upload tizen1.0 source
[kernel/linux-2.6.36.git] / drivers / net / mlx4 / en_tx.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 <asm/page.h>
35 #include <linux/mlx4/cq.h>
36 #include <linux/slab.h>
37 #include <linux/mlx4/qp.h>
38 #include <linux/skbuff.h>
39 #include <linux/if_vlan.h>
40 #include <linux/vmalloc.h>
41
42 #include "mlx4_en.h"
43
44 enum {
45         MAX_INLINE = 104, /* 128 - 16 - 4 - 4 */
46 };
47
48 static int inline_thold __read_mostly = MAX_INLINE;
49
50 module_param_named(inline_thold, inline_thold, int, 0444);
51 MODULE_PARM_DESC(inline_thold, "threshold for using inline data");
52
53 int mlx4_en_create_tx_ring(struct mlx4_en_priv *priv,
54                            struct mlx4_en_tx_ring *ring, u32 size,
55                            u16 stride)
56 {
57         struct mlx4_en_dev *mdev = priv->mdev;
58         int tmp;
59         int err;
60
61         ring->size = size;
62         ring->size_mask = size - 1;
63         ring->stride = stride;
64
65         inline_thold = min(inline_thold, MAX_INLINE);
66
67         spin_lock_init(&ring->comp_lock);
68
69         tmp = size * sizeof(struct mlx4_en_tx_info);
70         ring->tx_info = vmalloc(tmp);
71         if (!ring->tx_info) {
72                 en_err(priv, "Failed allocating tx_info ring\n");
73                 return -ENOMEM;
74         }
75         en_dbg(DRV, priv, "Allocated tx_info ring at addr:%p size:%d\n",
76                  ring->tx_info, tmp);
77
78         ring->bounce_buf = kmalloc(MAX_DESC_SIZE, GFP_KERNEL);
79         if (!ring->bounce_buf) {
80                 en_err(priv, "Failed allocating bounce buffer\n");
81                 err = -ENOMEM;
82                 goto err_tx;
83         }
84         ring->buf_size = ALIGN(size * ring->stride, MLX4_EN_PAGE_SIZE);
85
86         err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres, ring->buf_size,
87                                  2 * PAGE_SIZE);
88         if (err) {
89                 en_err(priv, "Failed allocating hwq resources\n");
90                 goto err_bounce;
91         }
92
93         err = mlx4_en_map_buffer(&ring->wqres.buf);
94         if (err) {
95                 en_err(priv, "Failed to map TX buffer\n");
96                 goto err_hwq_res;
97         }
98
99         ring->buf = ring->wqres.buf.direct.buf;
100
101         en_dbg(DRV, priv, "Allocated TX ring (addr:%p) - buf:%p size:%d "
102                "buf_size:%d dma:%llx\n", ring, ring->buf, ring->size,
103                ring->buf_size, (unsigned long long) ring->wqres.buf.direct.map);
104
105         err = mlx4_qp_reserve_range(mdev->dev, 1, 1, &ring->qpn);
106         if (err) {
107                 en_err(priv, "Failed reserving qp for tx ring.\n");
108                 goto err_map;
109         }
110
111         err = mlx4_qp_alloc(mdev->dev, ring->qpn, &ring->qp);
112         if (err) {
113                 en_err(priv, "Failed allocating qp %d\n", ring->qpn);
114                 goto err_reserve;
115         }
116         ring->qp.event = mlx4_en_sqp_event;
117
118         return 0;
119
120 err_reserve:
121         mlx4_qp_release_range(mdev->dev, ring->qpn, 1);
122 err_map:
123         mlx4_en_unmap_buffer(&ring->wqres.buf);
124 err_hwq_res:
125         mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
126 err_bounce:
127         kfree(ring->bounce_buf);
128         ring->bounce_buf = NULL;
129 err_tx:
130         vfree(ring->tx_info);
131         ring->tx_info = NULL;
132         return err;
133 }
134
135 void mlx4_en_destroy_tx_ring(struct mlx4_en_priv *priv,
136                              struct mlx4_en_tx_ring *ring)
137 {
138         struct mlx4_en_dev *mdev = priv->mdev;
139         en_dbg(DRV, priv, "Destroying tx ring, qpn: %d\n", ring->qpn);
140
141         mlx4_qp_remove(mdev->dev, &ring->qp);
142         mlx4_qp_free(mdev->dev, &ring->qp);
143         mlx4_qp_release_range(mdev->dev, ring->qpn, 1);
144         mlx4_en_unmap_buffer(&ring->wqres.buf);
145         mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
146         kfree(ring->bounce_buf);
147         ring->bounce_buf = NULL;
148         vfree(ring->tx_info);
149         ring->tx_info = NULL;
150 }
151
152 int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv,
153                              struct mlx4_en_tx_ring *ring,
154                              int cq)
155 {
156         struct mlx4_en_dev *mdev = priv->mdev;
157         int err;
158
159         ring->cqn = cq;
160         ring->prod = 0;
161         ring->cons = 0xffffffff;
162         ring->last_nr_txbb = 1;
163         ring->poll_cnt = 0;
164         ring->blocked = 0;
165         memset(ring->tx_info, 0, ring->size * sizeof(struct mlx4_en_tx_info));
166         memset(ring->buf, 0, ring->buf_size);
167
168         ring->qp_state = MLX4_QP_STATE_RST;
169         ring->doorbell_qpn = swab32(ring->qp.qpn << 8);
170
171         mlx4_en_fill_qp_context(priv, ring->size, ring->stride, 1, 0, ring->qpn,
172                                 ring->cqn, &ring->context);
173
174         err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, &ring->context,
175                                &ring->qp, &ring->qp_state);
176
177         return err;
178 }
179
180 void mlx4_en_deactivate_tx_ring(struct mlx4_en_priv *priv,
181                                 struct mlx4_en_tx_ring *ring)
182 {
183         struct mlx4_en_dev *mdev = priv->mdev;
184
185         mlx4_qp_modify(mdev->dev, NULL, ring->qp_state,
186                        MLX4_QP_STATE_RST, NULL, 0, 0, &ring->qp);
187 }
188
189
190 static u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
191                                 struct mlx4_en_tx_ring *ring,
192                                 int index, u8 owner)
193 {
194         struct mlx4_en_dev *mdev = priv->mdev;
195         struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
196         struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
197         struct mlx4_wqe_data_seg *data = (void *) tx_desc + tx_info->data_offset;
198         struct sk_buff *skb = tx_info->skb;
199         struct skb_frag_struct *frag;
200         void *end = ring->buf + ring->buf_size;
201         int frags = skb_shinfo(skb)->nr_frags;
202         int i;
203         __be32 *ptr = (__be32 *)tx_desc;
204         __be32 stamp = cpu_to_be32(STAMP_VAL | (!!owner << STAMP_SHIFT));
205
206         /* Optimize the common case when there are no wraparounds */
207         if (likely((void *) tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
208                 if (!tx_info->inl) {
209                         if (tx_info->linear) {
210                                 pci_unmap_single(mdev->pdev,
211                                         (dma_addr_t) be64_to_cpu(data->addr),
212                                          be32_to_cpu(data->byte_count),
213                                          PCI_DMA_TODEVICE);
214                                 ++data;
215                         }
216
217                         for (i = 0; i < frags; i++) {
218                                 frag = &skb_shinfo(skb)->frags[i];
219                                 pci_unmap_page(mdev->pdev,
220                                         (dma_addr_t) be64_to_cpu(data[i].addr),
221                                         frag->size, PCI_DMA_TODEVICE);
222                         }
223                 }
224                 /* Stamp the freed descriptor */
225                 for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE; i += STAMP_STRIDE) {
226                         *ptr = stamp;
227                         ptr += STAMP_DWORDS;
228                 }
229
230         } else {
231                 if (!tx_info->inl) {
232                         if ((void *) data >= end) {
233                                 data = (struct mlx4_wqe_data_seg *)
234                                                 (ring->buf + ((void *) data - end));
235                         }
236
237                         if (tx_info->linear) {
238                                 pci_unmap_single(mdev->pdev,
239                                         (dma_addr_t) be64_to_cpu(data->addr),
240                                          be32_to_cpu(data->byte_count),
241                                          PCI_DMA_TODEVICE);
242                                 ++data;
243                         }
244
245                         for (i = 0; i < frags; i++) {
246                                 /* Check for wraparound before unmapping */
247                                 if ((void *) data >= end)
248                                         data = (struct mlx4_wqe_data_seg *) ring->buf;
249                                 frag = &skb_shinfo(skb)->frags[i];
250                                 pci_unmap_page(mdev->pdev,
251                                         (dma_addr_t) be64_to_cpu(data->addr),
252                                          frag->size, PCI_DMA_TODEVICE);
253                                 ++data;
254                         }
255                 }
256                 /* Stamp the freed descriptor */
257                 for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE; i += STAMP_STRIDE) {
258                         *ptr = stamp;
259                         ptr += STAMP_DWORDS;
260                         if ((void *) ptr >= end) {
261                                 ptr = ring->buf;
262                                 stamp ^= cpu_to_be32(0x80000000);
263                         }
264                 }
265
266         }
267         dev_kfree_skb_any(skb);
268         return tx_info->nr_txbb;
269 }
270
271
272 int mlx4_en_free_tx_buf(struct net_device *dev, struct mlx4_en_tx_ring *ring)
273 {
274         struct mlx4_en_priv *priv = netdev_priv(dev);
275         int cnt = 0;
276
277         /* Skip last polled descriptor */
278         ring->cons += ring->last_nr_txbb;
279         en_dbg(DRV, priv, "Freeing Tx buf - cons:0x%x prod:0x%x\n",
280                  ring->cons, ring->prod);
281
282         if ((u32) (ring->prod - ring->cons) > ring->size) {
283                 if (netif_msg_tx_err(priv))
284                         en_warn(priv, "Tx consumer passed producer!\n");
285                 return 0;
286         }
287
288         while (ring->cons != ring->prod) {
289                 ring->last_nr_txbb = mlx4_en_free_tx_desc(priv, ring,
290                                                 ring->cons & ring->size_mask,
291                                                 !!(ring->cons & ring->size));
292                 ring->cons += ring->last_nr_txbb;
293                 cnt++;
294         }
295
296         if (cnt)
297                 en_dbg(DRV, priv, "Freed %d uncompleted tx descriptors\n", cnt);
298
299         return cnt;
300 }
301
302
303 static void mlx4_en_process_tx_cq(struct net_device *dev, struct mlx4_en_cq *cq)
304 {
305         struct mlx4_en_priv *priv = netdev_priv(dev);
306         struct mlx4_cq *mcq = &cq->mcq;
307         struct mlx4_en_tx_ring *ring = &priv->tx_ring[cq->ring];
308         struct mlx4_cqe *cqe = cq->buf;
309         u16 index;
310         u16 new_index;
311         u32 txbbs_skipped = 0;
312         u32 cq_last_sav;
313
314         /* index always points to the first TXBB of the last polled descriptor */
315         index = ring->cons & ring->size_mask;
316         new_index = be16_to_cpu(cqe->wqe_index) & ring->size_mask;
317         if (index == new_index)
318                 return;
319
320         if (!priv->port_up)
321                 return;
322
323         /*
324          * We use a two-stage loop:
325          * - the first samples the HW-updated CQE
326          * - the second frees TXBBs until the last sample
327          * This lets us amortize CQE cache misses, while still polling the CQ
328          * until is quiescent.
329          */
330         cq_last_sav = mcq->cons_index;
331         do {
332                 do {
333                         /* Skip over last polled CQE */
334                         index = (index + ring->last_nr_txbb) & ring->size_mask;
335                         txbbs_skipped += ring->last_nr_txbb;
336
337                         /* Poll next CQE */
338                         ring->last_nr_txbb = mlx4_en_free_tx_desc(
339                                                 priv, ring, index,
340                                                 !!((ring->cons + txbbs_skipped) &
341                                                    ring->size));
342                         ++mcq->cons_index;
343
344                 } while (index != new_index);
345
346                 new_index = be16_to_cpu(cqe->wqe_index) & ring->size_mask;
347         } while (index != new_index);
348         AVG_PERF_COUNTER(priv->pstats.tx_coal_avg,
349                          (u32) (mcq->cons_index - cq_last_sav));
350
351         /*
352          * To prevent CQ overflow we first update CQ consumer and only then
353          * the ring consumer.
354          */
355         mlx4_cq_set_ci(mcq);
356         wmb();
357         ring->cons += txbbs_skipped;
358
359         /* Wakeup Tx queue if this ring stopped it */
360         if (unlikely(ring->blocked)) {
361                 if ((u32) (ring->prod - ring->cons) <=
362                      ring->size - HEADROOM - MAX_DESC_TXBBS) {
363                         ring->blocked = 0;
364                         netif_tx_wake_queue(netdev_get_tx_queue(dev, cq->ring));
365                         priv->port_stats.wake_queue++;
366                 }
367         }
368 }
369
370 void mlx4_en_tx_irq(struct mlx4_cq *mcq)
371 {
372         struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
373         struct mlx4_en_priv *priv = netdev_priv(cq->dev);
374         struct mlx4_en_tx_ring *ring = &priv->tx_ring[cq->ring];
375
376         if (!spin_trylock(&ring->comp_lock))
377                 return;
378         mlx4_en_process_tx_cq(cq->dev, cq);
379         mod_timer(&cq->timer, jiffies + 1);
380         spin_unlock(&ring->comp_lock);
381 }
382
383
384 void mlx4_en_poll_tx_cq(unsigned long data)
385 {
386         struct mlx4_en_cq *cq = (struct mlx4_en_cq *) data;
387         struct mlx4_en_priv *priv = netdev_priv(cq->dev);
388         struct mlx4_en_tx_ring *ring = &priv->tx_ring[cq->ring];
389         u32 inflight;
390
391         INC_PERF_COUNTER(priv->pstats.tx_poll);
392
393         if (!spin_trylock_irq(&ring->comp_lock)) {
394                 mod_timer(&cq->timer, jiffies + MLX4_EN_TX_POLL_TIMEOUT);
395                 return;
396         }
397         mlx4_en_process_tx_cq(cq->dev, cq);
398         inflight = (u32) (ring->prod - ring->cons - ring->last_nr_txbb);
399
400         /* If there are still packets in flight and the timer has not already
401          * been scheduled by the Tx routine then schedule it here to guarantee
402          * completion processing of these packets */
403         if (inflight && priv->port_up)
404                 mod_timer(&cq->timer, jiffies + MLX4_EN_TX_POLL_TIMEOUT);
405
406         spin_unlock_irq(&ring->comp_lock);
407 }
408
409 static struct mlx4_en_tx_desc *mlx4_en_bounce_to_desc(struct mlx4_en_priv *priv,
410                                                       struct mlx4_en_tx_ring *ring,
411                                                       u32 index,
412                                                       unsigned int desc_size)
413 {
414         u32 copy = (ring->size - index) * TXBB_SIZE;
415         int i;
416
417         for (i = desc_size - copy - 4; i >= 0; i -= 4) {
418                 if ((i & (TXBB_SIZE - 1)) == 0)
419                         wmb();
420
421                 *((u32 *) (ring->buf + i)) =
422                         *((u32 *) (ring->bounce_buf + copy + i));
423         }
424
425         for (i = copy - 4; i >= 4 ; i -= 4) {
426                 if ((i & (TXBB_SIZE - 1)) == 0)
427                         wmb();
428
429                 *((u32 *) (ring->buf + index * TXBB_SIZE + i)) =
430                         *((u32 *) (ring->bounce_buf + i));
431         }
432
433         /* Return real descriptor location */
434         return ring->buf + index * TXBB_SIZE;
435 }
436
437 static inline void mlx4_en_xmit_poll(struct mlx4_en_priv *priv, int tx_ind)
438 {
439         struct mlx4_en_cq *cq = &priv->tx_cq[tx_ind];
440         struct mlx4_en_tx_ring *ring = &priv->tx_ring[tx_ind];
441         unsigned long flags;
442
443         /* If we don't have a pending timer, set one up to catch our recent
444            post in case the interface becomes idle */
445         if (!timer_pending(&cq->timer))
446                 mod_timer(&cq->timer, jiffies + MLX4_EN_TX_POLL_TIMEOUT);
447
448         /* Poll the CQ every mlx4_en_TX_MODER_POLL packets */
449         if ((++ring->poll_cnt & (MLX4_EN_TX_POLL_MODER - 1)) == 0)
450                 if (spin_trylock_irqsave(&ring->comp_lock, flags)) {
451                         mlx4_en_process_tx_cq(priv->dev, cq);
452                         spin_unlock_irqrestore(&ring->comp_lock, flags);
453                 }
454 }
455
456 static void *get_frag_ptr(struct sk_buff *skb)
457 {
458         struct skb_frag_struct *frag =  &skb_shinfo(skb)->frags[0];
459         struct page *page = frag->page;
460         void *ptr;
461
462         ptr = page_address(page);
463         if (unlikely(!ptr))
464                 return NULL;
465
466         return ptr + frag->page_offset;
467 }
468
469 static int is_inline(struct sk_buff *skb, void **pfrag)
470 {
471         void *ptr;
472
473         if (inline_thold && !skb_is_gso(skb) && skb->len <= inline_thold) {
474                 if (skb_shinfo(skb)->nr_frags == 1) {
475                         ptr = get_frag_ptr(skb);
476                         if (unlikely(!ptr))
477                                 return 0;
478
479                         if (pfrag)
480                                 *pfrag = ptr;
481
482                         return 1;
483                 } else if (unlikely(skb_shinfo(skb)->nr_frags))
484                         return 0;
485                 else
486                         return 1;
487         }
488
489         return 0;
490 }
491
492 static int inline_size(struct sk_buff *skb)
493 {
494         if (skb->len + CTRL_SIZE + sizeof(struct mlx4_wqe_inline_seg)
495             <= MLX4_INLINE_ALIGN)
496                 return ALIGN(skb->len + CTRL_SIZE +
497                              sizeof(struct mlx4_wqe_inline_seg), 16);
498         else
499                 return ALIGN(skb->len + CTRL_SIZE + 2 *
500                              sizeof(struct mlx4_wqe_inline_seg), 16);
501 }
502
503 static int get_real_size(struct sk_buff *skb, struct net_device *dev,
504                          int *lso_header_size)
505 {
506         struct mlx4_en_priv *priv = netdev_priv(dev);
507         int real_size;
508
509         if (skb_is_gso(skb)) {
510                 *lso_header_size = skb_transport_offset(skb) + tcp_hdrlen(skb);
511                 real_size = CTRL_SIZE + skb_shinfo(skb)->nr_frags * DS_SIZE +
512                         ALIGN(*lso_header_size + 4, DS_SIZE);
513                 if (unlikely(*lso_header_size != skb_headlen(skb))) {
514                         /* We add a segment for the skb linear buffer only if
515                          * it contains data */
516                         if (*lso_header_size < skb_headlen(skb))
517                                 real_size += DS_SIZE;
518                         else {
519                                 if (netif_msg_tx_err(priv))
520                                         en_warn(priv, "Non-linear headers\n");
521                                 return 0;
522                         }
523                 }
524         } else {
525                 *lso_header_size = 0;
526                 if (!is_inline(skb, NULL))
527                         real_size = CTRL_SIZE + (skb_shinfo(skb)->nr_frags + 1) * DS_SIZE;
528                 else
529                         real_size = inline_size(skb);
530         }
531
532         return real_size;
533 }
534
535 static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc, struct sk_buff *skb,
536                              int real_size, u16 *vlan_tag, int tx_ind, void *fragptr)
537 {
538         struct mlx4_wqe_inline_seg *inl = &tx_desc->inl;
539         int spc = MLX4_INLINE_ALIGN - CTRL_SIZE - sizeof *inl;
540
541         if (skb->len <= spc) {
542                 inl->byte_count = cpu_to_be32(1 << 31 | skb->len);
543                 skb_copy_from_linear_data(skb, inl + 1, skb_headlen(skb));
544                 if (skb_shinfo(skb)->nr_frags)
545                         memcpy(((void *)(inl + 1)) + skb_headlen(skb), fragptr,
546                                skb_shinfo(skb)->frags[0].size);
547
548         } else {
549                 inl->byte_count = cpu_to_be32(1 << 31 | spc);
550                 if (skb_headlen(skb) <= spc) {
551                         skb_copy_from_linear_data(skb, inl + 1, skb_headlen(skb));
552                         if (skb_headlen(skb) < spc) {
553                                 memcpy(((void *)(inl + 1)) + skb_headlen(skb),
554                                         fragptr, spc - skb_headlen(skb));
555                                 fragptr +=  spc - skb_headlen(skb);
556                         }
557                         inl = (void *) (inl + 1) + spc;
558                         memcpy(((void *)(inl + 1)), fragptr, skb->len - spc);
559                 } else {
560                         skb_copy_from_linear_data(skb, inl + 1, spc);
561                         inl = (void *) (inl + 1) + spc;
562                         skb_copy_from_linear_data_offset(skb, spc, inl + 1,
563                                         skb_headlen(skb) - spc);
564                         if (skb_shinfo(skb)->nr_frags)
565                                 memcpy(((void *)(inl + 1)) + skb_headlen(skb) - spc,
566                                         fragptr, skb_shinfo(skb)->frags[0].size);
567                 }
568
569                 wmb();
570                 inl->byte_count = cpu_to_be32(1 << 31 | (skb->len - spc));
571         }
572         tx_desc->ctrl.vlan_tag = cpu_to_be16(*vlan_tag);
573         tx_desc->ctrl.ins_vlan = MLX4_WQE_CTRL_INS_VLAN * !!(*vlan_tag);
574         tx_desc->ctrl.fence_size = (real_size / 16) & 0x3f;
575 }
576
577 u16 mlx4_en_select_queue(struct net_device *dev, struct sk_buff *skb)
578 {
579         struct mlx4_en_priv *priv = netdev_priv(dev);
580         u16 vlan_tag = 0;
581
582         /* If we support per priority flow control and the packet contains
583          * a vlan tag, send the packet to the TX ring assigned to that priority
584          */
585         if (priv->prof->rx_ppp && priv->vlgrp && vlan_tx_tag_present(skb)) {
586                 vlan_tag = vlan_tx_tag_get(skb);
587                 return MLX4_EN_NUM_TX_RINGS + (vlan_tag >> 13);
588         }
589
590         return skb_tx_hash(dev, skb);
591 }
592
593 netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
594 {
595         struct mlx4_en_priv *priv = netdev_priv(dev);
596         struct mlx4_en_dev *mdev = priv->mdev;
597         struct mlx4_en_tx_ring *ring;
598         struct mlx4_en_cq *cq;
599         struct mlx4_en_tx_desc *tx_desc;
600         struct mlx4_wqe_data_seg *data;
601         struct skb_frag_struct *frag;
602         struct mlx4_en_tx_info *tx_info;
603         int tx_ind = 0;
604         int nr_txbb;
605         int desc_size;
606         int real_size;
607         dma_addr_t dma;
608         u32 index;
609         __be32 op_own;
610         u16 vlan_tag = 0;
611         int i;
612         int lso_header_size;
613         void *fragptr;
614
615         real_size = get_real_size(skb, dev, &lso_header_size);
616         if (unlikely(!real_size))
617                 goto tx_drop;
618
619         /* Allign descriptor to TXBB size */
620         desc_size = ALIGN(real_size, TXBB_SIZE);
621         nr_txbb = desc_size / TXBB_SIZE;
622         if (unlikely(nr_txbb > MAX_DESC_TXBBS)) {
623                 if (netif_msg_tx_err(priv))
624                         en_warn(priv, "Oversized header or SG list\n");
625                 goto tx_drop;
626         }
627
628         tx_ind = skb->queue_mapping;
629         ring = &priv->tx_ring[tx_ind];
630         if (priv->vlgrp && vlan_tx_tag_present(skb))
631                 vlan_tag = vlan_tx_tag_get(skb);
632
633         /* Check available TXBBs And 2K spare for prefetch */
634         if (unlikely(((int)(ring->prod - ring->cons)) >
635                      ring->size - HEADROOM - MAX_DESC_TXBBS)) {
636                 /* every full Tx ring stops queue */
637                 netif_tx_stop_queue(netdev_get_tx_queue(dev, tx_ind));
638                 ring->blocked = 1;
639                 priv->port_stats.queue_stopped++;
640
641                 /* Use interrupts to find out when queue opened */
642                 cq = &priv->tx_cq[tx_ind];
643                 mlx4_en_arm_cq(priv, cq);
644                 return NETDEV_TX_BUSY;
645         }
646
647         /* Track current inflight packets for performance analysis */
648         AVG_PERF_COUNTER(priv->pstats.inflight_avg,
649                          (u32) (ring->prod - ring->cons - 1));
650
651         /* Packet is good - grab an index and transmit it */
652         index = ring->prod & ring->size_mask;
653
654         /* See if we have enough space for whole descriptor TXBB for setting
655          * SW ownership on next descriptor; if not, use a bounce buffer. */
656         if (likely(index + nr_txbb <= ring->size))
657                 tx_desc = ring->buf + index * TXBB_SIZE;
658         else
659                 tx_desc = (struct mlx4_en_tx_desc *) ring->bounce_buf;
660
661         /* Save skb in tx_info ring */
662         tx_info = &ring->tx_info[index];
663         tx_info->skb = skb;
664         tx_info->nr_txbb = nr_txbb;
665
666         /* Prepare ctrl segement apart opcode+ownership, which depends on
667          * whether LSO is used */
668         tx_desc->ctrl.vlan_tag = cpu_to_be16(vlan_tag);
669         tx_desc->ctrl.ins_vlan = MLX4_WQE_CTRL_INS_VLAN * !!vlan_tag;
670         tx_desc->ctrl.fence_size = (real_size / 16) & 0x3f;
671         tx_desc->ctrl.srcrb_flags = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE |
672                                                 MLX4_WQE_CTRL_SOLICITED);
673         if (likely(skb->ip_summed == CHECKSUM_PARTIAL)) {
674                 tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
675                                                          MLX4_WQE_CTRL_TCP_UDP_CSUM);
676                 priv->port_stats.tx_chksum_offload++;
677         }
678
679         /* Handle LSO (TSO) packets */
680         if (lso_header_size) {
681                 /* Mark opcode as LSO */
682                 op_own = cpu_to_be32(MLX4_OPCODE_LSO | (1 << 6)) |
683                         ((ring->prod & ring->size) ?
684                                 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
685
686                 /* Fill in the LSO prefix */
687                 tx_desc->lso.mss_hdr_size = cpu_to_be32(
688                         skb_shinfo(skb)->gso_size << 16 | lso_header_size);
689
690                 /* Copy headers;
691                  * note that we already verified that it is linear */
692                 memcpy(tx_desc->lso.header, skb->data, lso_header_size);
693                 data = ((void *) &tx_desc->lso +
694                         ALIGN(lso_header_size + 4, DS_SIZE));
695
696                 priv->port_stats.tso_packets++;
697                 i = ((skb->len - lso_header_size) / skb_shinfo(skb)->gso_size) +
698                         !!((skb->len - lso_header_size) % skb_shinfo(skb)->gso_size);
699                 ring->bytes += skb->len + (i - 1) * lso_header_size;
700                 ring->packets += i;
701         } else {
702                 /* Normal (Non LSO) packet */
703                 op_own = cpu_to_be32(MLX4_OPCODE_SEND) |
704                         ((ring->prod & ring->size) ?
705                          cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
706                 data = &tx_desc->data;
707                 ring->bytes += max(skb->len, (unsigned int) ETH_ZLEN);
708                 ring->packets++;
709
710         }
711         AVG_PERF_COUNTER(priv->pstats.tx_pktsz_avg, skb->len);
712
713
714         /* valid only for none inline segments */
715         tx_info->data_offset = (void *) data - (void *) tx_desc;
716
717         tx_info->linear = (lso_header_size < skb_headlen(skb) && !is_inline(skb, NULL)) ? 1 : 0;
718         data += skb_shinfo(skb)->nr_frags + tx_info->linear - 1;
719
720         if (!is_inline(skb, &fragptr)) {
721                 /* Map fragments */
722                 for (i = skb_shinfo(skb)->nr_frags - 1; i >= 0; i--) {
723                         frag = &skb_shinfo(skb)->frags[i];
724                         dma = pci_map_page(mdev->dev->pdev, frag->page, frag->page_offset,
725                                            frag->size, PCI_DMA_TODEVICE);
726                         data->addr = cpu_to_be64(dma);
727                         data->lkey = cpu_to_be32(mdev->mr.key);
728                         wmb();
729                         data->byte_count = cpu_to_be32(frag->size);
730                         --data;
731                 }
732
733                 /* Map linear part */
734                 if (tx_info->linear) {
735                         dma = pci_map_single(mdev->dev->pdev, skb->data + lso_header_size,
736                                              skb_headlen(skb) - lso_header_size, PCI_DMA_TODEVICE);
737                         data->addr = cpu_to_be64(dma);
738                         data->lkey = cpu_to_be32(mdev->mr.key);
739                         wmb();
740                         data->byte_count = cpu_to_be32(skb_headlen(skb) - lso_header_size);
741                 }
742                 tx_info->inl = 0;
743         } else {
744                 build_inline_wqe(tx_desc, skb, real_size, &vlan_tag, tx_ind, fragptr);
745                 tx_info->inl = 1;
746         }
747
748         ring->prod += nr_txbb;
749
750         /* If we used a bounce buffer then copy descriptor back into place */
751         if (tx_desc == (struct mlx4_en_tx_desc *) ring->bounce_buf)
752                 tx_desc = mlx4_en_bounce_to_desc(priv, ring, index, desc_size);
753
754         /* Run destructor before passing skb to HW */
755         if (likely(!skb_shared(skb)))
756                 skb_orphan(skb);
757
758         /* Ensure new descirptor hits memory
759          * before setting ownership of this descriptor to HW */
760         wmb();
761         tx_desc->ctrl.owner_opcode = op_own;
762
763         /* Ring doorbell! */
764         wmb();
765         writel(ring->doorbell_qpn, mdev->uar_map + MLX4_SEND_DOORBELL);
766
767         /* Poll CQ here */
768         mlx4_en_xmit_poll(priv, tx_ind);
769
770         return NETDEV_TX_OK;
771
772 tx_drop:
773         dev_kfree_skb_any(skb);
774         priv->stats.tx_dropped++;
775         return NETDEV_TX_OK;
776 }
777