mlx4_en: Handle page allocation failure during receive
[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 static 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
49 static void mlx4_en_srq_event(struct mlx4_srq *srq, enum mlx4_event type)
50 {
51         return;
52 }
53
54 static 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
66 static 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
103 static 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
122 out:
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
131 static 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
148 static 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
180 static 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
194 err:
195         while (i--)
196                 put_page(skb_frags[i].page);
197         return -ENOMEM;
198 }
199
200 static 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
205 static int mlx4_en_fill_rx_buffers(struct mlx4_en_priv *priv)
206 {
207         struct mlx4_en_dev *mdev = priv->mdev;
208         struct mlx4_en_rx_ring *ring;
209         int ring_ind;
210         int buf_ind;
211
212         for (buf_ind = 0; buf_ind < priv->prof->rx_ring_size; buf_ind++) {
213                 for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
214                         ring = &priv->rx_ring[ring_ind];
215
216                         if (mlx4_en_prepare_rx_desc(priv, ring,
217                                                     ring->actual_size)) {
218                                 if (ring->actual_size < MLX4_EN_MIN_RX_SIZE) {
219                                         mlx4_err(mdev, "Failed to allocate "
220                                                        "enough rx buffers\n");
221                                         return -ENOMEM;
222                                 } else {
223                                         if (netif_msg_rx_err(priv))
224                                                 mlx4_warn(mdev,
225                                                           "Only %d buffers allocated\n",
226                                                           ring->actual_size);
227                                         goto out;
228                                 }
229                         }
230                         ring->actual_size++;
231                         ring->prod++;
232                 }
233         }
234 out:
235         return 0;
236 }
237
238 static int mlx4_en_fill_rx_buf(struct net_device *dev,
239                                struct mlx4_en_rx_ring *ring)
240 {
241         struct mlx4_en_priv *priv = netdev_priv(dev);
242         int num = 0;
243         int err;
244
245         while ((u32) (ring->prod - ring->cons) < ring->actual_size) {
246                 err = mlx4_en_prepare_rx_desc(priv, ring, ring->prod &
247                                               ring->size_mask);
248                 if (err) {
249                         if (netif_msg_rx_err(priv))
250                                 mlx4_warn(priv->mdev,
251                                           "Failed preparing rx descriptor\n");
252                         priv->port_stats.rx_alloc_failed++;
253                         break;
254                 }
255                 ++num;
256                 ++ring->prod;
257         }
258         if ((u32) (ring->prod - ring->cons) == ring->size)
259                 ring->full = 1;
260
261         return num;
262 }
263
264 static void mlx4_en_free_rx_buf(struct mlx4_en_priv *priv,
265                                 struct mlx4_en_rx_ring *ring)
266 {
267         struct mlx4_en_dev *mdev = priv->mdev;
268         struct skb_frag_struct *skb_frags;
269         struct mlx4_en_rx_desc *rx_desc;
270         dma_addr_t dma;
271         int index;
272         int nr;
273
274         mlx4_dbg(DRV, priv, "Freeing Rx buf - cons:%d prod:%d\n",
275                         ring->cons, ring->prod);
276
277         /* Unmap and free Rx buffers */
278         BUG_ON((u32) (ring->prod - ring->cons) > ring->size);
279         while (ring->cons != ring->prod) {
280                 index = ring->cons & ring->size_mask;
281                 rx_desc = ring->buf + (index << ring->log_stride);
282                 skb_frags = ring->rx_info + (index << priv->log_rx_info);
283                 mlx4_dbg(DRV, priv, "Processing descriptor:%d\n", index);
284
285                 for (nr = 0; nr < priv->num_frags; nr++) {
286                         mlx4_dbg(DRV, priv, "Freeing fragment:%d\n", nr);
287                         dma = be64_to_cpu(rx_desc->data[nr].addr);
288
289                         mlx4_dbg(DRV, priv, "Unmaping buffer at dma:0x%llx\n", (u64) dma);
290                         pci_unmap_single(mdev->pdev, dma, skb_frags[nr].size,
291                                          PCI_DMA_FROMDEVICE);
292                         put_page(skb_frags[nr].page);
293                 }
294                 ++ring->cons;
295         }
296 }
297
298
299 void mlx4_en_rx_refill(struct work_struct *work)
300 {
301         struct delayed_work *delay = to_delayed_work(work);
302         struct mlx4_en_priv *priv = container_of(delay, struct mlx4_en_priv,
303                                                  refill_task);
304         struct mlx4_en_dev *mdev = priv->mdev;
305         struct net_device *dev = priv->dev;
306         struct mlx4_en_rx_ring *ring;
307         int need_refill = 0;
308         int i;
309
310         mutex_lock(&mdev->state_lock);
311         if (!mdev->device_up || !priv->port_up)
312                 goto out;
313
314         /* We only get here if there are no receive buffers, so we can't race
315          * with Rx interrupts while filling buffers */
316         for (i = 0; i < priv->rx_ring_num; i++) {
317                 ring = &priv->rx_ring[i];
318                 if (ring->need_refill) {
319                         if (mlx4_en_fill_rx_buf(dev, ring)) {
320                                 ring->need_refill = 0;
321                                 mlx4_en_update_rx_prod_db(ring);
322                         } else
323                                 need_refill = 1;
324                 }
325         }
326         if (need_refill)
327                 queue_delayed_work(mdev->workqueue, &priv->refill_task, HZ);
328
329 out:
330         mutex_unlock(&mdev->state_lock);
331 }
332
333
334 int mlx4_en_create_rx_ring(struct mlx4_en_priv *priv,
335                            struct mlx4_en_rx_ring *ring, u32 size, u16 stride)
336 {
337         struct mlx4_en_dev *mdev = priv->mdev;
338         int err;
339         int tmp;
340
341         /* Sanity check SRQ size before proceeding */
342         if (size >= mdev->dev->caps.max_srq_wqes)
343                 return -EINVAL;
344
345         ring->prod = 0;
346         ring->cons = 0;
347         ring->size = size;
348         ring->size_mask = size - 1;
349         ring->stride = stride;
350         ring->log_stride = ffs(ring->stride) - 1;
351         ring->buf_size = ring->size * ring->stride;
352
353         tmp = size * roundup_pow_of_two(MLX4_EN_MAX_RX_FRAGS *
354                                         sizeof(struct skb_frag_struct));
355         ring->rx_info = vmalloc(tmp);
356         if (!ring->rx_info) {
357                 mlx4_err(mdev, "Failed allocating rx_info ring\n");
358                 return -ENOMEM;
359         }
360         mlx4_dbg(DRV, priv, "Allocated rx_info ring at addr:%p size:%d\n",
361                  ring->rx_info, tmp);
362
363         err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres,
364                                  ring->buf_size, 2 * PAGE_SIZE);
365         if (err)
366                 goto err_ring;
367
368         err = mlx4_en_map_buffer(&ring->wqres.buf);
369         if (err) {
370                 mlx4_err(mdev, "Failed to map RX buffer\n");
371                 goto err_hwq;
372         }
373         ring->buf = ring->wqres.buf.direct.buf;
374
375         /* Configure lro mngr */
376         memset(&ring->lro, 0, sizeof(struct net_lro_mgr));
377         ring->lro.dev = priv->dev;
378         ring->lro.features = LRO_F_NAPI;
379         ring->lro.frag_align_pad = NET_IP_ALIGN;
380         ring->lro.ip_summed = CHECKSUM_UNNECESSARY;
381         ring->lro.ip_summed_aggr = CHECKSUM_UNNECESSARY;
382         ring->lro.max_desc = mdev->profile.num_lro;
383         ring->lro.max_aggr = MAX_SKB_FRAGS;
384         ring->lro.lro_arr = kzalloc(mdev->profile.num_lro *
385                                     sizeof(struct net_lro_desc),
386                                     GFP_KERNEL);
387         if (!ring->lro.lro_arr) {
388                 mlx4_err(mdev, "Failed to allocate lro array\n");
389                 goto err_map;
390         }
391         ring->lro.get_frag_header = mlx4_en_get_frag_header;
392
393         return 0;
394
395 err_map:
396         mlx4_en_unmap_buffer(&ring->wqres.buf);
397 err_hwq:
398         mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
399 err_ring:
400         vfree(ring->rx_info);
401         ring->rx_info = NULL;
402         return err;
403 }
404
405 int mlx4_en_activate_rx_rings(struct mlx4_en_priv *priv)
406 {
407         struct mlx4_en_dev *mdev = priv->mdev;
408         struct mlx4_wqe_srq_next_seg *next;
409         struct mlx4_en_rx_ring *ring;
410         int i;
411         int ring_ind;
412         int err;
413         int stride = roundup_pow_of_two(sizeof(struct mlx4_en_rx_desc) +
414                                         DS_SIZE * priv->num_frags);
415         int max_gs = (stride - sizeof(struct mlx4_wqe_srq_next_seg)) / DS_SIZE;
416
417         for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
418                 ring = &priv->rx_ring[ring_ind];
419
420                 ring->prod = 0;
421                 ring->cons = 0;
422                 ring->actual_size = 0;
423                 ring->cqn = priv->rx_cq[ring_ind].mcq.cqn;
424
425                 ring->stride = stride;
426                 ring->log_stride = ffs(ring->stride) - 1;
427                 ring->buf_size = ring->size * ring->stride;
428
429                 memset(ring->buf, 0, ring->buf_size);
430                 mlx4_en_update_rx_prod_db(ring);
431
432                 /* Initailize all descriptors */
433                 for (i = 0; i < ring->size; i++)
434                         mlx4_en_init_rx_desc(priv, ring, i);
435
436                 /* Initialize page allocators */
437                 err = mlx4_en_init_allocator(priv, ring);
438                 if (err) {
439                         mlx4_err(mdev, "Failed initializing ring allocator\n");
440                         ring_ind--;
441                         goto err_allocator;
442                 }
443
444                 /* Fill Rx buffers */
445                 ring->full = 0;
446         }
447         err = mlx4_en_fill_rx_buffers(priv);
448         if (err)
449                 goto err_buffers;
450
451         for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
452                 ring = &priv->rx_ring[ring_ind];
453
454                 mlx4_en_update_rx_prod_db(ring);
455
456                 /* Configure SRQ representing the ring */
457                 ring->srq.max    = ring->size;
458                 ring->srq.max_gs = max_gs;
459                 ring->srq.wqe_shift = ilog2(ring->stride);
460
461                 for (i = 0; i < ring->srq.max; ++i) {
462                         next = get_wqe(ring, i);
463                         next->next_wqe_index =
464                         cpu_to_be16((i + 1) & (ring->srq.max - 1));
465                 }
466
467                 err = mlx4_srq_alloc(mdev->dev, mdev->priv_pdn, &ring->wqres.mtt,
468                                      ring->wqres.db.dma, &ring->srq);
469                 if (err){
470                         mlx4_err(mdev, "Failed to allocate srq\n");
471                         ring_ind--;
472                         goto err_srq;
473                 }
474                 ring->srq.event = mlx4_en_srq_event;
475         }
476
477         return 0;
478
479 err_srq:
480         while (ring_ind >= 0) {
481                 ring = &priv->rx_ring[ring_ind];
482                 mlx4_srq_free(mdev->dev, &ring->srq);
483                 ring_ind--;
484         }
485
486 err_buffers:
487         for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++)
488                 mlx4_en_free_rx_buf(priv, &priv->rx_ring[ring_ind]);
489
490         ring_ind = priv->rx_ring_num - 1;
491 err_allocator:
492         while (ring_ind >= 0) {
493                 mlx4_en_destroy_allocator(priv, &priv->rx_ring[ring_ind]);
494                 ring_ind--;
495         }
496         return err;
497 }
498
499 void mlx4_en_destroy_rx_ring(struct mlx4_en_priv *priv,
500                              struct mlx4_en_rx_ring *ring)
501 {
502         struct mlx4_en_dev *mdev = priv->mdev;
503
504         kfree(ring->lro.lro_arr);
505         mlx4_en_unmap_buffer(&ring->wqres.buf);
506         mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
507         vfree(ring->rx_info);
508         ring->rx_info = NULL;
509 }
510
511 void mlx4_en_deactivate_rx_ring(struct mlx4_en_priv *priv,
512                                 struct mlx4_en_rx_ring *ring)
513 {
514         struct mlx4_en_dev *mdev = priv->mdev;
515
516         mlx4_srq_free(mdev->dev, &ring->srq);
517         mlx4_en_free_rx_buf(priv, ring);
518         mlx4_en_destroy_allocator(priv, ring);
519 }
520
521
522 /* Unmap a completed descriptor and free unused pages */
523 static int mlx4_en_complete_rx_desc(struct mlx4_en_priv *priv,
524                                     struct mlx4_en_rx_desc *rx_desc,
525                                     struct skb_frag_struct *skb_frags,
526                                     struct skb_frag_struct *skb_frags_rx,
527                                     struct mlx4_en_rx_alloc *page_alloc,
528                                     int length)
529 {
530         struct mlx4_en_dev *mdev = priv->mdev;
531         struct mlx4_en_frag_info *frag_info;
532         int nr;
533         dma_addr_t dma;
534
535         /* Collect used fragments while replacing them in the HW descirptors */
536         for (nr = 0; nr < priv->num_frags; nr++) {
537                 frag_info = &priv->frag_info[nr];
538                 if (length <= frag_info->frag_prefix_size)
539                         break;
540
541                 /* Save page reference in skb */
542                 skb_frags_rx[nr].page = skb_frags[nr].page;
543                 skb_frags_rx[nr].size = skb_frags[nr].size;
544                 skb_frags_rx[nr].page_offset = skb_frags[nr].page_offset;
545                 dma = be64_to_cpu(rx_desc->data[nr].addr);
546
547                 /* Allocate a replacement page */
548                 if (mlx4_en_alloc_frag(priv, rx_desc, skb_frags, page_alloc, nr))
549                         goto fail;
550
551                 /* Unmap buffer */
552                 pci_unmap_single(mdev->pdev, dma, skb_frags[nr].size,
553                                  PCI_DMA_FROMDEVICE);
554         }
555         /* Adjust size of last fragment to match actual length */
556         skb_frags_rx[nr - 1].size = length -
557                 priv->frag_info[nr - 1].frag_prefix_size;
558         return nr;
559
560 fail:
561         /* Drop all accumulated fragments (which have already been replaced in
562          * the descriptor) of this packet; remaining fragments are reused... */
563         while (nr > 0) {
564                 nr--;
565                 put_page(skb_frags_rx[nr].page);
566         }
567         return 0;
568 }
569
570
571 static struct sk_buff *mlx4_en_rx_skb(struct mlx4_en_priv *priv,
572                                       struct mlx4_en_rx_desc *rx_desc,
573                                       struct skb_frag_struct *skb_frags,
574                                       struct mlx4_en_rx_alloc *page_alloc,
575                                       unsigned int length)
576 {
577         struct mlx4_en_dev *mdev = priv->mdev;
578         struct sk_buff *skb;
579         void *va;
580         int used_frags;
581         dma_addr_t dma;
582
583         skb = dev_alloc_skb(SMALL_PACKET_SIZE + NET_IP_ALIGN);
584         if (!skb) {
585                 mlx4_dbg(RX_ERR, priv, "Failed allocating skb\n");
586                 return NULL;
587         }
588         skb->dev = priv->dev;
589         skb_reserve(skb, NET_IP_ALIGN);
590         skb->len = length;
591         skb->truesize = length + sizeof(struct sk_buff);
592
593         /* Get pointer to first fragment so we could copy the headers into the
594          * (linear part of the) skb */
595         va = page_address(skb_frags[0].page) + skb_frags[0].page_offset;
596
597         if (length <= SMALL_PACKET_SIZE) {
598                 /* We are copying all relevant data to the skb - temporarily
599                  * synch buffers for the copy */
600                 dma = be64_to_cpu(rx_desc->data[0].addr);
601                 dma_sync_single_range_for_cpu(&mdev->pdev->dev, dma, 0,
602                                               length, DMA_FROM_DEVICE);
603                 skb_copy_to_linear_data(skb, va, length);
604                 dma_sync_single_range_for_device(&mdev->pdev->dev, dma, 0,
605                                                  length, DMA_FROM_DEVICE);
606                 skb->tail += length;
607         } else {
608
609                 /* Move relevant fragments to skb */
610                 used_frags = mlx4_en_complete_rx_desc(priv, rx_desc, skb_frags,
611                                                       skb_shinfo(skb)->frags,
612                                                       page_alloc, length);
613                 if (unlikely(!used_frags)) {
614                         kfree_skb(skb);
615                         return NULL;
616                 }
617                 skb_shinfo(skb)->nr_frags = used_frags;
618
619                 /* Copy headers into the skb linear buffer */
620                 memcpy(skb->data, va, HEADER_COPY_SIZE);
621                 skb->tail += HEADER_COPY_SIZE;
622
623                 /* Skip headers in first fragment */
624                 skb_shinfo(skb)->frags[0].page_offset += HEADER_COPY_SIZE;
625
626                 /* Adjust size of first fragment */
627                 skb_shinfo(skb)->frags[0].size -= HEADER_COPY_SIZE;
628                 skb->data_len = length - HEADER_COPY_SIZE;
629         }
630         return skb;
631 }
632
633 static void mlx4_en_copy_desc(struct mlx4_en_priv *priv,
634                               struct mlx4_en_rx_ring *ring,
635                               int from, int to, int num)
636 {
637         struct skb_frag_struct *skb_frags_from;
638         struct skb_frag_struct *skb_frags_to;
639         struct mlx4_en_rx_desc *rx_desc_from;
640         struct mlx4_en_rx_desc *rx_desc_to;
641         int from_index, to_index;
642         int nr, i;
643
644         for (i = 0; i < num; i++) {
645                 from_index = (from + i) & ring->size_mask;
646                 to_index = (to + i) & ring->size_mask;
647                 skb_frags_from = ring->rx_info + (from_index << priv->log_rx_info);
648                 skb_frags_to = ring->rx_info + (to_index << priv->log_rx_info);
649                 rx_desc_from = ring->buf + (from_index << ring->log_stride);
650                 rx_desc_to = ring->buf + (to_index << ring->log_stride);
651
652                 for (nr = 0; nr < priv->num_frags; nr++) {
653                         skb_frags_to[nr].page = skb_frags_from[nr].page;
654                         skb_frags_to[nr].page_offset = skb_frags_from[nr].page_offset;
655                         rx_desc_to->data[nr].addr = rx_desc_from->data[nr].addr;
656                 }
657         }
658 }
659
660
661 int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int budget)
662 {
663         struct mlx4_en_priv *priv = netdev_priv(dev);
664         struct mlx4_en_dev *mdev = priv->mdev;
665         struct mlx4_cqe *cqe;
666         struct mlx4_en_rx_ring *ring = &priv->rx_ring[cq->ring];
667         struct skb_frag_struct *skb_frags;
668         struct skb_frag_struct lro_frags[MLX4_EN_MAX_RX_FRAGS];
669         struct mlx4_en_rx_desc *rx_desc;
670         struct sk_buff *skb;
671         int index;
672         int nr;
673         unsigned int length;
674         int polled = 0;
675         int ip_summed;
676
677         if (!priv->port_up)
678                 return 0;
679
680         /* We assume a 1:1 mapping between CQEs and Rx descriptors, so Rx
681          * descriptor offset can be deduced from the CQE index instead of
682          * reading 'cqe->index' */
683         index = cq->mcq.cons_index & ring->size_mask;
684         cqe = &cq->buf[index];
685
686         /* Process all completed CQEs */
687         while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
688                     cq->mcq.cons_index & cq->size)) {
689
690                 skb_frags = ring->rx_info + (index << priv->log_rx_info);
691                 rx_desc = ring->buf + (index << ring->log_stride);
692
693                 /*
694                  * make sure we read the CQE after we read the ownership bit
695                  */
696                 rmb();
697
698                 /* Drop packet on bad receive or bad checksum */
699                 if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
700                                                 MLX4_CQE_OPCODE_ERROR)) {
701                         mlx4_err(mdev, "CQE completed in error - vendor "
702                                   "syndrom:%d syndrom:%d\n",
703                                   ((struct mlx4_err_cqe *) cqe)->vendor_err_syndrome,
704                                   ((struct mlx4_err_cqe *) cqe)->syndrome);
705                         goto next;
706                 }
707                 if (unlikely(cqe->badfcs_enc & MLX4_CQE_BAD_FCS)) {
708                         mlx4_dbg(RX_ERR, priv, "Accepted frame with bad FCS\n");
709                         goto next;
710                 }
711
712                 /*
713                  * Packet is OK - process it.
714                  */
715                 length = be32_to_cpu(cqe->byte_cnt);
716                 ring->bytes += length;
717                 ring->packets++;
718
719                 if (likely(priv->rx_csum)) {
720                         if ((cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
721                             (cqe->checksum == cpu_to_be16(0xffff))) {
722                                 priv->port_stats.rx_chksum_good++;
723                                 /* This packet is eligible for LRO if it is:
724                                  * - DIX Ethernet (type interpretation)
725                                  * - TCP/IP (v4)
726                                  * - without IP options
727                                  * - not an IP fragment */
728                                 if (mlx4_en_can_lro(cqe->status) &&
729                                     dev->features & NETIF_F_LRO) {
730
731                                         nr = mlx4_en_complete_rx_desc(
732                                                 priv, rx_desc,
733                                                 skb_frags, lro_frags,
734                                                 ring->page_alloc, length);
735                                         if (!nr)
736                                                 goto next;
737
738                                         if (priv->vlgrp && (cqe->vlan_my_qpn &
739                                                             cpu_to_be32(MLX4_CQE_VLAN_PRESENT_MASK))) {
740                                                 lro_vlan_hwaccel_receive_frags(
741                                                        &ring->lro, lro_frags,
742                                                        length, length,
743                                                        priv->vlgrp,
744                                                        be16_to_cpu(cqe->sl_vid),
745                                                        NULL, 0);
746                                         } else
747                                                 lro_receive_frags(&ring->lro,
748                                                                   lro_frags,
749                                                                   length,
750                                                                   length,
751                                                                   NULL, 0);
752
753                                         goto next;
754                                 }
755
756                                 /* LRO not possible, complete processing here */
757                                 ip_summed = CHECKSUM_UNNECESSARY;
758                                 INC_PERF_COUNTER(priv->pstats.lro_misses);
759                         } else {
760                                 ip_summed = CHECKSUM_NONE;
761                                 priv->port_stats.rx_chksum_none++;
762                         }
763                 } else {
764                         ip_summed = CHECKSUM_NONE;
765                         priv->port_stats.rx_chksum_none++;
766                 }
767
768                 skb = mlx4_en_rx_skb(priv, rx_desc, skb_frags,
769                                      ring->page_alloc, length);
770                 if (!skb) {
771                         priv->stats.rx_dropped++;
772                         goto next;
773                 }
774
775                 skb->ip_summed = ip_summed;
776                 skb->protocol = eth_type_trans(skb, dev);
777                 skb_record_rx_queue(skb, cq->ring);
778
779                 /* Push it up the stack */
780                 if (priv->vlgrp && (be32_to_cpu(cqe->vlan_my_qpn) &
781                                     MLX4_CQE_VLAN_PRESENT_MASK)) {
782                         vlan_hwaccel_receive_skb(skb, priv->vlgrp,
783                                                 be16_to_cpu(cqe->sl_vid));
784                 } else
785                         netif_receive_skb(skb);
786
787 next:
788                 ++cq->mcq.cons_index;
789                 index = (cq->mcq.cons_index) & ring->size_mask;
790                 cqe = &cq->buf[index];
791                 if (++polled == budget) {
792                         /* We are here because we reached the NAPI budget -
793                          * flush only pending LRO sessions */
794                         lro_flush_all(&ring->lro);
795                         goto out;
796                 }
797         }
798
799         /* If CQ is empty flush all LRO sessions unconditionally */
800         lro_flush_all(&ring->lro);
801
802 out:
803         AVG_PERF_COUNTER(priv->pstats.rx_coal_avg, polled);
804         mlx4_cq_set_ci(&cq->mcq);
805         wmb(); /* ensure HW sees CQ consumer before we post new buffers */
806         ring->cons = cq->mcq.cons_index;
807         ring->prod += polled; /* Polled descriptors were realocated in place */
808         if (unlikely(!ring->full)) {
809                 mlx4_en_copy_desc(priv, ring, ring->cons - polled,
810                                   ring->prod - polled, polled);
811                 mlx4_en_fill_rx_buf(dev, ring);
812         }
813         mlx4_en_update_rx_prod_db(ring);
814         return polled;
815 }
816
817
818 void mlx4_en_rx_irq(struct mlx4_cq *mcq)
819 {
820         struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
821         struct mlx4_en_priv *priv = netdev_priv(cq->dev);
822
823         if (priv->port_up)
824                 napi_schedule(&cq->napi);
825         else
826                 mlx4_en_arm_cq(priv, cq);
827 }
828
829 /* Rx CQ polling - called by NAPI */
830 int mlx4_en_poll_rx_cq(struct napi_struct *napi, int budget)
831 {
832         struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
833         struct net_device *dev = cq->dev;
834         struct mlx4_en_priv *priv = netdev_priv(dev);
835         int done;
836
837         done = mlx4_en_process_rx_cq(dev, cq, budget);
838
839         /* If we used up all the quota - we're probably not done yet... */
840         if (done == budget)
841                 INC_PERF_COUNTER(priv->pstats.napi_quota);
842         else {
843                 /* Done for now */
844                 napi_complete(napi);
845                 mlx4_en_arm_cq(priv, cq);
846         }
847         return done;
848 }
849
850
851 /* Calculate the last offset position that accomodates a full fragment
852  * (assuming fagment size = stride-align) */
853 static int mlx4_en_last_alloc_offset(struct mlx4_en_priv *priv, u16 stride, u16 align)
854 {
855         u16 res = MLX4_EN_ALLOC_SIZE % stride;
856         u16 offset = MLX4_EN_ALLOC_SIZE - stride - res + align;
857
858         mlx4_dbg(DRV, priv, "Calculated last offset for stride:%d align:%d "
859                             "res:%d offset:%d\n", stride, align, res, offset);
860         return offset;
861 }
862
863
864 static int frag_sizes[] = {
865         FRAG_SZ0,
866         FRAG_SZ1,
867         FRAG_SZ2,
868         FRAG_SZ3
869 };
870
871 void mlx4_en_calc_rx_buf(struct net_device *dev)
872 {
873         struct mlx4_en_priv *priv = netdev_priv(dev);
874         int eff_mtu = dev->mtu + ETH_HLEN + VLAN_HLEN + ETH_LLC_SNAP_SIZE;
875         int buf_size = 0;
876         int i = 0;
877
878         while (buf_size < eff_mtu) {
879                 priv->frag_info[i].frag_size =
880                         (eff_mtu > buf_size + frag_sizes[i]) ?
881                                 frag_sizes[i] : eff_mtu - buf_size;
882                 priv->frag_info[i].frag_prefix_size = buf_size;
883                 if (!i) {
884                         priv->frag_info[i].frag_align = NET_IP_ALIGN;
885                         priv->frag_info[i].frag_stride =
886                                 ALIGN(frag_sizes[i] + NET_IP_ALIGN, SMP_CACHE_BYTES);
887                 } else {
888                         priv->frag_info[i].frag_align = 0;
889                         priv->frag_info[i].frag_stride =
890                                 ALIGN(frag_sizes[i], SMP_CACHE_BYTES);
891                 }
892                 priv->frag_info[i].last_offset = mlx4_en_last_alloc_offset(
893                                                 priv, priv->frag_info[i].frag_stride,
894                                                 priv->frag_info[i].frag_align);
895                 buf_size += priv->frag_info[i].frag_size;
896                 i++;
897         }
898
899         priv->num_frags = i;
900         priv->rx_skb_size = eff_mtu;
901         priv->log_rx_info = ROUNDUP_LOG2(i * sizeof(struct skb_frag_struct));
902
903         mlx4_dbg(DRV, priv, "Rx buffer scatter-list (effective-mtu:%d "
904                   "num_frags:%d):\n", eff_mtu, priv->num_frags);
905         for (i = 0; i < priv->num_frags; i++) {
906                 mlx4_dbg(DRV, priv, "  frag:%d - size:%d prefix:%d align:%d "
907                                 "stride:%d last_offset:%d\n", i,
908                                 priv->frag_info[i].frag_size,
909                                 priv->frag_info[i].frag_prefix_size,
910                                 priv->frag_info[i].frag_align,
911                                 priv->frag_info[i].frag_stride,
912                                 priv->frag_info[i].last_offset);
913         }
914 }
915
916 /* RSS related functions */
917
918 /* Calculate rss size and map each entry in rss table to rx ring */
919 void mlx4_en_set_default_rss_map(struct mlx4_en_priv *priv,
920                                  struct mlx4_en_rss_map *rss_map,
921                                  int num_entries, int num_rings)
922 {
923         int i;
924
925         rss_map->size = roundup_pow_of_two(num_entries);
926         mlx4_dbg(DRV, priv, "Setting default RSS map of %d entires\n",
927                  rss_map->size);
928
929         for (i = 0; i < rss_map->size; i++) {
930                 rss_map->map[i] = i % num_rings;
931                 mlx4_dbg(DRV, priv, "Entry %d ---> ring %d\n", i, rss_map->map[i]);
932         }
933 }
934
935 static int mlx4_en_config_rss_qp(struct mlx4_en_priv *priv,
936                                  int qpn, int srqn, int cqn,
937                                  enum mlx4_qp_state *state,
938                                  struct mlx4_qp *qp)
939 {
940         struct mlx4_en_dev *mdev = priv->mdev;
941         struct mlx4_qp_context *context;
942         int err = 0;
943
944         context = kmalloc(sizeof *context , GFP_KERNEL);
945         if (!context) {
946                 mlx4_err(mdev, "Failed to allocate qp context\n");
947                 return -ENOMEM;
948         }
949
950         err = mlx4_qp_alloc(mdev->dev, qpn, qp);
951         if (err) {
952                 mlx4_err(mdev, "Failed to allocate qp #%d\n", qpn);
953                 goto out;
954                 return err;
955         }
956         qp->event = mlx4_en_sqp_event;
957
958         memset(context, 0, sizeof *context);
959         mlx4_en_fill_qp_context(priv, 0, 0, 0, 0, qpn, cqn, srqn, context);
960
961         err = mlx4_qp_to_ready(mdev->dev, &priv->res.mtt, context, qp, state);
962         if (err) {
963                 mlx4_qp_remove(mdev->dev, qp);
964                 mlx4_qp_free(mdev->dev, qp);
965         }
966 out:
967         kfree(context);
968         return err;
969 }
970
971 /* Allocate rx qp's and configure them according to rss map */
972 int mlx4_en_config_rss_steer(struct mlx4_en_priv *priv)
973 {
974         struct mlx4_en_dev *mdev = priv->mdev;
975         struct mlx4_en_rss_map *rss_map = &priv->rss_map;
976         struct mlx4_qp_context context;
977         struct mlx4_en_rss_context *rss_context;
978         void *ptr;
979         int rss_xor = mdev->profile.rss_xor;
980         u8 rss_mask = mdev->profile.rss_mask;
981         int i, srqn, qpn, cqn;
982         int err = 0;
983         int good_qps = 0;
984
985         mlx4_dbg(DRV, priv, "Configuring rss steering for port %u\n", priv->port);
986         err = mlx4_qp_reserve_range(mdev->dev, rss_map->size,
987                                     rss_map->size, &rss_map->base_qpn);
988         if (err) {
989                 mlx4_err(mdev, "Failed reserving %d qps for port %u\n",
990                          rss_map->size, priv->port);
991                 return err;
992         }
993
994         for (i = 0; i < rss_map->size; i++) {
995                 cqn = priv->rx_ring[rss_map->map[i]].cqn;
996                 srqn = priv->rx_ring[rss_map->map[i]].srq.srqn;
997                 qpn = rss_map->base_qpn + i;
998                 err = mlx4_en_config_rss_qp(priv, qpn, srqn, cqn,
999                                             &rss_map->state[i],
1000                                             &rss_map->qps[i]);
1001                 if (err)
1002                         goto rss_err;
1003
1004                 ++good_qps;
1005         }
1006
1007         /* Configure RSS indirection qp */
1008         err = mlx4_qp_reserve_range(mdev->dev, 1, 1, &priv->base_qpn);
1009         if (err) {
1010                 mlx4_err(mdev, "Failed to reserve range for RSS "
1011                                "indirection qp\n");
1012                 goto rss_err;
1013         }
1014         err = mlx4_qp_alloc(mdev->dev, priv->base_qpn, &rss_map->indir_qp);
1015         if (err) {
1016                 mlx4_err(mdev, "Failed to allocate RSS indirection QP\n");
1017                 goto reserve_err;
1018         }
1019         rss_map->indir_qp.event = mlx4_en_sqp_event;
1020         mlx4_en_fill_qp_context(priv, 0, 0, 0, 1, priv->base_qpn,
1021                                 priv->rx_ring[0].cqn, 0, &context);
1022
1023         ptr = ((void *) &context) + 0x3c;
1024         rss_context = (struct mlx4_en_rss_context *) ptr;
1025         rss_context->base_qpn = cpu_to_be32(ilog2(rss_map->size) << 24 |
1026                                             (rss_map->base_qpn));
1027         rss_context->default_qpn = cpu_to_be32(rss_map->base_qpn);
1028         rss_context->hash_fn = rss_xor & 0x3;
1029         rss_context->flags = rss_mask << 2;
1030
1031         err = mlx4_qp_to_ready(mdev->dev, &priv->res.mtt, &context,
1032                                &rss_map->indir_qp, &rss_map->indir_state);
1033         if (err)
1034                 goto indir_err;
1035
1036         return 0;
1037
1038 indir_err:
1039         mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1040                        MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
1041         mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
1042         mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
1043 reserve_err:
1044         mlx4_qp_release_range(mdev->dev, priv->base_qpn, 1);
1045 rss_err:
1046         for (i = 0; i < good_qps; i++) {
1047                 mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1048                                MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1049                 mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1050                 mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1051         }
1052         mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, rss_map->size);
1053         return err;
1054 }
1055
1056 void mlx4_en_release_rss_steer(struct mlx4_en_priv *priv)
1057 {
1058         struct mlx4_en_dev *mdev = priv->mdev;
1059         struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1060         int i;
1061
1062         mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1063                        MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
1064         mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
1065         mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
1066         mlx4_qp_release_range(mdev->dev, priv->base_qpn, 1);
1067
1068         for (i = 0; i < rss_map->size; i++) {
1069                 mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1070                                MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1071                 mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1072                 mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1073         }
1074         mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, rss_map->size);
1075 }
1076
1077
1078
1079
1080