sfc: Fix some incorrect or redundant comments
[safe/jmp/linux-2.6] / 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_TX_QUEUE_OFFLOAD_CSUM;
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_TX_QUEUE_COUNT >= 128);
435
436                 efx_reado(efx, &reg, FR_AA_TX_CHKSM_CFG);
437                 if (tx_queue->queue == EFX_TX_QUEUE_OFFLOAD_CSUM)
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 void
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
665         if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
666                 /* Transmit completion */
667                 tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
668                 tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
669                 tx_queue = &efx->tx_queue[tx_ev_q_label];
670                 channel->irq_mod_score +=
671                         (tx_ev_desc_ptr - tx_queue->read_count) &
672                         EFX_TXQ_MASK;
673                 efx_xmit_done(tx_queue, tx_ev_desc_ptr);
674         } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
675                 /* Rewrite the FIFO write pointer */
676                 tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
677                 tx_queue = &efx->tx_queue[tx_ev_q_label];
678
679                 if (efx_dev_registered(efx))
680                         netif_tx_lock(efx->net_dev);
681                 efx_notify_tx_desc(tx_queue);
682                 if (efx_dev_registered(efx))
683                         netif_tx_unlock(efx->net_dev);
684         } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR) &&
685                    EFX_WORKAROUND_10727(efx)) {
686                 efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
687         } else {
688                 EFX_ERR(efx, "channel %d unexpected TX event "
689                         EFX_QWORD_FMT"\n", channel->channel,
690                         EFX_QWORD_VAL(*event));
691         }
692 }
693
694 /* Detect errors included in the rx_evt_pkt_ok bit. */
695 static void efx_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
696                                  const efx_qword_t *event,
697                                  bool *rx_ev_pkt_ok,
698                                  bool *discard)
699 {
700         struct efx_nic *efx = rx_queue->efx;
701         bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
702         bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
703         bool rx_ev_frm_trunc, rx_ev_drib_nib, rx_ev_tobe_disc;
704         bool rx_ev_other_err, rx_ev_pause_frm;
705         bool rx_ev_hdr_type, rx_ev_mcast_pkt;
706         unsigned rx_ev_pkt_type;
707
708         rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
709         rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
710         rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
711         rx_ev_pkt_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_TYPE);
712         rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
713                                                  FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
714         rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
715                                                   FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
716         rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
717                                                    FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
718         rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
719         rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
720         rx_ev_drib_nib = ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) ?
721                           0 : EFX_QWORD_FIELD(*event, FSF_AA_RX_EV_DRIB_NIB));
722         rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
723
724         /* Every error apart from tobe_disc and pause_frm */
725         rx_ev_other_err = (rx_ev_drib_nib | rx_ev_tcp_udp_chksum_err |
726                            rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
727                            rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
728
729         /* Count errors that are not in MAC stats.  Ignore expected
730          * checksum errors during self-test. */
731         if (rx_ev_frm_trunc)
732                 ++rx_queue->channel->n_rx_frm_trunc;
733         else if (rx_ev_tobe_disc)
734                 ++rx_queue->channel->n_rx_tobe_disc;
735         else if (!efx->loopback_selftest) {
736                 if (rx_ev_ip_hdr_chksum_err)
737                         ++rx_queue->channel->n_rx_ip_hdr_chksum_err;
738                 else if (rx_ev_tcp_udp_chksum_err)
739                         ++rx_queue->channel->n_rx_tcp_udp_chksum_err;
740         }
741
742         /* The frame must be discarded if any of these are true. */
743         *discard = (rx_ev_eth_crc_err | rx_ev_frm_trunc | rx_ev_drib_nib |
744                     rx_ev_tobe_disc | rx_ev_pause_frm);
745
746         /* TOBE_DISC is expected on unicast mismatches; don't print out an
747          * error message.  FRM_TRUNC indicates RXDP dropped the packet due
748          * to a FIFO overflow.
749          */
750 #ifdef EFX_ENABLE_DEBUG
751         if (rx_ev_other_err) {
752                 EFX_INFO_RL(efx, " RX queue %d unexpected RX event "
753                             EFX_QWORD_FMT "%s%s%s%s%s%s%s%s\n",
754                             rx_queue->queue, EFX_QWORD_VAL(*event),
755                             rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
756                             rx_ev_ip_hdr_chksum_err ?
757                             " [IP_HDR_CHKSUM_ERR]" : "",
758                             rx_ev_tcp_udp_chksum_err ?
759                             " [TCP_UDP_CHKSUM_ERR]" : "",
760                             rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
761                             rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
762                             rx_ev_drib_nib ? " [DRIB_NIB]" : "",
763                             rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
764                             rx_ev_pause_frm ? " [PAUSE]" : "");
765         }
766 #endif
767 }
768
769 /* Handle receive events that are not in-order. */
770 static void
771 efx_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
772 {
773         struct efx_nic *efx = rx_queue->efx;
774         unsigned expected, dropped;
775
776         expected = rx_queue->removed_count & EFX_RXQ_MASK;
777         dropped = (index - expected) & EFX_RXQ_MASK;
778         EFX_INFO(efx, "dropped %d events (index=%d expected=%d)\n",
779                 dropped, index, expected);
780
781         efx_schedule_reset(efx, EFX_WORKAROUND_5676(efx) ?
782                            RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
783 }
784
785 /* Handle a packet received event
786  *
787  * The NIC gives a "discard" flag if it's a unicast packet with the
788  * wrong destination address
789  * Also "is multicast" and "matches multicast filter" flags can be used to
790  * discard non-matching multicast packets.
791  */
792 static void
793 efx_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
794 {
795         unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
796         unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
797         unsigned expected_ptr;
798         bool rx_ev_pkt_ok, discard = false, checksummed;
799         struct efx_rx_queue *rx_queue;
800         struct efx_nic *efx = channel->efx;
801
802         /* Basic packet information */
803         rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
804         rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
805         rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
806         WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT));
807         WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP) != 1);
808         WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
809                 channel->channel);
810
811         rx_queue = &efx->rx_queue[channel->channel];
812
813         rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
814         expected_ptr = rx_queue->removed_count & EFX_RXQ_MASK;
815         if (unlikely(rx_ev_desc_ptr != expected_ptr))
816                 efx_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr);
817
818         if (likely(rx_ev_pkt_ok)) {
819                 /* If packet is marked as OK and packet type is TCP/IP or
820                  * UDP/IP, then we can rely on the hardware checksum.
821                  */
822                 checksummed =
823                         likely(efx->rx_checksum_enabled) &&
824                         (rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP ||
825                          rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP);
826         } else {
827                 efx_handle_rx_not_ok(rx_queue, event, &rx_ev_pkt_ok, &discard);
828                 checksummed = false;
829         }
830
831         /* Detect multicast packets that didn't match the filter */
832         rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
833         if (rx_ev_mcast_pkt) {
834                 unsigned int rx_ev_mcast_hash_match =
835                         EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
836
837                 if (unlikely(!rx_ev_mcast_hash_match)) {
838                         ++channel->n_rx_mcast_mismatch;
839                         discard = true;
840                 }
841         }
842
843         channel->irq_mod_score += 2;
844
845         /* Handle received packet */
846         efx_rx_packet(rx_queue, rx_ev_desc_ptr, rx_ev_byte_cnt,
847                       checksummed, discard);
848 }
849
850 /* Global events are basically PHY events */
851 static void
852 efx_handle_global_event(struct efx_channel *channel, efx_qword_t *event)
853 {
854         struct efx_nic *efx = channel->efx;
855         bool handled = false;
856
857         if (EFX_QWORD_FIELD(*event, FSF_AB_GLB_EV_G_PHY0_INTR) ||
858             EFX_QWORD_FIELD(*event, FSF_AB_GLB_EV_XG_PHY0_INTR) ||
859             EFX_QWORD_FIELD(*event, FSF_AB_GLB_EV_XFP_PHY0_INTR)) {
860                 /* Ignored */
861                 handled = true;
862         }
863
864         if ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) &&
865             EFX_QWORD_FIELD(*event, FSF_BB_GLB_EV_XG_MGT_INTR)) {
866                 efx->xmac_poll_required = true;
867                 handled = true;
868         }
869
870         if (efx_nic_rev(efx) <= EFX_REV_FALCON_A1 ?
871             EFX_QWORD_FIELD(*event, FSF_AA_GLB_EV_RX_RECOVERY) :
872             EFX_QWORD_FIELD(*event, FSF_BB_GLB_EV_RX_RECOVERY)) {
873                 EFX_ERR(efx, "channel %d seen global RX_RESET "
874                         "event. Resetting.\n", channel->channel);
875
876                 atomic_inc(&efx->rx_reset);
877                 efx_schedule_reset(efx, EFX_WORKAROUND_6555(efx) ?
878                                    RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
879                 handled = true;
880         }
881
882         if (!handled)
883                 EFX_ERR(efx, "channel %d unknown global event "
884                         EFX_QWORD_FMT "\n", channel->channel,
885                         EFX_QWORD_VAL(*event));
886 }
887
888 static void
889 efx_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
890 {
891         struct efx_nic *efx = channel->efx;
892         unsigned int ev_sub_code;
893         unsigned int ev_sub_data;
894
895         ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
896         ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
897
898         switch (ev_sub_code) {
899         case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
900                 EFX_TRACE(efx, "channel %d TXQ %d flushed\n",
901                           channel->channel, ev_sub_data);
902                 break;
903         case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
904                 EFX_TRACE(efx, "channel %d RXQ %d flushed\n",
905                           channel->channel, ev_sub_data);
906                 break;
907         case FSE_AZ_EVQ_INIT_DONE_EV:
908                 EFX_LOG(efx, "channel %d EVQ %d initialised\n",
909                         channel->channel, ev_sub_data);
910                 break;
911         case FSE_AZ_SRM_UPD_DONE_EV:
912                 EFX_TRACE(efx, "channel %d SRAM update done\n",
913                           channel->channel);
914                 break;
915         case FSE_AZ_WAKE_UP_EV:
916                 EFX_TRACE(efx, "channel %d RXQ %d wakeup event\n",
917                           channel->channel, ev_sub_data);
918                 break;
919         case FSE_AZ_TIMER_EV:
920                 EFX_TRACE(efx, "channel %d RX queue %d timer expired\n",
921                           channel->channel, ev_sub_data);
922                 break;
923         case FSE_AA_RX_RECOVER_EV:
924                 EFX_ERR(efx, "channel %d seen DRIVER RX_RESET event. "
925                         "Resetting.\n", channel->channel);
926                 atomic_inc(&efx->rx_reset);
927                 efx_schedule_reset(efx,
928                                    EFX_WORKAROUND_6555(efx) ?
929                                    RESET_TYPE_RX_RECOVERY :
930                                    RESET_TYPE_DISABLE);
931                 break;
932         case FSE_BZ_RX_DSC_ERROR_EV:
933                 EFX_ERR(efx, "RX DMA Q %d reports descriptor fetch error."
934                         " RX Q %d is disabled.\n", ev_sub_data, ev_sub_data);
935                 efx_schedule_reset(efx, RESET_TYPE_RX_DESC_FETCH);
936                 break;
937         case FSE_BZ_TX_DSC_ERROR_EV:
938                 EFX_ERR(efx, "TX DMA Q %d reports descriptor fetch error."
939                         " TX Q %d is disabled.\n", ev_sub_data, ev_sub_data);
940                 efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
941                 break;
942         default:
943                 EFX_TRACE(efx, "channel %d unknown driver event code %d "
944                           "data %04x\n", channel->channel, ev_sub_code,
945                           ev_sub_data);
946                 break;
947         }
948 }
949
950 int efx_nic_process_eventq(struct efx_channel *channel, int rx_quota)
951 {
952         unsigned int read_ptr;
953         efx_qword_t event, *p_event;
954         int ev_code;
955         int rx_packets = 0;
956
957         read_ptr = channel->eventq_read_ptr;
958
959         do {
960                 p_event = efx_event(channel, read_ptr);
961                 event = *p_event;
962
963                 if (!efx_event_present(&event))
964                         /* End of events */
965                         break;
966
967                 EFX_TRACE(channel->efx, "channel %d event is "EFX_QWORD_FMT"\n",
968                           channel->channel, EFX_QWORD_VAL(event));
969
970                 /* Clear this event by marking it all ones */
971                 EFX_SET_QWORD(*p_event);
972
973                 ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
974
975                 switch (ev_code) {
976                 case FSE_AZ_EV_CODE_RX_EV:
977                         efx_handle_rx_event(channel, &event);
978                         ++rx_packets;
979                         break;
980                 case FSE_AZ_EV_CODE_TX_EV:
981                         efx_handle_tx_event(channel, &event);
982                         break;
983                 case FSE_AZ_EV_CODE_DRV_GEN_EV:
984                         channel->eventq_magic = EFX_QWORD_FIELD(
985                                 event, FSF_AZ_DRV_GEN_EV_MAGIC);
986                         EFX_LOG(channel->efx, "channel %d received generated "
987                                 "event "EFX_QWORD_FMT"\n", channel->channel,
988                                 EFX_QWORD_VAL(event));
989                         break;
990                 case FSE_AZ_EV_CODE_GLOBAL_EV:
991                         efx_handle_global_event(channel, &event);
992                         break;
993                 case FSE_AZ_EV_CODE_DRIVER_EV:
994                         efx_handle_driver_event(channel, &event);
995                         break;
996                 case FSE_CZ_EV_CODE_MCDI_EV:
997                         efx_mcdi_process_event(channel, &event);
998                         break;
999                 default:
1000                         EFX_ERR(channel->efx, "channel %d unknown event type %d"
1001                                 " (data " EFX_QWORD_FMT ")\n", channel->channel,
1002                                 ev_code, EFX_QWORD_VAL(event));
1003                 }
1004
1005                 /* Increment read pointer */
1006                 read_ptr = (read_ptr + 1) & EFX_EVQ_MASK;
1007
1008         } while (rx_packets < rx_quota);
1009
1010         channel->eventq_read_ptr = read_ptr;
1011         return rx_packets;
1012 }
1013
1014
1015 /* Allocate buffer table entries for event queue */
1016 int efx_nic_probe_eventq(struct efx_channel *channel)
1017 {
1018         struct efx_nic *efx = channel->efx;
1019         BUILD_BUG_ON(EFX_EVQ_SIZE < 512 || EFX_EVQ_SIZE > 32768 ||
1020                      EFX_EVQ_SIZE & EFX_EVQ_MASK);
1021         return efx_alloc_special_buffer(efx, &channel->eventq,
1022                                         EFX_EVQ_SIZE * sizeof(efx_qword_t));
1023 }
1024
1025 void efx_nic_init_eventq(struct efx_channel *channel)
1026 {
1027         efx_oword_t reg;
1028         struct efx_nic *efx = channel->efx;
1029
1030         EFX_LOG(efx, "channel %d event queue in special buffers %d-%d\n",
1031                 channel->channel, channel->eventq.index,
1032                 channel->eventq.index + channel->eventq.entries - 1);
1033
1034         if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1035                 EFX_POPULATE_OWORD_3(reg,
1036                                      FRF_CZ_TIMER_Q_EN, 1,
1037                                      FRF_CZ_HOST_NOTIFY_MODE, 0,
1038                                      FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
1039                 efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1040         }
1041
1042         /* Pin event queue buffer */
1043         efx_init_special_buffer(efx, &channel->eventq);
1044
1045         /* Fill event queue with all ones (i.e. empty events) */
1046         memset(channel->eventq.addr, 0xff, channel->eventq.len);
1047
1048         /* Push event queue to card */
1049         EFX_POPULATE_OWORD_3(reg,
1050                              FRF_AZ_EVQ_EN, 1,
1051                              FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
1052                              FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
1053         efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1054                          channel->channel);
1055
1056         efx->type->push_irq_moderation(channel);
1057 }
1058
1059 void efx_nic_fini_eventq(struct efx_channel *channel)
1060 {
1061         efx_oword_t reg;
1062         struct efx_nic *efx = channel->efx;
1063
1064         /* Remove event queue from card */
1065         EFX_ZERO_OWORD(reg);
1066         efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1067                          channel->channel);
1068         if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1069                 efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1070
1071         /* Unpin event queue */
1072         efx_fini_special_buffer(efx, &channel->eventq);
1073 }
1074
1075 /* Free buffers backing event queue */
1076 void efx_nic_remove_eventq(struct efx_channel *channel)
1077 {
1078         efx_free_special_buffer(channel->efx, &channel->eventq);
1079 }
1080
1081
1082 /* Generates a test event on the event queue.  A subsequent call to
1083  * process_eventq() should pick up the event and place the value of
1084  * "magic" into channel->eventq_magic;
1085  */
1086 void efx_nic_generate_test_event(struct efx_channel *channel, unsigned int magic)
1087 {
1088         efx_qword_t test_event;
1089
1090         EFX_POPULATE_QWORD_2(test_event, FSF_AZ_EV_CODE,
1091                              FSE_AZ_EV_CODE_DRV_GEN_EV,
1092                              FSF_AZ_DRV_GEN_EV_MAGIC, magic);
1093         efx_generate_event(channel, &test_event);
1094 }
1095
1096 /**************************************************************************
1097  *
1098  * Flush handling
1099  *
1100  **************************************************************************/
1101
1102
1103 static void efx_poll_flush_events(struct efx_nic *efx)
1104 {
1105         struct efx_channel *channel = &efx->channel[0];
1106         struct efx_tx_queue *tx_queue;
1107         struct efx_rx_queue *rx_queue;
1108         unsigned int read_ptr = channel->eventq_read_ptr;
1109         unsigned int end_ptr = (read_ptr - 1) & EFX_EVQ_MASK;
1110
1111         do {
1112                 efx_qword_t *event = efx_event(channel, read_ptr);
1113                 int ev_code, ev_sub_code, ev_queue;
1114                 bool ev_failed;
1115
1116                 if (!efx_event_present(event))
1117                         break;
1118
1119                 ev_code = EFX_QWORD_FIELD(*event, FSF_AZ_EV_CODE);
1120                 ev_sub_code = EFX_QWORD_FIELD(*event,
1121                                               FSF_AZ_DRIVER_EV_SUBCODE);
1122                 if (ev_code == FSE_AZ_EV_CODE_DRIVER_EV &&
1123                     ev_sub_code == FSE_AZ_TX_DESCQ_FLS_DONE_EV) {
1124                         ev_queue = EFX_QWORD_FIELD(*event,
1125                                                    FSF_AZ_DRIVER_EV_SUBDATA);
1126                         if (ev_queue < EFX_TX_QUEUE_COUNT) {
1127                                 tx_queue = efx->tx_queue + ev_queue;
1128                                 tx_queue->flushed = FLUSH_DONE;
1129                         }
1130                 } else if (ev_code == FSE_AZ_EV_CODE_DRIVER_EV &&
1131                            ev_sub_code == FSE_AZ_RX_DESCQ_FLS_DONE_EV) {
1132                         ev_queue = EFX_QWORD_FIELD(
1133                                 *event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
1134                         ev_failed = EFX_QWORD_FIELD(
1135                                 *event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
1136                         if (ev_queue < efx->n_rx_queues) {
1137                                 rx_queue = efx->rx_queue + ev_queue;
1138                                 rx_queue->flushed =
1139                                         ev_failed ? FLUSH_FAILED : FLUSH_DONE;
1140                         }
1141                 }
1142
1143                 /* We're about to destroy the queue anyway, so
1144                  * it's ok to throw away every non-flush event */
1145                 EFX_SET_QWORD(*event);
1146
1147                 read_ptr = (read_ptr + 1) & EFX_EVQ_MASK;
1148         } while (read_ptr != end_ptr);
1149
1150         channel->eventq_read_ptr = read_ptr;
1151 }
1152
1153 /* Handle tx and rx flushes at the same time, since they run in
1154  * parallel in the hardware and there's no reason for us to
1155  * serialise them */
1156 int efx_nic_flush_queues(struct efx_nic *efx)
1157 {
1158         struct efx_rx_queue *rx_queue;
1159         struct efx_tx_queue *tx_queue;
1160         int i, tx_pending, rx_pending;
1161
1162         /* If necessary prepare the hardware for flushing */
1163         efx->type->prepare_flush(efx);
1164
1165         /* Flush all tx queues in parallel */
1166         efx_for_each_tx_queue(tx_queue, efx)
1167                 efx_flush_tx_queue(tx_queue);
1168
1169         /* The hardware supports four concurrent rx flushes, each of which may
1170          * need to be retried if there is an outstanding descriptor fetch */
1171         for (i = 0; i < EFX_FLUSH_POLL_COUNT; ++i) {
1172                 rx_pending = tx_pending = 0;
1173                 efx_for_each_rx_queue(rx_queue, efx) {
1174                         if (rx_queue->flushed == FLUSH_PENDING)
1175                                 ++rx_pending;
1176                 }
1177                 efx_for_each_rx_queue(rx_queue, efx) {
1178                         if (rx_pending == EFX_RX_FLUSH_COUNT)
1179                                 break;
1180                         if (rx_queue->flushed == FLUSH_FAILED ||
1181                             rx_queue->flushed == FLUSH_NONE) {
1182                                 efx_flush_rx_queue(rx_queue);
1183                                 ++rx_pending;
1184                         }
1185                 }
1186                 efx_for_each_tx_queue(tx_queue, efx) {
1187                         if (tx_queue->flushed != FLUSH_DONE)
1188                                 ++tx_pending;
1189                 }
1190
1191                 if (rx_pending == 0 && tx_pending == 0)
1192                         return 0;
1193
1194                 msleep(EFX_FLUSH_INTERVAL);
1195                 efx_poll_flush_events(efx);
1196         }
1197
1198         /* Mark the queues as all flushed. We're going to return failure
1199          * leading to a reset, or fake up success anyway */
1200         efx_for_each_tx_queue(tx_queue, efx) {
1201                 if (tx_queue->flushed != FLUSH_DONE)
1202                         EFX_ERR(efx, "tx queue %d flush command timed out\n",
1203                                 tx_queue->queue);
1204                 tx_queue->flushed = FLUSH_DONE;
1205         }
1206         efx_for_each_rx_queue(rx_queue, efx) {
1207                 if (rx_queue->flushed != FLUSH_DONE)
1208                         EFX_ERR(efx, "rx queue %d flush command timed out\n",
1209                                 rx_queue->queue);
1210                 rx_queue->flushed = FLUSH_DONE;
1211         }
1212
1213         if (EFX_WORKAROUND_7803(efx))
1214                 return 0;
1215
1216         return -ETIMEDOUT;
1217 }
1218
1219 /**************************************************************************
1220  *
1221  * Hardware interrupts
1222  * The hardware interrupt handler does very little work; all the event
1223  * queue processing is carried out by per-channel tasklets.
1224  *
1225  **************************************************************************/
1226
1227 /* Enable/disable/generate interrupts */
1228 static inline void efx_nic_interrupts(struct efx_nic *efx,
1229                                       bool enabled, bool force)
1230 {
1231         efx_oword_t int_en_reg_ker;
1232         unsigned int level = 0;
1233
1234         if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
1235                 /* Set the level always even if we're generating a test
1236                  * interrupt, because our legacy interrupt handler is safe */
1237                 level = 0x1f;
1238
1239         EFX_POPULATE_OWORD_3(int_en_reg_ker,
1240                              FRF_AZ_KER_INT_LEVE_SEL, level,
1241                              FRF_AZ_KER_INT_KER, force,
1242                              FRF_AZ_DRV_INT_EN_KER, enabled);
1243         efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
1244 }
1245
1246 void efx_nic_enable_interrupts(struct efx_nic *efx)
1247 {
1248         struct efx_channel *channel;
1249
1250         EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
1251         wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
1252
1253         /* Enable interrupts */
1254         efx_nic_interrupts(efx, true, false);
1255
1256         /* Force processing of all the channels to get the EVQ RPTRs up to
1257            date */
1258         efx_for_each_channel(channel, efx)
1259                 efx_schedule_channel(channel);
1260 }
1261
1262 void efx_nic_disable_interrupts(struct efx_nic *efx)
1263 {
1264         /* Disable interrupts */
1265         efx_nic_interrupts(efx, false, false);
1266 }
1267
1268 /* Generate a test interrupt
1269  * Interrupt must already have been enabled, otherwise nasty things
1270  * may happen.
1271  */
1272 void efx_nic_generate_interrupt(struct efx_nic *efx)
1273 {
1274         efx_nic_interrupts(efx, true, true);
1275 }
1276
1277 /* Process a fatal interrupt
1278  * Disable bus mastering ASAP and schedule a reset
1279  */
1280 irqreturn_t efx_nic_fatal_interrupt(struct efx_nic *efx)
1281 {
1282         struct falcon_nic_data *nic_data = efx->nic_data;
1283         efx_oword_t *int_ker = efx->irq_status.addr;
1284         efx_oword_t fatal_intr;
1285         int error, mem_perr;
1286
1287         efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
1288         error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
1289
1290         EFX_ERR(efx, "SYSTEM ERROR " EFX_OWORD_FMT " status "
1291                 EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
1292                 EFX_OWORD_VAL(fatal_intr),
1293                 error ? "disabling bus mastering" : "no recognised error");
1294         if (error == 0)
1295                 goto out;
1296
1297         /* If this is a memory parity error dump which blocks are offending */
1298         mem_perr = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER);
1299         if (mem_perr) {
1300                 efx_oword_t reg;
1301                 efx_reado(efx, &reg, FR_AZ_MEM_STAT);
1302                 EFX_ERR(efx, "SYSTEM ERROR: memory parity error "
1303                         EFX_OWORD_FMT "\n", EFX_OWORD_VAL(reg));
1304         }
1305
1306         /* Disable both devices */
1307         pci_clear_master(efx->pci_dev);
1308         if (efx_nic_is_dual_func(efx))
1309                 pci_clear_master(nic_data->pci_dev2);
1310         efx_nic_disable_interrupts(efx);
1311
1312         /* Count errors and reset or disable the NIC accordingly */
1313         if (efx->int_error_count == 0 ||
1314             time_after(jiffies, efx->int_error_expire)) {
1315                 efx->int_error_count = 0;
1316                 efx->int_error_expire =
1317                         jiffies + EFX_INT_ERROR_EXPIRE * HZ;
1318         }
1319         if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
1320                 EFX_ERR(efx, "SYSTEM ERROR - reset scheduled\n");
1321                 efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
1322         } else {
1323                 EFX_ERR(efx, "SYSTEM ERROR - max number of errors seen."
1324                         "NIC will be disabled\n");
1325                 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1326         }
1327 out:
1328         return IRQ_HANDLED;
1329 }
1330
1331 /* Handle a legacy interrupt
1332  * Acknowledges the interrupt and schedule event queue processing.
1333  */
1334 static irqreturn_t efx_legacy_interrupt(int irq, void *dev_id)
1335 {
1336         struct efx_nic *efx = dev_id;
1337         efx_oword_t *int_ker = efx->irq_status.addr;
1338         irqreturn_t result = IRQ_NONE;
1339         struct efx_channel *channel;
1340         efx_dword_t reg;
1341         u32 queues;
1342         int syserr;
1343
1344         /* Read the ISR which also ACKs the interrupts */
1345         efx_readd(efx, &reg, FR_BZ_INT_ISR0);
1346         queues = EFX_EXTRACT_DWORD(reg, 0, 31);
1347
1348         /* Check to see if we have a serious error condition */
1349         syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1350         if (unlikely(syserr))
1351                 return efx_nic_fatal_interrupt(efx);
1352
1353         if (queues != 0) {
1354                 if (EFX_WORKAROUND_15783(efx))
1355                         efx->irq_zero_count = 0;
1356
1357                 /* Schedule processing of any interrupting queues */
1358                 efx_for_each_channel(channel, efx) {
1359                         if (queues & 1)
1360                                 efx_schedule_channel(channel);
1361                         queues >>= 1;
1362                 }
1363                 result = IRQ_HANDLED;
1364
1365         } else if (EFX_WORKAROUND_15783(efx) &&
1366                    efx->irq_zero_count++ == 0) {
1367                 efx_qword_t *event;
1368
1369                 /* Ensure we rearm all event queues */
1370                 efx_for_each_channel(channel, efx) {
1371                         event = efx_event(channel, channel->eventq_read_ptr);
1372                         if (efx_event_present(event))
1373                                 efx_schedule_channel(channel);
1374                 }
1375
1376                 result = IRQ_HANDLED;
1377         }
1378
1379         if (result == IRQ_HANDLED) {
1380                 efx->last_irq_cpu = raw_smp_processor_id();
1381                 EFX_TRACE(efx, "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
1382                           irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
1383         } else if (EFX_WORKAROUND_15783(efx)) {
1384                 /* We can't return IRQ_HANDLED more than once on seeing ISR0=0
1385                  * because this might be a shared interrupt, but we do need to
1386                  * check the channel every time and preemptively rearm it if
1387                  * it's idle. */
1388                 efx_for_each_channel(channel, efx) {
1389                         if (!channel->work_pending)
1390                                 efx_nic_eventq_read_ack(channel);
1391                 }
1392         }
1393
1394         return result;
1395 }
1396
1397 /* Handle an MSI interrupt
1398  *
1399  * Handle an MSI hardware interrupt.  This routine schedules event
1400  * queue processing.  No interrupt acknowledgement cycle is necessary.
1401  * Also, we never need to check that the interrupt is for us, since
1402  * MSI interrupts cannot be shared.
1403  */
1404 static irqreturn_t efx_msi_interrupt(int irq, void *dev_id)
1405 {
1406         struct efx_channel *channel = dev_id;
1407         struct efx_nic *efx = channel->efx;
1408         efx_oword_t *int_ker = efx->irq_status.addr;
1409         int syserr;
1410
1411         efx->last_irq_cpu = raw_smp_processor_id();
1412         EFX_TRACE(efx, "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
1413                   irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
1414
1415         /* Check to see if we have a serious error condition */
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         /* Schedule processing of the channel */
1421         efx_schedule_channel(channel);
1422
1423         return IRQ_HANDLED;
1424 }
1425
1426
1427 /* Setup RSS indirection table.
1428  * This maps from the hash value of the packet to RXQ
1429  */
1430 static void efx_setup_rss_indir_table(struct efx_nic *efx)
1431 {
1432         int i = 0;
1433         unsigned long offset;
1434         efx_dword_t dword;
1435
1436         if (efx_nic_rev(efx) < EFX_REV_FALCON_B0)
1437                 return;
1438
1439         for (offset = FR_BZ_RX_INDIRECTION_TBL;
1440              offset < FR_BZ_RX_INDIRECTION_TBL + 0x800;
1441              offset += 0x10) {
1442                 EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
1443                                      i % efx->n_rx_queues);
1444                 efx_writed(efx, &dword, offset);
1445                 i++;
1446         }
1447 }
1448
1449 /* Hook interrupt handler(s)
1450  * Try MSI and then legacy interrupts.
1451  */
1452 int efx_nic_init_interrupt(struct efx_nic *efx)
1453 {
1454         struct efx_channel *channel;
1455         int rc;
1456
1457         if (!EFX_INT_MODE_USE_MSI(efx)) {
1458                 irq_handler_t handler;
1459                 if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1460                         handler = efx_legacy_interrupt;
1461                 else
1462                         handler = falcon_legacy_interrupt_a1;
1463
1464                 rc = request_irq(efx->legacy_irq, handler, IRQF_SHARED,
1465                                  efx->name, efx);
1466                 if (rc) {
1467                         EFX_ERR(efx, "failed to hook legacy IRQ %d\n",
1468                                 efx->pci_dev->irq);
1469                         goto fail1;
1470                 }
1471                 return 0;
1472         }
1473
1474         /* Hook MSI or MSI-X interrupt */
1475         efx_for_each_channel(channel, efx) {
1476                 rc = request_irq(channel->irq, efx_msi_interrupt,
1477                                  IRQF_PROBE_SHARED, /* Not shared */
1478                                  channel->name, channel);
1479                 if (rc) {
1480                         EFX_ERR(efx, "failed to hook IRQ %d\n", channel->irq);
1481                         goto fail2;
1482                 }
1483         }
1484
1485         return 0;
1486
1487  fail2:
1488         efx_for_each_channel(channel, efx)
1489                 free_irq(channel->irq, channel);
1490  fail1:
1491         return rc;
1492 }
1493
1494 void efx_nic_fini_interrupt(struct efx_nic *efx)
1495 {
1496         struct efx_channel *channel;
1497         efx_oword_t reg;
1498
1499         /* Disable MSI/MSI-X interrupts */
1500         efx_for_each_channel(channel, efx) {
1501                 if (channel->irq)
1502                         free_irq(channel->irq, channel);
1503         }
1504
1505         /* ACK legacy interrupt */
1506         if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1507                 efx_reado(efx, &reg, FR_BZ_INT_ISR0);
1508         else
1509                 falcon_irq_ack_a1(efx);
1510
1511         /* Disable legacy interrupt */
1512         if (efx->legacy_irq)
1513                 free_irq(efx->legacy_irq, efx);
1514 }
1515
1516 u32 efx_nic_fpga_ver(struct efx_nic *efx)
1517 {
1518         efx_oword_t altera_build;
1519         efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
1520         return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
1521 }
1522
1523 void efx_nic_init_common(struct efx_nic *efx)
1524 {
1525         efx_oword_t temp;
1526
1527         /* Set positions of descriptor caches in SRAM. */
1528         EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR,
1529                              efx->type->tx_dc_base / 8);
1530         efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
1531         EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR,
1532                              efx->type->rx_dc_base / 8);
1533         efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
1534
1535         /* Set TX descriptor cache size. */
1536         BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
1537         EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
1538         efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
1539
1540         /* Set RX descriptor cache size.  Set low watermark to size-8, as
1541          * this allows most efficient prefetching.
1542          */
1543         BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
1544         EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
1545         efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
1546         EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
1547         efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
1548
1549         /* Program INT_KER address */
1550         EFX_POPULATE_OWORD_2(temp,
1551                              FRF_AZ_NORM_INT_VEC_DIS_KER,
1552                              EFX_INT_MODE_USE_MSI(efx),
1553                              FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
1554         efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
1555
1556         /* Enable all the genuinely fatal interrupts.  (They are still
1557          * masked by the overall interrupt mask, controlled by
1558          * falcon_interrupts()).
1559          *
1560          * Note: All other fatal interrupts are enabled
1561          */
1562         EFX_POPULATE_OWORD_3(temp,
1563                              FRF_AZ_ILL_ADR_INT_KER_EN, 1,
1564                              FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
1565                              FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
1566         EFX_INVERT_OWORD(temp);
1567         efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
1568
1569         efx_setup_rss_indir_table(efx);
1570
1571         /* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
1572          * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
1573          */
1574         efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
1575         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
1576         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
1577         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
1578         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 0);
1579         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
1580         /* Enable SW_EV to inherit in char driver - assume harmless here */
1581         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
1582         /* Prefetch threshold 2 => fetch when descriptor cache half empty */
1583         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
1584         /* Disable hardware watchdog which can misfire */
1585         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
1586         /* Squash TX of packets of 16 bytes or less */
1587         if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1588                 EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
1589         efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
1590 }