projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
ipv4: New multicast-all socket option
[linux-2.6-block.git] / drivers / net / mlx4 / en_rx.c
CommitLineData
c27a02cd
YP		 1/*
2 * Copyright (c) 2007 Mellanox Technologies. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 *
32 */
33
34#include <linux/mlx4/cq.h>
35#include <linux/mlx4/qp.h>
36#include <linux/skbuff.h>
37#include <linux/if_ether.h>
38#include <linux/if_vlan.h>
39#include <linux/vmalloc.h>
40
41#include "mlx4_en.h"
42
43static void *get_wqe(struct mlx4_en_rx_ring *ring, int n)
44{
45 int offset = n << ring->srq.wqe_shift;
46 return ring->buf + offset;
47}
48
49static void mlx4_en_srq_event(struct mlx4_srq *srq, enum mlx4_event type)
50{
51 return;
52}
53
54static int mlx4_en_get_frag_header(struct skb_frag_struct *frags, void **mac_hdr,
55 void **ip_hdr, void **tcpudp_hdr,
56 u64 *hdr_flags, void *priv)
57{
58 *mac_hdr = page_address(frags->page) + frags->page_offset;
59 *ip_hdr = *mac_hdr + ETH_HLEN;
60 *tcpudp_hdr = (struct tcphdr *)(*ip_hdr + sizeof(struct iphdr));
61 *hdr_flags = LRO_IPV4 | LRO_TCP;
62
63 return 0;
64}
65
66static int mlx4_en_alloc_frag(struct mlx4_en_priv *priv,
67 struct mlx4_en_rx_desc *rx_desc,
68 struct skb_frag_struct *skb_frags,
69 struct mlx4_en_rx_alloc *ring_alloc,
70 int i)
71{
72 struct mlx4_en_dev *mdev = priv->mdev;
73 struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
74 struct mlx4_en_rx_alloc *page_alloc = &ring_alloc[i];
75 struct page *page;
76 dma_addr_t dma;
77
78 if (page_alloc->offset == frag_info->last_offset) {
79 /* Allocate new page */
80 page = alloc_pages(GFP_ATOMIC | __GFP_COMP, MLX4_EN_ALLOC_ORDER);
81 if (!page)
82 return -ENOMEM;
83
84 skb_frags[i].page = page_alloc->page;
85 skb_frags[i].page_offset = page_alloc->offset;
86 page_alloc->page = page;
87 page_alloc->offset = frag_info->frag_align;
88 } else {
89 page = page_alloc->page;
90 get_page(page);
91
92 skb_frags[i].page = page;
93 skb_frags[i].page_offset = page_alloc->offset;
94 page_alloc->offset += frag_info->frag_stride;
95 }
96 dma = pci_map_single(mdev->pdev, page_address(skb_frags[i].page) +
97 skb_frags[i].page_offset, frag_info->frag_size,
98 PCI_DMA_FROMDEVICE);
99 rx_desc->data[i].addr = cpu_to_be64(dma);
100 return 0;
101}
102
103static int mlx4_en_init_allocator(struct mlx4_en_priv *priv,
104 struct mlx4_en_rx_ring *ring)
105{
106 struct mlx4_en_rx_alloc *page_alloc;
107 int i;
108
109 for (i = 0; i < priv->num_frags; i++) {
110 page_alloc = &ring->page_alloc[i];
111 page_alloc->page = alloc_pages(GFP_ATOMIC | __GFP_COMP,
112 MLX4_EN_ALLOC_ORDER);
113 if (!page_alloc->page)
114 goto out;
115
116 page_alloc->offset = priv->frag_info[i].frag_align;
117 mlx4_dbg(DRV, priv, "Initialized allocator:%d with page:%p\n",
118 i, page_alloc->page);
119 }
120 return 0;
121
122out:
123 while (i--) {
124 page_alloc = &ring->page_alloc[i];
125 put_page(page_alloc->page);
126 page_alloc->page = NULL;
127 }
128 return -ENOMEM;
129}
130
131static void mlx4_en_destroy_allocator(struct mlx4_en_priv *priv,
132 struct mlx4_en_rx_ring *ring)
133{
134 struct mlx4_en_rx_alloc *page_alloc;
135 int i;
136
137 for (i = 0; i < priv->num_frags; i++) {
138 page_alloc = &ring->page_alloc[i];
139 mlx4_dbg(DRV, priv, "Freeing allocator:%d count:%d\n",
140 i, page_count(page_alloc->page));
141
142 put_page(page_alloc->page);
143 page_alloc->page = NULL;
144 }
145}
146
147
148static void mlx4_en_init_rx_desc(struct mlx4_en_priv *priv,
149 struct mlx4_en_rx_ring *ring, int index)
150{
151 struct mlx4_en_rx_desc *rx_desc = ring->buf + ring->stride * index;
152 struct skb_frag_struct *skb_frags = ring->rx_info +
153 (index << priv->log_rx_info);
154 int possible_frags;
155 int i;
156
157 /* Pre-link descriptor */
158 rx_desc->next.next_wqe_index = cpu_to_be16((index + 1) & ring->size_mask);
159
160 /* Set size and memtype fields */
161 for (i = 0; i < priv->num_frags; i++) {
162 skb_frags[i].size = priv->frag_info[i].frag_size;
163 rx_desc->data[i].byte_count =
164 cpu_to_be32(priv->frag_info[i].frag_size);
165 rx_desc->data[i].lkey = cpu_to_be32(priv->mdev->mr.key);
166 }
167
168 /* If the number of used fragments does not fill up the ring stride,
169 * remaining (unused) fragments must be padded with null address/size
170 * and a special memory key */
171 possible_frags = (ring->stride - sizeof(struct mlx4_en_rx_desc)) / DS_SIZE;
172 for (i = priv->num_frags; i < possible_frags; i++) {
173 rx_desc->data[i].byte_count = 0;
174 rx_desc->data[i].lkey = cpu_to_be32(MLX4_EN_MEMTYPE_PAD);
175 rx_desc->data[i].addr = 0;
176 }
177}
178
179
180static int mlx4_en_prepare_rx_desc(struct mlx4_en_priv *priv,
181 struct mlx4_en_rx_ring *ring, int index)
182{
183 struct mlx4_en_rx_desc *rx_desc = ring->buf + (index * ring->stride);
184 struct skb_frag_struct *skb_frags = ring->rx_info +
185 (index << priv->log_rx_info);
186 int i;
187
188 for (i = 0; i < priv->num_frags; i++)
189 if (mlx4_en_alloc_frag(priv, rx_desc, skb_frags, ring->page_alloc, i))
190 goto err;
191
192 return 0;
193
194err:
195 while (i--)
196 put_page(skb_frags[i].page);
197 return -ENOMEM;
198}
199
200static inline void mlx4_en_update_rx_prod_db(struct mlx4_en_rx_ring *ring)
201{
202 *ring->wqres.db.db = cpu_to_be32(ring->prod & 0xffff);
203}
204
38aab07c
YP		 205static void mlx4_en_free_rx_desc(struct mlx4_en_priv *priv,
206 struct mlx4_en_rx_ring *ring,
207 int index)
208{
209 struct mlx4_en_dev *mdev = priv->mdev;
210 struct skb_frag_struct *skb_frags;
211 struct mlx4_en_rx_desc *rx_desc = ring->buf + (index << ring->log_stride);
212 dma_addr_t dma;
213 int nr;
214
215 skb_frags = ring->rx_info + (index << priv->log_rx_info);
216 for (nr = 0; nr < priv->num_frags; nr++) {
217 mlx4_dbg(DRV, priv, "Freeing fragment:%d\n", nr);
218 dma = be64_to_cpu(rx_desc->data[nr].addr);
219
220 mlx4_dbg(DRV, priv, "Unmaping buffer at dma:0x%llx\n", (u64) dma);
221 pci_unmap_single(mdev->pdev, dma, skb_frags[nr].size,
222 PCI_DMA_FROMDEVICE);
223 put_page(skb_frags[nr].page);
224 }
225}
226
c27a02cd
YP		 227static int mlx4_en_fill_rx_buffers(struct mlx4_en_priv *priv)
228{
229 struct mlx4_en_dev *mdev = priv->mdev;
230 struct mlx4_en_rx_ring *ring;
231 int ring_ind;
232 int buf_ind;
38aab07c 233 int new_size;
c27a02cd
YP		 234
235 for (buf_ind = 0; buf_ind < priv->prof->rx_ring_size; buf_ind++) {
236 for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
237 ring = &priv->rx_ring[ring_ind];
238
239 if (mlx4_en_prepare_rx_desc(priv, ring,
240 ring->actual_size)) {
241 if (ring->actual_size < MLX4_EN_MIN_RX_SIZE) {
242 mlx4_err(mdev, "Failed to allocate "
243 "enough rx buffers\n");
244 return -ENOMEM;
245 } else {
38aab07c
YP		 246 new_size = rounddown_pow_of_two(ring->actual_size);
247 mlx4_warn(mdev, "Only %d buffers allocated "
248 "reducing ring size to %d",
249 ring->actual_size, new_size);
250 goto reduce_rings;
c27a02cd
YP		 251 }
252 }
253 ring->actual_size++;
254 ring->prod++;
255 }
256 }
38aab07c
YP		 257 return 0;
258
259reduce_rings:
260 for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
261 ring = &priv->rx_ring[ring_ind];
262 while (ring->actual_size > new_size) {
263 ring->actual_size--;
264 ring->prod--;
265 mlx4_en_free_rx_desc(priv, ring, ring->actual_size);
266 }
267 ring->size_mask = ring->actual_size - 1;
268 }
269
c27a02cd
YP		 270 return 0;
271}
272
273static int mlx4_en_fill_rx_buf(struct net_device *dev,
274 struct mlx4_en_rx_ring *ring)
275{
276 struct mlx4_en_priv *priv = netdev_priv(dev);
277 int num = 0;
278 int err;
279
280 while ((u32) (ring->prod - ring->cons) < ring->actual_size) {
281 err = mlx4_en_prepare_rx_desc(priv, ring, ring->prod &
282 ring->size_mask);
283 if (err) {
284 if (netif_msg_rx_err(priv))
285 mlx4_warn(priv->mdev,
286 "Failed preparing rx descriptor\n");
287 priv->port_stats.rx_alloc_failed++;
288 break;
289 }
290 ++num;
291 ++ring->prod;
292 }
38aab07c 293 if ((u32) (ring->prod - ring->cons) == ring->actual_size)
c27a02cd
YP		 294 ring->full = 1;
295
296 return num;
297}
298
299static void mlx4_en_free_rx_buf(struct mlx4_en_priv *priv,
300 struct mlx4_en_rx_ring *ring)
301{
c27a02cd 302 int index;
c27a02cd
YP		 303
304 mlx4_dbg(DRV, priv, "Freeing Rx buf - cons:%d prod:%d\n",
305 ring->cons, ring->prod);
306
307 /* Unmap and free Rx buffers */
38aab07c 308 BUG_ON((u32) (ring->prod - ring->cons) > ring->actual_size);
c27a02cd
YP		 309 while (ring->cons != ring->prod) {
310 index = ring->cons & ring->size_mask;
c27a02cd 311 mlx4_dbg(DRV, priv, "Processing descriptor:%d\n", index);
38aab07c 312 mlx4_en_free_rx_desc(priv, ring, index);
c27a02cd
YP		 313 ++ring->cons;
314 }
315}
316
317
318void mlx4_en_rx_refill(struct work_struct *work)
319{
bf6aede7 320 struct delayed_work *delay = to_delayed_work(work);
c27a02cd
YP		 321 struct mlx4_en_priv *priv = container_of(delay, struct mlx4_en_priv,
322 refill_task);
323 struct mlx4_en_dev *mdev = priv->mdev;
324 struct net_device *dev = priv->dev;
325 struct mlx4_en_rx_ring *ring;
326 int need_refill = 0;
327 int i;
328
329 mutex_lock(&mdev->state_lock);
330 if (!mdev->device_up || !priv->port_up)
331 goto out;
332
333 /* We only get here if there are no receive buffers, so we can't race
334 * with Rx interrupts while filling buffers */
335 for (i = 0; i < priv->rx_ring_num; i++) {
336 ring = &priv->rx_ring[i];
337 if (ring->need_refill) {
338 if (mlx4_en_fill_rx_buf(dev, ring)) {
339 ring->need_refill = 0;
340 mlx4_en_update_rx_prod_db(ring);
341 } else
342 need_refill = 1;
343 }
344 }
345 if (need_refill)
346 queue_delayed_work(mdev->workqueue, &priv->refill_task, HZ);
347
348out:
349 mutex_unlock(&mdev->state_lock);
350}
351
352
353int mlx4_en_create_rx_ring(struct mlx4_en_priv *priv,
354 struct mlx4_en_rx_ring *ring, u32 size, u16 stride)
355{
356 struct mlx4_en_dev *mdev = priv->mdev;
357 int err;
358 int tmp;
359
360 /* Sanity check SRQ size before proceeding */
361 if (size >= mdev->dev->caps.max_srq_wqes)
362 return -EINVAL;
363
364 ring->prod = 0;
365 ring->cons = 0;
366 ring->size = size;
367 ring->size_mask = size - 1;
368 ring->stride = stride;
369 ring->log_stride = ffs(ring->stride) - 1;
370 ring->buf_size = ring->size * ring->stride;
371
372 tmp = size * roundup_pow_of_two(MLX4_EN_MAX_RX_FRAGS *
373 sizeof(struct skb_frag_struct));
374 ring->rx_info = vmalloc(tmp);
375 if (!ring->rx_info) {
376 mlx4_err(mdev, "Failed allocating rx_info ring\n");
377 return -ENOMEM;
378 }
379 mlx4_dbg(DRV, priv, "Allocated rx_info ring at addr:%p size:%d\n",
380 ring->rx_info, tmp);
381
382 err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres,
383 ring->buf_size, 2 * PAGE_SIZE);
384 if (err)
385 goto err_ring;
386
387 err = mlx4_en_map_buffer(&ring->wqres.buf);
388 if (err) {
389 mlx4_err(mdev, "Failed to map RX buffer\n");
390 goto err_hwq;
391 }
392 ring->buf = ring->wqres.buf.direct.buf;
393
394 /* Configure lro mngr */
395 memset(&ring->lro, 0, sizeof(struct net_lro_mgr));
396 ring->lro.dev = priv->dev;
397 ring->lro.features = LRO_F_NAPI;
398 ring->lro.frag_align_pad = NET_IP_ALIGN;
399 ring->lro.ip_summed = CHECKSUM_UNNECESSARY;
400 ring->lro.ip_summed_aggr = CHECKSUM_UNNECESSARY;
401 ring->lro.max_desc = mdev->profile.num_lro;
402 ring->lro.max_aggr = MAX_SKB_FRAGS;
403 ring->lro.lro_arr = kzalloc(mdev->profile.num_lro *
404 sizeof(struct net_lro_desc),
405 GFP_KERNEL);
406 if (!ring->lro.lro_arr) {
407 mlx4_err(mdev, "Failed to allocate lro array\n");
408 goto err_map;
409 }
410 ring->lro.get_frag_header = mlx4_en_get_frag_header;
411
412 return 0;
413
414err_map:
415 mlx4_en_unmap_buffer(&ring->wqres.buf);
416err_hwq:
417 mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
418err_ring:
419 vfree(ring->rx_info);
420 ring->rx_info = NULL;
421 return err;
422}
423
424int mlx4_en_activate_rx_rings(struct mlx4_en_priv *priv)
425{
426 struct mlx4_en_dev *mdev = priv->mdev;
427 struct mlx4_wqe_srq_next_seg *next;
428 struct mlx4_en_rx_ring *ring;
429 int i;
430 int ring_ind;
431 int err;
432 int stride = roundup_pow_of_two(sizeof(struct mlx4_en_rx_desc) +
433 DS_SIZE * priv->num_frags);
434 int max_gs = (stride - sizeof(struct mlx4_wqe_srq_next_seg)) / DS_SIZE;
435
436 for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
437 ring = &priv->rx_ring[ring_ind];
438
439 ring->prod = 0;
440 ring->cons = 0;
441 ring->actual_size = 0;
442 ring->cqn = priv->rx_cq[ring_ind].mcq.cqn;
443
444 ring->stride = stride;
445 ring->log_stride = ffs(ring->stride) - 1;
446 ring->buf_size = ring->size * ring->stride;
447
448 memset(ring->buf, 0, ring->buf_size);
449 mlx4_en_update_rx_prod_db(ring);
450
451 /* Initailize all descriptors */
452 for (i = 0; i < ring->size; i++)
453 mlx4_en_init_rx_desc(priv, ring, i);
454
455 /* Initialize page allocators */
456 err = mlx4_en_init_allocator(priv, ring);
457 if (err) {
9a4f92a6
YP		 458 mlx4_err(mdev, "Failed initializing ring allocator\n");
459 ring_ind--;
460 goto err_allocator;
c27a02cd
YP		 461 }
462
463 /* Fill Rx buffers */
464 ring->full = 0;
465 }
b58515be
IM		 466 err = mlx4_en_fill_rx_buffers(priv);
467 if (err)
c27a02cd
YP		 468 goto err_buffers;
469
470 for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
471 ring = &priv->rx_ring[ring_ind];
472
473 mlx4_en_update_rx_prod_db(ring);
474
475 /* Configure SRQ representing the ring */
38aab07c 476 ring->srq.max = ring->actual_size;
c27a02cd
YP		 477 ring->srq.max_gs = max_gs;
478 ring->srq.wqe_shift = ilog2(ring->stride);
479
480 for (i = 0; i < ring->srq.max; ++i) {
481 next = get_wqe(ring, i);
482 next->next_wqe_index =
483 cpu_to_be16((i + 1) & (ring->srq.max - 1));
484 }
485
486 err = mlx4_srq_alloc(mdev->dev, mdev->priv_pdn, &ring->wqres.mtt,
487 ring->wqres.db.dma, &ring->srq);
488 if (err){
489 mlx4_err(mdev, "Failed to allocate srq\n");
9a4f92a6 490 ring_ind--;
c27a02cd
YP		 491 goto err_srq;
492 }
493 ring->srq.event = mlx4_en_srq_event;
494 }
495
496 return 0;
497
498err_srq:
499 while (ring_ind >= 0) {
500 ring = &priv->rx_ring[ring_ind];
501 mlx4_srq_free(mdev->dev, &ring->srq);
502 ring_ind--;
503 }
504
505err_buffers:
506 for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++)
507 mlx4_en_free_rx_buf(priv, &priv->rx_ring[ring_ind]);
508
509 ring_ind = priv->rx_ring_num - 1;
510err_allocator:
511 while (ring_ind >= 0) {
512 mlx4_en_destroy_allocator(priv, &priv->rx_ring[ring_ind]);
513 ring_ind--;
514 }
515 return err;
516}
517
518void mlx4_en_destroy_rx_ring(struct mlx4_en_priv *priv,
519 struct mlx4_en_rx_ring *ring)
520{
521 struct mlx4_en_dev *mdev = priv->mdev;
522
523 kfree(ring->lro.lro_arr);
524 mlx4_en_unmap_buffer(&ring->wqres.buf);
525 mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
526 vfree(ring->rx_info);
527 ring->rx_info = NULL;
528}
529
530void mlx4_en_deactivate_rx_ring(struct mlx4_en_priv *priv,
531 struct mlx4_en_rx_ring *ring)
532{
533 struct mlx4_en_dev *mdev = priv->mdev;
534
535 mlx4_srq_free(mdev->dev, &ring->srq);
536 mlx4_en_free_rx_buf(priv, ring);
537 mlx4_en_destroy_allocator(priv, ring);
538}
539
540
541/* Unmap a completed descriptor and free unused pages */
542static int mlx4_en_complete_rx_desc(struct mlx4_en_priv *priv,
543 struct mlx4_en_rx_desc *rx_desc,
544 struct skb_frag_struct *skb_frags,
545 struct skb_frag_struct *skb_frags_rx,
546 struct mlx4_en_rx_alloc *page_alloc,
547 int length)
548{
549 struct mlx4_en_dev *mdev = priv->mdev;
550 struct mlx4_en_frag_info *frag_info;
551 int nr;
552 dma_addr_t dma;
553
554 /* Collect used fragments while replacing them in the HW descirptors */
555 for (nr = 0; nr < priv->num_frags; nr++) {
556 frag_info = &priv->frag_info[nr];
557 if (length <= frag_info->frag_prefix_size)
558 break;
559
560 /* Save page reference in skb */
561 skb_frags_rx[nr].page = skb_frags[nr].page;
562 skb_frags_rx[nr].size = skb_frags[nr].size;
563 skb_frags_rx[nr].page_offset = skb_frags[nr].page_offset;
564 dma = be64_to_cpu(rx_desc->data[nr].addr);
565
566 /* Allocate a replacement page */
567 if (mlx4_en_alloc_frag(priv, rx_desc, skb_frags, page_alloc, nr))
568 goto fail;
569
570 /* Unmap buffer */
571 pci_unmap_single(mdev->pdev, dma, skb_frags[nr].size,
572 PCI_DMA_FROMDEVICE);
573 }
574 /* Adjust size of last fragment to match actual length */
575 skb_frags_rx[nr - 1].size = length -
576 priv->frag_info[nr - 1].frag_prefix_size;
577 return nr;
578
579fail:
580 /* Drop all accumulated fragments (which have already been replaced in
581 * the descriptor) of this packet; remaining fragments are reused... */
582 while (nr > 0) {
583 nr--;
584 put_page(skb_frags_rx[nr].page);
585 }
586 return 0;
587}
588
589
590static struct sk_buff *mlx4_en_rx_skb(struct mlx4_en_priv *priv,
591 struct mlx4_en_rx_desc *rx_desc,
592 struct skb_frag_struct *skb_frags,
593 struct mlx4_en_rx_alloc *page_alloc,
594 unsigned int length)
595{
596 struct mlx4_en_dev *mdev = priv->mdev;
597 struct sk_buff *skb;
598 void *va;
599 int used_frags;
600 dma_addr_t dma;
601
602 skb = dev_alloc_skb(SMALL_PACKET_SIZE + NET_IP_ALIGN);
603 if (!skb) {
604 mlx4_dbg(RX_ERR, priv, "Failed allocating skb\n");
605 return NULL;
606 }
607 skb->dev = priv->dev;
608 skb_reserve(skb, NET_IP_ALIGN);
609 skb->len = length;
610 skb->truesize = length + sizeof(struct sk_buff);
611
612 /* Get pointer to first fragment so we could copy the headers into the
613 * (linear part of the) skb */
614 va = page_address(skb_frags[0].page) + skb_frags[0].page_offset;
615
616 if (length <= SMALL_PACKET_SIZE) {
617 /* We are copying all relevant data to the skb - temporarily
618 * synch buffers for the copy */
619 dma = be64_to_cpu(rx_desc->data[0].addr);
620 dma_sync_single_range_for_cpu(&mdev->pdev->dev, dma, 0,
621 length, DMA_FROM_DEVICE);
622 skb_copy_to_linear_data(skb, va, length);
623 dma_sync_single_range_for_device(&mdev->pdev->dev, dma, 0,
624 length, DMA_FROM_DEVICE);
625 skb->tail += length;
626 } else {
627
628 /* Move relevant fragments to skb */
629 used_frags = mlx4_en_complete_rx_desc(priv, rx_desc, skb_frags,
630 skb_shinfo(skb)->frags,
631 page_alloc, length);
785a0982
YP		 632 if (unlikely(!used_frags)) {
633 kfree_skb(skb);
634 return NULL;
635 }
c27a02cd
YP		 636 skb_shinfo(skb)->nr_frags = used_frags;
637
638 /* Copy headers into the skb linear buffer */
639 memcpy(skb->data, va, HEADER_COPY_SIZE);
640 skb->tail += HEADER_COPY_SIZE;
641
642 /* Skip headers in first fragment */
643 skb_shinfo(skb)->frags[0].page_offset += HEADER_COPY_SIZE;
644
645 /* Adjust size of first fragment */
646 skb_shinfo(skb)->frags[0].size -= HEADER_COPY_SIZE;
647 skb->data_len = length - HEADER_COPY_SIZE;
648 }
649 return skb;
650}
651
652static void mlx4_en_copy_desc(struct mlx4_en_priv *priv,
653 struct mlx4_en_rx_ring *ring,
654 int from, int to, int num)
655{
656 struct skb_frag_struct *skb_frags_from;
657 struct skb_frag_struct *skb_frags_to;
658 struct mlx4_en_rx_desc *rx_desc_from;
659 struct mlx4_en_rx_desc *rx_desc_to;
660 int from_index, to_index;
661 int nr, i;
662
663 for (i = 0; i < num; i++) {
664 from_index = (from + i) & ring->size_mask;
665 to_index = (to + i) & ring->size_mask;
666 skb_frags_from = ring->rx_info + (from_index << priv->log_rx_info);
667 skb_frags_to = ring->rx_info + (to_index << priv->log_rx_info);
668 rx_desc_from = ring->buf + (from_index << ring->log_stride);
669 rx_desc_to = ring->buf + (to_index << ring->log_stride);
670
671 for (nr = 0; nr < priv->num_frags; nr++) {
672 skb_frags_to[nr].page = skb_frags_from[nr].page;
673 skb_frags_to[nr].page_offset = skb_frags_from[nr].page_offset;
674 rx_desc_to->data[nr].addr = rx_desc_from->data[nr].addr;
675 }
676 }
677}
678
679
680int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int budget)
681{
682 struct mlx4_en_priv *priv = netdev_priv(dev);
683 struct mlx4_en_dev *mdev = priv->mdev;
684 struct mlx4_cqe *cqe;
685 struct mlx4_en_rx_ring *ring = &priv->rx_ring[cq->ring];
686 struct skb_frag_struct *skb_frags;
687 struct skb_frag_struct lro_frags[MLX4_EN_MAX_RX_FRAGS];
688 struct mlx4_en_rx_desc *rx_desc;
689 struct sk_buff *skb;
690 int index;
691 int nr;
692 unsigned int length;
693 int polled = 0;
694 int ip_summed;
695
696 if (!priv->port_up)
697 return 0;
698
699 /* We assume a 1:1 mapping between CQEs and Rx descriptors, so Rx
700 * descriptor offset can be deduced from the CQE index instead of
701 * reading 'cqe->index' */
702 index = cq->mcq.cons_index & ring->size_mask;
703 cqe = &cq->buf[index];
704
705 /* Process all completed CQEs */
706 while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
707 cq->mcq.cons_index & cq->size)) {
708
709 skb_frags = ring->rx_info + (index << priv->log_rx_info);
710 rx_desc = ring->buf + (index << ring->log_stride);
711
712 /*
713 * make sure we read the CQE after we read the ownership bit
714 */
715 rmb();
716
717 /* Drop packet on bad receive or bad checksum */
718 if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
719 MLX4_CQE_OPCODE_ERROR)) {
720 mlx4_err(mdev, "CQE completed in error - vendor "
721 "syndrom:%d syndrom:%d\n",
722 ((struct mlx4_err_cqe *) cqe)->vendor_err_syndrome,
723 ((struct mlx4_err_cqe *) cqe)->syndrome);
724 goto next;
725 }
726 if (unlikely(cqe->badfcs_enc & MLX4_CQE_BAD_FCS)) {
727 mlx4_dbg(RX_ERR, priv, "Accepted frame with bad FCS\n");
728 goto next;
729 }
730
731 /*
732 * Packet is OK - process it.
733 */
734 length = be32_to_cpu(cqe->byte_cnt);
735 ring->bytes += length;
736 ring->packets++;
737
738 if (likely(priv->rx_csum)) {
739 if ((cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
740 (cqe->checksum == cpu_to_be16(0xffff))) {
741 priv->port_stats.rx_chksum_good++;
742 /* This packet is eligible for LRO if it is:
743 * - DIX Ethernet (type interpretation)
744 * - TCP/IP (v4)
745 * - without IP options
746 * - not an IP fragment */
747 if (mlx4_en_can_lro(cqe->status) &&
748 dev->features & NETIF_F_LRO) {
749
750 nr = mlx4_en_complete_rx_desc(
751 priv, rx_desc,
752 skb_frags, lro_frags,
753 ring->page_alloc, length);
754 if (!nr)
755 goto next;
756
757 if (priv->vlgrp && (cqe->vlan_my_qpn &
758 cpu_to_be32(MLX4_CQE_VLAN_PRESENT_MASK))) {
759 lro_vlan_hwaccel_receive_frags(
760 &ring->lro, lro_frags,
761 length, length,
762 priv->vlgrp,
763 be16_to_cpu(cqe->sl_vid),
764 NULL, 0);
765 } else
766 lro_receive_frags(&ring->lro,
767 lro_frags,
768 length,
769 length,
770 NULL, 0);
771
772 goto next;
773 }
774
775 /* LRO not possible, complete processing here */
776 ip_summed = CHECKSUM_UNNECESSARY;
777 INC_PERF_COUNTER(priv->pstats.lro_misses);
778 } else {
779 ip_summed = CHECKSUM_NONE;
780 priv->port_stats.rx_chksum_none++;
781 }
782 } else {
783 ip_summed = CHECKSUM_NONE;
784 priv->port_stats.rx_chksum_none++;
785 }
786
787 skb = mlx4_en_rx_skb(priv, rx_desc, skb_frags,
788 ring->page_alloc, length);
789 if (!skb) {
790 priv->stats.rx_dropped++;
791 goto next;
792 }
793
794 skb->ip_summed = ip_summed;
795 skb->protocol = eth_type_trans(skb, dev);
0c8dfc83 796 skb_record_rx_queue(skb, cq->ring);
c27a02cd
YP		 797
798 /* Push it up the stack */
799 if (priv->vlgrp && (be32_to_cpu(cqe->vlan_my_qpn) &
800 MLX4_CQE_VLAN_PRESENT_MASK)) {
801 vlan_hwaccel_receive_skb(skb, priv->vlgrp,
802 be16_to_cpu(cqe->sl_vid));
803 } else
804 netif_receive_skb(skb);
805
c27a02cd
YP		 806next:
807 ++cq->mcq.cons_index;
808 index = (cq->mcq.cons_index) & ring->size_mask;
809 cqe = &cq->buf[index];
810 if (++polled == budget) {
811 /* We are here because we reached the NAPI budget -
812 * flush only pending LRO sessions */
813 lro_flush_all(&ring->lro);
814 goto out;
815 }
816 }
817
818 /* If CQ is empty flush all LRO sessions unconditionally */
819 lro_flush_all(&ring->lro);
820
821out:
822 AVG_PERF_COUNTER(priv->pstats.rx_coal_avg, polled);
823 mlx4_cq_set_ci(&cq->mcq);
824 wmb(); /* ensure HW sees CQ consumer before we post new buffers */
825 ring->cons = cq->mcq.cons_index;
826 ring->prod += polled; /* Polled descriptors were realocated in place */
827 if (unlikely(!ring->full)) {
828 mlx4_en_copy_desc(priv, ring, ring->cons - polled,
829 ring->prod - polled, polled);
830 mlx4_en_fill_rx_buf(dev, ring);
831 }
832 mlx4_en_update_rx_prod_db(ring);
833 return polled;
834}
835
836
837void mlx4_en_rx_irq(struct mlx4_cq *mcq)
838{
839 struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
840 struct mlx4_en_priv *priv = netdev_priv(cq->dev);
841
842 if (priv->port_up)
288379f0 843 napi_schedule(&cq->napi);
c27a02cd
YP		 844 else
845 mlx4_en_arm_cq(priv, cq);
846}
847
848/* Rx CQ polling - called by NAPI */
849int mlx4_en_poll_rx_cq(struct napi_struct *napi, int budget)
850{
851 struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
852 struct net_device *dev = cq->dev;
853 struct mlx4_en_priv *priv = netdev_priv(dev);
854 int done;
855
856 done = mlx4_en_process_rx_cq(dev, cq, budget);
857
858 /* If we used up all the quota - we're probably not done yet... */
859 if (done == budget)
860 INC_PERF_COUNTER(priv->pstats.napi_quota);
861 else {
862 /* Done for now */
288379f0 863 napi_complete(napi);
c27a02cd
YP		 864 mlx4_en_arm_cq(priv, cq);
865 }
866 return done;
867}
868
869
870/* Calculate the last offset position that accomodates a full fragment
871 * (assuming fagment size = stride-align) */
872static int mlx4_en_last_alloc_offset(struct mlx4_en_priv *priv, u16 stride, u16 align)
873{
874 u16 res = MLX4_EN_ALLOC_SIZE % stride;
875 u16 offset = MLX4_EN_ALLOC_SIZE - stride - res + align;
876
877 mlx4_dbg(DRV, priv, "Calculated last offset for stride:%d align:%d "
878 "res:%d offset:%d\n", stride, align, res, offset);
879 return offset;
880}
881
882
883static int frag_sizes[] = {
884 FRAG_SZ0,
885 FRAG_SZ1,
886 FRAG_SZ2,
887 FRAG_SZ3
888};
889
890void mlx4_en_calc_rx_buf(struct net_device *dev)
891{
892 struct mlx4_en_priv *priv = netdev_priv(dev);
893 int eff_mtu = dev->mtu + ETH_HLEN + VLAN_HLEN + ETH_LLC_SNAP_SIZE;
894 int buf_size = 0;
895 int i = 0;
896
897 while (buf_size < eff_mtu) {
898 priv->frag_info[i].frag_size =
899 (eff_mtu > buf_size + frag_sizes[i]) ?
900 frag_sizes[i] : eff_mtu - buf_size;
901 priv->frag_info[i].frag_prefix_size = buf_size;
902 if (!i) {
903 priv->frag_info[i].frag_align = NET_IP_ALIGN;
904 priv->frag_info[i].frag_stride =
905 ALIGN(frag_sizes[i] + NET_IP_ALIGN, SMP_CACHE_BYTES);
906 } else {
907 priv->frag_info[i].frag_align = 0;
908 priv->frag_info[i].frag_stride =
909 ALIGN(frag_sizes[i], SMP_CACHE_BYTES);
910 }
911 priv->frag_info[i].last_offset = mlx4_en_last_alloc_offset(
912 priv, priv->frag_info[i].frag_stride,
913 priv->frag_info[i].frag_align);
914 buf_size += priv->frag_info[i].frag_size;
915 i++;
916 }
917
918 priv->num_frags = i;
919 priv->rx_skb_size = eff_mtu;
920 priv->log_rx_info = ROUNDUP_LOG2(i * sizeof(struct skb_frag_struct));
921
922 mlx4_dbg(DRV, priv, "Rx buffer scatter-list (effective-mtu:%d "
923 "num_frags:%d):\n", eff_mtu, priv->num_frags);
924 for (i = 0; i < priv->num_frags; i++) {
925 mlx4_dbg(DRV, priv, " frag:%d - size:%d prefix:%d align:%d "
926 "stride:%d last_offset:%d\n", i,
927 priv->frag_info[i].frag_size,
928 priv->frag_info[i].frag_prefix_size,
929 priv->frag_info[i].frag_align,
930 priv->frag_info[i].frag_stride,
931 priv->frag_info[i].last_offset);
932 }
933}
934
935/* RSS related functions */
936
937/* Calculate rss size and map each entry in rss table to rx ring */
938void mlx4_en_set_default_rss_map(struct mlx4_en_priv *priv,
939 struct mlx4_en_rss_map *rss_map,
940 int num_entries, int num_rings)
941{
942 int i;
943
944 rss_map->size = roundup_pow_of_two(num_entries);
945 mlx4_dbg(DRV, priv, "Setting default RSS map of %d entires\n",
946 rss_map->size);
947
948 for (i = 0; i < rss_map->size; i++) {
949 rss_map->map[i] = i % num_rings;
950 mlx4_dbg(DRV, priv, "Entry %d ---> ring %d\n", i, rss_map->map[i]);
951 }
952}
953
c27a02cd
YP		 954static int mlx4_en_config_rss_qp(struct mlx4_en_priv *priv,
955 int qpn, int srqn, int cqn,
956 enum mlx4_qp_state *state,
957 struct mlx4_qp *qp)
958{
959 struct mlx4_en_dev *mdev = priv->mdev;
960 struct mlx4_qp_context *context;
961 int err = 0;
962
963 context = kmalloc(sizeof *context , GFP_KERNEL);
964 if (!context) {
965 mlx4_err(mdev, "Failed to allocate qp context\n");
966 return -ENOMEM;
967 }
968
969 err = mlx4_qp_alloc(mdev->dev, qpn, qp);
970 if (err) {
971 mlx4_err(mdev, "Failed to allocate qp #%d\n", qpn);
972 goto out;
c27a02cd
YP		 973 }
974 qp->event = mlx4_en_sqp_event;
975
976 memset(context, 0, sizeof *context);
977 mlx4_en_fill_qp_context(priv, 0, 0, 0, 0, qpn, cqn, srqn, context);
978
979 err = mlx4_qp_to_ready(mdev->dev, &priv->res.mtt, context, qp, state);
980 if (err) {
981 mlx4_qp_remove(mdev->dev, qp);
982 mlx4_qp_free(mdev->dev, qp);
983 }
984out:
985 kfree(context);
986 return err;
987}
988
989/* Allocate rx qp's and configure them according to rss map */
990int mlx4_en_config_rss_steer(struct mlx4_en_priv *priv)
991{
992 struct mlx4_en_dev *mdev = priv->mdev;
993 struct mlx4_en_rss_map *rss_map = &priv->rss_map;
994 struct mlx4_qp_context context;
995 struct mlx4_en_rss_context *rss_context;
996 void *ptr;
997 int rss_xor = mdev->profile.rss_xor;
998 u8 rss_mask = mdev->profile.rss_mask;
999 int i, srqn, qpn, cqn;
1000 int err = 0;
1001 int good_qps = 0;
1002
1003 mlx4_dbg(DRV, priv, "Configuring rss steering for port %u\n", priv->port);
1004 err = mlx4_qp_reserve_range(mdev->dev, rss_map->size,
1005 rss_map->size, &rss_map->base_qpn);
1006 if (err) {
1007 mlx4_err(mdev, "Failed reserving %d qps for port %u\n",
1008 rss_map->size, priv->port);
1009 return err;
1010 }
1011
1012 for (i = 0; i < rss_map->size; i++) {
1013 cqn = priv->rx_ring[rss_map->map[i]].cqn;
1014 srqn = priv->rx_ring[rss_map->map[i]].srq.srqn;
1015 qpn = rss_map->base_qpn + i;
1016 err = mlx4_en_config_rss_qp(priv, qpn, srqn, cqn,
1017 &rss_map->state[i],
1018 &rss_map->qps[i]);
1019 if (err)
1020 goto rss_err;
1021
1022 ++good_qps;
1023 }
1024
1025 /* Configure RSS indirection qp */
1026 err = mlx4_qp_reserve_range(mdev->dev, 1, 1, &priv->base_qpn);
1027 if (err) {
1028 mlx4_err(mdev, "Failed to reserve range for RSS "
1029 "indirection qp\n");
1030 goto rss_err;
1031 }
1032 err = mlx4_qp_alloc(mdev->dev, priv->base_qpn, &rss_map->indir_qp);
1033 if (err) {
1034 mlx4_err(mdev, "Failed to allocate RSS indirection QP\n");
1035 goto reserve_err;
1036 }
1037 rss_map->indir_qp.event = mlx4_en_sqp_event;
1038 mlx4_en_fill_qp_context(priv, 0, 0, 0, 1, priv->base_qpn,
1039 priv->rx_ring[0].cqn, 0, &context);
1040
1041 ptr = ((void *) &context) + 0x3c;
1042 rss_context = (struct mlx4_en_rss_context *) ptr;
1043 rss_context->base_qpn = cpu_to_be32(ilog2(rss_map->size) << 24 |
1044 (rss_map->base_qpn));
1045 rss_context->default_qpn = cpu_to_be32(rss_map->base_qpn);
1046 rss_context->hash_fn = rss_xor & 0x3;
1047 rss_context->flags = rss_mask << 2;
1048
1049 err = mlx4_qp_to_ready(mdev->dev, &priv->res.mtt, &context,
1050 &rss_map->indir_qp, &rss_map->indir_state);
1051 if (err)
1052 goto indir_err;
1053
1054 return 0;
1055
1056indir_err:
1057 mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1058 MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
1059 mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
1060 mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
1061reserve_err:
1062 mlx4_qp_release_range(mdev->dev, priv->base_qpn, 1);
1063rss_err:
1064 for (i = 0; i < good_qps; i++) {
1065 mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1066 MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1067 mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1068 mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1069 }
1070 mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, rss_map->size);
1071 return err;
1072}
1073
1074void mlx4_en_release_rss_steer(struct mlx4_en_priv *priv)
1075{
1076 struct mlx4_en_dev *mdev = priv->mdev;
1077 struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1078 int i;
1079
1080 mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1081 MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
1082 mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
1083 mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
1084 mlx4_qp_release_range(mdev->dev, priv->base_qpn, 1);
1085
1086 for (i = 0; i < rss_map->size; i++) {
1087 mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1088 MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1089 mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1090 mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1091 }
1092 mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, rss_map->size);
1093}
1094
1095
1096
1097
1098
Linux 6.x block layer and io_uring tree(s)