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 	struct sge *sge;
243 };
244 
245 static void restart_sched(struct tasklet_struct *t);
246 
247 
248 /*
249  * Main SGE data structure
250  *
251  * Interrupts are handled by a single CPU and it is likely that on a MP system
252  * the application is migrated to another CPU. In that scenario, we try to
253  * separate the RX(in irq context) and TX state in order to decrease memory
254  * contention.
255  */
256 struct sge {
257 	struct adapter *adapter;	/* adapter backpointer */
258 	struct net_device *netdev;      /* netdevice backpointer */
259 	struct freelQ	freelQ[SGE_FREELQ_N]; /* buffer free lists */
260 	struct respQ	respQ;		/* response Q */
261 	unsigned long   stopped_tx_queues; /* bitmap of suspended Tx queues */
262 	unsigned int	rx_pkt_pad;     /* RX padding for L2 packets */
263 	unsigned int	jumbo_fl;       /* jumbo freelist Q index */
264 	unsigned int	intrtimer_nres;	/* no-resource interrupt timer */
265 	unsigned int    fixed_intrtimer;/* non-adaptive interrupt timer */
266 	struct timer_list tx_reclaim_timer; /* reclaims TX buffers */
267 	struct timer_list espibug_timer;
268 	unsigned long	espibug_timeout;
269 	struct sk_buff	*espibug_skb[MAX_NPORTS];
270 	u32		sge_control;	/* shadow value of sge control reg */
271 	struct sge_intr_counts stats;
272 	struct sge_port_stats __percpu *port_stats[MAX_NPORTS];
273 	struct sched	*tx_sched;
274 	struct cmdQ cmdQ[SGE_CMDQ_N] ____cacheline_aligned_in_smp;
275 };
276 
277 static const u8 ch_mac_addr[ETH_ALEN] = {
278 	0x0, 0x7, 0x43, 0x0, 0x0, 0x0
279 };
280 
281 /*
282  * stop tasklet and free all pending skb's
283  */
284 static void tx_sched_stop(struct sge *sge)
285 {
286 	struct sched *s = sge->tx_sched;
287 	int i;
288 
289 	tasklet_kill(&s->sched_tsk);
290 
291 	for (i = 0; i < MAX_NPORTS; i++)
292 		__skb_queue_purge(&s->p[s->port].skbq);
293 }
294 
295 /*
296  * t1_sched_update_parms() is called when the MTU or link speed changes. It
297  * re-computes scheduler parameters to scope with the change.
298  */
299 unsigned int t1_sched_update_parms(struct sge *sge, unsigned int port,
300 				   unsigned int mtu, unsigned int speed)
301 {
302 	struct sched *s = sge->tx_sched;
303 	struct sched_port *p = &s->p[port];
304 	unsigned int max_avail_segs;
305 
306 	pr_debug("%s mtu=%d speed=%d\n", __func__, mtu, speed);
307 	if (speed)
308 		p->speed = speed;
309 	if (mtu)
310 		p->mtu = mtu;
311 
312 	if (speed || mtu) {
313 		unsigned long long drain = 1024ULL * p->speed * (p->mtu - 40);
314 		do_div(drain, (p->mtu + 50) * 1000);
315 		p->drain_bits_per_1024ns = (unsigned int) drain;
316 
317 		if (p->speed < 1000)
318 			p->drain_bits_per_1024ns =
319 				90 * p->drain_bits_per_1024ns / 100;
320 	}
321 
322 	if (board_info(sge->adapter)->board == CHBT_BOARD_CHT204) {
323 		p->drain_bits_per_1024ns -= 16;
324 		s->max_avail = max(4096U, p->mtu + 16 + 14 + 4);
325 		max_avail_segs = max(1U, 4096 / (p->mtu - 40));
326 	} else {
327 		s->max_avail = 16384;
328 		max_avail_segs = max(1U, 9000 / (p->mtu - 40));
329 	}
330 
331 	pr_debug("t1_sched_update_parms: mtu %u speed %u max_avail %u "
332 		 "max_avail_segs %u drain_bits_per_1024ns %u\n", p->mtu,
333 		 p->speed, s->max_avail, max_avail_segs,
334 		 p->drain_bits_per_1024ns);
335 
336 	return max_avail_segs * (p->mtu - 40);
337 }
338 
339 #if 0
340 
341 /*
342  * t1_sched_max_avail_bytes() tells the scheduler the maximum amount of
343  * data that can be pushed per port.
344  */
345 void t1_sched_set_max_avail_bytes(struct sge *sge, unsigned int val)
346 {
347 	struct sched *s = sge->tx_sched;
348 	unsigned int i;
349 
350 	s->max_avail = val;
351 	for (i = 0; i < MAX_NPORTS; i++)
352 		t1_sched_update_parms(sge, i, 0, 0);
353 }
354 
355 /*
356  * t1_sched_set_drain_bits_per_us() tells the scheduler at which rate a port
357  * is draining.
358  */
359 void t1_sched_set_drain_bits_per_us(struct sge *sge, unsigned int port,
360 					 unsigned int val)
361 {
362 	struct sched *s = sge->tx_sched;
363 	struct sched_port *p = &s->p[port];
364 	p->drain_bits_per_1024ns = val * 1024 / 1000;
365 	t1_sched_update_parms(sge, port, 0, 0);
366 }
367 
368 #endif  /*  0  */
369 
370 /*
371  * tx_sched_init() allocates resources and does basic initialization.
372  */
373 static int tx_sched_init(struct sge *sge)
374 {
375 	struct sched *s;
376 	int i;
377 
378 	s = kzalloc(sizeof (struct sched), GFP_KERNEL);
379 	if (!s)
380 		return -ENOMEM;
381 
382 	pr_debug("tx_sched_init\n");
383 	tasklet_setup(&s->sched_tsk, restart_sched);
384 	s->sge = sge;
385 	sge->tx_sched = s;
386 
387 	for (i = 0; i < MAX_NPORTS; i++) {
388 		skb_queue_head_init(&s->p[i].skbq);
389 		t1_sched_update_parms(sge, i, 1500, 1000);
390 	}
391 
392 	return 0;
393 }
394 
395 /*
396  * sched_update_avail() computes the delta since the last time it was called
397  * and updates the per port quota (number of bits that can be sent to the any
398  * port).
399  */
400 static inline int sched_update_avail(struct sge *sge)
401 {
402 	struct sched *s = sge->tx_sched;
403 	ktime_t now = ktime_get();
404 	unsigned int i;
405 	long long delta_time_ns;
406 
407 	delta_time_ns = ktime_to_ns(ktime_sub(now, s->last_updated));
408 
409 	pr_debug("sched_update_avail delta=%lld\n", delta_time_ns);
410 	if (delta_time_ns < 15000)
411 		return 0;
412 
413 	for (i = 0; i < MAX_NPORTS; i++) {
414 		struct sched_port *p = &s->p[i];
415 		unsigned int delta_avail;
416 
417 		delta_avail = (p->drain_bits_per_1024ns * delta_time_ns) >> 13;
418 		p->avail = min(p->avail + delta_avail, s->max_avail);
419 	}
420 
421 	s->last_updated = now;
422 
423 	return 1;
424 }
425 
426 /*
427  * sched_skb() is called from two different places. In the tx path, any
428  * packet generating load on an output port will call sched_skb()
429  * (skb != NULL). In addition, sched_skb() is called from the irq/soft irq
430  * context (skb == NULL).
431  * The scheduler only returns a skb (which will then be sent) if the
432  * length of the skb is <= the current quota of the output port.
433  */
434 static struct sk_buff *sched_skb(struct sge *sge, struct sk_buff *skb,
435 				unsigned int credits)
436 {
437 	struct sched *s = sge->tx_sched;
438 	struct sk_buff_head *skbq;
439 	unsigned int i, len, update = 1;
440 
441 	pr_debug("sched_skb %p\n", skb);
442 	if (!skb) {
443 		if (!s->num)
444 			return NULL;
445 	} else {
446 		skbq = &s->p[skb->dev->if_port].skbq;
447 		__skb_queue_tail(skbq, skb);
448 		s->num++;
449 		skb = NULL;
450 	}
451 
452 	if (credits < MAX_SKB_FRAGS + 1)
453 		goto out;
454 
455 again:
456 	for (i = 0; i < MAX_NPORTS; i++) {
457 		s->port = (s->port + 1) & (MAX_NPORTS - 1);
458 		skbq = &s->p[s->port].skbq;
459 
460 		skb = skb_peek(skbq);
461 
462 		if (!skb)
463 			continue;
464 
465 		len = skb->len;
466 		if (len <= s->p[s->port].avail) {
467 			s->p[s->port].avail -= len;
468 			s->num--;
469 			__skb_unlink(skb, skbq);
470 			goto out;
471 		}
472 		skb = NULL;
473 	}
474 
475 	if (update-- && sched_update_avail(sge))
476 		goto again;
477 
478 out:
479 	/* If there are more pending skbs, we use the hardware to schedule us
480 	 * again.
481 	 */
482 	if (s->num && !skb) {
483 		struct cmdQ *q = &sge->cmdQ[0];
484 		clear_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
485 		if (test_and_set_bit(CMDQ_STAT_RUNNING, &q->status) == 0) {
486 			set_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
487 			writel(F_CMDQ0_ENABLE, sge->adapter->regs + A_SG_DOORBELL);
488 		}
489 	}
490 	pr_debug("sched_skb ret %p\n", skb);
491 
492 	return skb;
493 }
494 
495 /*
496  * PIO to indicate that memory mapped Q contains valid descriptor(s).
497  */
498 static inline void doorbell_pio(struct adapter *adapter, u32 val)
499 {
500 	wmb();
501 	writel(val, adapter->regs + A_SG_DOORBELL);
502 }
503 
504 /*
505  * Frees all RX buffers on the freelist Q. The caller must make sure that
506  * the SGE is turned off before calling this function.
507  */
508 static void free_freelQ_buffers(struct pci_dev *pdev, struct freelQ *q)
509 {
510 	unsigned int cidx = q->cidx;
511 
512 	while (q->credits--) {
513 		struct freelQ_ce *ce = &q->centries[cidx];
514 
515 		dma_unmap_single(&pdev->dev, dma_unmap_addr(ce, dma_addr),
516 				 dma_unmap_len(ce, dma_len), DMA_FROM_DEVICE);
517 		dev_kfree_skb(ce->skb);
518 		ce->skb = NULL;
519 		if (++cidx == q->size)
520 			cidx = 0;
521 	}
522 }
523 
524 /*
525  * Free RX free list and response queue resources.
526  */
527 static void free_rx_resources(struct sge *sge)
528 {
529 	struct pci_dev *pdev = sge->adapter->pdev;
530 	unsigned int size, i;
531 
532 	if (sge->respQ.entries) {
533 		size = sizeof(struct respQ_e) * sge->respQ.size;
534 		dma_free_coherent(&pdev->dev, size, sge->respQ.entries,
535 				  sge->respQ.dma_addr);
536 	}
537 
538 	for (i = 0; i < SGE_FREELQ_N; i++) {
539 		struct freelQ *q = &sge->freelQ[i];
540 
541 		if (q->centries) {
542 			free_freelQ_buffers(pdev, q);
543 			kfree(q->centries);
544 		}
545 		if (q->entries) {
546 			size = sizeof(struct freelQ_e) * q->size;
547 			dma_free_coherent(&pdev->dev, size, q->entries,
548 					  q->dma_addr);
549 		}
550 	}
551 }
552 
553 /*
554  * Allocates basic RX resources, consisting of memory mapped freelist Qs and a
555  * response queue.
556  */
557 static int alloc_rx_resources(struct sge *sge, struct sge_params *p)
558 {
559 	struct pci_dev *pdev = sge->adapter->pdev;
560 	unsigned int size, i;
561 
562 	for (i = 0; i < SGE_FREELQ_N; i++) {
563 		struct freelQ *q = &sge->freelQ[i];
564 
565 		q->genbit = 1;
566 		q->size = p->freelQ_size[i];
567 		q->dma_offset = sge->rx_pkt_pad ? 0 : NET_IP_ALIGN;
568 		size = sizeof(struct freelQ_e) * q->size;
569 		q->entries = dma_alloc_coherent(&pdev->dev, size,
570 						&q->dma_addr, GFP_KERNEL);
571 		if (!q->entries)
572 			goto err_no_mem;
573 
574 		size = sizeof(struct freelQ_ce) * q->size;
575 		q->centries = kzalloc(size, GFP_KERNEL);
576 		if (!q->centries)
577 			goto err_no_mem;
578 	}
579 
580 	/*
581 	 * Calculate the buffer sizes for the two free lists.  FL0 accommodates
582 	 * regular sized Ethernet frames, FL1 is sized not to exceed 16K,
583 	 * including all the sk_buff overhead.
584 	 *
585 	 * Note: For T2 FL0 and FL1 are reversed.
586 	 */
587 	sge->freelQ[!sge->jumbo_fl].rx_buffer_size = SGE_RX_SM_BUF_SIZE +
588 		sizeof(struct cpl_rx_data) +
589 		sge->freelQ[!sge->jumbo_fl].dma_offset;
590 
591 	size = (16 * 1024) - SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
592 
593 	sge->freelQ[sge->jumbo_fl].rx_buffer_size = size;
594 
595 	/*
596 	 * Setup which skb recycle Q should be used when recycling buffers from
597 	 * each free list.
598 	 */
599 	sge->freelQ[!sge->jumbo_fl].recycleq_idx = 0;
600 	sge->freelQ[sge->jumbo_fl].recycleq_idx = 1;
601 
602 	sge->respQ.genbit = 1;
603 	sge->respQ.size = SGE_RESPQ_E_N;
604 	sge->respQ.credits = 0;
605 	size = sizeof(struct respQ_e) * sge->respQ.size;
606 	sge->respQ.entries =
607 		dma_alloc_coherent(&pdev->dev, size, &sge->respQ.dma_addr,
608 				   GFP_KERNEL);
609 	if (!sge->respQ.entries)
610 		goto err_no_mem;
611 	return 0;
612 
613 err_no_mem:
614 	free_rx_resources(sge);
615 	return -ENOMEM;
616 }
617 
618 /*
619  * Reclaims n TX descriptors and frees the buffers associated with them.
620  */
621 static void free_cmdQ_buffers(struct sge *sge, struct cmdQ *q, unsigned int n)
622 {
623 	struct cmdQ_ce *ce;
624 	struct pci_dev *pdev = sge->adapter->pdev;
625 	unsigned int cidx = q->cidx;
626 
627 	q->in_use -= n;
628 	ce = &q->centries[cidx];
629 	while (n--) {
630 		if (likely(dma_unmap_len(ce, dma_len))) {
631 			dma_unmap_single(&pdev->dev,
632 					 dma_unmap_addr(ce, dma_addr),
633 					 dma_unmap_len(ce, dma_len),
634 					 DMA_TO_DEVICE);
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 			dma_free_coherent(&pdev->dev, 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 = dma_alloc_coherent(&pdev->dev, size,
698 						&q->dma_addr, GFP_KERNEL);
699 		if (!q->entries)
700 			goto err_no_mem;
701 
702 		size = sizeof(struct cmdQ_ce) * q->size;
703 		q->centries = kzalloc(size, GFP_KERNEL);
704 		if (!q->centries)
705 			goto err_no_mem;
706 	}
707 
708 	/*
709 	 * CommandQ 0 handles Ethernet and TOE packets, while queue 1 is TOE
710 	 * only.  For queue 0 set the stop threshold so we can handle one more
711 	 * packet from each port, plus reserve an additional 24 entries for
712 	 * Ethernet packets only.  Queue 1 never suspends nor do we reserve
713 	 * space for Ethernet packets.
714 	 */
715 	sge->cmdQ[0].stop_thres = sge->adapter->params.nports *
716 		(MAX_SKB_FRAGS + 1);
717 	return 0;
718 
719 err_no_mem:
720 	free_tx_resources(sge);
721 	return -ENOMEM;
722 }
723 
724 static inline void setup_ring_params(struct adapter *adapter, u64 addr,
725 				     u32 size, int base_reg_lo,
726 				     int base_reg_hi, int size_reg)
727 {
728 	writel((u32)addr, adapter->regs + base_reg_lo);
729 	writel(addr >> 32, adapter->regs + base_reg_hi);
730 	writel(size, adapter->regs + size_reg);
731 }
732 
733 /*
734  * Enable/disable VLAN acceleration.
735  */
736 void t1_vlan_mode(struct adapter *adapter, netdev_features_t features)
737 {
738 	struct sge *sge = adapter->sge;
739 
740 	if (features & NETIF_F_HW_VLAN_CTAG_RX)
741 		sge->sge_control |= F_VLAN_XTRACT;
742 	else
743 		sge->sge_control &= ~F_VLAN_XTRACT;
744 	if (adapter->open_device_map) {
745 		writel(sge->sge_control, adapter->regs + A_SG_CONTROL);
746 		readl(adapter->regs + A_SG_CONTROL);   /* flush */
747 	}
748 }
749 
750 /*
751  * Programs the various SGE registers. However, the engine is not yet enabled,
752  * but sge->sge_control is setup and ready to go.
753  */
754 static void configure_sge(struct sge *sge, struct sge_params *p)
755 {
756 	struct adapter *ap = sge->adapter;
757 
758 	writel(0, ap->regs + A_SG_CONTROL);
759 	setup_ring_params(ap, sge->cmdQ[0].dma_addr, sge->cmdQ[0].size,
760 			  A_SG_CMD0BASELWR, A_SG_CMD0BASEUPR, A_SG_CMD0SIZE);
761 	setup_ring_params(ap, sge->cmdQ[1].dma_addr, sge->cmdQ[1].size,
762 			  A_SG_CMD1BASELWR, A_SG_CMD1BASEUPR, A_SG_CMD1SIZE);
763 	setup_ring_params(ap, sge->freelQ[0].dma_addr,
764 			  sge->freelQ[0].size, A_SG_FL0BASELWR,
765 			  A_SG_FL0BASEUPR, A_SG_FL0SIZE);
766 	setup_ring_params(ap, sge->freelQ[1].dma_addr,
767 			  sge->freelQ[1].size, A_SG_FL1BASELWR,
768 			  A_SG_FL1BASEUPR, A_SG_FL1SIZE);
769 
770 	/* The threshold comparison uses <. */
771 	writel(SGE_RX_SM_BUF_SIZE + 1, ap->regs + A_SG_FLTHRESHOLD);
772 
773 	setup_ring_params(ap, sge->respQ.dma_addr, sge->respQ.size,
774 			  A_SG_RSPBASELWR, A_SG_RSPBASEUPR, A_SG_RSPSIZE);
775 	writel((u32)sge->respQ.size - 1, ap->regs + A_SG_RSPQUEUECREDIT);
776 
777 	sge->sge_control = F_CMDQ0_ENABLE | F_CMDQ1_ENABLE | F_FL0_ENABLE |
778 		F_FL1_ENABLE | F_CPL_ENABLE | F_RESPONSE_QUEUE_ENABLE |
779 		V_CMDQ_PRIORITY(2) | F_DISABLE_CMDQ1_GTS | F_ISCSI_COALESCE |
780 		V_RX_PKT_OFFSET(sge->rx_pkt_pad);
781 
782 #if defined(__BIG_ENDIAN_BITFIELD)
783 	sge->sge_control |= F_ENABLE_BIG_ENDIAN;
784 #endif
785 
786 	/* Initialize no-resource timer */
787 	sge->intrtimer_nres = SGE_INTRTIMER_NRES * core_ticks_per_usec(ap);
788 
789 	t1_sge_set_coalesce_params(sge, p);
790 }
791 
792 /*
793  * Return the payload capacity of the jumbo free-list buffers.
794  */
795 static inline unsigned int jumbo_payload_capacity(const struct sge *sge)
796 {
797 	return sge->freelQ[sge->jumbo_fl].rx_buffer_size -
798 		sge->freelQ[sge->jumbo_fl].dma_offset -
799 		sizeof(struct cpl_rx_data);
800 }
801 
802 /*
803  * Frees all SGE related resources and the sge structure itself
804  */
805 void t1_sge_destroy(struct sge *sge)
806 {
807 	int i;
808 
809 	for_each_port(sge->adapter, i)
810 		free_percpu(sge->port_stats[i]);
811 
812 	kfree(sge->tx_sched);
813 	free_tx_resources(sge);
814 	free_rx_resources(sge);
815 	kfree(sge);
816 }
817 
818 /*
819  * Allocates new RX buffers on the freelist Q (and tracks them on the freelist
820  * context Q) until the Q is full or alloc_skb fails.
821  *
822  * It is possible that the generation bits already match, indicating that the
823  * buffer is already valid and nothing needs to be done. This happens when we
824  * copied a received buffer into a new sk_buff during the interrupt processing.
825  *
826  * If the SGE doesn't automatically align packets properly (!sge->rx_pkt_pad),
827  * we specify a RX_OFFSET in order to make sure that the IP header is 4B
828  * aligned.
829  */
830 static void refill_free_list(struct sge *sge, struct freelQ *q)
831 {
832 	struct pci_dev *pdev = sge->adapter->pdev;
833 	struct freelQ_ce *ce = &q->centries[q->pidx];
834 	struct freelQ_e *e = &q->entries[q->pidx];
835 	unsigned int dma_len = q->rx_buffer_size - q->dma_offset;
836 
837 	while (q->credits < q->size) {
838 		struct sk_buff *skb;
839 		dma_addr_t mapping;
840 
841 		skb = dev_alloc_skb(q->rx_buffer_size);
842 		if (!skb)
843 			break;
844 
845 		skb_reserve(skb, q->dma_offset);
846 		mapping = dma_map_single(&pdev->dev, skb->data, dma_len,
847 					 DMA_FROM_DEVICE);
848 		skb_reserve(skb, sge->rx_pkt_pad);
849 
850 		ce->skb = skb;
851 		dma_unmap_addr_set(ce, dma_addr, mapping);
852 		dma_unmap_len_set(ce, dma_len, dma_len);
853 		e->addr_lo = (u32)mapping;
854 		e->addr_hi = (u64)mapping >> 32;
855 		e->len_gen = V_CMD_LEN(dma_len) | V_CMD_GEN1(q->genbit);
856 		wmb();
857 		e->gen2 = V_CMD_GEN2(q->genbit);
858 
859 		e++;
860 		ce++;
861 		if (++q->pidx == q->size) {
862 			q->pidx = 0;
863 			q->genbit ^= 1;
864 			ce = q->centries;
865 			e = q->entries;
866 		}
867 		q->credits++;
868 	}
869 }
870 
871 /*
872  * Calls refill_free_list for both free lists. If we cannot fill at least 1/4
873  * of both rings, we go into 'few interrupt mode' in order to give the system
874  * time to free up resources.
875  */
876 static void freelQs_empty(struct sge *sge)
877 {
878 	struct adapter *adapter = sge->adapter;
879 	u32 irq_reg = readl(adapter->regs + A_SG_INT_ENABLE);
880 	u32 irqholdoff_reg;
881 
882 	refill_free_list(sge, &sge->freelQ[0]);
883 	refill_free_list(sge, &sge->freelQ[1]);
884 
885 	if (sge->freelQ[0].credits > (sge->freelQ[0].size >> 2) &&
886 	    sge->freelQ[1].credits > (sge->freelQ[1].size >> 2)) {
887 		irq_reg |= F_FL_EXHAUSTED;
888 		irqholdoff_reg = sge->fixed_intrtimer;
889 	} else {
890 		/* Clear the F_FL_EXHAUSTED interrupts for now */
891 		irq_reg &= ~F_FL_EXHAUSTED;
892 		irqholdoff_reg = sge->intrtimer_nres;
893 	}
894 	writel(irqholdoff_reg, adapter->regs + A_SG_INTRTIMER);
895 	writel(irq_reg, adapter->regs + A_SG_INT_ENABLE);
896 
897 	/* We reenable the Qs to force a freelist GTS interrupt later */
898 	doorbell_pio(adapter, F_FL0_ENABLE | F_FL1_ENABLE);
899 }
900 
901 #define SGE_PL_INTR_MASK (F_PL_INTR_SGE_ERR | F_PL_INTR_SGE_DATA)
902 #define SGE_INT_FATAL (F_RESPQ_OVERFLOW | F_PACKET_TOO_BIG | F_PACKET_MISMATCH)
903 #define SGE_INT_ENABLE (F_RESPQ_EXHAUSTED | F_RESPQ_OVERFLOW | \
904 			F_FL_EXHAUSTED | F_PACKET_TOO_BIG | F_PACKET_MISMATCH)
905 
906 /*
907  * Disable SGE Interrupts
908  */
909 void t1_sge_intr_disable(struct sge *sge)
910 {
911 	u32 val = readl(sge->adapter->regs + A_PL_ENABLE);
912 
913 	writel(val & ~SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_ENABLE);
914 	writel(0, sge->adapter->regs + A_SG_INT_ENABLE);
915 }
916 
917 /*
918  * Enable SGE interrupts.
919  */
920 void t1_sge_intr_enable(struct sge *sge)
921 {
922 	u32 en = SGE_INT_ENABLE;
923 	u32 val = readl(sge->adapter->regs + A_PL_ENABLE);
924 
925 	if (sge->adapter->port[0].dev->hw_features & NETIF_F_TSO)
926 		en &= ~F_PACKET_TOO_BIG;
927 	writel(en, sge->adapter->regs + A_SG_INT_ENABLE);
928 	writel(val | SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_ENABLE);
929 }
930 
931 /*
932  * Clear SGE interrupts.
933  */
934 void t1_sge_intr_clear(struct sge *sge)
935 {
936 	writel(SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_CAUSE);
937 	writel(0xffffffff, sge->adapter->regs + A_SG_INT_CAUSE);
938 }
939 
940 /*
941  * SGE 'Error' interrupt handler
942  */
943 bool t1_sge_intr_error_handler(struct sge *sge)
944 {
945 	struct adapter *adapter = sge->adapter;
946 	u32 cause = readl(adapter->regs + A_SG_INT_CAUSE);
947 	bool wake = false;
948 
949 	if (adapter->port[0].dev->hw_features & NETIF_F_TSO)
950 		cause &= ~F_PACKET_TOO_BIG;
951 	if (cause & F_RESPQ_EXHAUSTED)
952 		sge->stats.respQ_empty++;
953 	if (cause & F_RESPQ_OVERFLOW) {
954 		sge->stats.respQ_overflow++;
955 		pr_alert("%s: SGE response queue overflow\n",
956 			 adapter->name);
957 	}
958 	if (cause & F_FL_EXHAUSTED) {
959 		sge->stats.freelistQ_empty++;
960 		freelQs_empty(sge);
961 	}
962 	if (cause & F_PACKET_TOO_BIG) {
963 		sge->stats.pkt_too_big++;
964 		pr_alert("%s: SGE max packet size exceeded\n",
965 			 adapter->name);
966 	}
967 	if (cause & F_PACKET_MISMATCH) {
968 		sge->stats.pkt_mismatch++;
969 		pr_alert("%s: SGE packet mismatch\n", adapter->name);
970 	}
971 	if (cause & SGE_INT_FATAL) {
972 		t1_interrupts_disable(adapter);
973 		adapter->pending_thread_intr |= F_PL_INTR_SGE_ERR;
974 		wake = true;
975 	}
976 
977 	writel(cause, adapter->regs + A_SG_INT_CAUSE);
978 	return wake;
979 }
980 
981 const struct sge_intr_counts *t1_sge_get_intr_counts(const struct sge *sge)
982 {
983 	return &sge->stats;
984 }
985 
986 void t1_sge_get_port_stats(const struct sge *sge, int port,
987 			   struct sge_port_stats *ss)
988 {
989 	int cpu;
990 
991 	memset(ss, 0, sizeof(*ss));
992 	for_each_possible_cpu(cpu) {
993 		struct sge_port_stats *st = per_cpu_ptr(sge->port_stats[port], cpu);
994 
995 		ss->rx_cso_good += st->rx_cso_good;
996 		ss->tx_cso += st->tx_cso;
997 		ss->tx_tso += st->tx_tso;
998 		ss->tx_need_hdrroom += st->tx_need_hdrroom;
999 		ss->vlan_xtract += st->vlan_xtract;
1000 		ss->vlan_insert += st->vlan_insert;
1001 	}
1002 }
1003 
1004 /**
1005  *	recycle_fl_buf - recycle a free list buffer
1006  *	@fl: the free list
1007  *	@idx: index of buffer to recycle
1008  *
1009  *	Recycles the specified buffer on the given free list by adding it at
1010  *	the next available slot on the list.
1011  */
1012 static void recycle_fl_buf(struct freelQ *fl, int idx)
1013 {
1014 	struct freelQ_e *from = &fl->entries[idx];
1015 	struct freelQ_e *to = &fl->entries[fl->pidx];
1016 
1017 	fl->centries[fl->pidx] = fl->centries[idx];
1018 	to->addr_lo = from->addr_lo;
1019 	to->addr_hi = from->addr_hi;
1020 	to->len_gen = G_CMD_LEN(from->len_gen) | V_CMD_GEN1(fl->genbit);
1021 	wmb();
1022 	to->gen2 = V_CMD_GEN2(fl->genbit);
1023 	fl->credits++;
1024 
1025 	if (++fl->pidx == fl->size) {
1026 		fl->pidx = 0;
1027 		fl->genbit ^= 1;
1028 	}
1029 }
1030 
1031 static int copybreak __read_mostly = 256;
1032 module_param(copybreak, int, 0);
1033 MODULE_PARM_DESC(copybreak, "Receive copy threshold");
1034 
1035 /**
1036  *	get_packet - return the next ingress packet buffer
1037  *	@adapter: the adapter that received the packet
1038  *	@fl: the SGE free list holding the packet
1039  *	@len: the actual packet length, excluding any SGE padding
1040  *
1041  *	Get the next packet from a free list and complete setup of the
1042  *	sk_buff.  If the packet is small we make a copy and recycle the
1043  *	original buffer, otherwise we use the original buffer itself.  If a
1044  *	positive drop threshold is supplied packets are dropped and their
1045  *	buffers recycled if (a) the number of remaining buffers is under the
1046  *	threshold and the packet is too big to copy, or (b) the packet should
1047  *	be copied but there is no memory for the copy.
1048  */
1049 static inline struct sk_buff *get_packet(struct adapter *adapter,
1050 					 struct freelQ *fl, unsigned int len)
1051 {
1052 	const struct freelQ_ce *ce = &fl->centries[fl->cidx];
1053 	struct pci_dev *pdev = adapter->pdev;
1054 	struct sk_buff *skb;
1055 
1056 	if (len < copybreak) {
1057 		skb = napi_alloc_skb(&adapter->napi, len);
1058 		if (!skb)
1059 			goto use_orig_buf;
1060 
1061 		skb_put(skb, len);
1062 		dma_sync_single_for_cpu(&pdev->dev,
1063 					dma_unmap_addr(ce, dma_addr),
1064 					dma_unmap_len(ce, dma_len),
1065 					DMA_FROM_DEVICE);
1066 		skb_copy_from_linear_data(ce->skb, skb->data, len);
1067 		dma_sync_single_for_device(&pdev->dev,
1068 					   dma_unmap_addr(ce, dma_addr),
1069 					   dma_unmap_len(ce, dma_len),
1070 					   DMA_FROM_DEVICE);
1071 		recycle_fl_buf(fl, fl->cidx);
1072 		return skb;
1073 	}
1074 
1075 use_orig_buf:
1076 	if (fl->credits < 2) {
1077 		recycle_fl_buf(fl, fl->cidx);
1078 		return NULL;
1079 	}
1080 
1081 	dma_unmap_single(&pdev->dev, dma_unmap_addr(ce, dma_addr),
1082 			 dma_unmap_len(ce, dma_len), DMA_FROM_DEVICE);
1083 	skb = ce->skb;
1084 	prefetch(skb->data);
1085 
1086 	skb_put(skb, len);
1087 	return skb;
1088 }
1089 
1090 /**
1091  *	unexpected_offload - handle an unexpected offload packet
1092  *	@adapter: the adapter
1093  *	@fl: the free list that received the packet
1094  *
1095  *	Called when we receive an unexpected offload packet (e.g., the TOE
1096  *	function is disabled or the card is a NIC).  Prints a message and
1097  *	recycles the buffer.
1098  */
1099 static void unexpected_offload(struct adapter *adapter, struct freelQ *fl)
1100 {
1101 	struct freelQ_ce *ce = &fl->centries[fl->cidx];
1102 	struct sk_buff *skb = ce->skb;
1103 
1104 	dma_sync_single_for_cpu(&adapter->pdev->dev,
1105 				dma_unmap_addr(ce, dma_addr),
1106 				dma_unmap_len(ce, dma_len), DMA_FROM_DEVICE);
1107 	pr_err("%s: unexpected offload packet, cmd %u\n",
1108 	       adapter->name, *skb->data);
1109 	recycle_fl_buf(fl, fl->cidx);
1110 }
1111 
1112 /*
1113  * T1/T2 SGE limits the maximum DMA size per TX descriptor to
1114  * SGE_TX_DESC_MAX_PLEN (16KB). If the PAGE_SIZE is larger than 16KB, the
1115  * stack might send more than SGE_TX_DESC_MAX_PLEN in a contiguous manner.
1116  * Note that the *_large_page_tx_descs stuff will be optimized out when
1117  * PAGE_SIZE <= SGE_TX_DESC_MAX_PLEN.
1118  *
1119  * compute_large_page_descs() computes how many additional descriptors are
1120  * required to break down the stack's request.
1121  */
1122 static inline unsigned int compute_large_page_tx_descs(struct sk_buff *skb)
1123 {
1124 	unsigned int count = 0;
1125 
1126 	if (PAGE_SIZE > SGE_TX_DESC_MAX_PLEN) {
1127 		unsigned int nfrags = skb_shinfo(skb)->nr_frags;
1128 		unsigned int i, len = skb_headlen(skb);
1129 		while (len > SGE_TX_DESC_MAX_PLEN) {
1130 			count++;
1131 			len -= SGE_TX_DESC_MAX_PLEN;
1132 		}
1133 		for (i = 0; nfrags--; i++) {
1134 			const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
1135 			len = skb_frag_size(frag);
1136 			while (len > SGE_TX_DESC_MAX_PLEN) {
1137 				count++;
1138 				len -= SGE_TX_DESC_MAX_PLEN;
1139 			}
1140 		}
1141 	}
1142 	return count;
1143 }
1144 
1145 /*
1146  * Write a cmdQ entry.
1147  *
1148  * Since this function writes the 'flags' field, it must not be used to
1149  * write the first cmdQ entry.
1150  */
1151 static inline void write_tx_desc(struct cmdQ_e *e, dma_addr_t mapping,
1152 				 unsigned int len, unsigned int gen,
1153 				 unsigned int eop)
1154 {
1155 	BUG_ON(len > SGE_TX_DESC_MAX_PLEN);
1156 
1157 	e->addr_lo = (u32)mapping;
1158 	e->addr_hi = (u64)mapping >> 32;
1159 	e->len_gen = V_CMD_LEN(len) | V_CMD_GEN1(gen);
1160 	e->flags = F_CMD_DATAVALID | V_CMD_EOP(eop) | V_CMD_GEN2(gen);
1161 }
1162 
1163 /*
1164  * See comment for previous function.
1165  *
1166  * write_tx_descs_large_page() writes additional SGE tx descriptors if
1167  * *desc_len exceeds HW's capability.
1168  */
1169 static inline unsigned int write_large_page_tx_descs(unsigned int pidx,
1170 						     struct cmdQ_e **e,
1171 						     struct cmdQ_ce **ce,
1172 						     unsigned int *gen,
1173 						     dma_addr_t *desc_mapping,
1174 						     unsigned int *desc_len,
1175 						     unsigned int nfrags,
1176 						     struct cmdQ *q)
1177 {
1178 	if (PAGE_SIZE > SGE_TX_DESC_MAX_PLEN) {
1179 		struct cmdQ_e *e1 = *e;
1180 		struct cmdQ_ce *ce1 = *ce;
1181 
1182 		while (*desc_len > SGE_TX_DESC_MAX_PLEN) {
1183 			*desc_len -= SGE_TX_DESC_MAX_PLEN;
1184 			write_tx_desc(e1, *desc_mapping, SGE_TX_DESC_MAX_PLEN,
1185 				      *gen, nfrags == 0 && *desc_len == 0);
1186 			ce1->skb = NULL;
1187 			dma_unmap_len_set(ce1, dma_len, 0);
1188 			*desc_mapping += SGE_TX_DESC_MAX_PLEN;
1189 			if (*desc_len) {
1190 				ce1++;
1191 				e1++;
1192 				if (++pidx == q->size) {
1193 					pidx = 0;
1194 					*gen ^= 1;
1195 					ce1 = q->centries;
1196 					e1 = q->entries;
1197 				}
1198 			}
1199 		}
1200 		*e = e1;
1201 		*ce = ce1;
1202 	}
1203 	return pidx;
1204 }
1205 
1206 /*
1207  * Write the command descriptors to transmit the given skb starting at
1208  * descriptor pidx with the given generation.
1209  */
1210 static inline void write_tx_descs(struct adapter *adapter, struct sk_buff *skb,
1211 				  unsigned int pidx, unsigned int gen,
1212 				  struct cmdQ *q)
1213 {
1214 	dma_addr_t mapping, desc_mapping;
1215 	struct cmdQ_e *e, *e1;
1216 	struct cmdQ_ce *ce;
1217 	unsigned int i, flags, first_desc_len, desc_len,
1218 	    nfrags = skb_shinfo(skb)->nr_frags;
1219 
1220 	e = e1 = &q->entries[pidx];
1221 	ce = &q->centries[pidx];
1222 
1223 	mapping = dma_map_single(&adapter->pdev->dev, skb->data,
1224 				 skb_headlen(skb), DMA_TO_DEVICE);
1225 
1226 	desc_mapping = mapping;
1227 	desc_len = skb_headlen(skb);
1228 
1229 	flags = F_CMD_DATAVALID | F_CMD_SOP |
1230 	    V_CMD_EOP(nfrags == 0 && desc_len <= SGE_TX_DESC_MAX_PLEN) |
1231 	    V_CMD_GEN2(gen);
1232 	first_desc_len = (desc_len <= SGE_TX_DESC_MAX_PLEN) ?
1233 	    desc_len : SGE_TX_DESC_MAX_PLEN;
1234 	e->addr_lo = (u32)desc_mapping;
1235 	e->addr_hi = (u64)desc_mapping >> 32;
1236 	e->len_gen = V_CMD_LEN(first_desc_len) | V_CMD_GEN1(gen);
1237 	ce->skb = NULL;
1238 	dma_unmap_len_set(ce, dma_len, 0);
1239 
1240 	if (PAGE_SIZE > SGE_TX_DESC_MAX_PLEN &&
1241 	    desc_len > SGE_TX_DESC_MAX_PLEN) {
1242 		desc_mapping += first_desc_len;
1243 		desc_len -= first_desc_len;
1244 		e1++;
1245 		ce++;
1246 		if (++pidx == q->size) {
1247 			pidx = 0;
1248 			gen ^= 1;
1249 			e1 = q->entries;
1250 			ce = q->centries;
1251 		}
1252 		pidx = write_large_page_tx_descs(pidx, &e1, &ce, &gen,
1253 						 &desc_mapping, &desc_len,
1254 						 nfrags, q);
1255 
1256 		if (likely(desc_len))
1257 			write_tx_desc(e1, desc_mapping, desc_len, gen,
1258 				      nfrags == 0);
1259 	}
1260 
1261 	ce->skb = NULL;
1262 	dma_unmap_addr_set(ce, dma_addr, mapping);
1263 	dma_unmap_len_set(ce, dma_len, skb_headlen(skb));
1264 
1265 	for (i = 0; nfrags--; i++) {
1266 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
1267 		e1++;
1268 		ce++;
1269 		if (++pidx == q->size) {
1270 			pidx = 0;
1271 			gen ^= 1;
1272 			e1 = q->entries;
1273 			ce = q->centries;
1274 		}
1275 
1276 		mapping = skb_frag_dma_map(&adapter->pdev->dev, frag, 0,
1277 					   skb_frag_size(frag), DMA_TO_DEVICE);
1278 		desc_mapping = mapping;
1279 		desc_len = skb_frag_size(frag);
1280 
1281 		pidx = write_large_page_tx_descs(pidx, &e1, &ce, &gen,
1282 						 &desc_mapping, &desc_len,
1283 						 nfrags, q);
1284 		if (likely(desc_len))
1285 			write_tx_desc(e1, desc_mapping, desc_len, gen,
1286 				      nfrags == 0);
1287 		ce->skb = NULL;
1288 		dma_unmap_addr_set(ce, dma_addr, mapping);
1289 		dma_unmap_len_set(ce, dma_len, skb_frag_size(frag));
1290 	}
1291 	ce->skb = skb;
1292 	wmb();
1293 	e->flags = flags;
1294 }
1295 
1296 /*
1297  * Clean up completed Tx buffers.
1298  */
1299 static inline void reclaim_completed_tx(struct sge *sge, struct cmdQ *q)
1300 {
1301 	unsigned int reclaim = q->processed - q->cleaned;
1302 
1303 	if (reclaim) {
1304 		pr_debug("reclaim_completed_tx processed:%d cleaned:%d\n",
1305 			 q->processed, q->cleaned);
1306 		free_cmdQ_buffers(sge, q, reclaim);
1307 		q->cleaned += reclaim;
1308 	}
1309 }
1310 
1311 /*
1312  * Called from tasklet. Checks the scheduler for any
1313  * pending skbs that can be sent.
1314  */
1315 static void restart_sched(struct tasklet_struct *t)
1316 {
1317 	struct sched *s = from_tasklet(s, t, sched_tsk);
1318 	struct sge *sge = s->sge;
1319 	struct adapter *adapter = sge->adapter;
1320 	struct cmdQ *q = &sge->cmdQ[0];
1321 	struct sk_buff *skb;
1322 	unsigned int credits, queued_skb = 0;
1323 
1324 	spin_lock(&q->lock);
1325 	reclaim_completed_tx(sge, q);
1326 
1327 	credits = q->size - q->in_use;
1328 	pr_debug("restart_sched credits=%d\n", credits);
1329 	while ((skb = sched_skb(sge, NULL, credits)) != NULL) {
1330 		unsigned int genbit, pidx, count;
1331 	        count = 1 + skb_shinfo(skb)->nr_frags;
1332 		count += compute_large_page_tx_descs(skb);
1333 		q->in_use += count;
1334 		genbit = q->genbit;
1335 		pidx = q->pidx;
1336 		q->pidx += count;
1337 		if (q->pidx >= q->size) {
1338 			q->pidx -= q->size;
1339 			q->genbit ^= 1;
1340 		}
1341 		write_tx_descs(adapter, skb, pidx, genbit, q);
1342 	        credits = q->size - q->in_use;
1343 		queued_skb = 1;
1344 	}
1345 
1346 	if (queued_skb) {
1347 		clear_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1348 		if (test_and_set_bit(CMDQ_STAT_RUNNING, &q->status) == 0) {
1349 			set_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1350 			writel(F_CMDQ0_ENABLE, adapter->regs + A_SG_DOORBELL);
1351 		}
1352 	}
1353 	spin_unlock(&q->lock);
1354 }
1355 
1356 /**
1357  *	sge_rx - process an ingress ethernet packet
1358  *	@sge: the sge structure
1359  *	@fl: the free list that contains the packet buffer
1360  *	@len: the packet length
1361  *
1362  *	Process an ingress ethernet packet and deliver it to the stack.
1363  */
1364 static void sge_rx(struct sge *sge, struct freelQ *fl, unsigned int len)
1365 {
1366 	struct sk_buff *skb;
1367 	const struct cpl_rx_pkt *p;
1368 	struct adapter *adapter = sge->adapter;
1369 	struct sge_port_stats *st;
1370 	struct net_device *dev;
1371 
1372 	skb = get_packet(adapter, fl, len - sge->rx_pkt_pad);
1373 	if (unlikely(!skb)) {
1374 		sge->stats.rx_drops++;
1375 		return;
1376 	}
1377 
1378 	p = (const struct cpl_rx_pkt *) skb->data;
1379 	if (p->iff >= adapter->params.nports) {
1380 		kfree_skb(skb);
1381 		return;
1382 	}
1383 	__skb_pull(skb, sizeof(*p));
1384 
1385 	st = this_cpu_ptr(sge->port_stats[p->iff]);
1386 	dev = adapter->port[p->iff].dev;
1387 
1388 	skb->protocol = eth_type_trans(skb, dev);
1389 	if ((dev->features & NETIF_F_RXCSUM) && p->csum == 0xffff &&
1390 	    skb->protocol == htons(ETH_P_IP) &&
1391 	    (skb->data[9] == IPPROTO_TCP || skb->data[9] == IPPROTO_UDP)) {
1392 		++st->rx_cso_good;
1393 		skb->ip_summed = CHECKSUM_UNNECESSARY;
1394 	} else
1395 		skb_checksum_none_assert(skb);
1396 
1397 	if (p->vlan_valid) {
1398 		st->vlan_xtract++;
1399 		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), ntohs(p->vlan));
1400 	}
1401 	netif_receive_skb(skb);
1402 }
1403 
1404 /*
1405  * Returns true if a command queue has enough available descriptors that
1406  * we can resume Tx operation after temporarily disabling its packet queue.
1407  */
1408 static inline int enough_free_Tx_descs(const struct cmdQ *q)
1409 {
1410 	unsigned int r = q->processed - q->cleaned;
1411 
1412 	return q->in_use - r < (q->size >> 1);
1413 }
1414 
1415 /*
1416  * Called when sufficient space has become available in the SGE command queues
1417  * after the Tx packet schedulers have been suspended to restart the Tx path.
1418  */
1419 static void restart_tx_queues(struct sge *sge)
1420 {
1421 	struct adapter *adap = sge->adapter;
1422 	int i;
1423 
1424 	if (!enough_free_Tx_descs(&sge->cmdQ[0]))
1425 		return;
1426 
1427 	for_each_port(adap, i) {
1428 		struct net_device *nd = adap->port[i].dev;
1429 
1430 		if (test_and_clear_bit(nd->if_port, &sge->stopped_tx_queues) &&
1431 		    netif_running(nd)) {
1432 			sge->stats.cmdQ_restarted[2]++;
1433 			netif_wake_queue(nd);
1434 		}
1435 	}
1436 }
1437 
1438 /*
1439  * update_tx_info is called from the interrupt handler/NAPI to return cmdQ0
1440  * information.
1441  */
1442 static unsigned int update_tx_info(struct adapter *adapter,
1443 					  unsigned int flags,
1444 					  unsigned int pr0)
1445 {
1446 	struct sge *sge = adapter->sge;
1447 	struct cmdQ *cmdq = &sge->cmdQ[0];
1448 
1449 	cmdq->processed += pr0;
1450 	if (flags & (F_FL0_ENABLE | F_FL1_ENABLE)) {
1451 		freelQs_empty(sge);
1452 		flags &= ~(F_FL0_ENABLE | F_FL1_ENABLE);
1453 	}
1454 	if (flags & F_CMDQ0_ENABLE) {
1455 		clear_bit(CMDQ_STAT_RUNNING, &cmdq->status);
1456 
1457 		if (cmdq->cleaned + cmdq->in_use != cmdq->processed &&
1458 		    !test_and_set_bit(CMDQ_STAT_LAST_PKT_DB, &cmdq->status)) {
1459 			set_bit(CMDQ_STAT_RUNNING, &cmdq->status);
1460 			writel(F_CMDQ0_ENABLE, adapter->regs + A_SG_DOORBELL);
1461 		}
1462 		if (sge->tx_sched)
1463 			tasklet_hi_schedule(&sge->tx_sched->sched_tsk);
1464 
1465 		flags &= ~F_CMDQ0_ENABLE;
1466 	}
1467 
1468 	if (unlikely(sge->stopped_tx_queues != 0))
1469 		restart_tx_queues(sge);
1470 
1471 	return flags;
1472 }
1473 
1474 /*
1475  * Process SGE responses, up to the supplied budget.  Returns the number of
1476  * responses processed.  A negative budget is effectively unlimited.
1477  */
1478 static int process_responses(struct adapter *adapter, int budget)
1479 {
1480 	struct sge *sge = adapter->sge;
1481 	struct respQ *q = &sge->respQ;
1482 	struct respQ_e *e = &q->entries[q->cidx];
1483 	int done = 0;
1484 	unsigned int flags = 0;
1485 	unsigned int cmdq_processed[SGE_CMDQ_N] = {0, 0};
1486 
1487 	while (done < budget && e->GenerationBit == q->genbit) {
1488 		flags |= e->Qsleeping;
1489 
1490 		cmdq_processed[0] += e->Cmdq0CreditReturn;
1491 		cmdq_processed[1] += e->Cmdq1CreditReturn;
1492 
1493 		/* We batch updates to the TX side to avoid cacheline
1494 		 * ping-pong of TX state information on MP where the sender
1495 		 * might run on a different CPU than this function...
1496 		 */
1497 		if (unlikely((flags & F_CMDQ0_ENABLE) || cmdq_processed[0] > 64)) {
1498 			flags = update_tx_info(adapter, flags, cmdq_processed[0]);
1499 			cmdq_processed[0] = 0;
1500 		}
1501 
1502 		if (unlikely(cmdq_processed[1] > 16)) {
1503 			sge->cmdQ[1].processed += cmdq_processed[1];
1504 			cmdq_processed[1] = 0;
1505 		}
1506 
1507 		if (likely(e->DataValid)) {
1508 			struct freelQ *fl = &sge->freelQ[e->FreelistQid];
1509 
1510 			BUG_ON(!e->Sop || !e->Eop);
1511 			if (unlikely(e->Offload))
1512 				unexpected_offload(adapter, fl);
1513 			else
1514 				sge_rx(sge, fl, e->BufferLength);
1515 
1516 			++done;
1517 
1518 			/*
1519 			 * Note: this depends on each packet consuming a
1520 			 * single free-list buffer; cf. the BUG above.
1521 			 */
1522 			if (++fl->cidx == fl->size)
1523 				fl->cidx = 0;
1524 			prefetch(fl->centries[fl->cidx].skb);
1525 
1526 			if (unlikely(--fl->credits <
1527 				     fl->size - SGE_FREEL_REFILL_THRESH))
1528 				refill_free_list(sge, fl);
1529 		} else
1530 			sge->stats.pure_rsps++;
1531 
1532 		e++;
1533 		if (unlikely(++q->cidx == q->size)) {
1534 			q->cidx = 0;
1535 			q->genbit ^= 1;
1536 			e = q->entries;
1537 		}
1538 		prefetch(e);
1539 
1540 		if (++q->credits > SGE_RESPQ_REPLENISH_THRES) {
1541 			writel(q->credits, adapter->regs + A_SG_RSPQUEUECREDIT);
1542 			q->credits = 0;
1543 		}
1544 	}
1545 
1546 	flags = update_tx_info(adapter, flags, cmdq_processed[0]);
1547 	sge->cmdQ[1].processed += cmdq_processed[1];
1548 
1549 	return done;
1550 }
1551 
1552 static inline int responses_pending(const struct adapter *adapter)
1553 {
1554 	const struct respQ *Q = &adapter->sge->respQ;
1555 	const struct respQ_e *e = &Q->entries[Q->cidx];
1556 
1557 	return e->GenerationBit == Q->genbit;
1558 }
1559 
1560 /*
1561  * A simpler version of process_responses() that handles only pure (i.e.,
1562  * non data-carrying) responses.  Such respones are too light-weight to justify
1563  * calling a softirq when using NAPI, so we handle them specially in hard
1564  * interrupt context.  The function is called with a pointer to a response,
1565  * which the caller must ensure is a valid pure response.  Returns 1 if it
1566  * encounters a valid data-carrying response, 0 otherwise.
1567  */
1568 static int process_pure_responses(struct adapter *adapter)
1569 {
1570 	struct sge *sge = adapter->sge;
1571 	struct respQ *q = &sge->respQ;
1572 	struct respQ_e *e = &q->entries[q->cidx];
1573 	const struct freelQ *fl = &sge->freelQ[e->FreelistQid];
1574 	unsigned int flags = 0;
1575 	unsigned int cmdq_processed[SGE_CMDQ_N] = {0, 0};
1576 
1577 	prefetch(fl->centries[fl->cidx].skb);
1578 	if (e->DataValid)
1579 		return 1;
1580 
1581 	do {
1582 		flags |= e->Qsleeping;
1583 
1584 		cmdq_processed[0] += e->Cmdq0CreditReturn;
1585 		cmdq_processed[1] += e->Cmdq1CreditReturn;
1586 
1587 		e++;
1588 		if (unlikely(++q->cidx == q->size)) {
1589 			q->cidx = 0;
1590 			q->genbit ^= 1;
1591 			e = q->entries;
1592 		}
1593 		prefetch(e);
1594 
1595 		if (++q->credits > SGE_RESPQ_REPLENISH_THRES) {
1596 			writel(q->credits, adapter->regs + A_SG_RSPQUEUECREDIT);
1597 			q->credits = 0;
1598 		}
1599 		sge->stats.pure_rsps++;
1600 	} while (e->GenerationBit == q->genbit && !e->DataValid);
1601 
1602 	flags = update_tx_info(adapter, flags, cmdq_processed[0]);
1603 	sge->cmdQ[1].processed += cmdq_processed[1];
1604 
1605 	return e->GenerationBit == q->genbit;
1606 }
1607 
1608 /*
1609  * Handler for new data events when using NAPI.  This does not need any locking
1610  * or protection from interrupts as data interrupts are off at this point and
1611  * other adapter interrupts do not interfere.
1612  */
1613 int t1_poll(struct napi_struct *napi, int budget)
1614 {
1615 	struct adapter *adapter = container_of(napi, struct adapter, napi);
1616 	int work_done = process_responses(adapter, budget);
1617 
1618 	if (likely(work_done < budget)) {
1619 		napi_complete_done(napi, work_done);
1620 		writel(adapter->sge->respQ.cidx,
1621 		       adapter->regs + A_SG_SLEEPING);
1622 	}
1623 	return work_done;
1624 }
1625 
1626 irqreturn_t t1_interrupt_thread(int irq, void *data)
1627 {
1628 	struct adapter *adapter = data;
1629 	u32 pending_thread_intr;
1630 
1631 	spin_lock_irq(&adapter->async_lock);
1632 	pending_thread_intr = adapter->pending_thread_intr;
1633 	adapter->pending_thread_intr = 0;
1634 	spin_unlock_irq(&adapter->async_lock);
1635 
1636 	if (!pending_thread_intr)
1637 		return IRQ_NONE;
1638 
1639 	if (pending_thread_intr & F_PL_INTR_EXT)
1640 		t1_elmer0_ext_intr_handler(adapter);
1641 
1642 	/* This error is fatal, interrupts remain off */
1643 	if (pending_thread_intr & F_PL_INTR_SGE_ERR) {
1644 		pr_alert("%s: encountered fatal error, operation suspended\n",
1645 			 adapter->name);
1646 		t1_sge_stop(adapter->sge);
1647 		return IRQ_HANDLED;
1648 	}
1649 
1650 	spin_lock_irq(&adapter->async_lock);
1651 	adapter->slow_intr_mask |= F_PL_INTR_EXT;
1652 
1653 	writel(F_PL_INTR_EXT, adapter->regs + A_PL_CAUSE);
1654 	writel(adapter->slow_intr_mask | F_PL_INTR_SGE_DATA,
1655 	       adapter->regs + A_PL_ENABLE);
1656 	spin_unlock_irq(&adapter->async_lock);
1657 
1658 	return IRQ_HANDLED;
1659 }
1660 
1661 irqreturn_t t1_interrupt(int irq, void *data)
1662 {
1663 	struct adapter *adapter = data;
1664 	struct sge *sge = adapter->sge;
1665 	irqreturn_t handled;
1666 
1667 	if (likely(responses_pending(adapter))) {
1668 		writel(F_PL_INTR_SGE_DATA, adapter->regs + A_PL_CAUSE);
1669 
1670 		if (napi_schedule_prep(&adapter->napi)) {
1671 			if (process_pure_responses(adapter))
1672 				__napi_schedule(&adapter->napi);
1673 			else {
1674 				/* no data, no NAPI needed */
1675 				writel(sge->respQ.cidx, adapter->regs + A_SG_SLEEPING);
1676 				/* undo schedule_prep */
1677 				napi_enable(&adapter->napi);
1678 			}
1679 		}
1680 		return IRQ_HANDLED;
1681 	}
1682 
1683 	spin_lock(&adapter->async_lock);
1684 	handled = t1_slow_intr_handler(adapter);
1685 	spin_unlock(&adapter->async_lock);
1686 
1687 	if (handled == IRQ_NONE)
1688 		sge->stats.unhandled_irqs++;
1689 
1690 	return handled;
1691 }
1692 
1693 /*
1694  * Enqueues the sk_buff onto the cmdQ[qid] and has hardware fetch it.
1695  *
1696  * The code figures out how many entries the sk_buff will require in the
1697  * cmdQ and updates the cmdQ data structure with the state once the enqueue
1698  * has complete. Then, it doesn't access the global structure anymore, but
1699  * uses the corresponding fields on the stack. In conjunction with a spinlock
1700  * around that code, we can make the function reentrant without holding the
1701  * lock when we actually enqueue (which might be expensive, especially on
1702  * architectures with IO MMUs).
1703  *
1704  * This runs with softirqs disabled.
1705  */
1706 static int t1_sge_tx(struct sk_buff *skb, struct adapter *adapter,
1707 		     unsigned int qid, struct net_device *dev)
1708 {
1709 	struct sge *sge = adapter->sge;
1710 	struct cmdQ *q = &sge->cmdQ[qid];
1711 	unsigned int credits, pidx, genbit, count, use_sched_skb = 0;
1712 
1713 	spin_lock(&q->lock);
1714 
1715 	reclaim_completed_tx(sge, q);
1716 
1717 	pidx = q->pidx;
1718 	credits = q->size - q->in_use;
1719 	count = 1 + skb_shinfo(skb)->nr_frags;
1720 	count += compute_large_page_tx_descs(skb);
1721 
1722 	/* Ethernet packet */
1723 	if (unlikely(credits < count)) {
1724 		if (!netif_queue_stopped(dev)) {
1725 			netif_stop_queue(dev);
1726 			set_bit(dev->if_port, &sge->stopped_tx_queues);
1727 			sge->stats.cmdQ_full[2]++;
1728 			pr_err("%s: Tx ring full while queue awake!\n",
1729 			       adapter->name);
1730 		}
1731 		spin_unlock(&q->lock);
1732 		return NETDEV_TX_BUSY;
1733 	}
1734 
1735 	if (unlikely(credits - count < q->stop_thres)) {
1736 		netif_stop_queue(dev);
1737 		set_bit(dev->if_port, &sge->stopped_tx_queues);
1738 		sge->stats.cmdQ_full[2]++;
1739 	}
1740 
1741 	/* T204 cmdQ0 skbs that are destined for a certain port have to go
1742 	 * through the scheduler.
1743 	 */
1744 	if (sge->tx_sched && !qid && skb->dev) {
1745 use_sched:
1746 		use_sched_skb = 1;
1747 		/* Note that the scheduler might return a different skb than
1748 		 * the one passed in.
1749 		 */
1750 		skb = sched_skb(sge, skb, credits);
1751 		if (!skb) {
1752 			spin_unlock(&q->lock);
1753 			return NETDEV_TX_OK;
1754 		}
1755 		pidx = q->pidx;
1756 		count = 1 + skb_shinfo(skb)->nr_frags;
1757 		count += compute_large_page_tx_descs(skb);
1758 	}
1759 
1760 	q->in_use += count;
1761 	genbit = q->genbit;
1762 	pidx = q->pidx;
1763 	q->pidx += count;
1764 	if (q->pidx >= q->size) {
1765 		q->pidx -= q->size;
1766 		q->genbit ^= 1;
1767 	}
1768 	spin_unlock(&q->lock);
1769 
1770 	write_tx_descs(adapter, skb, pidx, genbit, q);
1771 
1772 	/*
1773 	 * We always ring the doorbell for cmdQ1.  For cmdQ0, we only ring
1774 	 * the doorbell if the Q is asleep. There is a natural race, where
1775 	 * the hardware is going to sleep just after we checked, however,
1776 	 * then the interrupt handler will detect the outstanding TX packet
1777 	 * and ring the doorbell for us.
1778 	 */
1779 	if (qid)
1780 		doorbell_pio(adapter, F_CMDQ1_ENABLE);
1781 	else {
1782 		clear_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1783 		if (test_and_set_bit(CMDQ_STAT_RUNNING, &q->status) == 0) {
1784 			set_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1785 			writel(F_CMDQ0_ENABLE, adapter->regs + A_SG_DOORBELL);
1786 		}
1787 	}
1788 
1789 	if (use_sched_skb) {
1790 		if (spin_trylock(&q->lock)) {
1791 			credits = q->size - q->in_use;
1792 			skb = NULL;
1793 			goto use_sched;
1794 		}
1795 	}
1796 	return NETDEV_TX_OK;
1797 }
1798 
1799 #define MK_ETH_TYPE_MSS(type, mss) (((mss) & 0x3FFF) | ((type) << 14))
1800 
1801 /*
1802  *	eth_hdr_len - return the length of an Ethernet header
1803  *	@data: pointer to the start of the Ethernet header
1804  *
1805  *	Returns the length of an Ethernet header, including optional VLAN tag.
1806  */
1807 static inline int eth_hdr_len(const void *data)
1808 {
1809 	const struct ethhdr *e = data;
1810 
1811 	return e->h_proto == htons(ETH_P_8021Q) ? VLAN_ETH_HLEN : ETH_HLEN;
1812 }
1813 
1814 /*
1815  * Adds the CPL header to the sk_buff and passes it to t1_sge_tx.
1816  */
1817 netdev_tx_t t1_start_xmit(struct sk_buff *skb, struct net_device *dev)
1818 {
1819 	struct adapter *adapter = dev->ml_priv;
1820 	struct sge *sge = adapter->sge;
1821 	struct sge_port_stats *st = this_cpu_ptr(sge->port_stats[dev->if_port]);
1822 	struct cpl_tx_pkt *cpl;
1823 	struct sk_buff *orig_skb = skb;
1824 	int ret;
1825 
1826 	if (skb->protocol == htons(ETH_P_CPL5))
1827 		goto send;
1828 
1829 	/*
1830 	 * We are using a non-standard hard_header_len.
1831 	 * Allocate more header room in the rare cases it is not big enough.
1832 	 */
1833 	if (unlikely(skb_headroom(skb) < dev->hard_header_len - ETH_HLEN)) {
1834 		skb = skb_realloc_headroom(skb, sizeof(struct cpl_tx_pkt_lso));
1835 		++st->tx_need_hdrroom;
1836 		dev_kfree_skb_any(orig_skb);
1837 		if (!skb)
1838 			return NETDEV_TX_OK;
1839 	}
1840 
1841 	if (skb_shinfo(skb)->gso_size) {
1842 		int eth_type;
1843 		struct cpl_tx_pkt_lso *hdr;
1844 
1845 		++st->tx_tso;
1846 
1847 		eth_type = skb_network_offset(skb) == ETH_HLEN ?
1848 			CPL_ETH_II : CPL_ETH_II_VLAN;
1849 
1850 		hdr = skb_push(skb, sizeof(*hdr));
1851 		hdr->opcode = CPL_TX_PKT_LSO;
1852 		hdr->ip_csum_dis = hdr->l4_csum_dis = 0;
1853 		hdr->ip_hdr_words = ip_hdr(skb)->ihl;
1854 		hdr->tcp_hdr_words = tcp_hdr(skb)->doff;
1855 		hdr->eth_type_mss = htons(MK_ETH_TYPE_MSS(eth_type,
1856 							  skb_shinfo(skb)->gso_size));
1857 		hdr->len = htonl(skb->len - sizeof(*hdr));
1858 		cpl = (struct cpl_tx_pkt *)hdr;
1859 	} else {
1860 		/*
1861 		 * Packets shorter than ETH_HLEN can break the MAC, drop them
1862 		 * early.  Also, we may get oversized packets because some
1863 		 * parts of the kernel don't handle our unusual hard_header_len
1864 		 * right, drop those too.
1865 		 */
1866 		if (unlikely(skb->len < ETH_HLEN ||
1867 			     skb->len > dev->mtu + eth_hdr_len(skb->data))) {
1868 			netdev_dbg(dev, "packet size %d hdr %d mtu%d\n",
1869 				   skb->len, eth_hdr_len(skb->data), dev->mtu);
1870 			dev_kfree_skb_any(skb);
1871 			return NETDEV_TX_OK;
1872 		}
1873 
1874 		if (skb->ip_summed == CHECKSUM_PARTIAL &&
1875 		    ip_hdr(skb)->protocol == IPPROTO_UDP) {
1876 			if (unlikely(skb_checksum_help(skb))) {
1877 				netdev_dbg(dev, "unable to do udp checksum\n");
1878 				dev_kfree_skb_any(skb);
1879 				return NETDEV_TX_OK;
1880 			}
1881 		}
1882 
1883 		/* Hmmm, assuming to catch the gratious arp... and we'll use
1884 		 * it to flush out stuck espi packets...
1885 		 */
1886 		if ((unlikely(!adapter->sge->espibug_skb[dev->if_port]))) {
1887 			if (skb->protocol == htons(ETH_P_ARP) &&
1888 			    arp_hdr(skb)->ar_op == htons(ARPOP_REQUEST)) {
1889 				adapter->sge->espibug_skb[dev->if_port] = skb;
1890 				/* We want to re-use this skb later. We
1891 				 * simply bump the reference count and it
1892 				 * will not be freed...
1893 				 */
1894 				skb = skb_get(skb);
1895 			}
1896 		}
1897 
1898 		cpl = __skb_push(skb, sizeof(*cpl));
1899 		cpl->opcode = CPL_TX_PKT;
1900 		cpl->ip_csum_dis = 1;    /* SW calculates IP csum */
1901 		cpl->l4_csum_dis = skb->ip_summed == CHECKSUM_PARTIAL ? 0 : 1;
1902 		/* the length field isn't used so don't bother setting it */
1903 
1904 		st->tx_cso += (skb->ip_summed == CHECKSUM_PARTIAL);
1905 	}
1906 	cpl->iff = dev->if_port;
1907 
1908 	if (skb_vlan_tag_present(skb)) {
1909 		cpl->vlan_valid = 1;
1910 		cpl->vlan = htons(skb_vlan_tag_get(skb));
1911 		st->vlan_insert++;
1912 	} else
1913 		cpl->vlan_valid = 0;
1914 
1915 send:
1916 	ret = t1_sge_tx(skb, adapter, 0, dev);
1917 
1918 	/* If transmit busy, and we reallocated skb's due to headroom limit,
1919 	 * then silently discard to avoid leak.
1920 	 */
1921 	if (unlikely(ret != NETDEV_TX_OK && skb != orig_skb)) {
1922 		dev_kfree_skb_any(skb);
1923 		ret = NETDEV_TX_OK;
1924 	}
1925 	return ret;
1926 }
1927 
1928 /*
1929  * Callback for the Tx buffer reclaim timer.  Runs with softirqs disabled.
1930  */
1931 static void sge_tx_reclaim_cb(struct timer_list *t)
1932 {
1933 	int i;
1934 	struct sge *sge = from_timer(sge, t, tx_reclaim_timer);
1935 
1936 	for (i = 0; i < SGE_CMDQ_N; ++i) {
1937 		struct cmdQ *q = &sge->cmdQ[i];
1938 
1939 		if (!spin_trylock(&q->lock))
1940 			continue;
1941 
1942 		reclaim_completed_tx(sge, q);
1943 		if (i == 0 && q->in_use) {    /* flush pending credits */
1944 			writel(F_CMDQ0_ENABLE, sge->adapter->regs + A_SG_DOORBELL);
1945 		}
1946 		spin_unlock(&q->lock);
1947 	}
1948 	mod_timer(&sge->tx_reclaim_timer, jiffies + TX_RECLAIM_PERIOD);
1949 }
1950 
1951 /*
1952  * Propagate changes of the SGE coalescing parameters to the HW.
1953  */
1954 int t1_sge_set_coalesce_params(struct sge *sge, struct sge_params *p)
1955 {
1956 	sge->fixed_intrtimer = p->rx_coalesce_usecs *
1957 		core_ticks_per_usec(sge->adapter);
1958 	writel(sge->fixed_intrtimer, sge->adapter->regs + A_SG_INTRTIMER);
1959 	return 0;
1960 }
1961 
1962 /*
1963  * Allocates both RX and TX resources and configures the SGE. However,
1964  * the hardware is not enabled yet.
1965  */
1966 int t1_sge_configure(struct sge *sge, struct sge_params *p)
1967 {
1968 	if (alloc_rx_resources(sge, p))
1969 		return -ENOMEM;
1970 	if (alloc_tx_resources(sge, p)) {
1971 		free_rx_resources(sge);
1972 		return -ENOMEM;
1973 	}
1974 	configure_sge(sge, p);
1975 
1976 	/*
1977 	 * Now that we have sized the free lists calculate the payload
1978 	 * capacity of the large buffers.  Other parts of the driver use
1979 	 * this to set the max offload coalescing size so that RX packets
1980 	 * do not overflow our large buffers.
1981 	 */
1982 	p->large_buf_capacity = jumbo_payload_capacity(sge);
1983 	return 0;
1984 }
1985 
1986 /*
1987  * Disables the DMA engine.
1988  */
1989 void t1_sge_stop(struct sge *sge)
1990 {
1991 	int i;
1992 	writel(0, sge->adapter->regs + A_SG_CONTROL);
1993 	readl(sge->adapter->regs + A_SG_CONTROL); /* flush */
1994 
1995 	if (is_T2(sge->adapter))
1996 		del_timer_sync(&sge->espibug_timer);
1997 
1998 	del_timer_sync(&sge->tx_reclaim_timer);
1999 	if (sge->tx_sched)
2000 		tx_sched_stop(sge);
2001 
2002 	for (i = 0; i < MAX_NPORTS; i++)
2003 		kfree_skb(sge->espibug_skb[i]);
2004 }
2005 
2006 /*
2007  * Enables the DMA engine.
2008  */
2009 void t1_sge_start(struct sge *sge)
2010 {
2011 	refill_free_list(sge, &sge->freelQ[0]);
2012 	refill_free_list(sge, &sge->freelQ[1]);
2013 
2014 	writel(sge->sge_control, sge->adapter->regs + A_SG_CONTROL);
2015 	doorbell_pio(sge->adapter, F_FL0_ENABLE | F_FL1_ENABLE);
2016 	readl(sge->adapter->regs + A_SG_CONTROL); /* flush */
2017 
2018 	mod_timer(&sge->tx_reclaim_timer, jiffies + TX_RECLAIM_PERIOD);
2019 
2020 	if (is_T2(sge->adapter))
2021 		mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
2022 }
2023 
2024 /*
2025  * Callback for the T2 ESPI 'stuck packet feature' workaorund
2026  */
2027 static void espibug_workaround_t204(struct timer_list *t)
2028 {
2029 	struct sge *sge = from_timer(sge, t, espibug_timer);
2030 	struct adapter *adapter = sge->adapter;
2031 	unsigned int nports = adapter->params.nports;
2032 	u32 seop[MAX_NPORTS];
2033 
2034 	if (adapter->open_device_map & PORT_MASK) {
2035 		int i;
2036 
2037 		if (t1_espi_get_mon_t204(adapter, &(seop[0]), 0) < 0)
2038 			return;
2039 
2040 		for (i = 0; i < nports; i++) {
2041 			struct sk_buff *skb = sge->espibug_skb[i];
2042 
2043 			if (!netif_running(adapter->port[i].dev) ||
2044 			    netif_queue_stopped(adapter->port[i].dev) ||
2045 			    !seop[i] || ((seop[i] & 0xfff) != 0) || !skb)
2046 				continue;
2047 
2048 			if (!skb->cb[0]) {
2049 				skb_copy_to_linear_data_offset(skb,
2050 						    sizeof(struct cpl_tx_pkt),
2051 							       ch_mac_addr,
2052 							       ETH_ALEN);
2053 				skb_copy_to_linear_data_offset(skb,
2054 							       skb->len - 10,
2055 							       ch_mac_addr,
2056 							       ETH_ALEN);
2057 				skb->cb[0] = 0xff;
2058 			}
2059 
2060 			/* bump the reference count to avoid freeing of
2061 			 * the skb once the DMA has completed.
2062 			 */
2063 			skb = skb_get(skb);
2064 			t1_sge_tx(skb, adapter, 0, adapter->port[i].dev);
2065 		}
2066 	}
2067 	mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
2068 }
2069 
2070 static void espibug_workaround(struct timer_list *t)
2071 {
2072 	struct sge *sge = from_timer(sge, t, espibug_timer);
2073 	struct adapter *adapter = sge->adapter;
2074 
2075 	if (netif_running(adapter->port[0].dev)) {
2076 	        struct sk_buff *skb = sge->espibug_skb[0];
2077 	        u32 seop = t1_espi_get_mon(adapter, 0x930, 0);
2078 
2079 	        if ((seop & 0xfff0fff) == 0xfff && skb) {
2080 	                if (!skb->cb[0]) {
2081 	                        skb_copy_to_linear_data_offset(skb,
2082 						     sizeof(struct cpl_tx_pkt),
2083 							       ch_mac_addr,
2084 							       ETH_ALEN);
2085 	                        skb_copy_to_linear_data_offset(skb,
2086 							       skb->len - 10,
2087 							       ch_mac_addr,
2088 							       ETH_ALEN);
2089 	                        skb->cb[0] = 0xff;
2090 	                }
2091 
2092 	                /* bump the reference count to avoid freeing of the
2093 	                 * skb once the DMA has completed.
2094 	                 */
2095 	                skb = skb_get(skb);
2096 	                t1_sge_tx(skb, adapter, 0, adapter->port[0].dev);
2097 	        }
2098 	}
2099 	mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
2100 }
2101 
2102 /*
2103  * Creates a t1_sge structure and returns suggested resource parameters.
2104  */
2105 struct sge *t1_sge_create(struct adapter *adapter, struct sge_params *p)
2106 {
2107 	struct sge *sge = kzalloc(sizeof(*sge), GFP_KERNEL);
2108 	int i;
2109 
2110 	if (!sge)
2111 		return NULL;
2112 
2113 	sge->adapter = adapter;
2114 	sge->netdev = adapter->port[0].dev;
2115 	sge->rx_pkt_pad = t1_is_T1B(adapter) ? 0 : 2;
2116 	sge->jumbo_fl = t1_is_T1B(adapter) ? 1 : 0;
2117 
2118 	for_each_port(adapter, i) {
2119 		sge->port_stats[i] = alloc_percpu(struct sge_port_stats);
2120 		if (!sge->port_stats[i])
2121 			goto nomem_port;
2122 	}
2123 
2124 	timer_setup(&sge->tx_reclaim_timer, sge_tx_reclaim_cb, 0);
2125 
2126 	if (is_T2(sge->adapter)) {
2127 		timer_setup(&sge->espibug_timer,
2128 			    adapter->params.nports > 1 ? espibug_workaround_t204 : espibug_workaround,
2129 			    0);
2130 
2131 		if (adapter->params.nports > 1)
2132 			tx_sched_init(sge);
2133 
2134 		sge->espibug_timeout = 1;
2135 		/* for T204, every 10ms */
2136 		if (adapter->params.nports > 1)
2137 			sge->espibug_timeout = HZ/100;
2138 	}
2139 
2140 
2141 	p->cmdQ_size[0] = SGE_CMDQ0_E_N;
2142 	p->cmdQ_size[1] = SGE_CMDQ1_E_N;
2143 	p->freelQ_size[!sge->jumbo_fl] = SGE_FREEL_SIZE;
2144 	p->freelQ_size[sge->jumbo_fl] = SGE_JUMBO_FREEL_SIZE;
2145 	if (sge->tx_sched) {
2146 		if (board_info(sge->adapter)->board == CHBT_BOARD_CHT204)
2147 			p->rx_coalesce_usecs = 15;
2148 		else
2149 			p->rx_coalesce_usecs = 50;
2150 	} else
2151 		p->rx_coalesce_usecs = 50;
2152 
2153 	p->coalesce_enable = 0;
2154 	p->sample_interval_usecs = 0;
2155 
2156 	return sge;
2157 nomem_port:
2158 	while (i >= 0) {
2159 		free_percpu(sge->port_stats[i]);
2160 		--i;
2161 	}
2162 	kfree(sge);
2163 	return NULL;
2164 
2165 }
2166