xref: /openbmc/linux/drivers/net/ethernet/sfc/nic.c (revision 5b6262d0)
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 /* We poll for events every FLUSH_INTERVAL ms, and check FLUSH_POLL_COUNT times
53  */
54 #define EFX_FLUSH_INTERVAL 10
55 #define EFX_FLUSH_POLL_COUNT 100
56 
57 /* Depth of RX flush request fifo */
58 #define EFX_RX_FLUSH_COUNT 4
59 
60 /* Generated event code for efx_generate_test_event() */
61 #define EFX_CHANNEL_MAGIC_TEST(_channel)	\
62 	(0x00010100 + (_channel)->channel)
63 
64 /* Generated event code for efx_generate_fill_event() */
65 #define EFX_CHANNEL_MAGIC_FILL(_channel)	\
66 	(0x00010200 + (_channel)->channel)
67 
68 /**************************************************************************
69  *
70  * Solarstorm hardware access
71  *
72  **************************************************************************/
73 
74 static inline void efx_write_buf_tbl(struct efx_nic *efx, efx_qword_t *value,
75 				     unsigned int index)
76 {
77 	efx_sram_writeq(efx, efx->membase + efx->type->buf_tbl_base,
78 			value, index);
79 }
80 
81 /* Read the current event from the event queue */
82 static inline efx_qword_t *efx_event(struct efx_channel *channel,
83 				     unsigned int index)
84 {
85 	return ((efx_qword_t *) (channel->eventq.addr)) +
86 		(index & channel->eventq_mask);
87 }
88 
89 /* See if an event is present
90  *
91  * We check both the high and low dword of the event for all ones.  We
92  * wrote all ones when we cleared the event, and no valid event can
93  * have all ones in either its high or low dwords.  This approach is
94  * robust against reordering.
95  *
96  * Note that using a single 64-bit comparison is incorrect; even
97  * though the CPU read will be atomic, the DMA write may not be.
98  */
99 static inline int efx_event_present(efx_qword_t *event)
100 {
101 	return !(EFX_DWORD_IS_ALL_ONES(event->dword[0]) |
102 		  EFX_DWORD_IS_ALL_ONES(event->dword[1]));
103 }
104 
105 static bool efx_masked_compare_oword(const efx_oword_t *a, const efx_oword_t *b,
106 				     const efx_oword_t *mask)
107 {
108 	return ((a->u64[0] ^ b->u64[0]) & mask->u64[0]) ||
109 		((a->u64[1] ^ b->u64[1]) & mask->u64[1]);
110 }
111 
112 int efx_nic_test_registers(struct efx_nic *efx,
113 			   const struct efx_nic_register_test *regs,
114 			   size_t n_regs)
115 {
116 	unsigned address = 0, i, j;
117 	efx_oword_t mask, imask, original, reg, buf;
118 
119 	/* Falcon should be in loopback to isolate the XMAC from the PHY */
120 	WARN_ON(!LOOPBACK_INTERNAL(efx));
121 
122 	for (i = 0; i < n_regs; ++i) {
123 		address = regs[i].address;
124 		mask = imask = regs[i].mask;
125 		EFX_INVERT_OWORD(imask);
126 
127 		efx_reado(efx, &original, address);
128 
129 		/* bit sweep on and off */
130 		for (j = 0; j < 128; j++) {
131 			if (!EFX_EXTRACT_OWORD32(mask, j, j))
132 				continue;
133 
134 			/* Test this testable bit can be set in isolation */
135 			EFX_AND_OWORD(reg, original, mask);
136 			EFX_SET_OWORD32(reg, j, j, 1);
137 
138 			efx_writeo(efx, &reg, address);
139 			efx_reado(efx, &buf, address);
140 
141 			if (efx_masked_compare_oword(&reg, &buf, &mask))
142 				goto fail;
143 
144 			/* Test this testable bit can be cleared in isolation */
145 			EFX_OR_OWORD(reg, original, mask);
146 			EFX_SET_OWORD32(reg, j, j, 0);
147 
148 			efx_writeo(efx, &reg, address);
149 			efx_reado(efx, &buf, address);
150 
151 			if (efx_masked_compare_oword(&reg, &buf, &mask))
152 				goto fail;
153 		}
154 
155 		efx_writeo(efx, &original, address);
156 	}
157 
158 	return 0;
159 
160 fail:
161 	netif_err(efx, hw, efx->net_dev,
162 		  "wrote "EFX_OWORD_FMT" read "EFX_OWORD_FMT
163 		  " at address 0x%x mask "EFX_OWORD_FMT"\n", EFX_OWORD_VAL(reg),
164 		  EFX_OWORD_VAL(buf), address, EFX_OWORD_VAL(mask));
165 	return -EIO;
166 }
167 
168 /**************************************************************************
169  *
170  * Special buffer handling
171  * Special buffers are used for event queues and the TX and RX
172  * descriptor rings.
173  *
174  *************************************************************************/
175 
176 /*
177  * Initialise a special buffer
178  *
179  * This will define a buffer (previously allocated via
180  * efx_alloc_special_buffer()) in the buffer table, allowing
181  * it to be used for event queues, descriptor rings etc.
182  */
183 static void
184 efx_init_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
185 {
186 	efx_qword_t buf_desc;
187 	int index;
188 	dma_addr_t dma_addr;
189 	int i;
190 
191 	EFX_BUG_ON_PARANOID(!buffer->addr);
192 
193 	/* Write buffer descriptors to NIC */
194 	for (i = 0; i < buffer->entries; i++) {
195 		index = buffer->index + i;
196 		dma_addr = buffer->dma_addr + (i * EFX_BUF_SIZE);
197 		netif_dbg(efx, probe, efx->net_dev,
198 			  "mapping special buffer %d at %llx\n",
199 			  index, (unsigned long long)dma_addr);
200 		EFX_POPULATE_QWORD_3(buf_desc,
201 				     FRF_AZ_BUF_ADR_REGION, 0,
202 				     FRF_AZ_BUF_ADR_FBUF, dma_addr >> 12,
203 				     FRF_AZ_BUF_OWNER_ID_FBUF, 0);
204 		efx_write_buf_tbl(efx, &buf_desc, index);
205 	}
206 }
207 
208 /* Unmaps a buffer and clears the buffer table entries */
209 static void
210 efx_fini_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
211 {
212 	efx_oword_t buf_tbl_upd;
213 	unsigned int start = buffer->index;
214 	unsigned int end = (buffer->index + buffer->entries - 1);
215 
216 	if (!buffer->entries)
217 		return;
218 
219 	netif_dbg(efx, hw, efx->net_dev, "unmapping special buffers %d-%d\n",
220 		  buffer->index, buffer->index + buffer->entries - 1);
221 
222 	EFX_POPULATE_OWORD_4(buf_tbl_upd,
223 			     FRF_AZ_BUF_UPD_CMD, 0,
224 			     FRF_AZ_BUF_CLR_CMD, 1,
225 			     FRF_AZ_BUF_CLR_END_ID, end,
226 			     FRF_AZ_BUF_CLR_START_ID, start);
227 	efx_writeo(efx, &buf_tbl_upd, FR_AZ_BUF_TBL_UPD);
228 }
229 
230 /*
231  * Allocate a new special buffer
232  *
233  * This allocates memory for a new buffer, clears it and allocates a
234  * new buffer ID range.  It does not write into the buffer table.
235  *
236  * This call will allocate 4KB buffers, since 8KB buffers can't be
237  * used for event queues and descriptor rings.
238  */
239 static int efx_alloc_special_buffer(struct efx_nic *efx,
240 				    struct efx_special_buffer *buffer,
241 				    unsigned int len)
242 {
243 	len = ALIGN(len, EFX_BUF_SIZE);
244 
245 	buffer->addr = dma_alloc_coherent(&efx->pci_dev->dev, len,
246 					  &buffer->dma_addr, GFP_KERNEL);
247 	if (!buffer->addr)
248 		return -ENOMEM;
249 	buffer->len = len;
250 	buffer->entries = len / EFX_BUF_SIZE;
251 	BUG_ON(buffer->dma_addr & (EFX_BUF_SIZE - 1));
252 
253 	/* All zeros is a potentially valid event so memset to 0xff */
254 	memset(buffer->addr, 0xff, len);
255 
256 	/* Select new buffer ID */
257 	buffer->index = efx->next_buffer_table;
258 	efx->next_buffer_table += buffer->entries;
259 
260 	netif_dbg(efx, probe, efx->net_dev,
261 		  "allocating special buffers %d-%d at %llx+%x "
262 		  "(virt %p phys %llx)\n", buffer->index,
263 		  buffer->index + buffer->entries - 1,
264 		  (u64)buffer->dma_addr, len,
265 		  buffer->addr, (u64)virt_to_phys(buffer->addr));
266 
267 	return 0;
268 }
269 
270 static void
271 efx_free_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
272 {
273 	if (!buffer->addr)
274 		return;
275 
276 	netif_dbg(efx, hw, efx->net_dev,
277 		  "deallocating special buffers %d-%d at %llx+%x "
278 		  "(virt %p phys %llx)\n", buffer->index,
279 		  buffer->index + buffer->entries - 1,
280 		  (u64)buffer->dma_addr, buffer->len,
281 		  buffer->addr, (u64)virt_to_phys(buffer->addr));
282 
283 	dma_free_coherent(&efx->pci_dev->dev, buffer->len, buffer->addr,
284 			  buffer->dma_addr);
285 	buffer->addr = NULL;
286 	buffer->entries = 0;
287 }
288 
289 /**************************************************************************
290  *
291  * Generic buffer handling
292  * These buffers are used for interrupt status and MAC stats
293  *
294  **************************************************************************/
295 
296 int efx_nic_alloc_buffer(struct efx_nic *efx, struct efx_buffer *buffer,
297 			 unsigned int len)
298 {
299 	buffer->addr = pci_alloc_consistent(efx->pci_dev, len,
300 					    &buffer->dma_addr);
301 	if (!buffer->addr)
302 		return -ENOMEM;
303 	buffer->len = len;
304 	memset(buffer->addr, 0, len);
305 	return 0;
306 }
307 
308 void efx_nic_free_buffer(struct efx_nic *efx, struct efx_buffer *buffer)
309 {
310 	if (buffer->addr) {
311 		pci_free_consistent(efx->pci_dev, buffer->len,
312 				    buffer->addr, buffer->dma_addr);
313 		buffer->addr = NULL;
314 	}
315 }
316 
317 /**************************************************************************
318  *
319  * TX path
320  *
321  **************************************************************************/
322 
323 /* Returns a pointer to the specified transmit descriptor in the TX
324  * descriptor queue belonging to the specified channel.
325  */
326 static inline efx_qword_t *
327 efx_tx_desc(struct efx_tx_queue *tx_queue, unsigned int index)
328 {
329 	return ((efx_qword_t *) (tx_queue->txd.addr)) + index;
330 }
331 
332 /* This writes to the TX_DESC_WPTR; write pointer for TX descriptor ring */
333 static inline void efx_notify_tx_desc(struct efx_tx_queue *tx_queue)
334 {
335 	unsigned write_ptr;
336 	efx_dword_t reg;
337 
338 	write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
339 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_TX_DESC_WPTR_DWORD, write_ptr);
340 	efx_writed_page(tx_queue->efx, &reg,
341 			FR_AZ_TX_DESC_UPD_DWORD_P0, tx_queue->queue);
342 }
343 
344 /* Write pointer and first descriptor for TX descriptor ring */
345 static inline void efx_push_tx_desc(struct efx_tx_queue *tx_queue,
346 				    const efx_qword_t *txd)
347 {
348 	unsigned write_ptr;
349 	efx_oword_t reg;
350 
351 	BUILD_BUG_ON(FRF_AZ_TX_DESC_LBN != 0);
352 	BUILD_BUG_ON(FR_AA_TX_DESC_UPD_KER != FR_BZ_TX_DESC_UPD_P0);
353 
354 	write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
355 	EFX_POPULATE_OWORD_2(reg, FRF_AZ_TX_DESC_PUSH_CMD, true,
356 			     FRF_AZ_TX_DESC_WPTR, write_ptr);
357 	reg.qword[0] = *txd;
358 	efx_writeo_page(tx_queue->efx, &reg,
359 			FR_BZ_TX_DESC_UPD_P0, tx_queue->queue);
360 }
361 
362 static inline bool
363 efx_may_push_tx_desc(struct efx_tx_queue *tx_queue, unsigned int write_count)
364 {
365 	unsigned empty_read_count = ACCESS_ONCE(tx_queue->empty_read_count);
366 
367 	if (empty_read_count == 0)
368 		return false;
369 
370 	tx_queue->empty_read_count = 0;
371 	return ((empty_read_count ^ write_count) & ~EFX_EMPTY_COUNT_VALID) == 0;
372 }
373 
374 /* For each entry inserted into the software descriptor ring, create a
375  * descriptor in the hardware TX descriptor ring (in host memory), and
376  * write a doorbell.
377  */
378 void efx_nic_push_buffers(struct efx_tx_queue *tx_queue)
379 {
380 
381 	struct efx_tx_buffer *buffer;
382 	efx_qword_t *txd;
383 	unsigned write_ptr;
384 	unsigned old_write_count = tx_queue->write_count;
385 
386 	BUG_ON(tx_queue->write_count == tx_queue->insert_count);
387 
388 	do {
389 		write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
390 		buffer = &tx_queue->buffer[write_ptr];
391 		txd = efx_tx_desc(tx_queue, write_ptr);
392 		++tx_queue->write_count;
393 
394 		/* Create TX descriptor ring entry */
395 		EFX_POPULATE_QWORD_4(*txd,
396 				     FSF_AZ_TX_KER_CONT, buffer->continuation,
397 				     FSF_AZ_TX_KER_BYTE_COUNT, buffer->len,
398 				     FSF_AZ_TX_KER_BUF_REGION, 0,
399 				     FSF_AZ_TX_KER_BUF_ADDR, buffer->dma_addr);
400 	} while (tx_queue->write_count != tx_queue->insert_count);
401 
402 	wmb(); /* Ensure descriptors are written before they are fetched */
403 
404 	if (efx_may_push_tx_desc(tx_queue, old_write_count)) {
405 		txd = efx_tx_desc(tx_queue,
406 				  old_write_count & tx_queue->ptr_mask);
407 		efx_push_tx_desc(tx_queue, txd);
408 		++tx_queue->pushes;
409 	} else {
410 		efx_notify_tx_desc(tx_queue);
411 	}
412 }
413 
414 /* Allocate hardware resources for a TX queue */
415 int efx_nic_probe_tx(struct efx_tx_queue *tx_queue)
416 {
417 	struct efx_nic *efx = tx_queue->efx;
418 	unsigned entries;
419 
420 	entries = tx_queue->ptr_mask + 1;
421 	return efx_alloc_special_buffer(efx, &tx_queue->txd,
422 					entries * sizeof(efx_qword_t));
423 }
424 
425 void efx_nic_init_tx(struct efx_tx_queue *tx_queue)
426 {
427 	struct efx_nic *efx = tx_queue->efx;
428 	efx_oword_t reg;
429 
430 	tx_queue->flushed = FLUSH_NONE;
431 
432 	/* Pin TX descriptor ring */
433 	efx_init_special_buffer(efx, &tx_queue->txd);
434 
435 	/* Push TX descriptor ring to card */
436 	EFX_POPULATE_OWORD_10(reg,
437 			      FRF_AZ_TX_DESCQ_EN, 1,
438 			      FRF_AZ_TX_ISCSI_DDIG_EN, 0,
439 			      FRF_AZ_TX_ISCSI_HDIG_EN, 0,
440 			      FRF_AZ_TX_DESCQ_BUF_BASE_ID, tx_queue->txd.index,
441 			      FRF_AZ_TX_DESCQ_EVQ_ID,
442 			      tx_queue->channel->channel,
443 			      FRF_AZ_TX_DESCQ_OWNER_ID, 0,
444 			      FRF_AZ_TX_DESCQ_LABEL, tx_queue->queue,
445 			      FRF_AZ_TX_DESCQ_SIZE,
446 			      __ffs(tx_queue->txd.entries),
447 			      FRF_AZ_TX_DESCQ_TYPE, 0,
448 			      FRF_BZ_TX_NON_IP_DROP_DIS, 1);
449 
450 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
451 		int csum = tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD;
452 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_IP_CHKSM_DIS, !csum);
453 		EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_TCP_CHKSM_DIS,
454 				    !csum);
455 	}
456 
457 	efx_writeo_table(efx, &reg, efx->type->txd_ptr_tbl_base,
458 			 tx_queue->queue);
459 
460 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0) {
461 		/* Only 128 bits in this register */
462 		BUILD_BUG_ON(EFX_MAX_TX_QUEUES > 128);
463 
464 		efx_reado(efx, &reg, FR_AA_TX_CHKSM_CFG);
465 		if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD)
466 			clear_bit_le(tx_queue->queue, (void *)&reg);
467 		else
468 			set_bit_le(tx_queue->queue, (void *)&reg);
469 		efx_writeo(efx, &reg, FR_AA_TX_CHKSM_CFG);
470 	}
471 
472 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
473 		EFX_POPULATE_OWORD_1(reg,
474 				     FRF_BZ_TX_PACE,
475 				     (tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
476 				     FFE_BZ_TX_PACE_OFF :
477 				     FFE_BZ_TX_PACE_RESERVED);
478 		efx_writeo_table(efx, &reg, FR_BZ_TX_PACE_TBL,
479 				 tx_queue->queue);
480 	}
481 }
482 
483 static void efx_flush_tx_queue(struct efx_tx_queue *tx_queue)
484 {
485 	struct efx_nic *efx = tx_queue->efx;
486 	efx_oword_t tx_flush_descq;
487 
488 	tx_queue->flushed = FLUSH_PENDING;
489 
490 	/* Post a flush command */
491 	EFX_POPULATE_OWORD_2(tx_flush_descq,
492 			     FRF_AZ_TX_FLUSH_DESCQ_CMD, 1,
493 			     FRF_AZ_TX_FLUSH_DESCQ, tx_queue->queue);
494 	efx_writeo(efx, &tx_flush_descq, FR_AZ_TX_FLUSH_DESCQ);
495 }
496 
497 void efx_nic_fini_tx(struct efx_tx_queue *tx_queue)
498 {
499 	struct efx_nic *efx = tx_queue->efx;
500 	efx_oword_t tx_desc_ptr;
501 
502 	/* The queue should have been flushed */
503 	WARN_ON(tx_queue->flushed != FLUSH_DONE);
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 	rx_queue->flushed = FLUSH_NONE;
596 
597 	/* Pin RX descriptor ring */
598 	efx_init_special_buffer(efx, &rx_queue->rxd);
599 
600 	/* Push RX descriptor ring to card */
601 	EFX_POPULATE_OWORD_10(rx_desc_ptr,
602 			      FRF_AZ_RX_ISCSI_DDIG_EN, iscsi_digest_en,
603 			      FRF_AZ_RX_ISCSI_HDIG_EN, iscsi_digest_en,
604 			      FRF_AZ_RX_DESCQ_BUF_BASE_ID, rx_queue->rxd.index,
605 			      FRF_AZ_RX_DESCQ_EVQ_ID,
606 			      efx_rx_queue_channel(rx_queue)->channel,
607 			      FRF_AZ_RX_DESCQ_OWNER_ID, 0,
608 			      FRF_AZ_RX_DESCQ_LABEL,
609 			      efx_rx_queue_index(rx_queue),
610 			      FRF_AZ_RX_DESCQ_SIZE,
611 			      __ffs(rx_queue->rxd.entries),
612 			      FRF_AZ_RX_DESCQ_TYPE, 0 /* kernel queue */ ,
613 			      /* For >=B0 this is scatter so disable */
614 			      FRF_AZ_RX_DESCQ_JUMBO, !is_b0,
615 			      FRF_AZ_RX_DESCQ_EN, 1);
616 	efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
617 			 efx_rx_queue_index(rx_queue));
618 }
619 
620 static void efx_flush_rx_queue(struct efx_rx_queue *rx_queue)
621 {
622 	struct efx_nic *efx = rx_queue->efx;
623 	efx_oword_t rx_flush_descq;
624 
625 	rx_queue->flushed = FLUSH_PENDING;
626 
627 	/* Post a flush command */
628 	EFX_POPULATE_OWORD_2(rx_flush_descq,
629 			     FRF_AZ_RX_FLUSH_DESCQ_CMD, 1,
630 			     FRF_AZ_RX_FLUSH_DESCQ,
631 			     efx_rx_queue_index(rx_queue));
632 	efx_writeo(efx, &rx_flush_descq, FR_AZ_RX_FLUSH_DESCQ);
633 }
634 
635 void efx_nic_fini_rx(struct efx_rx_queue *rx_queue)
636 {
637 	efx_oword_t rx_desc_ptr;
638 	struct efx_nic *efx = rx_queue->efx;
639 
640 	/* The queue should already have been flushed */
641 	WARN_ON(rx_queue->flushed != FLUSH_DONE);
642 
643 	/* Remove RX descriptor ring from card */
644 	EFX_ZERO_OWORD(rx_desc_ptr);
645 	efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
646 			 efx_rx_queue_index(rx_queue));
647 
648 	/* Unpin RX descriptor ring */
649 	efx_fini_special_buffer(efx, &rx_queue->rxd);
650 }
651 
652 /* Free buffers backing RX queue */
653 void efx_nic_remove_rx(struct efx_rx_queue *rx_queue)
654 {
655 	efx_free_special_buffer(rx_queue->efx, &rx_queue->rxd);
656 }
657 
658 /**************************************************************************
659  *
660  * Event queue processing
661  * Event queues are processed by per-channel tasklets.
662  *
663  **************************************************************************/
664 
665 /* Update a channel's event queue's read pointer (RPTR) register
666  *
667  * This writes the EVQ_RPTR_REG register for the specified channel's
668  * event queue.
669  */
670 void efx_nic_eventq_read_ack(struct efx_channel *channel)
671 {
672 	efx_dword_t reg;
673 	struct efx_nic *efx = channel->efx;
674 
675 	EFX_POPULATE_DWORD_1(reg, FRF_AZ_EVQ_RPTR,
676 			     channel->eventq_read_ptr & channel->eventq_mask);
677 	efx_writed_table(efx, &reg, efx->type->evq_rptr_tbl_base,
678 			 channel->channel);
679 }
680 
681 /* Use HW to insert a SW defined event */
682 static void efx_generate_event(struct efx_channel *channel, efx_qword_t *event)
683 {
684 	efx_oword_t drv_ev_reg;
685 
686 	BUILD_BUG_ON(FRF_AZ_DRV_EV_DATA_LBN != 0 ||
687 		     FRF_AZ_DRV_EV_DATA_WIDTH != 64);
688 	drv_ev_reg.u32[0] = event->u32[0];
689 	drv_ev_reg.u32[1] = event->u32[1];
690 	drv_ev_reg.u32[2] = 0;
691 	drv_ev_reg.u32[3] = 0;
692 	EFX_SET_OWORD_FIELD(drv_ev_reg, FRF_AZ_DRV_EV_QID, channel->channel);
693 	efx_writeo(channel->efx, &drv_ev_reg, FR_AZ_DRV_EV);
694 }
695 
696 /* Handle a transmit completion event
697  *
698  * The NIC batches TX completion events; the message we receive is of
699  * the form "complete all TX events up to this index".
700  */
701 static int
702 efx_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
703 {
704 	unsigned int tx_ev_desc_ptr;
705 	unsigned int tx_ev_q_label;
706 	struct efx_tx_queue *tx_queue;
707 	struct efx_nic *efx = channel->efx;
708 	int tx_packets = 0;
709 
710 	if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
711 		/* Transmit completion */
712 		tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
713 		tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
714 		tx_queue = efx_channel_get_tx_queue(
715 			channel, tx_ev_q_label % EFX_TXQ_TYPES);
716 		tx_packets = ((tx_ev_desc_ptr - tx_queue->read_count) &
717 			      tx_queue->ptr_mask);
718 		channel->irq_mod_score += tx_packets;
719 		efx_xmit_done(tx_queue, tx_ev_desc_ptr);
720 	} else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
721 		/* Rewrite the FIFO write pointer */
722 		tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
723 		tx_queue = efx_channel_get_tx_queue(
724 			channel, tx_ev_q_label % EFX_TXQ_TYPES);
725 
726 		netif_tx_lock(efx->net_dev);
727 		efx_notify_tx_desc(tx_queue);
728 		netif_tx_unlock(efx->net_dev);
729 	} else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR) &&
730 		   EFX_WORKAROUND_10727(efx)) {
731 		efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
732 	} else {
733 		netif_err(efx, tx_err, efx->net_dev,
734 			  "channel %d unexpected TX event "
735 			  EFX_QWORD_FMT"\n", channel->channel,
736 			  EFX_QWORD_VAL(*event));
737 	}
738 
739 	return tx_packets;
740 }
741 
742 /* Detect errors included in the rx_evt_pkt_ok bit. */
743 static u16 efx_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
744 				const efx_qword_t *event)
745 {
746 	struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
747 	struct efx_nic *efx = rx_queue->efx;
748 	bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
749 	bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
750 	bool rx_ev_frm_trunc, rx_ev_drib_nib, rx_ev_tobe_disc;
751 	bool rx_ev_other_err, rx_ev_pause_frm;
752 	bool rx_ev_hdr_type, rx_ev_mcast_pkt;
753 	unsigned rx_ev_pkt_type;
754 
755 	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
756 	rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
757 	rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
758 	rx_ev_pkt_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_TYPE);
759 	rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
760 						 FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
761 	rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
762 						  FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
763 	rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
764 						   FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
765 	rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
766 	rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
767 	rx_ev_drib_nib = ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) ?
768 			  0 : EFX_QWORD_FIELD(*event, FSF_AA_RX_EV_DRIB_NIB));
769 	rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
770 
771 	/* Every error apart from tobe_disc and pause_frm */
772 	rx_ev_other_err = (rx_ev_drib_nib | rx_ev_tcp_udp_chksum_err |
773 			   rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
774 			   rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
775 
776 	/* Count errors that are not in MAC stats.  Ignore expected
777 	 * checksum errors during self-test. */
778 	if (rx_ev_frm_trunc)
779 		++channel->n_rx_frm_trunc;
780 	else if (rx_ev_tobe_disc)
781 		++channel->n_rx_tobe_disc;
782 	else if (!efx->loopback_selftest) {
783 		if (rx_ev_ip_hdr_chksum_err)
784 			++channel->n_rx_ip_hdr_chksum_err;
785 		else if (rx_ev_tcp_udp_chksum_err)
786 			++channel->n_rx_tcp_udp_chksum_err;
787 	}
788 
789 	/* TOBE_DISC is expected on unicast mismatches; don't print out an
790 	 * error message.  FRM_TRUNC indicates RXDP dropped the packet due
791 	 * to a FIFO overflow.
792 	 */
793 #ifdef DEBUG
794 	if (rx_ev_other_err && net_ratelimit()) {
795 		netif_dbg(efx, rx_err, efx->net_dev,
796 			  " RX queue %d unexpected RX event "
797 			  EFX_QWORD_FMT "%s%s%s%s%s%s%s%s\n",
798 			  efx_rx_queue_index(rx_queue), EFX_QWORD_VAL(*event),
799 			  rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
800 			  rx_ev_ip_hdr_chksum_err ?
801 			  " [IP_HDR_CHKSUM_ERR]" : "",
802 			  rx_ev_tcp_udp_chksum_err ?
803 			  " [TCP_UDP_CHKSUM_ERR]" : "",
804 			  rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
805 			  rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
806 			  rx_ev_drib_nib ? " [DRIB_NIB]" : "",
807 			  rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
808 			  rx_ev_pause_frm ? " [PAUSE]" : "");
809 	}
810 #endif
811 
812 	/* The frame must be discarded if any of these are true. */
813 	return (rx_ev_eth_crc_err | rx_ev_frm_trunc | rx_ev_drib_nib |
814 		rx_ev_tobe_disc | rx_ev_pause_frm) ?
815 		EFX_RX_PKT_DISCARD : 0;
816 }
817 
818 /* Handle receive events that are not in-order. */
819 static void
820 efx_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
821 {
822 	struct efx_nic *efx = rx_queue->efx;
823 	unsigned expected, dropped;
824 
825 	expected = rx_queue->removed_count & rx_queue->ptr_mask;
826 	dropped = (index - expected) & rx_queue->ptr_mask;
827 	netif_info(efx, rx_err, efx->net_dev,
828 		   "dropped %d events (index=%d expected=%d)\n",
829 		   dropped, index, expected);
830 
831 	efx_schedule_reset(efx, EFX_WORKAROUND_5676(efx) ?
832 			   RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
833 }
834 
835 /* Handle a packet received event
836  *
837  * The NIC gives a "discard" flag if it's a unicast packet with the
838  * wrong destination address
839  * Also "is multicast" and "matches multicast filter" flags can be used to
840  * discard non-matching multicast packets.
841  */
842 static void
843 efx_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
844 {
845 	unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
846 	unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
847 	unsigned expected_ptr;
848 	bool rx_ev_pkt_ok;
849 	u16 flags;
850 	struct efx_rx_queue *rx_queue;
851 
852 	/* Basic packet information */
853 	rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
854 	rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
855 	rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
856 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT));
857 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP) != 1);
858 	WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
859 		channel->channel);
860 
861 	rx_queue = efx_channel_get_rx_queue(channel);
862 
863 	rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
864 	expected_ptr = rx_queue->removed_count & rx_queue->ptr_mask;
865 	if (unlikely(rx_ev_desc_ptr != expected_ptr))
866 		efx_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr);
867 
868 	if (likely(rx_ev_pkt_ok)) {
869 		/* If packet is marked as OK and packet type is TCP/IP or
870 		 * UDP/IP, then we can rely on the hardware checksum.
871 		 */
872 		flags = (rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP ||
873 			 rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP) ?
874 			EFX_RX_PKT_CSUMMED : 0;
875 	} else {
876 		flags = efx_handle_rx_not_ok(rx_queue, event);
877 	}
878 
879 	/* Detect multicast packets that didn't match the filter */
880 	rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
881 	if (rx_ev_mcast_pkt) {
882 		unsigned int rx_ev_mcast_hash_match =
883 			EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
884 
885 		if (unlikely(!rx_ev_mcast_hash_match)) {
886 			++channel->n_rx_mcast_mismatch;
887 			flags |= EFX_RX_PKT_DISCARD;
888 		}
889 	}
890 
891 	channel->irq_mod_score += 2;
892 
893 	/* Handle received packet */
894 	efx_rx_packet(rx_queue, rx_ev_desc_ptr, rx_ev_byte_cnt, flags);
895 }
896 
897 static void
898 efx_handle_generated_event(struct efx_channel *channel, efx_qword_t *event)
899 {
900 	struct efx_nic *efx = channel->efx;
901 	unsigned code;
902 
903 	code = EFX_QWORD_FIELD(*event, FSF_AZ_DRV_GEN_EV_MAGIC);
904 	if (code == EFX_CHANNEL_MAGIC_TEST(channel))
905 		; /* ignore */
906 	else if (code == EFX_CHANNEL_MAGIC_FILL(channel))
907 		/* The queue must be empty, so we won't receive any rx
908 		 * events, so efx_process_channel() won't refill the
909 		 * queue. Refill it here */
910 		efx_fast_push_rx_descriptors(efx_channel_get_rx_queue(channel));
911 	else
912 		netif_dbg(efx, hw, efx->net_dev, "channel %d received "
913 			  "generated event "EFX_QWORD_FMT"\n",
914 			  channel->channel, EFX_QWORD_VAL(*event));
915 }
916 
917 static void
918 efx_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
919 {
920 	struct efx_nic *efx = channel->efx;
921 	unsigned int ev_sub_code;
922 	unsigned int ev_sub_data;
923 
924 	ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
925 	ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
926 
927 	switch (ev_sub_code) {
928 	case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
929 		netif_vdbg(efx, hw, efx->net_dev, "channel %d TXQ %d flushed\n",
930 			   channel->channel, ev_sub_data);
931 		break;
932 	case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
933 		netif_vdbg(efx, hw, efx->net_dev, "channel %d RXQ %d flushed\n",
934 			   channel->channel, ev_sub_data);
935 		break;
936 	case FSE_AZ_EVQ_INIT_DONE_EV:
937 		netif_dbg(efx, hw, efx->net_dev,
938 			  "channel %d EVQ %d initialised\n",
939 			  channel->channel, ev_sub_data);
940 		break;
941 	case FSE_AZ_SRM_UPD_DONE_EV:
942 		netif_vdbg(efx, hw, efx->net_dev,
943 			   "channel %d SRAM update done\n", channel->channel);
944 		break;
945 	case FSE_AZ_WAKE_UP_EV:
946 		netif_vdbg(efx, hw, efx->net_dev,
947 			   "channel %d RXQ %d wakeup event\n",
948 			   channel->channel, ev_sub_data);
949 		break;
950 	case FSE_AZ_TIMER_EV:
951 		netif_vdbg(efx, hw, efx->net_dev,
952 			   "channel %d RX queue %d timer expired\n",
953 			   channel->channel, ev_sub_data);
954 		break;
955 	case FSE_AA_RX_RECOVER_EV:
956 		netif_err(efx, rx_err, efx->net_dev,
957 			  "channel %d seen DRIVER RX_RESET event. "
958 			"Resetting.\n", channel->channel);
959 		atomic_inc(&efx->rx_reset);
960 		efx_schedule_reset(efx,
961 				   EFX_WORKAROUND_6555(efx) ?
962 				   RESET_TYPE_RX_RECOVERY :
963 				   RESET_TYPE_DISABLE);
964 		break;
965 	case FSE_BZ_RX_DSC_ERROR_EV:
966 		netif_err(efx, rx_err, efx->net_dev,
967 			  "RX DMA Q %d reports descriptor fetch error."
968 			  " RX Q %d is disabled.\n", ev_sub_data, ev_sub_data);
969 		efx_schedule_reset(efx, RESET_TYPE_RX_DESC_FETCH);
970 		break;
971 	case FSE_BZ_TX_DSC_ERROR_EV:
972 		netif_err(efx, tx_err, efx->net_dev,
973 			  "TX DMA Q %d reports descriptor fetch error."
974 			  " TX Q %d is disabled.\n", ev_sub_data, ev_sub_data);
975 		efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
976 		break;
977 	default:
978 		netif_vdbg(efx, hw, efx->net_dev,
979 			   "channel %d unknown driver event code %d "
980 			   "data %04x\n", channel->channel, ev_sub_code,
981 			   ev_sub_data);
982 		break;
983 	}
984 }
985 
986 int efx_nic_process_eventq(struct efx_channel *channel, int budget)
987 {
988 	struct efx_nic *efx = channel->efx;
989 	unsigned int read_ptr;
990 	efx_qword_t event, *p_event;
991 	int ev_code;
992 	int tx_packets = 0;
993 	int spent = 0;
994 
995 	read_ptr = channel->eventq_read_ptr;
996 
997 	for (;;) {
998 		p_event = efx_event(channel, read_ptr);
999 		event = *p_event;
1000 
1001 		if (!efx_event_present(&event))
1002 			/* End of events */
1003 			break;
1004 
1005 		netif_vdbg(channel->efx, intr, channel->efx->net_dev,
1006 			   "channel %d event is "EFX_QWORD_FMT"\n",
1007 			   channel->channel, EFX_QWORD_VAL(event));
1008 
1009 		/* Clear this event by marking it all ones */
1010 		EFX_SET_QWORD(*p_event);
1011 
1012 		++read_ptr;
1013 
1014 		ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
1015 
1016 		switch (ev_code) {
1017 		case FSE_AZ_EV_CODE_RX_EV:
1018 			efx_handle_rx_event(channel, &event);
1019 			if (++spent == budget)
1020 				goto out;
1021 			break;
1022 		case FSE_AZ_EV_CODE_TX_EV:
1023 			tx_packets += efx_handle_tx_event(channel, &event);
1024 			if (tx_packets > efx->txq_entries) {
1025 				spent = budget;
1026 				goto out;
1027 			}
1028 			break;
1029 		case FSE_AZ_EV_CODE_DRV_GEN_EV:
1030 			efx_handle_generated_event(channel, &event);
1031 			break;
1032 		case FSE_AZ_EV_CODE_DRIVER_EV:
1033 			efx_handle_driver_event(channel, &event);
1034 			break;
1035 		case FSE_CZ_EV_CODE_MCDI_EV:
1036 			efx_mcdi_process_event(channel, &event);
1037 			break;
1038 		case FSE_AZ_EV_CODE_GLOBAL_EV:
1039 			if (efx->type->handle_global_event &&
1040 			    efx->type->handle_global_event(channel, &event))
1041 				break;
1042 			/* else fall through */
1043 		default:
1044 			netif_err(channel->efx, hw, channel->efx->net_dev,
1045 				  "channel %d unknown event type %d (data "
1046 				  EFX_QWORD_FMT ")\n", channel->channel,
1047 				  ev_code, EFX_QWORD_VAL(event));
1048 		}
1049 	}
1050 
1051 out:
1052 	channel->eventq_read_ptr = read_ptr;
1053 	return spent;
1054 }
1055 
1056 /* Check whether an event is present in the eventq at the current
1057  * read pointer.  Only useful for self-test.
1058  */
1059 bool efx_nic_event_present(struct efx_channel *channel)
1060 {
1061 	return efx_event_present(efx_event(channel, channel->eventq_read_ptr));
1062 }
1063 
1064 /* Allocate buffer table entries for event queue */
1065 int efx_nic_probe_eventq(struct efx_channel *channel)
1066 {
1067 	struct efx_nic *efx = channel->efx;
1068 	unsigned entries;
1069 
1070 	entries = channel->eventq_mask + 1;
1071 	return efx_alloc_special_buffer(efx, &channel->eventq,
1072 					entries * sizeof(efx_qword_t));
1073 }
1074 
1075 void efx_nic_init_eventq(struct efx_channel *channel)
1076 {
1077 	efx_oword_t reg;
1078 	struct efx_nic *efx = channel->efx;
1079 
1080 	netif_dbg(efx, hw, efx->net_dev,
1081 		  "channel %d event queue in special buffers %d-%d\n",
1082 		  channel->channel, channel->eventq.index,
1083 		  channel->eventq.index + channel->eventq.entries - 1);
1084 
1085 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1086 		EFX_POPULATE_OWORD_3(reg,
1087 				     FRF_CZ_TIMER_Q_EN, 1,
1088 				     FRF_CZ_HOST_NOTIFY_MODE, 0,
1089 				     FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
1090 		efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1091 	}
1092 
1093 	/* Pin event queue buffer */
1094 	efx_init_special_buffer(efx, &channel->eventq);
1095 
1096 	/* Fill event queue with all ones (i.e. empty events) */
1097 	memset(channel->eventq.addr, 0xff, channel->eventq.len);
1098 
1099 	/* Push event queue to card */
1100 	EFX_POPULATE_OWORD_3(reg,
1101 			     FRF_AZ_EVQ_EN, 1,
1102 			     FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
1103 			     FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
1104 	efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1105 			 channel->channel);
1106 
1107 	efx->type->push_irq_moderation(channel);
1108 }
1109 
1110 void efx_nic_fini_eventq(struct efx_channel *channel)
1111 {
1112 	efx_oword_t reg;
1113 	struct efx_nic *efx = channel->efx;
1114 
1115 	/* Remove event queue from card */
1116 	EFX_ZERO_OWORD(reg);
1117 	efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1118 			 channel->channel);
1119 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1120 		efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1121 
1122 	/* Unpin event queue */
1123 	efx_fini_special_buffer(efx, &channel->eventq);
1124 }
1125 
1126 /* Free buffers backing event queue */
1127 void efx_nic_remove_eventq(struct efx_channel *channel)
1128 {
1129 	efx_free_special_buffer(channel->efx, &channel->eventq);
1130 }
1131 
1132 
1133 void efx_nic_generate_test_event(struct efx_channel *channel)
1134 {
1135 	unsigned int magic = EFX_CHANNEL_MAGIC_TEST(channel);
1136 	efx_qword_t test_event;
1137 
1138 	EFX_POPULATE_QWORD_2(test_event, FSF_AZ_EV_CODE,
1139 			     FSE_AZ_EV_CODE_DRV_GEN_EV,
1140 			     FSF_AZ_DRV_GEN_EV_MAGIC, magic);
1141 	efx_generate_event(channel, &test_event);
1142 }
1143 
1144 void efx_nic_generate_fill_event(struct efx_channel *channel)
1145 {
1146 	unsigned int magic = EFX_CHANNEL_MAGIC_FILL(channel);
1147 	efx_qword_t test_event;
1148 
1149 	EFX_POPULATE_QWORD_2(test_event, FSF_AZ_EV_CODE,
1150 			     FSE_AZ_EV_CODE_DRV_GEN_EV,
1151 			     FSF_AZ_DRV_GEN_EV_MAGIC, magic);
1152 	efx_generate_event(channel, &test_event);
1153 }
1154 
1155 /**************************************************************************
1156  *
1157  * Flush handling
1158  *
1159  **************************************************************************/
1160 
1161 
1162 static void efx_poll_flush_events(struct efx_nic *efx)
1163 {
1164 	struct efx_channel *channel = efx_get_channel(efx, 0);
1165 	struct efx_tx_queue *tx_queue;
1166 	struct efx_rx_queue *rx_queue;
1167 	unsigned int read_ptr = channel->eventq_read_ptr;
1168 	unsigned int end_ptr = read_ptr + channel->eventq_mask - 1;
1169 
1170 	do {
1171 		efx_qword_t *event = efx_event(channel, read_ptr);
1172 		int ev_code, ev_sub_code, ev_queue;
1173 		bool ev_failed;
1174 
1175 		if (!efx_event_present(event))
1176 			break;
1177 
1178 		ev_code = EFX_QWORD_FIELD(*event, FSF_AZ_EV_CODE);
1179 		ev_sub_code = EFX_QWORD_FIELD(*event,
1180 					      FSF_AZ_DRIVER_EV_SUBCODE);
1181 		if (ev_code == FSE_AZ_EV_CODE_DRIVER_EV &&
1182 		    ev_sub_code == FSE_AZ_TX_DESCQ_FLS_DONE_EV) {
1183 			ev_queue = EFX_QWORD_FIELD(*event,
1184 						   FSF_AZ_DRIVER_EV_SUBDATA);
1185 			if (ev_queue < EFX_TXQ_TYPES * efx->n_tx_channels) {
1186 				tx_queue = efx_get_tx_queue(
1187 					efx, ev_queue / EFX_TXQ_TYPES,
1188 					ev_queue % EFX_TXQ_TYPES);
1189 				tx_queue->flushed = FLUSH_DONE;
1190 			}
1191 		} else if (ev_code == FSE_AZ_EV_CODE_DRIVER_EV &&
1192 			   ev_sub_code == FSE_AZ_RX_DESCQ_FLS_DONE_EV) {
1193 			ev_queue = EFX_QWORD_FIELD(
1194 				*event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
1195 			ev_failed = EFX_QWORD_FIELD(
1196 				*event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
1197 			if (ev_queue < efx->n_rx_channels) {
1198 				rx_queue = efx_get_rx_queue(efx, ev_queue);
1199 				rx_queue->flushed =
1200 					ev_failed ? FLUSH_FAILED : FLUSH_DONE;
1201 			}
1202 		}
1203 
1204 		/* We're about to destroy the queue anyway, so
1205 		 * it's ok to throw away every non-flush event */
1206 		EFX_SET_QWORD(*event);
1207 
1208 		++read_ptr;
1209 	} while (read_ptr != end_ptr);
1210 
1211 	channel->eventq_read_ptr = read_ptr;
1212 }
1213 
1214 /* Handle tx and rx flushes at the same time, since they run in
1215  * parallel in the hardware and there's no reason for us to
1216  * serialise them */
1217 int efx_nic_flush_queues(struct efx_nic *efx)
1218 {
1219 	struct efx_channel *channel;
1220 	struct efx_rx_queue *rx_queue;
1221 	struct efx_tx_queue *tx_queue;
1222 	int i, tx_pending, rx_pending;
1223 
1224 	/* If necessary prepare the hardware for flushing */
1225 	efx->type->prepare_flush(efx);
1226 
1227 	/* Flush all tx queues in parallel */
1228 	efx_for_each_channel(channel, efx) {
1229 		efx_for_each_possible_channel_tx_queue(tx_queue, channel) {
1230 			if (tx_queue->initialised)
1231 				efx_flush_tx_queue(tx_queue);
1232 		}
1233 	}
1234 
1235 	/* The hardware supports four concurrent rx flushes, each of which may
1236 	 * need to be retried if there is an outstanding descriptor fetch */
1237 	for (i = 0; i < EFX_FLUSH_POLL_COUNT; ++i) {
1238 		rx_pending = tx_pending = 0;
1239 		efx_for_each_channel(channel, efx) {
1240 			efx_for_each_channel_rx_queue(rx_queue, channel) {
1241 				if (rx_queue->flushed == FLUSH_PENDING)
1242 					++rx_pending;
1243 			}
1244 		}
1245 		efx_for_each_channel(channel, efx) {
1246 			efx_for_each_channel_rx_queue(rx_queue, channel) {
1247 				if (rx_pending == EFX_RX_FLUSH_COUNT)
1248 					break;
1249 				if (rx_queue->flushed == FLUSH_FAILED ||
1250 				    rx_queue->flushed == FLUSH_NONE) {
1251 					efx_flush_rx_queue(rx_queue);
1252 					++rx_pending;
1253 				}
1254 			}
1255 			efx_for_each_possible_channel_tx_queue(tx_queue, channel) {
1256 				if (tx_queue->initialised &&
1257 				    tx_queue->flushed != FLUSH_DONE)
1258 					++tx_pending;
1259 			}
1260 		}
1261 
1262 		if (rx_pending == 0 && tx_pending == 0)
1263 			return 0;
1264 
1265 		msleep(EFX_FLUSH_INTERVAL);
1266 		efx_poll_flush_events(efx);
1267 	}
1268 
1269 	/* Mark the queues as all flushed. We're going to return failure
1270 	 * leading to a reset, or fake up success anyway */
1271 	efx_for_each_channel(channel, efx) {
1272 		efx_for_each_possible_channel_tx_queue(tx_queue, channel) {
1273 			if (tx_queue->initialised &&
1274 			    tx_queue->flushed != FLUSH_DONE)
1275 				netif_err(efx, hw, efx->net_dev,
1276 					  "tx queue %d flush command timed out\n",
1277 					  tx_queue->queue);
1278 			tx_queue->flushed = FLUSH_DONE;
1279 		}
1280 		efx_for_each_channel_rx_queue(rx_queue, channel) {
1281 			if (rx_queue->flushed != FLUSH_DONE)
1282 				netif_err(efx, hw, efx->net_dev,
1283 					  "rx queue %d flush command timed out\n",
1284 					  efx_rx_queue_index(rx_queue));
1285 			rx_queue->flushed = FLUSH_DONE;
1286 		}
1287 	}
1288 
1289 	return -ETIMEDOUT;
1290 }
1291 
1292 /**************************************************************************
1293  *
1294  * Hardware interrupts
1295  * The hardware interrupt handler does very little work; all the event
1296  * queue processing is carried out by per-channel tasklets.
1297  *
1298  **************************************************************************/
1299 
1300 /* Enable/disable/generate interrupts */
1301 static inline void efx_nic_interrupts(struct efx_nic *efx,
1302 				      bool enabled, bool force)
1303 {
1304 	efx_oword_t int_en_reg_ker;
1305 
1306 	EFX_POPULATE_OWORD_3(int_en_reg_ker,
1307 			     FRF_AZ_KER_INT_LEVE_SEL, efx->irq_level,
1308 			     FRF_AZ_KER_INT_KER, force,
1309 			     FRF_AZ_DRV_INT_EN_KER, enabled);
1310 	efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
1311 }
1312 
1313 void efx_nic_enable_interrupts(struct efx_nic *efx)
1314 {
1315 	struct efx_channel *channel;
1316 
1317 	EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
1318 	wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
1319 
1320 	/* Enable interrupts */
1321 	efx_nic_interrupts(efx, true, false);
1322 
1323 	/* Force processing of all the channels to get the EVQ RPTRs up to
1324 	   date */
1325 	efx_for_each_channel(channel, efx)
1326 		efx_schedule_channel(channel);
1327 }
1328 
1329 void efx_nic_disable_interrupts(struct efx_nic *efx)
1330 {
1331 	/* Disable interrupts */
1332 	efx_nic_interrupts(efx, false, false);
1333 }
1334 
1335 /* Generate a test interrupt
1336  * Interrupt must already have been enabled, otherwise nasty things
1337  * may happen.
1338  */
1339 void efx_nic_generate_interrupt(struct efx_nic *efx)
1340 {
1341 	efx_nic_interrupts(efx, true, true);
1342 }
1343 
1344 /* Process a fatal interrupt
1345  * Disable bus mastering ASAP and schedule a reset
1346  */
1347 irqreturn_t efx_nic_fatal_interrupt(struct efx_nic *efx)
1348 {
1349 	struct falcon_nic_data *nic_data = efx->nic_data;
1350 	efx_oword_t *int_ker = efx->irq_status.addr;
1351 	efx_oword_t fatal_intr;
1352 	int error, mem_perr;
1353 
1354 	efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
1355 	error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
1356 
1357 	netif_err(efx, hw, efx->net_dev, "SYSTEM ERROR "EFX_OWORD_FMT" status "
1358 		  EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
1359 		  EFX_OWORD_VAL(fatal_intr),
1360 		  error ? "disabling bus mastering" : "no recognised error");
1361 
1362 	/* If this is a memory parity error dump which blocks are offending */
1363 	mem_perr = (EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER) ||
1364 		    EFX_OWORD_FIELD(fatal_intr, FRF_AZ_SRM_PERR_INT_KER));
1365 	if (mem_perr) {
1366 		efx_oword_t reg;
1367 		efx_reado(efx, &reg, FR_AZ_MEM_STAT);
1368 		netif_err(efx, hw, efx->net_dev,
1369 			  "SYSTEM ERROR: memory parity error "EFX_OWORD_FMT"\n",
1370 			  EFX_OWORD_VAL(reg));
1371 	}
1372 
1373 	/* Disable both devices */
1374 	pci_clear_master(efx->pci_dev);
1375 	if (efx_nic_is_dual_func(efx))
1376 		pci_clear_master(nic_data->pci_dev2);
1377 	efx_nic_disable_interrupts(efx);
1378 
1379 	/* Count errors and reset or disable the NIC accordingly */
1380 	if (efx->int_error_count == 0 ||
1381 	    time_after(jiffies, efx->int_error_expire)) {
1382 		efx->int_error_count = 0;
1383 		efx->int_error_expire =
1384 			jiffies + EFX_INT_ERROR_EXPIRE * HZ;
1385 	}
1386 	if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
1387 		netif_err(efx, hw, efx->net_dev,
1388 			  "SYSTEM ERROR - reset scheduled\n");
1389 		efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
1390 	} else {
1391 		netif_err(efx, hw, efx->net_dev,
1392 			  "SYSTEM ERROR - max number of errors seen."
1393 			  "NIC will be disabled\n");
1394 		efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1395 	}
1396 
1397 	return IRQ_HANDLED;
1398 }
1399 
1400 /* Handle a legacy interrupt
1401  * Acknowledges the interrupt and schedule event queue processing.
1402  */
1403 static irqreturn_t efx_legacy_interrupt(int irq, void *dev_id)
1404 {
1405 	struct efx_nic *efx = dev_id;
1406 	efx_oword_t *int_ker = efx->irq_status.addr;
1407 	irqreturn_t result = IRQ_NONE;
1408 	struct efx_channel *channel;
1409 	efx_dword_t reg;
1410 	u32 queues;
1411 	int syserr;
1412 
1413 	/* Could this be ours?  If interrupts are disabled then the
1414 	 * channel state may not be valid.
1415 	 */
1416 	if (!efx->legacy_irq_enabled)
1417 		return result;
1418 
1419 	/* Read the ISR which also ACKs the interrupts */
1420 	efx_readd(efx, &reg, FR_BZ_INT_ISR0);
1421 	queues = EFX_EXTRACT_DWORD(reg, 0, 31);
1422 
1423 	/* Handle non-event-queue sources */
1424 	if (queues & (1U << efx->irq_level)) {
1425 		syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1426 		if (unlikely(syserr))
1427 			return efx_nic_fatal_interrupt(efx);
1428 		efx->last_irq_cpu = raw_smp_processor_id();
1429 	}
1430 
1431 	if (queues != 0) {
1432 		if (EFX_WORKAROUND_15783(efx))
1433 			efx->irq_zero_count = 0;
1434 
1435 		/* Schedule processing of any interrupting queues */
1436 		efx_for_each_channel(channel, efx) {
1437 			if (queues & 1)
1438 				efx_schedule_channel_irq(channel);
1439 			queues >>= 1;
1440 		}
1441 		result = IRQ_HANDLED;
1442 
1443 	} else if (EFX_WORKAROUND_15783(efx)) {
1444 		efx_qword_t *event;
1445 
1446 		/* We can't return IRQ_HANDLED more than once on seeing ISR=0
1447 		 * because this might be a shared interrupt. */
1448 		if (efx->irq_zero_count++ == 0)
1449 			result = IRQ_HANDLED;
1450 
1451 		/* Ensure we schedule or rearm all event queues */
1452 		efx_for_each_channel(channel, efx) {
1453 			event = efx_event(channel, channel->eventq_read_ptr);
1454 			if (efx_event_present(event))
1455 				efx_schedule_channel_irq(channel);
1456 			else
1457 				efx_nic_eventq_read_ack(channel);
1458 		}
1459 	}
1460 
1461 	if (result == IRQ_HANDLED)
1462 		netif_vdbg(efx, intr, efx->net_dev,
1463 			   "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
1464 			   irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
1465 
1466 	return result;
1467 }
1468 
1469 /* Handle an MSI interrupt
1470  *
1471  * Handle an MSI hardware interrupt.  This routine schedules event
1472  * queue processing.  No interrupt acknowledgement cycle is necessary.
1473  * Also, we never need to check that the interrupt is for us, since
1474  * MSI interrupts cannot be shared.
1475  */
1476 static irqreturn_t efx_msi_interrupt(int irq, void *dev_id)
1477 {
1478 	struct efx_channel *channel = *(struct efx_channel **)dev_id;
1479 	struct efx_nic *efx = channel->efx;
1480 	efx_oword_t *int_ker = efx->irq_status.addr;
1481 	int syserr;
1482 
1483 	netif_vdbg(efx, intr, efx->net_dev,
1484 		   "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
1485 		   irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
1486 
1487 	/* Handle non-event-queue sources */
1488 	if (channel->channel == efx->irq_level) {
1489 		syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1490 		if (unlikely(syserr))
1491 			return efx_nic_fatal_interrupt(efx);
1492 		efx->last_irq_cpu = raw_smp_processor_id();
1493 	}
1494 
1495 	/* Schedule processing of the channel */
1496 	efx_schedule_channel_irq(channel);
1497 
1498 	return IRQ_HANDLED;
1499 }
1500 
1501 
1502 /* Setup RSS indirection table.
1503  * This maps from the hash value of the packet to RXQ
1504  */
1505 void efx_nic_push_rx_indir_table(struct efx_nic *efx)
1506 {
1507 	size_t i = 0;
1508 	efx_dword_t dword;
1509 
1510 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0)
1511 		return;
1512 
1513 	BUILD_BUG_ON(ARRAY_SIZE(efx->rx_indir_table) !=
1514 		     FR_BZ_RX_INDIRECTION_TBL_ROWS);
1515 
1516 	for (i = 0; i < FR_BZ_RX_INDIRECTION_TBL_ROWS; i++) {
1517 		EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
1518 				     efx->rx_indir_table[i]);
1519 		efx_writed_table(efx, &dword, FR_BZ_RX_INDIRECTION_TBL, i);
1520 	}
1521 }
1522 
1523 /* Hook interrupt handler(s)
1524  * Try MSI and then legacy interrupts.
1525  */
1526 int efx_nic_init_interrupt(struct efx_nic *efx)
1527 {
1528 	struct efx_channel *channel;
1529 	int rc;
1530 
1531 	if (!EFX_INT_MODE_USE_MSI(efx)) {
1532 		irq_handler_t handler;
1533 		if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1534 			handler = efx_legacy_interrupt;
1535 		else
1536 			handler = falcon_legacy_interrupt_a1;
1537 
1538 		rc = request_irq(efx->legacy_irq, handler, IRQF_SHARED,
1539 				 efx->name, efx);
1540 		if (rc) {
1541 			netif_err(efx, drv, efx->net_dev,
1542 				  "failed to hook legacy IRQ %d\n",
1543 				  efx->pci_dev->irq);
1544 			goto fail1;
1545 		}
1546 		return 0;
1547 	}
1548 
1549 	/* Hook MSI or MSI-X interrupt */
1550 	efx_for_each_channel(channel, efx) {
1551 		rc = request_irq(channel->irq, efx_msi_interrupt,
1552 				 IRQF_PROBE_SHARED, /* Not shared */
1553 				 efx->channel_name[channel->channel],
1554 				 &efx->channel[channel->channel]);
1555 		if (rc) {
1556 			netif_err(efx, drv, efx->net_dev,
1557 				  "failed to hook IRQ %d\n", channel->irq);
1558 			goto fail2;
1559 		}
1560 	}
1561 
1562 	return 0;
1563 
1564  fail2:
1565 	efx_for_each_channel(channel, efx)
1566 		free_irq(channel->irq, &efx->channel[channel->channel]);
1567  fail1:
1568 	return rc;
1569 }
1570 
1571 void efx_nic_fini_interrupt(struct efx_nic *efx)
1572 {
1573 	struct efx_channel *channel;
1574 	efx_oword_t reg;
1575 
1576 	/* Disable MSI/MSI-X interrupts */
1577 	efx_for_each_channel(channel, efx) {
1578 		if (channel->irq)
1579 			free_irq(channel->irq, &efx->channel[channel->channel]);
1580 	}
1581 
1582 	/* ACK legacy interrupt */
1583 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1584 		efx_reado(efx, &reg, FR_BZ_INT_ISR0);
1585 	else
1586 		falcon_irq_ack_a1(efx);
1587 
1588 	/* Disable legacy interrupt */
1589 	if (efx->legacy_irq)
1590 		free_irq(efx->legacy_irq, efx);
1591 }
1592 
1593 u32 efx_nic_fpga_ver(struct efx_nic *efx)
1594 {
1595 	efx_oword_t altera_build;
1596 	efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
1597 	return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
1598 }
1599 
1600 void efx_nic_init_common(struct efx_nic *efx)
1601 {
1602 	efx_oword_t temp;
1603 
1604 	/* Set positions of descriptor caches in SRAM. */
1605 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR,
1606 			     efx->type->tx_dc_base / 8);
1607 	efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
1608 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR,
1609 			     efx->type->rx_dc_base / 8);
1610 	efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
1611 
1612 	/* Set TX descriptor cache size. */
1613 	BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
1614 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
1615 	efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
1616 
1617 	/* Set RX descriptor cache size.  Set low watermark to size-8, as
1618 	 * this allows most efficient prefetching.
1619 	 */
1620 	BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
1621 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
1622 	efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
1623 	EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
1624 	efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
1625 
1626 	/* Program INT_KER address */
1627 	EFX_POPULATE_OWORD_2(temp,
1628 			     FRF_AZ_NORM_INT_VEC_DIS_KER,
1629 			     EFX_INT_MODE_USE_MSI(efx),
1630 			     FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
1631 	efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
1632 
1633 	if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
1634 		/* Use an interrupt level unused by event queues */
1635 		efx->irq_level = 0x1f;
1636 	else
1637 		/* Use a valid MSI-X vector */
1638 		efx->irq_level = 0;
1639 
1640 	/* Enable all the genuinely fatal interrupts.  (They are still
1641 	 * masked by the overall interrupt mask, controlled by
1642 	 * falcon_interrupts()).
1643 	 *
1644 	 * Note: All other fatal interrupts are enabled
1645 	 */
1646 	EFX_POPULATE_OWORD_3(temp,
1647 			     FRF_AZ_ILL_ADR_INT_KER_EN, 1,
1648 			     FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
1649 			     FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
1650 	if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1651 		EFX_SET_OWORD_FIELD(temp, FRF_CZ_SRAM_PERR_INT_P_KER_EN, 1);
1652 	EFX_INVERT_OWORD(temp);
1653 	efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
1654 
1655 	efx_nic_push_rx_indir_table(efx);
1656 
1657 	/* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
1658 	 * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
1659 	 */
1660 	efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
1661 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
1662 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
1663 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
1664 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 1);
1665 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
1666 	/* Enable SW_EV to inherit in char driver - assume harmless here */
1667 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
1668 	/* Prefetch threshold 2 => fetch when descriptor cache half empty */
1669 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
1670 	/* Disable hardware watchdog which can misfire */
1671 	EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
1672 	/* Squash TX of packets of 16 bytes or less */
1673 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1674 		EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
1675 	efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
1676 
1677 	if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
1678 		EFX_POPULATE_OWORD_4(temp,
1679 				     /* Default values */
1680 				     FRF_BZ_TX_PACE_SB_NOT_AF, 0x15,
1681 				     FRF_BZ_TX_PACE_SB_AF, 0xb,
1682 				     FRF_BZ_TX_PACE_FB_BASE, 0,
1683 				     /* Allow large pace values in the
1684 				      * fast bin. */
1685 				     FRF_BZ_TX_PACE_BIN_TH,
1686 				     FFE_BZ_TX_PACE_RESERVED);
1687 		efx_writeo(efx, &temp, FR_BZ_TX_PACE);
1688 	}
1689 }
1690 
1691 /* Register dump */
1692 
1693 #define REGISTER_REVISION_A	1
1694 #define REGISTER_REVISION_B	2
1695 #define REGISTER_REVISION_C	3
1696 #define REGISTER_REVISION_Z	3	/* latest revision */
1697 
1698 struct efx_nic_reg {
1699 	u32 offset:24;
1700 	u32 min_revision:2, max_revision:2;
1701 };
1702 
1703 #define REGISTER(name, min_rev, max_rev) {				\
1704 	FR_ ## min_rev ## max_rev ## _ ## name,				\
1705 	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev	\
1706 }
1707 #define REGISTER_AA(name) REGISTER(name, A, A)
1708 #define REGISTER_AB(name) REGISTER(name, A, B)
1709 #define REGISTER_AZ(name) REGISTER(name, A, Z)
1710 #define REGISTER_BB(name) REGISTER(name, B, B)
1711 #define REGISTER_BZ(name) REGISTER(name, B, Z)
1712 #define REGISTER_CZ(name) REGISTER(name, C, Z)
1713 
1714 static const struct efx_nic_reg efx_nic_regs[] = {
1715 	REGISTER_AZ(ADR_REGION),
1716 	REGISTER_AZ(INT_EN_KER),
1717 	REGISTER_BZ(INT_EN_CHAR),
1718 	REGISTER_AZ(INT_ADR_KER),
1719 	REGISTER_BZ(INT_ADR_CHAR),
1720 	/* INT_ACK_KER is WO */
1721 	/* INT_ISR0 is RC */
1722 	REGISTER_AZ(HW_INIT),
1723 	REGISTER_CZ(USR_EV_CFG),
1724 	REGISTER_AB(EE_SPI_HCMD),
1725 	REGISTER_AB(EE_SPI_HADR),
1726 	REGISTER_AB(EE_SPI_HDATA),
1727 	REGISTER_AB(EE_BASE_PAGE),
1728 	REGISTER_AB(EE_VPD_CFG0),
1729 	/* EE_VPD_SW_CNTL and EE_VPD_SW_DATA are not used */
1730 	/* PMBX_DBG_IADDR and PBMX_DBG_IDATA are indirect */
1731 	/* PCIE_CORE_INDIRECT is indirect */
1732 	REGISTER_AB(NIC_STAT),
1733 	REGISTER_AB(GPIO_CTL),
1734 	REGISTER_AB(GLB_CTL),
1735 	/* FATAL_INTR_KER and FATAL_INTR_CHAR are partly RC */
1736 	REGISTER_BZ(DP_CTRL),
1737 	REGISTER_AZ(MEM_STAT),
1738 	REGISTER_AZ(CS_DEBUG),
1739 	REGISTER_AZ(ALTERA_BUILD),
1740 	REGISTER_AZ(CSR_SPARE),
1741 	REGISTER_AB(PCIE_SD_CTL0123),
1742 	REGISTER_AB(PCIE_SD_CTL45),
1743 	REGISTER_AB(PCIE_PCS_CTL_STAT),
1744 	/* DEBUG_DATA_OUT is not used */
1745 	/* DRV_EV is WO */
1746 	REGISTER_AZ(EVQ_CTL),
1747 	REGISTER_AZ(EVQ_CNT1),
1748 	REGISTER_AZ(EVQ_CNT2),
1749 	REGISTER_AZ(BUF_TBL_CFG),
1750 	REGISTER_AZ(SRM_RX_DC_CFG),
1751 	REGISTER_AZ(SRM_TX_DC_CFG),
1752 	REGISTER_AZ(SRM_CFG),
1753 	/* BUF_TBL_UPD is WO */
1754 	REGISTER_AZ(SRM_UPD_EVQ),
1755 	REGISTER_AZ(SRAM_PARITY),
1756 	REGISTER_AZ(RX_CFG),
1757 	REGISTER_BZ(RX_FILTER_CTL),
1758 	/* RX_FLUSH_DESCQ is WO */
1759 	REGISTER_AZ(RX_DC_CFG),
1760 	REGISTER_AZ(RX_DC_PF_WM),
1761 	REGISTER_BZ(RX_RSS_TKEY),
1762 	/* RX_NODESC_DROP is RC */
1763 	REGISTER_AA(RX_SELF_RST),
1764 	/* RX_DEBUG, RX_PUSH_DROP are not used */
1765 	REGISTER_CZ(RX_RSS_IPV6_REG1),
1766 	REGISTER_CZ(RX_RSS_IPV6_REG2),
1767 	REGISTER_CZ(RX_RSS_IPV6_REG3),
1768 	/* TX_FLUSH_DESCQ is WO */
1769 	REGISTER_AZ(TX_DC_CFG),
1770 	REGISTER_AA(TX_CHKSM_CFG),
1771 	REGISTER_AZ(TX_CFG),
1772 	/* TX_PUSH_DROP is not used */
1773 	REGISTER_AZ(TX_RESERVED),
1774 	REGISTER_BZ(TX_PACE),
1775 	/* TX_PACE_DROP_QID is RC */
1776 	REGISTER_BB(TX_VLAN),
1777 	REGISTER_BZ(TX_IPFIL_PORTEN),
1778 	REGISTER_AB(MD_TXD),
1779 	REGISTER_AB(MD_RXD),
1780 	REGISTER_AB(MD_CS),
1781 	REGISTER_AB(MD_PHY_ADR),
1782 	REGISTER_AB(MD_ID),
1783 	/* MD_STAT is RC */
1784 	REGISTER_AB(MAC_STAT_DMA),
1785 	REGISTER_AB(MAC_CTRL),
1786 	REGISTER_BB(GEN_MODE),
1787 	REGISTER_AB(MAC_MC_HASH_REG0),
1788 	REGISTER_AB(MAC_MC_HASH_REG1),
1789 	REGISTER_AB(GM_CFG1),
1790 	REGISTER_AB(GM_CFG2),
1791 	/* GM_IPG and GM_HD are not used */
1792 	REGISTER_AB(GM_MAX_FLEN),
1793 	/* GM_TEST is not used */
1794 	REGISTER_AB(GM_ADR1),
1795 	REGISTER_AB(GM_ADR2),
1796 	REGISTER_AB(GMF_CFG0),
1797 	REGISTER_AB(GMF_CFG1),
1798 	REGISTER_AB(GMF_CFG2),
1799 	REGISTER_AB(GMF_CFG3),
1800 	REGISTER_AB(GMF_CFG4),
1801 	REGISTER_AB(GMF_CFG5),
1802 	REGISTER_BB(TX_SRC_MAC_CTL),
1803 	REGISTER_AB(XM_ADR_LO),
1804 	REGISTER_AB(XM_ADR_HI),
1805 	REGISTER_AB(XM_GLB_CFG),
1806 	REGISTER_AB(XM_TX_CFG),
1807 	REGISTER_AB(XM_RX_CFG),
1808 	REGISTER_AB(XM_MGT_INT_MASK),
1809 	REGISTER_AB(XM_FC),
1810 	REGISTER_AB(XM_PAUSE_TIME),
1811 	REGISTER_AB(XM_TX_PARAM),
1812 	REGISTER_AB(XM_RX_PARAM),
1813 	/* XM_MGT_INT_MSK (note no 'A') is RC */
1814 	REGISTER_AB(XX_PWR_RST),
1815 	REGISTER_AB(XX_SD_CTL),
1816 	REGISTER_AB(XX_TXDRV_CTL),
1817 	/* XX_PRBS_CTL, XX_PRBS_CHK and XX_PRBS_ERR are not used */
1818 	/* XX_CORE_STAT is partly RC */
1819 };
1820 
1821 struct efx_nic_reg_table {
1822 	u32 offset:24;
1823 	u32 min_revision:2, max_revision:2;
1824 	u32 step:6, rows:21;
1825 };
1826 
1827 #define REGISTER_TABLE_DIMENSIONS(_, offset, min_rev, max_rev, step, rows) { \
1828 	offset,								\
1829 	REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev,	\
1830 	step, rows							\
1831 }
1832 #define REGISTER_TABLE(name, min_rev, max_rev)				\
1833 	REGISTER_TABLE_DIMENSIONS(					\
1834 		name, FR_ ## min_rev ## max_rev ## _ ## name,		\
1835 		min_rev, max_rev,					\
1836 		FR_ ## min_rev ## max_rev ## _ ## name ## _STEP,	\
1837 		FR_ ## min_rev ## max_rev ## _ ## name ## _ROWS)
1838 #define REGISTER_TABLE_AA(name) REGISTER_TABLE(name, A, A)
1839 #define REGISTER_TABLE_AZ(name) REGISTER_TABLE(name, A, Z)
1840 #define REGISTER_TABLE_BB(name) REGISTER_TABLE(name, B, B)
1841 #define REGISTER_TABLE_BZ(name) REGISTER_TABLE(name, B, Z)
1842 #define REGISTER_TABLE_BB_CZ(name)					\
1843 	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, B, B,		\
1844 				  FR_BZ_ ## name ## _STEP,		\
1845 				  FR_BB_ ## name ## _ROWS),		\
1846 	REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, C, Z,		\
1847 				  FR_BZ_ ## name ## _STEP,		\
1848 				  FR_CZ_ ## name ## _ROWS)
1849 #define REGISTER_TABLE_CZ(name) REGISTER_TABLE(name, C, Z)
1850 
1851 static const struct efx_nic_reg_table efx_nic_reg_tables[] = {
1852 	/* DRIVER is not used */
1853 	/* EVQ_RPTR, TIMER_COMMAND, USR_EV and {RX,TX}_DESC_UPD are WO */
1854 	REGISTER_TABLE_BB(TX_IPFIL_TBL),
1855 	REGISTER_TABLE_BB(TX_SRC_MAC_TBL),
1856 	REGISTER_TABLE_AA(RX_DESC_PTR_TBL_KER),
1857 	REGISTER_TABLE_BB_CZ(RX_DESC_PTR_TBL),
1858 	REGISTER_TABLE_AA(TX_DESC_PTR_TBL_KER),
1859 	REGISTER_TABLE_BB_CZ(TX_DESC_PTR_TBL),
1860 	REGISTER_TABLE_AA(EVQ_PTR_TBL_KER),
1861 	REGISTER_TABLE_BB_CZ(EVQ_PTR_TBL),
1862 	/* We can't reasonably read all of the buffer table (up to 8MB!).
1863 	 * However this driver will only use a few entries.  Reading
1864 	 * 1K entries allows for some expansion of queue count and
1865 	 * size before we need to change the version. */
1866 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL_KER, FR_AA_BUF_FULL_TBL_KER,
1867 				  A, A, 8, 1024),
1868 	REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL, FR_BZ_BUF_FULL_TBL,
1869 				  B, Z, 8, 1024),
1870 	REGISTER_TABLE_CZ(RX_MAC_FILTER_TBL0),
1871 	REGISTER_TABLE_BB_CZ(TIMER_TBL),
1872 	REGISTER_TABLE_BB_CZ(TX_PACE_TBL),
1873 	REGISTER_TABLE_BZ(RX_INDIRECTION_TBL),
1874 	/* TX_FILTER_TBL0 is huge and not used by this driver */
1875 	REGISTER_TABLE_CZ(TX_MAC_FILTER_TBL0),
1876 	REGISTER_TABLE_CZ(MC_TREG_SMEM),
1877 	/* MSIX_PBA_TABLE is not mapped */
1878 	/* SRM_DBG is not mapped (and is redundant with BUF_FLL_TBL) */
1879 	REGISTER_TABLE_BZ(RX_FILTER_TBL0),
1880 };
1881 
1882 size_t efx_nic_get_regs_len(struct efx_nic *efx)
1883 {
1884 	const struct efx_nic_reg *reg;
1885 	const struct efx_nic_reg_table *table;
1886 	size_t len = 0;
1887 
1888 	for (reg = efx_nic_regs;
1889 	     reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
1890 	     reg++)
1891 		if (efx->type->revision >= reg->min_revision &&
1892 		    efx->type->revision <= reg->max_revision)
1893 			len += sizeof(efx_oword_t);
1894 
1895 	for (table = efx_nic_reg_tables;
1896 	     table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
1897 	     table++)
1898 		if (efx->type->revision >= table->min_revision &&
1899 		    efx->type->revision <= table->max_revision)
1900 			len += table->rows * min_t(size_t, table->step, 16);
1901 
1902 	return len;
1903 }
1904 
1905 void efx_nic_get_regs(struct efx_nic *efx, void *buf)
1906 {
1907 	const struct efx_nic_reg *reg;
1908 	const struct efx_nic_reg_table *table;
1909 
1910 	for (reg = efx_nic_regs;
1911 	     reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
1912 	     reg++) {
1913 		if (efx->type->revision >= reg->min_revision &&
1914 		    efx->type->revision <= reg->max_revision) {
1915 			efx_reado(efx, (efx_oword_t *)buf, reg->offset);
1916 			buf += sizeof(efx_oword_t);
1917 		}
1918 	}
1919 
1920 	for (table = efx_nic_reg_tables;
1921 	     table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
1922 	     table++) {
1923 		size_t size, i;
1924 
1925 		if (!(efx->type->revision >= table->min_revision &&
1926 		      efx->type->revision <= table->max_revision))
1927 			continue;
1928 
1929 		size = min_t(size_t, table->step, 16);
1930 
1931 		for (i = 0; i < table->rows; i++) {
1932 			switch (table->step) {
1933 			case 4: /* 32-bit register or SRAM */
1934 				efx_readd_table(efx, buf, table->offset, i);
1935 				break;
1936 			case 8: /* 64-bit SRAM */
1937 				efx_sram_readq(efx,
1938 					       efx->membase + table->offset,
1939 					       buf, i);
1940 				break;
1941 			case 16: /* 128-bit register */
1942 				efx_reado_table(efx, buf, table->offset, i);
1943 				break;
1944 			case 32: /* 128-bit register, interleaved */
1945 				efx_reado_table(efx, buf, table->offset, 2 * i);
1946 				break;
1947 			default:
1948 				WARN_ON(1);
1949 				return;
1950 			}
1951 			buf += size;
1952 		}
1953 	}
1954 }
1955