1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* Copyright(c) 2013 - 2019 Intel Corporation. */ |
3 | |
4 | #include <linux/types.h> |
5 | #include <linux/module.h> |
6 | #include <net/ipv6.h> |
7 | #include <net/ip.h> |
8 | #include <net/tcp.h> |
9 | #include <linux/if_macvlan.h> |
10 | #include <linux/prefetch.h> |
11 | |
12 | #include "fm10k.h" |
13 | |
14 | #define DRV_SUMMARY "Intel(R) Ethernet Switch Host Interface Driver" |
15 | char fm10k_driver_name[] = "fm10k" ; |
16 | static const char fm10k_driver_string[] = DRV_SUMMARY; |
17 | static const char fm10k_copyright[] = |
18 | "Copyright(c) 2013 - 2019 Intel Corporation." ; |
19 | |
20 | MODULE_AUTHOR("Intel Corporation, <linux.nics@intel.com>" ); |
21 | MODULE_DESCRIPTION(DRV_SUMMARY); |
22 | MODULE_LICENSE("GPL v2" ); |
23 | |
24 | /* single workqueue for entire fm10k driver */ |
25 | struct workqueue_struct *fm10k_workqueue; |
26 | |
27 | /** |
28 | * fm10k_init_module - Driver Registration Routine |
29 | * |
30 | * fm10k_init_module is the first routine called when the driver is |
31 | * loaded. All it does is register with the PCI subsystem. |
32 | **/ |
33 | static int __init fm10k_init_module(void) |
34 | { |
35 | int ret; |
36 | |
37 | pr_info("%s\n" , fm10k_driver_string); |
38 | pr_info("%s\n" , fm10k_copyright); |
39 | |
40 | /* create driver workqueue */ |
41 | fm10k_workqueue = alloc_workqueue(fmt: "%s" , flags: WQ_MEM_RECLAIM, max_active: 0, |
42 | fm10k_driver_name); |
43 | if (!fm10k_workqueue) |
44 | return -ENOMEM; |
45 | |
46 | fm10k_dbg_init(); |
47 | |
48 | ret = fm10k_register_pci_driver(); |
49 | if (ret) { |
50 | fm10k_dbg_exit(); |
51 | destroy_workqueue(wq: fm10k_workqueue); |
52 | } |
53 | |
54 | return ret; |
55 | } |
56 | module_init(fm10k_init_module); |
57 | |
58 | /** |
59 | * fm10k_exit_module - Driver Exit Cleanup Routine |
60 | * |
61 | * fm10k_exit_module is called just before the driver is removed |
62 | * from memory. |
63 | **/ |
64 | static void __exit fm10k_exit_module(void) |
65 | { |
66 | fm10k_unregister_pci_driver(); |
67 | |
68 | fm10k_dbg_exit(); |
69 | |
70 | /* destroy driver workqueue */ |
71 | destroy_workqueue(wq: fm10k_workqueue); |
72 | } |
73 | module_exit(fm10k_exit_module); |
74 | |
75 | static bool fm10k_alloc_mapped_page(struct fm10k_ring *rx_ring, |
76 | struct fm10k_rx_buffer *bi) |
77 | { |
78 | struct page *page = bi->page; |
79 | dma_addr_t dma; |
80 | |
81 | /* Only page will be NULL if buffer was consumed */ |
82 | if (likely(page)) |
83 | return true; |
84 | |
85 | /* alloc new page for storage */ |
86 | page = dev_alloc_page(); |
87 | if (unlikely(!page)) { |
88 | rx_ring->rx_stats.alloc_failed++; |
89 | return false; |
90 | } |
91 | |
92 | /* map page for use */ |
93 | dma = dma_map_page(rx_ring->dev, page, 0, PAGE_SIZE, DMA_FROM_DEVICE); |
94 | |
95 | /* if mapping failed free memory back to system since |
96 | * there isn't much point in holding memory we can't use |
97 | */ |
98 | if (dma_mapping_error(dev: rx_ring->dev, dma_addr: dma)) { |
99 | __free_page(page); |
100 | |
101 | rx_ring->rx_stats.alloc_failed++; |
102 | return false; |
103 | } |
104 | |
105 | bi->dma = dma; |
106 | bi->page = page; |
107 | bi->page_offset = 0; |
108 | |
109 | return true; |
110 | } |
111 | |
112 | /** |
113 | * fm10k_alloc_rx_buffers - Replace used receive buffers |
114 | * @rx_ring: ring to place buffers on |
115 | * @cleaned_count: number of buffers to replace |
116 | **/ |
117 | void fm10k_alloc_rx_buffers(struct fm10k_ring *rx_ring, u16 cleaned_count) |
118 | { |
119 | union fm10k_rx_desc *rx_desc; |
120 | struct fm10k_rx_buffer *bi; |
121 | u16 i = rx_ring->next_to_use; |
122 | |
123 | /* nothing to do */ |
124 | if (!cleaned_count) |
125 | return; |
126 | |
127 | rx_desc = FM10K_RX_DESC(rx_ring, i); |
128 | bi = &rx_ring->rx_buffer[i]; |
129 | i -= rx_ring->count; |
130 | |
131 | do { |
132 | if (!fm10k_alloc_mapped_page(rx_ring, bi)) |
133 | break; |
134 | |
135 | /* Refresh the desc even if buffer_addrs didn't change |
136 | * because each write-back erases this info. |
137 | */ |
138 | rx_desc->q.pkt_addr = cpu_to_le64(bi->dma + bi->page_offset); |
139 | |
140 | rx_desc++; |
141 | bi++; |
142 | i++; |
143 | if (unlikely(!i)) { |
144 | rx_desc = FM10K_RX_DESC(rx_ring, 0); |
145 | bi = rx_ring->rx_buffer; |
146 | i -= rx_ring->count; |
147 | } |
148 | |
149 | /* clear the status bits for the next_to_use descriptor */ |
150 | rx_desc->d.staterr = 0; |
151 | |
152 | cleaned_count--; |
153 | } while (cleaned_count); |
154 | |
155 | i += rx_ring->count; |
156 | |
157 | if (rx_ring->next_to_use != i) { |
158 | /* record the next descriptor to use */ |
159 | rx_ring->next_to_use = i; |
160 | |
161 | /* update next to alloc since we have filled the ring */ |
162 | rx_ring->next_to_alloc = i; |
163 | |
164 | /* Force memory writes to complete before letting h/w |
165 | * know there are new descriptors to fetch. (Only |
166 | * applicable for weak-ordered memory model archs, |
167 | * such as IA-64). |
168 | */ |
169 | wmb(); |
170 | |
171 | /* notify hardware of new descriptors */ |
172 | writel(val: i, addr: rx_ring->tail); |
173 | } |
174 | } |
175 | |
176 | /** |
177 | * fm10k_reuse_rx_page - page flip buffer and store it back on the ring |
178 | * @rx_ring: rx descriptor ring to store buffers on |
179 | * @old_buff: donor buffer to have page reused |
180 | * |
181 | * Synchronizes page for reuse by the interface |
182 | **/ |
183 | static void fm10k_reuse_rx_page(struct fm10k_ring *rx_ring, |
184 | struct fm10k_rx_buffer *old_buff) |
185 | { |
186 | struct fm10k_rx_buffer *new_buff; |
187 | u16 nta = rx_ring->next_to_alloc; |
188 | |
189 | new_buff = &rx_ring->rx_buffer[nta]; |
190 | |
191 | /* update, and store next to alloc */ |
192 | nta++; |
193 | rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0; |
194 | |
195 | /* transfer page from old buffer to new buffer */ |
196 | *new_buff = *old_buff; |
197 | |
198 | /* sync the buffer for use by the device */ |
199 | dma_sync_single_range_for_device(dev: rx_ring->dev, addr: old_buff->dma, |
200 | offset: old_buff->page_offset, |
201 | FM10K_RX_BUFSZ, |
202 | dir: DMA_FROM_DEVICE); |
203 | } |
204 | |
205 | static bool fm10k_can_reuse_rx_page(struct fm10k_rx_buffer *rx_buffer, |
206 | struct page *page, |
207 | unsigned int __maybe_unused truesize) |
208 | { |
209 | /* avoid re-using remote and pfmemalloc pages */ |
210 | if (!dev_page_is_reusable(page)) |
211 | return false; |
212 | |
213 | #if (PAGE_SIZE < 8192) |
214 | /* if we are only owner of page we can reuse it */ |
215 | if (unlikely(page_count(page) != 1)) |
216 | return false; |
217 | |
218 | /* flip page offset to other buffer */ |
219 | rx_buffer->page_offset ^= FM10K_RX_BUFSZ; |
220 | #else |
221 | /* move offset up to the next cache line */ |
222 | rx_buffer->page_offset += truesize; |
223 | |
224 | if (rx_buffer->page_offset > (PAGE_SIZE - FM10K_RX_BUFSZ)) |
225 | return false; |
226 | #endif |
227 | |
228 | /* Even if we own the page, we are not allowed to use atomic_set() |
229 | * This would break get_page_unless_zero() users. |
230 | */ |
231 | page_ref_inc(page); |
232 | |
233 | return true; |
234 | } |
235 | |
236 | /** |
237 | * fm10k_add_rx_frag - Add contents of Rx buffer to sk_buff |
238 | * @rx_buffer: buffer containing page to add |
239 | * @size: packet size from rx_desc |
240 | * @rx_desc: descriptor containing length of buffer written by hardware |
241 | * @skb: sk_buff to place the data into |
242 | * |
243 | * This function will add the data contained in rx_buffer->page to the skb. |
244 | * This is done either through a direct copy if the data in the buffer is |
245 | * less than the skb header size, otherwise it will just attach the page as |
246 | * a frag to the skb. |
247 | * |
248 | * The function will then update the page offset if necessary and return |
249 | * true if the buffer can be reused by the interface. |
250 | **/ |
251 | static bool fm10k_add_rx_frag(struct fm10k_rx_buffer *rx_buffer, |
252 | unsigned int size, |
253 | union fm10k_rx_desc *rx_desc, |
254 | struct sk_buff *skb) |
255 | { |
256 | struct page *page = rx_buffer->page; |
257 | unsigned char *va = page_address(page) + rx_buffer->page_offset; |
258 | #if (PAGE_SIZE < 8192) |
259 | unsigned int truesize = FM10K_RX_BUFSZ; |
260 | #else |
261 | unsigned int truesize = ALIGN(size, 512); |
262 | #endif |
263 | unsigned int pull_len; |
264 | |
265 | if (unlikely(skb_is_nonlinear(skb))) |
266 | goto add_tail_frag; |
267 | |
268 | if (likely(size <= FM10K_RX_HDR_LEN)) { |
269 | memcpy(__skb_put(skb, size), va, ALIGN(size, sizeof(long))); |
270 | |
271 | /* page is reusable, we can reuse buffer as-is */ |
272 | if (dev_page_is_reusable(page)) |
273 | return true; |
274 | |
275 | /* this page cannot be reused so discard it */ |
276 | __free_page(page); |
277 | return false; |
278 | } |
279 | |
280 | /* we need the header to contain the greater of either ETH_HLEN or |
281 | * 60 bytes if the skb->len is less than 60 for skb_pad. |
282 | */ |
283 | pull_len = eth_get_headlen(dev: skb->dev, data: va, FM10K_RX_HDR_LEN); |
284 | |
285 | /* align pull length to size of long to optimize memcpy performance */ |
286 | memcpy(__skb_put(skb, pull_len), va, ALIGN(pull_len, sizeof(long))); |
287 | |
288 | /* update all of the pointers */ |
289 | va += pull_len; |
290 | size -= pull_len; |
291 | |
292 | add_tail_frag: |
293 | skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, |
294 | off: (unsigned long)va & ~PAGE_MASK, size, truesize); |
295 | |
296 | return fm10k_can_reuse_rx_page(rx_buffer, page, truesize); |
297 | } |
298 | |
299 | static struct sk_buff *fm10k_fetch_rx_buffer(struct fm10k_ring *rx_ring, |
300 | union fm10k_rx_desc *rx_desc, |
301 | struct sk_buff *skb) |
302 | { |
303 | unsigned int size = le16_to_cpu(rx_desc->w.length); |
304 | struct fm10k_rx_buffer *rx_buffer; |
305 | struct page *page; |
306 | |
307 | rx_buffer = &rx_ring->rx_buffer[rx_ring->next_to_clean]; |
308 | page = rx_buffer->page; |
309 | prefetchw(x: page); |
310 | |
311 | if (likely(!skb)) { |
312 | void *page_addr = page_address(page) + |
313 | rx_buffer->page_offset; |
314 | |
315 | /* prefetch first cache line of first page */ |
316 | net_prefetch(p: page_addr); |
317 | |
318 | /* allocate a skb to store the frags */ |
319 | skb = napi_alloc_skb(napi: &rx_ring->q_vector->napi, |
320 | FM10K_RX_HDR_LEN); |
321 | if (unlikely(!skb)) { |
322 | rx_ring->rx_stats.alloc_failed++; |
323 | return NULL; |
324 | } |
325 | |
326 | /* we will be copying header into skb->data in |
327 | * pskb_may_pull so it is in our interest to prefetch |
328 | * it now to avoid a possible cache miss |
329 | */ |
330 | prefetchw(x: skb->data); |
331 | } |
332 | |
333 | /* we are reusing so sync this buffer for CPU use */ |
334 | dma_sync_single_range_for_cpu(dev: rx_ring->dev, |
335 | addr: rx_buffer->dma, |
336 | offset: rx_buffer->page_offset, |
337 | size, |
338 | dir: DMA_FROM_DEVICE); |
339 | |
340 | /* pull page into skb */ |
341 | if (fm10k_add_rx_frag(rx_buffer, size, rx_desc, skb)) { |
342 | /* hand second half of page back to the ring */ |
343 | fm10k_reuse_rx_page(rx_ring, old_buff: rx_buffer); |
344 | } else { |
345 | /* we are not reusing the buffer so unmap it */ |
346 | dma_unmap_page(rx_ring->dev, rx_buffer->dma, |
347 | PAGE_SIZE, DMA_FROM_DEVICE); |
348 | } |
349 | |
350 | /* clear contents of rx_buffer */ |
351 | rx_buffer->page = NULL; |
352 | |
353 | return skb; |
354 | } |
355 | |
356 | static inline void fm10k_rx_checksum(struct fm10k_ring *ring, |
357 | union fm10k_rx_desc *rx_desc, |
358 | struct sk_buff *skb) |
359 | { |
360 | skb_checksum_none_assert(skb); |
361 | |
362 | /* Rx checksum disabled via ethtool */ |
363 | if (!(ring->netdev->features & NETIF_F_RXCSUM)) |
364 | return; |
365 | |
366 | /* TCP/UDP checksum error bit is set */ |
367 | if (fm10k_test_staterr(rx_desc, |
368 | FM10K_RXD_STATUS_L4E | |
369 | FM10K_RXD_STATUS_L4E2 | |
370 | FM10K_RXD_STATUS_IPE | |
371 | FM10K_RXD_STATUS_IPE2)) { |
372 | ring->rx_stats.csum_err++; |
373 | return; |
374 | } |
375 | |
376 | /* It must be a TCP or UDP packet with a valid checksum */ |
377 | if (fm10k_test_staterr(rx_desc, FM10K_RXD_STATUS_L4CS2)) |
378 | skb->encapsulation = true; |
379 | else if (!fm10k_test_staterr(rx_desc, FM10K_RXD_STATUS_L4CS)) |
380 | return; |
381 | |
382 | skb->ip_summed = CHECKSUM_UNNECESSARY; |
383 | |
384 | ring->rx_stats.csum_good++; |
385 | } |
386 | |
387 | #define \ |
388 | (BIT(FM10K_RSSTYPE_IPV4_TCP) | \ |
389 | BIT(FM10K_RSSTYPE_IPV4_UDP) | \ |
390 | BIT(FM10K_RSSTYPE_IPV6_TCP) | \ |
391 | BIT(FM10K_RSSTYPE_IPV6_UDP)) |
392 | |
393 | static inline void fm10k_rx_hash(struct fm10k_ring *ring, |
394 | union fm10k_rx_desc *rx_desc, |
395 | struct sk_buff *skb) |
396 | { |
397 | u16 ; |
398 | |
399 | if (!(ring->netdev->features & NETIF_F_RXHASH)) |
400 | return; |
401 | |
402 | rss_type = le16_to_cpu(rx_desc->w.pkt_info) & FM10K_RXD_RSSTYPE_MASK; |
403 | if (!rss_type) |
404 | return; |
405 | |
406 | skb_set_hash(skb, le32_to_cpu(rx_desc->d.rss), |
407 | type: (BIT(rss_type) & FM10K_RSS_L4_TYPES_MASK) ? |
408 | PKT_HASH_TYPE_L4 : PKT_HASH_TYPE_L3); |
409 | } |
410 | |
411 | static void fm10k_type_trans(struct fm10k_ring *rx_ring, |
412 | union fm10k_rx_desc __maybe_unused *rx_desc, |
413 | struct sk_buff *skb) |
414 | { |
415 | struct net_device *dev = rx_ring->netdev; |
416 | struct fm10k_l2_accel *l2_accel = rcu_dereference_bh(rx_ring->l2_accel); |
417 | |
418 | /* check to see if DGLORT belongs to a MACVLAN */ |
419 | if (l2_accel) { |
420 | u16 idx = le16_to_cpu(FM10K_CB(skb)->fi.w.dglort) - 1; |
421 | |
422 | idx -= l2_accel->dglort; |
423 | if (idx < l2_accel->size && l2_accel->macvlan[idx]) |
424 | dev = l2_accel->macvlan[idx]; |
425 | else |
426 | l2_accel = NULL; |
427 | } |
428 | |
429 | /* Record Rx queue, or update macvlan statistics */ |
430 | if (!l2_accel) |
431 | skb_record_rx_queue(skb, rx_queue: rx_ring->queue_index); |
432 | else |
433 | macvlan_count_rx(vlan: netdev_priv(dev), len: skb->len + ETH_HLEN, success: true, |
434 | multicast: false); |
435 | |
436 | skb->protocol = eth_type_trans(skb, dev); |
437 | } |
438 | |
439 | /** |
440 | * fm10k_process_skb_fields - Populate skb header fields from Rx descriptor |
441 | * @rx_ring: rx descriptor ring packet is being transacted on |
442 | * @rx_desc: pointer to the EOP Rx descriptor |
443 | * @skb: pointer to current skb being populated |
444 | * |
445 | * This function checks the ring, descriptor, and packet information in |
446 | * order to populate the hash, checksum, VLAN, timestamp, protocol, and |
447 | * other fields within the skb. |
448 | **/ |
449 | static unsigned int fm10k_process_skb_fields(struct fm10k_ring *rx_ring, |
450 | union fm10k_rx_desc *rx_desc, |
451 | struct sk_buff *skb) |
452 | { |
453 | unsigned int len = skb->len; |
454 | |
455 | fm10k_rx_hash(ring: rx_ring, rx_desc, skb); |
456 | |
457 | fm10k_rx_checksum(ring: rx_ring, rx_desc, skb); |
458 | |
459 | FM10K_CB(skb)->tstamp = rx_desc->q.timestamp; |
460 | |
461 | FM10K_CB(skb)->fi.w.vlan = rx_desc->w.vlan; |
462 | |
463 | FM10K_CB(skb)->fi.d.glort = rx_desc->d.glort; |
464 | |
465 | if (rx_desc->w.vlan) { |
466 | u16 vid = le16_to_cpu(rx_desc->w.vlan); |
467 | |
468 | if ((vid & VLAN_VID_MASK) != rx_ring->vid) |
469 | __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tci: vid); |
470 | else if (vid & VLAN_PRIO_MASK) |
471 | __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), |
472 | vlan_tci: vid & VLAN_PRIO_MASK); |
473 | } |
474 | |
475 | fm10k_type_trans(rx_ring, rx_desc, skb); |
476 | |
477 | return len; |
478 | } |
479 | |
480 | /** |
481 | * fm10k_is_non_eop - process handling of non-EOP buffers |
482 | * @rx_ring: Rx ring being processed |
483 | * @rx_desc: Rx descriptor for current buffer |
484 | * |
485 | * This function updates next to clean. If the buffer is an EOP buffer |
486 | * this function exits returning false, otherwise it will place the |
487 | * sk_buff in the next buffer to be chained and return true indicating |
488 | * that this is in fact a non-EOP buffer. |
489 | **/ |
490 | static bool fm10k_is_non_eop(struct fm10k_ring *rx_ring, |
491 | union fm10k_rx_desc *rx_desc) |
492 | { |
493 | u32 ntc = rx_ring->next_to_clean + 1; |
494 | |
495 | /* fetch, update, and store next to clean */ |
496 | ntc = (ntc < rx_ring->count) ? ntc : 0; |
497 | rx_ring->next_to_clean = ntc; |
498 | |
499 | prefetch(FM10K_RX_DESC(rx_ring, ntc)); |
500 | |
501 | if (likely(fm10k_test_staterr(rx_desc, FM10K_RXD_STATUS_EOP))) |
502 | return false; |
503 | |
504 | return true; |
505 | } |
506 | |
507 | /** |
508 | * fm10k_cleanup_headers - Correct corrupted or empty headers |
509 | * @rx_ring: rx descriptor ring packet is being transacted on |
510 | * @rx_desc: pointer to the EOP Rx descriptor |
511 | * @skb: pointer to current skb being fixed |
512 | * |
513 | * Address the case where we are pulling data in on pages only |
514 | * and as such no data is present in the skb header. |
515 | * |
516 | * In addition if skb is not at least 60 bytes we need to pad it so that |
517 | * it is large enough to qualify as a valid Ethernet frame. |
518 | * |
519 | * Returns true if an error was encountered and skb was freed. |
520 | **/ |
521 | static bool (struct fm10k_ring *rx_ring, |
522 | union fm10k_rx_desc *rx_desc, |
523 | struct sk_buff *skb) |
524 | { |
525 | if (unlikely((fm10k_test_staterr(rx_desc, |
526 | FM10K_RXD_STATUS_RXE)))) { |
527 | #define FM10K_TEST_RXD_BIT(rxd, bit) \ |
528 | ((rxd)->w.csum_err & cpu_to_le16(bit)) |
529 | if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_SWITCH_ERROR)) |
530 | rx_ring->rx_stats.switch_errors++; |
531 | if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_NO_DESCRIPTOR)) |
532 | rx_ring->rx_stats.drops++; |
533 | if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_PP_ERROR)) |
534 | rx_ring->rx_stats.pp_errors++; |
535 | if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_SWITCH_READY)) |
536 | rx_ring->rx_stats.link_errors++; |
537 | if (FM10K_TEST_RXD_BIT(rx_desc, FM10K_RXD_ERR_TOO_BIG)) |
538 | rx_ring->rx_stats.length_errors++; |
539 | dev_kfree_skb_any(skb); |
540 | rx_ring->rx_stats.errors++; |
541 | return true; |
542 | } |
543 | |
544 | /* if eth_skb_pad returns an error the skb was freed */ |
545 | if (eth_skb_pad(skb)) |
546 | return true; |
547 | |
548 | return false; |
549 | } |
550 | |
551 | /** |
552 | * fm10k_receive_skb - helper function to handle rx indications |
553 | * @q_vector: structure containing interrupt and ring information |
554 | * @skb: packet to send up |
555 | **/ |
556 | static void fm10k_receive_skb(struct fm10k_q_vector *q_vector, |
557 | struct sk_buff *skb) |
558 | { |
559 | napi_gro_receive(napi: &q_vector->napi, skb); |
560 | } |
561 | |
562 | static int fm10k_clean_rx_irq(struct fm10k_q_vector *q_vector, |
563 | struct fm10k_ring *rx_ring, |
564 | int budget) |
565 | { |
566 | struct sk_buff *skb = rx_ring->skb; |
567 | unsigned int total_bytes = 0, total_packets = 0; |
568 | u16 cleaned_count = fm10k_desc_unused(ring: rx_ring); |
569 | |
570 | while (likely(total_packets < budget)) { |
571 | union fm10k_rx_desc *rx_desc; |
572 | |
573 | /* return some buffers to hardware, one at a time is too slow */ |
574 | if (cleaned_count >= FM10K_RX_BUFFER_WRITE) { |
575 | fm10k_alloc_rx_buffers(rx_ring, cleaned_count); |
576 | cleaned_count = 0; |
577 | } |
578 | |
579 | rx_desc = FM10K_RX_DESC(rx_ring, rx_ring->next_to_clean); |
580 | |
581 | if (!rx_desc->d.staterr) |
582 | break; |
583 | |
584 | /* This memory barrier is needed to keep us from reading |
585 | * any other fields out of the rx_desc until we know the |
586 | * descriptor has been written back |
587 | */ |
588 | dma_rmb(); |
589 | |
590 | /* retrieve a buffer from the ring */ |
591 | skb = fm10k_fetch_rx_buffer(rx_ring, rx_desc, skb); |
592 | |
593 | /* exit if we failed to retrieve a buffer */ |
594 | if (!skb) |
595 | break; |
596 | |
597 | cleaned_count++; |
598 | |
599 | /* fetch next buffer in frame if non-eop */ |
600 | if (fm10k_is_non_eop(rx_ring, rx_desc)) |
601 | continue; |
602 | |
603 | /* verify the packet layout is correct */ |
604 | if (fm10k_cleanup_headers(rx_ring, rx_desc, skb)) { |
605 | skb = NULL; |
606 | continue; |
607 | } |
608 | |
609 | /* populate checksum, timestamp, VLAN, and protocol */ |
610 | total_bytes += fm10k_process_skb_fields(rx_ring, rx_desc, skb); |
611 | |
612 | fm10k_receive_skb(q_vector, skb); |
613 | |
614 | /* reset skb pointer */ |
615 | skb = NULL; |
616 | |
617 | /* update budget accounting */ |
618 | total_packets++; |
619 | } |
620 | |
621 | /* place incomplete frames back on ring for completion */ |
622 | rx_ring->skb = skb; |
623 | |
624 | u64_stats_update_begin(syncp: &rx_ring->syncp); |
625 | rx_ring->stats.packets += total_packets; |
626 | rx_ring->stats.bytes += total_bytes; |
627 | u64_stats_update_end(syncp: &rx_ring->syncp); |
628 | q_vector->rx.total_packets += total_packets; |
629 | q_vector->rx.total_bytes += total_bytes; |
630 | |
631 | return total_packets; |
632 | } |
633 | |
634 | #define VXLAN_HLEN (sizeof(struct udphdr) + 8) |
635 | static struct ethhdr *fm10k_port_is_vxlan(struct sk_buff *skb) |
636 | { |
637 | struct fm10k_intfc *interface = netdev_priv(dev: skb->dev); |
638 | |
639 | if (interface->vxlan_port != udp_hdr(skb)->dest) |
640 | return NULL; |
641 | |
642 | /* return offset of udp_hdr plus 8 bytes for VXLAN header */ |
643 | return (struct ethhdr *)(skb_transport_header(skb) + VXLAN_HLEN); |
644 | } |
645 | |
646 | #define FM10K_NVGRE_RESERVED0_FLAGS htons(0x9FFF) |
647 | #define NVGRE_TNI htons(0x2000) |
648 | struct fm10k_nvgre_hdr { |
649 | __be16 flags; |
650 | __be16 proto; |
651 | __be32 tni; |
652 | }; |
653 | |
654 | static struct ethhdr *fm10k_gre_is_nvgre(struct sk_buff *skb) |
655 | { |
656 | struct fm10k_nvgre_hdr *nvgre_hdr; |
657 | int hlen = ip_hdrlen(skb); |
658 | |
659 | /* currently only IPv4 is supported due to hlen above */ |
660 | if (vlan_get_protocol(skb) != htons(ETH_P_IP)) |
661 | return NULL; |
662 | |
663 | /* our transport header should be NVGRE */ |
664 | nvgre_hdr = (struct fm10k_nvgre_hdr *)(skb_network_header(skb) + hlen); |
665 | |
666 | /* verify all reserved flags are 0 */ |
667 | if (nvgre_hdr->flags & FM10K_NVGRE_RESERVED0_FLAGS) |
668 | return NULL; |
669 | |
670 | /* report start of ethernet header */ |
671 | if (nvgre_hdr->flags & NVGRE_TNI) |
672 | return (struct ethhdr *)(nvgre_hdr + 1); |
673 | |
674 | return (struct ethhdr *)(&nvgre_hdr->tni); |
675 | } |
676 | |
677 | __be16 fm10k_tx_encap_offload(struct sk_buff *skb) |
678 | { |
679 | u8 l4_hdr = 0, inner_l4_hdr = 0, inner_l4_hlen; |
680 | struct ethhdr *eth_hdr; |
681 | |
682 | if (skb->inner_protocol_type != ENCAP_TYPE_ETHER || |
683 | skb->inner_protocol != htons(ETH_P_TEB)) |
684 | return 0; |
685 | |
686 | switch (vlan_get_protocol(skb)) { |
687 | case htons(ETH_P_IP): |
688 | l4_hdr = ip_hdr(skb)->protocol; |
689 | break; |
690 | case htons(ETH_P_IPV6): |
691 | l4_hdr = ipv6_hdr(skb)->nexthdr; |
692 | break; |
693 | default: |
694 | return 0; |
695 | } |
696 | |
697 | switch (l4_hdr) { |
698 | case IPPROTO_UDP: |
699 | eth_hdr = fm10k_port_is_vxlan(skb); |
700 | break; |
701 | case IPPROTO_GRE: |
702 | eth_hdr = fm10k_gre_is_nvgre(skb); |
703 | break; |
704 | default: |
705 | return 0; |
706 | } |
707 | |
708 | if (!eth_hdr) |
709 | return 0; |
710 | |
711 | switch (eth_hdr->h_proto) { |
712 | case htons(ETH_P_IP): |
713 | inner_l4_hdr = inner_ip_hdr(skb)->protocol; |
714 | break; |
715 | case htons(ETH_P_IPV6): |
716 | inner_l4_hdr = inner_ipv6_hdr(skb)->nexthdr; |
717 | break; |
718 | default: |
719 | return 0; |
720 | } |
721 | |
722 | switch (inner_l4_hdr) { |
723 | case IPPROTO_TCP: |
724 | inner_l4_hlen = inner_tcp_hdrlen(skb); |
725 | break; |
726 | case IPPROTO_UDP: |
727 | inner_l4_hlen = 8; |
728 | break; |
729 | default: |
730 | return 0; |
731 | } |
732 | |
733 | /* The hardware allows tunnel offloads only if the combined inner and |
734 | * outer header is 184 bytes or less |
735 | */ |
736 | if (skb_inner_transport_header(skb) + inner_l4_hlen - |
737 | skb_mac_header(skb) > FM10K_TUNNEL_HEADER_LENGTH) |
738 | return 0; |
739 | |
740 | return eth_hdr->h_proto; |
741 | } |
742 | |
743 | static int fm10k_tso(struct fm10k_ring *tx_ring, |
744 | struct fm10k_tx_buffer *first) |
745 | { |
746 | struct sk_buff *skb = first->skb; |
747 | struct fm10k_tx_desc *tx_desc; |
748 | unsigned char *th; |
749 | u8 hdrlen; |
750 | |
751 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
752 | return 0; |
753 | |
754 | if (!skb_is_gso(skb)) |
755 | return 0; |
756 | |
757 | /* compute header lengths */ |
758 | if (skb->encapsulation) { |
759 | if (!fm10k_tx_encap_offload(skb)) |
760 | goto err_vxlan; |
761 | th = skb_inner_transport_header(skb); |
762 | } else { |
763 | th = skb_transport_header(skb); |
764 | } |
765 | |
766 | /* compute offset from SOF to transport header and add header len */ |
767 | hdrlen = (th - skb->data) + (((struct tcphdr *)th)->doff << 2); |
768 | |
769 | first->tx_flags |= FM10K_TX_FLAGS_CSUM; |
770 | |
771 | /* update gso size and bytecount with header size */ |
772 | first->gso_segs = skb_shinfo(skb)->gso_segs; |
773 | first->bytecount += (first->gso_segs - 1) * hdrlen; |
774 | |
775 | /* populate Tx descriptor header size and mss */ |
776 | tx_desc = FM10K_TX_DESC(tx_ring, tx_ring->next_to_use); |
777 | tx_desc->hdrlen = hdrlen; |
778 | tx_desc->mss = cpu_to_le16(skb_shinfo(skb)->gso_size); |
779 | |
780 | return 1; |
781 | |
782 | err_vxlan: |
783 | tx_ring->netdev->features &= ~NETIF_F_GSO_UDP_TUNNEL; |
784 | if (net_ratelimit()) |
785 | netdev_err(dev: tx_ring->netdev, |
786 | format: "TSO requested for unsupported tunnel, disabling offload\n" ); |
787 | return -1; |
788 | } |
789 | |
790 | static void fm10k_tx_csum(struct fm10k_ring *tx_ring, |
791 | struct fm10k_tx_buffer *first) |
792 | { |
793 | struct sk_buff *skb = first->skb; |
794 | struct fm10k_tx_desc *tx_desc; |
795 | union { |
796 | struct iphdr *ipv4; |
797 | struct ipv6hdr *ipv6; |
798 | u8 *raw; |
799 | } network_hdr; |
800 | u8 *transport_hdr; |
801 | __be16 frag_off; |
802 | __be16 protocol; |
803 | u8 l4_hdr = 0; |
804 | |
805 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
806 | goto no_csum; |
807 | |
808 | if (skb->encapsulation) { |
809 | protocol = fm10k_tx_encap_offload(skb); |
810 | if (!protocol) { |
811 | if (skb_checksum_help(skb)) { |
812 | dev_warn(tx_ring->dev, |
813 | "failed to offload encap csum!\n" ); |
814 | tx_ring->tx_stats.csum_err++; |
815 | } |
816 | goto no_csum; |
817 | } |
818 | network_hdr.raw = skb_inner_network_header(skb); |
819 | transport_hdr = skb_inner_transport_header(skb); |
820 | } else { |
821 | protocol = vlan_get_protocol(skb); |
822 | network_hdr.raw = skb_network_header(skb); |
823 | transport_hdr = skb_transport_header(skb); |
824 | } |
825 | |
826 | switch (protocol) { |
827 | case htons(ETH_P_IP): |
828 | l4_hdr = network_hdr.ipv4->protocol; |
829 | break; |
830 | case htons(ETH_P_IPV6): |
831 | l4_hdr = network_hdr.ipv6->nexthdr; |
832 | if (likely((transport_hdr - network_hdr.raw) == |
833 | sizeof(struct ipv6hdr))) |
834 | break; |
835 | ipv6_skip_exthdr(skb, start: network_hdr.raw - skb->data + |
836 | sizeof(struct ipv6hdr), |
837 | nexthdrp: &l4_hdr, frag_offp: &frag_off); |
838 | if (unlikely(frag_off)) |
839 | l4_hdr = NEXTHDR_FRAGMENT; |
840 | break; |
841 | default: |
842 | break; |
843 | } |
844 | |
845 | switch (l4_hdr) { |
846 | case IPPROTO_TCP: |
847 | case IPPROTO_UDP: |
848 | break; |
849 | case IPPROTO_GRE: |
850 | if (skb->encapsulation) |
851 | break; |
852 | fallthrough; |
853 | default: |
854 | if (unlikely(net_ratelimit())) { |
855 | dev_warn(tx_ring->dev, |
856 | "partial checksum, version=%d l4 proto=%x\n" , |
857 | protocol, l4_hdr); |
858 | } |
859 | skb_checksum_help(skb); |
860 | tx_ring->tx_stats.csum_err++; |
861 | goto no_csum; |
862 | } |
863 | |
864 | /* update TX checksum flag */ |
865 | first->tx_flags |= FM10K_TX_FLAGS_CSUM; |
866 | tx_ring->tx_stats.csum_good++; |
867 | |
868 | no_csum: |
869 | /* populate Tx descriptor header size and mss */ |
870 | tx_desc = FM10K_TX_DESC(tx_ring, tx_ring->next_to_use); |
871 | tx_desc->hdrlen = 0; |
872 | tx_desc->mss = 0; |
873 | } |
874 | |
875 | #define FM10K_SET_FLAG(_input, _flag, _result) \ |
876 | ((_flag <= _result) ? \ |
877 | ((u32)(_input & _flag) * (_result / _flag)) : \ |
878 | ((u32)(_input & _flag) / (_flag / _result))) |
879 | |
880 | static u8 fm10k_tx_desc_flags(struct sk_buff *skb, u32 tx_flags) |
881 | { |
882 | /* set type for advanced descriptor with frame checksum insertion */ |
883 | u32 desc_flags = 0; |
884 | |
885 | /* set checksum offload bits */ |
886 | desc_flags |= FM10K_SET_FLAG(tx_flags, FM10K_TX_FLAGS_CSUM, |
887 | FM10K_TXD_FLAG_CSUM); |
888 | |
889 | return desc_flags; |
890 | } |
891 | |
892 | static bool fm10k_tx_desc_push(struct fm10k_ring *tx_ring, |
893 | struct fm10k_tx_desc *tx_desc, u16 i, |
894 | dma_addr_t dma, unsigned int size, u8 desc_flags) |
895 | { |
896 | /* set RS and INT for last frame in a cache line */ |
897 | if ((++i & (FM10K_TXD_WB_FIFO_SIZE - 1)) == 0) |
898 | desc_flags |= FM10K_TXD_FLAG_RS | FM10K_TXD_FLAG_INT; |
899 | |
900 | /* record values to descriptor */ |
901 | tx_desc->buffer_addr = cpu_to_le64(dma); |
902 | tx_desc->flags = desc_flags; |
903 | tx_desc->buflen = cpu_to_le16(size); |
904 | |
905 | /* return true if we just wrapped the ring */ |
906 | return i == tx_ring->count; |
907 | } |
908 | |
909 | static int __fm10k_maybe_stop_tx(struct fm10k_ring *tx_ring, u16 size) |
910 | { |
911 | netif_stop_subqueue(dev: tx_ring->netdev, queue_index: tx_ring->queue_index); |
912 | |
913 | /* Memory barrier before checking head and tail */ |
914 | smp_mb(); |
915 | |
916 | /* Check again in a case another CPU has just made room available */ |
917 | if (likely(fm10k_desc_unused(tx_ring) < size)) |
918 | return -EBUSY; |
919 | |
920 | /* A reprieve! - use start_queue because it doesn't call schedule */ |
921 | netif_start_subqueue(dev: tx_ring->netdev, queue_index: tx_ring->queue_index); |
922 | ++tx_ring->tx_stats.restart_queue; |
923 | return 0; |
924 | } |
925 | |
926 | static inline int fm10k_maybe_stop_tx(struct fm10k_ring *tx_ring, u16 size) |
927 | { |
928 | if (likely(fm10k_desc_unused(tx_ring) >= size)) |
929 | return 0; |
930 | return __fm10k_maybe_stop_tx(tx_ring, size); |
931 | } |
932 | |
933 | static void fm10k_tx_map(struct fm10k_ring *tx_ring, |
934 | struct fm10k_tx_buffer *first) |
935 | { |
936 | struct sk_buff *skb = first->skb; |
937 | struct fm10k_tx_buffer *tx_buffer; |
938 | struct fm10k_tx_desc *tx_desc; |
939 | skb_frag_t *frag; |
940 | unsigned char *data; |
941 | dma_addr_t dma; |
942 | unsigned int data_len, size; |
943 | u32 tx_flags = first->tx_flags; |
944 | u16 i = tx_ring->next_to_use; |
945 | u8 flags = fm10k_tx_desc_flags(skb, tx_flags); |
946 | |
947 | tx_desc = FM10K_TX_DESC(tx_ring, i); |
948 | |
949 | /* add HW VLAN tag */ |
950 | if (skb_vlan_tag_present(skb)) |
951 | tx_desc->vlan = cpu_to_le16(skb_vlan_tag_get(skb)); |
952 | else |
953 | tx_desc->vlan = 0; |
954 | |
955 | size = skb_headlen(skb); |
956 | data = skb->data; |
957 | |
958 | dma = dma_map_single(tx_ring->dev, data, size, DMA_TO_DEVICE); |
959 | |
960 | data_len = skb->data_len; |
961 | tx_buffer = first; |
962 | |
963 | for (frag = &skb_shinfo(skb)->frags[0];; frag++) { |
964 | if (dma_mapping_error(dev: tx_ring->dev, dma_addr: dma)) |
965 | goto dma_error; |
966 | |
967 | /* record length, and DMA address */ |
968 | dma_unmap_len_set(tx_buffer, len, size); |
969 | dma_unmap_addr_set(tx_buffer, dma, dma); |
970 | |
971 | while (unlikely(size > FM10K_MAX_DATA_PER_TXD)) { |
972 | if (fm10k_tx_desc_push(tx_ring, tx_desc: tx_desc++, i: i++, dma, |
973 | FM10K_MAX_DATA_PER_TXD, desc_flags: flags)) { |
974 | tx_desc = FM10K_TX_DESC(tx_ring, 0); |
975 | i = 0; |
976 | } |
977 | |
978 | dma += FM10K_MAX_DATA_PER_TXD; |
979 | size -= FM10K_MAX_DATA_PER_TXD; |
980 | } |
981 | |
982 | if (likely(!data_len)) |
983 | break; |
984 | |
985 | if (fm10k_tx_desc_push(tx_ring, tx_desc: tx_desc++, i: i++, |
986 | dma, size, desc_flags: flags)) { |
987 | tx_desc = FM10K_TX_DESC(tx_ring, 0); |
988 | i = 0; |
989 | } |
990 | |
991 | size = skb_frag_size(frag); |
992 | data_len -= size; |
993 | |
994 | dma = skb_frag_dma_map(dev: tx_ring->dev, frag, offset: 0, size, |
995 | dir: DMA_TO_DEVICE); |
996 | |
997 | tx_buffer = &tx_ring->tx_buffer[i]; |
998 | } |
999 | |
1000 | /* write last descriptor with LAST bit set */ |
1001 | flags |= FM10K_TXD_FLAG_LAST; |
1002 | |
1003 | if (fm10k_tx_desc_push(tx_ring, tx_desc, i: i++, dma, size, desc_flags: flags)) |
1004 | i = 0; |
1005 | |
1006 | /* record bytecount for BQL */ |
1007 | netdev_tx_sent_queue(dev_queue: txring_txq(ring: tx_ring), bytes: first->bytecount); |
1008 | |
1009 | /* record SW timestamp if HW timestamp is not available */ |
1010 | skb_tx_timestamp(skb: first->skb); |
1011 | |
1012 | /* Force memory writes to complete before letting h/w know there |
1013 | * are new descriptors to fetch. (Only applicable for weak-ordered |
1014 | * memory model archs, such as IA-64). |
1015 | * |
1016 | * We also need this memory barrier to make certain all of the |
1017 | * status bits have been updated before next_to_watch is written. |
1018 | */ |
1019 | wmb(); |
1020 | |
1021 | /* set next_to_watch value indicating a packet is present */ |
1022 | first->next_to_watch = tx_desc; |
1023 | |
1024 | tx_ring->next_to_use = i; |
1025 | |
1026 | /* Make sure there is space in the ring for the next send. */ |
1027 | fm10k_maybe_stop_tx(tx_ring, DESC_NEEDED); |
1028 | |
1029 | /* notify HW of packet */ |
1030 | if (netif_xmit_stopped(dev_queue: txring_txq(ring: tx_ring)) || !netdev_xmit_more()) { |
1031 | writel(val: i, addr: tx_ring->tail); |
1032 | } |
1033 | |
1034 | return; |
1035 | dma_error: |
1036 | dev_err(tx_ring->dev, "TX DMA map failed\n" ); |
1037 | |
1038 | /* clear dma mappings for failed tx_buffer map */ |
1039 | for (;;) { |
1040 | tx_buffer = &tx_ring->tx_buffer[i]; |
1041 | fm10k_unmap_and_free_tx_resource(tx_ring, tx_buffer); |
1042 | if (tx_buffer == first) |
1043 | break; |
1044 | if (i == 0) |
1045 | i = tx_ring->count; |
1046 | i--; |
1047 | } |
1048 | |
1049 | tx_ring->next_to_use = i; |
1050 | } |
1051 | |
1052 | netdev_tx_t fm10k_xmit_frame_ring(struct sk_buff *skb, |
1053 | struct fm10k_ring *tx_ring) |
1054 | { |
1055 | u16 count = TXD_USE_COUNT(skb_headlen(skb)); |
1056 | struct fm10k_tx_buffer *first; |
1057 | unsigned short f; |
1058 | u32 tx_flags = 0; |
1059 | int tso; |
1060 | |
1061 | /* need: 1 descriptor per page * PAGE_SIZE/FM10K_MAX_DATA_PER_TXD, |
1062 | * + 1 desc for skb_headlen/FM10K_MAX_DATA_PER_TXD, |
1063 | * + 2 desc gap to keep tail from touching head |
1064 | * otherwise try next time |
1065 | */ |
1066 | for (f = 0; f < skb_shinfo(skb)->nr_frags; f++) { |
1067 | skb_frag_t *frag = &skb_shinfo(skb)->frags[f]; |
1068 | |
1069 | count += TXD_USE_COUNT(skb_frag_size(frag)); |
1070 | } |
1071 | |
1072 | if (fm10k_maybe_stop_tx(tx_ring, size: count + 3)) { |
1073 | tx_ring->tx_stats.tx_busy++; |
1074 | return NETDEV_TX_BUSY; |
1075 | } |
1076 | |
1077 | /* record the location of the first descriptor for this packet */ |
1078 | first = &tx_ring->tx_buffer[tx_ring->next_to_use]; |
1079 | first->skb = skb; |
1080 | first->bytecount = max_t(unsigned int, skb->len, ETH_ZLEN); |
1081 | first->gso_segs = 1; |
1082 | |
1083 | /* record initial flags and protocol */ |
1084 | first->tx_flags = tx_flags; |
1085 | |
1086 | tso = fm10k_tso(tx_ring, first); |
1087 | if (tso < 0) |
1088 | goto out_drop; |
1089 | else if (!tso) |
1090 | fm10k_tx_csum(tx_ring, first); |
1091 | |
1092 | fm10k_tx_map(tx_ring, first); |
1093 | |
1094 | return NETDEV_TX_OK; |
1095 | |
1096 | out_drop: |
1097 | dev_kfree_skb_any(skb: first->skb); |
1098 | first->skb = NULL; |
1099 | |
1100 | return NETDEV_TX_OK; |
1101 | } |
1102 | |
1103 | static u64 fm10k_get_tx_completed(struct fm10k_ring *ring) |
1104 | { |
1105 | return ring->stats.packets; |
1106 | } |
1107 | |
1108 | /** |
1109 | * fm10k_get_tx_pending - how many Tx descriptors not processed |
1110 | * @ring: the ring structure |
1111 | * @in_sw: is tx_pending being checked in SW or in HW? |
1112 | */ |
1113 | u64 fm10k_get_tx_pending(struct fm10k_ring *ring, bool in_sw) |
1114 | { |
1115 | struct fm10k_intfc *interface = ring->q_vector->interface; |
1116 | struct fm10k_hw *hw = &interface->hw; |
1117 | u32 head, tail; |
1118 | |
1119 | if (likely(in_sw)) { |
1120 | head = ring->next_to_clean; |
1121 | tail = ring->next_to_use; |
1122 | } else { |
1123 | head = fm10k_read_reg(hw, FM10K_TDH(ring->reg_idx)); |
1124 | tail = fm10k_read_reg(hw, FM10K_TDT(ring->reg_idx)); |
1125 | } |
1126 | |
1127 | return ((head <= tail) ? tail : tail + ring->count) - head; |
1128 | } |
1129 | |
1130 | bool fm10k_check_tx_hang(struct fm10k_ring *tx_ring) |
1131 | { |
1132 | u32 tx_done = fm10k_get_tx_completed(ring: tx_ring); |
1133 | u32 tx_done_old = tx_ring->tx_stats.tx_done_old; |
1134 | u32 tx_pending = fm10k_get_tx_pending(ring: tx_ring, in_sw: true); |
1135 | |
1136 | clear_check_for_tx_hang(tx_ring); |
1137 | |
1138 | /* Check for a hung queue, but be thorough. This verifies |
1139 | * that a transmit has been completed since the previous |
1140 | * check AND there is at least one packet pending. By |
1141 | * requiring this to fail twice we avoid races with |
1142 | * clearing the ARMED bit and conditions where we |
1143 | * run the check_tx_hang logic with a transmit completion |
1144 | * pending but without time to complete it yet. |
1145 | */ |
1146 | if (!tx_pending || (tx_done_old != tx_done)) { |
1147 | /* update completed stats and continue */ |
1148 | tx_ring->tx_stats.tx_done_old = tx_done; |
1149 | /* reset the countdown */ |
1150 | clear_bit(nr: __FM10K_HANG_CHECK_ARMED, addr: tx_ring->state); |
1151 | |
1152 | return false; |
1153 | } |
1154 | |
1155 | /* make sure it is true for two checks in a row */ |
1156 | return test_and_set_bit(nr: __FM10K_HANG_CHECK_ARMED, addr: tx_ring->state); |
1157 | } |
1158 | |
1159 | /** |
1160 | * fm10k_tx_timeout_reset - initiate reset due to Tx timeout |
1161 | * @interface: driver private struct |
1162 | **/ |
1163 | void fm10k_tx_timeout_reset(struct fm10k_intfc *interface) |
1164 | { |
1165 | /* Do the reset outside of interrupt context */ |
1166 | if (!test_bit(__FM10K_DOWN, interface->state)) { |
1167 | interface->tx_timeout_count++; |
1168 | set_bit(nr: FM10K_FLAG_RESET_REQUESTED, addr: interface->flags); |
1169 | fm10k_service_event_schedule(interface); |
1170 | } |
1171 | } |
1172 | |
1173 | /** |
1174 | * fm10k_clean_tx_irq - Reclaim resources after transmit completes |
1175 | * @q_vector: structure containing interrupt and ring information |
1176 | * @tx_ring: tx ring to clean |
1177 | * @napi_budget: Used to determine if we are in netpoll |
1178 | **/ |
1179 | static bool fm10k_clean_tx_irq(struct fm10k_q_vector *q_vector, |
1180 | struct fm10k_ring *tx_ring, int napi_budget) |
1181 | { |
1182 | struct fm10k_intfc *interface = q_vector->interface; |
1183 | struct fm10k_tx_buffer *tx_buffer; |
1184 | struct fm10k_tx_desc *tx_desc; |
1185 | unsigned int total_bytes = 0, total_packets = 0; |
1186 | unsigned int budget = q_vector->tx.work_limit; |
1187 | unsigned int i = tx_ring->next_to_clean; |
1188 | |
1189 | if (test_bit(__FM10K_DOWN, interface->state)) |
1190 | return true; |
1191 | |
1192 | tx_buffer = &tx_ring->tx_buffer[i]; |
1193 | tx_desc = FM10K_TX_DESC(tx_ring, i); |
1194 | i -= tx_ring->count; |
1195 | |
1196 | do { |
1197 | struct fm10k_tx_desc *eop_desc = tx_buffer->next_to_watch; |
1198 | |
1199 | /* if next_to_watch is not set then there is no work pending */ |
1200 | if (!eop_desc) |
1201 | break; |
1202 | |
1203 | /* prevent any other reads prior to eop_desc */ |
1204 | smp_rmb(); |
1205 | |
1206 | /* if DD is not set pending work has not been completed */ |
1207 | if (!(eop_desc->flags & FM10K_TXD_FLAG_DONE)) |
1208 | break; |
1209 | |
1210 | /* clear next_to_watch to prevent false hangs */ |
1211 | tx_buffer->next_to_watch = NULL; |
1212 | |
1213 | /* update the statistics for this packet */ |
1214 | total_bytes += tx_buffer->bytecount; |
1215 | total_packets += tx_buffer->gso_segs; |
1216 | |
1217 | /* free the skb */ |
1218 | napi_consume_skb(skb: tx_buffer->skb, budget: napi_budget); |
1219 | |
1220 | /* unmap skb header data */ |
1221 | dma_unmap_single(tx_ring->dev, |
1222 | dma_unmap_addr(tx_buffer, dma), |
1223 | dma_unmap_len(tx_buffer, len), |
1224 | DMA_TO_DEVICE); |
1225 | |
1226 | /* clear tx_buffer data */ |
1227 | tx_buffer->skb = NULL; |
1228 | dma_unmap_len_set(tx_buffer, len, 0); |
1229 | |
1230 | /* unmap remaining buffers */ |
1231 | while (tx_desc != eop_desc) { |
1232 | tx_buffer++; |
1233 | tx_desc++; |
1234 | i++; |
1235 | if (unlikely(!i)) { |
1236 | i -= tx_ring->count; |
1237 | tx_buffer = tx_ring->tx_buffer; |
1238 | tx_desc = FM10K_TX_DESC(tx_ring, 0); |
1239 | } |
1240 | |
1241 | /* unmap any remaining paged data */ |
1242 | if (dma_unmap_len(tx_buffer, len)) { |
1243 | dma_unmap_page(tx_ring->dev, |
1244 | dma_unmap_addr(tx_buffer, dma), |
1245 | dma_unmap_len(tx_buffer, len), |
1246 | DMA_TO_DEVICE); |
1247 | dma_unmap_len_set(tx_buffer, len, 0); |
1248 | } |
1249 | } |
1250 | |
1251 | /* move us one more past the eop_desc for start of next pkt */ |
1252 | tx_buffer++; |
1253 | tx_desc++; |
1254 | i++; |
1255 | if (unlikely(!i)) { |
1256 | i -= tx_ring->count; |
1257 | tx_buffer = tx_ring->tx_buffer; |
1258 | tx_desc = FM10K_TX_DESC(tx_ring, 0); |
1259 | } |
1260 | |
1261 | /* issue prefetch for next Tx descriptor */ |
1262 | prefetch(tx_desc); |
1263 | |
1264 | /* update budget accounting */ |
1265 | budget--; |
1266 | } while (likely(budget)); |
1267 | |
1268 | i += tx_ring->count; |
1269 | tx_ring->next_to_clean = i; |
1270 | u64_stats_update_begin(syncp: &tx_ring->syncp); |
1271 | tx_ring->stats.bytes += total_bytes; |
1272 | tx_ring->stats.packets += total_packets; |
1273 | u64_stats_update_end(syncp: &tx_ring->syncp); |
1274 | q_vector->tx.total_bytes += total_bytes; |
1275 | q_vector->tx.total_packets += total_packets; |
1276 | |
1277 | if (check_for_tx_hang(tx_ring) && fm10k_check_tx_hang(tx_ring)) { |
1278 | /* schedule immediate reset if we believe we hung */ |
1279 | struct fm10k_hw *hw = &interface->hw; |
1280 | |
1281 | netif_err(interface, drv, tx_ring->netdev, |
1282 | "Detected Tx Unit Hang\n" |
1283 | " Tx Queue <%d>\n" |
1284 | " TDH, TDT <%x>, <%x>\n" |
1285 | " next_to_use <%x>\n" |
1286 | " next_to_clean <%x>\n" , |
1287 | tx_ring->queue_index, |
1288 | fm10k_read_reg(hw, FM10K_TDH(tx_ring->reg_idx)), |
1289 | fm10k_read_reg(hw, FM10K_TDT(tx_ring->reg_idx)), |
1290 | tx_ring->next_to_use, i); |
1291 | |
1292 | netif_stop_subqueue(dev: tx_ring->netdev, |
1293 | queue_index: tx_ring->queue_index); |
1294 | |
1295 | netif_info(interface, probe, tx_ring->netdev, |
1296 | "tx hang %d detected on queue %d, resetting interface\n" , |
1297 | interface->tx_timeout_count + 1, |
1298 | tx_ring->queue_index); |
1299 | |
1300 | fm10k_tx_timeout_reset(interface); |
1301 | |
1302 | /* the netdev is about to reset, no point in enabling stuff */ |
1303 | return true; |
1304 | } |
1305 | |
1306 | /* notify netdev of completed buffers */ |
1307 | netdev_tx_completed_queue(dev_queue: txring_txq(ring: tx_ring), |
1308 | pkts: total_packets, bytes: total_bytes); |
1309 | |
1310 | #define TX_WAKE_THRESHOLD min_t(u16, FM10K_MIN_TXD - 1, DESC_NEEDED * 2) |
1311 | if (unlikely(total_packets && netif_carrier_ok(tx_ring->netdev) && |
1312 | (fm10k_desc_unused(tx_ring) >= TX_WAKE_THRESHOLD))) { |
1313 | /* Make sure that anybody stopping the queue after this |
1314 | * sees the new next_to_clean. |
1315 | */ |
1316 | smp_mb(); |
1317 | if (__netif_subqueue_stopped(dev: tx_ring->netdev, |
1318 | queue_index: tx_ring->queue_index) && |
1319 | !test_bit(__FM10K_DOWN, interface->state)) { |
1320 | netif_wake_subqueue(dev: tx_ring->netdev, |
1321 | queue_index: tx_ring->queue_index); |
1322 | ++tx_ring->tx_stats.restart_queue; |
1323 | } |
1324 | } |
1325 | |
1326 | return !!budget; |
1327 | } |
1328 | |
1329 | /** |
1330 | * fm10k_update_itr - update the dynamic ITR value based on packet size |
1331 | * |
1332 | * Stores a new ITR value based on strictly on packet size. The |
1333 | * divisors and thresholds used by this function were determined based |
1334 | * on theoretical maximum wire speed and testing data, in order to |
1335 | * minimize response time while increasing bulk throughput. |
1336 | * |
1337 | * @ring_container: Container for rings to have ITR updated |
1338 | **/ |
1339 | static void fm10k_update_itr(struct fm10k_ring_container *ring_container) |
1340 | { |
1341 | unsigned int avg_wire_size, packets, itr_round; |
1342 | |
1343 | /* Only update ITR if we are using adaptive setting */ |
1344 | if (!ITR_IS_ADAPTIVE(ring_container->itr)) |
1345 | goto clear_counts; |
1346 | |
1347 | packets = ring_container->total_packets; |
1348 | if (!packets) |
1349 | goto clear_counts; |
1350 | |
1351 | avg_wire_size = ring_container->total_bytes / packets; |
1352 | |
1353 | /* The following is a crude approximation of: |
1354 | * wmem_default / (size + overhead) = desired_pkts_per_int |
1355 | * rate / bits_per_byte / (size + ethernet overhead) = pkt_rate |
1356 | * (desired_pkt_rate / pkt_rate) * usecs_per_sec = ITR value |
1357 | * |
1358 | * Assuming wmem_default is 212992 and overhead is 640 bytes per |
1359 | * packet, (256 skb, 64 headroom, 320 shared info), we can reduce the |
1360 | * formula down to |
1361 | * |
1362 | * (34 * (size + 24)) / (size + 640) = ITR |
1363 | * |
1364 | * We first do some math on the packet size and then finally bitshift |
1365 | * by 8 after rounding up. We also have to account for PCIe link speed |
1366 | * difference as ITR scales based on this. |
1367 | */ |
1368 | if (avg_wire_size <= 360) { |
1369 | /* Start at 250K ints/sec and gradually drop to 77K ints/sec */ |
1370 | avg_wire_size *= 8; |
1371 | avg_wire_size += 376; |
1372 | } else if (avg_wire_size <= 1152) { |
1373 | /* 77K ints/sec to 45K ints/sec */ |
1374 | avg_wire_size *= 3; |
1375 | avg_wire_size += 2176; |
1376 | } else if (avg_wire_size <= 1920) { |
1377 | /* 45K ints/sec to 38K ints/sec */ |
1378 | avg_wire_size += 4480; |
1379 | } else { |
1380 | /* plateau at a limit of 38K ints/sec */ |
1381 | avg_wire_size = 6656; |
1382 | } |
1383 | |
1384 | /* Perform final bitshift for division after rounding up to ensure |
1385 | * that the calculation will never get below a 1. The bit shift |
1386 | * accounts for changes in the ITR due to PCIe link speed. |
1387 | */ |
1388 | itr_round = READ_ONCE(ring_container->itr_scale) + 8; |
1389 | avg_wire_size += BIT(itr_round) - 1; |
1390 | avg_wire_size >>= itr_round; |
1391 | |
1392 | /* write back value and retain adaptive flag */ |
1393 | ring_container->itr = avg_wire_size | FM10K_ITR_ADAPTIVE; |
1394 | |
1395 | clear_counts: |
1396 | ring_container->total_bytes = 0; |
1397 | ring_container->total_packets = 0; |
1398 | } |
1399 | |
1400 | static void fm10k_qv_enable(struct fm10k_q_vector *q_vector) |
1401 | { |
1402 | /* Enable auto-mask and clear the current mask */ |
1403 | u32 itr = FM10K_ITR_ENABLE; |
1404 | |
1405 | /* Update Tx ITR */ |
1406 | fm10k_update_itr(ring_container: &q_vector->tx); |
1407 | |
1408 | /* Update Rx ITR */ |
1409 | fm10k_update_itr(ring_container: &q_vector->rx); |
1410 | |
1411 | /* Store Tx itr in timer slot 0 */ |
1412 | itr |= (q_vector->tx.itr & FM10K_ITR_MAX); |
1413 | |
1414 | /* Shift Rx itr to timer slot 1 */ |
1415 | itr |= (q_vector->rx.itr & FM10K_ITR_MAX) << FM10K_ITR_INTERVAL1_SHIFT; |
1416 | |
1417 | /* Write the final value to the ITR register */ |
1418 | writel(val: itr, addr: q_vector->itr); |
1419 | } |
1420 | |
1421 | static int fm10k_poll(struct napi_struct *napi, int budget) |
1422 | { |
1423 | struct fm10k_q_vector *q_vector = |
1424 | container_of(napi, struct fm10k_q_vector, napi); |
1425 | struct fm10k_ring *ring; |
1426 | int per_ring_budget, work_done = 0; |
1427 | bool clean_complete = true; |
1428 | |
1429 | fm10k_for_each_ring(ring, q_vector->tx) { |
1430 | if (!fm10k_clean_tx_irq(q_vector, tx_ring: ring, napi_budget: budget)) |
1431 | clean_complete = false; |
1432 | } |
1433 | |
1434 | /* Handle case where we are called by netpoll with a budget of 0 */ |
1435 | if (budget <= 0) |
1436 | return budget; |
1437 | |
1438 | /* attempt to distribute budget to each queue fairly, but don't |
1439 | * allow the budget to go below 1 because we'll exit polling |
1440 | */ |
1441 | if (q_vector->rx.count > 1) |
1442 | per_ring_budget = max(budget / q_vector->rx.count, 1); |
1443 | else |
1444 | per_ring_budget = budget; |
1445 | |
1446 | fm10k_for_each_ring(ring, q_vector->rx) { |
1447 | int work = fm10k_clean_rx_irq(q_vector, rx_ring: ring, budget: per_ring_budget); |
1448 | |
1449 | work_done += work; |
1450 | if (work >= per_ring_budget) |
1451 | clean_complete = false; |
1452 | } |
1453 | |
1454 | /* If all work not completed, return budget and keep polling */ |
1455 | if (!clean_complete) |
1456 | return budget; |
1457 | |
1458 | /* Exit the polling mode, but don't re-enable interrupts if stack might |
1459 | * poll us due to busy-polling |
1460 | */ |
1461 | if (likely(napi_complete_done(napi, work_done))) |
1462 | fm10k_qv_enable(q_vector); |
1463 | |
1464 | return min(work_done, budget - 1); |
1465 | } |
1466 | |
1467 | /** |
1468 | * fm10k_set_qos_queues: Allocate queues for a QOS-enabled device |
1469 | * @interface: board private structure to initialize |
1470 | * |
1471 | * When QoS (Quality of Service) is enabled, allocate queues for |
1472 | * each traffic class. If multiqueue isn't available,then abort QoS |
1473 | * initialization. |
1474 | * |
1475 | * This function handles all combinations of Qos and RSS. |
1476 | * |
1477 | **/ |
1478 | static bool fm10k_set_qos_queues(struct fm10k_intfc *interface) |
1479 | { |
1480 | struct net_device *dev = interface->netdev; |
1481 | struct fm10k_ring_feature *f; |
1482 | int , i; |
1483 | int pcs; |
1484 | |
1485 | /* Map queue offset and counts onto allocated tx queues */ |
1486 | pcs = netdev_get_num_tc(dev); |
1487 | |
1488 | if (pcs <= 1) |
1489 | return false; |
1490 | |
1491 | /* set QoS mask and indices */ |
1492 | f = &interface->ring_feature[RING_F_QOS]; |
1493 | f->indices = pcs; |
1494 | f->mask = BIT(fls(pcs - 1)) - 1; |
1495 | |
1496 | /* determine the upper limit for our current DCB mode */ |
1497 | rss_i = interface->hw.mac.max_queues / pcs; |
1498 | rss_i = BIT(fls(rss_i) - 1); |
1499 | |
1500 | /* set RSS mask and indices */ |
1501 | f = &interface->ring_feature[RING_F_RSS]; |
1502 | rss_i = min_t(u16, rss_i, f->limit); |
1503 | f->indices = rss_i; |
1504 | f->mask = BIT(fls(rss_i - 1)) - 1; |
1505 | |
1506 | /* configure pause class to queue mapping */ |
1507 | for (i = 0; i < pcs; i++) |
1508 | netdev_set_tc_queue(dev, tc: i, count: rss_i, offset: rss_i * i); |
1509 | |
1510 | interface->num_rx_queues = rss_i * pcs; |
1511 | interface->num_tx_queues = rss_i * pcs; |
1512 | |
1513 | return true; |
1514 | } |
1515 | |
1516 | /** |
1517 | * fm10k_set_rss_queues: Allocate queues for RSS |
1518 | * @interface: board private structure to initialize |
1519 | * |
1520 | * This is our "base" multiqueue mode. RSS (Receive Side Scaling) will try |
1521 | * to allocate one Rx queue per CPU, and if available, one Tx queue per CPU. |
1522 | * |
1523 | **/ |
1524 | static bool (struct fm10k_intfc *interface) |
1525 | { |
1526 | struct fm10k_ring_feature *f; |
1527 | u16 ; |
1528 | |
1529 | f = &interface->ring_feature[RING_F_RSS]; |
1530 | rss_i = min_t(u16, interface->hw.mac.max_queues, f->limit); |
1531 | |
1532 | /* record indices and power of 2 mask for RSS */ |
1533 | f->indices = rss_i; |
1534 | f->mask = BIT(fls(rss_i - 1)) - 1; |
1535 | |
1536 | interface->num_rx_queues = rss_i; |
1537 | interface->num_tx_queues = rss_i; |
1538 | |
1539 | return true; |
1540 | } |
1541 | |
1542 | /** |
1543 | * fm10k_set_num_queues: Allocate queues for device, feature dependent |
1544 | * @interface: board private structure to initialize |
1545 | * |
1546 | * This is the top level queue allocation routine. The order here is very |
1547 | * important, starting with the "most" number of features turned on at once, |
1548 | * and ending with the smallest set of features. This way large combinations |
1549 | * can be allocated if they're turned on, and smaller combinations are the |
1550 | * fall through conditions. |
1551 | * |
1552 | **/ |
1553 | static void fm10k_set_num_queues(struct fm10k_intfc *interface) |
1554 | { |
1555 | /* Attempt to setup QoS and RSS first */ |
1556 | if (fm10k_set_qos_queues(interface)) |
1557 | return; |
1558 | |
1559 | /* If we don't have QoS, just fallback to only RSS. */ |
1560 | fm10k_set_rss_queues(interface); |
1561 | } |
1562 | |
1563 | /** |
1564 | * fm10k_reset_num_queues - Reset the number of queues to zero |
1565 | * @interface: board private structure |
1566 | * |
1567 | * This function should be called whenever we need to reset the number of |
1568 | * queues after an error condition. |
1569 | */ |
1570 | static void fm10k_reset_num_queues(struct fm10k_intfc *interface) |
1571 | { |
1572 | interface->num_tx_queues = 0; |
1573 | interface->num_rx_queues = 0; |
1574 | interface->num_q_vectors = 0; |
1575 | } |
1576 | |
1577 | /** |
1578 | * fm10k_alloc_q_vector - Allocate memory for a single interrupt vector |
1579 | * @interface: board private structure to initialize |
1580 | * @v_count: q_vectors allocated on interface, used for ring interleaving |
1581 | * @v_idx: index of vector in interface struct |
1582 | * @txr_count: total number of Tx rings to allocate |
1583 | * @txr_idx: index of first Tx ring to allocate |
1584 | * @rxr_count: total number of Rx rings to allocate |
1585 | * @rxr_idx: index of first Rx ring to allocate |
1586 | * |
1587 | * We allocate one q_vector. If allocation fails we return -ENOMEM. |
1588 | **/ |
1589 | static int fm10k_alloc_q_vector(struct fm10k_intfc *interface, |
1590 | unsigned int v_count, unsigned int v_idx, |
1591 | unsigned int txr_count, unsigned int txr_idx, |
1592 | unsigned int rxr_count, unsigned int rxr_idx) |
1593 | { |
1594 | struct fm10k_q_vector *q_vector; |
1595 | struct fm10k_ring *ring; |
1596 | int ring_count; |
1597 | |
1598 | ring_count = txr_count + rxr_count; |
1599 | |
1600 | /* allocate q_vector and rings */ |
1601 | q_vector = kzalloc(struct_size(q_vector, ring, ring_count), GFP_KERNEL); |
1602 | if (!q_vector) |
1603 | return -ENOMEM; |
1604 | |
1605 | /* initialize NAPI */ |
1606 | netif_napi_add(dev: interface->netdev, napi: &q_vector->napi, poll: fm10k_poll); |
1607 | |
1608 | /* tie q_vector and interface together */ |
1609 | interface->q_vector[v_idx] = q_vector; |
1610 | q_vector->interface = interface; |
1611 | q_vector->v_idx = v_idx; |
1612 | |
1613 | /* initialize pointer to rings */ |
1614 | ring = q_vector->ring; |
1615 | |
1616 | /* save Tx ring container info */ |
1617 | q_vector->tx.ring = ring; |
1618 | q_vector->tx.work_limit = FM10K_DEFAULT_TX_WORK; |
1619 | q_vector->tx.itr = interface->tx_itr; |
1620 | q_vector->tx.itr_scale = interface->hw.mac.itr_scale; |
1621 | q_vector->tx.count = txr_count; |
1622 | |
1623 | while (txr_count) { |
1624 | /* assign generic ring traits */ |
1625 | ring->dev = &interface->pdev->dev; |
1626 | ring->netdev = interface->netdev; |
1627 | |
1628 | /* configure backlink on ring */ |
1629 | ring->q_vector = q_vector; |
1630 | |
1631 | /* apply Tx specific ring traits */ |
1632 | ring->count = interface->tx_ring_count; |
1633 | ring->queue_index = txr_idx; |
1634 | |
1635 | /* assign ring to interface */ |
1636 | interface->tx_ring[txr_idx] = ring; |
1637 | |
1638 | /* update count and index */ |
1639 | txr_count--; |
1640 | txr_idx += v_count; |
1641 | |
1642 | /* push pointer to next ring */ |
1643 | ring++; |
1644 | } |
1645 | |
1646 | /* save Rx ring container info */ |
1647 | q_vector->rx.ring = ring; |
1648 | q_vector->rx.itr = interface->rx_itr; |
1649 | q_vector->rx.itr_scale = interface->hw.mac.itr_scale; |
1650 | q_vector->rx.count = rxr_count; |
1651 | |
1652 | while (rxr_count) { |
1653 | /* assign generic ring traits */ |
1654 | ring->dev = &interface->pdev->dev; |
1655 | ring->netdev = interface->netdev; |
1656 | rcu_assign_pointer(ring->l2_accel, interface->l2_accel); |
1657 | |
1658 | /* configure backlink on ring */ |
1659 | ring->q_vector = q_vector; |
1660 | |
1661 | /* apply Rx specific ring traits */ |
1662 | ring->count = interface->rx_ring_count; |
1663 | ring->queue_index = rxr_idx; |
1664 | |
1665 | /* assign ring to interface */ |
1666 | interface->rx_ring[rxr_idx] = ring; |
1667 | |
1668 | /* update count and index */ |
1669 | rxr_count--; |
1670 | rxr_idx += v_count; |
1671 | |
1672 | /* push pointer to next ring */ |
1673 | ring++; |
1674 | } |
1675 | |
1676 | fm10k_dbg_q_vector_init(q_vector); |
1677 | |
1678 | return 0; |
1679 | } |
1680 | |
1681 | /** |
1682 | * fm10k_free_q_vector - Free memory allocated for specific interrupt vector |
1683 | * @interface: board private structure to initialize |
1684 | * @v_idx: Index of vector to be freed |
1685 | * |
1686 | * This function frees the memory allocated to the q_vector. In addition if |
1687 | * NAPI is enabled it will delete any references to the NAPI struct prior |
1688 | * to freeing the q_vector. |
1689 | **/ |
1690 | static void fm10k_free_q_vector(struct fm10k_intfc *interface, int v_idx) |
1691 | { |
1692 | struct fm10k_q_vector *q_vector = interface->q_vector[v_idx]; |
1693 | struct fm10k_ring *ring; |
1694 | |
1695 | fm10k_dbg_q_vector_exit(q_vector); |
1696 | |
1697 | fm10k_for_each_ring(ring, q_vector->tx) |
1698 | interface->tx_ring[ring->queue_index] = NULL; |
1699 | |
1700 | fm10k_for_each_ring(ring, q_vector->rx) |
1701 | interface->rx_ring[ring->queue_index] = NULL; |
1702 | |
1703 | interface->q_vector[v_idx] = NULL; |
1704 | netif_napi_del(napi: &q_vector->napi); |
1705 | kfree_rcu(q_vector, rcu); |
1706 | } |
1707 | |
1708 | /** |
1709 | * fm10k_alloc_q_vectors - Allocate memory for interrupt vectors |
1710 | * @interface: board private structure to initialize |
1711 | * |
1712 | * We allocate one q_vector per queue interrupt. If allocation fails we |
1713 | * return -ENOMEM. |
1714 | **/ |
1715 | static int fm10k_alloc_q_vectors(struct fm10k_intfc *interface) |
1716 | { |
1717 | unsigned int q_vectors = interface->num_q_vectors; |
1718 | unsigned int rxr_remaining = interface->num_rx_queues; |
1719 | unsigned int txr_remaining = interface->num_tx_queues; |
1720 | unsigned int rxr_idx = 0, txr_idx = 0, v_idx = 0; |
1721 | int err; |
1722 | |
1723 | if (q_vectors >= (rxr_remaining + txr_remaining)) { |
1724 | for (; rxr_remaining; v_idx++) { |
1725 | err = fm10k_alloc_q_vector(interface, v_count: q_vectors, v_idx, |
1726 | txr_count: 0, txr_idx: 0, rxr_count: 1, rxr_idx); |
1727 | if (err) |
1728 | goto err_out; |
1729 | |
1730 | /* update counts and index */ |
1731 | rxr_remaining--; |
1732 | rxr_idx++; |
1733 | } |
1734 | } |
1735 | |
1736 | for (; v_idx < q_vectors; v_idx++) { |
1737 | int rqpv = DIV_ROUND_UP(rxr_remaining, q_vectors - v_idx); |
1738 | int tqpv = DIV_ROUND_UP(txr_remaining, q_vectors - v_idx); |
1739 | |
1740 | err = fm10k_alloc_q_vector(interface, v_count: q_vectors, v_idx, |
1741 | txr_count: tqpv, txr_idx, |
1742 | rxr_count: rqpv, rxr_idx); |
1743 | |
1744 | if (err) |
1745 | goto err_out; |
1746 | |
1747 | /* update counts and index */ |
1748 | rxr_remaining -= rqpv; |
1749 | txr_remaining -= tqpv; |
1750 | rxr_idx++; |
1751 | txr_idx++; |
1752 | } |
1753 | |
1754 | return 0; |
1755 | |
1756 | err_out: |
1757 | fm10k_reset_num_queues(interface); |
1758 | |
1759 | while (v_idx--) |
1760 | fm10k_free_q_vector(interface, v_idx); |
1761 | |
1762 | return -ENOMEM; |
1763 | } |
1764 | |
1765 | /** |
1766 | * fm10k_free_q_vectors - Free memory allocated for interrupt vectors |
1767 | * @interface: board private structure to initialize |
1768 | * |
1769 | * This function frees the memory allocated to the q_vectors. In addition if |
1770 | * NAPI is enabled it will delete any references to the NAPI struct prior |
1771 | * to freeing the q_vector. |
1772 | **/ |
1773 | static void fm10k_free_q_vectors(struct fm10k_intfc *interface) |
1774 | { |
1775 | int v_idx = interface->num_q_vectors; |
1776 | |
1777 | fm10k_reset_num_queues(interface); |
1778 | |
1779 | while (v_idx--) |
1780 | fm10k_free_q_vector(interface, v_idx); |
1781 | } |
1782 | |
1783 | /** |
1784 | * fm10k_reset_msix_capability - reset MSI-X capability |
1785 | * @interface: board private structure to initialize |
1786 | * |
1787 | * Reset the MSI-X capability back to its starting state |
1788 | **/ |
1789 | static void fm10k_reset_msix_capability(struct fm10k_intfc *interface) |
1790 | { |
1791 | pci_disable_msix(dev: interface->pdev); |
1792 | kfree(objp: interface->msix_entries); |
1793 | interface->msix_entries = NULL; |
1794 | } |
1795 | |
1796 | /** |
1797 | * fm10k_init_msix_capability - configure MSI-X capability |
1798 | * @interface: board private structure to initialize |
1799 | * |
1800 | * Attempt to configure the interrupts using the best available |
1801 | * capabilities of the hardware and the kernel. |
1802 | **/ |
1803 | static int fm10k_init_msix_capability(struct fm10k_intfc *interface) |
1804 | { |
1805 | struct fm10k_hw *hw = &interface->hw; |
1806 | int v_budget, vector; |
1807 | |
1808 | /* It's easy to be greedy for MSI-X vectors, but it really |
1809 | * doesn't do us much good if we have a lot more vectors |
1810 | * than CPU's. So let's be conservative and only ask for |
1811 | * (roughly) the same number of vectors as there are CPU's. |
1812 | * the default is to use pairs of vectors |
1813 | */ |
1814 | v_budget = max(interface->num_rx_queues, interface->num_tx_queues); |
1815 | v_budget = min_t(u16, v_budget, num_online_cpus()); |
1816 | |
1817 | /* account for vectors not related to queues */ |
1818 | v_budget += NON_Q_VECTORS; |
1819 | |
1820 | /* At the same time, hardware can only support a maximum of |
1821 | * hw.mac->max_msix_vectors vectors. With features |
1822 | * such as RSS and VMDq, we can easily surpass the number of Rx and Tx |
1823 | * descriptor queues supported by our device. Thus, we cap it off in |
1824 | * those rare cases where the cpu count also exceeds our vector limit. |
1825 | */ |
1826 | v_budget = min_t(int, v_budget, hw->mac.max_msix_vectors); |
1827 | |
1828 | /* A failure in MSI-X entry allocation is fatal. */ |
1829 | interface->msix_entries = kcalloc(n: v_budget, size: sizeof(struct msix_entry), |
1830 | GFP_KERNEL); |
1831 | if (!interface->msix_entries) |
1832 | return -ENOMEM; |
1833 | |
1834 | /* populate entry values */ |
1835 | for (vector = 0; vector < v_budget; vector++) |
1836 | interface->msix_entries[vector].entry = vector; |
1837 | |
1838 | /* Attempt to enable MSI-X with requested value */ |
1839 | v_budget = pci_enable_msix_range(dev: interface->pdev, |
1840 | entries: interface->msix_entries, |
1841 | MIN_MSIX_COUNT(hw), |
1842 | maxvec: v_budget); |
1843 | if (v_budget < 0) { |
1844 | kfree(objp: interface->msix_entries); |
1845 | interface->msix_entries = NULL; |
1846 | return v_budget; |
1847 | } |
1848 | |
1849 | /* record the number of queues available for q_vectors */ |
1850 | interface->num_q_vectors = v_budget - NON_Q_VECTORS; |
1851 | |
1852 | return 0; |
1853 | } |
1854 | |
1855 | /** |
1856 | * fm10k_cache_ring_qos - Descriptor ring to register mapping for QoS |
1857 | * @interface: Interface structure continaining rings and devices |
1858 | * |
1859 | * Cache the descriptor ring offsets for Qos |
1860 | **/ |
1861 | static bool fm10k_cache_ring_qos(struct fm10k_intfc *interface) |
1862 | { |
1863 | struct net_device *dev = interface->netdev; |
1864 | int pc, offset, , i; |
1865 | u16 pc_stride = interface->ring_feature[RING_F_QOS].mask + 1; |
1866 | u8 num_pcs = netdev_get_num_tc(dev); |
1867 | |
1868 | if (num_pcs <= 1) |
1869 | return false; |
1870 | |
1871 | rss_i = interface->ring_feature[RING_F_RSS].indices; |
1872 | |
1873 | for (pc = 0, offset = 0; pc < num_pcs; pc++, offset += rss_i) { |
1874 | int q_idx = pc; |
1875 | |
1876 | for (i = 0; i < rss_i; i++) { |
1877 | interface->tx_ring[offset + i]->reg_idx = q_idx; |
1878 | interface->tx_ring[offset + i]->qos_pc = pc; |
1879 | interface->rx_ring[offset + i]->reg_idx = q_idx; |
1880 | interface->rx_ring[offset + i]->qos_pc = pc; |
1881 | q_idx += pc_stride; |
1882 | } |
1883 | } |
1884 | |
1885 | return true; |
1886 | } |
1887 | |
1888 | /** |
1889 | * fm10k_cache_ring_rss - Descriptor ring to register mapping for RSS |
1890 | * @interface: Interface structure continaining rings and devices |
1891 | * |
1892 | * Cache the descriptor ring offsets for RSS |
1893 | **/ |
1894 | static void (struct fm10k_intfc *interface) |
1895 | { |
1896 | int i; |
1897 | |
1898 | for (i = 0; i < interface->num_rx_queues; i++) |
1899 | interface->rx_ring[i]->reg_idx = i; |
1900 | |
1901 | for (i = 0; i < interface->num_tx_queues; i++) |
1902 | interface->tx_ring[i]->reg_idx = i; |
1903 | } |
1904 | |
1905 | /** |
1906 | * fm10k_assign_rings - Map rings to network devices |
1907 | * @interface: Interface structure containing rings and devices |
1908 | * |
1909 | * This function is meant to go though and configure both the network |
1910 | * devices so that they contain rings, and configure the rings so that |
1911 | * they function with their network devices. |
1912 | **/ |
1913 | static void fm10k_assign_rings(struct fm10k_intfc *interface) |
1914 | { |
1915 | if (fm10k_cache_ring_qos(interface)) |
1916 | return; |
1917 | |
1918 | fm10k_cache_ring_rss(interface); |
1919 | } |
1920 | |
1921 | static void fm10k_init_reta(struct fm10k_intfc *interface) |
1922 | { |
1923 | u16 i, = interface->ring_feature[RING_F_RSS].indices; |
1924 | u32 reta; |
1925 | |
1926 | /* If the Rx flow indirection table has been configured manually, we |
1927 | * need to maintain it when possible. |
1928 | */ |
1929 | if (netif_is_rxfh_configured(dev: interface->netdev)) { |
1930 | for (i = FM10K_RETA_SIZE; i--;) { |
1931 | reta = interface->reta[i]; |
1932 | if ((((reta << 24) >> 24) < rss_i) && |
1933 | (((reta << 16) >> 24) < rss_i) && |
1934 | (((reta << 8) >> 24) < rss_i) && |
1935 | (((reta) >> 24) < rss_i)) |
1936 | continue; |
1937 | |
1938 | /* this should never happen */ |
1939 | dev_err(&interface->pdev->dev, |
1940 | "RSS indirection table assigned flows out of queue bounds. Reconfiguring.\n" ); |
1941 | goto repopulate_reta; |
1942 | } |
1943 | |
1944 | /* do nothing if all of the elements are in bounds */ |
1945 | return; |
1946 | } |
1947 | |
1948 | repopulate_reta: |
1949 | fm10k_write_reta(interface, NULL); |
1950 | } |
1951 | |
1952 | /** |
1953 | * fm10k_init_queueing_scheme - Determine proper queueing scheme |
1954 | * @interface: board private structure to initialize |
1955 | * |
1956 | * We determine which queueing scheme to use based on... |
1957 | * - Hardware queue count (num_*_queues) |
1958 | * - defined by miscellaneous hardware support/features (RSS, etc.) |
1959 | **/ |
1960 | int fm10k_init_queueing_scheme(struct fm10k_intfc *interface) |
1961 | { |
1962 | int err; |
1963 | |
1964 | /* Number of supported queues */ |
1965 | fm10k_set_num_queues(interface); |
1966 | |
1967 | /* Configure MSI-X capability */ |
1968 | err = fm10k_init_msix_capability(interface); |
1969 | if (err) { |
1970 | dev_err(&interface->pdev->dev, |
1971 | "Unable to initialize MSI-X capability\n" ); |
1972 | goto err_init_msix; |
1973 | } |
1974 | |
1975 | /* Allocate memory for queues */ |
1976 | err = fm10k_alloc_q_vectors(interface); |
1977 | if (err) { |
1978 | dev_err(&interface->pdev->dev, |
1979 | "Unable to allocate queue vectors\n" ); |
1980 | goto err_alloc_q_vectors; |
1981 | } |
1982 | |
1983 | /* Map rings to devices, and map devices to physical queues */ |
1984 | fm10k_assign_rings(interface); |
1985 | |
1986 | /* Initialize RSS redirection table */ |
1987 | fm10k_init_reta(interface); |
1988 | |
1989 | return 0; |
1990 | |
1991 | err_alloc_q_vectors: |
1992 | fm10k_reset_msix_capability(interface); |
1993 | err_init_msix: |
1994 | fm10k_reset_num_queues(interface); |
1995 | return err; |
1996 | } |
1997 | |
1998 | /** |
1999 | * fm10k_clear_queueing_scheme - Clear the current queueing scheme settings |
2000 | * @interface: board private structure to clear queueing scheme on |
2001 | * |
2002 | * We go through and clear queueing specific resources and reset the structure |
2003 | * to pre-load conditions |
2004 | **/ |
2005 | void fm10k_clear_queueing_scheme(struct fm10k_intfc *interface) |
2006 | { |
2007 | fm10k_free_q_vectors(interface); |
2008 | fm10k_reset_msix_capability(interface); |
2009 | } |
2010 | |