sfc: Survive ISR0=0 bug in the shared IRQ case
[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  * Note that EVQ_RPTR_REG contains the index of the "last read" event,
628  * whereas channel->eventq_read_ptr contains the index of the "next to
629  * read" event.
630  */
631 void efx_nic_eventq_read_ack(struct efx_channel *channel)
632 {
633         efx_dword_t reg;
634         struct efx_nic *efx = channel->efx;
635
636         EFX_POPULATE_DWORD_1(reg, FRF_AZ_EVQ_RPTR, channel->eventq_read_ptr);
637         efx_writed_table(efx, &reg, efx->type->evq_rptr_tbl_base,
638                          channel->channel);
639 }
640
641 /* Use HW to insert a SW defined event */
642 void efx_generate_event(struct efx_channel *channel, efx_qword_t *event)
643 {
644         efx_oword_t drv_ev_reg;
645
646         BUILD_BUG_ON(FRF_AZ_DRV_EV_DATA_LBN != 0 ||
647                      FRF_AZ_DRV_EV_DATA_WIDTH != 64);
648         drv_ev_reg.u32[0] = event->u32[0];
649         drv_ev_reg.u32[1] = event->u32[1];
650         drv_ev_reg.u32[2] = 0;
651         drv_ev_reg.u32[3] = 0;
652         EFX_SET_OWORD_FIELD(drv_ev_reg, FRF_AZ_DRV_EV_QID, channel->channel);
653         efx_writeo(channel->efx, &drv_ev_reg, FR_AZ_DRV_EV);
654 }
655
656 /* Handle a transmit completion event
657  *
658  * The NIC batches TX completion events; the message we receive is of
659  * the form "complete all TX events up to this index".
660  */
661 static void
662 efx_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
663 {
664         unsigned int tx_ev_desc_ptr;
665         unsigned int tx_ev_q_label;
666         struct efx_tx_queue *tx_queue;
667         struct efx_nic *efx = channel->efx;
668
669         if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
670                 /* Transmit completion */
671                 tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
672                 tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
673                 tx_queue = &efx->tx_queue[tx_ev_q_label];
674                 channel->irq_mod_score +=
675                         (tx_ev_desc_ptr - tx_queue->read_count) &
676                         EFX_TXQ_MASK;
677                 efx_xmit_done(tx_queue, tx_ev_desc_ptr);
678         } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
679                 /* Rewrite the FIFO write pointer */
680                 tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
681                 tx_queue = &efx->tx_queue[tx_ev_q_label];
682
683                 if (efx_dev_registered(efx))
684                         netif_tx_lock(efx->net_dev);
685                 efx_notify_tx_desc(tx_queue);
686                 if (efx_dev_registered(efx))
687                         netif_tx_unlock(efx->net_dev);
688         } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR) &&
689                    EFX_WORKAROUND_10727(efx)) {
690                 efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
691         } else {
692                 EFX_ERR(efx, "channel %d unexpected TX event "
693                         EFX_QWORD_FMT"\n", channel->channel,
694                         EFX_QWORD_VAL(*event));
695         }
696 }
697
698 /* Detect errors included in the rx_evt_pkt_ok bit. */
699 static void efx_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
700                                  const efx_qword_t *event,
701                                  bool *rx_ev_pkt_ok,
702                                  bool *discard)
703 {
704         struct efx_nic *efx = rx_queue->efx;
705         bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
706         bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
707         bool rx_ev_frm_trunc, rx_ev_drib_nib, rx_ev_tobe_disc;
708         bool rx_ev_other_err, rx_ev_pause_frm;
709         bool rx_ev_hdr_type, rx_ev_mcast_pkt;
710         unsigned rx_ev_pkt_type;
711
712         rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
713         rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
714         rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
715         rx_ev_pkt_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_TYPE);
716         rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
717                                                  FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
718         rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
719                                                   FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
720         rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
721                                                    FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
722         rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
723         rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
724         rx_ev_drib_nib = ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) ?
725                           0 : EFX_QWORD_FIELD(*event, FSF_AA_RX_EV_DRIB_NIB));
726         rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
727
728         /* Every error apart from tobe_disc and pause_frm */
729         rx_ev_other_err = (rx_ev_drib_nib | rx_ev_tcp_udp_chksum_err |
730                            rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
731                            rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
732
733         /* Count errors that are not in MAC stats.  Ignore expected
734          * checksum errors during self-test. */
735         if (rx_ev_frm_trunc)
736                 ++rx_queue->channel->n_rx_frm_trunc;
737         else if (rx_ev_tobe_disc)
738                 ++rx_queue->channel->n_rx_tobe_disc;
739         else if (!efx->loopback_selftest) {
740                 if (rx_ev_ip_hdr_chksum_err)
741                         ++rx_queue->channel->n_rx_ip_hdr_chksum_err;
742                 else if (rx_ev_tcp_udp_chksum_err)
743                         ++rx_queue->channel->n_rx_tcp_udp_chksum_err;
744         }
745
746         /* The frame must be discarded if any of these are true. */
747         *discard = (rx_ev_eth_crc_err | rx_ev_frm_trunc | rx_ev_drib_nib |
748                     rx_ev_tobe_disc | rx_ev_pause_frm);
749
750         /* TOBE_DISC is expected on unicast mismatches; don't print out an
751          * error message.  FRM_TRUNC indicates RXDP dropped the packet due
752          * to a FIFO overflow.
753          */
754 #ifdef EFX_ENABLE_DEBUG
755         if (rx_ev_other_err) {
756                 EFX_INFO_RL(efx, " RX queue %d unexpected RX event "
757                             EFX_QWORD_FMT "%s%s%s%s%s%s%s%s\n",
758                             rx_queue->queue, EFX_QWORD_VAL(*event),
759                             rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
760                             rx_ev_ip_hdr_chksum_err ?
761                             " [IP_HDR_CHKSUM_ERR]" : "",
762                             rx_ev_tcp_udp_chksum_err ?
763                             " [TCP_UDP_CHKSUM_ERR]" : "",
764                             rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
765                             rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
766                             rx_ev_drib_nib ? " [DRIB_NIB]" : "",
767                             rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
768                             rx_ev_pause_frm ? " [PAUSE]" : "");
769         }
770 #endif
771 }
772
773 /* Handle receive events that are not in-order. */
774 static void
775 efx_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
776 {
777         struct efx_nic *efx = rx_queue->efx;
778         unsigned expected, dropped;
779
780         expected = rx_queue->removed_count & EFX_RXQ_MASK;
781         dropped = (index - expected) & EFX_RXQ_MASK;
782         EFX_INFO(efx, "dropped %d events (index=%d expected=%d)\n",
783                 dropped, index, expected);
784
785         efx_schedule_reset(efx, EFX_WORKAROUND_5676(efx) ?
786                            RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
787 }
788
789 /* Handle a packet received event
790  *
791  * The NIC gives a "discard" flag if it's a unicast packet with the
792  * wrong destination address
793  * Also "is multicast" and "matches multicast filter" flags can be used to
794  * discard non-matching multicast packets.
795  */
796 static void
797 efx_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
798 {
799         unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
800         unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
801         unsigned expected_ptr;
802         bool rx_ev_pkt_ok, discard = false, checksummed;
803         struct efx_rx_queue *rx_queue;
804         struct efx_nic *efx = channel->efx;
805
806         /* Basic packet information */
807         rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
808         rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
809         rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
810         WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT));
811         WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP) != 1);
812         WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
813                 channel->channel);
814
815         rx_queue = &efx->rx_queue[channel->channel];
816
817         rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
818         expected_ptr = rx_queue->removed_count & EFX_RXQ_MASK;
819         if (unlikely(rx_ev_desc_ptr != expected_ptr))
820                 efx_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr);
821
822         if (likely(rx_ev_pkt_ok)) {
823                 /* If packet is marked as OK and packet type is TCP/IP or
824                  * UDP/IP, then we can rely on the hardware checksum.
825                  */
826                 checksummed =
827                         likely(efx->rx_checksum_enabled) &&
828                         (rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP ||
829                          rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP);
830         } else {
831                 efx_handle_rx_not_ok(rx_queue, event, &rx_ev_pkt_ok, &discard);
832                 checksummed = false;
833         }
834
835         /* Detect multicast packets that didn't match the filter */
836         rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
837         if (rx_ev_mcast_pkt) {
838                 unsigned int rx_ev_mcast_hash_match =
839                         EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
840
841                 if (unlikely(!rx_ev_mcast_hash_match)) {
842                         ++channel->n_rx_mcast_mismatch;
843                         discard = true;
844                 }
845         }
846
847         channel->irq_mod_score += 2;
848
849         /* Handle received packet */
850         efx_rx_packet(rx_queue, rx_ev_desc_ptr, rx_ev_byte_cnt,
851                       checksummed, discard);
852 }
853
854 /* Global events are basically PHY events */
855 static void
856 efx_handle_global_event(struct efx_channel *channel, efx_qword_t *event)
857 {
858         struct efx_nic *efx = channel->efx;
859         bool handled = false;
860
861         if (EFX_QWORD_FIELD(*event, FSF_AB_GLB_EV_G_PHY0_INTR) ||
862             EFX_QWORD_FIELD(*event, FSF_AB_GLB_EV_XG_PHY0_INTR) ||
863             EFX_QWORD_FIELD(*event, FSF_AB_GLB_EV_XFP_PHY0_INTR)) {
864                 /* Ignored */
865                 handled = true;
866         }
867
868         if ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) &&
869             EFX_QWORD_FIELD(*event, FSF_BB_GLB_EV_XG_MGT_INTR)) {
870                 efx->xmac_poll_required = true;
871                 handled = true;
872         }
873
874         if (efx_nic_rev(efx) <= EFX_REV_FALCON_A1 ?
875             EFX_QWORD_FIELD(*event, FSF_AA_GLB_EV_RX_RECOVERY) :
876             EFX_QWORD_FIELD(*event, FSF_BB_GLB_EV_RX_RECOVERY)) {
877                 EFX_ERR(efx, "channel %d seen global RX_RESET "
878                         "event. Resetting.\n", channel->channel);
879
880                 atomic_inc(&efx->rx_reset);
881                 efx_schedule_reset(efx, EFX_WORKAROUND_6555(efx) ?
882                                    RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
883                 handled = true;
884         }
885
886         if (!handled)
887                 EFX_ERR(efx, "channel %d unknown global event "
888                         EFX_QWORD_FMT "\n", channel->channel,
889                         EFX_QWORD_VAL(*event));
890 }
891
892 static void
893 efx_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
894 {
895         struct efx_nic *efx = channel->efx;
896         unsigned int ev_sub_code;
897         unsigned int ev_sub_data;
898
899         ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
900         ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
901
902         switch (ev_sub_code) {
903         case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
904                 EFX_TRACE(efx, "channel %d TXQ %d flushed\n",
905                           channel->channel, ev_sub_data);
906                 break;
907         case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
908                 EFX_TRACE(efx, "channel %d RXQ %d flushed\n",
909                           channel->channel, ev_sub_data);
910                 break;
911         case FSE_AZ_EVQ_INIT_DONE_EV:
912                 EFX_LOG(efx, "channel %d EVQ %d initialised\n",
913                         channel->channel, ev_sub_data);
914                 break;
915         case FSE_AZ_SRM_UPD_DONE_EV:
916                 EFX_TRACE(efx, "channel %d SRAM update done\n",
917                           channel->channel);
918                 break;
919         case FSE_AZ_WAKE_UP_EV:
920                 EFX_TRACE(efx, "channel %d RXQ %d wakeup event\n",
921                           channel->channel, ev_sub_data);
922                 break;
923         case FSE_AZ_TIMER_EV:
924                 EFX_TRACE(efx, "channel %d RX queue %d timer expired\n",
925                           channel->channel, ev_sub_data);
926                 break;
927         case FSE_AA_RX_RECOVER_EV:
928                 EFX_ERR(efx, "channel %d seen DRIVER RX_RESET event. "
929                         "Resetting.\n", channel->channel);
930                 atomic_inc(&efx->rx_reset);
931                 efx_schedule_reset(efx,
932                                    EFX_WORKAROUND_6555(efx) ?
933                                    RESET_TYPE_RX_RECOVERY :
934                                    RESET_TYPE_DISABLE);
935                 break;
936         case FSE_BZ_RX_DSC_ERROR_EV:
937                 EFX_ERR(efx, "RX DMA Q %d reports descriptor fetch error."
938                         " RX Q %d is disabled.\n", ev_sub_data, ev_sub_data);
939                 efx_schedule_reset(efx, RESET_TYPE_RX_DESC_FETCH);
940                 break;
941         case FSE_BZ_TX_DSC_ERROR_EV:
942                 EFX_ERR(efx, "TX DMA Q %d reports descriptor fetch error."
943                         " TX Q %d is disabled.\n", ev_sub_data, ev_sub_data);
944                 efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
945                 break;
946         default:
947                 EFX_TRACE(efx, "channel %d unknown driver event code %d "
948                           "data %04x\n", channel->channel, ev_sub_code,
949                           ev_sub_data);
950                 break;
951         }
952 }
953
954 int efx_nic_process_eventq(struct efx_channel *channel, int rx_quota)
955 {
956         unsigned int read_ptr;
957         efx_qword_t event, *p_event;
958         int ev_code;
959         int rx_packets = 0;
960
961         read_ptr = channel->eventq_read_ptr;
962
963         do {
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                 ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
978
979                 switch (ev_code) {
980                 case FSE_AZ_EV_CODE_RX_EV:
981                         efx_handle_rx_event(channel, &event);
982                         ++rx_packets;
983                         break;
984                 case FSE_AZ_EV_CODE_TX_EV:
985                         efx_handle_tx_event(channel, &event);
986                         break;
987                 case FSE_AZ_EV_CODE_DRV_GEN_EV:
988                         channel->eventq_magic = EFX_QWORD_FIELD(
989                                 event, FSF_AZ_DRV_GEN_EV_MAGIC);
990                         EFX_LOG(channel->efx, "channel %d received generated "
991                                 "event "EFX_QWORD_FMT"\n", channel->channel,
992                                 EFX_QWORD_VAL(event));
993                         break;
994                 case FSE_AZ_EV_CODE_GLOBAL_EV:
995                         efx_handle_global_event(channel, &event);
996                         break;
997                 case FSE_AZ_EV_CODE_DRIVER_EV:
998                         efx_handle_driver_event(channel, &event);
999                         break;
1000                 case FSE_CZ_EV_CODE_MCDI_EV:
1001                         efx_mcdi_process_event(channel, &event);
1002                         break;
1003                 default:
1004                         EFX_ERR(channel->efx, "channel %d unknown event type %d"
1005                                 " (data " EFX_QWORD_FMT ")\n", channel->channel,
1006                                 ev_code, EFX_QWORD_VAL(event));
1007                 }
1008
1009                 /* Increment read pointer */
1010                 read_ptr = (read_ptr + 1) & EFX_EVQ_MASK;
1011
1012         } while (rx_packets < rx_quota);
1013
1014         channel->eventq_read_ptr = read_ptr;
1015         return rx_packets;
1016 }
1017
1018
1019 /* Allocate buffer table entries for event queue */
1020 int efx_nic_probe_eventq(struct efx_channel *channel)
1021 {
1022         struct efx_nic *efx = channel->efx;
1023         BUILD_BUG_ON(EFX_EVQ_SIZE < 512 || EFX_EVQ_SIZE > 32768 ||
1024                      EFX_EVQ_SIZE & EFX_EVQ_MASK);
1025         return efx_alloc_special_buffer(efx, &channel->eventq,
1026                                         EFX_EVQ_SIZE * sizeof(efx_qword_t));
1027 }
1028
1029 void efx_nic_init_eventq(struct efx_channel *channel)
1030 {
1031         efx_oword_t reg;
1032         struct efx_nic *efx = channel->efx;
1033
1034         EFX_LOG(efx, "channel %d event queue in special buffers %d-%d\n",
1035                 channel->channel, channel->eventq.index,
1036                 channel->eventq.index + channel->eventq.entries - 1);
1037
1038         if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1039                 EFX_POPULATE_OWORD_3(reg,
1040                                      FRF_CZ_TIMER_Q_EN, 1,
1041                                      FRF_CZ_HOST_NOTIFY_MODE, 0,
1042                                      FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
1043                 efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1044         }
1045
1046         /* Pin event queue buffer */
1047         efx_init_special_buffer(efx, &channel->eventq);
1048
1049         /* Fill event queue with all ones (i.e. empty events) */
1050         memset(channel->eventq.addr, 0xff, channel->eventq.len);
1051
1052         /* Push event queue to card */
1053         EFX_POPULATE_OWORD_3(reg,
1054                              FRF_AZ_EVQ_EN, 1,
1055                              FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
1056                              FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
1057         efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1058                          channel->channel);
1059
1060         efx->type->push_irq_moderation(channel);
1061 }
1062
1063 void efx_nic_fini_eventq(struct efx_channel *channel)
1064 {
1065         efx_oword_t reg;
1066         struct efx_nic *efx = channel->efx;
1067
1068         /* Remove event queue from card */
1069         EFX_ZERO_OWORD(reg);
1070         efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1071                          channel->channel);
1072         if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1073                 efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1074
1075         /* Unpin event queue */
1076         efx_fini_special_buffer(efx, &channel->eventq);
1077 }
1078
1079 /* Free buffers backing event queue */
1080 void efx_nic_remove_eventq(struct efx_channel *channel)
1081 {
1082         efx_free_special_buffer(channel->efx, &channel->eventq);
1083 }
1084
1085
1086 /* Generates a test event on the event queue.  A subsequent call to
1087  * process_eventq() should pick up the event and place the value of
1088  * "magic" into channel->eventq_magic;
1089  */
1090 void efx_nic_generate_test_event(struct efx_channel *channel, unsigned int magic)
1091 {
1092         efx_qword_t test_event;
1093
1094         EFX_POPULATE_QWORD_2(test_event, FSF_AZ_EV_CODE,
1095                              FSE_AZ_EV_CODE_DRV_GEN_EV,
1096                              FSF_AZ_DRV_GEN_EV_MAGIC, magic);
1097         efx_generate_event(channel, &test_event);
1098 }
1099
1100 /**************************************************************************
1101  *
1102  * Flush handling
1103  *
1104  **************************************************************************/
1105
1106
1107 static void efx_poll_flush_events(struct efx_nic *efx)
1108 {
1109         struct efx_channel *channel = &efx->channel[0];
1110         struct efx_tx_queue *tx_queue;
1111         struct efx_rx_queue *rx_queue;
1112         unsigned int read_ptr = channel->eventq_read_ptr;
1113         unsigned int end_ptr = (read_ptr - 1) & EFX_EVQ_MASK;
1114
1115         do {
1116                 efx_qword_t *event = efx_event(channel, read_ptr);
1117                 int ev_code, ev_sub_code, ev_queue;
1118                 bool ev_failed;
1119
1120                 if (!efx_event_present(event))
1121                         break;
1122
1123                 ev_code = EFX_QWORD_FIELD(*event, FSF_AZ_EV_CODE);
1124                 ev_sub_code = EFX_QWORD_FIELD(*event,
1125                                               FSF_AZ_DRIVER_EV_SUBCODE);
1126                 if (ev_code == FSE_AZ_EV_CODE_DRIVER_EV &&
1127                     ev_sub_code == FSE_AZ_TX_DESCQ_FLS_DONE_EV) {
1128                         ev_queue = EFX_QWORD_FIELD(*event,
1129                                                    FSF_AZ_DRIVER_EV_SUBDATA);
1130                         if (ev_queue < EFX_TX_QUEUE_COUNT) {
1131                                 tx_queue = efx->tx_queue + ev_queue;
1132                                 tx_queue->flushed = FLUSH_DONE;
1133                         }
1134                 } else if (ev_code == FSE_AZ_EV_CODE_DRIVER_EV &&
1135                            ev_sub_code == FSE_AZ_RX_DESCQ_FLS_DONE_EV) {
1136                         ev_queue = EFX_QWORD_FIELD(
1137                                 *event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
1138                         ev_failed = EFX_QWORD_FIELD(
1139                                 *event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
1140                         if (ev_queue < efx->n_rx_queues) {
1141                                 rx_queue = efx->rx_queue + ev_queue;
1142                                 rx_queue->flushed =
1143                                         ev_failed ? FLUSH_FAILED : FLUSH_DONE;
1144                         }
1145                 }
1146
1147                 /* We're about to destroy the queue anyway, so
1148                  * it's ok to throw away every non-flush event */
1149                 EFX_SET_QWORD(*event);
1150
1151                 read_ptr = (read_ptr + 1) & EFX_EVQ_MASK;
1152         } while (read_ptr != end_ptr);
1153
1154         channel->eventq_read_ptr = read_ptr;
1155 }
1156
1157 /* Handle tx and rx flushes at the same time, since they run in
1158  * parallel in the hardware and there's no reason for us to
1159  * serialise them */
1160 int efx_nic_flush_queues(struct efx_nic *efx)
1161 {
1162         struct efx_rx_queue *rx_queue;
1163         struct efx_tx_queue *tx_queue;
1164         int i, tx_pending, rx_pending;
1165
1166         /* If necessary prepare the hardware for flushing */
1167         efx->type->prepare_flush(efx);
1168
1169         /* Flush all tx queues in parallel */
1170         efx_for_each_tx_queue(tx_queue, efx)
1171                 efx_flush_tx_queue(tx_queue);
1172
1173         /* The hardware supports four concurrent rx flushes, each of which may
1174          * need to be retried if there is an outstanding descriptor fetch */
1175         for (i = 0; i < EFX_FLUSH_POLL_COUNT; ++i) {
1176                 rx_pending = tx_pending = 0;
1177                 efx_for_each_rx_queue(rx_queue, efx) {
1178                         if (rx_queue->flushed == FLUSH_PENDING)
1179                                 ++rx_pending;
1180                 }
1181                 efx_for_each_rx_queue(rx_queue, efx) {
1182                         if (rx_pending == EFX_RX_FLUSH_COUNT)
1183                                 break;
1184                         if (rx_queue->flushed == FLUSH_FAILED ||
1185                             rx_queue->flushed == FLUSH_NONE) {
1186                                 efx_flush_rx_queue(rx_queue);
1187                                 ++rx_pending;
1188                         }
1189                 }
1190                 efx_for_each_tx_queue(tx_queue, efx) {
1191                         if (tx_queue->flushed != FLUSH_DONE)
1192                                 ++tx_pending;
1193                 }
1194
1195                 if (rx_pending == 0 && tx_pending == 0)
1196                         return 0;
1197
1198                 msleep(EFX_FLUSH_INTERVAL);
1199                 efx_poll_flush_events(efx);
1200         }
1201
1202         /* Mark the queues as all flushed. We're going to return failure
1203          * leading to a reset, or fake up success anyway */
1204         efx_for_each_tx_queue(tx_queue, efx) {
1205                 if (tx_queue->flushed != FLUSH_DONE)
1206                         EFX_ERR(efx, "tx queue %d flush command timed out\n",
1207                                 tx_queue->queue);
1208                 tx_queue->flushed = FLUSH_DONE;
1209         }
1210         efx_for_each_rx_queue(rx_queue, efx) {
1211                 if (rx_queue->flushed != FLUSH_DONE)
1212                         EFX_ERR(efx, "rx queue %d flush command timed out\n",
1213                                 rx_queue->queue);
1214                 rx_queue->flushed = FLUSH_DONE;
1215         }
1216
1217         if (EFX_WORKAROUND_7803(efx))
1218                 return 0;
1219
1220         return -ETIMEDOUT;
1221 }
1222
1223 /**************************************************************************
1224  *
1225  * Hardware interrupts
1226  * The hardware interrupt handler does very little work; all the event
1227  * queue processing is carried out by per-channel tasklets.
1228  *
1229  **************************************************************************/
1230
1231 /* Enable/disable/generate interrupts */
1232 static inline void efx_nic_interrupts(struct efx_nic *efx,
1233                                       bool enabled, bool force)
1234 {
1235         efx_oword_t int_en_reg_ker;
1236         unsigned int level = 0;
1237
1238         if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
1239                 /* Set the level always even if we're generating a test
1240                  * interrupt, because our legacy interrupt handler is safe */
1241                 level = 0x1f;
1242
1243         EFX_POPULATE_OWORD_3(int_en_reg_ker,
1244                              FRF_AZ_KER_INT_LEVE_SEL, level,
1245                              FRF_AZ_KER_INT_KER, force,
1246                              FRF_AZ_DRV_INT_EN_KER, enabled);
1247         efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
1248 }
1249
1250 void efx_nic_enable_interrupts(struct efx_nic *efx)
1251 {
1252         struct efx_channel *channel;
1253
1254         EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
1255         wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
1256
1257         /* Enable interrupts */
1258         efx_nic_interrupts(efx, true, false);
1259
1260         /* Force processing of all the channels to get the EVQ RPTRs up to
1261            date */
1262         efx_for_each_channel(channel, efx)
1263                 efx_schedule_channel(channel);
1264 }
1265
1266 void efx_nic_disable_interrupts(struct efx_nic *efx)
1267 {
1268         /* Disable interrupts */
1269         efx_nic_interrupts(efx, false, false);
1270 }
1271
1272 /* Generate a test interrupt
1273  * Interrupt must already have been enabled, otherwise nasty things
1274  * may happen.
1275  */
1276 void efx_nic_generate_interrupt(struct efx_nic *efx)
1277 {
1278         efx_nic_interrupts(efx, true, true);
1279 }
1280
1281 /* Process a fatal interrupt
1282  * Disable bus mastering ASAP and schedule a reset
1283  */
1284 irqreturn_t efx_nic_fatal_interrupt(struct efx_nic *efx)
1285 {
1286         struct falcon_nic_data *nic_data = efx->nic_data;
1287         efx_oword_t *int_ker = efx->irq_status.addr;
1288         efx_oword_t fatal_intr;
1289         int error, mem_perr;
1290
1291         efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
1292         error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
1293
1294         EFX_ERR(efx, "SYSTEM ERROR " EFX_OWORD_FMT " status "
1295                 EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
1296                 EFX_OWORD_VAL(fatal_intr),
1297                 error ? "disabling bus mastering" : "no recognised error");
1298         if (error == 0)
1299                 goto out;
1300
1301         /* If this is a memory parity error dump which blocks are offending */
1302         mem_perr = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER);
1303         if (mem_perr) {
1304                 efx_oword_t reg;
1305                 efx_reado(efx, &reg, FR_AZ_MEM_STAT);
1306                 EFX_ERR(efx, "SYSTEM ERROR: memory parity error "
1307                         EFX_OWORD_FMT "\n", EFX_OWORD_VAL(reg));
1308         }
1309
1310         /* Disable both devices */
1311         pci_clear_master(efx->pci_dev);
1312         if (efx_nic_is_dual_func(efx))
1313                 pci_clear_master(nic_data->pci_dev2);
1314         efx_nic_disable_interrupts(efx);
1315
1316         /* Count errors and reset or disable the NIC accordingly */
1317         if (efx->int_error_count == 0 ||
1318             time_after(jiffies, efx->int_error_expire)) {
1319                 efx->int_error_count = 0;
1320                 efx->int_error_expire =
1321                         jiffies + EFX_INT_ERROR_EXPIRE * HZ;
1322         }
1323         if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
1324                 EFX_ERR(efx, "SYSTEM ERROR - reset scheduled\n");
1325                 efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
1326         } else {
1327                 EFX_ERR(efx, "SYSTEM ERROR - max number of errors seen."
1328                         "NIC will be disabled\n");
1329                 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1330         }
1331 out:
1332         return IRQ_HANDLED;
1333 }
1334
1335 /* Handle a legacy interrupt
1336  * Acknowledges the interrupt and schedule event queue processing.
1337  */
1338 static irqreturn_t efx_legacy_interrupt(int irq, void *dev_id)
1339 {
1340         struct efx_nic *efx = dev_id;
1341         efx_oword_t *int_ker = efx->irq_status.addr;
1342         irqreturn_t result = IRQ_NONE;
1343         struct efx_channel *channel;
1344         efx_dword_t reg;
1345         u32 queues;
1346         int syserr;
1347
1348         /* Read the ISR which also ACKs the interrupts */
1349         efx_readd(efx, &reg, FR_BZ_INT_ISR0);
1350         queues = EFX_EXTRACT_DWORD(reg, 0, 31);
1351
1352         /* Check to see if we have a serious error condition */
1353         syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1354         if (unlikely(syserr))
1355                 return efx_nic_fatal_interrupt(efx);
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->irq_zero_count++ == 0) {
1371                 efx_qword_t *event;
1372
1373                 /* Ensure we rearm all event queues */
1374                 efx_for_each_channel(channel, efx) {
1375                         event = efx_event(channel, channel->eventq_read_ptr);
1376                         if (efx_event_present(event))
1377                                 efx_schedule_channel(channel);
1378                 }
1379
1380                 result = IRQ_HANDLED;
1381         }
1382
1383         if (result == IRQ_HANDLED) {
1384                 efx->last_irq_cpu = raw_smp_processor_id();
1385                 EFX_TRACE(efx, "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
1386                           irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
1387         } else if (EFX_WORKAROUND_15783(efx)) {
1388                 /* We can't return IRQ_HANDLED more than once on seeing ISR0=0
1389                  * because this might be a shared interrupt, but we do need to
1390                  * check the channel every time and preemptively rearm it if
1391                  * it's idle. */
1392                 efx_for_each_channel(channel, efx) {
1393                         if (!channel->work_pending)
1394                                 efx_nic_eventq_read_ack(channel);
1395                 }
1396         }
1397
1398         return result;
1399 }
1400
1401 /* Handle an MSI interrupt
1402  *
1403  * Handle an MSI hardware interrupt.  This routine schedules event
1404  * queue processing.  No interrupt acknowledgement cycle is necessary.
1405  * Also, we never need to check that the interrupt is for us, since
1406  * MSI interrupts cannot be shared.
1407  */
1408 static irqreturn_t efx_msi_interrupt(int irq, void *dev_id)
1409 {
1410         struct efx_channel *channel = dev_id;
1411         struct efx_nic *efx = channel->efx;
1412         efx_oword_t *int_ker = efx->irq_status.addr;
1413         int syserr;
1414
1415         efx->last_irq_cpu = raw_smp_processor_id();
1416         EFX_TRACE(efx, "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
1417                   irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
1418
1419         /* Check to see if we have a serious error condition */
1420         syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1421         if (unlikely(syserr))
1422                 return efx_nic_fatal_interrupt(efx);
1423
1424         /* Schedule processing of the channel */
1425         efx_schedule_channel(channel);
1426
1427         return IRQ_HANDLED;
1428 }
1429
1430
1431 /* Setup RSS indirection table.
1432  * This maps from the hash value of the packet to RXQ
1433  */
1434 static void efx_setup_rss_indir_table(struct efx_nic *efx)
1435 {
1436         int i = 0;
1437         unsigned long offset;
1438         efx_dword_t dword;
1439
1440         if (efx_nic_rev(efx) < EFX_REV_FALCON_B0)
1441                 return;
1442
1443         for (offset = FR_BZ_RX_INDIRECTION_TBL;
1444              offset < FR_BZ_RX_INDIRECTION_TBL + 0x800;
1445              offset += 0x10) {
1446                 EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
1447                                      i % efx->n_rx_queues);
1448                 efx_writed(efx, &dword, offset);
1449                 i++;
1450         }
1451 }
1452
1453 /* Hook interrupt handler(s)
1454  * Try MSI and then legacy interrupts.
1455  */
1456 int efx_nic_init_interrupt(struct efx_nic *efx)
1457 {
1458         struct efx_channel *channel;
1459         int rc;
1460
1461         if (!EFX_INT_MODE_USE_MSI(efx)) {
1462                 irq_handler_t handler;
1463                 if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1464                         handler = efx_legacy_interrupt;
1465                 else
1466                         handler = falcon_legacy_interrupt_a1;
1467
1468                 rc = request_irq(efx->legacy_irq, handler, IRQF_SHARED,
1469                                  efx->name, efx);
1470                 if (rc) {
1471                         EFX_ERR(efx, "failed to hook legacy IRQ %d\n",
1472                                 efx->pci_dev->irq);
1473                         goto fail1;
1474                 }
1475                 return 0;
1476         }
1477
1478         /* Hook MSI or MSI-X interrupt */
1479         efx_for_each_channel(channel, efx) {
1480                 rc = request_irq(channel->irq, efx_msi_interrupt,
1481                                  IRQF_PROBE_SHARED, /* Not shared */
1482                                  channel->name, channel);
1483                 if (rc) {
1484                         EFX_ERR(efx, "failed to hook IRQ %d\n", channel->irq);
1485                         goto fail2;
1486                 }
1487         }
1488
1489         return 0;
1490
1491  fail2:
1492         efx_for_each_channel(channel, efx)
1493                 free_irq(channel->irq, channel);
1494  fail1:
1495         return rc;
1496 }
1497
1498 void efx_nic_fini_interrupt(struct efx_nic *efx)
1499 {
1500         struct efx_channel *channel;
1501         efx_oword_t reg;
1502
1503         /* Disable MSI/MSI-X interrupts */
1504         efx_for_each_channel(channel, efx) {
1505                 if (channel->irq)
1506                         free_irq(channel->irq, channel);
1507         }
1508
1509         /* ACK legacy interrupt */
1510         if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1511                 efx_reado(efx, &reg, FR_BZ_INT_ISR0);
1512         else
1513                 falcon_irq_ack_a1(efx);
1514
1515         /* Disable legacy interrupt */
1516         if (efx->legacy_irq)
1517                 free_irq(efx->legacy_irq, efx);
1518 }
1519
1520 u32 efx_nic_fpga_ver(struct efx_nic *efx)
1521 {
1522         efx_oword_t altera_build;
1523         efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
1524         return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
1525 }
1526
1527 void efx_nic_init_common(struct efx_nic *efx)
1528 {
1529         efx_oword_t temp;
1530
1531         /* Set positions of descriptor caches in SRAM. */
1532         EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR,
1533                              efx->type->tx_dc_base / 8);
1534         efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
1535         EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR,
1536                              efx->type->rx_dc_base / 8);
1537         efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
1538
1539         /* Set TX descriptor cache size. */
1540         BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
1541         EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
1542         efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
1543
1544         /* Set RX descriptor cache size.  Set low watermark to size-8, as
1545          * this allows most efficient prefetching.
1546          */
1547         BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
1548         EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
1549         efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
1550         EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
1551         efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
1552
1553         /* Program INT_KER address */
1554         EFX_POPULATE_OWORD_2(temp,
1555                              FRF_AZ_NORM_INT_VEC_DIS_KER,
1556                              EFX_INT_MODE_USE_MSI(efx),
1557                              FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
1558         efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
1559
1560         /* Enable all the genuinely fatal interrupts.  (They are still
1561          * masked by the overall interrupt mask, controlled by
1562          * falcon_interrupts()).
1563          *
1564          * Note: All other fatal interrupts are enabled
1565          */
1566         EFX_POPULATE_OWORD_3(temp,
1567                              FRF_AZ_ILL_ADR_INT_KER_EN, 1,
1568                              FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
1569                              FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
1570         EFX_INVERT_OWORD(temp);
1571         efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
1572
1573         efx_setup_rss_indir_table(efx);
1574
1575         /* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
1576          * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
1577          */
1578         efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
1579         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
1580         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
1581         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
1582         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 0);
1583         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
1584         /* Enable SW_EV to inherit in char driver - assume harmless here */
1585         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
1586         /* Prefetch threshold 2 => fetch when descriptor cache half empty */
1587         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
1588         /* Disable hardware watchdog which can misfire */
1589         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
1590         /* Squash TX of packets of 16 bytes or less */
1591         if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1592                 EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
1593         efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
1594 }