xref: /openbmc/linux/drivers/net/ethernet/sfc/nic.c (revision 5bbe2f4f)
1 /****************************************************************************
2  * Driver for Solarflare Solarstorm network controllers and boards
3  * Copyright 2005-2006 Fen Systems Ltd.
4  * Copyright 2006-2011 Solarflare Communications Inc.
5  *
6  * This program is free software; you can redistribute it and/or modify it
7  * under the terms of the GNU General Public License version 2 as published
8  * by the Free Software Foundation, incorporated herein by reference.
9  */
10 
11 #include <linux/bitops.h>
12 #include <linux/delay.h>
13 #include <linux/interrupt.h>
14 #include <linux/pci.h>
15 #include <linux/module.h>
16 #include <linux/seq_file.h>
17 #include "net_driver.h"
18 #include "bitfield.h"
19 #include "efx.h"
20 #include "nic.h"
21 #include "regs.h"
22 #include "io.h"
23 #include "workarounds.h"
24 
25 /**************************************************************************
26  *
27  * Configurable values
28  *
29  **************************************************************************
30  */
31 
32 /* This is set to 16 for a good reason.  In summary, if larger than
33  * 16, the descriptor cache holds more than a default socket
34  * buffer's worth of packets (for UDP we can only have at most one
35  * socket buffer's worth outstanding).  This combined with the fact
36  * that we only get 1 TX event per descriptor cache means the NIC
37  * goes idle.
38  */
39 #define TX_DC_ENTRIES 16
40 #define TX_DC_ENTRIES_ORDER 1
41 
42 #define RX_DC_ENTRIES 64
43 #define RX_DC_ENTRIES_ORDER 3
44 
45 /* If EFX_MAX_INT_ERRORS internal errors occur within
46  * EFX_INT_ERROR_EXPIRE seconds, we consider the NIC broken and
47  * disable it.
48  */
49 #define EFX_INT_ERROR_EXPIRE 3600
50 #define EFX_MAX_INT_ERRORS 5
51 
52 /* Depth of RX flush request fifo */
53 #define EFX_RX_FLUSH_COUNT 4
54 
55 /* Driver generated events */
56 #define _EFX_CHANNEL_MAGIC_TEST		0x000101
57 #define _EFX_CHANNEL_MAGIC_FILL		0x000102
58 #define _EFX_CHANNEL_MAGIC_RX_DRAIN	0x000103
59 #define _EFX_CHANNEL_MAGIC_TX_DRAIN	0x000104
60 
61 #define _EFX_CHANNEL_MAGIC(_code, _data)	((_code) << 8 | (_data))
62 #define _EFX_CHANNEL_MAGIC_CODE(_magic)		((_magic) >> 8)
63 
64 #define EFX_CHANNEL_MAGIC_TEST(_channel)				\
65 	_EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_TEST, (_channel)->channel)
66 #define EFX_CHANNEL_MAGIC_FILL(_rx_queue)				\
67 	_EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_FILL,			\
68 			   efx_rx_queue_index(_rx_queue))
69 #define EFX_CHANNEL_MAGIC_RX_DRAIN(_rx_queue)				\
70 	_EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_RX_DRAIN,			\
71 			   efx_rx_queue_index(_rx_queue))
72 #define EFX_CHANNEL_MAGIC_TX_DRAIN(_tx_queue)				\
73 	_EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_TX_DRAIN,			\
74 			   (_tx_queue)->queue)
75 
76 /**************************************************************************
77  *
78  * Solarstorm hardware access
79  *
80  **************************************************************************/
81 
82 static inline void efx_write_buf_tbl(struct efx_nic *efx, efx_qword_t *value,
83 				     unsigned int index)
84 {
85 	efx_sram_writeq(efx, efx->membase + efx->type->buf_tbl_base,
86 			value, index);
87 }
88 
89 /* Read the current event from the event queue */
90 static inline efx_qword_t *efx_event(struct efx_channel *channel,
91 				     unsigned int index)
92 {
93 	return ((efx_qword_t *) (channel->eventq.addr)) +
94 		(index & channel->eventq_mask);
95 }
96 
97 /* See if an event is present
98  *
99  * We check both the high and low dword of the event for all ones.  We
100  * wrote all ones when we cleared the event, and no valid event can
101  * have all ones in either its high or low dwords.  This approach is
102  * robust against reordering.
103  *
104  * Note that using a single 64-bit comparison is incorrect; even
105  * though the CPU read will be atomic, the DMA write may not be.
106  */
107 static inline int efx_event_present(efx_qword_t *event)
108 {
109 	return !(EFX_DWORD_IS_ALL_ONES(event->dword[0]) |
110 		  EFX_DWORD_IS_ALL_ONES(event->dword[1]));
111 }
112 
113 static bool efx_masked_compare_oword(const efx_oword_t *a, const efx_oword_t *b,
114 				     const efx_oword_t *mask)
115 {
116 	return ((a->u64[0] ^ b->u64[0]) & mask->u64[0]) ||
117 		((a->u64[1] ^ b->u64[1]) & mask->u64[1]);
118 }
119 
120 int efx_nic_test_registers(struct efx_nic *efx,
121 			   const struct efx_nic_register_test *regs,
122 			   size_t n_regs)
123 {
124 	unsigned address = 0, i, j;
125 	efx_oword_t mask, imask, original, reg, buf;
126 
127 	/* Falcon should be in loopback to isolate the XMAC from the PHY */
128 	WARN_ON(!LOOPBACK_INTERNAL(efx));
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 	/* All zeros is a potentially valid event so memset to 0xff */
262 	memset(buffer->addr, 0xff, len);
263 
264 	/* Select new buffer ID */
265 	buffer->index = efx->next_buffer_table;
266 	efx->next_buffer_table += buffer->entries;
267 
268 	netif_dbg(efx, probe, efx->net_dev,
269 		  "allocating special buffers %d-%d at %llx+%x "
270 		  "(virt %p phys %llx)\n", buffer->index,
271 		  buffer->index + buffer->entries - 1,
272 		  (u64)buffer->dma_addr, len,
273 		  buffer->addr, (u64)virt_to_phys(buffer->addr));
274 
275 	return 0;
276 }
277 
278 static void
279 efx_free_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
280 {
281 	if (!buffer->addr)
282 		return;
283 
284 	netif_dbg(efx, hw, efx->net_dev,
285 		  "deallocating special buffers %d-%d at %llx+%x "
286 		  "(virt %p phys %llx)\n", buffer->index,
287 		  buffer->index + buffer->entries - 1,
288 		  (u64)buffer->dma_addr, buffer->len,
289 		  buffer->addr, (u64)virt_to_phys(buffer->addr));
290 
291 	dma_free_coherent(&efx->pci_dev->dev, buffer->len, buffer->addr,
292 			  buffer->dma_addr);
293 	buffer->addr = NULL;
294 	buffer->entries = 0;
295 }
296 
297 /**************************************************************************
298  *
299  * Generic buffer handling
300  * These buffers are used for interrupt status and MAC stats
301  *
302  **************************************************************************/
303 
304 int efx_nic_alloc_buffer(struct efx_nic *efx, struct efx_buffer *buffer,
305 			 unsigned int len)
306 {
307 	buffer->addr = pci_alloc_consistent(efx->pci_dev, len,
308 					    &buffer->dma_addr);
309 	if (!buffer->addr)
310 		return -ENOMEM;
311 	buffer->len = len;
312 	memset(buffer->addr, 0, len);
313 	return 0;
314 }
315 
316 void efx_nic_free_buffer(struct efx_nic *efx, struct efx_buffer *buffer)
317 {
318 	if (buffer->addr) {
319 		pci_free_consistent(efx->pci_dev, buffer->len,
320 				    buffer->addr, buffer->dma_addr);
321 		buffer->addr = NULL;
322 	}
323 }
324 
325 /**************************************************************************
326  *
327  * TX path
328  *
329  **************************************************************************/
330 
331 /* Returns a pointer to the specified transmit descriptor in the TX
332  * descriptor queue belonging to the specified channel.
333  */
334 static inline efx_qword_t *
335 efx_tx_desc(struct efx_tx_queue *tx_queue, unsigned int index)
336 {
337 	return ((efx_qword_t *) (tx_queue->txd.addr)) + index;
338 }
339 
340 /* This writes to the TX_DESC_WPTR; write pointer for TX descriptor ring */
341 static inline void efx_notify_tx_desc(struct efx_tx_queue *tx_queue)
342 {
343 	unsigned write_ptr;
344 	efx_dword_t reg;
345 
346 	write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
347 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_TX_DESC_WPTR_DWORD, write_ptr);
348 	efx_writed_page(tx_queue->efx, &reg,
349 			FR_AZ_TX_DESC_UPD_DWORD_P0, tx_queue->queue);
350 }
351 
352 /* Write pointer and first descriptor for TX descriptor ring */
353 static inline void efx_push_tx_desc(struct efx_tx_queue *tx_queue,
354 				    const efx_qword_t *txd)
355 {
356 	unsigned write_ptr;
357 	efx_oword_t reg;
358 
359 	BUILD_BUG_ON(FRF_AZ_TX_DESC_LBN != 0);
360 	BUILD_BUG_ON(FR_AA_TX_DESC_UPD_KER != FR_BZ_TX_DESC_UPD_P0);
361 
362 	write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
363 	EFX_POPULATE_OWORD_2(reg, FRF_AZ_TX_DESC_PUSH_CMD, true,
364 			     FRF_AZ_TX_DESC_WPTR, write_ptr);
365 	reg.qword[0] = *txd;
366 	efx_writeo_page(tx_queue->efx, &reg,
367 			FR_BZ_TX_DESC_UPD_P0, tx_queue->queue);
368 }
369 
370 static inline bool
371 efx_may_push_tx_desc(struct efx_tx_queue *tx_queue, unsigned int write_count)
372 {
373 	unsigned empty_read_count = ACCESS_ONCE(tx_queue->empty_read_count);
374 
375 	if (empty_read_count == 0)
376 		return false;
377 
378 	tx_queue->empty_read_count = 0;
379 	return ((empty_read_count ^ write_count) & ~EFX_EMPTY_COUNT_VALID) == 0;
380 }
381 
382 /* For each entry inserted into the software descriptor ring, create a
383  * descriptor in the hardware TX descriptor ring (in host memory), and
384  * write a doorbell.
385  */
386 void efx_nic_push_buffers(struct efx_tx_queue *tx_queue)
387 {
388 
389 	struct efx_tx_buffer *buffer;
390 	efx_qword_t *txd;
391 	unsigned write_ptr;
392 	unsigned old_write_count = tx_queue->write_count;
393 
394 	BUG_ON(tx_queue->write_count == tx_queue->insert_count);
395 
396 	do {
397 		write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
398 		buffer = &tx_queue->buffer[write_ptr];
399 		txd = efx_tx_desc(tx_queue, write_ptr);
400 		++tx_queue->write_count;
401 
402 		/* Create TX descriptor ring entry */
403 		EFX_POPULATE_QWORD_4(*txd,
404 				     FSF_AZ_TX_KER_CONT, buffer->continuation,
405 				     FSF_AZ_TX_KER_BYTE_COUNT, buffer->len,
406 				     FSF_AZ_TX_KER_BUF_REGION, 0,
407 				     FSF_AZ_TX_KER_BUF_ADDR, buffer->dma_addr);
408 	} while (tx_queue->write_count != tx_queue->insert_count);
409 
410 	wmb(); /* Ensure descriptors are written before they are fetched */
411 
412 	if (efx_may_push_tx_desc(tx_queue, old_write_count)) {
413 		txd = efx_tx_desc(tx_queue,
414 				  old_write_count & tx_queue->ptr_mask);
415 		efx_push_tx_desc(tx_queue, txd);
416 		++tx_queue->pushes;
417 	} else {
418 		efx_notify_tx_desc(tx_queue);
419 	}
420 }
421 
422 /* Allocate hardware resources for a TX queue */
423 int efx_nic_probe_tx(struct efx_tx_queue *tx_queue)
424 {
425 	struct efx_nic *efx = tx_queue->efx;
426 	unsigned entries;
427 
428 	entries = tx_queue->ptr_mask + 1;
429 	return efx_alloc_special_buffer(efx, &tx_queue->txd,
430 					entries * sizeof(efx_qword_t));
431 }
432 
433 void efx_nic_init_tx(struct efx_tx_queue *tx_queue)
434 {
435 	struct efx_nic *efx = tx_queue->efx;
436 	efx_oword_t reg;
437 
438 	/* Pin TX descriptor ring */
439 	efx_init_special_buffer(efx, &tx_queue->txd);
440 
441 	/* Push TX descriptor ring to card */
442 	EFX_POPULATE_OWORD_10(reg,
443 			      FRF_AZ_TX_DESCQ_EN, 1,
444 			      FRF_AZ_TX_ISCSI_DDIG_EN, 0,
445 			      FRF_AZ_TX_ISCSI_HDIG_EN, 0,
446 			      FRF_AZ_TX_DESCQ_BUF_BASE_ID, tx_queue->txd.index,
447 			      FRF_AZ_TX_DESCQ_EVQ_ID,
448 			      tx_queue->channel->channel,
449 			      FRF_AZ_TX_DESCQ_OWNER_ID, 0,
450 			      FRF_AZ_TX_DESCQ_LABEL, tx_queue->queue,
451 			      FRF_AZ_TX_DESCQ_SIZE,
452 			      __ffs(tx_queue->txd.entries),
453 			      FRF_AZ_TX_DESCQ_TYPE, 0,
454 			      FRF_BZ_TX_NON_IP_DROP_DIS, 1);
455 
456 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
457 		int csum = tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD;
458 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_IP_CHKSM_DIS, !csum);
459 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_TCP_CHKSM_DIS,
460 				    !csum);
461 	}
462 
463 	efx_writeo_table(efx, &reg, efx->type->txd_ptr_tbl_base,
464 			 tx_queue->queue);
465 
466 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0) {
467 		/* Only 128 bits in this register */
468 		BUILD_BUG_ON(EFX_MAX_TX_QUEUES > 128);
469 
470 		efx_reado(efx, &reg, FR_AA_TX_CHKSM_CFG);
471 		if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD)
472 			clear_bit_le(tx_queue->queue, (void *)&reg);
473 		else
474 			set_bit_le(tx_queue->queue, (void *)&reg);
475 		efx_writeo(efx, &reg, FR_AA_TX_CHKSM_CFG);
476 	}
477 
478 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
479 		EFX_POPULATE_OWORD_1(reg,
480 				     FRF_BZ_TX_PACE,
481 				     (tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
482 				     FFE_BZ_TX_PACE_OFF :
483 				     FFE_BZ_TX_PACE_RESERVED);
484 		efx_writeo_table(efx, &reg, FR_BZ_TX_PACE_TBL,
485 				 tx_queue->queue);
486 	}
487 }
488 
489 static void efx_flush_tx_queue(struct efx_tx_queue *tx_queue)
490 {
491 	struct efx_nic *efx = tx_queue->efx;
492 	efx_oword_t tx_flush_descq;
493 
494 	EFX_POPULATE_OWORD_2(tx_flush_descq,
495 			     FRF_AZ_TX_FLUSH_DESCQ_CMD, 1,
496 			     FRF_AZ_TX_FLUSH_DESCQ, tx_queue->queue);
497 	efx_writeo(efx, &tx_flush_descq, FR_AZ_TX_FLUSH_DESCQ);
498 }
499 
500 void efx_nic_fini_tx(struct efx_tx_queue *tx_queue)
501 {
502 	struct efx_nic *efx = tx_queue->efx;
503 	efx_oword_t tx_desc_ptr;
504 
505 	/* Remove TX descriptor ring from card */
506 	EFX_ZERO_OWORD(tx_desc_ptr);
507 	efx_writeo_table(efx, &tx_desc_ptr, efx->type->txd_ptr_tbl_base,
508 			 tx_queue->queue);
509 
510 	/* Unpin TX descriptor ring */
511 	efx_fini_special_buffer(efx, &tx_queue->txd);
512 }
513 
514 /* Free buffers backing TX queue */
515 void efx_nic_remove_tx(struct efx_tx_queue *tx_queue)
516 {
517 	efx_free_special_buffer(tx_queue->efx, &tx_queue->txd);
518 }
519 
520 /**************************************************************************
521  *
522  * RX path
523  *
524  **************************************************************************/
525 
526 /* Returns a pointer to the specified descriptor in the RX descriptor queue */
527 static inline efx_qword_t *
528 efx_rx_desc(struct efx_rx_queue *rx_queue, unsigned int index)
529 {
530 	return ((efx_qword_t *) (rx_queue->rxd.addr)) + index;
531 }
532 
533 /* This creates an entry in the RX descriptor queue */
534 static inline void
535 efx_build_rx_desc(struct efx_rx_queue *rx_queue, unsigned index)
536 {
537 	struct efx_rx_buffer *rx_buf;
538 	efx_qword_t *rxd;
539 
540 	rxd = efx_rx_desc(rx_queue, index);
541 	rx_buf = efx_rx_buffer(rx_queue, index);
542 	EFX_POPULATE_QWORD_3(*rxd,
543 			     FSF_AZ_RX_KER_BUF_SIZE,
544 			     rx_buf->len -
545 			     rx_queue->efx->type->rx_buffer_padding,
546 			     FSF_AZ_RX_KER_BUF_REGION, 0,
547 			     FSF_AZ_RX_KER_BUF_ADDR, rx_buf->dma_addr);
548 }
549 
550 /* This writes to the RX_DESC_WPTR register for the specified receive
551  * descriptor ring.
552  */
553 void efx_nic_notify_rx_desc(struct efx_rx_queue *rx_queue)
554 {
555 	struct efx_nic *efx = rx_queue->efx;
556 	efx_dword_t reg;
557 	unsigned write_ptr;
558 
559 	while (rx_queue->notified_count != rx_queue->added_count) {
560 		efx_build_rx_desc(
561 			rx_queue,
562 			rx_queue->notified_count & rx_queue->ptr_mask);
563 		++rx_queue->notified_count;
564 	}
565 
566 	wmb();
567 	write_ptr = rx_queue->added_count & rx_queue->ptr_mask;
568 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_RX_DESC_WPTR_DWORD, write_ptr);
569 	efx_writed_page(efx, &reg, FR_AZ_RX_DESC_UPD_DWORD_P0,
570 			efx_rx_queue_index(rx_queue));
571 }
572 
573 int efx_nic_probe_rx(struct efx_rx_queue *rx_queue)
574 {
575 	struct efx_nic *efx = rx_queue->efx;
576 	unsigned entries;
577 
578 	entries = rx_queue->ptr_mask + 1;
579 	return efx_alloc_special_buffer(efx, &rx_queue->rxd,
580 					entries * sizeof(efx_qword_t));
581 }
582 
583 void efx_nic_init_rx(struct efx_rx_queue *rx_queue)
584 {
585 	efx_oword_t rx_desc_ptr;
586 	struct efx_nic *efx = rx_queue->efx;
587 	bool is_b0 = efx_nic_rev(efx) >= EFX_REV_FALCON_B0;
588 	bool iscsi_digest_en = is_b0;
589 
590 	netif_dbg(efx, hw, efx->net_dev,
591 		  "RX queue %d ring in special buffers %d-%d\n",
592 		  efx_rx_queue_index(rx_queue), rx_queue->rxd.index,
593 		  rx_queue->rxd.index + rx_queue->rxd.entries - 1);
594 
595 	/* Pin RX descriptor ring */
596 	efx_init_special_buffer(efx, &rx_queue->rxd);
597 
598 	/* Push RX descriptor ring to card */
599 	EFX_POPULATE_OWORD_10(rx_desc_ptr,
600 			      FRF_AZ_RX_ISCSI_DDIG_EN, iscsi_digest_en,
601 			      FRF_AZ_RX_ISCSI_HDIG_EN, iscsi_digest_en,
602 			      FRF_AZ_RX_DESCQ_BUF_BASE_ID, rx_queue->rxd.index,
603 			      FRF_AZ_RX_DESCQ_EVQ_ID,
604 			      efx_rx_queue_channel(rx_queue)->channel,
605 			      FRF_AZ_RX_DESCQ_OWNER_ID, 0,
606 			      FRF_AZ_RX_DESCQ_LABEL,
607 			      efx_rx_queue_index(rx_queue),
608 			      FRF_AZ_RX_DESCQ_SIZE,
609 			      __ffs(rx_queue->rxd.entries),
610 			      FRF_AZ_RX_DESCQ_TYPE, 0 /* kernel queue */ ,
611 			      /* For >=B0 this is scatter so disable */
612 			      FRF_AZ_RX_DESCQ_JUMBO, !is_b0,
613 			      FRF_AZ_RX_DESCQ_EN, 1);
614 	efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
615 			 efx_rx_queue_index(rx_queue));
616 }
617 
618 static void efx_flush_rx_queue(struct efx_rx_queue *rx_queue)
619 {
620 	struct efx_nic *efx = rx_queue->efx;
621 	efx_oword_t rx_flush_descq;
622 
623 	EFX_POPULATE_OWORD_2(rx_flush_descq,
624 			     FRF_AZ_RX_FLUSH_DESCQ_CMD, 1,
625 			     FRF_AZ_RX_FLUSH_DESCQ,
626 			     efx_rx_queue_index(rx_queue));
627 	efx_writeo(efx, &rx_flush_descq, FR_AZ_RX_FLUSH_DESCQ);
628 }
629 
630 void efx_nic_fini_rx(struct efx_rx_queue *rx_queue)
631 {
632 	efx_oword_t rx_desc_ptr;
633 	struct efx_nic *efx = rx_queue->efx;
634 
635 	/* Remove RX descriptor ring from card */
636 	EFX_ZERO_OWORD(rx_desc_ptr);
637 	efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
638 			 efx_rx_queue_index(rx_queue));
639 
640 	/* Unpin RX descriptor ring */
641 	efx_fini_special_buffer(efx, &rx_queue->rxd);
642 }
643 
644 /* Free buffers backing RX queue */
645 void efx_nic_remove_rx(struct efx_rx_queue *rx_queue)
646 {
647 	efx_free_special_buffer(rx_queue->efx, &rx_queue->rxd);
648 }
649 
650 /**************************************************************************
651  *
652  * Flush handling
653  *
654  **************************************************************************/
655 
656 /* efx_nic_flush_queues() must be woken up when all flushes are completed,
657  * or more RX flushes can be kicked off.
658  */
659 static bool efx_flush_wake(struct efx_nic *efx)
660 {
661 	/* Ensure that all updates are visible to efx_nic_flush_queues() */
662 	smp_mb();
663 
664 	return (atomic_read(&efx->drain_pending) == 0 ||
665 		(atomic_read(&efx->rxq_flush_outstanding) < EFX_RX_FLUSH_COUNT
666 		 && atomic_read(&efx->rxq_flush_pending) > 0));
667 }
668 
669 /* Flush all the transmit queues, and continue flushing receive queues until
670  * they're all flushed. Wait for the DRAIN events to be recieved so that there
671  * are no more RX and TX events left on any channel. */
672 int efx_nic_flush_queues(struct efx_nic *efx)
673 {
674 	unsigned timeout = msecs_to_jiffies(5000); /* 5s for all flushes and drains */
675 	struct efx_channel *channel;
676 	struct efx_rx_queue *rx_queue;
677 	struct efx_tx_queue *tx_queue;
678 	int rc = 0;
679 
680 	efx->fc_disable++;
681 	efx->type->prepare_flush(efx);
682 
683 	efx_for_each_channel(channel, efx) {
684 		efx_for_each_channel_tx_queue(tx_queue, channel) {
685 			atomic_inc(&efx->drain_pending);
686 			efx_flush_tx_queue(tx_queue);
687 		}
688 		efx_for_each_channel_rx_queue(rx_queue, channel) {
689 			atomic_inc(&efx->drain_pending);
690 			rx_queue->flush_pending = true;
691 			atomic_inc(&efx->rxq_flush_pending);
692 		}
693 	}
694 
695 	while (timeout && atomic_read(&efx->drain_pending) > 0) {
696 		/* The hardware supports four concurrent rx flushes, each of
697 		 * which may need to be retried if there is an outstanding
698 		 * descriptor fetch
699 		 */
700 		efx_for_each_channel(channel, efx) {
701 			efx_for_each_channel_rx_queue(rx_queue, channel) {
702 				if (atomic_read(&efx->rxq_flush_outstanding) >=
703 				    EFX_RX_FLUSH_COUNT)
704 					break;
705 
706 				if (rx_queue->flush_pending) {
707 					rx_queue->flush_pending = false;
708 					atomic_dec(&efx->rxq_flush_pending);
709 					atomic_inc(&efx->rxq_flush_outstanding);
710 					efx_flush_rx_queue(rx_queue);
711 				}
712 			}
713 		}
714 
715 		timeout = wait_event_timeout(efx->flush_wq, efx_flush_wake(efx),
716 					     timeout);
717 	}
718 
719 	if (atomic_read(&efx->drain_pending)) {
720 		netif_err(efx, hw, efx->net_dev, "failed to flush %d queues "
721 			  "(rx %d+%d)\n", atomic_read(&efx->drain_pending),
722 			  atomic_read(&efx->rxq_flush_outstanding),
723 			  atomic_read(&efx->rxq_flush_pending));
724 		rc = -ETIMEDOUT;
725 
726 		atomic_set(&efx->drain_pending, 0);
727 		atomic_set(&efx->rxq_flush_pending, 0);
728 		atomic_set(&efx->rxq_flush_outstanding, 0);
729 	}
730 
731 	efx->fc_disable--;
732 
733 	return rc;
734 }
735 
736 /**************************************************************************
737  *
738  * Event queue processing
739  * Event queues are processed by per-channel tasklets.
740  *
741  **************************************************************************/
742 
743 /* Update a channel's event queue's read pointer (RPTR) register
744  *
745  * This writes the EVQ_RPTR_REG register for the specified channel's
746  * event queue.
747  */
748 void efx_nic_eventq_read_ack(struct efx_channel *channel)
749 {
750 	efx_dword_t reg;
751 	struct efx_nic *efx = channel->efx;
752 
753 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_EVQ_RPTR,
754 			     channel->eventq_read_ptr & channel->eventq_mask);
755 	efx_writed_table(efx, &reg, efx->type->evq_rptr_tbl_base,
756 			 channel->channel);
757 }
758 
759 /* Use HW to insert a SW defined event */
760 void efx_generate_event(struct efx_nic *efx, unsigned int evq,
761 			efx_qword_t *event)
762 {
763 	efx_oword_t drv_ev_reg;
764 
765 	BUILD_BUG_ON(FRF_AZ_DRV_EV_DATA_LBN != 0 ||
766 		     FRF_AZ_DRV_EV_DATA_WIDTH != 64);
767 	drv_ev_reg.u32[0] = event->u32[0];
768 	drv_ev_reg.u32[1] = event->u32[1];
769 	drv_ev_reg.u32[2] = 0;
770 	drv_ev_reg.u32[3] = 0;
771 	EFX_SET_OWORD_FIELD(drv_ev_reg, FRF_AZ_DRV_EV_QID, evq);
772 	efx_writeo(efx, &drv_ev_reg, FR_AZ_DRV_EV);
773 }
774 
775 static void efx_magic_event(struct efx_channel *channel, u32 magic)
776 {
777 	efx_qword_t event;
778 
779 	EFX_POPULATE_QWORD_2(event, FSF_AZ_EV_CODE,
780 			     FSE_AZ_EV_CODE_DRV_GEN_EV,
781 			     FSF_AZ_DRV_GEN_EV_MAGIC, magic);
782 	efx_generate_event(channel->efx, channel->channel, &event);
783 }
784 
785 /* Handle a transmit completion event
786  *
787  * The NIC batches TX completion events; the message we receive is of
788  * the form "complete all TX events up to this index".
789  */
790 static int
791 efx_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
792 {
793 	unsigned int tx_ev_desc_ptr;
794 	unsigned int tx_ev_q_label;
795 	struct efx_tx_queue *tx_queue;
796 	struct efx_nic *efx = channel->efx;
797 	int tx_packets = 0;
798 
799 	if (unlikely(ACCESS_ONCE(efx->reset_pending)))
800 		return 0;
801 
802 	if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
803 		/* Transmit completion */
804 		tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
805 		tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
806 		tx_queue = efx_channel_get_tx_queue(
807 			channel, tx_ev_q_label % EFX_TXQ_TYPES);
808 		tx_packets = ((tx_ev_desc_ptr - tx_queue->read_count) &
809 			      tx_queue->ptr_mask);
810 		channel->irq_mod_score += tx_packets;
811 		efx_xmit_done(tx_queue, tx_ev_desc_ptr);
812 	} else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
813 		/* Rewrite the FIFO write pointer */
814 		tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
815 		tx_queue = efx_channel_get_tx_queue(
816 			channel, tx_ev_q_label % EFX_TXQ_TYPES);
817 
818 		netif_tx_lock(efx->net_dev);
819 		efx_notify_tx_desc(tx_queue);
820 		netif_tx_unlock(efx->net_dev);
821 	} else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR) &&
822 		   EFX_WORKAROUND_10727(efx)) {
823 		efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
824 	} else {
825 		netif_err(efx, tx_err, efx->net_dev,
826 			  "channel %d unexpected TX event "
827 			  EFX_QWORD_FMT"\n", channel->channel,
828 			  EFX_QWORD_VAL(*event));
829 	}
830 
831 	return tx_packets;
832 }
833 
834 /* Detect errors included in the rx_evt_pkt_ok bit. */
835 static u16 efx_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
836 				const efx_qword_t *event)
837 {
838 	struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
839 	struct efx_nic *efx = rx_queue->efx;
840 	bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
841 	bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
842 	bool rx_ev_frm_trunc, rx_ev_drib_nib, rx_ev_tobe_disc;
843 	bool rx_ev_other_err, rx_ev_pause_frm;
844 	bool rx_ev_hdr_type, rx_ev_mcast_pkt;
845 	unsigned rx_ev_pkt_type;
846 
847 	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
848 	rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
849 	rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
850 	rx_ev_pkt_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_TYPE);
851 	rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
852 						 FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
853 	rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
854 						  FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
855 	rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
856 						   FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
857 	rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
858 	rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
859 	rx_ev_drib_nib = ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) ?
860 			  0 : EFX_QWORD_FIELD(*event, FSF_AA_RX_EV_DRIB_NIB));
861 	rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
862 
863 	/* Every error apart from tobe_disc and pause_frm */
864 	rx_ev_other_err = (rx_ev_drib_nib | rx_ev_tcp_udp_chksum_err |
865 			   rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
866 			   rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
867 
868 	/* Count errors that are not in MAC stats.  Ignore expected
869 	 * checksum errors during self-test. */
870 	if (rx_ev_frm_trunc)
871 		++channel->n_rx_frm_trunc;
872 	else if (rx_ev_tobe_disc)
873 		++channel->n_rx_tobe_disc;
874 	else if (!efx->loopback_selftest) {
875 		if (rx_ev_ip_hdr_chksum_err)
876 			++channel->n_rx_ip_hdr_chksum_err;
877 		else if (rx_ev_tcp_udp_chksum_err)
878 			++channel->n_rx_tcp_udp_chksum_err;
879 	}
880 
881 	/* TOBE_DISC is expected on unicast mismatches; don't print out an
882 	 * error message.  FRM_TRUNC indicates RXDP dropped the packet due
883 	 * to a FIFO overflow.
884 	 */
885 #ifdef DEBUG
886 	if (rx_ev_other_err && net_ratelimit()) {
887 		netif_dbg(efx, rx_err, efx->net_dev,
888 			  " RX queue %d unexpected RX event "
889 			  EFX_QWORD_FMT "%s%s%s%s%s%s%s%s\n",
890 			  efx_rx_queue_index(rx_queue), EFX_QWORD_VAL(*event),
891 			  rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
892 			  rx_ev_ip_hdr_chksum_err ?
893 			  " [IP_HDR_CHKSUM_ERR]" : "",
894 			  rx_ev_tcp_udp_chksum_err ?
895 			  " [TCP_UDP_CHKSUM_ERR]" : "",
896 			  rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
897 			  rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
898 			  rx_ev_drib_nib ? " [DRIB_NIB]" : "",
899 			  rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
900 			  rx_ev_pause_frm ? " [PAUSE]" : "");
901 	}
902 #endif
903 
904 	/* The frame must be discarded if any of these are true. */
905 	return (rx_ev_eth_crc_err | rx_ev_frm_trunc | rx_ev_drib_nib |
906 		rx_ev_tobe_disc | rx_ev_pause_frm) ?
907 		EFX_RX_PKT_DISCARD : 0;
908 }
909 
910 /* Handle receive events that are not in-order. */
911 static void
912 efx_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
913 {
914 	struct efx_nic *efx = rx_queue->efx;
915 	unsigned expected, dropped;
916 
917 	expected = rx_queue->removed_count & rx_queue->ptr_mask;
918 	dropped = (index - expected) & rx_queue->ptr_mask;
919 	netif_info(efx, rx_err, efx->net_dev,
920 		   "dropped %d events (index=%d expected=%d)\n",
921 		   dropped, index, expected);
922 
923 	efx_schedule_reset(efx, EFX_WORKAROUND_5676(efx) ?
924 			   RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
925 }
926 
927 /* Handle a packet received event
928  *
929  * The NIC gives a "discard" flag if it's a unicast packet with the
930  * wrong destination address
931  * Also "is multicast" and "matches multicast filter" flags can be used to
932  * discard non-matching multicast packets.
933  */
934 static void
935 efx_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
936 {
937 	unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
938 	unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
939 	unsigned expected_ptr;
940 	bool rx_ev_pkt_ok;
941 	u16 flags;
942 	struct efx_rx_queue *rx_queue;
943 	struct efx_nic *efx = channel->efx;
944 
945 	if (unlikely(ACCESS_ONCE(efx->reset_pending)))
946 		return;
947 
948 	/* Basic packet information */
949 	rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
950 	rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
951 	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
952 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT));
953 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP) != 1);
954 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
955 		channel->channel);
956 
957 	rx_queue = efx_channel_get_rx_queue(channel);
958 
959 	rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
960 	expected_ptr = rx_queue->removed_count & rx_queue->ptr_mask;
961 	if (unlikely(rx_ev_desc_ptr != expected_ptr))
962 		efx_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr);
963 
964 	if (likely(rx_ev_pkt_ok)) {
965 		/* If packet is marked as OK and packet type is TCP/IP or
966 		 * UDP/IP, then we can rely on the hardware checksum.
967 		 */
968 		flags = (rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP ||
969 			 rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP) ?
970 			EFX_RX_PKT_CSUMMED : 0;
971 	} else {
972 		flags = efx_handle_rx_not_ok(rx_queue, event);
973 	}
974 
975 	/* Detect multicast packets that didn't match the filter */
976 	rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
977 	if (rx_ev_mcast_pkt) {
978 		unsigned int rx_ev_mcast_hash_match =
979 			EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
980 
981 		if (unlikely(!rx_ev_mcast_hash_match)) {
982 			++channel->n_rx_mcast_mismatch;
983 			flags |= EFX_RX_PKT_DISCARD;
984 		}
985 	}
986 
987 	channel->irq_mod_score += 2;
988 
989 	/* Handle received packet */
990 	efx_rx_packet(rx_queue, rx_ev_desc_ptr, rx_ev_byte_cnt, flags);
991 }
992 
993 /* If this flush done event corresponds to a &struct efx_tx_queue, then
994  * send an %EFX_CHANNEL_MAGIC_TX_DRAIN event to drain the event queue
995  * of all transmit completions.
996  */
997 static void
998 efx_handle_tx_flush_done(struct efx_nic *efx, efx_qword_t *event)
999 {
1000 	struct efx_tx_queue *tx_queue;
1001 	int qid;
1002 
1003 	qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1004 	if (qid < EFX_TXQ_TYPES * efx->n_tx_channels) {
1005 		tx_queue = efx_get_tx_queue(efx, qid / EFX_TXQ_TYPES,
1006 					    qid % EFX_TXQ_TYPES);
1007 
1008 		efx_magic_event(tx_queue->channel,
1009 				EFX_CHANNEL_MAGIC_TX_DRAIN(tx_queue));
1010 	}
1011 }
1012 
1013 /* If this flush done event corresponds to a &struct efx_rx_queue: If the flush
1014  * was succesful then send an %EFX_CHANNEL_MAGIC_RX_DRAIN, otherwise add
1015  * the RX queue back to the mask of RX queues in need of flushing.
1016  */
1017 static void
1018 efx_handle_rx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1019 {
1020 	struct efx_channel *channel;
1021 	struct efx_rx_queue *rx_queue;
1022 	int qid;
1023 	bool failed;
1024 
1025 	qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
1026 	failed = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
1027 	if (qid >= efx->n_channels)
1028 		return;
1029 	channel = efx_get_channel(efx, qid);
1030 	if (!efx_channel_has_rx_queue(channel))
1031 		return;
1032 	rx_queue = efx_channel_get_rx_queue(channel);
1033 
1034 	if (failed) {
1035 		netif_info(efx, hw, efx->net_dev,
1036 			   "RXQ %d flush retry\n", qid);
1037 		rx_queue->flush_pending = true;
1038 		atomic_inc(&efx->rxq_flush_pending);
1039 	} else {
1040 		efx_magic_event(efx_rx_queue_channel(rx_queue),
1041 				EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue));
1042 	}
1043 	atomic_dec(&efx->rxq_flush_outstanding);
1044 	if (efx_flush_wake(efx))
1045 		wake_up(&efx->flush_wq);
1046 }
1047 
1048 static void
1049 efx_handle_drain_event(struct efx_channel *channel)
1050 {
1051 	struct efx_nic *efx = channel->efx;
1052 
1053 	WARN_ON(atomic_read(&efx->drain_pending) == 0);
1054 	atomic_dec(&efx->drain_pending);
1055 	if (efx_flush_wake(efx))
1056 		wake_up(&efx->flush_wq);
1057 }
1058 
1059 static void
1060 efx_handle_generated_event(struct efx_channel *channel, efx_qword_t *event)
1061 {
1062 	struct efx_nic *efx = channel->efx;
1063 	struct efx_rx_queue *rx_queue =
1064 		efx_channel_has_rx_queue(channel) ?
1065 		efx_channel_get_rx_queue(channel) : NULL;
1066 	unsigned magic, code;
1067 
1068 	magic = EFX_QWORD_FIELD(*event, FSF_AZ_DRV_GEN_EV_MAGIC);
1069 	code = _EFX_CHANNEL_MAGIC_CODE(magic);
1070 
1071 	if (magic == EFX_CHANNEL_MAGIC_TEST(channel)) {
1072 		/* ignore */
1073 	} else if (rx_queue && magic == EFX_CHANNEL_MAGIC_FILL(rx_queue)) {
1074 		/* The queue must be empty, so we won't receive any rx
1075 		 * events, so efx_process_channel() won't refill the
1076 		 * queue. Refill it here */
1077 		efx_fast_push_rx_descriptors(rx_queue);
1078 	} else if (rx_queue && magic == EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue)) {
1079 		rx_queue->enabled = false;
1080 		efx_handle_drain_event(channel);
1081 	} else if (code == _EFX_CHANNEL_MAGIC_TX_DRAIN) {
1082 		efx_handle_drain_event(channel);
1083 	} else {
1084 		netif_dbg(efx, hw, efx->net_dev, "channel %d received "
1085 			  "generated event "EFX_QWORD_FMT"\n",
1086 			  channel->channel, EFX_QWORD_VAL(*event));
1087 	}
1088 }
1089 
1090 static void
1091 efx_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
1092 {
1093 	struct efx_nic *efx = channel->efx;
1094 	unsigned int ev_sub_code;
1095 	unsigned int ev_sub_data;
1096 
1097 	ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
1098 	ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1099 
1100 	switch (ev_sub_code) {
1101 	case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
1102 		netif_vdbg(efx, hw, efx->net_dev, "channel %d TXQ %d flushed\n",
1103 			   channel->channel, ev_sub_data);
1104 		efx_handle_tx_flush_done(efx, event);
1105 		break;
1106 	case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
1107 		netif_vdbg(efx, hw, efx->net_dev, "channel %d RXQ %d flushed\n",
1108 			   channel->channel, ev_sub_data);
1109 		efx_handle_rx_flush_done(efx, event);
1110 		break;
1111 	case FSE_AZ_EVQ_INIT_DONE_EV:
1112 		netif_dbg(efx, hw, efx->net_dev,
1113 			  "channel %d EVQ %d initialised\n",
1114 			  channel->channel, ev_sub_data);
1115 		break;
1116 	case FSE_AZ_SRM_UPD_DONE_EV:
1117 		netif_vdbg(efx, hw, efx->net_dev,
1118 			   "channel %d SRAM update done\n", channel->channel);
1119 		break;
1120 	case FSE_AZ_WAKE_UP_EV:
1121 		netif_vdbg(efx, hw, efx->net_dev,
1122 			   "channel %d RXQ %d wakeup event\n",
1123 			   channel->channel, ev_sub_data);
1124 		break;
1125 	case FSE_AZ_TIMER_EV:
1126 		netif_vdbg(efx, hw, efx->net_dev,
1127 			   "channel %d RX queue %d timer expired\n",
1128 			   channel->channel, ev_sub_data);
1129 		break;
1130 	case FSE_AA_RX_RECOVER_EV:
1131 		netif_err(efx, rx_err, efx->net_dev,
1132 			  "channel %d seen DRIVER RX_RESET event. "
1133 			"Resetting.\n", channel->channel);
1134 		atomic_inc(&efx->rx_reset);
1135 		efx_schedule_reset(efx,
1136 				   EFX_WORKAROUND_6555(efx) ?
1137 				   RESET_TYPE_RX_RECOVERY :
1138 				   RESET_TYPE_DISABLE);
1139 		break;
1140 	case FSE_BZ_RX_DSC_ERROR_EV:
1141 		netif_err(efx, rx_err, efx->net_dev,
1142 			  "RX DMA Q %d reports descriptor fetch error."
1143 			  " RX Q %d is disabled.\n", ev_sub_data, ev_sub_data);
1144 		efx_schedule_reset(efx, RESET_TYPE_RX_DESC_FETCH);
1145 		break;
1146 	case FSE_BZ_TX_DSC_ERROR_EV:
1147 		netif_err(efx, tx_err, efx->net_dev,
1148 			  "TX DMA Q %d reports descriptor fetch error."
1149 			  " TX Q %d is disabled.\n", ev_sub_data, ev_sub_data);
1150 		efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
1151 		break;
1152 	default:
1153 		netif_vdbg(efx, hw, efx->net_dev,
1154 			   "channel %d unknown driver event code %d "
1155 			   "data %04x\n", channel->channel, ev_sub_code,
1156 			   ev_sub_data);
1157 		break;
1158 	}
1159 }
1160 
1161 int efx_nic_process_eventq(struct efx_channel *channel, int budget)
1162 {
1163 	struct efx_nic *efx = channel->efx;
1164 	unsigned int read_ptr;
1165 	efx_qword_t event, *p_event;
1166 	int ev_code;
1167 	int tx_packets = 0;
1168 	int spent = 0;
1169 
1170 	read_ptr = channel->eventq_read_ptr;
1171 
1172 	for (;;) {
1173 		p_event = efx_event(channel, read_ptr);
1174 		event = *p_event;
1175 
1176 		if (!efx_event_present(&event))
1177 			/* End of events */
1178 			break;
1179 
1180 		netif_vdbg(channel->efx, intr, channel->efx->net_dev,
1181 			   "channel %d event is "EFX_QWORD_FMT"\n",
1182 			   channel->channel, EFX_QWORD_VAL(event));
1183 
1184 		/* Clear this event by marking it all ones */
1185 		EFX_SET_QWORD(*p_event);
1186 
1187 		++read_ptr;
1188 
1189 		ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
1190 
1191 		switch (ev_code) {
1192 		case FSE_AZ_EV_CODE_RX_EV:
1193 			efx_handle_rx_event(channel, &event);
1194 			if (++spent == budget)
1195 				goto out;
1196 			break;
1197 		case FSE_AZ_EV_CODE_TX_EV:
1198 			tx_packets += efx_handle_tx_event(channel, &event);
1199 			if (tx_packets > efx->txq_entries) {
1200 				spent = budget;
1201 				goto out;
1202 			}
1203 			break;
1204 		case FSE_AZ_EV_CODE_DRV_GEN_EV:
1205 			efx_handle_generated_event(channel, &event);
1206 			break;
1207 		case FSE_AZ_EV_CODE_DRIVER_EV:
1208 			efx_handle_driver_event(channel, &event);
1209 			break;
1210 		case FSE_CZ_EV_CODE_MCDI_EV:
1211 			efx_mcdi_process_event(channel, &event);
1212 			break;
1213 		case FSE_AZ_EV_CODE_GLOBAL_EV:
1214 			if (efx->type->handle_global_event &&
1215 			    efx->type->handle_global_event(channel, &event))
1216 				break;
1217 			/* else fall through */
1218 		default:
1219 			netif_err(channel->efx, hw, channel->efx->net_dev,
1220 				  "channel %d unknown event type %d (data "
1221 				  EFX_QWORD_FMT ")\n", channel->channel,
1222 				  ev_code, EFX_QWORD_VAL(event));
1223 		}
1224 	}
1225 
1226 out:
1227 	channel->eventq_read_ptr = read_ptr;
1228 	return spent;
1229 }
1230 
1231 /* Check whether an event is present in the eventq at the current
1232  * read pointer.  Only useful for self-test.
1233  */
1234 bool efx_nic_event_present(struct efx_channel *channel)
1235 {
1236 	return efx_event_present(efx_event(channel, channel->eventq_read_ptr));
1237 }
1238 
1239 /* Allocate buffer table entries for event queue */
1240 int efx_nic_probe_eventq(struct efx_channel *channel)
1241 {
1242 	struct efx_nic *efx = channel->efx;
1243 	unsigned entries;
1244 
1245 	entries = channel->eventq_mask + 1;
1246 	return efx_alloc_special_buffer(efx, &channel->eventq,
1247 					entries * sizeof(efx_qword_t));
1248 }
1249 
1250 void efx_nic_init_eventq(struct efx_channel *channel)
1251 {
1252 	efx_oword_t reg;
1253 	struct efx_nic *efx = channel->efx;
1254 
1255 	netif_dbg(efx, hw, efx->net_dev,
1256 		  "channel %d event queue in special buffers %d-%d\n",
1257 		  channel->channel, channel->eventq.index,
1258 		  channel->eventq.index + channel->eventq.entries - 1);
1259 
1260 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1261 		EFX_POPULATE_OWORD_3(reg,
1262 				     FRF_CZ_TIMER_Q_EN, 1,
1263 				     FRF_CZ_HOST_NOTIFY_MODE, 0,
1264 				     FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
1265 		efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1266 	}
1267 
1268 	/* Pin event queue buffer */
1269 	efx_init_special_buffer(efx, &channel->eventq);
1270 
1271 	/* Fill event queue with all ones (i.e. empty events) */
1272 	memset(channel->eventq.addr, 0xff, channel->eventq.len);
1273 
1274 	/* Push event queue to card */
1275 	EFX_POPULATE_OWORD_3(reg,
1276 			     FRF_AZ_EVQ_EN, 1,
1277 			     FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
1278 			     FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
1279 	efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1280 			 channel->channel);
1281 
1282 	efx->type->push_irq_moderation(channel);
1283 }
1284 
1285 void efx_nic_fini_eventq(struct efx_channel *channel)
1286 {
1287 	efx_oword_t reg;
1288 	struct efx_nic *efx = channel->efx;
1289 
1290 	/* Remove event queue from card */
1291 	EFX_ZERO_OWORD(reg);
1292 	efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1293 			 channel->channel);
1294 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1295 		efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1296 
1297 	/* Unpin event queue */
1298 	efx_fini_special_buffer(efx, &channel->eventq);
1299 }
1300 
1301 /* Free buffers backing event queue */
1302 void efx_nic_remove_eventq(struct efx_channel *channel)
1303 {
1304 	efx_free_special_buffer(channel->efx, &channel->eventq);
1305 }
1306 
1307 
1308 void efx_nic_generate_test_event(struct efx_channel *channel)
1309 {
1310 	efx_magic_event(channel, EFX_CHANNEL_MAGIC_TEST(channel));
1311 }
1312 
1313 void efx_nic_generate_fill_event(struct efx_rx_queue *rx_queue)
1314 {
1315 	efx_magic_event(efx_rx_queue_channel(rx_queue),
1316 			EFX_CHANNEL_MAGIC_FILL(rx_queue));
1317 }
1318 
1319 /**************************************************************************
1320  *
1321  * Hardware interrupts
1322  * The hardware interrupt handler does very little work; all the event
1323  * queue processing is carried out by per-channel tasklets.
1324  *
1325  **************************************************************************/
1326 
1327 /* Enable/disable/generate interrupts */
1328 static inline void efx_nic_interrupts(struct efx_nic *efx,
1329 				      bool enabled, bool force)
1330 {
1331 	efx_oword_t int_en_reg_ker;
1332 
1333 	EFX_POPULATE_OWORD_3(int_en_reg_ker,
1334 			     FRF_AZ_KER_INT_LEVE_SEL, efx->irq_level,
1335 			     FRF_AZ_KER_INT_KER, force,
1336 			     FRF_AZ_DRV_INT_EN_KER, enabled);
1337 	efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
1338 }
1339 
1340 void efx_nic_enable_interrupts(struct efx_nic *efx)
1341 {
1342 	EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
1343 	wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
1344 
1345 	efx_nic_interrupts(efx, true, false);
1346 }
1347 
1348 void efx_nic_disable_interrupts(struct efx_nic *efx)
1349 {
1350 	/* Disable interrupts */
1351 	efx_nic_interrupts(efx, false, false);
1352 }
1353 
1354 /* Generate a test interrupt
1355  * Interrupt must already have been enabled, otherwise nasty things
1356  * may happen.
1357  */
1358 void efx_nic_generate_interrupt(struct efx_nic *efx)
1359 {
1360 	efx_nic_interrupts(efx, true, true);
1361 }
1362 
1363 /* Process a fatal interrupt
1364  * Disable bus mastering ASAP and schedule a reset
1365  */
1366 irqreturn_t efx_nic_fatal_interrupt(struct efx_nic *efx)
1367 {
1368 	struct falcon_nic_data *nic_data = efx->nic_data;
1369 	efx_oword_t *int_ker = efx->irq_status.addr;
1370 	efx_oword_t fatal_intr;
1371 	int error, mem_perr;
1372 
1373 	efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
1374 	error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
1375 
1376 	netif_err(efx, hw, efx->net_dev, "SYSTEM ERROR "EFX_OWORD_FMT" status "
1377 		  EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
1378 		  EFX_OWORD_VAL(fatal_intr),
1379 		  error ? "disabling bus mastering" : "no recognised error");
1380 
1381 	/* If this is a memory parity error dump which blocks are offending */
1382 	mem_perr = (EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER) ||
1383 		    EFX_OWORD_FIELD(fatal_intr, FRF_AZ_SRM_PERR_INT_KER));
1384 	if (mem_perr) {
1385 		efx_oword_t reg;
1386 		efx_reado(efx, &reg, FR_AZ_MEM_STAT);
1387 		netif_err(efx, hw, efx->net_dev,
1388 			  "SYSTEM ERROR: memory parity error "EFX_OWORD_FMT"\n",
1389 			  EFX_OWORD_VAL(reg));
1390 	}
1391 
1392 	/* Disable both devices */
1393 	pci_clear_master(efx->pci_dev);
1394 	if (efx_nic_is_dual_func(efx))
1395 		pci_clear_master(nic_data->pci_dev2);
1396 	efx_nic_disable_interrupts(efx);
1397 
1398 	/* Count errors and reset or disable the NIC accordingly */
1399 	if (efx->int_error_count == 0 ||
1400 	    time_after(jiffies, efx->int_error_expire)) {
1401 		efx->int_error_count = 0;
1402 		efx->int_error_expire =
1403 			jiffies + EFX_INT_ERROR_EXPIRE * HZ;
1404 	}
1405 	if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
1406 		netif_err(efx, hw, efx->net_dev,
1407 			  "SYSTEM ERROR - reset scheduled\n");
1408 		efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
1409 	} else {
1410 		netif_err(efx, hw, efx->net_dev,
1411 			  "SYSTEM ERROR - max number of errors seen."
1412 			  "NIC will be disabled\n");
1413 		efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1414 	}
1415 
1416 	return IRQ_HANDLED;
1417 }
1418 
1419 /* Handle a legacy interrupt
1420  * Acknowledges the interrupt and schedule event queue processing.
1421  */
1422 static irqreturn_t efx_legacy_interrupt(int irq, void *dev_id)
1423 {
1424 	struct efx_nic *efx = dev_id;
1425 	efx_oword_t *int_ker = efx->irq_status.addr;
1426 	irqreturn_t result = IRQ_NONE;
1427 	struct efx_channel *channel;
1428 	efx_dword_t reg;
1429 	u32 queues;
1430 	int syserr;
1431 
1432 	/* Could this be ours?  If interrupts are disabled then the
1433 	 * channel state may not be valid.
1434 	 */
1435 	if (!efx->legacy_irq_enabled)
1436 		return result;
1437 
1438 	/* Read the ISR which also ACKs the interrupts */
1439 	efx_readd(efx, &reg, FR_BZ_INT_ISR0);
1440 	queues = EFX_EXTRACT_DWORD(reg, 0, 31);
1441 
1442 	/* Handle non-event-queue sources */
1443 	if (queues & (1U << efx->irq_level)) {
1444 		syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1445 		if (unlikely(syserr))
1446 			return efx_nic_fatal_interrupt(efx);
1447 		efx->last_irq_cpu = raw_smp_processor_id();
1448 	}
1449 
1450 	if (queues != 0) {
1451 		if (EFX_WORKAROUND_15783(efx))
1452 			efx->irq_zero_count = 0;
1453 
1454 		/* Schedule processing of any interrupting queues */
1455 		efx_for_each_channel(channel, efx) {
1456 			if (queues & 1)
1457 				efx_schedule_channel_irq(channel);
1458 			queues >>= 1;
1459 		}
1460 		result = IRQ_HANDLED;
1461 
1462 	} else if (EFX_WORKAROUND_15783(efx)) {
1463 		efx_qword_t *event;
1464 
1465 		/* We can't return IRQ_HANDLED more than once on seeing ISR=0
1466 		 * because this might be a shared interrupt. */
1467 		if (efx->irq_zero_count++ == 0)
1468 			result = IRQ_HANDLED;
1469 
1470 		/* Ensure we schedule or rearm all event queues */
1471 		efx_for_each_channel(channel, efx) {
1472 			event = efx_event(channel, channel->eventq_read_ptr);
1473 			if (efx_event_present(event))
1474 				efx_schedule_channel_irq(channel);
1475 			else
1476 				efx_nic_eventq_read_ack(channel);
1477 		}
1478 	}
1479 
1480 	if (result == IRQ_HANDLED)
1481 		netif_vdbg(efx, intr, efx->net_dev,
1482 			   "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
1483 			   irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
1484 
1485 	return result;
1486 }
1487 
1488 /* Handle an MSI interrupt
1489  *
1490  * Handle an MSI hardware interrupt.  This routine schedules event
1491  * queue processing.  No interrupt acknowledgement cycle is necessary.
1492  * Also, we never need to check that the interrupt is for us, since
1493  * MSI interrupts cannot be shared.
1494  */
1495 static irqreturn_t efx_msi_interrupt(int irq, void *dev_id)
1496 {
1497 	struct efx_channel *channel = *(struct efx_channel **)dev_id;
1498 	struct efx_nic *efx = channel->efx;
1499 	efx_oword_t *int_ker = efx->irq_status.addr;
1500 	int syserr;
1501 
1502 	netif_vdbg(efx, intr, efx->net_dev,
1503 		   "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
1504 		   irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
1505 
1506 	/* Handle non-event-queue sources */
1507 	if (channel->channel == efx->irq_level) {
1508 		syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1509 		if (unlikely(syserr))
1510 			return efx_nic_fatal_interrupt(efx);
1511 		efx->last_irq_cpu = raw_smp_processor_id();
1512 	}
1513 
1514 	/* Schedule processing of the channel */
1515 	efx_schedule_channel_irq(channel);
1516 
1517 	return IRQ_HANDLED;
1518 }
1519 
1520 
1521 /* Setup RSS indirection table.
1522  * This maps from the hash value of the packet to RXQ
1523  */
1524 void efx_nic_push_rx_indir_table(struct efx_nic *efx)
1525 {
1526 	size_t i = 0;
1527 	efx_dword_t dword;
1528 
1529 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0)
1530 		return;
1531 
1532 	BUILD_BUG_ON(ARRAY_SIZE(efx->rx_indir_table) !=
1533 		     FR_BZ_RX_INDIRECTION_TBL_ROWS);
1534 
1535 	for (i = 0; i < FR_BZ_RX_INDIRECTION_TBL_ROWS; i++) {
1536 		EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
1537 				     efx->rx_indir_table[i]);
1538 		efx_writed_table(efx, &dword, FR_BZ_RX_INDIRECTION_TBL, i);
1539 	}
1540 }
1541 
1542 /* Hook interrupt handler(s)
1543  * Try MSI and then legacy interrupts.
1544  */
1545 int efx_nic_init_interrupt(struct efx_nic *efx)
1546 {
1547 	struct efx_channel *channel;
1548 	int rc;
1549 
1550 	if (!EFX_INT_MODE_USE_MSI(efx)) {
1551 		irq_handler_t handler;
1552 		if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1553 			handler = efx_legacy_interrupt;
1554 		else
1555 			handler = falcon_legacy_interrupt_a1;
1556 
1557 		rc = request_irq(efx->legacy_irq, handler, IRQF_SHARED,
1558 				 efx->name, efx);
1559 		if (rc) {
1560 			netif_err(efx, drv, efx->net_dev,
1561 				  "failed to hook legacy IRQ %d\n",
1562 				  efx->pci_dev->irq);
1563 			goto fail1;
1564 		}
1565 		return 0;
1566 	}
1567 
1568 	/* Hook MSI or MSI-X interrupt */
1569 	efx_for_each_channel(channel, efx) {
1570 		rc = request_irq(channel->irq, efx_msi_interrupt,
1571 				 IRQF_PROBE_SHARED, /* Not shared */
1572 				 efx->channel_name[channel->channel],
1573 				 &efx->channel[channel->channel]);
1574 		if (rc) {
1575 			netif_err(efx, drv, efx->net_dev,
1576 				  "failed to hook IRQ %d\n", channel->irq);
1577 			goto fail2;
1578 		}
1579 	}
1580 
1581 	return 0;
1582 
1583  fail2:
1584 	efx_for_each_channel(channel, efx)
1585 		free_irq(channel->irq, &efx->channel[channel->channel]);
1586  fail1:
1587 	return rc;
1588 }
1589 
1590 void efx_nic_fini_interrupt(struct efx_nic *efx)
1591 {
1592 	struct efx_channel *channel;
1593 	efx_oword_t reg;
1594 
1595 	/* Disable MSI/MSI-X interrupts */
1596 	efx_for_each_channel(channel, efx) {
1597 		if (channel->irq)
1598 			free_irq(channel->irq, &efx->channel[channel->channel]);
1599 	}
1600 
1601 	/* ACK legacy interrupt */
1602 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1603 		efx_reado(efx, &reg, FR_BZ_INT_ISR0);
1604 	else
1605 		falcon_irq_ack_a1(efx);
1606 
1607 	/* Disable legacy interrupt */
1608 	if (efx->legacy_irq)
1609 		free_irq(efx->legacy_irq, efx);
1610 }
1611 
1612 u32 efx_nic_fpga_ver(struct efx_nic *efx)
1613 {
1614 	efx_oword_t altera_build;
1615 	efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
1616 	return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
1617 }
1618 
1619 void efx_nic_init_common(struct efx_nic *efx)
1620 {
1621 	efx_oword_t temp;
1622 
1623 	/* Set positions of descriptor caches in SRAM. */
1624 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR,
1625 			     efx->type->tx_dc_base / 8);
1626 	efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
1627 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR,
1628 			     efx->type->rx_dc_base / 8);
1629 	efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
1630 
1631 	/* Set TX descriptor cache size. */
1632 	BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
1633 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
1634 	efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
1635 
1636 	/* Set RX descriptor cache size.  Set low watermark to size-8, as
1637 	 * this allows most efficient prefetching.
1638 	 */
1639 	BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
1640 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
1641 	efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
1642 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
1643 	efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
1644 
1645 	/* Program INT_KER address */
1646 	EFX_POPULATE_OWORD_2(temp,
1647 			     FRF_AZ_NORM_INT_VEC_DIS_KER,
1648 			     EFX_INT_MODE_USE_MSI(efx),
1649 			     FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
1650 	efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
1651 
1652 	if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
1653 		/* Use an interrupt level unused by event queues */
1654 		efx->irq_level = 0x1f;
1655 	else
1656 		/* Use a valid MSI-X vector */
1657 		efx->irq_level = 0;
1658 
1659 	/* Enable all the genuinely fatal interrupts.  (They are still
1660 	 * masked by the overall interrupt mask, controlled by
1661 	 * falcon_interrupts()).
1662 	 *
1663 	 * Note: All other fatal interrupts are enabled
1664 	 */
1665 	EFX_POPULATE_OWORD_3(temp,
1666 			     FRF_AZ_ILL_ADR_INT_KER_EN, 1,
1667 			     FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
1668 			     FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
1669 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1670 		EFX_SET_OWORD_FIELD(temp, FRF_CZ_SRAM_PERR_INT_P_KER_EN, 1);
1671 	EFX_INVERT_OWORD(temp);
1672 	efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
1673 
1674 	efx_nic_push_rx_indir_table(efx);
1675 
1676 	/* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
1677 	 * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
1678 	 */
1679 	efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
1680 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
1681 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
1682 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
1683 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 1);
1684 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
1685 	/* Enable SW_EV to inherit in char driver - assume harmless here */
1686 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
1687 	/* Prefetch threshold 2 => fetch when descriptor cache half empty */
1688 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
1689 	/* Disable hardware watchdog which can misfire */
1690 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
1691 	/* Squash TX of packets of 16 bytes or less */
1692 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1693 		EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
1694 	efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
1695 
1696 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
1697 		EFX_POPULATE_OWORD_4(temp,
1698 				     /* Default values */
1699 				     FRF_BZ_TX_PACE_SB_NOT_AF, 0x15,
1700 				     FRF_BZ_TX_PACE_SB_AF, 0xb,
1701 				     FRF_BZ_TX_PACE_FB_BASE, 0,
1702 				     /* Allow large pace values in the
1703 				      * fast bin. */
1704 				     FRF_BZ_TX_PACE_BIN_TH,
1705 				     FFE_BZ_TX_PACE_RESERVED);
1706 		efx_writeo(efx, &temp, FR_BZ_TX_PACE);
1707 	}
1708 }
1709 
1710 /* Register dump */
1711 
1712 #define REGISTER_REVISION_A	1
1713 #define REGISTER_REVISION_B	2
1714 #define REGISTER_REVISION_C	3
1715 #define REGISTER_REVISION_Z	3	/* latest revision */
1716 
1717 struct efx_nic_reg {
1718 	u32 offset:24;
1719 	u32 min_revision:2, max_revision:2;
1720 };
1721 
1722 #define REGISTER(name, min_rev, max_rev) {				\
1723 	FR_ ## min_rev ## max_rev ## _ ## name,				\
1724 	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev	\
1725 }
1726 #define REGISTER_AA(name) REGISTER(name, A, A)
1727 #define REGISTER_AB(name) REGISTER(name, A, B)
1728 #define REGISTER_AZ(name) REGISTER(name, A, Z)
1729 #define REGISTER_BB(name) REGISTER(name, B, B)
1730 #define REGISTER_BZ(name) REGISTER(name, B, Z)
1731 #define REGISTER_CZ(name) REGISTER(name, C, Z)
1732 
1733 static const struct efx_nic_reg efx_nic_regs[] = {
1734 	REGISTER_AZ(ADR_REGION),
1735 	REGISTER_AZ(INT_EN_KER),
1736 	REGISTER_BZ(INT_EN_CHAR),
1737 	REGISTER_AZ(INT_ADR_KER),
1738 	REGISTER_BZ(INT_ADR_CHAR),
1739 	/* INT_ACK_KER is WO */
1740 	/* INT_ISR0 is RC */
1741 	REGISTER_AZ(HW_INIT),
1742 	REGISTER_CZ(USR_EV_CFG),
1743 	REGISTER_AB(EE_SPI_HCMD),
1744 	REGISTER_AB(EE_SPI_HADR),
1745 	REGISTER_AB(EE_SPI_HDATA),
1746 	REGISTER_AB(EE_BASE_PAGE),
1747 	REGISTER_AB(EE_VPD_CFG0),
1748 	/* EE_VPD_SW_CNTL and EE_VPD_SW_DATA are not used */
1749 	/* PMBX_DBG_IADDR and PBMX_DBG_IDATA are indirect */
1750 	/* PCIE_CORE_INDIRECT is indirect */
1751 	REGISTER_AB(NIC_STAT),
1752 	REGISTER_AB(GPIO_CTL),
1753 	REGISTER_AB(GLB_CTL),
1754 	/* FATAL_INTR_KER and FATAL_INTR_CHAR are partly RC */
1755 	REGISTER_BZ(DP_CTRL),
1756 	REGISTER_AZ(MEM_STAT),
1757 	REGISTER_AZ(CS_DEBUG),
1758 	REGISTER_AZ(ALTERA_BUILD),
1759 	REGISTER_AZ(CSR_SPARE),
1760 	REGISTER_AB(PCIE_SD_CTL0123),
1761 	REGISTER_AB(PCIE_SD_CTL45),
1762 	REGISTER_AB(PCIE_PCS_CTL_STAT),
1763 	/* DEBUG_DATA_OUT is not used */
1764 	/* DRV_EV is WO */
1765 	REGISTER_AZ(EVQ_CTL),
1766 	REGISTER_AZ(EVQ_CNT1),
1767 	REGISTER_AZ(EVQ_CNT2),
1768 	REGISTER_AZ(BUF_TBL_CFG),
1769 	REGISTER_AZ(SRM_RX_DC_CFG),
1770 	REGISTER_AZ(SRM_TX_DC_CFG),
1771 	REGISTER_AZ(SRM_CFG),
1772 	/* BUF_TBL_UPD is WO */
1773 	REGISTER_AZ(SRM_UPD_EVQ),
1774 	REGISTER_AZ(SRAM_PARITY),
1775 	REGISTER_AZ(RX_CFG),
1776 	REGISTER_BZ(RX_FILTER_CTL),
1777 	/* RX_FLUSH_DESCQ is WO */
1778 	REGISTER_AZ(RX_DC_CFG),
1779 	REGISTER_AZ(RX_DC_PF_WM),
1780 	REGISTER_BZ(RX_RSS_TKEY),
1781 	/* RX_NODESC_DROP is RC */
1782 	REGISTER_AA(RX_SELF_RST),
1783 	/* RX_DEBUG, RX_PUSH_DROP are not used */
1784 	REGISTER_CZ(RX_RSS_IPV6_REG1),
1785 	REGISTER_CZ(RX_RSS_IPV6_REG2),
1786 	REGISTER_CZ(RX_RSS_IPV6_REG3),
1787 	/* TX_FLUSH_DESCQ is WO */
1788 	REGISTER_AZ(TX_DC_CFG),
1789 	REGISTER_AA(TX_CHKSM_CFG),
1790 	REGISTER_AZ(TX_CFG),
1791 	/* TX_PUSH_DROP is not used */
1792 	REGISTER_AZ(TX_RESERVED),
1793 	REGISTER_BZ(TX_PACE),
1794 	/* TX_PACE_DROP_QID is RC */
1795 	REGISTER_BB(TX_VLAN),
1796 	REGISTER_BZ(TX_IPFIL_PORTEN),
1797 	REGISTER_AB(MD_TXD),
1798 	REGISTER_AB(MD_RXD),
1799 	REGISTER_AB(MD_CS),
1800 	REGISTER_AB(MD_PHY_ADR),
1801 	REGISTER_AB(MD_ID),
1802 	/* MD_STAT is RC */
1803 	REGISTER_AB(MAC_STAT_DMA),
1804 	REGISTER_AB(MAC_CTRL),
1805 	REGISTER_BB(GEN_MODE),
1806 	REGISTER_AB(MAC_MC_HASH_REG0),
1807 	REGISTER_AB(MAC_MC_HASH_REG1),
1808 	REGISTER_AB(GM_CFG1),
1809 	REGISTER_AB(GM_CFG2),
1810 	/* GM_IPG and GM_HD are not used */
1811 	REGISTER_AB(GM_MAX_FLEN),
1812 	/* GM_TEST is not used */
1813 	REGISTER_AB(GM_ADR1),
1814 	REGISTER_AB(GM_ADR2),
1815 	REGISTER_AB(GMF_CFG0),
1816 	REGISTER_AB(GMF_CFG1),
1817 	REGISTER_AB(GMF_CFG2),
1818 	REGISTER_AB(GMF_CFG3),
1819 	REGISTER_AB(GMF_CFG4),
1820 	REGISTER_AB(GMF_CFG5),
1821 	REGISTER_BB(TX_SRC_MAC_CTL),
1822 	REGISTER_AB(XM_ADR_LO),
1823 	REGISTER_AB(XM_ADR_HI),
1824 	REGISTER_AB(XM_GLB_CFG),
1825 	REGISTER_AB(XM_TX_CFG),
1826 	REGISTER_AB(XM_RX_CFG),
1827 	REGISTER_AB(XM_MGT_INT_MASK),
1828 	REGISTER_AB(XM_FC),
1829 	REGISTER_AB(XM_PAUSE_TIME),
1830 	REGISTER_AB(XM_TX_PARAM),
1831 	REGISTER_AB(XM_RX_PARAM),
1832 	/* XM_MGT_INT_MSK (note no 'A') is RC */
1833 	REGISTER_AB(XX_PWR_RST),
1834 	REGISTER_AB(XX_SD_CTL),
1835 	REGISTER_AB(XX_TXDRV_CTL),
1836 	/* XX_PRBS_CTL, XX_PRBS_CHK and XX_PRBS_ERR are not used */
1837 	/* XX_CORE_STAT is partly RC */
1838 };
1839 
1840 struct efx_nic_reg_table {
1841 	u32 offset:24;
1842 	u32 min_revision:2, max_revision:2;
1843 	u32 step:6, rows:21;
1844 };
1845 
1846 #define REGISTER_TABLE_DIMENSIONS(_, offset, min_rev, max_rev, step, rows) { \
1847 	offset,								\
1848 	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev,	\
1849 	step, rows							\
1850 }
1851 #define REGISTER_TABLE(name, min_rev, max_rev)				\
1852 	REGISTER_TABLE_DIMENSIONS(					\
1853 		name, FR_ ## min_rev ## max_rev ## _ ## name,		\
1854 		min_rev, max_rev,					\
1855 		FR_ ## min_rev ## max_rev ## _ ## name ## _STEP,	\
1856 		FR_ ## min_rev ## max_rev ## _ ## name ## _ROWS)
1857 #define REGISTER_TABLE_AA(name) REGISTER_TABLE(name, A, A)
1858 #define REGISTER_TABLE_AZ(name) REGISTER_TABLE(name, A, Z)
1859 #define REGISTER_TABLE_BB(name) REGISTER_TABLE(name, B, B)
1860 #define REGISTER_TABLE_BZ(name) REGISTER_TABLE(name, B, Z)
1861 #define REGISTER_TABLE_BB_CZ(name)					\
1862 	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, B, B,		\
1863 				  FR_BZ_ ## name ## _STEP,		\
1864 				  FR_BB_ ## name ## _ROWS),		\
1865 	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, C, Z,		\
1866 				  FR_BZ_ ## name ## _STEP,		\
1867 				  FR_CZ_ ## name ## _ROWS)
1868 #define REGISTER_TABLE_CZ(name) REGISTER_TABLE(name, C, Z)
1869 
1870 static const struct efx_nic_reg_table efx_nic_reg_tables[] = {
1871 	/* DRIVER is not used */
1872 	/* EVQ_RPTR, TIMER_COMMAND, USR_EV and {RX,TX}_DESC_UPD are WO */
1873 	REGISTER_TABLE_BB(TX_IPFIL_TBL),
1874 	REGISTER_TABLE_BB(TX_SRC_MAC_TBL),
1875 	REGISTER_TABLE_AA(RX_DESC_PTR_TBL_KER),
1876 	REGISTER_TABLE_BB_CZ(RX_DESC_PTR_TBL),
1877 	REGISTER_TABLE_AA(TX_DESC_PTR_TBL_KER),
1878 	REGISTER_TABLE_BB_CZ(TX_DESC_PTR_TBL),
1879 	REGISTER_TABLE_AA(EVQ_PTR_TBL_KER),
1880 	REGISTER_TABLE_BB_CZ(EVQ_PTR_TBL),
1881 	/* We can't reasonably read all of the buffer table (up to 8MB!).
1882 	 * However this driver will only use a few entries.  Reading
1883 	 * 1K entries allows for some expansion of queue count and
1884 	 * size before we need to change the version. */
1885 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL_KER, FR_AA_BUF_FULL_TBL_KER,
1886 				  A, A, 8, 1024),
1887 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL, FR_BZ_BUF_FULL_TBL,
1888 				  B, Z, 8, 1024),
1889 	REGISTER_TABLE_CZ(RX_MAC_FILTER_TBL0),
1890 	REGISTER_TABLE_BB_CZ(TIMER_TBL),
1891 	REGISTER_TABLE_BB_CZ(TX_PACE_TBL),
1892 	REGISTER_TABLE_BZ(RX_INDIRECTION_TBL),
1893 	/* TX_FILTER_TBL0 is huge and not used by this driver */
1894 	REGISTER_TABLE_CZ(TX_MAC_FILTER_TBL0),
1895 	REGISTER_TABLE_CZ(MC_TREG_SMEM),
1896 	/* MSIX_PBA_TABLE is not mapped */
1897 	/* SRM_DBG is not mapped (and is redundant with BUF_FLL_TBL) */
1898 	REGISTER_TABLE_BZ(RX_FILTER_TBL0),
1899 };
1900 
1901 size_t efx_nic_get_regs_len(struct efx_nic *efx)
1902 {
1903 	const struct efx_nic_reg *reg;
1904 	const struct efx_nic_reg_table *table;
1905 	size_t len = 0;
1906 
1907 	for (reg = efx_nic_regs;
1908 	     reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
1909 	     reg++)
1910 		if (efx->type->revision >= reg->min_revision &&
1911 		    efx->type->revision <= reg->max_revision)
1912 			len += sizeof(efx_oword_t);
1913 
1914 	for (table = efx_nic_reg_tables;
1915 	     table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
1916 	     table++)
1917 		if (efx->type->revision >= table->min_revision &&
1918 		    efx->type->revision <= table->max_revision)
1919 			len += table->rows * min_t(size_t, table->step, 16);
1920 
1921 	return len;
1922 }
1923 
1924 void efx_nic_get_regs(struct efx_nic *efx, void *buf)
1925 {
1926 	const struct efx_nic_reg *reg;
1927 	const struct efx_nic_reg_table *table;
1928 
1929 	for (reg = efx_nic_regs;
1930 	     reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
1931 	     reg++) {
1932 		if (efx->type->revision >= reg->min_revision &&
1933 		    efx->type->revision <= reg->max_revision) {
1934 			efx_reado(efx, (efx_oword_t *)buf, reg->offset);
1935 			buf += sizeof(efx_oword_t);
1936 		}
1937 	}
1938 
1939 	for (table = efx_nic_reg_tables;
1940 	     table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
1941 	     table++) {
1942 		size_t size, i;
1943 
1944 		if (!(efx->type->revision >= table->min_revision &&
1945 		      efx->type->revision <= table->max_revision))
1946 			continue;
1947 
1948 		size = min_t(size_t, table->step, 16);
1949 
1950 		for (i = 0; i < table->rows; i++) {
1951 			switch (table->step) {
1952 			case 4: /* 32-bit register or SRAM */
1953 				efx_readd_table(efx, buf, table->offset, i);
1954 				break;
1955 			case 8: /* 64-bit SRAM */
1956 				efx_sram_readq(efx,
1957 					       efx->membase + table->offset,
1958 					       buf, i);
1959 				break;
1960 			case 16: /* 128-bit register */
1961 				efx_reado_table(efx, buf, table->offset, i);
1962 				break;
1963 			case 32: /* 128-bit register, interleaved */
1964 				efx_reado_table(efx, buf, table->offset, 2 * i);
1965 				break;
1966 			default:
1967 				WARN_ON(1);
1968 				return;
1969 			}
1970 			buf += size;
1971 		}
1972 	}
1973 }
1974