03b781a7a182f721fab29770fcaa9554acd6d0cc
[linux-2.6.git] / drivers / net / mlx4 / en_rx.c
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
43
44 static int mlx4_en_get_frag_header(struct skb_frag_struct *frags, void **mac_hdr,
45                                    void **ip_hdr, void **tcpudp_hdr,
46                                    u64 *hdr_flags, void *priv)
47 {
48         *mac_hdr = page_address(frags->page) + frags->page_offset;
49         *ip_hdr = *mac_hdr + ETH_HLEN;
50         *tcpudp_hdr = (struct tcphdr *)(*ip_hdr + sizeof(struct iphdr));
51         *hdr_flags = LRO_IPV4 | LRO_TCP;
52
53         return 0;
54 }
55
56 static int mlx4_en_alloc_frag(struct mlx4_en_priv *priv,
57                               struct mlx4_en_rx_desc *rx_desc,
58                               struct skb_frag_struct *skb_frags,
59                               struct mlx4_en_rx_alloc *ring_alloc,
60                               int i)
61 {
62         struct mlx4_en_dev *mdev = priv->mdev;
63         struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
64         struct mlx4_en_rx_alloc *page_alloc = &ring_alloc[i];
65         struct page *page;
66         dma_addr_t dma;
67
68         if (page_alloc->offset == frag_info->last_offset) {
69                 /* Allocate new page */
70                 page = alloc_pages(GFP_ATOMIC | __GFP_COMP, MLX4_EN_ALLOC_ORDER);
71                 if (!page)
72                         return -ENOMEM;
73
74                 skb_frags[i].page = page_alloc->page;
75                 skb_frags[i].page_offset = page_alloc->offset;
76                 page_alloc->page = page;
77                 page_alloc->offset = frag_info->frag_align;
78         } else {
79                 page = page_alloc->page;
80                 get_page(page);
81
82                 skb_frags[i].page = page;
83                 skb_frags[i].page_offset = page_alloc->offset;
84                 page_alloc->offset += frag_info->frag_stride;
85         }
86         dma = pci_map_single(mdev->pdev, page_address(skb_frags[i].page) +
87                              skb_frags[i].page_offset, frag_info->frag_size,
88                              PCI_DMA_FROMDEVICE);
89         rx_desc->data[i].addr = cpu_to_be64(dma);
90         return 0;
91 }
92
93 static int mlx4_en_init_allocator(struct mlx4_en_priv *priv,
94                                   struct mlx4_en_rx_ring *ring)
95 {
96         struct mlx4_en_rx_alloc *page_alloc;
97         int i;
98
99         for (i = 0; i < priv->num_frags; i++) {
100                 page_alloc = &ring->page_alloc[i];
101                 page_alloc->page = alloc_pages(GFP_ATOMIC | __GFP_COMP,
102                                                MLX4_EN_ALLOC_ORDER);
103                 if (!page_alloc->page)
104                         goto out;
105
106                 page_alloc->offset = priv->frag_info[i].frag_align;
107                 en_dbg(DRV, priv, "Initialized allocator:%d with page:%p\n",
108                        i, page_alloc->page);
109         }
110         return 0;
111
112 out:
113         while (i--) {
114                 page_alloc = &ring->page_alloc[i];
115                 put_page(page_alloc->page);
116                 page_alloc->page = NULL;
117         }
118         return -ENOMEM;
119 }
120
121 static void mlx4_en_destroy_allocator(struct mlx4_en_priv *priv,
122                                       struct mlx4_en_rx_ring *ring)
123 {
124         struct mlx4_en_rx_alloc *page_alloc;
125         int i;
126
127         for (i = 0; i < priv->num_frags; i++) {
128                 page_alloc = &ring->page_alloc[i];
129                 en_dbg(DRV, priv, "Freeing allocator:%d count:%d\n",
130                        i, page_count(page_alloc->page));
131
132                 put_page(page_alloc->page);
133                 page_alloc->page = NULL;
134         }
135 }
136
137
138 static void mlx4_en_init_rx_desc(struct mlx4_en_priv *priv,
139                                  struct mlx4_en_rx_ring *ring, int index)
140 {
141         struct mlx4_en_rx_desc *rx_desc = ring->buf + ring->stride * index;
142         struct skb_frag_struct *skb_frags = ring->rx_info +
143                                             (index << priv->log_rx_info);
144         int possible_frags;
145         int i;
146
147         /* Set size and memtype fields */
148         for (i = 0; i < priv->num_frags; i++) {
149                 skb_frags[i].size = priv->frag_info[i].frag_size;
150                 rx_desc->data[i].byte_count =
151                         cpu_to_be32(priv->frag_info[i].frag_size);
152                 rx_desc->data[i].lkey = cpu_to_be32(priv->mdev->mr.key);
153         }
154
155         /* If the number of used fragments does not fill up the ring stride,
156          * remaining (unused) fragments must be padded with null address/size
157          * and a special memory key */
158         possible_frags = (ring->stride - sizeof(struct mlx4_en_rx_desc)) / DS_SIZE;
159         for (i = priv->num_frags; i < possible_frags; i++) {
160                 rx_desc->data[i].byte_count = 0;
161                 rx_desc->data[i].lkey = cpu_to_be32(MLX4_EN_MEMTYPE_PAD);
162                 rx_desc->data[i].addr = 0;
163         }
164 }
165
166
167 static int mlx4_en_prepare_rx_desc(struct mlx4_en_priv *priv,
168                                    struct mlx4_en_rx_ring *ring, int index)
169 {
170         struct mlx4_en_rx_desc *rx_desc = ring->buf + (index * ring->stride);
171         struct skb_frag_struct *skb_frags = ring->rx_info +
172                                             (index << priv->log_rx_info);
173         int i;
174
175         for (i = 0; i < priv->num_frags; i++)
176                 if (mlx4_en_alloc_frag(priv, rx_desc, skb_frags, ring->page_alloc, i))
177                         goto err;
178
179         return 0;
180
181 err:
182         while (i--)
183                 put_page(skb_frags[i].page);
184         return -ENOMEM;
185 }
186
187 static inline void mlx4_en_update_rx_prod_db(struct mlx4_en_rx_ring *ring)
188 {
189         *ring->wqres.db.db = cpu_to_be32(ring->prod & 0xffff);
190 }
191
192 static void mlx4_en_free_rx_desc(struct mlx4_en_priv *priv,
193                                  struct mlx4_en_rx_ring *ring,
194                                  int index)
195 {
196         struct mlx4_en_dev *mdev = priv->mdev;
197         struct skb_frag_struct *skb_frags;
198         struct mlx4_en_rx_desc *rx_desc = ring->buf + (index << ring->log_stride);
199         dma_addr_t dma;
200         int nr;
201
202         skb_frags = ring->rx_info + (index << priv->log_rx_info);
203         for (nr = 0; nr < priv->num_frags; nr++) {
204                 en_dbg(DRV, priv, "Freeing fragment:%d\n", nr);
205                 dma = be64_to_cpu(rx_desc->data[nr].addr);
206
207                 en_dbg(DRV, priv, "Unmaping buffer at dma:0x%llx\n", (u64) dma);
208                 pci_unmap_single(mdev->pdev, dma, skb_frags[nr].size,
209                                  PCI_DMA_FROMDEVICE);
210                 put_page(skb_frags[nr].page);
211         }
212 }
213
214 static int mlx4_en_fill_rx_buffers(struct mlx4_en_priv *priv)
215 {
216         struct mlx4_en_rx_ring *ring;
217         int ring_ind;
218         int buf_ind;
219         int new_size;
220
221         for (buf_ind = 0; buf_ind < priv->prof->rx_ring_size; buf_ind++) {
222                 for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
223                         ring = &priv->rx_ring[ring_ind];
224
225                         if (mlx4_en_prepare_rx_desc(priv, ring,
226                                                     ring->actual_size)) {
227                                 if (ring->actual_size < MLX4_EN_MIN_RX_SIZE) {
228                                         en_err(priv, "Failed to allocate "
229                                                      "enough rx buffers\n");
230                                         return -ENOMEM;
231                                 } else {
232                                         new_size = rounddown_pow_of_two(ring->actual_size);
233                                         en_warn(priv, "Only %d buffers allocated "
234                                                       "reducing ring size to %d",
235                                                 ring->actual_size, new_size);
236                                         goto reduce_rings;
237                                 }
238                         }
239                         ring->actual_size++;
240                         ring->prod++;
241                 }
242         }
243         return 0;
244
245 reduce_rings:
246         for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
247                 ring = &priv->rx_ring[ring_ind];
248                 while (ring->actual_size > new_size) {
249                         ring->actual_size--;
250                         ring->prod--;
251                         mlx4_en_free_rx_desc(priv, ring, ring->actual_size);
252                 }
253                 ring->size_mask = ring->actual_size - 1;
254         }
255
256         return 0;
257 }
258
259 static void mlx4_en_free_rx_buf(struct mlx4_en_priv *priv,
260                                 struct mlx4_en_rx_ring *ring)
261 {
262         int index;
263
264         en_dbg(DRV, priv, "Freeing Rx buf - cons:%d prod:%d\n",
265                ring->cons, ring->prod);
266
267         /* Unmap and free Rx buffers */
268         BUG_ON((u32) (ring->prod - ring->cons) > ring->actual_size);
269         while (ring->cons != ring->prod) {
270                 index = ring->cons & ring->size_mask;
271                 en_dbg(DRV, priv, "Processing descriptor:%d\n", index);
272                 mlx4_en_free_rx_desc(priv, ring, index);
273                 ++ring->cons;
274         }
275 }
276
277 int mlx4_en_create_rx_ring(struct mlx4_en_priv *priv,
278                            struct mlx4_en_rx_ring *ring, u32 size, u16 stride)
279 {
280         struct mlx4_en_dev *mdev = priv->mdev;
281         int err;
282         int tmp;
283
284
285         ring->prod = 0;
286         ring->cons = 0;
287         ring->size = size;
288         ring->size_mask = size - 1;
289         ring->stride = stride;
290         ring->log_stride = ffs(ring->stride) - 1;
291         ring->buf_size = ring->size * ring->stride + TXBB_SIZE;
292
293         tmp = size * roundup_pow_of_two(MLX4_EN_MAX_RX_FRAGS *
294                                         sizeof(struct skb_frag_struct));
295         ring->rx_info = vmalloc(tmp);
296         if (!ring->rx_info) {
297                 en_err(priv, "Failed allocating rx_info ring\n");
298                 return -ENOMEM;
299         }
300         en_dbg(DRV, priv, "Allocated rx_info ring at addr:%p size:%d\n",
301                  ring->rx_info, tmp);
302
303         err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres,
304                                  ring->buf_size, 2 * PAGE_SIZE);
305         if (err)
306                 goto err_ring;
307
308         err = mlx4_en_map_buffer(&ring->wqres.buf);
309         if (err) {
310                 en_err(priv, "Failed to map RX buffer\n");
311                 goto err_hwq;
312         }
313         ring->buf = ring->wqres.buf.direct.buf;
314
315         /* Configure lro mngr */
316         memset(&ring->lro, 0, sizeof(struct net_lro_mgr));
317         ring->lro.dev = priv->dev;
318         ring->lro.features = LRO_F_NAPI;
319         ring->lro.frag_align_pad = NET_IP_ALIGN;
320         ring->lro.ip_summed = CHECKSUM_UNNECESSARY;
321         ring->lro.ip_summed_aggr = CHECKSUM_UNNECESSARY;
322         ring->lro.max_desc = mdev->profile.num_lro;
323         ring->lro.max_aggr = MAX_SKB_FRAGS;
324         ring->lro.lro_arr = kzalloc(mdev->profile.num_lro *
325                                     sizeof(struct net_lro_desc),
326                                     GFP_KERNEL);
327         if (!ring->lro.lro_arr) {
328                 en_err(priv, "Failed to allocate lro array\n");
329                 goto err_map;
330         }
331         ring->lro.get_frag_header = mlx4_en_get_frag_header;
332
333         return 0;
334
335 err_map:
336         mlx4_en_unmap_buffer(&ring->wqres.buf);
337 err_hwq:
338         mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
339 err_ring:
340         vfree(ring->rx_info);
341         ring->rx_info = NULL;
342         return err;
343 }
344
345 int mlx4_en_activate_rx_rings(struct mlx4_en_priv *priv)
346 {
347         struct mlx4_en_rx_ring *ring;
348         int i;
349         int ring_ind;
350         int err;
351         int stride = roundup_pow_of_two(sizeof(struct mlx4_en_rx_desc) +
352                                         DS_SIZE * priv->num_frags);
353
354         for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
355                 ring = &priv->rx_ring[ring_ind];
356
357                 ring->prod = 0;
358                 ring->cons = 0;
359                 ring->actual_size = 0;
360                 ring->cqn = priv->rx_cq[ring_ind].mcq.cqn;
361
362                 ring->stride = stride;
363                 if (ring->stride <= TXBB_SIZE)
364                         ring->buf += TXBB_SIZE;
365
366                 ring->log_stride = ffs(ring->stride) - 1;
367                 ring->buf_size = ring->size * ring->stride;
368
369                 memset(ring->buf, 0, ring->buf_size);
370                 mlx4_en_update_rx_prod_db(ring);
371
372                 /* Initailize all descriptors */
373                 for (i = 0; i < ring->size; i++)
374                         mlx4_en_init_rx_desc(priv, ring, i);
375
376                 /* Initialize page allocators */
377                 err = mlx4_en_init_allocator(priv, ring);
378                 if (err) {
379                         en_err(priv, "Failed initializing ring allocator\n");
380                         ring_ind--;
381                         goto err_allocator;
382                 }
383         }
384         err = mlx4_en_fill_rx_buffers(priv);
385         if (err)
386                 goto err_buffers;
387
388         for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
389                 ring = &priv->rx_ring[ring_ind];
390
391                 mlx4_en_update_rx_prod_db(ring);
392         }
393
394         return 0;
395
396 err_buffers:
397         for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++)
398                 mlx4_en_free_rx_buf(priv, &priv->rx_ring[ring_ind]);
399
400         ring_ind = priv->rx_ring_num - 1;
401 err_allocator:
402         while (ring_ind >= 0) {
403                 mlx4_en_destroy_allocator(priv, &priv->rx_ring[ring_ind]);
404                 ring_ind--;
405         }
406         return err;
407 }
408
409 void mlx4_en_destroy_rx_ring(struct mlx4_en_priv *priv,
410                              struct mlx4_en_rx_ring *ring)
411 {
412         struct mlx4_en_dev *mdev = priv->mdev;
413
414         kfree(ring->lro.lro_arr);
415         mlx4_en_unmap_buffer(&ring->wqres.buf);
416         mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size + TXBB_SIZE);
417         vfree(ring->rx_info);
418         ring->rx_info = NULL;
419 }
420
421 void mlx4_en_deactivate_rx_ring(struct mlx4_en_priv *priv,
422                                 struct mlx4_en_rx_ring *ring)
423 {
424         mlx4_en_free_rx_buf(priv, ring);
425         if (ring->stride <= TXBB_SIZE)
426                 ring->buf -= TXBB_SIZE;
427         mlx4_en_destroy_allocator(priv, ring);
428 }
429
430
431 /* Unmap a completed descriptor and free unused pages */
432 static int mlx4_en_complete_rx_desc(struct mlx4_en_priv *priv,
433                                     struct mlx4_en_rx_desc *rx_desc,
434                                     struct skb_frag_struct *skb_frags,
435                                     struct skb_frag_struct *skb_frags_rx,
436                                     struct mlx4_en_rx_alloc *page_alloc,
437                                     int length)
438 {
439         struct mlx4_en_dev *mdev = priv->mdev;
440         struct mlx4_en_frag_info *frag_info;
441         int nr;
442         dma_addr_t dma;
443
444         /* Collect used fragments while replacing them in the HW descirptors */
445         for (nr = 0; nr < priv->num_frags; nr++) {
446                 frag_info = &priv->frag_info[nr];
447                 if (length <= frag_info->frag_prefix_size)
448                         break;
449
450                 /* Save page reference in skb */
451                 skb_frags_rx[nr].page = skb_frags[nr].page;
452                 skb_frags_rx[nr].size = skb_frags[nr].size;
453                 skb_frags_rx[nr].page_offset = skb_frags[nr].page_offset;
454                 dma = be64_to_cpu(rx_desc->data[nr].addr);
455
456                 /* Allocate a replacement page */
457                 if (mlx4_en_alloc_frag(priv, rx_desc, skb_frags, page_alloc, nr))
458                         goto fail;
459
460                 /* Unmap buffer */
461                 pci_unmap_single(mdev->pdev, dma, skb_frags[nr].size,
462                                  PCI_DMA_FROMDEVICE);
463         }
464         /* Adjust size of last fragment to match actual length */
465         if (nr > 0)
466                 skb_frags_rx[nr - 1].size = length -
467                         priv->frag_info[nr - 1].frag_prefix_size;
468         return nr;
469
470 fail:
471         /* Drop all accumulated fragments (which have already been replaced in
472          * the descriptor) of this packet; remaining fragments are reused... */
473         while (nr > 0) {
474                 nr--;
475                 put_page(skb_frags_rx[nr].page);
476         }
477         return 0;
478 }
479
480
481 static struct sk_buff *mlx4_en_rx_skb(struct mlx4_en_priv *priv,
482                                       struct mlx4_en_rx_desc *rx_desc,
483                                       struct skb_frag_struct *skb_frags,
484                                       struct mlx4_en_rx_alloc *page_alloc,
485                                       unsigned int length)
486 {
487         struct mlx4_en_dev *mdev = priv->mdev;
488         struct sk_buff *skb;
489         void *va;
490         int used_frags;
491         dma_addr_t dma;
492
493         skb = dev_alloc_skb(SMALL_PACKET_SIZE + NET_IP_ALIGN);
494         if (!skb) {
495                 en_dbg(RX_ERR, priv, "Failed allocating skb\n");
496                 return NULL;
497         }
498         skb->dev = priv->dev;
499         skb_reserve(skb, NET_IP_ALIGN);
500         skb->len = length;
501         skb->truesize = length + sizeof(struct sk_buff);
502
503         /* Get pointer to first fragment so we could copy the headers into the
504          * (linear part of the) skb */
505         va = page_address(skb_frags[0].page) + skb_frags[0].page_offset;
506
507         if (length <= SMALL_PACKET_SIZE) {
508                 /* We are copying all relevant data to the skb - temporarily
509                  * synch buffers for the copy */
510                 dma = be64_to_cpu(rx_desc->data[0].addr);
511                 dma_sync_single_range_for_cpu(&mdev->pdev->dev, dma, 0,
512                                               length, DMA_FROM_DEVICE);
513                 skb_copy_to_linear_data(skb, va, length);
514                 dma_sync_single_range_for_device(&mdev->pdev->dev, dma, 0,
515                                                  length, DMA_FROM_DEVICE);
516                 skb->tail += length;
517         } else {
518
519                 /* Move relevant fragments to skb */
520                 used_frags = mlx4_en_complete_rx_desc(priv, rx_desc, skb_frags,
521                                                       skb_shinfo(skb)->frags,
522                                                       page_alloc, length);
523                 if (unlikely(!used_frags)) {
524                         kfree_skb(skb);
525                         return NULL;
526                 }
527                 skb_shinfo(skb)->nr_frags = used_frags;
528
529                 /* Copy headers into the skb linear buffer */
530                 memcpy(skb->data, va, HEADER_COPY_SIZE);
531                 skb->tail += HEADER_COPY_SIZE;
532
533                 /* Skip headers in first fragment */
534                 skb_shinfo(skb)->frags[0].page_offset += HEADER_COPY_SIZE;
535
536                 /* Adjust size of first fragment */
537                 skb_shinfo(skb)->frags[0].size -= HEADER_COPY_SIZE;
538                 skb->data_len = length - HEADER_COPY_SIZE;
539         }
540         return skb;
541 }
542
543
544 int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int budget)
545 {
546         struct mlx4_en_priv *priv = netdev_priv(dev);
547         struct mlx4_cqe *cqe;
548         struct mlx4_en_rx_ring *ring = &priv->rx_ring[cq->ring];
549         struct skb_frag_struct *skb_frags;
550         struct skb_frag_struct lro_frags[MLX4_EN_MAX_RX_FRAGS];
551         struct mlx4_en_rx_desc *rx_desc;
552         struct sk_buff *skb;
553         int index;
554         int nr;
555         unsigned int length;
556         int polled = 0;
557         int ip_summed;
558
559         if (!priv->port_up)
560                 return 0;
561
562         /* We assume a 1:1 mapping between CQEs and Rx descriptors, so Rx
563          * descriptor offset can be deduced from the CQE index instead of
564          * reading 'cqe->index' */
565         index = cq->mcq.cons_index & ring->size_mask;
566         cqe = &cq->buf[index];
567
568         /* Process all completed CQEs */
569         while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
570                     cq->mcq.cons_index & cq->size)) {
571
572                 skb_frags = ring->rx_info + (index << priv->log_rx_info);
573                 rx_desc = ring->buf + (index << ring->log_stride);
574
575                 /*
576                  * make sure we read the CQE after we read the ownership bit
577                  */
578                 rmb();
579
580                 /* Drop packet on bad receive or bad checksum */
581                 if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
582                                                 MLX4_CQE_OPCODE_ERROR)) {
583                         en_err(priv, "CQE completed in error - vendor "
584                                   "syndrom:%d syndrom:%d\n",
585                                   ((struct mlx4_err_cqe *) cqe)->vendor_err_syndrome,
586                                   ((struct mlx4_err_cqe *) cqe)->syndrome);
587                         goto next;
588                 }
589                 if (unlikely(cqe->badfcs_enc & MLX4_CQE_BAD_FCS)) {
590                         en_dbg(RX_ERR, priv, "Accepted frame with bad FCS\n");
591                         goto next;
592                 }
593
594                 /*
595                  * Packet is OK - process it.
596                  */
597                 length = be32_to_cpu(cqe->byte_cnt);
598                 ring->bytes += length;
599                 ring->packets++;
600
601                 if (likely(priv->rx_csum)) {
602                         if ((cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
603                             (cqe->checksum == cpu_to_be16(0xffff))) {
604                                 priv->port_stats.rx_chksum_good++;
605                                 /* This packet is eligible for LRO if it is:
606                                  * - DIX Ethernet (type interpretation)
607                                  * - TCP/IP (v4)
608                                  * - without IP options
609                                  * - not an IP fragment */
610                                 if (mlx4_en_can_lro(cqe->status) &&
611                                     dev->features & NETIF_F_LRO) {
612
613                                         nr = mlx4_en_complete_rx_desc(
614                                                 priv, rx_desc,
615                                                 skb_frags, lro_frags,
616                                                 ring->page_alloc, length);
617                                         if (!nr)
618                                                 goto next;
619
620                                         if (priv->vlgrp && (cqe->vlan_my_qpn &
621                                                             cpu_to_be32(MLX4_CQE_VLAN_PRESENT_MASK))) {
622                                                 lro_vlan_hwaccel_receive_frags(
623                                                        &ring->lro, lro_frags,
624                                                        length, length,
625                                                        priv->vlgrp,
626                                                        be16_to_cpu(cqe->sl_vid),
627                                                        NULL, 0);
628                                         } else
629                                                 lro_receive_frags(&ring->lro,
630                                                                   lro_frags,
631                                                                   length,
632                                                                   length,
633                                                                   NULL, 0);
634
635                                         goto next;
636                                 }
637
638                                 /* LRO not possible, complete processing here */
639                                 ip_summed = CHECKSUM_UNNECESSARY;
640                                 INC_PERF_COUNTER(priv->pstats.lro_misses);
641                         } else {
642                                 ip_summed = CHECKSUM_NONE;
643                                 priv->port_stats.rx_chksum_none++;
644                         }
645                 } else {
646                         ip_summed = CHECKSUM_NONE;
647                         priv->port_stats.rx_chksum_none++;
648                 }
649
650                 skb = mlx4_en_rx_skb(priv, rx_desc, skb_frags,
651                                      ring->page_alloc, length);
652                 if (!skb) {
653                         priv->stats.rx_dropped++;
654                         goto next;
655                 }
656
657                 skb->ip_summed = ip_summed;
658                 skb->protocol = eth_type_trans(skb, dev);
659                 skb_record_rx_queue(skb, cq->ring);
660
661                 /* Push it up the stack */
662                 if (priv->vlgrp && (be32_to_cpu(cqe->vlan_my_qpn) &
663                                     MLX4_CQE_VLAN_PRESENT_MASK)) {
664                         vlan_hwaccel_receive_skb(skb, priv->vlgrp,
665                                                 be16_to_cpu(cqe->sl_vid));
666                 } else
667                         netif_receive_skb(skb);
668
669 next:
670                 ++cq->mcq.cons_index;
671                 index = (cq->mcq.cons_index) & ring->size_mask;
672                 cqe = &cq->buf[index];
673                 if (++polled == budget) {
674                         /* We are here because we reached the NAPI budget -
675                          * flush only pending LRO sessions */
676                         lro_flush_all(&ring->lro);
677                         goto out;
678                 }
679         }
680
681         /* If CQ is empty flush all LRO sessions unconditionally */
682         lro_flush_all(&ring->lro);
683
684 out:
685         AVG_PERF_COUNTER(priv->pstats.rx_coal_avg, polled);
686         mlx4_cq_set_ci(&cq->mcq);
687         wmb(); /* ensure HW sees CQ consumer before we post new buffers */
688         ring->cons = cq->mcq.cons_index;
689         ring->prod += polled; /* Polled descriptors were realocated in place */
690         mlx4_en_update_rx_prod_db(ring);
691         return polled;
692 }
693
694
695 void mlx4_en_rx_irq(struct mlx4_cq *mcq)
696 {
697         struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
698         struct mlx4_en_priv *priv = netdev_priv(cq->dev);
699
700         if (priv->port_up)
701                 napi_schedule(&cq->napi);
702         else
703                 mlx4_en_arm_cq(priv, cq);
704 }
705
706 /* Rx CQ polling - called by NAPI */
707 int mlx4_en_poll_rx_cq(struct napi_struct *napi, int budget)
708 {
709         struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
710         struct net_device *dev = cq->dev;
711         struct mlx4_en_priv *priv = netdev_priv(dev);
712         int done;
713
714         done = mlx4_en_process_rx_cq(dev, cq, budget);
715
716         /* If we used up all the quota - we're probably not done yet... */
717         if (done == budget)
718                 INC_PERF_COUNTER(priv->pstats.napi_quota);
719         else {
720                 /* Done for now */
721                 napi_complete(napi);
722                 mlx4_en_arm_cq(priv, cq);
723         }
724         return done;
725 }
726
727
728 /* Calculate the last offset position that accomodates a full fragment
729  * (assuming fagment size = stride-align) */
730 static int mlx4_en_last_alloc_offset(struct mlx4_en_priv *priv, u16 stride, u16 align)
731 {
732         u16 res = MLX4_EN_ALLOC_SIZE % stride;
733         u16 offset = MLX4_EN_ALLOC_SIZE - stride - res + align;
734
735         en_dbg(DRV, priv, "Calculated last offset for stride:%d align:%d "
736                             "res:%d offset:%d\n", stride, align, res, offset);
737         return offset;
738 }
739
740
741 static int frag_sizes[] = {
742         FRAG_SZ0,
743         FRAG_SZ1,
744         FRAG_SZ2,
745         FRAG_SZ3
746 };
747
748 void mlx4_en_calc_rx_buf(struct net_device *dev)
749 {
750         struct mlx4_en_priv *priv = netdev_priv(dev);
751         int eff_mtu = dev->mtu + ETH_HLEN + VLAN_HLEN + ETH_LLC_SNAP_SIZE;
752         int buf_size = 0;
753         int i = 0;
754
755         while (buf_size < eff_mtu) {
756                 priv->frag_info[i].frag_size =
757                         (eff_mtu > buf_size + frag_sizes[i]) ?
758                                 frag_sizes[i] : eff_mtu - buf_size;
759                 priv->frag_info[i].frag_prefix_size = buf_size;
760                 if (!i) {
761                         priv->frag_info[i].frag_align = NET_IP_ALIGN;
762                         priv->frag_info[i].frag_stride =
763                                 ALIGN(frag_sizes[i] + NET_IP_ALIGN, SMP_CACHE_BYTES);
764                 } else {
765                         priv->frag_info[i].frag_align = 0;
766                         priv->frag_info[i].frag_stride =
767                                 ALIGN(frag_sizes[i], SMP_CACHE_BYTES);
768                 }
769                 priv->frag_info[i].last_offset = mlx4_en_last_alloc_offset(
770                                                 priv, priv->frag_info[i].frag_stride,
771                                                 priv->frag_info[i].frag_align);
772                 buf_size += priv->frag_info[i].frag_size;
773                 i++;
774         }
775
776         priv->num_frags = i;
777         priv->rx_skb_size = eff_mtu;
778         priv->log_rx_info = ROUNDUP_LOG2(i * sizeof(struct skb_frag_struct));
779
780         en_dbg(DRV, priv, "Rx buffer scatter-list (effective-mtu:%d "
781                   "num_frags:%d):\n", eff_mtu, priv->num_frags);
782         for (i = 0; i < priv->num_frags; i++) {
783                 en_dbg(DRV, priv, "  frag:%d - size:%d prefix:%d align:%d "
784                                 "stride:%d last_offset:%d\n", i,
785                                 priv->frag_info[i].frag_size,
786                                 priv->frag_info[i].frag_prefix_size,
787                                 priv->frag_info[i].frag_align,
788                                 priv->frag_info[i].frag_stride,
789                                 priv->frag_info[i].last_offset);
790         }
791 }
792
793 /* RSS related functions */
794
795 static int mlx4_en_config_rss_qp(struct mlx4_en_priv *priv, int qpn,
796                                  struct mlx4_en_rx_ring *ring,
797                                  enum mlx4_qp_state *state,
798                                  struct mlx4_qp *qp)
799 {
800         struct mlx4_en_dev *mdev = priv->mdev;
801         struct mlx4_qp_context *context;
802         int err = 0;
803
804         context = kmalloc(sizeof *context , GFP_KERNEL);
805         if (!context) {
806                 en_err(priv, "Failed to allocate qp context\n");
807                 return -ENOMEM;
808         }
809
810         err = mlx4_qp_alloc(mdev->dev, qpn, qp);
811         if (err) {
812                 en_err(priv, "Failed to allocate qp #%x\n", qpn);
813                 goto out;
814         }
815         qp->event = mlx4_en_sqp_event;
816
817         memset(context, 0, sizeof *context);
818         mlx4_en_fill_qp_context(priv, ring->size, ring->stride, 0, 0,
819                                 qpn, ring->cqn, context);
820         context->db_rec_addr = cpu_to_be64(ring->wqres.db.dma);
821
822         err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, context, qp, state);
823         if (err) {
824                 mlx4_qp_remove(mdev->dev, qp);
825                 mlx4_qp_free(mdev->dev, qp);
826         }
827         mlx4_en_update_rx_prod_db(ring);
828 out:
829         kfree(context);
830         return err;
831 }
832
833 /* Allocate rx qp's and configure them according to rss map */
834 int mlx4_en_config_rss_steer(struct mlx4_en_priv *priv)
835 {
836         struct mlx4_en_dev *mdev = priv->mdev;
837         struct mlx4_en_rss_map *rss_map = &priv->rss_map;
838         struct mlx4_qp_context context;
839         struct mlx4_en_rss_context *rss_context;
840         void *ptr;
841         int rss_xor = mdev->profile.rss_xor;
842         u8 rss_mask = mdev->profile.rss_mask;
843         int i, qpn;
844         int err = 0;
845         int good_qps = 0;
846
847         en_dbg(DRV, priv, "Configuring rss steering\n");
848         err = mlx4_qp_reserve_range(mdev->dev, priv->rx_ring_num,
849                                     priv->rx_ring_num,
850                                     &rss_map->base_qpn);
851         if (err) {
852                 en_err(priv, "Failed reserving %d qps\n", priv->rx_ring_num);
853                 return err;
854         }
855
856         for (i = 0; i < priv->rx_ring_num; i++) {
857                 qpn = rss_map->base_qpn + i;
858                 err = mlx4_en_config_rss_qp(priv, qpn, &priv->rx_ring[i],
859                                             &rss_map->state[i],
860                                             &rss_map->qps[i]);
861                 if (err)
862                         goto rss_err;
863
864                 ++good_qps;
865         }
866
867         /* Configure RSS indirection qp */
868         err = mlx4_qp_reserve_range(mdev->dev, 1, 1, &priv->base_qpn);
869         if (err) {
870                 en_err(priv, "Failed to reserve range for RSS "
871                              "indirection qp\n");
872                 goto rss_err;
873         }
874         err = mlx4_qp_alloc(mdev->dev, priv->base_qpn, &rss_map->indir_qp);
875         if (err) {
876                 en_err(priv, "Failed to allocate RSS indirection QP\n");
877                 goto reserve_err;
878         }
879         rss_map->indir_qp.event = mlx4_en_sqp_event;
880         mlx4_en_fill_qp_context(priv, 0, 0, 0, 1, priv->base_qpn,
881                                 priv->rx_ring[0].cqn, &context);
882
883         ptr = ((void *) &context) + 0x3c;
884         rss_context = (struct mlx4_en_rss_context *) ptr;
885         rss_context->base_qpn = cpu_to_be32(ilog2(priv->rx_ring_num) << 24 |
886                                             (rss_map->base_qpn));
887         rss_context->default_qpn = cpu_to_be32(rss_map->base_qpn);
888         rss_context->hash_fn = rss_xor & 0x3;
889         rss_context->flags = rss_mask << 2;
890
891         err = mlx4_qp_to_ready(mdev->dev, &priv->res.mtt, &context,
892                                &rss_map->indir_qp, &rss_map->indir_state);
893         if (err)
894                 goto indir_err;
895
896         return 0;
897
898 indir_err:
899         mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
900                        MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
901         mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
902         mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
903 reserve_err:
904         mlx4_qp_release_range(mdev->dev, priv->base_qpn, 1);
905 rss_err:
906         for (i = 0; i < good_qps; i++) {
907                 mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
908                                MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
909                 mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
910                 mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
911         }
912         mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
913         return err;
914 }
915
916 void mlx4_en_release_rss_steer(struct mlx4_en_priv *priv)
917 {
918         struct mlx4_en_dev *mdev = priv->mdev;
919         struct mlx4_en_rss_map *rss_map = &priv->rss_map;
920         int i;
921
922         mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
923                        MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
924         mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
925         mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
926         mlx4_qp_release_range(mdev->dev, priv->base_qpn, 1);
927
928         for (i = 0; i < priv->rx_ring_num; i++) {
929                 mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
930                                MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
931                 mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
932                 mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
933         }
934         mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
935 }
936
937
938
939
940