xref: /openbmc/linux/drivers/block/aoe/aoecmd.c (revision ae213c44)
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/blk-mq.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 		f->sent = ktime_get();
402 		__skb_queue_head_init(&queue);
403 		__skb_queue_tail(&queue, skb);
404 		aoenet_xmit(&queue);
405 	}
406 	return 1;
407 }
408 
409 /* some callers cannot sleep, and they can call this function,
410  * transmitting the packets later, when interrupts are on
411  */
412 static void
413 aoecmd_cfg_pkts(ushort aoemajor, unsigned char aoeminor, struct sk_buff_head *queue)
414 {
415 	struct aoe_hdr *h;
416 	struct aoe_cfghdr *ch;
417 	struct sk_buff *skb;
418 	struct net_device *ifp;
419 
420 	rcu_read_lock();
421 	for_each_netdev_rcu(&init_net, ifp) {
422 		dev_hold(ifp);
423 		if (!is_aoe_netif(ifp))
424 			goto cont;
425 
426 		skb = new_skb(sizeof *h + sizeof *ch);
427 		if (skb == NULL) {
428 			printk(KERN_INFO "aoe: skb alloc failure\n");
429 			goto cont;
430 		}
431 		skb_put(skb, sizeof *h + sizeof *ch);
432 		skb->dev = ifp;
433 		__skb_queue_tail(queue, skb);
434 		h = (struct aoe_hdr *) skb_mac_header(skb);
435 		memset(h, 0, sizeof *h + sizeof *ch);
436 
437 		memset(h->dst, 0xff, sizeof h->dst);
438 		memcpy(h->src, ifp->dev_addr, sizeof h->src);
439 		h->type = __constant_cpu_to_be16(ETH_P_AOE);
440 		h->verfl = AOE_HVER;
441 		h->major = cpu_to_be16(aoemajor);
442 		h->minor = aoeminor;
443 		h->cmd = AOECMD_CFG;
444 
445 cont:
446 		dev_put(ifp);
447 	}
448 	rcu_read_unlock();
449 }
450 
451 static void
452 resend(struct aoedev *d, struct frame *f)
453 {
454 	struct sk_buff *skb;
455 	struct sk_buff_head queue;
456 	struct aoe_hdr *h;
457 	struct aoetgt *t;
458 	char buf[128];
459 	u32 n;
460 
461 	t = f->t;
462 	n = newtag(d);
463 	skb = f->skb;
464 	if (ifrotate(t) == NULL) {
465 		/* probably can't happen, but set it up to fail anyway */
466 		pr_info("aoe: resend: no interfaces to rotate to.\n");
467 		ktcomplete(f, NULL);
468 		return;
469 	}
470 	h = (struct aoe_hdr *) skb_mac_header(skb);
471 
472 	if (!(f->flags & FFL_PROBE)) {
473 		snprintf(buf, sizeof(buf),
474 			"%15s e%ld.%d oldtag=%08x@%08lx newtag=%08x s=%pm d=%pm nout=%d\n",
475 			"retransmit", d->aoemajor, d->aoeminor,
476 			f->tag, jiffies, n,
477 			h->src, h->dst, t->nout);
478 		aoechr_error(buf);
479 	}
480 
481 	f->tag = n;
482 	fhash(f);
483 	h->tag = cpu_to_be32(n);
484 	memcpy(h->dst, t->addr, sizeof h->dst);
485 	memcpy(h->src, t->ifp->nd->dev_addr, sizeof h->src);
486 
487 	skb->dev = t->ifp->nd;
488 	skb = skb_clone(skb, GFP_ATOMIC);
489 	if (skb == NULL)
490 		return;
491 	f->sent = ktime_get();
492 	__skb_queue_head_init(&queue);
493 	__skb_queue_tail(&queue, skb);
494 	aoenet_xmit(&queue);
495 }
496 
497 static int
498 tsince_hr(struct frame *f)
499 {
500 	u64 delta = ktime_to_ns(ktime_sub(ktime_get(), f->sent));
501 
502 	/* delta is normally under 4.2 seconds, avoid 64-bit division */
503 	if (likely(delta <= UINT_MAX))
504 		return (u32)delta / NSEC_PER_USEC;
505 
506 	/* avoid overflow after 71 minutes */
507 	if (delta > ((u64)INT_MAX * NSEC_PER_USEC))
508 		return INT_MAX;
509 
510 	return div_u64(delta, NSEC_PER_USEC);
511 }
512 
513 static int
514 tsince(u32 tag)
515 {
516 	int n;
517 
518 	n = jiffies & 0xffff;
519 	n -= tag & 0xffff;
520 	if (n < 0)
521 		n += 1<<16;
522 	return jiffies_to_usecs(n + 1);
523 }
524 
525 static struct aoeif *
526 getif(struct aoetgt *t, struct net_device *nd)
527 {
528 	struct aoeif *p, *e;
529 
530 	p = t->ifs;
531 	e = p + NAOEIFS;
532 	for (; p < e; p++)
533 		if (p->nd == nd)
534 			return p;
535 	return NULL;
536 }
537 
538 static void
539 ejectif(struct aoetgt *t, struct aoeif *ifp)
540 {
541 	struct aoeif *e;
542 	struct net_device *nd;
543 	ulong n;
544 
545 	nd = ifp->nd;
546 	e = t->ifs + NAOEIFS - 1;
547 	n = (e - ifp) * sizeof *ifp;
548 	memmove(ifp, ifp+1, n);
549 	e->nd = NULL;
550 	dev_put(nd);
551 }
552 
553 static struct frame *
554 reassign_frame(struct frame *f)
555 {
556 	struct frame *nf;
557 	struct sk_buff *skb;
558 
559 	nf = newframe(f->t->d);
560 	if (!nf)
561 		return NULL;
562 	if (nf->t == f->t) {
563 		aoe_freetframe(nf);
564 		return NULL;
565 	}
566 
567 	skb = nf->skb;
568 	nf->skb = f->skb;
569 	nf->buf = f->buf;
570 	nf->iter = f->iter;
571 	nf->waited = 0;
572 	nf->waited_total = f->waited_total;
573 	nf->sent = f->sent;
574 	f->skb = skb;
575 
576 	return nf;
577 }
578 
579 static void
580 probe(struct aoetgt *t)
581 {
582 	struct aoedev *d;
583 	struct frame *f;
584 	struct sk_buff *skb;
585 	struct sk_buff_head queue;
586 	size_t n, m;
587 	int frag;
588 
589 	d = t->d;
590 	f = newtframe(d, t);
591 	if (!f) {
592 		pr_err("%s %pm for e%ld.%d: %s\n",
593 			"aoe: cannot probe remote address",
594 			t->addr,
595 			(long) d->aoemajor, d->aoeminor,
596 			"no frame available");
597 		return;
598 	}
599 	f->flags |= FFL_PROBE;
600 	ifrotate(t);
601 	f->iter.bi_size = t->d->maxbcnt ? t->d->maxbcnt : DEFAULTBCNT;
602 	ata_rw_frameinit(f);
603 	skb = f->skb;
604 	for (frag = 0, n = f->iter.bi_size; n > 0; ++frag, n -= m) {
605 		if (n < PAGE_SIZE)
606 			m = n;
607 		else
608 			m = PAGE_SIZE;
609 		skb_fill_page_desc(skb, frag, empty_page, 0, m);
610 	}
611 	skb->len += f->iter.bi_size;
612 	skb->data_len = f->iter.bi_size;
613 	skb->truesize += f->iter.bi_size;
614 
615 	skb = skb_clone(f->skb, GFP_ATOMIC);
616 	if (skb) {
617 		f->sent = ktime_get();
618 		__skb_queue_head_init(&queue);
619 		__skb_queue_tail(&queue, skb);
620 		aoenet_xmit(&queue);
621 	}
622 }
623 
624 static long
625 rto(struct aoedev *d)
626 {
627 	long t;
628 
629 	t = 2 * d->rttavg >> RTTSCALE;
630 	t += 8 * d->rttdev >> RTTDSCALE;
631 	if (t == 0)
632 		t = 1;
633 
634 	return t;
635 }
636 
637 static void
638 rexmit_deferred(struct aoedev *d)
639 {
640 	struct aoetgt *t;
641 	struct frame *f;
642 	struct frame *nf;
643 	struct list_head *pos, *nx, *head;
644 	int since;
645 	int untainted;
646 
647 	count_targets(d, &untainted);
648 
649 	head = &d->rexmitq;
650 	list_for_each_safe(pos, nx, head) {
651 		f = list_entry(pos, struct frame, head);
652 		t = f->t;
653 		if (t->taint) {
654 			if (!(f->flags & FFL_PROBE)) {
655 				nf = reassign_frame(f);
656 				if (nf) {
657 					if (t->nout_probes == 0
658 					&& untainted > 0) {
659 						probe(t);
660 						t->nout_probes++;
661 					}
662 					list_replace(&f->head, &nf->head);
663 					pos = &nf->head;
664 					aoe_freetframe(f);
665 					f = nf;
666 					t = f->t;
667 				}
668 			} else if (untainted < 1) {
669 				/* don't probe w/o other untainted aoetgts */
670 				goto stop_probe;
671 			} else if (tsince_hr(f) < t->taint * rto(d)) {
672 				/* reprobe slowly when taint is high */
673 				continue;
674 			}
675 		} else if (f->flags & FFL_PROBE) {
676 stop_probe:		/* don't probe untainted aoetgts */
677 			list_del(pos);
678 			aoe_freetframe(f);
679 			/* leaving d->kicked, because this is routine */
680 			f->t->d->flags |= DEVFL_KICKME;
681 			continue;
682 		}
683 		if (t->nout >= t->maxout)
684 			continue;
685 		list_del(pos);
686 		t->nout++;
687 		if (f->flags & FFL_PROBE)
688 			t->nout_probes++;
689 		since = tsince_hr(f);
690 		f->waited += since;
691 		f->waited_total += since;
692 		resend(d, f);
693 	}
694 }
695 
696 /* An aoetgt accumulates demerits quickly, and successful
697  * probing redeems the aoetgt slowly.
698  */
699 static void
700 scorn(struct aoetgt *t)
701 {
702 	int n;
703 
704 	n = t->taint++;
705 	t->taint += t->taint * 2;
706 	if (n > t->taint)
707 		t->taint = n;
708 	if (t->taint > MAX_TAINT)
709 		t->taint = MAX_TAINT;
710 }
711 
712 static int
713 count_targets(struct aoedev *d, int *untainted)
714 {
715 	int i, good;
716 
717 	for (i = good = 0; i < d->ntargets && d->targets[i]; ++i)
718 		if (d->targets[i]->taint == 0)
719 			good++;
720 
721 	if (untainted)
722 		*untainted = good;
723 	return i;
724 }
725 
726 static void
727 rexmit_timer(struct timer_list *timer)
728 {
729 	struct aoedev *d;
730 	struct aoetgt *t;
731 	struct aoeif *ifp;
732 	struct frame *f;
733 	struct list_head *head, *pos, *nx;
734 	LIST_HEAD(flist);
735 	register long timeout;
736 	ulong flags, n;
737 	int i;
738 	int utgts;	/* number of aoetgt descriptors (not slots) */
739 	int since;
740 
741 	d = from_timer(d, timer, timer);
742 
743 	spin_lock_irqsave(&d->lock, flags);
744 
745 	/* timeout based on observed timings and variations */
746 	timeout = rto(d);
747 
748 	utgts = count_targets(d, NULL);
749 
750 	if (d->flags & DEVFL_TKILL) {
751 		spin_unlock_irqrestore(&d->lock, flags);
752 		return;
753 	}
754 
755 	/* collect all frames to rexmit into flist */
756 	for (i = 0; i < NFACTIVE; i++) {
757 		head = &d->factive[i];
758 		list_for_each_safe(pos, nx, head) {
759 			f = list_entry(pos, struct frame, head);
760 			if (tsince_hr(f) < timeout)
761 				break;	/* end of expired frames */
762 			/* move to flist for later processing */
763 			list_move_tail(pos, &flist);
764 		}
765 	}
766 
767 	/* process expired frames */
768 	while (!list_empty(&flist)) {
769 		pos = flist.next;
770 		f = list_entry(pos, struct frame, head);
771 		since = tsince_hr(f);
772 		n = f->waited_total + since;
773 		n /= USEC_PER_SEC;
774 		if (aoe_deadsecs
775 		&& n > aoe_deadsecs
776 		&& !(f->flags & FFL_PROBE)) {
777 			/* Waited too long.  Device failure.
778 			 * Hang all frames on first hash bucket for downdev
779 			 * to clean up.
780 			 */
781 			list_splice(&flist, &d->factive[0]);
782 			aoedev_downdev(d);
783 			goto out;
784 		}
785 
786 		t = f->t;
787 		n = f->waited + since;
788 		n /= USEC_PER_SEC;
789 		if (aoe_deadsecs && utgts > 0
790 		&& (n > aoe_deadsecs / utgts || n > HARD_SCORN_SECS))
791 			scorn(t); /* avoid this target */
792 
793 		if (t->maxout != 1) {
794 			t->ssthresh = t->maxout / 2;
795 			t->maxout = 1;
796 		}
797 
798 		if (f->flags & FFL_PROBE) {
799 			t->nout_probes--;
800 		} else {
801 			ifp = getif(t, f->skb->dev);
802 			if (ifp && ++ifp->lost > (t->nframes << 1)
803 			&& (ifp != t->ifs || t->ifs[1].nd)) {
804 				ejectif(t, ifp);
805 				ifp = NULL;
806 			}
807 		}
808 		list_move_tail(pos, &d->rexmitq);
809 		t->nout--;
810 	}
811 	rexmit_deferred(d);
812 
813 out:
814 	if ((d->flags & DEVFL_KICKME) && d->blkq) {
815 		d->flags &= ~DEVFL_KICKME;
816 		blk_mq_run_hw_queues(d->blkq, true);
817 	}
818 
819 	d->timer.expires = jiffies + TIMERTICK;
820 	add_timer(&d->timer);
821 
822 	spin_unlock_irqrestore(&d->lock, flags);
823 }
824 
825 static void
826 bufinit(struct buf *buf, struct request *rq, struct bio *bio)
827 {
828 	memset(buf, 0, sizeof(*buf));
829 	buf->rq = rq;
830 	buf->bio = bio;
831 	buf->iter = bio->bi_iter;
832 }
833 
834 static struct buf *
835 nextbuf(struct aoedev *d)
836 {
837 	struct request *rq;
838 	struct request_queue *q;
839 	struct aoe_req *req;
840 	struct buf *buf;
841 	struct bio *bio;
842 
843 	q = d->blkq;
844 	if (q == NULL)
845 		return NULL;	/* initializing */
846 	if (d->ip.buf)
847 		return d->ip.buf;
848 	rq = d->ip.rq;
849 	if (rq == NULL) {
850 		rq = list_first_entry_or_null(&d->rq_list, struct request,
851 						queuelist);
852 		if (rq == NULL)
853 			return NULL;
854 		list_del_init(&rq->queuelist);
855 		blk_mq_start_request(rq);
856 		d->ip.rq = rq;
857 		d->ip.nxbio = rq->bio;
858 
859 		req = blk_mq_rq_to_pdu(rq);
860 		req->nr_bios = 0;
861 		__rq_for_each_bio(bio, rq)
862 			req->nr_bios++;
863 	}
864 	buf = mempool_alloc(d->bufpool, GFP_ATOMIC);
865 	if (buf == NULL) {
866 		pr_err("aoe: nextbuf: unable to mempool_alloc!\n");
867 		return NULL;
868 	}
869 	bio = d->ip.nxbio;
870 	bufinit(buf, rq, bio);
871 	bio = bio->bi_next;
872 	d->ip.nxbio = bio;
873 	if (bio == NULL)
874 		d->ip.rq = NULL;
875 	return d->ip.buf = buf;
876 }
877 
878 /* enters with d->lock held */
879 void
880 aoecmd_work(struct aoedev *d)
881 {
882 	rexmit_deferred(d);
883 	while (aoecmd_ata_rw(d))
884 		;
885 }
886 
887 /* this function performs work that has been deferred until sleeping is OK
888  */
889 void
890 aoecmd_sleepwork(struct work_struct *work)
891 {
892 	struct aoedev *d = container_of(work, struct aoedev, work);
893 	struct block_device *bd;
894 	u64 ssize;
895 
896 	if (d->flags & DEVFL_GDALLOC)
897 		aoeblk_gdalloc(d);
898 
899 	if (d->flags & DEVFL_NEWSIZE) {
900 		ssize = get_capacity(d->gd);
901 		bd = bdget_disk(d->gd, 0);
902 		if (bd) {
903 			inode_lock(bd->bd_inode);
904 			i_size_write(bd->bd_inode, (loff_t)ssize<<9);
905 			inode_unlock(bd->bd_inode);
906 			bdput(bd);
907 		}
908 		spin_lock_irq(&d->lock);
909 		d->flags |= DEVFL_UP;
910 		d->flags &= ~DEVFL_NEWSIZE;
911 		spin_unlock_irq(&d->lock);
912 	}
913 }
914 
915 static void
916 ata_ident_fixstring(u16 *id, int ns)
917 {
918 	u16 s;
919 
920 	while (ns-- > 0) {
921 		s = *id;
922 		*id++ = s >> 8 | s << 8;
923 	}
924 }
925 
926 static void
927 ataid_complete(struct aoedev *d, struct aoetgt *t, unsigned char *id)
928 {
929 	u64 ssize;
930 	u16 n;
931 
932 	/* word 83: command set supported */
933 	n = get_unaligned_le16(&id[83 << 1]);
934 
935 	/* word 86: command set/feature enabled */
936 	n |= get_unaligned_le16(&id[86 << 1]);
937 
938 	if (n & (1<<10)) {	/* bit 10: LBA 48 */
939 		d->flags |= DEVFL_EXT;
940 
941 		/* word 100: number lba48 sectors */
942 		ssize = get_unaligned_le64(&id[100 << 1]);
943 
944 		/* set as in ide-disk.c:init_idedisk_capacity */
945 		d->geo.cylinders = ssize;
946 		d->geo.cylinders /= (255 * 63);
947 		d->geo.heads = 255;
948 		d->geo.sectors = 63;
949 	} else {
950 		d->flags &= ~DEVFL_EXT;
951 
952 		/* number lba28 sectors */
953 		ssize = get_unaligned_le32(&id[60 << 1]);
954 
955 		/* NOTE: obsolete in ATA 6 */
956 		d->geo.cylinders = get_unaligned_le16(&id[54 << 1]);
957 		d->geo.heads = get_unaligned_le16(&id[55 << 1]);
958 		d->geo.sectors = get_unaligned_le16(&id[56 << 1]);
959 	}
960 
961 	ata_ident_fixstring((u16 *) &id[10<<1], 10);	/* serial */
962 	ata_ident_fixstring((u16 *) &id[23<<1], 4);	/* firmware */
963 	ata_ident_fixstring((u16 *) &id[27<<1], 20);	/* model */
964 	memcpy(d->ident, id, sizeof(d->ident));
965 
966 	if (d->ssize != ssize)
967 		printk(KERN_INFO
968 			"aoe: %pm e%ld.%d v%04x has %llu sectors\n",
969 			t->addr,
970 			d->aoemajor, d->aoeminor,
971 			d->fw_ver, (long long)ssize);
972 	d->ssize = ssize;
973 	d->geo.start = 0;
974 	if (d->flags & (DEVFL_GDALLOC|DEVFL_NEWSIZE))
975 		return;
976 	if (d->gd != NULL) {
977 		set_capacity(d->gd, ssize);
978 		d->flags |= DEVFL_NEWSIZE;
979 	} else
980 		d->flags |= DEVFL_GDALLOC;
981 	schedule_work(&d->work);
982 }
983 
984 static void
985 calc_rttavg(struct aoedev *d, struct aoetgt *t, int rtt)
986 {
987 	register long n;
988 
989 	n = rtt;
990 
991 	/* cf. Congestion Avoidance and Control, Jacobson & Karels, 1988 */
992 	n -= d->rttavg >> RTTSCALE;
993 	d->rttavg += n;
994 	if (n < 0)
995 		n = -n;
996 	n -= d->rttdev >> RTTDSCALE;
997 	d->rttdev += n;
998 
999 	if (!t || t->maxout >= t->nframes)
1000 		return;
1001 	if (t->maxout < t->ssthresh)
1002 		t->maxout += 1;
1003 	else if (t->nout == t->maxout && t->next_cwnd-- == 0) {
1004 		t->maxout += 1;
1005 		t->next_cwnd = t->maxout;
1006 	}
1007 }
1008 
1009 static struct aoetgt *
1010 gettgt(struct aoedev *d, char *addr)
1011 {
1012 	struct aoetgt **t, **e;
1013 
1014 	t = d->targets;
1015 	e = t + d->ntargets;
1016 	for (; t < e && *t; t++)
1017 		if (memcmp((*t)->addr, addr, sizeof((*t)->addr)) == 0)
1018 			return *t;
1019 	return NULL;
1020 }
1021 
1022 static void
1023 bvcpy(struct sk_buff *skb, struct bio *bio, struct bvec_iter iter, long cnt)
1024 {
1025 	int soff = 0;
1026 	struct bio_vec bv;
1027 
1028 	iter.bi_size = cnt;
1029 
1030 	__bio_for_each_segment(bv, bio, iter, iter) {
1031 		char *p = kmap_atomic(bv.bv_page) + bv.bv_offset;
1032 		skb_copy_bits(skb, soff, p, bv.bv_len);
1033 		kunmap_atomic(p);
1034 		soff += bv.bv_len;
1035 	}
1036 }
1037 
1038 void
1039 aoe_end_request(struct aoedev *d, struct request *rq, int fastfail)
1040 {
1041 	struct bio *bio;
1042 	int bok;
1043 	struct request_queue *q;
1044 	blk_status_t err = BLK_STS_OK;
1045 
1046 	q = d->blkq;
1047 	if (rq == d->ip.rq)
1048 		d->ip.rq = NULL;
1049 	do {
1050 		bio = rq->bio;
1051 		bok = !fastfail && !bio->bi_status;
1052 		if (!bok)
1053 			err = BLK_STS_IOERR;
1054 	} while (blk_update_request(rq, bok ? BLK_STS_OK : BLK_STS_IOERR, bio->bi_iter.bi_size));
1055 
1056 	__blk_mq_end_request(rq, err);
1057 
1058 	/* cf. http://lkml.org/lkml/2006/10/31/28 */
1059 	if (!fastfail)
1060 		blk_mq_run_hw_queues(q, true);
1061 }
1062 
1063 static void
1064 aoe_end_buf(struct aoedev *d, struct buf *buf)
1065 {
1066 	struct request *rq = buf->rq;
1067 	struct aoe_req *req = blk_mq_rq_to_pdu(rq);
1068 
1069 	if (buf == d->ip.buf)
1070 		d->ip.buf = NULL;
1071 	mempool_free(buf, d->bufpool);
1072 	if (--req->nr_bios == 0)
1073 		aoe_end_request(d, rq, 0);
1074 }
1075 
1076 static void
1077 ktiocomplete(struct frame *f)
1078 {
1079 	struct aoe_hdr *hin, *hout;
1080 	struct aoe_atahdr *ahin, *ahout;
1081 	struct buf *buf;
1082 	struct sk_buff *skb;
1083 	struct aoetgt *t;
1084 	struct aoeif *ifp;
1085 	struct aoedev *d;
1086 	long n;
1087 	int untainted;
1088 
1089 	if (f == NULL)
1090 		return;
1091 
1092 	t = f->t;
1093 	d = t->d;
1094 	skb = f->r_skb;
1095 	buf = f->buf;
1096 	if (f->flags & FFL_PROBE)
1097 		goto out;
1098 	if (!skb)		/* just fail the buf. */
1099 		goto noskb;
1100 
1101 	hout = (struct aoe_hdr *) skb_mac_header(f->skb);
1102 	ahout = (struct aoe_atahdr *) (hout+1);
1103 
1104 	hin = (struct aoe_hdr *) skb->data;
1105 	skb_pull(skb, sizeof(*hin));
1106 	ahin = (struct aoe_atahdr *) skb->data;
1107 	skb_pull(skb, sizeof(*ahin));
1108 	if (ahin->cmdstat & 0xa9) {	/* these bits cleared on success */
1109 		pr_err("aoe: ata error cmd=%2.2Xh stat=%2.2Xh from e%ld.%d\n",
1110 			ahout->cmdstat, ahin->cmdstat,
1111 			d->aoemajor, d->aoeminor);
1112 noskb:		if (buf)
1113 			buf->bio->bi_status = BLK_STS_IOERR;
1114 		goto out;
1115 	}
1116 
1117 	n = ahout->scnt << 9;
1118 	switch (ahout->cmdstat) {
1119 	case ATA_CMD_PIO_READ:
1120 	case ATA_CMD_PIO_READ_EXT:
1121 		if (skb->len < n) {
1122 			pr_err("%s e%ld.%d.  skb->len=%d need=%ld\n",
1123 				"aoe: runt data size in read from",
1124 				(long) d->aoemajor, d->aoeminor,
1125 			       skb->len, n);
1126 			buf->bio->bi_status = BLK_STS_IOERR;
1127 			break;
1128 		}
1129 		if (n > f->iter.bi_size) {
1130 			pr_err_ratelimited("%s e%ld.%d.  bytes=%ld need=%u\n",
1131 				"aoe: too-large data size in read from",
1132 				(long) d->aoemajor, d->aoeminor,
1133 				n, f->iter.bi_size);
1134 			buf->bio->bi_status = BLK_STS_IOERR;
1135 			break;
1136 		}
1137 		bvcpy(skb, f->buf->bio, f->iter, n);
1138 		/* fall through */
1139 	case ATA_CMD_PIO_WRITE:
1140 	case ATA_CMD_PIO_WRITE_EXT:
1141 		spin_lock_irq(&d->lock);
1142 		ifp = getif(t, skb->dev);
1143 		if (ifp)
1144 			ifp->lost = 0;
1145 		spin_unlock_irq(&d->lock);
1146 		break;
1147 	case ATA_CMD_ID_ATA:
1148 		if (skb->len < 512) {
1149 			pr_info("%s e%ld.%d.  skb->len=%d need=512\n",
1150 				"aoe: runt data size in ataid from",
1151 				(long) d->aoemajor, d->aoeminor,
1152 				skb->len);
1153 			break;
1154 		}
1155 		if (skb_linearize(skb))
1156 			break;
1157 		spin_lock_irq(&d->lock);
1158 		ataid_complete(d, t, skb->data);
1159 		spin_unlock_irq(&d->lock);
1160 		break;
1161 	default:
1162 		pr_info("aoe: unrecognized ata command %2.2Xh for %d.%d\n",
1163 			ahout->cmdstat,
1164 			be16_to_cpu(get_unaligned(&hin->major)),
1165 			hin->minor);
1166 	}
1167 out:
1168 	spin_lock_irq(&d->lock);
1169 	if (t->taint > 0
1170 	&& --t->taint > 0
1171 	&& t->nout_probes == 0) {
1172 		count_targets(d, &untainted);
1173 		if (untainted > 0) {
1174 			probe(t);
1175 			t->nout_probes++;
1176 		}
1177 	}
1178 
1179 	aoe_freetframe(f);
1180 
1181 	if (buf && --buf->nframesout == 0 && buf->iter.bi_size == 0)
1182 		aoe_end_buf(d, buf);
1183 
1184 	spin_unlock_irq(&d->lock);
1185 	aoedev_put(d);
1186 	dev_kfree_skb(skb);
1187 }
1188 
1189 /* Enters with iocq.lock held.
1190  * Returns true iff responses needing processing remain.
1191  */
1192 static int
1193 ktio(int id)
1194 {
1195 	struct frame *f;
1196 	struct list_head *pos;
1197 	int i;
1198 	int actual_id;
1199 
1200 	for (i = 0; ; ++i) {
1201 		if (i == MAXIOC)
1202 			return 1;
1203 		if (list_empty(&iocq[id].head))
1204 			return 0;
1205 		pos = iocq[id].head.next;
1206 		list_del(pos);
1207 		f = list_entry(pos, struct frame, head);
1208 		spin_unlock_irq(&iocq[id].lock);
1209 		ktiocomplete(f);
1210 
1211 		/* Figure out if extra threads are required. */
1212 		actual_id = f->t->d->aoeminor % ncpus;
1213 
1214 		if (!kts[actual_id].active) {
1215 			BUG_ON(id != 0);
1216 			mutex_lock(&ktio_spawn_lock);
1217 			if (!kts[actual_id].active
1218 				&& aoe_ktstart(&kts[actual_id]) == 0)
1219 				kts[actual_id].active = 1;
1220 			mutex_unlock(&ktio_spawn_lock);
1221 		}
1222 		spin_lock_irq(&iocq[id].lock);
1223 	}
1224 }
1225 
1226 static int
1227 kthread(void *vp)
1228 {
1229 	struct ktstate *k;
1230 	DECLARE_WAITQUEUE(wait, current);
1231 	int more;
1232 
1233 	k = vp;
1234 	current->flags |= PF_NOFREEZE;
1235 	set_user_nice(current, -10);
1236 	complete(&k->rendez);	/* tell spawner we're running */
1237 	do {
1238 		spin_lock_irq(k->lock);
1239 		more = k->fn(k->id);
1240 		if (!more) {
1241 			add_wait_queue(k->waitq, &wait);
1242 			__set_current_state(TASK_INTERRUPTIBLE);
1243 		}
1244 		spin_unlock_irq(k->lock);
1245 		if (!more) {
1246 			schedule();
1247 			remove_wait_queue(k->waitq, &wait);
1248 		} else
1249 			cond_resched();
1250 	} while (!kthread_should_stop());
1251 	complete(&k->rendez);	/* tell spawner we're stopping */
1252 	return 0;
1253 }
1254 
1255 void
1256 aoe_ktstop(struct ktstate *k)
1257 {
1258 	kthread_stop(k->task);
1259 	wait_for_completion(&k->rendez);
1260 }
1261 
1262 int
1263 aoe_ktstart(struct ktstate *k)
1264 {
1265 	struct task_struct *task;
1266 
1267 	init_completion(&k->rendez);
1268 	task = kthread_run(kthread, k, "%s", k->name);
1269 	if (task == NULL || IS_ERR(task))
1270 		return -ENOMEM;
1271 	k->task = task;
1272 	wait_for_completion(&k->rendez); /* allow kthread to start */
1273 	init_completion(&k->rendez);	/* for waiting for exit later */
1274 	return 0;
1275 }
1276 
1277 /* pass it off to kthreads for processing */
1278 static void
1279 ktcomplete(struct frame *f, struct sk_buff *skb)
1280 {
1281 	int id;
1282 	ulong flags;
1283 
1284 	f->r_skb = skb;
1285 	id = f->t->d->aoeminor % ncpus;
1286 	spin_lock_irqsave(&iocq[id].lock, flags);
1287 	if (!kts[id].active) {
1288 		spin_unlock_irqrestore(&iocq[id].lock, flags);
1289 		/* The thread with id has not been spawned yet,
1290 		 * so delegate the work to the main thread and
1291 		 * try spawning a new thread.
1292 		 */
1293 		id = 0;
1294 		spin_lock_irqsave(&iocq[id].lock, flags);
1295 	}
1296 	list_add_tail(&f->head, &iocq[id].head);
1297 	spin_unlock_irqrestore(&iocq[id].lock, flags);
1298 	wake_up(&ktiowq[id]);
1299 }
1300 
1301 struct sk_buff *
1302 aoecmd_ata_rsp(struct sk_buff *skb)
1303 {
1304 	struct aoedev *d;
1305 	struct aoe_hdr *h;
1306 	struct frame *f;
1307 	u32 n;
1308 	ulong flags;
1309 	char ebuf[128];
1310 	u16 aoemajor;
1311 
1312 	h = (struct aoe_hdr *) skb->data;
1313 	aoemajor = be16_to_cpu(get_unaligned(&h->major));
1314 	d = aoedev_by_aoeaddr(aoemajor, h->minor, 0);
1315 	if (d == NULL) {
1316 		snprintf(ebuf, sizeof ebuf, "aoecmd_ata_rsp: ata response "
1317 			"for unknown device %d.%d\n",
1318 			aoemajor, h->minor);
1319 		aoechr_error(ebuf);
1320 		return skb;
1321 	}
1322 
1323 	spin_lock_irqsave(&d->lock, flags);
1324 
1325 	n = be32_to_cpu(get_unaligned(&h->tag));
1326 	f = getframe(d, n);
1327 	if (f) {
1328 		calc_rttavg(d, f->t, tsince_hr(f));
1329 		f->t->nout--;
1330 		if (f->flags & FFL_PROBE)
1331 			f->t->nout_probes--;
1332 	} else {
1333 		f = getframe_deferred(d, n);
1334 		if (f) {
1335 			calc_rttavg(d, NULL, tsince_hr(f));
1336 		} else {
1337 			calc_rttavg(d, NULL, tsince(n));
1338 			spin_unlock_irqrestore(&d->lock, flags);
1339 			aoedev_put(d);
1340 			snprintf(ebuf, sizeof(ebuf),
1341 				 "%15s e%d.%d    tag=%08x@%08lx s=%pm d=%pm\n",
1342 				 "unexpected rsp",
1343 				 get_unaligned_be16(&h->major),
1344 				 h->minor,
1345 				 get_unaligned_be32(&h->tag),
1346 				 jiffies,
1347 				 h->src,
1348 				 h->dst);
1349 			aoechr_error(ebuf);
1350 			return skb;
1351 		}
1352 	}
1353 	aoecmd_work(d);
1354 
1355 	spin_unlock_irqrestore(&d->lock, flags);
1356 
1357 	ktcomplete(f, skb);
1358 
1359 	/*
1360 	 * Note here that we do not perform an aoedev_put, as we are
1361 	 * leaving this reference for the ktio to release.
1362 	 */
1363 	return NULL;
1364 }
1365 
1366 void
1367 aoecmd_cfg(ushort aoemajor, unsigned char aoeminor)
1368 {
1369 	struct sk_buff_head queue;
1370 
1371 	__skb_queue_head_init(&queue);
1372 	aoecmd_cfg_pkts(aoemajor, aoeminor, &queue);
1373 	aoenet_xmit(&queue);
1374 }
1375 
1376 struct sk_buff *
1377 aoecmd_ata_id(struct aoedev *d)
1378 {
1379 	struct aoe_hdr *h;
1380 	struct aoe_atahdr *ah;
1381 	struct frame *f;
1382 	struct sk_buff *skb;
1383 	struct aoetgt *t;
1384 
1385 	f = newframe(d);
1386 	if (f == NULL)
1387 		return NULL;
1388 
1389 	t = *d->tgt;
1390 
1391 	/* initialize the headers & frame */
1392 	skb = f->skb;
1393 	h = (struct aoe_hdr *) skb_mac_header(skb);
1394 	ah = (struct aoe_atahdr *) (h+1);
1395 	skb_put(skb, sizeof *h + sizeof *ah);
1396 	memset(h, 0, skb->len);
1397 	f->tag = aoehdr_atainit(d, t, h);
1398 	fhash(f);
1399 	t->nout++;
1400 	f->waited = 0;
1401 	f->waited_total = 0;
1402 
1403 	/* set up ata header */
1404 	ah->scnt = 1;
1405 	ah->cmdstat = ATA_CMD_ID_ATA;
1406 	ah->lba3 = 0xa0;
1407 
1408 	skb->dev = t->ifp->nd;
1409 
1410 	d->rttavg = RTTAVG_INIT;
1411 	d->rttdev = RTTDEV_INIT;
1412 	d->timer.function = rexmit_timer;
1413 
1414 	skb = skb_clone(skb, GFP_ATOMIC);
1415 	if (skb)
1416 		f->sent = ktime_get();
1417 
1418 	return skb;
1419 }
1420 
1421 static struct aoetgt **
1422 grow_targets(struct aoedev *d)
1423 {
1424 	ulong oldn, newn;
1425 	struct aoetgt **tt;
1426 
1427 	oldn = d->ntargets;
1428 	newn = oldn * 2;
1429 	tt = kcalloc(newn, sizeof(*d->targets), GFP_ATOMIC);
1430 	if (!tt)
1431 		return NULL;
1432 	memmove(tt, d->targets, sizeof(*d->targets) * oldn);
1433 	d->tgt = tt + (d->tgt - d->targets);
1434 	kfree(d->targets);
1435 	d->targets = tt;
1436 	d->ntargets = newn;
1437 
1438 	return &d->targets[oldn];
1439 }
1440 
1441 static struct aoetgt *
1442 addtgt(struct aoedev *d, char *addr, ulong nframes)
1443 {
1444 	struct aoetgt *t, **tt, **te;
1445 
1446 	tt = d->targets;
1447 	te = tt + d->ntargets;
1448 	for (; tt < te && *tt; tt++)
1449 		;
1450 
1451 	if (tt == te) {
1452 		tt = grow_targets(d);
1453 		if (!tt)
1454 			goto nomem;
1455 	}
1456 	t = kzalloc(sizeof(*t), GFP_ATOMIC);
1457 	if (!t)
1458 		goto nomem;
1459 	t->nframes = nframes;
1460 	t->d = d;
1461 	memcpy(t->addr, addr, sizeof t->addr);
1462 	t->ifp = t->ifs;
1463 	aoecmd_wreset(t);
1464 	t->maxout = t->nframes / 2;
1465 	INIT_LIST_HEAD(&t->ffree);
1466 	return *tt = t;
1467 
1468  nomem:
1469 	pr_info("aoe: cannot allocate memory to add target\n");
1470 	return NULL;
1471 }
1472 
1473 static void
1474 setdbcnt(struct aoedev *d)
1475 {
1476 	struct aoetgt **t, **e;
1477 	int bcnt = 0;
1478 
1479 	t = d->targets;
1480 	e = t + d->ntargets;
1481 	for (; t < e && *t; t++)
1482 		if (bcnt == 0 || bcnt > (*t)->minbcnt)
1483 			bcnt = (*t)->minbcnt;
1484 	if (bcnt != d->maxbcnt) {
1485 		d->maxbcnt = bcnt;
1486 		pr_info("aoe: e%ld.%d: setting %d byte data frames\n",
1487 			d->aoemajor, d->aoeminor, bcnt);
1488 	}
1489 }
1490 
1491 static void
1492 setifbcnt(struct aoetgt *t, struct net_device *nd, int bcnt)
1493 {
1494 	struct aoedev *d;
1495 	struct aoeif *p, *e;
1496 	int minbcnt;
1497 
1498 	d = t->d;
1499 	minbcnt = bcnt;
1500 	p = t->ifs;
1501 	e = p + NAOEIFS;
1502 	for (; p < e; p++) {
1503 		if (p->nd == NULL)
1504 			break;		/* end of the valid interfaces */
1505 		if (p->nd == nd) {
1506 			p->bcnt = bcnt;	/* we're updating */
1507 			nd = NULL;
1508 		} else if (minbcnt > p->bcnt)
1509 			minbcnt = p->bcnt; /* find the min interface */
1510 	}
1511 	if (nd) {
1512 		if (p == e) {
1513 			pr_err("aoe: device setifbcnt failure; too many interfaces.\n");
1514 			return;
1515 		}
1516 		dev_hold(nd);
1517 		p->nd = nd;
1518 		p->bcnt = bcnt;
1519 	}
1520 	t->minbcnt = minbcnt;
1521 	setdbcnt(d);
1522 }
1523 
1524 void
1525 aoecmd_cfg_rsp(struct sk_buff *skb)
1526 {
1527 	struct aoedev *d;
1528 	struct aoe_hdr *h;
1529 	struct aoe_cfghdr *ch;
1530 	struct aoetgt *t;
1531 	ulong flags, aoemajor;
1532 	struct sk_buff *sl;
1533 	struct sk_buff_head queue;
1534 	u16 n;
1535 
1536 	sl = NULL;
1537 	h = (struct aoe_hdr *) skb_mac_header(skb);
1538 	ch = (struct aoe_cfghdr *) (h+1);
1539 
1540 	/*
1541 	 * Enough people have their dip switches set backwards to
1542 	 * warrant a loud message for this special case.
1543 	 */
1544 	aoemajor = get_unaligned_be16(&h->major);
1545 	if (aoemajor == 0xfff) {
1546 		printk(KERN_ERR "aoe: Warning: shelf address is all ones.  "
1547 			"Check shelf dip switches.\n");
1548 		return;
1549 	}
1550 	if (aoemajor == 0xffff) {
1551 		pr_info("aoe: e%ld.%d: broadcast shelf number invalid\n",
1552 			aoemajor, (int) h->minor);
1553 		return;
1554 	}
1555 	if (h->minor == 0xff) {
1556 		pr_info("aoe: e%ld.%d: broadcast slot number invalid\n",
1557 			aoemajor, (int) h->minor);
1558 		return;
1559 	}
1560 
1561 	n = be16_to_cpu(ch->bufcnt);
1562 	if (n > aoe_maxout)	/* keep it reasonable */
1563 		n = aoe_maxout;
1564 
1565 	d = aoedev_by_aoeaddr(aoemajor, h->minor, 1);
1566 	if (d == NULL) {
1567 		pr_info("aoe: device allocation failure\n");
1568 		return;
1569 	}
1570 
1571 	spin_lock_irqsave(&d->lock, flags);
1572 
1573 	t = gettgt(d, h->src);
1574 	if (t) {
1575 		t->nframes = n;
1576 		if (n < t->maxout)
1577 			aoecmd_wreset(t);
1578 	} else {
1579 		t = addtgt(d, h->src, n);
1580 		if (!t)
1581 			goto bail;
1582 	}
1583 	n = skb->dev->mtu;
1584 	n -= sizeof(struct aoe_hdr) + sizeof(struct aoe_atahdr);
1585 	n /= 512;
1586 	if (n > ch->scnt)
1587 		n = ch->scnt;
1588 	n = n ? n * 512 : DEFAULTBCNT;
1589 	setifbcnt(t, skb->dev, n);
1590 
1591 	/* don't change users' perspective */
1592 	if (d->nopen == 0) {
1593 		d->fw_ver = be16_to_cpu(ch->fwver);
1594 		sl = aoecmd_ata_id(d);
1595 	}
1596 bail:
1597 	spin_unlock_irqrestore(&d->lock, flags);
1598 	aoedev_put(d);
1599 	if (sl) {
1600 		__skb_queue_head_init(&queue);
1601 		__skb_queue_tail(&queue, sl);
1602 		aoenet_xmit(&queue);
1603 	}
1604 }
1605 
1606 void
1607 aoecmd_wreset(struct aoetgt *t)
1608 {
1609 	t->maxout = 1;
1610 	t->ssthresh = t->nframes / 2;
1611 	t->next_cwnd = t->nframes;
1612 }
1613 
1614 void
1615 aoecmd_cleanslate(struct aoedev *d)
1616 {
1617 	struct aoetgt **t, **te;
1618 
1619 	d->rttavg = RTTAVG_INIT;
1620 	d->rttdev = RTTDEV_INIT;
1621 	d->maxbcnt = 0;
1622 
1623 	t = d->targets;
1624 	te = t + d->ntargets;
1625 	for (; t < te && *t; t++)
1626 		aoecmd_wreset(*t);
1627 }
1628 
1629 void
1630 aoe_failbuf(struct aoedev *d, struct buf *buf)
1631 {
1632 	if (buf == NULL)
1633 		return;
1634 	buf->iter.bi_size = 0;
1635 	buf->bio->bi_status = BLK_STS_IOERR;
1636 	if (buf->nframesout == 0)
1637 		aoe_end_buf(d, buf);
1638 }
1639 
1640 void
1641 aoe_flush_iocq(void)
1642 {
1643 	int i;
1644 
1645 	for (i = 0; i < ncpus; i++) {
1646 		if (kts[i].active)
1647 			aoe_flush_iocq_by_index(i);
1648 	}
1649 }
1650 
1651 void
1652 aoe_flush_iocq_by_index(int id)
1653 {
1654 	struct frame *f;
1655 	struct aoedev *d;
1656 	LIST_HEAD(flist);
1657 	struct list_head *pos;
1658 	struct sk_buff *skb;
1659 	ulong flags;
1660 
1661 	spin_lock_irqsave(&iocq[id].lock, flags);
1662 	list_splice_init(&iocq[id].head, &flist);
1663 	spin_unlock_irqrestore(&iocq[id].lock, flags);
1664 	while (!list_empty(&flist)) {
1665 		pos = flist.next;
1666 		list_del(pos);
1667 		f = list_entry(pos, struct frame, head);
1668 		d = f->t->d;
1669 		skb = f->r_skb;
1670 		spin_lock_irqsave(&d->lock, flags);
1671 		if (f->buf) {
1672 			f->buf->nframesout--;
1673 			aoe_failbuf(d, f->buf);
1674 		}
1675 		aoe_freetframe(f);
1676 		spin_unlock_irqrestore(&d->lock, flags);
1677 		dev_kfree_skb(skb);
1678 		aoedev_put(d);
1679 	}
1680 }
1681 
1682 int __init
1683 aoecmd_init(void)
1684 {
1685 	void *p;
1686 	int i;
1687 	int ret;
1688 
1689 	/* get_zeroed_page returns page with ref count 1 */
1690 	p = (void *) get_zeroed_page(GFP_KERNEL);
1691 	if (!p)
1692 		return -ENOMEM;
1693 	empty_page = virt_to_page(p);
1694 
1695 	ncpus = num_online_cpus();
1696 
1697 	iocq = kcalloc(ncpus, sizeof(struct iocq_ktio), GFP_KERNEL);
1698 	if (!iocq)
1699 		return -ENOMEM;
1700 
1701 	kts = kcalloc(ncpus, sizeof(struct ktstate), GFP_KERNEL);
1702 	if (!kts) {
1703 		ret = -ENOMEM;
1704 		goto kts_fail;
1705 	}
1706 
1707 	ktiowq = kcalloc(ncpus, sizeof(wait_queue_head_t), GFP_KERNEL);
1708 	if (!ktiowq) {
1709 		ret = -ENOMEM;
1710 		goto ktiowq_fail;
1711 	}
1712 
1713 	mutex_init(&ktio_spawn_lock);
1714 
1715 	for (i = 0; i < ncpus; i++) {
1716 		INIT_LIST_HEAD(&iocq[i].head);
1717 		spin_lock_init(&iocq[i].lock);
1718 		init_waitqueue_head(&ktiowq[i]);
1719 		snprintf(kts[i].name, sizeof(kts[i].name), "aoe_ktio%d", i);
1720 		kts[i].fn = ktio;
1721 		kts[i].waitq = &ktiowq[i];
1722 		kts[i].lock = &iocq[i].lock;
1723 		kts[i].id = i;
1724 		kts[i].active = 0;
1725 	}
1726 	kts[0].active = 1;
1727 	if (aoe_ktstart(&kts[0])) {
1728 		ret = -ENOMEM;
1729 		goto ktstart_fail;
1730 	}
1731 	return 0;
1732 
1733 ktstart_fail:
1734 	kfree(ktiowq);
1735 ktiowq_fail:
1736 	kfree(kts);
1737 kts_fail:
1738 	kfree(iocq);
1739 
1740 	return ret;
1741 }
1742 
1743 void
1744 aoecmd_exit(void)
1745 {
1746 	int i;
1747 
1748 	for (i = 0; i < ncpus; i++)
1749 		if (kts[i].active)
1750 			aoe_ktstop(&kts[i]);
1751 
1752 	aoe_flush_iocq();
1753 
1754 	/* Free up the iocq and thread speicific configuration
1755 	* allocated during startup.
1756 	*/
1757 	kfree(iocq);
1758 	kfree(kts);
1759 	kfree(ktiowq);
1760 
1761 	free_page((unsigned long) page_address(empty_page));
1762 	empty_page = NULL;
1763 }
1764