1 /*
2  * This file is part of the Chelsio T4 Ethernet driver for Linux.
3  *
4  * Copyright (c) 2003-2014 Chelsio Communications, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34 
35 #include <linux/skbuff.h>
36 #include <linux/netdevice.h>
37 #include <linux/etherdevice.h>
38 #include <linux/if_vlan.h>
39 #include <linux/ip.h>
40 #include <linux/dma-mapping.h>
41 #include <linux/jiffies.h>
42 #include <linux/prefetch.h>
43 #include <linux/export.h>
44 #include <net/xfrm.h>
45 #include <net/ipv6.h>
46 #include <net/tcp.h>
47 #include <net/busy_poll.h>
48 #ifdef CONFIG_CHELSIO_T4_FCOE
49 #include <scsi/fc/fc_fcoe.h>
50 #endif /* CONFIG_CHELSIO_T4_FCOE */
51 #include "cxgb4.h"
52 #include "t4_regs.h"
53 #include "t4_values.h"
54 #include "t4_msg.h"
55 #include "t4fw_api.h"
56 #include "cxgb4_ptp.h"
57 #include "cxgb4_uld.h"
58 #include "cxgb4_tc_mqprio.h"
59 #include "sched.h"
60 
61 /*
62  * Rx buffer size.  We use largish buffers if possible but settle for single
63  * pages under memory shortage.
64  */
65 #if PAGE_SHIFT >= 16
66 # define FL_PG_ORDER 0
67 #else
68 # define FL_PG_ORDER (16 - PAGE_SHIFT)
69 #endif
70 
71 /* RX_PULL_LEN should be <= RX_COPY_THRES */
72 #define RX_COPY_THRES    256
73 #define RX_PULL_LEN      128
74 
75 /*
76  * Main body length for sk_buffs used for Rx Ethernet packets with fragments.
77  * Should be >= RX_PULL_LEN but possibly bigger to give pskb_may_pull some room.
78  */
79 #define RX_PKT_SKB_LEN   512
80 
81 /*
82  * Max number of Tx descriptors we clean up at a time.  Should be modest as
83  * freeing skbs isn't cheap and it happens while holding locks.  We just need
84  * to free packets faster than they arrive, we eventually catch up and keep
85  * the amortized cost reasonable.  Must be >= 2 * TXQ_STOP_THRES.  It should
86  * also match the CIDX Flush Threshold.
87  */
88 #define MAX_TX_RECLAIM 32
89 
90 /*
91  * Max number of Rx buffers we replenish at a time.  Again keep this modest,
92  * allocating buffers isn't cheap either.
93  */
94 #define MAX_RX_REFILL 16U
95 
96 /*
97  * Period of the Rx queue check timer.  This timer is infrequent as it has
98  * something to do only when the system experiences severe memory shortage.
99  */
100 #define RX_QCHECK_PERIOD (HZ / 2)
101 
102 /*
103  * Period of the Tx queue check timer.
104  */
105 #define TX_QCHECK_PERIOD (HZ / 2)
106 
107 /*
108  * Max number of Tx descriptors to be reclaimed by the Tx timer.
109  */
110 #define MAX_TIMER_TX_RECLAIM 100
111 
112 /*
113  * Timer index used when backing off due to memory shortage.
114  */
115 #define NOMEM_TMR_IDX (SGE_NTIMERS - 1)
116 
117 /*
118  * Suspension threshold for non-Ethernet Tx queues.  We require enough room
119  * for a full sized WR.
120  */
121 #define TXQ_STOP_THRES (SGE_MAX_WR_LEN / sizeof(struct tx_desc))
122 
123 /*
124  * Max Tx descriptor space we allow for an Ethernet packet to be inlined
125  * into a WR.
126  */
127 #define MAX_IMM_TX_PKT_LEN 256
128 
129 /*
130  * Max size of a WR sent through a control Tx queue.
131  */
132 #define MAX_CTRL_WR_LEN SGE_MAX_WR_LEN
133 
134 struct rx_sw_desc {                /* SW state per Rx descriptor */
135 	struct page *page;
136 	dma_addr_t dma_addr;
137 };
138 
139 /*
140  * Rx buffer sizes for "useskbs" Free List buffers (one ingress packet pe skb
141  * buffer).  We currently only support two sizes for 1500- and 9000-byte MTUs.
142  * We could easily support more but there doesn't seem to be much need for
143  * that ...
144  */
145 #define FL_MTU_SMALL 1500
146 #define FL_MTU_LARGE 9000
147 
148 static inline unsigned int fl_mtu_bufsize(struct adapter *adapter,
149 					  unsigned int mtu)
150 {
151 	struct sge *s = &adapter->sge;
152 
153 	return ALIGN(s->pktshift + ETH_HLEN + VLAN_HLEN + mtu, s->fl_align);
154 }
155 
156 #define FL_MTU_SMALL_BUFSIZE(adapter) fl_mtu_bufsize(adapter, FL_MTU_SMALL)
157 #define FL_MTU_LARGE_BUFSIZE(adapter) fl_mtu_bufsize(adapter, FL_MTU_LARGE)
158 
159 /*
160  * Bits 0..3 of rx_sw_desc.dma_addr have special meaning.  The hardware uses
161  * these to specify the buffer size as an index into the SGE Free List Buffer
162  * Size register array.  We also use bit 4, when the buffer has been unmapped
163  * for DMA, but this is of course never sent to the hardware and is only used
164  * to prevent double unmappings.  All of the above requires that the Free List
165  * Buffers which we allocate have the bottom 5 bits free (0) -- i.e. are
166  * 32-byte or or a power of 2 greater in alignment.  Since the SGE's minimal
167  * Free List Buffer alignment is 32 bytes, this works out for us ...
168  */
169 enum {
170 	RX_BUF_FLAGS     = 0x1f,   /* bottom five bits are special */
171 	RX_BUF_SIZE      = 0x0f,   /* bottom three bits are for buf sizes */
172 	RX_UNMAPPED_BUF  = 0x10,   /* buffer is not mapped */
173 
174 	/*
175 	 * XXX We shouldn't depend on being able to use these indices.
176 	 * XXX Especially when some other Master PF has initialized the
177 	 * XXX adapter or we use the Firmware Configuration File.  We
178 	 * XXX should really search through the Host Buffer Size register
179 	 * XXX array for the appropriately sized buffer indices.
180 	 */
181 	RX_SMALL_PG_BUF  = 0x0,   /* small (PAGE_SIZE) page buffer */
182 	RX_LARGE_PG_BUF  = 0x1,   /* buffer large (FL_PG_ORDER) page buffer */
183 
184 	RX_SMALL_MTU_BUF = 0x2,   /* small MTU buffer */
185 	RX_LARGE_MTU_BUF = 0x3,   /* large MTU buffer */
186 };
187 
188 static int timer_pkt_quota[] = {1, 1, 2, 3, 4, 5};
189 #define MIN_NAPI_WORK  1
190 
191 static inline dma_addr_t get_buf_addr(const struct rx_sw_desc *d)
192 {
193 	return d->dma_addr & ~(dma_addr_t)RX_BUF_FLAGS;
194 }
195 
196 static inline bool is_buf_mapped(const struct rx_sw_desc *d)
197 {
198 	return !(d->dma_addr & RX_UNMAPPED_BUF);
199 }
200 
201 /**
202  *	txq_avail - return the number of available slots in a Tx queue
203  *	@q: the Tx queue
204  *
205  *	Returns the number of descriptors in a Tx queue available to write new
206  *	packets.
207  */
208 static inline unsigned int txq_avail(const struct sge_txq *q)
209 {
210 	return q->size - 1 - q->in_use;
211 }
212 
213 /**
214  *	fl_cap - return the capacity of a free-buffer list
215  *	@fl: the FL
216  *
217  *	Returns the capacity of a free-buffer list.  The capacity is less than
218  *	the size because one descriptor needs to be left unpopulated, otherwise
219  *	HW will think the FL is empty.
220  */
221 static inline unsigned int fl_cap(const struct sge_fl *fl)
222 {
223 	return fl->size - 8;   /* 1 descriptor = 8 buffers */
224 }
225 
226 /**
227  *	fl_starving - return whether a Free List is starving.
228  *	@adapter: pointer to the adapter
229  *	@fl: the Free List
230  *
231  *	Tests specified Free List to see whether the number of buffers
232  *	available to the hardware has falled below our "starvation"
233  *	threshold.
234  */
235 static inline bool fl_starving(const struct adapter *adapter,
236 			       const struct sge_fl *fl)
237 {
238 	const struct sge *s = &adapter->sge;
239 
240 	return fl->avail - fl->pend_cred <= s->fl_starve_thres;
241 }
242 
243 int cxgb4_map_skb(struct device *dev, const struct sk_buff *skb,
244 		  dma_addr_t *addr)
245 {
246 	const skb_frag_t *fp, *end;
247 	const struct skb_shared_info *si;
248 
249 	*addr = dma_map_single(dev, skb->data, skb_headlen(skb), DMA_TO_DEVICE);
250 	if (dma_mapping_error(dev, *addr))
251 		goto out_err;
252 
253 	si = skb_shinfo(skb);
254 	end = &si->frags[si->nr_frags];
255 
256 	for (fp = si->frags; fp < end; fp++) {
257 		*++addr = skb_frag_dma_map(dev, fp, 0, skb_frag_size(fp),
258 					   DMA_TO_DEVICE);
259 		if (dma_mapping_error(dev, *addr))
260 			goto unwind;
261 	}
262 	return 0;
263 
264 unwind:
265 	while (fp-- > si->frags)
266 		dma_unmap_page(dev, *--addr, skb_frag_size(fp), DMA_TO_DEVICE);
267 
268 	dma_unmap_single(dev, addr[-1], skb_headlen(skb), DMA_TO_DEVICE);
269 out_err:
270 	return -ENOMEM;
271 }
272 EXPORT_SYMBOL(cxgb4_map_skb);
273 
274 static void unmap_skb(struct device *dev, const struct sk_buff *skb,
275 		      const dma_addr_t *addr)
276 {
277 	const skb_frag_t *fp, *end;
278 	const struct skb_shared_info *si;
279 
280 	dma_unmap_single(dev, *addr++, skb_headlen(skb), DMA_TO_DEVICE);
281 
282 	si = skb_shinfo(skb);
283 	end = &si->frags[si->nr_frags];
284 	for (fp = si->frags; fp < end; fp++)
285 		dma_unmap_page(dev, *addr++, skb_frag_size(fp), DMA_TO_DEVICE);
286 }
287 
288 #ifdef CONFIG_NEED_DMA_MAP_STATE
289 /**
290  *	deferred_unmap_destructor - unmap a packet when it is freed
291  *	@skb: the packet
292  *
293  *	This is the packet destructor used for Tx packets that need to remain
294  *	mapped until they are freed rather than until their Tx descriptors are
295  *	freed.
296  */
297 static void deferred_unmap_destructor(struct sk_buff *skb)
298 {
299 	unmap_skb(skb->dev->dev.parent, skb, (dma_addr_t *)skb->head);
300 }
301 #endif
302 
303 /**
304  *	free_tx_desc - reclaims Tx descriptors and their buffers
305  *	@adap: the adapter
306  *	@q: the Tx queue to reclaim descriptors from
307  *	@n: the number of descriptors to reclaim
308  *	@unmap: whether the buffers should be unmapped for DMA
309  *
310  *	Reclaims Tx descriptors from an SGE Tx queue and frees the associated
311  *	Tx buffers.  Called with the Tx queue lock held.
312  */
313 void free_tx_desc(struct adapter *adap, struct sge_txq *q,
314 		  unsigned int n, bool unmap)
315 {
316 	unsigned int cidx = q->cidx;
317 	struct tx_sw_desc *d;
318 
319 	d = &q->sdesc[cidx];
320 	while (n--) {
321 		if (d->skb) {                       /* an SGL is present */
322 			if (unmap && d->addr[0]) {
323 				unmap_skb(adap->pdev_dev, d->skb, d->addr);
324 				memset(d->addr, 0, sizeof(d->addr));
325 			}
326 			dev_consume_skb_any(d->skb);
327 			d->skb = NULL;
328 		}
329 		++d;
330 		if (++cidx == q->size) {
331 			cidx = 0;
332 			d = q->sdesc;
333 		}
334 	}
335 	q->cidx = cidx;
336 }
337 
338 /*
339  * Return the number of reclaimable descriptors in a Tx queue.
340  */
341 static inline int reclaimable(const struct sge_txq *q)
342 {
343 	int hw_cidx = ntohs(READ_ONCE(q->stat->cidx));
344 	hw_cidx -= q->cidx;
345 	return hw_cidx < 0 ? hw_cidx + q->size : hw_cidx;
346 }
347 
348 /**
349  *	reclaim_completed_tx - reclaims completed TX Descriptors
350  *	@adap: the adapter
351  *	@q: the Tx queue to reclaim completed descriptors from
352  *	@maxreclaim: the maximum number of TX Descriptors to reclaim or -1
353  *	@unmap: whether the buffers should be unmapped for DMA
354  *
355  *	Reclaims Tx Descriptors that the SGE has indicated it has processed,
356  *	and frees the associated buffers if possible.  If @max == -1, then
357  *	we'll use a defaiult maximum.  Called with the TX Queue locked.
358  */
359 static inline int reclaim_completed_tx(struct adapter *adap, struct sge_txq *q,
360 				       int maxreclaim, bool unmap)
361 {
362 	int reclaim = reclaimable(q);
363 
364 	if (reclaim) {
365 		/*
366 		 * Limit the amount of clean up work we do at a time to keep
367 		 * the Tx lock hold time O(1).
368 		 */
369 		if (maxreclaim < 0)
370 			maxreclaim = MAX_TX_RECLAIM;
371 		if (reclaim > maxreclaim)
372 			reclaim = maxreclaim;
373 
374 		free_tx_desc(adap, q, reclaim, unmap);
375 		q->in_use -= reclaim;
376 	}
377 
378 	return reclaim;
379 }
380 
381 /**
382  *	cxgb4_reclaim_completed_tx - reclaims completed Tx descriptors
383  *	@adap: the adapter
384  *	@q: the Tx queue to reclaim completed descriptors from
385  *	@unmap: whether the buffers should be unmapped for DMA
386  *
387  *	Reclaims Tx descriptors that the SGE has indicated it has processed,
388  *	and frees the associated buffers if possible.  Called with the Tx
389  *	queue locked.
390  */
391 void cxgb4_reclaim_completed_tx(struct adapter *adap, struct sge_txq *q,
392 				bool unmap)
393 {
394 	(void)reclaim_completed_tx(adap, q, -1, unmap);
395 }
396 EXPORT_SYMBOL(cxgb4_reclaim_completed_tx);
397 
398 static inline int get_buf_size(struct adapter *adapter,
399 			       const struct rx_sw_desc *d)
400 {
401 	struct sge *s = &adapter->sge;
402 	unsigned int rx_buf_size_idx = d->dma_addr & RX_BUF_SIZE;
403 	int buf_size;
404 
405 	switch (rx_buf_size_idx) {
406 	case RX_SMALL_PG_BUF:
407 		buf_size = PAGE_SIZE;
408 		break;
409 
410 	case RX_LARGE_PG_BUF:
411 		buf_size = PAGE_SIZE << s->fl_pg_order;
412 		break;
413 
414 	case RX_SMALL_MTU_BUF:
415 		buf_size = FL_MTU_SMALL_BUFSIZE(adapter);
416 		break;
417 
418 	case RX_LARGE_MTU_BUF:
419 		buf_size = FL_MTU_LARGE_BUFSIZE(adapter);
420 		break;
421 
422 	default:
423 		BUG();
424 	}
425 
426 	return buf_size;
427 }
428 
429 /**
430  *	free_rx_bufs - free the Rx buffers on an SGE free list
431  *	@adap: the adapter
432  *	@q: the SGE free list to free buffers from
433  *	@n: how many buffers to free
434  *
435  *	Release the next @n buffers on an SGE free-buffer Rx queue.   The
436  *	buffers must be made inaccessible to HW before calling this function.
437  */
438 static void free_rx_bufs(struct adapter *adap, struct sge_fl *q, int n)
439 {
440 	while (n--) {
441 		struct rx_sw_desc *d = &q->sdesc[q->cidx];
442 
443 		if (is_buf_mapped(d))
444 			dma_unmap_page(adap->pdev_dev, get_buf_addr(d),
445 				       get_buf_size(adap, d),
446 				       DMA_FROM_DEVICE);
447 		put_page(d->page);
448 		d->page = NULL;
449 		if (++q->cidx == q->size)
450 			q->cidx = 0;
451 		q->avail--;
452 	}
453 }
454 
455 /**
456  *	unmap_rx_buf - unmap the current Rx buffer on an SGE free list
457  *	@adap: the adapter
458  *	@q: the SGE free list
459  *
460  *	Unmap the current buffer on an SGE free-buffer Rx queue.   The
461  *	buffer must be made inaccessible to HW before calling this function.
462  *
463  *	This is similar to @free_rx_bufs above but does not free the buffer.
464  *	Do note that the FL still loses any further access to the buffer.
465  */
466 static void unmap_rx_buf(struct adapter *adap, struct sge_fl *q)
467 {
468 	struct rx_sw_desc *d = &q->sdesc[q->cidx];
469 
470 	if (is_buf_mapped(d))
471 		dma_unmap_page(adap->pdev_dev, get_buf_addr(d),
472 			       get_buf_size(adap, d), DMA_FROM_DEVICE);
473 	d->page = NULL;
474 	if (++q->cidx == q->size)
475 		q->cidx = 0;
476 	q->avail--;
477 }
478 
479 static inline void ring_fl_db(struct adapter *adap, struct sge_fl *q)
480 {
481 	if (q->pend_cred >= 8) {
482 		u32 val = adap->params.arch.sge_fl_db;
483 
484 		if (is_t4(adap->params.chip))
485 			val |= PIDX_V(q->pend_cred / 8);
486 		else
487 			val |= PIDX_T5_V(q->pend_cred / 8);
488 
489 		/* Make sure all memory writes to the Free List queue are
490 		 * committed before we tell the hardware about them.
491 		 */
492 		wmb();
493 
494 		/* If we don't have access to the new User Doorbell (T5+), use
495 		 * the old doorbell mechanism; otherwise use the new BAR2
496 		 * mechanism.
497 		 */
498 		if (unlikely(q->bar2_addr == NULL)) {
499 			t4_write_reg(adap, MYPF_REG(SGE_PF_KDOORBELL_A),
500 				     val | QID_V(q->cntxt_id));
501 		} else {
502 			writel(val | QID_V(q->bar2_qid),
503 			       q->bar2_addr + SGE_UDB_KDOORBELL);
504 
505 			/* This Write memory Barrier will force the write to
506 			 * the User Doorbell area to be flushed.
507 			 */
508 			wmb();
509 		}
510 		q->pend_cred &= 7;
511 	}
512 }
513 
514 static inline void set_rx_sw_desc(struct rx_sw_desc *sd, struct page *pg,
515 				  dma_addr_t mapping)
516 {
517 	sd->page = pg;
518 	sd->dma_addr = mapping;      /* includes size low bits */
519 }
520 
521 /**
522  *	refill_fl - refill an SGE Rx buffer ring
523  *	@adap: the adapter
524  *	@q: the ring to refill
525  *	@n: the number of new buffers to allocate
526  *	@gfp: the gfp flags for the allocations
527  *
528  *	(Re)populate an SGE free-buffer queue with up to @n new packet buffers,
529  *	allocated with the supplied gfp flags.  The caller must assure that
530  *	@n does not exceed the queue's capacity.  If afterwards the queue is
531  *	found critically low mark it as starving in the bitmap of starving FLs.
532  *
533  *	Returns the number of buffers allocated.
534  */
535 static unsigned int refill_fl(struct adapter *adap, struct sge_fl *q, int n,
536 			      gfp_t gfp)
537 {
538 	struct sge *s = &adap->sge;
539 	struct page *pg;
540 	dma_addr_t mapping;
541 	unsigned int cred = q->avail;
542 	__be64 *d = &q->desc[q->pidx];
543 	struct rx_sw_desc *sd = &q->sdesc[q->pidx];
544 	int node;
545 
546 #ifdef CONFIG_DEBUG_FS
547 	if (test_bit(q->cntxt_id - adap->sge.egr_start, adap->sge.blocked_fl))
548 		goto out;
549 #endif
550 
551 	gfp |= __GFP_NOWARN;
552 	node = dev_to_node(adap->pdev_dev);
553 
554 	if (s->fl_pg_order == 0)
555 		goto alloc_small_pages;
556 
557 	/*
558 	 * Prefer large buffers
559 	 */
560 	while (n) {
561 		pg = alloc_pages_node(node, gfp | __GFP_COMP, s->fl_pg_order);
562 		if (unlikely(!pg)) {
563 			q->large_alloc_failed++;
564 			break;       /* fall back to single pages */
565 		}
566 
567 		mapping = dma_map_page(adap->pdev_dev, pg, 0,
568 				       PAGE_SIZE << s->fl_pg_order,
569 				       DMA_FROM_DEVICE);
570 		if (unlikely(dma_mapping_error(adap->pdev_dev, mapping))) {
571 			__free_pages(pg, s->fl_pg_order);
572 			q->mapping_err++;
573 			goto out;   /* do not try small pages for this error */
574 		}
575 		mapping |= RX_LARGE_PG_BUF;
576 		*d++ = cpu_to_be64(mapping);
577 
578 		set_rx_sw_desc(sd, pg, mapping);
579 		sd++;
580 
581 		q->avail++;
582 		if (++q->pidx == q->size) {
583 			q->pidx = 0;
584 			sd = q->sdesc;
585 			d = q->desc;
586 		}
587 		n--;
588 	}
589 
590 alloc_small_pages:
591 	while (n--) {
592 		pg = alloc_pages_node(node, gfp, 0);
593 		if (unlikely(!pg)) {
594 			q->alloc_failed++;
595 			break;
596 		}
597 
598 		mapping = dma_map_page(adap->pdev_dev, pg, 0, PAGE_SIZE,
599 				       DMA_FROM_DEVICE);
600 		if (unlikely(dma_mapping_error(adap->pdev_dev, mapping))) {
601 			put_page(pg);
602 			q->mapping_err++;
603 			goto out;
604 		}
605 		*d++ = cpu_to_be64(mapping);
606 
607 		set_rx_sw_desc(sd, pg, mapping);
608 		sd++;
609 
610 		q->avail++;
611 		if (++q->pidx == q->size) {
612 			q->pidx = 0;
613 			sd = q->sdesc;
614 			d = q->desc;
615 		}
616 	}
617 
618 out:	cred = q->avail - cred;
619 	q->pend_cred += cred;
620 	ring_fl_db(adap, q);
621 
622 	if (unlikely(fl_starving(adap, q))) {
623 		smp_wmb();
624 		q->low++;
625 		set_bit(q->cntxt_id - adap->sge.egr_start,
626 			adap->sge.starving_fl);
627 	}
628 
629 	return cred;
630 }
631 
632 static inline void __refill_fl(struct adapter *adap, struct sge_fl *fl)
633 {
634 	refill_fl(adap, fl, min(MAX_RX_REFILL, fl_cap(fl) - fl->avail),
635 		  GFP_ATOMIC);
636 }
637 
638 /**
639  *	alloc_ring - allocate resources for an SGE descriptor ring
640  *	@dev: the PCI device's core device
641  *	@nelem: the number of descriptors
642  *	@elem_size: the size of each descriptor
643  *	@sw_size: the size of the SW state associated with each ring element
644  *	@phys: the physical address of the allocated ring
645  *	@metadata: address of the array holding the SW state for the ring
646  *	@stat_size: extra space in HW ring for status information
647  *	@node: preferred node for memory allocations
648  *
649  *	Allocates resources for an SGE descriptor ring, such as Tx queues,
650  *	free buffer lists, or response queues.  Each SGE ring requires
651  *	space for its HW descriptors plus, optionally, space for the SW state
652  *	associated with each HW entry (the metadata).  The function returns
653  *	three values: the virtual address for the HW ring (the return value
654  *	of the function), the bus address of the HW ring, and the address
655  *	of the SW ring.
656  */
657 static void *alloc_ring(struct device *dev, size_t nelem, size_t elem_size,
658 			size_t sw_size, dma_addr_t *phys, void *metadata,
659 			size_t stat_size, int node)
660 {
661 	size_t len = nelem * elem_size + stat_size;
662 	void *s = NULL;
663 	void *p = dma_alloc_coherent(dev, len, phys, GFP_KERNEL);
664 
665 	if (!p)
666 		return NULL;
667 	if (sw_size) {
668 		s = kcalloc_node(sw_size, nelem, GFP_KERNEL, node);
669 
670 		if (!s) {
671 			dma_free_coherent(dev, len, p, *phys);
672 			return NULL;
673 		}
674 	}
675 	if (metadata)
676 		*(void **)metadata = s;
677 	return p;
678 }
679 
680 /**
681  *	sgl_len - calculates the size of an SGL of the given capacity
682  *	@n: the number of SGL entries
683  *
684  *	Calculates the number of flits needed for a scatter/gather list that
685  *	can hold the given number of entries.
686  */
687 static inline unsigned int sgl_len(unsigned int n)
688 {
689 	/* A Direct Scatter Gather List uses 32-bit lengths and 64-bit PCI DMA
690 	 * addresses.  The DSGL Work Request starts off with a 32-bit DSGL
691 	 * ULPTX header, then Length0, then Address0, then, for 1 <= i <= N,
692 	 * repeated sequences of { Length[i], Length[i+1], Address[i],
693 	 * Address[i+1] } (this ensures that all addresses are on 64-bit
694 	 * boundaries).  If N is even, then Length[N+1] should be set to 0 and
695 	 * Address[N+1] is omitted.
696 	 *
697 	 * The following calculation incorporates all of the above.  It's
698 	 * somewhat hard to follow but, briefly: the "+2" accounts for the
699 	 * first two flits which include the DSGL header, Length0 and
700 	 * Address0; the "(3*(n-1))/2" covers the main body of list entries (3
701 	 * flits for every pair of the remaining N) +1 if (n-1) is odd; and
702 	 * finally the "+((n-1)&1)" adds the one remaining flit needed if
703 	 * (n-1) is odd ...
704 	 */
705 	n--;
706 	return (3 * n) / 2 + (n & 1) + 2;
707 }
708 
709 /**
710  *	flits_to_desc - returns the num of Tx descriptors for the given flits
711  *	@n: the number of flits
712  *
713  *	Returns the number of Tx descriptors needed for the supplied number
714  *	of flits.
715  */
716 static inline unsigned int flits_to_desc(unsigned int n)
717 {
718 	BUG_ON(n > SGE_MAX_WR_LEN / 8);
719 	return DIV_ROUND_UP(n, 8);
720 }
721 
722 /**
723  *	is_eth_imm - can an Ethernet packet be sent as immediate data?
724  *	@skb: the packet
725  *	@chip_ver: chip version
726  *
727  *	Returns whether an Ethernet packet is small enough to fit as
728  *	immediate data. Return value corresponds to headroom required.
729  */
730 static inline int is_eth_imm(const struct sk_buff *skb, unsigned int chip_ver)
731 {
732 	int hdrlen = 0;
733 
734 	if (skb->encapsulation && skb_shinfo(skb)->gso_size &&
735 	    chip_ver > CHELSIO_T5) {
736 		hdrlen = sizeof(struct cpl_tx_tnl_lso);
737 		hdrlen += sizeof(struct cpl_tx_pkt_core);
738 	} else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) {
739 		return 0;
740 	} else {
741 		hdrlen = skb_shinfo(skb)->gso_size ?
742 			 sizeof(struct cpl_tx_pkt_lso_core) : 0;
743 		hdrlen += sizeof(struct cpl_tx_pkt);
744 	}
745 	if (skb->len <= MAX_IMM_TX_PKT_LEN - hdrlen)
746 		return hdrlen;
747 	return 0;
748 }
749 
750 /**
751  *	calc_tx_flits - calculate the number of flits for a packet Tx WR
752  *	@skb: the packet
753  *	@chip_ver: chip version
754  *
755  *	Returns the number of flits needed for a Tx WR for the given Ethernet
756  *	packet, including the needed WR and CPL headers.
757  */
758 static inline unsigned int calc_tx_flits(const struct sk_buff *skb,
759 					 unsigned int chip_ver)
760 {
761 	unsigned int flits;
762 	int hdrlen = is_eth_imm(skb, chip_ver);
763 
764 	/* If the skb is small enough, we can pump it out as a work request
765 	 * with only immediate data.  In that case we just have to have the
766 	 * TX Packet header plus the skb data in the Work Request.
767 	 */
768 
769 	if (hdrlen)
770 		return DIV_ROUND_UP(skb->len + hdrlen, sizeof(__be64));
771 
772 	/* Otherwise, we're going to have to construct a Scatter gather list
773 	 * of the skb body and fragments.  We also include the flits necessary
774 	 * for the TX Packet Work Request and CPL.  We always have a firmware
775 	 * Write Header (incorporated as part of the cpl_tx_pkt_lso and
776 	 * cpl_tx_pkt structures), followed by either a TX Packet Write CPL
777 	 * message or, if we're doing a Large Send Offload, an LSO CPL message
778 	 * with an embedded TX Packet Write CPL message.
779 	 */
780 	flits = sgl_len(skb_shinfo(skb)->nr_frags + 1);
781 	if (skb_shinfo(skb)->gso_size) {
782 		if (skb->encapsulation && chip_ver > CHELSIO_T5) {
783 			hdrlen = sizeof(struct fw_eth_tx_pkt_wr) +
784 				 sizeof(struct cpl_tx_tnl_lso);
785 		} else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) {
786 			u32 pkt_hdrlen;
787 
788 			pkt_hdrlen = eth_get_headlen(skb->dev, skb->data,
789 						     skb_headlen(skb));
790 			hdrlen = sizeof(struct fw_eth_tx_eo_wr) +
791 				 round_up(pkt_hdrlen, 16);
792 		} else {
793 			hdrlen = sizeof(struct fw_eth_tx_pkt_wr) +
794 				 sizeof(struct cpl_tx_pkt_lso_core);
795 		}
796 
797 		hdrlen += sizeof(struct cpl_tx_pkt_core);
798 		flits += (hdrlen / sizeof(__be64));
799 	} else {
800 		flits += (sizeof(struct fw_eth_tx_pkt_wr) +
801 			  sizeof(struct cpl_tx_pkt_core)) / sizeof(__be64);
802 	}
803 	return flits;
804 }
805 
806 /**
807  *	calc_tx_descs - calculate the number of Tx descriptors for a packet
808  *	@skb: the packet
809  *	@chip_ver: chip version
810  *
811  *	Returns the number of Tx descriptors needed for the given Ethernet
812  *	packet, including the needed WR and CPL headers.
813  */
814 static inline unsigned int calc_tx_descs(const struct sk_buff *skb,
815 					 unsigned int chip_ver)
816 {
817 	return flits_to_desc(calc_tx_flits(skb, chip_ver));
818 }
819 
820 /**
821  *	cxgb4_write_sgl - populate a scatter/gather list for a packet
822  *	@skb: the packet
823  *	@q: the Tx queue we are writing into
824  *	@sgl: starting location for writing the SGL
825  *	@end: points right after the end of the SGL
826  *	@start: start offset into skb main-body data to include in the SGL
827  *	@addr: the list of bus addresses for the SGL elements
828  *
829  *	Generates a gather list for the buffers that make up a packet.
830  *	The caller must provide adequate space for the SGL that will be written.
831  *	The SGL includes all of the packet's page fragments and the data in its
832  *	main body except for the first @start bytes.  @sgl must be 16-byte
833  *	aligned and within a Tx descriptor with available space.  @end points
834  *	right after the end of the SGL but does not account for any potential
835  *	wrap around, i.e., @end > @sgl.
836  */
837 void cxgb4_write_sgl(const struct sk_buff *skb, struct sge_txq *q,
838 		     struct ulptx_sgl *sgl, u64 *end, unsigned int start,
839 		     const dma_addr_t *addr)
840 {
841 	unsigned int i, len;
842 	struct ulptx_sge_pair *to;
843 	const struct skb_shared_info *si = skb_shinfo(skb);
844 	unsigned int nfrags = si->nr_frags;
845 	struct ulptx_sge_pair buf[MAX_SKB_FRAGS / 2 + 1];
846 
847 	len = skb_headlen(skb) - start;
848 	if (likely(len)) {
849 		sgl->len0 = htonl(len);
850 		sgl->addr0 = cpu_to_be64(addr[0] + start);
851 		nfrags++;
852 	} else {
853 		sgl->len0 = htonl(skb_frag_size(&si->frags[0]));
854 		sgl->addr0 = cpu_to_be64(addr[1]);
855 	}
856 
857 	sgl->cmd_nsge = htonl(ULPTX_CMD_V(ULP_TX_SC_DSGL) |
858 			      ULPTX_NSGE_V(nfrags));
859 	if (likely(--nfrags == 0))
860 		return;
861 	/*
862 	 * Most of the complexity below deals with the possibility we hit the
863 	 * end of the queue in the middle of writing the SGL.  For this case
864 	 * only we create the SGL in a temporary buffer and then copy it.
865 	 */
866 	to = (u8 *)end > (u8 *)q->stat ? buf : sgl->sge;
867 
868 	for (i = (nfrags != si->nr_frags); nfrags >= 2; nfrags -= 2, to++) {
869 		to->len[0] = cpu_to_be32(skb_frag_size(&si->frags[i]));
870 		to->len[1] = cpu_to_be32(skb_frag_size(&si->frags[++i]));
871 		to->addr[0] = cpu_to_be64(addr[i]);
872 		to->addr[1] = cpu_to_be64(addr[++i]);
873 	}
874 	if (nfrags) {
875 		to->len[0] = cpu_to_be32(skb_frag_size(&si->frags[i]));
876 		to->len[1] = cpu_to_be32(0);
877 		to->addr[0] = cpu_to_be64(addr[i + 1]);
878 	}
879 	if (unlikely((u8 *)end > (u8 *)q->stat)) {
880 		unsigned int part0 = (u8 *)q->stat - (u8 *)sgl->sge, part1;
881 
882 		if (likely(part0))
883 			memcpy(sgl->sge, buf, part0);
884 		part1 = (u8 *)end - (u8 *)q->stat;
885 		memcpy(q->desc, (u8 *)buf + part0, part1);
886 		end = (void *)q->desc + part1;
887 	}
888 	if ((uintptr_t)end & 8)           /* 0-pad to multiple of 16 */
889 		*end = 0;
890 }
891 EXPORT_SYMBOL(cxgb4_write_sgl);
892 
893 /*	cxgb4_write_partial_sgl - populate SGL for partial packet
894  *	@skb: the packet
895  *	@q: the Tx queue we are writing into
896  *	@sgl: starting location for writing the SGL
897  *	@end: points right after the end of the SGL
898  *	@addr: the list of bus addresses for the SGL elements
899  *	@start: start offset in the SKB where partial data starts
900  *	@len: length of data from @start to send out
901  *
902  *	This API will handle sending out partial data of a skb if required.
903  *	Unlike cxgb4_write_sgl, @start can be any offset into the skb data,
904  *	and @len will decide how much data after @start offset to send out.
905  */
906 void cxgb4_write_partial_sgl(const struct sk_buff *skb, struct sge_txq *q,
907 			     struct ulptx_sgl *sgl, u64 *end,
908 			     const dma_addr_t *addr, u32 start, u32 len)
909 {
910 	struct ulptx_sge_pair buf[MAX_SKB_FRAGS / 2 + 1] = {0}, *to;
911 	u32 frag_size, skb_linear_data_len = skb_headlen(skb);
912 	struct skb_shared_info *si = skb_shinfo(skb);
913 	u8 i = 0, frag_idx = 0, nfrags = 0;
914 	skb_frag_t *frag;
915 
916 	/* Fill the first SGL either from linear data or from partial
917 	 * frag based on @start.
918 	 */
919 	if (unlikely(start < skb_linear_data_len)) {
920 		frag_size = min(len, skb_linear_data_len - start);
921 		sgl->len0 = htonl(frag_size);
922 		sgl->addr0 = cpu_to_be64(addr[0] + start);
923 		len -= frag_size;
924 		nfrags++;
925 	} else {
926 		start -= skb_linear_data_len;
927 		frag = &si->frags[frag_idx];
928 		frag_size = skb_frag_size(frag);
929 		/* find the first frag */
930 		while (start >= frag_size) {
931 			start -= frag_size;
932 			frag_idx++;
933 			frag = &si->frags[frag_idx];
934 			frag_size = skb_frag_size(frag);
935 		}
936 
937 		frag_size = min(len, skb_frag_size(frag) - start);
938 		sgl->len0 = cpu_to_be32(frag_size);
939 		sgl->addr0 = cpu_to_be64(addr[frag_idx + 1] + start);
940 		len -= frag_size;
941 		nfrags++;
942 		frag_idx++;
943 	}
944 
945 	/* If the entire partial data fit in one SGL, then send it out
946 	 * now.
947 	 */
948 	if (!len)
949 		goto done;
950 
951 	/* Most of the complexity below deals with the possibility we hit the
952 	 * end of the queue in the middle of writing the SGL.  For this case
953 	 * only we create the SGL in a temporary buffer and then copy it.
954 	 */
955 	to = (u8 *)end > (u8 *)q->stat ? buf : sgl->sge;
956 
957 	/* If the skb couldn't fit in first SGL completely, fill the
958 	 * rest of the frags in subsequent SGLs. Note that each SGL
959 	 * pair can store 2 frags.
960 	 */
961 	while (len) {
962 		frag_size = min(len, skb_frag_size(&si->frags[frag_idx]));
963 		to->len[i & 1] = cpu_to_be32(frag_size);
964 		to->addr[i & 1] = cpu_to_be64(addr[frag_idx + 1]);
965 		if (i && (i & 1))
966 			to++;
967 		nfrags++;
968 		frag_idx++;
969 		i++;
970 		len -= frag_size;
971 	}
972 
973 	/* If we ended in an odd boundary, then set the second SGL's
974 	 * length in the pair to 0.
975 	 */
976 	if (i & 1)
977 		to->len[1] = cpu_to_be32(0);
978 
979 	/* Copy from temporary buffer to Tx ring, in case we hit the
980 	 * end of the queue in the middle of writing the SGL.
981 	 */
982 	if (unlikely((u8 *)end > (u8 *)q->stat)) {
983 		u32 part0 = (u8 *)q->stat - (u8 *)sgl->sge, part1;
984 
985 		if (likely(part0))
986 			memcpy(sgl->sge, buf, part0);
987 		part1 = (u8 *)end - (u8 *)q->stat;
988 		memcpy(q->desc, (u8 *)buf + part0, part1);
989 		end = (void *)q->desc + part1;
990 	}
991 
992 	/* 0-pad to multiple of 16 */
993 	if ((uintptr_t)end & 8)
994 		*end = 0;
995 done:
996 	sgl->cmd_nsge = htonl(ULPTX_CMD_V(ULP_TX_SC_DSGL) |
997 			ULPTX_NSGE_V(nfrags));
998 }
999 EXPORT_SYMBOL(cxgb4_write_partial_sgl);
1000 
1001 /* This function copies 64 byte coalesced work request to
1002  * memory mapped BAR2 space. For coalesced WR SGE fetches
1003  * data from the FIFO instead of from Host.
1004  */
1005 static void cxgb_pio_copy(u64 __iomem *dst, u64 *src)
1006 {
1007 	int count = 8;
1008 
1009 	while (count) {
1010 		writeq(*src, dst);
1011 		src++;
1012 		dst++;
1013 		count--;
1014 	}
1015 }
1016 
1017 /**
1018  *	cxgb4_ring_tx_db - check and potentially ring a Tx queue's doorbell
1019  *	@adap: the adapter
1020  *	@q: the Tx queue
1021  *	@n: number of new descriptors to give to HW
1022  *
1023  *	Ring the doorbel for a Tx queue.
1024  */
1025 inline void cxgb4_ring_tx_db(struct adapter *adap, struct sge_txq *q, int n)
1026 {
1027 	/* Make sure that all writes to the TX Descriptors are committed
1028 	 * before we tell the hardware about them.
1029 	 */
1030 	wmb();
1031 
1032 	/* If we don't have access to the new User Doorbell (T5+), use the old
1033 	 * doorbell mechanism; otherwise use the new BAR2 mechanism.
1034 	 */
1035 	if (unlikely(q->bar2_addr == NULL)) {
1036 		u32 val = PIDX_V(n);
1037 		unsigned long flags;
1038 
1039 		/* For T4 we need to participate in the Doorbell Recovery
1040 		 * mechanism.
1041 		 */
1042 		spin_lock_irqsave(&q->db_lock, flags);
1043 		if (!q->db_disabled)
1044 			t4_write_reg(adap, MYPF_REG(SGE_PF_KDOORBELL_A),
1045 				     QID_V(q->cntxt_id) | val);
1046 		else
1047 			q->db_pidx_inc += n;
1048 		q->db_pidx = q->pidx;
1049 		spin_unlock_irqrestore(&q->db_lock, flags);
1050 	} else {
1051 		u32 val = PIDX_T5_V(n);
1052 
1053 		/* T4 and later chips share the same PIDX field offset within
1054 		 * the doorbell, but T5 and later shrank the field in order to
1055 		 * gain a bit for Doorbell Priority.  The field was absurdly
1056 		 * large in the first place (14 bits) so we just use the T5
1057 		 * and later limits and warn if a Queue ID is too large.
1058 		 */
1059 		WARN_ON(val & DBPRIO_F);
1060 
1061 		/* If we're only writing a single TX Descriptor and we can use
1062 		 * Inferred QID registers, we can use the Write Combining
1063 		 * Gather Buffer; otherwise we use the simple doorbell.
1064 		 */
1065 		if (n == 1 && q->bar2_qid == 0) {
1066 			int index = (q->pidx
1067 				     ? (q->pidx - 1)
1068 				     : (q->size - 1));
1069 			u64 *wr = (u64 *)&q->desc[index];
1070 
1071 			cxgb_pio_copy((u64 __iomem *)
1072 				      (q->bar2_addr + SGE_UDB_WCDOORBELL),
1073 				      wr);
1074 		} else {
1075 			writel(val | QID_V(q->bar2_qid),
1076 			       q->bar2_addr + SGE_UDB_KDOORBELL);
1077 		}
1078 
1079 		/* This Write Memory Barrier will force the write to the User
1080 		 * Doorbell area to be flushed.  This is needed to prevent
1081 		 * writes on different CPUs for the same queue from hitting
1082 		 * the adapter out of order.  This is required when some Work
1083 		 * Requests take the Write Combine Gather Buffer path (user
1084 		 * doorbell area offset [SGE_UDB_WCDOORBELL..+63]) and some
1085 		 * take the traditional path where we simply increment the
1086 		 * PIDX (User Doorbell area SGE_UDB_KDOORBELL) and have the
1087 		 * hardware DMA read the actual Work Request.
1088 		 */
1089 		wmb();
1090 	}
1091 }
1092 EXPORT_SYMBOL(cxgb4_ring_tx_db);
1093 
1094 /**
1095  *	cxgb4_inline_tx_skb - inline a packet's data into Tx descriptors
1096  *	@skb: the packet
1097  *	@q: the Tx queue where the packet will be inlined
1098  *	@pos: starting position in the Tx queue where to inline the packet
1099  *
1100  *	Inline a packet's contents directly into Tx descriptors, starting at
1101  *	the given position within the Tx DMA ring.
1102  *	Most of the complexity of this operation is dealing with wrap arounds
1103  *	in the middle of the packet we want to inline.
1104  */
1105 void cxgb4_inline_tx_skb(const struct sk_buff *skb,
1106 			 const struct sge_txq *q, void *pos)
1107 {
1108 	int left = (void *)q->stat - pos;
1109 	u64 *p;
1110 
1111 	if (likely(skb->len <= left)) {
1112 		if (likely(!skb->data_len))
1113 			skb_copy_from_linear_data(skb, pos, skb->len);
1114 		else
1115 			skb_copy_bits(skb, 0, pos, skb->len);
1116 		pos += skb->len;
1117 	} else {
1118 		skb_copy_bits(skb, 0, pos, left);
1119 		skb_copy_bits(skb, left, q->desc, skb->len - left);
1120 		pos = (void *)q->desc + (skb->len - left);
1121 	}
1122 
1123 	/* 0-pad to multiple of 16 */
1124 	p = PTR_ALIGN(pos, 8);
1125 	if ((uintptr_t)p & 8)
1126 		*p = 0;
1127 }
1128 EXPORT_SYMBOL(cxgb4_inline_tx_skb);
1129 
1130 static void *inline_tx_skb_header(const struct sk_buff *skb,
1131 				  const struct sge_txq *q,  void *pos,
1132 				  int length)
1133 {
1134 	u64 *p;
1135 	int left = (void *)q->stat - pos;
1136 
1137 	if (likely(length <= left)) {
1138 		memcpy(pos, skb->data, length);
1139 		pos += length;
1140 	} else {
1141 		memcpy(pos, skb->data, left);
1142 		memcpy(q->desc, skb->data + left, length - left);
1143 		pos = (void *)q->desc + (length - left);
1144 	}
1145 	/* 0-pad to multiple of 16 */
1146 	p = PTR_ALIGN(pos, 8);
1147 	if ((uintptr_t)p & 8) {
1148 		*p = 0;
1149 		return p + 1;
1150 	}
1151 	return p;
1152 }
1153 
1154 /*
1155  * Figure out what HW csum a packet wants and return the appropriate control
1156  * bits.
1157  */
1158 static u64 hwcsum(enum chip_type chip, const struct sk_buff *skb)
1159 {
1160 	int csum_type;
1161 	bool inner_hdr_csum = false;
1162 	u16 proto, ver;
1163 
1164 	if (skb->encapsulation &&
1165 	    (CHELSIO_CHIP_VERSION(chip) > CHELSIO_T5))
1166 		inner_hdr_csum = true;
1167 
1168 	if (inner_hdr_csum) {
1169 		ver = inner_ip_hdr(skb)->version;
1170 		proto = (ver == 4) ? inner_ip_hdr(skb)->protocol :
1171 			inner_ipv6_hdr(skb)->nexthdr;
1172 	} else {
1173 		ver = ip_hdr(skb)->version;
1174 		proto = (ver == 4) ? ip_hdr(skb)->protocol :
1175 			ipv6_hdr(skb)->nexthdr;
1176 	}
1177 
1178 	if (ver == 4) {
1179 		if (proto == IPPROTO_TCP)
1180 			csum_type = TX_CSUM_TCPIP;
1181 		else if (proto == IPPROTO_UDP)
1182 			csum_type = TX_CSUM_UDPIP;
1183 		else {
1184 nocsum:			/*
1185 			 * unknown protocol, disable HW csum
1186 			 * and hope a bad packet is detected
1187 			 */
1188 			return TXPKT_L4CSUM_DIS_F;
1189 		}
1190 	} else {
1191 		/*
1192 		 * this doesn't work with extension headers
1193 		 */
1194 		if (proto == IPPROTO_TCP)
1195 			csum_type = TX_CSUM_TCPIP6;
1196 		else if (proto == IPPROTO_UDP)
1197 			csum_type = TX_CSUM_UDPIP6;
1198 		else
1199 			goto nocsum;
1200 	}
1201 
1202 	if (likely(csum_type >= TX_CSUM_TCPIP)) {
1203 		int eth_hdr_len, l4_len;
1204 		u64 hdr_len;
1205 
1206 		if (inner_hdr_csum) {
1207 			/* This allows checksum offload for all encapsulated
1208 			 * packets like GRE etc..
1209 			 */
1210 			l4_len = skb_inner_network_header_len(skb);
1211 			eth_hdr_len = skb_inner_network_offset(skb) - ETH_HLEN;
1212 		} else {
1213 			l4_len = skb_network_header_len(skb);
1214 			eth_hdr_len = skb_network_offset(skb) - ETH_HLEN;
1215 		}
1216 		hdr_len = TXPKT_IPHDR_LEN_V(l4_len);
1217 
1218 		if (CHELSIO_CHIP_VERSION(chip) <= CHELSIO_T5)
1219 			hdr_len |= TXPKT_ETHHDR_LEN_V(eth_hdr_len);
1220 		else
1221 			hdr_len |= T6_TXPKT_ETHHDR_LEN_V(eth_hdr_len);
1222 		return TXPKT_CSUM_TYPE_V(csum_type) | hdr_len;
1223 	} else {
1224 		int start = skb_transport_offset(skb);
1225 
1226 		return TXPKT_CSUM_TYPE_V(csum_type) |
1227 			TXPKT_CSUM_START_V(start) |
1228 			TXPKT_CSUM_LOC_V(start + skb->csum_offset);
1229 	}
1230 }
1231 
1232 static void eth_txq_stop(struct sge_eth_txq *q)
1233 {
1234 	netif_tx_stop_queue(q->txq);
1235 	q->q.stops++;
1236 }
1237 
1238 static inline void txq_advance(struct sge_txq *q, unsigned int n)
1239 {
1240 	q->in_use += n;
1241 	q->pidx += n;
1242 	if (q->pidx >= q->size)
1243 		q->pidx -= q->size;
1244 }
1245 
1246 #ifdef CONFIG_CHELSIO_T4_FCOE
1247 static inline int
1248 cxgb_fcoe_offload(struct sk_buff *skb, struct adapter *adap,
1249 		  const struct port_info *pi, u64 *cntrl)
1250 {
1251 	const struct cxgb_fcoe *fcoe = &pi->fcoe;
1252 
1253 	if (!(fcoe->flags & CXGB_FCOE_ENABLED))
1254 		return 0;
1255 
1256 	if (skb->protocol != htons(ETH_P_FCOE))
1257 		return 0;
1258 
1259 	skb_reset_mac_header(skb);
1260 	skb->mac_len = sizeof(struct ethhdr);
1261 
1262 	skb_set_network_header(skb, skb->mac_len);
1263 	skb_set_transport_header(skb, skb->mac_len + sizeof(struct fcoe_hdr));
1264 
1265 	if (!cxgb_fcoe_sof_eof_supported(adap, skb))
1266 		return -ENOTSUPP;
1267 
1268 	/* FC CRC offload */
1269 	*cntrl = TXPKT_CSUM_TYPE_V(TX_CSUM_FCOE) |
1270 		     TXPKT_L4CSUM_DIS_F | TXPKT_IPCSUM_DIS_F |
1271 		     TXPKT_CSUM_START_V(CXGB_FCOE_TXPKT_CSUM_START) |
1272 		     TXPKT_CSUM_END_V(CXGB_FCOE_TXPKT_CSUM_END) |
1273 		     TXPKT_CSUM_LOC_V(CXGB_FCOE_TXPKT_CSUM_END);
1274 	return 0;
1275 }
1276 #endif /* CONFIG_CHELSIO_T4_FCOE */
1277 
1278 /* Returns tunnel type if hardware supports offloading of the same.
1279  * It is called only for T5 and onwards.
1280  */
1281 enum cpl_tx_tnl_lso_type cxgb_encap_offload_supported(struct sk_buff *skb)
1282 {
1283 	u8 l4_hdr = 0;
1284 	enum cpl_tx_tnl_lso_type tnl_type = TX_TNL_TYPE_OPAQUE;
1285 	struct port_info *pi = netdev_priv(skb->dev);
1286 	struct adapter *adapter = pi->adapter;
1287 
1288 	if (skb->inner_protocol_type != ENCAP_TYPE_ETHER ||
1289 	    skb->inner_protocol != htons(ETH_P_TEB))
1290 		return tnl_type;
1291 
1292 	switch (vlan_get_protocol(skb)) {
1293 	case htons(ETH_P_IP):
1294 		l4_hdr = ip_hdr(skb)->protocol;
1295 		break;
1296 	case htons(ETH_P_IPV6):
1297 		l4_hdr = ipv6_hdr(skb)->nexthdr;
1298 		break;
1299 	default:
1300 		return tnl_type;
1301 	}
1302 
1303 	switch (l4_hdr) {
1304 	case IPPROTO_UDP:
1305 		if (adapter->vxlan_port == udp_hdr(skb)->dest)
1306 			tnl_type = TX_TNL_TYPE_VXLAN;
1307 		else if (adapter->geneve_port == udp_hdr(skb)->dest)
1308 			tnl_type = TX_TNL_TYPE_GENEVE;
1309 		break;
1310 	default:
1311 		return tnl_type;
1312 	}
1313 
1314 	return tnl_type;
1315 }
1316 
1317 static inline void t6_fill_tnl_lso(struct sk_buff *skb,
1318 				   struct cpl_tx_tnl_lso *tnl_lso,
1319 				   enum cpl_tx_tnl_lso_type tnl_type)
1320 {
1321 	u32 val;
1322 	int in_eth_xtra_len;
1323 	int l3hdr_len = skb_network_header_len(skb);
1324 	int eth_xtra_len = skb_network_offset(skb) - ETH_HLEN;
1325 	const struct skb_shared_info *ssi = skb_shinfo(skb);
1326 	bool v6 = (ip_hdr(skb)->version == 6);
1327 
1328 	val = CPL_TX_TNL_LSO_OPCODE_V(CPL_TX_TNL_LSO) |
1329 	      CPL_TX_TNL_LSO_FIRST_F |
1330 	      CPL_TX_TNL_LSO_LAST_F |
1331 	      (v6 ? CPL_TX_TNL_LSO_IPV6OUT_F : 0) |
1332 	      CPL_TX_TNL_LSO_ETHHDRLENOUT_V(eth_xtra_len / 4) |
1333 	      CPL_TX_TNL_LSO_IPHDRLENOUT_V(l3hdr_len / 4) |
1334 	      (v6 ? 0 : CPL_TX_TNL_LSO_IPHDRCHKOUT_F) |
1335 	      CPL_TX_TNL_LSO_IPLENSETOUT_F |
1336 	      (v6 ? 0 : CPL_TX_TNL_LSO_IPIDINCOUT_F);
1337 	tnl_lso->op_to_IpIdSplitOut = htonl(val);
1338 
1339 	tnl_lso->IpIdOffsetOut = 0;
1340 
1341 	/* Get the tunnel header length */
1342 	val = skb_inner_mac_header(skb) - skb_mac_header(skb);
1343 	in_eth_xtra_len = skb_inner_network_header(skb) -
1344 			  skb_inner_mac_header(skb) - ETH_HLEN;
1345 
1346 	switch (tnl_type) {
1347 	case TX_TNL_TYPE_VXLAN:
1348 	case TX_TNL_TYPE_GENEVE:
1349 		tnl_lso->UdpLenSetOut_to_TnlHdrLen =
1350 			htons(CPL_TX_TNL_LSO_UDPCHKCLROUT_F |
1351 			CPL_TX_TNL_LSO_UDPLENSETOUT_F);
1352 		break;
1353 	default:
1354 		tnl_lso->UdpLenSetOut_to_TnlHdrLen = 0;
1355 		break;
1356 	}
1357 
1358 	tnl_lso->UdpLenSetOut_to_TnlHdrLen |=
1359 		 htons(CPL_TX_TNL_LSO_TNLHDRLEN_V(val) |
1360 		       CPL_TX_TNL_LSO_TNLTYPE_V(tnl_type));
1361 
1362 	tnl_lso->r1 = 0;
1363 
1364 	val = CPL_TX_TNL_LSO_ETHHDRLEN_V(in_eth_xtra_len / 4) |
1365 	      CPL_TX_TNL_LSO_IPV6_V(inner_ip_hdr(skb)->version == 6) |
1366 	      CPL_TX_TNL_LSO_IPHDRLEN_V(skb_inner_network_header_len(skb) / 4) |
1367 	      CPL_TX_TNL_LSO_TCPHDRLEN_V(inner_tcp_hdrlen(skb) / 4);
1368 	tnl_lso->Flow_to_TcpHdrLen = htonl(val);
1369 
1370 	tnl_lso->IpIdOffset = htons(0);
1371 
1372 	tnl_lso->IpIdSplit_to_Mss = htons(CPL_TX_TNL_LSO_MSS_V(ssi->gso_size));
1373 	tnl_lso->TCPSeqOffset = htonl(0);
1374 	tnl_lso->EthLenOffset_Size = htonl(CPL_TX_TNL_LSO_SIZE_V(skb->len));
1375 }
1376 
1377 static inline void *write_tso_wr(struct adapter *adap, struct sk_buff *skb,
1378 				 struct cpl_tx_pkt_lso_core *lso)
1379 {
1380 	int eth_xtra_len = skb_network_offset(skb) - ETH_HLEN;
1381 	int l3hdr_len = skb_network_header_len(skb);
1382 	const struct skb_shared_info *ssi;
1383 	bool ipv6 = false;
1384 
1385 	ssi = skb_shinfo(skb);
1386 	if (ssi->gso_type & SKB_GSO_TCPV6)
1387 		ipv6 = true;
1388 
1389 	lso->lso_ctrl = htonl(LSO_OPCODE_V(CPL_TX_PKT_LSO) |
1390 			      LSO_FIRST_SLICE_F | LSO_LAST_SLICE_F |
1391 			      LSO_IPV6_V(ipv6) |
1392 			      LSO_ETHHDR_LEN_V(eth_xtra_len / 4) |
1393 			      LSO_IPHDR_LEN_V(l3hdr_len / 4) |
1394 			      LSO_TCPHDR_LEN_V(tcp_hdr(skb)->doff));
1395 	lso->ipid_ofst = htons(0);
1396 	lso->mss = htons(ssi->gso_size);
1397 	lso->seqno_offset = htonl(0);
1398 	if (is_t4(adap->params.chip))
1399 		lso->len = htonl(skb->len);
1400 	else
1401 		lso->len = htonl(LSO_T5_XFER_SIZE_V(skb->len));
1402 
1403 	return (void *)(lso + 1);
1404 }
1405 
1406 /**
1407  *	t4_sge_eth_txq_egress_update - handle Ethernet TX Queue update
1408  *	@adap: the adapter
1409  *	@eq: the Ethernet TX Queue
1410  *	@maxreclaim: the maximum number of TX Descriptors to reclaim or -1
1411  *
1412  *	We're typically called here to update the state of an Ethernet TX
1413  *	Queue with respect to the hardware's progress in consuming the TX
1414  *	Work Requests that we've put on that Egress Queue.  This happens
1415  *	when we get Egress Queue Update messages and also prophylactically
1416  *	in regular timer-based Ethernet TX Queue maintenance.
1417  */
1418 int t4_sge_eth_txq_egress_update(struct adapter *adap, struct sge_eth_txq *eq,
1419 				 int maxreclaim)
1420 {
1421 	unsigned int reclaimed, hw_cidx;
1422 	struct sge_txq *q = &eq->q;
1423 	int hw_in_use;
1424 
1425 	if (!q->in_use || !__netif_tx_trylock(eq->txq))
1426 		return 0;
1427 
1428 	/* Reclaim pending completed TX Descriptors. */
1429 	reclaimed = reclaim_completed_tx(adap, &eq->q, maxreclaim, true);
1430 
1431 	hw_cidx = ntohs(READ_ONCE(q->stat->cidx));
1432 	hw_in_use = q->pidx - hw_cidx;
1433 	if (hw_in_use < 0)
1434 		hw_in_use += q->size;
1435 
1436 	/* If the TX Queue is currently stopped and there's now more than half
1437 	 * the queue available, restart it.  Otherwise bail out since the rest
1438 	 * of what we want do here is with the possibility of shipping any
1439 	 * currently buffered Coalesced TX Work Request.
1440 	 */
1441 	if (netif_tx_queue_stopped(eq->txq) && hw_in_use < (q->size / 2)) {
1442 		netif_tx_wake_queue(eq->txq);
1443 		eq->q.restarts++;
1444 	}
1445 
1446 	__netif_tx_unlock(eq->txq);
1447 	return reclaimed;
1448 }
1449 
1450 static inline int cxgb4_validate_skb(struct sk_buff *skb,
1451 				     struct net_device *dev,
1452 				     u32 min_pkt_len)
1453 {
1454 	u32 max_pkt_len;
1455 
1456 	/* The chip min packet length is 10 octets but some firmware
1457 	 * commands have a minimum packet length requirement. So, play
1458 	 * safe and reject anything shorter than @min_pkt_len.
1459 	 */
1460 	if (unlikely(skb->len < min_pkt_len))
1461 		return -EINVAL;
1462 
1463 	/* Discard the packet if the length is greater than mtu */
1464 	max_pkt_len = ETH_HLEN + dev->mtu;
1465 
1466 	if (skb_vlan_tagged(skb))
1467 		max_pkt_len += VLAN_HLEN;
1468 
1469 	if (!skb_shinfo(skb)->gso_size && (unlikely(skb->len > max_pkt_len)))
1470 		return -EINVAL;
1471 
1472 	return 0;
1473 }
1474 
1475 static void *write_eo_udp_wr(struct sk_buff *skb, struct fw_eth_tx_eo_wr *wr,
1476 			     u32 hdr_len)
1477 {
1478 	wr->u.udpseg.type = FW_ETH_TX_EO_TYPE_UDPSEG;
1479 	wr->u.udpseg.ethlen = skb_network_offset(skb);
1480 	wr->u.udpseg.iplen = cpu_to_be16(skb_network_header_len(skb));
1481 	wr->u.udpseg.udplen = sizeof(struct udphdr);
1482 	wr->u.udpseg.rtplen = 0;
1483 	wr->u.udpseg.r4 = 0;
1484 	if (skb_shinfo(skb)->gso_size)
1485 		wr->u.udpseg.mss = cpu_to_be16(skb_shinfo(skb)->gso_size);
1486 	else
1487 		wr->u.udpseg.mss = cpu_to_be16(skb->len - hdr_len);
1488 	wr->u.udpseg.schedpktsize = wr->u.udpseg.mss;
1489 	wr->u.udpseg.plen = cpu_to_be32(skb->len - hdr_len);
1490 
1491 	return (void *)(wr + 1);
1492 }
1493 
1494 /**
1495  *	cxgb4_eth_xmit - add a packet to an Ethernet Tx queue
1496  *	@skb: the packet
1497  *	@dev: the egress net device
1498  *
1499  *	Add a packet to an SGE Ethernet Tx queue.  Runs with softirqs disabled.
1500  */
1501 static netdev_tx_t cxgb4_eth_xmit(struct sk_buff *skb, struct net_device *dev)
1502 {
1503 	enum cpl_tx_tnl_lso_type tnl_type = TX_TNL_TYPE_OPAQUE;
1504 	bool ptp_enabled = is_ptp_enabled(skb, dev);
1505 	unsigned int last_desc, flits, ndesc;
1506 	u32 wr_mid, ctrl0, op, sgl_off = 0;
1507 	const struct skb_shared_info *ssi;
1508 	int len, qidx, credits, ret, left;
1509 	struct tx_sw_desc *sgl_sdesc;
1510 	struct fw_eth_tx_eo_wr *eowr;
1511 	struct fw_eth_tx_pkt_wr *wr;
1512 	struct cpl_tx_pkt_core *cpl;
1513 	const struct port_info *pi;
1514 	bool immediate = false;
1515 	u64 cntrl, *end, *sgl;
1516 	struct sge_eth_txq *q;
1517 	unsigned int chip_ver;
1518 	struct adapter *adap;
1519 
1520 	ret = cxgb4_validate_skb(skb, dev, ETH_HLEN);
1521 	if (ret)
1522 		goto out_free;
1523 
1524 	pi = netdev_priv(dev);
1525 	adap = pi->adapter;
1526 	ssi = skb_shinfo(skb);
1527 #if IS_ENABLED(CONFIG_CHELSIO_IPSEC_INLINE)
1528 	if (xfrm_offload(skb) && !ssi->gso_size)
1529 		return adap->uld[CXGB4_ULD_IPSEC].tx_handler(skb, dev);
1530 #endif /* CHELSIO_IPSEC_INLINE */
1531 
1532 #if IS_ENABLED(CONFIG_CHELSIO_TLS_DEVICE)
1533 	if (cxgb4_is_ktls_skb(skb) &&
1534 	    (skb->len - skb_tcp_all_headers(skb)))
1535 		return adap->uld[CXGB4_ULD_KTLS].tx_handler(skb, dev);
1536 #endif /* CHELSIO_TLS_DEVICE */
1537 
1538 	qidx = skb_get_queue_mapping(skb);
1539 	if (ptp_enabled) {
1540 		if (!(adap->ptp_tx_skb)) {
1541 			skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
1542 			adap->ptp_tx_skb = skb_get(skb);
1543 		} else {
1544 			goto out_free;
1545 		}
1546 		q = &adap->sge.ptptxq;
1547 	} else {
1548 		q = &adap->sge.ethtxq[qidx + pi->first_qset];
1549 	}
1550 	skb_tx_timestamp(skb);
1551 
1552 	reclaim_completed_tx(adap, &q->q, -1, true);
1553 	cntrl = TXPKT_L4CSUM_DIS_F | TXPKT_IPCSUM_DIS_F;
1554 
1555 #ifdef CONFIG_CHELSIO_T4_FCOE
1556 	ret = cxgb_fcoe_offload(skb, adap, pi, &cntrl);
1557 	if (unlikely(ret == -EOPNOTSUPP))
1558 		goto out_free;
1559 #endif /* CONFIG_CHELSIO_T4_FCOE */
1560 
1561 	chip_ver = CHELSIO_CHIP_VERSION(adap->params.chip);
1562 	flits = calc_tx_flits(skb, chip_ver);
1563 	ndesc = flits_to_desc(flits);
1564 	credits = txq_avail(&q->q) - ndesc;
1565 
1566 	if (unlikely(credits < 0)) {
1567 		eth_txq_stop(q);
1568 		dev_err(adap->pdev_dev,
1569 			"%s: Tx ring %u full while queue awake!\n",
1570 			dev->name, qidx);
1571 		return NETDEV_TX_BUSY;
1572 	}
1573 
1574 	if (is_eth_imm(skb, chip_ver))
1575 		immediate = true;
1576 
1577 	if (skb->encapsulation && chip_ver > CHELSIO_T5)
1578 		tnl_type = cxgb_encap_offload_supported(skb);
1579 
1580 	last_desc = q->q.pidx + ndesc - 1;
1581 	if (last_desc >= q->q.size)
1582 		last_desc -= q->q.size;
1583 	sgl_sdesc = &q->q.sdesc[last_desc];
1584 
1585 	if (!immediate &&
1586 	    unlikely(cxgb4_map_skb(adap->pdev_dev, skb, sgl_sdesc->addr) < 0)) {
1587 		memset(sgl_sdesc->addr, 0, sizeof(sgl_sdesc->addr));
1588 		q->mapping_err++;
1589 		goto out_free;
1590 	}
1591 
1592 	wr_mid = FW_WR_LEN16_V(DIV_ROUND_UP(flits, 2));
1593 	if (unlikely(credits < ETHTXQ_STOP_THRES)) {
1594 		/* After we're done injecting the Work Request for this
1595 		 * packet, we'll be below our "stop threshold" so stop the TX
1596 		 * Queue now and schedule a request for an SGE Egress Queue
1597 		 * Update message. The queue will get started later on when
1598 		 * the firmware processes this Work Request and sends us an
1599 		 * Egress Queue Status Update message indicating that space
1600 		 * has opened up.
1601 		 */
1602 		eth_txq_stop(q);
1603 		if (chip_ver > CHELSIO_T5)
1604 			wr_mid |= FW_WR_EQUEQ_F | FW_WR_EQUIQ_F;
1605 	}
1606 
1607 	wr = (void *)&q->q.desc[q->q.pidx];
1608 	eowr = (void *)&q->q.desc[q->q.pidx];
1609 	wr->equiq_to_len16 = htonl(wr_mid);
1610 	wr->r3 = cpu_to_be64(0);
1611 	if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4)
1612 		end = (u64 *)eowr + flits;
1613 	else
1614 		end = (u64 *)wr + flits;
1615 
1616 	len = immediate ? skb->len : 0;
1617 	len += sizeof(*cpl);
1618 	if (ssi->gso_size && !(ssi->gso_type & SKB_GSO_UDP_L4)) {
1619 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
1620 		struct cpl_tx_tnl_lso *tnl_lso = (void *)(wr + 1);
1621 
1622 		if (tnl_type)
1623 			len += sizeof(*tnl_lso);
1624 		else
1625 			len += sizeof(*lso);
1626 
1627 		wr->op_immdlen = htonl(FW_WR_OP_V(FW_ETH_TX_PKT_WR) |
1628 				       FW_WR_IMMDLEN_V(len));
1629 		if (tnl_type) {
1630 			struct iphdr *iph = ip_hdr(skb);
1631 
1632 			t6_fill_tnl_lso(skb, tnl_lso, tnl_type);
1633 			cpl = (void *)(tnl_lso + 1);
1634 			/* Driver is expected to compute partial checksum that
1635 			 * does not include the IP Total Length.
1636 			 */
1637 			if (iph->version == 4) {
1638 				iph->check = 0;
1639 				iph->tot_len = 0;
1640 				iph->check = ~ip_fast_csum((u8 *)iph, iph->ihl);
1641 			}
1642 			if (skb->ip_summed == CHECKSUM_PARTIAL)
1643 				cntrl = hwcsum(adap->params.chip, skb);
1644 		} else {
1645 			cpl = write_tso_wr(adap, skb, lso);
1646 			cntrl = hwcsum(adap->params.chip, skb);
1647 		}
1648 		sgl = (u64 *)(cpl + 1); /* sgl start here */
1649 		q->tso++;
1650 		q->tx_cso += ssi->gso_segs;
1651 	} else if (ssi->gso_size) {
1652 		u64 *start;
1653 		u32 hdrlen;
1654 
1655 		hdrlen = eth_get_headlen(dev, skb->data, skb_headlen(skb));
1656 		len += hdrlen;
1657 		wr->op_immdlen = cpu_to_be32(FW_WR_OP_V(FW_ETH_TX_EO_WR) |
1658 					     FW_ETH_TX_EO_WR_IMMDLEN_V(len));
1659 		cpl = write_eo_udp_wr(skb, eowr, hdrlen);
1660 		cntrl = hwcsum(adap->params.chip, skb);
1661 
1662 		start = (u64 *)(cpl + 1);
1663 		sgl = (u64 *)inline_tx_skb_header(skb, &q->q, (void *)start,
1664 						  hdrlen);
1665 		if (unlikely(start > sgl)) {
1666 			left = (u8 *)end - (u8 *)q->q.stat;
1667 			end = (void *)q->q.desc + left;
1668 		}
1669 		sgl_off = hdrlen;
1670 		q->uso++;
1671 		q->tx_cso += ssi->gso_segs;
1672 	} else {
1673 		if (ptp_enabled)
1674 			op = FW_PTP_TX_PKT_WR;
1675 		else
1676 			op = FW_ETH_TX_PKT_WR;
1677 		wr->op_immdlen = htonl(FW_WR_OP_V(op) |
1678 				       FW_WR_IMMDLEN_V(len));
1679 		cpl = (void *)(wr + 1);
1680 		sgl = (u64 *)(cpl + 1);
1681 		if (skb->ip_summed == CHECKSUM_PARTIAL) {
1682 			cntrl = hwcsum(adap->params.chip, skb) |
1683 				TXPKT_IPCSUM_DIS_F;
1684 			q->tx_cso++;
1685 		}
1686 	}
1687 
1688 	if (unlikely((u8 *)sgl >= (u8 *)q->q.stat)) {
1689 		/* If current position is already at the end of the
1690 		 * txq, reset the current to point to start of the queue
1691 		 * and update the end ptr as well.
1692 		 */
1693 		left = (u8 *)end - (u8 *)q->q.stat;
1694 		end = (void *)q->q.desc + left;
1695 		sgl = (void *)q->q.desc;
1696 	}
1697 
1698 	if (skb_vlan_tag_present(skb)) {
1699 		q->vlan_ins++;
1700 		cntrl |= TXPKT_VLAN_VLD_F | TXPKT_VLAN_V(skb_vlan_tag_get(skb));
1701 #ifdef CONFIG_CHELSIO_T4_FCOE
1702 		if (skb->protocol == htons(ETH_P_FCOE))
1703 			cntrl |= TXPKT_VLAN_V(
1704 				 ((skb->priority & 0x7) << VLAN_PRIO_SHIFT));
1705 #endif /* CONFIG_CHELSIO_T4_FCOE */
1706 	}
1707 
1708 	ctrl0 = TXPKT_OPCODE_V(CPL_TX_PKT_XT) | TXPKT_INTF_V(pi->tx_chan) |
1709 		TXPKT_PF_V(adap->pf);
1710 	if (ptp_enabled)
1711 		ctrl0 |= TXPKT_TSTAMP_F;
1712 #ifdef CONFIG_CHELSIO_T4_DCB
1713 	if (is_t4(adap->params.chip))
1714 		ctrl0 |= TXPKT_OVLAN_IDX_V(q->dcb_prio);
1715 	else
1716 		ctrl0 |= TXPKT_T5_OVLAN_IDX_V(q->dcb_prio);
1717 #endif
1718 	cpl->ctrl0 = htonl(ctrl0);
1719 	cpl->pack = htons(0);
1720 	cpl->len = htons(skb->len);
1721 	cpl->ctrl1 = cpu_to_be64(cntrl);
1722 
1723 	if (immediate) {
1724 		cxgb4_inline_tx_skb(skb, &q->q, sgl);
1725 		dev_consume_skb_any(skb);
1726 	} else {
1727 		cxgb4_write_sgl(skb, &q->q, (void *)sgl, end, sgl_off,
1728 				sgl_sdesc->addr);
1729 		skb_orphan(skb);
1730 		sgl_sdesc->skb = skb;
1731 	}
1732 
1733 	txq_advance(&q->q, ndesc);
1734 
1735 	cxgb4_ring_tx_db(adap, &q->q, ndesc);
1736 	return NETDEV_TX_OK;
1737 
1738 out_free:
1739 	dev_kfree_skb_any(skb);
1740 	return NETDEV_TX_OK;
1741 }
1742 
1743 /* Constants ... */
1744 enum {
1745 	/* Egress Queue sizes, producer and consumer indices are all in units
1746 	 * of Egress Context Units bytes.  Note that as far as the hardware is
1747 	 * concerned, the free list is an Egress Queue (the host produces free
1748 	 * buffers which the hardware consumes) and free list entries are
1749 	 * 64-bit PCI DMA addresses.
1750 	 */
1751 	EQ_UNIT = SGE_EQ_IDXSIZE,
1752 	FL_PER_EQ_UNIT = EQ_UNIT / sizeof(__be64),
1753 	TXD_PER_EQ_UNIT = EQ_UNIT / sizeof(__be64),
1754 
1755 	T4VF_ETHTXQ_MAX_HDR = (sizeof(struct fw_eth_tx_pkt_vm_wr) +
1756 			       sizeof(struct cpl_tx_pkt_lso_core) +
1757 			       sizeof(struct cpl_tx_pkt_core)) / sizeof(__be64),
1758 };
1759 
1760 /**
1761  *	t4vf_is_eth_imm - can an Ethernet packet be sent as immediate data?
1762  *	@skb: the packet
1763  *
1764  *	Returns whether an Ethernet packet is small enough to fit completely as
1765  *	immediate data.
1766  */
1767 static inline int t4vf_is_eth_imm(const struct sk_buff *skb)
1768 {
1769 	/* The VF Driver uses the FW_ETH_TX_PKT_VM_WR firmware Work Request
1770 	 * which does not accommodate immediate data.  We could dike out all
1771 	 * of the support code for immediate data but that would tie our hands
1772 	 * too much if we ever want to enhace the firmware.  It would also
1773 	 * create more differences between the PF and VF Drivers.
1774 	 */
1775 	return false;
1776 }
1777 
1778 /**
1779  *	t4vf_calc_tx_flits - calculate the number of flits for a packet TX WR
1780  *	@skb: the packet
1781  *
1782  *	Returns the number of flits needed for a TX Work Request for the
1783  *	given Ethernet packet, including the needed WR and CPL headers.
1784  */
1785 static inline unsigned int t4vf_calc_tx_flits(const struct sk_buff *skb)
1786 {
1787 	unsigned int flits;
1788 
1789 	/* If the skb is small enough, we can pump it out as a work request
1790 	 * with only immediate data.  In that case we just have to have the
1791 	 * TX Packet header plus the skb data in the Work Request.
1792 	 */
1793 	if (t4vf_is_eth_imm(skb))
1794 		return DIV_ROUND_UP(skb->len + sizeof(struct cpl_tx_pkt),
1795 				    sizeof(__be64));
1796 
1797 	/* Otherwise, we're going to have to construct a Scatter gather list
1798 	 * of the skb body and fragments.  We also include the flits necessary
1799 	 * for the TX Packet Work Request and CPL.  We always have a firmware
1800 	 * Write Header (incorporated as part of the cpl_tx_pkt_lso and
1801 	 * cpl_tx_pkt structures), followed by either a TX Packet Write CPL
1802 	 * message or, if we're doing a Large Send Offload, an LSO CPL message
1803 	 * with an embedded TX Packet Write CPL message.
1804 	 */
1805 	flits = sgl_len(skb_shinfo(skb)->nr_frags + 1);
1806 	if (skb_shinfo(skb)->gso_size)
1807 		flits += (sizeof(struct fw_eth_tx_pkt_vm_wr) +
1808 			  sizeof(struct cpl_tx_pkt_lso_core) +
1809 			  sizeof(struct cpl_tx_pkt_core)) / sizeof(__be64);
1810 	else
1811 		flits += (sizeof(struct fw_eth_tx_pkt_vm_wr) +
1812 			  sizeof(struct cpl_tx_pkt_core)) / sizeof(__be64);
1813 	return flits;
1814 }
1815 
1816 /**
1817  *	cxgb4_vf_eth_xmit - add a packet to an Ethernet TX queue
1818  *	@skb: the packet
1819  *	@dev: the egress net device
1820  *
1821  *	Add a packet to an SGE Ethernet TX queue.  Runs with softirqs disabled.
1822  */
1823 static netdev_tx_t cxgb4_vf_eth_xmit(struct sk_buff *skb,
1824 				     struct net_device *dev)
1825 {
1826 	unsigned int last_desc, flits, ndesc;
1827 	const struct skb_shared_info *ssi;
1828 	struct fw_eth_tx_pkt_vm_wr *wr;
1829 	struct tx_sw_desc *sgl_sdesc;
1830 	struct cpl_tx_pkt_core *cpl;
1831 	const struct port_info *pi;
1832 	struct sge_eth_txq *txq;
1833 	struct adapter *adapter;
1834 	int qidx, credits, ret;
1835 	size_t fw_hdr_copy_len;
1836 	unsigned int chip_ver;
1837 	u64 cntrl, *end;
1838 	u32 wr_mid;
1839 
1840 	/* The chip minimum packet length is 10 octets but the firmware
1841 	 * command that we are using requires that we copy the Ethernet header
1842 	 * (including the VLAN tag) into the header so we reject anything
1843 	 * smaller than that ...
1844 	 */
1845 	BUILD_BUG_ON(sizeof(wr->firmware) !=
1846 		     (sizeof(wr->ethmacdst) + sizeof(wr->ethmacsrc) +
1847 		      sizeof(wr->ethtype) + sizeof(wr->vlantci)));
1848 	fw_hdr_copy_len = sizeof(wr->firmware);
1849 	ret = cxgb4_validate_skb(skb, dev, fw_hdr_copy_len);
1850 	if (ret)
1851 		goto out_free;
1852 
1853 	/* Figure out which TX Queue we're going to use. */
1854 	pi = netdev_priv(dev);
1855 	adapter = pi->adapter;
1856 	qidx = skb_get_queue_mapping(skb);
1857 	WARN_ON(qidx >= pi->nqsets);
1858 	txq = &adapter->sge.ethtxq[pi->first_qset + qidx];
1859 
1860 	/* Take this opportunity to reclaim any TX Descriptors whose DMA
1861 	 * transfers have completed.
1862 	 */
1863 	reclaim_completed_tx(adapter, &txq->q, -1, true);
1864 
1865 	/* Calculate the number of flits and TX Descriptors we're going to
1866 	 * need along with how many TX Descriptors will be left over after
1867 	 * we inject our Work Request.
1868 	 */
1869 	flits = t4vf_calc_tx_flits(skb);
1870 	ndesc = flits_to_desc(flits);
1871 	credits = txq_avail(&txq->q) - ndesc;
1872 
1873 	if (unlikely(credits < 0)) {
1874 		/* Not enough room for this packet's Work Request.  Stop the
1875 		 * TX Queue and return a "busy" condition.  The queue will get
1876 		 * started later on when the firmware informs us that space
1877 		 * has opened up.
1878 		 */
1879 		eth_txq_stop(txq);
1880 		dev_err(adapter->pdev_dev,
1881 			"%s: TX ring %u full while queue awake!\n",
1882 			dev->name, qidx);
1883 		return NETDEV_TX_BUSY;
1884 	}
1885 
1886 	last_desc = txq->q.pidx + ndesc - 1;
1887 	if (last_desc >= txq->q.size)
1888 		last_desc -= txq->q.size;
1889 	sgl_sdesc = &txq->q.sdesc[last_desc];
1890 
1891 	if (!t4vf_is_eth_imm(skb) &&
1892 	    unlikely(cxgb4_map_skb(adapter->pdev_dev, skb,
1893 				   sgl_sdesc->addr) < 0)) {
1894 		/* We need to map the skb into PCI DMA space (because it can't
1895 		 * be in-lined directly into the Work Request) and the mapping
1896 		 * operation failed.  Record the error and drop the packet.
1897 		 */
1898 		memset(sgl_sdesc->addr, 0, sizeof(sgl_sdesc->addr));
1899 		txq->mapping_err++;
1900 		goto out_free;
1901 	}
1902 
1903 	chip_ver = CHELSIO_CHIP_VERSION(adapter->params.chip);
1904 	wr_mid = FW_WR_LEN16_V(DIV_ROUND_UP(flits, 2));
1905 	if (unlikely(credits < ETHTXQ_STOP_THRES)) {
1906 		/* After we're done injecting the Work Request for this
1907 		 * packet, we'll be below our "stop threshold" so stop the TX
1908 		 * Queue now and schedule a request for an SGE Egress Queue
1909 		 * Update message.  The queue will get started later on when
1910 		 * the firmware processes this Work Request and sends us an
1911 		 * Egress Queue Status Update message indicating that space
1912 		 * has opened up.
1913 		 */
1914 		eth_txq_stop(txq);
1915 		if (chip_ver > CHELSIO_T5)
1916 			wr_mid |= FW_WR_EQUEQ_F | FW_WR_EQUIQ_F;
1917 	}
1918 
1919 	/* Start filling in our Work Request.  Note that we do _not_ handle
1920 	 * the WR Header wrapping around the TX Descriptor Ring.  If our
1921 	 * maximum header size ever exceeds one TX Descriptor, we'll need to
1922 	 * do something else here.
1923 	 */
1924 	WARN_ON(DIV_ROUND_UP(T4VF_ETHTXQ_MAX_HDR, TXD_PER_EQ_UNIT) > 1);
1925 	wr = (void *)&txq->q.desc[txq->q.pidx];
1926 	wr->equiq_to_len16 = cpu_to_be32(wr_mid);
1927 	wr->r3[0] = cpu_to_be32(0);
1928 	wr->r3[1] = cpu_to_be32(0);
1929 	skb_copy_from_linear_data(skb, &wr->firmware, fw_hdr_copy_len);
1930 	end = (u64 *)wr + flits;
1931 
1932 	/* If this is a Large Send Offload packet we'll put in an LSO CPL
1933 	 * message with an encapsulated TX Packet CPL message.  Otherwise we
1934 	 * just use a TX Packet CPL message.
1935 	 */
1936 	ssi = skb_shinfo(skb);
1937 	if (ssi->gso_size) {
1938 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
1939 		bool v6 = (ssi->gso_type & SKB_GSO_TCPV6) != 0;
1940 		int l3hdr_len = skb_network_header_len(skb);
1941 		int eth_xtra_len = skb_network_offset(skb) - ETH_HLEN;
1942 
1943 		wr->op_immdlen =
1944 			cpu_to_be32(FW_WR_OP_V(FW_ETH_TX_PKT_VM_WR) |
1945 				    FW_WR_IMMDLEN_V(sizeof(*lso) +
1946 						    sizeof(*cpl)));
1947 		 /* Fill in the LSO CPL message. */
1948 		lso->lso_ctrl =
1949 			cpu_to_be32(LSO_OPCODE_V(CPL_TX_PKT_LSO) |
1950 				    LSO_FIRST_SLICE_F |
1951 				    LSO_LAST_SLICE_F |
1952 				    LSO_IPV6_V(v6) |
1953 				    LSO_ETHHDR_LEN_V(eth_xtra_len / 4) |
1954 				    LSO_IPHDR_LEN_V(l3hdr_len / 4) |
1955 				    LSO_TCPHDR_LEN_V(tcp_hdr(skb)->doff));
1956 		lso->ipid_ofst = cpu_to_be16(0);
1957 		lso->mss = cpu_to_be16(ssi->gso_size);
1958 		lso->seqno_offset = cpu_to_be32(0);
1959 		if (is_t4(adapter->params.chip))
1960 			lso->len = cpu_to_be32(skb->len);
1961 		else
1962 			lso->len = cpu_to_be32(LSO_T5_XFER_SIZE_V(skb->len));
1963 
1964 		/* Set up TX Packet CPL pointer, control word and perform
1965 		 * accounting.
1966 		 */
1967 		cpl = (void *)(lso + 1);
1968 
1969 		if (chip_ver <= CHELSIO_T5)
1970 			cntrl = TXPKT_ETHHDR_LEN_V(eth_xtra_len);
1971 		else
1972 			cntrl = T6_TXPKT_ETHHDR_LEN_V(eth_xtra_len);
1973 
1974 		cntrl |= TXPKT_CSUM_TYPE_V(v6 ?
1975 					   TX_CSUM_TCPIP6 : TX_CSUM_TCPIP) |
1976 			 TXPKT_IPHDR_LEN_V(l3hdr_len);
1977 		txq->tso++;
1978 		txq->tx_cso += ssi->gso_segs;
1979 	} else {
1980 		int len;
1981 
1982 		len = (t4vf_is_eth_imm(skb)
1983 		       ? skb->len + sizeof(*cpl)
1984 		       : sizeof(*cpl));
1985 		wr->op_immdlen =
1986 			cpu_to_be32(FW_WR_OP_V(FW_ETH_TX_PKT_VM_WR) |
1987 				    FW_WR_IMMDLEN_V(len));
1988 
1989 		/* Set up TX Packet CPL pointer, control word and perform
1990 		 * accounting.
1991 		 */
1992 		cpl = (void *)(wr + 1);
1993 		if (skb->ip_summed == CHECKSUM_PARTIAL) {
1994 			cntrl = hwcsum(adapter->params.chip, skb) |
1995 				TXPKT_IPCSUM_DIS_F;
1996 			txq->tx_cso++;
1997 		} else {
1998 			cntrl = TXPKT_L4CSUM_DIS_F | TXPKT_IPCSUM_DIS_F;
1999 		}
2000 	}
2001 
2002 	/* If there's a VLAN tag present, add that to the list of things to
2003 	 * do in this Work Request.
2004 	 */
2005 	if (skb_vlan_tag_present(skb)) {
2006 		txq->vlan_ins++;
2007 		cntrl |= TXPKT_VLAN_VLD_F | TXPKT_VLAN_V(skb_vlan_tag_get(skb));
2008 	}
2009 
2010 	 /* Fill in the TX Packet CPL message header. */
2011 	cpl->ctrl0 = cpu_to_be32(TXPKT_OPCODE_V(CPL_TX_PKT_XT) |
2012 				 TXPKT_INTF_V(pi->port_id) |
2013 				 TXPKT_PF_V(0));
2014 	cpl->pack = cpu_to_be16(0);
2015 	cpl->len = cpu_to_be16(skb->len);
2016 	cpl->ctrl1 = cpu_to_be64(cntrl);
2017 
2018 	/* Fill in the body of the TX Packet CPL message with either in-lined
2019 	 * data or a Scatter/Gather List.
2020 	 */
2021 	if (t4vf_is_eth_imm(skb)) {
2022 		/* In-line the packet's data and free the skb since we don't
2023 		 * need it any longer.
2024 		 */
2025 		cxgb4_inline_tx_skb(skb, &txq->q, cpl + 1);
2026 		dev_consume_skb_any(skb);
2027 	} else {
2028 		/* Write the skb's Scatter/Gather list into the TX Packet CPL
2029 		 * message and retain a pointer to the skb so we can free it
2030 		 * later when its DMA completes.  (We store the skb pointer
2031 		 * in the Software Descriptor corresponding to the last TX
2032 		 * Descriptor used by the Work Request.)
2033 		 *
2034 		 * The retained skb will be freed when the corresponding TX
2035 		 * Descriptors are reclaimed after their DMAs complete.
2036 		 * However, this could take quite a while since, in general,
2037 		 * the hardware is set up to be lazy about sending DMA
2038 		 * completion notifications to us and we mostly perform TX
2039 		 * reclaims in the transmit routine.
2040 		 *
2041 		 * This is good for performamce but means that we rely on new
2042 		 * TX packets arriving to run the destructors of completed
2043 		 * packets, which open up space in their sockets' send queues.
2044 		 * Sometimes we do not get such new packets causing TX to
2045 		 * stall.  A single UDP transmitter is a good example of this
2046 		 * situation.  We have a clean up timer that periodically
2047 		 * reclaims completed packets but it doesn't run often enough
2048 		 * (nor do we want it to) to prevent lengthy stalls.  A
2049 		 * solution to this problem is to run the destructor early,
2050 		 * after the packet is queued but before it's DMAd.  A con is
2051 		 * that we lie to socket memory accounting, but the amount of
2052 		 * extra memory is reasonable (limited by the number of TX
2053 		 * descriptors), the packets do actually get freed quickly by
2054 		 * new packets almost always, and for protocols like TCP that
2055 		 * wait for acks to really free up the data the extra memory
2056 		 * is even less.  On the positive side we run the destructors
2057 		 * on the sending CPU rather than on a potentially different
2058 		 * completing CPU, usually a good thing.
2059 		 *
2060 		 * Run the destructor before telling the DMA engine about the
2061 		 * packet to make sure it doesn't complete and get freed
2062 		 * prematurely.
2063 		 */
2064 		struct ulptx_sgl *sgl = (struct ulptx_sgl *)(cpl + 1);
2065 		struct sge_txq *tq = &txq->q;
2066 
2067 		/* If the Work Request header was an exact multiple of our TX
2068 		 * Descriptor length, then it's possible that the starting SGL
2069 		 * pointer lines up exactly with the end of our TX Descriptor
2070 		 * ring.  If that's the case, wrap around to the beginning
2071 		 * here ...
2072 		 */
2073 		if (unlikely((void *)sgl == (void *)tq->stat)) {
2074 			sgl = (void *)tq->desc;
2075 			end = (void *)((void *)tq->desc +
2076 				       ((void *)end - (void *)tq->stat));
2077 		}
2078 
2079 		cxgb4_write_sgl(skb, tq, sgl, end, 0, sgl_sdesc->addr);
2080 		skb_orphan(skb);
2081 		sgl_sdesc->skb = skb;
2082 	}
2083 
2084 	/* Advance our internal TX Queue state, tell the hardware about
2085 	 * the new TX descriptors and return success.
2086 	 */
2087 	txq_advance(&txq->q, ndesc);
2088 
2089 	cxgb4_ring_tx_db(adapter, &txq->q, ndesc);
2090 	return NETDEV_TX_OK;
2091 
2092 out_free:
2093 	/* An error of some sort happened.  Free the TX skb and tell the
2094 	 * OS that we've "dealt" with the packet ...
2095 	 */
2096 	dev_kfree_skb_any(skb);
2097 	return NETDEV_TX_OK;
2098 }
2099 
2100 /**
2101  * reclaim_completed_tx_imm - reclaim completed control-queue Tx descs
2102  * @q: the SGE control Tx queue
2103  *
2104  * This is a variant of cxgb4_reclaim_completed_tx() that is used
2105  * for Tx queues that send only immediate data (presently just
2106  * the control queues) and	thus do not have any sk_buffs to release.
2107  */
2108 static inline void reclaim_completed_tx_imm(struct sge_txq *q)
2109 {
2110 	int hw_cidx = ntohs(READ_ONCE(q->stat->cidx));
2111 	int reclaim = hw_cidx - q->cidx;
2112 
2113 	if (reclaim < 0)
2114 		reclaim += q->size;
2115 
2116 	q->in_use -= reclaim;
2117 	q->cidx = hw_cidx;
2118 }
2119 
2120 static inline void eosw_txq_advance_index(u32 *idx, u32 n, u32 max)
2121 {
2122 	u32 val = *idx + n;
2123 
2124 	if (val >= max)
2125 		val -= max;
2126 
2127 	*idx = val;
2128 }
2129 
2130 void cxgb4_eosw_txq_free_desc(struct adapter *adap,
2131 			      struct sge_eosw_txq *eosw_txq, u32 ndesc)
2132 {
2133 	struct tx_sw_desc *d;
2134 
2135 	d = &eosw_txq->desc[eosw_txq->last_cidx];
2136 	while (ndesc--) {
2137 		if (d->skb) {
2138 			if (d->addr[0]) {
2139 				unmap_skb(adap->pdev_dev, d->skb, d->addr);
2140 				memset(d->addr, 0, sizeof(d->addr));
2141 			}
2142 			dev_consume_skb_any(d->skb);
2143 			d->skb = NULL;
2144 		}
2145 		eosw_txq_advance_index(&eosw_txq->last_cidx, 1,
2146 				       eosw_txq->ndesc);
2147 		d = &eosw_txq->desc[eosw_txq->last_cidx];
2148 	}
2149 }
2150 
2151 static inline void eosw_txq_advance(struct sge_eosw_txq *eosw_txq, u32 n)
2152 {
2153 	eosw_txq_advance_index(&eosw_txq->pidx, n, eosw_txq->ndesc);
2154 	eosw_txq->inuse += n;
2155 }
2156 
2157 static inline int eosw_txq_enqueue(struct sge_eosw_txq *eosw_txq,
2158 				   struct sk_buff *skb)
2159 {
2160 	if (eosw_txq->inuse == eosw_txq->ndesc)
2161 		return -ENOMEM;
2162 
2163 	eosw_txq->desc[eosw_txq->pidx].skb = skb;
2164 	return 0;
2165 }
2166 
2167 static inline struct sk_buff *eosw_txq_peek(struct sge_eosw_txq *eosw_txq)
2168 {
2169 	return eosw_txq->desc[eosw_txq->last_pidx].skb;
2170 }
2171 
2172 static inline u8 ethofld_calc_tx_flits(struct adapter *adap,
2173 				       struct sk_buff *skb, u32 hdr_len)
2174 {
2175 	u8 flits, nsgl = 0;
2176 	u32 wrlen;
2177 
2178 	wrlen = sizeof(struct fw_eth_tx_eo_wr) + sizeof(struct cpl_tx_pkt_core);
2179 	if (skb_shinfo(skb)->gso_size &&
2180 	    !(skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4))
2181 		wrlen += sizeof(struct cpl_tx_pkt_lso_core);
2182 
2183 	wrlen += roundup(hdr_len, 16);
2184 
2185 	/* Packet headers + WR + CPLs */
2186 	flits = DIV_ROUND_UP(wrlen, 8);
2187 
2188 	if (skb_shinfo(skb)->nr_frags > 0) {
2189 		if (skb_headlen(skb) - hdr_len)
2190 			nsgl = sgl_len(skb_shinfo(skb)->nr_frags + 1);
2191 		else
2192 			nsgl = sgl_len(skb_shinfo(skb)->nr_frags);
2193 	} else if (skb->len - hdr_len) {
2194 		nsgl = sgl_len(1);
2195 	}
2196 
2197 	return flits + nsgl;
2198 }
2199 
2200 static void *write_eo_wr(struct adapter *adap, struct sge_eosw_txq *eosw_txq,
2201 			 struct sk_buff *skb, struct fw_eth_tx_eo_wr *wr,
2202 			 u32 hdr_len, u32 wrlen)
2203 {
2204 	const struct skb_shared_info *ssi = skb_shinfo(skb);
2205 	struct cpl_tx_pkt_core *cpl;
2206 	u32 immd_len, wrlen16;
2207 	bool compl = false;
2208 	u8 ver, proto;
2209 
2210 	ver = ip_hdr(skb)->version;
2211 	proto = (ver == 6) ? ipv6_hdr(skb)->nexthdr : ip_hdr(skb)->protocol;
2212 
2213 	wrlen16 = DIV_ROUND_UP(wrlen, 16);
2214 	immd_len = sizeof(struct cpl_tx_pkt_core);
2215 	if (skb_shinfo(skb)->gso_size &&
2216 	    !(skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4))
2217 		immd_len += sizeof(struct cpl_tx_pkt_lso_core);
2218 	immd_len += hdr_len;
2219 
2220 	if (!eosw_txq->ncompl ||
2221 	    (eosw_txq->last_compl + wrlen16) >=
2222 	    (adap->params.ofldq_wr_cred / 2)) {
2223 		compl = true;
2224 		eosw_txq->ncompl++;
2225 		eosw_txq->last_compl = 0;
2226 	}
2227 
2228 	wr->op_immdlen = cpu_to_be32(FW_WR_OP_V(FW_ETH_TX_EO_WR) |
2229 				     FW_ETH_TX_EO_WR_IMMDLEN_V(immd_len) |
2230 				     FW_WR_COMPL_V(compl));
2231 	wr->equiq_to_len16 = cpu_to_be32(FW_WR_LEN16_V(wrlen16) |
2232 					 FW_WR_FLOWID_V(eosw_txq->hwtid));
2233 	wr->r3 = 0;
2234 	if (proto == IPPROTO_UDP) {
2235 		cpl = write_eo_udp_wr(skb, wr, hdr_len);
2236 	} else {
2237 		wr->u.tcpseg.type = FW_ETH_TX_EO_TYPE_TCPSEG;
2238 		wr->u.tcpseg.ethlen = skb_network_offset(skb);
2239 		wr->u.tcpseg.iplen = cpu_to_be16(skb_network_header_len(skb));
2240 		wr->u.tcpseg.tcplen = tcp_hdrlen(skb);
2241 		wr->u.tcpseg.tsclk_tsoff = 0;
2242 		wr->u.tcpseg.r4 = 0;
2243 		wr->u.tcpseg.r5 = 0;
2244 		wr->u.tcpseg.plen = cpu_to_be32(skb->len - hdr_len);
2245 
2246 		if (ssi->gso_size) {
2247 			struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
2248 
2249 			wr->u.tcpseg.mss = cpu_to_be16(ssi->gso_size);
2250 			cpl = write_tso_wr(adap, skb, lso);
2251 		} else {
2252 			wr->u.tcpseg.mss = cpu_to_be16(0xffff);
2253 			cpl = (void *)(wr + 1);
2254 		}
2255 	}
2256 
2257 	eosw_txq->cred -= wrlen16;
2258 	eosw_txq->last_compl += wrlen16;
2259 	return cpl;
2260 }
2261 
2262 static int ethofld_hard_xmit(struct net_device *dev,
2263 			     struct sge_eosw_txq *eosw_txq)
2264 {
2265 	struct port_info *pi = netdev2pinfo(dev);
2266 	struct adapter *adap = netdev2adap(dev);
2267 	u32 wrlen, wrlen16, hdr_len, data_len;
2268 	enum sge_eosw_state next_state;
2269 	u64 cntrl, *start, *end, *sgl;
2270 	struct sge_eohw_txq *eohw_txq;
2271 	struct cpl_tx_pkt_core *cpl;
2272 	struct fw_eth_tx_eo_wr *wr;
2273 	bool skip_eotx_wr = false;
2274 	struct tx_sw_desc *d;
2275 	struct sk_buff *skb;
2276 	int left, ret = 0;
2277 	u8 flits, ndesc;
2278 
2279 	eohw_txq = &adap->sge.eohw_txq[eosw_txq->hwqid];
2280 	spin_lock(&eohw_txq->lock);
2281 	reclaim_completed_tx_imm(&eohw_txq->q);
2282 
2283 	d = &eosw_txq->desc[eosw_txq->last_pidx];
2284 	skb = d->skb;
2285 	skb_tx_timestamp(skb);
2286 
2287 	wr = (struct fw_eth_tx_eo_wr *)&eohw_txq->q.desc[eohw_txq->q.pidx];
2288 	if (unlikely(eosw_txq->state != CXGB4_EO_STATE_ACTIVE &&
2289 		     eosw_txq->last_pidx == eosw_txq->flowc_idx)) {
2290 		hdr_len = skb->len;
2291 		data_len = 0;
2292 		flits = DIV_ROUND_UP(hdr_len, 8);
2293 		if (eosw_txq->state == CXGB4_EO_STATE_FLOWC_OPEN_SEND)
2294 			next_state = CXGB4_EO_STATE_FLOWC_OPEN_REPLY;
2295 		else
2296 			next_state = CXGB4_EO_STATE_FLOWC_CLOSE_REPLY;
2297 		skip_eotx_wr = true;
2298 	} else {
2299 		hdr_len = eth_get_headlen(dev, skb->data, skb_headlen(skb));
2300 		data_len = skb->len - hdr_len;
2301 		flits = ethofld_calc_tx_flits(adap, skb, hdr_len);
2302 	}
2303 	ndesc = flits_to_desc(flits);
2304 	wrlen = flits * 8;
2305 	wrlen16 = DIV_ROUND_UP(wrlen, 16);
2306 
2307 	left = txq_avail(&eohw_txq->q) - ndesc;
2308 
2309 	/* If there are no descriptors left in hardware queues or no
2310 	 * CPL credits left in software queues, then wait for them
2311 	 * to come back and retry again. Note that we always request
2312 	 * for credits update via interrupt for every half credits
2313 	 * consumed. So, the interrupt will eventually restore the
2314 	 * credits and invoke the Tx path again.
2315 	 */
2316 	if (unlikely(left < 0 || wrlen16 > eosw_txq->cred)) {
2317 		ret = -ENOMEM;
2318 		goto out_unlock;
2319 	}
2320 
2321 	if (unlikely(skip_eotx_wr)) {
2322 		start = (u64 *)wr;
2323 		eosw_txq->state = next_state;
2324 		eosw_txq->cred -= wrlen16;
2325 		eosw_txq->ncompl++;
2326 		eosw_txq->last_compl = 0;
2327 		goto write_wr_headers;
2328 	}
2329 
2330 	cpl = write_eo_wr(adap, eosw_txq, skb, wr, hdr_len, wrlen);
2331 	cntrl = hwcsum(adap->params.chip, skb);
2332 	if (skb_vlan_tag_present(skb))
2333 		cntrl |= TXPKT_VLAN_VLD_F | TXPKT_VLAN_V(skb_vlan_tag_get(skb));
2334 
2335 	cpl->ctrl0 = cpu_to_be32(TXPKT_OPCODE_V(CPL_TX_PKT_XT) |
2336 				 TXPKT_INTF_V(pi->tx_chan) |
2337 				 TXPKT_PF_V(adap->pf));
2338 	cpl->pack = 0;
2339 	cpl->len = cpu_to_be16(skb->len);
2340 	cpl->ctrl1 = cpu_to_be64(cntrl);
2341 
2342 	start = (u64 *)(cpl + 1);
2343 
2344 write_wr_headers:
2345 	sgl = (u64 *)inline_tx_skb_header(skb, &eohw_txq->q, (void *)start,
2346 					  hdr_len);
2347 	if (data_len) {
2348 		ret = cxgb4_map_skb(adap->pdev_dev, skb, d->addr);
2349 		if (unlikely(ret)) {
2350 			memset(d->addr, 0, sizeof(d->addr));
2351 			eohw_txq->mapping_err++;
2352 			goto out_unlock;
2353 		}
2354 
2355 		end = (u64 *)wr + flits;
2356 		if (unlikely(start > sgl)) {
2357 			left = (u8 *)end - (u8 *)eohw_txq->q.stat;
2358 			end = (void *)eohw_txq->q.desc + left;
2359 		}
2360 
2361 		if (unlikely((u8 *)sgl >= (u8 *)eohw_txq->q.stat)) {
2362 			/* If current position is already at the end of the
2363 			 * txq, reset the current to point to start of the queue
2364 			 * and update the end ptr as well.
2365 			 */
2366 			left = (u8 *)end - (u8 *)eohw_txq->q.stat;
2367 
2368 			end = (void *)eohw_txq->q.desc + left;
2369 			sgl = (void *)eohw_txq->q.desc;
2370 		}
2371 
2372 		cxgb4_write_sgl(skb, &eohw_txq->q, (void *)sgl, end, hdr_len,
2373 				d->addr);
2374 	}
2375 
2376 	if (skb_shinfo(skb)->gso_size) {
2377 		if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4)
2378 			eohw_txq->uso++;
2379 		else
2380 			eohw_txq->tso++;
2381 		eohw_txq->tx_cso += skb_shinfo(skb)->gso_segs;
2382 	} else if (skb->ip_summed == CHECKSUM_PARTIAL) {
2383 		eohw_txq->tx_cso++;
2384 	}
2385 
2386 	if (skb_vlan_tag_present(skb))
2387 		eohw_txq->vlan_ins++;
2388 
2389 	txq_advance(&eohw_txq->q, ndesc);
2390 	cxgb4_ring_tx_db(adap, &eohw_txq->q, ndesc);
2391 	eosw_txq_advance_index(&eosw_txq->last_pidx, 1, eosw_txq->ndesc);
2392 
2393 out_unlock:
2394 	spin_unlock(&eohw_txq->lock);
2395 	return ret;
2396 }
2397 
2398 static void ethofld_xmit(struct net_device *dev, struct sge_eosw_txq *eosw_txq)
2399 {
2400 	struct sk_buff *skb;
2401 	int pktcount, ret;
2402 
2403 	switch (eosw_txq->state) {
2404 	case CXGB4_EO_STATE_ACTIVE:
2405 	case CXGB4_EO_STATE_FLOWC_OPEN_SEND:
2406 	case CXGB4_EO_STATE_FLOWC_CLOSE_SEND:
2407 		pktcount = eosw_txq->pidx - eosw_txq->last_pidx;
2408 		if (pktcount < 0)
2409 			pktcount += eosw_txq->ndesc;
2410 		break;
2411 	case CXGB4_EO_STATE_FLOWC_OPEN_REPLY:
2412 	case CXGB4_EO_STATE_FLOWC_CLOSE_REPLY:
2413 	case CXGB4_EO_STATE_CLOSED:
2414 	default:
2415 		return;
2416 	}
2417 
2418 	while (pktcount--) {
2419 		skb = eosw_txq_peek(eosw_txq);
2420 		if (!skb) {
2421 			eosw_txq_advance_index(&eosw_txq->last_pidx, 1,
2422 					       eosw_txq->ndesc);
2423 			continue;
2424 		}
2425 
2426 		ret = ethofld_hard_xmit(dev, eosw_txq);
2427 		if (ret)
2428 			break;
2429 	}
2430 }
2431 
2432 static netdev_tx_t cxgb4_ethofld_xmit(struct sk_buff *skb,
2433 				      struct net_device *dev)
2434 {
2435 	struct cxgb4_tc_port_mqprio *tc_port_mqprio;
2436 	struct port_info *pi = netdev2pinfo(dev);
2437 	struct adapter *adap = netdev2adap(dev);
2438 	struct sge_eosw_txq *eosw_txq;
2439 	u32 qid;
2440 	int ret;
2441 
2442 	ret = cxgb4_validate_skb(skb, dev, ETH_HLEN);
2443 	if (ret)
2444 		goto out_free;
2445 
2446 	tc_port_mqprio = &adap->tc_mqprio->port_mqprio[pi->port_id];
2447 	qid = skb_get_queue_mapping(skb) - pi->nqsets;
2448 	eosw_txq = &tc_port_mqprio->eosw_txq[qid];
2449 	spin_lock_bh(&eosw_txq->lock);
2450 	if (eosw_txq->state != CXGB4_EO_STATE_ACTIVE)
2451 		goto out_unlock;
2452 
2453 	ret = eosw_txq_enqueue(eosw_txq, skb);
2454 	if (ret)
2455 		goto out_unlock;
2456 
2457 	/* SKB is queued for processing until credits are available.
2458 	 * So, call the destructor now and we'll free the skb later
2459 	 * after it has been successfully transmitted.
2460 	 */
2461 	skb_orphan(skb);
2462 
2463 	eosw_txq_advance(eosw_txq, 1);
2464 	ethofld_xmit(dev, eosw_txq);
2465 	spin_unlock_bh(&eosw_txq->lock);
2466 	return NETDEV_TX_OK;
2467 
2468 out_unlock:
2469 	spin_unlock_bh(&eosw_txq->lock);
2470 out_free:
2471 	dev_kfree_skb_any(skb);
2472 	return NETDEV_TX_OK;
2473 }
2474 
2475 netdev_tx_t t4_start_xmit(struct sk_buff *skb, struct net_device *dev)
2476 {
2477 	struct port_info *pi = netdev_priv(dev);
2478 	u16 qid = skb_get_queue_mapping(skb);
2479 
2480 	if (unlikely(pi->eth_flags & PRIV_FLAG_PORT_TX_VM))
2481 		return cxgb4_vf_eth_xmit(skb, dev);
2482 
2483 	if (unlikely(qid >= pi->nqsets))
2484 		return cxgb4_ethofld_xmit(skb, dev);
2485 
2486 	if (is_ptp_enabled(skb, dev)) {
2487 		struct adapter *adap = netdev2adap(dev);
2488 		netdev_tx_t ret;
2489 
2490 		spin_lock(&adap->ptp_lock);
2491 		ret = cxgb4_eth_xmit(skb, dev);
2492 		spin_unlock(&adap->ptp_lock);
2493 		return ret;
2494 	}
2495 
2496 	return cxgb4_eth_xmit(skb, dev);
2497 }
2498 
2499 static void eosw_txq_flush_pending_skbs(struct sge_eosw_txq *eosw_txq)
2500 {
2501 	int pktcount = eosw_txq->pidx - eosw_txq->last_pidx;
2502 	int pidx = eosw_txq->pidx;
2503 	struct sk_buff *skb;
2504 
2505 	if (!pktcount)
2506 		return;
2507 
2508 	if (pktcount < 0)
2509 		pktcount += eosw_txq->ndesc;
2510 
2511 	while (pktcount--) {
2512 		pidx--;
2513 		if (pidx < 0)
2514 			pidx += eosw_txq->ndesc;
2515 
2516 		skb = eosw_txq->desc[pidx].skb;
2517 		if (skb) {
2518 			dev_consume_skb_any(skb);
2519 			eosw_txq->desc[pidx].skb = NULL;
2520 			eosw_txq->inuse--;
2521 		}
2522 	}
2523 
2524 	eosw_txq->pidx = eosw_txq->last_pidx + 1;
2525 }
2526 
2527 /**
2528  * cxgb4_ethofld_send_flowc - Send ETHOFLD flowc request to bind eotid to tc.
2529  * @dev: netdevice
2530  * @eotid: ETHOFLD tid to bind/unbind
2531  * @tc: traffic class. If set to FW_SCHED_CLS_NONE, then unbinds the @eotid
2532  *
2533  * Send a FLOWC work request to bind an ETHOFLD TID to a traffic class.
2534  * If @tc is set to FW_SCHED_CLS_NONE, then the @eotid is unbound from
2535  * a traffic class.
2536  */
2537 int cxgb4_ethofld_send_flowc(struct net_device *dev, u32 eotid, u32 tc)
2538 {
2539 	struct port_info *pi = netdev2pinfo(dev);
2540 	struct adapter *adap = netdev2adap(dev);
2541 	enum sge_eosw_state next_state;
2542 	struct sge_eosw_txq *eosw_txq;
2543 	u32 len, len16, nparams = 6;
2544 	struct fw_flowc_wr *flowc;
2545 	struct eotid_entry *entry;
2546 	struct sge_ofld_rxq *rxq;
2547 	struct sk_buff *skb;
2548 	int ret = 0;
2549 
2550 	len = struct_size(flowc, mnemval, nparams);
2551 	len16 = DIV_ROUND_UP(len, 16);
2552 
2553 	entry = cxgb4_lookup_eotid(&adap->tids, eotid);
2554 	if (!entry)
2555 		return -ENOMEM;
2556 
2557 	eosw_txq = (struct sge_eosw_txq *)entry->data;
2558 	if (!eosw_txq)
2559 		return -ENOMEM;
2560 
2561 	if (!(adap->flags & CXGB4_FW_OK)) {
2562 		/* Don't stall caller when access to FW is lost */
2563 		complete(&eosw_txq->completion);
2564 		return -EIO;
2565 	}
2566 
2567 	skb = alloc_skb(len, GFP_KERNEL);
2568 	if (!skb)
2569 		return -ENOMEM;
2570 
2571 	spin_lock_bh(&eosw_txq->lock);
2572 	if (tc != FW_SCHED_CLS_NONE) {
2573 		if (eosw_txq->state != CXGB4_EO_STATE_CLOSED)
2574 			goto out_free_skb;
2575 
2576 		next_state = CXGB4_EO_STATE_FLOWC_OPEN_SEND;
2577 	} else {
2578 		if (eosw_txq->state != CXGB4_EO_STATE_ACTIVE)
2579 			goto out_free_skb;
2580 
2581 		next_state = CXGB4_EO_STATE_FLOWC_CLOSE_SEND;
2582 	}
2583 
2584 	flowc = __skb_put(skb, len);
2585 	memset(flowc, 0, len);
2586 
2587 	rxq = &adap->sge.eohw_rxq[eosw_txq->hwqid];
2588 	flowc->flowid_len16 = cpu_to_be32(FW_WR_LEN16_V(len16) |
2589 					  FW_WR_FLOWID_V(eosw_txq->hwtid));
2590 	flowc->op_to_nparams = cpu_to_be32(FW_WR_OP_V(FW_FLOWC_WR) |
2591 					   FW_FLOWC_WR_NPARAMS_V(nparams) |
2592 					   FW_WR_COMPL_V(1));
2593 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
2594 	flowc->mnemval[0].val = cpu_to_be32(FW_PFVF_CMD_PFN_V(adap->pf));
2595 	flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
2596 	flowc->mnemval[1].val = cpu_to_be32(pi->tx_chan);
2597 	flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
2598 	flowc->mnemval[2].val = cpu_to_be32(pi->tx_chan);
2599 	flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
2600 	flowc->mnemval[3].val = cpu_to_be32(rxq->rspq.abs_id);
2601 	flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
2602 	flowc->mnemval[4].val = cpu_to_be32(tc);
2603 	flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_EOSTATE;
2604 	flowc->mnemval[5].val = cpu_to_be32(tc == FW_SCHED_CLS_NONE ?
2605 					    FW_FLOWC_MNEM_EOSTATE_CLOSING :
2606 					    FW_FLOWC_MNEM_EOSTATE_ESTABLISHED);
2607 
2608 	/* Free up any pending skbs to ensure there's room for
2609 	 * termination FLOWC.
2610 	 */
2611 	if (tc == FW_SCHED_CLS_NONE)
2612 		eosw_txq_flush_pending_skbs(eosw_txq);
2613 
2614 	ret = eosw_txq_enqueue(eosw_txq, skb);
2615 	if (ret)
2616 		goto out_free_skb;
2617 
2618 	eosw_txq->state = next_state;
2619 	eosw_txq->flowc_idx = eosw_txq->pidx;
2620 	eosw_txq_advance(eosw_txq, 1);
2621 	ethofld_xmit(dev, eosw_txq);
2622 
2623 	spin_unlock_bh(&eosw_txq->lock);
2624 	return 0;
2625 
2626 out_free_skb:
2627 	dev_consume_skb_any(skb);
2628 	spin_unlock_bh(&eosw_txq->lock);
2629 	return ret;
2630 }
2631 
2632 /**
2633  *	is_imm - check whether a packet can be sent as immediate data
2634  *	@skb: the packet
2635  *
2636  *	Returns true if a packet can be sent as a WR with immediate data.
2637  */
2638 static inline int is_imm(const struct sk_buff *skb)
2639 {
2640 	return skb->len <= MAX_CTRL_WR_LEN;
2641 }
2642 
2643 /**
2644  *	ctrlq_check_stop - check if a control queue is full and should stop
2645  *	@q: the queue
2646  *	@wr: most recent WR written to the queue
2647  *
2648  *	Check if a control queue has become full and should be stopped.
2649  *	We clean up control queue descriptors very lazily, only when we are out.
2650  *	If the queue is still full after reclaiming any completed descriptors
2651  *	we suspend it and have the last WR wake it up.
2652  */
2653 static void ctrlq_check_stop(struct sge_ctrl_txq *q, struct fw_wr_hdr *wr)
2654 {
2655 	reclaim_completed_tx_imm(&q->q);
2656 	if (unlikely(txq_avail(&q->q) < TXQ_STOP_THRES)) {
2657 		wr->lo |= htonl(FW_WR_EQUEQ_F | FW_WR_EQUIQ_F);
2658 		q->q.stops++;
2659 		q->full = 1;
2660 	}
2661 }
2662 
2663 #define CXGB4_SELFTEST_LB_STR "CHELSIO_SELFTEST"
2664 
2665 int cxgb4_selftest_lb_pkt(struct net_device *netdev)
2666 {
2667 	struct port_info *pi = netdev_priv(netdev);
2668 	struct adapter *adap = pi->adapter;
2669 	struct cxgb4_ethtool_lb_test *lb;
2670 	int ret, i = 0, pkt_len, credits;
2671 	struct fw_eth_tx_pkt_wr *wr;
2672 	struct cpl_tx_pkt_core *cpl;
2673 	u32 ctrl0, ndesc, flits;
2674 	struct sge_eth_txq *q;
2675 	u8 *sgl;
2676 
2677 	pkt_len = ETH_HLEN + sizeof(CXGB4_SELFTEST_LB_STR);
2678 
2679 	flits = DIV_ROUND_UP(pkt_len + sizeof(*cpl) + sizeof(*wr),
2680 			     sizeof(__be64));
2681 	ndesc = flits_to_desc(flits);
2682 
2683 	lb = &pi->ethtool_lb;
2684 	lb->loopback = 1;
2685 
2686 	q = &adap->sge.ethtxq[pi->first_qset];
2687 	__netif_tx_lock(q->txq, smp_processor_id());
2688 
2689 	reclaim_completed_tx(adap, &q->q, -1, true);
2690 	credits = txq_avail(&q->q) - ndesc;
2691 	if (unlikely(credits < 0)) {
2692 		__netif_tx_unlock(q->txq);
2693 		return -ENOMEM;
2694 	}
2695 
2696 	wr = (void *)&q->q.desc[q->q.pidx];
2697 	memset(wr, 0, sizeof(struct tx_desc));
2698 
2699 	wr->op_immdlen = htonl(FW_WR_OP_V(FW_ETH_TX_PKT_WR) |
2700 			       FW_WR_IMMDLEN_V(pkt_len +
2701 			       sizeof(*cpl)));
2702 	wr->equiq_to_len16 = htonl(FW_WR_LEN16_V(DIV_ROUND_UP(flits, 2)));
2703 	wr->r3 = cpu_to_be64(0);
2704 
2705 	cpl = (void *)(wr + 1);
2706 	sgl = (u8 *)(cpl + 1);
2707 
2708 	ctrl0 = TXPKT_OPCODE_V(CPL_TX_PKT_XT) | TXPKT_PF_V(adap->pf) |
2709 		TXPKT_INTF_V(pi->tx_chan + 4);
2710 
2711 	cpl->ctrl0 = htonl(ctrl0);
2712 	cpl->pack = htons(0);
2713 	cpl->len = htons(pkt_len);
2714 	cpl->ctrl1 = cpu_to_be64(TXPKT_L4CSUM_DIS_F | TXPKT_IPCSUM_DIS_F);
2715 
2716 	eth_broadcast_addr(sgl);
2717 	i += ETH_ALEN;
2718 	ether_addr_copy(&sgl[i], netdev->dev_addr);
2719 	i += ETH_ALEN;
2720 
2721 	snprintf(&sgl[i], sizeof(CXGB4_SELFTEST_LB_STR), "%s",
2722 		 CXGB4_SELFTEST_LB_STR);
2723 
2724 	init_completion(&lb->completion);
2725 	txq_advance(&q->q, ndesc);
2726 	cxgb4_ring_tx_db(adap, &q->q, ndesc);
2727 	__netif_tx_unlock(q->txq);
2728 
2729 	/* wait for the pkt to return */
2730 	ret = wait_for_completion_timeout(&lb->completion, 10 * HZ);
2731 	if (!ret)
2732 		ret = -ETIMEDOUT;
2733 	else
2734 		ret = lb->result;
2735 
2736 	lb->loopback = 0;
2737 
2738 	return ret;
2739 }
2740 
2741 /**
2742  *	ctrl_xmit - send a packet through an SGE control Tx queue
2743  *	@q: the control queue
2744  *	@skb: the packet
2745  *
2746  *	Send a packet through an SGE control Tx queue.  Packets sent through
2747  *	a control queue must fit entirely as immediate data.
2748  */
2749 static int ctrl_xmit(struct sge_ctrl_txq *q, struct sk_buff *skb)
2750 {
2751 	unsigned int ndesc;
2752 	struct fw_wr_hdr *wr;
2753 
2754 	if (unlikely(!is_imm(skb))) {
2755 		WARN_ON(1);
2756 		dev_kfree_skb(skb);
2757 		return NET_XMIT_DROP;
2758 	}
2759 
2760 	ndesc = DIV_ROUND_UP(skb->len, sizeof(struct tx_desc));
2761 	spin_lock(&q->sendq.lock);
2762 
2763 	if (unlikely(q->full)) {
2764 		skb->priority = ndesc;                  /* save for restart */
2765 		__skb_queue_tail(&q->sendq, skb);
2766 		spin_unlock(&q->sendq.lock);
2767 		return NET_XMIT_CN;
2768 	}
2769 
2770 	wr = (struct fw_wr_hdr *)&q->q.desc[q->q.pidx];
2771 	cxgb4_inline_tx_skb(skb, &q->q, wr);
2772 
2773 	txq_advance(&q->q, ndesc);
2774 	if (unlikely(txq_avail(&q->q) < TXQ_STOP_THRES))
2775 		ctrlq_check_stop(q, wr);
2776 
2777 	cxgb4_ring_tx_db(q->adap, &q->q, ndesc);
2778 	spin_unlock(&q->sendq.lock);
2779 
2780 	kfree_skb(skb);
2781 	return NET_XMIT_SUCCESS;
2782 }
2783 
2784 /**
2785  *	restart_ctrlq - restart a suspended control queue
2786  *	@t: pointer to the tasklet associated with this handler
2787  *
2788  *	Resumes transmission on a suspended Tx control queue.
2789  */
2790 static void restart_ctrlq(struct tasklet_struct *t)
2791 {
2792 	struct sk_buff *skb;
2793 	unsigned int written = 0;
2794 	struct sge_ctrl_txq *q = from_tasklet(q, t, qresume_tsk);
2795 
2796 	spin_lock(&q->sendq.lock);
2797 	reclaim_completed_tx_imm(&q->q);
2798 	BUG_ON(txq_avail(&q->q) < TXQ_STOP_THRES);  /* q should be empty */
2799 
2800 	while ((skb = __skb_dequeue(&q->sendq)) != NULL) {
2801 		struct fw_wr_hdr *wr;
2802 		unsigned int ndesc = skb->priority;     /* previously saved */
2803 
2804 		written += ndesc;
2805 		/* Write descriptors and free skbs outside the lock to limit
2806 		 * wait times.  q->full is still set so new skbs will be queued.
2807 		 */
2808 		wr = (struct fw_wr_hdr *)&q->q.desc[q->q.pidx];
2809 		txq_advance(&q->q, ndesc);
2810 		spin_unlock(&q->sendq.lock);
2811 
2812 		cxgb4_inline_tx_skb(skb, &q->q, wr);
2813 		kfree_skb(skb);
2814 
2815 		if (unlikely(txq_avail(&q->q) < TXQ_STOP_THRES)) {
2816 			unsigned long old = q->q.stops;
2817 
2818 			ctrlq_check_stop(q, wr);
2819 			if (q->q.stops != old) {          /* suspended anew */
2820 				spin_lock(&q->sendq.lock);
2821 				goto ringdb;
2822 			}
2823 		}
2824 		if (written > 16) {
2825 			cxgb4_ring_tx_db(q->adap, &q->q, written);
2826 			written = 0;
2827 		}
2828 		spin_lock(&q->sendq.lock);
2829 	}
2830 	q->full = 0;
2831 ringdb:
2832 	if (written)
2833 		cxgb4_ring_tx_db(q->adap, &q->q, written);
2834 	spin_unlock(&q->sendq.lock);
2835 }
2836 
2837 /**
2838  *	t4_mgmt_tx - send a management message
2839  *	@adap: the adapter
2840  *	@skb: the packet containing the management message
2841  *
2842  *	Send a management message through control queue 0.
2843  */
2844 int t4_mgmt_tx(struct adapter *adap, struct sk_buff *skb)
2845 {
2846 	int ret;
2847 
2848 	local_bh_disable();
2849 	ret = ctrl_xmit(&adap->sge.ctrlq[0], skb);
2850 	local_bh_enable();
2851 	return ret;
2852 }
2853 
2854 /**
2855  *	is_ofld_imm - check whether a packet can be sent as immediate data
2856  *	@skb: the packet
2857  *
2858  *	Returns true if a packet can be sent as an offload WR with immediate
2859  *	data.
2860  *	FW_OFLD_TX_DATA_WR limits the payload to 255 bytes due to 8-bit field.
2861  *      However, FW_ULPTX_WR commands have a 256 byte immediate only
2862  *      payload limit.
2863  */
2864 static inline int is_ofld_imm(const struct sk_buff *skb)
2865 {
2866 	struct work_request_hdr *req = (struct work_request_hdr *)skb->data;
2867 	unsigned long opcode = FW_WR_OP_G(ntohl(req->wr_hi));
2868 
2869 	if (unlikely(opcode == FW_ULPTX_WR))
2870 		return skb->len <= MAX_IMM_ULPTX_WR_LEN;
2871 	else if (opcode == FW_CRYPTO_LOOKASIDE_WR)
2872 		return skb->len <= SGE_MAX_WR_LEN;
2873 	else
2874 		return skb->len <= MAX_IMM_OFLD_TX_DATA_WR_LEN;
2875 }
2876 
2877 /**
2878  *	calc_tx_flits_ofld - calculate # of flits for an offload packet
2879  *	@skb: the packet
2880  *
2881  *	Returns the number of flits needed for the given offload packet.
2882  *	These packets are already fully constructed and no additional headers
2883  *	will be added.
2884  */
2885 static inline unsigned int calc_tx_flits_ofld(const struct sk_buff *skb)
2886 {
2887 	unsigned int flits, cnt;
2888 
2889 	if (is_ofld_imm(skb))
2890 		return DIV_ROUND_UP(skb->len, 8);
2891 
2892 	flits = skb_transport_offset(skb) / 8U;   /* headers */
2893 	cnt = skb_shinfo(skb)->nr_frags;
2894 	if (skb_tail_pointer(skb) != skb_transport_header(skb))
2895 		cnt++;
2896 	return flits + sgl_len(cnt);
2897 }
2898 
2899 /**
2900  *	txq_stop_maperr - stop a Tx queue due to I/O MMU exhaustion
2901  *	@q: the queue to stop
2902  *
2903  *	Mark a Tx queue stopped due to I/O MMU exhaustion and resulting
2904  *	inability to map packets.  A periodic timer attempts to restart
2905  *	queues so marked.
2906  */
2907 static void txq_stop_maperr(struct sge_uld_txq *q)
2908 {
2909 	q->mapping_err++;
2910 	q->q.stops++;
2911 	set_bit(q->q.cntxt_id - q->adap->sge.egr_start,
2912 		q->adap->sge.txq_maperr);
2913 }
2914 
2915 /**
2916  *	ofldtxq_stop - stop an offload Tx queue that has become full
2917  *	@q: the queue to stop
2918  *	@wr: the Work Request causing the queue to become full
2919  *
2920  *	Stops an offload Tx queue that has become full and modifies the packet
2921  *	being written to request a wakeup.
2922  */
2923 static void ofldtxq_stop(struct sge_uld_txq *q, struct fw_wr_hdr *wr)
2924 {
2925 	wr->lo |= htonl(FW_WR_EQUEQ_F | FW_WR_EQUIQ_F);
2926 	q->q.stops++;
2927 	q->full = 1;
2928 }
2929 
2930 /**
2931  *	service_ofldq - service/restart a suspended offload queue
2932  *	@q: the offload queue
2933  *
2934  *	Services an offload Tx queue by moving packets from its Pending Send
2935  *	Queue to the Hardware TX ring.  The function starts and ends with the
2936  *	Send Queue locked, but drops the lock while putting the skb at the
2937  *	head of the Send Queue onto the Hardware TX Ring.  Dropping the lock
2938  *	allows more skbs to be added to the Send Queue by other threads.
2939  *	The packet being processed at the head of the Pending Send Queue is
2940  *	left on the queue in case we experience DMA Mapping errors, etc.
2941  *	and need to give up and restart later.
2942  *
2943  *	service_ofldq() can be thought of as a task which opportunistically
2944  *	uses other threads execution contexts.  We use the Offload Queue
2945  *	boolean "service_ofldq_running" to make sure that only one instance
2946  *	is ever running at a time ...
2947  */
2948 static void service_ofldq(struct sge_uld_txq *q)
2949 	__must_hold(&q->sendq.lock)
2950 {
2951 	u64 *pos, *before, *end;
2952 	int credits;
2953 	struct sk_buff *skb;
2954 	struct sge_txq *txq;
2955 	unsigned int left;
2956 	unsigned int written = 0;
2957 	unsigned int flits, ndesc;
2958 
2959 	/* If another thread is currently in service_ofldq() processing the
2960 	 * Pending Send Queue then there's nothing to do. Otherwise, flag
2961 	 * that we're doing the work and continue.  Examining/modifying
2962 	 * the Offload Queue boolean "service_ofldq_running" must be done
2963 	 * while holding the Pending Send Queue Lock.
2964 	 */
2965 	if (q->service_ofldq_running)
2966 		return;
2967 	q->service_ofldq_running = true;
2968 
2969 	while ((skb = skb_peek(&q->sendq)) != NULL && !q->full) {
2970 		/* We drop the lock while we're working with the skb at the
2971 		 * head of the Pending Send Queue.  This allows more skbs to
2972 		 * be added to the Pending Send Queue while we're working on
2973 		 * this one.  We don't need to lock to guard the TX Ring
2974 		 * updates because only one thread of execution is ever
2975 		 * allowed into service_ofldq() at a time.
2976 		 */
2977 		spin_unlock(&q->sendq.lock);
2978 
2979 		cxgb4_reclaim_completed_tx(q->adap, &q->q, false);
2980 
2981 		flits = skb->priority;                /* previously saved */
2982 		ndesc = flits_to_desc(flits);
2983 		credits = txq_avail(&q->q) - ndesc;
2984 		BUG_ON(credits < 0);
2985 		if (unlikely(credits < TXQ_STOP_THRES))
2986 			ofldtxq_stop(q, (struct fw_wr_hdr *)skb->data);
2987 
2988 		pos = (u64 *)&q->q.desc[q->q.pidx];
2989 		if (is_ofld_imm(skb))
2990 			cxgb4_inline_tx_skb(skb, &q->q, pos);
2991 		else if (cxgb4_map_skb(q->adap->pdev_dev, skb,
2992 				       (dma_addr_t *)skb->head)) {
2993 			txq_stop_maperr(q);
2994 			spin_lock(&q->sendq.lock);
2995 			break;
2996 		} else {
2997 			int last_desc, hdr_len = skb_transport_offset(skb);
2998 
2999 			/* The WR headers  may not fit within one descriptor.
3000 			 * So we need to deal with wrap-around here.
3001 			 */
3002 			before = (u64 *)pos;
3003 			end = (u64 *)pos + flits;
3004 			txq = &q->q;
3005 			pos = (void *)inline_tx_skb_header(skb, &q->q,
3006 							   (void *)pos,
3007 							   hdr_len);
3008 			if (before > (u64 *)pos) {
3009 				left = (u8 *)end - (u8 *)txq->stat;
3010 				end = (void *)txq->desc + left;
3011 			}
3012 
3013 			/* If current position is already at the end of the
3014 			 * ofld queue, reset the current to point to
3015 			 * start of the queue and update the end ptr as well.
3016 			 */
3017 			if (pos == (u64 *)txq->stat) {
3018 				left = (u8 *)end - (u8 *)txq->stat;
3019 				end = (void *)txq->desc + left;
3020 				pos = (void *)txq->desc;
3021 			}
3022 
3023 			cxgb4_write_sgl(skb, &q->q, (void *)pos,
3024 					end, hdr_len,
3025 					(dma_addr_t *)skb->head);
3026 #ifdef CONFIG_NEED_DMA_MAP_STATE
3027 			skb->dev = q->adap->port[0];
3028 			skb->destructor = deferred_unmap_destructor;
3029 #endif
3030 			last_desc = q->q.pidx + ndesc - 1;
3031 			if (last_desc >= q->q.size)
3032 				last_desc -= q->q.size;
3033 			q->q.sdesc[last_desc].skb = skb;
3034 		}
3035 
3036 		txq_advance(&q->q, ndesc);
3037 		written += ndesc;
3038 		if (unlikely(written > 32)) {
3039 			cxgb4_ring_tx_db(q->adap, &q->q, written);
3040 			written = 0;
3041 		}
3042 
3043 		/* Reacquire the Pending Send Queue Lock so we can unlink the
3044 		 * skb we've just successfully transferred to the TX Ring and
3045 		 * loop for the next skb which may be at the head of the
3046 		 * Pending Send Queue.
3047 		 */
3048 		spin_lock(&q->sendq.lock);
3049 		__skb_unlink(skb, &q->sendq);
3050 		if (is_ofld_imm(skb))
3051 			kfree_skb(skb);
3052 	}
3053 	if (likely(written))
3054 		cxgb4_ring_tx_db(q->adap, &q->q, written);
3055 
3056 	/*Indicate that no thread is processing the Pending Send Queue
3057 	 * currently.
3058 	 */
3059 	q->service_ofldq_running = false;
3060 }
3061 
3062 /**
3063  *	ofld_xmit - send a packet through an offload queue
3064  *	@q: the Tx offload queue
3065  *	@skb: the packet
3066  *
3067  *	Send an offload packet through an SGE offload queue.
3068  */
3069 static int ofld_xmit(struct sge_uld_txq *q, struct sk_buff *skb)
3070 {
3071 	skb->priority = calc_tx_flits_ofld(skb);       /* save for restart */
3072 	spin_lock(&q->sendq.lock);
3073 
3074 	/* Queue the new skb onto the Offload Queue's Pending Send Queue.  If
3075 	 * that results in this new skb being the only one on the queue, start
3076 	 * servicing it.  If there are other skbs already on the list, then
3077 	 * either the queue is currently being processed or it's been stopped
3078 	 * for some reason and it'll be restarted at a later time.  Restart
3079 	 * paths are triggered by events like experiencing a DMA Mapping Error
3080 	 * or filling the Hardware TX Ring.
3081 	 */
3082 	__skb_queue_tail(&q->sendq, skb);
3083 	if (q->sendq.qlen == 1)
3084 		service_ofldq(q);
3085 
3086 	spin_unlock(&q->sendq.lock);
3087 	return NET_XMIT_SUCCESS;
3088 }
3089 
3090 /**
3091  *	restart_ofldq - restart a suspended offload queue
3092  *	@t: pointer to the tasklet associated with this handler
3093  *
3094  *	Resumes transmission on a suspended Tx offload queue.
3095  */
3096 static void restart_ofldq(struct tasklet_struct *t)
3097 {
3098 	struct sge_uld_txq *q = from_tasklet(q, t, qresume_tsk);
3099 
3100 	spin_lock(&q->sendq.lock);
3101 	q->full = 0;            /* the queue actually is completely empty now */
3102 	service_ofldq(q);
3103 	spin_unlock(&q->sendq.lock);
3104 }
3105 
3106 /**
3107  *	skb_txq - return the Tx queue an offload packet should use
3108  *	@skb: the packet
3109  *
3110  *	Returns the Tx queue an offload packet should use as indicated by bits
3111  *	1-15 in the packet's queue_mapping.
3112  */
3113 static inline unsigned int skb_txq(const struct sk_buff *skb)
3114 {
3115 	return skb->queue_mapping >> 1;
3116 }
3117 
3118 /**
3119  *	is_ctrl_pkt - return whether an offload packet is a control packet
3120  *	@skb: the packet
3121  *
3122  *	Returns whether an offload packet should use an OFLD or a CTRL
3123  *	Tx queue as indicated by bit 0 in the packet's queue_mapping.
3124  */
3125 static inline unsigned int is_ctrl_pkt(const struct sk_buff *skb)
3126 {
3127 	return skb->queue_mapping & 1;
3128 }
3129 
3130 static inline int uld_send(struct adapter *adap, struct sk_buff *skb,
3131 			   unsigned int tx_uld_type)
3132 {
3133 	struct sge_uld_txq_info *txq_info;
3134 	struct sge_uld_txq *txq;
3135 	unsigned int idx = skb_txq(skb);
3136 
3137 	if (unlikely(is_ctrl_pkt(skb))) {
3138 		/* Single ctrl queue is a requirement for LE workaround path */
3139 		if (adap->tids.nsftids)
3140 			idx = 0;
3141 		return ctrl_xmit(&adap->sge.ctrlq[idx], skb);
3142 	}
3143 
3144 	txq_info = adap->sge.uld_txq_info[tx_uld_type];
3145 	if (unlikely(!txq_info)) {
3146 		WARN_ON(true);
3147 		kfree_skb(skb);
3148 		return NET_XMIT_DROP;
3149 	}
3150 
3151 	txq = &txq_info->uldtxq[idx];
3152 	return ofld_xmit(txq, skb);
3153 }
3154 
3155 /**
3156  *	t4_ofld_send - send an offload packet
3157  *	@adap: the adapter
3158  *	@skb: the packet
3159  *
3160  *	Sends an offload packet.  We use the packet queue_mapping to select the
3161  *	appropriate Tx queue as follows: bit 0 indicates whether the packet
3162  *	should be sent as regular or control, bits 1-15 select the queue.
3163  */
3164 int t4_ofld_send(struct adapter *adap, struct sk_buff *skb)
3165 {
3166 	int ret;
3167 
3168 	local_bh_disable();
3169 	ret = uld_send(adap, skb, CXGB4_TX_OFLD);
3170 	local_bh_enable();
3171 	return ret;
3172 }
3173 
3174 /**
3175  *	cxgb4_ofld_send - send an offload packet
3176  *	@dev: the net device
3177  *	@skb: the packet
3178  *
3179  *	Sends an offload packet.  This is an exported version of @t4_ofld_send,
3180  *	intended for ULDs.
3181  */
3182 int cxgb4_ofld_send(struct net_device *dev, struct sk_buff *skb)
3183 {
3184 	return t4_ofld_send(netdev2adap(dev), skb);
3185 }
3186 EXPORT_SYMBOL(cxgb4_ofld_send);
3187 
3188 static void *inline_tx_header(const void *src,
3189 			      const struct sge_txq *q,
3190 			      void *pos, int length)
3191 {
3192 	int left = (void *)q->stat - pos;
3193 	u64 *p;
3194 
3195 	if (likely(length <= left)) {
3196 		memcpy(pos, src, length);
3197 		pos += length;
3198 	} else {
3199 		memcpy(pos, src, left);
3200 		memcpy(q->desc, src + left, length - left);
3201 		pos = (void *)q->desc + (length - left);
3202 	}
3203 	/* 0-pad to multiple of 16 */
3204 	p = PTR_ALIGN(pos, 8);
3205 	if ((uintptr_t)p & 8) {
3206 		*p = 0;
3207 		return p + 1;
3208 	}
3209 	return p;
3210 }
3211 
3212 /**
3213  *      ofld_xmit_direct - copy a WR into offload queue
3214  *      @q: the Tx offload queue
3215  *      @src: location of WR
3216  *      @len: WR length
3217  *
3218  *      Copy an immediate WR into an uncontended SGE offload queue.
3219  */
3220 static int ofld_xmit_direct(struct sge_uld_txq *q, const void *src,
3221 			    unsigned int len)
3222 {
3223 	unsigned int ndesc;
3224 	int credits;
3225 	u64 *pos;
3226 
3227 	/* Use the lower limit as the cut-off */
3228 	if (len > MAX_IMM_OFLD_TX_DATA_WR_LEN) {
3229 		WARN_ON(1);
3230 		return NET_XMIT_DROP;
3231 	}
3232 
3233 	/* Don't return NET_XMIT_CN here as the current
3234 	 * implementation doesn't queue the request
3235 	 * using an skb when the following conditions not met
3236 	 */
3237 	if (!spin_trylock(&q->sendq.lock))
3238 		return NET_XMIT_DROP;
3239 
3240 	if (q->full || !skb_queue_empty(&q->sendq) ||
3241 	    q->service_ofldq_running) {
3242 		spin_unlock(&q->sendq.lock);
3243 		return NET_XMIT_DROP;
3244 	}
3245 	ndesc = flits_to_desc(DIV_ROUND_UP(len, 8));
3246 	credits = txq_avail(&q->q) - ndesc;
3247 	pos = (u64 *)&q->q.desc[q->q.pidx];
3248 
3249 	/* ofldtxq_stop modifies WR header in-situ */
3250 	inline_tx_header(src, &q->q, pos, len);
3251 	if (unlikely(credits < TXQ_STOP_THRES))
3252 		ofldtxq_stop(q, (struct fw_wr_hdr *)pos);
3253 	txq_advance(&q->q, ndesc);
3254 	cxgb4_ring_tx_db(q->adap, &q->q, ndesc);
3255 
3256 	spin_unlock(&q->sendq.lock);
3257 	return NET_XMIT_SUCCESS;
3258 }
3259 
3260 int cxgb4_immdata_send(struct net_device *dev, unsigned int idx,
3261 		       const void *src, unsigned int len)
3262 {
3263 	struct sge_uld_txq_info *txq_info;
3264 	struct sge_uld_txq *txq;
3265 	struct adapter *adap;
3266 	int ret;
3267 
3268 	adap = netdev2adap(dev);
3269 
3270 	local_bh_disable();
3271 	txq_info = adap->sge.uld_txq_info[CXGB4_TX_OFLD];
3272 	if (unlikely(!txq_info)) {
3273 		WARN_ON(true);
3274 		local_bh_enable();
3275 		return NET_XMIT_DROP;
3276 	}
3277 	txq = &txq_info->uldtxq[idx];
3278 
3279 	ret = ofld_xmit_direct(txq, src, len);
3280 	local_bh_enable();
3281 	return net_xmit_eval(ret);
3282 }
3283 EXPORT_SYMBOL(cxgb4_immdata_send);
3284 
3285 /**
3286  *	t4_crypto_send - send crypto packet
3287  *	@adap: the adapter
3288  *	@skb: the packet
3289  *
3290  *	Sends crypto packet.  We use the packet queue_mapping to select the
3291  *	appropriate Tx queue as follows: bit 0 indicates whether the packet
3292  *	should be sent as regular or control, bits 1-15 select the queue.
3293  */
3294 static int t4_crypto_send(struct adapter *adap, struct sk_buff *skb)
3295 {
3296 	int ret;
3297 
3298 	local_bh_disable();
3299 	ret = uld_send(adap, skb, CXGB4_TX_CRYPTO);
3300 	local_bh_enable();
3301 	return ret;
3302 }
3303 
3304 /**
3305  *	cxgb4_crypto_send - send crypto packet
3306  *	@dev: the net device
3307  *	@skb: the packet
3308  *
3309  *	Sends crypto packet.  This is an exported version of @t4_crypto_send,
3310  *	intended for ULDs.
3311  */
3312 int cxgb4_crypto_send(struct net_device *dev, struct sk_buff *skb)
3313 {
3314 	return t4_crypto_send(netdev2adap(dev), skb);
3315 }
3316 EXPORT_SYMBOL(cxgb4_crypto_send);
3317 
3318 static inline void copy_frags(struct sk_buff *skb,
3319 			      const struct pkt_gl *gl, unsigned int offset)
3320 {
3321 	int i;
3322 
3323 	/* usually there's just one frag */
3324 	__skb_fill_page_desc(skb, 0, gl->frags[0].page,
3325 			     gl->frags[0].offset + offset,
3326 			     gl->frags[0].size - offset);
3327 	skb_shinfo(skb)->nr_frags = gl->nfrags;
3328 	for (i = 1; i < gl->nfrags; i++)
3329 		__skb_fill_page_desc(skb, i, gl->frags[i].page,
3330 				     gl->frags[i].offset,
3331 				     gl->frags[i].size);
3332 
3333 	/* get a reference to the last page, we don't own it */
3334 	get_page(gl->frags[gl->nfrags - 1].page);
3335 }
3336 
3337 /**
3338  *	cxgb4_pktgl_to_skb - build an sk_buff from a packet gather list
3339  *	@gl: the gather list
3340  *	@skb_len: size of sk_buff main body if it carries fragments
3341  *	@pull_len: amount of data to move to the sk_buff's main body
3342  *
3343  *	Builds an sk_buff from the given packet gather list.  Returns the
3344  *	sk_buff or %NULL if sk_buff allocation failed.
3345  */
3346 struct sk_buff *cxgb4_pktgl_to_skb(const struct pkt_gl *gl,
3347 				   unsigned int skb_len, unsigned int pull_len)
3348 {
3349 	struct sk_buff *skb;
3350 
3351 	/*
3352 	 * Below we rely on RX_COPY_THRES being less than the smallest Rx buffer
3353 	 * size, which is expected since buffers are at least PAGE_SIZEd.
3354 	 * In this case packets up to RX_COPY_THRES have only one fragment.
3355 	 */
3356 	if (gl->tot_len <= RX_COPY_THRES) {
3357 		skb = dev_alloc_skb(gl->tot_len);
3358 		if (unlikely(!skb))
3359 			goto out;
3360 		__skb_put(skb, gl->tot_len);
3361 		skb_copy_to_linear_data(skb, gl->va, gl->tot_len);
3362 	} else {
3363 		skb = dev_alloc_skb(skb_len);
3364 		if (unlikely(!skb))
3365 			goto out;
3366 		__skb_put(skb, pull_len);
3367 		skb_copy_to_linear_data(skb, gl->va, pull_len);
3368 
3369 		copy_frags(skb, gl, pull_len);
3370 		skb->len = gl->tot_len;
3371 		skb->data_len = skb->len - pull_len;
3372 		skb->truesize += skb->data_len;
3373 	}
3374 out:	return skb;
3375 }
3376 EXPORT_SYMBOL(cxgb4_pktgl_to_skb);
3377 
3378 /**
3379  *	t4_pktgl_free - free a packet gather list
3380  *	@gl: the gather list
3381  *
3382  *	Releases the pages of a packet gather list.  We do not own the last
3383  *	page on the list and do not free it.
3384  */
3385 static void t4_pktgl_free(const struct pkt_gl *gl)
3386 {
3387 	int n;
3388 	const struct page_frag *p;
3389 
3390 	for (p = gl->frags, n = gl->nfrags - 1; n--; p++)
3391 		put_page(p->page);
3392 }
3393 
3394 /*
3395  * Process an MPS trace packet.  Give it an unused protocol number so it won't
3396  * be delivered to anyone and send it to the stack for capture.
3397  */
3398 static noinline int handle_trace_pkt(struct adapter *adap,
3399 				     const struct pkt_gl *gl)
3400 {
3401 	struct sk_buff *skb;
3402 
3403 	skb = cxgb4_pktgl_to_skb(gl, RX_PULL_LEN, RX_PULL_LEN);
3404 	if (unlikely(!skb)) {
3405 		t4_pktgl_free(gl);
3406 		return 0;
3407 	}
3408 
3409 	if (is_t4(adap->params.chip))
3410 		__skb_pull(skb, sizeof(struct cpl_trace_pkt));
3411 	else
3412 		__skb_pull(skb, sizeof(struct cpl_t5_trace_pkt));
3413 
3414 	skb_reset_mac_header(skb);
3415 	skb->protocol = htons(0xffff);
3416 	skb->dev = adap->port[0];
3417 	netif_receive_skb(skb);
3418 	return 0;
3419 }
3420 
3421 /**
3422  * cxgb4_sgetim_to_hwtstamp - convert sge time stamp to hw time stamp
3423  * @adap: the adapter
3424  * @hwtstamps: time stamp structure to update
3425  * @sgetstamp: 60bit iqe timestamp
3426  *
3427  * Every ingress queue entry has the 60-bit timestamp, convert that timestamp
3428  * which is in Core Clock ticks into ktime_t and assign it
3429  **/
3430 static void cxgb4_sgetim_to_hwtstamp(struct adapter *adap,
3431 				     struct skb_shared_hwtstamps *hwtstamps,
3432 				     u64 sgetstamp)
3433 {
3434 	u64 ns;
3435 	u64 tmp = (sgetstamp * 1000 * 1000 + adap->params.vpd.cclk / 2);
3436 
3437 	ns = div_u64(tmp, adap->params.vpd.cclk);
3438 
3439 	memset(hwtstamps, 0, sizeof(*hwtstamps));
3440 	hwtstamps->hwtstamp = ns_to_ktime(ns);
3441 }
3442 
3443 static void do_gro(struct sge_eth_rxq *rxq, const struct pkt_gl *gl,
3444 		   const struct cpl_rx_pkt *pkt, unsigned long tnl_hdr_len)
3445 {
3446 	struct adapter *adapter = rxq->rspq.adap;
3447 	struct sge *s = &adapter->sge;
3448 	struct port_info *pi;
3449 	int ret;
3450 	struct sk_buff *skb;
3451 
3452 	skb = napi_get_frags(&rxq->rspq.napi);
3453 	if (unlikely(!skb)) {
3454 		t4_pktgl_free(gl);
3455 		rxq->stats.rx_drops++;
3456 		return;
3457 	}
3458 
3459 	copy_frags(skb, gl, s->pktshift);
3460 	if (tnl_hdr_len)
3461 		skb->csum_level = 1;
3462 	skb->len = gl->tot_len - s->pktshift;
3463 	skb->data_len = skb->len;
3464 	skb->truesize += skb->data_len;
3465 	skb->ip_summed = CHECKSUM_UNNECESSARY;
3466 	skb_record_rx_queue(skb, rxq->rspq.idx);
3467 	pi = netdev_priv(skb->dev);
3468 	if (pi->rxtstamp)
3469 		cxgb4_sgetim_to_hwtstamp(adapter, skb_hwtstamps(skb),
3470 					 gl->sgetstamp);
3471 	if (rxq->rspq.netdev->features & NETIF_F_RXHASH)
3472 		skb_set_hash(skb, (__force u32)pkt->rsshdr.hash_val,
3473 			     PKT_HASH_TYPE_L3);
3474 
3475 	if (unlikely(pkt->vlan_ex)) {
3476 		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), ntohs(pkt->vlan));
3477 		rxq->stats.vlan_ex++;
3478 	}
3479 	ret = napi_gro_frags(&rxq->rspq.napi);
3480 	if (ret == GRO_HELD)
3481 		rxq->stats.lro_pkts++;
3482 	else if (ret == GRO_MERGED || ret == GRO_MERGED_FREE)
3483 		rxq->stats.lro_merged++;
3484 	rxq->stats.pkts++;
3485 	rxq->stats.rx_cso++;
3486 }
3487 
3488 enum {
3489 	RX_NON_PTP_PKT = 0,
3490 	RX_PTP_PKT_SUC = 1,
3491 	RX_PTP_PKT_ERR = 2
3492 };
3493 
3494 /**
3495  *     t4_systim_to_hwstamp - read hardware time stamp
3496  *     @adapter: the adapter
3497  *     @skb: the packet
3498  *
3499  *     Read Time Stamp from MPS packet and insert in skb which
3500  *     is forwarded to PTP application
3501  */
3502 static noinline int t4_systim_to_hwstamp(struct adapter *adapter,
3503 					 struct sk_buff *skb)
3504 {
3505 	struct skb_shared_hwtstamps *hwtstamps;
3506 	struct cpl_rx_mps_pkt *cpl = NULL;
3507 	unsigned char *data;
3508 	int offset;
3509 
3510 	cpl = (struct cpl_rx_mps_pkt *)skb->data;
3511 	if (!(CPL_RX_MPS_PKT_TYPE_G(ntohl(cpl->op_to_r1_hi)) &
3512 	     X_CPL_RX_MPS_PKT_TYPE_PTP))
3513 		return RX_PTP_PKT_ERR;
3514 
3515 	data = skb->data + sizeof(*cpl);
3516 	skb_pull(skb, 2 * sizeof(u64) + sizeof(struct cpl_rx_mps_pkt));
3517 	offset = ETH_HLEN + IPV4_HLEN(skb->data) + UDP_HLEN;
3518 	if (skb->len < offset + OFF_PTP_SEQUENCE_ID + sizeof(short))
3519 		return RX_PTP_PKT_ERR;
3520 
3521 	hwtstamps = skb_hwtstamps(skb);
3522 	memset(hwtstamps, 0, sizeof(*hwtstamps));
3523 	hwtstamps->hwtstamp = ns_to_ktime(get_unaligned_be64(data));
3524 
3525 	return RX_PTP_PKT_SUC;
3526 }
3527 
3528 /**
3529  *     t4_rx_hststamp - Recv PTP Event Message
3530  *     @adapter: the adapter
3531  *     @rsp: the response queue descriptor holding the RX_PKT message
3532  *     @rxq: the response queue holding the RX_PKT message
3533  *     @skb: the packet
3534  *
3535  *     PTP enabled and MPS packet, read HW timestamp
3536  */
3537 static int t4_rx_hststamp(struct adapter *adapter, const __be64 *rsp,
3538 			  struct sge_eth_rxq *rxq, struct sk_buff *skb)
3539 {
3540 	int ret;
3541 
3542 	if (unlikely((*(u8 *)rsp == CPL_RX_MPS_PKT) &&
3543 		     !is_t4(adapter->params.chip))) {
3544 		ret = t4_systim_to_hwstamp(adapter, skb);
3545 		if (ret == RX_PTP_PKT_ERR) {
3546 			kfree_skb(skb);
3547 			rxq->stats.rx_drops++;
3548 		}
3549 		return ret;
3550 	}
3551 	return RX_NON_PTP_PKT;
3552 }
3553 
3554 /**
3555  *      t4_tx_hststamp - Loopback PTP Transmit Event Message
3556  *      @adapter: the adapter
3557  *      @skb: the packet
3558  *      @dev: the ingress net device
3559  *
3560  *      Read hardware timestamp for the loopback PTP Tx event message
3561  */
3562 static int t4_tx_hststamp(struct adapter *adapter, struct sk_buff *skb,
3563 			  struct net_device *dev)
3564 {
3565 	struct port_info *pi = netdev_priv(dev);
3566 
3567 	if (!is_t4(adapter->params.chip) && adapter->ptp_tx_skb) {
3568 		cxgb4_ptp_read_hwstamp(adapter, pi);
3569 		kfree_skb(skb);
3570 		return 0;
3571 	}
3572 	return 1;
3573 }
3574 
3575 /**
3576  *	t4_tx_completion_handler - handle CPL_SGE_EGR_UPDATE messages
3577  *	@rspq: Ethernet RX Response Queue associated with Ethernet TX Queue
3578  *	@rsp: Response Entry pointer into Response Queue
3579  *	@gl: Gather List pointer
3580  *
3581  *	For adapters which support the SGE Doorbell Queue Timer facility,
3582  *	we configure the Ethernet TX Queues to send CIDX Updates to the
3583  *	Associated Ethernet RX Response Queue with CPL_SGE_EGR_UPDATE
3584  *	messages.  This adds a small load to PCIe Link RX bandwidth and,
3585  *	potentially, higher CPU Interrupt load, but allows us to respond
3586  *	much more quickly to the CIDX Updates.  This is important for
3587  *	Upper Layer Software which isn't willing to have a large amount
3588  *	of TX Data outstanding before receiving DMA Completions.
3589  */
3590 static void t4_tx_completion_handler(struct sge_rspq *rspq,
3591 				     const __be64 *rsp,
3592 				     const struct pkt_gl *gl)
3593 {
3594 	u8 opcode = ((const struct rss_header *)rsp)->opcode;
3595 	struct port_info *pi = netdev_priv(rspq->netdev);
3596 	struct adapter *adapter = rspq->adap;
3597 	struct sge *s = &adapter->sge;
3598 	struct sge_eth_txq *txq;
3599 
3600 	/* skip RSS header */
3601 	rsp++;
3602 
3603 	/* FW can send EGR_UPDATEs encapsulated in a CPL_FW4_MSG.
3604 	 */
3605 	if (unlikely(opcode == CPL_FW4_MSG &&
3606 		     ((const struct cpl_fw4_msg *)rsp)->type ==
3607 							FW_TYPE_RSSCPL)) {
3608 		rsp++;
3609 		opcode = ((const struct rss_header *)rsp)->opcode;
3610 		rsp++;
3611 	}
3612 
3613 	if (unlikely(opcode != CPL_SGE_EGR_UPDATE)) {
3614 		pr_info("%s: unexpected FW4/CPL %#x on Rx queue\n",
3615 			__func__, opcode);
3616 		return;
3617 	}
3618 
3619 	txq = &s->ethtxq[pi->first_qset + rspq->idx];
3620 
3621 	/* We've got the Hardware Consumer Index Update in the Egress Update
3622 	 * message. These Egress Update messages will be our sole CIDX Updates
3623 	 * we get since we don't want to chew up PCIe bandwidth for both Ingress
3624 	 * Messages and Status Page writes.  However, The code which manages
3625 	 * reclaiming successfully DMA'ed TX Work Requests uses the CIDX value
3626 	 * stored in the Status Page at the end of the TX Queue.  It's easiest
3627 	 * to simply copy the CIDX Update value from the Egress Update message
3628 	 * to the Status Page.  Also note that no Endian issues need to be
3629 	 * considered here since both are Big Endian and we're just copying
3630 	 * bytes consistently ...
3631 	 */
3632 	if (CHELSIO_CHIP_VERSION(adapter->params.chip) <= CHELSIO_T5) {
3633 		struct cpl_sge_egr_update *egr;
3634 
3635 		egr = (struct cpl_sge_egr_update *)rsp;
3636 		WRITE_ONCE(txq->q.stat->cidx, egr->cidx);
3637 	}
3638 
3639 	t4_sge_eth_txq_egress_update(adapter, txq, -1);
3640 }
3641 
3642 static int cxgb4_validate_lb_pkt(struct port_info *pi, const struct pkt_gl *si)
3643 {
3644 	struct adapter *adap = pi->adapter;
3645 	struct cxgb4_ethtool_lb_test *lb;
3646 	struct sge *s = &adap->sge;
3647 	struct net_device *netdev;
3648 	u8 *data;
3649 	int i;
3650 
3651 	netdev = adap->port[pi->port_id];
3652 	lb = &pi->ethtool_lb;
3653 	data = si->va + s->pktshift;
3654 
3655 	i = ETH_ALEN;
3656 	if (!ether_addr_equal(data + i, netdev->dev_addr))
3657 		return -1;
3658 
3659 	i += ETH_ALEN;
3660 	if (strcmp(&data[i], CXGB4_SELFTEST_LB_STR))
3661 		lb->result = -EIO;
3662 
3663 	complete(&lb->completion);
3664 	return 0;
3665 }
3666 
3667 /**
3668  *	t4_ethrx_handler - process an ingress ethernet packet
3669  *	@q: the response queue that received the packet
3670  *	@rsp: the response queue descriptor holding the RX_PKT message
3671  *	@si: the gather list of packet fragments
3672  *
3673  *	Process an ingress ethernet packet and deliver it to the stack.
3674  */
3675 int t4_ethrx_handler(struct sge_rspq *q, const __be64 *rsp,
3676 		     const struct pkt_gl *si)
3677 {
3678 	bool csum_ok;
3679 	struct sk_buff *skb;
3680 	const struct cpl_rx_pkt *pkt;
3681 	struct sge_eth_rxq *rxq = container_of(q, struct sge_eth_rxq, rspq);
3682 	struct adapter *adapter = q->adap;
3683 	struct sge *s = &q->adap->sge;
3684 	int cpl_trace_pkt = is_t4(q->adap->params.chip) ?
3685 			    CPL_TRACE_PKT : CPL_TRACE_PKT_T5;
3686 	u16 err_vec, tnl_hdr_len = 0;
3687 	struct port_info *pi;
3688 	int ret = 0;
3689 
3690 	pi = netdev_priv(q->netdev);
3691 	/* If we're looking at TX Queue CIDX Update, handle that separately
3692 	 * and return.
3693 	 */
3694 	if (unlikely((*(u8 *)rsp == CPL_FW4_MSG) ||
3695 		     (*(u8 *)rsp == CPL_SGE_EGR_UPDATE))) {
3696 		t4_tx_completion_handler(q, rsp, si);
3697 		return 0;
3698 	}
3699 
3700 	if (unlikely(*(u8 *)rsp == cpl_trace_pkt))
3701 		return handle_trace_pkt(q->adap, si);
3702 
3703 	pkt = (const struct cpl_rx_pkt *)rsp;
3704 	/* Compressed error vector is enabled for T6 only */
3705 	if (q->adap->params.tp.rx_pkt_encap) {
3706 		err_vec = T6_COMPR_RXERR_VEC_G(be16_to_cpu(pkt->err_vec));
3707 		tnl_hdr_len = T6_RX_TNLHDR_LEN_G(ntohs(pkt->err_vec));
3708 	} else {
3709 		err_vec = be16_to_cpu(pkt->err_vec);
3710 	}
3711 
3712 	csum_ok = pkt->csum_calc && !err_vec &&
3713 		  (q->netdev->features & NETIF_F_RXCSUM);
3714 
3715 	if (err_vec)
3716 		rxq->stats.bad_rx_pkts++;
3717 
3718 	if (unlikely(pi->ethtool_lb.loopback && pkt->iff >= NCHAN)) {
3719 		ret = cxgb4_validate_lb_pkt(pi, si);
3720 		if (!ret)
3721 			return 0;
3722 	}
3723 
3724 	if (((pkt->l2info & htonl(RXF_TCP_F)) ||
3725 	     tnl_hdr_len) &&
3726 	    (q->netdev->features & NETIF_F_GRO) && csum_ok && !pkt->ip_frag) {
3727 		do_gro(rxq, si, pkt, tnl_hdr_len);
3728 		return 0;
3729 	}
3730 
3731 	skb = cxgb4_pktgl_to_skb(si, RX_PKT_SKB_LEN, RX_PULL_LEN);
3732 	if (unlikely(!skb)) {
3733 		t4_pktgl_free(si);
3734 		rxq->stats.rx_drops++;
3735 		return 0;
3736 	}
3737 
3738 	/* Handle PTP Event Rx packet */
3739 	if (unlikely(pi->ptp_enable)) {
3740 		ret = t4_rx_hststamp(adapter, rsp, rxq, skb);
3741 		if (ret == RX_PTP_PKT_ERR)
3742 			return 0;
3743 	}
3744 	if (likely(!ret))
3745 		__skb_pull(skb, s->pktshift); /* remove ethernet header pad */
3746 
3747 	/* Handle the PTP Event Tx Loopback packet */
3748 	if (unlikely(pi->ptp_enable && !ret &&
3749 		     (pkt->l2info & htonl(RXF_UDP_F)) &&
3750 		     cxgb4_ptp_is_ptp_rx(skb))) {
3751 		if (!t4_tx_hststamp(adapter, skb, q->netdev))
3752 			return 0;
3753 	}
3754 
3755 	skb->protocol = eth_type_trans(skb, q->netdev);
3756 	skb_record_rx_queue(skb, q->idx);
3757 	if (skb->dev->features & NETIF_F_RXHASH)
3758 		skb_set_hash(skb, (__force u32)pkt->rsshdr.hash_val,
3759 			     PKT_HASH_TYPE_L3);
3760 
3761 	rxq->stats.pkts++;
3762 
3763 	if (pi->rxtstamp)
3764 		cxgb4_sgetim_to_hwtstamp(q->adap, skb_hwtstamps(skb),
3765 					 si->sgetstamp);
3766 	if (csum_ok && (pkt->l2info & htonl(RXF_UDP_F | RXF_TCP_F))) {
3767 		if (!pkt->ip_frag) {
3768 			skb->ip_summed = CHECKSUM_UNNECESSARY;
3769 			rxq->stats.rx_cso++;
3770 		} else if (pkt->l2info & htonl(RXF_IP_F)) {
3771 			__sum16 c = (__force __sum16)pkt->csum;
3772 			skb->csum = csum_unfold(c);
3773 
3774 			if (tnl_hdr_len) {
3775 				skb->ip_summed = CHECKSUM_UNNECESSARY;
3776 				skb->csum_level = 1;
3777 			} else {
3778 				skb->ip_summed = CHECKSUM_COMPLETE;
3779 			}
3780 			rxq->stats.rx_cso++;
3781 		}
3782 	} else {
3783 		skb_checksum_none_assert(skb);
3784 #ifdef CONFIG_CHELSIO_T4_FCOE
3785 #define CPL_RX_PKT_FLAGS (RXF_PSH_F | RXF_SYN_F | RXF_UDP_F | \
3786 			  RXF_TCP_F | RXF_IP_F | RXF_IP6_F | RXF_LRO_F)
3787 
3788 		if (!(pkt->l2info & cpu_to_be32(CPL_RX_PKT_FLAGS))) {
3789 			if ((pkt->l2info & cpu_to_be32(RXF_FCOE_F)) &&
3790 			    (pi->fcoe.flags & CXGB_FCOE_ENABLED)) {
3791 				if (q->adap->params.tp.rx_pkt_encap)
3792 					csum_ok = err_vec &
3793 						  T6_COMPR_RXERR_SUM_F;
3794 				else
3795 					csum_ok = err_vec & RXERR_CSUM_F;
3796 				if (!csum_ok)
3797 					skb->ip_summed = CHECKSUM_UNNECESSARY;
3798 			}
3799 		}
3800 
3801 #undef CPL_RX_PKT_FLAGS
3802 #endif /* CONFIG_CHELSIO_T4_FCOE */
3803 	}
3804 
3805 	if (unlikely(pkt->vlan_ex)) {
3806 		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), ntohs(pkt->vlan));
3807 		rxq->stats.vlan_ex++;
3808 	}
3809 	skb_mark_napi_id(skb, &q->napi);
3810 	netif_receive_skb(skb);
3811 	return 0;
3812 }
3813 
3814 /**
3815  *	restore_rx_bufs - put back a packet's Rx buffers
3816  *	@si: the packet gather list
3817  *	@q: the SGE free list
3818  *	@frags: number of FL buffers to restore
3819  *
3820  *	Puts back on an FL the Rx buffers associated with @si.  The buffers
3821  *	have already been unmapped and are left unmapped, we mark them so to
3822  *	prevent further unmapping attempts.
3823  *
3824  *	This function undoes a series of @unmap_rx_buf calls when we find out
3825  *	that the current packet can't be processed right away afterall and we
3826  *	need to come back to it later.  This is a very rare event and there's
3827  *	no effort to make this particularly efficient.
3828  */
3829 static void restore_rx_bufs(const struct pkt_gl *si, struct sge_fl *q,
3830 			    int frags)
3831 {
3832 	struct rx_sw_desc *d;
3833 
3834 	while (frags--) {
3835 		if (q->cidx == 0)
3836 			q->cidx = q->size - 1;
3837 		else
3838 			q->cidx--;
3839 		d = &q->sdesc[q->cidx];
3840 		d->page = si->frags[frags].page;
3841 		d->dma_addr |= RX_UNMAPPED_BUF;
3842 		q->avail++;
3843 	}
3844 }
3845 
3846 /**
3847  *	is_new_response - check if a response is newly written
3848  *	@r: the response descriptor
3849  *	@q: the response queue
3850  *
3851  *	Returns true if a response descriptor contains a yet unprocessed
3852  *	response.
3853  */
3854 static inline bool is_new_response(const struct rsp_ctrl *r,
3855 				   const struct sge_rspq *q)
3856 {
3857 	return (r->type_gen >> RSPD_GEN_S) == q->gen;
3858 }
3859 
3860 /**
3861  *	rspq_next - advance to the next entry in a response queue
3862  *	@q: the queue
3863  *
3864  *	Updates the state of a response queue to advance it to the next entry.
3865  */
3866 static inline void rspq_next(struct sge_rspq *q)
3867 {
3868 	q->cur_desc = (void *)q->cur_desc + q->iqe_len;
3869 	if (unlikely(++q->cidx == q->size)) {
3870 		q->cidx = 0;
3871 		q->gen ^= 1;
3872 		q->cur_desc = q->desc;
3873 	}
3874 }
3875 
3876 /**
3877  *	process_responses - process responses from an SGE response queue
3878  *	@q: the ingress queue to process
3879  *	@budget: how many responses can be processed in this round
3880  *
3881  *	Process responses from an SGE response queue up to the supplied budget.
3882  *	Responses include received packets as well as control messages from FW
3883  *	or HW.
3884  *
3885  *	Additionally choose the interrupt holdoff time for the next interrupt
3886  *	on this queue.  If the system is under memory shortage use a fairly
3887  *	long delay to help recovery.
3888  */
3889 static int process_responses(struct sge_rspq *q, int budget)
3890 {
3891 	int ret, rsp_type;
3892 	int budget_left = budget;
3893 	const struct rsp_ctrl *rc;
3894 	struct sge_eth_rxq *rxq = container_of(q, struct sge_eth_rxq, rspq);
3895 	struct adapter *adapter = q->adap;
3896 	struct sge *s = &adapter->sge;
3897 
3898 	while (likely(budget_left)) {
3899 		rc = (void *)q->cur_desc + (q->iqe_len - sizeof(*rc));
3900 		if (!is_new_response(rc, q)) {
3901 			if (q->flush_handler)
3902 				q->flush_handler(q);
3903 			break;
3904 		}
3905 
3906 		dma_rmb();
3907 		rsp_type = RSPD_TYPE_G(rc->type_gen);
3908 		if (likely(rsp_type == RSPD_TYPE_FLBUF_X)) {
3909 			struct page_frag *fp;
3910 			struct pkt_gl si;
3911 			const struct rx_sw_desc *rsd;
3912 			u32 len = ntohl(rc->pldbuflen_qid), bufsz, frags;
3913 
3914 			if (len & RSPD_NEWBUF_F) {
3915 				if (likely(q->offset > 0)) {
3916 					free_rx_bufs(q->adap, &rxq->fl, 1);
3917 					q->offset = 0;
3918 				}
3919 				len = RSPD_LEN_G(len);
3920 			}
3921 			si.tot_len = len;
3922 
3923 			/* gather packet fragments */
3924 			for (frags = 0, fp = si.frags; ; frags++, fp++) {
3925 				rsd = &rxq->fl.sdesc[rxq->fl.cidx];
3926 				bufsz = get_buf_size(adapter, rsd);
3927 				fp->page = rsd->page;
3928 				fp->offset = q->offset;
3929 				fp->size = min(bufsz, len);
3930 				len -= fp->size;
3931 				if (!len)
3932 					break;
3933 				unmap_rx_buf(q->adap, &rxq->fl);
3934 			}
3935 
3936 			si.sgetstamp = SGE_TIMESTAMP_G(
3937 					be64_to_cpu(rc->last_flit));
3938 			/*
3939 			 * Last buffer remains mapped so explicitly make it
3940 			 * coherent for CPU access.
3941 			 */
3942 			dma_sync_single_for_cpu(q->adap->pdev_dev,
3943 						get_buf_addr(rsd),
3944 						fp->size, DMA_FROM_DEVICE);
3945 
3946 			si.va = page_address(si.frags[0].page) +
3947 				si.frags[0].offset;
3948 			prefetch(si.va);
3949 
3950 			si.nfrags = frags + 1;
3951 			ret = q->handler(q, q->cur_desc, &si);
3952 			if (likely(ret == 0))
3953 				q->offset += ALIGN(fp->size, s->fl_align);
3954 			else
3955 				restore_rx_bufs(&si, &rxq->fl, frags);
3956 		} else if (likely(rsp_type == RSPD_TYPE_CPL_X)) {
3957 			ret = q->handler(q, q->cur_desc, NULL);
3958 		} else {
3959 			ret = q->handler(q, (const __be64 *)rc, CXGB4_MSG_AN);
3960 		}
3961 
3962 		if (unlikely(ret)) {
3963 			/* couldn't process descriptor, back off for recovery */
3964 			q->next_intr_params = QINTR_TIMER_IDX_V(NOMEM_TMR_IDX);
3965 			break;
3966 		}
3967 
3968 		rspq_next(q);
3969 		budget_left--;
3970 	}
3971 
3972 	if (q->offset >= 0 && fl_cap(&rxq->fl) - rxq->fl.avail >= 16)
3973 		__refill_fl(q->adap, &rxq->fl);
3974 	return budget - budget_left;
3975 }
3976 
3977 /**
3978  *	napi_rx_handler - the NAPI handler for Rx processing
3979  *	@napi: the napi instance
3980  *	@budget: how many packets we can process in this round
3981  *
3982  *	Handler for new data events when using NAPI.  This does not need any
3983  *	locking or protection from interrupts as data interrupts are off at
3984  *	this point and other adapter interrupts do not interfere (the latter
3985  *	in not a concern at all with MSI-X as non-data interrupts then have
3986  *	a separate handler).
3987  */
3988 static int napi_rx_handler(struct napi_struct *napi, int budget)
3989 {
3990 	unsigned int params;
3991 	struct sge_rspq *q = container_of(napi, struct sge_rspq, napi);
3992 	int work_done;
3993 	u32 val;
3994 
3995 	work_done = process_responses(q, budget);
3996 	if (likely(work_done < budget)) {
3997 		int timer_index;
3998 
3999 		napi_complete_done(napi, work_done);
4000 		timer_index = QINTR_TIMER_IDX_G(q->next_intr_params);
4001 
4002 		if (q->adaptive_rx) {
4003 			if (work_done > max(timer_pkt_quota[timer_index],
4004 					    MIN_NAPI_WORK))
4005 				timer_index = (timer_index + 1);
4006 			else
4007 				timer_index = timer_index - 1;
4008 
4009 			timer_index = clamp(timer_index, 0, SGE_TIMERREGS - 1);
4010 			q->next_intr_params =
4011 					QINTR_TIMER_IDX_V(timer_index) |
4012 					QINTR_CNT_EN_V(0);
4013 			params = q->next_intr_params;
4014 		} else {
4015 			params = q->next_intr_params;
4016 			q->next_intr_params = q->intr_params;
4017 		}
4018 	} else
4019 		params = QINTR_TIMER_IDX_V(7);
4020 
4021 	val = CIDXINC_V(work_done) | SEINTARM_V(params);
4022 
4023 	/* If we don't have access to the new User GTS (T5+), use the old
4024 	 * doorbell mechanism; otherwise use the new BAR2 mechanism.
4025 	 */
4026 	if (unlikely(q->bar2_addr == NULL)) {
4027 		t4_write_reg(q->adap, MYPF_REG(SGE_PF_GTS_A),
4028 			     val | INGRESSQID_V((u32)q->cntxt_id));
4029 	} else {
4030 		writel(val | INGRESSQID_V(q->bar2_qid),
4031 		       q->bar2_addr + SGE_UDB_GTS);
4032 		wmb();
4033 	}
4034 	return work_done;
4035 }
4036 
4037 void cxgb4_ethofld_restart(struct tasklet_struct *t)
4038 {
4039 	struct sge_eosw_txq *eosw_txq = from_tasklet(eosw_txq, t,
4040 						     qresume_tsk);
4041 	int pktcount;
4042 
4043 	spin_lock(&eosw_txq->lock);
4044 	pktcount = eosw_txq->cidx - eosw_txq->last_cidx;
4045 	if (pktcount < 0)
4046 		pktcount += eosw_txq->ndesc;
4047 
4048 	if (pktcount) {
4049 		cxgb4_eosw_txq_free_desc(netdev2adap(eosw_txq->netdev),
4050 					 eosw_txq, pktcount);
4051 		eosw_txq->inuse -= pktcount;
4052 	}
4053 
4054 	/* There may be some packets waiting for completions. So,
4055 	 * attempt to send these packets now.
4056 	 */
4057 	ethofld_xmit(eosw_txq->netdev, eosw_txq);
4058 	spin_unlock(&eosw_txq->lock);
4059 }
4060 
4061 /* cxgb4_ethofld_rx_handler - Process ETHOFLD Tx completions
4062  * @q: the response queue that received the packet
4063  * @rsp: the response queue descriptor holding the CPL message
4064  * @si: the gather list of packet fragments
4065  *
4066  * Process a ETHOFLD Tx completion. Increment the cidx here, but
4067  * free up the descriptors in a tasklet later.
4068  */
4069 int cxgb4_ethofld_rx_handler(struct sge_rspq *q, const __be64 *rsp,
4070 			     const struct pkt_gl *si)
4071 {
4072 	u8 opcode = ((const struct rss_header *)rsp)->opcode;
4073 
4074 	/* skip RSS header */
4075 	rsp++;
4076 
4077 	if (opcode == CPL_FW4_ACK) {
4078 		const struct cpl_fw4_ack *cpl;
4079 		struct sge_eosw_txq *eosw_txq;
4080 		struct eotid_entry *entry;
4081 		struct sk_buff *skb;
4082 		u32 hdr_len, eotid;
4083 		u8 flits, wrlen16;
4084 		int credits;
4085 
4086 		cpl = (const struct cpl_fw4_ack *)rsp;
4087 		eotid = CPL_FW4_ACK_FLOWID_G(ntohl(OPCODE_TID(cpl))) -
4088 			q->adap->tids.eotid_base;
4089 		entry = cxgb4_lookup_eotid(&q->adap->tids, eotid);
4090 		if (!entry)
4091 			goto out_done;
4092 
4093 		eosw_txq = (struct sge_eosw_txq *)entry->data;
4094 		if (!eosw_txq)
4095 			goto out_done;
4096 
4097 		spin_lock(&eosw_txq->lock);
4098 		credits = cpl->credits;
4099 		while (credits > 0) {
4100 			skb = eosw_txq->desc[eosw_txq->cidx].skb;
4101 			if (!skb)
4102 				break;
4103 
4104 			if (unlikely((eosw_txq->state ==
4105 				      CXGB4_EO_STATE_FLOWC_OPEN_REPLY ||
4106 				      eosw_txq->state ==
4107 				      CXGB4_EO_STATE_FLOWC_CLOSE_REPLY) &&
4108 				     eosw_txq->cidx == eosw_txq->flowc_idx)) {
4109 				flits = DIV_ROUND_UP(skb->len, 8);
4110 				if (eosw_txq->state ==
4111 				    CXGB4_EO_STATE_FLOWC_OPEN_REPLY)
4112 					eosw_txq->state = CXGB4_EO_STATE_ACTIVE;
4113 				else
4114 					eosw_txq->state = CXGB4_EO_STATE_CLOSED;
4115 				complete(&eosw_txq->completion);
4116 			} else {
4117 				hdr_len = eth_get_headlen(eosw_txq->netdev,
4118 							  skb->data,
4119 							  skb_headlen(skb));
4120 				flits = ethofld_calc_tx_flits(q->adap, skb,
4121 							      hdr_len);
4122 			}
4123 			eosw_txq_advance_index(&eosw_txq->cidx, 1,
4124 					       eosw_txq->ndesc);
4125 			wrlen16 = DIV_ROUND_UP(flits * 8, 16);
4126 			credits -= wrlen16;
4127 		}
4128 
4129 		eosw_txq->cred += cpl->credits;
4130 		eosw_txq->ncompl--;
4131 
4132 		spin_unlock(&eosw_txq->lock);
4133 
4134 		/* Schedule a tasklet to reclaim SKBs and restart ETHOFLD Tx,
4135 		 * if there were packets waiting for completion.
4136 		 */
4137 		tasklet_schedule(&eosw_txq->qresume_tsk);
4138 	}
4139 
4140 out_done:
4141 	return 0;
4142 }
4143 
4144 /*
4145  * The MSI-X interrupt handler for an SGE response queue.
4146  */
4147 irqreturn_t t4_sge_intr_msix(int irq, void *cookie)
4148 {
4149 	struct sge_rspq *q = cookie;
4150 
4151 	napi_schedule(&q->napi);
4152 	return IRQ_HANDLED;
4153 }
4154 
4155 /*
4156  * Process the indirect interrupt entries in the interrupt queue and kick off
4157  * NAPI for each queue that has generated an entry.
4158  */
4159 static unsigned int process_intrq(struct adapter *adap)
4160 {
4161 	unsigned int credits;
4162 	const struct rsp_ctrl *rc;
4163 	struct sge_rspq *q = &adap->sge.intrq;
4164 	u32 val;
4165 
4166 	spin_lock(&adap->sge.intrq_lock);
4167 	for (credits = 0; ; credits++) {
4168 		rc = (void *)q->cur_desc + (q->iqe_len - sizeof(*rc));
4169 		if (!is_new_response(rc, q))
4170 			break;
4171 
4172 		dma_rmb();
4173 		if (RSPD_TYPE_G(rc->type_gen) == RSPD_TYPE_INTR_X) {
4174 			unsigned int qid = ntohl(rc->pldbuflen_qid);
4175 
4176 			qid -= adap->sge.ingr_start;
4177 			napi_schedule(&adap->sge.ingr_map[qid]->napi);
4178 		}
4179 
4180 		rspq_next(q);
4181 	}
4182 
4183 	val =  CIDXINC_V(credits) | SEINTARM_V(q->intr_params);
4184 
4185 	/* If we don't have access to the new User GTS (T5+), use the old
4186 	 * doorbell mechanism; otherwise use the new BAR2 mechanism.
4187 	 */
4188 	if (unlikely(q->bar2_addr == NULL)) {
4189 		t4_write_reg(adap, MYPF_REG(SGE_PF_GTS_A),
4190 			     val | INGRESSQID_V(q->cntxt_id));
4191 	} else {
4192 		writel(val | INGRESSQID_V(q->bar2_qid),
4193 		       q->bar2_addr + SGE_UDB_GTS);
4194 		wmb();
4195 	}
4196 	spin_unlock(&adap->sge.intrq_lock);
4197 	return credits;
4198 }
4199 
4200 /*
4201  * The MSI interrupt handler, which handles data events from SGE response queues
4202  * as well as error and other async events as they all use the same MSI vector.
4203  */
4204 static irqreturn_t t4_intr_msi(int irq, void *cookie)
4205 {
4206 	struct adapter *adap = cookie;
4207 
4208 	if (adap->flags & CXGB4_MASTER_PF)
4209 		t4_slow_intr_handler(adap);
4210 	process_intrq(adap);
4211 	return IRQ_HANDLED;
4212 }
4213 
4214 /*
4215  * Interrupt handler for legacy INTx interrupts.
4216  * Handles data events from SGE response queues as well as error and other
4217  * async events as they all use the same interrupt line.
4218  */
4219 static irqreturn_t t4_intr_intx(int irq, void *cookie)
4220 {
4221 	struct adapter *adap = cookie;
4222 
4223 	t4_write_reg(adap, MYPF_REG(PCIE_PF_CLI_A), 0);
4224 	if (((adap->flags & CXGB4_MASTER_PF) && t4_slow_intr_handler(adap)) |
4225 	    process_intrq(adap))
4226 		return IRQ_HANDLED;
4227 	return IRQ_NONE;             /* probably shared interrupt */
4228 }
4229 
4230 /**
4231  *	t4_intr_handler - select the top-level interrupt handler
4232  *	@adap: the adapter
4233  *
4234  *	Selects the top-level interrupt handler based on the type of interrupts
4235  *	(MSI-X, MSI, or INTx).
4236  */
4237 irq_handler_t t4_intr_handler(struct adapter *adap)
4238 {
4239 	if (adap->flags & CXGB4_USING_MSIX)
4240 		return t4_sge_intr_msix;
4241 	if (adap->flags & CXGB4_USING_MSI)
4242 		return t4_intr_msi;
4243 	return t4_intr_intx;
4244 }
4245 
4246 static void sge_rx_timer_cb(struct timer_list *t)
4247 {
4248 	unsigned long m;
4249 	unsigned int i;
4250 	struct adapter *adap = from_timer(adap, t, sge.rx_timer);
4251 	struct sge *s = &adap->sge;
4252 
4253 	for (i = 0; i < BITS_TO_LONGS(s->egr_sz); i++)
4254 		for (m = s->starving_fl[i]; m; m &= m - 1) {
4255 			struct sge_eth_rxq *rxq;
4256 			unsigned int id = __ffs(m) + i * BITS_PER_LONG;
4257 			struct sge_fl *fl = s->egr_map[id];
4258 
4259 			clear_bit(id, s->starving_fl);
4260 			smp_mb__after_atomic();
4261 
4262 			if (fl_starving(adap, fl)) {
4263 				rxq = container_of(fl, struct sge_eth_rxq, fl);
4264 				if (napi_reschedule(&rxq->rspq.napi))
4265 					fl->starving++;
4266 				else
4267 					set_bit(id, s->starving_fl);
4268 			}
4269 		}
4270 	/* The remainder of the SGE RX Timer Callback routine is dedicated to
4271 	 * global Master PF activities like checking for chip ingress stalls,
4272 	 * etc.
4273 	 */
4274 	if (!(adap->flags & CXGB4_MASTER_PF))
4275 		goto done;
4276 
4277 	t4_idma_monitor(adap, &s->idma_monitor, HZ, RX_QCHECK_PERIOD);
4278 
4279 done:
4280 	mod_timer(&s->rx_timer, jiffies + RX_QCHECK_PERIOD);
4281 }
4282 
4283 static void sge_tx_timer_cb(struct timer_list *t)
4284 {
4285 	struct adapter *adap = from_timer(adap, t, sge.tx_timer);
4286 	struct sge *s = &adap->sge;
4287 	unsigned long m, period;
4288 	unsigned int i, budget;
4289 
4290 	for (i = 0; i < BITS_TO_LONGS(s->egr_sz); i++)
4291 		for (m = s->txq_maperr[i]; m; m &= m - 1) {
4292 			unsigned long id = __ffs(m) + i * BITS_PER_LONG;
4293 			struct sge_uld_txq *txq = s->egr_map[id];
4294 
4295 			clear_bit(id, s->txq_maperr);
4296 			tasklet_schedule(&txq->qresume_tsk);
4297 		}
4298 
4299 	if (!is_t4(adap->params.chip)) {
4300 		struct sge_eth_txq *q = &s->ptptxq;
4301 		int avail;
4302 
4303 		spin_lock(&adap->ptp_lock);
4304 		avail = reclaimable(&q->q);
4305 
4306 		if (avail) {
4307 			free_tx_desc(adap, &q->q, avail, false);
4308 			q->q.in_use -= avail;
4309 		}
4310 		spin_unlock(&adap->ptp_lock);
4311 	}
4312 
4313 	budget = MAX_TIMER_TX_RECLAIM;
4314 	i = s->ethtxq_rover;
4315 	do {
4316 		budget -= t4_sge_eth_txq_egress_update(adap, &s->ethtxq[i],
4317 						       budget);
4318 		if (!budget)
4319 			break;
4320 
4321 		if (++i >= s->ethqsets)
4322 			i = 0;
4323 	} while (i != s->ethtxq_rover);
4324 	s->ethtxq_rover = i;
4325 
4326 	if (budget == 0) {
4327 		/* If we found too many reclaimable packets schedule a timer
4328 		 * in the near future to continue where we left off.
4329 		 */
4330 		period = 2;
4331 	} else {
4332 		/* We reclaimed all reclaimable TX Descriptors, so reschedule
4333 		 * at the normal period.
4334 		 */
4335 		period = TX_QCHECK_PERIOD;
4336 	}
4337 
4338 	mod_timer(&s->tx_timer, jiffies + period);
4339 }
4340 
4341 /**
4342  *	bar2_address - return the BAR2 address for an SGE Queue's Registers
4343  *	@adapter: the adapter
4344  *	@qid: the SGE Queue ID
4345  *	@qtype: the SGE Queue Type (Egress or Ingress)
4346  *	@pbar2_qid: BAR2 Queue ID or 0 for Queue ID inferred SGE Queues
4347  *
4348  *	Returns the BAR2 address for the SGE Queue Registers associated with
4349  *	@qid.  If BAR2 SGE Registers aren't available, returns NULL.  Also
4350  *	returns the BAR2 Queue ID to be used with writes to the BAR2 SGE
4351  *	Queue Registers.  If the BAR2 Queue ID is 0, then "Inferred Queue ID"
4352  *	Registers are supported (e.g. the Write Combining Doorbell Buffer).
4353  */
4354 static void __iomem *bar2_address(struct adapter *adapter,
4355 				  unsigned int qid,
4356 				  enum t4_bar2_qtype qtype,
4357 				  unsigned int *pbar2_qid)
4358 {
4359 	u64 bar2_qoffset;
4360 	int ret;
4361 
4362 	ret = t4_bar2_sge_qregs(adapter, qid, qtype, 0,
4363 				&bar2_qoffset, pbar2_qid);
4364 	if (ret)
4365 		return NULL;
4366 
4367 	return adapter->bar2 + bar2_qoffset;
4368 }
4369 
4370 /* @intr_idx: MSI/MSI-X vector if >=0, -(absolute qid + 1) if < 0
4371  * @cong: < 0 -> no congestion feedback, >= 0 -> congestion channel map
4372  */
4373 int t4_sge_alloc_rxq(struct adapter *adap, struct sge_rspq *iq, bool fwevtq,
4374 		     struct net_device *dev, int intr_idx,
4375 		     struct sge_fl *fl, rspq_handler_t hnd,
4376 		     rspq_flush_handler_t flush_hnd, int cong)
4377 {
4378 	int ret, flsz = 0;
4379 	struct fw_iq_cmd c;
4380 	struct sge *s = &adap->sge;
4381 	struct port_info *pi = netdev_priv(dev);
4382 	int relaxed = !(adap->flags & CXGB4_ROOT_NO_RELAXED_ORDERING);
4383 
4384 	/* Size needs to be multiple of 16, including status entry. */
4385 	iq->size = roundup(iq->size, 16);
4386 
4387 	iq->desc = alloc_ring(adap->pdev_dev, iq->size, iq->iqe_len, 0,
4388 			      &iq->phys_addr, NULL, 0,
4389 			      dev_to_node(adap->pdev_dev));
4390 	if (!iq->desc)
4391 		return -ENOMEM;
4392 
4393 	memset(&c, 0, sizeof(c));
4394 	c.op_to_vfn = htonl(FW_CMD_OP_V(FW_IQ_CMD) | FW_CMD_REQUEST_F |
4395 			    FW_CMD_WRITE_F | FW_CMD_EXEC_F |
4396 			    FW_IQ_CMD_PFN_V(adap->pf) | FW_IQ_CMD_VFN_V(0));
4397 	c.alloc_to_len16 = htonl(FW_IQ_CMD_ALLOC_F | FW_IQ_CMD_IQSTART_F |
4398 				 FW_LEN16(c));
4399 	c.type_to_iqandstindex = htonl(FW_IQ_CMD_TYPE_V(FW_IQ_TYPE_FL_INT_CAP) |
4400 		FW_IQ_CMD_IQASYNCH_V(fwevtq) | FW_IQ_CMD_VIID_V(pi->viid) |
4401 		FW_IQ_CMD_IQANDST_V(intr_idx < 0) |
4402 		FW_IQ_CMD_IQANUD_V(UPDATEDELIVERY_INTERRUPT_X) |
4403 		FW_IQ_CMD_IQANDSTINDEX_V(intr_idx >= 0 ? intr_idx :
4404 							-intr_idx - 1));
4405 	c.iqdroprss_to_iqesize = htons(FW_IQ_CMD_IQPCIECH_V(pi->tx_chan) |
4406 		FW_IQ_CMD_IQGTSMODE_F |
4407 		FW_IQ_CMD_IQINTCNTTHRESH_V(iq->pktcnt_idx) |
4408 		FW_IQ_CMD_IQESIZE_V(ilog2(iq->iqe_len) - 4));
4409 	c.iqsize = htons(iq->size);
4410 	c.iqaddr = cpu_to_be64(iq->phys_addr);
4411 	if (cong >= 0)
4412 		c.iqns_to_fl0congen = htonl(FW_IQ_CMD_IQFLINTCONGEN_F |
4413 				FW_IQ_CMD_IQTYPE_V(cong ? FW_IQ_IQTYPE_NIC
4414 							:  FW_IQ_IQTYPE_OFLD));
4415 
4416 	if (fl) {
4417 		unsigned int chip_ver =
4418 			CHELSIO_CHIP_VERSION(adap->params.chip);
4419 
4420 		/* Allocate the ring for the hardware free list (with space
4421 		 * for its status page) along with the associated software
4422 		 * descriptor ring.  The free list size needs to be a multiple
4423 		 * of the Egress Queue Unit and at least 2 Egress Units larger
4424 		 * than the SGE's Egress Congrestion Threshold
4425 		 * (fl_starve_thres - 1).
4426 		 */
4427 		if (fl->size < s->fl_starve_thres - 1 + 2 * 8)
4428 			fl->size = s->fl_starve_thres - 1 + 2 * 8;
4429 		fl->size = roundup(fl->size, 8);
4430 		fl->desc = alloc_ring(adap->pdev_dev, fl->size, sizeof(__be64),
4431 				      sizeof(struct rx_sw_desc), &fl->addr,
4432 				      &fl->sdesc, s->stat_len,
4433 				      dev_to_node(adap->pdev_dev));
4434 		if (!fl->desc)
4435 			goto fl_nomem;
4436 
4437 		flsz = fl->size / 8 + s->stat_len / sizeof(struct tx_desc);
4438 		c.iqns_to_fl0congen |= htonl(FW_IQ_CMD_FL0PACKEN_F |
4439 					     FW_IQ_CMD_FL0FETCHRO_V(relaxed) |
4440 					     FW_IQ_CMD_FL0DATARO_V(relaxed) |
4441 					     FW_IQ_CMD_FL0PADEN_F);
4442 		if (cong >= 0)
4443 			c.iqns_to_fl0congen |=
4444 				htonl(FW_IQ_CMD_FL0CNGCHMAP_V(cong) |
4445 				      FW_IQ_CMD_FL0CONGCIF_F |
4446 				      FW_IQ_CMD_FL0CONGEN_F);
4447 		/* In T6, for egress queue type FL there is internal overhead
4448 		 * of 16B for header going into FLM module.  Hence the maximum
4449 		 * allowed burst size is 448 bytes.  For T4/T5, the hardware
4450 		 * doesn't coalesce fetch requests if more than 64 bytes of
4451 		 * Free List pointers are provided, so we use a 128-byte Fetch
4452 		 * Burst Minimum there (T6 implements coalescing so we can use
4453 		 * the smaller 64-byte value there).
4454 		 */
4455 		c.fl0dcaen_to_fl0cidxfthresh =
4456 			htons(FW_IQ_CMD_FL0FBMIN_V(chip_ver <= CHELSIO_T5 ?
4457 						   FETCHBURSTMIN_128B_X :
4458 						   FETCHBURSTMIN_64B_T6_X) |
4459 			      FW_IQ_CMD_FL0FBMAX_V((chip_ver <= CHELSIO_T5) ?
4460 						   FETCHBURSTMAX_512B_X :
4461 						   FETCHBURSTMAX_256B_X));
4462 		c.fl0size = htons(flsz);
4463 		c.fl0addr = cpu_to_be64(fl->addr);
4464 	}
4465 
4466 	ret = t4_wr_mbox(adap, adap->mbox, &c, sizeof(c), &c);
4467 	if (ret)
4468 		goto err;
4469 
4470 	netif_napi_add(dev, &iq->napi, napi_rx_handler);
4471 	iq->cur_desc = iq->desc;
4472 	iq->cidx = 0;
4473 	iq->gen = 1;
4474 	iq->next_intr_params = iq->intr_params;
4475 	iq->cntxt_id = ntohs(c.iqid);
4476 	iq->abs_id = ntohs(c.physiqid);
4477 	iq->bar2_addr = bar2_address(adap,
4478 				     iq->cntxt_id,
4479 				     T4_BAR2_QTYPE_INGRESS,
4480 				     &iq->bar2_qid);
4481 	iq->size--;                           /* subtract status entry */
4482 	iq->netdev = dev;
4483 	iq->handler = hnd;
4484 	iq->flush_handler = flush_hnd;
4485 
4486 	memset(&iq->lro_mgr, 0, sizeof(struct t4_lro_mgr));
4487 	skb_queue_head_init(&iq->lro_mgr.lroq);
4488 
4489 	/* set offset to -1 to distinguish ingress queues without FL */
4490 	iq->offset = fl ? 0 : -1;
4491 
4492 	adap->sge.ingr_map[iq->cntxt_id - adap->sge.ingr_start] = iq;
4493 
4494 	if (fl) {
4495 		fl->cntxt_id = ntohs(c.fl0id);
4496 		fl->avail = fl->pend_cred = 0;
4497 		fl->pidx = fl->cidx = 0;
4498 		fl->alloc_failed = fl->large_alloc_failed = fl->starving = 0;
4499 		adap->sge.egr_map[fl->cntxt_id - adap->sge.egr_start] = fl;
4500 
4501 		/* Note, we must initialize the BAR2 Free List User Doorbell
4502 		 * information before refilling the Free List!
4503 		 */
4504 		fl->bar2_addr = bar2_address(adap,
4505 					     fl->cntxt_id,
4506 					     T4_BAR2_QTYPE_EGRESS,
4507 					     &fl->bar2_qid);
4508 		refill_fl(adap, fl, fl_cap(fl), GFP_KERNEL);
4509 	}
4510 
4511 	/* For T5 and later we attempt to set up the Congestion Manager values
4512 	 * of the new RX Ethernet Queue.  This should really be handled by
4513 	 * firmware because it's more complex than any host driver wants to
4514 	 * get involved with and it's different per chip and this is almost
4515 	 * certainly wrong.  Firmware would be wrong as well, but it would be
4516 	 * a lot easier to fix in one place ...  For now we do something very
4517 	 * simple (and hopefully less wrong).
4518 	 */
4519 	if (!is_t4(adap->params.chip) && cong >= 0) {
4520 		u32 param, val, ch_map = 0;
4521 		int i;
4522 		u16 cng_ch_bits_log = adap->params.arch.cng_ch_bits_log;
4523 
4524 		param = (FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_DMAQ) |
4525 			 FW_PARAMS_PARAM_X_V(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
4526 			 FW_PARAMS_PARAM_YZ_V(iq->cntxt_id));
4527 		if (cong == 0) {
4528 			val = CONMCTXT_CNGTPMODE_V(CONMCTXT_CNGTPMODE_QUEUE_X);
4529 		} else {
4530 			val =
4531 			    CONMCTXT_CNGTPMODE_V(CONMCTXT_CNGTPMODE_CHANNEL_X);
4532 			for (i = 0; i < 4; i++) {
4533 				if (cong & (1 << i))
4534 					ch_map |= 1 << (i << cng_ch_bits_log);
4535 			}
4536 			val |= CONMCTXT_CNGCHMAP_V(ch_map);
4537 		}
4538 		ret = t4_set_params(adap, adap->mbox, adap->pf, 0, 1,
4539 				    &param, &val);
4540 		if (ret)
4541 			dev_warn(adap->pdev_dev, "Failed to set Congestion"
4542 				 " Manager Context for Ingress Queue %d: %d\n",
4543 				 iq->cntxt_id, -ret);
4544 	}
4545 
4546 	return 0;
4547 
4548 fl_nomem:
4549 	ret = -ENOMEM;
4550 err:
4551 	if (iq->desc) {
4552 		dma_free_coherent(adap->pdev_dev, iq->size * iq->iqe_len,
4553 				  iq->desc, iq->phys_addr);
4554 		iq->desc = NULL;
4555 	}
4556 	if (fl && fl->desc) {
4557 		kfree(fl->sdesc);
4558 		fl->sdesc = NULL;
4559 		dma_free_coherent(adap->pdev_dev, flsz * sizeof(struct tx_desc),
4560 				  fl->desc, fl->addr);
4561 		fl->desc = NULL;
4562 	}
4563 	return ret;
4564 }
4565 
4566 static void init_txq(struct adapter *adap, struct sge_txq *q, unsigned int id)
4567 {
4568 	q->cntxt_id = id;
4569 	q->bar2_addr = bar2_address(adap,
4570 				    q->cntxt_id,
4571 				    T4_BAR2_QTYPE_EGRESS,
4572 				    &q->bar2_qid);
4573 	q->in_use = 0;
4574 	q->cidx = q->pidx = 0;
4575 	q->stops = q->restarts = 0;
4576 	q->stat = (void *)&q->desc[q->size];
4577 	spin_lock_init(&q->db_lock);
4578 	adap->sge.egr_map[id - adap->sge.egr_start] = q;
4579 }
4580 
4581 /**
4582  *	t4_sge_alloc_eth_txq - allocate an Ethernet TX Queue
4583  *	@adap: the adapter
4584  *	@txq: the SGE Ethernet TX Queue to initialize
4585  *	@dev: the Linux Network Device
4586  *	@netdevq: the corresponding Linux TX Queue
4587  *	@iqid: the Ingress Queue to which to deliver CIDX Update messages
4588  *	@dbqt: whether this TX Queue will use the SGE Doorbell Queue Timers
4589  */
4590 int t4_sge_alloc_eth_txq(struct adapter *adap, struct sge_eth_txq *txq,
4591 			 struct net_device *dev, struct netdev_queue *netdevq,
4592 			 unsigned int iqid, u8 dbqt)
4593 {
4594 	unsigned int chip_ver = CHELSIO_CHIP_VERSION(adap->params.chip);
4595 	struct port_info *pi = netdev_priv(dev);
4596 	struct sge *s = &adap->sge;
4597 	struct fw_eq_eth_cmd c;
4598 	int ret, nentries;
4599 
4600 	/* Add status entries */
4601 	nentries = txq->q.size + s->stat_len / sizeof(struct tx_desc);
4602 
4603 	txq->q.desc = alloc_ring(adap->pdev_dev, txq->q.size,
4604 			sizeof(struct tx_desc), sizeof(struct tx_sw_desc),
4605 			&txq->q.phys_addr, &txq->q.sdesc, s->stat_len,
4606 			netdev_queue_numa_node_read(netdevq));
4607 	if (!txq->q.desc)
4608 		return -ENOMEM;
4609 
4610 	memset(&c, 0, sizeof(c));
4611 	c.op_to_vfn = htonl(FW_CMD_OP_V(FW_EQ_ETH_CMD) | FW_CMD_REQUEST_F |
4612 			    FW_CMD_WRITE_F | FW_CMD_EXEC_F |
4613 			    FW_EQ_ETH_CMD_PFN_V(adap->pf) |
4614 			    FW_EQ_ETH_CMD_VFN_V(0));
4615 	c.alloc_to_len16 = htonl(FW_EQ_ETH_CMD_ALLOC_F |
4616 				 FW_EQ_ETH_CMD_EQSTART_F | FW_LEN16(c));
4617 
4618 	/* For TX Ethernet Queues using the SGE Doorbell Queue Timer
4619 	 * mechanism, we use Ingress Queue messages for Hardware Consumer
4620 	 * Index Updates on the TX Queue.  Otherwise we have the Hardware
4621 	 * write the CIDX Updates into the Status Page at the end of the
4622 	 * TX Queue.
4623 	 */
4624 	c.autoequiqe_to_viid = htonl(((chip_ver <= CHELSIO_T5) ?
4625 				      FW_EQ_ETH_CMD_AUTOEQUIQE_F :
4626 				      FW_EQ_ETH_CMD_AUTOEQUEQE_F) |
4627 				     FW_EQ_ETH_CMD_VIID_V(pi->viid));
4628 
4629 	c.fetchszm_to_iqid =
4630 		htonl(FW_EQ_ETH_CMD_HOSTFCMODE_V((chip_ver <= CHELSIO_T5) ?
4631 						 HOSTFCMODE_INGRESS_QUEUE_X :
4632 						 HOSTFCMODE_STATUS_PAGE_X) |
4633 		      FW_EQ_ETH_CMD_PCIECHN_V(pi->tx_chan) |
4634 		      FW_EQ_ETH_CMD_FETCHRO_F | FW_EQ_ETH_CMD_IQID_V(iqid));
4635 
4636 	/* Note that the CIDX Flush Threshold should match MAX_TX_RECLAIM. */
4637 	c.dcaen_to_eqsize =
4638 		htonl(FW_EQ_ETH_CMD_FBMIN_V(chip_ver <= CHELSIO_T5
4639 					    ? FETCHBURSTMIN_64B_X
4640 					    : FETCHBURSTMIN_64B_T6_X) |
4641 		      FW_EQ_ETH_CMD_FBMAX_V(FETCHBURSTMAX_512B_X) |
4642 		      FW_EQ_ETH_CMD_CIDXFTHRESH_V(CIDXFLUSHTHRESH_32_X) |
4643 		      FW_EQ_ETH_CMD_CIDXFTHRESHO_V(chip_ver == CHELSIO_T5) |
4644 		      FW_EQ_ETH_CMD_EQSIZE_V(nentries));
4645 
4646 	c.eqaddr = cpu_to_be64(txq->q.phys_addr);
4647 
4648 	/* If we're using the SGE Doorbell Queue Timer mechanism, pass in the
4649 	 * currently configured Timer Index.  THis can be changed later via an
4650 	 * ethtool -C tx-usecs {Timer Val} command.  Note that the SGE
4651 	 * Doorbell Queue mode is currently automatically enabled in the
4652 	 * Firmware by setting either AUTOEQUEQE or AUTOEQUIQE ...
4653 	 */
4654 	if (dbqt)
4655 		c.timeren_timerix =
4656 			cpu_to_be32(FW_EQ_ETH_CMD_TIMEREN_F |
4657 				    FW_EQ_ETH_CMD_TIMERIX_V(txq->dbqtimerix));
4658 
4659 	ret = t4_wr_mbox(adap, adap->mbox, &c, sizeof(c), &c);
4660 	if (ret) {
4661 		kfree(txq->q.sdesc);
4662 		txq->q.sdesc = NULL;
4663 		dma_free_coherent(adap->pdev_dev,
4664 				  nentries * sizeof(struct tx_desc),
4665 				  txq->q.desc, txq->q.phys_addr);
4666 		txq->q.desc = NULL;
4667 		return ret;
4668 	}
4669 
4670 	txq->q.q_type = CXGB4_TXQ_ETH;
4671 	init_txq(adap, &txq->q, FW_EQ_ETH_CMD_EQID_G(ntohl(c.eqid_pkd)));
4672 	txq->txq = netdevq;
4673 	txq->tso = 0;
4674 	txq->uso = 0;
4675 	txq->tx_cso = 0;
4676 	txq->vlan_ins = 0;
4677 	txq->mapping_err = 0;
4678 	txq->dbqt = dbqt;
4679 
4680 	return 0;
4681 }
4682 
4683 int t4_sge_alloc_ctrl_txq(struct adapter *adap, struct sge_ctrl_txq *txq,
4684 			  struct net_device *dev, unsigned int iqid,
4685 			  unsigned int cmplqid)
4686 {
4687 	unsigned int chip_ver = CHELSIO_CHIP_VERSION(adap->params.chip);
4688 	struct port_info *pi = netdev_priv(dev);
4689 	struct sge *s = &adap->sge;
4690 	struct fw_eq_ctrl_cmd c;
4691 	int ret, nentries;
4692 
4693 	/* Add status entries */
4694 	nentries = txq->q.size + s->stat_len / sizeof(struct tx_desc);
4695 
4696 	txq->q.desc = alloc_ring(adap->pdev_dev, nentries,
4697 				 sizeof(struct tx_desc), 0, &txq->q.phys_addr,
4698 				 NULL, 0, dev_to_node(adap->pdev_dev));
4699 	if (!txq->q.desc)
4700 		return -ENOMEM;
4701 
4702 	c.op_to_vfn = htonl(FW_CMD_OP_V(FW_EQ_CTRL_CMD) | FW_CMD_REQUEST_F |
4703 			    FW_CMD_WRITE_F | FW_CMD_EXEC_F |
4704 			    FW_EQ_CTRL_CMD_PFN_V(adap->pf) |
4705 			    FW_EQ_CTRL_CMD_VFN_V(0));
4706 	c.alloc_to_len16 = htonl(FW_EQ_CTRL_CMD_ALLOC_F |
4707 				 FW_EQ_CTRL_CMD_EQSTART_F | FW_LEN16(c));
4708 	c.cmpliqid_eqid = htonl(FW_EQ_CTRL_CMD_CMPLIQID_V(cmplqid));
4709 	c.physeqid_pkd = htonl(0);
4710 	c.fetchszm_to_iqid =
4711 		htonl(FW_EQ_CTRL_CMD_HOSTFCMODE_V(HOSTFCMODE_STATUS_PAGE_X) |
4712 		      FW_EQ_CTRL_CMD_PCIECHN_V(pi->tx_chan) |
4713 		      FW_EQ_CTRL_CMD_FETCHRO_F | FW_EQ_CTRL_CMD_IQID_V(iqid));
4714 	c.dcaen_to_eqsize =
4715 		htonl(FW_EQ_CTRL_CMD_FBMIN_V(chip_ver <= CHELSIO_T5
4716 					     ? FETCHBURSTMIN_64B_X
4717 					     : FETCHBURSTMIN_64B_T6_X) |
4718 		      FW_EQ_CTRL_CMD_FBMAX_V(FETCHBURSTMAX_512B_X) |
4719 		      FW_EQ_CTRL_CMD_CIDXFTHRESH_V(CIDXFLUSHTHRESH_32_X) |
4720 		      FW_EQ_CTRL_CMD_EQSIZE_V(nentries));
4721 	c.eqaddr = cpu_to_be64(txq->q.phys_addr);
4722 
4723 	ret = t4_wr_mbox(adap, adap->mbox, &c, sizeof(c), &c);
4724 	if (ret) {
4725 		dma_free_coherent(adap->pdev_dev,
4726 				  nentries * sizeof(struct tx_desc),
4727 				  txq->q.desc, txq->q.phys_addr);
4728 		txq->q.desc = NULL;
4729 		return ret;
4730 	}
4731 
4732 	txq->q.q_type = CXGB4_TXQ_CTRL;
4733 	init_txq(adap, &txq->q, FW_EQ_CTRL_CMD_EQID_G(ntohl(c.cmpliqid_eqid)));
4734 	txq->adap = adap;
4735 	skb_queue_head_init(&txq->sendq);
4736 	tasklet_setup(&txq->qresume_tsk, restart_ctrlq);
4737 	txq->full = 0;
4738 	return 0;
4739 }
4740 
4741 int t4_sge_mod_ctrl_txq(struct adapter *adap, unsigned int eqid,
4742 			unsigned int cmplqid)
4743 {
4744 	u32 param, val;
4745 
4746 	param = (FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_DMAQ) |
4747 		 FW_PARAMS_PARAM_X_V(FW_PARAMS_PARAM_DMAQ_EQ_CMPLIQID_CTRL) |
4748 		 FW_PARAMS_PARAM_YZ_V(eqid));
4749 	val = cmplqid;
4750 	return t4_set_params(adap, adap->mbox, adap->pf, 0, 1, &param, &val);
4751 }
4752 
4753 static int t4_sge_alloc_ofld_txq(struct adapter *adap, struct sge_txq *q,
4754 				 struct net_device *dev, u32 cmd, u32 iqid)
4755 {
4756 	unsigned int chip_ver = CHELSIO_CHIP_VERSION(adap->params.chip);
4757 	struct port_info *pi = netdev_priv(dev);
4758 	struct sge *s = &adap->sge;
4759 	struct fw_eq_ofld_cmd c;
4760 	u32 fb_min, nentries;
4761 	int ret;
4762 
4763 	/* Add status entries */
4764 	nentries = q->size + s->stat_len / sizeof(struct tx_desc);
4765 	q->desc = alloc_ring(adap->pdev_dev, q->size, sizeof(struct tx_desc),
4766 			     sizeof(struct tx_sw_desc), &q->phys_addr,
4767 			     &q->sdesc, s->stat_len, NUMA_NO_NODE);
4768 	if (!q->desc)
4769 		return -ENOMEM;
4770 
4771 	if (chip_ver <= CHELSIO_T5)
4772 		fb_min = FETCHBURSTMIN_64B_X;
4773 	else
4774 		fb_min = FETCHBURSTMIN_64B_T6_X;
4775 
4776 	memset(&c, 0, sizeof(c));
4777 	c.op_to_vfn = htonl(FW_CMD_OP_V(cmd) | FW_CMD_REQUEST_F |
4778 			    FW_CMD_WRITE_F | FW_CMD_EXEC_F |
4779 			    FW_EQ_OFLD_CMD_PFN_V(adap->pf) |
4780 			    FW_EQ_OFLD_CMD_VFN_V(0));
4781 	c.alloc_to_len16 = htonl(FW_EQ_OFLD_CMD_ALLOC_F |
4782 				 FW_EQ_OFLD_CMD_EQSTART_F | FW_LEN16(c));
4783 	c.fetchszm_to_iqid =
4784 		htonl(FW_EQ_OFLD_CMD_HOSTFCMODE_V(HOSTFCMODE_STATUS_PAGE_X) |
4785 		      FW_EQ_OFLD_CMD_PCIECHN_V(pi->tx_chan) |
4786 		      FW_EQ_OFLD_CMD_FETCHRO_F | FW_EQ_OFLD_CMD_IQID_V(iqid));
4787 	c.dcaen_to_eqsize =
4788 		htonl(FW_EQ_OFLD_CMD_FBMIN_V(fb_min) |
4789 		      FW_EQ_OFLD_CMD_FBMAX_V(FETCHBURSTMAX_512B_X) |
4790 		      FW_EQ_OFLD_CMD_CIDXFTHRESH_V(CIDXFLUSHTHRESH_32_X) |
4791 		      FW_EQ_OFLD_CMD_EQSIZE_V(nentries));
4792 	c.eqaddr = cpu_to_be64(q->phys_addr);
4793 
4794 	ret = t4_wr_mbox(adap, adap->mbox, &c, sizeof(c), &c);
4795 	if (ret) {
4796 		kfree(q->sdesc);
4797 		q->sdesc = NULL;
4798 		dma_free_coherent(adap->pdev_dev,
4799 				  nentries * sizeof(struct tx_desc),
4800 				  q->desc, q->phys_addr);
4801 		q->desc = NULL;
4802 		return ret;
4803 	}
4804 
4805 	init_txq(adap, q, FW_EQ_OFLD_CMD_EQID_G(ntohl(c.eqid_pkd)));
4806 	return 0;
4807 }
4808 
4809 int t4_sge_alloc_uld_txq(struct adapter *adap, struct sge_uld_txq *txq,
4810 			 struct net_device *dev, unsigned int iqid,
4811 			 unsigned int uld_type)
4812 {
4813 	u32 cmd = FW_EQ_OFLD_CMD;
4814 	int ret;
4815 
4816 	if (unlikely(uld_type == CXGB4_TX_CRYPTO))
4817 		cmd = FW_EQ_CTRL_CMD;
4818 
4819 	ret = t4_sge_alloc_ofld_txq(adap, &txq->q, dev, cmd, iqid);
4820 	if (ret)
4821 		return ret;
4822 
4823 	txq->q.q_type = CXGB4_TXQ_ULD;
4824 	txq->adap = adap;
4825 	skb_queue_head_init(&txq->sendq);
4826 	tasklet_setup(&txq->qresume_tsk, restart_ofldq);
4827 	txq->full = 0;
4828 	txq->mapping_err = 0;
4829 	return 0;
4830 }
4831 
4832 int t4_sge_alloc_ethofld_txq(struct adapter *adap, struct sge_eohw_txq *txq,
4833 			     struct net_device *dev, u32 iqid)
4834 {
4835 	int ret;
4836 
4837 	ret = t4_sge_alloc_ofld_txq(adap, &txq->q, dev, FW_EQ_OFLD_CMD, iqid);
4838 	if (ret)
4839 		return ret;
4840 
4841 	txq->q.q_type = CXGB4_TXQ_ULD;
4842 	spin_lock_init(&txq->lock);
4843 	txq->adap = adap;
4844 	txq->tso = 0;
4845 	txq->uso = 0;
4846 	txq->tx_cso = 0;
4847 	txq->vlan_ins = 0;
4848 	txq->mapping_err = 0;
4849 	return 0;
4850 }
4851 
4852 void free_txq(struct adapter *adap, struct sge_txq *q)
4853 {
4854 	struct sge *s = &adap->sge;
4855 
4856 	dma_free_coherent(adap->pdev_dev,
4857 			  q->size * sizeof(struct tx_desc) + s->stat_len,
4858 			  q->desc, q->phys_addr);
4859 	q->cntxt_id = 0;
4860 	q->sdesc = NULL;
4861 	q->desc = NULL;
4862 }
4863 
4864 void free_rspq_fl(struct adapter *adap, struct sge_rspq *rq,
4865 		  struct sge_fl *fl)
4866 {
4867 	struct sge *s = &adap->sge;
4868 	unsigned int fl_id = fl ? fl->cntxt_id : 0xffff;
4869 
4870 	adap->sge.ingr_map[rq->cntxt_id - adap->sge.ingr_start] = NULL;
4871 	t4_iq_free(adap, adap->mbox, adap->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
4872 		   rq->cntxt_id, fl_id, 0xffff);
4873 	dma_free_coherent(adap->pdev_dev, (rq->size + 1) * rq->iqe_len,
4874 			  rq->desc, rq->phys_addr);
4875 	netif_napi_del(&rq->napi);
4876 	rq->netdev = NULL;
4877 	rq->cntxt_id = rq->abs_id = 0;
4878 	rq->desc = NULL;
4879 
4880 	if (fl) {
4881 		free_rx_bufs(adap, fl, fl->avail);
4882 		dma_free_coherent(adap->pdev_dev, fl->size * 8 + s->stat_len,
4883 				  fl->desc, fl->addr);
4884 		kfree(fl->sdesc);
4885 		fl->sdesc = NULL;
4886 		fl->cntxt_id = 0;
4887 		fl->desc = NULL;
4888 	}
4889 }
4890 
4891 /**
4892  *      t4_free_ofld_rxqs - free a block of consecutive Rx queues
4893  *      @adap: the adapter
4894  *      @n: number of queues
4895  *      @q: pointer to first queue
4896  *
4897  *      Release the resources of a consecutive block of offload Rx queues.
4898  */
4899 void t4_free_ofld_rxqs(struct adapter *adap, int n, struct sge_ofld_rxq *q)
4900 {
4901 	for ( ; n; n--, q++)
4902 		if (q->rspq.desc)
4903 			free_rspq_fl(adap, &q->rspq,
4904 				     q->fl.size ? &q->fl : NULL);
4905 }
4906 
4907 void t4_sge_free_ethofld_txq(struct adapter *adap, struct sge_eohw_txq *txq)
4908 {
4909 	if (txq->q.desc) {
4910 		t4_ofld_eq_free(adap, adap->mbox, adap->pf, 0,
4911 				txq->q.cntxt_id);
4912 		free_tx_desc(adap, &txq->q, txq->q.in_use, false);
4913 		kfree(txq->q.sdesc);
4914 		free_txq(adap, &txq->q);
4915 	}
4916 }
4917 
4918 /**
4919  *	t4_free_sge_resources - free SGE resources
4920  *	@adap: the adapter
4921  *
4922  *	Frees resources used by the SGE queue sets.
4923  */
4924 void t4_free_sge_resources(struct adapter *adap)
4925 {
4926 	int i;
4927 	struct sge_eth_rxq *eq;
4928 	struct sge_eth_txq *etq;
4929 
4930 	/* stop all Rx queues in order to start them draining */
4931 	for (i = 0; i < adap->sge.ethqsets; i++) {
4932 		eq = &adap->sge.ethrxq[i];
4933 		if (eq->rspq.desc)
4934 			t4_iq_stop(adap, adap->mbox, adap->pf, 0,
4935 				   FW_IQ_TYPE_FL_INT_CAP,
4936 				   eq->rspq.cntxt_id,
4937 				   eq->fl.size ? eq->fl.cntxt_id : 0xffff,
4938 				   0xffff);
4939 	}
4940 
4941 	/* clean up Ethernet Tx/Rx queues */
4942 	for (i = 0; i < adap->sge.ethqsets; i++) {
4943 		eq = &adap->sge.ethrxq[i];
4944 		if (eq->rspq.desc)
4945 			free_rspq_fl(adap, &eq->rspq,
4946 				     eq->fl.size ? &eq->fl : NULL);
4947 		if (eq->msix) {
4948 			cxgb4_free_msix_idx_in_bmap(adap, eq->msix->idx);
4949 			eq->msix = NULL;
4950 		}
4951 
4952 		etq = &adap->sge.ethtxq[i];
4953 		if (etq->q.desc) {
4954 			t4_eth_eq_free(adap, adap->mbox, adap->pf, 0,
4955 				       etq->q.cntxt_id);
4956 			__netif_tx_lock_bh(etq->txq);
4957 			free_tx_desc(adap, &etq->q, etq->q.in_use, true);
4958 			__netif_tx_unlock_bh(etq->txq);
4959 			kfree(etq->q.sdesc);
4960 			free_txq(adap, &etq->q);
4961 		}
4962 	}
4963 
4964 	/* clean up control Tx queues */
4965 	for (i = 0; i < ARRAY_SIZE(adap->sge.ctrlq); i++) {
4966 		struct sge_ctrl_txq *cq = &adap->sge.ctrlq[i];
4967 
4968 		if (cq->q.desc) {
4969 			tasklet_kill(&cq->qresume_tsk);
4970 			t4_ctrl_eq_free(adap, adap->mbox, adap->pf, 0,
4971 					cq->q.cntxt_id);
4972 			__skb_queue_purge(&cq->sendq);
4973 			free_txq(adap, &cq->q);
4974 		}
4975 	}
4976 
4977 	if (adap->sge.fw_evtq.desc) {
4978 		free_rspq_fl(adap, &adap->sge.fw_evtq, NULL);
4979 		if (adap->sge.fwevtq_msix_idx >= 0)
4980 			cxgb4_free_msix_idx_in_bmap(adap,
4981 						    adap->sge.fwevtq_msix_idx);
4982 	}
4983 
4984 	if (adap->sge.nd_msix_idx >= 0)
4985 		cxgb4_free_msix_idx_in_bmap(adap, adap->sge.nd_msix_idx);
4986 
4987 	if (adap->sge.intrq.desc)
4988 		free_rspq_fl(adap, &adap->sge.intrq, NULL);
4989 
4990 	if (!is_t4(adap->params.chip)) {
4991 		etq = &adap->sge.ptptxq;
4992 		if (etq->q.desc) {
4993 			t4_eth_eq_free(adap, adap->mbox, adap->pf, 0,
4994 				       etq->q.cntxt_id);
4995 			spin_lock_bh(&adap->ptp_lock);
4996 			free_tx_desc(adap, &etq->q, etq->q.in_use, true);
4997 			spin_unlock_bh(&adap->ptp_lock);
4998 			kfree(etq->q.sdesc);
4999 			free_txq(adap, &etq->q);
5000 		}
5001 	}
5002 
5003 	/* clear the reverse egress queue map */
5004 	memset(adap->sge.egr_map, 0,
5005 	       adap->sge.egr_sz * sizeof(*adap->sge.egr_map));
5006 }
5007 
5008 void t4_sge_start(struct adapter *adap)
5009 {
5010 	adap->sge.ethtxq_rover = 0;
5011 	mod_timer(&adap->sge.rx_timer, jiffies + RX_QCHECK_PERIOD);
5012 	mod_timer(&adap->sge.tx_timer, jiffies + TX_QCHECK_PERIOD);
5013 }
5014 
5015 /**
5016  *	t4_sge_stop - disable SGE operation
5017  *	@adap: the adapter
5018  *
5019  *	Stop tasklets and timers associated with the DMA engine.  Note that
5020  *	this is effective only if measures have been taken to disable any HW
5021  *	events that may restart them.
5022  */
5023 void t4_sge_stop(struct adapter *adap)
5024 {
5025 	int i;
5026 	struct sge *s = &adap->sge;
5027 
5028 	if (s->rx_timer.function)
5029 		del_timer_sync(&s->rx_timer);
5030 	if (s->tx_timer.function)
5031 		del_timer_sync(&s->tx_timer);
5032 
5033 	if (is_offload(adap)) {
5034 		struct sge_uld_txq_info *txq_info;
5035 
5036 		txq_info = adap->sge.uld_txq_info[CXGB4_TX_OFLD];
5037 		if (txq_info) {
5038 			struct sge_uld_txq *txq = txq_info->uldtxq;
5039 
5040 			for_each_ofldtxq(&adap->sge, i) {
5041 				if (txq->q.desc)
5042 					tasklet_kill(&txq->qresume_tsk);
5043 			}
5044 		}
5045 	}
5046 
5047 	if (is_pci_uld(adap)) {
5048 		struct sge_uld_txq_info *txq_info;
5049 
5050 		txq_info = adap->sge.uld_txq_info[CXGB4_TX_CRYPTO];
5051 		if (txq_info) {
5052 			struct sge_uld_txq *txq = txq_info->uldtxq;
5053 
5054 			for_each_ofldtxq(&adap->sge, i) {
5055 				if (txq->q.desc)
5056 					tasklet_kill(&txq->qresume_tsk);
5057 			}
5058 		}
5059 	}
5060 
5061 	for (i = 0; i < ARRAY_SIZE(s->ctrlq); i++) {
5062 		struct sge_ctrl_txq *cq = &s->ctrlq[i];
5063 
5064 		if (cq->q.desc)
5065 			tasklet_kill(&cq->qresume_tsk);
5066 	}
5067 }
5068 
5069 /**
5070  *	t4_sge_init_soft - grab core SGE values needed by SGE code
5071  *	@adap: the adapter
5072  *
5073  *	We need to grab the SGE operating parameters that we need to have
5074  *	in order to do our job and make sure we can live with them.
5075  */
5076 
5077 static int t4_sge_init_soft(struct adapter *adap)
5078 {
5079 	struct sge *s = &adap->sge;
5080 	u32 fl_small_pg, fl_large_pg, fl_small_mtu, fl_large_mtu;
5081 	u32 timer_value_0_and_1, timer_value_2_and_3, timer_value_4_and_5;
5082 	u32 ingress_rx_threshold;
5083 
5084 	/*
5085 	 * Verify that CPL messages are going to the Ingress Queue for
5086 	 * process_responses() and that only packet data is going to the
5087 	 * Free Lists.
5088 	 */
5089 	if ((t4_read_reg(adap, SGE_CONTROL_A) & RXPKTCPLMODE_F) !=
5090 	    RXPKTCPLMODE_V(RXPKTCPLMODE_SPLIT_X)) {
5091 		dev_err(adap->pdev_dev, "bad SGE CPL MODE\n");
5092 		return -EINVAL;
5093 	}
5094 
5095 	/*
5096 	 * Validate the Host Buffer Register Array indices that we want to
5097 	 * use ...
5098 	 *
5099 	 * XXX Note that we should really read through the Host Buffer Size
5100 	 * XXX register array and find the indices of the Buffer Sizes which
5101 	 * XXX meet our needs!
5102 	 */
5103 	#define READ_FL_BUF(x) \
5104 		t4_read_reg(adap, SGE_FL_BUFFER_SIZE0_A+(x)*sizeof(u32))
5105 
5106 	fl_small_pg = READ_FL_BUF(RX_SMALL_PG_BUF);
5107 	fl_large_pg = READ_FL_BUF(RX_LARGE_PG_BUF);
5108 	fl_small_mtu = READ_FL_BUF(RX_SMALL_MTU_BUF);
5109 	fl_large_mtu = READ_FL_BUF(RX_LARGE_MTU_BUF);
5110 
5111 	/* We only bother using the Large Page logic if the Large Page Buffer
5112 	 * is larger than our Page Size Buffer.
5113 	 */
5114 	if (fl_large_pg <= fl_small_pg)
5115 		fl_large_pg = 0;
5116 
5117 	#undef READ_FL_BUF
5118 
5119 	/* The Page Size Buffer must be exactly equal to our Page Size and the
5120 	 * Large Page Size Buffer should be 0 (per above) or a power of 2.
5121 	 */
5122 	if (fl_small_pg != PAGE_SIZE ||
5123 	    (fl_large_pg & (fl_large_pg-1)) != 0) {
5124 		dev_err(adap->pdev_dev, "bad SGE FL page buffer sizes [%d, %d]\n",
5125 			fl_small_pg, fl_large_pg);
5126 		return -EINVAL;
5127 	}
5128 	if (fl_large_pg)
5129 		s->fl_pg_order = ilog2(fl_large_pg) - PAGE_SHIFT;
5130 
5131 	if (fl_small_mtu < FL_MTU_SMALL_BUFSIZE(adap) ||
5132 	    fl_large_mtu < FL_MTU_LARGE_BUFSIZE(adap)) {
5133 		dev_err(adap->pdev_dev, "bad SGE FL MTU sizes [%d, %d]\n",
5134 			fl_small_mtu, fl_large_mtu);
5135 		return -EINVAL;
5136 	}
5137 
5138 	/*
5139 	 * Retrieve our RX interrupt holdoff timer values and counter
5140 	 * threshold values from the SGE parameters.
5141 	 */
5142 	timer_value_0_and_1 = t4_read_reg(adap, SGE_TIMER_VALUE_0_AND_1_A);
5143 	timer_value_2_and_3 = t4_read_reg(adap, SGE_TIMER_VALUE_2_AND_3_A);
5144 	timer_value_4_and_5 = t4_read_reg(adap, SGE_TIMER_VALUE_4_AND_5_A);
5145 	s->timer_val[0] = core_ticks_to_us(adap,
5146 		TIMERVALUE0_G(timer_value_0_and_1));
5147 	s->timer_val[1] = core_ticks_to_us(adap,
5148 		TIMERVALUE1_G(timer_value_0_and_1));
5149 	s->timer_val[2] = core_ticks_to_us(adap,
5150 		TIMERVALUE2_G(timer_value_2_and_3));
5151 	s->timer_val[3] = core_ticks_to_us(adap,
5152 		TIMERVALUE3_G(timer_value_2_and_3));
5153 	s->timer_val[4] = core_ticks_to_us(adap,
5154 		TIMERVALUE4_G(timer_value_4_and_5));
5155 	s->timer_val[5] = core_ticks_to_us(adap,
5156 		TIMERVALUE5_G(timer_value_4_and_5));
5157 
5158 	ingress_rx_threshold = t4_read_reg(adap, SGE_INGRESS_RX_THRESHOLD_A);
5159 	s->counter_val[0] = THRESHOLD_0_G(ingress_rx_threshold);
5160 	s->counter_val[1] = THRESHOLD_1_G(ingress_rx_threshold);
5161 	s->counter_val[2] = THRESHOLD_2_G(ingress_rx_threshold);
5162 	s->counter_val[3] = THRESHOLD_3_G(ingress_rx_threshold);
5163 
5164 	return 0;
5165 }
5166 
5167 /**
5168  *     t4_sge_init - initialize SGE
5169  *     @adap: the adapter
5170  *
5171  *     Perform low-level SGE code initialization needed every time after a
5172  *     chip reset.
5173  */
5174 int t4_sge_init(struct adapter *adap)
5175 {
5176 	struct sge *s = &adap->sge;
5177 	u32 sge_control, sge_conm_ctrl;
5178 	int ret, egress_threshold;
5179 
5180 	/*
5181 	 * Ingress Padding Boundary and Egress Status Page Size are set up by
5182 	 * t4_fixup_host_params().
5183 	 */
5184 	sge_control = t4_read_reg(adap, SGE_CONTROL_A);
5185 	s->pktshift = PKTSHIFT_G(sge_control);
5186 	s->stat_len = (sge_control & EGRSTATUSPAGESIZE_F) ? 128 : 64;
5187 
5188 	s->fl_align = t4_fl_pkt_align(adap);
5189 	ret = t4_sge_init_soft(adap);
5190 	if (ret < 0)
5191 		return ret;
5192 
5193 	/*
5194 	 * A FL with <= fl_starve_thres buffers is starving and a periodic
5195 	 * timer will attempt to refill it.  This needs to be larger than the
5196 	 * SGE's Egress Congestion Threshold.  If it isn't, then we can get
5197 	 * stuck waiting for new packets while the SGE is waiting for us to
5198 	 * give it more Free List entries.  (Note that the SGE's Egress
5199 	 * Congestion Threshold is in units of 2 Free List pointers.) For T4,
5200 	 * there was only a single field to control this.  For T5 there's the
5201 	 * original field which now only applies to Unpacked Mode Free List
5202 	 * buffers and a new field which only applies to Packed Mode Free List
5203 	 * buffers.
5204 	 */
5205 	sge_conm_ctrl = t4_read_reg(adap, SGE_CONM_CTRL_A);
5206 	switch (CHELSIO_CHIP_VERSION(adap->params.chip)) {
5207 	case CHELSIO_T4:
5208 		egress_threshold = EGRTHRESHOLD_G(sge_conm_ctrl);
5209 		break;
5210 	case CHELSIO_T5:
5211 		egress_threshold = EGRTHRESHOLDPACKING_G(sge_conm_ctrl);
5212 		break;
5213 	case CHELSIO_T6:
5214 		egress_threshold = T6_EGRTHRESHOLDPACKING_G(sge_conm_ctrl);
5215 		break;
5216 	default:
5217 		dev_err(adap->pdev_dev, "Unsupported Chip version %d\n",
5218 			CHELSIO_CHIP_VERSION(adap->params.chip));
5219 		return -EINVAL;
5220 	}
5221 	s->fl_starve_thres = 2*egress_threshold + 1;
5222 
5223 	t4_idma_monitor_init(adap, &s->idma_monitor);
5224 
5225 	/* Set up timers used for recuring callbacks to process RX and TX
5226 	 * administrative tasks.
5227 	 */
5228 	timer_setup(&s->rx_timer, sge_rx_timer_cb, 0);
5229 	timer_setup(&s->tx_timer, sge_tx_timer_cb, 0);
5230 
5231 	spin_lock_init(&s->intrq_lock);
5232 
5233 	return 0;
5234 }
5235