xref: /openbmc/linux/drivers/net/ethernet/sfc/nic.c (revision 90893000)
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 	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->type->prepare_flush(efx);
681 
682 	efx_for_each_channel(channel, efx) {
683 		efx_for_each_channel_tx_queue(tx_queue, channel) {
684 			atomic_inc(&efx->drain_pending);
685 			efx_flush_tx_queue(tx_queue);
686 		}
687 		efx_for_each_channel_rx_queue(rx_queue, channel) {
688 			atomic_inc(&efx->drain_pending);
689 			rx_queue->flush_pending = true;
690 			atomic_inc(&efx->rxq_flush_pending);
691 		}
692 	}
693 
694 	while (timeout && atomic_read(&efx->drain_pending) > 0) {
695 		/* The hardware supports four concurrent rx flushes, each of
696 		 * which may need to be retried if there is an outstanding
697 		 * descriptor fetch
698 		 */
699 		efx_for_each_channel(channel, efx) {
700 			efx_for_each_channel_rx_queue(rx_queue, channel) {
701 				if (atomic_read(&efx->rxq_flush_outstanding) >=
702 				    EFX_RX_FLUSH_COUNT)
703 					break;
704 
705 				if (rx_queue->flush_pending) {
706 					rx_queue->flush_pending = false;
707 					atomic_dec(&efx->rxq_flush_pending);
708 					atomic_inc(&efx->rxq_flush_outstanding);
709 					efx_flush_rx_queue(rx_queue);
710 				}
711 			}
712 		}
713 
714 		timeout = wait_event_timeout(efx->flush_wq, efx_flush_wake(efx),
715 					     timeout);
716 	}
717 
718 	if (atomic_read(&efx->drain_pending)) {
719 		netif_err(efx, hw, efx->net_dev, "failed to flush %d queues "
720 			  "(rx %d+%d)\n", atomic_read(&efx->drain_pending),
721 			  atomic_read(&efx->rxq_flush_outstanding),
722 			  atomic_read(&efx->rxq_flush_pending));
723 		rc = -ETIMEDOUT;
724 
725 		atomic_set(&efx->drain_pending, 0);
726 		atomic_set(&efx->rxq_flush_pending, 0);
727 		atomic_set(&efx->rxq_flush_outstanding, 0);
728 	}
729 
730 	return rc;
731 }
732 
733 /**************************************************************************
734  *
735  * Event queue processing
736  * Event queues are processed by per-channel tasklets.
737  *
738  **************************************************************************/
739 
740 /* Update a channel's event queue's read pointer (RPTR) register
741  *
742  * This writes the EVQ_RPTR_REG register for the specified channel's
743  * event queue.
744  */
745 void efx_nic_eventq_read_ack(struct efx_channel *channel)
746 {
747 	efx_dword_t reg;
748 	struct efx_nic *efx = channel->efx;
749 
750 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_EVQ_RPTR,
751 			     channel->eventq_read_ptr & channel->eventq_mask);
752 	efx_writed_table(efx, &reg, efx->type->evq_rptr_tbl_base,
753 			 channel->channel);
754 }
755 
756 /* Use HW to insert a SW defined event */
757 void efx_generate_event(struct efx_nic *efx, unsigned int evq,
758 			efx_qword_t *event)
759 {
760 	efx_oword_t drv_ev_reg;
761 
762 	BUILD_BUG_ON(FRF_AZ_DRV_EV_DATA_LBN != 0 ||
763 		     FRF_AZ_DRV_EV_DATA_WIDTH != 64);
764 	drv_ev_reg.u32[0] = event->u32[0];
765 	drv_ev_reg.u32[1] = event->u32[1];
766 	drv_ev_reg.u32[2] = 0;
767 	drv_ev_reg.u32[3] = 0;
768 	EFX_SET_OWORD_FIELD(drv_ev_reg, FRF_AZ_DRV_EV_QID, evq);
769 	efx_writeo(efx, &drv_ev_reg, FR_AZ_DRV_EV);
770 }
771 
772 static void efx_magic_event(struct efx_channel *channel, u32 magic)
773 {
774 	efx_qword_t event;
775 
776 	EFX_POPULATE_QWORD_2(event, FSF_AZ_EV_CODE,
777 			     FSE_AZ_EV_CODE_DRV_GEN_EV,
778 			     FSF_AZ_DRV_GEN_EV_MAGIC, magic);
779 	efx_generate_event(channel->efx, channel->channel, &event);
780 }
781 
782 /* Handle a transmit completion event
783  *
784  * The NIC batches TX completion events; the message we receive is of
785  * the form "complete all TX events up to this index".
786  */
787 static int
788 efx_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
789 {
790 	unsigned int tx_ev_desc_ptr;
791 	unsigned int tx_ev_q_label;
792 	struct efx_tx_queue *tx_queue;
793 	struct efx_nic *efx = channel->efx;
794 	int tx_packets = 0;
795 
796 	if (unlikely(ACCESS_ONCE(efx->reset_pending)))
797 		return 0;
798 
799 	if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
800 		/* Transmit completion */
801 		tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
802 		tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
803 		tx_queue = efx_channel_get_tx_queue(
804 			channel, tx_ev_q_label % EFX_TXQ_TYPES);
805 		tx_packets = ((tx_ev_desc_ptr - tx_queue->read_count) &
806 			      tx_queue->ptr_mask);
807 		channel->irq_mod_score += tx_packets;
808 		efx_xmit_done(tx_queue, tx_ev_desc_ptr);
809 	} else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
810 		/* Rewrite the FIFO write pointer */
811 		tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
812 		tx_queue = efx_channel_get_tx_queue(
813 			channel, tx_ev_q_label % EFX_TXQ_TYPES);
814 
815 		netif_tx_lock(efx->net_dev);
816 		efx_notify_tx_desc(tx_queue);
817 		netif_tx_unlock(efx->net_dev);
818 	} else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR) &&
819 		   EFX_WORKAROUND_10727(efx)) {
820 		efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
821 	} else {
822 		netif_err(efx, tx_err, efx->net_dev,
823 			  "channel %d unexpected TX event "
824 			  EFX_QWORD_FMT"\n", channel->channel,
825 			  EFX_QWORD_VAL(*event));
826 	}
827 
828 	return tx_packets;
829 }
830 
831 /* Detect errors included in the rx_evt_pkt_ok bit. */
832 static u16 efx_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
833 				const efx_qword_t *event)
834 {
835 	struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
836 	struct efx_nic *efx = rx_queue->efx;
837 	bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
838 	bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
839 	bool rx_ev_frm_trunc, rx_ev_drib_nib, rx_ev_tobe_disc;
840 	bool rx_ev_other_err, rx_ev_pause_frm;
841 	bool rx_ev_hdr_type, rx_ev_mcast_pkt;
842 	unsigned rx_ev_pkt_type;
843 
844 	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
845 	rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
846 	rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
847 	rx_ev_pkt_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_TYPE);
848 	rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
849 						 FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
850 	rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
851 						  FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
852 	rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
853 						   FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
854 	rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
855 	rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
856 	rx_ev_drib_nib = ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) ?
857 			  0 : EFX_QWORD_FIELD(*event, FSF_AA_RX_EV_DRIB_NIB));
858 	rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
859 
860 	/* Every error apart from tobe_disc and pause_frm */
861 	rx_ev_other_err = (rx_ev_drib_nib | rx_ev_tcp_udp_chksum_err |
862 			   rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
863 			   rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
864 
865 	/* Count errors that are not in MAC stats.  Ignore expected
866 	 * checksum errors during self-test. */
867 	if (rx_ev_frm_trunc)
868 		++channel->n_rx_frm_trunc;
869 	else if (rx_ev_tobe_disc)
870 		++channel->n_rx_tobe_disc;
871 	else if (!efx->loopback_selftest) {
872 		if (rx_ev_ip_hdr_chksum_err)
873 			++channel->n_rx_ip_hdr_chksum_err;
874 		else if (rx_ev_tcp_udp_chksum_err)
875 			++channel->n_rx_tcp_udp_chksum_err;
876 	}
877 
878 	/* TOBE_DISC is expected on unicast mismatches; don't print out an
879 	 * error message.  FRM_TRUNC indicates RXDP dropped the packet due
880 	 * to a FIFO overflow.
881 	 */
882 #ifdef DEBUG
883 	if (rx_ev_other_err && net_ratelimit()) {
884 		netif_dbg(efx, rx_err, efx->net_dev,
885 			  " RX queue %d unexpected RX event "
886 			  EFX_QWORD_FMT "%s%s%s%s%s%s%s%s\n",
887 			  efx_rx_queue_index(rx_queue), EFX_QWORD_VAL(*event),
888 			  rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
889 			  rx_ev_ip_hdr_chksum_err ?
890 			  " [IP_HDR_CHKSUM_ERR]" : "",
891 			  rx_ev_tcp_udp_chksum_err ?
892 			  " [TCP_UDP_CHKSUM_ERR]" : "",
893 			  rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
894 			  rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
895 			  rx_ev_drib_nib ? " [DRIB_NIB]" : "",
896 			  rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
897 			  rx_ev_pause_frm ? " [PAUSE]" : "");
898 	}
899 #endif
900 
901 	/* The frame must be discarded if any of these are true. */
902 	return (rx_ev_eth_crc_err | rx_ev_frm_trunc | rx_ev_drib_nib |
903 		rx_ev_tobe_disc | rx_ev_pause_frm) ?
904 		EFX_RX_PKT_DISCARD : 0;
905 }
906 
907 /* Handle receive events that are not in-order. */
908 static void
909 efx_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
910 {
911 	struct efx_nic *efx = rx_queue->efx;
912 	unsigned expected, dropped;
913 
914 	expected = rx_queue->removed_count & rx_queue->ptr_mask;
915 	dropped = (index - expected) & rx_queue->ptr_mask;
916 	netif_info(efx, rx_err, efx->net_dev,
917 		   "dropped %d events (index=%d expected=%d)\n",
918 		   dropped, index, expected);
919 
920 	efx_schedule_reset(efx, EFX_WORKAROUND_5676(efx) ?
921 			   RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
922 }
923 
924 /* Handle a packet received event
925  *
926  * The NIC gives a "discard" flag if it's a unicast packet with the
927  * wrong destination address
928  * Also "is multicast" and "matches multicast filter" flags can be used to
929  * discard non-matching multicast packets.
930  */
931 static void
932 efx_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
933 {
934 	unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
935 	unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
936 	unsigned expected_ptr;
937 	bool rx_ev_pkt_ok;
938 	u16 flags;
939 	struct efx_rx_queue *rx_queue;
940 	struct efx_nic *efx = channel->efx;
941 
942 	if (unlikely(ACCESS_ONCE(efx->reset_pending)))
943 		return;
944 
945 	/* Basic packet information */
946 	rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
947 	rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
948 	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
949 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT));
950 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP) != 1);
951 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
952 		channel->channel);
953 
954 	rx_queue = efx_channel_get_rx_queue(channel);
955 
956 	rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
957 	expected_ptr = rx_queue->removed_count & rx_queue->ptr_mask;
958 	if (unlikely(rx_ev_desc_ptr != expected_ptr))
959 		efx_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr);
960 
961 	if (likely(rx_ev_pkt_ok)) {
962 		/* If packet is marked as OK and packet type is TCP/IP or
963 		 * UDP/IP, then we can rely on the hardware checksum.
964 		 */
965 		flags = (rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP ||
966 			 rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP) ?
967 			EFX_RX_PKT_CSUMMED : 0;
968 	} else {
969 		flags = efx_handle_rx_not_ok(rx_queue, event);
970 	}
971 
972 	/* Detect multicast packets that didn't match the filter */
973 	rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
974 	if (rx_ev_mcast_pkt) {
975 		unsigned int rx_ev_mcast_hash_match =
976 			EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
977 
978 		if (unlikely(!rx_ev_mcast_hash_match)) {
979 			++channel->n_rx_mcast_mismatch;
980 			flags |= EFX_RX_PKT_DISCARD;
981 		}
982 	}
983 
984 	channel->irq_mod_score += 2;
985 
986 	/* Handle received packet */
987 	efx_rx_packet(rx_queue, rx_ev_desc_ptr, rx_ev_byte_cnt, flags);
988 }
989 
990 /* If this flush done event corresponds to a &struct efx_tx_queue, then
991  * send an %EFX_CHANNEL_MAGIC_TX_DRAIN event to drain the event queue
992  * of all transmit completions.
993  */
994 static void
995 efx_handle_tx_flush_done(struct efx_nic *efx, efx_qword_t *event)
996 {
997 	struct efx_tx_queue *tx_queue;
998 	int qid;
999 
1000 	qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1001 	if (qid < EFX_TXQ_TYPES * efx->n_tx_channels) {
1002 		tx_queue = efx_get_tx_queue(efx, qid / EFX_TXQ_TYPES,
1003 					    qid % EFX_TXQ_TYPES);
1004 
1005 		efx_magic_event(tx_queue->channel,
1006 				EFX_CHANNEL_MAGIC_TX_DRAIN(tx_queue));
1007 	}
1008 }
1009 
1010 /* If this flush done event corresponds to a &struct efx_rx_queue: If the flush
1011  * was succesful then send an %EFX_CHANNEL_MAGIC_RX_DRAIN, otherwise add
1012  * the RX queue back to the mask of RX queues in need of flushing.
1013  */
1014 static void
1015 efx_handle_rx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1016 {
1017 	struct efx_channel *channel;
1018 	struct efx_rx_queue *rx_queue;
1019 	int qid;
1020 	bool failed;
1021 
1022 	qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
1023 	failed = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
1024 	if (qid >= efx->n_channels)
1025 		return;
1026 	channel = efx_get_channel(efx, qid);
1027 	if (!efx_channel_has_rx_queue(channel))
1028 		return;
1029 	rx_queue = efx_channel_get_rx_queue(channel);
1030 
1031 	if (failed) {
1032 		netif_info(efx, hw, efx->net_dev,
1033 			   "RXQ %d flush retry\n", qid);
1034 		rx_queue->flush_pending = true;
1035 		atomic_inc(&efx->rxq_flush_pending);
1036 	} else {
1037 		efx_magic_event(efx_rx_queue_channel(rx_queue),
1038 				EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue));
1039 	}
1040 	atomic_dec(&efx->rxq_flush_outstanding);
1041 	if (efx_flush_wake(efx))
1042 		wake_up(&efx->flush_wq);
1043 }
1044 
1045 static void
1046 efx_handle_drain_event(struct efx_channel *channel)
1047 {
1048 	struct efx_nic *efx = channel->efx;
1049 
1050 	WARN_ON(atomic_read(&efx->drain_pending) == 0);
1051 	atomic_dec(&efx->drain_pending);
1052 	if (efx_flush_wake(efx))
1053 		wake_up(&efx->flush_wq);
1054 }
1055 
1056 static void
1057 efx_handle_generated_event(struct efx_channel *channel, efx_qword_t *event)
1058 {
1059 	struct efx_nic *efx = channel->efx;
1060 	struct efx_rx_queue *rx_queue =
1061 		efx_channel_has_rx_queue(channel) ?
1062 		efx_channel_get_rx_queue(channel) : NULL;
1063 	unsigned magic, code;
1064 
1065 	magic = EFX_QWORD_FIELD(*event, FSF_AZ_DRV_GEN_EV_MAGIC);
1066 	code = _EFX_CHANNEL_MAGIC_CODE(magic);
1067 
1068 	if (magic == EFX_CHANNEL_MAGIC_TEST(channel)) {
1069 		/* ignore */
1070 	} else if (rx_queue && magic == EFX_CHANNEL_MAGIC_FILL(rx_queue)) {
1071 		/* The queue must be empty, so we won't receive any rx
1072 		 * events, so efx_process_channel() won't refill the
1073 		 * queue. Refill it here */
1074 		efx_fast_push_rx_descriptors(rx_queue);
1075 	} else if (rx_queue && magic == EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue)) {
1076 		rx_queue->enabled = false;
1077 		efx_handle_drain_event(channel);
1078 	} else if (code == _EFX_CHANNEL_MAGIC_TX_DRAIN) {
1079 		efx_handle_drain_event(channel);
1080 	} else {
1081 		netif_dbg(efx, hw, efx->net_dev, "channel %d received "
1082 			  "generated event "EFX_QWORD_FMT"\n",
1083 			  channel->channel, EFX_QWORD_VAL(*event));
1084 	}
1085 }
1086 
1087 static void
1088 efx_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
1089 {
1090 	struct efx_nic *efx = channel->efx;
1091 	unsigned int ev_sub_code;
1092 	unsigned int ev_sub_data;
1093 
1094 	ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
1095 	ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1096 
1097 	switch (ev_sub_code) {
1098 	case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
1099 		netif_vdbg(efx, hw, efx->net_dev, "channel %d TXQ %d flushed\n",
1100 			   channel->channel, ev_sub_data);
1101 		efx_handle_tx_flush_done(efx, event);
1102 		break;
1103 	case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
1104 		netif_vdbg(efx, hw, efx->net_dev, "channel %d RXQ %d flushed\n",
1105 			   channel->channel, ev_sub_data);
1106 		efx_handle_rx_flush_done(efx, event);
1107 		break;
1108 	case FSE_AZ_EVQ_INIT_DONE_EV:
1109 		netif_dbg(efx, hw, efx->net_dev,
1110 			  "channel %d EVQ %d initialised\n",
1111 			  channel->channel, ev_sub_data);
1112 		break;
1113 	case FSE_AZ_SRM_UPD_DONE_EV:
1114 		netif_vdbg(efx, hw, efx->net_dev,
1115 			   "channel %d SRAM update done\n", channel->channel);
1116 		break;
1117 	case FSE_AZ_WAKE_UP_EV:
1118 		netif_vdbg(efx, hw, efx->net_dev,
1119 			   "channel %d RXQ %d wakeup event\n",
1120 			   channel->channel, ev_sub_data);
1121 		break;
1122 	case FSE_AZ_TIMER_EV:
1123 		netif_vdbg(efx, hw, efx->net_dev,
1124 			   "channel %d RX queue %d timer expired\n",
1125 			   channel->channel, ev_sub_data);
1126 		break;
1127 	case FSE_AA_RX_RECOVER_EV:
1128 		netif_err(efx, rx_err, efx->net_dev,
1129 			  "channel %d seen DRIVER RX_RESET event. "
1130 			"Resetting.\n", channel->channel);
1131 		atomic_inc(&efx->rx_reset);
1132 		efx_schedule_reset(efx,
1133 				   EFX_WORKAROUND_6555(efx) ?
1134 				   RESET_TYPE_RX_RECOVERY :
1135 				   RESET_TYPE_DISABLE);
1136 		break;
1137 	case FSE_BZ_RX_DSC_ERROR_EV:
1138 		netif_err(efx, rx_err, efx->net_dev,
1139 			  "RX DMA Q %d reports descriptor fetch error."
1140 			  " RX Q %d is disabled.\n", ev_sub_data, ev_sub_data);
1141 		efx_schedule_reset(efx, RESET_TYPE_RX_DESC_FETCH);
1142 		break;
1143 	case FSE_BZ_TX_DSC_ERROR_EV:
1144 		netif_err(efx, tx_err, efx->net_dev,
1145 			  "TX DMA Q %d reports descriptor fetch error."
1146 			  " TX Q %d is disabled.\n", ev_sub_data, ev_sub_data);
1147 		efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
1148 		break;
1149 	default:
1150 		netif_vdbg(efx, hw, efx->net_dev,
1151 			   "channel %d unknown driver event code %d "
1152 			   "data %04x\n", channel->channel, ev_sub_code,
1153 			   ev_sub_data);
1154 		break;
1155 	}
1156 }
1157 
1158 int efx_nic_process_eventq(struct efx_channel *channel, int budget)
1159 {
1160 	struct efx_nic *efx = channel->efx;
1161 	unsigned int read_ptr;
1162 	efx_qword_t event, *p_event;
1163 	int ev_code;
1164 	int tx_packets = 0;
1165 	int spent = 0;
1166 
1167 	read_ptr = channel->eventq_read_ptr;
1168 
1169 	for (;;) {
1170 		p_event = efx_event(channel, read_ptr);
1171 		event = *p_event;
1172 
1173 		if (!efx_event_present(&event))
1174 			/* End of events */
1175 			break;
1176 
1177 		netif_vdbg(channel->efx, intr, channel->efx->net_dev,
1178 			   "channel %d event is "EFX_QWORD_FMT"\n",
1179 			   channel->channel, EFX_QWORD_VAL(event));
1180 
1181 		/* Clear this event by marking it all ones */
1182 		EFX_SET_QWORD(*p_event);
1183 
1184 		++read_ptr;
1185 
1186 		ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
1187 
1188 		switch (ev_code) {
1189 		case FSE_AZ_EV_CODE_RX_EV:
1190 			efx_handle_rx_event(channel, &event);
1191 			if (++spent == budget)
1192 				goto out;
1193 			break;
1194 		case FSE_AZ_EV_CODE_TX_EV:
1195 			tx_packets += efx_handle_tx_event(channel, &event);
1196 			if (tx_packets > efx->txq_entries) {
1197 				spent = budget;
1198 				goto out;
1199 			}
1200 			break;
1201 		case FSE_AZ_EV_CODE_DRV_GEN_EV:
1202 			efx_handle_generated_event(channel, &event);
1203 			break;
1204 		case FSE_AZ_EV_CODE_DRIVER_EV:
1205 			efx_handle_driver_event(channel, &event);
1206 			break;
1207 		case FSE_CZ_EV_CODE_MCDI_EV:
1208 			efx_mcdi_process_event(channel, &event);
1209 			break;
1210 		case FSE_AZ_EV_CODE_GLOBAL_EV:
1211 			if (efx->type->handle_global_event &&
1212 			    efx->type->handle_global_event(channel, &event))
1213 				break;
1214 			/* else fall through */
1215 		default:
1216 			netif_err(channel->efx, hw, channel->efx->net_dev,
1217 				  "channel %d unknown event type %d (data "
1218 				  EFX_QWORD_FMT ")\n", channel->channel,
1219 				  ev_code, EFX_QWORD_VAL(event));
1220 		}
1221 	}
1222 
1223 out:
1224 	channel->eventq_read_ptr = read_ptr;
1225 	return spent;
1226 }
1227 
1228 /* Check whether an event is present in the eventq at the current
1229  * read pointer.  Only useful for self-test.
1230  */
1231 bool efx_nic_event_present(struct efx_channel *channel)
1232 {
1233 	return efx_event_present(efx_event(channel, channel->eventq_read_ptr));
1234 }
1235 
1236 /* Allocate buffer table entries for event queue */
1237 int efx_nic_probe_eventq(struct efx_channel *channel)
1238 {
1239 	struct efx_nic *efx = channel->efx;
1240 	unsigned entries;
1241 
1242 	entries = channel->eventq_mask + 1;
1243 	return efx_alloc_special_buffer(efx, &channel->eventq,
1244 					entries * sizeof(efx_qword_t));
1245 }
1246 
1247 void efx_nic_init_eventq(struct efx_channel *channel)
1248 {
1249 	efx_oword_t reg;
1250 	struct efx_nic *efx = channel->efx;
1251 
1252 	netif_dbg(efx, hw, efx->net_dev,
1253 		  "channel %d event queue in special buffers %d-%d\n",
1254 		  channel->channel, channel->eventq.index,
1255 		  channel->eventq.index + channel->eventq.entries - 1);
1256 
1257 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1258 		EFX_POPULATE_OWORD_3(reg,
1259 				     FRF_CZ_TIMER_Q_EN, 1,
1260 				     FRF_CZ_HOST_NOTIFY_MODE, 0,
1261 				     FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
1262 		efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1263 	}
1264 
1265 	/* Pin event queue buffer */
1266 	efx_init_special_buffer(efx, &channel->eventq);
1267 
1268 	/* Fill event queue with all ones (i.e. empty events) */
1269 	memset(channel->eventq.addr, 0xff, channel->eventq.len);
1270 
1271 	/* Push event queue to card */
1272 	EFX_POPULATE_OWORD_3(reg,
1273 			     FRF_AZ_EVQ_EN, 1,
1274 			     FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
1275 			     FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
1276 	efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1277 			 channel->channel);
1278 
1279 	efx->type->push_irq_moderation(channel);
1280 }
1281 
1282 void efx_nic_fini_eventq(struct efx_channel *channel)
1283 {
1284 	efx_oword_t reg;
1285 	struct efx_nic *efx = channel->efx;
1286 
1287 	/* Remove event queue from card */
1288 	EFX_ZERO_OWORD(reg);
1289 	efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1290 			 channel->channel);
1291 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1292 		efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1293 
1294 	/* Unpin event queue */
1295 	efx_fini_special_buffer(efx, &channel->eventq);
1296 }
1297 
1298 /* Free buffers backing event queue */
1299 void efx_nic_remove_eventq(struct efx_channel *channel)
1300 {
1301 	efx_free_special_buffer(channel->efx, &channel->eventq);
1302 }
1303 
1304 
1305 void efx_nic_generate_test_event(struct efx_channel *channel)
1306 {
1307 	efx_magic_event(channel, EFX_CHANNEL_MAGIC_TEST(channel));
1308 }
1309 
1310 void efx_nic_generate_fill_event(struct efx_rx_queue *rx_queue)
1311 {
1312 	efx_magic_event(efx_rx_queue_channel(rx_queue),
1313 			EFX_CHANNEL_MAGIC_FILL(rx_queue));
1314 }
1315 
1316 /**************************************************************************
1317  *
1318  * Hardware interrupts
1319  * The hardware interrupt handler does very little work; all the event
1320  * queue processing is carried out by per-channel tasklets.
1321  *
1322  **************************************************************************/
1323 
1324 /* Enable/disable/generate interrupts */
1325 static inline void efx_nic_interrupts(struct efx_nic *efx,
1326 				      bool enabled, bool force)
1327 {
1328 	efx_oword_t int_en_reg_ker;
1329 
1330 	EFX_POPULATE_OWORD_3(int_en_reg_ker,
1331 			     FRF_AZ_KER_INT_LEVE_SEL, efx->irq_level,
1332 			     FRF_AZ_KER_INT_KER, force,
1333 			     FRF_AZ_DRV_INT_EN_KER, enabled);
1334 	efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
1335 }
1336 
1337 void efx_nic_enable_interrupts(struct efx_nic *efx)
1338 {
1339 	EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
1340 	wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
1341 
1342 	efx_nic_interrupts(efx, true, false);
1343 }
1344 
1345 void efx_nic_disable_interrupts(struct efx_nic *efx)
1346 {
1347 	/* Disable interrupts */
1348 	efx_nic_interrupts(efx, false, false);
1349 }
1350 
1351 /* Generate a test interrupt
1352  * Interrupt must already have been enabled, otherwise nasty things
1353  * may happen.
1354  */
1355 void efx_nic_generate_interrupt(struct efx_nic *efx)
1356 {
1357 	efx_nic_interrupts(efx, true, true);
1358 }
1359 
1360 /* Process a fatal interrupt
1361  * Disable bus mastering ASAP and schedule a reset
1362  */
1363 irqreturn_t efx_nic_fatal_interrupt(struct efx_nic *efx)
1364 {
1365 	struct falcon_nic_data *nic_data = efx->nic_data;
1366 	efx_oword_t *int_ker = efx->irq_status.addr;
1367 	efx_oword_t fatal_intr;
1368 	int error, mem_perr;
1369 
1370 	efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
1371 	error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
1372 
1373 	netif_err(efx, hw, efx->net_dev, "SYSTEM ERROR "EFX_OWORD_FMT" status "
1374 		  EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
1375 		  EFX_OWORD_VAL(fatal_intr),
1376 		  error ? "disabling bus mastering" : "no recognised error");
1377 
1378 	/* If this is a memory parity error dump which blocks are offending */
1379 	mem_perr = (EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER) ||
1380 		    EFX_OWORD_FIELD(fatal_intr, FRF_AZ_SRM_PERR_INT_KER));
1381 	if (mem_perr) {
1382 		efx_oword_t reg;
1383 		efx_reado(efx, &reg, FR_AZ_MEM_STAT);
1384 		netif_err(efx, hw, efx->net_dev,
1385 			  "SYSTEM ERROR: memory parity error "EFX_OWORD_FMT"\n",
1386 			  EFX_OWORD_VAL(reg));
1387 	}
1388 
1389 	/* Disable both devices */
1390 	pci_clear_master(efx->pci_dev);
1391 	if (efx_nic_is_dual_func(efx))
1392 		pci_clear_master(nic_data->pci_dev2);
1393 	efx_nic_disable_interrupts(efx);
1394 
1395 	/* Count errors and reset or disable the NIC accordingly */
1396 	if (efx->int_error_count == 0 ||
1397 	    time_after(jiffies, efx->int_error_expire)) {
1398 		efx->int_error_count = 0;
1399 		efx->int_error_expire =
1400 			jiffies + EFX_INT_ERROR_EXPIRE * HZ;
1401 	}
1402 	if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
1403 		netif_err(efx, hw, efx->net_dev,
1404 			  "SYSTEM ERROR - reset scheduled\n");
1405 		efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
1406 	} else {
1407 		netif_err(efx, hw, efx->net_dev,
1408 			  "SYSTEM ERROR - max number of errors seen."
1409 			  "NIC will be disabled\n");
1410 		efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1411 	}
1412 
1413 	return IRQ_HANDLED;
1414 }
1415 
1416 /* Handle a legacy interrupt
1417  * Acknowledges the interrupt and schedule event queue processing.
1418  */
1419 static irqreturn_t efx_legacy_interrupt(int irq, void *dev_id)
1420 {
1421 	struct efx_nic *efx = dev_id;
1422 	efx_oword_t *int_ker = efx->irq_status.addr;
1423 	irqreturn_t result = IRQ_NONE;
1424 	struct efx_channel *channel;
1425 	efx_dword_t reg;
1426 	u32 queues;
1427 	int syserr;
1428 
1429 	/* Could this be ours?  If interrupts are disabled then the
1430 	 * channel state may not be valid.
1431 	 */
1432 	if (!efx->legacy_irq_enabled)
1433 		return result;
1434 
1435 	/* Read the ISR which also ACKs the interrupts */
1436 	efx_readd(efx, &reg, FR_BZ_INT_ISR0);
1437 	queues = EFX_EXTRACT_DWORD(reg, 0, 31);
1438 
1439 	/* Handle non-event-queue sources */
1440 	if (queues & (1U << efx->irq_level)) {
1441 		syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1442 		if (unlikely(syserr))
1443 			return efx_nic_fatal_interrupt(efx);
1444 		efx->last_irq_cpu = raw_smp_processor_id();
1445 	}
1446 
1447 	if (queues != 0) {
1448 		if (EFX_WORKAROUND_15783(efx))
1449 			efx->irq_zero_count = 0;
1450 
1451 		/* Schedule processing of any interrupting queues */
1452 		efx_for_each_channel(channel, efx) {
1453 			if (queues & 1)
1454 				efx_schedule_channel_irq(channel);
1455 			queues >>= 1;
1456 		}
1457 		result = IRQ_HANDLED;
1458 
1459 	} else if (EFX_WORKAROUND_15783(efx)) {
1460 		efx_qword_t *event;
1461 
1462 		/* We can't return IRQ_HANDLED more than once on seeing ISR=0
1463 		 * because this might be a shared interrupt. */
1464 		if (efx->irq_zero_count++ == 0)
1465 			result = IRQ_HANDLED;
1466 
1467 		/* Ensure we schedule or rearm all event queues */
1468 		efx_for_each_channel(channel, efx) {
1469 			event = efx_event(channel, channel->eventq_read_ptr);
1470 			if (efx_event_present(event))
1471 				efx_schedule_channel_irq(channel);
1472 			else
1473 				efx_nic_eventq_read_ack(channel);
1474 		}
1475 	}
1476 
1477 	if (result == IRQ_HANDLED)
1478 		netif_vdbg(efx, intr, efx->net_dev,
1479 			   "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
1480 			   irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
1481 
1482 	return result;
1483 }
1484 
1485 /* Handle an MSI interrupt
1486  *
1487  * Handle an MSI hardware interrupt.  This routine schedules event
1488  * queue processing.  No interrupt acknowledgement cycle is necessary.
1489  * Also, we never need to check that the interrupt is for us, since
1490  * MSI interrupts cannot be shared.
1491  */
1492 static irqreturn_t efx_msi_interrupt(int irq, void *dev_id)
1493 {
1494 	struct efx_channel *channel = *(struct efx_channel **)dev_id;
1495 	struct efx_nic *efx = channel->efx;
1496 	efx_oword_t *int_ker = efx->irq_status.addr;
1497 	int syserr;
1498 
1499 	netif_vdbg(efx, intr, efx->net_dev,
1500 		   "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
1501 		   irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
1502 
1503 	/* Handle non-event-queue sources */
1504 	if (channel->channel == efx->irq_level) {
1505 		syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1506 		if (unlikely(syserr))
1507 			return efx_nic_fatal_interrupt(efx);
1508 		efx->last_irq_cpu = raw_smp_processor_id();
1509 	}
1510 
1511 	/* Schedule processing of the channel */
1512 	efx_schedule_channel_irq(channel);
1513 
1514 	return IRQ_HANDLED;
1515 }
1516 
1517 
1518 /* Setup RSS indirection table.
1519  * This maps from the hash value of the packet to RXQ
1520  */
1521 void efx_nic_push_rx_indir_table(struct efx_nic *efx)
1522 {
1523 	size_t i = 0;
1524 	efx_dword_t dword;
1525 
1526 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0)
1527 		return;
1528 
1529 	BUILD_BUG_ON(ARRAY_SIZE(efx->rx_indir_table) !=
1530 		     FR_BZ_RX_INDIRECTION_TBL_ROWS);
1531 
1532 	for (i = 0; i < FR_BZ_RX_INDIRECTION_TBL_ROWS; i++) {
1533 		EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
1534 				     efx->rx_indir_table[i]);
1535 		efx_writed_table(efx, &dword, FR_BZ_RX_INDIRECTION_TBL, i);
1536 	}
1537 }
1538 
1539 /* Hook interrupt handler(s)
1540  * Try MSI and then legacy interrupts.
1541  */
1542 int efx_nic_init_interrupt(struct efx_nic *efx)
1543 {
1544 	struct efx_channel *channel;
1545 	int rc;
1546 
1547 	if (!EFX_INT_MODE_USE_MSI(efx)) {
1548 		irq_handler_t handler;
1549 		if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1550 			handler = efx_legacy_interrupt;
1551 		else
1552 			handler = falcon_legacy_interrupt_a1;
1553 
1554 		rc = request_irq(efx->legacy_irq, handler, IRQF_SHARED,
1555 				 efx->name, efx);
1556 		if (rc) {
1557 			netif_err(efx, drv, efx->net_dev,
1558 				  "failed to hook legacy IRQ %d\n",
1559 				  efx->pci_dev->irq);
1560 			goto fail1;
1561 		}
1562 		return 0;
1563 	}
1564 
1565 	/* Hook MSI or MSI-X interrupt */
1566 	efx_for_each_channel(channel, efx) {
1567 		rc = request_irq(channel->irq, efx_msi_interrupt,
1568 				 IRQF_PROBE_SHARED, /* Not shared */
1569 				 efx->channel_name[channel->channel],
1570 				 &efx->channel[channel->channel]);
1571 		if (rc) {
1572 			netif_err(efx, drv, efx->net_dev,
1573 				  "failed to hook IRQ %d\n", channel->irq);
1574 			goto fail2;
1575 		}
1576 	}
1577 
1578 	return 0;
1579 
1580  fail2:
1581 	efx_for_each_channel(channel, efx)
1582 		free_irq(channel->irq, &efx->channel[channel->channel]);
1583  fail1:
1584 	return rc;
1585 }
1586 
1587 void efx_nic_fini_interrupt(struct efx_nic *efx)
1588 {
1589 	struct efx_channel *channel;
1590 	efx_oword_t reg;
1591 
1592 	/* Disable MSI/MSI-X interrupts */
1593 	efx_for_each_channel(channel, efx) {
1594 		if (channel->irq)
1595 			free_irq(channel->irq, &efx->channel[channel->channel]);
1596 	}
1597 
1598 	/* ACK legacy interrupt */
1599 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1600 		efx_reado(efx, &reg, FR_BZ_INT_ISR0);
1601 	else
1602 		falcon_irq_ack_a1(efx);
1603 
1604 	/* Disable legacy interrupt */
1605 	if (efx->legacy_irq)
1606 		free_irq(efx->legacy_irq, efx);
1607 }
1608 
1609 u32 efx_nic_fpga_ver(struct efx_nic *efx)
1610 {
1611 	efx_oword_t altera_build;
1612 	efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
1613 	return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
1614 }
1615 
1616 void efx_nic_init_common(struct efx_nic *efx)
1617 {
1618 	efx_oword_t temp;
1619 
1620 	/* Set positions of descriptor caches in SRAM. */
1621 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR,
1622 			     efx->type->tx_dc_base / 8);
1623 	efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
1624 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR,
1625 			     efx->type->rx_dc_base / 8);
1626 	efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
1627 
1628 	/* Set TX descriptor cache size. */
1629 	BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
1630 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
1631 	efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
1632 
1633 	/* Set RX descriptor cache size.  Set low watermark to size-8, as
1634 	 * this allows most efficient prefetching.
1635 	 */
1636 	BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
1637 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
1638 	efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
1639 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
1640 	efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
1641 
1642 	/* Program INT_KER address */
1643 	EFX_POPULATE_OWORD_2(temp,
1644 			     FRF_AZ_NORM_INT_VEC_DIS_KER,
1645 			     EFX_INT_MODE_USE_MSI(efx),
1646 			     FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
1647 	efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
1648 
1649 	if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
1650 		/* Use an interrupt level unused by event queues */
1651 		efx->irq_level = 0x1f;
1652 	else
1653 		/* Use a valid MSI-X vector */
1654 		efx->irq_level = 0;
1655 
1656 	/* Enable all the genuinely fatal interrupts.  (They are still
1657 	 * masked by the overall interrupt mask, controlled by
1658 	 * falcon_interrupts()).
1659 	 *
1660 	 * Note: All other fatal interrupts are enabled
1661 	 */
1662 	EFX_POPULATE_OWORD_3(temp,
1663 			     FRF_AZ_ILL_ADR_INT_KER_EN, 1,
1664 			     FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
1665 			     FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
1666 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1667 		EFX_SET_OWORD_FIELD(temp, FRF_CZ_SRAM_PERR_INT_P_KER_EN, 1);
1668 	EFX_INVERT_OWORD(temp);
1669 	efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
1670 
1671 	efx_nic_push_rx_indir_table(efx);
1672 
1673 	/* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
1674 	 * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
1675 	 */
1676 	efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
1677 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
1678 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
1679 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
1680 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 1);
1681 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
1682 	/* Enable SW_EV to inherit in char driver - assume harmless here */
1683 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
1684 	/* Prefetch threshold 2 => fetch when descriptor cache half empty */
1685 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
1686 	/* Disable hardware watchdog which can misfire */
1687 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
1688 	/* Squash TX of packets of 16 bytes or less */
1689 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1690 		EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
1691 	efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
1692 
1693 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
1694 		EFX_POPULATE_OWORD_4(temp,
1695 				     /* Default values */
1696 				     FRF_BZ_TX_PACE_SB_NOT_AF, 0x15,
1697 				     FRF_BZ_TX_PACE_SB_AF, 0xb,
1698 				     FRF_BZ_TX_PACE_FB_BASE, 0,
1699 				     /* Allow large pace values in the
1700 				      * fast bin. */
1701 				     FRF_BZ_TX_PACE_BIN_TH,
1702 				     FFE_BZ_TX_PACE_RESERVED);
1703 		efx_writeo(efx, &temp, FR_BZ_TX_PACE);
1704 	}
1705 }
1706 
1707 /* Register dump */
1708 
1709 #define REGISTER_REVISION_A	1
1710 #define REGISTER_REVISION_B	2
1711 #define REGISTER_REVISION_C	3
1712 #define REGISTER_REVISION_Z	3	/* latest revision */
1713 
1714 struct efx_nic_reg {
1715 	u32 offset:24;
1716 	u32 min_revision:2, max_revision:2;
1717 };
1718 
1719 #define REGISTER(name, min_rev, max_rev) {				\
1720 	FR_ ## min_rev ## max_rev ## _ ## name,				\
1721 	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev	\
1722 }
1723 #define REGISTER_AA(name) REGISTER(name, A, A)
1724 #define REGISTER_AB(name) REGISTER(name, A, B)
1725 #define REGISTER_AZ(name) REGISTER(name, A, Z)
1726 #define REGISTER_BB(name) REGISTER(name, B, B)
1727 #define REGISTER_BZ(name) REGISTER(name, B, Z)
1728 #define REGISTER_CZ(name) REGISTER(name, C, Z)
1729 
1730 static const struct efx_nic_reg efx_nic_regs[] = {
1731 	REGISTER_AZ(ADR_REGION),
1732 	REGISTER_AZ(INT_EN_KER),
1733 	REGISTER_BZ(INT_EN_CHAR),
1734 	REGISTER_AZ(INT_ADR_KER),
1735 	REGISTER_BZ(INT_ADR_CHAR),
1736 	/* INT_ACK_KER is WO */
1737 	/* INT_ISR0 is RC */
1738 	REGISTER_AZ(HW_INIT),
1739 	REGISTER_CZ(USR_EV_CFG),
1740 	REGISTER_AB(EE_SPI_HCMD),
1741 	REGISTER_AB(EE_SPI_HADR),
1742 	REGISTER_AB(EE_SPI_HDATA),
1743 	REGISTER_AB(EE_BASE_PAGE),
1744 	REGISTER_AB(EE_VPD_CFG0),
1745 	/* EE_VPD_SW_CNTL and EE_VPD_SW_DATA are not used */
1746 	/* PMBX_DBG_IADDR and PBMX_DBG_IDATA are indirect */
1747 	/* PCIE_CORE_INDIRECT is indirect */
1748 	REGISTER_AB(NIC_STAT),
1749 	REGISTER_AB(GPIO_CTL),
1750 	REGISTER_AB(GLB_CTL),
1751 	/* FATAL_INTR_KER and FATAL_INTR_CHAR are partly RC */
1752 	REGISTER_BZ(DP_CTRL),
1753 	REGISTER_AZ(MEM_STAT),
1754 	REGISTER_AZ(CS_DEBUG),
1755 	REGISTER_AZ(ALTERA_BUILD),
1756 	REGISTER_AZ(CSR_SPARE),
1757 	REGISTER_AB(PCIE_SD_CTL0123),
1758 	REGISTER_AB(PCIE_SD_CTL45),
1759 	REGISTER_AB(PCIE_PCS_CTL_STAT),
1760 	/* DEBUG_DATA_OUT is not used */
1761 	/* DRV_EV is WO */
1762 	REGISTER_AZ(EVQ_CTL),
1763 	REGISTER_AZ(EVQ_CNT1),
1764 	REGISTER_AZ(EVQ_CNT2),
1765 	REGISTER_AZ(BUF_TBL_CFG),
1766 	REGISTER_AZ(SRM_RX_DC_CFG),
1767 	REGISTER_AZ(SRM_TX_DC_CFG),
1768 	REGISTER_AZ(SRM_CFG),
1769 	/* BUF_TBL_UPD is WO */
1770 	REGISTER_AZ(SRM_UPD_EVQ),
1771 	REGISTER_AZ(SRAM_PARITY),
1772 	REGISTER_AZ(RX_CFG),
1773 	REGISTER_BZ(RX_FILTER_CTL),
1774 	/* RX_FLUSH_DESCQ is WO */
1775 	REGISTER_AZ(RX_DC_CFG),
1776 	REGISTER_AZ(RX_DC_PF_WM),
1777 	REGISTER_BZ(RX_RSS_TKEY),
1778 	/* RX_NODESC_DROP is RC */
1779 	REGISTER_AA(RX_SELF_RST),
1780 	/* RX_DEBUG, RX_PUSH_DROP are not used */
1781 	REGISTER_CZ(RX_RSS_IPV6_REG1),
1782 	REGISTER_CZ(RX_RSS_IPV6_REG2),
1783 	REGISTER_CZ(RX_RSS_IPV6_REG3),
1784 	/* TX_FLUSH_DESCQ is WO */
1785 	REGISTER_AZ(TX_DC_CFG),
1786 	REGISTER_AA(TX_CHKSM_CFG),
1787 	REGISTER_AZ(TX_CFG),
1788 	/* TX_PUSH_DROP is not used */
1789 	REGISTER_AZ(TX_RESERVED),
1790 	REGISTER_BZ(TX_PACE),
1791 	/* TX_PACE_DROP_QID is RC */
1792 	REGISTER_BB(TX_VLAN),
1793 	REGISTER_BZ(TX_IPFIL_PORTEN),
1794 	REGISTER_AB(MD_TXD),
1795 	REGISTER_AB(MD_RXD),
1796 	REGISTER_AB(MD_CS),
1797 	REGISTER_AB(MD_PHY_ADR),
1798 	REGISTER_AB(MD_ID),
1799 	/* MD_STAT is RC */
1800 	REGISTER_AB(MAC_STAT_DMA),
1801 	REGISTER_AB(MAC_CTRL),
1802 	REGISTER_BB(GEN_MODE),
1803 	REGISTER_AB(MAC_MC_HASH_REG0),
1804 	REGISTER_AB(MAC_MC_HASH_REG1),
1805 	REGISTER_AB(GM_CFG1),
1806 	REGISTER_AB(GM_CFG2),
1807 	/* GM_IPG and GM_HD are not used */
1808 	REGISTER_AB(GM_MAX_FLEN),
1809 	/* GM_TEST is not used */
1810 	REGISTER_AB(GM_ADR1),
1811 	REGISTER_AB(GM_ADR2),
1812 	REGISTER_AB(GMF_CFG0),
1813 	REGISTER_AB(GMF_CFG1),
1814 	REGISTER_AB(GMF_CFG2),
1815 	REGISTER_AB(GMF_CFG3),
1816 	REGISTER_AB(GMF_CFG4),
1817 	REGISTER_AB(GMF_CFG5),
1818 	REGISTER_BB(TX_SRC_MAC_CTL),
1819 	REGISTER_AB(XM_ADR_LO),
1820 	REGISTER_AB(XM_ADR_HI),
1821 	REGISTER_AB(XM_GLB_CFG),
1822 	REGISTER_AB(XM_TX_CFG),
1823 	REGISTER_AB(XM_RX_CFG),
1824 	REGISTER_AB(XM_MGT_INT_MASK),
1825 	REGISTER_AB(XM_FC),
1826 	REGISTER_AB(XM_PAUSE_TIME),
1827 	REGISTER_AB(XM_TX_PARAM),
1828 	REGISTER_AB(XM_RX_PARAM),
1829 	/* XM_MGT_INT_MSK (note no 'A') is RC */
1830 	REGISTER_AB(XX_PWR_RST),
1831 	REGISTER_AB(XX_SD_CTL),
1832 	REGISTER_AB(XX_TXDRV_CTL),
1833 	/* XX_PRBS_CTL, XX_PRBS_CHK and XX_PRBS_ERR are not used */
1834 	/* XX_CORE_STAT is partly RC */
1835 };
1836 
1837 struct efx_nic_reg_table {
1838 	u32 offset:24;
1839 	u32 min_revision:2, max_revision:2;
1840 	u32 step:6, rows:21;
1841 };
1842 
1843 #define REGISTER_TABLE_DIMENSIONS(_, offset, min_rev, max_rev, step, rows) { \
1844 	offset,								\
1845 	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev,	\
1846 	step, rows							\
1847 }
1848 #define REGISTER_TABLE(name, min_rev, max_rev)				\
1849 	REGISTER_TABLE_DIMENSIONS(					\
1850 		name, FR_ ## min_rev ## max_rev ## _ ## name,		\
1851 		min_rev, max_rev,					\
1852 		FR_ ## min_rev ## max_rev ## _ ## name ## _STEP,	\
1853 		FR_ ## min_rev ## max_rev ## _ ## name ## _ROWS)
1854 #define REGISTER_TABLE_AA(name) REGISTER_TABLE(name, A, A)
1855 #define REGISTER_TABLE_AZ(name) REGISTER_TABLE(name, A, Z)
1856 #define REGISTER_TABLE_BB(name) REGISTER_TABLE(name, B, B)
1857 #define REGISTER_TABLE_BZ(name) REGISTER_TABLE(name, B, Z)
1858 #define REGISTER_TABLE_BB_CZ(name)					\
1859 	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, B, B,		\
1860 				  FR_BZ_ ## name ## _STEP,		\
1861 				  FR_BB_ ## name ## _ROWS),		\
1862 	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, C, Z,		\
1863 				  FR_BZ_ ## name ## _STEP,		\
1864 				  FR_CZ_ ## name ## _ROWS)
1865 #define REGISTER_TABLE_CZ(name) REGISTER_TABLE(name, C, Z)
1866 
1867 static const struct efx_nic_reg_table efx_nic_reg_tables[] = {
1868 	/* DRIVER is not used */
1869 	/* EVQ_RPTR, TIMER_COMMAND, USR_EV and {RX,TX}_DESC_UPD are WO */
1870 	REGISTER_TABLE_BB(TX_IPFIL_TBL),
1871 	REGISTER_TABLE_BB(TX_SRC_MAC_TBL),
1872 	REGISTER_TABLE_AA(RX_DESC_PTR_TBL_KER),
1873 	REGISTER_TABLE_BB_CZ(RX_DESC_PTR_TBL),
1874 	REGISTER_TABLE_AA(TX_DESC_PTR_TBL_KER),
1875 	REGISTER_TABLE_BB_CZ(TX_DESC_PTR_TBL),
1876 	REGISTER_TABLE_AA(EVQ_PTR_TBL_KER),
1877 	REGISTER_TABLE_BB_CZ(EVQ_PTR_TBL),
1878 	/* We can't reasonably read all of the buffer table (up to 8MB!).
1879 	 * However this driver will only use a few entries.  Reading
1880 	 * 1K entries allows for some expansion of queue count and
1881 	 * size before we need to change the version. */
1882 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL_KER, FR_AA_BUF_FULL_TBL_KER,
1883 				  A, A, 8, 1024),
1884 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL, FR_BZ_BUF_FULL_TBL,
1885 				  B, Z, 8, 1024),
1886 	REGISTER_TABLE_CZ(RX_MAC_FILTER_TBL0),
1887 	REGISTER_TABLE_BB_CZ(TIMER_TBL),
1888 	REGISTER_TABLE_BB_CZ(TX_PACE_TBL),
1889 	REGISTER_TABLE_BZ(RX_INDIRECTION_TBL),
1890 	/* TX_FILTER_TBL0 is huge and not used by this driver */
1891 	REGISTER_TABLE_CZ(TX_MAC_FILTER_TBL0),
1892 	REGISTER_TABLE_CZ(MC_TREG_SMEM),
1893 	/* MSIX_PBA_TABLE is not mapped */
1894 	/* SRM_DBG is not mapped (and is redundant with BUF_FLL_TBL) */
1895 	REGISTER_TABLE_BZ(RX_FILTER_TBL0),
1896 };
1897 
1898 size_t efx_nic_get_regs_len(struct efx_nic *efx)
1899 {
1900 	const struct efx_nic_reg *reg;
1901 	const struct efx_nic_reg_table *table;
1902 	size_t len = 0;
1903 
1904 	for (reg = efx_nic_regs;
1905 	     reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
1906 	     reg++)
1907 		if (efx->type->revision >= reg->min_revision &&
1908 		    efx->type->revision <= reg->max_revision)
1909 			len += sizeof(efx_oword_t);
1910 
1911 	for (table = efx_nic_reg_tables;
1912 	     table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
1913 	     table++)
1914 		if (efx->type->revision >= table->min_revision &&
1915 		    efx->type->revision <= table->max_revision)
1916 			len += table->rows * min_t(size_t, table->step, 16);
1917 
1918 	return len;
1919 }
1920 
1921 void efx_nic_get_regs(struct efx_nic *efx, void *buf)
1922 {
1923 	const struct efx_nic_reg *reg;
1924 	const struct efx_nic_reg_table *table;
1925 
1926 	for (reg = efx_nic_regs;
1927 	     reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
1928 	     reg++) {
1929 		if (efx->type->revision >= reg->min_revision &&
1930 		    efx->type->revision <= reg->max_revision) {
1931 			efx_reado(efx, (efx_oword_t *)buf, reg->offset);
1932 			buf += sizeof(efx_oword_t);
1933 		}
1934 	}
1935 
1936 	for (table = efx_nic_reg_tables;
1937 	     table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
1938 	     table++) {
1939 		size_t size, i;
1940 
1941 		if (!(efx->type->revision >= table->min_revision &&
1942 		      efx->type->revision <= table->max_revision))
1943 			continue;
1944 
1945 		size = min_t(size_t, table->step, 16);
1946 
1947 		for (i = 0; i < table->rows; i++) {
1948 			switch (table->step) {
1949 			case 4: /* 32-bit register or SRAM */
1950 				efx_readd_table(efx, buf, table->offset, i);
1951 				break;
1952 			case 8: /* 64-bit SRAM */
1953 				efx_sram_readq(efx,
1954 					       efx->membase + table->offset,
1955 					       buf, i);
1956 				break;
1957 			case 16: /* 128-bit register */
1958 				efx_reado_table(efx, buf, table->offset, i);
1959 				break;
1960 			case 32: /* 128-bit register, interleaved */
1961 				efx_reado_table(efx, buf, table->offset, 2 * i);
1962 				break;
1963 			default:
1964 				WARN_ON(1);
1965 				return;
1966 			}
1967 			buf += size;
1968 		}
1969 	}
1970 }
1971