[SK_BUFF]: Introduce skb_copy_from_linear_data{_offset}
[linux-3.10.git] / drivers / net / chelsio / sge.c
1 /*****************************************************************************
2  *                                                                           *
3  * File: sge.c                                                               *
4  * $Revision: 1.26 $                                                         *
5  * $Date: 2005/06/21 18:29:48 $                                              *
6  * Description:                                                              *
7  *  DMA engine.                                                              *
8  *  part of the Chelsio 10Gb Ethernet Driver.                                *
9  *                                                                           *
10  * This program is free software; you can redistribute it and/or modify      *
11  * it under the terms of the GNU General Public License, version 2, as       *
12  * published by the Free Software Foundation.                                *
13  *                                                                           *
14  * You should have received a copy of the GNU General Public License along   *
15  * with this program; if not, write to the Free Software Foundation, Inc.,   *
16  * 59 Temple Place - Suite 330, Boston, MA  02111-1307, USA.                 *
17  *                                                                           *
18  * THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY EXPRESS OR IMPLIED    *
19  * WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED WARRANTIES OF      *
20  * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.                     *
21  *                                                                           *
22  * http://www.chelsio.com                                                    *
23  *                                                                           *
24  * Copyright (c) 2003 - 2005 Chelsio Communications, Inc.                    *
25  * All rights reserved.                                                      *
26  *                                                                           *
27  * Maintainers: maintainers@chelsio.com                                      *
28  *                                                                           *
29  * Authors: Dimitrios Michailidis   <dm@chelsio.com>                         *
30  *          Tina Yang               <tainay@chelsio.com>                     *
31  *          Felix Marti             <felix@chelsio.com>                      *
32  *          Scott Bardone           <sbardone@chelsio.com>                   *
33  *          Kurt Ottaway            <kottaway@chelsio.com>                   *
34  *          Frank DiMambro          <frank@chelsio.com>                      *
35  *                                                                           *
36  * History:                                                                  *
37  *                                                                           *
38  ****************************************************************************/
39
40 #include "common.h"
41
42 #include <linux/types.h>
43 #include <linux/errno.h>
44 #include <linux/pci.h>
45 #include <linux/ktime.h>
46 #include <linux/netdevice.h>
47 #include <linux/etherdevice.h>
48 #include <linux/if_vlan.h>
49 #include <linux/skbuff.h>
50 #include <linux/init.h>
51 #include <linux/mm.h>
52 #include <linux/tcp.h>
53 #include <linux/ip.h>
54 #include <linux/in.h>
55 #include <linux/if_arp.h>
56
57 #include "cpl5_cmd.h"
58 #include "sge.h"
59 #include "regs.h"
60 #include "espi.h"
61
62 /* This belongs in if_ether.h */
63 #define ETH_P_CPL5 0xf
64
65 #define SGE_CMDQ_N              2
66 #define SGE_FREELQ_N            2
67 #define SGE_CMDQ0_E_N           1024
68 #define SGE_CMDQ1_E_N           128
69 #define SGE_FREEL_SIZE          4096
70 #define SGE_JUMBO_FREEL_SIZE    512
71 #define SGE_FREEL_REFILL_THRESH 16
72 #define SGE_RESPQ_E_N           1024
73 #define SGE_INTRTIMER_NRES      1000
74 #define SGE_RX_SM_BUF_SIZE      1536
75 #define SGE_TX_DESC_MAX_PLEN    16384
76
77 #define SGE_RESPQ_REPLENISH_THRES (SGE_RESPQ_E_N / 4)
78
79 /*
80  * Period of the TX buffer reclaim timer.  This timer does not need to run
81  * frequently as TX buffers are usually reclaimed by new TX packets.
82  */
83 #define TX_RECLAIM_PERIOD (HZ / 4)
84
85 #define M_CMD_LEN       0x7fffffff
86 #define V_CMD_LEN(v)    (v)
87 #define G_CMD_LEN(v)    ((v) & M_CMD_LEN)
88 #define V_CMD_GEN1(v)   ((v) << 31)
89 #define V_CMD_GEN2(v)   (v)
90 #define F_CMD_DATAVALID (1 << 1)
91 #define F_CMD_SOP       (1 << 2)
92 #define V_CMD_EOP(v)    ((v) << 3)
93
94 /*
95  * Command queue, receive buffer list, and response queue descriptors.
96  */
97 #if defined(__BIG_ENDIAN_BITFIELD)
98 struct cmdQ_e {
99         u32 addr_lo;
100         u32 len_gen;
101         u32 flags;
102         u32 addr_hi;
103 };
104
105 struct freelQ_e {
106         u32 addr_lo;
107         u32 len_gen;
108         u32 gen2;
109         u32 addr_hi;
110 };
111
112 struct respQ_e {
113         u32 Qsleeping           : 4;
114         u32 Cmdq1CreditReturn   : 5;
115         u32 Cmdq1DmaComplete    : 5;
116         u32 Cmdq0CreditReturn   : 5;
117         u32 Cmdq0DmaComplete    : 5;
118         u32 FreelistQid         : 2;
119         u32 CreditValid         : 1;
120         u32 DataValid           : 1;
121         u32 Offload             : 1;
122         u32 Eop                 : 1;
123         u32 Sop                 : 1;
124         u32 GenerationBit       : 1;
125         u32 BufferLength;
126 };
127 #elif defined(__LITTLE_ENDIAN_BITFIELD)
128 struct cmdQ_e {
129         u32 len_gen;
130         u32 addr_lo;
131         u32 addr_hi;
132         u32 flags;
133 };
134
135 struct freelQ_e {
136         u32 len_gen;
137         u32 addr_lo;
138         u32 addr_hi;
139         u32 gen2;
140 };
141
142 struct respQ_e {
143         u32 BufferLength;
144         u32 GenerationBit       : 1;
145         u32 Sop                 : 1;
146         u32 Eop                 : 1;
147         u32 Offload             : 1;
148         u32 DataValid           : 1;
149         u32 CreditValid         : 1;
150         u32 FreelistQid         : 2;
151         u32 Cmdq0DmaComplete    : 5;
152         u32 Cmdq0CreditReturn   : 5;
153         u32 Cmdq1DmaComplete    : 5;
154         u32 Cmdq1CreditReturn   : 5;
155         u32 Qsleeping           : 4;
156 } ;
157 #endif
158
159 /*
160  * SW Context Command and Freelist Queue Descriptors
161  */
162 struct cmdQ_ce {
163         struct sk_buff *skb;
164         DECLARE_PCI_UNMAP_ADDR(dma_addr);
165         DECLARE_PCI_UNMAP_LEN(dma_len);
166 };
167
168 struct freelQ_ce {
169         struct sk_buff *skb;
170         DECLARE_PCI_UNMAP_ADDR(dma_addr);
171         DECLARE_PCI_UNMAP_LEN(dma_len);
172 };
173
174 /*
175  * SW command, freelist and response rings
176  */
177 struct cmdQ {
178         unsigned long   status;         /* HW DMA fetch status */
179         unsigned int    in_use;         /* # of in-use command descriptors */
180         unsigned int    size;           /* # of descriptors */
181         unsigned int    processed;      /* total # of descs HW has processed */
182         unsigned int    cleaned;        /* total # of descs SW has reclaimed */
183         unsigned int    stop_thres;     /* SW TX queue suspend threshold */
184         u16             pidx;           /* producer index (SW) */
185         u16             cidx;           /* consumer index (HW) */
186         u8              genbit;         /* current generation (=valid) bit */
187         u8              sop;            /* is next entry start of packet? */
188         struct cmdQ_e  *entries;        /* HW command descriptor Q */
189         struct cmdQ_ce *centries;       /* SW command context descriptor Q */
190         dma_addr_t      dma_addr;       /* DMA addr HW command descriptor Q */
191         spinlock_t      lock;           /* Lock to protect cmdQ enqueuing */
192 };
193
194 struct freelQ {
195         unsigned int    credits;        /* # of available RX buffers */
196         unsigned int    size;           /* free list capacity */
197         u16             pidx;           /* producer index (SW) */
198         u16             cidx;           /* consumer index (HW) */
199         u16             rx_buffer_size; /* Buffer size on this free list */
200         u16             dma_offset;     /* DMA offset to align IP headers */
201         u16             recycleq_idx;   /* skb recycle q to use */
202         u8              genbit;         /* current generation (=valid) bit */
203         struct freelQ_e *entries;       /* HW freelist descriptor Q */
204         struct freelQ_ce *centries;     /* SW freelist context descriptor Q */
205         dma_addr_t      dma_addr;       /* DMA addr HW freelist descriptor Q */
206 };
207
208 struct respQ {
209         unsigned int    credits;        /* credits to be returned to SGE */
210         unsigned int    size;           /* # of response Q descriptors */
211         u16             cidx;           /* consumer index (SW) */
212         u8              genbit;         /* current generation(=valid) bit */
213         struct respQ_e *entries;        /* HW response descriptor Q */
214         dma_addr_t      dma_addr;       /* DMA addr HW response descriptor Q */
215 };
216
217 /* Bit flags for cmdQ.status */
218 enum {
219         CMDQ_STAT_RUNNING = 1,          /* fetch engine is running */
220         CMDQ_STAT_LAST_PKT_DB = 2       /* last packet rung the doorbell */
221 };
222
223 /* T204 TX SW scheduler */
224
225 /* Per T204 TX port */
226 struct sched_port {
227         unsigned int    avail;          /* available bits - quota */
228         unsigned int    drain_bits_per_1024ns; /* drain rate */
229         unsigned int    speed;          /* drain rate, mbps */
230         unsigned int    mtu;            /* mtu size */
231         struct sk_buff_head skbq;       /* pending skbs */
232 };
233
234 /* Per T204 device */
235 struct sched {
236         ktime_t         last_updated;   /* last time quotas were computed */
237         unsigned int    max_avail;      /* max bits to be sent to any port */
238         unsigned int    port;           /* port index (round robin ports) */
239         unsigned int    num;            /* num skbs in per port queues */
240         struct sched_port p[MAX_NPORTS];
241         struct tasklet_struct sched_tsk;/* tasklet used to run scheduler */
242 };
243 static void restart_sched(unsigned long);
244
245
246 /*
247  * Main SGE data structure
248  *
249  * Interrupts are handled by a single CPU and it is likely that on a MP system
250  * the application is migrated to another CPU. In that scenario, we try to
251  * seperate the RX(in irq context) and TX state in order to decrease memory
252  * contention.
253  */
254 struct sge {
255         struct adapter *adapter;        /* adapter backpointer */
256         struct net_device *netdev;      /* netdevice backpointer */
257         struct freelQ   freelQ[SGE_FREELQ_N]; /* buffer free lists */
258         struct respQ    respQ;          /* response Q */
259         unsigned long   stopped_tx_queues; /* bitmap of suspended Tx queues */
260         unsigned int    rx_pkt_pad;     /* RX padding for L2 packets */
261         unsigned int    jumbo_fl;       /* jumbo freelist Q index */
262         unsigned int    intrtimer_nres; /* no-resource interrupt timer */
263         unsigned int    fixed_intrtimer;/* non-adaptive interrupt timer */
264         struct timer_list tx_reclaim_timer; /* reclaims TX buffers */
265         struct timer_list espibug_timer;
266         unsigned long   espibug_timeout;
267         struct sk_buff  *espibug_skb[MAX_NPORTS];
268         u32             sge_control;    /* shadow value of sge control reg */
269         struct sge_intr_counts stats;
270         struct sge_port_stats *port_stats[MAX_NPORTS];
271         struct sched    *tx_sched;
272         struct cmdQ cmdQ[SGE_CMDQ_N] ____cacheline_aligned_in_smp;
273 };
274
275 /*
276  * stop tasklet and free all pending skb's
277  */
278 static void tx_sched_stop(struct sge *sge)
279 {
280         struct sched *s = sge->tx_sched;
281         int i;
282
283         tasklet_kill(&s->sched_tsk);
284
285         for (i = 0; i < MAX_NPORTS; i++)
286                 __skb_queue_purge(&s->p[s->port].skbq);
287 }
288
289 /*
290  * t1_sched_update_parms() is called when the MTU or link speed changes. It
291  * re-computes scheduler parameters to scope with the change.
292  */
293 unsigned int t1_sched_update_parms(struct sge *sge, unsigned int port,
294                                    unsigned int mtu, unsigned int speed)
295 {
296         struct sched *s = sge->tx_sched;
297         struct sched_port *p = &s->p[port];
298         unsigned int max_avail_segs;
299
300         pr_debug("t1_sched_update_params mtu=%d speed=%d\n", mtu, speed);
301         if (speed)
302                 p->speed = speed;
303         if (mtu)
304                 p->mtu = mtu;
305
306         if (speed || mtu) {
307                 unsigned long long drain = 1024ULL * p->speed * (p->mtu - 40);
308                 do_div(drain, (p->mtu + 50) * 1000);
309                 p->drain_bits_per_1024ns = (unsigned int) drain;
310
311                 if (p->speed < 1000)
312                         p->drain_bits_per_1024ns =
313                                 90 * p->drain_bits_per_1024ns / 100;
314         }
315
316         if (board_info(sge->adapter)->board == CHBT_BOARD_CHT204) {
317                 p->drain_bits_per_1024ns -= 16;
318                 s->max_avail = max(4096U, p->mtu + 16 + 14 + 4);
319                 max_avail_segs = max(1U, 4096 / (p->mtu - 40));
320         } else {
321                 s->max_avail = 16384;
322                 max_avail_segs = max(1U, 9000 / (p->mtu - 40));
323         }
324
325         pr_debug("t1_sched_update_parms: mtu %u speed %u max_avail %u "
326                  "max_avail_segs %u drain_bits_per_1024ns %u\n", p->mtu,
327                  p->speed, s->max_avail, max_avail_segs,
328                  p->drain_bits_per_1024ns);
329
330         return max_avail_segs * (p->mtu - 40);
331 }
332
333 /*
334  * t1_sched_max_avail_bytes() tells the scheduler the maximum amount of
335  * data that can be pushed per port.
336  */
337 void t1_sched_set_max_avail_bytes(struct sge *sge, unsigned int val)
338 {
339         struct sched *s = sge->tx_sched;
340         unsigned int i;
341
342         s->max_avail = val;
343         for (i = 0; i < MAX_NPORTS; i++)
344                 t1_sched_update_parms(sge, i, 0, 0);
345 }
346
347 /*
348  * t1_sched_set_drain_bits_per_us() tells the scheduler at which rate a port
349  * is draining.
350  */
351 void t1_sched_set_drain_bits_per_us(struct sge *sge, unsigned int port,
352                                          unsigned int val)
353 {
354         struct sched *s = sge->tx_sched;
355         struct sched_port *p = &s->p[port];
356         p->drain_bits_per_1024ns = val * 1024 / 1000;
357         t1_sched_update_parms(sge, port, 0, 0);
358 }
359
360
361 /*
362  * get_clock() implements a ns clock (see ktime_get)
363  */
364 static inline ktime_t get_clock(void)
365 {
366         struct timespec ts;
367
368         ktime_get_ts(&ts);
369         return timespec_to_ktime(ts);
370 }
371
372 /*
373  * tx_sched_init() allocates resources and does basic initialization.
374  */
375 static int tx_sched_init(struct sge *sge)
376 {
377         struct sched *s;
378         int i;
379
380         s = kzalloc(sizeof (struct sched), GFP_KERNEL);
381         if (!s)
382                 return -ENOMEM;
383
384         pr_debug("tx_sched_init\n");
385         tasklet_init(&s->sched_tsk, restart_sched, (unsigned long) sge);
386         sge->tx_sched = s;
387
388         for (i = 0; i < MAX_NPORTS; i++) {
389                 skb_queue_head_init(&s->p[i].skbq);
390                 t1_sched_update_parms(sge, i, 1500, 1000);
391         }
392
393         return 0;
394 }
395
396 /*
397  * sched_update_avail() computes the delta since the last time it was called
398  * and updates the per port quota (number of bits that can be sent to the any
399  * port).
400  */
401 static inline int sched_update_avail(struct sge *sge)
402 {
403         struct sched *s = sge->tx_sched;
404         ktime_t now = get_clock();
405         unsigned int i;
406         long long delta_time_ns;
407
408         delta_time_ns = ktime_to_ns(ktime_sub(now, s->last_updated));
409
410         pr_debug("sched_update_avail delta=%lld\n", delta_time_ns);
411         if (delta_time_ns < 15000)
412                 return 0;
413
414         for (i = 0; i < MAX_NPORTS; i++) {
415                 struct sched_port *p = &s->p[i];
416                 unsigned int delta_avail;
417
418                 delta_avail = (p->drain_bits_per_1024ns * delta_time_ns) >> 13;
419                 p->avail = min(p->avail + delta_avail, s->max_avail);
420         }
421
422         s->last_updated = now;
423
424         return 1;
425 }
426
427 /*
428  * sched_skb() is called from two different places. In the tx path, any
429  * packet generating load on an output port will call sched_skb()
430  * (skb != NULL). In addition, sched_skb() is called from the irq/soft irq
431  * context (skb == NULL).
432  * The scheduler only returns a skb (which will then be sent) if the
433  * length of the skb is <= the current quota of the output port.
434  */
435 static struct sk_buff *sched_skb(struct sge *sge, struct sk_buff *skb,
436                                 unsigned int credits)
437 {
438         struct sched *s = sge->tx_sched;
439         struct sk_buff_head *skbq;
440         unsigned int i, len, update = 1;
441
442         pr_debug("sched_skb %p\n", skb);
443         if (!skb) {
444                 if (!s->num)
445                         return NULL;
446         } else {
447                 skbq = &s->p[skb->dev->if_port].skbq;
448                 __skb_queue_tail(skbq, skb);
449                 s->num++;
450                 skb = NULL;
451         }
452
453         if (credits < MAX_SKB_FRAGS + 1)
454                 goto out;
455
456 again:
457         for (i = 0; i < MAX_NPORTS; i++) {
458                 s->port = ++s->port & (MAX_NPORTS - 1);
459                 skbq = &s->p[s->port].skbq;
460
461                 skb = skb_peek(skbq);
462
463                 if (!skb)
464                         continue;
465
466                 len = skb->len;
467                 if (len <= s->p[s->port].avail) {
468                         s->p[s->port].avail -= len;
469                         s->num--;
470                         __skb_unlink(skb, skbq);
471                         goto out;
472                 }
473                 skb = NULL;
474         }
475
476         if (update-- && sched_update_avail(sge))
477                 goto again;
478
479 out:
480         /* If there are more pending skbs, we use the hardware to schedule us
481          * again.
482          */
483         if (s->num && !skb) {
484                 struct cmdQ *q = &sge->cmdQ[0];
485                 clear_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
486                 if (test_and_set_bit(CMDQ_STAT_RUNNING, &q->status) == 0) {
487                         set_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
488                         writel(F_CMDQ0_ENABLE, sge->adapter->regs + A_SG_DOORBELL);
489                 }
490         }
491         pr_debug("sched_skb ret %p\n", skb);
492
493         return skb;
494 }
495
496 /*
497  * PIO to indicate that memory mapped Q contains valid descriptor(s).
498  */
499 static inline void doorbell_pio(struct adapter *adapter, u32 val)
500 {
501         wmb();
502         writel(val, adapter->regs + A_SG_DOORBELL);
503 }
504
505 /*
506  * Frees all RX buffers on the freelist Q. The caller must make sure that
507  * the SGE is turned off before calling this function.
508  */
509 static void free_freelQ_buffers(struct pci_dev *pdev, struct freelQ *q)
510 {
511         unsigned int cidx = q->cidx;
512
513         while (q->credits--) {
514                 struct freelQ_ce *ce = &q->centries[cidx];
515
516                 pci_unmap_single(pdev, pci_unmap_addr(ce, dma_addr),
517                                  pci_unmap_len(ce, dma_len),
518                                  PCI_DMA_FROMDEVICE);
519                 dev_kfree_skb(ce->skb);
520                 ce->skb = NULL;
521                 if (++cidx == q->size)
522                         cidx = 0;
523         }
524 }
525
526 /*
527  * Free RX free list and response queue resources.
528  */
529 static void free_rx_resources(struct sge *sge)
530 {
531         struct pci_dev *pdev = sge->adapter->pdev;
532         unsigned int size, i;
533
534         if (sge->respQ.entries) {
535                 size = sizeof(struct respQ_e) * sge->respQ.size;
536                 pci_free_consistent(pdev, size, sge->respQ.entries,
537                                     sge->respQ.dma_addr);
538         }
539
540         for (i = 0; i < SGE_FREELQ_N; i++) {
541                 struct freelQ *q = &sge->freelQ[i];
542
543                 if (q->centries) {
544                         free_freelQ_buffers(pdev, q);
545                         kfree(q->centries);
546                 }
547                 if (q->entries) {
548                         size = sizeof(struct freelQ_e) * q->size;
549                         pci_free_consistent(pdev, size, q->entries,
550                                             q->dma_addr);
551                 }
552         }
553 }
554
555 /*
556  * Allocates basic RX resources, consisting of memory mapped freelist Qs and a
557  * response queue.
558  */
559 static int alloc_rx_resources(struct sge *sge, struct sge_params *p)
560 {
561         struct pci_dev *pdev = sge->adapter->pdev;
562         unsigned int size, i;
563
564         for (i = 0; i < SGE_FREELQ_N; i++) {
565                 struct freelQ *q = &sge->freelQ[i];
566
567                 q->genbit = 1;
568                 q->size = p->freelQ_size[i];
569                 q->dma_offset = sge->rx_pkt_pad ? 0 : NET_IP_ALIGN;
570                 size = sizeof(struct freelQ_e) * q->size;
571                 q->entries = pci_alloc_consistent(pdev, size, &q->dma_addr);
572                 if (!q->entries)
573                         goto err_no_mem;
574
575                 size = sizeof(struct freelQ_ce) * q->size;
576                 q->centries = kzalloc(size, GFP_KERNEL);
577                 if (!q->centries)
578                         goto err_no_mem;
579         }
580
581         /*
582          * Calculate the buffer sizes for the two free lists.  FL0 accommodates
583          * regular sized Ethernet frames, FL1 is sized not to exceed 16K,
584          * including all the sk_buff overhead.
585          *
586          * Note: For T2 FL0 and FL1 are reversed.
587          */
588         sge->freelQ[!sge->jumbo_fl].rx_buffer_size = SGE_RX_SM_BUF_SIZE +
589                 sizeof(struct cpl_rx_data) +
590                 sge->freelQ[!sge->jumbo_fl].dma_offset;
591
592                 size = (16 * 1024) -
593                     SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
594
595         sge->freelQ[sge->jumbo_fl].rx_buffer_size = size;
596
597         /*
598          * Setup which skb recycle Q should be used when recycling buffers from
599          * each free list.
600          */
601         sge->freelQ[!sge->jumbo_fl].recycleq_idx = 0;
602         sge->freelQ[sge->jumbo_fl].recycleq_idx = 1;
603
604         sge->respQ.genbit = 1;
605         sge->respQ.size = SGE_RESPQ_E_N;
606         sge->respQ.credits = 0;
607         size = sizeof(struct respQ_e) * sge->respQ.size;
608         sge->respQ.entries =
609                 pci_alloc_consistent(pdev, size, &sge->respQ.dma_addr);
610         if (!sge->respQ.entries)
611                 goto err_no_mem;
612         return 0;
613
614 err_no_mem:
615         free_rx_resources(sge);
616         return -ENOMEM;
617 }
618
619 /*
620  * Reclaims n TX descriptors and frees the buffers associated with them.
621  */
622 static void free_cmdQ_buffers(struct sge *sge, struct cmdQ *q, unsigned int n)
623 {
624         struct cmdQ_ce *ce;
625         struct pci_dev *pdev = sge->adapter->pdev;
626         unsigned int cidx = q->cidx;
627
628         q->in_use -= n;
629         ce = &q->centries[cidx];
630         while (n--) {
631                 if (likely(pci_unmap_len(ce, dma_len))) {
632                         pci_unmap_single(pdev, pci_unmap_addr(ce, dma_addr),
633                                          pci_unmap_len(ce, dma_len),
634                                          PCI_DMA_TODEVICE);
635                         if (q->sop)
636                                 q->sop = 0;
637                 }
638                 if (ce->skb) {
639                         dev_kfree_skb_any(ce->skb);
640                         q->sop = 1;
641                 }
642                 ce++;
643                 if (++cidx == q->size) {
644                         cidx = 0;
645                         ce = q->centries;
646                 }
647         }
648         q->cidx = cidx;
649 }
650
651 /*
652  * Free TX resources.
653  *
654  * Assumes that SGE is stopped and all interrupts are disabled.
655  */
656 static void free_tx_resources(struct sge *sge)
657 {
658         struct pci_dev *pdev = sge->adapter->pdev;
659         unsigned int size, i;
660
661         for (i = 0; i < SGE_CMDQ_N; i++) {
662                 struct cmdQ *q = &sge->cmdQ[i];
663
664                 if (q->centries) {
665                         if (q->in_use)
666                                 free_cmdQ_buffers(sge, q, q->in_use);
667                         kfree(q->centries);
668                 }
669                 if (q->entries) {
670                         size = sizeof(struct cmdQ_e) * q->size;
671                         pci_free_consistent(pdev, size, q->entries,
672                                             q->dma_addr);
673                 }
674         }
675 }
676
677 /*
678  * Allocates basic TX resources, consisting of memory mapped command Qs.
679  */
680 static int alloc_tx_resources(struct sge *sge, struct sge_params *p)
681 {
682         struct pci_dev *pdev = sge->adapter->pdev;
683         unsigned int size, i;
684
685         for (i = 0; i < SGE_CMDQ_N; i++) {
686                 struct cmdQ *q = &sge->cmdQ[i];
687
688                 q->genbit = 1;
689                 q->sop = 1;
690                 q->size = p->cmdQ_size[i];
691                 q->in_use = 0;
692                 q->status = 0;
693                 q->processed = q->cleaned = 0;
694                 q->stop_thres = 0;
695                 spin_lock_init(&q->lock);
696                 size = sizeof(struct cmdQ_e) * q->size;
697                 q->entries = pci_alloc_consistent(pdev, size, &q->dma_addr);
698                 if (!q->entries)
699                         goto err_no_mem;
700
701                 size = sizeof(struct cmdQ_ce) * q->size;
702                 q->centries = kzalloc(size, GFP_KERNEL);
703                 if (!q->centries)
704                         goto err_no_mem;
705         }
706
707         /*
708          * CommandQ 0 handles Ethernet and TOE packets, while queue 1 is TOE
709          * only.  For queue 0 set the stop threshold so we can handle one more
710          * packet from each port, plus reserve an additional 24 entries for
711          * Ethernet packets only.  Queue 1 never suspends nor do we reserve
712          * space for Ethernet packets.
713          */
714         sge->cmdQ[0].stop_thres = sge->adapter->params.nports *
715                 (MAX_SKB_FRAGS + 1);
716         return 0;
717
718 err_no_mem:
719         free_tx_resources(sge);
720         return -ENOMEM;
721 }
722
723 static inline void setup_ring_params(struct adapter *adapter, u64 addr,
724                                      u32 size, int base_reg_lo,
725                                      int base_reg_hi, int size_reg)
726 {
727         writel((u32)addr, adapter->regs + base_reg_lo);
728         writel(addr >> 32, adapter->regs + base_reg_hi);
729         writel(size, adapter->regs + size_reg);
730 }
731
732 /*
733  * Enable/disable VLAN acceleration.
734  */
735 void t1_set_vlan_accel(struct adapter *adapter, int on_off)
736 {
737         struct sge *sge = adapter->sge;
738
739         sge->sge_control &= ~F_VLAN_XTRACT;
740         if (on_off)
741                 sge->sge_control |= F_VLAN_XTRACT;
742         if (adapter->open_device_map) {
743                 writel(sge->sge_control, adapter->regs + A_SG_CONTROL);
744                 readl(adapter->regs + A_SG_CONTROL);   /* flush */
745         }
746 }
747
748 /*
749  * Programs the various SGE registers. However, the engine is not yet enabled,
750  * but sge->sge_control is setup and ready to go.
751  */
752 static void configure_sge(struct sge *sge, struct sge_params *p)
753 {
754         struct adapter *ap = sge->adapter;
755
756         writel(0, ap->regs + A_SG_CONTROL);
757         setup_ring_params(ap, sge->cmdQ[0].dma_addr, sge->cmdQ[0].size,
758                           A_SG_CMD0BASELWR, A_SG_CMD0BASEUPR, A_SG_CMD0SIZE);
759         setup_ring_params(ap, sge->cmdQ[1].dma_addr, sge->cmdQ[1].size,
760                           A_SG_CMD1BASELWR, A_SG_CMD1BASEUPR, A_SG_CMD1SIZE);
761         setup_ring_params(ap, sge->freelQ[0].dma_addr,
762                           sge->freelQ[0].size, A_SG_FL0BASELWR,
763                           A_SG_FL0BASEUPR, A_SG_FL0SIZE);
764         setup_ring_params(ap, sge->freelQ[1].dma_addr,
765                           sge->freelQ[1].size, A_SG_FL1BASELWR,
766                           A_SG_FL1BASEUPR, A_SG_FL1SIZE);
767
768         /* The threshold comparison uses <. */
769         writel(SGE_RX_SM_BUF_SIZE + 1, ap->regs + A_SG_FLTHRESHOLD);
770
771         setup_ring_params(ap, sge->respQ.dma_addr, sge->respQ.size,
772                           A_SG_RSPBASELWR, A_SG_RSPBASEUPR, A_SG_RSPSIZE);
773         writel((u32)sge->respQ.size - 1, ap->regs + A_SG_RSPQUEUECREDIT);
774
775         sge->sge_control = F_CMDQ0_ENABLE | F_CMDQ1_ENABLE | F_FL0_ENABLE |
776                 F_FL1_ENABLE | F_CPL_ENABLE | F_RESPONSE_QUEUE_ENABLE |
777                 V_CMDQ_PRIORITY(2) | F_DISABLE_CMDQ1_GTS | F_ISCSI_COALESCE |
778                 V_RX_PKT_OFFSET(sge->rx_pkt_pad);
779
780 #if defined(__BIG_ENDIAN_BITFIELD)
781         sge->sge_control |= F_ENABLE_BIG_ENDIAN;
782 #endif
783
784         /* Initialize no-resource timer */
785         sge->intrtimer_nres = SGE_INTRTIMER_NRES * core_ticks_per_usec(ap);
786
787         t1_sge_set_coalesce_params(sge, p);
788 }
789
790 /*
791  * Return the payload capacity of the jumbo free-list buffers.
792  */
793 static inline unsigned int jumbo_payload_capacity(const struct sge *sge)
794 {
795         return sge->freelQ[sge->jumbo_fl].rx_buffer_size -
796                 sge->freelQ[sge->jumbo_fl].dma_offset -
797                 sizeof(struct cpl_rx_data);
798 }
799
800 /*
801  * Frees all SGE related resources and the sge structure itself
802  */
803 void t1_sge_destroy(struct sge *sge)
804 {
805         int i;
806
807         for_each_port(sge->adapter, i)
808                 free_percpu(sge->port_stats[i]);
809
810         kfree(sge->tx_sched);
811         free_tx_resources(sge);
812         free_rx_resources(sge);
813         kfree(sge);
814 }
815
816 /*
817  * Allocates new RX buffers on the freelist Q (and tracks them on the freelist
818  * context Q) until the Q is full or alloc_skb fails.
819  *
820  * It is possible that the generation bits already match, indicating that the
821  * buffer is already valid and nothing needs to be done. This happens when we
822  * copied a received buffer into a new sk_buff during the interrupt processing.
823  *
824  * If the SGE doesn't automatically align packets properly (!sge->rx_pkt_pad),
825  * we specify a RX_OFFSET in order to make sure that the IP header is 4B
826  * aligned.
827  */
828 static void refill_free_list(struct sge *sge, struct freelQ *q)
829 {
830         struct pci_dev *pdev = sge->adapter->pdev;
831         struct freelQ_ce *ce = &q->centries[q->pidx];
832         struct freelQ_e *e = &q->entries[q->pidx];
833         unsigned int dma_len = q->rx_buffer_size - q->dma_offset;
834
835         while (q->credits < q->size) {
836                 struct sk_buff *skb;
837                 dma_addr_t mapping;
838
839                 skb = alloc_skb(q->rx_buffer_size, GFP_ATOMIC);
840                 if (!skb)
841                         break;
842
843                 skb_reserve(skb, q->dma_offset);
844                 mapping = pci_map_single(pdev, skb->data, dma_len,
845                                          PCI_DMA_FROMDEVICE);
846                 skb_reserve(skb, sge->rx_pkt_pad);
847
848                 ce->skb = skb;
849                 pci_unmap_addr_set(ce, dma_addr, mapping);
850                 pci_unmap_len_set(ce, dma_len, dma_len);
851                 e->addr_lo = (u32)mapping;
852                 e->addr_hi = (u64)mapping >> 32;
853                 e->len_gen = V_CMD_LEN(dma_len) | V_CMD_GEN1(q->genbit);
854                 wmb();
855                 e->gen2 = V_CMD_GEN2(q->genbit);
856
857                 e++;
858                 ce++;
859                 if (++q->pidx == q->size) {
860                         q->pidx = 0;
861                         q->genbit ^= 1;
862                         ce = q->centries;
863                         e = q->entries;
864                 }
865                 q->credits++;
866         }
867 }
868
869 /*
870  * Calls refill_free_list for both free lists. If we cannot fill at least 1/4
871  * of both rings, we go into 'few interrupt mode' in order to give the system
872  * time to free up resources.
873  */
874 static void freelQs_empty(struct sge *sge)
875 {
876         struct adapter *adapter = sge->adapter;
877         u32 irq_reg = readl(adapter->regs + A_SG_INT_ENABLE);
878         u32 irqholdoff_reg;
879
880         refill_free_list(sge, &sge->freelQ[0]);
881         refill_free_list(sge, &sge->freelQ[1]);
882
883         if (sge->freelQ[0].credits > (sge->freelQ[0].size >> 2) &&
884             sge->freelQ[1].credits > (sge->freelQ[1].size >> 2)) {
885                 irq_reg |= F_FL_EXHAUSTED;
886                 irqholdoff_reg = sge->fixed_intrtimer;
887         } else {
888                 /* Clear the F_FL_EXHAUSTED interrupts for now */
889                 irq_reg &= ~F_FL_EXHAUSTED;
890                 irqholdoff_reg = sge->intrtimer_nres;
891         }
892         writel(irqholdoff_reg, adapter->regs + A_SG_INTRTIMER);
893         writel(irq_reg, adapter->regs + A_SG_INT_ENABLE);
894
895         /* We reenable the Qs to force a freelist GTS interrupt later */
896         doorbell_pio(adapter, F_FL0_ENABLE | F_FL1_ENABLE);
897 }
898
899 #define SGE_PL_INTR_MASK (F_PL_INTR_SGE_ERR | F_PL_INTR_SGE_DATA)
900 #define SGE_INT_FATAL (F_RESPQ_OVERFLOW | F_PACKET_TOO_BIG | F_PACKET_MISMATCH)
901 #define SGE_INT_ENABLE (F_RESPQ_EXHAUSTED | F_RESPQ_OVERFLOW | \
902                         F_FL_EXHAUSTED | F_PACKET_TOO_BIG | F_PACKET_MISMATCH)
903
904 /*
905  * Disable SGE Interrupts
906  */
907 void t1_sge_intr_disable(struct sge *sge)
908 {
909         u32 val = readl(sge->adapter->regs + A_PL_ENABLE);
910
911         writel(val & ~SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_ENABLE);
912         writel(0, sge->adapter->regs + A_SG_INT_ENABLE);
913 }
914
915 /*
916  * Enable SGE interrupts.
917  */
918 void t1_sge_intr_enable(struct sge *sge)
919 {
920         u32 en = SGE_INT_ENABLE;
921         u32 val = readl(sge->adapter->regs + A_PL_ENABLE);
922
923         if (sge->adapter->flags & TSO_CAPABLE)
924                 en &= ~F_PACKET_TOO_BIG;
925         writel(en, sge->adapter->regs + A_SG_INT_ENABLE);
926         writel(val | SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_ENABLE);
927 }
928
929 /*
930  * Clear SGE interrupts.
931  */
932 void t1_sge_intr_clear(struct sge *sge)
933 {
934         writel(SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_CAUSE);
935         writel(0xffffffff, sge->adapter->regs + A_SG_INT_CAUSE);
936 }
937
938 /*
939  * SGE 'Error' interrupt handler
940  */
941 int t1_sge_intr_error_handler(struct sge *sge)
942 {
943         struct adapter *adapter = sge->adapter;
944         u32 cause = readl(adapter->regs + A_SG_INT_CAUSE);
945
946         if (adapter->flags & TSO_CAPABLE)
947                 cause &= ~F_PACKET_TOO_BIG;
948         if (cause & F_RESPQ_EXHAUSTED)
949                 sge->stats.respQ_empty++;
950         if (cause & F_RESPQ_OVERFLOW) {
951                 sge->stats.respQ_overflow++;
952                 CH_ALERT("%s: SGE response queue overflow\n",
953                          adapter->name);
954         }
955         if (cause & F_FL_EXHAUSTED) {
956                 sge->stats.freelistQ_empty++;
957                 freelQs_empty(sge);
958         }
959         if (cause & F_PACKET_TOO_BIG) {
960                 sge->stats.pkt_too_big++;
961                 CH_ALERT("%s: SGE max packet size exceeded\n",
962                          adapter->name);
963         }
964         if (cause & F_PACKET_MISMATCH) {
965                 sge->stats.pkt_mismatch++;
966                 CH_ALERT("%s: SGE packet mismatch\n", adapter->name);
967         }
968         if (cause & SGE_INT_FATAL)
969                 t1_fatal_err(adapter);
970
971         writel(cause, adapter->regs + A_SG_INT_CAUSE);
972         return 0;
973 }
974
975 const struct sge_intr_counts *t1_sge_get_intr_counts(const struct sge *sge)
976 {
977         return &sge->stats;
978 }
979
980 void t1_sge_get_port_stats(const struct sge *sge, int port,
981                            struct sge_port_stats *ss)
982 {
983         int cpu;
984
985         memset(ss, 0, sizeof(*ss));
986         for_each_possible_cpu(cpu) {
987                 struct sge_port_stats *st = per_cpu_ptr(sge->port_stats[port], cpu);
988
989                 ss->rx_packets += st->rx_packets;
990                 ss->rx_cso_good += st->rx_cso_good;
991                 ss->tx_packets += st->tx_packets;
992                 ss->tx_cso += st->tx_cso;
993                 ss->tx_tso += st->tx_tso;
994                 ss->vlan_xtract += st->vlan_xtract;
995                 ss->vlan_insert += st->vlan_insert;
996         }
997 }
998
999 /**
1000  *      recycle_fl_buf - recycle a free list buffer
1001  *      @fl: the free list
1002  *      @idx: index of buffer to recycle
1003  *
1004  *      Recycles the specified buffer on the given free list by adding it at
1005  *      the next available slot on the list.
1006  */
1007 static void recycle_fl_buf(struct freelQ *fl, int idx)
1008 {
1009         struct freelQ_e *from = &fl->entries[idx];
1010         struct freelQ_e *to = &fl->entries[fl->pidx];
1011
1012         fl->centries[fl->pidx] = fl->centries[idx];
1013         to->addr_lo = from->addr_lo;
1014         to->addr_hi = from->addr_hi;
1015         to->len_gen = G_CMD_LEN(from->len_gen) | V_CMD_GEN1(fl->genbit);
1016         wmb();
1017         to->gen2 = V_CMD_GEN2(fl->genbit);
1018         fl->credits++;
1019
1020         if (++fl->pidx == fl->size) {
1021                 fl->pidx = 0;
1022                 fl->genbit ^= 1;
1023         }
1024 }
1025
1026 static int copybreak __read_mostly = 256;
1027 module_param(copybreak, int, 0);
1028 MODULE_PARM_DESC(copybreak, "Receive copy threshold");
1029
1030 /**
1031  *      get_packet - return the next ingress packet buffer
1032  *      @pdev: the PCI device that received the packet
1033  *      @fl: the SGE free list holding the packet
1034  *      @len: the actual packet length, excluding any SGE padding
1035  *      @dma_pad: padding at beginning of buffer left by SGE DMA
1036  *      @skb_pad: padding to be used if the packet is copied
1037  *      @copy_thres: length threshold under which a packet should be copied
1038  *      @drop_thres: # of remaining buffers before we start dropping packets
1039  *
1040  *      Get the next packet from a free list and complete setup of the
1041  *      sk_buff.  If the packet is small we make a copy and recycle the
1042  *      original buffer, otherwise we use the original buffer itself.  If a
1043  *      positive drop threshold is supplied packets are dropped and their
1044  *      buffers recycled if (a) the number of remaining buffers is under the
1045  *      threshold and the packet is too big to copy, or (b) the packet should
1046  *      be copied but there is no memory for the copy.
1047  */
1048 static inline struct sk_buff *get_packet(struct pci_dev *pdev,
1049                                          struct freelQ *fl, unsigned int len)
1050 {
1051         struct sk_buff *skb;
1052         const struct freelQ_ce *ce = &fl->centries[fl->cidx];
1053
1054         if (len < copybreak) {
1055                 skb = alloc_skb(len + 2, GFP_ATOMIC);
1056                 if (!skb)
1057                         goto use_orig_buf;
1058
1059                 skb_reserve(skb, 2);    /* align IP header */
1060                 skb_put(skb, len);
1061                 pci_dma_sync_single_for_cpu(pdev,
1062                                             pci_unmap_addr(ce, dma_addr),
1063                                             pci_unmap_len(ce, dma_len),
1064                                             PCI_DMA_FROMDEVICE);
1065                 skb_copy_from_linear_data(ce->skb, skb->data, len);
1066                 pci_dma_sync_single_for_device(pdev,
1067                                                pci_unmap_addr(ce, dma_addr),
1068                                                pci_unmap_len(ce, dma_len),
1069                                                PCI_DMA_FROMDEVICE);
1070                 recycle_fl_buf(fl, fl->cidx);
1071                 return skb;
1072         }
1073
1074 use_orig_buf:
1075         if (fl->credits < 2) {
1076                 recycle_fl_buf(fl, fl->cidx);
1077                 return NULL;
1078         }
1079
1080         pci_unmap_single(pdev, pci_unmap_addr(ce, dma_addr),
1081                          pci_unmap_len(ce, dma_len), PCI_DMA_FROMDEVICE);
1082         skb = ce->skb;
1083         prefetch(skb->data);
1084
1085         skb_put(skb, len);
1086         return skb;
1087 }
1088
1089 /**
1090  *      unexpected_offload - handle an unexpected offload packet
1091  *      @adapter: the adapter
1092  *      @fl: the free list that received the packet
1093  *
1094  *      Called when we receive an unexpected offload packet (e.g., the TOE
1095  *      function is disabled or the card is a NIC).  Prints a message and
1096  *      recycles the buffer.
1097  */
1098 static void unexpected_offload(struct adapter *adapter, struct freelQ *fl)
1099 {
1100         struct freelQ_ce *ce = &fl->centries[fl->cidx];
1101         struct sk_buff *skb = ce->skb;
1102
1103         pci_dma_sync_single_for_cpu(adapter->pdev, pci_unmap_addr(ce, dma_addr),
1104                             pci_unmap_len(ce, dma_len), PCI_DMA_FROMDEVICE);
1105         CH_ERR("%s: unexpected offload packet, cmd %u\n",
1106                adapter->name, *skb->data);
1107         recycle_fl_buf(fl, fl->cidx);
1108 }
1109
1110 /*
1111  * T1/T2 SGE limits the maximum DMA size per TX descriptor to
1112  * SGE_TX_DESC_MAX_PLEN (16KB). If the PAGE_SIZE is larger than 16KB, the
1113  * stack might send more than SGE_TX_DESC_MAX_PLEN in a contiguous manner.
1114  * Note that the *_large_page_tx_descs stuff will be optimized out when
1115  * PAGE_SIZE <= SGE_TX_DESC_MAX_PLEN.
1116  *
1117  * compute_large_page_descs() computes how many additional descriptors are
1118  * required to break down the stack's request.
1119  */
1120 static inline unsigned int compute_large_page_tx_descs(struct sk_buff *skb)
1121 {
1122         unsigned int count = 0;
1123
1124         if (PAGE_SIZE > SGE_TX_DESC_MAX_PLEN) {
1125                 unsigned int nfrags = skb_shinfo(skb)->nr_frags;
1126                 unsigned int i, len = skb->len - skb->data_len;
1127                 while (len > SGE_TX_DESC_MAX_PLEN) {
1128                         count++;
1129                         len -= SGE_TX_DESC_MAX_PLEN;
1130                 }
1131                 for (i = 0; nfrags--; i++) {
1132                         skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
1133                         len = frag->size;
1134                         while (len > SGE_TX_DESC_MAX_PLEN) {
1135                                 count++;
1136                                 len -= SGE_TX_DESC_MAX_PLEN;
1137                         }
1138                 }
1139         }
1140         return count;
1141 }
1142
1143 /*
1144  * Write a cmdQ entry.
1145  *
1146  * Since this function writes the 'flags' field, it must not be used to
1147  * write the first cmdQ entry.
1148  */
1149 static inline void write_tx_desc(struct cmdQ_e *e, dma_addr_t mapping,
1150                                  unsigned int len, unsigned int gen,
1151                                  unsigned int eop)
1152 {
1153         if (unlikely(len > SGE_TX_DESC_MAX_PLEN))
1154                 BUG();
1155         e->addr_lo = (u32)mapping;
1156         e->addr_hi = (u64)mapping >> 32;
1157         e->len_gen = V_CMD_LEN(len) | V_CMD_GEN1(gen);
1158         e->flags = F_CMD_DATAVALID | V_CMD_EOP(eop) | V_CMD_GEN2(gen);
1159 }
1160
1161 /*
1162  * See comment for previous function.
1163  *
1164  * write_tx_descs_large_page() writes additional SGE tx descriptors if
1165  * *desc_len exceeds HW's capability.
1166  */
1167 static inline unsigned int write_large_page_tx_descs(unsigned int pidx,
1168                                                      struct cmdQ_e **e,
1169                                                      struct cmdQ_ce **ce,
1170                                                      unsigned int *gen,
1171                                                      dma_addr_t *desc_mapping,
1172                                                      unsigned int *desc_len,
1173                                                      unsigned int nfrags,
1174                                                      struct cmdQ *q)
1175 {
1176         if (PAGE_SIZE > SGE_TX_DESC_MAX_PLEN) {
1177                 struct cmdQ_e *e1 = *e;
1178                 struct cmdQ_ce *ce1 = *ce;
1179
1180                 while (*desc_len > SGE_TX_DESC_MAX_PLEN) {
1181                         *desc_len -= SGE_TX_DESC_MAX_PLEN;
1182                         write_tx_desc(e1, *desc_mapping, SGE_TX_DESC_MAX_PLEN,
1183                                       *gen, nfrags == 0 && *desc_len == 0);
1184                         ce1->skb = NULL;
1185                         pci_unmap_len_set(ce1, dma_len, 0);
1186                         *desc_mapping += SGE_TX_DESC_MAX_PLEN;
1187                         if (*desc_len) {
1188                                 ce1++;
1189                                 e1++;
1190                                 if (++pidx == q->size) {
1191                                         pidx = 0;
1192                                         *gen ^= 1;
1193                                         ce1 = q->centries;
1194                                         e1 = q->entries;
1195                                 }
1196                         }
1197                 }
1198                 *e = e1;
1199                 *ce = ce1;
1200         }
1201         return pidx;
1202 }
1203
1204 /*
1205  * Write the command descriptors to transmit the given skb starting at
1206  * descriptor pidx with the given generation.
1207  */
1208 static inline void write_tx_descs(struct adapter *adapter, struct sk_buff *skb,
1209                                   unsigned int pidx, unsigned int gen,
1210                                   struct cmdQ *q)
1211 {
1212         dma_addr_t mapping, desc_mapping;
1213         struct cmdQ_e *e, *e1;
1214         struct cmdQ_ce *ce;
1215         unsigned int i, flags, first_desc_len, desc_len,
1216             nfrags = skb_shinfo(skb)->nr_frags;
1217
1218         e = e1 = &q->entries[pidx];
1219         ce = &q->centries[pidx];
1220
1221         mapping = pci_map_single(adapter->pdev, skb->data,
1222                                 skb->len - skb->data_len, PCI_DMA_TODEVICE);
1223
1224         desc_mapping = mapping;
1225         desc_len = skb->len - skb->data_len;
1226
1227         flags = F_CMD_DATAVALID | F_CMD_SOP |
1228             V_CMD_EOP(nfrags == 0 && desc_len <= SGE_TX_DESC_MAX_PLEN) |
1229             V_CMD_GEN2(gen);
1230         first_desc_len = (desc_len <= SGE_TX_DESC_MAX_PLEN) ?
1231             desc_len : SGE_TX_DESC_MAX_PLEN;
1232         e->addr_lo = (u32)desc_mapping;
1233         e->addr_hi = (u64)desc_mapping >> 32;
1234         e->len_gen = V_CMD_LEN(first_desc_len) | V_CMD_GEN1(gen);
1235         ce->skb = NULL;
1236         pci_unmap_len_set(ce, dma_len, 0);
1237
1238         if (PAGE_SIZE > SGE_TX_DESC_MAX_PLEN &&
1239             desc_len > SGE_TX_DESC_MAX_PLEN) {
1240                 desc_mapping += first_desc_len;
1241                 desc_len -= first_desc_len;
1242                 e1++;
1243                 ce++;
1244                 if (++pidx == q->size) {
1245                         pidx = 0;
1246                         gen ^= 1;
1247                         e1 = q->entries;
1248                         ce = q->centries;
1249                 }
1250                 pidx = write_large_page_tx_descs(pidx, &e1, &ce, &gen,
1251                                                  &desc_mapping, &desc_len,
1252                                                  nfrags, q);
1253
1254                 if (likely(desc_len))
1255                         write_tx_desc(e1, desc_mapping, desc_len, gen,
1256                                       nfrags == 0);
1257         }
1258
1259         ce->skb = NULL;
1260         pci_unmap_addr_set(ce, dma_addr, mapping);
1261         pci_unmap_len_set(ce, dma_len, skb->len - skb->data_len);
1262
1263         for (i = 0; nfrags--; i++) {
1264                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
1265                 e1++;
1266                 ce++;
1267                 if (++pidx == q->size) {
1268                         pidx = 0;
1269                         gen ^= 1;
1270                         e1 = q->entries;
1271                         ce = q->centries;
1272                 }
1273
1274                 mapping = pci_map_page(adapter->pdev, frag->page,
1275                                        frag->page_offset, frag->size,
1276                                        PCI_DMA_TODEVICE);
1277                 desc_mapping = mapping;
1278                 desc_len = frag->size;
1279
1280                 pidx = write_large_page_tx_descs(pidx, &e1, &ce, &gen,
1281                                                  &desc_mapping, &desc_len,
1282                                                  nfrags, q);
1283                 if (likely(desc_len))
1284                         write_tx_desc(e1, desc_mapping, desc_len, gen,
1285                                       nfrags == 0);
1286                 ce->skb = NULL;
1287                 pci_unmap_addr_set(ce, dma_addr, mapping);
1288                 pci_unmap_len_set(ce, dma_len, frag->size);
1289         }
1290         ce->skb = skb;
1291         wmb();
1292         e->flags = flags;
1293 }
1294
1295 /*
1296  * Clean up completed Tx buffers.
1297  */
1298 static inline void reclaim_completed_tx(struct sge *sge, struct cmdQ *q)
1299 {
1300         unsigned int reclaim = q->processed - q->cleaned;
1301
1302         if (reclaim) {
1303                 pr_debug("reclaim_completed_tx processed:%d cleaned:%d\n",
1304                          q->processed, q->cleaned);
1305                 free_cmdQ_buffers(sge, q, reclaim);
1306                 q->cleaned += reclaim;
1307         }
1308 }
1309
1310 /*
1311  * Called from tasklet. Checks the scheduler for any
1312  * pending skbs that can be sent.
1313  */
1314 static void restart_sched(unsigned long arg)
1315 {
1316         struct sge *sge = (struct sge *) arg;
1317         struct adapter *adapter = sge->adapter;
1318         struct cmdQ *q = &sge->cmdQ[0];
1319         struct sk_buff *skb;
1320         unsigned int credits, queued_skb = 0;
1321
1322         spin_lock(&q->lock);
1323         reclaim_completed_tx(sge, q);
1324
1325         credits = q->size - q->in_use;
1326         pr_debug("restart_sched credits=%d\n", credits);
1327         while ((skb = sched_skb(sge, NULL, credits)) != NULL) {
1328                 unsigned int genbit, pidx, count;
1329                 count = 1 + skb_shinfo(skb)->nr_frags;
1330                 count += compute_large_page_tx_descs(skb);
1331                 q->in_use += count;
1332                 genbit = q->genbit;
1333                 pidx = q->pidx;
1334                 q->pidx += count;
1335                 if (q->pidx >= q->size) {
1336                         q->pidx -= q->size;
1337                         q->genbit ^= 1;
1338                 }
1339                 write_tx_descs(adapter, skb, pidx, genbit, q);
1340                 credits = q->size - q->in_use;
1341                 queued_skb = 1;
1342         }
1343
1344         if (queued_skb) {
1345                 clear_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1346                 if (test_and_set_bit(CMDQ_STAT_RUNNING, &q->status) == 0) {
1347                         set_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1348                         writel(F_CMDQ0_ENABLE, adapter->regs + A_SG_DOORBELL);
1349                 }
1350         }
1351         spin_unlock(&q->lock);
1352 }
1353
1354 /**
1355  *      sge_rx - process an ingress ethernet packet
1356  *      @sge: the sge structure
1357  *      @fl: the free list that contains the packet buffer
1358  *      @len: the packet length
1359  *
1360  *      Process an ingress ethernet pakcet and deliver it to the stack.
1361  */
1362 static void sge_rx(struct sge *sge, struct freelQ *fl, unsigned int len)
1363 {
1364         struct sk_buff *skb;
1365         const struct cpl_rx_pkt *p;
1366         struct adapter *adapter = sge->adapter;
1367         struct sge_port_stats *st;
1368
1369         skb = get_packet(adapter->pdev, fl, len - sge->rx_pkt_pad);
1370         if (unlikely(!skb)) {
1371                 sge->stats.rx_drops++;
1372                 return;
1373         }
1374
1375         p = (const struct cpl_rx_pkt *) skb->data;
1376         if (p->iff >= adapter->params.nports) {
1377                 kfree_skb(skb);
1378                 return;
1379         }
1380         __skb_pull(skb, sizeof(*p));
1381
1382         skb->dev->last_rx = jiffies;
1383         st = per_cpu_ptr(sge->port_stats[p->iff], smp_processor_id());
1384         st->rx_packets++;
1385
1386         skb->protocol = eth_type_trans(skb, adapter->port[p->iff].dev);
1387         if ((adapter->flags & RX_CSUM_ENABLED) && p->csum == 0xffff &&
1388             skb->protocol == htons(ETH_P_IP) &&
1389             (skb->data[9] == IPPROTO_TCP || skb->data[9] == IPPROTO_UDP)) {
1390                 ++st->rx_cso_good;
1391                 skb->ip_summed = CHECKSUM_UNNECESSARY;
1392         } else
1393                 skb->ip_summed = CHECKSUM_NONE;
1394
1395         if (unlikely(adapter->vlan_grp && p->vlan_valid)) {
1396                 st->vlan_xtract++;
1397 #ifdef CONFIG_CHELSIO_T1_NAPI
1398                         vlan_hwaccel_receive_skb(skb, adapter->vlan_grp,
1399                                                  ntohs(p->vlan));
1400 #else
1401                         vlan_hwaccel_rx(skb, adapter->vlan_grp,
1402                                         ntohs(p->vlan));
1403 #endif
1404         } else {
1405 #ifdef CONFIG_CHELSIO_T1_NAPI
1406                 netif_receive_skb(skb);
1407 #else
1408                 netif_rx(skb);
1409 #endif
1410         }
1411 }
1412
1413 /*
1414  * Returns true if a command queue has enough available descriptors that
1415  * we can resume Tx operation after temporarily disabling its packet queue.
1416  */
1417 static inline int enough_free_Tx_descs(const struct cmdQ *q)
1418 {
1419         unsigned int r = q->processed - q->cleaned;
1420
1421         return q->in_use - r < (q->size >> 1);
1422 }
1423
1424 /*
1425  * Called when sufficient space has become available in the SGE command queues
1426  * after the Tx packet schedulers have been suspended to restart the Tx path.
1427  */
1428 static void restart_tx_queues(struct sge *sge)
1429 {
1430         struct adapter *adap = sge->adapter;
1431         int i;
1432
1433         if (!enough_free_Tx_descs(&sge->cmdQ[0]))
1434                 return;
1435
1436         for_each_port(adap, i) {
1437                 struct net_device *nd = adap->port[i].dev;
1438
1439                 if (test_and_clear_bit(nd->if_port, &sge->stopped_tx_queues) &&
1440                     netif_running(nd)) {
1441                         sge->stats.cmdQ_restarted[2]++;
1442                         netif_wake_queue(nd);
1443                 }
1444         }
1445 }
1446
1447 /*
1448  * update_tx_info is called from the interrupt handler/NAPI to return cmdQ0
1449  * information.
1450  */
1451 static unsigned int update_tx_info(struct adapter *adapter,
1452                                           unsigned int flags,
1453                                           unsigned int pr0)
1454 {
1455         struct sge *sge = adapter->sge;
1456         struct cmdQ *cmdq = &sge->cmdQ[0];
1457
1458         cmdq->processed += pr0;
1459         if (flags & (F_FL0_ENABLE | F_FL1_ENABLE)) {
1460                 freelQs_empty(sge);
1461                 flags &= ~(F_FL0_ENABLE | F_FL1_ENABLE);
1462         }
1463         if (flags & F_CMDQ0_ENABLE) {
1464                 clear_bit(CMDQ_STAT_RUNNING, &cmdq->status);
1465
1466                 if (cmdq->cleaned + cmdq->in_use != cmdq->processed &&
1467                     !test_and_set_bit(CMDQ_STAT_LAST_PKT_DB, &cmdq->status)) {
1468                         set_bit(CMDQ_STAT_RUNNING, &cmdq->status);
1469                         writel(F_CMDQ0_ENABLE, adapter->regs + A_SG_DOORBELL);
1470                 }
1471                 if (sge->tx_sched)
1472                         tasklet_hi_schedule(&sge->tx_sched->sched_tsk);
1473
1474                 flags &= ~F_CMDQ0_ENABLE;
1475         }
1476
1477         if (unlikely(sge->stopped_tx_queues != 0))
1478                 restart_tx_queues(sge);
1479
1480         return flags;
1481 }
1482
1483 /*
1484  * Process SGE responses, up to the supplied budget.  Returns the number of
1485  * responses processed.  A negative budget is effectively unlimited.
1486  */
1487 static int process_responses(struct adapter *adapter, int budget)
1488 {
1489         struct sge *sge = adapter->sge;
1490         struct respQ *q = &sge->respQ;
1491         struct respQ_e *e = &q->entries[q->cidx];
1492         int done = 0;
1493         unsigned int flags = 0;
1494         unsigned int cmdq_processed[SGE_CMDQ_N] = {0, 0};
1495
1496         while (done < budget && e->GenerationBit == q->genbit) {
1497                 flags |= e->Qsleeping;
1498
1499                 cmdq_processed[0] += e->Cmdq0CreditReturn;
1500                 cmdq_processed[1] += e->Cmdq1CreditReturn;
1501
1502                 /* We batch updates to the TX side to avoid cacheline
1503                  * ping-pong of TX state information on MP where the sender
1504                  * might run on a different CPU than this function...
1505                  */
1506                 if (unlikely((flags & F_CMDQ0_ENABLE) || cmdq_processed[0] > 64)) {
1507                         flags = update_tx_info(adapter, flags, cmdq_processed[0]);
1508                         cmdq_processed[0] = 0;
1509                 }
1510
1511                 if (unlikely(cmdq_processed[1] > 16)) {
1512                         sge->cmdQ[1].processed += cmdq_processed[1];
1513                         cmdq_processed[1] = 0;
1514                 }
1515
1516                 if (likely(e->DataValid)) {
1517                         struct freelQ *fl = &sge->freelQ[e->FreelistQid];
1518
1519                         BUG_ON(!e->Sop || !e->Eop);
1520                         if (unlikely(e->Offload))
1521                                 unexpected_offload(adapter, fl);
1522                         else
1523                                 sge_rx(sge, fl, e->BufferLength);
1524
1525                         ++done;
1526
1527                         /*
1528                          * Note: this depends on each packet consuming a
1529                          * single free-list buffer; cf. the BUG above.
1530                          */
1531                         if (++fl->cidx == fl->size)
1532                                 fl->cidx = 0;
1533                         prefetch(fl->centries[fl->cidx].skb);
1534
1535                         if (unlikely(--fl->credits <
1536                                      fl->size - SGE_FREEL_REFILL_THRESH))
1537                                 refill_free_list(sge, fl);
1538                 } else
1539                         sge->stats.pure_rsps++;
1540
1541                 e++;
1542                 if (unlikely(++q->cidx == q->size)) {
1543                         q->cidx = 0;
1544                         q->genbit ^= 1;
1545                         e = q->entries;
1546                 }
1547                 prefetch(e);
1548
1549                 if (++q->credits > SGE_RESPQ_REPLENISH_THRES) {
1550                         writel(q->credits, adapter->regs + A_SG_RSPQUEUECREDIT);
1551                         q->credits = 0;
1552                 }
1553         }
1554
1555         flags = update_tx_info(adapter, flags, cmdq_processed[0]);
1556         sge->cmdQ[1].processed += cmdq_processed[1];
1557
1558         return done;
1559 }
1560
1561 static inline int responses_pending(const struct adapter *adapter)
1562 {
1563         const struct respQ *Q = &adapter->sge->respQ;
1564         const struct respQ_e *e = &Q->entries[Q->cidx];
1565
1566         return (e->GenerationBit == Q->genbit);
1567 }
1568
1569 #ifdef CONFIG_CHELSIO_T1_NAPI
1570 /*
1571  * A simpler version of process_responses() that handles only pure (i.e.,
1572  * non data-carrying) responses.  Such respones are too light-weight to justify
1573  * calling a softirq when using NAPI, so we handle them specially in hard
1574  * interrupt context.  The function is called with a pointer to a response,
1575  * which the caller must ensure is a valid pure response.  Returns 1 if it
1576  * encounters a valid data-carrying response, 0 otherwise.
1577  */
1578 static int process_pure_responses(struct adapter *adapter)
1579 {
1580         struct sge *sge = adapter->sge;
1581         struct respQ *q = &sge->respQ;
1582         struct respQ_e *e = &q->entries[q->cidx];
1583         const struct freelQ *fl = &sge->freelQ[e->FreelistQid];
1584         unsigned int flags = 0;
1585         unsigned int cmdq_processed[SGE_CMDQ_N] = {0, 0};
1586
1587         prefetch(fl->centries[fl->cidx].skb);
1588         if (e->DataValid)
1589                 return 1;
1590
1591         do {
1592                 flags |= e->Qsleeping;
1593
1594                 cmdq_processed[0] += e->Cmdq0CreditReturn;
1595                 cmdq_processed[1] += e->Cmdq1CreditReturn;
1596
1597                 e++;
1598                 if (unlikely(++q->cidx == q->size)) {
1599                         q->cidx = 0;
1600                         q->genbit ^= 1;
1601                         e = q->entries;
1602                 }
1603                 prefetch(e);
1604
1605                 if (++q->credits > SGE_RESPQ_REPLENISH_THRES) {
1606                         writel(q->credits, adapter->regs + A_SG_RSPQUEUECREDIT);
1607                         q->credits = 0;
1608                 }
1609                 sge->stats.pure_rsps++;
1610         } while (e->GenerationBit == q->genbit && !e->DataValid);
1611
1612         flags = update_tx_info(adapter, flags, cmdq_processed[0]);
1613         sge->cmdQ[1].processed += cmdq_processed[1];
1614
1615         return e->GenerationBit == q->genbit;
1616 }
1617
1618 /*
1619  * Handler for new data events when using NAPI.  This does not need any locking
1620  * or protection from interrupts as data interrupts are off at this point and
1621  * other adapter interrupts do not interfere.
1622  */
1623 int t1_poll(struct net_device *dev, int *budget)
1624 {
1625         struct adapter *adapter = dev->priv;
1626         int work_done;
1627
1628         work_done = process_responses(adapter, min(*budget, dev->quota));
1629         *budget -= work_done;
1630         dev->quota -= work_done;
1631
1632         if (unlikely(responses_pending(adapter)))
1633                 return 1;
1634
1635         netif_rx_complete(dev);
1636         writel(adapter->sge->respQ.cidx, adapter->regs + A_SG_SLEEPING);
1637
1638         return 0;
1639
1640 }
1641
1642 /*
1643  * NAPI version of the main interrupt handler.
1644  */
1645 irqreturn_t t1_interrupt(int irq, void *data)
1646 {
1647         struct adapter *adapter = data;
1648         struct sge *sge = adapter->sge;
1649         int handled;
1650
1651         if (likely(responses_pending(adapter))) {
1652                 struct net_device *dev = sge->netdev;
1653
1654                 writel(F_PL_INTR_SGE_DATA, adapter->regs + A_PL_CAUSE);
1655
1656                 if (__netif_rx_schedule_prep(dev)) {
1657                         if (process_pure_responses(adapter))
1658                                 __netif_rx_schedule(dev);
1659                         else {
1660                                 /* no data, no NAPI needed */
1661                                 writel(sge->respQ.cidx, adapter->regs + A_SG_SLEEPING);
1662                                 netif_poll_enable(dev); /* undo schedule_prep */
1663                         }
1664                 }
1665                 return IRQ_HANDLED;
1666         }
1667
1668         spin_lock(&adapter->async_lock);
1669         handled = t1_slow_intr_handler(adapter);
1670         spin_unlock(&adapter->async_lock);
1671
1672         if (!handled)
1673                 sge->stats.unhandled_irqs++;
1674
1675         return IRQ_RETVAL(handled != 0);
1676 }
1677
1678 #else
1679 /*
1680  * Main interrupt handler, optimized assuming that we took a 'DATA'
1681  * interrupt.
1682  *
1683  * 1. Clear the interrupt
1684  * 2. Loop while we find valid descriptors and process them; accumulate
1685  *      information that can be processed after the loop
1686  * 3. Tell the SGE at which index we stopped processing descriptors
1687  * 4. Bookkeeping; free TX buffers, ring doorbell if there are any
1688  *      outstanding TX buffers waiting, replenish RX buffers, potentially
1689  *      reenable upper layers if they were turned off due to lack of TX
1690  *      resources which are available again.
1691  * 5. If we took an interrupt, but no valid respQ descriptors was found we
1692  *      let the slow_intr_handler run and do error handling.
1693  */
1694 irqreturn_t t1_interrupt(int irq, void *cookie)
1695 {
1696         int work_done;
1697         struct adapter *adapter = cookie;
1698         struct respQ *Q = &adapter->sge->respQ;
1699
1700         spin_lock(&adapter->async_lock);
1701
1702         writel(F_PL_INTR_SGE_DATA, adapter->regs + A_PL_CAUSE);
1703
1704         if (likely(responses_pending(adapter)))
1705                 work_done = process_responses(adapter, -1);
1706         else
1707                 work_done = t1_slow_intr_handler(adapter);
1708
1709         /*
1710          * The unconditional clearing of the PL_CAUSE above may have raced
1711          * with DMA completion and the corresponding generation of a response
1712          * to cause us to miss the resulting data interrupt.  The next write
1713          * is also unconditional to recover the missed interrupt and render
1714          * this race harmless.
1715          */
1716         writel(Q->cidx, adapter->regs + A_SG_SLEEPING);
1717
1718         if (!work_done)
1719                 adapter->sge->stats.unhandled_irqs++;
1720         spin_unlock(&adapter->async_lock);
1721         return IRQ_RETVAL(work_done != 0);
1722 }
1723 #endif
1724
1725 /*
1726  * Enqueues the sk_buff onto the cmdQ[qid] and has hardware fetch it.
1727  *
1728  * The code figures out how many entries the sk_buff will require in the
1729  * cmdQ and updates the cmdQ data structure with the state once the enqueue
1730  * has complete. Then, it doesn't access the global structure anymore, but
1731  * uses the corresponding fields on the stack. In conjuction with a spinlock
1732  * around that code, we can make the function reentrant without holding the
1733  * lock when we actually enqueue (which might be expensive, especially on
1734  * architectures with IO MMUs).
1735  *
1736  * This runs with softirqs disabled.
1737  */
1738 static int t1_sge_tx(struct sk_buff *skb, struct adapter *adapter,
1739                      unsigned int qid, struct net_device *dev)
1740 {
1741         struct sge *sge = adapter->sge;
1742         struct cmdQ *q = &sge->cmdQ[qid];
1743         unsigned int credits, pidx, genbit, count, use_sched_skb = 0;
1744
1745         if (!spin_trylock(&q->lock))
1746                 return NETDEV_TX_LOCKED;
1747
1748         reclaim_completed_tx(sge, q);
1749
1750         pidx = q->pidx;
1751         credits = q->size - q->in_use;
1752         count = 1 + skb_shinfo(skb)->nr_frags;
1753         count += compute_large_page_tx_descs(skb);
1754
1755         /* Ethernet packet */
1756         if (unlikely(credits < count)) {
1757                 if (!netif_queue_stopped(dev)) {
1758                         netif_stop_queue(dev);
1759                         set_bit(dev->if_port, &sge->stopped_tx_queues);
1760                         sge->stats.cmdQ_full[2]++;
1761                         CH_ERR("%s: Tx ring full while queue awake!\n",
1762                                adapter->name);
1763                 }
1764                 spin_unlock(&q->lock);
1765                 return NETDEV_TX_BUSY;
1766         }
1767
1768         if (unlikely(credits - count < q->stop_thres)) {
1769                 netif_stop_queue(dev);
1770                 set_bit(dev->if_port, &sge->stopped_tx_queues);
1771                 sge->stats.cmdQ_full[2]++;
1772         }
1773
1774         /* T204 cmdQ0 skbs that are destined for a certain port have to go
1775          * through the scheduler.
1776          */
1777         if (sge->tx_sched && !qid && skb->dev) {
1778 use_sched:
1779                 use_sched_skb = 1;
1780                 /* Note that the scheduler might return a different skb than
1781                  * the one passed in.
1782                  */
1783                 skb = sched_skb(sge, skb, credits);
1784                 if (!skb) {
1785                         spin_unlock(&q->lock);
1786                         return NETDEV_TX_OK;
1787                 }
1788                 pidx = q->pidx;
1789                 count = 1 + skb_shinfo(skb)->nr_frags;
1790                 count += compute_large_page_tx_descs(skb);
1791         }
1792
1793         q->in_use += count;
1794         genbit = q->genbit;
1795         pidx = q->pidx;
1796         q->pidx += count;
1797         if (q->pidx >= q->size) {
1798                 q->pidx -= q->size;
1799                 q->genbit ^= 1;
1800         }
1801         spin_unlock(&q->lock);
1802
1803         write_tx_descs(adapter, skb, pidx, genbit, q);
1804
1805         /*
1806          * We always ring the doorbell for cmdQ1.  For cmdQ0, we only ring
1807          * the doorbell if the Q is asleep. There is a natural race, where
1808          * the hardware is going to sleep just after we checked, however,
1809          * then the interrupt handler will detect the outstanding TX packet
1810          * and ring the doorbell for us.
1811          */
1812         if (qid)
1813                 doorbell_pio(adapter, F_CMDQ1_ENABLE);
1814         else {
1815                 clear_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1816                 if (test_and_set_bit(CMDQ_STAT_RUNNING, &q->status) == 0) {
1817                         set_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1818                         writel(F_CMDQ0_ENABLE, adapter->regs + A_SG_DOORBELL);
1819                 }
1820         }
1821
1822         if (use_sched_skb) {
1823                 if (spin_trylock(&q->lock)) {
1824                         credits = q->size - q->in_use;
1825                         skb = NULL;
1826                         goto use_sched;
1827                 }
1828         }
1829         return NETDEV_TX_OK;
1830 }
1831
1832 #define MK_ETH_TYPE_MSS(type, mss) (((mss) & 0x3FFF) | ((type) << 14))
1833
1834 /*
1835  *      eth_hdr_len - return the length of an Ethernet header
1836  *      @data: pointer to the start of the Ethernet header
1837  *
1838  *      Returns the length of an Ethernet header, including optional VLAN tag.
1839  */
1840 static inline int eth_hdr_len(const void *data)
1841 {
1842         const struct ethhdr *e = data;
1843
1844         return e->h_proto == htons(ETH_P_8021Q) ? VLAN_ETH_HLEN : ETH_HLEN;
1845 }
1846
1847 /*
1848  * Adds the CPL header to the sk_buff and passes it to t1_sge_tx.
1849  */
1850 int t1_start_xmit(struct sk_buff *skb, struct net_device *dev)
1851 {
1852         struct adapter *adapter = dev->priv;
1853         struct sge *sge = adapter->sge;
1854         struct sge_port_stats *st = per_cpu_ptr(sge->port_stats[dev->if_port], smp_processor_id());
1855         struct cpl_tx_pkt *cpl;
1856         struct sk_buff *orig_skb = skb;
1857         int ret;
1858
1859         if (skb->protocol == htons(ETH_P_CPL5))
1860                 goto send;
1861
1862         if (skb_shinfo(skb)->gso_size) {
1863                 int eth_type;
1864                 struct cpl_tx_pkt_lso *hdr;
1865
1866                 ++st->tx_tso;
1867
1868                 eth_type = skb_network_offset(skb) == ETH_HLEN ?
1869                         CPL_ETH_II : CPL_ETH_II_VLAN;
1870
1871                 hdr = (struct cpl_tx_pkt_lso *)skb_push(skb, sizeof(*hdr));
1872                 hdr->opcode = CPL_TX_PKT_LSO;
1873                 hdr->ip_csum_dis = hdr->l4_csum_dis = 0;
1874                 hdr->ip_hdr_words = ip_hdr(skb)->ihl;
1875                 hdr->tcp_hdr_words = tcp_hdr(skb)->doff;
1876                 hdr->eth_type_mss = htons(MK_ETH_TYPE_MSS(eth_type,
1877                                                           skb_shinfo(skb)->gso_size));
1878                 hdr->len = htonl(skb->len - sizeof(*hdr));
1879                 cpl = (struct cpl_tx_pkt *)hdr;
1880         } else {
1881                 /*
1882                  * Packets shorter than ETH_HLEN can break the MAC, drop them
1883                  * early.  Also, we may get oversized packets because some
1884                  * parts of the kernel don't handle our unusual hard_header_len
1885                  * right, drop those too.
1886                  */
1887                 if (unlikely(skb->len < ETH_HLEN ||
1888                              skb->len > dev->mtu + eth_hdr_len(skb->data))) {
1889                         pr_debug("%s: packet size %d hdr %d mtu%d\n", dev->name,
1890                                  skb->len, eth_hdr_len(skb->data), dev->mtu);
1891                         dev_kfree_skb_any(skb);
1892                         return NETDEV_TX_OK;
1893                 }
1894
1895                 /*
1896                  * We are using a non-standard hard_header_len and some kernel
1897                  * components, such as pktgen, do not handle it right.
1898                  * Complain when this happens but try to fix things up.
1899                  */
1900                 if (unlikely(skb_headroom(skb) < dev->hard_header_len - ETH_HLEN)) {
1901                         pr_debug("%s: headroom %d header_len %d\n", dev->name,
1902                                  skb_headroom(skb), dev->hard_header_len);
1903
1904                         if (net_ratelimit())
1905                                 printk(KERN_ERR "%s: inadequate headroom in "
1906                                        "Tx packet\n", dev->name);
1907                         skb = skb_realloc_headroom(skb, sizeof(*cpl));
1908                         dev_kfree_skb_any(orig_skb);
1909                         if (!skb)
1910                                 return NETDEV_TX_OK;
1911                 }
1912
1913                 if (!(adapter->flags & UDP_CSUM_CAPABLE) &&
1914                     skb->ip_summed == CHECKSUM_PARTIAL &&
1915                     ip_hdr(skb)->protocol == IPPROTO_UDP) {
1916                         if (unlikely(skb_checksum_help(skb))) {
1917                                 pr_debug("%s: unable to do udp checksum\n", dev->name);
1918                                 dev_kfree_skb_any(skb);
1919                                 return NETDEV_TX_OK;
1920                         }
1921                 }
1922
1923                 /* Hmmm, assuming to catch the gratious arp... and we'll use
1924                  * it to flush out stuck espi packets...
1925                  */
1926                 if ((unlikely(!adapter->sge->espibug_skb[dev->if_port]))) {
1927                         if (skb->protocol == htons(ETH_P_ARP) &&
1928                             arp_hdr(skb)->ar_op == htons(ARPOP_REQUEST)) {
1929                                 adapter->sge->espibug_skb[dev->if_port] = skb;
1930                                 /* We want to re-use this skb later. We
1931                                  * simply bump the reference count and it
1932                                  * will not be freed...
1933                                  */
1934                                 skb = skb_get(skb);
1935                         }
1936                 }
1937
1938                 cpl = (struct cpl_tx_pkt *)__skb_push(skb, sizeof(*cpl));
1939                 cpl->opcode = CPL_TX_PKT;
1940                 cpl->ip_csum_dis = 1;    /* SW calculates IP csum */
1941                 cpl->l4_csum_dis = skb->ip_summed == CHECKSUM_PARTIAL ? 0 : 1;
1942                 /* the length field isn't used so don't bother setting it */
1943
1944                 st->tx_cso += (skb->ip_summed == CHECKSUM_PARTIAL);
1945         }
1946         cpl->iff = dev->if_port;
1947
1948 #if defined(CONFIG_VLAN_8021Q) || defined(CONFIG_VLAN_8021Q_MODULE)
1949         if (adapter->vlan_grp && vlan_tx_tag_present(skb)) {
1950                 cpl->vlan_valid = 1;
1951                 cpl->vlan = htons(vlan_tx_tag_get(skb));
1952                 st->vlan_insert++;
1953         } else
1954 #endif
1955                 cpl->vlan_valid = 0;
1956
1957 send:
1958         st->tx_packets++;
1959         dev->trans_start = jiffies;
1960         ret = t1_sge_tx(skb, adapter, 0, dev);
1961
1962         /* If transmit busy, and we reallocated skb's due to headroom limit,
1963          * then silently discard to avoid leak.
1964          */
1965         if (unlikely(ret != NETDEV_TX_OK && skb != orig_skb)) {
1966                 dev_kfree_skb_any(skb);
1967                 ret = NETDEV_TX_OK;
1968         }
1969         return ret;
1970 }
1971
1972 /*
1973  * Callback for the Tx buffer reclaim timer.  Runs with softirqs disabled.
1974  */
1975 static void sge_tx_reclaim_cb(unsigned long data)
1976 {
1977         int i;
1978         struct sge *sge = (struct sge *)data;
1979
1980         for (i = 0; i < SGE_CMDQ_N; ++i) {
1981                 struct cmdQ *q = &sge->cmdQ[i];
1982
1983                 if (!spin_trylock(&q->lock))
1984                         continue;
1985
1986                 reclaim_completed_tx(sge, q);
1987                 if (i == 0 && q->in_use) {    /* flush pending credits */
1988                         writel(F_CMDQ0_ENABLE, sge->adapter->regs + A_SG_DOORBELL);
1989                 }
1990                 spin_unlock(&q->lock);
1991         }
1992         mod_timer(&sge->tx_reclaim_timer, jiffies + TX_RECLAIM_PERIOD);
1993 }
1994
1995 /*
1996  * Propagate changes of the SGE coalescing parameters to the HW.
1997  */
1998 int t1_sge_set_coalesce_params(struct sge *sge, struct sge_params *p)
1999 {
2000         sge->fixed_intrtimer = p->rx_coalesce_usecs *
2001                 core_ticks_per_usec(sge->adapter);
2002         writel(sge->fixed_intrtimer, sge->adapter->regs + A_SG_INTRTIMER);
2003         return 0;
2004 }
2005
2006 /*
2007  * Allocates both RX and TX resources and configures the SGE. However,
2008  * the hardware is not enabled yet.
2009  */
2010 int t1_sge_configure(struct sge *sge, struct sge_params *p)
2011 {
2012         if (alloc_rx_resources(sge, p))
2013                 return -ENOMEM;
2014         if (alloc_tx_resources(sge, p)) {
2015                 free_rx_resources(sge);
2016                 return -ENOMEM;
2017         }
2018         configure_sge(sge, p);
2019
2020         /*
2021          * Now that we have sized the free lists calculate the payload
2022          * capacity of the large buffers.  Other parts of the driver use
2023          * this to set the max offload coalescing size so that RX packets
2024          * do not overflow our large buffers.
2025          */
2026         p->large_buf_capacity = jumbo_payload_capacity(sge);
2027         return 0;
2028 }
2029
2030 /*
2031  * Disables the DMA engine.
2032  */
2033 void t1_sge_stop(struct sge *sge)
2034 {
2035         int i;
2036         writel(0, sge->adapter->regs + A_SG_CONTROL);
2037         readl(sge->adapter->regs + A_SG_CONTROL); /* flush */
2038
2039         if (is_T2(sge->adapter))
2040                 del_timer_sync(&sge->espibug_timer);
2041
2042         del_timer_sync(&sge->tx_reclaim_timer);
2043         if (sge->tx_sched)
2044                 tx_sched_stop(sge);
2045
2046         for (i = 0; i < MAX_NPORTS; i++)
2047                 if (sge->espibug_skb[i])
2048                         kfree_skb(sge->espibug_skb[i]);
2049 }
2050
2051 /*
2052  * Enables the DMA engine.
2053  */
2054 void t1_sge_start(struct sge *sge)
2055 {
2056         refill_free_list(sge, &sge->freelQ[0]);
2057         refill_free_list(sge, &sge->freelQ[1]);
2058
2059         writel(sge->sge_control, sge->adapter->regs + A_SG_CONTROL);
2060         doorbell_pio(sge->adapter, F_FL0_ENABLE | F_FL1_ENABLE);
2061         readl(sge->adapter->regs + A_SG_CONTROL); /* flush */
2062
2063         mod_timer(&sge->tx_reclaim_timer, jiffies + TX_RECLAIM_PERIOD);
2064
2065         if (is_T2(sge->adapter))
2066                 mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
2067 }
2068
2069 /*
2070  * Callback for the T2 ESPI 'stuck packet feature' workaorund
2071  */
2072 static void espibug_workaround_t204(unsigned long data)
2073 {
2074         struct adapter *adapter = (struct adapter *)data;
2075         struct sge *sge = adapter->sge;
2076         unsigned int nports = adapter->params.nports;
2077         u32 seop[MAX_NPORTS];
2078
2079         if (adapter->open_device_map & PORT_MASK) {
2080                 int i;
2081
2082                 if (t1_espi_get_mon_t204(adapter, &(seop[0]), 0) < 0)
2083                         return;
2084
2085                 for (i = 0; i < nports; i++) {
2086                         struct sk_buff *skb = sge->espibug_skb[i];
2087
2088                         if (!netif_running(adapter->port[i].dev) ||
2089                             netif_queue_stopped(adapter->port[i].dev) ||
2090                             !seop[i] || ((seop[i] & 0xfff) != 0) || !skb)
2091                                 continue;
2092
2093                         if (!skb->cb[0]) {
2094                                 u8 ch_mac_addr[ETH_ALEN] = {
2095                                         0x0, 0x7, 0x43, 0x0, 0x0, 0x0
2096                                 };
2097
2098                                 memcpy(skb->data + sizeof(struct cpl_tx_pkt),
2099                                         ch_mac_addr, ETH_ALEN);
2100                                 memcpy(skb->data + skb->len - 10,
2101                                         ch_mac_addr, ETH_ALEN);
2102                                 skb->cb[0] = 0xff;
2103                         }
2104
2105                         /* bump the reference count to avoid freeing of
2106                          * the skb once the DMA has completed.
2107                          */
2108                         skb = skb_get(skb);
2109                         t1_sge_tx(skb, adapter, 0, adapter->port[i].dev);
2110                 }
2111         }
2112         mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
2113 }
2114
2115 static void espibug_workaround(unsigned long data)
2116 {
2117         struct adapter *adapter = (struct adapter *)data;
2118         struct sge *sge = adapter->sge;
2119
2120         if (netif_running(adapter->port[0].dev)) {
2121                 struct sk_buff *skb = sge->espibug_skb[0];
2122                 u32 seop = t1_espi_get_mon(adapter, 0x930, 0);
2123
2124                 if ((seop & 0xfff0fff) == 0xfff && skb) {
2125                         if (!skb->cb[0]) {
2126                                 u8 ch_mac_addr[ETH_ALEN] =
2127                                     {0x0, 0x7, 0x43, 0x0, 0x0, 0x0};
2128                                 memcpy(skb->data + sizeof(struct cpl_tx_pkt),
2129                                        ch_mac_addr, ETH_ALEN);
2130                                 memcpy(skb->data + skb->len - 10, ch_mac_addr,
2131                                        ETH_ALEN);
2132                                 skb->cb[0] = 0xff;
2133                         }
2134
2135                         /* bump the reference count to avoid freeing of the
2136                          * skb once the DMA has completed.
2137                          */
2138                         skb = skb_get(skb);
2139                         t1_sge_tx(skb, adapter, 0, adapter->port[0].dev);
2140                 }
2141         }
2142         mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
2143 }
2144
2145 /*
2146  * Creates a t1_sge structure and returns suggested resource parameters.
2147  */
2148 struct sge * __devinit t1_sge_create(struct adapter *adapter,
2149                                      struct sge_params *p)
2150 {
2151         struct sge *sge = kzalloc(sizeof(*sge), GFP_KERNEL);
2152         int i;
2153
2154         if (!sge)
2155                 return NULL;
2156
2157         sge->adapter = adapter;
2158         sge->netdev = adapter->port[0].dev;
2159         sge->rx_pkt_pad = t1_is_T1B(adapter) ? 0 : 2;
2160         sge->jumbo_fl = t1_is_T1B(adapter) ? 1 : 0;
2161
2162         for_each_port(adapter, i) {
2163                 sge->port_stats[i] = alloc_percpu(struct sge_port_stats);
2164                 if (!sge->port_stats[i])
2165                         goto nomem_port;
2166         }
2167
2168         init_timer(&sge->tx_reclaim_timer);
2169         sge->tx_reclaim_timer.data = (unsigned long)sge;
2170         sge->tx_reclaim_timer.function = sge_tx_reclaim_cb;
2171
2172         if (is_T2(sge->adapter)) {
2173                 init_timer(&sge->espibug_timer);
2174
2175                 if (adapter->params.nports > 1) {
2176                         tx_sched_init(sge);
2177                         sge->espibug_timer.function = espibug_workaround_t204;
2178                 } else
2179                         sge->espibug_timer.function = espibug_workaround;
2180                 sge->espibug_timer.data = (unsigned long)sge->adapter;
2181
2182                 sge->espibug_timeout = 1;
2183                 /* for T204, every 10ms */
2184                 if (adapter->params.nports > 1)
2185                         sge->espibug_timeout = HZ/100;
2186         }
2187
2188
2189         p->cmdQ_size[0] = SGE_CMDQ0_E_N;
2190         p->cmdQ_size[1] = SGE_CMDQ1_E_N;
2191         p->freelQ_size[!sge->jumbo_fl] = SGE_FREEL_SIZE;
2192         p->freelQ_size[sge->jumbo_fl] = SGE_JUMBO_FREEL_SIZE;
2193         if (sge->tx_sched) {
2194                 if (board_info(sge->adapter)->board == CHBT_BOARD_CHT204)
2195                         p->rx_coalesce_usecs = 15;
2196                 else
2197                         p->rx_coalesce_usecs = 50;
2198         } else
2199                 p->rx_coalesce_usecs = 50;
2200
2201         p->coalesce_enable = 0;
2202         p->sample_interval_usecs = 0;
2203
2204         return sge;
2205 nomem_port:
2206         while (i >= 0) {
2207                 free_percpu(sge->port_stats[i]);
2208                 --i;
2209         }
2210         kfree(sge);
2211         return NULL;
2212
2213 }