xref: /linux/drivers/net/ethernet/chelsio/cxgb/sge.c (revision 2a906db2824b75444982f5e9df870106982afca8)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*****************************************************************************
3  *                                                                           *
4  * File: sge.c                                                               *
5  * $Revision: 1.26 $                                                         *
6  * $Date: 2005/06/21 18:29:48 $                                              *
7  * Description:                                                              *
8  *  DMA engine.                                                              *
9  *  part of the Chelsio 10Gb Ethernet Driver.                                *
10  *                                                                           *
11  *                                                                           *
12  * http://www.chelsio.com                                                    *
13  *                                                                           *
14  * Copyright (c) 2003 - 2005 Chelsio Communications, Inc.                    *
15  * All rights reserved.                                                      *
16  *                                                                           *
17  * Maintainers: maintainers@chelsio.com                                      *
18  *                                                                           *
19  * Authors: Dimitrios Michailidis   <dm@chelsio.com>                         *
20  *          Tina Yang               <tainay@chelsio.com>                     *
21  *          Felix Marti             <felix@chelsio.com>                      *
22  *          Scott Bardone           <sbardone@chelsio.com>                   *
23  *          Kurt Ottaway            <kottaway@chelsio.com>                   *
24  *          Frank DiMambro          <frank@chelsio.com>                      *
25  *                                                                           *
26  * History:                                                                  *
27  *                                                                           *
28  ****************************************************************************/
29 
30 #include "common.h"
31 
32 #include <linux/types.h>
33 #include <linux/errno.h>
34 #include <linux/pci.h>
35 #include <linux/ktime.h>
36 #include <linux/netdevice.h>
37 #include <linux/etherdevice.h>
38 #include <linux/if_vlan.h>
39 #include <linux/skbuff.h>
40 #include <linux/mm.h>
41 #include <linux/tcp.h>
42 #include <linux/ip.h>
43 #include <linux/in.h>
44 #include <linux/if_arp.h>
45 #include <linux/slab.h>
46 #include <linux/prefetch.h>
47 
48 #include "cpl5_cmd.h"
49 #include "sge.h"
50 #include "regs.h"
51 #include "espi.h"
52 
53 /* This belongs in if_ether.h */
54 #define ETH_P_CPL5 0xf
55 
56 #define SGE_CMDQ_N		2
57 #define SGE_FREELQ_N		2
58 #define SGE_CMDQ0_E_N		1024
59 #define SGE_CMDQ1_E_N		128
60 #define SGE_FREEL_SIZE		4096
61 #define SGE_JUMBO_FREEL_SIZE	512
62 #define SGE_FREEL_REFILL_THRESH	16
63 #define SGE_RESPQ_E_N		1024
64 #define SGE_INTRTIMER_NRES	1000
65 #define SGE_RX_SM_BUF_SIZE	1536
66 #define SGE_TX_DESC_MAX_PLEN	16384
67 
68 #define SGE_RESPQ_REPLENISH_THRES (SGE_RESPQ_E_N / 4)
69 
70 /*
71  * Period of the TX buffer reclaim timer.  This timer does not need to run
72  * frequently as TX buffers are usually reclaimed by new TX packets.
73  */
74 #define TX_RECLAIM_PERIOD (HZ / 4)
75 
76 #define M_CMD_LEN       0x7fffffff
77 #define V_CMD_LEN(v)    (v)
78 #define G_CMD_LEN(v)    ((v) & M_CMD_LEN)
79 #define V_CMD_GEN1(v)   ((v) << 31)
80 #define V_CMD_GEN2(v)   (v)
81 #define F_CMD_DATAVALID (1 << 1)
82 #define F_CMD_SOP       (1 << 2)
83 #define V_CMD_EOP(v)    ((v) << 3)
84 
85 /*
86  * Command queue, receive buffer list, and response queue descriptors.
87  */
88 #if defined(__BIG_ENDIAN_BITFIELD)
89 struct cmdQ_e {
90 	u32 addr_lo;
91 	u32 len_gen;
92 	u32 flags;
93 	u32 addr_hi;
94 };
95 
96 struct freelQ_e {
97 	u32 addr_lo;
98 	u32 len_gen;
99 	u32 gen2;
100 	u32 addr_hi;
101 };
102 
103 struct respQ_e {
104 	u32 Qsleeping		: 4;
105 	u32 Cmdq1CreditReturn	: 5;
106 	u32 Cmdq1DmaComplete	: 5;
107 	u32 Cmdq0CreditReturn	: 5;
108 	u32 Cmdq0DmaComplete	: 5;
109 	u32 FreelistQid		: 2;
110 	u32 CreditValid		: 1;
111 	u32 DataValid		: 1;
112 	u32 Offload		: 1;
113 	u32 Eop			: 1;
114 	u32 Sop			: 1;
115 	u32 GenerationBit	: 1;
116 	u32 BufferLength;
117 };
118 #elif defined(__LITTLE_ENDIAN_BITFIELD)
119 struct cmdQ_e {
120 	u32 len_gen;
121 	u32 addr_lo;
122 	u32 addr_hi;
123 	u32 flags;
124 };
125 
126 struct freelQ_e {
127 	u32 len_gen;
128 	u32 addr_lo;
129 	u32 addr_hi;
130 	u32 gen2;
131 };
132 
133 struct respQ_e {
134 	u32 BufferLength;
135 	u32 GenerationBit	: 1;
136 	u32 Sop			: 1;
137 	u32 Eop			: 1;
138 	u32 Offload		: 1;
139 	u32 DataValid		: 1;
140 	u32 CreditValid		: 1;
141 	u32 FreelistQid		: 2;
142 	u32 Cmdq0DmaComplete	: 5;
143 	u32 Cmdq0CreditReturn	: 5;
144 	u32 Cmdq1DmaComplete	: 5;
145 	u32 Cmdq1CreditReturn	: 5;
146 	u32 Qsleeping		: 4;
147 } ;
148 #endif
149 
150 /*
151  * SW Context Command and Freelist Queue Descriptors
152  */
153 struct cmdQ_ce {
154 	struct sk_buff *skb;
155 	DEFINE_DMA_UNMAP_ADDR(dma_addr);
156 	DEFINE_DMA_UNMAP_LEN(dma_len);
157 };
158 
159 struct freelQ_ce {
160 	struct sk_buff *skb;
161 	DEFINE_DMA_UNMAP_ADDR(dma_addr);
162 	DEFINE_DMA_UNMAP_LEN(dma_len);
163 };
164 
165 /*
166  * SW command, freelist and response rings
167  */
168 struct cmdQ {
169 	unsigned long   status;         /* HW DMA fetch status */
170 	unsigned int    in_use;         /* # of in-use command descriptors */
171 	unsigned int	size;	        /* # of descriptors */
172 	unsigned int    processed;      /* total # of descs HW has processed */
173 	unsigned int    cleaned;        /* total # of descs SW has reclaimed */
174 	unsigned int    stop_thres;     /* SW TX queue suspend threshold */
175 	u16		pidx;           /* producer index (SW) */
176 	u16		cidx;           /* consumer index (HW) */
177 	u8		genbit;         /* current generation (=valid) bit */
178 	u8              sop;            /* is next entry start of packet? */
179 	struct cmdQ_e  *entries;        /* HW command descriptor Q */
180 	struct cmdQ_ce *centries;       /* SW command context descriptor Q */
181 	dma_addr_t	dma_addr;       /* DMA addr HW command descriptor Q */
182 	spinlock_t	lock;           /* Lock to protect cmdQ enqueuing */
183 };
184 
185 struct freelQ {
186 	unsigned int	credits;        /* # of available RX buffers */
187 	unsigned int	size;	        /* free list capacity */
188 	u16		pidx;           /* producer index (SW) */
189 	u16		cidx;           /* consumer index (HW) */
190 	u16		rx_buffer_size; /* Buffer size on this free list */
191 	u16             dma_offset;     /* DMA offset to align IP headers */
192 	u16             recycleq_idx;   /* skb recycle q to use */
193 	u8		genbit;	        /* current generation (=valid) bit */
194 	struct freelQ_e	*entries;       /* HW freelist descriptor Q */
195 	struct freelQ_ce *centries;     /* SW freelist context descriptor Q */
196 	dma_addr_t	dma_addr;       /* DMA addr HW freelist descriptor Q */
197 };
198 
199 struct respQ {
200 	unsigned int	credits;        /* credits to be returned to SGE */
201 	unsigned int	size;	        /* # of response Q descriptors */
202 	u16		cidx;	        /* consumer index (SW) */
203 	u8		genbit;	        /* current generation(=valid) bit */
204 	struct respQ_e *entries;        /* HW response descriptor Q */
205 	dma_addr_t	dma_addr;       /* DMA addr HW response descriptor Q */
206 };
207 
208 /* Bit flags for cmdQ.status */
209 enum {
210 	CMDQ_STAT_RUNNING = 1,          /* fetch engine is running */
211 	CMDQ_STAT_LAST_PKT_DB = 2       /* last packet rung the doorbell */
212 };
213 
214 /* T204 TX SW scheduler */
215 
216 /* Per T204 TX port */
217 struct sched_port {
218 	unsigned int	avail;		/* available bits - quota */
219 	unsigned int	drain_bits_per_1024ns; /* drain rate */
220 	unsigned int	speed;		/* drain rate, mbps */
221 	unsigned int	mtu;		/* mtu size */
222 	struct sk_buff_head skbq;	/* pending skbs */
223 };
224 
225 /* Per T204 device */
226 struct sched {
227 	ktime_t         last_updated;   /* last time quotas were computed */
228 	unsigned int	max_avail;	/* max bits to be sent to any port */
229 	unsigned int	port;		/* port index (round robin ports) */
230 	unsigned int	num;		/* num skbs in per port queues */
231 	struct sched_port p[MAX_NPORTS];
232 	struct tasklet_struct sched_tsk;/* tasklet used to run scheduler */
233 	struct sge *sge;
234 };
235 
236 static void restart_sched(struct tasklet_struct *t);
237 
238 
239 /*
240  * Main SGE data structure
241  *
242  * Interrupts are handled by a single CPU and it is likely that on a MP system
243  * the application is migrated to another CPU. In that scenario, we try to
244  * separate the RX(in irq context) and TX state in order to decrease memory
245  * contention.
246  */
247 struct sge {
248 	struct adapter *adapter;	/* adapter backpointer */
249 	struct net_device *netdev;      /* netdevice backpointer */
250 	struct freelQ	freelQ[SGE_FREELQ_N]; /* buffer free lists */
251 	struct respQ	respQ;		/* response Q */
252 	unsigned long   stopped_tx_queues; /* bitmap of suspended Tx queues */
253 	unsigned int	rx_pkt_pad;     /* RX padding for L2 packets */
254 	unsigned int	jumbo_fl;       /* jumbo freelist Q index */
255 	unsigned int	intrtimer_nres;	/* no-resource interrupt timer */
256 	unsigned int    fixed_intrtimer;/* non-adaptive interrupt timer */
257 	struct timer_list tx_reclaim_timer; /* reclaims TX buffers */
258 	struct timer_list espibug_timer;
259 	unsigned long	espibug_timeout;
260 	struct sk_buff	*espibug_skb[MAX_NPORTS];
261 	u32		sge_control;	/* shadow value of sge control reg */
262 	struct sge_intr_counts stats;
263 	struct sge_port_stats __percpu *port_stats[MAX_NPORTS];
264 	struct sched	*tx_sched;
265 	struct cmdQ cmdQ[SGE_CMDQ_N] ____cacheline_aligned_in_smp;
266 };
267 
268 static const u8 ch_mac_addr[ETH_ALEN] = {
269 	0x0, 0x7, 0x43, 0x0, 0x0, 0x0
270 };
271 
272 /*
273  * stop tasklet and free all pending skb's
274  */
275 static void tx_sched_stop(struct sge *sge)
276 {
277 	struct sched *s = sge->tx_sched;
278 	int i;
279 
280 	tasklet_kill(&s->sched_tsk);
281 
282 	for (i = 0; i < MAX_NPORTS; i++)
283 		__skb_queue_purge(&s->p[s->port].skbq);
284 }
285 
286 /*
287  * t1_sched_update_parms() is called when the MTU or link speed changes. It
288  * re-computes scheduler parameters to scope with the change.
289  */
290 unsigned int t1_sched_update_parms(struct sge *sge, unsigned int port,
291 				   unsigned int mtu, unsigned int speed)
292 {
293 	struct sched *s = sge->tx_sched;
294 	struct sched_port *p = &s->p[port];
295 	unsigned int max_avail_segs;
296 
297 	pr_debug("%s mtu=%d speed=%d\n", __func__, mtu, speed);
298 	if (speed)
299 		p->speed = speed;
300 	if (mtu)
301 		p->mtu = mtu;
302 
303 	if (speed || mtu) {
304 		unsigned long long drain = 1024ULL * p->speed * (p->mtu - 40);
305 		do_div(drain, (p->mtu + 50) * 1000);
306 		p->drain_bits_per_1024ns = (unsigned int) drain;
307 
308 		if (p->speed < 1000)
309 			p->drain_bits_per_1024ns =
310 				90 * p->drain_bits_per_1024ns / 100;
311 	}
312 
313 	if (board_info(sge->adapter)->board == CHBT_BOARD_CHT204) {
314 		p->drain_bits_per_1024ns -= 16;
315 		s->max_avail = max(4096U, p->mtu + 16 + 14 + 4);
316 		max_avail_segs = max(1U, 4096 / (p->mtu - 40));
317 	} else {
318 		s->max_avail = 16384;
319 		max_avail_segs = max(1U, 9000 / (p->mtu - 40));
320 	}
321 
322 	pr_debug("t1_sched_update_parms: mtu %u speed %u max_avail %u "
323 		 "max_avail_segs %u drain_bits_per_1024ns %u\n", p->mtu,
324 		 p->speed, s->max_avail, max_avail_segs,
325 		 p->drain_bits_per_1024ns);
326 
327 	return max_avail_segs * (p->mtu - 40);
328 }
329 
330 #if 0
331 
332 /*
333  * t1_sched_max_avail_bytes() tells the scheduler the maximum amount of
334  * data that can be pushed per port.
335  */
336 void t1_sched_set_max_avail_bytes(struct sge *sge, unsigned int val)
337 {
338 	struct sched *s = sge->tx_sched;
339 	unsigned int i;
340 
341 	s->max_avail = val;
342 	for (i = 0; i < MAX_NPORTS; i++)
343 		t1_sched_update_parms(sge, i, 0, 0);
344 }
345 
346 /*
347  * t1_sched_set_drain_bits_per_us() tells the scheduler at which rate a port
348  * is draining.
349  */
350 void t1_sched_set_drain_bits_per_us(struct sge *sge, unsigned int port,
351 					 unsigned int val)
352 {
353 	struct sched *s = sge->tx_sched;
354 	struct sched_port *p = &s->p[port];
355 	p->drain_bits_per_1024ns = val * 1024 / 1000;
356 	t1_sched_update_parms(sge, port, 0, 0);
357 }
358 
359 #endif  /*  0  */
360 
361 /*
362  * tx_sched_init() allocates resources and does basic initialization.
363  */
364 static int tx_sched_init(struct sge *sge)
365 {
366 	struct sched *s;
367 	int i;
368 
369 	s = kzalloc(sizeof (struct sched), GFP_KERNEL);
370 	if (!s)
371 		return -ENOMEM;
372 
373 	pr_debug("tx_sched_init\n");
374 	tasklet_setup(&s->sched_tsk, restart_sched);
375 	s->sge = sge;
376 	sge->tx_sched = s;
377 
378 	for (i = 0; i < MAX_NPORTS; i++) {
379 		skb_queue_head_init(&s->p[i].skbq);
380 		t1_sched_update_parms(sge, i, 1500, 1000);
381 	}
382 
383 	return 0;
384 }
385 
386 /*
387  * sched_update_avail() computes the delta since the last time it was called
388  * and updates the per port quota (number of bits that can be sent to the any
389  * port).
390  */
391 static inline int sched_update_avail(struct sge *sge)
392 {
393 	struct sched *s = sge->tx_sched;
394 	ktime_t now = ktime_get();
395 	unsigned int i;
396 	long long delta_time_ns;
397 
398 	delta_time_ns = ktime_to_ns(ktime_sub(now, s->last_updated));
399 
400 	pr_debug("sched_update_avail delta=%lld\n", delta_time_ns);
401 	if (delta_time_ns < 15000)
402 		return 0;
403 
404 	for (i = 0; i < MAX_NPORTS; i++) {
405 		struct sched_port *p = &s->p[i];
406 		unsigned int delta_avail;
407 
408 		delta_avail = (p->drain_bits_per_1024ns * delta_time_ns) >> 13;
409 		p->avail = min(p->avail + delta_avail, s->max_avail);
410 	}
411 
412 	s->last_updated = now;
413 
414 	return 1;
415 }
416 
417 /*
418  * sched_skb() is called from two different places. In the tx path, any
419  * packet generating load on an output port will call sched_skb()
420  * (skb != NULL). In addition, sched_skb() is called from the irq/soft irq
421  * context (skb == NULL).
422  * The scheduler only returns a skb (which will then be sent) if the
423  * length of the skb is <= the current quota of the output port.
424  */
425 static struct sk_buff *sched_skb(struct sge *sge, struct sk_buff *skb,
426 				unsigned int credits)
427 {
428 	struct sched *s = sge->tx_sched;
429 	struct sk_buff_head *skbq;
430 	unsigned int i, len, update = 1;
431 
432 	pr_debug("sched_skb %p\n", skb);
433 	if (!skb) {
434 		if (!s->num)
435 			return NULL;
436 	} else {
437 		skbq = &s->p[skb->dev->if_port].skbq;
438 		__skb_queue_tail(skbq, skb);
439 		s->num++;
440 		skb = NULL;
441 	}
442 
443 	if (credits < MAX_SKB_FRAGS + 1)
444 		goto out;
445 
446 again:
447 	for (i = 0; i < MAX_NPORTS; i++) {
448 		s->port = (s->port + 1) & (MAX_NPORTS - 1);
449 		skbq = &s->p[s->port].skbq;
450 
451 		skb = skb_peek(skbq);
452 
453 		if (!skb)
454 			continue;
455 
456 		len = skb->len;
457 		if (len <= s->p[s->port].avail) {
458 			s->p[s->port].avail -= len;
459 			s->num--;
460 			__skb_unlink(skb, skbq);
461 			goto out;
462 		}
463 		skb = NULL;
464 	}
465 
466 	if (update-- && sched_update_avail(sge))
467 		goto again;
468 
469 out:
470 	/* If there are more pending skbs, we use the hardware to schedule us
471 	 * again.
472 	 */
473 	if (s->num && !skb) {
474 		struct cmdQ *q = &sge->cmdQ[0];
475 		clear_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
476 		if (test_and_set_bit(CMDQ_STAT_RUNNING, &q->status) == 0) {
477 			set_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
478 			writel(F_CMDQ0_ENABLE, sge->adapter->regs + A_SG_DOORBELL);
479 		}
480 	}
481 	pr_debug("sched_skb ret %p\n", skb);
482 
483 	return skb;
484 }
485 
486 /*
487  * PIO to indicate that memory mapped Q contains valid descriptor(s).
488  */
489 static inline void doorbell_pio(struct adapter *adapter, u32 val)
490 {
491 	wmb();
492 	writel(val, adapter->regs + A_SG_DOORBELL);
493 }
494 
495 /*
496  * Frees all RX buffers on the freelist Q. The caller must make sure that
497  * the SGE is turned off before calling this function.
498  */
499 static void free_freelQ_buffers(struct pci_dev *pdev, struct freelQ *q)
500 {
501 	unsigned int cidx = q->cidx;
502 
503 	while (q->credits--) {
504 		struct freelQ_ce *ce = &q->centries[cidx];
505 
506 		dma_unmap_single(&pdev->dev, dma_unmap_addr(ce, dma_addr),
507 				 dma_unmap_len(ce, dma_len), DMA_FROM_DEVICE);
508 		dev_kfree_skb(ce->skb);
509 		ce->skb = NULL;
510 		if (++cidx == q->size)
511 			cidx = 0;
512 	}
513 }
514 
515 /*
516  * Free RX free list and response queue resources.
517  */
518 static void free_rx_resources(struct sge *sge)
519 {
520 	struct pci_dev *pdev = sge->adapter->pdev;
521 	unsigned int size, i;
522 
523 	if (sge->respQ.entries) {
524 		size = sizeof(struct respQ_e) * sge->respQ.size;
525 		dma_free_coherent(&pdev->dev, size, sge->respQ.entries,
526 				  sge->respQ.dma_addr);
527 	}
528 
529 	for (i = 0; i < SGE_FREELQ_N; i++) {
530 		struct freelQ *q = &sge->freelQ[i];
531 
532 		if (q->centries) {
533 			free_freelQ_buffers(pdev, q);
534 			kfree(q->centries);
535 		}
536 		if (q->entries) {
537 			size = sizeof(struct freelQ_e) * q->size;
538 			dma_free_coherent(&pdev->dev, size, q->entries,
539 					  q->dma_addr);
540 		}
541 	}
542 }
543 
544 /*
545  * Allocates basic RX resources, consisting of memory mapped freelist Qs and a
546  * response queue.
547  */
548 static int alloc_rx_resources(struct sge *sge, struct sge_params *p)
549 {
550 	struct pci_dev *pdev = sge->adapter->pdev;
551 	unsigned int size, i;
552 
553 	for (i = 0; i < SGE_FREELQ_N; i++) {
554 		struct freelQ *q = &sge->freelQ[i];
555 
556 		q->genbit = 1;
557 		q->size = p->freelQ_size[i];
558 		q->dma_offset = sge->rx_pkt_pad ? 0 : NET_IP_ALIGN;
559 		size = sizeof(struct freelQ_e) * q->size;
560 		q->entries = dma_alloc_coherent(&pdev->dev, size,
561 						&q->dma_addr, GFP_KERNEL);
562 		if (!q->entries)
563 			goto err_no_mem;
564 
565 		size = sizeof(struct freelQ_ce) * q->size;
566 		q->centries = kzalloc(size, GFP_KERNEL);
567 		if (!q->centries)
568 			goto err_no_mem;
569 	}
570 
571 	/*
572 	 * Calculate the buffer sizes for the two free lists.  FL0 accommodates
573 	 * regular sized Ethernet frames, FL1 is sized not to exceed 16K,
574 	 * including all the sk_buff overhead.
575 	 *
576 	 * Note: For T2 FL0 and FL1 are reversed.
577 	 */
578 	sge->freelQ[!sge->jumbo_fl].rx_buffer_size = SGE_RX_SM_BUF_SIZE +
579 		sizeof(struct cpl_rx_data) +
580 		sge->freelQ[!sge->jumbo_fl].dma_offset;
581 
582 	size = (16 * 1024) - SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
583 
584 	sge->freelQ[sge->jumbo_fl].rx_buffer_size = size;
585 
586 	/*
587 	 * Setup which skb recycle Q should be used when recycling buffers from
588 	 * each free list.
589 	 */
590 	sge->freelQ[!sge->jumbo_fl].recycleq_idx = 0;
591 	sge->freelQ[sge->jumbo_fl].recycleq_idx = 1;
592 
593 	sge->respQ.genbit = 1;
594 	sge->respQ.size = SGE_RESPQ_E_N;
595 	sge->respQ.credits = 0;
596 	size = sizeof(struct respQ_e) * sge->respQ.size;
597 	sge->respQ.entries =
598 		dma_alloc_coherent(&pdev->dev, size, &sge->respQ.dma_addr,
599 				   GFP_KERNEL);
600 	if (!sge->respQ.entries)
601 		goto err_no_mem;
602 	return 0;
603 
604 err_no_mem:
605 	free_rx_resources(sge);
606 	return -ENOMEM;
607 }
608 
609 /*
610  * Reclaims n TX descriptors and frees the buffers associated with them.
611  */
612 static void free_cmdQ_buffers(struct sge *sge, struct cmdQ *q, unsigned int n)
613 {
614 	struct cmdQ_ce *ce;
615 	struct pci_dev *pdev = sge->adapter->pdev;
616 	unsigned int cidx = q->cidx;
617 
618 	q->in_use -= n;
619 	ce = &q->centries[cidx];
620 	while (n--) {
621 		if (likely(dma_unmap_len(ce, dma_len))) {
622 			dma_unmap_single(&pdev->dev,
623 					 dma_unmap_addr(ce, dma_addr),
624 					 dma_unmap_len(ce, dma_len),
625 					 DMA_TO_DEVICE);
626 			if (q->sop)
627 				q->sop = 0;
628 		}
629 		if (ce->skb) {
630 			dev_kfree_skb_any(ce->skb);
631 			q->sop = 1;
632 		}
633 		ce++;
634 		if (++cidx == q->size) {
635 			cidx = 0;
636 			ce = q->centries;
637 		}
638 	}
639 	q->cidx = cidx;
640 }
641 
642 /*
643  * Free TX resources.
644  *
645  * Assumes that SGE is stopped and all interrupts are disabled.
646  */
647 static void free_tx_resources(struct sge *sge)
648 {
649 	struct pci_dev *pdev = sge->adapter->pdev;
650 	unsigned int size, i;
651 
652 	for (i = 0; i < SGE_CMDQ_N; i++) {
653 		struct cmdQ *q = &sge->cmdQ[i];
654 
655 		if (q->centries) {
656 			if (q->in_use)
657 				free_cmdQ_buffers(sge, q, q->in_use);
658 			kfree(q->centries);
659 		}
660 		if (q->entries) {
661 			size = sizeof(struct cmdQ_e) * q->size;
662 			dma_free_coherent(&pdev->dev, size, q->entries,
663 					  q->dma_addr);
664 		}
665 	}
666 }
667 
668 /*
669  * Allocates basic TX resources, consisting of memory mapped command Qs.
670  */
671 static int alloc_tx_resources(struct sge *sge, struct sge_params *p)
672 {
673 	struct pci_dev *pdev = sge->adapter->pdev;
674 	unsigned int size, i;
675 
676 	for (i = 0; i < SGE_CMDQ_N; i++) {
677 		struct cmdQ *q = &sge->cmdQ[i];
678 
679 		q->genbit = 1;
680 		q->sop = 1;
681 		q->size = p->cmdQ_size[i];
682 		q->in_use = 0;
683 		q->status = 0;
684 		q->processed = q->cleaned = 0;
685 		q->stop_thres = 0;
686 		spin_lock_init(&q->lock);
687 		size = sizeof(struct cmdQ_e) * q->size;
688 		q->entries = dma_alloc_coherent(&pdev->dev, size,
689 						&q->dma_addr, GFP_KERNEL);
690 		if (!q->entries)
691 			goto err_no_mem;
692 
693 		size = sizeof(struct cmdQ_ce) * q->size;
694 		q->centries = kzalloc(size, GFP_KERNEL);
695 		if (!q->centries)
696 			goto err_no_mem;
697 	}
698 
699 	/*
700 	 * CommandQ 0 handles Ethernet and TOE packets, while queue 1 is TOE
701 	 * only.  For queue 0 set the stop threshold so we can handle one more
702 	 * packet from each port, plus reserve an additional 24 entries for
703 	 * Ethernet packets only.  Queue 1 never suspends nor do we reserve
704 	 * space for Ethernet packets.
705 	 */
706 	sge->cmdQ[0].stop_thres = sge->adapter->params.nports *
707 		(MAX_SKB_FRAGS + 1);
708 	return 0;
709 
710 err_no_mem:
711 	free_tx_resources(sge);
712 	return -ENOMEM;
713 }
714 
715 static inline void setup_ring_params(struct adapter *adapter, u64 addr,
716 				     u32 size, int base_reg_lo,
717 				     int base_reg_hi, int size_reg)
718 {
719 	writel((u32)addr, adapter->regs + base_reg_lo);
720 	writel(addr >> 32, adapter->regs + base_reg_hi);
721 	writel(size, adapter->regs + size_reg);
722 }
723 
724 /*
725  * Enable/disable VLAN acceleration.
726  */
727 void t1_vlan_mode(struct adapter *adapter, netdev_features_t features)
728 {
729 	struct sge *sge = adapter->sge;
730 
731 	if (features & NETIF_F_HW_VLAN_CTAG_RX)
732 		sge->sge_control |= F_VLAN_XTRACT;
733 	else
734 		sge->sge_control &= ~F_VLAN_XTRACT;
735 	if (adapter->open_device_map) {
736 		writel(sge->sge_control, adapter->regs + A_SG_CONTROL);
737 		readl(adapter->regs + A_SG_CONTROL);   /* flush */
738 	}
739 }
740 
741 /*
742  * Programs the various SGE registers. However, the engine is not yet enabled,
743  * but sge->sge_control is setup and ready to go.
744  */
745 static void configure_sge(struct sge *sge, struct sge_params *p)
746 {
747 	struct adapter *ap = sge->adapter;
748 
749 	writel(0, ap->regs + A_SG_CONTROL);
750 	setup_ring_params(ap, sge->cmdQ[0].dma_addr, sge->cmdQ[0].size,
751 			  A_SG_CMD0BASELWR, A_SG_CMD0BASEUPR, A_SG_CMD0SIZE);
752 	setup_ring_params(ap, sge->cmdQ[1].dma_addr, sge->cmdQ[1].size,
753 			  A_SG_CMD1BASELWR, A_SG_CMD1BASEUPR, A_SG_CMD1SIZE);
754 	setup_ring_params(ap, sge->freelQ[0].dma_addr,
755 			  sge->freelQ[0].size, A_SG_FL0BASELWR,
756 			  A_SG_FL0BASEUPR, A_SG_FL0SIZE);
757 	setup_ring_params(ap, sge->freelQ[1].dma_addr,
758 			  sge->freelQ[1].size, A_SG_FL1BASELWR,
759 			  A_SG_FL1BASEUPR, A_SG_FL1SIZE);
760 
761 	/* The threshold comparison uses <. */
762 	writel(SGE_RX_SM_BUF_SIZE + 1, ap->regs + A_SG_FLTHRESHOLD);
763 
764 	setup_ring_params(ap, sge->respQ.dma_addr, sge->respQ.size,
765 			  A_SG_RSPBASELWR, A_SG_RSPBASEUPR, A_SG_RSPSIZE);
766 	writel((u32)sge->respQ.size - 1, ap->regs + A_SG_RSPQUEUECREDIT);
767 
768 	sge->sge_control = F_CMDQ0_ENABLE | F_CMDQ1_ENABLE | F_FL0_ENABLE |
769 		F_FL1_ENABLE | F_CPL_ENABLE | F_RESPONSE_QUEUE_ENABLE |
770 		V_CMDQ_PRIORITY(2) | F_DISABLE_CMDQ1_GTS | F_ISCSI_COALESCE |
771 		V_RX_PKT_OFFSET(sge->rx_pkt_pad);
772 
773 #if defined(__BIG_ENDIAN_BITFIELD)
774 	sge->sge_control |= F_ENABLE_BIG_ENDIAN;
775 #endif
776 
777 	/* Initialize no-resource timer */
778 	sge->intrtimer_nres = SGE_INTRTIMER_NRES * core_ticks_per_usec(ap);
779 
780 	t1_sge_set_coalesce_params(sge, p);
781 }
782 
783 /*
784  * Return the payload capacity of the jumbo free-list buffers.
785  */
786 static inline unsigned int jumbo_payload_capacity(const struct sge *sge)
787 {
788 	return sge->freelQ[sge->jumbo_fl].rx_buffer_size -
789 		sge->freelQ[sge->jumbo_fl].dma_offset -
790 		sizeof(struct cpl_rx_data);
791 }
792 
793 /*
794  * Frees all SGE related resources and the sge structure itself
795  */
796 void t1_sge_destroy(struct sge *sge)
797 {
798 	int i;
799 
800 	for_each_port(sge->adapter, i)
801 		free_percpu(sge->port_stats[i]);
802 
803 	kfree(sge->tx_sched);
804 	free_tx_resources(sge);
805 	free_rx_resources(sge);
806 	kfree(sge);
807 }
808 
809 /*
810  * Allocates new RX buffers on the freelist Q (and tracks them on the freelist
811  * context Q) until the Q is full or alloc_skb fails.
812  *
813  * It is possible that the generation bits already match, indicating that the
814  * buffer is already valid and nothing needs to be done. This happens when we
815  * copied a received buffer into a new sk_buff during the interrupt processing.
816  *
817  * If the SGE doesn't automatically align packets properly (!sge->rx_pkt_pad),
818  * we specify a RX_OFFSET in order to make sure that the IP header is 4B
819  * aligned.
820  */
821 static void refill_free_list(struct sge *sge, struct freelQ *q)
822 {
823 	struct pci_dev *pdev = sge->adapter->pdev;
824 	struct freelQ_ce *ce = &q->centries[q->pidx];
825 	struct freelQ_e *e = &q->entries[q->pidx];
826 	unsigned int dma_len = q->rx_buffer_size - q->dma_offset;
827 
828 	while (q->credits < q->size) {
829 		struct sk_buff *skb;
830 		dma_addr_t mapping;
831 
832 		skb = dev_alloc_skb(q->rx_buffer_size);
833 		if (!skb)
834 			break;
835 
836 		skb_reserve(skb, q->dma_offset);
837 		mapping = dma_map_single(&pdev->dev, skb->data, dma_len,
838 					 DMA_FROM_DEVICE);
839 		skb_reserve(skb, sge->rx_pkt_pad);
840 
841 		ce->skb = skb;
842 		dma_unmap_addr_set(ce, dma_addr, mapping);
843 		dma_unmap_len_set(ce, dma_len, dma_len);
844 		e->addr_lo = (u32)mapping;
845 		e->addr_hi = (u64)mapping >> 32;
846 		e->len_gen = V_CMD_LEN(dma_len) | V_CMD_GEN1(q->genbit);
847 		wmb();
848 		e->gen2 = V_CMD_GEN2(q->genbit);
849 
850 		e++;
851 		ce++;
852 		if (++q->pidx == q->size) {
853 			q->pidx = 0;
854 			q->genbit ^= 1;
855 			ce = q->centries;
856 			e = q->entries;
857 		}
858 		q->credits++;
859 	}
860 }
861 
862 /*
863  * Calls refill_free_list for both free lists. If we cannot fill at least 1/4
864  * of both rings, we go into 'few interrupt mode' in order to give the system
865  * time to free up resources.
866  */
867 static void freelQs_empty(struct sge *sge)
868 {
869 	struct adapter *adapter = sge->adapter;
870 	u32 irq_reg = readl(adapter->regs + A_SG_INT_ENABLE);
871 	u32 irqholdoff_reg;
872 
873 	refill_free_list(sge, &sge->freelQ[0]);
874 	refill_free_list(sge, &sge->freelQ[1]);
875 
876 	if (sge->freelQ[0].credits > (sge->freelQ[0].size >> 2) &&
877 	    sge->freelQ[1].credits > (sge->freelQ[1].size >> 2)) {
878 		irq_reg |= F_FL_EXHAUSTED;
879 		irqholdoff_reg = sge->fixed_intrtimer;
880 	} else {
881 		/* Clear the F_FL_EXHAUSTED interrupts for now */
882 		irq_reg &= ~F_FL_EXHAUSTED;
883 		irqholdoff_reg = sge->intrtimer_nres;
884 	}
885 	writel(irqholdoff_reg, adapter->regs + A_SG_INTRTIMER);
886 	writel(irq_reg, adapter->regs + A_SG_INT_ENABLE);
887 
888 	/* We reenable the Qs to force a freelist GTS interrupt later */
889 	doorbell_pio(adapter, F_FL0_ENABLE | F_FL1_ENABLE);
890 }
891 
892 #define SGE_PL_INTR_MASK (F_PL_INTR_SGE_ERR | F_PL_INTR_SGE_DATA)
893 #define SGE_INT_FATAL (F_RESPQ_OVERFLOW | F_PACKET_TOO_BIG | F_PACKET_MISMATCH)
894 #define SGE_INT_ENABLE (F_RESPQ_EXHAUSTED | F_RESPQ_OVERFLOW | \
895 			F_FL_EXHAUSTED | F_PACKET_TOO_BIG | F_PACKET_MISMATCH)
896 
897 /*
898  * Disable SGE Interrupts
899  */
900 void t1_sge_intr_disable(struct sge *sge)
901 {
902 	u32 val = readl(sge->adapter->regs + A_PL_ENABLE);
903 
904 	writel(val & ~SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_ENABLE);
905 	writel(0, sge->adapter->regs + A_SG_INT_ENABLE);
906 }
907 
908 /*
909  * Enable SGE interrupts.
910  */
911 void t1_sge_intr_enable(struct sge *sge)
912 {
913 	u32 en = SGE_INT_ENABLE;
914 	u32 val = readl(sge->adapter->regs + A_PL_ENABLE);
915 
916 	if (sge->adapter->port[0].dev->hw_features & NETIF_F_TSO)
917 		en &= ~F_PACKET_TOO_BIG;
918 	writel(en, sge->adapter->regs + A_SG_INT_ENABLE);
919 	writel(val | SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_ENABLE);
920 }
921 
922 /*
923  * Clear SGE interrupts.
924  */
925 void t1_sge_intr_clear(struct sge *sge)
926 {
927 	writel(SGE_PL_INTR_MASK, sge->adapter->regs + A_PL_CAUSE);
928 	writel(0xffffffff, sge->adapter->regs + A_SG_INT_CAUSE);
929 }
930 
931 /*
932  * SGE 'Error' interrupt handler
933  */
934 bool t1_sge_intr_error_handler(struct sge *sge)
935 {
936 	struct adapter *adapter = sge->adapter;
937 	u32 cause = readl(adapter->regs + A_SG_INT_CAUSE);
938 	bool wake = false;
939 
940 	if (adapter->port[0].dev->hw_features & NETIF_F_TSO)
941 		cause &= ~F_PACKET_TOO_BIG;
942 	if (cause & F_RESPQ_EXHAUSTED)
943 		sge->stats.respQ_empty++;
944 	if (cause & F_RESPQ_OVERFLOW) {
945 		sge->stats.respQ_overflow++;
946 		pr_alert("%s: SGE response queue overflow\n",
947 			 adapter->name);
948 	}
949 	if (cause & F_FL_EXHAUSTED) {
950 		sge->stats.freelistQ_empty++;
951 		freelQs_empty(sge);
952 	}
953 	if (cause & F_PACKET_TOO_BIG) {
954 		sge->stats.pkt_too_big++;
955 		pr_alert("%s: SGE max packet size exceeded\n",
956 			 adapter->name);
957 	}
958 	if (cause & F_PACKET_MISMATCH) {
959 		sge->stats.pkt_mismatch++;
960 		pr_alert("%s: SGE packet mismatch\n", adapter->name);
961 	}
962 	if (cause & SGE_INT_FATAL) {
963 		t1_interrupts_disable(adapter);
964 		adapter->pending_thread_intr |= F_PL_INTR_SGE_ERR;
965 		wake = true;
966 	}
967 
968 	writel(cause, adapter->regs + A_SG_INT_CAUSE);
969 	return wake;
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  *	@adapter: the adapter 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 adapter *adapter,
1041 					 struct freelQ *fl, unsigned int len)
1042 {
1043 	const struct freelQ_ce *ce = &fl->centries[fl->cidx];
1044 	struct pci_dev *pdev = adapter->pdev;
1045 	struct sk_buff *skb;
1046 
1047 	if (len < copybreak) {
1048 		skb = napi_alloc_skb(&adapter->napi, len);
1049 		if (!skb)
1050 			goto use_orig_buf;
1051 
1052 		skb_put(skb, len);
1053 		dma_sync_single_for_cpu(&pdev->dev,
1054 					dma_unmap_addr(ce, dma_addr),
1055 					dma_unmap_len(ce, dma_len),
1056 					DMA_FROM_DEVICE);
1057 		skb_copy_from_linear_data(ce->skb, skb->data, len);
1058 		dma_sync_single_for_device(&pdev->dev,
1059 					   dma_unmap_addr(ce, dma_addr),
1060 					   dma_unmap_len(ce, dma_len),
1061 					   DMA_FROM_DEVICE);
1062 		recycle_fl_buf(fl, fl->cidx);
1063 		return skb;
1064 	}
1065 
1066 use_orig_buf:
1067 	if (fl->credits < 2) {
1068 		recycle_fl_buf(fl, fl->cidx);
1069 		return NULL;
1070 	}
1071 
1072 	dma_unmap_single(&pdev->dev, dma_unmap_addr(ce, dma_addr),
1073 			 dma_unmap_len(ce, dma_len), DMA_FROM_DEVICE);
1074 	skb = ce->skb;
1075 	prefetch(skb->data);
1076 
1077 	skb_put(skb, len);
1078 	return skb;
1079 }
1080 
1081 /**
1082  *	unexpected_offload - handle an unexpected offload packet
1083  *	@adapter: the adapter
1084  *	@fl: the free list that received the packet
1085  *
1086  *	Called when we receive an unexpected offload packet (e.g., the TOE
1087  *	function is disabled or the card is a NIC).  Prints a message and
1088  *	recycles the buffer.
1089  */
1090 static void unexpected_offload(struct adapter *adapter, struct freelQ *fl)
1091 {
1092 	struct freelQ_ce *ce = &fl->centries[fl->cidx];
1093 	struct sk_buff *skb = ce->skb;
1094 
1095 	dma_sync_single_for_cpu(&adapter->pdev->dev,
1096 				dma_unmap_addr(ce, dma_addr),
1097 				dma_unmap_len(ce, dma_len), DMA_FROM_DEVICE);
1098 	pr_err("%s: unexpected offload packet, cmd %u\n",
1099 	       adapter->name, *skb->data);
1100 	recycle_fl_buf(fl, fl->cidx);
1101 }
1102 
1103 /*
1104  * T1/T2 SGE limits the maximum DMA size per TX descriptor to
1105  * SGE_TX_DESC_MAX_PLEN (16KB). If the PAGE_SIZE is larger than 16KB, the
1106  * stack might send more than SGE_TX_DESC_MAX_PLEN in a contiguous manner.
1107  * Note that the *_large_page_tx_descs stuff will be optimized out when
1108  * PAGE_SIZE <= SGE_TX_DESC_MAX_PLEN.
1109  *
1110  * compute_large_page_descs() computes how many additional descriptors are
1111  * required to break down the stack's request.
1112  */
1113 static inline unsigned int compute_large_page_tx_descs(struct sk_buff *skb)
1114 {
1115 	unsigned int count = 0;
1116 
1117 	if (PAGE_SIZE > SGE_TX_DESC_MAX_PLEN) {
1118 		unsigned int nfrags = skb_shinfo(skb)->nr_frags;
1119 		unsigned int i, len = skb_headlen(skb);
1120 		while (len > SGE_TX_DESC_MAX_PLEN) {
1121 			count++;
1122 			len -= SGE_TX_DESC_MAX_PLEN;
1123 		}
1124 		for (i = 0; nfrags--; i++) {
1125 			const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
1126 			len = skb_frag_size(frag);
1127 			while (len > SGE_TX_DESC_MAX_PLEN) {
1128 				count++;
1129 				len -= SGE_TX_DESC_MAX_PLEN;
1130 			}
1131 		}
1132 	}
1133 	return count;
1134 }
1135 
1136 /*
1137  * Write a cmdQ entry.
1138  *
1139  * Since this function writes the 'flags' field, it must not be used to
1140  * write the first cmdQ entry.
1141  */
1142 static inline void write_tx_desc(struct cmdQ_e *e, dma_addr_t mapping,
1143 				 unsigned int len, unsigned int gen,
1144 				 unsigned int eop)
1145 {
1146 	BUG_ON(len > SGE_TX_DESC_MAX_PLEN);
1147 
1148 	e->addr_lo = (u32)mapping;
1149 	e->addr_hi = (u64)mapping >> 32;
1150 	e->len_gen = V_CMD_LEN(len) | V_CMD_GEN1(gen);
1151 	e->flags = F_CMD_DATAVALID | V_CMD_EOP(eop) | V_CMD_GEN2(gen);
1152 }
1153 
1154 /*
1155  * See comment for previous function.
1156  *
1157  * write_tx_descs_large_page() writes additional SGE tx descriptors if
1158  * *desc_len exceeds HW's capability.
1159  */
1160 static inline unsigned int write_large_page_tx_descs(unsigned int pidx,
1161 						     struct cmdQ_e **e,
1162 						     struct cmdQ_ce **ce,
1163 						     unsigned int *gen,
1164 						     dma_addr_t *desc_mapping,
1165 						     unsigned int *desc_len,
1166 						     unsigned int nfrags,
1167 						     struct cmdQ *q)
1168 {
1169 	if (PAGE_SIZE > SGE_TX_DESC_MAX_PLEN) {
1170 		struct cmdQ_e *e1 = *e;
1171 		struct cmdQ_ce *ce1 = *ce;
1172 
1173 		while (*desc_len > SGE_TX_DESC_MAX_PLEN) {
1174 			*desc_len -= SGE_TX_DESC_MAX_PLEN;
1175 			write_tx_desc(e1, *desc_mapping, SGE_TX_DESC_MAX_PLEN,
1176 				      *gen, nfrags == 0 && *desc_len == 0);
1177 			ce1->skb = NULL;
1178 			dma_unmap_len_set(ce1, dma_len, 0);
1179 			*desc_mapping += SGE_TX_DESC_MAX_PLEN;
1180 			if (*desc_len) {
1181 				ce1++;
1182 				e1++;
1183 				if (++pidx == q->size) {
1184 					pidx = 0;
1185 					*gen ^= 1;
1186 					ce1 = q->centries;
1187 					e1 = q->entries;
1188 				}
1189 			}
1190 		}
1191 		*e = e1;
1192 		*ce = ce1;
1193 	}
1194 	return pidx;
1195 }
1196 
1197 /*
1198  * Write the command descriptors to transmit the given skb starting at
1199  * descriptor pidx with the given generation.
1200  */
1201 static inline void write_tx_descs(struct adapter *adapter, struct sk_buff *skb,
1202 				  unsigned int pidx, unsigned int gen,
1203 				  struct cmdQ *q)
1204 {
1205 	dma_addr_t mapping, desc_mapping;
1206 	struct cmdQ_e *e, *e1;
1207 	struct cmdQ_ce *ce;
1208 	unsigned int i, flags, first_desc_len, desc_len,
1209 	    nfrags = skb_shinfo(skb)->nr_frags;
1210 
1211 	e = e1 = &q->entries[pidx];
1212 	ce = &q->centries[pidx];
1213 
1214 	mapping = dma_map_single(&adapter->pdev->dev, skb->data,
1215 				 skb_headlen(skb), DMA_TO_DEVICE);
1216 
1217 	desc_mapping = mapping;
1218 	desc_len = skb_headlen(skb);
1219 
1220 	flags = F_CMD_DATAVALID | F_CMD_SOP |
1221 	    V_CMD_EOP(nfrags == 0 && desc_len <= SGE_TX_DESC_MAX_PLEN) |
1222 	    V_CMD_GEN2(gen);
1223 	first_desc_len = (desc_len <= SGE_TX_DESC_MAX_PLEN) ?
1224 	    desc_len : SGE_TX_DESC_MAX_PLEN;
1225 	e->addr_lo = (u32)desc_mapping;
1226 	e->addr_hi = (u64)desc_mapping >> 32;
1227 	e->len_gen = V_CMD_LEN(first_desc_len) | V_CMD_GEN1(gen);
1228 	ce->skb = NULL;
1229 	dma_unmap_len_set(ce, dma_len, 0);
1230 
1231 	if (PAGE_SIZE > SGE_TX_DESC_MAX_PLEN &&
1232 	    desc_len > SGE_TX_DESC_MAX_PLEN) {
1233 		desc_mapping += first_desc_len;
1234 		desc_len -= first_desc_len;
1235 		e1++;
1236 		ce++;
1237 		if (++pidx == q->size) {
1238 			pidx = 0;
1239 			gen ^= 1;
1240 			e1 = q->entries;
1241 			ce = q->centries;
1242 		}
1243 		pidx = write_large_page_tx_descs(pidx, &e1, &ce, &gen,
1244 						 &desc_mapping, &desc_len,
1245 						 nfrags, q);
1246 
1247 		if (likely(desc_len))
1248 			write_tx_desc(e1, desc_mapping, desc_len, gen,
1249 				      nfrags == 0);
1250 	}
1251 
1252 	ce->skb = NULL;
1253 	dma_unmap_addr_set(ce, dma_addr, mapping);
1254 	dma_unmap_len_set(ce, dma_len, skb_headlen(skb));
1255 
1256 	for (i = 0; nfrags--; i++) {
1257 		skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
1258 		e1++;
1259 		ce++;
1260 		if (++pidx == q->size) {
1261 			pidx = 0;
1262 			gen ^= 1;
1263 			e1 = q->entries;
1264 			ce = q->centries;
1265 		}
1266 
1267 		mapping = skb_frag_dma_map(&adapter->pdev->dev, frag, 0,
1268 					   skb_frag_size(frag), DMA_TO_DEVICE);
1269 		desc_mapping = mapping;
1270 		desc_len = skb_frag_size(frag);
1271 
1272 		pidx = write_large_page_tx_descs(pidx, &e1, &ce, &gen,
1273 						 &desc_mapping, &desc_len,
1274 						 nfrags, q);
1275 		if (likely(desc_len))
1276 			write_tx_desc(e1, desc_mapping, desc_len, gen,
1277 				      nfrags == 0);
1278 		ce->skb = NULL;
1279 		dma_unmap_addr_set(ce, dma_addr, mapping);
1280 		dma_unmap_len_set(ce, dma_len, skb_frag_size(frag));
1281 	}
1282 	ce->skb = skb;
1283 	wmb();
1284 	e->flags = flags;
1285 }
1286 
1287 /*
1288  * Clean up completed Tx buffers.
1289  */
1290 static inline void reclaim_completed_tx(struct sge *sge, struct cmdQ *q)
1291 {
1292 	unsigned int reclaim = q->processed - q->cleaned;
1293 
1294 	if (reclaim) {
1295 		pr_debug("reclaim_completed_tx processed:%d cleaned:%d\n",
1296 			 q->processed, q->cleaned);
1297 		free_cmdQ_buffers(sge, q, reclaim);
1298 		q->cleaned += reclaim;
1299 	}
1300 }
1301 
1302 /*
1303  * Called from tasklet. Checks the scheduler for any
1304  * pending skbs that can be sent.
1305  */
1306 static void restart_sched(struct tasklet_struct *t)
1307 {
1308 	struct sched *s = from_tasklet(s, t, sched_tsk);
1309 	struct sge *sge = s->sge;
1310 	struct adapter *adapter = sge->adapter;
1311 	struct cmdQ *q = &sge->cmdQ[0];
1312 	struct sk_buff *skb;
1313 	unsigned int credits, queued_skb = 0;
1314 
1315 	spin_lock(&q->lock);
1316 	reclaim_completed_tx(sge, q);
1317 
1318 	credits = q->size - q->in_use;
1319 	pr_debug("restart_sched credits=%d\n", credits);
1320 	while ((skb = sched_skb(sge, NULL, credits)) != NULL) {
1321 		unsigned int genbit, pidx, count;
1322 	        count = 1 + skb_shinfo(skb)->nr_frags;
1323 		count += compute_large_page_tx_descs(skb);
1324 		q->in_use += count;
1325 		genbit = q->genbit;
1326 		pidx = q->pidx;
1327 		q->pidx += count;
1328 		if (q->pidx >= q->size) {
1329 			q->pidx -= q->size;
1330 			q->genbit ^= 1;
1331 		}
1332 		write_tx_descs(adapter, skb, pidx, genbit, q);
1333 	        credits = q->size - q->in_use;
1334 		queued_skb = 1;
1335 	}
1336 
1337 	if (queued_skb) {
1338 		clear_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1339 		if (test_and_set_bit(CMDQ_STAT_RUNNING, &q->status) == 0) {
1340 			set_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1341 			writel(F_CMDQ0_ENABLE, adapter->regs + A_SG_DOORBELL);
1342 		}
1343 	}
1344 	spin_unlock(&q->lock);
1345 }
1346 
1347 /**
1348  *	sge_rx - process an ingress ethernet packet
1349  *	@sge: the sge structure
1350  *	@fl: the free list that contains the packet buffer
1351  *	@len: the packet length
1352  *
1353  *	Process an ingress ethernet packet and deliver it to the stack.
1354  */
1355 static void sge_rx(struct sge *sge, struct freelQ *fl, unsigned int len)
1356 {
1357 	struct sk_buff *skb;
1358 	const struct cpl_rx_pkt *p;
1359 	struct adapter *adapter = sge->adapter;
1360 	struct sge_port_stats *st;
1361 	struct net_device *dev;
1362 
1363 	skb = get_packet(adapter, fl, len - sge->rx_pkt_pad);
1364 	if (unlikely(!skb)) {
1365 		sge->stats.rx_drops++;
1366 		return;
1367 	}
1368 
1369 	p = (const struct cpl_rx_pkt *) skb->data;
1370 	if (p->iff >= adapter->params.nports) {
1371 		kfree_skb(skb);
1372 		return;
1373 	}
1374 	__skb_pull(skb, sizeof(*p));
1375 
1376 	st = this_cpu_ptr(sge->port_stats[p->iff]);
1377 	dev = adapter->port[p->iff].dev;
1378 
1379 	skb->protocol = eth_type_trans(skb, dev);
1380 	if ((dev->features & NETIF_F_RXCSUM) && p->csum == 0xffff &&
1381 	    skb->protocol == htons(ETH_P_IP) &&
1382 	    (skb->data[9] == IPPROTO_TCP || skb->data[9] == IPPROTO_UDP)) {
1383 		++st->rx_cso_good;
1384 		skb->ip_summed = CHECKSUM_UNNECESSARY;
1385 	} else
1386 		skb_checksum_none_assert(skb);
1387 
1388 	if (p->vlan_valid) {
1389 		st->vlan_xtract++;
1390 		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), ntohs(p->vlan));
1391 	}
1392 	netif_receive_skb(skb);
1393 }
1394 
1395 /*
1396  * Returns true if a command queue has enough available descriptors that
1397  * we can resume Tx operation after temporarily disabling its packet queue.
1398  */
1399 static inline int enough_free_Tx_descs(const struct cmdQ *q)
1400 {
1401 	unsigned int r = q->processed - q->cleaned;
1402 
1403 	return q->in_use - r < (q->size >> 1);
1404 }
1405 
1406 /*
1407  * Called when sufficient space has become available in the SGE command queues
1408  * after the Tx packet schedulers have been suspended to restart the Tx path.
1409  */
1410 static void restart_tx_queues(struct sge *sge)
1411 {
1412 	struct adapter *adap = sge->adapter;
1413 	int i;
1414 
1415 	if (!enough_free_Tx_descs(&sge->cmdQ[0]))
1416 		return;
1417 
1418 	for_each_port(adap, i) {
1419 		struct net_device *nd = adap->port[i].dev;
1420 
1421 		if (test_and_clear_bit(nd->if_port, &sge->stopped_tx_queues) &&
1422 		    netif_running(nd)) {
1423 			sge->stats.cmdQ_restarted[2]++;
1424 			netif_wake_queue(nd);
1425 		}
1426 	}
1427 }
1428 
1429 /*
1430  * update_tx_info is called from the interrupt handler/NAPI to return cmdQ0
1431  * information.
1432  */
1433 static unsigned int update_tx_info(struct adapter *adapter,
1434 					  unsigned int flags,
1435 					  unsigned int pr0)
1436 {
1437 	struct sge *sge = adapter->sge;
1438 	struct cmdQ *cmdq = &sge->cmdQ[0];
1439 
1440 	cmdq->processed += pr0;
1441 	if (flags & (F_FL0_ENABLE | F_FL1_ENABLE)) {
1442 		freelQs_empty(sge);
1443 		flags &= ~(F_FL0_ENABLE | F_FL1_ENABLE);
1444 	}
1445 	if (flags & F_CMDQ0_ENABLE) {
1446 		clear_bit(CMDQ_STAT_RUNNING, &cmdq->status);
1447 
1448 		if (cmdq->cleaned + cmdq->in_use != cmdq->processed &&
1449 		    !test_and_set_bit(CMDQ_STAT_LAST_PKT_DB, &cmdq->status)) {
1450 			set_bit(CMDQ_STAT_RUNNING, &cmdq->status);
1451 			writel(F_CMDQ0_ENABLE, adapter->regs + A_SG_DOORBELL);
1452 		}
1453 		if (sge->tx_sched)
1454 			tasklet_hi_schedule(&sge->tx_sched->sched_tsk);
1455 
1456 		flags &= ~F_CMDQ0_ENABLE;
1457 	}
1458 
1459 	if (unlikely(sge->stopped_tx_queues != 0))
1460 		restart_tx_queues(sge);
1461 
1462 	return flags;
1463 }
1464 
1465 /*
1466  * Process SGE responses, up to the supplied budget.  Returns the number of
1467  * responses processed.  A negative budget is effectively unlimited.
1468  */
1469 static int process_responses(struct adapter *adapter, int budget)
1470 {
1471 	struct sge *sge = adapter->sge;
1472 	struct respQ *q = &sge->respQ;
1473 	struct respQ_e *e = &q->entries[q->cidx];
1474 	int done = 0;
1475 	unsigned int flags = 0;
1476 	unsigned int cmdq_processed[SGE_CMDQ_N] = {0, 0};
1477 
1478 	while (done < budget && e->GenerationBit == q->genbit) {
1479 		flags |= e->Qsleeping;
1480 
1481 		cmdq_processed[0] += e->Cmdq0CreditReturn;
1482 		cmdq_processed[1] += e->Cmdq1CreditReturn;
1483 
1484 		/* We batch updates to the TX side to avoid cacheline
1485 		 * ping-pong of TX state information on MP where the sender
1486 		 * might run on a different CPU than this function...
1487 		 */
1488 		if (unlikely((flags & F_CMDQ0_ENABLE) || cmdq_processed[0] > 64)) {
1489 			flags = update_tx_info(adapter, flags, cmdq_processed[0]);
1490 			cmdq_processed[0] = 0;
1491 		}
1492 
1493 		if (unlikely(cmdq_processed[1] > 16)) {
1494 			sge->cmdQ[1].processed += cmdq_processed[1];
1495 			cmdq_processed[1] = 0;
1496 		}
1497 
1498 		if (likely(e->DataValid)) {
1499 			struct freelQ *fl = &sge->freelQ[e->FreelistQid];
1500 
1501 			BUG_ON(!e->Sop || !e->Eop);
1502 			if (unlikely(e->Offload))
1503 				unexpected_offload(adapter, fl);
1504 			else
1505 				sge_rx(sge, fl, e->BufferLength);
1506 
1507 			++done;
1508 
1509 			/*
1510 			 * Note: this depends on each packet consuming a
1511 			 * single free-list buffer; cf. the BUG above.
1512 			 */
1513 			if (++fl->cidx == fl->size)
1514 				fl->cidx = 0;
1515 			prefetch(fl->centries[fl->cidx].skb);
1516 
1517 			if (unlikely(--fl->credits <
1518 				     fl->size - SGE_FREEL_REFILL_THRESH))
1519 				refill_free_list(sge, fl);
1520 		} else
1521 			sge->stats.pure_rsps++;
1522 
1523 		e++;
1524 		if (unlikely(++q->cidx == q->size)) {
1525 			q->cidx = 0;
1526 			q->genbit ^= 1;
1527 			e = q->entries;
1528 		}
1529 		prefetch(e);
1530 
1531 		if (++q->credits > SGE_RESPQ_REPLENISH_THRES) {
1532 			writel(q->credits, adapter->regs + A_SG_RSPQUEUECREDIT);
1533 			q->credits = 0;
1534 		}
1535 	}
1536 
1537 	flags = update_tx_info(adapter, flags, cmdq_processed[0]);
1538 	sge->cmdQ[1].processed += cmdq_processed[1];
1539 
1540 	return done;
1541 }
1542 
1543 static inline int responses_pending(const struct adapter *adapter)
1544 {
1545 	const struct respQ *Q = &adapter->sge->respQ;
1546 	const struct respQ_e *e = &Q->entries[Q->cidx];
1547 
1548 	return e->GenerationBit == Q->genbit;
1549 }
1550 
1551 /*
1552  * A simpler version of process_responses() that handles only pure (i.e.,
1553  * non data-carrying) responses.  Such respones are too light-weight to justify
1554  * calling a softirq when using NAPI, so we handle them specially in hard
1555  * interrupt context.  The function is called with a pointer to a response,
1556  * which the caller must ensure is a valid pure response.  Returns 1 if it
1557  * encounters a valid data-carrying response, 0 otherwise.
1558  */
1559 static int process_pure_responses(struct adapter *adapter)
1560 {
1561 	struct sge *sge = adapter->sge;
1562 	struct respQ *q = &sge->respQ;
1563 	struct respQ_e *e = &q->entries[q->cidx];
1564 	const struct freelQ *fl = &sge->freelQ[e->FreelistQid];
1565 	unsigned int flags = 0;
1566 	unsigned int cmdq_processed[SGE_CMDQ_N] = {0, 0};
1567 
1568 	prefetch(fl->centries[fl->cidx].skb);
1569 	if (e->DataValid)
1570 		return 1;
1571 
1572 	do {
1573 		flags |= e->Qsleeping;
1574 
1575 		cmdq_processed[0] += e->Cmdq0CreditReturn;
1576 		cmdq_processed[1] += e->Cmdq1CreditReturn;
1577 
1578 		e++;
1579 		if (unlikely(++q->cidx == q->size)) {
1580 			q->cidx = 0;
1581 			q->genbit ^= 1;
1582 			e = q->entries;
1583 		}
1584 		prefetch(e);
1585 
1586 		if (++q->credits > SGE_RESPQ_REPLENISH_THRES) {
1587 			writel(q->credits, adapter->regs + A_SG_RSPQUEUECREDIT);
1588 			q->credits = 0;
1589 		}
1590 		sge->stats.pure_rsps++;
1591 	} while (e->GenerationBit == q->genbit && !e->DataValid);
1592 
1593 	flags = update_tx_info(adapter, flags, cmdq_processed[0]);
1594 	sge->cmdQ[1].processed += cmdq_processed[1];
1595 
1596 	return e->GenerationBit == q->genbit;
1597 }
1598 
1599 /*
1600  * Handler for new data events when using NAPI.  This does not need any locking
1601  * or protection from interrupts as data interrupts are off at this point and
1602  * other adapter interrupts do not interfere.
1603  */
1604 int t1_poll(struct napi_struct *napi, int budget)
1605 {
1606 	struct adapter *adapter = container_of(napi, struct adapter, napi);
1607 	int work_done = process_responses(adapter, budget);
1608 
1609 	if (likely(work_done < budget)) {
1610 		napi_complete_done(napi, work_done);
1611 		writel(adapter->sge->respQ.cidx,
1612 		       adapter->regs + A_SG_SLEEPING);
1613 	}
1614 	return work_done;
1615 }
1616 
1617 irqreturn_t t1_interrupt_thread(int irq, void *data)
1618 {
1619 	struct adapter *adapter = data;
1620 	u32 pending_thread_intr;
1621 
1622 	spin_lock_irq(&adapter->async_lock);
1623 	pending_thread_intr = adapter->pending_thread_intr;
1624 	adapter->pending_thread_intr = 0;
1625 	spin_unlock_irq(&adapter->async_lock);
1626 
1627 	if (!pending_thread_intr)
1628 		return IRQ_NONE;
1629 
1630 	if (pending_thread_intr & F_PL_INTR_EXT)
1631 		t1_elmer0_ext_intr_handler(adapter);
1632 
1633 	/* This error is fatal, interrupts remain off */
1634 	if (pending_thread_intr & F_PL_INTR_SGE_ERR) {
1635 		pr_alert("%s: encountered fatal error, operation suspended\n",
1636 			 adapter->name);
1637 		t1_sge_stop(adapter->sge);
1638 		return IRQ_HANDLED;
1639 	}
1640 
1641 	spin_lock_irq(&adapter->async_lock);
1642 	adapter->slow_intr_mask |= F_PL_INTR_EXT;
1643 
1644 	writel(F_PL_INTR_EXT, adapter->regs + A_PL_CAUSE);
1645 	writel(adapter->slow_intr_mask | F_PL_INTR_SGE_DATA,
1646 	       adapter->regs + A_PL_ENABLE);
1647 	spin_unlock_irq(&adapter->async_lock);
1648 
1649 	return IRQ_HANDLED;
1650 }
1651 
1652 irqreturn_t t1_interrupt(int irq, void *data)
1653 {
1654 	struct adapter *adapter = data;
1655 	struct sge *sge = adapter->sge;
1656 	irqreturn_t handled;
1657 
1658 	if (likely(responses_pending(adapter))) {
1659 		writel(F_PL_INTR_SGE_DATA, adapter->regs + A_PL_CAUSE);
1660 
1661 		if (napi_schedule_prep(&adapter->napi)) {
1662 			if (process_pure_responses(adapter))
1663 				__napi_schedule(&adapter->napi);
1664 			else {
1665 				/* no data, no NAPI needed */
1666 				writel(sge->respQ.cidx, adapter->regs + A_SG_SLEEPING);
1667 				/* undo schedule_prep */
1668 				napi_enable(&adapter->napi);
1669 			}
1670 		}
1671 		return IRQ_HANDLED;
1672 	}
1673 
1674 	spin_lock(&adapter->async_lock);
1675 	handled = t1_slow_intr_handler(adapter);
1676 	spin_unlock(&adapter->async_lock);
1677 
1678 	if (handled == IRQ_NONE)
1679 		sge->stats.unhandled_irqs++;
1680 
1681 	return handled;
1682 }
1683 
1684 /*
1685  * Enqueues the sk_buff onto the cmdQ[qid] and has hardware fetch it.
1686  *
1687  * The code figures out how many entries the sk_buff will require in the
1688  * cmdQ and updates the cmdQ data structure with the state once the enqueue
1689  * has complete. Then, it doesn't access the global structure anymore, but
1690  * uses the corresponding fields on the stack. In conjunction with a spinlock
1691  * around that code, we can make the function reentrant without holding the
1692  * lock when we actually enqueue (which might be expensive, especially on
1693  * architectures with IO MMUs).
1694  *
1695  * This runs with softirqs disabled.
1696  */
1697 static int t1_sge_tx(struct sk_buff *skb, struct adapter *adapter,
1698 		     unsigned int qid, struct net_device *dev)
1699 {
1700 	struct sge *sge = adapter->sge;
1701 	struct cmdQ *q = &sge->cmdQ[qid];
1702 	unsigned int credits, pidx, genbit, count, use_sched_skb = 0;
1703 
1704 	spin_lock(&q->lock);
1705 
1706 	reclaim_completed_tx(sge, q);
1707 
1708 	pidx = q->pidx;
1709 	credits = q->size - q->in_use;
1710 	count = 1 + skb_shinfo(skb)->nr_frags;
1711 	count += compute_large_page_tx_descs(skb);
1712 
1713 	/* Ethernet packet */
1714 	if (unlikely(credits < count)) {
1715 		if (!netif_queue_stopped(dev)) {
1716 			netif_stop_queue(dev);
1717 			set_bit(dev->if_port, &sge->stopped_tx_queues);
1718 			sge->stats.cmdQ_full[2]++;
1719 			pr_err("%s: Tx ring full while queue awake!\n",
1720 			       adapter->name);
1721 		}
1722 		spin_unlock(&q->lock);
1723 		return NETDEV_TX_BUSY;
1724 	}
1725 
1726 	if (unlikely(credits - count < q->stop_thres)) {
1727 		netif_stop_queue(dev);
1728 		set_bit(dev->if_port, &sge->stopped_tx_queues);
1729 		sge->stats.cmdQ_full[2]++;
1730 	}
1731 
1732 	/* T204 cmdQ0 skbs that are destined for a certain port have to go
1733 	 * through the scheduler.
1734 	 */
1735 	if (sge->tx_sched && !qid && skb->dev) {
1736 use_sched:
1737 		use_sched_skb = 1;
1738 		/* Note that the scheduler might return a different skb than
1739 		 * the one passed in.
1740 		 */
1741 		skb = sched_skb(sge, skb, credits);
1742 		if (!skb) {
1743 			spin_unlock(&q->lock);
1744 			return NETDEV_TX_OK;
1745 		}
1746 		pidx = q->pidx;
1747 		count = 1 + skb_shinfo(skb)->nr_frags;
1748 		count += compute_large_page_tx_descs(skb);
1749 	}
1750 
1751 	q->in_use += count;
1752 	genbit = q->genbit;
1753 	pidx = q->pidx;
1754 	q->pidx += count;
1755 	if (q->pidx >= q->size) {
1756 		q->pidx -= q->size;
1757 		q->genbit ^= 1;
1758 	}
1759 	spin_unlock(&q->lock);
1760 
1761 	write_tx_descs(adapter, skb, pidx, genbit, q);
1762 
1763 	/*
1764 	 * We always ring the doorbell for cmdQ1.  For cmdQ0, we only ring
1765 	 * the doorbell if the Q is asleep. There is a natural race, where
1766 	 * the hardware is going to sleep just after we checked, however,
1767 	 * then the interrupt handler will detect the outstanding TX packet
1768 	 * and ring the doorbell for us.
1769 	 */
1770 	if (qid)
1771 		doorbell_pio(adapter, F_CMDQ1_ENABLE);
1772 	else {
1773 		clear_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1774 		if (test_and_set_bit(CMDQ_STAT_RUNNING, &q->status) == 0) {
1775 			set_bit(CMDQ_STAT_LAST_PKT_DB, &q->status);
1776 			writel(F_CMDQ0_ENABLE, adapter->regs + A_SG_DOORBELL);
1777 		}
1778 	}
1779 
1780 	if (use_sched_skb) {
1781 		if (spin_trylock(&q->lock)) {
1782 			credits = q->size - q->in_use;
1783 			skb = NULL;
1784 			goto use_sched;
1785 		}
1786 	}
1787 	return NETDEV_TX_OK;
1788 }
1789 
1790 #define MK_ETH_TYPE_MSS(type, mss) (((mss) & 0x3FFF) | ((type) << 14))
1791 
1792 /*
1793  *	eth_hdr_len - return the length of an Ethernet header
1794  *	@data: pointer to the start of the Ethernet header
1795  *
1796  *	Returns the length of an Ethernet header, including optional VLAN tag.
1797  */
1798 static inline int eth_hdr_len(const void *data)
1799 {
1800 	const struct ethhdr *e = data;
1801 
1802 	return e->h_proto == htons(ETH_P_8021Q) ? VLAN_ETH_HLEN : ETH_HLEN;
1803 }
1804 
1805 /*
1806  * Adds the CPL header to the sk_buff and passes it to t1_sge_tx.
1807  */
1808 netdev_tx_t t1_start_xmit(struct sk_buff *skb, struct net_device *dev)
1809 {
1810 	struct adapter *adapter = dev->ml_priv;
1811 	struct sge *sge = adapter->sge;
1812 	struct sge_port_stats *st = this_cpu_ptr(sge->port_stats[dev->if_port]);
1813 	struct cpl_tx_pkt *cpl;
1814 	struct sk_buff *orig_skb = skb;
1815 	int ret;
1816 
1817 	if (skb->protocol == htons(ETH_P_CPL5))
1818 		goto send;
1819 
1820 	/*
1821 	 * We are using a non-standard hard_header_len.
1822 	 * Allocate more header room in the rare cases it is not big enough.
1823 	 */
1824 	if (unlikely(skb_headroom(skb) < dev->hard_header_len - ETH_HLEN)) {
1825 		skb = skb_realloc_headroom(skb, sizeof(struct cpl_tx_pkt_lso));
1826 		++st->tx_need_hdrroom;
1827 		dev_kfree_skb_any(orig_skb);
1828 		if (!skb)
1829 			return NETDEV_TX_OK;
1830 	}
1831 
1832 	if (skb_shinfo(skb)->gso_size) {
1833 		int eth_type;
1834 		struct cpl_tx_pkt_lso *hdr;
1835 
1836 		++st->tx_tso;
1837 
1838 		eth_type = skb_network_offset(skb) == ETH_HLEN ?
1839 			CPL_ETH_II : CPL_ETH_II_VLAN;
1840 
1841 		hdr = skb_push(skb, sizeof(*hdr));
1842 		hdr->opcode = CPL_TX_PKT_LSO;
1843 		hdr->ip_csum_dis = hdr->l4_csum_dis = 0;
1844 		hdr->ip_hdr_words = ip_hdr(skb)->ihl;
1845 		hdr->tcp_hdr_words = tcp_hdr(skb)->doff;
1846 		hdr->eth_type_mss = htons(MK_ETH_TYPE_MSS(eth_type,
1847 							  skb_shinfo(skb)->gso_size));
1848 		hdr->len = htonl(skb->len - sizeof(*hdr));
1849 		cpl = (struct cpl_tx_pkt *)hdr;
1850 	} else {
1851 		/*
1852 		 * Packets shorter than ETH_HLEN can break the MAC, drop them
1853 		 * early.  Also, we may get oversized packets because some
1854 		 * parts of the kernel don't handle our unusual hard_header_len
1855 		 * right, drop those too.
1856 		 */
1857 		if (unlikely(skb->len < ETH_HLEN ||
1858 			     skb->len > dev->mtu + eth_hdr_len(skb->data))) {
1859 			netdev_dbg(dev, "packet size %d hdr %d mtu%d\n",
1860 				   skb->len, eth_hdr_len(skb->data), dev->mtu);
1861 			dev_kfree_skb_any(skb);
1862 			return NETDEV_TX_OK;
1863 		}
1864 
1865 		if (skb->ip_summed == CHECKSUM_PARTIAL &&
1866 		    ip_hdr(skb)->protocol == IPPROTO_UDP) {
1867 			if (unlikely(skb_checksum_help(skb))) {
1868 				netdev_dbg(dev, "unable to do udp checksum\n");
1869 				dev_kfree_skb_any(skb);
1870 				return NETDEV_TX_OK;
1871 			}
1872 		}
1873 
1874 		/* Hmmm, assuming to catch the gratious arp... and we'll use
1875 		 * it to flush out stuck espi packets...
1876 		 */
1877 		if ((unlikely(!adapter->sge->espibug_skb[dev->if_port]))) {
1878 			if (skb->protocol == htons(ETH_P_ARP) &&
1879 			    arp_hdr(skb)->ar_op == htons(ARPOP_REQUEST)) {
1880 				adapter->sge->espibug_skb[dev->if_port] = skb;
1881 				/* We want to re-use this skb later. We
1882 				 * simply bump the reference count and it
1883 				 * will not be freed...
1884 				 */
1885 				skb = skb_get(skb);
1886 			}
1887 		}
1888 
1889 		cpl = __skb_push(skb, sizeof(*cpl));
1890 		cpl->opcode = CPL_TX_PKT;
1891 		cpl->ip_csum_dis = 1;    /* SW calculates IP csum */
1892 		cpl->l4_csum_dis = skb->ip_summed == CHECKSUM_PARTIAL ? 0 : 1;
1893 		/* the length field isn't used so don't bother setting it */
1894 
1895 		st->tx_cso += (skb->ip_summed == CHECKSUM_PARTIAL);
1896 	}
1897 	cpl->iff = dev->if_port;
1898 
1899 	if (skb_vlan_tag_present(skb)) {
1900 		cpl->vlan_valid = 1;
1901 		cpl->vlan = htons(skb_vlan_tag_get(skb));
1902 		st->vlan_insert++;
1903 	} else
1904 		cpl->vlan_valid = 0;
1905 
1906 send:
1907 	ret = t1_sge_tx(skb, adapter, 0, dev);
1908 
1909 	/* If transmit busy, and we reallocated skb's due to headroom limit,
1910 	 * then silently discard to avoid leak.
1911 	 */
1912 	if (unlikely(ret != NETDEV_TX_OK && skb != orig_skb)) {
1913 		dev_kfree_skb_any(skb);
1914 		ret = NETDEV_TX_OK;
1915 	}
1916 	return ret;
1917 }
1918 
1919 /*
1920  * Callback for the Tx buffer reclaim timer.  Runs with softirqs disabled.
1921  */
1922 static void sge_tx_reclaim_cb(struct timer_list *t)
1923 {
1924 	int i;
1925 	struct sge *sge = from_timer(sge, t, tx_reclaim_timer);
1926 
1927 	for (i = 0; i < SGE_CMDQ_N; ++i) {
1928 		struct cmdQ *q = &sge->cmdQ[i];
1929 
1930 		if (!spin_trylock(&q->lock))
1931 			continue;
1932 
1933 		reclaim_completed_tx(sge, q);
1934 		if (i == 0 && q->in_use) {    /* flush pending credits */
1935 			writel(F_CMDQ0_ENABLE, sge->adapter->regs + A_SG_DOORBELL);
1936 		}
1937 		spin_unlock(&q->lock);
1938 	}
1939 	mod_timer(&sge->tx_reclaim_timer, jiffies + TX_RECLAIM_PERIOD);
1940 }
1941 
1942 /*
1943  * Propagate changes of the SGE coalescing parameters to the HW.
1944  */
1945 int t1_sge_set_coalesce_params(struct sge *sge, struct sge_params *p)
1946 {
1947 	sge->fixed_intrtimer = p->rx_coalesce_usecs *
1948 		core_ticks_per_usec(sge->adapter);
1949 	writel(sge->fixed_intrtimer, sge->adapter->regs + A_SG_INTRTIMER);
1950 	return 0;
1951 }
1952 
1953 /*
1954  * Allocates both RX and TX resources and configures the SGE. However,
1955  * the hardware is not enabled yet.
1956  */
1957 int t1_sge_configure(struct sge *sge, struct sge_params *p)
1958 {
1959 	if (alloc_rx_resources(sge, p))
1960 		return -ENOMEM;
1961 	if (alloc_tx_resources(sge, p)) {
1962 		free_rx_resources(sge);
1963 		return -ENOMEM;
1964 	}
1965 	configure_sge(sge, p);
1966 
1967 	/*
1968 	 * Now that we have sized the free lists calculate the payload
1969 	 * capacity of the large buffers.  Other parts of the driver use
1970 	 * this to set the max offload coalescing size so that RX packets
1971 	 * do not overflow our large buffers.
1972 	 */
1973 	p->large_buf_capacity = jumbo_payload_capacity(sge);
1974 	return 0;
1975 }
1976 
1977 /*
1978  * Disables the DMA engine.
1979  */
1980 void t1_sge_stop(struct sge *sge)
1981 {
1982 	int i;
1983 	writel(0, sge->adapter->regs + A_SG_CONTROL);
1984 	readl(sge->adapter->regs + A_SG_CONTROL); /* flush */
1985 
1986 	if (is_T2(sge->adapter))
1987 		del_timer_sync(&sge->espibug_timer);
1988 
1989 	del_timer_sync(&sge->tx_reclaim_timer);
1990 	if (sge->tx_sched)
1991 		tx_sched_stop(sge);
1992 
1993 	for (i = 0; i < MAX_NPORTS; i++)
1994 		kfree_skb(sge->espibug_skb[i]);
1995 }
1996 
1997 /*
1998  * Enables the DMA engine.
1999  */
2000 void t1_sge_start(struct sge *sge)
2001 {
2002 	refill_free_list(sge, &sge->freelQ[0]);
2003 	refill_free_list(sge, &sge->freelQ[1]);
2004 
2005 	writel(sge->sge_control, sge->adapter->regs + A_SG_CONTROL);
2006 	doorbell_pio(sge->adapter, F_FL0_ENABLE | F_FL1_ENABLE);
2007 	readl(sge->adapter->regs + A_SG_CONTROL); /* flush */
2008 
2009 	mod_timer(&sge->tx_reclaim_timer, jiffies + TX_RECLAIM_PERIOD);
2010 
2011 	if (is_T2(sge->adapter))
2012 		mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
2013 }
2014 
2015 /*
2016  * Callback for the T2 ESPI 'stuck packet feature' workaorund
2017  */
2018 static void espibug_workaround_t204(struct timer_list *t)
2019 {
2020 	struct sge *sge = from_timer(sge, t, espibug_timer);
2021 	struct adapter *adapter = sge->adapter;
2022 	unsigned int nports = adapter->params.nports;
2023 	u32 seop[MAX_NPORTS];
2024 
2025 	if (adapter->open_device_map & PORT_MASK) {
2026 		int i;
2027 
2028 		if (t1_espi_get_mon_t204(adapter, &(seop[0]), 0) < 0)
2029 			return;
2030 
2031 		for (i = 0; i < nports; i++) {
2032 			struct sk_buff *skb = sge->espibug_skb[i];
2033 
2034 			if (!netif_running(adapter->port[i].dev) ||
2035 			    netif_queue_stopped(adapter->port[i].dev) ||
2036 			    !seop[i] || ((seop[i] & 0xfff) != 0) || !skb)
2037 				continue;
2038 
2039 			if (!skb->cb[0]) {
2040 				skb_copy_to_linear_data_offset(skb,
2041 						    sizeof(struct cpl_tx_pkt),
2042 							       ch_mac_addr,
2043 							       ETH_ALEN);
2044 				skb_copy_to_linear_data_offset(skb,
2045 							       skb->len - 10,
2046 							       ch_mac_addr,
2047 							       ETH_ALEN);
2048 				skb->cb[0] = 0xff;
2049 			}
2050 
2051 			/* bump the reference count to avoid freeing of
2052 			 * the skb once the DMA has completed.
2053 			 */
2054 			skb = skb_get(skb);
2055 			t1_sge_tx(skb, adapter, 0, adapter->port[i].dev);
2056 		}
2057 	}
2058 	mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
2059 }
2060 
2061 static void espibug_workaround(struct timer_list *t)
2062 {
2063 	struct sge *sge = from_timer(sge, t, espibug_timer);
2064 	struct adapter *adapter = sge->adapter;
2065 
2066 	if (netif_running(adapter->port[0].dev)) {
2067 	        struct sk_buff *skb = sge->espibug_skb[0];
2068 	        u32 seop = t1_espi_get_mon(adapter, 0x930, 0);
2069 
2070 	        if ((seop & 0xfff0fff) == 0xfff && skb) {
2071 	                if (!skb->cb[0]) {
2072 	                        skb_copy_to_linear_data_offset(skb,
2073 						     sizeof(struct cpl_tx_pkt),
2074 							       ch_mac_addr,
2075 							       ETH_ALEN);
2076 	                        skb_copy_to_linear_data_offset(skb,
2077 							       skb->len - 10,
2078 							       ch_mac_addr,
2079 							       ETH_ALEN);
2080 	                        skb->cb[0] = 0xff;
2081 	                }
2082 
2083 	                /* bump the reference count to avoid freeing of the
2084 	                 * skb once the DMA has completed.
2085 	                 */
2086 	                skb = skb_get(skb);
2087 	                t1_sge_tx(skb, adapter, 0, adapter->port[0].dev);
2088 	        }
2089 	}
2090 	mod_timer(&sge->espibug_timer, jiffies + sge->espibug_timeout);
2091 }
2092 
2093 /*
2094  * Creates a t1_sge structure and returns suggested resource parameters.
2095  */
2096 struct sge *t1_sge_create(struct adapter *adapter, struct sge_params *p)
2097 {
2098 	struct sge *sge = kzalloc(sizeof(*sge), GFP_KERNEL);
2099 	int i;
2100 
2101 	if (!sge)
2102 		return NULL;
2103 
2104 	sge->adapter = adapter;
2105 	sge->netdev = adapter->port[0].dev;
2106 	sge->rx_pkt_pad = t1_is_T1B(adapter) ? 0 : 2;
2107 	sge->jumbo_fl = t1_is_T1B(adapter) ? 1 : 0;
2108 
2109 	for_each_port(adapter, i) {
2110 		sge->port_stats[i] = alloc_percpu(struct sge_port_stats);
2111 		if (!sge->port_stats[i])
2112 			goto nomem_port;
2113 	}
2114 
2115 	timer_setup(&sge->tx_reclaim_timer, sge_tx_reclaim_cb, 0);
2116 
2117 	if (is_T2(sge->adapter)) {
2118 		timer_setup(&sge->espibug_timer,
2119 			    adapter->params.nports > 1 ? espibug_workaround_t204 : espibug_workaround,
2120 			    0);
2121 
2122 		if (adapter->params.nports > 1)
2123 			tx_sched_init(sge);
2124 
2125 		sge->espibug_timeout = 1;
2126 		/* for T204, every 10ms */
2127 		if (adapter->params.nports > 1)
2128 			sge->espibug_timeout = HZ/100;
2129 	}
2130 
2131 
2132 	p->cmdQ_size[0] = SGE_CMDQ0_E_N;
2133 	p->cmdQ_size[1] = SGE_CMDQ1_E_N;
2134 	p->freelQ_size[!sge->jumbo_fl] = SGE_FREEL_SIZE;
2135 	p->freelQ_size[sge->jumbo_fl] = SGE_JUMBO_FREEL_SIZE;
2136 	if (sge->tx_sched) {
2137 		if (board_info(sge->adapter)->board == CHBT_BOARD_CHT204)
2138 			p->rx_coalesce_usecs = 15;
2139 		else
2140 			p->rx_coalesce_usecs = 50;
2141 	} else
2142 		p->rx_coalesce_usecs = 50;
2143 
2144 	p->coalesce_enable = 0;
2145 	p->sample_interval_usecs = 0;
2146 
2147 	return sge;
2148 nomem_port:
2149 	while (i >= 0) {
2150 		free_percpu(sge->port_stats[i]);
2151 		--i;
2152 	}
2153 	kfree(sge);
2154 	return NULL;
2155 
2156 }
2157