xref: /openbmc/linux/drivers/net/ethernet/sfc/nic.c (revision d8aec745)
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 "farch_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.buf.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->buf.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->buf.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 	if (efx_nic_alloc_buffer(efx, &buffer->buf, len, GFP_KERNEL))
254 		return -ENOMEM;
255 	buffer->entries = len / EFX_BUF_SIZE;
256 	BUG_ON(buffer->buf.dma_addr & (EFX_BUF_SIZE - 1));
257 
258 	/* Select new buffer ID */
259 	buffer->index = efx->next_buffer_table;
260 	efx->next_buffer_table += buffer->entries;
261 #ifdef CONFIG_SFC_SRIOV
262 	BUG_ON(efx_sriov_enabled(efx) &&
263 	       efx->vf_buftbl_base < efx->next_buffer_table);
264 #endif
265 
266 	netif_dbg(efx, probe, efx->net_dev,
267 		  "allocating special buffers %d-%d at %llx+%x "
268 		  "(virt %p phys %llx)\n", buffer->index,
269 		  buffer->index + buffer->entries - 1,
270 		  (u64)buffer->buf.dma_addr, len,
271 		  buffer->buf.addr, (u64)virt_to_phys(buffer->buf.addr));
272 
273 	return 0;
274 }
275 
276 static void
277 efx_free_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
278 {
279 	if (!buffer->buf.addr)
280 		return;
281 
282 	netif_dbg(efx, hw, efx->net_dev,
283 		  "deallocating special buffers %d-%d at %llx+%x "
284 		  "(virt %p phys %llx)\n", buffer->index,
285 		  buffer->index + buffer->entries - 1,
286 		  (u64)buffer->buf.dma_addr, buffer->buf.len,
287 		  buffer->buf.addr, (u64)virt_to_phys(buffer->buf.addr));
288 
289 	efx_nic_free_buffer(efx, &buffer->buf);
290 	buffer->entries = 0;
291 }
292 
293 /**************************************************************************
294  *
295  * Generic buffer handling
296  * These buffers are used for interrupt status, MAC stats, etc.
297  *
298  **************************************************************************/
299 
300 int efx_nic_alloc_buffer(struct efx_nic *efx, struct efx_buffer *buffer,
301 			 unsigned int len, gfp_t gfp_flags)
302 {
303 	buffer->addr = dma_alloc_coherent(&efx->pci_dev->dev, len,
304 					  &buffer->dma_addr,
305 					  gfp_flags | __GFP_ZERO);
306 	if (!buffer->addr)
307 		return -ENOMEM;
308 	buffer->len = len;
309 	return 0;
310 }
311 
312 void efx_nic_free_buffer(struct efx_nic *efx, struct efx_buffer *buffer)
313 {
314 	if (buffer->addr) {
315 		dma_free_coherent(&efx->pci_dev->dev, buffer->len,
316 				  buffer->addr, buffer->dma_addr);
317 		buffer->addr = NULL;
318 	}
319 }
320 
321 /**************************************************************************
322  *
323  * TX path
324  *
325  **************************************************************************/
326 
327 /* Returns a pointer to the specified transmit descriptor in the TX
328  * descriptor queue belonging to the specified channel.
329  */
330 static inline efx_qword_t *
331 efx_tx_desc(struct efx_tx_queue *tx_queue, unsigned int index)
332 {
333 	return ((efx_qword_t *) (tx_queue->txd.buf.addr)) + index;
334 }
335 
336 /* This writes to the TX_DESC_WPTR; write pointer for TX descriptor ring */
337 static inline void efx_notify_tx_desc(struct efx_tx_queue *tx_queue)
338 {
339 	unsigned write_ptr;
340 	efx_dword_t reg;
341 
342 	write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
343 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_TX_DESC_WPTR_DWORD, write_ptr);
344 	efx_writed_page(tx_queue->efx, &reg,
345 			FR_AZ_TX_DESC_UPD_DWORD_P0, tx_queue->queue);
346 }
347 
348 /* Write pointer and first descriptor for TX descriptor ring */
349 static inline void efx_push_tx_desc(struct efx_tx_queue *tx_queue,
350 				    const efx_qword_t *txd)
351 {
352 	unsigned write_ptr;
353 	efx_oword_t reg;
354 
355 	BUILD_BUG_ON(FRF_AZ_TX_DESC_LBN != 0);
356 	BUILD_BUG_ON(FR_AA_TX_DESC_UPD_KER != FR_BZ_TX_DESC_UPD_P0);
357 
358 	write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
359 	EFX_POPULATE_OWORD_2(reg, FRF_AZ_TX_DESC_PUSH_CMD, true,
360 			     FRF_AZ_TX_DESC_WPTR, write_ptr);
361 	reg.qword[0] = *txd;
362 	efx_writeo_page(tx_queue->efx, &reg,
363 			FR_BZ_TX_DESC_UPD_P0, tx_queue->queue);
364 }
365 
366 static inline bool
367 efx_may_push_tx_desc(struct efx_tx_queue *tx_queue, unsigned int write_count)
368 {
369 	unsigned empty_read_count = ACCESS_ONCE(tx_queue->empty_read_count);
370 
371 	if (empty_read_count == 0)
372 		return false;
373 
374 	tx_queue->empty_read_count = 0;
375 	return ((empty_read_count ^ write_count) & ~EFX_EMPTY_COUNT_VALID) == 0
376 		&& tx_queue->write_count - write_count == 1;
377 }
378 
379 /* For each entry inserted into the software descriptor ring, create a
380  * descriptor in the hardware TX descriptor ring (in host memory), and
381  * write a doorbell.
382  */
383 void efx_nic_push_buffers(struct efx_tx_queue *tx_queue)
384 {
385 
386 	struct efx_tx_buffer *buffer;
387 	efx_qword_t *txd;
388 	unsigned write_ptr;
389 	unsigned old_write_count = tx_queue->write_count;
390 
391 	BUG_ON(tx_queue->write_count == tx_queue->insert_count);
392 
393 	do {
394 		write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
395 		buffer = &tx_queue->buffer[write_ptr];
396 		txd = efx_tx_desc(tx_queue, write_ptr);
397 		++tx_queue->write_count;
398 
399 		/* Create TX descriptor ring entry */
400 		BUILD_BUG_ON(EFX_TX_BUF_CONT != 1);
401 		EFX_POPULATE_QWORD_4(*txd,
402 				     FSF_AZ_TX_KER_CONT,
403 				     buffer->flags & EFX_TX_BUF_CONT,
404 				     FSF_AZ_TX_KER_BYTE_COUNT, buffer->len,
405 				     FSF_AZ_TX_KER_BUF_REGION, 0,
406 				     FSF_AZ_TX_KER_BUF_ADDR, buffer->dma_addr);
407 	} while (tx_queue->write_count != tx_queue->insert_count);
408 
409 	wmb(); /* Ensure descriptors are written before they are fetched */
410 
411 	if (efx_may_push_tx_desc(tx_queue, old_write_count)) {
412 		txd = efx_tx_desc(tx_queue,
413 				  old_write_count & tx_queue->ptr_mask);
414 		efx_push_tx_desc(tx_queue, txd);
415 		++tx_queue->pushes;
416 	} else {
417 		efx_notify_tx_desc(tx_queue);
418 	}
419 }
420 
421 /* Allocate hardware resources for a TX queue */
422 int efx_nic_probe_tx(struct efx_tx_queue *tx_queue)
423 {
424 	struct efx_nic *efx = tx_queue->efx;
425 	unsigned entries;
426 
427 	entries = tx_queue->ptr_mask + 1;
428 	return efx_alloc_special_buffer(efx, &tx_queue->txd,
429 					entries * sizeof(efx_qword_t));
430 }
431 
432 void efx_nic_init_tx(struct efx_tx_queue *tx_queue)
433 {
434 	struct efx_nic *efx = tx_queue->efx;
435 	efx_oword_t reg;
436 
437 	/* Pin TX descriptor ring */
438 	efx_init_special_buffer(efx, &tx_queue->txd);
439 
440 	/* Push TX descriptor ring to card */
441 	EFX_POPULATE_OWORD_10(reg,
442 			      FRF_AZ_TX_DESCQ_EN, 1,
443 			      FRF_AZ_TX_ISCSI_DDIG_EN, 0,
444 			      FRF_AZ_TX_ISCSI_HDIG_EN, 0,
445 			      FRF_AZ_TX_DESCQ_BUF_BASE_ID, tx_queue->txd.index,
446 			      FRF_AZ_TX_DESCQ_EVQ_ID,
447 			      tx_queue->channel->channel,
448 			      FRF_AZ_TX_DESCQ_OWNER_ID, 0,
449 			      FRF_AZ_TX_DESCQ_LABEL, tx_queue->queue,
450 			      FRF_AZ_TX_DESCQ_SIZE,
451 			      __ffs(tx_queue->txd.entries),
452 			      FRF_AZ_TX_DESCQ_TYPE, 0,
453 			      FRF_BZ_TX_NON_IP_DROP_DIS, 1);
454 
455 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
456 		int csum = tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD;
457 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_IP_CHKSM_DIS, !csum);
458 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_TCP_CHKSM_DIS,
459 				    !csum);
460 	}
461 
462 	efx_writeo_table(efx, &reg, efx->type->txd_ptr_tbl_base,
463 			 tx_queue->queue);
464 
465 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0) {
466 		/* Only 128 bits in this register */
467 		BUILD_BUG_ON(EFX_MAX_TX_QUEUES > 128);
468 
469 		efx_reado(efx, &reg, FR_AA_TX_CHKSM_CFG);
470 		if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD)
471 			__clear_bit_le(tx_queue->queue, &reg);
472 		else
473 			__set_bit_le(tx_queue->queue, &reg);
474 		efx_writeo(efx, &reg, FR_AA_TX_CHKSM_CFG);
475 	}
476 
477 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
478 		EFX_POPULATE_OWORD_1(reg,
479 				     FRF_BZ_TX_PACE,
480 				     (tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
481 				     FFE_BZ_TX_PACE_OFF :
482 				     FFE_BZ_TX_PACE_RESERVED);
483 		efx_writeo_table(efx, &reg, FR_BZ_TX_PACE_TBL,
484 				 tx_queue->queue);
485 	}
486 }
487 
488 static void efx_flush_tx_queue(struct efx_tx_queue *tx_queue)
489 {
490 	struct efx_nic *efx = tx_queue->efx;
491 	efx_oword_t tx_flush_descq;
492 
493 	WARN_ON(atomic_read(&tx_queue->flush_outstanding));
494 	atomic_set(&tx_queue->flush_outstanding, 1);
495 
496 	EFX_POPULATE_OWORD_2(tx_flush_descq,
497 			     FRF_AZ_TX_FLUSH_DESCQ_CMD, 1,
498 			     FRF_AZ_TX_FLUSH_DESCQ, tx_queue->queue);
499 	efx_writeo(efx, &tx_flush_descq, FR_AZ_TX_FLUSH_DESCQ);
500 }
501 
502 void efx_nic_fini_tx(struct efx_tx_queue *tx_queue)
503 {
504 	struct efx_nic *efx = tx_queue->efx;
505 	efx_oword_t tx_desc_ptr;
506 
507 	/* Remove TX descriptor ring from card */
508 	EFX_ZERO_OWORD(tx_desc_ptr);
509 	efx_writeo_table(efx, &tx_desc_ptr, efx->type->txd_ptr_tbl_base,
510 			 tx_queue->queue);
511 
512 	/* Unpin TX descriptor ring */
513 	efx_fini_special_buffer(efx, &tx_queue->txd);
514 }
515 
516 /* Free buffers backing TX queue */
517 void efx_nic_remove_tx(struct efx_tx_queue *tx_queue)
518 {
519 	efx_free_special_buffer(tx_queue->efx, &tx_queue->txd);
520 }
521 
522 /**************************************************************************
523  *
524  * RX path
525  *
526  **************************************************************************/
527 
528 /* Returns a pointer to the specified descriptor in the RX descriptor queue */
529 static inline efx_qword_t *
530 efx_rx_desc(struct efx_rx_queue *rx_queue, unsigned int index)
531 {
532 	return ((efx_qword_t *) (rx_queue->rxd.buf.addr)) + index;
533 }
534 
535 /* This creates an entry in the RX descriptor queue */
536 static inline void
537 efx_build_rx_desc(struct efx_rx_queue *rx_queue, unsigned index)
538 {
539 	struct efx_rx_buffer *rx_buf;
540 	efx_qword_t *rxd;
541 
542 	rxd = efx_rx_desc(rx_queue, index);
543 	rx_buf = efx_rx_buffer(rx_queue, index);
544 	EFX_POPULATE_QWORD_3(*rxd,
545 			     FSF_AZ_RX_KER_BUF_SIZE,
546 			     rx_buf->len -
547 			     rx_queue->efx->type->rx_buffer_padding,
548 			     FSF_AZ_RX_KER_BUF_REGION, 0,
549 			     FSF_AZ_RX_KER_BUF_ADDR, rx_buf->dma_addr);
550 }
551 
552 /* This writes to the RX_DESC_WPTR register for the specified receive
553  * descriptor ring.
554  */
555 void efx_nic_notify_rx_desc(struct efx_rx_queue *rx_queue)
556 {
557 	struct efx_nic *efx = rx_queue->efx;
558 	efx_dword_t reg;
559 	unsigned write_ptr;
560 
561 	while (rx_queue->notified_count != rx_queue->added_count) {
562 		efx_build_rx_desc(
563 			rx_queue,
564 			rx_queue->notified_count & rx_queue->ptr_mask);
565 		++rx_queue->notified_count;
566 	}
567 
568 	wmb();
569 	write_ptr = rx_queue->added_count & rx_queue->ptr_mask;
570 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_RX_DESC_WPTR_DWORD, write_ptr);
571 	efx_writed_page(efx, &reg, FR_AZ_RX_DESC_UPD_DWORD_P0,
572 			efx_rx_queue_index(rx_queue));
573 }
574 
575 int efx_nic_probe_rx(struct efx_rx_queue *rx_queue)
576 {
577 	struct efx_nic *efx = rx_queue->efx;
578 	unsigned entries;
579 
580 	entries = rx_queue->ptr_mask + 1;
581 	return efx_alloc_special_buffer(efx, &rx_queue->rxd,
582 					entries * sizeof(efx_qword_t));
583 }
584 
585 void efx_nic_init_rx(struct efx_rx_queue *rx_queue)
586 {
587 	efx_oword_t rx_desc_ptr;
588 	struct efx_nic *efx = rx_queue->efx;
589 	bool is_b0 = efx_nic_rev(efx) >= EFX_REV_FALCON_B0;
590 	bool iscsi_digest_en = is_b0;
591 	bool jumbo_en;
592 
593 	/* For kernel-mode queues in Falcon A1, the JUMBO flag enables
594 	 * DMA to continue after a PCIe page boundary (and scattering
595 	 * is not possible).  In Falcon B0 and Siena, it enables
596 	 * scatter.
597 	 */
598 	jumbo_en = !is_b0 || efx->rx_scatter;
599 
600 	netif_dbg(efx, hw, efx->net_dev,
601 		  "RX queue %d ring in special buffers %d-%d\n",
602 		  efx_rx_queue_index(rx_queue), rx_queue->rxd.index,
603 		  rx_queue->rxd.index + rx_queue->rxd.entries - 1);
604 
605 	rx_queue->scatter_n = 0;
606 
607 	/* Pin RX descriptor ring */
608 	efx_init_special_buffer(efx, &rx_queue->rxd);
609 
610 	/* Push RX descriptor ring to card */
611 	EFX_POPULATE_OWORD_10(rx_desc_ptr,
612 			      FRF_AZ_RX_ISCSI_DDIG_EN, iscsi_digest_en,
613 			      FRF_AZ_RX_ISCSI_HDIG_EN, iscsi_digest_en,
614 			      FRF_AZ_RX_DESCQ_BUF_BASE_ID, rx_queue->rxd.index,
615 			      FRF_AZ_RX_DESCQ_EVQ_ID,
616 			      efx_rx_queue_channel(rx_queue)->channel,
617 			      FRF_AZ_RX_DESCQ_OWNER_ID, 0,
618 			      FRF_AZ_RX_DESCQ_LABEL,
619 			      efx_rx_queue_index(rx_queue),
620 			      FRF_AZ_RX_DESCQ_SIZE,
621 			      __ffs(rx_queue->rxd.entries),
622 			      FRF_AZ_RX_DESCQ_TYPE, 0 /* kernel queue */ ,
623 			      FRF_AZ_RX_DESCQ_JUMBO, jumbo_en,
624 			      FRF_AZ_RX_DESCQ_EN, 1);
625 	efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
626 			 efx_rx_queue_index(rx_queue));
627 }
628 
629 static void efx_flush_rx_queue(struct efx_rx_queue *rx_queue)
630 {
631 	struct efx_nic *efx = rx_queue->efx;
632 	efx_oword_t rx_flush_descq;
633 
634 	EFX_POPULATE_OWORD_2(rx_flush_descq,
635 			     FRF_AZ_RX_FLUSH_DESCQ_CMD, 1,
636 			     FRF_AZ_RX_FLUSH_DESCQ,
637 			     efx_rx_queue_index(rx_queue));
638 	efx_writeo(efx, &rx_flush_descq, FR_AZ_RX_FLUSH_DESCQ);
639 }
640 
641 void efx_nic_fini_rx(struct efx_rx_queue *rx_queue)
642 {
643 	efx_oword_t rx_desc_ptr;
644 	struct efx_nic *efx = rx_queue->efx;
645 
646 	/* Remove RX descriptor ring from card */
647 	EFX_ZERO_OWORD(rx_desc_ptr);
648 	efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
649 			 efx_rx_queue_index(rx_queue));
650 
651 	/* Unpin RX descriptor ring */
652 	efx_fini_special_buffer(efx, &rx_queue->rxd);
653 }
654 
655 /* Free buffers backing RX queue */
656 void efx_nic_remove_rx(struct efx_rx_queue *rx_queue)
657 {
658 	efx_free_special_buffer(rx_queue->efx, &rx_queue->rxd);
659 }
660 
661 /**************************************************************************
662  *
663  * Flush handling
664  *
665  **************************************************************************/
666 
667 /* efx_nic_flush_queues() must be woken up when all flushes are completed,
668  * or more RX flushes can be kicked off.
669  */
670 static bool efx_flush_wake(struct efx_nic *efx)
671 {
672 	/* Ensure that all updates are visible to efx_nic_flush_queues() */
673 	smp_mb();
674 
675 	return (atomic_read(&efx->drain_pending) == 0 ||
676 		(atomic_read(&efx->rxq_flush_outstanding) < EFX_RX_FLUSH_COUNT
677 		 && atomic_read(&efx->rxq_flush_pending) > 0));
678 }
679 
680 static bool efx_check_tx_flush_complete(struct efx_nic *efx)
681 {
682 	bool i = true;
683 	efx_oword_t txd_ptr_tbl;
684 	struct efx_channel *channel;
685 	struct efx_tx_queue *tx_queue;
686 
687 	efx_for_each_channel(channel, efx) {
688 		efx_for_each_channel_tx_queue(tx_queue, channel) {
689 			efx_reado_table(efx, &txd_ptr_tbl,
690 					FR_BZ_TX_DESC_PTR_TBL, tx_queue->queue);
691 			if (EFX_OWORD_FIELD(txd_ptr_tbl,
692 					    FRF_AZ_TX_DESCQ_FLUSH) ||
693 			    EFX_OWORD_FIELD(txd_ptr_tbl,
694 					    FRF_AZ_TX_DESCQ_EN)) {
695 				netif_dbg(efx, hw, efx->net_dev,
696 					  "flush did not complete on TXQ %d\n",
697 					  tx_queue->queue);
698 				i = false;
699 			} else if (atomic_cmpxchg(&tx_queue->flush_outstanding,
700 						  1, 0)) {
701 				/* The flush is complete, but we didn't
702 				 * receive a flush completion event
703 				 */
704 				netif_dbg(efx, hw, efx->net_dev,
705 					  "flush complete on TXQ %d, so drain "
706 					  "the queue\n", tx_queue->queue);
707 				/* Don't need to increment drain_pending as it
708 				 * has already been incremented for the queues
709 				 * which did not drain
710 				 */
711 				efx_magic_event(channel,
712 						EFX_CHANNEL_MAGIC_TX_DRAIN(
713 							tx_queue));
714 			}
715 		}
716 	}
717 
718 	return i;
719 }
720 
721 /* Flush all the transmit queues, and continue flushing receive queues until
722  * they're all flushed. Wait for the DRAIN events to be recieved so that there
723  * are no more RX and TX events left on any channel. */
724 int efx_nic_flush_queues(struct efx_nic *efx)
725 {
726 	unsigned timeout = msecs_to_jiffies(5000); /* 5s for all flushes and drains */
727 	struct efx_channel *channel;
728 	struct efx_rx_queue *rx_queue;
729 	struct efx_tx_queue *tx_queue;
730 	int rc = 0;
731 
732 	efx->type->prepare_flush(efx);
733 
734 	efx_for_each_channel(channel, efx) {
735 		efx_for_each_channel_tx_queue(tx_queue, channel) {
736 			atomic_inc(&efx->drain_pending);
737 			efx_flush_tx_queue(tx_queue);
738 		}
739 		efx_for_each_channel_rx_queue(rx_queue, channel) {
740 			atomic_inc(&efx->drain_pending);
741 			rx_queue->flush_pending = true;
742 			atomic_inc(&efx->rxq_flush_pending);
743 		}
744 	}
745 
746 	while (timeout && atomic_read(&efx->drain_pending) > 0) {
747 		/* If SRIOV is enabled, then offload receive queue flushing to
748 		 * the firmware (though we will still have to poll for
749 		 * completion). If that fails, fall back to the old scheme.
750 		 */
751 		if (efx_sriov_enabled(efx)) {
752 			rc = efx_mcdi_flush_rxqs(efx);
753 			if (!rc)
754 				goto wait;
755 		}
756 
757 		/* The hardware supports four concurrent rx flushes, each of
758 		 * which may need to be retried if there is an outstanding
759 		 * descriptor fetch
760 		 */
761 		efx_for_each_channel(channel, efx) {
762 			efx_for_each_channel_rx_queue(rx_queue, channel) {
763 				if (atomic_read(&efx->rxq_flush_outstanding) >=
764 				    EFX_RX_FLUSH_COUNT)
765 					break;
766 
767 				if (rx_queue->flush_pending) {
768 					rx_queue->flush_pending = false;
769 					atomic_dec(&efx->rxq_flush_pending);
770 					atomic_inc(&efx->rxq_flush_outstanding);
771 					efx_flush_rx_queue(rx_queue);
772 				}
773 			}
774 		}
775 
776 	wait:
777 		timeout = wait_event_timeout(efx->flush_wq, efx_flush_wake(efx),
778 					     timeout);
779 	}
780 
781 	if (atomic_read(&efx->drain_pending) &&
782 	    !efx_check_tx_flush_complete(efx)) {
783 		netif_err(efx, hw, efx->net_dev, "failed to flush %d queues "
784 			  "(rx %d+%d)\n", atomic_read(&efx->drain_pending),
785 			  atomic_read(&efx->rxq_flush_outstanding),
786 			  atomic_read(&efx->rxq_flush_pending));
787 		rc = -ETIMEDOUT;
788 
789 		atomic_set(&efx->drain_pending, 0);
790 		atomic_set(&efx->rxq_flush_pending, 0);
791 		atomic_set(&efx->rxq_flush_outstanding, 0);
792 	}
793 
794 	efx->type->finish_flush(efx);
795 
796 	return rc;
797 }
798 
799 /**************************************************************************
800  *
801  * Event queue processing
802  * Event queues are processed by per-channel tasklets.
803  *
804  **************************************************************************/
805 
806 /* Update a channel's event queue's read pointer (RPTR) register
807  *
808  * This writes the EVQ_RPTR_REG register for the specified channel's
809  * event queue.
810  */
811 void efx_nic_eventq_read_ack(struct efx_channel *channel)
812 {
813 	efx_dword_t reg;
814 	struct efx_nic *efx = channel->efx;
815 
816 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_EVQ_RPTR,
817 			     channel->eventq_read_ptr & channel->eventq_mask);
818 
819 	/* For Falcon A1, EVQ_RPTR_KER is documented as having a step size
820 	 * of 4 bytes, but it is really 16 bytes just like later revisions.
821 	 */
822 	efx_writed(efx, &reg,
823 		   efx->type->evq_rptr_tbl_base +
824 		   FR_BZ_EVQ_RPTR_STEP * channel->channel);
825 }
826 
827 /* Use HW to insert a SW defined event */
828 void efx_generate_event(struct efx_nic *efx, unsigned int evq,
829 			efx_qword_t *event)
830 {
831 	efx_oword_t drv_ev_reg;
832 
833 	BUILD_BUG_ON(FRF_AZ_DRV_EV_DATA_LBN != 0 ||
834 		     FRF_AZ_DRV_EV_DATA_WIDTH != 64);
835 	drv_ev_reg.u32[0] = event->u32[0];
836 	drv_ev_reg.u32[1] = event->u32[1];
837 	drv_ev_reg.u32[2] = 0;
838 	drv_ev_reg.u32[3] = 0;
839 	EFX_SET_OWORD_FIELD(drv_ev_reg, FRF_AZ_DRV_EV_QID, evq);
840 	efx_writeo(efx, &drv_ev_reg, FR_AZ_DRV_EV);
841 }
842 
843 static void efx_magic_event(struct efx_channel *channel, u32 magic)
844 {
845 	efx_qword_t event;
846 
847 	EFX_POPULATE_QWORD_2(event, FSF_AZ_EV_CODE,
848 			     FSE_AZ_EV_CODE_DRV_GEN_EV,
849 			     FSF_AZ_DRV_GEN_EV_MAGIC, magic);
850 	efx_generate_event(channel->efx, channel->channel, &event);
851 }
852 
853 /* Handle a transmit completion event
854  *
855  * The NIC batches TX completion events; the message we receive is of
856  * the form "complete all TX events up to this index".
857  */
858 static int
859 efx_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
860 {
861 	unsigned int tx_ev_desc_ptr;
862 	unsigned int tx_ev_q_label;
863 	struct efx_tx_queue *tx_queue;
864 	struct efx_nic *efx = channel->efx;
865 	int tx_packets = 0;
866 
867 	if (unlikely(ACCESS_ONCE(efx->reset_pending)))
868 		return 0;
869 
870 	if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
871 		/* Transmit completion */
872 		tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
873 		tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
874 		tx_queue = efx_channel_get_tx_queue(
875 			channel, tx_ev_q_label % EFX_TXQ_TYPES);
876 		tx_packets = ((tx_ev_desc_ptr - tx_queue->read_count) &
877 			      tx_queue->ptr_mask);
878 		efx_xmit_done(tx_queue, tx_ev_desc_ptr);
879 	} else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
880 		/* Rewrite the FIFO write pointer */
881 		tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
882 		tx_queue = efx_channel_get_tx_queue(
883 			channel, tx_ev_q_label % EFX_TXQ_TYPES);
884 
885 		netif_tx_lock(efx->net_dev);
886 		efx_notify_tx_desc(tx_queue);
887 		netif_tx_unlock(efx->net_dev);
888 	} else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR) &&
889 		   EFX_WORKAROUND_10727(efx)) {
890 		efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
891 	} else {
892 		netif_err(efx, tx_err, efx->net_dev,
893 			  "channel %d unexpected TX event "
894 			  EFX_QWORD_FMT"\n", channel->channel,
895 			  EFX_QWORD_VAL(*event));
896 	}
897 
898 	return tx_packets;
899 }
900 
901 /* Detect errors included in the rx_evt_pkt_ok bit. */
902 static u16 efx_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
903 				const efx_qword_t *event)
904 {
905 	struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
906 	struct efx_nic *efx = rx_queue->efx;
907 	bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
908 	bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
909 	bool rx_ev_frm_trunc, rx_ev_drib_nib, rx_ev_tobe_disc;
910 	bool rx_ev_other_err, rx_ev_pause_frm;
911 	bool rx_ev_hdr_type, rx_ev_mcast_pkt;
912 	unsigned rx_ev_pkt_type;
913 
914 	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
915 	rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
916 	rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
917 	rx_ev_pkt_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_TYPE);
918 	rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
919 						 FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
920 	rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
921 						  FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
922 	rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
923 						   FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
924 	rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
925 	rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
926 	rx_ev_drib_nib = ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) ?
927 			  0 : EFX_QWORD_FIELD(*event, FSF_AA_RX_EV_DRIB_NIB));
928 	rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
929 
930 	/* Every error apart from tobe_disc and pause_frm */
931 	rx_ev_other_err = (rx_ev_drib_nib | rx_ev_tcp_udp_chksum_err |
932 			   rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
933 			   rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
934 
935 	/* Count errors that are not in MAC stats.  Ignore expected
936 	 * checksum errors during self-test. */
937 	if (rx_ev_frm_trunc)
938 		++channel->n_rx_frm_trunc;
939 	else if (rx_ev_tobe_disc)
940 		++channel->n_rx_tobe_disc;
941 	else if (!efx->loopback_selftest) {
942 		if (rx_ev_ip_hdr_chksum_err)
943 			++channel->n_rx_ip_hdr_chksum_err;
944 		else if (rx_ev_tcp_udp_chksum_err)
945 			++channel->n_rx_tcp_udp_chksum_err;
946 	}
947 
948 	/* TOBE_DISC is expected on unicast mismatches; don't print out an
949 	 * error message.  FRM_TRUNC indicates RXDP dropped the packet due
950 	 * to a FIFO overflow.
951 	 */
952 #ifdef DEBUG
953 	if (rx_ev_other_err && net_ratelimit()) {
954 		netif_dbg(efx, rx_err, efx->net_dev,
955 			  " RX queue %d unexpected RX event "
956 			  EFX_QWORD_FMT "%s%s%s%s%s%s%s%s\n",
957 			  efx_rx_queue_index(rx_queue), EFX_QWORD_VAL(*event),
958 			  rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
959 			  rx_ev_ip_hdr_chksum_err ?
960 			  " [IP_HDR_CHKSUM_ERR]" : "",
961 			  rx_ev_tcp_udp_chksum_err ?
962 			  " [TCP_UDP_CHKSUM_ERR]" : "",
963 			  rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
964 			  rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
965 			  rx_ev_drib_nib ? " [DRIB_NIB]" : "",
966 			  rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
967 			  rx_ev_pause_frm ? " [PAUSE]" : "");
968 	}
969 #endif
970 
971 	/* The frame must be discarded if any of these are true. */
972 	return (rx_ev_eth_crc_err | rx_ev_frm_trunc | rx_ev_drib_nib |
973 		rx_ev_tobe_disc | rx_ev_pause_frm) ?
974 		EFX_RX_PKT_DISCARD : 0;
975 }
976 
977 /* Handle receive events that are not in-order. Return true if this
978  * can be handled as a partial packet discard, false if it's more
979  * serious.
980  */
981 static bool
982 efx_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
983 {
984 	struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
985 	struct efx_nic *efx = rx_queue->efx;
986 	unsigned expected, dropped;
987 
988 	if (rx_queue->scatter_n &&
989 	    index == ((rx_queue->removed_count + rx_queue->scatter_n - 1) &
990 		      rx_queue->ptr_mask)) {
991 		++channel->n_rx_nodesc_trunc;
992 		return true;
993 	}
994 
995 	expected = rx_queue->removed_count & rx_queue->ptr_mask;
996 	dropped = (index - expected) & rx_queue->ptr_mask;
997 	netif_info(efx, rx_err, efx->net_dev,
998 		   "dropped %d events (index=%d expected=%d)\n",
999 		   dropped, index, expected);
1000 
1001 	efx_schedule_reset(efx, EFX_WORKAROUND_5676(efx) ?
1002 			   RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
1003 	return false;
1004 }
1005 
1006 /* Handle a packet received event
1007  *
1008  * The NIC gives a "discard" flag if it's a unicast packet with the
1009  * wrong destination address
1010  * Also "is multicast" and "matches multicast filter" flags can be used to
1011  * discard non-matching multicast packets.
1012  */
1013 static void
1014 efx_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
1015 {
1016 	unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
1017 	unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
1018 	unsigned expected_ptr;
1019 	bool rx_ev_pkt_ok, rx_ev_sop, rx_ev_cont;
1020 	u16 flags;
1021 	struct efx_rx_queue *rx_queue;
1022 	struct efx_nic *efx = channel->efx;
1023 
1024 	if (unlikely(ACCESS_ONCE(efx->reset_pending)))
1025 		return;
1026 
1027 	rx_ev_cont = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT);
1028 	rx_ev_sop = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP);
1029 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
1030 		channel->channel);
1031 
1032 	rx_queue = efx_channel_get_rx_queue(channel);
1033 
1034 	rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
1035 	expected_ptr = ((rx_queue->removed_count + rx_queue->scatter_n) &
1036 			rx_queue->ptr_mask);
1037 
1038 	/* Check for partial drops and other errors */
1039 	if (unlikely(rx_ev_desc_ptr != expected_ptr) ||
1040 	    unlikely(rx_ev_sop != (rx_queue->scatter_n == 0))) {
1041 		if (rx_ev_desc_ptr != expected_ptr &&
1042 		    !efx_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr))
1043 			return;
1044 
1045 		/* Discard all pending fragments */
1046 		if (rx_queue->scatter_n) {
1047 			efx_rx_packet(
1048 				rx_queue,
1049 				rx_queue->removed_count & rx_queue->ptr_mask,
1050 				rx_queue->scatter_n, 0, EFX_RX_PKT_DISCARD);
1051 			rx_queue->removed_count += rx_queue->scatter_n;
1052 			rx_queue->scatter_n = 0;
1053 		}
1054 
1055 		/* Return if there is no new fragment */
1056 		if (rx_ev_desc_ptr != expected_ptr)
1057 			return;
1058 
1059 		/* Discard new fragment if not SOP */
1060 		if (!rx_ev_sop) {
1061 			efx_rx_packet(
1062 				rx_queue,
1063 				rx_queue->removed_count & rx_queue->ptr_mask,
1064 				1, 0, EFX_RX_PKT_DISCARD);
1065 			++rx_queue->removed_count;
1066 			return;
1067 		}
1068 	}
1069 
1070 	++rx_queue->scatter_n;
1071 	if (rx_ev_cont)
1072 		return;
1073 
1074 	rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
1075 	rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
1076 	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
1077 
1078 	if (likely(rx_ev_pkt_ok)) {
1079 		/* If packet is marked as OK then we can rely on the
1080 		 * hardware checksum and classification.
1081 		 */
1082 		flags = 0;
1083 		switch (rx_ev_hdr_type) {
1084 		case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP:
1085 			flags |= EFX_RX_PKT_TCP;
1086 			/* fall through */
1087 		case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP:
1088 			flags |= EFX_RX_PKT_CSUMMED;
1089 			/* fall through */
1090 		case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_OTHER:
1091 		case FSE_AZ_RX_EV_HDR_TYPE_OTHER:
1092 			break;
1093 		}
1094 	} else {
1095 		flags = efx_handle_rx_not_ok(rx_queue, event);
1096 	}
1097 
1098 	/* Detect multicast packets that didn't match the filter */
1099 	rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
1100 	if (rx_ev_mcast_pkt) {
1101 		unsigned int rx_ev_mcast_hash_match =
1102 			EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
1103 
1104 		if (unlikely(!rx_ev_mcast_hash_match)) {
1105 			++channel->n_rx_mcast_mismatch;
1106 			flags |= EFX_RX_PKT_DISCARD;
1107 		}
1108 	}
1109 
1110 	channel->irq_mod_score += 2;
1111 
1112 	/* Handle received packet */
1113 	efx_rx_packet(rx_queue,
1114 		      rx_queue->removed_count & rx_queue->ptr_mask,
1115 		      rx_queue->scatter_n, rx_ev_byte_cnt, flags);
1116 	rx_queue->removed_count += rx_queue->scatter_n;
1117 	rx_queue->scatter_n = 0;
1118 }
1119 
1120 /* If this flush done event corresponds to a &struct efx_tx_queue, then
1121  * send an %EFX_CHANNEL_MAGIC_TX_DRAIN event to drain the event queue
1122  * of all transmit completions.
1123  */
1124 static void
1125 efx_handle_tx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1126 {
1127 	struct efx_tx_queue *tx_queue;
1128 	int qid;
1129 
1130 	qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1131 	if (qid < EFX_TXQ_TYPES * efx->n_tx_channels) {
1132 		tx_queue = efx_get_tx_queue(efx, qid / EFX_TXQ_TYPES,
1133 					    qid % EFX_TXQ_TYPES);
1134 		if (atomic_cmpxchg(&tx_queue->flush_outstanding, 1, 0)) {
1135 			efx_magic_event(tx_queue->channel,
1136 					EFX_CHANNEL_MAGIC_TX_DRAIN(tx_queue));
1137 		}
1138 	}
1139 }
1140 
1141 /* If this flush done event corresponds to a &struct efx_rx_queue: If the flush
1142  * was succesful then send an %EFX_CHANNEL_MAGIC_RX_DRAIN, otherwise add
1143  * the RX queue back to the mask of RX queues in need of flushing.
1144  */
1145 static void
1146 efx_handle_rx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1147 {
1148 	struct efx_channel *channel;
1149 	struct efx_rx_queue *rx_queue;
1150 	int qid;
1151 	bool failed;
1152 
1153 	qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
1154 	failed = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
1155 	if (qid >= efx->n_channels)
1156 		return;
1157 	channel = efx_get_channel(efx, qid);
1158 	if (!efx_channel_has_rx_queue(channel))
1159 		return;
1160 	rx_queue = efx_channel_get_rx_queue(channel);
1161 
1162 	if (failed) {
1163 		netif_info(efx, hw, efx->net_dev,
1164 			   "RXQ %d flush retry\n", qid);
1165 		rx_queue->flush_pending = true;
1166 		atomic_inc(&efx->rxq_flush_pending);
1167 	} else {
1168 		efx_magic_event(efx_rx_queue_channel(rx_queue),
1169 				EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue));
1170 	}
1171 	atomic_dec(&efx->rxq_flush_outstanding);
1172 	if (efx_flush_wake(efx))
1173 		wake_up(&efx->flush_wq);
1174 }
1175 
1176 static void
1177 efx_handle_drain_event(struct efx_channel *channel)
1178 {
1179 	struct efx_nic *efx = channel->efx;
1180 
1181 	WARN_ON(atomic_read(&efx->drain_pending) == 0);
1182 	atomic_dec(&efx->drain_pending);
1183 	if (efx_flush_wake(efx))
1184 		wake_up(&efx->flush_wq);
1185 }
1186 
1187 static void
1188 efx_handle_generated_event(struct efx_channel *channel, efx_qword_t *event)
1189 {
1190 	struct efx_nic *efx = channel->efx;
1191 	struct efx_rx_queue *rx_queue =
1192 		efx_channel_has_rx_queue(channel) ?
1193 		efx_channel_get_rx_queue(channel) : NULL;
1194 	unsigned magic, code;
1195 
1196 	magic = EFX_QWORD_FIELD(*event, FSF_AZ_DRV_GEN_EV_MAGIC);
1197 	code = _EFX_CHANNEL_MAGIC_CODE(magic);
1198 
1199 	if (magic == EFX_CHANNEL_MAGIC_TEST(channel)) {
1200 		channel->event_test_cpu = raw_smp_processor_id();
1201 	} else if (rx_queue && magic == EFX_CHANNEL_MAGIC_FILL(rx_queue)) {
1202 		/* The queue must be empty, so we won't receive any rx
1203 		 * events, so efx_process_channel() won't refill the
1204 		 * queue. Refill it here */
1205 		efx_fast_push_rx_descriptors(rx_queue);
1206 	} else if (rx_queue && magic == EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue)) {
1207 		efx_handle_drain_event(channel);
1208 	} else if (code == _EFX_CHANNEL_MAGIC_TX_DRAIN) {
1209 		efx_handle_drain_event(channel);
1210 	} else {
1211 		netif_dbg(efx, hw, efx->net_dev, "channel %d received "
1212 			  "generated event "EFX_QWORD_FMT"\n",
1213 			  channel->channel, EFX_QWORD_VAL(*event));
1214 	}
1215 }
1216 
1217 static void
1218 efx_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
1219 {
1220 	struct efx_nic *efx = channel->efx;
1221 	unsigned int ev_sub_code;
1222 	unsigned int ev_sub_data;
1223 
1224 	ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
1225 	ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1226 
1227 	switch (ev_sub_code) {
1228 	case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
1229 		netif_vdbg(efx, hw, efx->net_dev, "channel %d TXQ %d flushed\n",
1230 			   channel->channel, ev_sub_data);
1231 		efx_handle_tx_flush_done(efx, event);
1232 		efx_sriov_tx_flush_done(efx, event);
1233 		break;
1234 	case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
1235 		netif_vdbg(efx, hw, efx->net_dev, "channel %d RXQ %d flushed\n",
1236 			   channel->channel, ev_sub_data);
1237 		efx_handle_rx_flush_done(efx, event);
1238 		efx_sriov_rx_flush_done(efx, event);
1239 		break;
1240 	case FSE_AZ_EVQ_INIT_DONE_EV:
1241 		netif_dbg(efx, hw, efx->net_dev,
1242 			  "channel %d EVQ %d initialised\n",
1243 			  channel->channel, ev_sub_data);
1244 		break;
1245 	case FSE_AZ_SRM_UPD_DONE_EV:
1246 		netif_vdbg(efx, hw, efx->net_dev,
1247 			   "channel %d SRAM update done\n", channel->channel);
1248 		break;
1249 	case FSE_AZ_WAKE_UP_EV:
1250 		netif_vdbg(efx, hw, efx->net_dev,
1251 			   "channel %d RXQ %d wakeup event\n",
1252 			   channel->channel, ev_sub_data);
1253 		break;
1254 	case FSE_AZ_TIMER_EV:
1255 		netif_vdbg(efx, hw, efx->net_dev,
1256 			   "channel %d RX queue %d timer expired\n",
1257 			   channel->channel, ev_sub_data);
1258 		break;
1259 	case FSE_AA_RX_RECOVER_EV:
1260 		netif_err(efx, rx_err, efx->net_dev,
1261 			  "channel %d seen DRIVER RX_RESET event. "
1262 			"Resetting.\n", channel->channel);
1263 		atomic_inc(&efx->rx_reset);
1264 		efx_schedule_reset(efx,
1265 				   EFX_WORKAROUND_6555(efx) ?
1266 				   RESET_TYPE_RX_RECOVERY :
1267 				   RESET_TYPE_DISABLE);
1268 		break;
1269 	case FSE_BZ_RX_DSC_ERROR_EV:
1270 		if (ev_sub_data < EFX_VI_BASE) {
1271 			netif_err(efx, rx_err, efx->net_dev,
1272 				  "RX DMA Q %d reports descriptor fetch error."
1273 				  " RX Q %d is disabled.\n", ev_sub_data,
1274 				  ev_sub_data);
1275 			efx_schedule_reset(efx, RESET_TYPE_RX_DESC_FETCH);
1276 		} else
1277 			efx_sriov_desc_fetch_err(efx, ev_sub_data);
1278 		break;
1279 	case FSE_BZ_TX_DSC_ERROR_EV:
1280 		if (ev_sub_data < EFX_VI_BASE) {
1281 			netif_err(efx, tx_err, efx->net_dev,
1282 				  "TX DMA Q %d reports descriptor fetch error."
1283 				  " TX Q %d is disabled.\n", ev_sub_data,
1284 				  ev_sub_data);
1285 			efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
1286 		} else
1287 			efx_sriov_desc_fetch_err(efx, ev_sub_data);
1288 		break;
1289 	default:
1290 		netif_vdbg(efx, hw, efx->net_dev,
1291 			   "channel %d unknown driver event code %d "
1292 			   "data %04x\n", channel->channel, ev_sub_code,
1293 			   ev_sub_data);
1294 		break;
1295 	}
1296 }
1297 
1298 int efx_nic_process_eventq(struct efx_channel *channel, int budget)
1299 {
1300 	struct efx_nic *efx = channel->efx;
1301 	unsigned int read_ptr;
1302 	efx_qword_t event, *p_event;
1303 	int ev_code;
1304 	int tx_packets = 0;
1305 	int spent = 0;
1306 
1307 	read_ptr = channel->eventq_read_ptr;
1308 
1309 	for (;;) {
1310 		p_event = efx_event(channel, read_ptr);
1311 		event = *p_event;
1312 
1313 		if (!efx_event_present(&event))
1314 			/* End of events */
1315 			break;
1316 
1317 		netif_vdbg(channel->efx, intr, channel->efx->net_dev,
1318 			   "channel %d event is "EFX_QWORD_FMT"\n",
1319 			   channel->channel, EFX_QWORD_VAL(event));
1320 
1321 		/* Clear this event by marking it all ones */
1322 		EFX_SET_QWORD(*p_event);
1323 
1324 		++read_ptr;
1325 
1326 		ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
1327 
1328 		switch (ev_code) {
1329 		case FSE_AZ_EV_CODE_RX_EV:
1330 			efx_handle_rx_event(channel, &event);
1331 			if (++spent == budget)
1332 				goto out;
1333 			break;
1334 		case FSE_AZ_EV_CODE_TX_EV:
1335 			tx_packets += efx_handle_tx_event(channel, &event);
1336 			if (tx_packets > efx->txq_entries) {
1337 				spent = budget;
1338 				goto out;
1339 			}
1340 			break;
1341 		case FSE_AZ_EV_CODE_DRV_GEN_EV:
1342 			efx_handle_generated_event(channel, &event);
1343 			break;
1344 		case FSE_AZ_EV_CODE_DRIVER_EV:
1345 			efx_handle_driver_event(channel, &event);
1346 			break;
1347 		case FSE_CZ_EV_CODE_USER_EV:
1348 			efx_sriov_event(channel, &event);
1349 			break;
1350 		case FSE_CZ_EV_CODE_MCDI_EV:
1351 			efx_mcdi_process_event(channel, &event);
1352 			break;
1353 		case FSE_AZ_EV_CODE_GLOBAL_EV:
1354 			if (efx->type->handle_global_event &&
1355 			    efx->type->handle_global_event(channel, &event))
1356 				break;
1357 			/* else fall through */
1358 		default:
1359 			netif_err(channel->efx, hw, channel->efx->net_dev,
1360 				  "channel %d unknown event type %d (data "
1361 				  EFX_QWORD_FMT ")\n", channel->channel,
1362 				  ev_code, EFX_QWORD_VAL(event));
1363 		}
1364 	}
1365 
1366 out:
1367 	channel->eventq_read_ptr = read_ptr;
1368 	return spent;
1369 }
1370 
1371 /* Check whether an event is present in the eventq at the current
1372  * read pointer.  Only useful for self-test.
1373  */
1374 bool efx_nic_event_present(struct efx_channel *channel)
1375 {
1376 	return efx_event_present(efx_event(channel, channel->eventq_read_ptr));
1377 }
1378 
1379 /* Allocate buffer table entries for event queue */
1380 int efx_nic_probe_eventq(struct efx_channel *channel)
1381 {
1382 	struct efx_nic *efx = channel->efx;
1383 	unsigned entries;
1384 
1385 	entries = channel->eventq_mask + 1;
1386 	return efx_alloc_special_buffer(efx, &channel->eventq,
1387 					entries * sizeof(efx_qword_t));
1388 }
1389 
1390 void efx_nic_init_eventq(struct efx_channel *channel)
1391 {
1392 	efx_oword_t reg;
1393 	struct efx_nic *efx = channel->efx;
1394 
1395 	netif_dbg(efx, hw, efx->net_dev,
1396 		  "channel %d event queue in special buffers %d-%d\n",
1397 		  channel->channel, channel->eventq.index,
1398 		  channel->eventq.index + channel->eventq.entries - 1);
1399 
1400 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1401 		EFX_POPULATE_OWORD_3(reg,
1402 				     FRF_CZ_TIMER_Q_EN, 1,
1403 				     FRF_CZ_HOST_NOTIFY_MODE, 0,
1404 				     FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
1405 		efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1406 	}
1407 
1408 	/* Pin event queue buffer */
1409 	efx_init_special_buffer(efx, &channel->eventq);
1410 
1411 	/* Fill event queue with all ones (i.e. empty events) */
1412 	memset(channel->eventq.buf.addr, 0xff, channel->eventq.buf.len);
1413 
1414 	/* Push event queue to card */
1415 	EFX_POPULATE_OWORD_3(reg,
1416 			     FRF_AZ_EVQ_EN, 1,
1417 			     FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
1418 			     FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
1419 	efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1420 			 channel->channel);
1421 
1422 	efx->type->push_irq_moderation(channel);
1423 }
1424 
1425 void efx_nic_fini_eventq(struct efx_channel *channel)
1426 {
1427 	efx_oword_t reg;
1428 	struct efx_nic *efx = channel->efx;
1429 
1430 	/* Remove event queue from card */
1431 	EFX_ZERO_OWORD(reg);
1432 	efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1433 			 channel->channel);
1434 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1435 		efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1436 
1437 	/* Unpin event queue */
1438 	efx_fini_special_buffer(efx, &channel->eventq);
1439 }
1440 
1441 /* Free buffers backing event queue */
1442 void efx_nic_remove_eventq(struct efx_channel *channel)
1443 {
1444 	efx_free_special_buffer(channel->efx, &channel->eventq);
1445 }
1446 
1447 
1448 void efx_nic_event_test_start(struct efx_channel *channel)
1449 {
1450 	channel->event_test_cpu = -1;
1451 	smp_wmb();
1452 	efx_magic_event(channel, EFX_CHANNEL_MAGIC_TEST(channel));
1453 }
1454 
1455 void efx_nic_generate_fill_event(struct efx_rx_queue *rx_queue)
1456 {
1457 	efx_magic_event(efx_rx_queue_channel(rx_queue),
1458 			EFX_CHANNEL_MAGIC_FILL(rx_queue));
1459 }
1460 
1461 /**************************************************************************
1462  *
1463  * Hardware interrupts
1464  * The hardware interrupt handler does very little work; all the event
1465  * queue processing is carried out by per-channel tasklets.
1466  *
1467  **************************************************************************/
1468 
1469 /* Enable/disable/generate interrupts */
1470 static inline void efx_nic_interrupts(struct efx_nic *efx,
1471 				      bool enabled, bool force)
1472 {
1473 	efx_oword_t int_en_reg_ker;
1474 
1475 	EFX_POPULATE_OWORD_3(int_en_reg_ker,
1476 			     FRF_AZ_KER_INT_LEVE_SEL, efx->irq_level,
1477 			     FRF_AZ_KER_INT_KER, force,
1478 			     FRF_AZ_DRV_INT_EN_KER, enabled);
1479 	efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
1480 }
1481 
1482 void efx_nic_enable_interrupts(struct efx_nic *efx)
1483 {
1484 	EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
1485 	wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
1486 
1487 	efx_nic_interrupts(efx, true, false);
1488 }
1489 
1490 void efx_nic_disable_interrupts(struct efx_nic *efx)
1491 {
1492 	/* Disable interrupts */
1493 	efx_nic_interrupts(efx, false, false);
1494 }
1495 
1496 /* Generate a test interrupt
1497  * Interrupt must already have been enabled, otherwise nasty things
1498  * may happen.
1499  */
1500 void efx_nic_irq_test_start(struct efx_nic *efx)
1501 {
1502 	efx->last_irq_cpu = -1;
1503 	smp_wmb();
1504 	efx_nic_interrupts(efx, true, true);
1505 }
1506 
1507 /* Process a fatal interrupt
1508  * Disable bus mastering ASAP and schedule a reset
1509  */
1510 irqreturn_t efx_nic_fatal_interrupt(struct efx_nic *efx)
1511 {
1512 	struct falcon_nic_data *nic_data = efx->nic_data;
1513 	efx_oword_t *int_ker = efx->irq_status.addr;
1514 	efx_oword_t fatal_intr;
1515 	int error, mem_perr;
1516 
1517 	efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
1518 	error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
1519 
1520 	netif_err(efx, hw, efx->net_dev, "SYSTEM ERROR "EFX_OWORD_FMT" status "
1521 		  EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
1522 		  EFX_OWORD_VAL(fatal_intr),
1523 		  error ? "disabling bus mastering" : "no recognised error");
1524 
1525 	/* If this is a memory parity error dump which blocks are offending */
1526 	mem_perr = (EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER) ||
1527 		    EFX_OWORD_FIELD(fatal_intr, FRF_AZ_SRM_PERR_INT_KER));
1528 	if (mem_perr) {
1529 		efx_oword_t reg;
1530 		efx_reado(efx, &reg, FR_AZ_MEM_STAT);
1531 		netif_err(efx, hw, efx->net_dev,
1532 			  "SYSTEM ERROR: memory parity error "EFX_OWORD_FMT"\n",
1533 			  EFX_OWORD_VAL(reg));
1534 	}
1535 
1536 	/* Disable both devices */
1537 	pci_clear_master(efx->pci_dev);
1538 	if (efx_nic_is_dual_func(efx))
1539 		pci_clear_master(nic_data->pci_dev2);
1540 	efx_nic_disable_interrupts(efx);
1541 
1542 	/* Count errors and reset or disable the NIC accordingly */
1543 	if (efx->int_error_count == 0 ||
1544 	    time_after(jiffies, efx->int_error_expire)) {
1545 		efx->int_error_count = 0;
1546 		efx->int_error_expire =
1547 			jiffies + EFX_INT_ERROR_EXPIRE * HZ;
1548 	}
1549 	if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
1550 		netif_err(efx, hw, efx->net_dev,
1551 			  "SYSTEM ERROR - reset scheduled\n");
1552 		efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
1553 	} else {
1554 		netif_err(efx, hw, efx->net_dev,
1555 			  "SYSTEM ERROR - max number of errors seen."
1556 			  "NIC will be disabled\n");
1557 		efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1558 	}
1559 
1560 	return IRQ_HANDLED;
1561 }
1562 
1563 /* Handle a legacy interrupt
1564  * Acknowledges the interrupt and schedule event queue processing.
1565  */
1566 static irqreturn_t efx_legacy_interrupt(int irq, void *dev_id)
1567 {
1568 	struct efx_nic *efx = dev_id;
1569 	bool soft_enabled = ACCESS_ONCE(efx->irq_soft_enabled);
1570 	efx_oword_t *int_ker = efx->irq_status.addr;
1571 	irqreturn_t result = IRQ_NONE;
1572 	struct efx_channel *channel;
1573 	efx_dword_t reg;
1574 	u32 queues;
1575 	int syserr;
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 	/* Legacy interrupts are disabled too late by the EEH kernel
1582 	 * code. Disable them earlier.
1583 	 * If an EEH error occurred, the read will have returned all ones.
1584 	 */
1585 	if (EFX_DWORD_IS_ALL_ONES(reg) && efx_try_recovery(efx) &&
1586 	    !efx->eeh_disabled_legacy_irq) {
1587 		disable_irq_nosync(efx->legacy_irq);
1588 		efx->eeh_disabled_legacy_irq = true;
1589 	}
1590 
1591 	/* Handle non-event-queue sources */
1592 	if (queues & (1U << efx->irq_level) && soft_enabled) {
1593 		syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1594 		if (unlikely(syserr))
1595 			return efx_nic_fatal_interrupt(efx);
1596 		efx->last_irq_cpu = raw_smp_processor_id();
1597 	}
1598 
1599 	if (queues != 0) {
1600 		if (EFX_WORKAROUND_15783(efx))
1601 			efx->irq_zero_count = 0;
1602 
1603 		/* Schedule processing of any interrupting queues */
1604 		if (likely(soft_enabled)) {
1605 			efx_for_each_channel(channel, efx) {
1606 				if (queues & 1)
1607 					efx_schedule_channel_irq(channel);
1608 				queues >>= 1;
1609 			}
1610 		}
1611 		result = IRQ_HANDLED;
1612 
1613 	} else if (EFX_WORKAROUND_15783(efx)) {
1614 		efx_qword_t *event;
1615 
1616 		/* We can't return IRQ_HANDLED more than once on seeing ISR=0
1617 		 * because this might be a shared interrupt. */
1618 		if (efx->irq_zero_count++ == 0)
1619 			result = IRQ_HANDLED;
1620 
1621 		/* Ensure we schedule or rearm all event queues */
1622 		if (likely(soft_enabled)) {
1623 			efx_for_each_channel(channel, efx) {
1624 				event = efx_event(channel,
1625 						  channel->eventq_read_ptr);
1626 				if (efx_event_present(event))
1627 					efx_schedule_channel_irq(channel);
1628 				else
1629 					efx_nic_eventq_read_ack(channel);
1630 			}
1631 		}
1632 	}
1633 
1634 	if (result == IRQ_HANDLED)
1635 		netif_vdbg(efx, intr, efx->net_dev,
1636 			   "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
1637 			   irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
1638 
1639 	return result;
1640 }
1641 
1642 /* Handle an MSI interrupt
1643  *
1644  * Handle an MSI hardware interrupt.  This routine schedules event
1645  * queue processing.  No interrupt acknowledgement cycle is necessary.
1646  * Also, we never need to check that the interrupt is for us, since
1647  * MSI interrupts cannot be shared.
1648  */
1649 static irqreturn_t efx_msi_interrupt(int irq, void *dev_id)
1650 {
1651 	struct efx_msi_context *context = dev_id;
1652 	struct efx_nic *efx = context->efx;
1653 	efx_oword_t *int_ker = efx->irq_status.addr;
1654 	int syserr;
1655 
1656 	netif_vdbg(efx, intr, efx->net_dev,
1657 		   "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
1658 		   irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
1659 
1660 	if (!likely(ACCESS_ONCE(efx->irq_soft_enabled)))
1661 		return IRQ_HANDLED;
1662 
1663 	/* Handle non-event-queue sources */
1664 	if (context->index == efx->irq_level) {
1665 		syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1666 		if (unlikely(syserr))
1667 			return efx_nic_fatal_interrupt(efx);
1668 		efx->last_irq_cpu = raw_smp_processor_id();
1669 	}
1670 
1671 	/* Schedule processing of the channel */
1672 	efx_schedule_channel_irq(efx->channel[context->index]);
1673 
1674 	return IRQ_HANDLED;
1675 }
1676 
1677 
1678 /* Setup RSS indirection table.
1679  * This maps from the hash value of the packet to RXQ
1680  */
1681 void efx_nic_push_rx_indir_table(struct efx_nic *efx)
1682 {
1683 	size_t i = 0;
1684 	efx_dword_t dword;
1685 
1686 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0)
1687 		return;
1688 
1689 	BUILD_BUG_ON(ARRAY_SIZE(efx->rx_indir_table) !=
1690 		     FR_BZ_RX_INDIRECTION_TBL_ROWS);
1691 
1692 	for (i = 0; i < FR_BZ_RX_INDIRECTION_TBL_ROWS; i++) {
1693 		EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
1694 				     efx->rx_indir_table[i]);
1695 		efx_writed(efx, &dword,
1696 			   FR_BZ_RX_INDIRECTION_TBL +
1697 			   FR_BZ_RX_INDIRECTION_TBL_STEP * i);
1698 	}
1699 }
1700 
1701 /* Hook interrupt handler(s)
1702  * Try MSI and then legacy interrupts.
1703  */
1704 int efx_nic_init_interrupt(struct efx_nic *efx)
1705 {
1706 	struct efx_channel *channel;
1707 	unsigned int n_irqs;
1708 	int rc;
1709 
1710 	if (!EFX_INT_MODE_USE_MSI(efx)) {
1711 		irq_handler_t handler;
1712 		if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1713 			handler = efx_legacy_interrupt;
1714 		else
1715 			handler = falcon_legacy_interrupt_a1;
1716 
1717 		rc = request_irq(efx->legacy_irq, handler, IRQF_SHARED,
1718 				 efx->name, efx);
1719 		if (rc) {
1720 			netif_err(efx, drv, efx->net_dev,
1721 				  "failed to hook legacy IRQ %d\n",
1722 				  efx->pci_dev->irq);
1723 			goto fail1;
1724 		}
1725 		return 0;
1726 	}
1727 
1728 #ifdef CONFIG_RFS_ACCEL
1729 	if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
1730 		efx->net_dev->rx_cpu_rmap =
1731 			alloc_irq_cpu_rmap(efx->n_rx_channels);
1732 		if (!efx->net_dev->rx_cpu_rmap) {
1733 			rc = -ENOMEM;
1734 			goto fail1;
1735 		}
1736 	}
1737 #endif
1738 
1739 	/* Hook MSI or MSI-X interrupt */
1740 	n_irqs = 0;
1741 	efx_for_each_channel(channel, efx) {
1742 		rc = request_irq(channel->irq, efx_msi_interrupt,
1743 				 IRQF_PROBE_SHARED, /* Not shared */
1744 				 efx->msi_context[channel->channel].name,
1745 				 &efx->msi_context[channel->channel]);
1746 		if (rc) {
1747 			netif_err(efx, drv, efx->net_dev,
1748 				  "failed to hook IRQ %d\n", channel->irq);
1749 			goto fail2;
1750 		}
1751 		++n_irqs;
1752 
1753 #ifdef CONFIG_RFS_ACCEL
1754 		if (efx->interrupt_mode == EFX_INT_MODE_MSIX &&
1755 		    channel->channel < efx->n_rx_channels) {
1756 			rc = irq_cpu_rmap_add(efx->net_dev->rx_cpu_rmap,
1757 					      channel->irq);
1758 			if (rc)
1759 				goto fail2;
1760 		}
1761 #endif
1762 	}
1763 
1764 	return 0;
1765 
1766  fail2:
1767 #ifdef CONFIG_RFS_ACCEL
1768 	free_irq_cpu_rmap(efx->net_dev->rx_cpu_rmap);
1769 	efx->net_dev->rx_cpu_rmap = NULL;
1770 #endif
1771 	efx_for_each_channel(channel, efx) {
1772 		if (n_irqs-- == 0)
1773 			break;
1774 		free_irq(channel->irq, &efx->msi_context[channel->channel]);
1775 	}
1776  fail1:
1777 	return rc;
1778 }
1779 
1780 void efx_nic_fini_interrupt(struct efx_nic *efx)
1781 {
1782 	struct efx_channel *channel;
1783 
1784 #ifdef CONFIG_RFS_ACCEL
1785 	free_irq_cpu_rmap(efx->net_dev->rx_cpu_rmap);
1786 	efx->net_dev->rx_cpu_rmap = NULL;
1787 #endif
1788 
1789 	/* Disable MSI/MSI-X interrupts */
1790 	efx_for_each_channel(channel, efx)
1791 		free_irq(channel->irq, &efx->msi_context[channel->channel]);
1792 
1793 	/* Disable legacy interrupt */
1794 	if (efx->legacy_irq)
1795 		free_irq(efx->legacy_irq, efx);
1796 }
1797 
1798 /* Looks at available SRAM resources and works out how many queues we
1799  * can support, and where things like descriptor caches should live.
1800  *
1801  * SRAM is split up as follows:
1802  * 0                          buftbl entries for channels
1803  * efx->vf_buftbl_base        buftbl entries for SR-IOV
1804  * efx->rx_dc_base            RX descriptor caches
1805  * efx->tx_dc_base            TX descriptor caches
1806  */
1807 void efx_nic_dimension_resources(struct efx_nic *efx, unsigned sram_lim_qw)
1808 {
1809 	unsigned vi_count, buftbl_min;
1810 
1811 	/* Account for the buffer table entries backing the datapath channels
1812 	 * and the descriptor caches for those channels.
1813 	 */
1814 	buftbl_min = ((efx->n_rx_channels * EFX_MAX_DMAQ_SIZE +
1815 		       efx->n_tx_channels * EFX_TXQ_TYPES * EFX_MAX_DMAQ_SIZE +
1816 		       efx->n_channels * EFX_MAX_EVQ_SIZE)
1817 		      * sizeof(efx_qword_t) / EFX_BUF_SIZE);
1818 	vi_count = max(efx->n_channels, efx->n_tx_channels * EFX_TXQ_TYPES);
1819 
1820 #ifdef CONFIG_SFC_SRIOV
1821 	if (efx_sriov_wanted(efx)) {
1822 		unsigned vi_dc_entries, buftbl_free, entries_per_vf, vf_limit;
1823 
1824 		efx->vf_buftbl_base = buftbl_min;
1825 
1826 		vi_dc_entries = RX_DC_ENTRIES + TX_DC_ENTRIES;
1827 		vi_count = max(vi_count, EFX_VI_BASE);
1828 		buftbl_free = (sram_lim_qw - buftbl_min -
1829 			       vi_count * vi_dc_entries);
1830 
1831 		entries_per_vf = ((vi_dc_entries + EFX_VF_BUFTBL_PER_VI) *
1832 				  efx_vf_size(efx));
1833 		vf_limit = min(buftbl_free / entries_per_vf,
1834 			       (1024U - EFX_VI_BASE) >> efx->vi_scale);
1835 
1836 		if (efx->vf_count > vf_limit) {
1837 			netif_err(efx, probe, efx->net_dev,
1838 				  "Reducing VF count from from %d to %d\n",
1839 				  efx->vf_count, vf_limit);
1840 			efx->vf_count = vf_limit;
1841 		}
1842 		vi_count += efx->vf_count * efx_vf_size(efx);
1843 	}
1844 #endif
1845 
1846 	efx->tx_dc_base = sram_lim_qw - vi_count * TX_DC_ENTRIES;
1847 	efx->rx_dc_base = efx->tx_dc_base - vi_count * RX_DC_ENTRIES;
1848 }
1849 
1850 u32 efx_nic_fpga_ver(struct efx_nic *efx)
1851 {
1852 	efx_oword_t altera_build;
1853 	efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
1854 	return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
1855 }
1856 
1857 void efx_nic_init_common(struct efx_nic *efx)
1858 {
1859 	efx_oword_t temp;
1860 
1861 	/* Set positions of descriptor caches in SRAM. */
1862 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR, efx->tx_dc_base);
1863 	efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
1864 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR, efx->rx_dc_base);
1865 	efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
1866 
1867 	/* Set TX descriptor cache size. */
1868 	BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
1869 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
1870 	efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
1871 
1872 	/* Set RX descriptor cache size.  Set low watermark to size-8, as
1873 	 * this allows most efficient prefetching.
1874 	 */
1875 	BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
1876 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
1877 	efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
1878 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
1879 	efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
1880 
1881 	/* Program INT_KER address */
1882 	EFX_POPULATE_OWORD_2(temp,
1883 			     FRF_AZ_NORM_INT_VEC_DIS_KER,
1884 			     EFX_INT_MODE_USE_MSI(efx),
1885 			     FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
1886 	efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
1887 
1888 	if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
1889 		/* Use an interrupt level unused by event queues */
1890 		efx->irq_level = 0x1f;
1891 	else
1892 		/* Use a valid MSI-X vector */
1893 		efx->irq_level = 0;
1894 
1895 	/* Enable all the genuinely fatal interrupts.  (They are still
1896 	 * masked by the overall interrupt mask, controlled by
1897 	 * falcon_interrupts()).
1898 	 *
1899 	 * Note: All other fatal interrupts are enabled
1900 	 */
1901 	EFX_POPULATE_OWORD_3(temp,
1902 			     FRF_AZ_ILL_ADR_INT_KER_EN, 1,
1903 			     FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
1904 			     FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
1905 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1906 		EFX_SET_OWORD_FIELD(temp, FRF_CZ_SRAM_PERR_INT_P_KER_EN, 1);
1907 	EFX_INVERT_OWORD(temp);
1908 	efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
1909 
1910 	efx_nic_push_rx_indir_table(efx);
1911 
1912 	/* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
1913 	 * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
1914 	 */
1915 	efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
1916 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
1917 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
1918 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
1919 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 1);
1920 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
1921 	/* Enable SW_EV to inherit in char driver - assume harmless here */
1922 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
1923 	/* Prefetch threshold 2 => fetch when descriptor cache half empty */
1924 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
1925 	/* Disable hardware watchdog which can misfire */
1926 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
1927 	/* Squash TX of packets of 16 bytes or less */
1928 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1929 		EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
1930 	efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
1931 
1932 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
1933 		EFX_POPULATE_OWORD_4(temp,
1934 				     /* Default values */
1935 				     FRF_BZ_TX_PACE_SB_NOT_AF, 0x15,
1936 				     FRF_BZ_TX_PACE_SB_AF, 0xb,
1937 				     FRF_BZ_TX_PACE_FB_BASE, 0,
1938 				     /* Allow large pace values in the
1939 				      * fast bin. */
1940 				     FRF_BZ_TX_PACE_BIN_TH,
1941 				     FFE_BZ_TX_PACE_RESERVED);
1942 		efx_writeo(efx, &temp, FR_BZ_TX_PACE);
1943 	}
1944 }
1945 
1946 /* Register dump */
1947 
1948 #define REGISTER_REVISION_A	1
1949 #define REGISTER_REVISION_B	2
1950 #define REGISTER_REVISION_C	3
1951 #define REGISTER_REVISION_Z	3	/* latest revision */
1952 
1953 struct efx_nic_reg {
1954 	u32 offset:24;
1955 	u32 min_revision:2, max_revision:2;
1956 };
1957 
1958 #define REGISTER(name, min_rev, max_rev) {				\
1959 	FR_ ## min_rev ## max_rev ## _ ## name,				\
1960 	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev	\
1961 }
1962 #define REGISTER_AA(name) REGISTER(name, A, A)
1963 #define REGISTER_AB(name) REGISTER(name, A, B)
1964 #define REGISTER_AZ(name) REGISTER(name, A, Z)
1965 #define REGISTER_BB(name) REGISTER(name, B, B)
1966 #define REGISTER_BZ(name) REGISTER(name, B, Z)
1967 #define REGISTER_CZ(name) REGISTER(name, C, Z)
1968 
1969 static const struct efx_nic_reg efx_nic_regs[] = {
1970 	REGISTER_AZ(ADR_REGION),
1971 	REGISTER_AZ(INT_EN_KER),
1972 	REGISTER_BZ(INT_EN_CHAR),
1973 	REGISTER_AZ(INT_ADR_KER),
1974 	REGISTER_BZ(INT_ADR_CHAR),
1975 	/* INT_ACK_KER is WO */
1976 	/* INT_ISR0 is RC */
1977 	REGISTER_AZ(HW_INIT),
1978 	REGISTER_CZ(USR_EV_CFG),
1979 	REGISTER_AB(EE_SPI_HCMD),
1980 	REGISTER_AB(EE_SPI_HADR),
1981 	REGISTER_AB(EE_SPI_HDATA),
1982 	REGISTER_AB(EE_BASE_PAGE),
1983 	REGISTER_AB(EE_VPD_CFG0),
1984 	/* EE_VPD_SW_CNTL and EE_VPD_SW_DATA are not used */
1985 	/* PMBX_DBG_IADDR and PBMX_DBG_IDATA are indirect */
1986 	/* PCIE_CORE_INDIRECT is indirect */
1987 	REGISTER_AB(NIC_STAT),
1988 	REGISTER_AB(GPIO_CTL),
1989 	REGISTER_AB(GLB_CTL),
1990 	/* FATAL_INTR_KER and FATAL_INTR_CHAR are partly RC */
1991 	REGISTER_BZ(DP_CTRL),
1992 	REGISTER_AZ(MEM_STAT),
1993 	REGISTER_AZ(CS_DEBUG),
1994 	REGISTER_AZ(ALTERA_BUILD),
1995 	REGISTER_AZ(CSR_SPARE),
1996 	REGISTER_AB(PCIE_SD_CTL0123),
1997 	REGISTER_AB(PCIE_SD_CTL45),
1998 	REGISTER_AB(PCIE_PCS_CTL_STAT),
1999 	/* DEBUG_DATA_OUT is not used */
2000 	/* DRV_EV is WO */
2001 	REGISTER_AZ(EVQ_CTL),
2002 	REGISTER_AZ(EVQ_CNT1),
2003 	REGISTER_AZ(EVQ_CNT2),
2004 	REGISTER_AZ(BUF_TBL_CFG),
2005 	REGISTER_AZ(SRM_RX_DC_CFG),
2006 	REGISTER_AZ(SRM_TX_DC_CFG),
2007 	REGISTER_AZ(SRM_CFG),
2008 	/* BUF_TBL_UPD is WO */
2009 	REGISTER_AZ(SRM_UPD_EVQ),
2010 	REGISTER_AZ(SRAM_PARITY),
2011 	REGISTER_AZ(RX_CFG),
2012 	REGISTER_BZ(RX_FILTER_CTL),
2013 	/* RX_FLUSH_DESCQ is WO */
2014 	REGISTER_AZ(RX_DC_CFG),
2015 	REGISTER_AZ(RX_DC_PF_WM),
2016 	REGISTER_BZ(RX_RSS_TKEY),
2017 	/* RX_NODESC_DROP is RC */
2018 	REGISTER_AA(RX_SELF_RST),
2019 	/* RX_DEBUG, RX_PUSH_DROP are not used */
2020 	REGISTER_CZ(RX_RSS_IPV6_REG1),
2021 	REGISTER_CZ(RX_RSS_IPV6_REG2),
2022 	REGISTER_CZ(RX_RSS_IPV6_REG3),
2023 	/* TX_FLUSH_DESCQ is WO */
2024 	REGISTER_AZ(TX_DC_CFG),
2025 	REGISTER_AA(TX_CHKSM_CFG),
2026 	REGISTER_AZ(TX_CFG),
2027 	/* TX_PUSH_DROP is not used */
2028 	REGISTER_AZ(TX_RESERVED),
2029 	REGISTER_BZ(TX_PACE),
2030 	/* TX_PACE_DROP_QID is RC */
2031 	REGISTER_BB(TX_VLAN),
2032 	REGISTER_BZ(TX_IPFIL_PORTEN),
2033 	REGISTER_AB(MD_TXD),
2034 	REGISTER_AB(MD_RXD),
2035 	REGISTER_AB(MD_CS),
2036 	REGISTER_AB(MD_PHY_ADR),
2037 	REGISTER_AB(MD_ID),
2038 	/* MD_STAT is RC */
2039 	REGISTER_AB(MAC_STAT_DMA),
2040 	REGISTER_AB(MAC_CTRL),
2041 	REGISTER_BB(GEN_MODE),
2042 	REGISTER_AB(MAC_MC_HASH_REG0),
2043 	REGISTER_AB(MAC_MC_HASH_REG1),
2044 	REGISTER_AB(GM_CFG1),
2045 	REGISTER_AB(GM_CFG2),
2046 	/* GM_IPG and GM_HD are not used */
2047 	REGISTER_AB(GM_MAX_FLEN),
2048 	/* GM_TEST is not used */
2049 	REGISTER_AB(GM_ADR1),
2050 	REGISTER_AB(GM_ADR2),
2051 	REGISTER_AB(GMF_CFG0),
2052 	REGISTER_AB(GMF_CFG1),
2053 	REGISTER_AB(GMF_CFG2),
2054 	REGISTER_AB(GMF_CFG3),
2055 	REGISTER_AB(GMF_CFG4),
2056 	REGISTER_AB(GMF_CFG5),
2057 	REGISTER_BB(TX_SRC_MAC_CTL),
2058 	REGISTER_AB(XM_ADR_LO),
2059 	REGISTER_AB(XM_ADR_HI),
2060 	REGISTER_AB(XM_GLB_CFG),
2061 	REGISTER_AB(XM_TX_CFG),
2062 	REGISTER_AB(XM_RX_CFG),
2063 	REGISTER_AB(XM_MGT_INT_MASK),
2064 	REGISTER_AB(XM_FC),
2065 	REGISTER_AB(XM_PAUSE_TIME),
2066 	REGISTER_AB(XM_TX_PARAM),
2067 	REGISTER_AB(XM_RX_PARAM),
2068 	/* XM_MGT_INT_MSK (note no 'A') is RC */
2069 	REGISTER_AB(XX_PWR_RST),
2070 	REGISTER_AB(XX_SD_CTL),
2071 	REGISTER_AB(XX_TXDRV_CTL),
2072 	/* XX_PRBS_CTL, XX_PRBS_CHK and XX_PRBS_ERR are not used */
2073 	/* XX_CORE_STAT is partly RC */
2074 };
2075 
2076 struct efx_nic_reg_table {
2077 	u32 offset:24;
2078 	u32 min_revision:2, max_revision:2;
2079 	u32 step:6, rows:21;
2080 };
2081 
2082 #define REGISTER_TABLE_DIMENSIONS(_, offset, min_rev, max_rev, step, rows) { \
2083 	offset,								\
2084 	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev,	\
2085 	step, rows							\
2086 }
2087 #define REGISTER_TABLE(name, min_rev, max_rev)				\
2088 	REGISTER_TABLE_DIMENSIONS(					\
2089 		name, FR_ ## min_rev ## max_rev ## _ ## name,		\
2090 		min_rev, max_rev,					\
2091 		FR_ ## min_rev ## max_rev ## _ ## name ## _STEP,	\
2092 		FR_ ## min_rev ## max_rev ## _ ## name ## _ROWS)
2093 #define REGISTER_TABLE_AA(name) REGISTER_TABLE(name, A, A)
2094 #define REGISTER_TABLE_AZ(name) REGISTER_TABLE(name, A, Z)
2095 #define REGISTER_TABLE_BB(name) REGISTER_TABLE(name, B, B)
2096 #define REGISTER_TABLE_BZ(name) REGISTER_TABLE(name, B, Z)
2097 #define REGISTER_TABLE_BB_CZ(name)					\
2098 	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, B, B,		\
2099 				  FR_BZ_ ## name ## _STEP,		\
2100 				  FR_BB_ ## name ## _ROWS),		\
2101 	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, C, Z,		\
2102 				  FR_BZ_ ## name ## _STEP,		\
2103 				  FR_CZ_ ## name ## _ROWS)
2104 #define REGISTER_TABLE_CZ(name) REGISTER_TABLE(name, C, Z)
2105 
2106 static const struct efx_nic_reg_table efx_nic_reg_tables[] = {
2107 	/* DRIVER is not used */
2108 	/* EVQ_RPTR, TIMER_COMMAND, USR_EV and {RX,TX}_DESC_UPD are WO */
2109 	REGISTER_TABLE_BB(TX_IPFIL_TBL),
2110 	REGISTER_TABLE_BB(TX_SRC_MAC_TBL),
2111 	REGISTER_TABLE_AA(RX_DESC_PTR_TBL_KER),
2112 	REGISTER_TABLE_BB_CZ(RX_DESC_PTR_TBL),
2113 	REGISTER_TABLE_AA(TX_DESC_PTR_TBL_KER),
2114 	REGISTER_TABLE_BB_CZ(TX_DESC_PTR_TBL),
2115 	REGISTER_TABLE_AA(EVQ_PTR_TBL_KER),
2116 	REGISTER_TABLE_BB_CZ(EVQ_PTR_TBL),
2117 	/* We can't reasonably read all of the buffer table (up to 8MB!).
2118 	 * However this driver will only use a few entries.  Reading
2119 	 * 1K entries allows for some expansion of queue count and
2120 	 * size before we need to change the version. */
2121 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL_KER, FR_AA_BUF_FULL_TBL_KER,
2122 				  A, A, 8, 1024),
2123 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL, FR_BZ_BUF_FULL_TBL,
2124 				  B, Z, 8, 1024),
2125 	REGISTER_TABLE_CZ(RX_MAC_FILTER_TBL0),
2126 	REGISTER_TABLE_BB_CZ(TIMER_TBL),
2127 	REGISTER_TABLE_BB_CZ(TX_PACE_TBL),
2128 	REGISTER_TABLE_BZ(RX_INDIRECTION_TBL),
2129 	/* TX_FILTER_TBL0 is huge and not used by this driver */
2130 	REGISTER_TABLE_CZ(TX_MAC_FILTER_TBL0),
2131 	REGISTER_TABLE_CZ(MC_TREG_SMEM),
2132 	/* MSIX_PBA_TABLE is not mapped */
2133 	/* SRM_DBG is not mapped (and is redundant with BUF_FLL_TBL) */
2134 	REGISTER_TABLE_BZ(RX_FILTER_TBL0),
2135 };
2136 
2137 size_t efx_nic_get_regs_len(struct efx_nic *efx)
2138 {
2139 	const struct efx_nic_reg *reg;
2140 	const struct efx_nic_reg_table *table;
2141 	size_t len = 0;
2142 
2143 	for (reg = efx_nic_regs;
2144 	     reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
2145 	     reg++)
2146 		if (efx->type->revision >= reg->min_revision &&
2147 		    efx->type->revision <= reg->max_revision)
2148 			len += sizeof(efx_oword_t);
2149 
2150 	for (table = efx_nic_reg_tables;
2151 	     table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
2152 	     table++)
2153 		if (efx->type->revision >= table->min_revision &&
2154 		    efx->type->revision <= table->max_revision)
2155 			len += table->rows * min_t(size_t, table->step, 16);
2156 
2157 	return len;
2158 }
2159 
2160 void efx_nic_get_regs(struct efx_nic *efx, void *buf)
2161 {
2162 	const struct efx_nic_reg *reg;
2163 	const struct efx_nic_reg_table *table;
2164 
2165 	for (reg = efx_nic_regs;
2166 	     reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
2167 	     reg++) {
2168 		if (efx->type->revision >= reg->min_revision &&
2169 		    efx->type->revision <= reg->max_revision) {
2170 			efx_reado(efx, (efx_oword_t *)buf, reg->offset);
2171 			buf += sizeof(efx_oword_t);
2172 		}
2173 	}
2174 
2175 	for (table = efx_nic_reg_tables;
2176 	     table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
2177 	     table++) {
2178 		size_t size, i;
2179 
2180 		if (!(efx->type->revision >= table->min_revision &&
2181 		      efx->type->revision <= table->max_revision))
2182 			continue;
2183 
2184 		size = min_t(size_t, table->step, 16);
2185 
2186 		for (i = 0; i < table->rows; i++) {
2187 			switch (table->step) {
2188 			case 4: /* 32-bit SRAM */
2189 				efx_readd(efx, buf, table->offset + 4 * i);
2190 				break;
2191 			case 8: /* 64-bit SRAM */
2192 				efx_sram_readq(efx,
2193 					       efx->membase + table->offset,
2194 					       buf, i);
2195 				break;
2196 			case 16: /* 128-bit-readable register */
2197 				efx_reado_table(efx, buf, table->offset, i);
2198 				break;
2199 			case 32: /* 128-bit register, interleaved */
2200 				efx_reado_table(efx, buf, table->offset, 2 * i);
2201 				break;
2202 			default:
2203 				WARN_ON(1);
2204 				return;
2205 			}
2206 			buf += size;
2207 		}
2208 	}
2209 }
2210