]> bbs.cooldavid.org Git - net-next-2.6.git/blob - drivers/net/sfc/nic.c
sfc: Create multiple TX queues
[net-next-2.6.git] / drivers / net / sfc / nic.c
1 /****************************************************************************
2  * Driver for Solarflare Solarstorm network controllers and boards
3  * Copyright 2005-2006 Fen Systems Ltd.
4  * Copyright 2006-2009 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/pci.h>
14 #include <linux/module.h>
15 #include <linux/seq_file.h>
16 #include "net_driver.h"
17 #include "bitfield.h"
18 #include "efx.h"
19 #include "nic.h"
20 #include "regs.h"
21 #include "io.h"
22 #include "workarounds.h"
23
24 /**************************************************************************
25  *
26  * Configurable values
27  *
28  **************************************************************************
29  */
30
31 /* This is set to 16 for a good reason.  In summary, if larger than
32  * 16, the descriptor cache holds more than a default socket
33  * buffer's worth of packets (for UDP we can only have at most one
34  * socket buffer's worth outstanding).  This combined with the fact
35  * that we only get 1 TX event per descriptor cache means the NIC
36  * goes idle.
37  */
38 #define TX_DC_ENTRIES 16
39 #define TX_DC_ENTRIES_ORDER 1
40
41 #define RX_DC_ENTRIES 64
42 #define RX_DC_ENTRIES_ORDER 3
43
44 /* RX FIFO XOFF watermark
45  *
46  * When the amount of the RX FIFO increases used increases past this
47  * watermark send XOFF. Only used if RX flow control is enabled (ethtool -A)
48  * This also has an effect on RX/TX arbitration
49  */
50 int efx_nic_rx_xoff_thresh = -1;
51 module_param_named(rx_xoff_thresh_bytes, efx_nic_rx_xoff_thresh, int, 0644);
52 MODULE_PARM_DESC(rx_xoff_thresh_bytes, "RX fifo XOFF threshold");
53
54 /* RX FIFO XON watermark
55  *
56  * When the amount of the RX FIFO used decreases below this
57  * watermark send XON. Only used if TX flow control is enabled (ethtool -A)
58  * This also has an effect on RX/TX arbitration
59  */
60 int efx_nic_rx_xon_thresh = -1;
61 module_param_named(rx_xon_thresh_bytes, efx_nic_rx_xon_thresh, int, 0644);
62 MODULE_PARM_DESC(rx_xon_thresh_bytes, "RX fifo XON threshold");
63
64 /* If EFX_MAX_INT_ERRORS internal errors occur within
65  * EFX_INT_ERROR_EXPIRE seconds, we consider the NIC broken and
66  * disable it.
67  */
68 #define EFX_INT_ERROR_EXPIRE 3600
69 #define EFX_MAX_INT_ERRORS 5
70
71 /* We poll for events every FLUSH_INTERVAL ms, and check FLUSH_POLL_COUNT times
72  */
73 #define EFX_FLUSH_INTERVAL 10
74 #define EFX_FLUSH_POLL_COUNT 100
75
76 /* Size and alignment of special buffers (4KB) */
77 #define EFX_BUF_SIZE 4096
78
79 /* Depth of RX flush request fifo */
80 #define EFX_RX_FLUSH_COUNT 4
81
82 /**************************************************************************
83  *
84  * Solarstorm hardware access
85  *
86  **************************************************************************/
87
88 static inline void efx_write_buf_tbl(struct efx_nic *efx, efx_qword_t *value,
89                                      unsigned int index)
90 {
91         efx_sram_writeq(efx, efx->membase + efx->type->buf_tbl_base,
92                         value, index);
93 }
94
95 /* Read the current event from the event queue */
96 static inline efx_qword_t *efx_event(struct efx_channel *channel,
97                                      unsigned int index)
98 {
99         return (((efx_qword_t *) (channel->eventq.addr)) + index);
100 }
101
102 /* See if an event is present
103  *
104  * We check both the high and low dword of the event for all ones.  We
105  * wrote all ones when we cleared the event, and no valid event can
106  * have all ones in either its high or low dwords.  This approach is
107  * robust against reordering.
108  *
109  * Note that using a single 64-bit comparison is incorrect; even
110  * though the CPU read will be atomic, the DMA write may not be.
111  */
112 static inline int efx_event_present(efx_qword_t *event)
113 {
114         return (!(EFX_DWORD_IS_ALL_ONES(event->dword[0]) |
115                   EFX_DWORD_IS_ALL_ONES(event->dword[1])));
116 }
117
118 static bool efx_masked_compare_oword(const efx_oword_t *a, const efx_oword_t *b,
119                                      const efx_oword_t *mask)
120 {
121         return ((a->u64[0] ^ b->u64[0]) & mask->u64[0]) ||
122                 ((a->u64[1] ^ b->u64[1]) & mask->u64[1]);
123 }
124
125 int efx_nic_test_registers(struct efx_nic *efx,
126                            const struct efx_nic_register_test *regs,
127                            size_t n_regs)
128 {
129         unsigned address = 0, i, j;
130         efx_oword_t mask, imask, original, reg, buf;
131
132         /* Falcon should be in loopback to isolate the XMAC from the PHY */
133         WARN_ON(!LOOPBACK_INTERNAL(efx));
134
135         for (i = 0; i < n_regs; ++i) {
136                 address = regs[i].address;
137                 mask = imask = regs[i].mask;
138                 EFX_INVERT_OWORD(imask);
139
140                 efx_reado(efx, &original, address);
141
142                 /* bit sweep on and off */
143                 for (j = 0; j < 128; j++) {
144                         if (!EFX_EXTRACT_OWORD32(mask, j, j))
145                                 continue;
146
147                         /* Test this testable bit can be set in isolation */
148                         EFX_AND_OWORD(reg, original, mask);
149                         EFX_SET_OWORD32(reg, j, j, 1);
150
151                         efx_writeo(efx, &reg, address);
152                         efx_reado(efx, &buf, address);
153
154                         if (efx_masked_compare_oword(&reg, &buf, &mask))
155                                 goto fail;
156
157                         /* Test this testable bit can be cleared in isolation */
158                         EFX_OR_OWORD(reg, original, mask);
159                         EFX_SET_OWORD32(reg, j, j, 0);
160
161                         efx_writeo(efx, &reg, address);
162                         efx_reado(efx, &buf, address);
163
164                         if (efx_masked_compare_oword(&reg, &buf, &mask))
165                                 goto fail;
166                 }
167
168                 efx_writeo(efx, &original, address);
169         }
170
171         return 0;
172
173 fail:
174         EFX_ERR(efx, "wrote "EFX_OWORD_FMT" read "EFX_OWORD_FMT
175                 " at address 0x%x mask "EFX_OWORD_FMT"\n", EFX_OWORD_VAL(reg),
176                 EFX_OWORD_VAL(buf), address, EFX_OWORD_VAL(mask));
177         return -EIO;
178 }
179
180 /**************************************************************************
181  *
182  * Special buffer handling
183  * Special buffers are used for event queues and the TX and RX
184  * descriptor rings.
185  *
186  *************************************************************************/
187
188 /*
189  * Initialise a special buffer
190  *
191  * This will define a buffer (previously allocated via
192  * efx_alloc_special_buffer()) in the buffer table, allowing
193  * it to be used for event queues, descriptor rings etc.
194  */
195 static void
196 efx_init_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
197 {
198         efx_qword_t buf_desc;
199         int index;
200         dma_addr_t dma_addr;
201         int i;
202
203         EFX_BUG_ON_PARANOID(!buffer->addr);
204
205         /* Write buffer descriptors to NIC */
206         for (i = 0; i < buffer->entries; i++) {
207                 index = buffer->index + i;
208                 dma_addr = buffer->dma_addr + (i * 4096);
209                 EFX_LOG(efx, "mapping special buffer %d at %llx\n",
210                         index, (unsigned long long)dma_addr);
211                 EFX_POPULATE_QWORD_3(buf_desc,
212                                      FRF_AZ_BUF_ADR_REGION, 0,
213                                      FRF_AZ_BUF_ADR_FBUF, dma_addr >> 12,
214                                      FRF_AZ_BUF_OWNER_ID_FBUF, 0);
215                 efx_write_buf_tbl(efx, &buf_desc, index);
216         }
217 }
218
219 /* Unmaps a buffer and clears the buffer table entries */
220 static void
221 efx_fini_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
222 {
223         efx_oword_t buf_tbl_upd;
224         unsigned int start = buffer->index;
225         unsigned int end = (buffer->index + buffer->entries - 1);
226
227         if (!buffer->entries)
228                 return;
229
230         EFX_LOG(efx, "unmapping special buffers %d-%d\n",
231                 buffer->index, buffer->index + buffer->entries - 1);
232
233         EFX_POPULATE_OWORD_4(buf_tbl_upd,
234                              FRF_AZ_BUF_UPD_CMD, 0,
235                              FRF_AZ_BUF_CLR_CMD, 1,
236                              FRF_AZ_BUF_CLR_END_ID, end,
237                              FRF_AZ_BUF_CLR_START_ID, start);
238         efx_writeo(efx, &buf_tbl_upd, FR_AZ_BUF_TBL_UPD);
239 }
240
241 /*
242  * Allocate a new special buffer
243  *
244  * This allocates memory for a new buffer, clears it and allocates a
245  * new buffer ID range.  It does not write into the buffer table.
246  *
247  * This call will allocate 4KB buffers, since 8KB buffers can't be
248  * used for event queues and descriptor rings.
249  */
250 static int efx_alloc_special_buffer(struct efx_nic *efx,
251                                     struct efx_special_buffer *buffer,
252                                     unsigned int len)
253 {
254         len = ALIGN(len, EFX_BUF_SIZE);
255
256         buffer->addr = pci_alloc_consistent(efx->pci_dev, len,
257                                             &buffer->dma_addr);
258         if (!buffer->addr)
259                 return -ENOMEM;
260         buffer->len = len;
261         buffer->entries = len / EFX_BUF_SIZE;
262         BUG_ON(buffer->dma_addr & (EFX_BUF_SIZE - 1));
263
264         /* All zeros is a potentially valid event so memset to 0xff */
265         memset(buffer->addr, 0xff, len);
266
267         /* Select new buffer ID */
268         buffer->index = efx->next_buffer_table;
269         efx->next_buffer_table += buffer->entries;
270
271         EFX_LOG(efx, "allocating special buffers %d-%d at %llx+%x "
272                 "(virt %p phys %llx)\n", buffer->index,
273                 buffer->index + buffer->entries - 1,
274                 (u64)buffer->dma_addr, len,
275                 buffer->addr, (u64)virt_to_phys(buffer->addr));
276
277         return 0;
278 }
279
280 static void
281 efx_free_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
282 {
283         if (!buffer->addr)
284                 return;
285
286         EFX_LOG(efx, "deallocating special buffers %d-%d at %llx+%x "
287                 "(virt %p phys %llx)\n", buffer->index,
288                 buffer->index + buffer->entries - 1,
289                 (u64)buffer->dma_addr, buffer->len,
290                 buffer->addr, (u64)virt_to_phys(buffer->addr));
291
292         pci_free_consistent(efx->pci_dev, buffer->len, buffer->addr,
293                             buffer->dma_addr);
294         buffer->addr = NULL;
295         buffer->entries = 0;
296 }
297
298 /**************************************************************************
299  *
300  * Generic buffer handling
301  * These buffers are used for interrupt status and MAC stats
302  *
303  **************************************************************************/
304
305 int efx_nic_alloc_buffer(struct efx_nic *efx, struct efx_buffer *buffer,
306                          unsigned int len)
307 {
308         buffer->addr = pci_alloc_consistent(efx->pci_dev, len,
309                                             &buffer->dma_addr);
310         if (!buffer->addr)
311                 return -ENOMEM;
312         buffer->len = len;
313         memset(buffer->addr, 0, len);
314         return 0;
315 }
316
317 void efx_nic_free_buffer(struct efx_nic *efx, struct efx_buffer *buffer)
318 {
319         if (buffer->addr) {
320                 pci_free_consistent(efx->pci_dev, buffer->len,
321                                     buffer->addr, buffer->dma_addr);
322                 buffer->addr = NULL;
323         }
324 }
325
326 /**************************************************************************
327  *
328  * TX path
329  *
330  **************************************************************************/
331
332 /* Returns a pointer to the specified transmit descriptor in the TX
333  * descriptor queue belonging to the specified channel.
334  */
335 static inline efx_qword_t *
336 efx_tx_desc(struct efx_tx_queue *tx_queue, unsigned int index)
337 {
338         return (((efx_qword_t *) (tx_queue->txd.addr)) + index);
339 }
340
341 /* This writes to the TX_DESC_WPTR; write pointer for TX descriptor ring */
342 static inline void efx_notify_tx_desc(struct efx_tx_queue *tx_queue)
343 {
344         unsigned write_ptr;
345         efx_dword_t reg;
346
347         write_ptr = tx_queue->write_count & EFX_TXQ_MASK;
348         EFX_POPULATE_DWORD_1(reg, FRF_AZ_TX_DESC_WPTR_DWORD, write_ptr);
349         efx_writed_page(tx_queue->efx, &reg,
350                         FR_AZ_TX_DESC_UPD_DWORD_P0, tx_queue->queue);
351 }
352
353
354 /* For each entry inserted into the software descriptor ring, create a
355  * descriptor in the hardware TX descriptor ring (in host memory), and
356  * write a doorbell.
357  */
358 void efx_nic_push_buffers(struct efx_tx_queue *tx_queue)
359 {
360
361         struct efx_tx_buffer *buffer;
362         efx_qword_t *txd;
363         unsigned write_ptr;
364
365         BUG_ON(tx_queue->write_count == tx_queue->insert_count);
366
367         do {
368                 write_ptr = tx_queue->write_count & EFX_TXQ_MASK;
369                 buffer = &tx_queue->buffer[write_ptr];
370                 txd = efx_tx_desc(tx_queue, write_ptr);
371                 ++tx_queue->write_count;
372
373                 /* Create TX descriptor ring entry */
374                 EFX_POPULATE_QWORD_4(*txd,
375                                      FSF_AZ_TX_KER_CONT, buffer->continuation,
376                                      FSF_AZ_TX_KER_BYTE_COUNT, buffer->len,
377                                      FSF_AZ_TX_KER_BUF_REGION, 0,
378                                      FSF_AZ_TX_KER_BUF_ADDR, buffer->dma_addr);
379         } while (tx_queue->write_count != tx_queue->insert_count);
380
381         wmb(); /* Ensure descriptors are written before they are fetched */
382         efx_notify_tx_desc(tx_queue);
383 }
384
385 /* Allocate hardware resources for a TX queue */
386 int efx_nic_probe_tx(struct efx_tx_queue *tx_queue)
387 {
388         struct efx_nic *efx = tx_queue->efx;
389         BUILD_BUG_ON(EFX_TXQ_SIZE < 512 || EFX_TXQ_SIZE > 4096 ||
390                      EFX_TXQ_SIZE & EFX_TXQ_MASK);
391         return efx_alloc_special_buffer(efx, &tx_queue->txd,
392                                         EFX_TXQ_SIZE * sizeof(efx_qword_t));
393 }
394
395 void efx_nic_init_tx(struct efx_tx_queue *tx_queue)
396 {
397         efx_oword_t tx_desc_ptr;
398         struct efx_nic *efx = tx_queue->efx;
399
400         tx_queue->flushed = FLUSH_NONE;
401
402         /* Pin TX descriptor ring */
403         efx_init_special_buffer(efx, &tx_queue->txd);
404
405         /* Push TX descriptor ring to card */
406         EFX_POPULATE_OWORD_10(tx_desc_ptr,
407                               FRF_AZ_TX_DESCQ_EN, 1,
408                               FRF_AZ_TX_ISCSI_DDIG_EN, 0,
409                               FRF_AZ_TX_ISCSI_HDIG_EN, 0,
410                               FRF_AZ_TX_DESCQ_BUF_BASE_ID, tx_queue->txd.index,
411                               FRF_AZ_TX_DESCQ_EVQ_ID,
412                               tx_queue->channel->channel,
413                               FRF_AZ_TX_DESCQ_OWNER_ID, 0,
414                               FRF_AZ_TX_DESCQ_LABEL, tx_queue->queue,
415                               FRF_AZ_TX_DESCQ_SIZE,
416                               __ffs(tx_queue->txd.entries),
417                               FRF_AZ_TX_DESCQ_TYPE, 0,
418                               FRF_BZ_TX_NON_IP_DROP_DIS, 1);
419
420         if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
421                 int csum = tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD;
422                 EFX_SET_OWORD_FIELD(tx_desc_ptr, FRF_BZ_TX_IP_CHKSM_DIS, !csum);
423                 EFX_SET_OWORD_FIELD(tx_desc_ptr, FRF_BZ_TX_TCP_CHKSM_DIS,
424                                     !csum);
425         }
426
427         efx_writeo_table(efx, &tx_desc_ptr, efx->type->txd_ptr_tbl_base,
428                          tx_queue->queue);
429
430         if (efx_nic_rev(efx) < EFX_REV_FALCON_B0) {
431                 efx_oword_t reg;
432
433                 /* Only 128 bits in this register */
434                 BUILD_BUG_ON(EFX_MAX_TX_QUEUES > 128);
435
436                 efx_reado(efx, &reg, FR_AA_TX_CHKSM_CFG);
437                 if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD)
438                         clear_bit_le(tx_queue->queue, (void *)&reg);
439                 else
440                         set_bit_le(tx_queue->queue, (void *)&reg);
441                 efx_writeo(efx, &reg, FR_AA_TX_CHKSM_CFG);
442         }
443 }
444
445 static void efx_flush_tx_queue(struct efx_tx_queue *tx_queue)
446 {
447         struct efx_nic *efx = tx_queue->efx;
448         efx_oword_t tx_flush_descq;
449
450         tx_queue->flushed = FLUSH_PENDING;
451
452         /* Post a flush command */
453         EFX_POPULATE_OWORD_2(tx_flush_descq,
454                              FRF_AZ_TX_FLUSH_DESCQ_CMD, 1,
455                              FRF_AZ_TX_FLUSH_DESCQ, tx_queue->queue);
456         efx_writeo(efx, &tx_flush_descq, FR_AZ_TX_FLUSH_DESCQ);
457 }
458
459 void efx_nic_fini_tx(struct efx_tx_queue *tx_queue)
460 {
461         struct efx_nic *efx = tx_queue->efx;
462         efx_oword_t tx_desc_ptr;
463
464         /* The queue should have been flushed */
465         WARN_ON(tx_queue->flushed != FLUSH_DONE);
466
467         /* Remove TX descriptor ring from card */
468         EFX_ZERO_OWORD(tx_desc_ptr);
469         efx_writeo_table(efx, &tx_desc_ptr, efx->type->txd_ptr_tbl_base,
470                          tx_queue->queue);
471
472         /* Unpin TX descriptor ring */
473         efx_fini_special_buffer(efx, &tx_queue->txd);
474 }
475
476 /* Free buffers backing TX queue */
477 void efx_nic_remove_tx(struct efx_tx_queue *tx_queue)
478 {
479         efx_free_special_buffer(tx_queue->efx, &tx_queue->txd);
480 }
481
482 /**************************************************************************
483  *
484  * RX path
485  *
486  **************************************************************************/
487
488 /* Returns a pointer to the specified descriptor in the RX descriptor queue */
489 static inline efx_qword_t *
490 efx_rx_desc(struct efx_rx_queue *rx_queue, unsigned int index)
491 {
492         return (((efx_qword_t *) (rx_queue->rxd.addr)) + index);
493 }
494
495 /* This creates an entry in the RX descriptor queue */
496 static inline void
497 efx_build_rx_desc(struct efx_rx_queue *rx_queue, unsigned index)
498 {
499         struct efx_rx_buffer *rx_buf;
500         efx_qword_t *rxd;
501
502         rxd = efx_rx_desc(rx_queue, index);
503         rx_buf = efx_rx_buffer(rx_queue, index);
504         EFX_POPULATE_QWORD_3(*rxd,
505                              FSF_AZ_RX_KER_BUF_SIZE,
506                              rx_buf->len -
507                              rx_queue->efx->type->rx_buffer_padding,
508                              FSF_AZ_RX_KER_BUF_REGION, 0,
509                              FSF_AZ_RX_KER_BUF_ADDR, rx_buf->dma_addr);
510 }
511
512 /* This writes to the RX_DESC_WPTR register for the specified receive
513  * descriptor ring.
514  */
515 void efx_nic_notify_rx_desc(struct efx_rx_queue *rx_queue)
516 {
517         efx_dword_t reg;
518         unsigned write_ptr;
519
520         while (rx_queue->notified_count != rx_queue->added_count) {
521                 efx_build_rx_desc(rx_queue,
522                                   rx_queue->notified_count &
523                                   EFX_RXQ_MASK);
524                 ++rx_queue->notified_count;
525         }
526
527         wmb();
528         write_ptr = rx_queue->added_count & EFX_RXQ_MASK;
529         EFX_POPULATE_DWORD_1(reg, FRF_AZ_RX_DESC_WPTR_DWORD, write_ptr);
530         efx_writed_page(rx_queue->efx, &reg,
531                         FR_AZ_RX_DESC_UPD_DWORD_P0, rx_queue->queue);
532 }
533
534 int efx_nic_probe_rx(struct efx_rx_queue *rx_queue)
535 {
536         struct efx_nic *efx = rx_queue->efx;
537         BUILD_BUG_ON(EFX_RXQ_SIZE < 512 || EFX_RXQ_SIZE > 4096 ||
538                      EFX_RXQ_SIZE & EFX_RXQ_MASK);
539         return efx_alloc_special_buffer(efx, &rx_queue->rxd,
540                                         EFX_RXQ_SIZE * sizeof(efx_qword_t));
541 }
542
543 void efx_nic_init_rx(struct efx_rx_queue *rx_queue)
544 {
545         efx_oword_t rx_desc_ptr;
546         struct efx_nic *efx = rx_queue->efx;
547         bool is_b0 = efx_nic_rev(efx) >= EFX_REV_FALCON_B0;
548         bool iscsi_digest_en = is_b0;
549
550         EFX_LOG(efx, "RX queue %d ring in special buffers %d-%d\n",
551                 rx_queue->queue, rx_queue->rxd.index,
552                 rx_queue->rxd.index + rx_queue->rxd.entries - 1);
553
554         rx_queue->flushed = FLUSH_NONE;
555
556         /* Pin RX descriptor ring */
557         efx_init_special_buffer(efx, &rx_queue->rxd);
558
559         /* Push RX descriptor ring to card */
560         EFX_POPULATE_OWORD_10(rx_desc_ptr,
561                               FRF_AZ_RX_ISCSI_DDIG_EN, iscsi_digest_en,
562                               FRF_AZ_RX_ISCSI_HDIG_EN, iscsi_digest_en,
563                               FRF_AZ_RX_DESCQ_BUF_BASE_ID, rx_queue->rxd.index,
564                               FRF_AZ_RX_DESCQ_EVQ_ID,
565                               rx_queue->channel->channel,
566                               FRF_AZ_RX_DESCQ_OWNER_ID, 0,
567                               FRF_AZ_RX_DESCQ_LABEL, rx_queue->queue,
568                               FRF_AZ_RX_DESCQ_SIZE,
569                               __ffs(rx_queue->rxd.entries),
570                               FRF_AZ_RX_DESCQ_TYPE, 0 /* kernel queue */ ,
571                               /* For >=B0 this is scatter so disable */
572                               FRF_AZ_RX_DESCQ_JUMBO, !is_b0,
573                               FRF_AZ_RX_DESCQ_EN, 1);
574         efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
575                          rx_queue->queue);
576 }
577
578 static void efx_flush_rx_queue(struct efx_rx_queue *rx_queue)
579 {
580         struct efx_nic *efx = rx_queue->efx;
581         efx_oword_t rx_flush_descq;
582
583         rx_queue->flushed = FLUSH_PENDING;
584
585         /* Post a flush command */
586         EFX_POPULATE_OWORD_2(rx_flush_descq,
587                              FRF_AZ_RX_FLUSH_DESCQ_CMD, 1,
588                              FRF_AZ_RX_FLUSH_DESCQ, rx_queue->queue);
589         efx_writeo(efx, &rx_flush_descq, FR_AZ_RX_FLUSH_DESCQ);
590 }
591
592 void efx_nic_fini_rx(struct efx_rx_queue *rx_queue)
593 {
594         efx_oword_t rx_desc_ptr;
595         struct efx_nic *efx = rx_queue->efx;
596
597         /* The queue should already have been flushed */
598         WARN_ON(rx_queue->flushed != FLUSH_DONE);
599
600         /* Remove RX descriptor ring from card */
601         EFX_ZERO_OWORD(rx_desc_ptr);
602         efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
603                          rx_queue->queue);
604
605         /* Unpin RX descriptor ring */
606         efx_fini_special_buffer(efx, &rx_queue->rxd);
607 }
608
609 /* Free buffers backing RX queue */
610 void efx_nic_remove_rx(struct efx_rx_queue *rx_queue)
611 {
612         efx_free_special_buffer(rx_queue->efx, &rx_queue->rxd);
613 }
614
615 /**************************************************************************
616  *
617  * Event queue processing
618  * Event queues are processed by per-channel tasklets.
619  *
620  **************************************************************************/
621
622 /* Update a channel's event queue's read pointer (RPTR) register
623  *
624  * This writes the EVQ_RPTR_REG register for the specified channel's
625  * event queue.
626  */
627 void efx_nic_eventq_read_ack(struct efx_channel *channel)
628 {
629         efx_dword_t reg;
630         struct efx_nic *efx = channel->efx;
631
632         EFX_POPULATE_DWORD_1(reg, FRF_AZ_EVQ_RPTR, channel->eventq_read_ptr);
633         efx_writed_table(efx, &reg, efx->type->evq_rptr_tbl_base,
634                          channel->channel);
635 }
636
637 /* Use HW to insert a SW defined event */
638 void efx_generate_event(struct efx_channel *channel, efx_qword_t *event)
639 {
640         efx_oword_t drv_ev_reg;
641
642         BUILD_BUG_ON(FRF_AZ_DRV_EV_DATA_LBN != 0 ||
643                      FRF_AZ_DRV_EV_DATA_WIDTH != 64);
644         drv_ev_reg.u32[0] = event->u32[0];
645         drv_ev_reg.u32[1] = event->u32[1];
646         drv_ev_reg.u32[2] = 0;
647         drv_ev_reg.u32[3] = 0;
648         EFX_SET_OWORD_FIELD(drv_ev_reg, FRF_AZ_DRV_EV_QID, channel->channel);
649         efx_writeo(channel->efx, &drv_ev_reg, FR_AZ_DRV_EV);
650 }
651
652 /* Handle a transmit completion event
653  *
654  * The NIC batches TX completion events; the message we receive is of
655  * the form "complete all TX events up to this index".
656  */
657 static int
658 efx_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
659 {
660         unsigned int tx_ev_desc_ptr;
661         unsigned int tx_ev_q_label;
662         struct efx_tx_queue *tx_queue;
663         struct efx_nic *efx = channel->efx;
664         int tx_packets = 0;
665
666         if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
667                 /* Transmit completion */
668                 tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
669                 tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
670                 tx_queue = &efx->tx_queue[tx_ev_q_label];
671                 tx_packets = ((tx_ev_desc_ptr - tx_queue->read_count) &
672                               EFX_TXQ_MASK);
673                 channel->irq_mod_score += tx_packets;
674                 efx_xmit_done(tx_queue, tx_ev_desc_ptr);
675         } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
676                 /* Rewrite the FIFO write pointer */
677                 tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
678                 tx_queue = &efx->tx_queue[tx_ev_q_label];
679
680                 if (efx_dev_registered(efx))
681                         netif_tx_lock(efx->net_dev);
682                 efx_notify_tx_desc(tx_queue);
683                 if (efx_dev_registered(efx))
684                         netif_tx_unlock(efx->net_dev);
685         } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR) &&
686                    EFX_WORKAROUND_10727(efx)) {
687                 efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
688         } else {
689                 EFX_ERR(efx, "channel %d unexpected TX event "
690                         EFX_QWORD_FMT"\n", channel->channel,
691                         EFX_QWORD_VAL(*event));
692         }
693
694         return tx_packets;
695 }
696
697 /* Detect errors included in the rx_evt_pkt_ok bit. */
698 static void efx_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
699                                  const efx_qword_t *event,
700                                  bool *rx_ev_pkt_ok,
701                                  bool *discard)
702 {
703         struct efx_nic *efx = rx_queue->efx;
704         bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
705         bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
706         bool rx_ev_frm_trunc, rx_ev_drib_nib, rx_ev_tobe_disc;
707         bool rx_ev_other_err, rx_ev_pause_frm;
708         bool rx_ev_hdr_type, rx_ev_mcast_pkt;
709         unsigned rx_ev_pkt_type;
710
711         rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
712         rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
713         rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
714         rx_ev_pkt_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_TYPE);
715         rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
716                                                  FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
717         rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
718                                                   FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
719         rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
720                                                    FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
721         rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
722         rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
723         rx_ev_drib_nib = ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) ?
724                           0 : EFX_QWORD_FIELD(*event, FSF_AA_RX_EV_DRIB_NIB));
725         rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
726
727         /* Every error apart from tobe_disc and pause_frm */
728         rx_ev_other_err = (rx_ev_drib_nib | rx_ev_tcp_udp_chksum_err |
729                            rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
730                            rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
731
732         /* Count errors that are not in MAC stats.  Ignore expected
733          * checksum errors during self-test. */
734         if (rx_ev_frm_trunc)
735                 ++rx_queue->channel->n_rx_frm_trunc;
736         else if (rx_ev_tobe_disc)
737                 ++rx_queue->channel->n_rx_tobe_disc;
738         else if (!efx->loopback_selftest) {
739                 if (rx_ev_ip_hdr_chksum_err)
740                         ++rx_queue->channel->n_rx_ip_hdr_chksum_err;
741                 else if (rx_ev_tcp_udp_chksum_err)
742                         ++rx_queue->channel->n_rx_tcp_udp_chksum_err;
743         }
744
745         /* The frame must be discarded if any of these are true. */
746         *discard = (rx_ev_eth_crc_err | rx_ev_frm_trunc | rx_ev_drib_nib |
747                     rx_ev_tobe_disc | rx_ev_pause_frm);
748
749         /* TOBE_DISC is expected on unicast mismatches; don't print out an
750          * error message.  FRM_TRUNC indicates RXDP dropped the packet due
751          * to a FIFO overflow.
752          */
753 #ifdef EFX_ENABLE_DEBUG
754         if (rx_ev_other_err) {
755                 EFX_INFO_RL(efx, " RX queue %d unexpected RX event "
756                             EFX_QWORD_FMT "%s%s%s%s%s%s%s%s\n",
757                             rx_queue->queue, EFX_QWORD_VAL(*event),
758                             rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
759                             rx_ev_ip_hdr_chksum_err ?
760                             " [IP_HDR_CHKSUM_ERR]" : "",
761                             rx_ev_tcp_udp_chksum_err ?
762                             " [TCP_UDP_CHKSUM_ERR]" : "",
763                             rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
764                             rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
765                             rx_ev_drib_nib ? " [DRIB_NIB]" : "",
766                             rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
767                             rx_ev_pause_frm ? " [PAUSE]" : "");
768         }
769 #endif
770 }
771
772 /* Handle receive events that are not in-order. */
773 static void
774 efx_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
775 {
776         struct efx_nic *efx = rx_queue->efx;
777         unsigned expected, dropped;
778
779         expected = rx_queue->removed_count & EFX_RXQ_MASK;
780         dropped = (index - expected) & EFX_RXQ_MASK;
781         EFX_INFO(efx, "dropped %d events (index=%d expected=%d)\n",
782                 dropped, index, expected);
783
784         efx_schedule_reset(efx, EFX_WORKAROUND_5676(efx) ?
785                            RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
786 }
787
788 /* Handle a packet received event
789  *
790  * The NIC gives a "discard" flag if it's a unicast packet with the
791  * wrong destination address
792  * Also "is multicast" and "matches multicast filter" flags can be used to
793  * discard non-matching multicast packets.
794  */
795 static void
796 efx_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
797 {
798         unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
799         unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
800         unsigned expected_ptr;
801         bool rx_ev_pkt_ok, discard = false, checksummed;
802         struct efx_rx_queue *rx_queue;
803         struct efx_nic *efx = channel->efx;
804
805         /* Basic packet information */
806         rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
807         rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
808         rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
809         WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT));
810         WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP) != 1);
811         WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
812                 channel->channel);
813
814         rx_queue = &efx->rx_queue[channel->channel];
815
816         rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
817         expected_ptr = rx_queue->removed_count & EFX_RXQ_MASK;
818         if (unlikely(rx_ev_desc_ptr != expected_ptr))
819                 efx_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr);
820
821         if (likely(rx_ev_pkt_ok)) {
822                 /* If packet is marked as OK and packet type is TCP/IP or
823                  * UDP/IP, then we can rely on the hardware checksum.
824                  */
825                 checksummed =
826                         likely(efx->rx_checksum_enabled) &&
827                         (rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP ||
828                          rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP);
829         } else {
830                 efx_handle_rx_not_ok(rx_queue, event, &rx_ev_pkt_ok, &discard);
831                 checksummed = false;
832         }
833
834         /* Detect multicast packets that didn't match the filter */
835         rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
836         if (rx_ev_mcast_pkt) {
837                 unsigned int rx_ev_mcast_hash_match =
838                         EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
839
840                 if (unlikely(!rx_ev_mcast_hash_match)) {
841                         ++channel->n_rx_mcast_mismatch;
842                         discard = true;
843                 }
844         }
845
846         channel->irq_mod_score += 2;
847
848         /* Handle received packet */
849         efx_rx_packet(rx_queue, rx_ev_desc_ptr, rx_ev_byte_cnt,
850                       checksummed, discard);
851 }
852
853 /* Global events are basically PHY events */
854 static void
855 efx_handle_global_event(struct efx_channel *channel, efx_qword_t *event)
856 {
857         struct efx_nic *efx = channel->efx;
858         bool handled = false;
859
860         if (EFX_QWORD_FIELD(*event, FSF_AB_GLB_EV_G_PHY0_INTR) ||
861             EFX_QWORD_FIELD(*event, FSF_AB_GLB_EV_XG_PHY0_INTR) ||
862             EFX_QWORD_FIELD(*event, FSF_AB_GLB_EV_XFP_PHY0_INTR)) {
863                 /* Ignored */
864                 handled = true;
865         }
866
867         if ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) &&
868             EFX_QWORD_FIELD(*event, FSF_BB_GLB_EV_XG_MGT_INTR)) {
869                 efx->xmac_poll_required = true;
870                 handled = true;
871         }
872
873         if (efx_nic_rev(efx) <= EFX_REV_FALCON_A1 ?
874             EFX_QWORD_FIELD(*event, FSF_AA_GLB_EV_RX_RECOVERY) :
875             EFX_QWORD_FIELD(*event, FSF_BB_GLB_EV_RX_RECOVERY)) {
876                 EFX_ERR(efx, "channel %d seen global RX_RESET "
877                         "event. Resetting.\n", channel->channel);
878
879                 atomic_inc(&efx->rx_reset);
880                 efx_schedule_reset(efx, EFX_WORKAROUND_6555(efx) ?
881                                    RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
882                 handled = true;
883         }
884
885         if (!handled)
886                 EFX_ERR(efx, "channel %d unknown global event "
887                         EFX_QWORD_FMT "\n", channel->channel,
888                         EFX_QWORD_VAL(*event));
889 }
890
891 static void
892 efx_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
893 {
894         struct efx_nic *efx = channel->efx;
895         unsigned int ev_sub_code;
896         unsigned int ev_sub_data;
897
898         ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
899         ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
900
901         switch (ev_sub_code) {
902         case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
903                 EFX_TRACE(efx, "channel %d TXQ %d flushed\n",
904                           channel->channel, ev_sub_data);
905                 break;
906         case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
907                 EFX_TRACE(efx, "channel %d RXQ %d flushed\n",
908                           channel->channel, ev_sub_data);
909                 break;
910         case FSE_AZ_EVQ_INIT_DONE_EV:
911                 EFX_LOG(efx, "channel %d EVQ %d initialised\n",
912                         channel->channel, ev_sub_data);
913                 break;
914         case FSE_AZ_SRM_UPD_DONE_EV:
915                 EFX_TRACE(efx, "channel %d SRAM update done\n",
916                           channel->channel);
917                 break;
918         case FSE_AZ_WAKE_UP_EV:
919                 EFX_TRACE(efx, "channel %d RXQ %d wakeup event\n",
920                           channel->channel, ev_sub_data);
921                 break;
922         case FSE_AZ_TIMER_EV:
923                 EFX_TRACE(efx, "channel %d RX queue %d timer expired\n",
924                           channel->channel, ev_sub_data);
925                 break;
926         case FSE_AA_RX_RECOVER_EV:
927                 EFX_ERR(efx, "channel %d seen DRIVER RX_RESET event. "
928                         "Resetting.\n", channel->channel);
929                 atomic_inc(&efx->rx_reset);
930                 efx_schedule_reset(efx,
931                                    EFX_WORKAROUND_6555(efx) ?
932                                    RESET_TYPE_RX_RECOVERY :
933                                    RESET_TYPE_DISABLE);
934                 break;
935         case FSE_BZ_RX_DSC_ERROR_EV:
936                 EFX_ERR(efx, "RX DMA Q %d reports descriptor fetch error."
937                         " RX Q %d is disabled.\n", ev_sub_data, ev_sub_data);
938                 efx_schedule_reset(efx, RESET_TYPE_RX_DESC_FETCH);
939                 break;
940         case FSE_BZ_TX_DSC_ERROR_EV:
941                 EFX_ERR(efx, "TX DMA Q %d reports descriptor fetch error."
942                         " TX Q %d is disabled.\n", ev_sub_data, ev_sub_data);
943                 efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
944                 break;
945         default:
946                 EFX_TRACE(efx, "channel %d unknown driver event code %d "
947                           "data %04x\n", channel->channel, ev_sub_code,
948                           ev_sub_data);
949                 break;
950         }
951 }
952
953 int efx_nic_process_eventq(struct efx_channel *channel, int budget)
954 {
955         unsigned int read_ptr;
956         efx_qword_t event, *p_event;
957         int ev_code;
958         int tx_packets = 0;
959         int spent = 0;
960
961         read_ptr = channel->eventq_read_ptr;
962
963         for (;;) {
964                 p_event = efx_event(channel, read_ptr);
965                 event = *p_event;
966
967                 if (!efx_event_present(&event))
968                         /* End of events */
969                         break;
970
971                 EFX_TRACE(channel->efx, "channel %d event is "EFX_QWORD_FMT"\n",
972                           channel->channel, EFX_QWORD_VAL(event));
973
974                 /* Clear this event by marking it all ones */
975                 EFX_SET_QWORD(*p_event);
976
977                 /* Increment read pointer */
978                 read_ptr = (read_ptr + 1) & EFX_EVQ_MASK;
979
980                 ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
981
982                 switch (ev_code) {
983                 case FSE_AZ_EV_CODE_RX_EV:
984                         efx_handle_rx_event(channel, &event);
985                         if (++spent == budget)
986                                 goto out;
987                         break;
988                 case FSE_AZ_EV_CODE_TX_EV:
989                         tx_packets += efx_handle_tx_event(channel, &event);
990                         if (tx_packets >= EFX_TXQ_SIZE) {
991                                 spent = budget;
992                                 goto out;
993                         }
994                         break;
995                 case FSE_AZ_EV_CODE_DRV_GEN_EV:
996                         channel->eventq_magic = EFX_QWORD_FIELD(
997                                 event, FSF_AZ_DRV_GEN_EV_MAGIC);
998                         EFX_LOG(channel->efx, "channel %d received generated "
999                                 "event "EFX_QWORD_FMT"\n", channel->channel,
1000                                 EFX_QWORD_VAL(event));
1001                         break;
1002                 case FSE_AZ_EV_CODE_GLOBAL_EV:
1003                         efx_handle_global_event(channel, &event);
1004                         break;
1005                 case FSE_AZ_EV_CODE_DRIVER_EV:
1006                         efx_handle_driver_event(channel, &event);
1007                         break;
1008                 case FSE_CZ_EV_CODE_MCDI_EV:
1009                         efx_mcdi_process_event(channel, &event);
1010                         break;
1011                 default:
1012                         EFX_ERR(channel->efx, "channel %d unknown event type %d"
1013                                 " (data " EFX_QWORD_FMT ")\n", channel->channel,
1014                                 ev_code, EFX_QWORD_VAL(event));
1015                 }
1016         }
1017
1018 out:
1019         channel->eventq_read_ptr = read_ptr;
1020         return spent;
1021 }
1022
1023
1024 /* Allocate buffer table entries for event queue */
1025 int efx_nic_probe_eventq(struct efx_channel *channel)
1026 {
1027         struct efx_nic *efx = channel->efx;
1028         BUILD_BUG_ON(EFX_EVQ_SIZE < 512 || EFX_EVQ_SIZE > 32768 ||
1029                      EFX_EVQ_SIZE & EFX_EVQ_MASK);
1030         return efx_alloc_special_buffer(efx, &channel->eventq,
1031                                         EFX_EVQ_SIZE * sizeof(efx_qword_t));
1032 }
1033
1034 void efx_nic_init_eventq(struct efx_channel *channel)
1035 {
1036         efx_oword_t reg;
1037         struct efx_nic *efx = channel->efx;
1038
1039         EFX_LOG(efx, "channel %d event queue in special buffers %d-%d\n",
1040                 channel->channel, channel->eventq.index,
1041                 channel->eventq.index + channel->eventq.entries - 1);
1042
1043         if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1044                 EFX_POPULATE_OWORD_3(reg,
1045                                      FRF_CZ_TIMER_Q_EN, 1,
1046                                      FRF_CZ_HOST_NOTIFY_MODE, 0,
1047                                      FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
1048                 efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1049         }
1050
1051         /* Pin event queue buffer */
1052         efx_init_special_buffer(efx, &channel->eventq);
1053
1054         /* Fill event queue with all ones (i.e. empty events) */
1055         memset(channel->eventq.addr, 0xff, channel->eventq.len);
1056
1057         /* Push event queue to card */
1058         EFX_POPULATE_OWORD_3(reg,
1059                              FRF_AZ_EVQ_EN, 1,
1060                              FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
1061                              FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
1062         efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1063                          channel->channel);
1064
1065         efx->type->push_irq_moderation(channel);
1066 }
1067
1068 void efx_nic_fini_eventq(struct efx_channel *channel)
1069 {
1070         efx_oword_t reg;
1071         struct efx_nic *efx = channel->efx;
1072
1073         /* Remove event queue from card */
1074         EFX_ZERO_OWORD(reg);
1075         efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1076                          channel->channel);
1077         if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1078                 efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1079
1080         /* Unpin event queue */
1081         efx_fini_special_buffer(efx, &channel->eventq);
1082 }
1083
1084 /* Free buffers backing event queue */
1085 void efx_nic_remove_eventq(struct efx_channel *channel)
1086 {
1087         efx_free_special_buffer(channel->efx, &channel->eventq);
1088 }
1089
1090
1091 /* Generates a test event on the event queue.  A subsequent call to
1092  * process_eventq() should pick up the event and place the value of
1093  * "magic" into channel->eventq_magic;
1094  */
1095 void efx_nic_generate_test_event(struct efx_channel *channel, unsigned int magic)
1096 {
1097         efx_qword_t test_event;
1098
1099         EFX_POPULATE_QWORD_2(test_event, FSF_AZ_EV_CODE,
1100                              FSE_AZ_EV_CODE_DRV_GEN_EV,
1101                              FSF_AZ_DRV_GEN_EV_MAGIC, magic);
1102         efx_generate_event(channel, &test_event);
1103 }
1104
1105 /**************************************************************************
1106  *
1107  * Flush handling
1108  *
1109  **************************************************************************/
1110
1111
1112 static void efx_poll_flush_events(struct efx_nic *efx)
1113 {
1114         struct efx_channel *channel = &efx->channel[0];
1115         struct efx_tx_queue *tx_queue;
1116         struct efx_rx_queue *rx_queue;
1117         unsigned int read_ptr = channel->eventq_read_ptr;
1118         unsigned int end_ptr = (read_ptr - 1) & EFX_EVQ_MASK;
1119
1120         do {
1121                 efx_qword_t *event = efx_event(channel, read_ptr);
1122                 int ev_code, ev_sub_code, ev_queue;
1123                 bool ev_failed;
1124
1125                 if (!efx_event_present(event))
1126                         break;
1127
1128                 ev_code = EFX_QWORD_FIELD(*event, FSF_AZ_EV_CODE);
1129                 ev_sub_code = EFX_QWORD_FIELD(*event,
1130                                               FSF_AZ_DRIVER_EV_SUBCODE);
1131                 if (ev_code == FSE_AZ_EV_CODE_DRIVER_EV &&
1132                     ev_sub_code == FSE_AZ_TX_DESCQ_FLS_DONE_EV) {
1133                         ev_queue = EFX_QWORD_FIELD(*event,
1134                                                    FSF_AZ_DRIVER_EV_SUBDATA);
1135                         if (ev_queue < EFX_TXQ_TYPES * efx->n_tx_channels) {
1136                                 tx_queue = efx->tx_queue + ev_queue;
1137                                 tx_queue->flushed = FLUSH_DONE;
1138                         }
1139                 } else if (ev_code == FSE_AZ_EV_CODE_DRIVER_EV &&
1140                            ev_sub_code == FSE_AZ_RX_DESCQ_FLS_DONE_EV) {
1141                         ev_queue = EFX_QWORD_FIELD(
1142                                 *event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
1143                         ev_failed = EFX_QWORD_FIELD(
1144                                 *event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
1145                         if (ev_queue < efx->n_rx_channels) {
1146                                 rx_queue = efx->rx_queue + ev_queue;
1147                                 rx_queue->flushed =
1148                                         ev_failed ? FLUSH_FAILED : FLUSH_DONE;
1149                         }
1150                 }
1151
1152                 /* We're about to destroy the queue anyway, so
1153                  * it's ok to throw away every non-flush event */
1154                 EFX_SET_QWORD(*event);
1155
1156                 read_ptr = (read_ptr + 1) & EFX_EVQ_MASK;
1157         } while (read_ptr != end_ptr);
1158
1159         channel->eventq_read_ptr = read_ptr;
1160 }
1161
1162 /* Handle tx and rx flushes at the same time, since they run in
1163  * parallel in the hardware and there's no reason for us to
1164  * serialise them */
1165 int efx_nic_flush_queues(struct efx_nic *efx)
1166 {
1167         struct efx_rx_queue *rx_queue;
1168         struct efx_tx_queue *tx_queue;
1169         int i, tx_pending, rx_pending;
1170
1171         /* If necessary prepare the hardware for flushing */
1172         efx->type->prepare_flush(efx);
1173
1174         /* Flush all tx queues in parallel */
1175         efx_for_each_tx_queue(tx_queue, efx)
1176                 efx_flush_tx_queue(tx_queue);
1177
1178         /* The hardware supports four concurrent rx flushes, each of which may
1179          * need to be retried if there is an outstanding descriptor fetch */
1180         for (i = 0; i < EFX_FLUSH_POLL_COUNT; ++i) {
1181                 rx_pending = tx_pending = 0;
1182                 efx_for_each_rx_queue(rx_queue, efx) {
1183                         if (rx_queue->flushed == FLUSH_PENDING)
1184                                 ++rx_pending;
1185                 }
1186                 efx_for_each_rx_queue(rx_queue, efx) {
1187                         if (rx_pending == EFX_RX_FLUSH_COUNT)
1188                                 break;
1189                         if (rx_queue->flushed == FLUSH_FAILED ||
1190                             rx_queue->flushed == FLUSH_NONE) {
1191                                 efx_flush_rx_queue(rx_queue);
1192                                 ++rx_pending;
1193                         }
1194                 }
1195                 efx_for_each_tx_queue(tx_queue, efx) {
1196                         if (tx_queue->flushed != FLUSH_DONE)
1197                                 ++tx_pending;
1198                 }
1199
1200                 if (rx_pending == 0 && tx_pending == 0)
1201                         return 0;
1202
1203                 msleep(EFX_FLUSH_INTERVAL);
1204                 efx_poll_flush_events(efx);
1205         }
1206
1207         /* Mark the queues as all flushed. We're going to return failure
1208          * leading to a reset, or fake up success anyway */
1209         efx_for_each_tx_queue(tx_queue, efx) {
1210                 if (tx_queue->flushed != FLUSH_DONE)
1211                         EFX_ERR(efx, "tx queue %d flush command timed out\n",
1212                                 tx_queue->queue);
1213                 tx_queue->flushed = FLUSH_DONE;
1214         }
1215         efx_for_each_rx_queue(rx_queue, efx) {
1216                 if (rx_queue->flushed != FLUSH_DONE)
1217                         EFX_ERR(efx, "rx queue %d flush command timed out\n",
1218                                 rx_queue->queue);
1219                 rx_queue->flushed = FLUSH_DONE;
1220         }
1221
1222         if (EFX_WORKAROUND_7803(efx))
1223                 return 0;
1224
1225         return -ETIMEDOUT;
1226 }
1227
1228 /**************************************************************************
1229  *
1230  * Hardware interrupts
1231  * The hardware interrupt handler does very little work; all the event
1232  * queue processing is carried out by per-channel tasklets.
1233  *
1234  **************************************************************************/
1235
1236 /* Enable/disable/generate interrupts */
1237 static inline void efx_nic_interrupts(struct efx_nic *efx,
1238                                       bool enabled, bool force)
1239 {
1240         efx_oword_t int_en_reg_ker;
1241
1242         EFX_POPULATE_OWORD_3(int_en_reg_ker,
1243                              FRF_AZ_KER_INT_LEVE_SEL, efx->fatal_irq_level,
1244                              FRF_AZ_KER_INT_KER, force,
1245                              FRF_AZ_DRV_INT_EN_KER, enabled);
1246         efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
1247 }
1248
1249 void efx_nic_enable_interrupts(struct efx_nic *efx)
1250 {
1251         struct efx_channel *channel;
1252
1253         EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
1254         wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
1255
1256         /* Enable interrupts */
1257         efx_nic_interrupts(efx, true, false);
1258
1259         /* Force processing of all the channels to get the EVQ RPTRs up to
1260            date */
1261         efx_for_each_channel(channel, efx)
1262                 efx_schedule_channel(channel);
1263 }
1264
1265 void efx_nic_disable_interrupts(struct efx_nic *efx)
1266 {
1267         /* Disable interrupts */
1268         efx_nic_interrupts(efx, false, false);
1269 }
1270
1271 /* Generate a test interrupt
1272  * Interrupt must already have been enabled, otherwise nasty things
1273  * may happen.
1274  */
1275 void efx_nic_generate_interrupt(struct efx_nic *efx)
1276 {
1277         efx_nic_interrupts(efx, true, true);
1278 }
1279
1280 /* Process a fatal interrupt
1281  * Disable bus mastering ASAP and schedule a reset
1282  */
1283 irqreturn_t efx_nic_fatal_interrupt(struct efx_nic *efx)
1284 {
1285         struct falcon_nic_data *nic_data = efx->nic_data;
1286         efx_oword_t *int_ker = efx->irq_status.addr;
1287         efx_oword_t fatal_intr;
1288         int error, mem_perr;
1289
1290         efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
1291         error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
1292
1293         EFX_ERR(efx, "SYSTEM ERROR " EFX_OWORD_FMT " status "
1294                 EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
1295                 EFX_OWORD_VAL(fatal_intr),
1296                 error ? "disabling bus mastering" : "no recognised error");
1297
1298         /* If this is a memory parity error dump which blocks are offending */
1299         mem_perr = (EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER) ||
1300                     EFX_OWORD_FIELD(fatal_intr, FRF_AZ_SRM_PERR_INT_KER));
1301         if (mem_perr) {
1302                 efx_oword_t reg;
1303                 efx_reado(efx, &reg, FR_AZ_MEM_STAT);
1304                 EFX_ERR(efx, "SYSTEM ERROR: memory parity error "
1305                         EFX_OWORD_FMT "\n", EFX_OWORD_VAL(reg));
1306         }
1307
1308         /* Disable both devices */
1309         pci_clear_master(efx->pci_dev);
1310         if (efx_nic_is_dual_func(efx))
1311                 pci_clear_master(nic_data->pci_dev2);
1312         efx_nic_disable_interrupts(efx);
1313
1314         /* Count errors and reset or disable the NIC accordingly */
1315         if (efx->int_error_count == 0 ||
1316             time_after(jiffies, efx->int_error_expire)) {
1317                 efx->int_error_count = 0;
1318                 efx->int_error_expire =
1319                         jiffies + EFX_INT_ERROR_EXPIRE * HZ;
1320         }
1321         if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
1322                 EFX_ERR(efx, "SYSTEM ERROR - reset scheduled\n");
1323                 efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
1324         } else {
1325                 EFX_ERR(efx, "SYSTEM ERROR - max number of errors seen."
1326                         "NIC will be disabled\n");
1327                 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1328         }
1329
1330         return IRQ_HANDLED;
1331 }
1332
1333 /* Handle a legacy interrupt
1334  * Acknowledges the interrupt and schedule event queue processing.
1335  */
1336 static irqreturn_t efx_legacy_interrupt(int irq, void *dev_id)
1337 {
1338         struct efx_nic *efx = dev_id;
1339         efx_oword_t *int_ker = efx->irq_status.addr;
1340         irqreturn_t result = IRQ_NONE;
1341         struct efx_channel *channel;
1342         efx_dword_t reg;
1343         u32 queues;
1344         int syserr;
1345
1346         /* Read the ISR which also ACKs the interrupts */
1347         efx_readd(efx, &reg, FR_BZ_INT_ISR0);
1348         queues = EFX_EXTRACT_DWORD(reg, 0, 31);
1349
1350         /* Check to see if we have a serious error condition */
1351         if (queues & (1U << efx->fatal_irq_level)) {
1352                 syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1353                 if (unlikely(syserr))
1354                         return efx_nic_fatal_interrupt(efx);
1355         }
1356
1357         if (queues != 0) {
1358                 if (EFX_WORKAROUND_15783(efx))
1359                         efx->irq_zero_count = 0;
1360
1361                 /* Schedule processing of any interrupting queues */
1362                 efx_for_each_channel(channel, efx) {
1363                         if (queues & 1)
1364                                 efx_schedule_channel(channel);
1365                         queues >>= 1;
1366                 }
1367                 result = IRQ_HANDLED;
1368
1369         } else if (EFX_WORKAROUND_15783(efx)) {
1370                 efx_qword_t *event;
1371
1372                 /* We can't return IRQ_HANDLED more than once on seeing ISR=0
1373                  * because this might be a shared interrupt. */
1374                 if (efx->irq_zero_count++ == 0)
1375                         result = IRQ_HANDLED;
1376
1377                 /* Ensure we schedule or rearm all event queues */
1378                 efx_for_each_channel(channel, efx) {
1379                         event = efx_event(channel, channel->eventq_read_ptr);
1380                         if (efx_event_present(event))
1381                                 efx_schedule_channel(channel);
1382                         else
1383                                 efx_nic_eventq_read_ack(channel);
1384                 }
1385         }
1386
1387         if (result == IRQ_HANDLED) {
1388                 efx->last_irq_cpu = raw_smp_processor_id();
1389                 EFX_TRACE(efx, "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
1390                           irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
1391         }
1392
1393         return result;
1394 }
1395
1396 /* Handle an MSI interrupt
1397  *
1398  * Handle an MSI hardware interrupt.  This routine schedules event
1399  * queue processing.  No interrupt acknowledgement cycle is necessary.
1400  * Also, we never need to check that the interrupt is for us, since
1401  * MSI interrupts cannot be shared.
1402  */
1403 static irqreturn_t efx_msi_interrupt(int irq, void *dev_id)
1404 {
1405         struct efx_channel *channel = dev_id;
1406         struct efx_nic *efx = channel->efx;
1407         efx_oword_t *int_ker = efx->irq_status.addr;
1408         int syserr;
1409
1410         efx->last_irq_cpu = raw_smp_processor_id();
1411         EFX_TRACE(efx, "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
1412                   irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
1413
1414         /* Check to see if we have a serious error condition */
1415         if (channel->channel == efx->fatal_irq_level) {
1416                 syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1417                 if (unlikely(syserr))
1418                         return efx_nic_fatal_interrupt(efx);
1419         }
1420
1421         /* Schedule processing of the channel */
1422         efx_schedule_channel(channel);
1423
1424         return IRQ_HANDLED;
1425 }
1426
1427
1428 /* Setup RSS indirection table.
1429  * This maps from the hash value of the packet to RXQ
1430  */
1431 static void efx_setup_rss_indir_table(struct efx_nic *efx)
1432 {
1433         int i = 0;
1434         unsigned long offset;
1435         efx_dword_t dword;
1436
1437         if (efx_nic_rev(efx) < EFX_REV_FALCON_B0)
1438                 return;
1439
1440         for (offset = FR_BZ_RX_INDIRECTION_TBL;
1441              offset < FR_BZ_RX_INDIRECTION_TBL + 0x800;
1442              offset += 0x10) {
1443                 EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
1444                                      i % efx->n_rx_channels);
1445                 efx_writed(efx, &dword, offset);
1446                 i++;
1447         }
1448 }
1449
1450 /* Hook interrupt handler(s)
1451  * Try MSI and then legacy interrupts.
1452  */
1453 int efx_nic_init_interrupt(struct efx_nic *efx)
1454 {
1455         struct efx_channel *channel;
1456         int rc;
1457
1458         if (!EFX_INT_MODE_USE_MSI(efx)) {
1459                 irq_handler_t handler;
1460                 if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1461                         handler = efx_legacy_interrupt;
1462                 else
1463                         handler = falcon_legacy_interrupt_a1;
1464
1465                 rc = request_irq(efx->legacy_irq, handler, IRQF_SHARED,
1466                                  efx->name, efx);
1467                 if (rc) {
1468                         EFX_ERR(efx, "failed to hook legacy IRQ %d\n",
1469                                 efx->pci_dev->irq);
1470                         goto fail1;
1471                 }
1472                 return 0;
1473         }
1474
1475         /* Hook MSI or MSI-X interrupt */
1476         efx_for_each_channel(channel, efx) {
1477                 rc = request_irq(channel->irq, efx_msi_interrupt,
1478                                  IRQF_PROBE_SHARED, /* Not shared */
1479                                  channel->name, channel);
1480                 if (rc) {
1481                         EFX_ERR(efx, "failed to hook IRQ %d\n", channel->irq);
1482                         goto fail2;
1483                 }
1484         }
1485
1486         return 0;
1487
1488  fail2:
1489         efx_for_each_channel(channel, efx)
1490                 free_irq(channel->irq, channel);
1491  fail1:
1492         return rc;
1493 }
1494
1495 void efx_nic_fini_interrupt(struct efx_nic *efx)
1496 {
1497         struct efx_channel *channel;
1498         efx_oword_t reg;
1499
1500         /* Disable MSI/MSI-X interrupts */
1501         efx_for_each_channel(channel, efx) {
1502                 if (channel->irq)
1503                         free_irq(channel->irq, channel);
1504         }
1505
1506         /* ACK legacy interrupt */
1507         if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1508                 efx_reado(efx, &reg, FR_BZ_INT_ISR0);
1509         else
1510                 falcon_irq_ack_a1(efx);
1511
1512         /* Disable legacy interrupt */
1513         if (efx->legacy_irq)
1514                 free_irq(efx->legacy_irq, efx);
1515 }
1516
1517 u32 efx_nic_fpga_ver(struct efx_nic *efx)
1518 {
1519         efx_oword_t altera_build;
1520         efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
1521         return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
1522 }
1523
1524 void efx_nic_init_common(struct efx_nic *efx)
1525 {
1526         efx_oword_t temp;
1527
1528         /* Set positions of descriptor caches in SRAM. */
1529         EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR,
1530                              efx->type->tx_dc_base / 8);
1531         efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
1532         EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR,
1533                              efx->type->rx_dc_base / 8);
1534         efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
1535
1536         /* Set TX descriptor cache size. */
1537         BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
1538         EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
1539         efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
1540
1541         /* Set RX descriptor cache size.  Set low watermark to size-8, as
1542          * this allows most efficient prefetching.
1543          */
1544         BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
1545         EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
1546         efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
1547         EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
1548         efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
1549
1550         /* Program INT_KER address */
1551         EFX_POPULATE_OWORD_2(temp,
1552                              FRF_AZ_NORM_INT_VEC_DIS_KER,
1553                              EFX_INT_MODE_USE_MSI(efx),
1554                              FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
1555         efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
1556
1557         if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
1558                 /* Use an interrupt level unused by event queues */
1559                 efx->fatal_irq_level = 0x1f;
1560         else
1561                 /* Use a valid MSI-X vector */
1562                 efx->fatal_irq_level = 0;
1563
1564         /* Enable all the genuinely fatal interrupts.  (They are still
1565          * masked by the overall interrupt mask, controlled by
1566          * falcon_interrupts()).
1567          *
1568          * Note: All other fatal interrupts are enabled
1569          */
1570         EFX_POPULATE_OWORD_3(temp,
1571                              FRF_AZ_ILL_ADR_INT_KER_EN, 1,
1572                              FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
1573                              FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
1574         if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1575                 EFX_SET_OWORD_FIELD(temp, FRF_CZ_SRAM_PERR_INT_P_KER_EN, 1);
1576         EFX_INVERT_OWORD(temp);
1577         efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
1578
1579         efx_setup_rss_indir_table(efx);
1580
1581         /* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
1582          * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
1583          */
1584         efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
1585         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
1586         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
1587         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
1588         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 0);
1589         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
1590         /* Enable SW_EV to inherit in char driver - assume harmless here */
1591         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
1592         /* Prefetch threshold 2 => fetch when descriptor cache half empty */
1593         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
1594         /* Disable hardware watchdog which can misfire */
1595         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
1596         /* Squash TX of packets of 16 bytes or less */
1597         if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1598                 EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
1599         efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
1600 }