rx_ev_frm_trunc, rx_ev_drib_nib, rx_ev_tobe_disc;
915 	bool rx_ev_other_err, rx_ev_pause_frm;
916 	bool rx_ev_hdr_type, rx_ev_mcast_pkt;
917 	unsigned rx_ev_pkt_type;
918 
919 	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
920 	rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
921 	rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
922 	rx_ev_pkt_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_TYPE);
923 	rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
924 						 FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
925 	rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
926 						  FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
927 	rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
928 						   FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
929 	rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
930 	rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
931 	rx_ev_drib_nib = ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) ?
932 			  0 : EFX_QWORD_FIELD(*event, FSF_AA_RX_EV_DRIB_NIB));
933 	rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
934 
935 	/* Every error apart from tobe_disc and pause_frm */
936 	rx_ev_other_err = (rx_ev_drib_nib | rx_ev_tcp_udp_chksum_err |
937 			   rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
938 			   rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
939 
940 	/* Count errors that are not in MAC stats.  Ignore expected
941 	 * checksum errors during self-test. */
942 	if (rx_ev_frm_trunc)
943 		++channel->n_rx_frm_trunc;
944 	else if (rx_ev_tobe_disc)
945 		++channel->n_rx_tobe_disc;
946 	else if (!efx->loopback_selftest) {
947 		if (rx_ev_ip_hdr_chksum_err)
948 			++channel->n_rx_ip_hdr_chksum_err;
949 		else if (rx_ev_tcp_udp_chksum_err)
950 			++channel->n_rx_tcp_udp_chksum_err;
951 	}
952 
953 	/* TOBE_DISC is expected on unicast mismatches; don't print out an
954 	 * error message.  FRM_TRUNC indicates RXDP dropped the packet due
955 	 * to a FIFO overflow.
956 	 */
957 #ifdef DEBUG
958 	if (rx_ev_other_err && net_ratelimit()) {
959 		netif_dbg(efx, rx_err, efx->net_dev,
960 			  " RX queue %d unexpected RX event "
961 			  EFX_QWORD_FMT "%s%s%s%s%s%s%s%s\n",
962 			  efx_rx_queue_index(rx_queue), EFX_QWORD_VAL(*event),
963 			  rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
964 			  rx_ev_ip_hdr_chksum_err ?
965 			  " [IP_HDR_CHKSUM_ERR]" : "",
966 			  rx_ev_tcp_udp_chksum_err ?
967 			  " [TCP_UDP_CHKSUM_ERR]" : "",
968 			  rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
969 			  rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
970 			  rx_ev_drib_nib ? " [DRIB_NIB]" : "",
971 			  rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
972 			  rx_ev_pause_frm ? " [PAUSE]" : "");
973 	}
974 #endif
975 
976 	/* The frame must be discarded if any of these are true. */
977 	return (rx_ev_eth_crc_err | rx_ev_frm_trunc | rx_ev_drib_nib |
978 		rx_ev_tobe_disc | rx_ev_pause_frm) ?
979 		EFX_RX_PKT_DISCARD : 0;
980 }
981 
982 /* Handle receive events that are not in-order. Return true if this
983  * can be handled as a partial packet discard, false if it's more
984  * serious.
985  */
986 static bool
987 efx_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
988 {
989 	struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
990 	struct efx_nic *efx = rx_queue->efx;
991 	unsigned expected, dropped;
992 
993 	if (rx_queue->scatter_n &&
994 	    index == ((rx_queue->removed_count + rx_queue->scatter_n - 1) &
995 		      rx_queue->ptr_mask)) {
996 		++channel->n_rx_nodesc_trunc;
997 		return true;
998 	}
999 
1000 	expected = rx_queue->removed_count & rx_queue->ptr_mask;
1001 	dropped = (index - expected) & rx_queue->ptr_mask;
1002 	netif_info(efx, rx_err, efx->net_dev,
1003 		   "dropped %d events (index=%d expected=%d)\n",
1004 		   dropped, index, expected);
1005 
1006 	efx_schedule_reset(efx, EFX_WORKAROUND_5676(efx) ?
1007 			   RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
1008 	return false;
1009 }
1010 
1011 /* Handle a packet received event
1012  *
1013  * The NIC gives a "discard" flag if it's a unicast packet with the
1014  * wrong destination address
1015  * Also "is multicast" and "matches multicast filter" flags can be used to
1016  * discard non-matching multicast packets.
1017  */
1018 static void
1019 efx_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
1020 {
1021 	unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
1022 	unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
1023 	unsigned expected_ptr;
1024 	bool rx_ev_pkt_ok, rx_ev_sop, rx_ev_cont;
1025 	u16 flags;
1026 	struct efx_rx_queue *rx_queue;
1027 	struct efx_nic *efx = channel->efx;
1028 
1029 	if (unlikely(ACCESS_ONCE(efx->reset_pending)))
1030 		return;
1031 
1032 	rx_ev_cont = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT);
1033 	rx_ev_sop = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP);
1034 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
1035 		channel->channel);
1036 
1037 	rx_queue = efx_channel_get_rx_queue(channel);
1038 
1039 	rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
1040 	expected_ptr = ((rx_queue->removed_count + rx_queue->scatter_n) &
1041 			rx_queue->ptr_mask);
1042 
1043 	/* Check for partial drops and other errors */
1044 	if (unlikely(rx_ev_desc_ptr != expected_ptr) ||
1045 	    unlikely(rx_ev_sop != (rx_queue->scatter_n == 0))) {
1046 		if (rx_ev_desc_ptr != expected_ptr &&
1047 		    !efx_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr))
1048 			return;
1049 
1050 		/* Discard all pending fragments */
1051 		if (rx_queue->scatter_n) {
1052 			efx_rx_packet(
1053 				rx_queue,
1054 				rx_queue->removed_count & rx_queue->ptr_mask,
1055 				rx_queue->scatter_n, 0, EFX_RX_PKT_DISCARD);
1056 			rx_queue->removed_count += rx_queue->scatter_n;
1057 			rx_queue->scatter_n = 0;
1058 		}
1059 
1060 		/* Return if there is no new fragment */
1061 		if (rx_ev_desc_ptr != expected_ptr)
1062 			return;
1063 
1064 		/* Discard new fragment if not SOP */
1065 		if (!rx_ev_sop) {
1066 			efx_rx_packet(
1067 				rx_queue,
1068 				rx_queue->removed_count & rx_queue->ptr_mask,
1069 				1, 0, EFX_RX_PKT_DISCARD);
1070 			++rx_queue->removed_count;
1071 			return;
1072 		}
1073 	}
1074 
1075 	++rx_queue->scatter_n;
1076 	if (rx_ev_cont)
1077 		return;
1078 
1079 	rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
1080 	rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
1081 	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
1082 
1083 	if (likely(rx_ev_pkt_ok)) {
1084 		/* If packet is marked as OK then we can rely on the
1085 		 * hardware checksum and classification.
1086 		 */
1087 		flags = 0;
1088 		switch (rx_ev_hdr_type) {
1089 		case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP:
1090 			flags |= EFX_RX_PKT_TCP;
1091 			/* fall through */
1092 		case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP:
1093 			flags |= EFX_RX_PKT_CSUMMED;
1094 			/* fall through */
1095 		case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_OTHER:
1096 		case FSE_AZ_RX_EV_HDR_TYPE_OTHER:
1097 			break;
1098 		}
1099 	} else {
1100 		flags = efx_handle_rx_not_ok(rx_queue, event);
1101 	}
1102 
1103 	/* Detect multicast packets that didn't match the filter */
1104 	rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
1105 	if (rx_ev_mcast_pkt) {
1106 		unsigned int rx_ev_mcast_hash_match =
1107 			EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
1108 
1109 		if (unlikely(!rx_ev_mcast_hash_match)) {
1110 			++channel->n_rx_mcast_mismatch;
1111 			flags |= EFX_RX_PKT_DISCARD;
1112 		}
1113 	}
1114 
1115 	channel->irq_mod_score += 2;
1116 
1117 	/* Handle received packet */
1118 	efx_rx_packet(rx_queue,
1119 		      rx_queue->removed_count & rx_queue->ptr_mask,
1120 		      rx_queue->scatter_n, rx_ev_byte_cnt, flags);
1121 	rx_queue->removed_count += rx_queue->scatter_n;
1122 	rx_queue->scatter_n = 0;
1123 }
1124 
1125 /* If this flush done event corresponds to a &struct efx_tx_queue, then
1126  * send an %EFX_CHANNEL_MAGIC_TX_DRAIN event to drain the event queue
1127  * of all transmit completions.
1128  */
1129 static void
1130 efx_handle_tx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1131 {
1132 	struct efx_tx_queue *tx_queue;
1133 	int qid;
1134 
1135 	qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1136 	if (qid < EFX_TXQ_TYPES * efx->n_tx_channels) {
1137 		tx_queue = efx_get_tx_queue(efx, qid / EFX_TXQ_TYPES,
1138 					    qid % EFX_TXQ_TYPES);
1139 		if (atomic_cmpxchg(&tx_queue->flush_outstanding, 1, 0)) {
1140 			efx_magic_event(tx_queue->channel,
1141 					EFX_CHANNEL_MAGIC_TX_DRAIN(tx_queue));
1142 		}
1143 	}
1144 }
1145 
1146 /* If this flush done event corresponds to a &struct efx_rx_queue: If the flush
1147  * was succesful then send an %EFX_CHANNEL_MAGIC_RX_DRAIN, otherwise add
1148  * the RX queue back to the mask of RX queues in need of flushing.
1149  */
1150 static void
1151 efx_handle_rx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1152 {
1153 	struct efx_channel *channel;
1154 	struct efx_rx_queue *rx_queue;
1155 	int qid;
1156 	bool failed;
1157 
1158 	qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
1159 	failed = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
1160 	if (qid >= efx->n_channels)
1161 		return;
1162 	channel = efx_get_channel(efx, qid);
1163 	if (!efx_channel_has_rx_queue(channel))
1164 		return;
1165 	rx_queue = efx_channel_get_rx_queue(channel);
1166 
1167 	if (failed) {
1168 		netif_info(efx, hw, efx->net_dev,
1169 			   "RXQ %d flush retry\n", qid);
1170 		rx_queue->flush_pending = true;
1171 		atomic_inc(&efx->rxq_flush_pending);
1172 	} else {
1173 		efx_magic_event(efx_rx_queue_channel(rx_queue),
1174 				EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue));
1175 	}
1176 	atomic_dec(&efx->rxq_flush_outstanding);
1177 	if (efx_flush_wake(efx))
1178 		wake_up(&efx->flush_wq);
1179 }
1180 
1181 static void
1182 efx_handle_drain_event(struct efx_channel *channel)
1183 {
1184 	struct efx_nic *efx = channel->efx;
1185 
1186 	WARN_ON(atomic_read(&efx->drain_pending) == 0);
1187 	atomic_dec(&efx->drain_pending);
1188 	if (efx_flush_wake(efx))
1189 		wake_up(&efx->flush_wq);
1190 }
1191 
1192 static void
1193 efx_handle_generated_event(struct efx_channel *channel, efx_qword_t *event)
1194 {
1195 	struct efx_nic *efx = channel->efx;
1196 	struct efx_rx_queue *rx_queue =
1197 		efx_channel_has_rx_queue(channel) ?
1198 		efx_channel_get_rx_queue(channel) : NULL;
1199 	unsigned magic, code;
1200 
1201 	magic = EFX_QWORD_FIELD(*event, FSF_AZ_DRV_GEN_EV_MAGIC);
1202 	code = _EFX_CHANNEL_MAGIC_CODE(magic);
1203 
1204 	if (magic == EFX_CHANNEL_MAGIC_TEST(channel)) {
1205 		channel->event_test_cpu = raw_smp_processor_id();
1206 	} else if (rx_queue && magic == EFX_CHANNEL_MAGIC_FILL(rx_queue)) {
1207 		/* The queue must be empty, so we won't receive any rx
1208 		 * events, so efx_process_channel() won't refill the
1209 		 * queue. Refill it here */
1210 		efx_fast_push_rx_descriptors(rx_queue);
1211 	} else if (rx_queue && magic == EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue)) {
1212 		rx_queue->enabled = false;
1213 		efx_handle_drain_event(channel);
1214 	} else if (code == _EFX_CHANNEL_MAGIC_TX_DRAIN) {
1215 		efx_handle_drain_event(channel);
1216 	} else {
1217 		netif_dbg(efx, hw, efx->net_dev, "channel %d received "
1218 			  "generated event "EFX_QWORD_FMT"\n",
1219 			  channel->channel, EFX_QWORD_VAL(*event));
1220 	}
1221 }
1222 
1223 static void
1224 efx_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
1225 {
1226 	struct efx_nic *efx = channel->efx;
1227 	unsigned int ev_sub_code;
1228 	unsigned int ev_sub_data;
1229 
1230 	ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
1231 	ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1232 
1233 	switch (ev_sub_code) {
1234 	case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
1235 		netif_vdbg(efx, hw, efx->net_dev, "channel %d TXQ %d flushed\n",
1236 			   channel->channel, ev_sub_data);
1237 		efx_handle_tx_flush_done(efx, event);
1238 		efx_sriov_tx_flush_done(efx, event);
1239 		break;
1240 	case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
1241 		netif_vdbg(efx, hw, efx->net_dev, "channel %d RXQ %d flushed\n",
1242 			   channel->channel, ev_sub_data);
1243 		efx_handle_rx_flush_done(efx, event);
1244 		efx_sriov_rx_flush_done(efx, event);
1245 		break;
1246 	case FSE_AZ_EVQ_INIT_DONE_EV:
1247 		netif_dbg(efx, hw, efx->net_dev,
1248 			  "channel %d EVQ %d initialised\n",
1249 			  channel->channel, ev_sub_data);
1250 		break;
1251 	case FSE_AZ_SRM_UPD_DONE_EV:
1252 		netif_vdbg(efx, hw, efx->net_dev,
1253 			   "channel %d SRAM update done\n", channel->channel);
1254 		break;
1255 	case FSE_AZ_WAKE_UP_EV:
1256 		netif_vdbg(efx, hw, efx->net_dev,
1257 			   "channel %d RXQ %d wakeup event\n",
1258 			   channel->channel, ev_sub_data);
1259 		break;
1260 	case FSE_AZ_TIMER_EV:
1261 		netif_vdbg(efx, hw, efx->net_dev,
1262 			   "channel %d RX queue %d timer expired\n",
1263 			   channel->channel, ev_sub_data);
1264 		break;
1265 	case FSE_AA_RX_RECOVER_EV:
1266 		netif_err(efx, rx_err, efx->net_dev,
1267 			  "channel %d seen DRIVER RX_RESET event. "
1268 			"Resetting.\n", channel->channel);
1269 		atomic_inc(&efx->rx_reset);
1270 		efx_schedule_reset(efx,
1271 				   EFX_WORKAROUND_6555(efx) ?
1272 				   RESET_TYPE_RX_RECOVERY :
1273 				   RESET_TYPE_DISABLE);
1274 		break;
1275 	case FSE_BZ_RX_DSC_ERROR_EV:
1276 		if (ev_sub_data < EFX_VI_BASE) {
1277 			netif_err(efx, rx_err, efx->net_dev,
1278 				  "RX DMA Q %d reports descriptor fetch error."
1279 				  " RX Q %d is disabled.\n", ev_sub_data,
1280 				  ev_sub_data);
1281 			efx_schedule_reset(efx, RESET_TYPE_RX_DESC_FETCH);
1282 		} else
1283 			efx_sriov_desc_fetch_err(efx, ev_sub_data);
1284 		break;
1285 	case FSE_BZ_TX_DSC_ERROR_EV:
1286 		if (ev_sub_data < EFX_VI_BASE) {
1287 			netif_err(efx, tx_err, efx->net_dev,
1288 				  "TX DMA Q %d reports descriptor fetch error."
1289 				  " TX Q %d is disabled.\n", ev_sub_data,
1290 				  ev_sub_data);
1291 			efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
1292 		} else
1293 			efx_sriov_desc_fetch_err(efx, ev_sub_data);
1294 		break;
1295 	default:
1296 		netif_vdbg(efx, hw, efx->net_dev,
1297 			   "channel %d unknown driver event code %d "
1298 			   "data %04x\n", channel->channel, ev_sub_code,
1299 			   ev_sub_data);
1300 		break;
1301 	}
1302 }
1303 
1304 int efx_nic_process_eventq(struct efx_channel *channel, int budget)
1305 {
1306 	struct efx_nic *efx = channel->efx;
1307 	unsigned int read_ptr;
1308 	efx_qword_t event, *p_event;
1309 	int ev_code;
1310 	int tx_packets = 0;
1311 	int spent = 0;
1312 
1313 	read_ptr = channel->eventq_read_ptr;
1314 
1315 	for (;;) {
1316 		p_event = efx_event(channel, read_ptr);
1317 		event = *p_event;
1318 
1319 		if (!efx_event_present(&event))
1320 			/* End of events */
1321 			break;
1322 
1323 		netif_vdbg(channel->efx, intr, channel->efx->net_dev,
1324 			   "channel %d event is "EFX_QWORD_FMT"\n",
1325 			   channel->channel, EFX_QWORD_VAL(event));
1326 
1327 		/* Clear this event by marking it all ones */
1328 		EFX_SET_QWORD(*p_event);
1329 
1330 		++read_ptr;
1331 
1332 		ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
1333 
1334 		switch (ev_code) {
1335 		case FSE_AZ_EV_CODE_RX_EV:
1336 			efx_handle_rx_event(channel, &event);
1337 			if (++spent == budget)
1338 				goto out;
1339 			break;
1340 		case FSE_AZ_EV_CODE_TX_EV:
1341 			tx_packets += efx_handle_tx_event(channel, &event);
1342 			if (tx_packets > efx->txq_entries) {
1343 				spent = budget;
1344 				goto out;
1345 			}
1346 			break;
1347 		case FSE_AZ_EV_CODE_DRV_GEN_EV:
1348 			efx_handle_generated_event(channel, &event);
1349 			break;
1350 		case FSE_AZ_EV_CODE_DRIVER_EV:
1351 			efx_handle_driver_event(channel, &event);
1352 			break;
1353 		case FSE_CZ_EV_CODE_USER_EV:
1354 			efx_sriov_event(channel, &event);
1355 			break;
1356 		case FSE_CZ_EV_CODE_MCDI_EV:
1357 			efx_mcdi_process_event(channel, &event);
1358 			break;
1359 		case FSE_AZ_EV_CODE_GLOBAL_EV:
1360 			if (efx->type->handle_global_event &&
1361 			    efx->type->handle_global_event(channel, &event))
1362 				break;
1363 			/* else fall through */
1364 		default:
1365 			netif_err(channel->efx, hw, channel->efx->net_dev,
1366 				  "channel %d unknown event type %d (data "
1367 				  EFX_QWORD_FMT ")\n", channel->channel,
1368 				  ev_code, EFX_QWORD_VAL(event));
1369 		}
1370 	}
1371 
1372 out:
1373 	channel->eventq_read_ptr = read_ptr;
1374 	return spent;
1375 }
1376 
1377 /* Check whether an event is present in the eventq at the current
1378  * read pointer.  Only useful for self-test.
1379  */
1380 bool efx_nic_event_present(struct efx_channel *channel)
1381 {
1382 	return efx_event_present(efx_event(channel, channel->eventq_read_ptr));
1383 }
1384 
1385 /* Allocate buffer table entries for event queue */
1386 int efx_nic_probe_eventq(struct efx_channel *channel)
1387 {
1388 	struct efx_nic *efx = channel->efx;
1389 	unsigned entries;
1390 
1391 	entries = channel->eventq_mask + 1;
1392 	return efx_alloc_special_buffer(efx, &channel->eventq,
1393 					entries * sizeof(efx_qword_t));
1394 }
1395 
1396 void efx_nic_init_eventq(struct efx_channel *channel)
1397 {
1398 	efx_oword_t reg;
1399 	struct efx_nic *efx = channel->efx;
1400 
1401 	netif_dbg(efx, hw, efx->net_dev,
1402 		  "channel %d event queue in special buffers %d-%d\n",
1403 		  channel->channel, channel->eventq.index,
1404 		  channel->eventq.index + channel->eventq.entries - 1);
1405 
1406 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1407 		EFX_POPULATE_OWORD_3(reg,
1408 				     FRF_CZ_TIMER_Q_EN, 1,
1409 				     FRF_CZ_HOST_NOTIFY_MODE, 0,
1410 				     FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
1411 		efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1412 	}
1413 
1414 	/* Pin event queue buffer */
1415 	efx_init_special_buffer(efx, &channel->eventq);
1416 
1417 	/* Fill event queue with all ones (i.e. empty events) */
1418 	memset(channel->eventq.addr, 0xff, channel->eventq.len);
1419 
1420 	/* Push event queue to card */
1421 	EFX_POPULATE_OWORD_3(reg,
1422 			     FRF_AZ_EVQ_EN, 1,
1423 			     FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
1424 			     FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
1425 	efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1426 			 channel->channel);
1427 
1428 	efx->type->push_irq_moderation(channel);
1429 }
1430 
1431 void efx_nic_fini_eventq(struct efx_channel *channel)
1432 {
1433 	efx_oword_t reg;
1434 	struct efx_nic *efx = channel->efx;
1435 
1436 	/* Remove event queue from card */
1437 	EFX_ZERO_OWORD(reg);
1438 	efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1439 			 channel->channel);
1440 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1441 		efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1442 
1443 	/* Unpin event queue */
1444 	efx_fini_special_buffer(efx, &channel->eventq);
1445 }
1446 
1447 /* Free buffers backing event queue */
1448 void efx_nic_remove_eventq(struct efx_channel *channel)
1449 {
1450 	efx_free_special_buffer(channel->efx, &channel->eventq);
1451 }
1452 
1453 
1454 void efx_nic_event_test_start(struct efx_channel *channel)
1455 {
1456 	channel->event_test_cpu = -1;
1457 	smp_wmb();
1458 	efx_magic_event(channel, EFX_CHANNEL_MAGIC_TEST(channel));
1459 }
1460 
1461 void efx_nic_generate_fill_event(struct efx_rx_queue *rx_queue)
1462 {
1463 	efx_magic_event(efx_rx_queue_channel(rx_queue),
1464 			EFX_CHANNEL_MAGIC_FILL(rx_queue));
1465 }
1466 
1467 /**************************************************************************
1468  *
1469  * Hardware interrupts
1470  * The hardware interrupt handler does very little work; all the event
1471  * queue processing is carried out by per-channel tasklets.
1472  *
1473  **************************************************************************/
1474 
1475 /* Enable/disable/generate interrupts */
1476 static inline void efx_nic_interrupts(struct efx_nic *efx,
1477 				      bool enabled, bool force)
1478 {
1479 	efx_oword_t int_en_reg_ker;
1480 
1481 	EFX_POPULATE_OWORD_3(int_en_reg_ker,
1482 			     FRF_AZ_KER_INT_LEVE_SEL, efx->irq_level,
1483 			     FRF_AZ_KER_INT_KER, force,
1484 			     FRF_AZ_DRV_INT_EN_KER, enabled);
1485 	efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
1486 }
1487 
1488 void efx_nic_enable_interrupts(struct efx_nic *efx)
1489 {
1490 	EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
1491 	wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
1492 
1493 	efx_nic_interrupts(efx, true, false);
1494 }
1495 
1496 void efx_nic_disable_interrupts(struct efx_nic *efx)
1497 {
1498 	/* Disable interrupts */
1499 	efx_nic_interrupts(efx, false, false);
1500 }
1501 
1502 /* Generate a test interrupt
1503  * Interrupt must already have been enabled, otherwise nasty things
1504  * may happen.
1505  */
1506 void efx_nic_irq_test_start(struct efx_nic *efx)
1507 {
1508 	efx->last_irq_cpu = -1;
1509 	smp_wmb();
1510 	efx_nic_interrupts(efx, true, true);
1511 }
1512 
1513 /* Process a fatal interrupt
1514  * Disable bus mastering ASAP and schedule a reset
1515  */
1516 irqreturn_t efx_nic_fatal_interrupt(struct efx_nic *efx)
1517 {
1518 	struct falcon_nic_data *nic_data = efx->nic_data;
1519 	efx_oword_t *int_ker = efx->irq_status.addr;
1520 	efx_oword_t fatal_intr;
1521 	int error, mem_perr;
1522 
1523 	efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
1524 	error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
1525 
1526 	netif_err(efx, hw, efx->net_dev, "SYSTEM ERROR "EFX_OWORD_FMT" status "
1527 		  EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
1528 		  EFX_OWORD_VAL(fatal_intr),
1529 		  error ? "disabling bus mastering" : "no recognised error");
1530 
1531 	/* If this is a memory parity error dump which blocks are offending */
1532 	mem_perr = (EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER) ||
1533 		    EFX_OWORD_FIELD(fatal_intr, FRF_AZ_SRM_PERR_INT_KER));
1534 	if (mem_perr) {
1535 		efx_oword_t reg;
1536 		efx_reado(efx, &reg, FR_AZ_MEM_STAT);
1537 		netif_err(efx, hw, efx->net_dev,
1538 			  "SYSTEM ERROR: memory parity error "EFX_OWORD_FMT"\n",
1539 			  EFX_OWORD_VAL(reg));
1540 	}
1541 
1542 	/* Disable both devices */
1543 	pci_clear_master(efx->pci_dev);
1544 	if (efx_nic_is_dual_func(efx))
1545 		pci_clear_master(nic_data->pci_dev2);
1546 	efx_nic_disable_interrupts(efx);
1547 
1548 	/* Count errors and reset or disable the NIC accordingly */
1549 	if (efx->int_error_count == 0 ||
1550 	    time_after(jiffies, efx->int_error_expire)) {
1551 		efx->int_error_count = 0;
1552 		efx->int_error_expire =
1553 			jiffies + EFX_INT_ERROR_EXPIRE * HZ;
1554 	}
1555 	if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
1556 		netif_err(efx, hw, efx->net_dev,
1557 			  "SYSTEM ERROR - reset scheduled\n");
1558 		efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
1559 	} else {
1560 		netif_err(efx, hw, efx->net_dev,
1561 			  "SYSTEM ERROR - max number of errors seen."
1562 			  "NIC will be disabled\n");
1563 		efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1564 	}
1565 
1566 	return IRQ_HANDLED;
1567 }
1568 
1569 /* Handle a legacy interrupt
1570  * Acknowledges the interrupt and schedule event queue processing.
1571  */
1572 static irqreturn_t efx_legacy_interrupt(int irq, void *dev_id)
1573 {
1574 	struct efx_nic *efx = dev_id;
1575 	efx_oword_t *int_ker = efx->irq_status.addr;
1576 	irqreturn_t result = IRQ_NONE;
1577 	struct efx_channel *channel;
1578 	efx_dword_t reg;
1579 	u32 queues;
1580 	int syserr;
1581 
1582 	/* Could this be ours?  If interrupts are disabled then the
1583 	 * channel state may not be valid.
1584 	 */
1585 	if (!efx->legacy_irq_enabled)
1586 		return result;
1587 
1588 	/* Read the ISR which also ACKs the interrupts */
1589 	efx_readd(efx, &reg, FR_BZ_INT_ISR0);
1590 	queues = EFX_EXTRACT_DWORD(reg, 0, 31);
1591 
1592 	/* Legacy interrupts are disabled too late by the EEH kernel
1593 	 * code. Disable them earlier.
1594 	 * If an EEH error occurred, the read will have returned all ones.
1595 	 */
1596 	if (EFX_DWORD_IS_ALL_ONES(reg) && efx_try_recovery(efx) &&
1597 	    !efx->eeh_disabled_legacy_irq) {
1598 		disable_irq_nosync(efx->legacy_irq);
1599 		efx->eeh_disabled_legacy_irq = true;
1600 	}
1601 
1602 	/* Handle non-event-queue sources */
1603 	if (queues & (1U << efx->irq_level)) {
1604 		syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1605 		if (unlikely(syserr))
1606 			return efx_nic_fatal_interrupt(efx);
1607 		efx->last_irq_cpu = raw_smp_processor_id();
1608 	}
1609 
1610 	if (queues != 0) {
1611 		if (EFX_WORKAROUND_15783(efx))
1612 			efx->irq_zero_count = 0;
1613 
1614 		/* Schedule processing of any interrupting queues */
1615 		efx_for_each_channel(channel, efx) {
1616 			if (queues & 1)
1617 				efx_schedule_channel_irq(channel);
1618 			queues >>= 1;
1619 		}
1620 		result = IRQ_HANDLED;
1621 
1622 	} else if (EFX_WORKAROUND_15783(efx)) {
1623 		efx_qword_t *event;
1624 
1625 		/* We can't return IRQ_HANDLED more than once on seeing ISR=0
1626 		 * because this might be a shared interrupt. */
1627 		if (efx->irq_zero_count++ == 0)
1628 			result = IRQ_HANDLED;
1629 
1630 		/* Ensure we schedule or rearm all event queues */
1631 		efx_for_each_channel(channel, efx) {
1632 			event = efx_event(channel, channel->eventq_read_ptr);
1633 			if (efx_event_present(event))
1634 				efx_schedule_channel_irq(channel);
1635 			else
1636 				efx_nic_eventq_read_ack(channel);
1637 		}
1638 	}
1639 
1640 	if (result == IRQ_HANDLED)
1641 		netif_vdbg(efx, intr, efx->net_dev,
1642 			   "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
1643 			   irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
1644 
1645 	return result;
1646 }
1647 
1648 /* Handle an MSI interrupt
1649  *
1650  * Handle an MSI hardware interrupt.  This routine schedules event
1651  * queue processing.  No interrupt acknowledgement cycle is necessary.
1652  * Also, we never need to check that the interrupt is for us, since
1653  * MSI interrupts cannot be shared.
1654  */
1655 static irqreturn_t efx_msi_interrupt(int irq, void *dev_id)
1656 {
1657 	struct efx_channel *channel = *(struct efx_channel **)dev_id;
1658 	struct efx_nic *efx = channel->efx;
1659 	efx_oword_t *int_ker = efx->irq_status.addr;
1660 	int syserr;
1661 
1662 	netif_vdbg(efx, intr, efx->net_dev,
1663 		   "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
1664 		   irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
1665 
1666 	/* Handle non-event-queue sources */
1667 	if (channel->channel == efx->irq_level) {
1668 		syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1669 		if (unlikely(syserr))
1670 			return efx_nic_fatal_interrupt(efx);
1671 		efx->last_irq_cpu = raw_smp_processor_id();
1672 	}
1673 
1674 	/* Schedule processing of the channel */
1675 	efx_schedule_channel_irq(channel);
1676 
1677 	return IRQ_HANDLED;
1678 }
1679 
1680 
1681 /* Setup RSS indirection table.
1682  * This maps from the hash value of the packet to RXQ
1683  */
1684 void efx_nic_push_rx_indir_table(struct efx_nic *efx)
1685 {
1686 	size_t i = 0;
1687 	efx_dword_t dword;
1688 
1689 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0)
1690 		return;
1691 
1692 	BUILD_BUG_ON(ARRAY_SIZE(efx->rx_indir_table) !=
1693 		     FR_BZ_RX_INDIRECTION_TBL_ROWS);
1694 
1695 	for (i = 0; i < FR_BZ_RX_INDIRECTION_TBL_ROWS; i++) {
1696 		EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
1697 				     efx->rx_indir_table[i]);
1698 		efx_writed(efx, &dword,
1699 			   FR_BZ_RX_INDIRECTION_TBL +
1700 			   FR_BZ_RX_INDIRECTION_TBL_STEP * i);
1701 	}
1702 }
1703 
1704 /* Hook interrupt handler(s)
1705  * Try MSI and then legacy interrupts.
1706  */
1707 int efx_nic_init_interrupt(struct efx_nic *efx)
1708 {
1709 	struct efx_channel *channel;
1710 	unsigned int n_irqs;
1711 	int rc;
1712 
1713 	if (!EFX_INT_MODE_USE_MSI(efx)) {
1714 		irq_handler_t handler;
1715 		if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1716 			handler = efx_legacy_interrupt;
1717 		else
1718 			handler = falcon_legacy_interrupt_a1;
1719 
1720 		rc = request_irq(efx->legacy_irq, handler, IRQF_SHARED,
1721 				 efx->name, efx);
1722 		if (rc) {
1723 			netif_err(efx, drv, efx->net_dev,
1724 				  "failed to hook legacy IRQ %d\n",
1725 				  efx->pci_dev->irq);
1726 			goto fail1;
1727 		}
1728 		return 0;
1729 	}
1730 
1731 #ifdef CONFIG_RFS_ACCEL
1732 	if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
1733 		efx->net_dev->rx_cpu_rmap =
1734 			alloc_irq_cpu_rmap(efx->n_rx_channels);
1735 		if (!efx->net_dev->rx_cpu_rmap) {
1736 			rc = -ENOMEM;
1737 			goto fail1;
1738 		}
1739 	}
1740 #endif
1741 
1742 	/* Hook MSI or MSI-X interrupt */
1743 	n_irqs = 0;
1744 	efx_for_each_channel(channel, efx) {
1745 		rc = request_irq(channel->irq, efx_msi_interrupt,
1746 				 IRQF_PROBE_SHARED, /* Not shared */
1747 				 efx->channel_name[channel->channel],
1748 				 &efx->channel[channel->channel]);
1749 		if (rc) {
1750 			netif_err(efx, drv, efx->net_dev,
1751 				  "failed to hook IRQ %d\n", channel->irq);
1752 			goto fail2;
1753 		}
1754 		++n_irqs;
1755 
1756 #ifdef CONFIG_RFS_ACCEL
1757 		if (efx->interrupt_mode == EFX_INT_MODE_MSIX &&
1758 		    channel->channel < efx->n_rx_channels) {
1759 			rc = irq_cpu_rmap_add(efx->net_dev->rx_cpu_rmap,
1760 					      channel->irq);
1761 			if (rc)
1762 				goto fail2;
1763 		}
1764 #endif
1765 	}
1766 
1767 	return 0;
1768 
1769  fail2:
1770 #ifdef CONFIG_RFS_ACCEL
1771 	free_irq_cpu_rmap(efx->net_dev->rx_cpu_rmap);
1772 	efx->net_dev->rx_cpu_rmap = NULL;
1773 #endif
1774 	efx_for_each_channel(channel, efx) {
1775 		if (n_irqs-- == 0)
1776 			break;
1777 		free_irq(channel->irq, &efx->channel[channel->channel]);
1778 	}
1779  fail1:
1780 	return rc;
1781 }
1782 
1783 void efx_nic_fini_interrupt(struct efx_nic *efx)
1784 {
1785 	struct efx_channel *channel;
1786 	efx_oword_t reg;
1787 
1788 #ifdef CONFIG_RFS_ACCEL
1789 	free_irq_cpu_rmap(efx->net_dev->rx_cpu_rmap);
1790 	efx->net_dev->rx_cpu_rmap = NULL;
1791 #endif
1792 
1793 	/* Disable MSI/MSI-X interrupts */
1794 	efx_for_each_channel(channel, efx)
1795 		free_irq(channel->irq, &efx->channel[channel->channel]);
1796 
1797 	/* ACK legacy interrupt */
1798 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1799 		efx_reado(efx, &reg, FR_BZ_INT_ISR0);
1800 	else
1801 		falcon_irq_ack_a1(efx);
1802 
1803 	/* Disable legacy interrupt */
1804 	if (efx->legacy_irq)
1805 		free_irq(efx->legacy_irq, efx);
1806 }
1807 
1808 /* Looks at available SRAM resources and works out how many queues we
1809  * can support, and where things like descriptor caches should live.
1810  *
1811  * SRAM is split up as follows:
1812  * 0                          buftbl entries for channels
1813  * efx->vf_buftbl_base        buftbl entries for SR-IOV
1814  * efx->rx_dc_base            RX descriptor caches
1815  * efx->tx_dc_base            TX descriptor caches
1816  */
1817 void efx_nic_dimension_resources(struct efx_nic *efx, unsigned sram_lim_qw)
1818 {
1819 	unsigned vi_count, buftbl_min;
1820 
1821 	/* Account for the buffer table entries backing the datapath channels
1822 	 * and the descriptor caches for those channels.
1823 	 */
1824 	buftbl_min = ((efx->n_rx_channels * EFX_MAX_DMAQ_SIZE +
1825 		       efx->n_tx_channels * EFX_TXQ_TYPES * EFX_MAX_DMAQ_SIZE +
1826 		       efx->n_channels * EFX_MAX_EVQ_SIZE)
1827 		      * sizeof(efx_qword_t) / EFX_BUF_SIZE);
1828 	vi_count = max(efx->n_channels, efx->n_tx_channels * EFX_TXQ_TYPES);
1829 
1830 #ifdef CONFIG_SFC_SRIOV
1831 	if (efx_sriov_wanted(efx)) {
1832 		unsigned vi_dc_entries, buftbl_free, entries_per_vf, vf_limit;
1833 
1834 		efx->vf_buftbl_base = buftbl_min;
1835 
1836 		vi_dc_entries = RX_DC_ENTRIES + TX_DC_ENTRIES;
1837 		vi_count = max(vi_count, EFX_VI_BASE);
1838 		buftbl_free = (sram_lim_qw - buftbl_min -
1839 			       vi_count * vi_dc_entries);
1840 
1841 		entries_per_vf = ((vi_dc_entries + EFX_VF_BUFTBL_PER_VI) *
1842 				  efx_vf_size(efx));
1843 		vf_limit = min(buftbl_free / entries_per_vf,
1844 			       (1024U - EFX_VI_BASE) >> efx->vi_scale);
1845 
1846 		if (efx->vf_count > vf_limit) {
1847 			netif_err(efx, probe, efx->net_dev,
1848 				  "Reducing VF count from from %d to %d\n",
1849 				  efx->vf_count, vf_limit);
1850 			efx->vf_count = vf_limit;
1851 		}
1852 		vi_count += efx->vf_count * efx_vf_size(efx);
1853 	}
1854 #endif
1855 
1856 	efx->tx_dc_base = sram_lim_qw - vi_count * TX_DC_ENTRIES;
1857 	efx->rx_dc_base = efx->tx_dc_base - vi_count * RX_DC_ENTRIES;
1858 }
1859 
1860 u32 efx_nic_fpga_ver(struct efx_nic *efx)
1861 {
1862 	efx_oword_t altera_build;
1863 	efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
1864 	return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
1865 }
1866 
1867 void efx_nic_init_common(struct efx_nic *efx)
1868 {
1869 	efx_oword_t temp;
1870 
1871 	/* Set positions of descriptor caches in SRAM. */
1872 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR, efx->tx_dc_base);
1873 	efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
1874 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR, efx->rx_dc_base);
1875 	efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
1876 
1877 	/* Set TX descriptor cache size. */
1878 	BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
1879 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
1880 	efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
1881 
1882 	/* Set RX descriptor cache size.  Set low watermark to size-8, as
1883 	 * this allows most efficient prefetching.
1884 	 */
1885 	BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
1886 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
1887 	efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
1888 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
1889 	efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
1890 
1891 	/* Program INT_KER address */
1892 	EFX_POPULATE_OWORD_2(temp,
1893 			     FRF_AZ_NORM_INT_VEC_DIS_KER,
1894 			     EFX_INT_MODE_USE_MSI(efx),
1895 			     FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
1896 	efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
1897 
1898 	if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
1899 		/* Use an interrupt level unused by event queues */
1900 		efx->irq_level = 0x1f;
1901 	else
1902 		/* Use a valid MSI-X vector */
1903 		efx->irq_level = 0;
1904 
1905 	/* Enable all the genuinely fatal interrupts.  (They are still
1906 	 * masked by the overall interrupt mask, controlled by
1907 	 * falcon_interrupts()).
1908 	 *
1909 	 * Note: All other fatal interrupts are enabled
1910 	 */
1911 	EFX_POPULATE_OWORD_3(temp,
1912 			     FRF_AZ_ILL_ADR_INT_KER_EN, 1,
1913 			     FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
1914 			     FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
1915 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1916 		EFX_SET_OWORD_FIELD(temp, FRF_CZ_SRAM_PERR_INT_P_KER_EN, 1);
1917 	EFX_INVERT_OWORD(temp);
1918 	efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
1919 
1920 	efx_nic_push_rx_indir_table(efx);
1921 
1922 	/* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
1923 	 * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
1924 	 */
1925 	efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
1926 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
1927 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
1928 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
1929 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 1);
1930 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
1931 	/* Enable SW_EV to inherit in char driver - assume harmless here */
1932 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
1933 	/* Prefetch threshold 2 => fetch when descriptor cache half empty */
1934 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
1935 	/* Disable hardware watchdog which can misfire */
1936 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
1937 	/* Squash TX of packets of 16 bytes or less */
1938 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1939 		EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
1940 	efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
1941 
1942 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
1943 		EFX_POPULATE_OWORD_4(temp,
1944 				     /* Default values */
1945 				     FRF_BZ_TX_PACE_SB_NOT_AF, 0x15,
1946 				     FRF_BZ_TX_PACE_SB_AF, 0xb,
1947 				     FRF_BZ_TX_PACE_FB_BASE, 0,
1948 				     /* Allow large pace values in the
1949 				      * fast bin. */
1950 				     FRF_BZ_TX_PACE_BIN_TH,
1951 				     FFE_BZ_TX_PACE_RESERVED);
1952 		efx_writeo(efx, &temp, FR_BZ_TX_PACE);
1953 	}
1954 }
1955 
1956 /* Register dump */
1957 
1958 #define REGISTER_REVISION_A	1
1959 #define REGISTER_REVISION_B	2
1960 #define REGISTER_REVISION_C	3
1961 #define REGISTER_REVISION_Z	3	/* latest revision */
1962 
1963 struct efx_nic_reg {
1964 	u32 offset:24;
1965 	u32 min_revision:2, max_revision:2;
1966 };
1967 
1968 #define REGISTER(name, min_rev, max_rev) {				\
1969 	FR_ ## min_rev ## max_rev ## _ ## name,				\
1970 	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev	\
1971 }
1972 #define REGISTER_AA(name) REGISTER(name, A, A)
1973 #define REGISTER_AB(name) REGISTER(name, A, B)
1974 #define REGISTER_AZ(name) REGISTER(name, A, Z)
1975 #define REGISTER_BB(name) REGISTER(name, B, B)
1976 #define REGISTER_BZ(name) REGISTER(name, B, Z)
1977 #define REGISTER_CZ(name) REGISTER(name, C, Z)
1978 
1979 static const struct efx_nic_reg efx_nic_regs[] = {
1980 	REGISTER_AZ(ADR_REGION),
1981 	REGISTER_AZ(INT_EN_KER),
1982 	REGISTER_BZ(INT_EN_CHAR),
1983 	REGISTER_AZ(INT_ADR_KER),
1984 	REGISTER_BZ(INT_ADR_CHAR),
1985 	/* INT_ACK_KER is WO */
1986 	/* INT_ISR0 is RC */
1987 	REGISTER_AZ(HW_INIT),
1988 	REGISTER_CZ(USR_EV_CFG),
1989 	REGISTER_AB(EE_SPI_HCMD),
1990 	REGISTER_AB(EE_SPI_HADR),
1991 	REGISTER_AB(EE_SPI_HDATA),
1992 	REGISTER_AB(EE_BASE_PAGE),
1993 	REGISTER_AB(EE_VPD_CFG0),
1994 	/* EE_VPD_SW_CNTL and EE_VPD_SW_DATA are not used */
1995 	/* PMBX_DBG_IADDR and PBMX_DBG_IDATA are indirect */
1996 	/* PCIE_CORE_INDIRECT is indirect */
1997 	REGISTER_AB(NIC_STAT),
1998 	REGISTER_AB(GPIO_CTL),
1999 	REGISTER_AB(GLB_CTL),
2000 	/* FATAL_INTR_KER and FATAL_INTR_CHAR are partly RC */
2001 	REGISTER_BZ(DP_CTRL),
2002 	REGISTER_AZ(MEM_STAT),
2003 	REGISTER_AZ(CS_DEBUG),
2004 	REGISTER_AZ(ALTERA_BUILD),
2005 	REGISTER_AZ(CSR_SPARE),
2006 	REGISTER_AB(PCIE_SD_CTL0123),
2007 	REGISTER_AB(PCIE_SD_CTL45),
2008 	REGISTER_AB(PCIE_PCS_CTL_STAT),
2009 	/* DEBUG_DATA_OUT is not used */
2010 	/* DRV_EV is WO */
2011 	REGISTER_AZ(EVQ_CTL),
2012 	REGISTER_AZ(EVQ_CNT1),
2013 	REGISTER_AZ(EVQ_CNT2),
2014 	REGISTER_AZ(BUF_TBL_CFG),
2015 	REGISTER_AZ(SRM_RX_DC_CFG),
2016 	REGISTER_AZ(SRM_TX_DC_CFG),
2017 	REGISTER_AZ(SRM_CFG),
2018 	/* BUF_TBL_UPD is WO */
2019 	REGISTER_AZ(SRM_UPD_EVQ),
2020 	REGISTER_AZ(SRAM_PARITY),
2021 	REGISTER_AZ(RX_CFG),
2022 	REGISTER_BZ(RX_FILTER_CTL),
2023 	/* RX_FLUSH_DESCQ is WO */
2024 	REGISTER_AZ(RX_DC_CFG),
2025 	REGISTER_AZ(RX_DC_PF_WM),
2026 	REGISTER_BZ(RX_RSS_TKEY),
2027 	/* RX_NODESC_DROP is RC */
2028 	REGISTER_AA(RX_SELF_RST),
2029 	/* RX_DEBUG, RX_PUSH_DROP are not used */
2030 	REGISTER_CZ(RX_RSS_IPV6_REG1),
2031 	REGISTER_CZ(RX_RSS_IPV6_REG2),
2032 	REGISTER_CZ(RX_RSS_IPV6_REG3),
2033 	/* TX_FLUSH_DESCQ is WO */
2034 	REGISTER_AZ(TX_DC_CFG),
2035 	REGISTER_AA(TX_CHKSM_CFG),
2036 	REGISTER_AZ(TX_CFG),
2037 	/* TX_PUSH_DROP is not used */
2038 	REGISTER_AZ(TX_RESERVED),
2039 	REGISTER_BZ(TX_PACE),
2040 	/* TX_PACE_DROP_QID is RC */
2041 	REGISTER_BB(TX_VLAN),
2042 	REGISTER_BZ(TX_IPFIL_PORTEN),
2043 	REGISTER_AB(MD_TXD),
2044 	REGISTER_AB(MD_RXD),
2045 	REGISTER_AB(MD_CS),
2046 	REGISTER_AB(MD_PHY_ADR),
2047 	REGISTER_AB(MD_ID),
2048 	/* MD_STAT is RC */
2049 	REGISTER_AB(MAC_STAT_DMA),
2050 	REGISTER_AB(MAC_CTRL),
2051 	REGISTER_BB(GEN_MODE),
2052 	REGISTER_AB(MAC_MC_HASH_REG0),
2053 	REGISTER_AB(MAC_MC_HASH_REG1),
2054 	REGISTER_AB(GM_CFG1),
2055 	REGISTER_AB(GM_CFG2),
2056 	/* GM_IPG and GM_HD are not used */
2057 	REGISTER_AB(GM_MAX_FLEN),
2058 	/* GM_TEST is not used */
2059 	REGISTER_AB(GM_ADR1),
2060 	REGISTER_AB(GM_ADR2),
2061 	REGISTER_AB(GMF_CFG0),
2062 	REGISTER_AB(GMF_CFG1),
2063 	REGISTER_AB(GMF_CFG2),
2064 	REGISTER_AB(GMF_CFG3),
2065 	REGISTER_AB(GMF_CFG4),
2066 	REGISTER_AB(GMF_CFG5),
2067 	REGISTER_BB(TX_SRC_MAC_CTL),
2068 	REGISTER_AB(XM_ADR_LO),
2069 	REGISTER_AB(XM_ADR_HI),
2070 	REGISTER_AB(XM_GLB_CFG),
2071 	REGISTER_AB(XM_TX_CFG),
2072 	REGISTER_AB(XM_RX_CFG),
2073 	REGISTER_AB(XM_MGT_INT_MASK),
2074 	REGISTER_AB(XM_FC),
2075 	REGISTER_AB(XM_PAUSE_TIME),
2076 	REGISTER_AB(XM_TX_PARAM),
2077 	REGISTER_AB(XM_RX_PARAM),
2078 	/* XM_MGT_INT_MSK (note no 'A') is RC */
2079 	REGISTER_AB(XX_PWR_RST),
2080 	REGISTER_AB(XX_SD_CTL),
2081 	REGISTER_AB(XX_TXDRV_CTL),
2082 	/* XX_PRBS_CTL, XX_PRBS_CHK and XX_PRBS_ERR are not used */
2083 	/* XX_CORE_STAT is partly RC */
2084 };
2085 
2086 struct efx_nic_reg_table {
2087 	u32 offset:24;
2088 	u32 min_revision:2, max_revision:2;
2089 	u32 step:6, rows:21;
2090 };
2091 
2092 #define REGISTER_TABLE_DIMENSIONS(_, offset, min_rev, max_rev, step, rows) { \
2093 	offset,								\
2094 	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev,	\
2095 	step, rows							\
2096 }
2097 #define REGISTER_TABLE(name, min_rev, max_rev)				\
2098 	REGISTER_TABLE_DIMENSIONS(					\
2099 		name, FR_ ## min_rev ## max_rev ## _ ## name,		\
2100 		min_rev, max_rev,					\
2101 		FR_ ## min_rev ## max_rev ## _ ## name ## _STEP,	\
2102 		FR_ ## min_rev ## max_rev ## _ ## name ## _ROWS)
2103 #define REGISTER_TABLE_AA(name) REGISTER_TABLE(name, A, A)
2104 #define REGISTER_TABLE_AZ(name) REGISTER_TABLE(name, A, Z)
2105 #define REGISTER_TABLE_BB(name) REGISTER_TABLE(name, B, B)
2106 #define REGISTER_TABLE_BZ(name) REGISTER_TABLE(name, B, Z)
2107 #define REGISTER_TABLE_BB_CZ(name)					\
2108 	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, B, B,		\
2109 				  FR_BZ_ ## name ## _STEP,		\
2110 				  FR_BB_ ## name ## _ROWS),		\
2111 	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, C, Z,		\
2112 				  FR_BZ_ ## name ## _STEP,		\
2113 				  FR_CZ_ ## name ## _ROWS)
2114 #define REGISTER_TABLE_CZ(name) REGISTER_TABLE(name, C, Z)
2115 
2116 static const struct efx_nic_reg_table efx_nic_reg_tables[] = {
2117 	/* DRIVER is not used */
2118 	/* EVQ_RPTR, TIMER_COMMAND, USR_EV and {RX,TX}_DESC_UPD are WO */
2119 	REGISTER_TABLE_BB(TX_IPFIL_TBL),
2120 	REGISTER_TABLE_BB(TX_SRC_MAC_TBL),
2121 	REGISTER_TABLE_AA(RX_DESC_PTR_TBL_KER),
2122 	REGISTER_TABLE_BB_CZ(RX_DESC_PTR_TBL),
2123 	REGISTER_TABLE_AA(TX_DESC_PTR_TBL_KER),
2124 	REGISTER_TABLE_BB_CZ(TX_DESC_PTR_TBL),
2125 	REGISTER_TABLE_AA(EVQ_PTR_TBL_KER),
2126 	REGISTER_TABLE_BB_CZ(EVQ_PTR_TBL),
2127 	/* We can't reasonably read all of the buffer table (up to 8MB!).
2128 	 * However this driver will only use a few entries.  Reading
2129 	 * 1K entries allows for some expansion of queue count and
2130 	 * size before we need to change the version. */
2131 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL_KER, FR_AA_BUF_FULL_TBL_KER,
2132 				  A, A, 8, 1024),
2133 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL, FR_BZ_BUF_FULL_TBL,
2134 				  B, Z, 8, 1024),
2135 	REGISTER_TABLE_CZ(RX_MAC_FILTER_TBL0),
2136 	REGISTER_TABLE_BB_CZ(TIMER_TBL),
2137 	REGISTER_TABLE_BB_CZ(TX_PACE_TBL),
2138 	REGISTER_TABLE_BZ(RX_INDIRECTION_TBL),
2139 	/* TX_FILTER_TBL0 is huge and not used by this driver */
2140 	REGISTER_TABLE_CZ(TX_MAC_FILTER_TBL0),
2141 	REGISTER_TABLE_CZ(MC_TREG_SMEM),
2142 	/* MSIX_PBA_TABLE is not mapped */
2143 	/* SRM_DBG is not mapped (and is redundant with BUF_FLL_TBL) */
2144 	REGISTER_TABLE_BZ(RX_FILTER_TBL0),
2145 };
2146 
2147 size_t efx_nic_get_regs_len(struct efx_nic *efx)
2148 {
2149 	const struct efx_nic_reg *reg;
2150 	const struct efx_nic_reg_table *table;
2151 	size_t len = 0;
2152 
2153 	for (reg = efx_nic_regs;
2154 	     reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
2155 	     reg++)
2156 		if (efx->type->revision >= reg->min_revision &&
2157 		    efx->type->revision <= reg->max_revision)
2158 			len += sizeof(efx_oword_t);
2159 
2160 	for (table = efx_nic_reg_tables;
2161 	     table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
2162 	     table++)
2163 		if (efx->type->revision >= table->min_revision &&
2164 		    efx->type->revision <= table->max_revision)
2165 			len += table->rows * min_t(size_t, table->step, 16);
2166 
2167 	return len;
2168 }
2169 
2170 void efx_nic_get_regs(struct efx_nic *efx, void *buf)
2171 {
2172 	const struct efx_nic_reg *reg;
2173 	const struct efx_nic_reg_table *table;
2174 
2175 	for (reg = efx_nic_regs;
2176 	     reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
2177 	     reg++) {
2178 		if (efx->type->revision >= reg->min_revision &&
2179 		    efx->type->revision <= reg->max_revision) {
2180 			efx_reado(efx, (efx_oword_t *)buf, reg->offset);
2181 			buf += sizeof(efx_oword_t);
2182 		}
2183 	}
2184 
2185 	for (table = efx_nic_reg_tables;
2186 	     table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
2187 	     table++) {
2188 		size_t size, i;
2189 
2190 		if (!(efx->type->revision >= table->min_revision &&
2191 		      efx->type->revision <= table->max_revision))
2192 			continue;
2193 
2194 		size = min_t(size_t, table->step, 16);
2195 
2196 		for (i = 0; i < table->rows; i++) {
2197 			switch (table->step) {
2198 			case 4: /* 32-bit SRAM */
2199 				efx_readd(efx, buf, table->offset + 4 * i);
2200 				break;
2201 			case 8: /* 64-bit SRAM */
2202 				efx_sram_readq(efx,
2203 					       efx->membase + table->offset,
2204 					       buf, i);
2205 				break;
2206 			case 16: /* 128-bit-readable register */
2207 				efx_reado_table(efx, buf, table->offset, i);
2208 				break;
2209 			case 32: /* 128-bit register, interleaved */
2210 				efx_reado_table(efx, buf, table->offset, 2 * i);
2211 				break;
2212 			default:
2213 				WARN_ON(1);
2214 				return;
2215 			}
2216 			buf += size;
2217 		}
2218 	}
2219 }
2220