1 /*
2  * This file is part of the Chelsio T4 PCI-E SR-IOV Virtual Function Ethernet
3  * driver for Linux.
4  *
5  * Copyright (c) 2009-2010 Chelsio Communications, Inc. All rights reserved.
6  *
7  * This software is available to you under a choice of one of two
8  * licenses.  You may choose to be licensed under the terms of the GNU
9  * General Public License (GPL) Version 2, available from the file
10  * COPYING in the main directory of this source tree, or the
11  * OpenIB.org BSD license below:
12  *
13  *     Redistribution and use in source and binary forms, with or
14  *     without modification, are permitted provided that the following
15  *     conditions are met:
16  *
17  *      - Redistributions of source code must retain the above
18  *        copyright notice, this list of conditions and the following
19  *        disclaimer.
20  *
21  *      - Redistributions in binary form must reproduce the above
22  *        copyright notice, this list of conditions and the following
23  *        disclaimer in the documentation and/or other materials
24  *        provided with the distribution.
25  *
26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33  * SOFTWARE.
34  */
35 
36 #include <linux/skbuff.h>
37 #include <linux/netdevice.h>
38 #include <linux/etherdevice.h>
39 #include <linux/if_vlan.h>
40 #include <linux/ip.h>
41 #include <net/ipv6.h>
42 #include <net/tcp.h>
43 #include <linux/dma-mapping.h>
44 #include <linux/prefetch.h>
45 
46 #include "t4vf_common.h"
47 #include "t4vf_defs.h"
48 
49 #include "../cxgb4/t4_regs.h"
50 #include "../cxgb4/t4_values.h"
51 #include "../cxgb4/t4fw_api.h"
52 #include "../cxgb4/t4_msg.h"
53 
54 /*
55  * Constants ...
56  */
57 enum {
58 	/*
59 	 * Egress Queue sizes, producer and consumer indices are all in units
60 	 * of Egress Context Units bytes.  Note that as far as the hardware is
61 	 * concerned, the free list is an Egress Queue (the host produces free
62 	 * buffers which the hardware consumes) and free list entries are
63 	 * 64-bit PCI DMA addresses.
64 	 */
65 	EQ_UNIT = SGE_EQ_IDXSIZE,
66 	FL_PER_EQ_UNIT = EQ_UNIT / sizeof(__be64),
67 	TXD_PER_EQ_UNIT = EQ_UNIT / sizeof(__be64),
68 
69 	/*
70 	 * Max number of TX descriptors we clean up at a time.  Should be
71 	 * modest as freeing skbs isn't cheap and it happens while holding
72 	 * locks.  We just need to free packets faster than they arrive, we
73 	 * eventually catch up and keep the amortized cost reasonable.
74 	 */
75 	MAX_TX_RECLAIM = 16,
76 
77 	/*
78 	 * Max number of Rx buffers we replenish at a time.  Again keep this
79 	 * modest, allocating buffers isn't cheap either.
80 	 */
81 	MAX_RX_REFILL = 16,
82 
83 	/*
84 	 * Period of the Rx queue check timer.  This timer is infrequent as it
85 	 * has something to do only when the system experiences severe memory
86 	 * shortage.
87 	 */
88 	RX_QCHECK_PERIOD = (HZ / 2),
89 
90 	/*
91 	 * Period of the TX queue check timer and the maximum number of TX
92 	 * descriptors to be reclaimed by the TX timer.
93 	 */
94 	TX_QCHECK_PERIOD = (HZ / 2),
95 	MAX_TIMER_TX_RECLAIM = 100,
96 
97 	/*
98 	 * Suspend an Ethernet TX queue with fewer available descriptors than
99 	 * this.  We always want to have room for a maximum sized packet:
100 	 * inline immediate data + MAX_SKB_FRAGS. This is the same as
101 	 * calc_tx_flits() for a TSO packet with nr_frags == MAX_SKB_FRAGS
102 	 * (see that function and its helpers for a description of the
103 	 * calculation).
104 	 */
105 	ETHTXQ_MAX_FRAGS = MAX_SKB_FRAGS + 1,
106 	ETHTXQ_MAX_SGL_LEN = ((3 * (ETHTXQ_MAX_FRAGS-1))/2 +
107 				   ((ETHTXQ_MAX_FRAGS-1) & 1) +
108 				   2),
109 	ETHTXQ_MAX_HDR = (sizeof(struct fw_eth_tx_pkt_vm_wr) +
110 			  sizeof(struct cpl_tx_pkt_lso_core) +
111 			  sizeof(struct cpl_tx_pkt_core)) / sizeof(__be64),
112 	ETHTXQ_MAX_FLITS = ETHTXQ_MAX_SGL_LEN + ETHTXQ_MAX_HDR,
113 
114 	ETHTXQ_STOP_THRES = 1 + DIV_ROUND_UP(ETHTXQ_MAX_FLITS, TXD_PER_EQ_UNIT),
115 
116 	/*
117 	 * Max TX descriptor space we allow for an Ethernet packet to be
118 	 * inlined into a WR.  This is limited by the maximum value which
119 	 * we can specify for immediate data in the firmware Ethernet TX
120 	 * Work Request.
121 	 */
122 	MAX_IMM_TX_PKT_LEN = FW_WR_IMMDLEN_M,
123 
124 	/*
125 	 * Max size of a WR sent through a control TX queue.
126 	 */
127 	MAX_CTRL_WR_LEN = 256,
128 
129 	/*
130 	 * Maximum amount of data which we'll ever need to inline into a
131 	 * TX ring: max(MAX_IMM_TX_PKT_LEN, MAX_CTRL_WR_LEN).
132 	 */
133 	MAX_IMM_TX_LEN = (MAX_IMM_TX_PKT_LEN > MAX_CTRL_WR_LEN
134 			  ? MAX_IMM_TX_PKT_LEN
135 			  : MAX_CTRL_WR_LEN),
136 
137 	/*
138 	 * For incoming packets less than RX_COPY_THRES, we copy the data into
139 	 * an skb rather than referencing the data.  We allocate enough
140 	 * in-line room in skb's to accommodate pulling in RX_PULL_LEN bytes
141 	 * of the data (header).
142 	 */
143 	RX_COPY_THRES = 256,
144 	RX_PULL_LEN = 128,
145 
146 	/*
147 	 * Main body length for sk_buffs used for RX Ethernet packets with
148 	 * fragments.  Should be >= RX_PULL_LEN but possibly bigger to give
149 	 * pskb_may_pull() some room.
150 	 */
151 	RX_SKB_LEN = 512,
152 };
153 
154 /*
155  * Software state per TX descriptor.
156  */
157 struct tx_sw_desc {
158 	struct sk_buff *skb;		/* socket buffer of TX data source */
159 	struct ulptx_sgl *sgl;		/* scatter/gather list in TX Queue */
160 };
161 
162 /*
163  * Software state per RX Free List descriptor.  We keep track of the allocated
164  * FL page, its size, and its PCI DMA address (if the page is mapped).  The FL
165  * page size and its PCI DMA mapped state are stored in the low bits of the
166  * PCI DMA address as per below.
167  */
168 struct rx_sw_desc {
169 	struct page *page;		/* Free List page buffer */
170 	dma_addr_t dma_addr;		/* PCI DMA address (if mapped) */
171 					/*   and flags (see below) */
172 };
173 
174 /*
175  * The low bits of rx_sw_desc.dma_addr have special meaning.  Note that the
176  * SGE also uses the low 4 bits to determine the size of the buffer.  It uses
177  * those bits to index into the SGE_FL_BUFFER_SIZE[index] register array.
178  * Since we only use SGE_FL_BUFFER_SIZE0 and SGE_FL_BUFFER_SIZE1, these low 4
179  * bits can only contain a 0 or a 1 to indicate which size buffer we're giving
180  * to the SGE.  Thus, our software state of "is the buffer mapped for DMA" is
181  * maintained in an inverse sense so the hardware never sees that bit high.
182  */
183 enum {
184 	RX_LARGE_BUF    = 1 << 0,	/* buffer is SGE_FL_BUFFER_SIZE[1] */
185 	RX_UNMAPPED_BUF = 1 << 1,	/* buffer is not mapped */
186 };
187 
188 /**
189  *	get_buf_addr - return DMA buffer address of software descriptor
190  *	@sdesc: pointer to the software buffer descriptor
191  *
192  *	Return the DMA buffer address of a software descriptor (stripping out
193  *	our low-order flag bits).
194  */
195 static inline dma_addr_t get_buf_addr(const struct rx_sw_desc *sdesc)
196 {
197 	return sdesc->dma_addr & ~(dma_addr_t)(RX_LARGE_BUF | RX_UNMAPPED_BUF);
198 }
199 
200 /**
201  *	is_buf_mapped - is buffer mapped for DMA?
202  *	@sdesc: pointer to the software buffer descriptor
203  *
204  *	Determine whether the buffer associated with a software descriptor in
205  *	mapped for DMA or not.
206  */
207 static inline bool is_buf_mapped(const struct rx_sw_desc *sdesc)
208 {
209 	return !(sdesc->dma_addr & RX_UNMAPPED_BUF);
210 }
211 
212 /**
213  *	need_skb_unmap - does the platform need unmapping of sk_buffs?
214  *
215  *	Returns true if the platform needs sk_buff unmapping.  The compiler
216  *	optimizes away unnecessary code if this returns true.
217  */
218 static inline int need_skb_unmap(void)
219 {
220 #ifdef CONFIG_NEED_DMA_MAP_STATE
221 	return 1;
222 #else
223 	return 0;
224 #endif
225 }
226 
227 /**
228  *	txq_avail - return the number of available slots in a TX queue
229  *	@tq: the TX queue
230  *
231  *	Returns the number of available descriptors in a TX queue.
232  */
233 static inline unsigned int txq_avail(const struct sge_txq *tq)
234 {
235 	return tq->size - 1 - tq->in_use;
236 }
237 
238 /**
239  *	fl_cap - return the capacity of a Free List
240  *	@fl: the Free List
241  *
242  *	Returns the capacity of a Free List.  The capacity is less than the
243  *	size because an Egress Queue Index Unit worth of descriptors needs to
244  *	be left unpopulated, otherwise the Producer and Consumer indices PIDX
245  *	and CIDX will match and the hardware will think the FL is empty.
246  */
247 static inline unsigned int fl_cap(const struct sge_fl *fl)
248 {
249 	return fl->size - FL_PER_EQ_UNIT;
250 }
251 
252 /**
253  *	fl_starving - return whether a Free List is starving.
254  *	@adapter: pointer to the adapter
255  *	@fl: the Free List
256  *
257  *	Tests specified Free List to see whether the number of buffers
258  *	available to the hardware has falled below our "starvation"
259  *	threshold.
260  */
261 static inline bool fl_starving(const struct adapter *adapter,
262 			       const struct sge_fl *fl)
263 {
264 	const struct sge *s = &adapter->sge;
265 
266 	return fl->avail - fl->pend_cred <= s->fl_starve_thres;
267 }
268 
269 /**
270  *	map_skb -  map an skb for DMA to the device
271  *	@dev: the egress net device
272  *	@skb: the packet to map
273  *	@addr: a pointer to the base of the DMA mapping array
274  *
275  *	Map an skb for DMA to the device and return an array of DMA addresses.
276  */
277 static int map_skb(struct device *dev, const struct sk_buff *skb,
278 		   dma_addr_t *addr)
279 {
280 	const skb_frag_t *fp, *end;
281 	const struct skb_shared_info *si;
282 
283 	*addr = dma_map_single(dev, skb->data, skb_headlen(skb), DMA_TO_DEVICE);
284 	if (dma_mapping_error(dev, *addr))
285 		goto out_err;
286 
287 	si = skb_shinfo(skb);
288 	end = &si->frags[si->nr_frags];
289 	for (fp = si->frags; fp < end; fp++) {
290 		*++addr = skb_frag_dma_map(dev, fp, 0, skb_frag_size(fp),
291 					   DMA_TO_DEVICE);
292 		if (dma_mapping_error(dev, *addr))
293 			goto unwind;
294 	}
295 	return 0;
296 
297 unwind:
298 	while (fp-- > si->frags)
299 		dma_unmap_page(dev, *--addr, skb_frag_size(fp), DMA_TO_DEVICE);
300 	dma_unmap_single(dev, addr[-1], skb_headlen(skb), DMA_TO_DEVICE);
301 
302 out_err:
303 	return -ENOMEM;
304 }
305 
306 static void unmap_sgl(struct device *dev, const struct sk_buff *skb,
307 		      const struct ulptx_sgl *sgl, const struct sge_txq *tq)
308 {
309 	const struct ulptx_sge_pair *p;
310 	unsigned int nfrags = skb_shinfo(skb)->nr_frags;
311 
312 	if (likely(skb_headlen(skb)))
313 		dma_unmap_single(dev, be64_to_cpu(sgl->addr0),
314 				 be32_to_cpu(sgl->len0), DMA_TO_DEVICE);
315 	else {
316 		dma_unmap_page(dev, be64_to_cpu(sgl->addr0),
317 			       be32_to_cpu(sgl->len0), DMA_TO_DEVICE);
318 		nfrags--;
319 	}
320 
321 	/*
322 	 * the complexity below is because of the possibility of a wrap-around
323 	 * in the middle of an SGL
324 	 */
325 	for (p = sgl->sge; nfrags >= 2; nfrags -= 2) {
326 		if (likely((u8 *)(p + 1) <= (u8 *)tq->stat)) {
327 unmap:
328 			dma_unmap_page(dev, be64_to_cpu(p->addr[0]),
329 				       be32_to_cpu(p->len[0]), DMA_TO_DEVICE);
330 			dma_unmap_page(dev, be64_to_cpu(p->addr[1]),
331 				       be32_to_cpu(p->len[1]), DMA_TO_DEVICE);
332 			p++;
333 		} else if ((u8 *)p == (u8 *)tq->stat) {
334 			p = (const struct ulptx_sge_pair *)tq->desc;
335 			goto unmap;
336 		} else if ((u8 *)p + 8 == (u8 *)tq->stat) {
337 			const __be64 *addr = (const __be64 *)tq->desc;
338 
339 			dma_unmap_page(dev, be64_to_cpu(addr[0]),
340 				       be32_to_cpu(p->len[0]), DMA_TO_DEVICE);
341 			dma_unmap_page(dev, be64_to_cpu(addr[1]),
342 				       be32_to_cpu(p->len[1]), DMA_TO_DEVICE);
343 			p = (const struct ulptx_sge_pair *)&addr[2];
344 		} else {
345 			const __be64 *addr = (const __be64 *)tq->desc;
346 
347 			dma_unmap_page(dev, be64_to_cpu(p->addr[0]),
348 				       be32_to_cpu(p->len[0]), DMA_TO_DEVICE);
349 			dma_unmap_page(dev, be64_to_cpu(addr[0]),
350 				       be32_to_cpu(p->len[1]), DMA_TO_DEVICE);
351 			p = (const struct ulptx_sge_pair *)&addr[1];
352 		}
353 	}
354 	if (nfrags) {
355 		__be64 addr;
356 
357 		if ((u8 *)p == (u8 *)tq->stat)
358 			p = (const struct ulptx_sge_pair *)tq->desc;
359 		addr = ((u8 *)p + 16 <= (u8 *)tq->stat
360 			? p->addr[0]
361 			: *(const __be64 *)tq->desc);
362 		dma_unmap_page(dev, be64_to_cpu(addr), be32_to_cpu(p->len[0]),
363 			       DMA_TO_DEVICE);
364 	}
365 }
366 
367 /**
368  *	free_tx_desc - reclaims TX descriptors and their buffers
369  *	@adapter: the adapter
370  *	@tq: the TX queue to reclaim descriptors from
371  *	@n: the number of descriptors to reclaim
372  *	@unmap: whether the buffers should be unmapped for DMA
373  *
374  *	Reclaims TX descriptors from an SGE TX queue and frees the associated
375  *	TX buffers.  Called with the TX queue lock held.
376  */
377 static void free_tx_desc(struct adapter *adapter, struct sge_txq *tq,
378 			 unsigned int n, bool unmap)
379 {
380 	struct tx_sw_desc *sdesc;
381 	unsigned int cidx = tq->cidx;
382 	struct device *dev = adapter->pdev_dev;
383 
384 	const int need_unmap = need_skb_unmap() && unmap;
385 
386 	sdesc = &tq->sdesc[cidx];
387 	while (n--) {
388 		/*
389 		 * If we kept a reference to the original TX skb, we need to
390 		 * unmap it from PCI DMA space (if required) and free it.
391 		 */
392 		if (sdesc->skb) {
393 			if (need_unmap)
394 				unmap_sgl(dev, sdesc->skb, sdesc->sgl, tq);
395 			dev_consume_skb_any(sdesc->skb);
396 			sdesc->skb = NULL;
397 		}
398 
399 		sdesc++;
400 		if (++cidx == tq->size) {
401 			cidx = 0;
402 			sdesc = tq->sdesc;
403 		}
404 	}
405 	tq->cidx = cidx;
406 }
407 
408 /*
409  * Return the number of reclaimable descriptors in a TX queue.
410  */
411 static inline int reclaimable(const struct sge_txq *tq)
412 {
413 	int hw_cidx = be16_to_cpu(tq->stat->cidx);
414 	int reclaimable = hw_cidx - tq->cidx;
415 	if (reclaimable < 0)
416 		reclaimable += tq->size;
417 	return reclaimable;
418 }
419 
420 /**
421  *	reclaim_completed_tx - reclaims completed TX descriptors
422  *	@adapter: the adapter
423  *	@tq: the TX queue to reclaim completed descriptors from
424  *	@unmap: whether the buffers should be unmapped for DMA
425  *
426  *	Reclaims TX descriptors that the SGE has indicated it has processed,
427  *	and frees the associated buffers if possible.  Called with the TX
428  *	queue locked.
429  */
430 static inline void reclaim_completed_tx(struct adapter *adapter,
431 					struct sge_txq *tq,
432 					bool unmap)
433 {
434 	int avail = reclaimable(tq);
435 
436 	if (avail) {
437 		/*
438 		 * Limit the amount of clean up work we do at a time to keep
439 		 * the TX lock hold time O(1).
440 		 */
441 		if (avail > MAX_TX_RECLAIM)
442 			avail = MAX_TX_RECLAIM;
443 
444 		free_tx_desc(adapter, tq, avail, unmap);
445 		tq->in_use -= avail;
446 	}
447 }
448 
449 /**
450  *	get_buf_size - return the size of an RX Free List buffer.
451  *	@adapter: pointer to the associated adapter
452  *	@sdesc: pointer to the software buffer descriptor
453  */
454 static inline int get_buf_size(const struct adapter *adapter,
455 			       const struct rx_sw_desc *sdesc)
456 {
457 	const struct sge *s = &adapter->sge;
458 
459 	return (s->fl_pg_order > 0 && (sdesc->dma_addr & RX_LARGE_BUF)
460 		? (PAGE_SIZE << s->fl_pg_order) : PAGE_SIZE);
461 }
462 
463 /**
464  *	free_rx_bufs - free RX buffers on an SGE Free List
465  *	@adapter: the adapter
466  *	@fl: the SGE Free List to free buffers from
467  *	@n: how many buffers to free
468  *
469  *	Release the next @n buffers on an SGE Free List RX queue.   The
470  *	buffers must be made inaccessible to hardware before calling this
471  *	function.
472  */
473 static void free_rx_bufs(struct adapter *adapter, struct sge_fl *fl, int n)
474 {
475 	while (n--) {
476 		struct rx_sw_desc *sdesc = &fl->sdesc[fl->cidx];
477 
478 		if (is_buf_mapped(sdesc))
479 			dma_unmap_page(adapter->pdev_dev, get_buf_addr(sdesc),
480 				       get_buf_size(adapter, sdesc),
481 				       PCI_DMA_FROMDEVICE);
482 		put_page(sdesc->page);
483 		sdesc->page = NULL;
484 		if (++fl->cidx == fl->size)
485 			fl->cidx = 0;
486 		fl->avail--;
487 	}
488 }
489 
490 /**
491  *	unmap_rx_buf - unmap the current RX buffer on an SGE Free List
492  *	@adapter: the adapter
493  *	@fl: the SGE Free List
494  *
495  *	Unmap the current buffer on an SGE Free List RX queue.   The
496  *	buffer must be made inaccessible to HW before calling this function.
497  *
498  *	This is similar to @free_rx_bufs above but does not free the buffer.
499  *	Do note that the FL still loses any further access to the buffer.
500  *	This is used predominantly to "transfer ownership" of an FL buffer
501  *	to another entity (typically an skb's fragment list).
502  */
503 static void unmap_rx_buf(struct adapter *adapter, struct sge_fl *fl)
504 {
505 	struct rx_sw_desc *sdesc = &fl->sdesc[fl->cidx];
506 
507 	if (is_buf_mapped(sdesc))
508 		dma_unmap_page(adapter->pdev_dev, get_buf_addr(sdesc),
509 			       get_buf_size(adapter, sdesc),
510 			       PCI_DMA_FROMDEVICE);
511 	sdesc->page = NULL;
512 	if (++fl->cidx == fl->size)
513 		fl->cidx = 0;
514 	fl->avail--;
515 }
516 
517 /**
518  *	ring_fl_db - righ doorbell on free list
519  *	@adapter: the adapter
520  *	@fl: the Free List whose doorbell should be rung ...
521  *
522  *	Tell the Scatter Gather Engine that there are new free list entries
523  *	available.
524  */
525 static inline void ring_fl_db(struct adapter *adapter, struct sge_fl *fl)
526 {
527 	u32 val = adapter->params.arch.sge_fl_db;
528 
529 	/* The SGE keeps track of its Producer and Consumer Indices in terms
530 	 * of Egress Queue Units so we can only tell it about integral numbers
531 	 * of multiples of Free List Entries per Egress Queue Units ...
532 	 */
533 	if (fl->pend_cred >= FL_PER_EQ_UNIT) {
534 		if (is_t4(adapter->params.chip))
535 			val |= PIDX_V(fl->pend_cred / FL_PER_EQ_UNIT);
536 		else
537 			val |= PIDX_T5_V(fl->pend_cred / FL_PER_EQ_UNIT);
538 
539 		/* Make sure all memory writes to the Free List queue are
540 		 * committed before we tell the hardware about them.
541 		 */
542 		wmb();
543 
544 		/* If we don't have access to the new User Doorbell (T5+), use
545 		 * the old doorbell mechanism; otherwise use the new BAR2
546 		 * mechanism.
547 		 */
548 		if (unlikely(fl->bar2_addr == NULL)) {
549 			t4_write_reg(adapter,
550 				     T4VF_SGE_BASE_ADDR + SGE_VF_KDOORBELL,
551 				     QID_V(fl->cntxt_id) | val);
552 		} else {
553 			writel(val | QID_V(fl->bar2_qid),
554 			       fl->bar2_addr + SGE_UDB_KDOORBELL);
555 
556 			/* This Write memory Barrier will force the write to
557 			 * the User Doorbell area to be flushed.
558 			 */
559 			wmb();
560 		}
561 		fl->pend_cred %= FL_PER_EQ_UNIT;
562 	}
563 }
564 
565 /**
566  *	set_rx_sw_desc - initialize software RX buffer descriptor
567  *	@sdesc: pointer to the softwore RX buffer descriptor
568  *	@page: pointer to the page data structure backing the RX buffer
569  *	@dma_addr: PCI DMA address (possibly with low-bit flags)
570  */
571 static inline void set_rx_sw_desc(struct rx_sw_desc *sdesc, struct page *page,
572 				  dma_addr_t dma_addr)
573 {
574 	sdesc->page = page;
575 	sdesc->dma_addr = dma_addr;
576 }
577 
578 /*
579  * Support for poisoning RX buffers ...
580  */
581 #define POISON_BUF_VAL -1
582 
583 static inline void poison_buf(struct page *page, size_t sz)
584 {
585 #if POISON_BUF_VAL >= 0
586 	memset(page_address(page), POISON_BUF_VAL, sz);
587 #endif
588 }
589 
590 /**
591  *	refill_fl - refill an SGE RX buffer ring
592  *	@adapter: the adapter
593  *	@fl: the Free List ring to refill
594  *	@n: the number of new buffers to allocate
595  *	@gfp: the gfp flags for the allocations
596  *
597  *	(Re)populate an SGE free-buffer queue with up to @n new packet buffers,
598  *	allocated with the supplied gfp flags.  The caller must assure that
599  *	@n does not exceed the queue's capacity -- i.e. (cidx == pidx) _IN
600  *	EGRESS QUEUE UNITS_ indicates an empty Free List!  Returns the number
601  *	of buffers allocated.  If afterwards the queue is found critically low,
602  *	mark it as starving in the bitmap of starving FLs.
603  */
604 static unsigned int refill_fl(struct adapter *adapter, struct sge_fl *fl,
605 			      int n, gfp_t gfp)
606 {
607 	struct sge *s = &adapter->sge;
608 	struct page *page;
609 	dma_addr_t dma_addr;
610 	unsigned int cred = fl->avail;
611 	__be64 *d = &fl->desc[fl->pidx];
612 	struct rx_sw_desc *sdesc = &fl->sdesc[fl->pidx];
613 
614 	/*
615 	 * Sanity: ensure that the result of adding n Free List buffers
616 	 * won't result in wrapping the SGE's Producer Index around to
617 	 * it's Consumer Index thereby indicating an empty Free List ...
618 	 */
619 	BUG_ON(fl->avail + n > fl->size - FL_PER_EQ_UNIT);
620 
621 	gfp |= __GFP_NOWARN;
622 
623 	/*
624 	 * If we support large pages, prefer large buffers and fail over to
625 	 * small pages if we can't allocate large pages to satisfy the refill.
626 	 * If we don't support large pages, drop directly into the small page
627 	 * allocation code.
628 	 */
629 	if (s->fl_pg_order == 0)
630 		goto alloc_small_pages;
631 
632 	while (n) {
633 		page = __dev_alloc_pages(gfp, s->fl_pg_order);
634 		if (unlikely(!page)) {
635 			/*
636 			 * We've failed inour attempt to allocate a "large
637 			 * page".  Fail over to the "small page" allocation
638 			 * below.
639 			 */
640 			fl->large_alloc_failed++;
641 			break;
642 		}
643 		poison_buf(page, PAGE_SIZE << s->fl_pg_order);
644 
645 		dma_addr = dma_map_page(adapter->pdev_dev, page, 0,
646 					PAGE_SIZE << s->fl_pg_order,
647 					PCI_DMA_FROMDEVICE);
648 		if (unlikely(dma_mapping_error(adapter->pdev_dev, dma_addr))) {
649 			/*
650 			 * We've run out of DMA mapping space.  Free up the
651 			 * buffer and return with what we've managed to put
652 			 * into the free list.  We don't want to fail over to
653 			 * the small page allocation below in this case
654 			 * because DMA mapping resources are typically
655 			 * critical resources once they become scarse.
656 			 */
657 			__free_pages(page, s->fl_pg_order);
658 			goto out;
659 		}
660 		dma_addr |= RX_LARGE_BUF;
661 		*d++ = cpu_to_be64(dma_addr);
662 
663 		set_rx_sw_desc(sdesc, page, dma_addr);
664 		sdesc++;
665 
666 		fl->avail++;
667 		if (++fl->pidx == fl->size) {
668 			fl->pidx = 0;
669 			sdesc = fl->sdesc;
670 			d = fl->desc;
671 		}
672 		n--;
673 	}
674 
675 alloc_small_pages:
676 	while (n--) {
677 		page = __dev_alloc_page(gfp);
678 		if (unlikely(!page)) {
679 			fl->alloc_failed++;
680 			break;
681 		}
682 		poison_buf(page, PAGE_SIZE);
683 
684 		dma_addr = dma_map_page(adapter->pdev_dev, page, 0, PAGE_SIZE,
685 				       PCI_DMA_FROMDEVICE);
686 		if (unlikely(dma_mapping_error(adapter->pdev_dev, dma_addr))) {
687 			put_page(page);
688 			break;
689 		}
690 		*d++ = cpu_to_be64(dma_addr);
691 
692 		set_rx_sw_desc(sdesc, page, dma_addr);
693 		sdesc++;
694 
695 		fl->avail++;
696 		if (++fl->pidx == fl->size) {
697 			fl->pidx = 0;
698 			sdesc = fl->sdesc;
699 			d = fl->desc;
700 		}
701 	}
702 
703 out:
704 	/*
705 	 * Update our accounting state to incorporate the new Free List
706 	 * buffers, tell the hardware about them and return the number of
707 	 * buffers which we were able to allocate.
708 	 */
709 	cred = fl->avail - cred;
710 	fl->pend_cred += cred;
711 	ring_fl_db(adapter, fl);
712 
713 	if (unlikely(fl_starving(adapter, fl))) {
714 		smp_wmb();
715 		set_bit(fl->cntxt_id, adapter->sge.starving_fl);
716 	}
717 
718 	return cred;
719 }
720 
721 /*
722  * Refill a Free List to its capacity or the Maximum Refill Increment,
723  * whichever is smaller ...
724  */
725 static inline void __refill_fl(struct adapter *adapter, struct sge_fl *fl)
726 {
727 	refill_fl(adapter, fl,
728 		  min((unsigned int)MAX_RX_REFILL, fl_cap(fl) - fl->avail),
729 		  GFP_ATOMIC);
730 }
731 
732 /**
733  *	alloc_ring - allocate resources for an SGE descriptor ring
734  *	@dev: the PCI device's core device
735  *	@nelem: the number of descriptors
736  *	@hwsize: the size of each hardware descriptor
737  *	@swsize: the size of each software descriptor
738  *	@busaddrp: the physical PCI bus address of the allocated ring
739  *	@swringp: return address pointer for software ring
740  *	@stat_size: extra space in hardware ring for status information
741  *
742  *	Allocates resources for an SGE descriptor ring, such as TX queues,
743  *	free buffer lists, response queues, etc.  Each SGE ring requires
744  *	space for its hardware descriptors plus, optionally, space for software
745  *	state associated with each hardware entry (the metadata).  The function
746  *	returns three values: the virtual address for the hardware ring (the
747  *	return value of the function), the PCI bus address of the hardware
748  *	ring (in *busaddrp), and the address of the software ring (in swringp).
749  *	Both the hardware and software rings are returned zeroed out.
750  */
751 static void *alloc_ring(struct device *dev, size_t nelem, size_t hwsize,
752 			size_t swsize, dma_addr_t *busaddrp, void *swringp,
753 			size_t stat_size)
754 {
755 	/*
756 	 * Allocate the hardware ring and PCI DMA bus address space for said.
757 	 */
758 	size_t hwlen = nelem * hwsize + stat_size;
759 	void *hwring = dma_alloc_coherent(dev, hwlen, busaddrp, GFP_KERNEL);
760 
761 	if (!hwring)
762 		return NULL;
763 
764 	/*
765 	 * If the caller wants a software ring, allocate it and return a
766 	 * pointer to it in *swringp.
767 	 */
768 	BUG_ON((swsize != 0) != (swringp != NULL));
769 	if (swsize) {
770 		void *swring = kcalloc(nelem, swsize, GFP_KERNEL);
771 
772 		if (!swring) {
773 			dma_free_coherent(dev, hwlen, hwring, *busaddrp);
774 			return NULL;
775 		}
776 		*(void **)swringp = swring;
777 	}
778 
779 	return hwring;
780 }
781 
782 /**
783  *	sgl_len - calculates the size of an SGL of the given capacity
784  *	@n: the number of SGL entries
785  *
786  *	Calculates the number of flits (8-byte units) needed for a Direct
787  *	Scatter/Gather List that can hold the given number of entries.
788  */
789 static inline unsigned int sgl_len(unsigned int n)
790 {
791 	/*
792 	 * A Direct Scatter Gather List uses 32-bit lengths and 64-bit PCI DMA
793 	 * addresses.  The DSGL Work Request starts off with a 32-bit DSGL
794 	 * ULPTX header, then Length0, then Address0, then, for 1 <= i <= N,
795 	 * repeated sequences of { Length[i], Length[i+1], Address[i],
796 	 * Address[i+1] } (this ensures that all addresses are on 64-bit
797 	 * boundaries).  If N is even, then Length[N+1] should be set to 0 and
798 	 * Address[N+1] is omitted.
799 	 *
800 	 * The following calculation incorporates all of the above.  It's
801 	 * somewhat hard to follow but, briefly: the "+2" accounts for the
802 	 * first two flits which include the DSGL header, Length0 and
803 	 * Address0; the "(3*(n-1))/2" covers the main body of list entries (3
804 	 * flits for every pair of the remaining N) +1 if (n-1) is odd; and
805 	 * finally the "+((n-1)&1)" adds the one remaining flit needed if
806 	 * (n-1) is odd ...
807 	 */
808 	n--;
809 	return (3 * n) / 2 + (n & 1) + 2;
810 }
811 
812 /**
813  *	flits_to_desc - returns the num of TX descriptors for the given flits
814  *	@flits: the number of flits
815  *
816  *	Returns the number of TX descriptors needed for the supplied number
817  *	of flits.
818  */
819 static inline unsigned int flits_to_desc(unsigned int flits)
820 {
821 	BUG_ON(flits > SGE_MAX_WR_LEN / sizeof(__be64));
822 	return DIV_ROUND_UP(flits, TXD_PER_EQ_UNIT);
823 }
824 
825 /**
826  *	is_eth_imm - can an Ethernet packet be sent as immediate data?
827  *	@skb: the packet
828  *
829  *	Returns whether an Ethernet packet is small enough to fit completely as
830  *	immediate data.
831  */
832 static inline int is_eth_imm(const struct sk_buff *skb)
833 {
834 	/*
835 	 * The VF Driver uses the FW_ETH_TX_PKT_VM_WR firmware Work Request
836 	 * which does not accommodate immediate data.  We could dike out all
837 	 * of the support code for immediate data but that would tie our hands
838 	 * too much if we ever want to enhace the firmware.  It would also
839 	 * create more differences between the PF and VF Drivers.
840 	 */
841 	return false;
842 }
843 
844 /**
845  *	calc_tx_flits - calculate the number of flits for a packet TX WR
846  *	@skb: the packet
847  *
848  *	Returns the number of flits needed for a TX Work Request for the
849  *	given Ethernet packet, including the needed WR and CPL headers.
850  */
851 static inline unsigned int calc_tx_flits(const struct sk_buff *skb)
852 {
853 	unsigned int flits;
854 
855 	/*
856 	 * If the skb is small enough, we can pump it out as a work request
857 	 * with only immediate data.  In that case we just have to have the
858 	 * TX Packet header plus the skb data in the Work Request.
859 	 */
860 	if (is_eth_imm(skb))
861 		return DIV_ROUND_UP(skb->len + sizeof(struct cpl_tx_pkt),
862 				    sizeof(__be64));
863 
864 	/*
865 	 * Otherwise, we're going to have to construct a Scatter gather list
866 	 * of the skb body and fragments.  We also include the flits necessary
867 	 * for the TX Packet Work Request and CPL.  We always have a firmware
868 	 * Write Header (incorporated as part of the cpl_tx_pkt_lso and
869 	 * cpl_tx_pkt structures), followed by either a TX Packet Write CPL
870 	 * message or, if we're doing a Large Send Offload, an LSO CPL message
871 	 * with an embedded TX Packet Write CPL message.
872 	 */
873 	flits = sgl_len(skb_shinfo(skb)->nr_frags + 1);
874 	if (skb_shinfo(skb)->gso_size)
875 		flits += (sizeof(struct fw_eth_tx_pkt_vm_wr) +
876 			  sizeof(struct cpl_tx_pkt_lso_core) +
877 			  sizeof(struct cpl_tx_pkt_core)) / sizeof(__be64);
878 	else
879 		flits += (sizeof(struct fw_eth_tx_pkt_vm_wr) +
880 			  sizeof(struct cpl_tx_pkt_core)) / sizeof(__be64);
881 	return flits;
882 }
883 
884 /**
885  *	write_sgl - populate a Scatter/Gather List for a packet
886  *	@skb: the packet
887  *	@tq: the TX queue we are writing into
888  *	@sgl: starting location for writing the SGL
889  *	@end: points right after the end of the SGL
890  *	@start: start offset into skb main-body data to include in the SGL
891  *	@addr: the list of DMA bus addresses for the SGL elements
892  *
893  *	Generates a Scatter/Gather List for the buffers that make up a packet.
894  *	The caller must provide adequate space for the SGL that will be written.
895  *	The SGL includes all of the packet's page fragments and the data in its
896  *	main body except for the first @start bytes.  @pos must be 16-byte
897  *	aligned and within a TX descriptor with available space.  @end points
898  *	write after the end of the SGL but does not account for any potential
899  *	wrap around, i.e., @end > @tq->stat.
900  */
901 static void write_sgl(const struct sk_buff *skb, struct sge_txq *tq,
902 		      struct ulptx_sgl *sgl, u64 *end, unsigned int start,
903 		      const dma_addr_t *addr)
904 {
905 	unsigned int i, len;
906 	struct ulptx_sge_pair *to;
907 	const struct skb_shared_info *si = skb_shinfo(skb);
908 	unsigned int nfrags = si->nr_frags;
909 	struct ulptx_sge_pair buf[MAX_SKB_FRAGS / 2 + 1];
910 
911 	len = skb_headlen(skb) - start;
912 	if (likely(len)) {
913 		sgl->len0 = htonl(len);
914 		sgl->addr0 = cpu_to_be64(addr[0] + start);
915 		nfrags++;
916 	} else {
917 		sgl->len0 = htonl(skb_frag_size(&si->frags[0]));
918 		sgl->addr0 = cpu_to_be64(addr[1]);
919 	}
920 
921 	sgl->cmd_nsge = htonl(ULPTX_CMD_V(ULP_TX_SC_DSGL) |
922 			      ULPTX_NSGE_V(nfrags));
923 	if (likely(--nfrags == 0))
924 		return;
925 	/*
926 	 * Most of the complexity below deals with the possibility we hit the
927 	 * end of the queue in the middle of writing the SGL.  For this case
928 	 * only we create the SGL in a temporary buffer and then copy it.
929 	 */
930 	to = (u8 *)end > (u8 *)tq->stat ? buf : sgl->sge;
931 
932 	for (i = (nfrags != si->nr_frags); nfrags >= 2; nfrags -= 2, to++) {
933 		to->len[0] = cpu_to_be32(skb_frag_size(&si->frags[i]));
934 		to->len[1] = cpu_to_be32(skb_frag_size(&si->frags[++i]));
935 		to->addr[0] = cpu_to_be64(addr[i]);
936 		to->addr[1] = cpu_to_be64(addr[++i]);
937 	}
938 	if (nfrags) {
939 		to->len[0] = cpu_to_be32(skb_frag_size(&si->frags[i]));
940 		to->len[1] = cpu_to_be32(0);
941 		to->addr[0] = cpu_to_be64(addr[i + 1]);
942 	}
943 	if (unlikely((u8 *)end > (u8 *)tq->stat)) {
944 		unsigned int part0 = (u8 *)tq->stat - (u8 *)sgl->sge, part1;
945 
946 		if (likely(part0))
947 			memcpy(sgl->sge, buf, part0);
948 		part1 = (u8 *)end - (u8 *)tq->stat;
949 		memcpy(tq->desc, (u8 *)buf + part0, part1);
950 		end = (void *)tq->desc + part1;
951 	}
952 	if ((uintptr_t)end & 8)           /* 0-pad to multiple of 16 */
953 		*end = 0;
954 }
955 
956 /**
957  *	check_ring_tx_db - check and potentially ring a TX queue's doorbell
958  *	@adapter: the adapter
959  *	@tq: the TX queue
960  *	@n: number of new descriptors to give to HW
961  *
962  *	Ring the doorbel for a TX queue.
963  */
964 static inline void ring_tx_db(struct adapter *adapter, struct sge_txq *tq,
965 			      int n)
966 {
967 	/* Make sure that all writes to the TX Descriptors are committed
968 	 * before we tell the hardware about them.
969 	 */
970 	wmb();
971 
972 	/* If we don't have access to the new User Doorbell (T5+), use the old
973 	 * doorbell mechanism; otherwise use the new BAR2 mechanism.
974 	 */
975 	if (unlikely(tq->bar2_addr == NULL)) {
976 		u32 val = PIDX_V(n);
977 
978 		t4_write_reg(adapter, T4VF_SGE_BASE_ADDR + SGE_VF_KDOORBELL,
979 			     QID_V(tq->cntxt_id) | val);
980 	} else {
981 		u32 val = PIDX_T5_V(n);
982 
983 		/* T4 and later chips share the same PIDX field offset within
984 		 * the doorbell, but T5 and later shrank the field in order to
985 		 * gain a bit for Doorbell Priority.  The field was absurdly
986 		 * large in the first place (14 bits) so we just use the T5
987 		 * and later limits and warn if a Queue ID is too large.
988 		 */
989 		WARN_ON(val & DBPRIO_F);
990 
991 		/* If we're only writing a single Egress Unit and the BAR2
992 		 * Queue ID is 0, we can use the Write Combining Doorbell
993 		 * Gather Buffer; otherwise we use the simple doorbell.
994 		 */
995 		if (n == 1 && tq->bar2_qid == 0) {
996 			unsigned int index = (tq->pidx
997 					      ? (tq->pidx - 1)
998 					      : (tq->size - 1));
999 			__be64 *src = (__be64 *)&tq->desc[index];
1000 			__be64 __iomem *dst = (__be64 __iomem *)(tq->bar2_addr +
1001 							 SGE_UDB_WCDOORBELL);
1002 			unsigned int count = EQ_UNIT / sizeof(__be64);
1003 
1004 			/* Copy the TX Descriptor in a tight loop in order to
1005 			 * try to get it to the adapter in a single Write
1006 			 * Combined transfer on the PCI-E Bus.  If the Write
1007 			 * Combine fails (say because of an interrupt, etc.)
1008 			 * the hardware will simply take the last write as a
1009 			 * simple doorbell write with a PIDX Increment of 1
1010 			 * and will fetch the TX Descriptor from memory via
1011 			 * DMA.
1012 			 */
1013 			while (count) {
1014 				/* the (__force u64) is because the compiler
1015 				 * doesn't understand the endian swizzling
1016 				 * going on
1017 				 */
1018 				writeq((__force u64)*src, dst);
1019 				src++;
1020 				dst++;
1021 				count--;
1022 			}
1023 		} else
1024 			writel(val | QID_V(tq->bar2_qid),
1025 			       tq->bar2_addr + SGE_UDB_KDOORBELL);
1026 
1027 		/* This Write Memory Barrier will force the write to the User
1028 		 * Doorbell area to be flushed.  This is needed to prevent
1029 		 * writes on different CPUs for the same queue from hitting
1030 		 * the adapter out of order.  This is required when some Work
1031 		 * Requests take the Write Combine Gather Buffer path (user
1032 		 * doorbell area offset [SGE_UDB_WCDOORBELL..+63]) and some
1033 		 * take the traditional path where we simply increment the
1034 		 * PIDX (User Doorbell area SGE_UDB_KDOORBELL) and have the
1035 		 * hardware DMA read the actual Work Request.
1036 		 */
1037 		wmb();
1038 	}
1039 }
1040 
1041 /**
1042  *	inline_tx_skb - inline a packet's data into TX descriptors
1043  *	@skb: the packet
1044  *	@tq: the TX queue where the packet will be inlined
1045  *	@pos: starting position in the TX queue to inline the packet
1046  *
1047  *	Inline a packet's contents directly into TX descriptors, starting at
1048  *	the given position within the TX DMA ring.
1049  *	Most of the complexity of this operation is dealing with wrap arounds
1050  *	in the middle of the packet we want to inline.
1051  */
1052 static void inline_tx_skb(const struct sk_buff *skb, const struct sge_txq *tq,
1053 			  void *pos)
1054 {
1055 	u64 *p;
1056 	int left = (void *)tq->stat - pos;
1057 
1058 	if (likely(skb->len <= left)) {
1059 		if (likely(!skb->data_len))
1060 			skb_copy_from_linear_data(skb, pos, skb->len);
1061 		else
1062 			skb_copy_bits(skb, 0, pos, skb->len);
1063 		pos += skb->len;
1064 	} else {
1065 		skb_copy_bits(skb, 0, pos, left);
1066 		skb_copy_bits(skb, left, tq->desc, skb->len - left);
1067 		pos = (void *)tq->desc + (skb->len - left);
1068 	}
1069 
1070 	/* 0-pad to multiple of 16 */
1071 	p = PTR_ALIGN(pos, 8);
1072 	if ((uintptr_t)p & 8)
1073 		*p = 0;
1074 }
1075 
1076 /*
1077  * Figure out what HW csum a packet wants and return the appropriate control
1078  * bits.
1079  */
1080 static u64 hwcsum(enum chip_type chip, const struct sk_buff *skb)
1081 {
1082 	int csum_type;
1083 	const struct iphdr *iph = ip_hdr(skb);
1084 
1085 	if (iph->version == 4) {
1086 		if (iph->protocol == IPPROTO_TCP)
1087 			csum_type = TX_CSUM_TCPIP;
1088 		else if (iph->protocol == IPPROTO_UDP)
1089 			csum_type = TX_CSUM_UDPIP;
1090 		else {
1091 nocsum:
1092 			/*
1093 			 * unknown protocol, disable HW csum
1094 			 * and hope a bad packet is detected
1095 			 */
1096 			return TXPKT_L4CSUM_DIS_F;
1097 		}
1098 	} else {
1099 		/*
1100 		 * this doesn't work with extension headers
1101 		 */
1102 		const struct ipv6hdr *ip6h = (const struct ipv6hdr *)iph;
1103 
1104 		if (ip6h->nexthdr == IPPROTO_TCP)
1105 			csum_type = TX_CSUM_TCPIP6;
1106 		else if (ip6h->nexthdr == IPPROTO_UDP)
1107 			csum_type = TX_CSUM_UDPIP6;
1108 		else
1109 			goto nocsum;
1110 	}
1111 
1112 	if (likely(csum_type >= TX_CSUM_TCPIP)) {
1113 		u64 hdr_len = TXPKT_IPHDR_LEN_V(skb_network_header_len(skb));
1114 		int eth_hdr_len = skb_network_offset(skb) - ETH_HLEN;
1115 
1116 		if (chip <= CHELSIO_T5)
1117 			hdr_len |= TXPKT_ETHHDR_LEN_V(eth_hdr_len);
1118 		else
1119 			hdr_len |= T6_TXPKT_ETHHDR_LEN_V(eth_hdr_len);
1120 		return TXPKT_CSUM_TYPE_V(csum_type) | hdr_len;
1121 	} else {
1122 		int start = skb_transport_offset(skb);
1123 
1124 		return TXPKT_CSUM_TYPE_V(csum_type) |
1125 			TXPKT_CSUM_START_V(start) |
1126 			TXPKT_CSUM_LOC_V(start + skb->csum_offset);
1127 	}
1128 }
1129 
1130 /*
1131  * Stop an Ethernet TX queue and record that state change.
1132  */
1133 static void txq_stop(struct sge_eth_txq *txq)
1134 {
1135 	netif_tx_stop_queue(txq->txq);
1136 	txq->q.stops++;
1137 }
1138 
1139 /*
1140  * Advance our software state for a TX queue by adding n in use descriptors.
1141  */
1142 static inline void txq_advance(struct sge_txq *tq, unsigned int n)
1143 {
1144 	tq->in_use += n;
1145 	tq->pidx += n;
1146 	if (tq->pidx >= tq->size)
1147 		tq->pidx -= tq->size;
1148 }
1149 
1150 /**
1151  *	t4vf_eth_xmit - add a packet to an Ethernet TX queue
1152  *	@skb: the packet
1153  *	@dev: the egress net device
1154  *
1155  *	Add a packet to an SGE Ethernet TX queue.  Runs with softirqs disabled.
1156  */
1157 int t4vf_eth_xmit(struct sk_buff *skb, struct net_device *dev)
1158 {
1159 	u32 wr_mid;
1160 	u64 cntrl, *end;
1161 	int qidx, credits, max_pkt_len;
1162 	unsigned int flits, ndesc;
1163 	struct adapter *adapter;
1164 	struct sge_eth_txq *txq;
1165 	const struct port_info *pi;
1166 	struct fw_eth_tx_pkt_vm_wr *wr;
1167 	struct cpl_tx_pkt_core *cpl;
1168 	const struct skb_shared_info *ssi;
1169 	dma_addr_t addr[MAX_SKB_FRAGS + 1];
1170 	const size_t fw_hdr_copy_len = (sizeof(wr->ethmacdst) +
1171 					sizeof(wr->ethmacsrc) +
1172 					sizeof(wr->ethtype) +
1173 					sizeof(wr->vlantci));
1174 
1175 	/*
1176 	 * The chip minimum packet length is 10 octets but the firmware
1177 	 * command that we are using requires that we copy the Ethernet header
1178 	 * (including the VLAN tag) into the header so we reject anything
1179 	 * smaller than that ...
1180 	 */
1181 	if (unlikely(skb->len < fw_hdr_copy_len))
1182 		goto out_free;
1183 
1184 	/* Discard the packet if the length is greater than mtu */
1185 	max_pkt_len = ETH_HLEN + dev->mtu;
1186 	if (skb_vlan_tagged(skb))
1187 		max_pkt_len += VLAN_HLEN;
1188 	if (!skb_shinfo(skb)->gso_size && (unlikely(skb->len > max_pkt_len)))
1189 		goto out_free;
1190 
1191 	/*
1192 	 * Figure out which TX Queue we're going to use.
1193 	 */
1194 	pi = netdev_priv(dev);
1195 	adapter = pi->adapter;
1196 	qidx = skb_get_queue_mapping(skb);
1197 	BUG_ON(qidx >= pi->nqsets);
1198 	txq = &adapter->sge.ethtxq[pi->first_qset + qidx];
1199 
1200 	if (pi->vlan_id && !skb_vlan_tag_present(skb))
1201 		__vlan_hwaccel_put_tag(skb, cpu_to_be16(ETH_P_8021Q),
1202 				       pi->vlan_id);
1203 
1204 	/*
1205 	 * Take this opportunity to reclaim any TX Descriptors whose DMA
1206 	 * transfers have completed.
1207 	 */
1208 	reclaim_completed_tx(adapter, &txq->q, true);
1209 
1210 	/*
1211 	 * Calculate the number of flits and TX Descriptors we're going to
1212 	 * need along with how many TX Descriptors will be left over after
1213 	 * we inject our Work Request.
1214 	 */
1215 	flits = calc_tx_flits(skb);
1216 	ndesc = flits_to_desc(flits);
1217 	credits = txq_avail(&txq->q) - ndesc;
1218 
1219 	if (unlikely(credits < 0)) {
1220 		/*
1221 		 * Not enough room for this packet's Work Request.  Stop the
1222 		 * TX Queue and return a "busy" condition.  The queue will get
1223 		 * started later on when the firmware informs us that space
1224 		 * has opened up.
1225 		 */
1226 		txq_stop(txq);
1227 		dev_err(adapter->pdev_dev,
1228 			"%s: TX ring %u full while queue awake!\n",
1229 			dev->name, qidx);
1230 		return NETDEV_TX_BUSY;
1231 	}
1232 
1233 	if (!is_eth_imm(skb) &&
1234 	    unlikely(map_skb(adapter->pdev_dev, skb, addr) < 0)) {
1235 		/*
1236 		 * We need to map the skb into PCI DMA space (because it can't
1237 		 * be in-lined directly into the Work Request) and the mapping
1238 		 * operation failed.  Record the error and drop the packet.
1239 		 */
1240 		txq->mapping_err++;
1241 		goto out_free;
1242 	}
1243 
1244 	wr_mid = FW_WR_LEN16_V(DIV_ROUND_UP(flits, 2));
1245 	if (unlikely(credits < ETHTXQ_STOP_THRES)) {
1246 		/*
1247 		 * After we're done injecting the Work Request for this
1248 		 * packet, we'll be below our "stop threshold" so stop the TX
1249 		 * Queue now and schedule a request for an SGE Egress Queue
1250 		 * Update message.  The queue will get started later on when
1251 		 * the firmware processes this Work Request and sends us an
1252 		 * Egress Queue Status Update message indicating that space
1253 		 * has opened up.
1254 		 */
1255 		txq_stop(txq);
1256 		wr_mid |= FW_WR_EQUEQ_F | FW_WR_EQUIQ_F;
1257 	}
1258 
1259 	/*
1260 	 * Start filling in our Work Request.  Note that we do _not_ handle
1261 	 * the WR Header wrapping around the TX Descriptor Ring.  If our
1262 	 * maximum header size ever exceeds one TX Descriptor, we'll need to
1263 	 * do something else here.
1264 	 */
1265 	BUG_ON(DIV_ROUND_UP(ETHTXQ_MAX_HDR, TXD_PER_EQ_UNIT) > 1);
1266 	wr = (void *)&txq->q.desc[txq->q.pidx];
1267 	wr->equiq_to_len16 = cpu_to_be32(wr_mid);
1268 	wr->r3[0] = cpu_to_be32(0);
1269 	wr->r3[1] = cpu_to_be32(0);
1270 	skb_copy_from_linear_data(skb, (void *)wr->ethmacdst, fw_hdr_copy_len);
1271 	end = (u64 *)wr + flits;
1272 
1273 	/*
1274 	 * If this is a Large Send Offload packet we'll put in an LSO CPL
1275 	 * message with an encapsulated TX Packet CPL message.  Otherwise we
1276 	 * just use a TX Packet CPL message.
1277 	 */
1278 	ssi = skb_shinfo(skb);
1279 	if (ssi->gso_size) {
1280 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
1281 		bool v6 = (ssi->gso_type & SKB_GSO_TCPV6) != 0;
1282 		int l3hdr_len = skb_network_header_len(skb);
1283 		int eth_xtra_len = skb_network_offset(skb) - ETH_HLEN;
1284 
1285 		wr->op_immdlen =
1286 			cpu_to_be32(FW_WR_OP_V(FW_ETH_TX_PKT_VM_WR) |
1287 				    FW_WR_IMMDLEN_V(sizeof(*lso) +
1288 						    sizeof(*cpl)));
1289 		/*
1290 		 * Fill in the LSO CPL message.
1291 		 */
1292 		lso->lso_ctrl =
1293 			cpu_to_be32(LSO_OPCODE_V(CPL_TX_PKT_LSO) |
1294 				    LSO_FIRST_SLICE_F |
1295 				    LSO_LAST_SLICE_F |
1296 				    LSO_IPV6_V(v6) |
1297 				    LSO_ETHHDR_LEN_V(eth_xtra_len / 4) |
1298 				    LSO_IPHDR_LEN_V(l3hdr_len / 4) |
1299 				    LSO_TCPHDR_LEN_V(tcp_hdr(skb)->doff));
1300 		lso->ipid_ofst = cpu_to_be16(0);
1301 		lso->mss = cpu_to_be16(ssi->gso_size);
1302 		lso->seqno_offset = cpu_to_be32(0);
1303 		if (is_t4(adapter->params.chip))
1304 			lso->len = cpu_to_be32(skb->len);
1305 		else
1306 			lso->len = cpu_to_be32(LSO_T5_XFER_SIZE_V(skb->len));
1307 
1308 		/*
1309 		 * Set up TX Packet CPL pointer, control word and perform
1310 		 * accounting.
1311 		 */
1312 		cpl = (void *)(lso + 1);
1313 
1314 		if (CHELSIO_CHIP_VERSION(adapter->params.chip) <= CHELSIO_T5)
1315 			cntrl = TXPKT_ETHHDR_LEN_V(eth_xtra_len);
1316 		else
1317 			cntrl = T6_TXPKT_ETHHDR_LEN_V(eth_xtra_len);
1318 
1319 		cntrl |= TXPKT_CSUM_TYPE_V(v6 ?
1320 					   TX_CSUM_TCPIP6 : TX_CSUM_TCPIP) |
1321 			 TXPKT_IPHDR_LEN_V(l3hdr_len);
1322 		txq->tso++;
1323 		txq->tx_cso += ssi->gso_segs;
1324 	} else {
1325 		int len;
1326 
1327 		len = is_eth_imm(skb) ? skb->len + sizeof(*cpl) : sizeof(*cpl);
1328 		wr->op_immdlen =
1329 			cpu_to_be32(FW_WR_OP_V(FW_ETH_TX_PKT_VM_WR) |
1330 				    FW_WR_IMMDLEN_V(len));
1331 
1332 		/*
1333 		 * Set up TX Packet CPL pointer, control word and perform
1334 		 * accounting.
1335 		 */
1336 		cpl = (void *)(wr + 1);
1337 		if (skb->ip_summed == CHECKSUM_PARTIAL) {
1338 			cntrl = hwcsum(adapter->params.chip, skb) |
1339 				TXPKT_IPCSUM_DIS_F;
1340 			txq->tx_cso++;
1341 		} else
1342 			cntrl = TXPKT_L4CSUM_DIS_F | TXPKT_IPCSUM_DIS_F;
1343 	}
1344 
1345 	/*
1346 	 * If there's a VLAN tag present, add that to the list of things to
1347 	 * do in this Work Request.
1348 	 */
1349 	if (skb_vlan_tag_present(skb)) {
1350 		txq->vlan_ins++;
1351 		cntrl |= TXPKT_VLAN_VLD_F | TXPKT_VLAN_V(skb_vlan_tag_get(skb));
1352 	}
1353 
1354 	/*
1355 	 * Fill in the TX Packet CPL message header.
1356 	 */
1357 	cpl->ctrl0 = cpu_to_be32(TXPKT_OPCODE_V(CPL_TX_PKT_XT) |
1358 				 TXPKT_INTF_V(pi->port_id) |
1359 				 TXPKT_PF_V(0));
1360 	cpl->pack = cpu_to_be16(0);
1361 	cpl->len = cpu_to_be16(skb->len);
1362 	cpl->ctrl1 = cpu_to_be64(cntrl);
1363 
1364 #ifdef T4_TRACE
1365 	T4_TRACE5(adapter->tb[txq->q.cntxt_id & 7],
1366 		  "eth_xmit: ndesc %u, credits %u, pidx %u, len %u, frags %u",
1367 		  ndesc, credits, txq->q.pidx, skb->len, ssi->nr_frags);
1368 #endif
1369 
1370 	/*
1371 	 * Fill in the body of the TX Packet CPL message with either in-lined
1372 	 * data or a Scatter/Gather List.
1373 	 */
1374 	if (is_eth_imm(skb)) {
1375 		/*
1376 		 * In-line the packet's data and free the skb since we don't
1377 		 * need it any longer.
1378 		 */
1379 		inline_tx_skb(skb, &txq->q, cpl + 1);
1380 		dev_consume_skb_any(skb);
1381 	} else {
1382 		/*
1383 		 * Write the skb's Scatter/Gather list into the TX Packet CPL
1384 		 * message and retain a pointer to the skb so we can free it
1385 		 * later when its DMA completes.  (We store the skb pointer
1386 		 * in the Software Descriptor corresponding to the last TX
1387 		 * Descriptor used by the Work Request.)
1388 		 *
1389 		 * The retained skb will be freed when the corresponding TX
1390 		 * Descriptors are reclaimed after their DMAs complete.
1391 		 * However, this could take quite a while since, in general,
1392 		 * the hardware is set up to be lazy about sending DMA
1393 		 * completion notifications to us and we mostly perform TX
1394 		 * reclaims in the transmit routine.
1395 		 *
1396 		 * This is good for performamce but means that we rely on new
1397 		 * TX packets arriving to run the destructors of completed
1398 		 * packets, which open up space in their sockets' send queues.
1399 		 * Sometimes we do not get such new packets causing TX to
1400 		 * stall.  A single UDP transmitter is a good example of this
1401 		 * situation.  We have a clean up timer that periodically
1402 		 * reclaims completed packets but it doesn't run often enough
1403 		 * (nor do we want it to) to prevent lengthy stalls.  A
1404 		 * solution to this problem is to run the destructor early,
1405 		 * after the packet is queued but before it's DMAd.  A con is
1406 		 * that we lie to socket memory accounting, but the amount of
1407 		 * extra memory is reasonable (limited by the number of TX
1408 		 * descriptors), the packets do actually get freed quickly by
1409 		 * new packets almost always, and for protocols like TCP that
1410 		 * wait for acks to really free up the data the extra memory
1411 		 * is even less.  On the positive side we run the destructors
1412 		 * on the sending CPU rather than on a potentially different
1413 		 * completing CPU, usually a good thing.
1414 		 *
1415 		 * Run the destructor before telling the DMA engine about the
1416 		 * packet to make sure it doesn't complete and get freed
1417 		 * prematurely.
1418 		 */
1419 		struct ulptx_sgl *sgl = (struct ulptx_sgl *)(cpl + 1);
1420 		struct sge_txq *tq = &txq->q;
1421 		int last_desc;
1422 
1423 		/*
1424 		 * If the Work Request header was an exact multiple of our TX
1425 		 * Descriptor length, then it's possible that the starting SGL
1426 		 * pointer lines up exactly with the end of our TX Descriptor
1427 		 * ring.  If that's the case, wrap around to the beginning
1428 		 * here ...
1429 		 */
1430 		if (unlikely((void *)sgl == (void *)tq->stat)) {
1431 			sgl = (void *)tq->desc;
1432 			end = ((void *)tq->desc + ((void *)end - (void *)tq->stat));
1433 		}
1434 
1435 		write_sgl(skb, tq, sgl, end, 0, addr);
1436 		skb_orphan(skb);
1437 
1438 		last_desc = tq->pidx + ndesc - 1;
1439 		if (last_desc >= tq->size)
1440 			last_desc -= tq->size;
1441 		tq->sdesc[last_desc].skb = skb;
1442 		tq->sdesc[last_desc].sgl = sgl;
1443 	}
1444 
1445 	/*
1446 	 * Advance our internal TX Queue state, tell the hardware about
1447 	 * the new TX descriptors and return success.
1448 	 */
1449 	txq_advance(&txq->q, ndesc);
1450 	netif_trans_update(dev);
1451 	ring_tx_db(adapter, &txq->q, ndesc);
1452 	return NETDEV_TX_OK;
1453 
1454 out_free:
1455 	/*
1456 	 * An error of some sort happened.  Free the TX skb and tell the
1457 	 * OS that we've "dealt" with the packet ...
1458 	 */
1459 	dev_kfree_skb_any(skb);
1460 	return NETDEV_TX_OK;
1461 }
1462 
1463 /**
1464  *	copy_frags - copy fragments from gather list into skb_shared_info
1465  *	@skb: destination skb
1466  *	@gl: source internal packet gather list
1467  *	@offset: packet start offset in first page
1468  *
1469  *	Copy an internal packet gather list into a Linux skb_shared_info
1470  *	structure.
1471  */
1472 static inline void copy_frags(struct sk_buff *skb,
1473 			      const struct pkt_gl *gl,
1474 			      unsigned int offset)
1475 {
1476 	int i;
1477 
1478 	/* usually there's just one frag */
1479 	__skb_fill_page_desc(skb, 0, gl->frags[0].page,
1480 			     gl->frags[0].offset + offset,
1481 			     gl->frags[0].size - offset);
1482 	skb_shinfo(skb)->nr_frags = gl->nfrags;
1483 	for (i = 1; i < gl->nfrags; i++)
1484 		__skb_fill_page_desc(skb, i, gl->frags[i].page,
1485 				     gl->frags[i].offset,
1486 				     gl->frags[i].size);
1487 
1488 	/* get a reference to the last page, we don't own it */
1489 	get_page(gl->frags[gl->nfrags - 1].page);
1490 }
1491 
1492 /**
1493  *	t4vf_pktgl_to_skb - build an sk_buff from a packet gather list
1494  *	@gl: the gather list
1495  *	@skb_len: size of sk_buff main body if it carries fragments
1496  *	@pull_len: amount of data to move to the sk_buff's main body
1497  *
1498  *	Builds an sk_buff from the given packet gather list.  Returns the
1499  *	sk_buff or %NULL if sk_buff allocation failed.
1500  */
1501 static struct sk_buff *t4vf_pktgl_to_skb(const struct pkt_gl *gl,
1502 					 unsigned int skb_len,
1503 					 unsigned int pull_len)
1504 {
1505 	struct sk_buff *skb;
1506 
1507 	/*
1508 	 * If the ingress packet is small enough, allocate an skb large enough
1509 	 * for all of the data and copy it inline.  Otherwise, allocate an skb
1510 	 * with enough room to pull in the header and reference the rest of
1511 	 * the data via the skb fragment list.
1512 	 *
1513 	 * Below we rely on RX_COPY_THRES being less than the smallest Rx
1514 	 * buff!  size, which is expected since buffers are at least
1515 	 * PAGE_SIZEd.  In this case packets up to RX_COPY_THRES have only one
1516 	 * fragment.
1517 	 */
1518 	if (gl->tot_len <= RX_COPY_THRES) {
1519 		/* small packets have only one fragment */
1520 		skb = alloc_skb(gl->tot_len, GFP_ATOMIC);
1521 		if (unlikely(!skb))
1522 			goto out;
1523 		__skb_put(skb, gl->tot_len);
1524 		skb_copy_to_linear_data(skb, gl->va, gl->tot_len);
1525 	} else {
1526 		skb = alloc_skb(skb_len, GFP_ATOMIC);
1527 		if (unlikely(!skb))
1528 			goto out;
1529 		__skb_put(skb, pull_len);
1530 		skb_copy_to_linear_data(skb, gl->va, pull_len);
1531 
1532 		copy_frags(skb, gl, pull_len);
1533 		skb->len = gl->tot_len;
1534 		skb->data_len = skb->len - pull_len;
1535 		skb->truesize += skb->data_len;
1536 	}
1537 
1538 out:
1539 	return skb;
1540 }
1541 
1542 /**
1543  *	t4vf_pktgl_free - free a packet gather list
1544  *	@gl: the gather list
1545  *
1546  *	Releases the pages of a packet gather list.  We do not own the last
1547  *	page on the list and do not free it.
1548  */
1549 static void t4vf_pktgl_free(const struct pkt_gl *gl)
1550 {
1551 	int frag;
1552 
1553 	frag = gl->nfrags - 1;
1554 	while (frag--)
1555 		put_page(gl->frags[frag].page);
1556 }
1557 
1558 /**
1559  *	do_gro - perform Generic Receive Offload ingress packet processing
1560  *	@rxq: ingress RX Ethernet Queue
1561  *	@gl: gather list for ingress packet
1562  *	@pkt: CPL header for last packet fragment
1563  *
1564  *	Perform Generic Receive Offload (GRO) ingress packet processing.
1565  *	We use the standard Linux GRO interfaces for this.
1566  */
1567 static void do_gro(struct sge_eth_rxq *rxq, const struct pkt_gl *gl,
1568 		   const struct cpl_rx_pkt *pkt)
1569 {
1570 	struct adapter *adapter = rxq->rspq.adapter;
1571 	struct sge *s = &adapter->sge;
1572 	struct port_info *pi;
1573 	int ret;
1574 	struct sk_buff *skb;
1575 
1576 	skb = napi_get_frags(&rxq->rspq.napi);
1577 	if (unlikely(!skb)) {
1578 		t4vf_pktgl_free(gl);
1579 		rxq->stats.rx_drops++;
1580 		return;
1581 	}
1582 
1583 	copy_frags(skb, gl, s->pktshift);
1584 	skb->len = gl->tot_len - s->pktshift;
1585 	skb->data_len = skb->len;
1586 	skb->truesize += skb->data_len;
1587 	skb->ip_summed = CHECKSUM_UNNECESSARY;
1588 	skb_record_rx_queue(skb, rxq->rspq.idx);
1589 	pi = netdev_priv(skb->dev);
1590 
1591 	if (pkt->vlan_ex && !pi->vlan_id) {
1592 		__vlan_hwaccel_put_tag(skb, cpu_to_be16(ETH_P_8021Q),
1593 					be16_to_cpu(pkt->vlan));
1594 		rxq->stats.vlan_ex++;
1595 	}
1596 	ret = napi_gro_frags(&rxq->rspq.napi);
1597 
1598 	if (ret == GRO_HELD)
1599 		rxq->stats.lro_pkts++;
1600 	else if (ret == GRO_MERGED || ret == GRO_MERGED_FREE)
1601 		rxq->stats.lro_merged++;
1602 	rxq->stats.pkts++;
1603 	rxq->stats.rx_cso++;
1604 }
1605 
1606 /**
1607  *	t4vf_ethrx_handler - process an ingress ethernet packet
1608  *	@rspq: the response queue that received the packet
1609  *	@rsp: the response queue descriptor holding the RX_PKT message
1610  *	@gl: the gather list of packet fragments
1611  *
1612  *	Process an ingress ethernet packet and deliver it to the stack.
1613  */
1614 int t4vf_ethrx_handler(struct sge_rspq *rspq, const __be64 *rsp,
1615 		       const struct pkt_gl *gl)
1616 {
1617 	struct sk_buff *skb;
1618 	const struct cpl_rx_pkt *pkt = (void *)rsp;
1619 	bool csum_ok = pkt->csum_calc && !pkt->err_vec &&
1620 		       (rspq->netdev->features & NETIF_F_RXCSUM);
1621 	struct sge_eth_rxq *rxq = container_of(rspq, struct sge_eth_rxq, rspq);
1622 	struct adapter *adapter = rspq->adapter;
1623 	struct sge *s = &adapter->sge;
1624 	struct port_info *pi;
1625 
1626 	/*
1627 	 * If this is a good TCP packet and we have Generic Receive Offload
1628 	 * enabled, handle the packet in the GRO path.
1629 	 */
1630 	if ((pkt->l2info & cpu_to_be32(RXF_TCP_F)) &&
1631 	    (rspq->netdev->features & NETIF_F_GRO) && csum_ok &&
1632 	    !pkt->ip_frag) {
1633 		do_gro(rxq, gl, pkt);
1634 		return 0;
1635 	}
1636 
1637 	/*
1638 	 * Convert the Packet Gather List into an skb.
1639 	 */
1640 	skb = t4vf_pktgl_to_skb(gl, RX_SKB_LEN, RX_PULL_LEN);
1641 	if (unlikely(!skb)) {
1642 		t4vf_pktgl_free(gl);
1643 		rxq->stats.rx_drops++;
1644 		return 0;
1645 	}
1646 	__skb_pull(skb, s->pktshift);
1647 	skb->protocol = eth_type_trans(skb, rspq->netdev);
1648 	skb_record_rx_queue(skb, rspq->idx);
1649 	pi = netdev_priv(skb->dev);
1650 	rxq->stats.pkts++;
1651 
1652 	if (csum_ok && !pkt->err_vec &&
1653 	    (be32_to_cpu(pkt->l2info) & (RXF_UDP_F | RXF_TCP_F))) {
1654 		if (!pkt->ip_frag) {
1655 			skb->ip_summed = CHECKSUM_UNNECESSARY;
1656 			rxq->stats.rx_cso++;
1657 		} else if (pkt->l2info & htonl(RXF_IP_F)) {
1658 			__sum16 c = (__force __sum16)pkt->csum;
1659 			skb->csum = csum_unfold(c);
1660 			skb->ip_summed = CHECKSUM_COMPLETE;
1661 			rxq->stats.rx_cso++;
1662 		}
1663 	} else
1664 		skb_checksum_none_assert(skb);
1665 
1666 	if (pkt->vlan_ex && !pi->vlan_id) {
1667 		rxq->stats.vlan_ex++;
1668 		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
1669 				       be16_to_cpu(pkt->vlan));
1670 	}
1671 
1672 	netif_receive_skb(skb);
1673 
1674 	return 0;
1675 }
1676 
1677 /**
1678  *	is_new_response - check if a response is newly written
1679  *	@rc: the response control descriptor
1680  *	@rspq: the response queue
1681  *
1682  *	Returns true if a response descriptor contains a yet unprocessed
1683  *	response.
1684  */
1685 static inline bool is_new_response(const struct rsp_ctrl *rc,
1686 				   const struct sge_rspq *rspq)
1687 {
1688 	return ((rc->type_gen >> RSPD_GEN_S) & 0x1) == rspq->gen;
1689 }
1690 
1691 /**
1692  *	restore_rx_bufs - put back a packet's RX buffers
1693  *	@gl: the packet gather list
1694  *	@fl: the SGE Free List
1695  *	@frags: how many fragments in @si
1696  *
1697  *	Called when we find out that the current packet, @si, can't be
1698  *	processed right away for some reason.  This is a very rare event and
1699  *	there's no effort to make this suspension/resumption process
1700  *	particularly efficient.
1701  *
1702  *	We implement the suspension by putting all of the RX buffers associated
1703  *	with the current packet back on the original Free List.  The buffers
1704  *	have already been unmapped and are left unmapped, we mark them as
1705  *	unmapped in order to prevent further unmapping attempts.  (Effectively
1706  *	this function undoes the series of @unmap_rx_buf calls which were done
1707  *	to create the current packet's gather list.)  This leaves us ready to
1708  *	restart processing of the packet the next time we start processing the
1709  *	RX Queue ...
1710  */
1711 static void restore_rx_bufs(const struct pkt_gl *gl, struct sge_fl *fl,
1712 			    int frags)
1713 {
1714 	struct rx_sw_desc *sdesc;
1715 
1716 	while (frags--) {
1717 		if (fl->cidx == 0)
1718 			fl->cidx = fl->size - 1;
1719 		else
1720 			fl->cidx--;
1721 		sdesc = &fl->sdesc[fl->cidx];
1722 		sdesc->page = gl->frags[frags].page;
1723 		sdesc->dma_addr |= RX_UNMAPPED_BUF;
1724 		fl->avail++;
1725 	}
1726 }
1727 
1728 /**
1729  *	rspq_next - advance to the next entry in a response queue
1730  *	@rspq: the queue
1731  *
1732  *	Updates the state of a response queue to advance it to the next entry.
1733  */
1734 static inline void rspq_next(struct sge_rspq *rspq)
1735 {
1736 	rspq->cur_desc = (void *)rspq->cur_desc + rspq->iqe_len;
1737 	if (unlikely(++rspq->cidx == rspq->size)) {
1738 		rspq->cidx = 0;
1739 		rspq->gen ^= 1;
1740 		rspq->cur_desc = rspq->desc;
1741 	}
1742 }
1743 
1744 /**
1745  *	process_responses - process responses from an SGE response queue
1746  *	@rspq: the ingress response queue to process
1747  *	@budget: how many responses can be processed in this round
1748  *
1749  *	Process responses from a Scatter Gather Engine response queue up to
1750  *	the supplied budget.  Responses include received packets as well as
1751  *	control messages from firmware or hardware.
1752  *
1753  *	Additionally choose the interrupt holdoff time for the next interrupt
1754  *	on this queue.  If the system is under memory shortage use a fairly
1755  *	long delay to help recovery.
1756  */
1757 static int process_responses(struct sge_rspq *rspq, int budget)
1758 {
1759 	struct sge_eth_rxq *rxq = container_of(rspq, struct sge_eth_rxq, rspq);
1760 	struct adapter *adapter = rspq->adapter;
1761 	struct sge *s = &adapter->sge;
1762 	int budget_left = budget;
1763 
1764 	while (likely(budget_left)) {
1765 		int ret, rsp_type;
1766 		const struct rsp_ctrl *rc;
1767 
1768 		rc = (void *)rspq->cur_desc + (rspq->iqe_len - sizeof(*rc));
1769 		if (!is_new_response(rc, rspq))
1770 			break;
1771 
1772 		/*
1773 		 * Figure out what kind of response we've received from the
1774 		 * SGE.
1775 		 */
1776 		dma_rmb();
1777 		rsp_type = RSPD_TYPE_G(rc->type_gen);
1778 		if (likely(rsp_type == RSPD_TYPE_FLBUF_X)) {
1779 			struct page_frag *fp;
1780 			struct pkt_gl gl;
1781 			const struct rx_sw_desc *sdesc;
1782 			u32 bufsz, frag;
1783 			u32 len = be32_to_cpu(rc->pldbuflen_qid);
1784 
1785 			/*
1786 			 * If we get a "new buffer" message from the SGE we
1787 			 * need to move on to the next Free List buffer.
1788 			 */
1789 			if (len & RSPD_NEWBUF_F) {
1790 				/*
1791 				 * We get one "new buffer" message when we
1792 				 * first start up a queue so we need to ignore
1793 				 * it when our offset into the buffer is 0.
1794 				 */
1795 				if (likely(rspq->offset > 0)) {
1796 					free_rx_bufs(rspq->adapter, &rxq->fl,
1797 						     1);
1798 					rspq->offset = 0;
1799 				}
1800 				len = RSPD_LEN_G(len);
1801 			}
1802 			gl.tot_len = len;
1803 
1804 			/*
1805 			 * Gather packet fragments.
1806 			 */
1807 			for (frag = 0, fp = gl.frags; /**/; frag++, fp++) {
1808 				BUG_ON(frag >= MAX_SKB_FRAGS);
1809 				BUG_ON(rxq->fl.avail == 0);
1810 				sdesc = &rxq->fl.sdesc[rxq->fl.cidx];
1811 				bufsz = get_buf_size(adapter, sdesc);
1812 				fp->page = sdesc->page;
1813 				fp->offset = rspq->offset;
1814 				fp->size = min(bufsz, len);
1815 				len -= fp->size;
1816 				if (!len)
1817 					break;
1818 				unmap_rx_buf(rspq->adapter, &rxq->fl);
1819 			}
1820 			gl.nfrags = frag+1;
1821 
1822 			/*
1823 			 * Last buffer remains mapped so explicitly make it
1824 			 * coherent for CPU access and start preloading first
1825 			 * cache line ...
1826 			 */
1827 			dma_sync_single_for_cpu(rspq->adapter->pdev_dev,
1828 						get_buf_addr(sdesc),
1829 						fp->size, DMA_FROM_DEVICE);
1830 			gl.va = (page_address(gl.frags[0].page) +
1831 				 gl.frags[0].offset);
1832 			prefetch(gl.va);
1833 
1834 			/*
1835 			 * Hand the new ingress packet to the handler for
1836 			 * this Response Queue.
1837 			 */
1838 			ret = rspq->handler(rspq, rspq->cur_desc, &gl);
1839 			if (likely(ret == 0))
1840 				rspq->offset += ALIGN(fp->size, s->fl_align);
1841 			else
1842 				restore_rx_bufs(&gl, &rxq->fl, frag);
1843 		} else if (likely(rsp_type == RSPD_TYPE_CPL_X)) {
1844 			ret = rspq->handler(rspq, rspq->cur_desc, NULL);
1845 		} else {
1846 			WARN_ON(rsp_type > RSPD_TYPE_CPL_X);
1847 			ret = 0;
1848 		}
1849 
1850 		if (unlikely(ret)) {
1851 			/*
1852 			 * Couldn't process descriptor, back off for recovery.
1853 			 * We use the SGE's last timer which has the longest
1854 			 * interrupt coalescing value ...
1855 			 */
1856 			const int NOMEM_TIMER_IDX = SGE_NTIMERS-1;
1857 			rspq->next_intr_params =
1858 				QINTR_TIMER_IDX_V(NOMEM_TIMER_IDX);
1859 			break;
1860 		}
1861 
1862 		rspq_next(rspq);
1863 		budget_left--;
1864 	}
1865 
1866 	/*
1867 	 * If this is a Response Queue with an associated Free List and
1868 	 * at least two Egress Queue units available in the Free List
1869 	 * for new buffer pointers, refill the Free List.
1870 	 */
1871 	if (rspq->offset >= 0 &&
1872 	    fl_cap(&rxq->fl) - rxq->fl.avail >= 2*FL_PER_EQ_UNIT)
1873 		__refill_fl(rspq->adapter, &rxq->fl);
1874 	return budget - budget_left;
1875 }
1876 
1877 /**
1878  *	napi_rx_handler - the NAPI handler for RX processing
1879  *	@napi: the napi instance
1880  *	@budget: how many packets we can process in this round
1881  *
1882  *	Handler for new data events when using NAPI.  This does not need any
1883  *	locking or protection from interrupts as data interrupts are off at
1884  *	this point and other adapter interrupts do not interfere (the latter
1885  *	in not a concern at all with MSI-X as non-data interrupts then have
1886  *	a separate handler).
1887  */
1888 static int napi_rx_handler(struct napi_struct *napi, int budget)
1889 {
1890 	unsigned int intr_params;
1891 	struct sge_rspq *rspq = container_of(napi, struct sge_rspq, napi);
1892 	int work_done = process_responses(rspq, budget);
1893 	u32 val;
1894 
1895 	if (likely(work_done < budget)) {
1896 		napi_complete_done(napi, work_done);
1897 		intr_params = rspq->next_intr_params;
1898 		rspq->next_intr_params = rspq->intr_params;
1899 	} else
1900 		intr_params = QINTR_TIMER_IDX_V(SGE_TIMER_UPD_CIDX);
1901 
1902 	if (unlikely(work_done == 0))
1903 		rspq->unhandled_irqs++;
1904 
1905 	val = CIDXINC_V(work_done) | SEINTARM_V(intr_params);
1906 	/* If we don't have access to the new User GTS (T5+), use the old
1907 	 * doorbell mechanism; otherwise use the new BAR2 mechanism.
1908 	 */
1909 	if (unlikely(!rspq->bar2_addr)) {
1910 		t4_write_reg(rspq->adapter,
1911 			     T4VF_SGE_BASE_ADDR + SGE_VF_GTS,
1912 			     val | INGRESSQID_V((u32)rspq->cntxt_id));
1913 	} else {
1914 		writel(val | INGRESSQID_V(rspq->bar2_qid),
1915 		       rspq->bar2_addr + SGE_UDB_GTS);
1916 		wmb();
1917 	}
1918 	return work_done;
1919 }
1920 
1921 /*
1922  * The MSI-X interrupt handler for an SGE response queue for the NAPI case
1923  * (i.e., response queue serviced by NAPI polling).
1924  */
1925 irqreturn_t t4vf_sge_intr_msix(int irq, void *cookie)
1926 {
1927 	struct sge_rspq *rspq = cookie;
1928 
1929 	napi_schedule(&rspq->napi);
1930 	return IRQ_HANDLED;
1931 }
1932 
1933 /*
1934  * Process the indirect interrupt entries in the interrupt queue and kick off
1935  * NAPI for each queue that has generated an entry.
1936  */
1937 static unsigned int process_intrq(struct adapter *adapter)
1938 {
1939 	struct sge *s = &adapter->sge;
1940 	struct sge_rspq *intrq = &s->intrq;
1941 	unsigned int work_done;
1942 	u32 val;
1943 
1944 	spin_lock(&adapter->sge.intrq_lock);
1945 	for (work_done = 0; ; work_done++) {
1946 		const struct rsp_ctrl *rc;
1947 		unsigned int qid, iq_idx;
1948 		struct sge_rspq *rspq;
1949 
1950 		/*
1951 		 * Grab the next response from the interrupt queue and bail
1952 		 * out if it's not a new response.
1953 		 */
1954 		rc = (void *)intrq->cur_desc + (intrq->iqe_len - sizeof(*rc));
1955 		if (!is_new_response(rc, intrq))
1956 			break;
1957 
1958 		/*
1959 		 * If the response isn't a forwarded interrupt message issue a
1960 		 * error and go on to the next response message.  This should
1961 		 * never happen ...
1962 		 */
1963 		dma_rmb();
1964 		if (unlikely(RSPD_TYPE_G(rc->type_gen) != RSPD_TYPE_INTR_X)) {
1965 			dev_err(adapter->pdev_dev,
1966 				"Unexpected INTRQ response type %d\n",
1967 				RSPD_TYPE_G(rc->type_gen));
1968 			continue;
1969 		}
1970 
1971 		/*
1972 		 * Extract the Queue ID from the interrupt message and perform
1973 		 * sanity checking to make sure it really refers to one of our
1974 		 * Ingress Queues which is active and matches the queue's ID.
1975 		 * None of these error conditions should ever happen so we may
1976 		 * want to either make them fatal and/or conditionalized under
1977 		 * DEBUG.
1978 		 */
1979 		qid = RSPD_QID_G(be32_to_cpu(rc->pldbuflen_qid));
1980 		iq_idx = IQ_IDX(s, qid);
1981 		if (unlikely(iq_idx >= MAX_INGQ)) {
1982 			dev_err(adapter->pdev_dev,
1983 				"Ingress QID %d out of range\n", qid);
1984 			continue;
1985 		}
1986 		rspq = s->ingr_map[iq_idx];
1987 		if (unlikely(rspq == NULL)) {
1988 			dev_err(adapter->pdev_dev,
1989 				"Ingress QID %d RSPQ=NULL\n", qid);
1990 			continue;
1991 		}
1992 		if (unlikely(rspq->abs_id != qid)) {
1993 			dev_err(adapter->pdev_dev,
1994 				"Ingress QID %d refers to RSPQ %d\n",
1995 				qid, rspq->abs_id);
1996 			continue;
1997 		}
1998 
1999 		/*
2000 		 * Schedule NAPI processing on the indicated Response Queue
2001 		 * and move on to the next entry in the Forwarded Interrupt
2002 		 * Queue.
2003 		 */
2004 		napi_schedule(&rspq->napi);
2005 		rspq_next(intrq);
2006 	}
2007 
2008 	val = CIDXINC_V(work_done) | SEINTARM_V(intrq->intr_params);
2009 	/* If we don't have access to the new User GTS (T5+), use the old
2010 	 * doorbell mechanism; otherwise use the new BAR2 mechanism.
2011 	 */
2012 	if (unlikely(!intrq->bar2_addr)) {
2013 		t4_write_reg(adapter, T4VF_SGE_BASE_ADDR + SGE_VF_GTS,
2014 			     val | INGRESSQID_V(intrq->cntxt_id));
2015 	} else {
2016 		writel(val | INGRESSQID_V(intrq->bar2_qid),
2017 		       intrq->bar2_addr + SGE_UDB_GTS);
2018 		wmb();
2019 	}
2020 
2021 	spin_unlock(&adapter->sge.intrq_lock);
2022 
2023 	return work_done;
2024 }
2025 
2026 /*
2027  * The MSI interrupt handler handles data events from SGE response queues as
2028  * well as error and other async events as they all use the same MSI vector.
2029  */
2030 static irqreturn_t t4vf_intr_msi(int irq, void *cookie)
2031 {
2032 	struct adapter *adapter = cookie;
2033 
2034 	process_intrq(adapter);
2035 	return IRQ_HANDLED;
2036 }
2037 
2038 /**
2039  *	t4vf_intr_handler - select the top-level interrupt handler
2040  *	@adapter: the adapter
2041  *
2042  *	Selects the top-level interrupt handler based on the type of interrupts
2043  *	(MSI-X or MSI).
2044  */
2045 irq_handler_t t4vf_intr_handler(struct adapter *adapter)
2046 {
2047 	BUG_ON((adapter->flags &
2048 	       (CXGB4VF_USING_MSIX | CXGB4VF_USING_MSI)) == 0);
2049 	if (adapter->flags & CXGB4VF_USING_MSIX)
2050 		return t4vf_sge_intr_msix;
2051 	else
2052 		return t4vf_intr_msi;
2053 }
2054 
2055 /**
2056  *	sge_rx_timer_cb - perform periodic maintenance of SGE RX queues
2057  *	@t: Rx timer
2058  *
2059  *	Runs periodically from a timer to perform maintenance of SGE RX queues.
2060  *
2061  *	a) Replenishes RX queues that have run out due to memory shortage.
2062  *	Normally new RX buffers are added when existing ones are consumed but
2063  *	when out of memory a queue can become empty.  We schedule NAPI to do
2064  *	the actual refill.
2065  */
2066 static void sge_rx_timer_cb(struct timer_list *t)
2067 {
2068 	struct adapter *adapter = from_timer(adapter, t, sge.rx_timer);
2069 	struct sge *s = &adapter->sge;
2070 	unsigned int i;
2071 
2072 	/*
2073 	 * Scan the "Starving Free Lists" flag array looking for any Free
2074 	 * Lists in need of more free buffers.  If we find one and it's not
2075 	 * being actively polled, then bump its "starving" counter and attempt
2076 	 * to refill it.  If we're successful in adding enough buffers to push
2077 	 * the Free List over the starving threshold, then we can clear its
2078 	 * "starving" status.
2079 	 */
2080 	for (i = 0; i < ARRAY_SIZE(s->starving_fl); i++) {
2081 		unsigned long m;
2082 
2083 		for (m = s->starving_fl[i]; m; m &= m - 1) {
2084 			unsigned int id = __ffs(m) + i * BITS_PER_LONG;
2085 			struct sge_fl *fl = s->egr_map[id];
2086 
2087 			clear_bit(id, s->starving_fl);
2088 			smp_mb__after_atomic();
2089 
2090 			/*
2091 			 * Since we are accessing fl without a lock there's a
2092 			 * small probability of a false positive where we
2093 			 * schedule napi but the FL is no longer starving.
2094 			 * No biggie.
2095 			 */
2096 			if (fl_starving(adapter, fl)) {
2097 				struct sge_eth_rxq *rxq;
2098 
2099 				rxq = container_of(fl, struct sge_eth_rxq, fl);
2100 				if (napi_reschedule(&rxq->rspq.napi))
2101 					fl->starving++;
2102 				else
2103 					set_bit(id, s->starving_fl);
2104 			}
2105 		}
2106 	}
2107 
2108 	/*
2109 	 * Reschedule the next scan for starving Free Lists ...
2110 	 */
2111 	mod_timer(&s->rx_timer, jiffies + RX_QCHECK_PERIOD);
2112 }
2113 
2114 /**
2115  *	sge_tx_timer_cb - perform periodic maintenance of SGE Tx queues
2116  *	@t: Tx timer
2117  *
2118  *	Runs periodically from a timer to perform maintenance of SGE TX queues.
2119  *
2120  *	b) Reclaims completed Tx packets for the Ethernet queues.  Normally
2121  *	packets are cleaned up by new Tx packets, this timer cleans up packets
2122  *	when no new packets are being submitted.  This is essential for pktgen,
2123  *	at least.
2124  */
2125 static void sge_tx_timer_cb(struct timer_list *t)
2126 {
2127 	struct adapter *adapter = from_timer(adapter, t, sge.tx_timer);
2128 	struct sge *s = &adapter->sge;
2129 	unsigned int i, budget;
2130 
2131 	budget = MAX_TIMER_TX_RECLAIM;
2132 	i = s->ethtxq_rover;
2133 	do {
2134 		struct sge_eth_txq *txq = &s->ethtxq[i];
2135 
2136 		if (reclaimable(&txq->q) && __netif_tx_trylock(txq->txq)) {
2137 			int avail = reclaimable(&txq->q);
2138 
2139 			if (avail > budget)
2140 				avail = budget;
2141 
2142 			free_tx_desc(adapter, &txq->q, avail, true);
2143 			txq->q.in_use -= avail;
2144 			__netif_tx_unlock(txq->txq);
2145 
2146 			budget -= avail;
2147 			if (!budget)
2148 				break;
2149 		}
2150 
2151 		i++;
2152 		if (i >= s->ethqsets)
2153 			i = 0;
2154 	} while (i != s->ethtxq_rover);
2155 	s->ethtxq_rover = i;
2156 
2157 	/*
2158 	 * If we found too many reclaimable packets schedule a timer in the
2159 	 * near future to continue where we left off.  Otherwise the next timer
2160 	 * will be at its normal interval.
2161 	 */
2162 	mod_timer(&s->tx_timer, jiffies + (budget ? TX_QCHECK_PERIOD : 2));
2163 }
2164 
2165 /**
2166  *	bar2_address - return the BAR2 address for an SGE Queue's Registers
2167  *	@adapter: the adapter
2168  *	@qid: the SGE Queue ID
2169  *	@qtype: the SGE Queue Type (Egress or Ingress)
2170  *	@pbar2_qid: BAR2 Queue ID or 0 for Queue ID inferred SGE Queues
2171  *
2172  *	Returns the BAR2 address for the SGE Queue Registers associated with
2173  *	@qid.  If BAR2 SGE Registers aren't available, returns NULL.  Also
2174  *	returns the BAR2 Queue ID to be used with writes to the BAR2 SGE
2175  *	Queue Registers.  If the BAR2 Queue ID is 0, then "Inferred Queue ID"
2176  *	Registers are supported (e.g. the Write Combining Doorbell Buffer).
2177  */
2178 static void __iomem *bar2_address(struct adapter *adapter,
2179 				  unsigned int qid,
2180 				  enum t4_bar2_qtype qtype,
2181 				  unsigned int *pbar2_qid)
2182 {
2183 	u64 bar2_qoffset;
2184 	int ret;
2185 
2186 	ret = t4vf_bar2_sge_qregs(adapter, qid, qtype,
2187 				  &bar2_qoffset, pbar2_qid);
2188 	if (ret)
2189 		return NULL;
2190 
2191 	return adapter->bar2 + bar2_qoffset;
2192 }
2193 
2194 /**
2195  *	t4vf_sge_alloc_rxq - allocate an SGE RX Queue
2196  *	@adapter: the adapter
2197  *	@rspq: pointer to to the new rxq's Response Queue to be filled in
2198  *	@iqasynch: if 0, a normal rspq; if 1, an asynchronous event queue
2199  *	@dev: the network device associated with the new rspq
2200  *	@intr_dest: MSI-X vector index (overriden in MSI mode)
2201  *	@fl: pointer to the new rxq's Free List to be filled in
2202  *	@hnd: the interrupt handler to invoke for the rspq
2203  */
2204 int t4vf_sge_alloc_rxq(struct adapter *adapter, struct sge_rspq *rspq,
2205 		       bool iqasynch, struct net_device *dev,
2206 		       int intr_dest,
2207 		       struct sge_fl *fl, rspq_handler_t hnd)
2208 {
2209 	struct sge *s = &adapter->sge;
2210 	struct port_info *pi = netdev_priv(dev);
2211 	struct fw_iq_cmd cmd, rpl;
2212 	int ret, iqandst, flsz = 0;
2213 	int relaxed = !(adapter->flags & CXGB4VF_ROOT_NO_RELAXED_ORDERING);
2214 
2215 	/*
2216 	 * If we're using MSI interrupts and we're not initializing the
2217 	 * Forwarded Interrupt Queue itself, then set up this queue for
2218 	 * indirect interrupts to the Forwarded Interrupt Queue.  Obviously
2219 	 * the Forwarded Interrupt Queue must be set up before any other
2220 	 * ingress queue ...
2221 	 */
2222 	if ((adapter->flags & CXGB4VF_USING_MSI) &&
2223 	    rspq != &adapter->sge.intrq) {
2224 		iqandst = SGE_INTRDST_IQ;
2225 		intr_dest = adapter->sge.intrq.abs_id;
2226 	} else
2227 		iqandst = SGE_INTRDST_PCI;
2228 
2229 	/*
2230 	 * Allocate the hardware ring for the Response Queue.  The size needs
2231 	 * to be a multiple of 16 which includes the mandatory status entry
2232 	 * (regardless of whether the Status Page capabilities are enabled or
2233 	 * not).
2234 	 */
2235 	rspq->size = roundup(rspq->size, 16);
2236 	rspq->desc = alloc_ring(adapter->pdev_dev, rspq->size, rspq->iqe_len,
2237 				0, &rspq->phys_addr, NULL, 0);
2238 	if (!rspq->desc)
2239 		return -ENOMEM;
2240 
2241 	/*
2242 	 * Fill in the Ingress Queue Command.  Note: Ideally this code would
2243 	 * be in t4vf_hw.c but there are so many parameters and dependencies
2244 	 * on our Linux SGE state that we would end up having to pass tons of
2245 	 * parameters.  We'll have to think about how this might be migrated
2246 	 * into OS-independent common code ...
2247 	 */
2248 	memset(&cmd, 0, sizeof(cmd));
2249 	cmd.op_to_vfn = cpu_to_be32(FW_CMD_OP_V(FW_IQ_CMD) |
2250 				    FW_CMD_REQUEST_F |
2251 				    FW_CMD_WRITE_F |
2252 				    FW_CMD_EXEC_F);
2253 	cmd.alloc_to_len16 = cpu_to_be32(FW_IQ_CMD_ALLOC_F |
2254 					 FW_IQ_CMD_IQSTART_F |
2255 					 FW_LEN16(cmd));
2256 	cmd.type_to_iqandstindex =
2257 		cpu_to_be32(FW_IQ_CMD_TYPE_V(FW_IQ_TYPE_FL_INT_CAP) |
2258 			    FW_IQ_CMD_IQASYNCH_V(iqasynch) |
2259 			    FW_IQ_CMD_VIID_V(pi->viid) |
2260 			    FW_IQ_CMD_IQANDST_V(iqandst) |
2261 			    FW_IQ_CMD_IQANUS_V(1) |
2262 			    FW_IQ_CMD_IQANUD_V(SGE_UPDATEDEL_INTR) |
2263 			    FW_IQ_CMD_IQANDSTINDEX_V(intr_dest));
2264 	cmd.iqdroprss_to_iqesize =
2265 		cpu_to_be16(FW_IQ_CMD_IQPCIECH_V(pi->port_id) |
2266 			    FW_IQ_CMD_IQGTSMODE_F |
2267 			    FW_IQ_CMD_IQINTCNTTHRESH_V(rspq->pktcnt_idx) |
2268 			    FW_IQ_CMD_IQESIZE_V(ilog2(rspq->iqe_len) - 4));
2269 	cmd.iqsize = cpu_to_be16(rspq->size);
2270 	cmd.iqaddr = cpu_to_be64(rspq->phys_addr);
2271 
2272 	if (fl) {
2273 		unsigned int chip_ver =
2274 			CHELSIO_CHIP_VERSION(adapter->params.chip);
2275 		/*
2276 		 * Allocate the ring for the hardware free list (with space
2277 		 * for its status page) along with the associated software
2278 		 * descriptor ring.  The free list size needs to be a multiple
2279 		 * of the Egress Queue Unit and at least 2 Egress Units larger
2280 		 * than the SGE's Egress Congrestion Threshold
2281 		 * (fl_starve_thres - 1).
2282 		 */
2283 		if (fl->size < s->fl_starve_thres - 1 + 2 * FL_PER_EQ_UNIT)
2284 			fl->size = s->fl_starve_thres - 1 + 2 * FL_PER_EQ_UNIT;
2285 		fl->size = roundup(fl->size, FL_PER_EQ_UNIT);
2286 		fl->desc = alloc_ring(adapter->pdev_dev, fl->size,
2287 				      sizeof(__be64), sizeof(struct rx_sw_desc),
2288 				      &fl->addr, &fl->sdesc, s->stat_len);
2289 		if (!fl->desc) {
2290 			ret = -ENOMEM;
2291 			goto err;
2292 		}
2293 
2294 		/*
2295 		 * Calculate the size of the hardware free list ring plus
2296 		 * Status Page (which the SGE will place after the end of the
2297 		 * free list ring) in Egress Queue Units.
2298 		 */
2299 		flsz = (fl->size / FL_PER_EQ_UNIT +
2300 			s->stat_len / EQ_UNIT);
2301 
2302 		/*
2303 		 * Fill in all the relevant firmware Ingress Queue Command
2304 		 * fields for the free list.
2305 		 */
2306 		cmd.iqns_to_fl0congen =
2307 			cpu_to_be32(
2308 				FW_IQ_CMD_FL0HOSTFCMODE_V(SGE_HOSTFCMODE_NONE) |
2309 				FW_IQ_CMD_FL0PACKEN_F |
2310 				FW_IQ_CMD_FL0FETCHRO_V(relaxed) |
2311 				FW_IQ_CMD_FL0DATARO_V(relaxed) |
2312 				FW_IQ_CMD_FL0PADEN_F);
2313 
2314 		/* In T6, for egress queue type FL there is internal overhead
2315 		 * of 16B for header going into FLM module.  Hence the maximum
2316 		 * allowed burst size is 448 bytes.  For T4/T5, the hardware
2317 		 * doesn't coalesce fetch requests if more than 64 bytes of
2318 		 * Free List pointers are provided, so we use a 128-byte Fetch
2319 		 * Burst Minimum there (T6 implements coalescing so we can use
2320 		 * the smaller 64-byte value there).
2321 		 */
2322 		cmd.fl0dcaen_to_fl0cidxfthresh =
2323 			cpu_to_be16(
2324 				FW_IQ_CMD_FL0FBMIN_V(chip_ver <= CHELSIO_T5
2325 						     ? FETCHBURSTMIN_128B_X
2326 						     : FETCHBURSTMIN_64B_T6_X) |
2327 				FW_IQ_CMD_FL0FBMAX_V((chip_ver <= CHELSIO_T5) ?
2328 						     FETCHBURSTMAX_512B_X :
2329 						     FETCHBURSTMAX_256B_X));
2330 		cmd.fl0size = cpu_to_be16(flsz);
2331 		cmd.fl0addr = cpu_to_be64(fl->addr);
2332 	}
2333 
2334 	/*
2335 	 * Issue the firmware Ingress Queue Command and extract the results if
2336 	 * it completes successfully.
2337 	 */
2338 	ret = t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), &rpl);
2339 	if (ret)
2340 		goto err;
2341 
2342 	netif_napi_add(dev, &rspq->napi, napi_rx_handler, 64);
2343 	rspq->cur_desc = rspq->desc;
2344 	rspq->cidx = 0;
2345 	rspq->gen = 1;
2346 	rspq->next_intr_params = rspq->intr_params;
2347 	rspq->cntxt_id = be16_to_cpu(rpl.iqid);
2348 	rspq->bar2_addr = bar2_address(adapter,
2349 				       rspq->cntxt_id,
2350 				       T4_BAR2_QTYPE_INGRESS,
2351 				       &rspq->bar2_qid);
2352 	rspq->abs_id = be16_to_cpu(rpl.physiqid);
2353 	rspq->size--;			/* subtract status entry */
2354 	rspq->adapter = adapter;
2355 	rspq->netdev = dev;
2356 	rspq->handler = hnd;
2357 
2358 	/* set offset to -1 to distinguish ingress queues without FL */
2359 	rspq->offset = fl ? 0 : -1;
2360 
2361 	if (fl) {
2362 		fl->cntxt_id = be16_to_cpu(rpl.fl0id);
2363 		fl->avail = 0;
2364 		fl->pend_cred = 0;
2365 		fl->pidx = 0;
2366 		fl->cidx = 0;
2367 		fl->alloc_failed = 0;
2368 		fl->large_alloc_failed = 0;
2369 		fl->starving = 0;
2370 
2371 		/* Note, we must initialize the BAR2 Free List User Doorbell
2372 		 * information before refilling the Free List!
2373 		 */
2374 		fl->bar2_addr = bar2_address(adapter,
2375 					     fl->cntxt_id,
2376 					     T4_BAR2_QTYPE_EGRESS,
2377 					     &fl->bar2_qid);
2378 
2379 		refill_fl(adapter, fl, fl_cap(fl), GFP_KERNEL);
2380 	}
2381 
2382 	return 0;
2383 
2384 err:
2385 	/*
2386 	 * An error occurred.  Clean up our partial allocation state and
2387 	 * return the error.
2388 	 */
2389 	if (rspq->desc) {
2390 		dma_free_coherent(adapter->pdev_dev, rspq->size * rspq->iqe_len,
2391 				  rspq->desc, rspq->phys_addr);
2392 		rspq->desc = NULL;
2393 	}
2394 	if (fl && fl->desc) {
2395 		kfree(fl->sdesc);
2396 		fl->sdesc = NULL;
2397 		dma_free_coherent(adapter->pdev_dev, flsz * EQ_UNIT,
2398 				  fl->desc, fl->addr);
2399 		fl->desc = NULL;
2400 	}
2401 	return ret;
2402 }
2403 
2404 /**
2405  *	t4vf_sge_alloc_eth_txq - allocate an SGE Ethernet TX Queue
2406  *	@adapter: the adapter
2407  *	@txq: pointer to the new txq to be filled in
2408  *	@dev: the network device
2409  *	@devq: the network TX queue associated with the new txq
2410  *	@iqid: the relative ingress queue ID to which events relating to
2411  *		the new txq should be directed
2412  */
2413 int t4vf_sge_alloc_eth_txq(struct adapter *adapter, struct sge_eth_txq *txq,
2414 			   struct net_device *dev, struct netdev_queue *devq,
2415 			   unsigned int iqid)
2416 {
2417 	unsigned int chip_ver = CHELSIO_CHIP_VERSION(adapter->params.chip);
2418 	struct port_info *pi = netdev_priv(dev);
2419 	struct fw_eq_eth_cmd cmd, rpl;
2420 	struct sge *s = &adapter->sge;
2421 	int ret, nentries;
2422 
2423 	/*
2424 	 * Calculate the size of the hardware TX Queue (including the Status
2425 	 * Page on the end of the TX Queue) in units of TX Descriptors.
2426 	 */
2427 	nentries = txq->q.size + s->stat_len / sizeof(struct tx_desc);
2428 
2429 	/*
2430 	 * Allocate the hardware ring for the TX ring (with space for its
2431 	 * status page) along with the associated software descriptor ring.
2432 	 */
2433 	txq->q.desc = alloc_ring(adapter->pdev_dev, txq->q.size,
2434 				 sizeof(struct tx_desc),
2435 				 sizeof(struct tx_sw_desc),
2436 				 &txq->q.phys_addr, &txq->q.sdesc, s->stat_len);
2437 	if (!txq->q.desc)
2438 		return -ENOMEM;
2439 
2440 	/*
2441 	 * Fill in the Egress Queue Command.  Note: As with the direct use of
2442 	 * the firmware Ingress Queue COmmand above in our RXQ allocation
2443 	 * routine, ideally, this code would be in t4vf_hw.c.  Again, we'll
2444 	 * have to see if there's some reasonable way to parameterize it
2445 	 * into the common code ...
2446 	 */
2447 	memset(&cmd, 0, sizeof(cmd));
2448 	cmd.op_to_vfn = cpu_to_be32(FW_CMD_OP_V(FW_EQ_ETH_CMD) |
2449 				    FW_CMD_REQUEST_F |
2450 				    FW_CMD_WRITE_F |
2451 				    FW_CMD_EXEC_F);
2452 	cmd.alloc_to_len16 = cpu_to_be32(FW_EQ_ETH_CMD_ALLOC_F |
2453 					 FW_EQ_ETH_CMD_EQSTART_F |
2454 					 FW_LEN16(cmd));
2455 	cmd.autoequiqe_to_viid = cpu_to_be32(FW_EQ_ETH_CMD_AUTOEQUEQE_F |
2456 					     FW_EQ_ETH_CMD_VIID_V(pi->viid));
2457 	cmd.fetchszm_to_iqid =
2458 		cpu_to_be32(FW_EQ_ETH_CMD_HOSTFCMODE_V(SGE_HOSTFCMODE_STPG) |
2459 			    FW_EQ_ETH_CMD_PCIECHN_V(pi->port_id) |
2460 			    FW_EQ_ETH_CMD_IQID_V(iqid));
2461 	cmd.dcaen_to_eqsize =
2462 		cpu_to_be32(FW_EQ_ETH_CMD_FBMIN_V(chip_ver <= CHELSIO_T5
2463 						  ? FETCHBURSTMIN_64B_X
2464 						  : FETCHBURSTMIN_64B_T6_X) |
2465 			    FW_EQ_ETH_CMD_FBMAX_V(FETCHBURSTMAX_512B_X) |
2466 			    FW_EQ_ETH_CMD_CIDXFTHRESH_V(
2467 						CIDXFLUSHTHRESH_32_X) |
2468 			    FW_EQ_ETH_CMD_EQSIZE_V(nentries));
2469 	cmd.eqaddr = cpu_to_be64(txq->q.phys_addr);
2470 
2471 	/*
2472 	 * Issue the firmware Egress Queue Command and extract the results if
2473 	 * it completes successfully.
2474 	 */
2475 	ret = t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), &rpl);
2476 	if (ret) {
2477 		/*
2478 		 * The girmware Ingress Queue Command failed for some reason.
2479 		 * Free up our partial allocation state and return the error.
2480 		 */
2481 		kfree(txq->q.sdesc);
2482 		txq->q.sdesc = NULL;
2483 		dma_free_coherent(adapter->pdev_dev,
2484 				  nentries * sizeof(struct tx_desc),
2485 				  txq->q.desc, txq->q.phys_addr);
2486 		txq->q.desc = NULL;
2487 		return ret;
2488 	}
2489 
2490 	txq->q.in_use = 0;
2491 	txq->q.cidx = 0;
2492 	txq->q.pidx = 0;
2493 	txq->q.stat = (void *)&txq->q.desc[txq->q.size];
2494 	txq->q.cntxt_id = FW_EQ_ETH_CMD_EQID_G(be32_to_cpu(rpl.eqid_pkd));
2495 	txq->q.bar2_addr = bar2_address(adapter,
2496 					txq->q.cntxt_id,
2497 					T4_BAR2_QTYPE_EGRESS,
2498 					&txq->q.bar2_qid);
2499 	txq->q.abs_id =
2500 		FW_EQ_ETH_CMD_PHYSEQID_G(be32_to_cpu(rpl.physeqid_pkd));
2501 	txq->txq = devq;
2502 	txq->tso = 0;
2503 	txq->tx_cso = 0;
2504 	txq->vlan_ins = 0;
2505 	txq->q.stops = 0;
2506 	txq->q.restarts = 0;
2507 	txq->mapping_err = 0;
2508 	return 0;
2509 }
2510 
2511 /*
2512  * Free the DMA map resources associated with a TX queue.
2513  */
2514 static void free_txq(struct adapter *adapter, struct sge_txq *tq)
2515 {
2516 	struct sge *s = &adapter->sge;
2517 
2518 	dma_free_coherent(adapter->pdev_dev,
2519 			  tq->size * sizeof(*tq->desc) + s->stat_len,
2520 			  tq->desc, tq->phys_addr);
2521 	tq->cntxt_id = 0;
2522 	tq->sdesc = NULL;
2523 	tq->desc = NULL;
2524 }
2525 
2526 /*
2527  * Free the resources associated with a response queue (possibly including a
2528  * free list).
2529  */
2530 static void free_rspq_fl(struct adapter *adapter, struct sge_rspq *rspq,
2531 			 struct sge_fl *fl)
2532 {
2533 	struct sge *s = &adapter->sge;
2534 	unsigned int flid = fl ? fl->cntxt_id : 0xffff;
2535 
2536 	t4vf_iq_free(adapter, FW_IQ_TYPE_FL_INT_CAP,
2537 		     rspq->cntxt_id, flid, 0xffff);
2538 	dma_free_coherent(adapter->pdev_dev, (rspq->size + 1) * rspq->iqe_len,
2539 			  rspq->desc, rspq->phys_addr);
2540 	netif_napi_del(&rspq->napi);
2541 	rspq->netdev = NULL;
2542 	rspq->cntxt_id = 0;
2543 	rspq->abs_id = 0;
2544 	rspq->desc = NULL;
2545 
2546 	if (fl) {
2547 		free_rx_bufs(adapter, fl, fl->avail);
2548 		dma_free_coherent(adapter->pdev_dev,
2549 				  fl->size * sizeof(*fl->desc) + s->stat_len,
2550 				  fl->desc, fl->addr);
2551 		kfree(fl->sdesc);
2552 		fl->sdesc = NULL;
2553 		fl->cntxt_id = 0;
2554 		fl->desc = NULL;
2555 	}
2556 }
2557 
2558 /**
2559  *	t4vf_free_sge_resources - free SGE resources
2560  *	@adapter: the adapter
2561  *
2562  *	Frees resources used by the SGE queue sets.
2563  */
2564 void t4vf_free_sge_resources(struct adapter *adapter)
2565 {
2566 	struct sge *s = &adapter->sge;
2567 	struct sge_eth_rxq *rxq = s->ethrxq;
2568 	struct sge_eth_txq *txq = s->ethtxq;
2569 	struct sge_rspq *evtq = &s->fw_evtq;
2570 	struct sge_rspq *intrq = &s->intrq;
2571 	int qs;
2572 
2573 	for (qs = 0; qs < adapter->sge.ethqsets; qs++, rxq++, txq++) {
2574 		if (rxq->rspq.desc)
2575 			free_rspq_fl(adapter, &rxq->rspq, &rxq->fl);
2576 		if (txq->q.desc) {
2577 			t4vf_eth_eq_free(adapter, txq->q.cntxt_id);
2578 			free_tx_desc(adapter, &txq->q, txq->q.in_use, true);
2579 			kfree(txq->q.sdesc);
2580 			free_txq(adapter, &txq->q);
2581 		}
2582 	}
2583 	if (evtq->desc)
2584 		free_rspq_fl(adapter, evtq, NULL);
2585 	if (intrq->desc)
2586 		free_rspq_fl(adapter, intrq, NULL);
2587 }
2588 
2589 /**
2590  *	t4vf_sge_start - enable SGE operation
2591  *	@adapter: the adapter
2592  *
2593  *	Start tasklets and timers associated with the DMA engine.
2594  */
2595 void t4vf_sge_start(struct adapter *adapter)
2596 {
2597 	adapter->sge.ethtxq_rover = 0;
2598 	mod_timer(&adapter->sge.rx_timer, jiffies + RX_QCHECK_PERIOD);
2599 	mod_timer(&adapter->sge.tx_timer, jiffies + TX_QCHECK_PERIOD);
2600 }
2601 
2602 /**
2603  *	t4vf_sge_stop - disable SGE operation
2604  *	@adapter: the adapter
2605  *
2606  *	Stop tasklets and timers associated with the DMA engine.  Note that
2607  *	this is effective only if measures have been taken to disable any HW
2608  *	events that may restart them.
2609  */
2610 void t4vf_sge_stop(struct adapter *adapter)
2611 {
2612 	struct sge *s = &adapter->sge;
2613 
2614 	if (s->rx_timer.function)
2615 		del_timer_sync(&s->rx_timer);
2616 	if (s->tx_timer.function)
2617 		del_timer_sync(&s->tx_timer);
2618 }
2619 
2620 /**
2621  *	t4vf_sge_init - initialize SGE
2622  *	@adapter: the adapter
2623  *
2624  *	Performs SGE initialization needed every time after a chip reset.
2625  *	We do not initialize any of the queue sets here, instead the driver
2626  *	top-level must request those individually.  We also do not enable DMA
2627  *	here, that should be done after the queues have been set up.
2628  */
2629 int t4vf_sge_init(struct adapter *adapter)
2630 {
2631 	struct sge_params *sge_params = &adapter->params.sge;
2632 	u32 fl_small_pg = sge_params->sge_fl_buffer_size[0];
2633 	u32 fl_large_pg = sge_params->sge_fl_buffer_size[1];
2634 	struct sge *s = &adapter->sge;
2635 
2636 	/*
2637 	 * Start by vetting the basic SGE parameters which have been set up by
2638 	 * the Physical Function Driver.  Ideally we should be able to deal
2639 	 * with _any_ configuration.  Practice is different ...
2640 	 */
2641 
2642 	/* We only bother using the Large Page logic if the Large Page Buffer
2643 	 * is larger than our Page Size Buffer.
2644 	 */
2645 	if (fl_large_pg <= fl_small_pg)
2646 		fl_large_pg = 0;
2647 
2648 	/* The Page Size Buffer must be exactly equal to our Page Size and the
2649 	 * Large Page Size Buffer should be 0 (per above) or a power of 2.
2650 	 */
2651 	if (fl_small_pg != PAGE_SIZE ||
2652 	    (fl_large_pg & (fl_large_pg - 1)) != 0) {
2653 		dev_err(adapter->pdev_dev, "bad SGE FL buffer sizes [%d, %d]\n",
2654 			fl_small_pg, fl_large_pg);
2655 		return -EINVAL;
2656 	}
2657 	if ((sge_params->sge_control & RXPKTCPLMODE_F) !=
2658 	    RXPKTCPLMODE_V(RXPKTCPLMODE_SPLIT_X)) {
2659 		dev_err(adapter->pdev_dev, "bad SGE CPL MODE\n");
2660 		return -EINVAL;
2661 	}
2662 
2663 	/*
2664 	 * Now translate the adapter parameters into our internal forms.
2665 	 */
2666 	if (fl_large_pg)
2667 		s->fl_pg_order = ilog2(fl_large_pg) - PAGE_SHIFT;
2668 	s->stat_len = ((sge_params->sge_control & EGRSTATUSPAGESIZE_F)
2669 			? 128 : 64);
2670 	s->pktshift = PKTSHIFT_G(sge_params->sge_control);
2671 	s->fl_align = t4vf_fl_pkt_align(adapter);
2672 
2673 	/* A FL with <= fl_starve_thres buffers is starving and a periodic
2674 	 * timer will attempt to refill it.  This needs to be larger than the
2675 	 * SGE's Egress Congestion Threshold.  If it isn't, then we can get
2676 	 * stuck waiting for new packets while the SGE is waiting for us to
2677 	 * give it more Free List entries.  (Note that the SGE's Egress
2678 	 * Congestion Threshold is in units of 2 Free List pointers.)
2679 	 */
2680 	switch (CHELSIO_CHIP_VERSION(adapter->params.chip)) {
2681 	case CHELSIO_T4:
2682 		s->fl_starve_thres =
2683 		   EGRTHRESHOLD_G(sge_params->sge_congestion_control);
2684 		break;
2685 	case CHELSIO_T5:
2686 		s->fl_starve_thres =
2687 		   EGRTHRESHOLDPACKING_G(sge_params->sge_congestion_control);
2688 		break;
2689 	case CHELSIO_T6:
2690 	default:
2691 		s->fl_starve_thres =
2692 		   T6_EGRTHRESHOLDPACKING_G(sge_params->sge_congestion_control);
2693 		break;
2694 	}
2695 	s->fl_starve_thres = s->fl_starve_thres * 2 + 1;
2696 
2697 	/*
2698 	 * Set up tasklet timers.
2699 	 */
2700 	timer_setup(&s->rx_timer, sge_rx_timer_cb, 0);
2701 	timer_setup(&s->tx_timer, sge_tx_timer_cb, 0);
2702 
2703 	/*
2704 	 * Initialize Forwarded Interrupt Queue lock.
2705 	 */
2706 	spin_lock_init(&s->intrq_lock);
2707 
2708 	return 0;
2709 }
2710