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, see <http://www.gnu.org/licenses/>.            *
16  *                                                                           *
17  * THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY EXPRESS OR IMPLIED    *
18  * WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED WARRANTIES OF      *
19  * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.                     *
20  *                                                                           *
21  * http://www.chelsio.com                                                    *
22  *                                                                           *
23  * Copyright (c) 2003 - 2005 Chelsio Communications, Inc.                    *
24  * All rights reserved.                                                      *
25  *                                                                           *
26  * Maintainers: maintainers@chelsio.com                                      *
27  *                                                                           *
28  * Authors: Dimitrios Michailidis   <dm@chelsio.com>                         *
29  *          Tina Yang               <tainay@chelsio.com>                     *
30  *          Felix Marti             <felix@chelsio.com>                      *
31  *          Scott Bardone           <sbardone@chelsio.com>                   *
32  *          Kurt Ottaway            <kottaway@chelsio.com>                   *
33  *          Frank DiMambro          <frank@chelsio.com>                      *
34  *                                                                           *
35  * History:                                                                  *
36  *                                                                           *
37  ****************************************************************************/
38 
39 #include "common.h"
40 
41 #include <linux/types.h>
42 #include <linux/errno.h>
43 #include <linux/pci.h>
44 #include <linux/ktime.h>
45 #include <linux/netdevice.h>
46 #include <linux/etherdevice.h>
47 #include <linux/if_vlan.h>
48 #include <linux/skbuff.h>
49 #include <linux/mm.h>
50 #include <linux/tcp.h>
51 #include <linux/ip.h>
52 #include <linux/in.h>
53 #include <linux/if_arp.h>
54 #include <linux/slab.h>
55 #include <linux/prefetch.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 	DEFINE_DMA_UNMAP_ADDR(dma_addr);
165 	DEFINE_DMA_UNMAP_LEN(dma_len);
166 };
167 
168 struct freelQ_ce {
169 	struct sk_buff *skb;
170 	DEFINE_DMA_UNMAP_ADDR(dma_addr);
171 	DEFINE_DMA_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  * separate 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 __percpu *port_stats[MAX_NPORTS];
271 	struct sched	*tx_sched;
272 	struct cmdQ cmdQ[SGE_CMDQ_N] ____cacheline_aligned_in_smp;
273 };
274 
275 static const u8 ch_mac_addr[ETH_ALEN] = {
276 	0x0, 0x7, 0x43, 0x0, 0x0, 0x0
277 };
278 
279 /*
280  * stop tasklet and free all pending skb's
281  */
282 static void tx_sched_stop(struct sge *sge)
283 {
284 	struct sched *s = sge->tx_sched;
285 	int i;
286 
287 	tasklet_kill(&s->sched_tsk);
288 
289 	for (i = 0; i < MAX_NPORTS; i++)
290 		__skb_queue_purge(&s->p[s->port].skbq);
291 }
292 
293 /*
294  * t1_sched_update_parms() is called when the MTU or link speed changes. It
295  * re-computes scheduler parameters to scope with the change.
296  */
297 unsigned int t1_sched_update_parms(struct sge *sge, unsigned int port,
298 				   unsigned int mtu, unsigned int speed)
299 {
300 	struct sched *s = sge->tx_sched;
301 	struct sched_port *p = &s->p[port];
302 	unsigned int max_avail_segs;
303 
304 	pr_debug("t1_sched_update_params mtu=%d speed=%d\n", mtu, speed);
305 	if (speed)
306 		p->speed = speed;
307 	if (mtu)
308 		p->mtu = mtu;
309 
310 	if (speed || mtu) {
311 		unsigned long long drain = 1024ULL * p->speed * (p->mtu - 40);
312 		do_div(drain, (p->mtu + 50) * 1000);
313 		p->drain_bits_per_1024ns = (unsigned int) drain;
314 
315 		if (p->speed < 1000)
316 			p->drain_bits_per_1024ns =
317 				90 * p->drain_bits_per_1024ns / 100;
318 	}
319 
320 	if (board_info(sge->adapter)->board == CHBT_BOARD_CHT204) {
321 		p->drain_bits_per_1024ns -= 16;
322 		s->max_avail = max(4096U, p->mtu + 16 + 14 + 4);
323 		max_avail_segs = max(1U, 4096 / (p->mtu - 40));
324 	} else {
325 		s->max_avail = 16384;
326 		max_avail_segs = max(1U, 9000 / (p->mtu - 40));
327 	}
328 
329 	pr_debug("t1_sched_update_parms: mtu %u speed %u max_avail %u "
330 		 "max_avail_segs %u drain_bits_per_1024ns %u\n", p->mtu,
331 		 p->speed, s->max_avail, max_avail_segs,
332 		 p->drain_bits_per_1024ns);
333 
334 	return max_avail_segs * (p->mtu - 40);
335 }
336 
337 #if 0
338 
339 /*
340  * t1_sched_max_avail_bytes() tells the scheduler the maximum amount of
341  * data that can be pushed per port.
342  */
343 void t1_sched_set_max_avail_bytes(struct sge *sge, unsigned int val)
344 {
345 	struct sched *s = sge->tx_sched;
346 	unsigned int i;
347 
348 	s->max_avail = val;
349 	for (i = 0; i < MAX_NPORTS; i++)
350 		t1_sched_update_parms(sge, i, 0, 0);
351 }
352 
353 /*
354  * t1_sched_set_drain_bits_per_us() tells the scheduler at which rate a port
355  * is draining.
356  */
357 void t1_sched_set_drain_bits_per_us(struct sge *sge, unsigned int port,
358 					 unsigned int val)
359 {
360 	struct sched *s = sge->tx_sched;
361 	struct sched_port *p = &s->p[port];
362 	p->drain_bits_per_1024ns = val * 1024 / 1000;
363 	t1_sched_update_parms(sge, port, 0, 0);
364 }
365 
366 #endif  /*  0  */
367 
368 /*
369  * tx_sched_init() allocates resources and does basic initialization.
370  */
371 static int tx_sched_init(struct sge *sge)
372 {
373 	struct sched *s;
374 	int i;
375 
376 	s = kzalloc(sizeof (struct sched), GFP_KERNEL);
377 	if (!s)
378 		return -ENOMEM;
379 
380 	pr_debug("tx_sched_init\n");
381 	tasklet_init(&s->sched_tsk, restart_sched, (unsigned long) sge);
382 	sge->tx_sched = s;
383 
384 	for (i = 0; i < MAX_NPORTS; i++) {
385 		skb_queue_head_init(&s->p[i].skbq);
386 		t1_sched_update_parms(sge, i, 1500, 1000);
387 	}
388 
389 	return 0;
390 }
391 
392 /*
393  * sched_update_avail() computes the delta since the last time it was called
394  * and updates the per port quota (number of bits that can be sent to the any
395  * port).
396  */
397 static inline int sched_update_avail(struct sge *sge)
398 {
399 	struct sched *s = sge->tx_sched;
400 	ktime_t now = ktime_get();
401 	unsigned int i;
402 	long long delta_time_ns;
403 
404 	delta_time_ns = ktime_to_ns(ktime_sub(now, s->last_updated));
405 
406 	pr_debug("sched_update_avail delta=%lld\n", delta_time_ns);
407 	if (delta_time_ns < 15000)
408 		return 0;
409 
410 	for (i = 0; i < MAX_NPORTS; i++) {
411 		struct sched_port *p = &s->p[i];
412 		unsigned int delta_avail;
413 
414 		delta_avail = (p->drain_bits_per_1024ns * delta_time_ns) >> 13;
415 		p->avail = min(p->avail + delta_avail, s->max_avail);
416 	}
417 
418 	s->last_updated = now;
419 
420 	return 1;
421 }
422 
423 /*
424  * sched_skb() is called from two different places. In the tx path, any
425  * packet generating load on an output port will call sched_skb()
426  * (skb != NULL). In addition, sched_skb() is called from the irq/soft irq
427  * context (skb == NULL).
428  * The scheduler only returns a skb (which will then be sent) if the
429  * length of the skb is <= the current quota of the output port.
430  */
431 static struct sk_buff *sched_skb(struct sge *sge, struct sk_buff *skb,
432 				unsigned int credits)
433 {
434 	struct sched *s = sge->tx_sched;
435 	struct sk_buff_head *skbq;
436 	unsigned int i, len, update = 1;
437 
438 	pr_debug("sched_skb %p\n", skb);
439 	if (!skb) {
440 		if (!s->num)
441 			return NULL;
442 	} else {
443 		skbq = &s->p[skb->dev->if_port].skbq;
444 		__skb_queue_tail(skbq, skb);
445 		s->num++;
446 		skb = NULL;
447 	}
448 
449 	if (credits < MAX_SKB_FRAGS + 1)
450 		goto out;
451 
452 again:
453 	for (i = 0; i < MAX_NPORTS; i++) {
454 		s->port = (s->port + 1) & (MAX_NPORTS - 1);
455 		skbq = &s->p[s->port].skbq;
456 
457 		skb = skb_peek(skbq);
458 
459 		if (!skb)
460 			continue;
461 
462 		len = skb->len;
463 		if (len <= s->p[s->port].avail) {
464 			s->p[s->port].avail -= len;
465 			s->num--;
466 			__skb_unlink(skb, skbq);
467 			goto out;
468 		}
469 		skb = NULL;
470 	}
471 
472 	if (update-- && sched_update_avail(sge))
473 		goto again;
474 
475 out:
476 	/* If there are more pending skbs, we use the hardware to schedule us
477 	 * again.
478 	 */
479 	if (s->num && !skb) {
480 		struct cmdQ *q = &sge->cmdQ[0];
481 		clear_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
482 		if (test_and_set_bit(CMDQ_STAT_RUNNING, &q->status) == 0) {
483 			set_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
484 			writel(F_CMDQ0_ENABLE, sge->adapter->regs + A_SG_DOORBELL);
485 		}
486 	}
487 	pr_debug("sched_skb ret %p\n", skb);
488 
489 	return skb;
490 }
491 
492 /*
493  * PIO to indicate that memory mapped Q contains valid descriptor(s).
494  */
495 static inline void doorbell_pio(struct adapter *adapter, u32 val)
496 {
497 	wmb();
498 	writel(val, adapter->regs + A_SG_DOORBELL);
499 }
500 
501 /*
502  * Frees all RX buffers on the freelist Q. The caller must make sure that
503  * the SGE is turned off before calling this function.
504  */
505 static void free_freelQ_buffers(struct pci_dev *pdev, struct freelQ *q)
506 {
507 	unsigned int cidx = q->cidx;
508 
509 	while (q->credits--) {
510 		struct freelQ_ce *ce = &q->centries[cidx];
511 
512 		pci_unmap_single(pdev, dma_unmap_addr(ce, dma_addr),
513 				 dma_unmap_len(ce, dma_len),
514 				 PCI_DMA_FROMDEVICE);
515 		dev_kfree_skb(ce->skb);
516 		ce->skb = NULL;
517 		if (++cidx == q->size)
518 			cidx = 0;
519 	}
520 }
521 
522 /*
523  * Free RX free list and response queue resources.
524  */
525 static void free_rx_resources(struct sge *sge)
526 {
527 	struct pci_dev *pdev = sge->adapter->pdev;
528 	unsigned int size, i;
529 
530 	if (sge->respQ.entries) {
531 		size = sizeof(struct respQ_e) * sge->respQ.size;
532 		pci_free_consistent(pdev, size, sge->respQ.entries,
533 				    sge->respQ.dma_addr);
534 	}
535 
536 	for (i = 0; i < SGE_FREELQ_N; i++) {
537 		struct freelQ *q = &sge->freelQ[i];
538 
539 		if (q->centries) {
540 			free_freelQ_buffers(pdev, q);
541 			kfree(q->centries);
542 		}
543 		if (q->entries) {
544 			size = sizeof(struct freelQ_e) * q->size;
545 			pci_free_consistent(pdev, size, q->entries,
546 					    q->dma_addr);
547 		}
548 	}
549 }
550 
551 /*
552  * Allocates basic RX resources, consisting of memory mapped freelist Qs and a
553  * response queue.
554  */
555 static int alloc_rx_resources(struct sge *sge, struct sge_params *p)
556 {
557 	struct pci_dev *pdev = sge->adapter->pdev;
558 	unsigned int size, i;
559 
560 	for (i = 0; i < SGE_FREELQ_N; i++) {
561 		struct freelQ *q = &sge->freelQ[i];
562 
563 		q->genbit = 1;
564 		q->size = p->freelQ_size[i];
565 		q->dma_offset = sge->rx_pkt_pad ? 0 : NET_IP_ALIGN;
566 		size = sizeof(struct freelQ_e) * q->size;
567 		q->entries = pci_alloc_consistent(pdev, size, &q->dma_addr);
568 		if (!q->entries)
569 			goto err_no_mem;
570 
571 		size = sizeof(struct freelQ_ce) * q->size;
572 		q->centries = kzalloc(size, GFP_KERNEL);
573 		if (!q->centries)
574 			goto err_no_mem;
575 	}
576 
577 	/*
578 	 * Calculate the buffer sizes for the two free lists.  FL0 accommodates
579 	 * regular sized Ethernet frames, FL1 is sized not to exceed 16K,
580 	 * including all the sk_buff overhead.
581 	 *
582 	 * Note: For T2 FL0 and FL1 are reversed.
583 	 */
584 	sge->freelQ[!sge->jumbo_fl].rx_buffer_size = SGE_RX_SM_BUF_SIZE +
585 		sizeof(struct cpl_rx_data) +
586 		sge->freelQ[!sge->jumbo_fl].dma_offset;
587 
588 		size = (16 * 1024) -
589 		    SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
590 
591 	sge->freelQ[sge->jumbo_fl].rx_buffer_size = size;
592 
593 	/*
594 	 * Setup which skb recycle Q should be used when recycling buffers from
595 	 * each free list.
596 	 */
597 	sge->freelQ[!sge->jumbo_fl].recycleq_idx = 0;
598 	sge->freelQ[sge->jumbo_fl].recycleq_idx = 1;
599 
600 	sge->respQ.genbit = 1;
601 	sge->respQ.size = SGE_RESPQ_E_N;
602 	sge->respQ.credits = 0;
603 	size = sizeof(struct respQ_e) * sge->respQ.size;
604 	sge->respQ.entries =
605 		pci_alloc_consistent(pdev, size, &sge->respQ.dma_addr);
606 	if (!sge->respQ.entries)
607 		goto err_no_mem;
608 	return 0;
609 
610 err_no_mem:
611 	free_rx_resources(sge);
612 	return -ENOMEM;
613 }
614 
615 /*
616  * Reclaims n TX descriptors and frees the buffers associated with them.
617  */
618 static void free_cmdQ_buffers(struct sge *sge, struct cmdQ *q, unsigned int n)
619 {
620 	struct cmdQ_ce *ce;
621 	struct pci_dev *pdev = sge->adapter->pdev;
622 	unsigned int cidx = q->cidx;
623 
624 	q->in_use -= n;
625 	ce = &q->centries[cidx];
626 	while (n--) {
627 		if (likely(dma_unmap_len(ce, dma_len))) {
628 			pci_unmap_single(pdev, dma_unmap_addr(ce, dma_addr),
629 					 dma_unmap_len(ce, dma_len),
630 					 PCI_DMA_TODEVICE);
631 			if (q->sop)
632 				q->sop = 0;
633 		}
634 		if (ce->skb) {
635 			dev_kfree_skb_any(ce->skb);
636 			q->sop = 1;
637 		}
638 		ce++;
639 		if (++cidx == q->size) {
640 			cidx = 0;
641 			ce = q->centries;
642 		}
643 	}
644 	q->cidx = cidx;
645 }
646 
647 /*
648  * Free TX resources.
649  *
650  * Assumes that SGE is stopped and all interrupts are disabled.
651  */
652 static void free_tx_resources(struct sge *sge)
653 {
654 	struct pci_dev *pdev = sge->adapter->pdev;
655 	unsigned int size, i;
656 
657 	for (i = 0; i < SGE_CMDQ_N; i++) {
658 		struct cmdQ *q = &sge->cmdQ[i];
659 
660 		if (q->centries) {
661 			if (q->in_use)
662 				free_cmdQ_buffers(sge, q, q->in_use);
663 			kfree(q->centries);
664 		}
665 		if (q->entries) {
666 			size = sizeof(struct cmdQ_e) * q->size;
667 			pci_free_consistent(pdev, size, q->entries,
668 					    q->dma_addr);
669 		}
670 	}
671 }
672 
673 /*
674  * Allocates basic TX resources, consisting of memory mapped command Qs.
675  */
676 static int alloc_tx_resources(struct sge *sge, struct sge_params *p)
677 {
678 	struct pci_dev *pdev = sge->adapter->pdev;
679 	unsigned int size, i;
680 
681 	for (i = 0; i < SGE_CMDQ_N; i++) {
682 		struct cmdQ *q = &sge->cmdQ[i];
683 
684 		q->genbit = 1;
685 		q->sop = 1;
686 		q->size = p->cmdQ_size[i];
687 		q->in_use = 0;
688 		q->status = 0;
689 		q->processed = q->cleaned = 0;
690 		q->stop_thres = 0;
691 		spin_lock_init(&q->lock);
692 		size = sizeof(struct cmdQ_e) * q->size;
693 		q->entries = pci_alloc_consistent(pdev, size, &q->dma_addr);
694 		if (!q->entries)
695 			goto err_no_mem;
696 
697 		size = sizeof(struct cmdQ_ce) * q->size;
698 		q->centries = kzalloc(size, GFP_KERNEL);
699 		if (!q->centries)
700 			goto err_no_mem;
701 	}
702 
703 	/*
704 	 * CommandQ 0 handles Ethernet and TOE packets, while queue 1 is TOE
705 	 * only.  For queue 0 set the stop threshold so we can handle one more
706 	 * packet from each port, plus reserve an additional 24 entries for
707 	 * Ethernet packets only.  Queue 1 never suspends nor do we reserve
708 	 * space for Ethernet packets.
709 	 */
710 	sge->cmdQ[0].stop_thres = sge->adapter->params.nports *
711 		(MAX_SKB_FRAGS + 1);
712 	return 0;
713 
714 err_no_mem:
715 	free_tx_resources(sge);
716 	return -ENOMEM;
717 }
718 
719 static inline void setup_ring_params(struct adapter *adapter, u64 addr,
720 				     u32 size, int base_reg_lo,
721 				     int base_reg_hi, int size_reg)
722 {
723 	writel((u32)addr, adapter->regs + base_reg_lo);
724 	writel(addr >> 32, adapter->regs + base_reg_hi);
725 	writel(size, adapter->regs + size_reg);
726 }
727 
728 /*
729  * Enable/disable VLAN acceleration.
730  */
731 void t1_vlan_mode(struct adapter *adapter, netdev_features_t features)
732 {
733 	struct sge *sge = adapter->sge;
734 
735 	if (features & NETIF_F_HW_VLAN_CTAG_RX)
736 		sge->sge_control |= F_VLAN_XTRACT;
737 	else
738 		sge->sge_control &= ~F_VLAN_XTRACT;
739 	if (adapter->open_device_map) {
740 		writel(sge->sge_control, adapter->regs + A_SG_CONTROL);
741 		readl(adapter->regs + A_SG_CONTROL);   /* flush */
742 	}
743 }
744 
745 /*
746  * Programs the various SGE registers. However, the engine is not yet enabled,
747  * but sge->sge_control is setup and ready to go.
748  */
749 static void configure_sge(struct sge *sge, struct sge_params *p)
750 {
751 	struct adapter *ap = sge->adapter;
752 
753 	writel(0, ap->regs + A_SG_CONTROL);
754 	setup_ring_params(ap, sge->cmdQ[0].dma_addr, sge->cmdQ[0].size,
755 			  A_SG_CMD0BASELWR, A_SG_CMD0BASEUPR, A_SG_CMD0SIZE);
756 	setup_ring_params(ap, sge->cmdQ[1].dma_addr, sge->cmdQ[1].size,
757 			  A_SG_CMD1BASELWR, A_SG_CMD1BASEUPR, A_SG_CMD1SIZE);
758 	setup_ring_params(ap, sge->freelQ[0].dma_addr,
759 			  sge->freelQ[0].size, A_SG_FL0BASELWR,
760 			  A_SG_FL0BASEUPR, A_SG_FL0SIZE);
761 	setup_ring_params(ap, sge->freelQ[1].dma_addr,
762 			  sge->freelQ[1].size, A_SG_FL1BASELWR,
763 			  A_SG_FL1BASEUPR, A_SG_FL1SIZE);
764 
765 	/* The threshold comparison uses <. */
766 	writel(SGE_RX_SM_BUF_SIZE + 1, ap->regs + A_SG_FLTHRESHOLD);
767 
768 	setup_ring_params(ap, sge->respQ.dma_addr, sge->respQ.size,
769 			  A_SG_RSPBASELWR, A_SG_RSPBASEUPR, A_SG_RSPSIZE);
770 	writel((u32)sge->respQ.size - 1, ap->regs + A_SG_RSPQUEUECREDIT);
771 
772 	sge->sge_control = F_CMDQ0_ENABLE | F_CMDQ1_ENABLE | F_FL0_ENABLE |
773 		F_FL1_ENABLE | F_CPL_ENABLE | F_RESPONSE_QUEUE_ENABLE |
774 		V_CMDQ_PRIORITY(2) | F_DISABLE_CMDQ1_GTS | F_ISCSI_COALESCE |
775 		V_RX_PKT_OFFSET(sge->rx_pkt_pad);
776 
777 #if defined(__BIG_ENDIAN_BITFIELD)
778 	sge->sge_control |= F_ENABLE_BIG_ENDIAN;
779 #endif
780 
781 	/* Initialize no-resource timer */
782 	sge->intrtimer_nres = SGE_INTRTIMER_NRES * core_ticks_per_usec(ap);
783 
784 	t1_sge_set_coalesce_params(sge, p);
785 }
786 
787 /*
788  * Return the payload capacity of the jumbo free-list buffers.
789  */
790 static inline unsigned int jumbo_payload_capacity(const struct sge *sge)
791 {
792 	return sge->freelQ[sge->jumbo_fl].rx_buffer_size -
793 		sge->freelQ[sge->jumbo_fl].dma_offset -
794 		sizeof(struct cpl_rx_data);
795 }
796 
797 /*
798  * Frees all SGE related resources and the sge structure itself
799  */
800 void t1_sge_destroy(struct sge *sge)
801 {
802 	int i;
803 
804 	for_each_port(sge->adapter, i)
805 		free_percpu(sge->port_stats[i]);
806 
807 	kfree(sge->tx_sched);
808 	free_tx_resources(sge);
809 	free_rx_resources(sge);
810 	kfree(sge);
811 }
812 
813 /*
814  * Allocates new RX buffers on the freelist Q (and tracks them on the freelist
815  * context Q) until the Q is full or alloc_skb fails.
816  *
817  * It is possible that the generation bits already match, indicating that the
818  * buffer is already valid and nothing needs to be done. This happens when we
819  * copied a received buffer into a new sk_buff during the interrupt processing.
820  *
821  * If the SGE doesn't automatically align packets properly (!sge->rx_pkt_pad),
822  * we specify a RX_OFFSET in order to make sure that the IP header is 4B
823  * aligned.
824  */
825 static void refill_free_list(struct sge *sge, struct freelQ *q)
826 {
827 	struct pci_dev *pdev = sge->adapter->pdev;
828 	struct freelQ_ce *ce = &q->centries[q->pidx];
829 	struct freelQ_e *e = &q->entries[q->pidx];
830 	unsigned int dma_len = q->rx_buffer_size - q->dma_offset;
831 
832 	while (q->credits < q->size) {
833 		struct sk_buff *skb;
834 		dma_addr_t mapping;
835 
836 		skb = dev_alloc_skb(q->rx_buffer_size);
837 		if (!skb)
838 			break;
839 
840 		skb_reserve(skb, q->dma_offset);
841 		mapping = pci_map_single(pdev, skb->data, dma_len,
842 					 PCI_DMA_FROMDEVICE);
843 		skb_reserve(skb, sge->rx_pkt_pad);
844 
845 		ce->skb = skb;
846 		dma_unmap_addr_set(ce, dma_addr, mapping);
847 		dma_unmap_len_set(ce, dma_len, dma_len);
848 		e->addr_lo = (u32)mapping;
849 		e->addr_hi = (u64)mapping >> 32;
850 		e->len_gen = V_CMD_LEN(dma_len) | V_CMD_GEN1(q->genbit);
851 		wmb();
852 		e->gen2 = V_CMD_GEN2(q->genbit);
853 
854 		e++;
855 		ce++;
856 		if (++q->pidx == q->size) {
857 			q->pidx = 0;
858 			q->genbit ^= 1;
859 			ce = q->centries;
860 			e = q->entries;
861 		}
862 		q->credits++;
863 	}
864 }
865 
866 /*
867  * Calls refill_free_list for both free lists. If we cannot fill at least 1/4
868  * of both rings, we go into 'few interrupt mode' in order to give the system
869  * time to free up resources.
870  */
871 static void freelQs_empty(struct sge *sge)
872 {
873 	struct adapter *adapter = sge->adapter;
874 	u32 irq_reg = readl(adapter->regs + A_SG_INT_ENABLE);
875 	u32 irqholdoff_reg;
876 
877 	refill_free_list(sge, &sge->freelQ[0]);
878 	refill_free_list(sge, &sge->freelQ[1]);
879 
880 	if (sge->freelQ[0].credits > (sge->freelQ[0].size >> 2) &&
881 	    sge->freelQ[1].credits > (sge->freelQ[1].size >> 2)) {
882 		irq_reg |= F_FL_EXHAUSTED;
883 		irqholdoff_reg = sge->fixed_intrtimer;
884 	} else {
885 		/* Clear the F_FL_EXHAUSTED interrupts for now */
886 		irq_reg &= ~F_FL_EXHAUSTED;
887 		irqholdoff_reg = sge->intrtimer_nres;
888 	}
889 	writel(irqholdoff_reg, adapter->regs + A_SG_INTRTIMER);
890 	writel(irq_reg, adapter->regs + A_SG_INT_ENABLE);
891 
892 	/* We reenable the Qs to force a freelist GTS interrupt later */
893 	doorbell_pio(adapter, F_FL0_ENABLE | F_FL1_ENABLE);
894 }
895 
896 #define SGE_PL_INTR_MASK (F_PL_INTR_SGE_ERR | F_PL_INTR_SGE_DATA)
897 #define SGE_INT_FATAL (F_RESPQ_OVERFLOW | F_PACKET_TOO_BIG | F_PACKET_MISMATCH)
898 #define SGE_INT_ENABLE (F_RESPQ_EXHAUSTED | F_RESPQ_OVERFLOW | \
899 			F_FL_EXHAUSTED | F_PACKET_TOO_BIG | F_PACKET_MISMATCH)
900 
901 /*
902  * Disable SGE Interrupts
903  */
904 void t1_sge_intr_disable(struct sge *sge)
905 {
906 	u32 val = readl(sge->adapter->regs + A_PL_ENABLE);
907 
908 	writel(val & ~SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_ENABLE);
909 	writel(0, sge->adapter->regs + A_SG_INT_ENABLE);
910 }
911 
912 /*
913  * Enable SGE interrupts.
914  */
915 void t1_sge_intr_enable(struct sge *sge)
916 {
917 	u32 en = SGE_INT_ENABLE;
918 	u32 val = readl(sge->adapter->regs + A_PL_ENABLE);
919 
920 	if (sge->adapter->port[0].dev->hw_features & NETIF_F_TSO)
921 		en &= ~F_PACKET_TOO_BIG;
922 	writel(en, sge->adapter->regs + A_SG_INT_ENABLE);
923 	writel(val | SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_ENABLE);
924 }
925 
926 /*
927  * Clear SGE interrupts.
928  */
929 void t1_sge_intr_clear(struct sge *sge)
930 {
931 	writel(SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_CAUSE);
932 	writel(0xffffffff, sge->adapter->regs + A_SG_INT_CAUSE);
933 }
934 
935 /*
936  * SGE 'Error' interrupt handler
937  */
938 int t1_sge_intr_error_handler(struct sge *sge)
939 {
940 	struct adapter *adapter = sge->adapter;
941 	u32 cause = readl(adapter->regs + A_SG_INT_CAUSE);
942 
943 	if (adapter->port[0].dev->hw_features & NETIF_F_TSO)
944 		cause &= ~F_PACKET_TOO_BIG;
945 	if (cause & F_RESPQ_EXHAUSTED)
946 		sge->stats.respQ_empty++;
947 	if (cause & F_RESPQ_OVERFLOW) {
948 		sge->stats.respQ_overflow++;
949 		pr_alert("%s: SGE response queue overflow\n",
950 			 adapter->name);
951 	}
952 	if (cause & F_FL_EXHAUSTED) {
953 		sge->stats.freelistQ_empty++;
954 		freelQs_empty(sge);
955 	}
956 	if (cause & F_PACKET_TOO_BIG) {
957 		sge->stats.pkt_too_big++;
958 		pr_alert("%s: SGE max packet size exceeded\n",
959 			 adapter->name);
960 	}
961 	if (cause & F_PACKET_MISMATCH) {
962 		sge->stats.pkt_mismatch++;
963 		pr_alert("%s: SGE packet mismatch\n", adapter->name);
964 	}
965 	if (cause & SGE_INT_FATAL)
966 		t1_fatal_err(adapter);
967 
968 	writel(cause, adapter->regs + A_SG_INT_CAUSE);
969 	return 0;
970 }
971 
972 const struct sge_intr_counts *t1_sge_get_intr_counts(const struct sge *sge)
973 {
974 	return &sge->stats;
975 }
976 
977 void t1_sge_get_port_stats(const struct sge *sge, int port,
978 			   struct sge_port_stats *ss)
979 {
980 	int cpu;
981 
982 	memset(ss, 0, sizeof(*ss));
983 	for_each_possible_cpu(cpu) {
984 		struct sge_port_stats *st = per_cpu_ptr(sge->port_stats[port], cpu);
985 
986 		ss->rx_cso_good += st->rx_cso_good;
987 		ss->tx_cso += st->tx_cso;
988 		ss->tx_tso += st->tx_tso;
989 		ss->tx_need_hdrroom += st->tx_need_hdrroom;
990 		ss->vlan_xtract += st->vlan_xtract;
991 		ss->vlan_insert += st->vlan_insert;
992 	}
993 }
994 
995 /**
996  *	recycle_fl_buf - recycle a free list buffer
997  *	@fl: the free list
998  *	@idx: index of buffer to recycle
999  *
1000  *	Recycles the specified buffer on the given free list by adding it at
1001  *	the next available slot on the list.
1002  */
1003 static void recycle_fl_buf(struct freelQ *fl, int idx)
1004 {
1005 	struct freelQ_e *from = &fl->entries[idx];
1006 	struct freelQ_e *to = &fl->entries[fl->pidx];
1007 
1008 	fl->centries[fl->pidx] = fl->centries[idx];
1009 	to->addr_lo = from->addr_lo;
1010 	to->addr_hi = from->addr_hi;
1011 	to->len_gen = G_CMD_LEN(from->len_gen) | V_CMD_GEN1(fl->genbit);
1012 	wmb();
1013 	to->gen2 = V_CMD_GEN2(fl->genbit);
1014 	fl->credits++;
1015 
1016 	if (++fl->pidx == fl->size) {
1017 		fl->pidx = 0;
1018 		fl->genbit ^= 1;
1019 	}
1020 }
1021 
1022 static int copybreak __read_mostly = 256;
1023 module_param(copybreak, int, 0);
1024 MODULE_PARM_DESC(copybreak, "Receive copy threshold");
1025 
1026 /**
1027  *	get_packet - return the next ingress packet buffer
1028  *	@pdev: the PCI device that received the packet
1029  *	@fl: the SGE free list holding the packet
1030  *	@len: the actual packet length, excluding any SGE padding
1031  *
1032  *	Get the next packet from a free list and complete setup of the
1033  *	sk_buff.  If the packet is small we make a copy and recycle the
1034  *	original buffer, otherwise we use the original buffer itself.  If a
1035  *	positive drop threshold is supplied packets are dropped and their
1036  *	buffers recycled if (a) the number of remaining buffers is under the
1037  *	threshold and the packet is too big to copy, or (b) the packet should
1038  *	be copied but there is no memory for the copy.
1039  */
1040 static inline struct sk_buff *get_packet(struct pci_dev *pdev,
1041 					 struct freelQ *fl, unsigned int len)
1042 {
1043 	struct sk_buff *skb;
1044 	const struct freelQ_ce *ce = &fl->centries[fl->cidx];
1045 
1046 	if (len < copybreak) {
1047 		skb = netdev_alloc_skb_ip_align(NULL, len);
1048 		if (!skb)
1049 			goto use_orig_buf;
1050 
1051 		skb_put(skb, len);
1052 		pci_dma_sync_single_for_cpu(pdev,
1053 					    dma_unmap_addr(ce, dma_addr),
1054 					    dma_unmap_len(ce, dma_len),
1055 					    PCI_DMA_FROMDEVICE);
1056 		skb_copy_from_linear_data(ce->skb, skb->data, len);
1057 		pci_dma_sync_single_for_device(pdev,
1058 					       dma_unmap_addr(ce, dma_addr),
1059 					       dma_unmap_len(ce, dma_len),
1060 					       PCI_DMA_FROMDEVICE);
1061 		recycle_fl_buf(fl, fl->cidx);
1062 		return skb;
1063 	}
1064 
1065 use_orig_buf:
1066 	if (fl->credits < 2) {
1067 		recycle_fl_buf(fl, fl->cidx);
1068 		return NULL;
1069 	}
1070 
1071 	pci_unmap_single(pdev, dma_unmap_addr(ce, dma_addr),
1072 			 dma_unmap_len(ce, dma_len), PCI_DMA_FROMDEVICE);
1073 	skb = ce->skb;
1074 	prefetch(skb->data);
1075 
1076 	skb_put(skb, len);
1077 	return skb;
1078 }
1079 
1080 /**
1081  *	unexpected_offload - handle an unexpected offload packet
1082  *	@adapter: the adapter
1083  *	@fl: the free list that received the packet
1084  *
1085  *	Called when we receive an unexpected offload packet (e.g., the TOE
1086  *	function is disabled or the card is a NIC).  Prints a message and
1087  *	recycles the buffer.
1088  */
1089 static void unexpected_offload(struct adapter *adapter, struct freelQ *fl)
1090 {
1091 	struct freelQ_ce *ce = &fl->centries[fl->cidx];
1092 	struct sk_buff *skb = ce->skb;
1093 
1094 	pci_dma_sync_single_for_cpu(adapter->pdev, dma_unmap_addr(ce, dma_addr),
1095 			    dma_unmap_len(ce, dma_len), PCI_DMA_FROMDEVICE);
1096 	pr_err("%s: unexpected offload packet, cmd %u\n",
1097 	       adapter->name, *skb->data);
1098 	recycle_fl_buf(fl, fl->cidx);
1099 }
1100 
1101 /*
1102  * T1/T2 SGE limits the maximum DMA size per TX descriptor to
1103  * SGE_TX_DESC_MAX_PLEN (16KB). If the PAGE_SIZE is larger than 16KB, the
1104  * stack might send more than SGE_TX_DESC_MAX_PLEN in a contiguous manner.
1105  * Note that the *_large_page_tx_descs stuff will be optimized out when
1106  * PAGE_SIZE <= SGE_TX_DESC_MAX_PLEN.
1107  *
1108  * compute_large_page_descs() computes how many additional descriptors are
1109  * required to break down the stack's request.
1110  */
1111 static inline unsigned int compute_large_page_tx_descs(struct sk_buff *skb)
1112 {
1113 	unsigned int count = 0;
1114 
1115 	if (PAGE_SIZE > SGE_TX_DESC_MAX_PLEN) {
1116 		unsigned int nfrags = skb_shinfo(skb)->nr_frags;
1117 		unsigned int i, len = skb_headlen(skb);
1118 		while (len > SGE_TX_DESC_MAX_PLEN) {
1119 			count++;
1120 			len -= SGE_TX_DESC_MAX_PLEN;
1121 		}
1122 		for (i = 0; nfrags--; i++) {
1123 			const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
1124 			len = skb_frag_size(frag);
1125 			while (len > SGE_TX_DESC_MAX_PLEN) {
1126 				count++;
1127 				len -= SGE_TX_DESC_MAX_PLEN;
1128 			}
1129 		}
1130 	}
1131 	return count;
1132 }
1133 
1134 /*
1135  * Write a cmdQ entry.
1136  *
1137  * Since this function writes the 'flags' field, it must not be used to
1138  * write the first cmdQ entry.
1139  */
1140 static inline void write_tx_desc(struct cmdQ_e *e, dma_addr_t mapping,
1141 				 unsigned int len, unsigned int gen,
1142 				 unsigned int eop)
1143 {
1144 	BUG_ON(len > SGE_TX_DESC_MAX_PLEN);
1145 
1146 	e->addr_lo = (u32)mapping;
1147 	e->addr_hi = (u64)mapping >> 32;
1148 	e->len_gen = V_CMD_LEN(len) | V_CMD_GEN1(gen);
1149 	e->flags = F_CMD_DATAVALID | V_CMD_EOP(eop) | V_CMD_GEN2(gen);
1150 }
1151 
1152 /*
1153  * See comment for previous function.
1154  *
1155  * write_tx_descs_large_page() writes additional SGE tx descriptors if
1156  * *desc_len exceeds HW's capability.
1157  */
1158 static inline unsigned int write_large_page_tx_descs(unsigned int pidx,
1159 						     struct cmdQ_e **e,
1160 						     struct cmdQ_ce **ce,
1161 						     unsigned int *gen,
1162 						     dma_addr_t *desc_mapping,
1163 						     unsigned int *desc_len,
1164 						     unsigned int nfrags,
1165 						     struct cmdQ *q)
1166 {
1167 	if (PAGE_SIZE > SGE_TX_DESC_MAX_PLEN) {
1168 		struct cmdQ_e *e1 = *e;
1169 		struct cmdQ_ce *ce1 = *ce;
1170 
1171 		while (*desc_len > SGE_TX_DESC_MAX_PLEN) {
1172 			*desc_len -= SGE_TX_DESC_MAX_PLEN;
1173 			write_tx_desc(e1, *desc_mapping, SGE_TX_DESC_MAX_PLEN,
1174 				      *gen, nfrags == 0 && *desc_len == 0);
1175 			ce1->skb = NULL;
1176 			dma_unmap_len_set(ce1, dma_len, 0);
1177 			*desc_mapping += SGE_TX_DESC_MAX_PLEN;
1178 			if (*desc_len) {
1179 				ce1++;
1180 				e1++;
1181 				if (++pidx == q->size) {
1182 					pidx = 0;
1183 					*gen ^= 1;
1184 					ce1 = q->centries;
1185 					e1 = q->entries;
1186 				}
1187 			}
1188 		}
1189 		*e = e1;
1190 		*ce = ce1;
1191 	}
1192 	return pidx;
1193 }
1194 
1195 /*
1196  * Write the command descriptors to transmit the given skb starting at
1197  * descriptor pidx with the given generation.
1198  */
1199 static inline void write_tx_descs(struct adapter *adapter, struct sk_buff *skb,
1200 				  unsigned int pidx, unsigned int gen,
1201 				  struct cmdQ *q)
1202 {
1203 	dma_addr_t mapping, desc_mapping;
1204 	struct cmdQ_e *e, *e1;
1205 	struct cmdQ_ce *ce;
1206 	unsigned int i, flags, first_desc_len, desc_len,
1207 	    nfrags = skb_shinfo(skb)->nr_frags;
1208 
1209 	e = e1 = &q->entries[pidx];
1210 	ce = &q->centries[pidx];
1211 
1212 	mapping = pci_map_single(adapter->pdev, skb->data,
1213 				 skb_headlen(skb), PCI_DMA_TODEVICE);
1214 
1215 	desc_mapping = mapping;
1216 	desc_len = skb_headlen(skb);
1217 
1218 	flags = F_CMD_DATAVALID | F_CMD_SOP |
1219 	    V_CMD_EOP(nfrags == 0 && desc_len <= SGE_TX_DESC_MAX_PLEN) |
1220 	    V_CMD_GEN2(gen);
1221 	first_desc_len = (desc_len <= SGE_TX_DESC_MAX_PLEN) ?
1222 	    desc_len : SGE_TX_DESC_MAX_PLEN;
1223 	e->addr_lo = (u32)desc_mapping;
1224 	e->addr_hi = (u64)desc_mapping >> 32;
1225 	e->len_gen = V_CMD_LEN(first_desc_len) | V_CMD_GEN1(gen);
1226 	ce->skb = NULL;
1227 	dma_unmap_len_set(ce, dma_len, 0);
1228 
1229 	if (PAGE_SIZE > SGE_TX_DESC_MAX_PLEN &&
1230 	    desc_len > SGE_TX_DESC_MAX_PLEN) {
1231 		desc_mapping += first_desc_len;
1232 		desc_len -= first_desc_len;
1233 		e1++;
1234 		ce++;
1235 		if (++pidx == q->size) {
1236 			pidx = 0;
1237 			gen ^= 1;
1238 			e1 = q->entries;
1239 			ce = q->centries;
1240 		}
1241 		pidx = write_large_page_tx_descs(pidx, &e1, &ce, &gen,
1242 						 &desc_mapping, &desc_len,
1243 						 nfrags, q);
1244 
1245 		if (likely(desc_len))
1246 			write_tx_desc(e1, desc_mapping, desc_len, gen,
1247 				      nfrags == 0);
1248 	}
1249 
1250 	ce->skb = NULL;
1251 	dma_unmap_addr_set(ce, dma_addr, mapping);
1252 	dma_unmap_len_set(ce, dma_len, skb_headlen(skb));
1253 
1254 	for (i = 0; nfrags--; i++) {
1255 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
1256 		e1++;
1257 		ce++;
1258 		if (++pidx == q->size) {
1259 			pidx = 0;
1260 			gen ^= 1;
1261 			e1 = q->entries;
1262 			ce = q->centries;
1263 		}
1264 
1265 		mapping = skb_frag_dma_map(&adapter->pdev->dev, frag, 0,
1266 					   skb_frag_size(frag), DMA_TO_DEVICE);
1267 		desc_mapping = mapping;
1268 		desc_len = skb_frag_size(frag);
1269 
1270 		pidx = write_large_page_tx_descs(pidx, &e1, &ce, &gen,
1271 						 &desc_mapping, &desc_len,
1272 						 nfrags, q);
1273 		if (likely(desc_len))
1274 			write_tx_desc(e1, desc_mapping, desc_len, gen,
1275 				      nfrags == 0);
1276 		ce->skb = NULL;
1277 		dma_unmap_addr_set(ce, dma_addr, mapping);
1278 		dma_unmap_len_set(ce, dma_len, skb_frag_size(frag));
1279 	}
1280 	ce->skb = skb;
1281 	wmb();
1282 	e->flags = flags;
1283 }
1284 
1285 /*
1286  * Clean up completed Tx buffers.
1287  */
1288 static inline void reclaim_completed_tx(struct sge *sge, struct cmdQ *q)
1289 {
1290 	unsigned int reclaim = q->processed - q->cleaned;
1291 
1292 	if (reclaim) {
1293 		pr_debug("reclaim_completed_tx processed:%d cleaned:%d\n",
1294 			 q->processed, q->cleaned);
1295 		free_cmdQ_buffers(sge, q, reclaim);
1296 		q->cleaned += reclaim;
1297 	}
1298 }
1299 
1300 /*
1301  * Called from tasklet. Checks the scheduler for any
1302  * pending skbs that can be sent.
1303  */
1304 static void restart_sched(unsigned long arg)
1305 {
1306 	struct sge *sge = (struct sge *) arg;
1307 	struct adapter *adapter = sge->adapter;
1308 	struct cmdQ *q = &sge->cmdQ[0];
1309 	struct sk_buff *skb;
1310 	unsigned int credits, queued_skb = 0;
1311 
1312 	spin_lock(&q->lock);
1313 	reclaim_completed_tx(sge, q);
1314 
1315 	credits = q->size - q->in_use;
1316 	pr_debug("restart_sched credits=%d\n", credits);
1317 	while ((skb = sched_skb(sge, NULL, credits)) != NULL) {
1318 		unsigned int genbit, pidx, count;
1319 	        count = 1 + skb_shinfo(skb)->nr_frags;
1320 		count += compute_large_page_tx_descs(skb);
1321 		q->in_use += count;
1322 		genbit = q->genbit;
1323 		pidx = q->pidx;
1324 		q->pidx += count;
1325 		if (q->pidx >= q->size) {
1326 			q->pidx -= q->size;
1327 			q->genbit ^= 1;
1328 		}
1329 		write_tx_descs(adapter, skb, pidx, genbit, q);
1330 	        credits = q->size - q->in_use;
1331 		queued_skb = 1;
1332 	}
1333 
1334 	if (queued_skb) {
1335 		clear_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1336 		if (test_and_set_bit(CMDQ_STAT_RUNNING, &q->status) == 0) {
1337 			set_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1338 			writel(F_CMDQ0_ENABLE, adapter->regs + A_SG_DOORBELL);
1339 		}
1340 	}
1341 	spin_unlock(&q->lock);
1342 }
1343 
1344 /**
1345  *	sge_rx - process an ingress ethernet packet
1346  *	@sge: the sge structure
1347  *	@fl: the free list that contains the packet buffer
1348  *	@len: the packet length
1349  *
1350  *	Process an ingress ethernet pakcet and deliver it to the stack.
1351  */
1352 static void sge_rx(struct sge *sge, struct freelQ *fl, unsigned int len)
1353 {
1354 	struct sk_buff *skb;
1355 	const struct cpl_rx_pkt *p;
1356 	struct adapter *adapter = sge->adapter;
1357 	struct sge_port_stats *st;
1358 	struct net_device *dev;
1359 
1360 	skb = get_packet(adapter->pdev, fl, len - sge->rx_pkt_pad);
1361 	if (unlikely(!skb)) {
1362 		sge->stats.rx_drops++;
1363 		return;
1364 	}
1365 
1366 	p = (const struct cpl_rx_pkt *) skb->data;
1367 	if (p->iff >= adapter->params.nports) {
1368 		kfree_skb(skb);
1369 		return;
1370 	}
1371 	__skb_pull(skb, sizeof(*p));
1372 
1373 	st = this_cpu_ptr(sge->port_stats[p->iff]);
1374 	dev = adapter->port[p->iff].dev;
1375 
1376 	skb->protocol = eth_type_trans(skb, dev);
1377 	if ((dev->features & NETIF_F_RXCSUM) && p->csum == 0xffff &&
1378 	    skb->protocol == htons(ETH_P_IP) &&
1379 	    (skb->data[9] == IPPROTO_TCP || skb->data[9] == IPPROTO_UDP)) {
1380 		++st->rx_cso_good;
1381 		skb->ip_summed = CHECKSUM_UNNECESSARY;
1382 	} else
1383 		skb_checksum_none_assert(skb);
1384 
1385 	if (p->vlan_valid) {
1386 		st->vlan_xtract++;
1387 		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), ntohs(p->vlan));
1388 	}
1389 	netif_receive_skb(skb);
1390 }
1391 
1392 /*
1393  * Returns true if a command queue has enough available descriptors that
1394  * we can resume Tx operation after temporarily disabling its packet queue.
1395  */
1396 static inline int enough_free_Tx_descs(const struct cmdQ *q)
1397 {
1398 	unsigned int r = q->processed - q->cleaned;
1399 
1400 	return q->in_use - r < (q->size >> 1);
1401 }
1402 
1403 /*
1404  * Called when sufficient space has become available in the SGE command queues
1405  * after the Tx packet schedulers have been suspended to restart the Tx path.
1406  */
1407 static void restart_tx_queues(struct sge *sge)
1408 {
1409 	struct adapter *adap = sge->adapter;
1410 	int i;
1411 
1412 	if (!enough_free_Tx_descs(&sge->cmdQ[0]))
1413 		return;
1414 
1415 	for_each_port(adap, i) {
1416 		struct net_device *nd = adap->port[i].dev;
1417 
1418 		if (test_and_clear_bit(nd->if_port, &sge->stopped_tx_queues) &&
1419 		    netif_running(nd)) {
1420 			sge->stats.cmdQ_restarted[2]++;
1421 			netif_wake_queue(nd);
1422 		}
1423 	}
1424 }
1425 
1426 /*
1427  * update_tx_info is called from the interrupt handler/NAPI to return cmdQ0
1428  * information.
1429  */
1430 static unsigned int update_tx_info(struct adapter *adapter,
1431 					  unsigned int flags,
1432 					  unsigned int pr0)
1433 {
1434 	struct sge *sge = adapter->sge;
1435 	struct cmdQ *cmdq = &sge->cmdQ[0];
1436 
1437 	cmdq->processed += pr0;
1438 	if (flags & (F_FL0_ENABLE | F_FL1_ENABLE)) {
1439 		freelQs_empty(sge);
1440 		flags &= ~(F_FL0_ENABLE | F_FL1_ENABLE);
1441 	}
1442 	if (flags & F_CMDQ0_ENABLE) {
1443 		clear_bit(CMDQ_STAT_RUNNING, &cmdq->status);
1444 
1445 		if (cmdq->cleaned + cmdq->in_use != cmdq->processed &&
1446 		    !test_and_set_bit(CMDQ_STAT_LAST_PKT_DB, &cmdq->status)) {
1447 			set_bit(CMDQ_STAT_RUNNING, &cmdq->status);
1448 			writel(F_CMDQ0_ENABLE, adapter->regs + A_SG_DOORBELL);
1449 		}
1450 		if (sge->tx_sched)
1451 			tasklet_hi_schedule(&sge->tx_sched->sched_tsk);
1452 
1453 		flags &= ~F_CMDQ0_ENABLE;
1454 	}
1455 
1456 	if (unlikely(sge->stopped_tx_queues != 0))
1457 		restart_tx_queues(sge);
1458 
1459 	return flags;
1460 }
1461 
1462 /*
1463  * Process SGE responses, up to the supplied budget.  Returns the number of
1464  * responses processed.  A negative budget is effectively unlimited.
1465  */
1466 static int process_responses(struct adapter *adapter, int budget)
1467 {
1468 	struct sge *sge = adapter->sge;
1469 	struct respQ *q = &sge->respQ;
1470 	struct respQ_e *e = &q->entries[q->cidx];
1471 	int done = 0;
1472 	unsigned int flags = 0;
1473 	unsigned int cmdq_processed[SGE_CMDQ_N] = {0, 0};
1474 
1475 	while (done < budget && e->GenerationBit == q->genbit) {
1476 		flags |= e->Qsleeping;
1477 
1478 		cmdq_processed[0] += e->Cmdq0CreditReturn;
1479 		cmdq_processed[1] += e->Cmdq1CreditReturn;
1480 
1481 		/* We batch updates to the TX side to avoid cacheline
1482 		 * ping-pong of TX state information on MP where the sender
1483 		 * might run on a different CPU than this function...
1484 		 */
1485 		if (unlikely((flags & F_CMDQ0_ENABLE) || cmdq_processed[0] > 64)) {
1486 			flags = update_tx_info(adapter, flags, cmdq_processed[0]);
1487 			cmdq_processed[0] = 0;
1488 		}
1489 
1490 		if (unlikely(cmdq_processed[1] > 16)) {
1491 			sge->cmdQ[1].processed += cmdq_processed[1];
1492 			cmdq_processed[1] = 0;
1493 		}
1494 
1495 		if (likely(e->DataValid)) {
1496 			struct freelQ *fl = &sge->freelQ[e->FreelistQid];
1497 
1498 			BUG_ON(!e->Sop || !e->Eop);
1499 			if (unlikely(e->Offload))
1500 				unexpected_offload(adapter, fl);
1501 			else
1502 				sge_rx(sge, fl, e->BufferLength);
1503 
1504 			++done;
1505 
1506 			/*
1507 			 * Note: this depends on each packet consuming a
1508 			 * single free-list buffer; cf. the BUG above.
1509 			 */
1510 			if (++fl->cidx == fl->size)
1511 				fl->cidx = 0;
1512 			prefetch(fl->centries[fl->cidx].skb);
1513 
1514 			if (unlikely(--fl->credits <
1515 				     fl->size - SGE_FREEL_REFILL_THRESH))
1516 				refill_free_list(sge, fl);
1517 		} else
1518 			sge->stats.pure_rsps++;
1519 
1520 		e++;
1521 		if (unlikely(++q->cidx == q->size)) {
1522 			q->cidx = 0;
1523 			q->genbit ^= 1;
1524 			e = q->entries;
1525 		}
1526 		prefetch(e);
1527 
1528 		if (++q->credits > SGE_RESPQ_REPLENISH_THRES) {
1529 			writel(q->credits, adapter->regs + A_SG_RSPQUEUECREDIT);
1530 			q->credits = 0;
1531 		}
1532 	}
1533 
1534 	flags = update_tx_info(adapter, flags, cmdq_processed[0]);
1535 	sge->cmdQ[1].processed += cmdq_processed[1];
1536 
1537 	return done;
1538 }
1539 
1540 static inline int responses_pending(const struct adapter *adapter)
1541 {
1542 	const struct respQ *Q = &adapter->sge->respQ;
1543 	const struct respQ_e *e = &Q->entries[Q->cidx];
1544 
1545 	return e->GenerationBit == Q->genbit;
1546 }
1547 
1548 /*
1549  * A simpler version of process_responses() that handles only pure (i.e.,
1550  * non data-carrying) responses.  Such respones are too light-weight to justify
1551  * calling a softirq when using NAPI, so we handle them specially in hard
1552  * interrupt context.  The function is called with a pointer to a response,
1553  * which the caller must ensure is a valid pure response.  Returns 1 if it
1554  * encounters a valid data-carrying response, 0 otherwise.
1555  */
1556 static int process_pure_responses(struct adapter *adapter)
1557 {
1558 	struct sge *sge = adapter->sge;
1559 	struct respQ *q = &sge->respQ;
1560 	struct respQ_e *e = &q->entries[q->cidx];
1561 	const struct freelQ *fl = &sge->freelQ[e->FreelistQid];
1562 	unsigned int flags = 0;
1563 	unsigned int cmdq_processed[SGE_CMDQ_N] = {0, 0};
1564 
1565 	prefetch(fl->centries[fl->cidx].skb);
1566 	if (e->DataValid)
1567 		return 1;
1568 
1569 	do {
1570 		flags |= e->Qsleeping;
1571 
1572 		cmdq_processed[0] += e->Cmdq0CreditReturn;
1573 		cmdq_processed[1] += e->Cmdq1CreditReturn;
1574 
1575 		e++;
1576 		if (unlikely(++q->cidx == q->size)) {
1577 			q->cidx = 0;
1578 			q->genbit ^= 1;
1579 			e = q->entries;
1580 		}
1581 		prefetch(e);
1582 
1583 		if (++q->credits > SGE_RESPQ_REPLENISH_THRES) {
1584 			writel(q->credits, adapter->regs + A_SG_RSPQUEUECREDIT);
1585 			q->credits = 0;
1586 		}
1587 		sge->stats.pure_rsps++;
1588 	} while (e->GenerationBit == q->genbit && !e->DataValid);
1589 
1590 	flags = update_tx_info(adapter, flags, cmdq_processed[0]);
1591 	sge->cmdQ[1].processed += cmdq_processed[1];
1592 
1593 	return e->GenerationBit == q->genbit;
1594 }
1595 
1596 /*
1597  * Handler for new data events when using NAPI.  This does not need any locking
1598  * or protection from interrupts as data interrupts are off at this point and
1599  * other adapter interrupts do not interfere.
1600  */
1601 int t1_poll(struct napi_struct *napi, int budget)
1602 {
1603 	struct adapter *adapter = container_of(napi, struct adapter, napi);
1604 	int work_done = process_responses(adapter, budget);
1605 
1606 	if (likely(work_done < budget)) {
1607 		napi_complete(napi);
1608 		writel(adapter->sge->respQ.cidx,
1609 		       adapter->regs + A_SG_SLEEPING);
1610 	}
1611 	return work_done;
1612 }
1613 
1614 irqreturn_t t1_interrupt(int irq, void *data)
1615 {
1616 	struct adapter *adapter = data;
1617 	struct sge *sge = adapter->sge;
1618 	int handled;
1619 
1620 	if (likely(responses_pending(adapter))) {
1621 		writel(F_PL_INTR_SGE_DATA, adapter->regs + A_PL_CAUSE);
1622 
1623 		if (napi_schedule_prep(&adapter->napi)) {
1624 			if (process_pure_responses(adapter))
1625 				__napi_schedule(&adapter->napi);
1626 			else {
1627 				/* no data, no NAPI needed */
1628 				writel(sge->respQ.cidx, adapter->regs + A_SG_SLEEPING);
1629 				/* undo schedule_prep */
1630 				napi_enable(&adapter->napi);
1631 			}
1632 		}
1633 		return IRQ_HANDLED;
1634 	}
1635 
1636 	spin_lock(&adapter->async_lock);
1637 	handled = t1_slow_intr_handler(adapter);
1638 	spin_unlock(&adapter->async_lock);
1639 
1640 	if (!handled)
1641 		sge->stats.unhandled_irqs++;
1642 
1643 	return IRQ_RETVAL(handled != 0);
1644 }
1645 
1646 /*
1647  * Enqueues the sk_buff onto the cmdQ[qid] and has hardware fetch it.
1648  *
1649  * The code figures out how many entries the sk_buff will require in the
1650  * cmdQ and updates the cmdQ data structure with the state once the enqueue
1651  * has complete. Then, it doesn't access the global structure anymore, but
1652  * uses the corresponding fields on the stack. In conjunction with a spinlock
1653  * around that code, we can make the function reentrant without holding the
1654  * lock when we actually enqueue (which might be expensive, especially on
1655  * architectures with IO MMUs).
1656  *
1657  * This runs with softirqs disabled.
1658  */
1659 static int t1_sge_tx(struct sk_buff *skb, struct adapter *adapter,
1660 		     unsigned int qid, struct net_device *dev)
1661 {
1662 	struct sge *sge = adapter->sge;
1663 	struct cmdQ *q = &sge->cmdQ[qid];
1664 	unsigned int credits, pidx, genbit, count, use_sched_skb = 0;
1665 
1666 	if (!spin_trylock(&q->lock))
1667 		return NETDEV_TX_LOCKED;
1668 
1669 	reclaim_completed_tx(sge, q);
1670 
1671 	pidx = q->pidx;
1672 	credits = q->size - q->in_use;
1673 	count = 1 + skb_shinfo(skb)->nr_frags;
1674 	count += compute_large_page_tx_descs(skb);
1675 
1676 	/* Ethernet packet */
1677 	if (unlikely(credits < count)) {
1678 		if (!netif_queue_stopped(dev)) {
1679 			netif_stop_queue(dev);
1680 			set_bit(dev->if_port, &sge->stopped_tx_queues);
1681 			sge->stats.cmdQ_full[2]++;
1682 			pr_err("%s: Tx ring full while queue awake!\n",
1683 			       adapter->name);
1684 		}
1685 		spin_unlock(&q->lock);
1686 		return NETDEV_TX_BUSY;
1687 	}
1688 
1689 	if (unlikely(credits - count < q->stop_thres)) {
1690 		netif_stop_queue(dev);
1691 		set_bit(dev->if_port, &sge->stopped_tx_queues);
1692 		sge->stats.cmdQ_full[2]++;
1693 	}
1694 
1695 	/* T204 cmdQ0 skbs that are destined for a certain port have to go
1696 	 * through the scheduler.
1697 	 */
1698 	if (sge->tx_sched && !qid && skb->dev) {
1699 use_sched:
1700 		use_sched_skb = 1;
1701 		/* Note that the scheduler might return a different skb than
1702 		 * the one passed in.
1703 		 */
1704 		skb = sched_skb(sge, skb, credits);
1705 		if (!skb) {
1706 			spin_unlock(&q->lock);
1707 			return NETDEV_TX_OK;
1708 		}
1709 		pidx = q->pidx;
1710 		count = 1 + skb_shinfo(skb)->nr_frags;
1711 		count += compute_large_page_tx_descs(skb);
1712 	}
1713 
1714 	q->in_use += count;
1715 	genbit = q->genbit;
1716 	pidx = q->pidx;
1717 	q->pidx += count;
1718 	if (q->pidx >= q->size) {
1719 		q->pidx -= q->size;
1720 		q->genbit ^= 1;
1721 	}
1722 	spin_unlock(&q->lock);
1723 
1724 	write_tx_descs(adapter, skb, pidx, genbit, q);
1725 
1726 	/*
1727 	 * We always ring the doorbell for cmdQ1.  For cmdQ0, we only ring
1728 	 * the doorbell if the Q is asleep. There is a natural race, where
1729 	 * the hardware is going to sleep just after we checked, however,
1730 	 * then the interrupt handler will detect the outstanding TX packet
1731 	 * and ring the doorbell for us.
1732 	 */
1733 	if (qid)
1734 		doorbell_pio(adapter, F_CMDQ1_ENABLE);
1735 	else {
1736 		clear_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1737 		if (test_and_set_bit(CMDQ_STAT_RUNNING, &q->status) == 0) {
1738 			set_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1739 			writel(F_CMDQ0_ENABLE, adapter->regs + A_SG_DOORBELL);
1740 		}
1741 	}
1742 
1743 	if (use_sched_skb) {
1744 		if (spin_trylock(&q->lock)) {
1745 			credits = q->size - q->in_use;
1746 			skb = NULL;
1747 			goto use_sched;
1748 		}
1749 	}
1750 	return NETDEV_TX_OK;
1751 }
1752 
1753 #define MK_ETH_TYPE_MSS(type, mss) (((mss) & 0x3FFF) | ((type) << 14))
1754 
1755 /*
1756  *	eth_hdr_len - return the length of an Ethernet header
1757  *	@data: pointer to the start of the Ethernet header
1758  *
1759  *	Returns the length of an Ethernet header, including optional VLAN tag.
1760  */
1761 static inline int eth_hdr_len(const void *data)
1762 {
1763 	const struct ethhdr *e = data;
1764 
1765 	return e->h_proto == htons(ETH_P_8021Q) ? VLAN_ETH_HLEN : ETH_HLEN;
1766 }
1767 
1768 /*
1769  * Adds the CPL header to the sk_buff and passes it to t1_sge_tx.
1770  */
1771 netdev_tx_t t1_start_xmit(struct sk_buff *skb, struct net_device *dev)
1772 {
1773 	struct adapter *adapter = dev->ml_priv;
1774 	struct sge *sge = adapter->sge;
1775 	struct sge_port_stats *st = this_cpu_ptr(sge->port_stats[dev->if_port]);
1776 	struct cpl_tx_pkt *cpl;
1777 	struct sk_buff *orig_skb = skb;
1778 	int ret;
1779 
1780 	if (skb->protocol == htons(ETH_P_CPL5))
1781 		goto send;
1782 
1783 	/*
1784 	 * We are using a non-standard hard_header_len.
1785 	 * Allocate more header room in the rare cases it is not big enough.
1786 	 */
1787 	if (unlikely(skb_headroom(skb) < dev->hard_header_len - ETH_HLEN)) {
1788 		skb = skb_realloc_headroom(skb, sizeof(struct cpl_tx_pkt_lso));
1789 		++st->tx_need_hdrroom;
1790 		dev_kfree_skb_any(orig_skb);
1791 		if (!skb)
1792 			return NETDEV_TX_OK;
1793 	}
1794 
1795 	if (skb_shinfo(skb)->gso_size) {
1796 		int eth_type;
1797 		struct cpl_tx_pkt_lso *hdr;
1798 
1799 		++st->tx_tso;
1800 
1801 		eth_type = skb_network_offset(skb) == ETH_HLEN ?
1802 			CPL_ETH_II : CPL_ETH_II_VLAN;
1803 
1804 		hdr = (struct cpl_tx_pkt_lso *)skb_push(skb, sizeof(*hdr));
1805 		hdr->opcode = CPL_TX_PKT_LSO;
1806 		hdr->ip_csum_dis = hdr->l4_csum_dis = 0;
1807 		hdr->ip_hdr_words = ip_hdr(skb)->ihl;
1808 		hdr->tcp_hdr_words = tcp_hdr(skb)->doff;
1809 		hdr->eth_type_mss = htons(MK_ETH_TYPE_MSS(eth_type,
1810 							  skb_shinfo(skb)->gso_size));
1811 		hdr->len = htonl(skb->len - sizeof(*hdr));
1812 		cpl = (struct cpl_tx_pkt *)hdr;
1813 	} else {
1814 		/*
1815 		 * Packets shorter than ETH_HLEN can break the MAC, drop them
1816 		 * early.  Also, we may get oversized packets because some
1817 		 * parts of the kernel don't handle our unusual hard_header_len
1818 		 * right, drop those too.
1819 		 */
1820 		if (unlikely(skb->len < ETH_HLEN ||
1821 			     skb->len > dev->mtu + eth_hdr_len(skb->data))) {
1822 			netdev_dbg(dev, "packet size %d hdr %d mtu%d\n",
1823 				   skb->len, eth_hdr_len(skb->data), dev->mtu);
1824 			dev_kfree_skb_any(skb);
1825 			return NETDEV_TX_OK;
1826 		}
1827 
1828 		if (skb->ip_summed == CHECKSUM_PARTIAL &&
1829 		    ip_hdr(skb)->protocol == IPPROTO_UDP) {
1830 			if (unlikely(skb_checksum_help(skb))) {
1831 				netdev_dbg(dev, "unable to do udp checksum\n");
1832 				dev_kfree_skb_any(skb);
1833 				return NETDEV_TX_OK;
1834 			}
1835 		}
1836 
1837 		/* Hmmm, assuming to catch the gratious arp... and we'll use
1838 		 * it to flush out stuck espi packets...
1839 		 */
1840 		if ((unlikely(!adapter->sge->espibug_skb[dev->if_port]))) {
1841 			if (skb->protocol == htons(ETH_P_ARP) &&
1842 			    arp_hdr(skb)->ar_op == htons(ARPOP_REQUEST)) {
1843 				adapter->sge->espibug_skb[dev->if_port] = skb;
1844 				/* We want to re-use this skb later. We
1845 				 * simply bump the reference count and it
1846 				 * will not be freed...
1847 				 */
1848 				skb = skb_get(skb);
1849 			}
1850 		}
1851 
1852 		cpl = (struct cpl_tx_pkt *)__skb_push(skb, sizeof(*cpl));
1853 		cpl->opcode = CPL_TX_PKT;
1854 		cpl->ip_csum_dis = 1;    /* SW calculates IP csum */
1855 		cpl->l4_csum_dis = skb->ip_summed == CHECKSUM_PARTIAL ? 0 : 1;
1856 		/* the length field isn't used so don't bother setting it */
1857 
1858 		st->tx_cso += (skb->ip_summed == CHECKSUM_PARTIAL);
1859 	}
1860 	cpl->iff = dev->if_port;
1861 
1862 	if (vlan_tx_tag_present(skb)) {
1863 		cpl->vlan_valid = 1;
1864 		cpl->vlan = htons(vlan_tx_tag_get(skb));
1865 		st->vlan_insert++;
1866 	} else
1867 		cpl->vlan_valid = 0;
1868 
1869 send:
1870 	ret = t1_sge_tx(skb, adapter, 0, dev);
1871 
1872 	/* If transmit busy, and we reallocated skb's due to headroom limit,
1873 	 * then silently discard to avoid leak.
1874 	 */
1875 	if (unlikely(ret != NETDEV_TX_OK && skb != orig_skb)) {
1876 		dev_kfree_skb_any(skb);
1877 		ret = NETDEV_TX_OK;
1878 	}
1879 	return ret;
1880 }
1881 
1882 /*
1883  * Callback for the Tx buffer reclaim timer.  Runs with softirqs disabled.
1884  */
1885 static void sge_tx_reclaim_cb(unsigned long data)
1886 {
1887 	int i;
1888 	struct sge *sge = (struct sge *)data;
1889 
1890 	for (i = 0; i < SGE_CMDQ_N; ++i) {
1891 		struct cmdQ *q = &sge->cmdQ[i];
1892 
1893 		if (!spin_trylock(&q->lock))
1894 			continue;
1895 
1896 		reclaim_completed_tx(sge, q);
1897 		if (i == 0 && q->in_use) {    /* flush pending credits */
1898 			writel(F_CMDQ0_ENABLE, sge->adapter->regs + A_SG_DOORBELL);
1899 		}
1900 		spin_unlock(&q->lock);
1901 	}
1902 	mod_timer(&sge->tx_reclaim_timer, jiffies + TX_RECLAIM_PERIOD);
1903 }
1904 
1905 /*
1906  * Propagate changes of the SGE coalescing parameters to the HW.
1907  */
1908 int t1_sge_set_coalesce_params(struct sge *sge, struct sge_params *p)
1909 {
1910 	sge->fixed_intrtimer = p->rx_coalesce_usecs *
1911 		core_ticks_per_usec(sge->adapter);
1912 	writel(sge->fixed_intrtimer, sge->adapter->regs + A_SG_INTRTIMER);
1913 	return 0;
1914 }
1915 
1916 /*
1917  * Allocates both RX and TX resources and configures the SGE. However,
1918  * the hardware is not enabled yet.
1919  */
1920 int t1_sge_configure(struct sge *sge, struct sge_params *p)
1921 {
1922 	if (alloc_rx_resources(sge, p))
1923 		return -ENOMEM;
1924 	if (alloc_tx_resources(sge, p)) {
1925 		free_rx_resources(sge);
1926 		return -ENOMEM;
1927 	}
1928 	configure_sge(sge, p);
1929 
1930 	/*
1931 	 * Now that we have sized the free lists calculate the payload
1932 	 * capacity of the large buffers.  Other parts of the driver use
1933 	 * this to set the max offload coalescing size so that RX packets
1934 	 * do not overflow our large buffers.
1935 	 */
1936 	p->large_buf_capacity = jumbo_payload_capacity(sge);
1937 	return 0;
1938 }
1939 
1940 /*
1941  * Disables the DMA engine.
1942  */
1943 void t1_sge_stop(struct sge *sge)
1944 {
1945 	int i;
1946 	writel(0, sge->adapter->regs + A_SG_CONTROL);
1947 	readl(sge->adapter->regs + A_SG_CONTROL); /* flush */
1948 
1949 	if (is_T2(sge->adapter))
1950 		del_timer_sync(&sge->espibug_timer);
1951 
1952 	del_timer_sync(&sge->tx_reclaim_timer);
1953 	if (sge->tx_sched)
1954 		tx_sched_stop(sge);
1955 
1956 	for (i = 0; i < MAX_NPORTS; i++)
1957 		kfree_skb(sge->espibug_skb[i]);
1958 }
1959 
1960 /*
1961  * Enables the DMA engine.
1962  */
1963 void t1_sge_start(struct sge *sge)
1964 {
1965 	refill_free_list(sge, &sge->freelQ[0]);
1966 	refill_free_list(sge, &sge->freelQ[1]);
1967 
1968 	writel(sge->sge_control, sge->adapter->regs + A_SG_CONTROL);
1969 	doorbell_pio(sge->adapter, F_FL0_ENABLE | F_FL1_ENABLE);
1970 	readl(sge->adapter->regs + A_SG_CONTROL); /* flush */
1971 
1972 	mod_timer(&sge->tx_reclaim_timer, jiffies + TX_RECLAIM_PERIOD);
1973 
1974 	if (is_T2(sge->adapter))
1975 		mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
1976 }
1977 
1978 /*
1979  * Callback for the T2 ESPI 'stuck packet feature' workaorund
1980  */
1981 static void espibug_workaround_t204(unsigned long data)
1982 {
1983 	struct adapter *adapter = (struct adapter *)data;
1984 	struct sge *sge = adapter->sge;
1985 	unsigned int nports = adapter->params.nports;
1986 	u32 seop[MAX_NPORTS];
1987 
1988 	if (adapter->open_device_map & PORT_MASK) {
1989 		int i;
1990 
1991 		if (t1_espi_get_mon_t204(adapter, &(seop[0]), 0) < 0)
1992 			return;
1993 
1994 		for (i = 0; i < nports; i++) {
1995 			struct sk_buff *skb = sge->espibug_skb[i];
1996 
1997 			if (!netif_running(adapter->port[i].dev) ||
1998 			    netif_queue_stopped(adapter->port[i].dev) ||
1999 			    !seop[i] || ((seop[i] & 0xfff) != 0) || !skb)
2000 				continue;
2001 
2002 			if (!skb->cb[0]) {
2003 				skb_copy_to_linear_data_offset(skb,
2004 						    sizeof(struct cpl_tx_pkt),
2005 							       ch_mac_addr,
2006 							       ETH_ALEN);
2007 				skb_copy_to_linear_data_offset(skb,
2008 							       skb->len - 10,
2009 							       ch_mac_addr,
2010 							       ETH_ALEN);
2011 				skb->cb[0] = 0xff;
2012 			}
2013 
2014 			/* bump the reference count to avoid freeing of
2015 			 * the skb once the DMA has completed.
2016 			 */
2017 			skb = skb_get(skb);
2018 			t1_sge_tx(skb, adapter, 0, adapter->port[i].dev);
2019 		}
2020 	}
2021 	mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
2022 }
2023 
2024 static void espibug_workaround(unsigned long data)
2025 {
2026 	struct adapter *adapter = (struct adapter *)data;
2027 	struct sge *sge = adapter->sge;
2028 
2029 	if (netif_running(adapter->port[0].dev)) {
2030 	        struct sk_buff *skb = sge->espibug_skb[0];
2031 	        u32 seop = t1_espi_get_mon(adapter, 0x930, 0);
2032 
2033 	        if ((seop & 0xfff0fff) == 0xfff && skb) {
2034 	                if (!skb->cb[0]) {
2035 	                        skb_copy_to_linear_data_offset(skb,
2036 						     sizeof(struct cpl_tx_pkt),
2037 							       ch_mac_addr,
2038 							       ETH_ALEN);
2039 	                        skb_copy_to_linear_data_offset(skb,
2040 							       skb->len - 10,
2041 							       ch_mac_addr,
2042 							       ETH_ALEN);
2043 	                        skb->cb[0] = 0xff;
2044 	                }
2045 
2046 	                /* bump the reference count to avoid freeing of the
2047 	                 * skb once the DMA has completed.
2048 	                 */
2049 	                skb = skb_get(skb);
2050 	                t1_sge_tx(skb, adapter, 0, adapter->port[0].dev);
2051 	        }
2052 	}
2053 	mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
2054 }
2055 
2056 /*
2057  * Creates a t1_sge structure and returns suggested resource parameters.
2058  */
2059 struct sge *t1_sge_create(struct adapter *adapter, struct sge_params *p)
2060 {
2061 	struct sge *sge = kzalloc(sizeof(*sge), GFP_KERNEL);
2062 	int i;
2063 
2064 	if (!sge)
2065 		return NULL;
2066 
2067 	sge->adapter = adapter;
2068 	sge->netdev = adapter->port[0].dev;
2069 	sge->rx_pkt_pad = t1_is_T1B(adapter) ? 0 : 2;
2070 	sge->jumbo_fl = t1_is_T1B(adapter) ? 1 : 0;
2071 
2072 	for_each_port(adapter, i) {
2073 		sge->port_stats[i] = alloc_percpu(struct sge_port_stats);
2074 		if (!sge->port_stats[i])
2075 			goto nomem_port;
2076 	}
2077 
2078 	init_timer(&sge->tx_reclaim_timer);
2079 	sge->tx_reclaim_timer.data = (unsigned long)sge;
2080 	sge->tx_reclaim_timer.function = sge_tx_reclaim_cb;
2081 
2082 	if (is_T2(sge->adapter)) {
2083 		init_timer(&sge->espibug_timer);
2084 
2085 		if (adapter->params.nports > 1) {
2086 			tx_sched_init(sge);
2087 			sge->espibug_timer.function = espibug_workaround_t204;
2088 		} else
2089 			sge->espibug_timer.function = espibug_workaround;
2090 		sge->espibug_timer.data = (unsigned long)sge->adapter;
2091 
2092 		sge->espibug_timeout = 1;
2093 		/* for T204, every 10ms */
2094 		if (adapter->params.nports > 1)
2095 			sge->espibug_timeout = HZ/100;
2096 	}
2097 
2098 
2099 	p->cmdQ_size[0] = SGE_CMDQ0_E_N;
2100 	p->cmdQ_size[1] = SGE_CMDQ1_E_N;
2101 	p->freelQ_size[!sge->jumbo_fl] = SGE_FREEL_SIZE;
2102 	p->freelQ_size[sge->jumbo_fl] = SGE_JUMBO_FREEL_SIZE;
2103 	if (sge->tx_sched) {
2104 		if (board_info(sge->adapter)->board == CHBT_BOARD_CHT204)
2105 			p->rx_coalesce_usecs = 15;
2106 		else
2107 			p->rx_coalesce_usecs = 50;
2108 	} else
2109 		p->rx_coalesce_usecs = 50;
2110 
2111 	p->coalesce_enable = 0;
2112 	p->sample_interval_usecs = 0;
2113 
2114 	return sge;
2115 nomem_port:
2116 	while (i >= 0) {
2117 		free_percpu(sge->port_stats[i]);
2118 		--i;
2119 	}
2120 	kfree(sge);
2121 	return NULL;
2122 
2123 }
2124