xref: /linux/drivers/block/aoe/aoecmd.c (revision 3932b9ca55b0be314a36d3e84faff3e823c081f5)
1 /* Copyright (c) 2013 Coraid, Inc.  See COPYING for GPL terms. */
2 /*
3  * aoecmd.c
4  * Filesystem request handling methods
5  */
6 
7 #include <linux/ata.h>
8 #include <linux/slab.h>
9 #include <linux/hdreg.h>
10 #include <linux/blkdev.h>
11 #include <linux/skbuff.h>
12 #include <linux/netdevice.h>
13 #include <linux/genhd.h>
14 #include <linux/moduleparam.h>
15 #include <linux/workqueue.h>
16 #include <linux/kthread.h>
17 #include <net/net_namespace.h>
18 #include <asm/unaligned.h>
19 #include <linux/uio.h>
20 #include "aoe.h"
21 
22 #define MAXIOC (8192)	/* default meant to avoid most soft lockups */
23 
24 static void ktcomplete(struct frame *, struct sk_buff *);
25 static int count_targets(struct aoedev *d, int *untainted);
26 
27 static struct buf *nextbuf(struct aoedev *);
28 
29 static int aoe_deadsecs = 60 * 3;
30 module_param(aoe_deadsecs, int, 0644);
31 MODULE_PARM_DESC(aoe_deadsecs, "After aoe_deadsecs seconds, give up and fail dev.");
32 
33 static int aoe_maxout = 64;
34 module_param(aoe_maxout, int, 0644);
35 MODULE_PARM_DESC(aoe_maxout,
36 	"Only aoe_maxout outstanding packets for every MAC on eX.Y.");
37 
38 /* The number of online cpus during module initialization gives us a
39  * convenient heuristic cap on the parallelism used for ktio threads
40  * doing I/O completion.  It is not important that the cap equal the
41  * actual number of running CPUs at any given time, but because of CPU
42  * hotplug, we take care to use ncpus instead of using
43  * num_online_cpus() after module initialization.
44  */
45 static int ncpus;
46 
47 /* mutex lock used for synchronization while thread spawning */
48 static DEFINE_MUTEX(ktio_spawn_lock);
49 
50 static wait_queue_head_t *ktiowq;
51 static struct ktstate *kts;
52 
53 /* io completion queue */
54 struct iocq_ktio {
55 	struct list_head head;
56 	spinlock_t lock;
57 };
58 static struct iocq_ktio *iocq;
59 
60 static struct page *empty_page;
61 
62 static struct sk_buff *
63 new_skb(ulong len)
64 {
65 	struct sk_buff *skb;
66 
67 	skb = alloc_skb(len + MAX_HEADER, GFP_ATOMIC);
68 	if (skb) {
69 		skb_reserve(skb, MAX_HEADER);
70 		skb_reset_mac_header(skb);
71 		skb_reset_network_header(skb);
72 		skb->protocol = __constant_htons(ETH_P_AOE);
73 		skb_checksum_none_assert(skb);
74 	}
75 	return skb;
76 }
77 
78 static struct frame *
79 getframe_deferred(struct aoedev *d, u32 tag)
80 {
81 	struct list_head *head, *pos, *nx;
82 	struct frame *f;
83 
84 	head = &d->rexmitq;
85 	list_for_each_safe(pos, nx, head) {
86 		f = list_entry(pos, struct frame, head);
87 		if (f->tag == tag) {
88 			list_del(pos);
89 			return f;
90 		}
91 	}
92 	return NULL;
93 }
94 
95 static struct frame *
96 getframe(struct aoedev *d, u32 tag)
97 {
98 	struct frame *f;
99 	struct list_head *head, *pos, *nx;
100 	u32 n;
101 
102 	n = tag % NFACTIVE;
103 	head = &d->factive[n];
104 	list_for_each_safe(pos, nx, head) {
105 		f = list_entry(pos, struct frame, head);
106 		if (f->tag == tag) {
107 			list_del(pos);
108 			return f;
109 		}
110 	}
111 	return NULL;
112 }
113 
114 /*
115  * Leave the top bit clear so we have tagspace for userland.
116  * The bottom 16 bits are the xmit tick for rexmit/rttavg processing.
117  * This driver reserves tag -1 to mean "unused frame."
118  */
119 static int
120 newtag(struct aoedev *d)
121 {
122 	register ulong n;
123 
124 	n = jiffies & 0xffff;
125 	return n |= (++d->lasttag & 0x7fff) << 16;
126 }
127 
128 static u32
129 aoehdr_atainit(struct aoedev *d, struct aoetgt *t, struct aoe_hdr *h)
130 {
131 	u32 host_tag = newtag(d);
132 
133 	memcpy(h->src, t->ifp->nd->dev_addr, sizeof h->src);
134 	memcpy(h->dst, t->addr, sizeof h->dst);
135 	h->type = __constant_cpu_to_be16(ETH_P_AOE);
136 	h->verfl = AOE_HVER;
137 	h->major = cpu_to_be16(d->aoemajor);
138 	h->minor = d->aoeminor;
139 	h->cmd = AOECMD_ATA;
140 	h->tag = cpu_to_be32(host_tag);
141 
142 	return host_tag;
143 }
144 
145 static inline void
146 put_lba(struct aoe_atahdr *ah, sector_t lba)
147 {
148 	ah->lba0 = lba;
149 	ah->lba1 = lba >>= 8;
150 	ah->lba2 = lba >>= 8;
151 	ah->lba3 = lba >>= 8;
152 	ah->lba4 = lba >>= 8;
153 	ah->lba5 = lba >>= 8;
154 }
155 
156 static struct aoeif *
157 ifrotate(struct aoetgt *t)
158 {
159 	struct aoeif *ifp;
160 
161 	ifp = t->ifp;
162 	ifp++;
163 	if (ifp >= &t->ifs[NAOEIFS] || ifp->nd == NULL)
164 		ifp = t->ifs;
165 	if (ifp->nd == NULL)
166 		return NULL;
167 	return t->ifp = ifp;
168 }
169 
170 static void
171 skb_pool_put(struct aoedev *d, struct sk_buff *skb)
172 {
173 	__skb_queue_tail(&d->skbpool, skb);
174 }
175 
176 static struct sk_buff *
177 skb_pool_get(struct aoedev *d)
178 {
179 	struct sk_buff *skb = skb_peek(&d->skbpool);
180 
181 	if (skb && atomic_read(&skb_shinfo(skb)->dataref) == 1) {
182 		__skb_unlink(skb, &d->skbpool);
183 		return skb;
184 	}
185 	if (skb_queue_len(&d->skbpool) < NSKBPOOLMAX &&
186 	    (skb = new_skb(ETH_ZLEN)))
187 		return skb;
188 
189 	return NULL;
190 }
191 
192 void
193 aoe_freetframe(struct frame *f)
194 {
195 	struct aoetgt *t;
196 
197 	t = f->t;
198 	f->buf = NULL;
199 	memset(&f->iter, 0, sizeof(f->iter));
200 	f->r_skb = NULL;
201 	f->flags = 0;
202 	list_add(&f->head, &t->ffree);
203 }
204 
205 static struct frame *
206 newtframe(struct aoedev *d, struct aoetgt *t)
207 {
208 	struct frame *f;
209 	struct sk_buff *skb;
210 	struct list_head *pos;
211 
212 	if (list_empty(&t->ffree)) {
213 		if (t->falloc >= NSKBPOOLMAX*2)
214 			return NULL;
215 		f = kcalloc(1, sizeof(*f), GFP_ATOMIC);
216 		if (f == NULL)
217 			return NULL;
218 		t->falloc++;
219 		f->t = t;
220 	} else {
221 		pos = t->ffree.next;
222 		list_del(pos);
223 		f = list_entry(pos, struct frame, head);
224 	}
225 
226 	skb = f->skb;
227 	if (skb == NULL) {
228 		f->skb = skb = new_skb(ETH_ZLEN);
229 		if (!skb) {
230 bail:			aoe_freetframe(f);
231 			return NULL;
232 		}
233 	}
234 
235 	if (atomic_read(&skb_shinfo(skb)->dataref) != 1) {
236 		skb = skb_pool_get(d);
237 		if (skb == NULL)
238 			goto bail;
239 		skb_pool_put(d, f->skb);
240 		f->skb = skb;
241 	}
242 
243 	skb->truesize -= skb->data_len;
244 	skb_shinfo(skb)->nr_frags = skb->data_len = 0;
245 	skb_trim(skb, 0);
246 	return f;
247 }
248 
249 static struct frame *
250 newframe(struct aoedev *d)
251 {
252 	struct frame *f;
253 	struct aoetgt *t, **tt;
254 	int totout = 0;
255 	int use_tainted;
256 	int has_untainted;
257 
258 	if (!d->targets || !d->targets[0]) {
259 		printk(KERN_ERR "aoe: NULL TARGETS!\n");
260 		return NULL;
261 	}
262 	tt = d->tgt;	/* last used target */
263 	for (use_tainted = 0, has_untainted = 0;;) {
264 		tt++;
265 		if (tt >= &d->targets[d->ntargets] || !*tt)
266 			tt = d->targets;
267 		t = *tt;
268 		if (!t->taint) {
269 			has_untainted = 1;
270 			totout += t->nout;
271 		}
272 		if (t->nout < t->maxout
273 		&& (use_tainted || !t->taint)
274 		&& t->ifp->nd) {
275 			f = newtframe(d, t);
276 			if (f) {
277 				ifrotate(t);
278 				d->tgt = tt;
279 				return f;
280 			}
281 		}
282 		if (tt == d->tgt) {	/* we've looped and found nada */
283 			if (!use_tainted && !has_untainted)
284 				use_tainted = 1;
285 			else
286 				break;
287 		}
288 	}
289 	if (totout == 0) {
290 		d->kicked++;
291 		d->flags |= DEVFL_KICKME;
292 	}
293 	return NULL;
294 }
295 
296 static void
297 skb_fillup(struct sk_buff *skb, struct bio *bio, struct bvec_iter iter)
298 {
299 	int frag = 0;
300 	struct bio_vec bv;
301 
302 	__bio_for_each_segment(bv, bio, iter, iter)
303 		skb_fill_page_desc(skb, frag++, bv.bv_page,
304 				   bv.bv_offset, bv.bv_len);
305 }
306 
307 static void
308 fhash(struct frame *f)
309 {
310 	struct aoedev *d = f->t->d;
311 	u32 n;
312 
313 	n = f->tag % NFACTIVE;
314 	list_add_tail(&f->head, &d->factive[n]);
315 }
316 
317 static void
318 ata_rw_frameinit(struct frame *f)
319 {
320 	struct aoetgt *t;
321 	struct aoe_hdr *h;
322 	struct aoe_atahdr *ah;
323 	struct sk_buff *skb;
324 	char writebit, extbit;
325 
326 	skb = f->skb;
327 	h = (struct aoe_hdr *) skb_mac_header(skb);
328 	ah = (struct aoe_atahdr *) (h + 1);
329 	skb_put(skb, sizeof(*h) + sizeof(*ah));
330 	memset(h, 0, skb->len);
331 
332 	writebit = 0x10;
333 	extbit = 0x4;
334 
335 	t = f->t;
336 	f->tag = aoehdr_atainit(t->d, t, h);
337 	fhash(f);
338 	t->nout++;
339 	f->waited = 0;
340 	f->waited_total = 0;
341 
342 	/* set up ata header */
343 	ah->scnt = f->iter.bi_size >> 9;
344 	put_lba(ah, f->iter.bi_sector);
345 	if (t->d->flags & DEVFL_EXT) {
346 		ah->aflags |= AOEAFL_EXT;
347 	} else {
348 		extbit = 0;
349 		ah->lba3 &= 0x0f;
350 		ah->lba3 |= 0xe0;	/* LBA bit + obsolete 0xa0 */
351 	}
352 	if (f->buf && bio_data_dir(f->buf->bio) == WRITE) {
353 		skb_fillup(skb, f->buf->bio, f->iter);
354 		ah->aflags |= AOEAFL_WRITE;
355 		skb->len += f->iter.bi_size;
356 		skb->data_len = f->iter.bi_size;
357 		skb->truesize += f->iter.bi_size;
358 		t->wpkts++;
359 	} else {
360 		t->rpkts++;
361 		writebit = 0;
362 	}
363 
364 	ah->cmdstat = ATA_CMD_PIO_READ | writebit | extbit;
365 	skb->dev = t->ifp->nd;
366 }
367 
368 static int
369 aoecmd_ata_rw(struct aoedev *d)
370 {
371 	struct frame *f;
372 	struct buf *buf;
373 	struct sk_buff *skb;
374 	struct sk_buff_head queue;
375 
376 	buf = nextbuf(d);
377 	if (buf == NULL)
378 		return 0;
379 	f = newframe(d);
380 	if (f == NULL)
381 		return 0;
382 
383 	/* initialize the headers & frame */
384 	f->buf = buf;
385 	f->iter = buf->iter;
386 	f->iter.bi_size = min_t(unsigned long,
387 				d->maxbcnt ?: DEFAULTBCNT,
388 				f->iter.bi_size);
389 	bio_advance_iter(buf->bio, &buf->iter, f->iter.bi_size);
390 
391 	if (!buf->iter.bi_size)
392 		d->ip.buf = NULL;
393 
394 	/* mark all tracking fields and load out */
395 	buf->nframesout += 1;
396 
397 	ata_rw_frameinit(f);
398 
399 	skb = skb_clone(f->skb, GFP_ATOMIC);
400 	if (skb) {
401 		do_gettimeofday(&f->sent);
402 		f->sent_jiffs = (u32) jiffies;
403 		__skb_queue_head_init(&queue);
404 		__skb_queue_tail(&queue, skb);
405 		aoenet_xmit(&queue);
406 	}
407 	return 1;
408 }
409 
410 /* some callers cannot sleep, and they can call this function,
411  * transmitting the packets later, when interrupts are on
412  */
413 static void
414 aoecmd_cfg_pkts(ushort aoemajor, unsigned char aoeminor, struct sk_buff_head *queue)
415 {
416 	struct aoe_hdr *h;
417 	struct aoe_cfghdr *ch;
418 	struct sk_buff *skb;
419 	struct net_device *ifp;
420 
421 	rcu_read_lock();
422 	for_each_netdev_rcu(&init_net, ifp) {
423 		dev_hold(ifp);
424 		if (!is_aoe_netif(ifp))
425 			goto cont;
426 
427 		skb = new_skb(sizeof *h + sizeof *ch);
428 		if (skb == NULL) {
429 			printk(KERN_INFO "aoe: skb alloc failure\n");
430 			goto cont;
431 		}
432 		skb_put(skb, sizeof *h + sizeof *ch);
433 		skb->dev = ifp;
434 		__skb_queue_tail(queue, skb);
435 		h = (struct aoe_hdr *) skb_mac_header(skb);
436 		memset(h, 0, sizeof *h + sizeof *ch);
437 
438 		memset(h->dst, 0xff, sizeof h->dst);
439 		memcpy(h->src, ifp->dev_addr, sizeof h->src);
440 		h->type = __constant_cpu_to_be16(ETH_P_AOE);
441 		h->verfl = AOE_HVER;
442 		h->major = cpu_to_be16(aoemajor);
443 		h->minor = aoeminor;
444 		h->cmd = AOECMD_CFG;
445 
446 cont:
447 		dev_put(ifp);
448 	}
449 	rcu_read_unlock();
450 }
451 
452 static void
453 resend(struct aoedev *d, struct frame *f)
454 {
455 	struct sk_buff *skb;
456 	struct sk_buff_head queue;
457 	struct aoe_hdr *h;
458 	struct aoetgt *t;
459 	char buf[128];
460 	u32 n;
461 
462 	t = f->t;
463 	n = newtag(d);
464 	skb = f->skb;
465 	if (ifrotate(t) == NULL) {
466 		/* probably can't happen, but set it up to fail anyway */
467 		pr_info("aoe: resend: no interfaces to rotate to.\n");
468 		ktcomplete(f, NULL);
469 		return;
470 	}
471 	h = (struct aoe_hdr *) skb_mac_header(skb);
472 
473 	if (!(f->flags & FFL_PROBE)) {
474 		snprintf(buf, sizeof(buf),
475 			"%15s e%ld.%d oldtag=%08x@%08lx newtag=%08x s=%pm d=%pm nout=%d\n",
476 			"retransmit", d->aoemajor, d->aoeminor,
477 			f->tag, jiffies, n,
478 			h->src, h->dst, t->nout);
479 		aoechr_error(buf);
480 	}
481 
482 	f->tag = n;
483 	fhash(f);
484 	h->tag = cpu_to_be32(n);
485 	memcpy(h->dst, t->addr, sizeof h->dst);
486 	memcpy(h->src, t->ifp->nd->dev_addr, sizeof h->src);
487 
488 	skb->dev = t->ifp->nd;
489 	skb = skb_clone(skb, GFP_ATOMIC);
490 	if (skb == NULL)
491 		return;
492 	do_gettimeofday(&f->sent);
493 	f->sent_jiffs = (u32) jiffies;
494 	__skb_queue_head_init(&queue);
495 	__skb_queue_tail(&queue, skb);
496 	aoenet_xmit(&queue);
497 }
498 
499 static int
500 tsince_hr(struct frame *f)
501 {
502 	struct timeval now;
503 	int n;
504 
505 	do_gettimeofday(&now);
506 	n = now.tv_usec - f->sent.tv_usec;
507 	n += (now.tv_sec - f->sent.tv_sec) * USEC_PER_SEC;
508 
509 	if (n < 0)
510 		n = -n;
511 
512 	/* For relatively long periods, use jiffies to avoid
513 	 * discrepancies caused by updates to the system time.
514 	 *
515 	 * On system with HZ of 1000, 32-bits is over 49 days
516 	 * worth of jiffies, or over 71 minutes worth of usecs.
517 	 *
518 	 * Jiffies overflow is handled by subtraction of unsigned ints:
519 	 * (gdb) print (unsigned) 2 - (unsigned) 0xfffffffe
520 	 * $3 = 4
521 	 * (gdb)
522 	 */
523 	if (n > USEC_PER_SEC / 4) {
524 		n = ((u32) jiffies) - f->sent_jiffs;
525 		n *= USEC_PER_SEC / HZ;
526 	}
527 
528 	return n;
529 }
530 
531 static int
532 tsince(u32 tag)
533 {
534 	int n;
535 
536 	n = jiffies & 0xffff;
537 	n -= tag & 0xffff;
538 	if (n < 0)
539 		n += 1<<16;
540 	return jiffies_to_usecs(n + 1);
541 }
542 
543 static struct aoeif *
544 getif(struct aoetgt *t, struct net_device *nd)
545 {
546 	struct aoeif *p, *e;
547 
548 	p = t->ifs;
549 	e = p + NAOEIFS;
550 	for (; p < e; p++)
551 		if (p->nd == nd)
552 			return p;
553 	return NULL;
554 }
555 
556 static void
557 ejectif(struct aoetgt *t, struct aoeif *ifp)
558 {
559 	struct aoeif *e;
560 	struct net_device *nd;
561 	ulong n;
562 
563 	nd = ifp->nd;
564 	e = t->ifs + NAOEIFS - 1;
565 	n = (e - ifp) * sizeof *ifp;
566 	memmove(ifp, ifp+1, n);
567 	e->nd = NULL;
568 	dev_put(nd);
569 }
570 
571 static struct frame *
572 reassign_frame(struct frame *f)
573 {
574 	struct frame *nf;
575 	struct sk_buff *skb;
576 
577 	nf = newframe(f->t->d);
578 	if (!nf)
579 		return NULL;
580 	if (nf->t == f->t) {
581 		aoe_freetframe(nf);
582 		return NULL;
583 	}
584 
585 	skb = nf->skb;
586 	nf->skb = f->skb;
587 	nf->buf = f->buf;
588 	nf->iter = f->iter;
589 	nf->waited = 0;
590 	nf->waited_total = f->waited_total;
591 	nf->sent = f->sent;
592 	nf->sent_jiffs = f->sent_jiffs;
593 	f->skb = skb;
594 
595 	return nf;
596 }
597 
598 static void
599 probe(struct aoetgt *t)
600 {
601 	struct aoedev *d;
602 	struct frame *f;
603 	struct sk_buff *skb;
604 	struct sk_buff_head queue;
605 	size_t n, m;
606 	int frag;
607 
608 	d = t->d;
609 	f = newtframe(d, t);
610 	if (!f) {
611 		pr_err("%s %pm for e%ld.%d: %s\n",
612 			"aoe: cannot probe remote address",
613 			t->addr,
614 			(long) d->aoemajor, d->aoeminor,
615 			"no frame available");
616 		return;
617 	}
618 	f->flags |= FFL_PROBE;
619 	ifrotate(t);
620 	f->iter.bi_size = t->d->maxbcnt ? t->d->maxbcnt : DEFAULTBCNT;
621 	ata_rw_frameinit(f);
622 	skb = f->skb;
623 	for (frag = 0, n = f->iter.bi_size; n > 0; ++frag, n -= m) {
624 		if (n < PAGE_SIZE)
625 			m = n;
626 		else
627 			m = PAGE_SIZE;
628 		skb_fill_page_desc(skb, frag, empty_page, 0, m);
629 	}
630 	skb->len += f->iter.bi_size;
631 	skb->data_len = f->iter.bi_size;
632 	skb->truesize += f->iter.bi_size;
633 
634 	skb = skb_clone(f->skb, GFP_ATOMIC);
635 	if (skb) {
636 		do_gettimeofday(&f->sent);
637 		f->sent_jiffs = (u32) jiffies;
638 		__skb_queue_head_init(&queue);
639 		__skb_queue_tail(&queue, skb);
640 		aoenet_xmit(&queue);
641 	}
642 }
643 
644 static long
645 rto(struct aoedev *d)
646 {
647 	long t;
648 
649 	t = 2 * d->rttavg >> RTTSCALE;
650 	t += 8 * d->rttdev >> RTTDSCALE;
651 	if (t == 0)
652 		t = 1;
653 
654 	return t;
655 }
656 
657 static void
658 rexmit_deferred(struct aoedev *d)
659 {
660 	struct aoetgt *t;
661 	struct frame *f;
662 	struct frame *nf;
663 	struct list_head *pos, *nx, *head;
664 	int since;
665 	int untainted;
666 
667 	count_targets(d, &untainted);
668 
669 	head = &d->rexmitq;
670 	list_for_each_safe(pos, nx, head) {
671 		f = list_entry(pos, struct frame, head);
672 		t = f->t;
673 		if (t->taint) {
674 			if (!(f->flags & FFL_PROBE)) {
675 				nf = reassign_frame(f);
676 				if (nf) {
677 					if (t->nout_probes == 0
678 					&& untainted > 0) {
679 						probe(t);
680 						t->nout_probes++;
681 					}
682 					list_replace(&f->head, &nf->head);
683 					pos = &nf->head;
684 					aoe_freetframe(f);
685 					f = nf;
686 					t = f->t;
687 				}
688 			} else if (untainted < 1) {
689 				/* don't probe w/o other untainted aoetgts */
690 				goto stop_probe;
691 			} else if (tsince_hr(f) < t->taint * rto(d)) {
692 				/* reprobe slowly when taint is high */
693 				continue;
694 			}
695 		} else if (f->flags & FFL_PROBE) {
696 stop_probe:		/* don't probe untainted aoetgts */
697 			list_del(pos);
698 			aoe_freetframe(f);
699 			/* leaving d->kicked, because this is routine */
700 			f->t->d->flags |= DEVFL_KICKME;
701 			continue;
702 		}
703 		if (t->nout >= t->maxout)
704 			continue;
705 		list_del(pos);
706 		t->nout++;
707 		if (f->flags & FFL_PROBE)
708 			t->nout_probes++;
709 		since = tsince_hr(f);
710 		f->waited += since;
711 		f->waited_total += since;
712 		resend(d, f);
713 	}
714 }
715 
716 /* An aoetgt accumulates demerits quickly, and successful
717  * probing redeems the aoetgt slowly.
718  */
719 static void
720 scorn(struct aoetgt *t)
721 {
722 	int n;
723 
724 	n = t->taint++;
725 	t->taint += t->taint * 2;
726 	if (n > t->taint)
727 		t->taint = n;
728 	if (t->taint > MAX_TAINT)
729 		t->taint = MAX_TAINT;
730 }
731 
732 static int
733 count_targets(struct aoedev *d, int *untainted)
734 {
735 	int i, good;
736 
737 	for (i = good = 0; i < d->ntargets && d->targets[i]; ++i)
738 		if (d->targets[i]->taint == 0)
739 			good++;
740 
741 	if (untainted)
742 		*untainted = good;
743 	return i;
744 }
745 
746 static void
747 rexmit_timer(ulong vp)
748 {
749 	struct aoedev *d;
750 	struct aoetgt *t;
751 	struct aoeif *ifp;
752 	struct frame *f;
753 	struct list_head *head, *pos, *nx;
754 	LIST_HEAD(flist);
755 	register long timeout;
756 	ulong flags, n;
757 	int i;
758 	int utgts;	/* number of aoetgt descriptors (not slots) */
759 	int since;
760 
761 	d = (struct aoedev *) vp;
762 
763 	spin_lock_irqsave(&d->lock, flags);
764 
765 	/* timeout based on observed timings and variations */
766 	timeout = rto(d);
767 
768 	utgts = count_targets(d, NULL);
769 
770 	if (d->flags & DEVFL_TKILL) {
771 		spin_unlock_irqrestore(&d->lock, flags);
772 		return;
773 	}
774 
775 	/* collect all frames to rexmit into flist */
776 	for (i = 0; i < NFACTIVE; i++) {
777 		head = &d->factive[i];
778 		list_for_each_safe(pos, nx, head) {
779 			f = list_entry(pos, struct frame, head);
780 			if (tsince_hr(f) < timeout)
781 				break;	/* end of expired frames */
782 			/* move to flist for later processing */
783 			list_move_tail(pos, &flist);
784 		}
785 	}
786 
787 	/* process expired frames */
788 	while (!list_empty(&flist)) {
789 		pos = flist.next;
790 		f = list_entry(pos, struct frame, head);
791 		since = tsince_hr(f);
792 		n = f->waited_total + since;
793 		n /= USEC_PER_SEC;
794 		if (aoe_deadsecs
795 		&& n > aoe_deadsecs
796 		&& !(f->flags & FFL_PROBE)) {
797 			/* Waited too long.  Device failure.
798 			 * Hang all frames on first hash bucket for downdev
799 			 * to clean up.
800 			 */
801 			list_splice(&flist, &d->factive[0]);
802 			aoedev_downdev(d);
803 			goto out;
804 		}
805 
806 		t = f->t;
807 		n = f->waited + since;
808 		n /= USEC_PER_SEC;
809 		if (aoe_deadsecs && utgts > 0
810 		&& (n > aoe_deadsecs / utgts || n > HARD_SCORN_SECS))
811 			scorn(t); /* avoid this target */
812 
813 		if (t->maxout != 1) {
814 			t->ssthresh = t->maxout / 2;
815 			t->maxout = 1;
816 		}
817 
818 		if (f->flags & FFL_PROBE) {
819 			t->nout_probes--;
820 		} else {
821 			ifp = getif(t, f->skb->dev);
822 			if (ifp && ++ifp->lost > (t->nframes << 1)
823 			&& (ifp != t->ifs || t->ifs[1].nd)) {
824 				ejectif(t, ifp);
825 				ifp = NULL;
826 			}
827 		}
828 		list_move_tail(pos, &d->rexmitq);
829 		t->nout--;
830 	}
831 	rexmit_deferred(d);
832 
833 out:
834 	if ((d->flags & DEVFL_KICKME) && d->blkq) {
835 		d->flags &= ~DEVFL_KICKME;
836 		d->blkq->request_fn(d->blkq);
837 	}
838 
839 	d->timer.expires = jiffies + TIMERTICK;
840 	add_timer(&d->timer);
841 
842 	spin_unlock_irqrestore(&d->lock, flags);
843 }
844 
845 static unsigned long
846 rqbiocnt(struct request *r)
847 {
848 	struct bio *bio;
849 	unsigned long n = 0;
850 
851 	__rq_for_each_bio(bio, r)
852 		n++;
853 	return n;
854 }
855 
856 /* This can be removed if we are certain that no users of the block
857  * layer will ever use zero-count pages in bios.  Otherwise we have to
858  * protect against the put_page sometimes done by the network layer.
859  *
860  * See http://oss.sgi.com/archives/xfs/2007-01/msg00594.html for
861  * discussion.
862  *
863  * We cannot use get_page in the workaround, because it insists on a
864  * positive page count as a precondition.  So we use _count directly.
865  */
866 static void
867 bio_pageinc(struct bio *bio)
868 {
869 	struct bio_vec bv;
870 	struct page *page;
871 	struct bvec_iter iter;
872 
873 	bio_for_each_segment(bv, bio, iter) {
874 		/* Non-zero page count for non-head members of
875 		 * compound pages is no longer allowed by the kernel.
876 		 */
877 		page = compound_head(bv.bv_page);
878 		atomic_inc(&page->_count);
879 	}
880 }
881 
882 static void
883 bio_pagedec(struct bio *bio)
884 {
885 	struct page *page;
886 	struct bio_vec bv;
887 	struct bvec_iter iter;
888 
889 	bio_for_each_segment(bv, bio, iter) {
890 		page = compound_head(bv.bv_page);
891 		atomic_dec(&page->_count);
892 	}
893 }
894 
895 static void
896 bufinit(struct buf *buf, struct request *rq, struct bio *bio)
897 {
898 	memset(buf, 0, sizeof(*buf));
899 	buf->rq = rq;
900 	buf->bio = bio;
901 	buf->iter = bio->bi_iter;
902 	bio_pageinc(bio);
903 }
904 
905 static struct buf *
906 nextbuf(struct aoedev *d)
907 {
908 	struct request *rq;
909 	struct request_queue *q;
910 	struct buf *buf;
911 	struct bio *bio;
912 
913 	q = d->blkq;
914 	if (q == NULL)
915 		return NULL;	/* initializing */
916 	if (d->ip.buf)
917 		return d->ip.buf;
918 	rq = d->ip.rq;
919 	if (rq == NULL) {
920 		rq = blk_peek_request(q);
921 		if (rq == NULL)
922 			return NULL;
923 		blk_start_request(rq);
924 		d->ip.rq = rq;
925 		d->ip.nxbio = rq->bio;
926 		rq->special = (void *) rqbiocnt(rq);
927 	}
928 	buf = mempool_alloc(d->bufpool, GFP_ATOMIC);
929 	if (buf == NULL) {
930 		pr_err("aoe: nextbuf: unable to mempool_alloc!\n");
931 		return NULL;
932 	}
933 	bio = d->ip.nxbio;
934 	bufinit(buf, rq, bio);
935 	bio = bio->bi_next;
936 	d->ip.nxbio = bio;
937 	if (bio == NULL)
938 		d->ip.rq = NULL;
939 	return d->ip.buf = buf;
940 }
941 
942 /* enters with d->lock held */
943 void
944 aoecmd_work(struct aoedev *d)
945 {
946 	rexmit_deferred(d);
947 	while (aoecmd_ata_rw(d))
948 		;
949 }
950 
951 /* this function performs work that has been deferred until sleeping is OK
952  */
953 void
954 aoecmd_sleepwork(struct work_struct *work)
955 {
956 	struct aoedev *d = container_of(work, struct aoedev, work);
957 	struct block_device *bd;
958 	u64 ssize;
959 
960 	if (d->flags & DEVFL_GDALLOC)
961 		aoeblk_gdalloc(d);
962 
963 	if (d->flags & DEVFL_NEWSIZE) {
964 		ssize = get_capacity(d->gd);
965 		bd = bdget_disk(d->gd, 0);
966 		if (bd) {
967 			mutex_lock(&bd->bd_inode->i_mutex);
968 			i_size_write(bd->bd_inode, (loff_t)ssize<<9);
969 			mutex_unlock(&bd->bd_inode->i_mutex);
970 			bdput(bd);
971 		}
972 		spin_lock_irq(&d->lock);
973 		d->flags |= DEVFL_UP;
974 		d->flags &= ~DEVFL_NEWSIZE;
975 		spin_unlock_irq(&d->lock);
976 	}
977 }
978 
979 static void
980 ata_ident_fixstring(u16 *id, int ns)
981 {
982 	u16 s;
983 
984 	while (ns-- > 0) {
985 		s = *id;
986 		*id++ = s >> 8 | s << 8;
987 	}
988 }
989 
990 static void
991 ataid_complete(struct aoedev *d, struct aoetgt *t, unsigned char *id)
992 {
993 	u64 ssize;
994 	u16 n;
995 
996 	/* word 83: command set supported */
997 	n = get_unaligned_le16(&id[83 << 1]);
998 
999 	/* word 86: command set/feature enabled */
1000 	n |= get_unaligned_le16(&id[86 << 1]);
1001 
1002 	if (n & (1<<10)) {	/* bit 10: LBA 48 */
1003 		d->flags |= DEVFL_EXT;
1004 
1005 		/* word 100: number lba48 sectors */
1006 		ssize = get_unaligned_le64(&id[100 << 1]);
1007 
1008 		/* set as in ide-disk.c:init_idedisk_capacity */
1009 		d->geo.cylinders = ssize;
1010 		d->geo.cylinders /= (255 * 63);
1011 		d->geo.heads = 255;
1012 		d->geo.sectors = 63;
1013 	} else {
1014 		d->flags &= ~DEVFL_EXT;
1015 
1016 		/* number lba28 sectors */
1017 		ssize = get_unaligned_le32(&id[60 << 1]);
1018 
1019 		/* NOTE: obsolete in ATA 6 */
1020 		d->geo.cylinders = get_unaligned_le16(&id[54 << 1]);
1021 		d->geo.heads = get_unaligned_le16(&id[55 << 1]);
1022 		d->geo.sectors = get_unaligned_le16(&id[56 << 1]);
1023 	}
1024 
1025 	ata_ident_fixstring((u16 *) &id[10<<1], 10);	/* serial */
1026 	ata_ident_fixstring((u16 *) &id[23<<1], 4);	/* firmware */
1027 	ata_ident_fixstring((u16 *) &id[27<<1], 20);	/* model */
1028 	memcpy(d->ident, id, sizeof(d->ident));
1029 
1030 	if (d->ssize != ssize)
1031 		printk(KERN_INFO
1032 			"aoe: %pm e%ld.%d v%04x has %llu sectors\n",
1033 			t->addr,
1034 			d->aoemajor, d->aoeminor,
1035 			d->fw_ver, (long long)ssize);
1036 	d->ssize = ssize;
1037 	d->geo.start = 0;
1038 	if (d->flags & (DEVFL_GDALLOC|DEVFL_NEWSIZE))
1039 		return;
1040 	if (d->gd != NULL) {
1041 		set_capacity(d->gd, ssize);
1042 		d->flags |= DEVFL_NEWSIZE;
1043 	} else
1044 		d->flags |= DEVFL_GDALLOC;
1045 	schedule_work(&d->work);
1046 }
1047 
1048 static void
1049 calc_rttavg(struct aoedev *d, struct aoetgt *t, int rtt)
1050 {
1051 	register long n;
1052 
1053 	n = rtt;
1054 
1055 	/* cf. Congestion Avoidance and Control, Jacobson & Karels, 1988 */
1056 	n -= d->rttavg >> RTTSCALE;
1057 	d->rttavg += n;
1058 	if (n < 0)
1059 		n = -n;
1060 	n -= d->rttdev >> RTTDSCALE;
1061 	d->rttdev += n;
1062 
1063 	if (!t || t->maxout >= t->nframes)
1064 		return;
1065 	if (t->maxout < t->ssthresh)
1066 		t->maxout += 1;
1067 	else if (t->nout == t->maxout && t->next_cwnd-- == 0) {
1068 		t->maxout += 1;
1069 		t->next_cwnd = t->maxout;
1070 	}
1071 }
1072 
1073 static struct aoetgt *
1074 gettgt(struct aoedev *d, char *addr)
1075 {
1076 	struct aoetgt **t, **e;
1077 
1078 	t = d->targets;
1079 	e = t + d->ntargets;
1080 	for (; t < e && *t; t++)
1081 		if (memcmp((*t)->addr, addr, sizeof((*t)->addr)) == 0)
1082 			return *t;
1083 	return NULL;
1084 }
1085 
1086 static void
1087 bvcpy(struct sk_buff *skb, struct bio *bio, struct bvec_iter iter, long cnt)
1088 {
1089 	int soff = 0;
1090 	struct bio_vec bv;
1091 
1092 	iter.bi_size = cnt;
1093 
1094 	__bio_for_each_segment(bv, bio, iter, iter) {
1095 		char *p = page_address(bv.bv_page) + bv.bv_offset;
1096 		skb_copy_bits(skb, soff, p, bv.bv_len);
1097 		soff += bv.bv_len;
1098 	}
1099 }
1100 
1101 void
1102 aoe_end_request(struct aoedev *d, struct request *rq, int fastfail)
1103 {
1104 	struct bio *bio;
1105 	int bok;
1106 	struct request_queue *q;
1107 
1108 	q = d->blkq;
1109 	if (rq == d->ip.rq)
1110 		d->ip.rq = NULL;
1111 	do {
1112 		bio = rq->bio;
1113 		bok = !fastfail && test_bit(BIO_UPTODATE, &bio->bi_flags);
1114 	} while (__blk_end_request(rq, bok ? 0 : -EIO, bio->bi_iter.bi_size));
1115 
1116 	/* cf. http://lkml.org/lkml/2006/10/31/28 */
1117 	if (!fastfail)
1118 		__blk_run_queue(q);
1119 }
1120 
1121 static void
1122 aoe_end_buf(struct aoedev *d, struct buf *buf)
1123 {
1124 	struct request *rq;
1125 	unsigned long n;
1126 
1127 	if (buf == d->ip.buf)
1128 		d->ip.buf = NULL;
1129 	rq = buf->rq;
1130 	bio_pagedec(buf->bio);
1131 	mempool_free(buf, d->bufpool);
1132 	n = (unsigned long) rq->special;
1133 	rq->special = (void *) --n;
1134 	if (n == 0)
1135 		aoe_end_request(d, rq, 0);
1136 }
1137 
1138 static void
1139 ktiocomplete(struct frame *f)
1140 {
1141 	struct aoe_hdr *hin, *hout;
1142 	struct aoe_atahdr *ahin, *ahout;
1143 	struct buf *buf;
1144 	struct sk_buff *skb;
1145 	struct aoetgt *t;
1146 	struct aoeif *ifp;
1147 	struct aoedev *d;
1148 	long n;
1149 	int untainted;
1150 
1151 	if (f == NULL)
1152 		return;
1153 
1154 	t = f->t;
1155 	d = t->d;
1156 	skb = f->r_skb;
1157 	buf = f->buf;
1158 	if (f->flags & FFL_PROBE)
1159 		goto out;
1160 	if (!skb)		/* just fail the buf. */
1161 		goto noskb;
1162 
1163 	hout = (struct aoe_hdr *) skb_mac_header(f->skb);
1164 	ahout = (struct aoe_atahdr *) (hout+1);
1165 
1166 	hin = (struct aoe_hdr *) skb->data;
1167 	skb_pull(skb, sizeof(*hin));
1168 	ahin = (struct aoe_atahdr *) skb->data;
1169 	skb_pull(skb, sizeof(*ahin));
1170 	if (ahin->cmdstat & 0xa9) {	/* these bits cleared on success */
1171 		pr_err("aoe: ata error cmd=%2.2Xh stat=%2.2Xh from e%ld.%d\n",
1172 			ahout->cmdstat, ahin->cmdstat,
1173 			d->aoemajor, d->aoeminor);
1174 noskb:		if (buf)
1175 			clear_bit(BIO_UPTODATE, &buf->bio->bi_flags);
1176 		goto out;
1177 	}
1178 
1179 	n = ahout->scnt << 9;
1180 	switch (ahout->cmdstat) {
1181 	case ATA_CMD_PIO_READ:
1182 	case ATA_CMD_PIO_READ_EXT:
1183 		if (skb->len < n) {
1184 			pr_err("%s e%ld.%d.  skb->len=%d need=%ld\n",
1185 				"aoe: runt data size in read from",
1186 				(long) d->aoemajor, d->aoeminor,
1187 			       skb->len, n);
1188 			clear_bit(BIO_UPTODATE, &buf->bio->bi_flags);
1189 			break;
1190 		}
1191 		if (n > f->iter.bi_size) {
1192 			pr_err_ratelimited("%s e%ld.%d.  bytes=%ld need=%u\n",
1193 				"aoe: too-large data size in read from",
1194 				(long) d->aoemajor, d->aoeminor,
1195 				n, f->iter.bi_size);
1196 			clear_bit(BIO_UPTODATE, &buf->bio->bi_flags);
1197 			break;
1198 		}
1199 		bvcpy(skb, f->buf->bio, f->iter, n);
1200 	case ATA_CMD_PIO_WRITE:
1201 	case ATA_CMD_PIO_WRITE_EXT:
1202 		spin_lock_irq(&d->lock);
1203 		ifp = getif(t, skb->dev);
1204 		if (ifp)
1205 			ifp->lost = 0;
1206 		spin_unlock_irq(&d->lock);
1207 		break;
1208 	case ATA_CMD_ID_ATA:
1209 		if (skb->len < 512) {
1210 			pr_info("%s e%ld.%d.  skb->len=%d need=512\n",
1211 				"aoe: runt data size in ataid from",
1212 				(long) d->aoemajor, d->aoeminor,
1213 				skb->len);
1214 			break;
1215 		}
1216 		if (skb_linearize(skb))
1217 			break;
1218 		spin_lock_irq(&d->lock);
1219 		ataid_complete(d, t, skb->data);
1220 		spin_unlock_irq(&d->lock);
1221 		break;
1222 	default:
1223 		pr_info("aoe: unrecognized ata command %2.2Xh for %d.%d\n",
1224 			ahout->cmdstat,
1225 			be16_to_cpu(get_unaligned(&hin->major)),
1226 			hin->minor);
1227 	}
1228 out:
1229 	spin_lock_irq(&d->lock);
1230 	if (t->taint > 0
1231 	&& --t->taint > 0
1232 	&& t->nout_probes == 0) {
1233 		count_targets(d, &untainted);
1234 		if (untainted > 0) {
1235 			probe(t);
1236 			t->nout_probes++;
1237 		}
1238 	}
1239 
1240 	aoe_freetframe(f);
1241 
1242 	if (buf && --buf->nframesout == 0 && buf->iter.bi_size == 0)
1243 		aoe_end_buf(d, buf);
1244 
1245 	spin_unlock_irq(&d->lock);
1246 	aoedev_put(d);
1247 	dev_kfree_skb(skb);
1248 }
1249 
1250 /* Enters with iocq.lock held.
1251  * Returns true iff responses needing processing remain.
1252  */
1253 static int
1254 ktio(int id)
1255 {
1256 	struct frame *f;
1257 	struct list_head *pos;
1258 	int i;
1259 	int actual_id;
1260 
1261 	for (i = 0; ; ++i) {
1262 		if (i == MAXIOC)
1263 			return 1;
1264 		if (list_empty(&iocq[id].head))
1265 			return 0;
1266 		pos = iocq[id].head.next;
1267 		list_del(pos);
1268 		f = list_entry(pos, struct frame, head);
1269 		spin_unlock_irq(&iocq[id].lock);
1270 		ktiocomplete(f);
1271 
1272 		/* Figure out if extra threads are required. */
1273 		actual_id = f->t->d->aoeminor % ncpus;
1274 
1275 		if (!kts[actual_id].active) {
1276 			BUG_ON(id != 0);
1277 			mutex_lock(&ktio_spawn_lock);
1278 			if (!kts[actual_id].active
1279 				&& aoe_ktstart(&kts[actual_id]) == 0)
1280 				kts[actual_id].active = 1;
1281 			mutex_unlock(&ktio_spawn_lock);
1282 		}
1283 		spin_lock_irq(&iocq[id].lock);
1284 	}
1285 }
1286 
1287 static int
1288 kthread(void *vp)
1289 {
1290 	struct ktstate *k;
1291 	DECLARE_WAITQUEUE(wait, current);
1292 	int more;
1293 
1294 	k = vp;
1295 	current->flags |= PF_NOFREEZE;
1296 	set_user_nice(current, -10);
1297 	complete(&k->rendez);	/* tell spawner we're running */
1298 	do {
1299 		spin_lock_irq(k->lock);
1300 		more = k->fn(k->id);
1301 		if (!more) {
1302 			add_wait_queue(k->waitq, &wait);
1303 			__set_current_state(TASK_INTERRUPTIBLE);
1304 		}
1305 		spin_unlock_irq(k->lock);
1306 		if (!more) {
1307 			schedule();
1308 			remove_wait_queue(k->waitq, &wait);
1309 		} else
1310 			cond_resched();
1311 	} while (!kthread_should_stop());
1312 	complete(&k->rendez);	/* tell spawner we're stopping */
1313 	return 0;
1314 }
1315 
1316 void
1317 aoe_ktstop(struct ktstate *k)
1318 {
1319 	kthread_stop(k->task);
1320 	wait_for_completion(&k->rendez);
1321 }
1322 
1323 int
1324 aoe_ktstart(struct ktstate *k)
1325 {
1326 	struct task_struct *task;
1327 
1328 	init_completion(&k->rendez);
1329 	task = kthread_run(kthread, k, "%s", k->name);
1330 	if (task == NULL || IS_ERR(task))
1331 		return -ENOMEM;
1332 	k->task = task;
1333 	wait_for_completion(&k->rendez); /* allow kthread to start */
1334 	init_completion(&k->rendez);	/* for waiting for exit later */
1335 	return 0;
1336 }
1337 
1338 /* pass it off to kthreads for processing */
1339 static void
1340 ktcomplete(struct frame *f, struct sk_buff *skb)
1341 {
1342 	int id;
1343 	ulong flags;
1344 
1345 	f->r_skb = skb;
1346 	id = f->t->d->aoeminor % ncpus;
1347 	spin_lock_irqsave(&iocq[id].lock, flags);
1348 	if (!kts[id].active) {
1349 		spin_unlock_irqrestore(&iocq[id].lock, flags);
1350 		/* The thread with id has not been spawned yet,
1351 		 * so delegate the work to the main thread and
1352 		 * try spawning a new thread.
1353 		 */
1354 		id = 0;
1355 		spin_lock_irqsave(&iocq[id].lock, flags);
1356 	}
1357 	list_add_tail(&f->head, &iocq[id].head);
1358 	spin_unlock_irqrestore(&iocq[id].lock, flags);
1359 	wake_up(&ktiowq[id]);
1360 }
1361 
1362 struct sk_buff *
1363 aoecmd_ata_rsp(struct sk_buff *skb)
1364 {
1365 	struct aoedev *d;
1366 	struct aoe_hdr *h;
1367 	struct frame *f;
1368 	u32 n;
1369 	ulong flags;
1370 	char ebuf[128];
1371 	u16 aoemajor;
1372 
1373 	h = (struct aoe_hdr *) skb->data;
1374 	aoemajor = be16_to_cpu(get_unaligned(&h->major));
1375 	d = aoedev_by_aoeaddr(aoemajor, h->minor, 0);
1376 	if (d == NULL) {
1377 		snprintf(ebuf, sizeof ebuf, "aoecmd_ata_rsp: ata response "
1378 			"for unknown device %d.%d\n",
1379 			aoemajor, h->minor);
1380 		aoechr_error(ebuf);
1381 		return skb;
1382 	}
1383 
1384 	spin_lock_irqsave(&d->lock, flags);
1385 
1386 	n = be32_to_cpu(get_unaligned(&h->tag));
1387 	f = getframe(d, n);
1388 	if (f) {
1389 		calc_rttavg(d, f->t, tsince_hr(f));
1390 		f->t->nout--;
1391 		if (f->flags & FFL_PROBE)
1392 			f->t->nout_probes--;
1393 	} else {
1394 		f = getframe_deferred(d, n);
1395 		if (f) {
1396 			calc_rttavg(d, NULL, tsince_hr(f));
1397 		} else {
1398 			calc_rttavg(d, NULL, tsince(n));
1399 			spin_unlock_irqrestore(&d->lock, flags);
1400 			aoedev_put(d);
1401 			snprintf(ebuf, sizeof(ebuf),
1402 				 "%15s e%d.%d    tag=%08x@%08lx s=%pm d=%pm\n",
1403 				 "unexpected rsp",
1404 				 get_unaligned_be16(&h->major),
1405 				 h->minor,
1406 				 get_unaligned_be32(&h->tag),
1407 				 jiffies,
1408 				 h->src,
1409 				 h->dst);
1410 			aoechr_error(ebuf);
1411 			return skb;
1412 		}
1413 	}
1414 	aoecmd_work(d);
1415 
1416 	spin_unlock_irqrestore(&d->lock, flags);
1417 
1418 	ktcomplete(f, skb);
1419 
1420 	/*
1421 	 * Note here that we do not perform an aoedev_put, as we are
1422 	 * leaving this reference for the ktio to release.
1423 	 */
1424 	return NULL;
1425 }
1426 
1427 void
1428 aoecmd_cfg(ushort aoemajor, unsigned char aoeminor)
1429 {
1430 	struct sk_buff_head queue;
1431 
1432 	__skb_queue_head_init(&queue);
1433 	aoecmd_cfg_pkts(aoemajor, aoeminor, &queue);
1434 	aoenet_xmit(&queue);
1435 }
1436 
1437 struct sk_buff *
1438 aoecmd_ata_id(struct aoedev *d)
1439 {
1440 	struct aoe_hdr *h;
1441 	struct aoe_atahdr *ah;
1442 	struct frame *f;
1443 	struct sk_buff *skb;
1444 	struct aoetgt *t;
1445 
1446 	f = newframe(d);
1447 	if (f == NULL)
1448 		return NULL;
1449 
1450 	t = *d->tgt;
1451 
1452 	/* initialize the headers & frame */
1453 	skb = f->skb;
1454 	h = (struct aoe_hdr *) skb_mac_header(skb);
1455 	ah = (struct aoe_atahdr *) (h+1);
1456 	skb_put(skb, sizeof *h + sizeof *ah);
1457 	memset(h, 0, skb->len);
1458 	f->tag = aoehdr_atainit(d, t, h);
1459 	fhash(f);
1460 	t->nout++;
1461 	f->waited = 0;
1462 	f->waited_total = 0;
1463 
1464 	/* set up ata header */
1465 	ah->scnt = 1;
1466 	ah->cmdstat = ATA_CMD_ID_ATA;
1467 	ah->lba3 = 0xa0;
1468 
1469 	skb->dev = t->ifp->nd;
1470 
1471 	d->rttavg = RTTAVG_INIT;
1472 	d->rttdev = RTTDEV_INIT;
1473 	d->timer.function = rexmit_timer;
1474 
1475 	skb = skb_clone(skb, GFP_ATOMIC);
1476 	if (skb) {
1477 		do_gettimeofday(&f->sent);
1478 		f->sent_jiffs = (u32) jiffies;
1479 	}
1480 
1481 	return skb;
1482 }
1483 
1484 static struct aoetgt **
1485 grow_targets(struct aoedev *d)
1486 {
1487 	ulong oldn, newn;
1488 	struct aoetgt **tt;
1489 
1490 	oldn = d->ntargets;
1491 	newn = oldn * 2;
1492 	tt = kcalloc(newn, sizeof(*d->targets), GFP_ATOMIC);
1493 	if (!tt)
1494 		return NULL;
1495 	memmove(tt, d->targets, sizeof(*d->targets) * oldn);
1496 	d->tgt = tt + (d->tgt - d->targets);
1497 	kfree(d->targets);
1498 	d->targets = tt;
1499 	d->ntargets = newn;
1500 
1501 	return &d->targets[oldn];
1502 }
1503 
1504 static struct aoetgt *
1505 addtgt(struct aoedev *d, char *addr, ulong nframes)
1506 {
1507 	struct aoetgt *t, **tt, **te;
1508 
1509 	tt = d->targets;
1510 	te = tt + d->ntargets;
1511 	for (; tt < te && *tt; tt++)
1512 		;
1513 
1514 	if (tt == te) {
1515 		tt = grow_targets(d);
1516 		if (!tt)
1517 			goto nomem;
1518 	}
1519 	t = kzalloc(sizeof(*t), GFP_ATOMIC);
1520 	if (!t)
1521 		goto nomem;
1522 	t->nframes = nframes;
1523 	t->d = d;
1524 	memcpy(t->addr, addr, sizeof t->addr);
1525 	t->ifp = t->ifs;
1526 	aoecmd_wreset(t);
1527 	t->maxout = t->nframes / 2;
1528 	INIT_LIST_HEAD(&t->ffree);
1529 	return *tt = t;
1530 
1531  nomem:
1532 	pr_info("aoe: cannot allocate memory to add target\n");
1533 	return NULL;
1534 }
1535 
1536 static void
1537 setdbcnt(struct aoedev *d)
1538 {
1539 	struct aoetgt **t, **e;
1540 	int bcnt = 0;
1541 
1542 	t = d->targets;
1543 	e = t + d->ntargets;
1544 	for (; t < e && *t; t++)
1545 		if (bcnt == 0 || bcnt > (*t)->minbcnt)
1546 			bcnt = (*t)->minbcnt;
1547 	if (bcnt != d->maxbcnt) {
1548 		d->maxbcnt = bcnt;
1549 		pr_info("aoe: e%ld.%d: setting %d byte data frames\n",
1550 			d->aoemajor, d->aoeminor, bcnt);
1551 	}
1552 }
1553 
1554 static void
1555 setifbcnt(struct aoetgt *t, struct net_device *nd, int bcnt)
1556 {
1557 	struct aoedev *d;
1558 	struct aoeif *p, *e;
1559 	int minbcnt;
1560 
1561 	d = t->d;
1562 	minbcnt = bcnt;
1563 	p = t->ifs;
1564 	e = p + NAOEIFS;
1565 	for (; p < e; p++) {
1566 		if (p->nd == NULL)
1567 			break;		/* end of the valid interfaces */
1568 		if (p->nd == nd) {
1569 			p->bcnt = bcnt;	/* we're updating */
1570 			nd = NULL;
1571 		} else if (minbcnt > p->bcnt)
1572 			minbcnt = p->bcnt; /* find the min interface */
1573 	}
1574 	if (nd) {
1575 		if (p == e) {
1576 			pr_err("aoe: device setifbcnt failure; too many interfaces.\n");
1577 			return;
1578 		}
1579 		dev_hold(nd);
1580 		p->nd = nd;
1581 		p->bcnt = bcnt;
1582 	}
1583 	t->minbcnt = minbcnt;
1584 	setdbcnt(d);
1585 }
1586 
1587 void
1588 aoecmd_cfg_rsp(struct sk_buff *skb)
1589 {
1590 	struct aoedev *d;
1591 	struct aoe_hdr *h;
1592 	struct aoe_cfghdr *ch;
1593 	struct aoetgt *t;
1594 	ulong flags, aoemajor;
1595 	struct sk_buff *sl;
1596 	struct sk_buff_head queue;
1597 	u16 n;
1598 
1599 	sl = NULL;
1600 	h = (struct aoe_hdr *) skb_mac_header(skb);
1601 	ch = (struct aoe_cfghdr *) (h+1);
1602 
1603 	/*
1604 	 * Enough people have their dip switches set backwards to
1605 	 * warrant a loud message for this special case.
1606 	 */
1607 	aoemajor = get_unaligned_be16(&h->major);
1608 	if (aoemajor == 0xfff) {
1609 		printk(KERN_ERR "aoe: Warning: shelf address is all ones.  "
1610 			"Check shelf dip switches.\n");
1611 		return;
1612 	}
1613 	if (aoemajor == 0xffff) {
1614 		pr_info("aoe: e%ld.%d: broadcast shelf number invalid\n",
1615 			aoemajor, (int) h->minor);
1616 		return;
1617 	}
1618 	if (h->minor == 0xff) {
1619 		pr_info("aoe: e%ld.%d: broadcast slot number invalid\n",
1620 			aoemajor, (int) h->minor);
1621 		return;
1622 	}
1623 
1624 	n = be16_to_cpu(ch->bufcnt);
1625 	if (n > aoe_maxout)	/* keep it reasonable */
1626 		n = aoe_maxout;
1627 
1628 	d = aoedev_by_aoeaddr(aoemajor, h->minor, 1);
1629 	if (d == NULL) {
1630 		pr_info("aoe: device allocation failure\n");
1631 		return;
1632 	}
1633 
1634 	spin_lock_irqsave(&d->lock, flags);
1635 
1636 	t = gettgt(d, h->src);
1637 	if (t) {
1638 		t->nframes = n;
1639 		if (n < t->maxout)
1640 			aoecmd_wreset(t);
1641 	} else {
1642 		t = addtgt(d, h->src, n);
1643 		if (!t)
1644 			goto bail;
1645 	}
1646 	n = skb->dev->mtu;
1647 	n -= sizeof(struct aoe_hdr) + sizeof(struct aoe_atahdr);
1648 	n /= 512;
1649 	if (n > ch->scnt)
1650 		n = ch->scnt;
1651 	n = n ? n * 512 : DEFAULTBCNT;
1652 	setifbcnt(t, skb->dev, n);
1653 
1654 	/* don't change users' perspective */
1655 	if (d->nopen == 0) {
1656 		d->fw_ver = be16_to_cpu(ch->fwver);
1657 		sl = aoecmd_ata_id(d);
1658 	}
1659 bail:
1660 	spin_unlock_irqrestore(&d->lock, flags);
1661 	aoedev_put(d);
1662 	if (sl) {
1663 		__skb_queue_head_init(&queue);
1664 		__skb_queue_tail(&queue, sl);
1665 		aoenet_xmit(&queue);
1666 	}
1667 }
1668 
1669 void
1670 aoecmd_wreset(struct aoetgt *t)
1671 {
1672 	t->maxout = 1;
1673 	t->ssthresh = t->nframes / 2;
1674 	t->next_cwnd = t->nframes;
1675 }
1676 
1677 void
1678 aoecmd_cleanslate(struct aoedev *d)
1679 {
1680 	struct aoetgt **t, **te;
1681 
1682 	d->rttavg = RTTAVG_INIT;
1683 	d->rttdev = RTTDEV_INIT;
1684 	d->maxbcnt = 0;
1685 
1686 	t = d->targets;
1687 	te = t + d->ntargets;
1688 	for (; t < te && *t; t++)
1689 		aoecmd_wreset(*t);
1690 }
1691 
1692 void
1693 aoe_failbuf(struct aoedev *d, struct buf *buf)
1694 {
1695 	if (buf == NULL)
1696 		return;
1697 	buf->iter.bi_size = 0;
1698 	clear_bit(BIO_UPTODATE, &buf->bio->bi_flags);
1699 	if (buf->nframesout == 0)
1700 		aoe_end_buf(d, buf);
1701 }
1702 
1703 void
1704 aoe_flush_iocq(void)
1705 {
1706 	int i;
1707 
1708 	for (i = 0; i < ncpus; i++) {
1709 		if (kts[i].active)
1710 			aoe_flush_iocq_by_index(i);
1711 	}
1712 }
1713 
1714 void
1715 aoe_flush_iocq_by_index(int id)
1716 {
1717 	struct frame *f;
1718 	struct aoedev *d;
1719 	LIST_HEAD(flist);
1720 	struct list_head *pos;
1721 	struct sk_buff *skb;
1722 	ulong flags;
1723 
1724 	spin_lock_irqsave(&iocq[id].lock, flags);
1725 	list_splice_init(&iocq[id].head, &flist);
1726 	spin_unlock_irqrestore(&iocq[id].lock, flags);
1727 	while (!list_empty(&flist)) {
1728 		pos = flist.next;
1729 		list_del(pos);
1730 		f = list_entry(pos, struct frame, head);
1731 		d = f->t->d;
1732 		skb = f->r_skb;
1733 		spin_lock_irqsave(&d->lock, flags);
1734 		if (f->buf) {
1735 			f->buf->nframesout--;
1736 			aoe_failbuf(d, f->buf);
1737 		}
1738 		aoe_freetframe(f);
1739 		spin_unlock_irqrestore(&d->lock, flags);
1740 		dev_kfree_skb(skb);
1741 		aoedev_put(d);
1742 	}
1743 }
1744 
1745 int __init
1746 aoecmd_init(void)
1747 {
1748 	void *p;
1749 	int i;
1750 	int ret;
1751 
1752 	/* get_zeroed_page returns page with ref count 1 */
1753 	p = (void *) get_zeroed_page(GFP_KERNEL | __GFP_REPEAT);
1754 	if (!p)
1755 		return -ENOMEM;
1756 	empty_page = virt_to_page(p);
1757 
1758 	ncpus = num_online_cpus();
1759 
1760 	iocq = kcalloc(ncpus, sizeof(struct iocq_ktio), GFP_KERNEL);
1761 	if (!iocq)
1762 		return -ENOMEM;
1763 
1764 	kts = kcalloc(ncpus, sizeof(struct ktstate), GFP_KERNEL);
1765 	if (!kts) {
1766 		ret = -ENOMEM;
1767 		goto kts_fail;
1768 	}
1769 
1770 	ktiowq = kcalloc(ncpus, sizeof(wait_queue_head_t), GFP_KERNEL);
1771 	if (!ktiowq) {
1772 		ret = -ENOMEM;
1773 		goto ktiowq_fail;
1774 	}
1775 
1776 	mutex_init(&ktio_spawn_lock);
1777 
1778 	for (i = 0; i < ncpus; i++) {
1779 		INIT_LIST_HEAD(&iocq[i].head);
1780 		spin_lock_init(&iocq[i].lock);
1781 		init_waitqueue_head(&ktiowq[i]);
1782 		snprintf(kts[i].name, sizeof(kts[i].name), "aoe_ktio%d", i);
1783 		kts[i].fn = ktio;
1784 		kts[i].waitq = &ktiowq[i];
1785 		kts[i].lock = &iocq[i].lock;
1786 		kts[i].id = i;
1787 		kts[i].active = 0;
1788 	}
1789 	kts[0].active = 1;
1790 	if (aoe_ktstart(&kts[0])) {
1791 		ret = -ENOMEM;
1792 		goto ktstart_fail;
1793 	}
1794 	return 0;
1795 
1796 ktstart_fail:
1797 	kfree(ktiowq);
1798 ktiowq_fail:
1799 	kfree(kts);
1800 kts_fail:
1801 	kfree(iocq);
1802 
1803 	return ret;
1804 }
1805 
1806 void
1807 aoecmd_exit(void)
1808 {
1809 	int i;
1810 
1811 	for (i = 0; i < ncpus; i++)
1812 		if (kts[i].active)
1813 			aoe_ktstop(&kts[i]);
1814 
1815 	aoe_flush_iocq();
1816 
1817 	/* Free up the iocq and thread speicific configuration
1818 	* allocated during startup.
1819 	*/
1820 	kfree(iocq);
1821 	kfree(kts);
1822 	kfree(ktiowq);
1823 
1824 	free_page((unsigned long) page_address(empty_page));
1825 	empty_page = NULL;
1826 }
1827