xref: /openbmc/linux/drivers/block/aoe/aoecmd.c (revision f35e839a)
1 /* Copyright (c) 2012 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 static wait_queue_head_t ktiowq;
39 static struct ktstate kts;
40 
41 /* io completion queue */
42 static struct {
43 	struct list_head head;
44 	spinlock_t lock;
45 } iocq;
46 
47 static struct page *empty_page;
48 
49 static struct sk_buff *
50 new_skb(ulong len)
51 {
52 	struct sk_buff *skb;
53 
54 	skb = alloc_skb(len + MAX_HEADER, GFP_ATOMIC);
55 	if (skb) {
56 		skb_reserve(skb, MAX_HEADER);
57 		skb_reset_mac_header(skb);
58 		skb_reset_network_header(skb);
59 		skb->protocol = __constant_htons(ETH_P_AOE);
60 		skb_checksum_none_assert(skb);
61 	}
62 	return skb;
63 }
64 
65 static struct frame *
66 getframe_deferred(struct aoedev *d, u32 tag)
67 {
68 	struct list_head *head, *pos, *nx;
69 	struct frame *f;
70 
71 	head = &d->rexmitq;
72 	list_for_each_safe(pos, nx, head) {
73 		f = list_entry(pos, struct frame, head);
74 		if (f->tag == tag) {
75 			list_del(pos);
76 			return f;
77 		}
78 	}
79 	return NULL;
80 }
81 
82 static struct frame *
83 getframe(struct aoedev *d, u32 tag)
84 {
85 	struct frame *f;
86 	struct list_head *head, *pos, *nx;
87 	u32 n;
88 
89 	n = tag % NFACTIVE;
90 	head = &d->factive[n];
91 	list_for_each_safe(pos, nx, head) {
92 		f = list_entry(pos, struct frame, head);
93 		if (f->tag == tag) {
94 			list_del(pos);
95 			return f;
96 		}
97 	}
98 	return NULL;
99 }
100 
101 /*
102  * Leave the top bit clear so we have tagspace for userland.
103  * The bottom 16 bits are the xmit tick for rexmit/rttavg processing.
104  * This driver reserves tag -1 to mean "unused frame."
105  */
106 static int
107 newtag(struct aoedev *d)
108 {
109 	register ulong n;
110 
111 	n = jiffies & 0xffff;
112 	return n |= (++d->lasttag & 0x7fff) << 16;
113 }
114 
115 static u32
116 aoehdr_atainit(struct aoedev *d, struct aoetgt *t, struct aoe_hdr *h)
117 {
118 	u32 host_tag = newtag(d);
119 
120 	memcpy(h->src, t->ifp->nd->dev_addr, sizeof h->src);
121 	memcpy(h->dst, t->addr, sizeof h->dst);
122 	h->type = __constant_cpu_to_be16(ETH_P_AOE);
123 	h->verfl = AOE_HVER;
124 	h->major = cpu_to_be16(d->aoemajor);
125 	h->minor = d->aoeminor;
126 	h->cmd = AOECMD_ATA;
127 	h->tag = cpu_to_be32(host_tag);
128 
129 	return host_tag;
130 }
131 
132 static inline void
133 put_lba(struct aoe_atahdr *ah, sector_t lba)
134 {
135 	ah->lba0 = lba;
136 	ah->lba1 = lba >>= 8;
137 	ah->lba2 = lba >>= 8;
138 	ah->lba3 = lba >>= 8;
139 	ah->lba4 = lba >>= 8;
140 	ah->lba5 = lba >>= 8;
141 }
142 
143 static struct aoeif *
144 ifrotate(struct aoetgt *t)
145 {
146 	struct aoeif *ifp;
147 
148 	ifp = t->ifp;
149 	ifp++;
150 	if (ifp >= &t->ifs[NAOEIFS] || ifp->nd == NULL)
151 		ifp = t->ifs;
152 	if (ifp->nd == NULL)
153 		return NULL;
154 	return t->ifp = ifp;
155 }
156 
157 static void
158 skb_pool_put(struct aoedev *d, struct sk_buff *skb)
159 {
160 	__skb_queue_tail(&d->skbpool, skb);
161 }
162 
163 static struct sk_buff *
164 skb_pool_get(struct aoedev *d)
165 {
166 	struct sk_buff *skb = skb_peek(&d->skbpool);
167 
168 	if (skb && atomic_read(&skb_shinfo(skb)->dataref) == 1) {
169 		__skb_unlink(skb, &d->skbpool);
170 		return skb;
171 	}
172 	if (skb_queue_len(&d->skbpool) < NSKBPOOLMAX &&
173 	    (skb = new_skb(ETH_ZLEN)))
174 		return skb;
175 
176 	return NULL;
177 }
178 
179 void
180 aoe_freetframe(struct frame *f)
181 {
182 	struct aoetgt *t;
183 
184 	t = f->t;
185 	f->buf = NULL;
186 	f->lba = 0;
187 	f->bv = NULL;
188 	f->r_skb = NULL;
189 	f->flags = 0;
190 	list_add(&f->head, &t->ffree);
191 }
192 
193 static struct frame *
194 newtframe(struct aoedev *d, struct aoetgt *t)
195 {
196 	struct frame *f;
197 	struct sk_buff *skb;
198 	struct list_head *pos;
199 
200 	if (list_empty(&t->ffree)) {
201 		if (t->falloc >= NSKBPOOLMAX*2)
202 			return NULL;
203 		f = kcalloc(1, sizeof(*f), GFP_ATOMIC);
204 		if (f == NULL)
205 			return NULL;
206 		t->falloc++;
207 		f->t = t;
208 	} else {
209 		pos = t->ffree.next;
210 		list_del(pos);
211 		f = list_entry(pos, struct frame, head);
212 	}
213 
214 	skb = f->skb;
215 	if (skb == NULL) {
216 		f->skb = skb = new_skb(ETH_ZLEN);
217 		if (!skb) {
218 bail:			aoe_freetframe(f);
219 			return NULL;
220 		}
221 	}
222 
223 	if (atomic_read(&skb_shinfo(skb)->dataref) != 1) {
224 		skb = skb_pool_get(d);
225 		if (skb == NULL)
226 			goto bail;
227 		skb_pool_put(d, f->skb);
228 		f->skb = skb;
229 	}
230 
231 	skb->truesize -= skb->data_len;
232 	skb_shinfo(skb)->nr_frags = skb->data_len = 0;
233 	skb_trim(skb, 0);
234 	return f;
235 }
236 
237 static struct frame *
238 newframe(struct aoedev *d)
239 {
240 	struct frame *f;
241 	struct aoetgt *t, **tt;
242 	int totout = 0;
243 	int use_tainted;
244 	int has_untainted;
245 
246 	if (!d->targets || !d->targets[0]) {
247 		printk(KERN_ERR "aoe: NULL TARGETS!\n");
248 		return NULL;
249 	}
250 	tt = d->tgt;	/* last used target */
251 	for (use_tainted = 0, has_untainted = 0;;) {
252 		tt++;
253 		if (tt >= &d->targets[d->ntargets] || !*tt)
254 			tt = d->targets;
255 		t = *tt;
256 		if (!t->taint) {
257 			has_untainted = 1;
258 			totout += t->nout;
259 		}
260 		if (t->nout < t->maxout
261 		&& (use_tainted || !t->taint)
262 		&& t->ifp->nd) {
263 			f = newtframe(d, t);
264 			if (f) {
265 				ifrotate(t);
266 				d->tgt = tt;
267 				return f;
268 			}
269 		}
270 		if (tt == d->tgt) {	/* we've looped and found nada */
271 			if (!use_tainted && !has_untainted)
272 				use_tainted = 1;
273 			else
274 				break;
275 		}
276 	}
277 	if (totout == 0) {
278 		d->kicked++;
279 		d->flags |= DEVFL_KICKME;
280 	}
281 	return NULL;
282 }
283 
284 static void
285 skb_fillup(struct sk_buff *skb, struct bio_vec *bv, ulong off, ulong cnt)
286 {
287 	int frag = 0;
288 	ulong fcnt;
289 loop:
290 	fcnt = bv->bv_len - (off - bv->bv_offset);
291 	if (fcnt > cnt)
292 		fcnt = cnt;
293 	skb_fill_page_desc(skb, frag++, bv->bv_page, off, fcnt);
294 	cnt -= fcnt;
295 	if (cnt <= 0)
296 		return;
297 	bv++;
298 	off = bv->bv_offset;
299 	goto loop;
300 }
301 
302 static void
303 fhash(struct frame *f)
304 {
305 	struct aoedev *d = f->t->d;
306 	u32 n;
307 
308 	n = f->tag % NFACTIVE;
309 	list_add_tail(&f->head, &d->factive[n]);
310 }
311 
312 static void
313 ata_rw_frameinit(struct frame *f)
314 {
315 	struct aoetgt *t;
316 	struct aoe_hdr *h;
317 	struct aoe_atahdr *ah;
318 	struct sk_buff *skb;
319 	char writebit, extbit;
320 
321 	skb = f->skb;
322 	h = (struct aoe_hdr *) skb_mac_header(skb);
323 	ah = (struct aoe_atahdr *) (h + 1);
324 	skb_put(skb, sizeof(*h) + sizeof(*ah));
325 	memset(h, 0, skb->len);
326 
327 	writebit = 0x10;
328 	extbit = 0x4;
329 
330 	t = f->t;
331 	f->tag = aoehdr_atainit(t->d, t, h);
332 	fhash(f);
333 	t->nout++;
334 	f->waited = 0;
335 	f->waited_total = 0;
336 	if (f->buf)
337 		f->lba = f->buf->sector;
338 
339 	/* set up ata header */
340 	ah->scnt = f->bcnt >> 9;
341 	put_lba(ah, f->lba);
342 	if (t->d->flags & DEVFL_EXT) {
343 		ah->aflags |= AOEAFL_EXT;
344 	} else {
345 		extbit = 0;
346 		ah->lba3 &= 0x0f;
347 		ah->lba3 |= 0xe0;	/* LBA bit + obsolete 0xa0 */
348 	}
349 	if (f->buf && bio_data_dir(f->buf->bio) == WRITE) {
350 		skb_fillup(skb, f->bv, f->bv_off, f->bcnt);
351 		ah->aflags |= AOEAFL_WRITE;
352 		skb->len += f->bcnt;
353 		skb->data_len = f->bcnt;
354 		skb->truesize += f->bcnt;
355 		t->wpkts++;
356 	} else {
357 		t->rpkts++;
358 		writebit = 0;
359 	}
360 
361 	ah->cmdstat = ATA_CMD_PIO_READ | writebit | extbit;
362 	skb->dev = t->ifp->nd;
363 }
364 
365 static int
366 aoecmd_ata_rw(struct aoedev *d)
367 {
368 	struct frame *f;
369 	struct buf *buf;
370 	struct aoetgt *t;
371 	struct sk_buff *skb;
372 	struct sk_buff_head queue;
373 	ulong bcnt, fbcnt;
374 
375 	buf = nextbuf(d);
376 	if (buf == NULL)
377 		return 0;
378 	f = newframe(d);
379 	if (f == NULL)
380 		return 0;
381 	t = *d->tgt;
382 	bcnt = d->maxbcnt;
383 	if (bcnt == 0)
384 		bcnt = DEFAULTBCNT;
385 	if (bcnt > buf->resid)
386 		bcnt = buf->resid;
387 	fbcnt = bcnt;
388 	f->bv = buf->bv;
389 	f->bv_off = f->bv->bv_offset + (f->bv->bv_len - buf->bv_resid);
390 	do {
391 		if (fbcnt < buf->bv_resid) {
392 			buf->bv_resid -= fbcnt;
393 			buf->resid -= fbcnt;
394 			break;
395 		}
396 		fbcnt -= buf->bv_resid;
397 		buf->resid -= buf->bv_resid;
398 		if (buf->resid == 0) {
399 			d->ip.buf = NULL;
400 			break;
401 		}
402 		buf->bv++;
403 		buf->bv_resid = buf->bv->bv_len;
404 		WARN_ON(buf->bv_resid == 0);
405 	} while (fbcnt);
406 
407 	/* initialize the headers & frame */
408 	f->buf = buf;
409 	f->bcnt = bcnt;
410 	ata_rw_frameinit(f);
411 
412 	/* mark all tracking fields and load out */
413 	buf->nframesout += 1;
414 	buf->sector += bcnt >> 9;
415 
416 	skb = skb_clone(f->skb, GFP_ATOMIC);
417 	if (skb) {
418 		do_gettimeofday(&f->sent);
419 		f->sent_jiffs = (u32) jiffies;
420 		__skb_queue_head_init(&queue);
421 		__skb_queue_tail(&queue, skb);
422 		aoenet_xmit(&queue);
423 	}
424 	return 1;
425 }
426 
427 /* some callers cannot sleep, and they can call this function,
428  * transmitting the packets later, when interrupts are on
429  */
430 static void
431 aoecmd_cfg_pkts(ushort aoemajor, unsigned char aoeminor, struct sk_buff_head *queue)
432 {
433 	struct aoe_hdr *h;
434 	struct aoe_cfghdr *ch;
435 	struct sk_buff *skb;
436 	struct net_device *ifp;
437 
438 	rcu_read_lock();
439 	for_each_netdev_rcu(&init_net, ifp) {
440 		dev_hold(ifp);
441 		if (!is_aoe_netif(ifp))
442 			goto cont;
443 
444 		skb = new_skb(sizeof *h + sizeof *ch);
445 		if (skb == NULL) {
446 			printk(KERN_INFO "aoe: skb alloc failure\n");
447 			goto cont;
448 		}
449 		skb_put(skb, sizeof *h + sizeof *ch);
450 		skb->dev = ifp;
451 		__skb_queue_tail(queue, skb);
452 		h = (struct aoe_hdr *) skb_mac_header(skb);
453 		memset(h, 0, sizeof *h + sizeof *ch);
454 
455 		memset(h->dst, 0xff, sizeof h->dst);
456 		memcpy(h->src, ifp->dev_addr, sizeof h->src);
457 		h->type = __constant_cpu_to_be16(ETH_P_AOE);
458 		h->verfl = AOE_HVER;
459 		h->major = cpu_to_be16(aoemajor);
460 		h->minor = aoeminor;
461 		h->cmd = AOECMD_CFG;
462 
463 cont:
464 		dev_put(ifp);
465 	}
466 	rcu_read_unlock();
467 }
468 
469 static void
470 resend(struct aoedev *d, struct frame *f)
471 {
472 	struct sk_buff *skb;
473 	struct sk_buff_head queue;
474 	struct aoe_hdr *h;
475 	struct aoe_atahdr *ah;
476 	struct aoetgt *t;
477 	char buf[128];
478 	u32 n;
479 
480 	t = f->t;
481 	n = newtag(d);
482 	skb = f->skb;
483 	if (ifrotate(t) == NULL) {
484 		/* probably can't happen, but set it up to fail anyway */
485 		pr_info("aoe: resend: no interfaces to rotate to.\n");
486 		ktcomplete(f, NULL);
487 		return;
488 	}
489 	h = (struct aoe_hdr *) skb_mac_header(skb);
490 	ah = (struct aoe_atahdr *) (h+1);
491 
492 	if (!(f->flags & FFL_PROBE)) {
493 		snprintf(buf, sizeof(buf),
494 			"%15s e%ld.%d oldtag=%08x@%08lx newtag=%08x s=%pm d=%pm nout=%d\n",
495 			"retransmit", d->aoemajor, d->aoeminor,
496 			f->tag, jiffies, n,
497 			h->src, h->dst, t->nout);
498 		aoechr_error(buf);
499 	}
500 
501 	f->tag = n;
502 	fhash(f);
503 	h->tag = cpu_to_be32(n);
504 	memcpy(h->dst, t->addr, sizeof h->dst);
505 	memcpy(h->src, t->ifp->nd->dev_addr, sizeof h->src);
506 
507 	skb->dev = t->ifp->nd;
508 	skb = skb_clone(skb, GFP_ATOMIC);
509 	if (skb == NULL)
510 		return;
511 	do_gettimeofday(&f->sent);
512 	f->sent_jiffs = (u32) jiffies;
513 	__skb_queue_head_init(&queue);
514 	__skb_queue_tail(&queue, skb);
515 	aoenet_xmit(&queue);
516 }
517 
518 static int
519 tsince_hr(struct frame *f)
520 {
521 	struct timeval now;
522 	int n;
523 
524 	do_gettimeofday(&now);
525 	n = now.tv_usec - f->sent.tv_usec;
526 	n += (now.tv_sec - f->sent.tv_sec) * USEC_PER_SEC;
527 
528 	if (n < 0)
529 		n = -n;
530 
531 	/* For relatively long periods, use jiffies to avoid
532 	 * discrepancies caused by updates to the system time.
533 	 *
534 	 * On system with HZ of 1000, 32-bits is over 49 days
535 	 * worth of jiffies, or over 71 minutes worth of usecs.
536 	 *
537 	 * Jiffies overflow is handled by subtraction of unsigned ints:
538 	 * (gdb) print (unsigned) 2 - (unsigned) 0xfffffffe
539 	 * $3 = 4
540 	 * (gdb)
541 	 */
542 	if (n > USEC_PER_SEC / 4) {
543 		n = ((u32) jiffies) - f->sent_jiffs;
544 		n *= USEC_PER_SEC / HZ;
545 	}
546 
547 	return n;
548 }
549 
550 static int
551 tsince(u32 tag)
552 {
553 	int n;
554 
555 	n = jiffies & 0xffff;
556 	n -= tag & 0xffff;
557 	if (n < 0)
558 		n += 1<<16;
559 	return jiffies_to_usecs(n + 1);
560 }
561 
562 static struct aoeif *
563 getif(struct aoetgt *t, struct net_device *nd)
564 {
565 	struct aoeif *p, *e;
566 
567 	p = t->ifs;
568 	e = p + NAOEIFS;
569 	for (; p < e; p++)
570 		if (p->nd == nd)
571 			return p;
572 	return NULL;
573 }
574 
575 static void
576 ejectif(struct aoetgt *t, struct aoeif *ifp)
577 {
578 	struct aoeif *e;
579 	struct net_device *nd;
580 	ulong n;
581 
582 	nd = ifp->nd;
583 	e = t->ifs + NAOEIFS - 1;
584 	n = (e - ifp) * sizeof *ifp;
585 	memmove(ifp, ifp+1, n);
586 	e->nd = NULL;
587 	dev_put(nd);
588 }
589 
590 static struct frame *
591 reassign_frame(struct frame *f)
592 {
593 	struct frame *nf;
594 	struct sk_buff *skb;
595 
596 	nf = newframe(f->t->d);
597 	if (!nf)
598 		return NULL;
599 	if (nf->t == f->t) {
600 		aoe_freetframe(nf);
601 		return NULL;
602 	}
603 
604 	skb = nf->skb;
605 	nf->skb = f->skb;
606 	nf->buf = f->buf;
607 	nf->bcnt = f->bcnt;
608 	nf->lba = f->lba;
609 	nf->bv = f->bv;
610 	nf->bv_off = f->bv_off;
611 	nf->waited = 0;
612 	nf->waited_total = f->waited_total;
613 	nf->sent = f->sent;
614 	nf->sent_jiffs = f->sent_jiffs;
615 	f->skb = skb;
616 
617 	return nf;
618 }
619 
620 static void
621 probe(struct aoetgt *t)
622 {
623 	struct aoedev *d;
624 	struct frame *f;
625 	struct sk_buff *skb;
626 	struct sk_buff_head queue;
627 	size_t n, m;
628 	int frag;
629 
630 	d = t->d;
631 	f = newtframe(d, t);
632 	if (!f) {
633 		pr_err("%s %pm for e%ld.%d: %s\n",
634 			"aoe: cannot probe remote address",
635 			t->addr,
636 			(long) d->aoemajor, d->aoeminor,
637 			"no frame available");
638 		return;
639 	}
640 	f->flags |= FFL_PROBE;
641 	ifrotate(t);
642 	f->bcnt = t->d->maxbcnt ? t->d->maxbcnt : DEFAULTBCNT;
643 	ata_rw_frameinit(f);
644 	skb = f->skb;
645 	for (frag = 0, n = f->bcnt; n > 0; ++frag, n -= m) {
646 		if (n < PAGE_SIZE)
647 			m = n;
648 		else
649 			m = PAGE_SIZE;
650 		skb_fill_page_desc(skb, frag, empty_page, 0, m);
651 	}
652 	skb->len += f->bcnt;
653 	skb->data_len = f->bcnt;
654 	skb->truesize += f->bcnt;
655 
656 	skb = skb_clone(f->skb, GFP_ATOMIC);
657 	if (skb) {
658 		do_gettimeofday(&f->sent);
659 		f->sent_jiffs = (u32) jiffies;
660 		__skb_queue_head_init(&queue);
661 		__skb_queue_tail(&queue, skb);
662 		aoenet_xmit(&queue);
663 	}
664 }
665 
666 static long
667 rto(struct aoedev *d)
668 {
669 	long t;
670 
671 	t = 2 * d->rttavg >> RTTSCALE;
672 	t += 8 * d->rttdev >> RTTDSCALE;
673 	if (t == 0)
674 		t = 1;
675 
676 	return t;
677 }
678 
679 static void
680 rexmit_deferred(struct aoedev *d)
681 {
682 	struct aoetgt *t;
683 	struct frame *f;
684 	struct frame *nf;
685 	struct list_head *pos, *nx, *head;
686 	int since;
687 	int untainted;
688 
689 	count_targets(d, &untainted);
690 
691 	head = &d->rexmitq;
692 	list_for_each_safe(pos, nx, head) {
693 		f = list_entry(pos, struct frame, head);
694 		t = f->t;
695 		if (t->taint) {
696 			if (!(f->flags & FFL_PROBE)) {
697 				nf = reassign_frame(f);
698 				if (nf) {
699 					if (t->nout_probes == 0
700 					&& untainted > 0) {
701 						probe(t);
702 						t->nout_probes++;
703 					}
704 					list_replace(&f->head, &nf->head);
705 					pos = &nf->head;
706 					aoe_freetframe(f);
707 					f = nf;
708 					t = f->t;
709 				}
710 			} else if (untainted < 1) {
711 				/* don't probe w/o other untainted aoetgts */
712 				goto stop_probe;
713 			} else if (tsince_hr(f) < t->taint * rto(d)) {
714 				/* reprobe slowly when taint is high */
715 				continue;
716 			}
717 		} else if (f->flags & FFL_PROBE) {
718 stop_probe:		/* don't probe untainted aoetgts */
719 			list_del(pos);
720 			aoe_freetframe(f);
721 			/* leaving d->kicked, because this is routine */
722 			f->t->d->flags |= DEVFL_KICKME;
723 			continue;
724 		}
725 		if (t->nout >= t->maxout)
726 			continue;
727 		list_del(pos);
728 		t->nout++;
729 		if (f->flags & FFL_PROBE)
730 			t->nout_probes++;
731 		since = tsince_hr(f);
732 		f->waited += since;
733 		f->waited_total += since;
734 		resend(d, f);
735 	}
736 }
737 
738 /* An aoetgt accumulates demerits quickly, and successful
739  * probing redeems the aoetgt slowly.
740  */
741 static void
742 scorn(struct aoetgt *t)
743 {
744 	int n;
745 
746 	n = t->taint++;
747 	t->taint += t->taint * 2;
748 	if (n > t->taint)
749 		t->taint = n;
750 	if (t->taint > MAX_TAINT)
751 		t->taint = MAX_TAINT;
752 }
753 
754 static int
755 count_targets(struct aoedev *d, int *untainted)
756 {
757 	int i, good;
758 
759 	for (i = good = 0; i < d->ntargets && d->targets[i]; ++i)
760 		if (d->targets[i]->taint == 0)
761 			good++;
762 
763 	if (untainted)
764 		*untainted = good;
765 	return i;
766 }
767 
768 static void
769 rexmit_timer(ulong vp)
770 {
771 	struct aoedev *d;
772 	struct aoetgt *t;
773 	struct aoeif *ifp;
774 	struct frame *f;
775 	struct list_head *head, *pos, *nx;
776 	LIST_HEAD(flist);
777 	register long timeout;
778 	ulong flags, n;
779 	int i;
780 	int utgts;	/* number of aoetgt descriptors (not slots) */
781 	int since;
782 
783 	d = (struct aoedev *) vp;
784 
785 	spin_lock_irqsave(&d->lock, flags);
786 
787 	/* timeout based on observed timings and variations */
788 	timeout = rto(d);
789 
790 	utgts = count_targets(d, NULL);
791 
792 	if (d->flags & DEVFL_TKILL) {
793 		spin_unlock_irqrestore(&d->lock, flags);
794 		return;
795 	}
796 
797 	/* collect all frames to rexmit into flist */
798 	for (i = 0; i < NFACTIVE; i++) {
799 		head = &d->factive[i];
800 		list_for_each_safe(pos, nx, head) {
801 			f = list_entry(pos, struct frame, head);
802 			if (tsince_hr(f) < timeout)
803 				break;	/* end of expired frames */
804 			/* move to flist for later processing */
805 			list_move_tail(pos, &flist);
806 		}
807 	}
808 
809 	/* process expired frames */
810 	while (!list_empty(&flist)) {
811 		pos = flist.next;
812 		f = list_entry(pos, struct frame, head);
813 		since = tsince_hr(f);
814 		n = f->waited_total + since;
815 		n /= USEC_PER_SEC;
816 		if (aoe_deadsecs
817 		&& n > aoe_deadsecs
818 		&& !(f->flags & FFL_PROBE)) {
819 			/* Waited too long.  Device failure.
820 			 * Hang all frames on first hash bucket for downdev
821 			 * to clean up.
822 			 */
823 			list_splice(&flist, &d->factive[0]);
824 			aoedev_downdev(d);
825 			goto out;
826 		}
827 
828 		t = f->t;
829 		n = f->waited + since;
830 		n /= USEC_PER_SEC;
831 		if (aoe_deadsecs && utgts > 0
832 		&& (n > aoe_deadsecs / utgts || n > HARD_SCORN_SECS))
833 			scorn(t); /* avoid this target */
834 
835 		if (t->maxout != 1) {
836 			t->ssthresh = t->maxout / 2;
837 			t->maxout = 1;
838 		}
839 
840 		if (f->flags & FFL_PROBE) {
841 			t->nout_probes--;
842 		} else {
843 			ifp = getif(t, f->skb->dev);
844 			if (ifp && ++ifp->lost > (t->nframes << 1)
845 			&& (ifp != t->ifs || t->ifs[1].nd)) {
846 				ejectif(t, ifp);
847 				ifp = NULL;
848 			}
849 		}
850 		list_move_tail(pos, &d->rexmitq);
851 		t->nout--;
852 	}
853 	rexmit_deferred(d);
854 
855 out:
856 	if ((d->flags & DEVFL_KICKME) && d->blkq) {
857 		d->flags &= ~DEVFL_KICKME;
858 		d->blkq->request_fn(d->blkq);
859 	}
860 
861 	d->timer.expires = jiffies + TIMERTICK;
862 	add_timer(&d->timer);
863 
864 	spin_unlock_irqrestore(&d->lock, flags);
865 }
866 
867 static unsigned long
868 rqbiocnt(struct request *r)
869 {
870 	struct bio *bio;
871 	unsigned long n = 0;
872 
873 	__rq_for_each_bio(bio, r)
874 		n++;
875 	return n;
876 }
877 
878 /* This can be removed if we are certain that no users of the block
879  * layer will ever use zero-count pages in bios.  Otherwise we have to
880  * protect against the put_page sometimes done by the network layer.
881  *
882  * See http://oss.sgi.com/archives/xfs/2007-01/msg00594.html for
883  * discussion.
884  *
885  * We cannot use get_page in the workaround, because it insists on a
886  * positive page count as a precondition.  So we use _count directly.
887  */
888 static void
889 bio_pageinc(struct bio *bio)
890 {
891 	struct bio_vec *bv;
892 	struct page *page;
893 	int i;
894 
895 	bio_for_each_segment(bv, bio, i) {
896 		page = bv->bv_page;
897 		/* Non-zero page count for non-head members of
898 		 * compound pages is no longer allowed by the kernel,
899 		 * but this has never been seen here.
900 		 */
901 		if (unlikely(PageCompound(page)))
902 			if (compound_trans_head(page) != page) {
903 				pr_crit("page tail used for block I/O\n");
904 				BUG();
905 			}
906 		atomic_inc(&page->_count);
907 	}
908 }
909 
910 static void
911 bio_pagedec(struct bio *bio)
912 {
913 	struct bio_vec *bv;
914 	int i;
915 
916 	bio_for_each_segment(bv, bio, i)
917 		atomic_dec(&bv->bv_page->_count);
918 }
919 
920 static void
921 bufinit(struct buf *buf, struct request *rq, struct bio *bio)
922 {
923 	memset(buf, 0, sizeof(*buf));
924 	buf->rq = rq;
925 	buf->bio = bio;
926 	buf->resid = bio->bi_size;
927 	buf->sector = bio->bi_sector;
928 	bio_pageinc(bio);
929 	buf->bv = bio_iovec(bio);
930 	buf->bv_resid = buf->bv->bv_len;
931 	WARN_ON(buf->bv_resid == 0);
932 }
933 
934 static struct buf *
935 nextbuf(struct aoedev *d)
936 {
937 	struct request *rq;
938 	struct request_queue *q;
939 	struct buf *buf;
940 	struct bio *bio;
941 
942 	q = d->blkq;
943 	if (q == NULL)
944 		return NULL;	/* initializing */
945 	if (d->ip.buf)
946 		return d->ip.buf;
947 	rq = d->ip.rq;
948 	if (rq == NULL) {
949 		rq = blk_peek_request(q);
950 		if (rq == NULL)
951 			return NULL;
952 		blk_start_request(rq);
953 		d->ip.rq = rq;
954 		d->ip.nxbio = rq->bio;
955 		rq->special = (void *) rqbiocnt(rq);
956 	}
957 	buf = mempool_alloc(d->bufpool, GFP_ATOMIC);
958 	if (buf == NULL) {
959 		pr_err("aoe: nextbuf: unable to mempool_alloc!\n");
960 		return NULL;
961 	}
962 	bio = d->ip.nxbio;
963 	bufinit(buf, rq, bio);
964 	bio = bio->bi_next;
965 	d->ip.nxbio = bio;
966 	if (bio == NULL)
967 		d->ip.rq = NULL;
968 	return d->ip.buf = buf;
969 }
970 
971 /* enters with d->lock held */
972 void
973 aoecmd_work(struct aoedev *d)
974 {
975 	rexmit_deferred(d);
976 	while (aoecmd_ata_rw(d))
977 		;
978 }
979 
980 /* this function performs work that has been deferred until sleeping is OK
981  */
982 void
983 aoecmd_sleepwork(struct work_struct *work)
984 {
985 	struct aoedev *d = container_of(work, struct aoedev, work);
986 	struct block_device *bd;
987 	u64 ssize;
988 
989 	if (d->flags & DEVFL_GDALLOC)
990 		aoeblk_gdalloc(d);
991 
992 	if (d->flags & DEVFL_NEWSIZE) {
993 		ssize = get_capacity(d->gd);
994 		bd = bdget_disk(d->gd, 0);
995 		if (bd) {
996 			mutex_lock(&bd->bd_inode->i_mutex);
997 			i_size_write(bd->bd_inode, (loff_t)ssize<<9);
998 			mutex_unlock(&bd->bd_inode->i_mutex);
999 			bdput(bd);
1000 		}
1001 		spin_lock_irq(&d->lock);
1002 		d->flags |= DEVFL_UP;
1003 		d->flags &= ~DEVFL_NEWSIZE;
1004 		spin_unlock_irq(&d->lock);
1005 	}
1006 }
1007 
1008 static void
1009 ata_ident_fixstring(u16 *id, int ns)
1010 {
1011 	u16 s;
1012 
1013 	while (ns-- > 0) {
1014 		s = *id;
1015 		*id++ = s >> 8 | s << 8;
1016 	}
1017 }
1018 
1019 static void
1020 ataid_complete(struct aoedev *d, struct aoetgt *t, unsigned char *id)
1021 {
1022 	u64 ssize;
1023 	u16 n;
1024 
1025 	/* word 83: command set supported */
1026 	n = get_unaligned_le16(&id[83 << 1]);
1027 
1028 	/* word 86: command set/feature enabled */
1029 	n |= get_unaligned_le16(&id[86 << 1]);
1030 
1031 	if (n & (1<<10)) {	/* bit 10: LBA 48 */
1032 		d->flags |= DEVFL_EXT;
1033 
1034 		/* word 100: number lba48 sectors */
1035 		ssize = get_unaligned_le64(&id[100 << 1]);
1036 
1037 		/* set as in ide-disk.c:init_idedisk_capacity */
1038 		d->geo.cylinders = ssize;
1039 		d->geo.cylinders /= (255 * 63);
1040 		d->geo.heads = 255;
1041 		d->geo.sectors = 63;
1042 	} else {
1043 		d->flags &= ~DEVFL_EXT;
1044 
1045 		/* number lba28 sectors */
1046 		ssize = get_unaligned_le32(&id[60 << 1]);
1047 
1048 		/* NOTE: obsolete in ATA 6 */
1049 		d->geo.cylinders = get_unaligned_le16(&id[54 << 1]);
1050 		d->geo.heads = get_unaligned_le16(&id[55 << 1]);
1051 		d->geo.sectors = get_unaligned_le16(&id[56 << 1]);
1052 	}
1053 
1054 	ata_ident_fixstring((u16 *) &id[10<<1], 10);	/* serial */
1055 	ata_ident_fixstring((u16 *) &id[23<<1], 4);	/* firmware */
1056 	ata_ident_fixstring((u16 *) &id[27<<1], 20);	/* model */
1057 	memcpy(d->ident, id, sizeof(d->ident));
1058 
1059 	if (d->ssize != ssize)
1060 		printk(KERN_INFO
1061 			"aoe: %pm e%ld.%d v%04x has %llu sectors\n",
1062 			t->addr,
1063 			d->aoemajor, d->aoeminor,
1064 			d->fw_ver, (long long)ssize);
1065 	d->ssize = ssize;
1066 	d->geo.start = 0;
1067 	if (d->flags & (DEVFL_GDALLOC|DEVFL_NEWSIZE))
1068 		return;
1069 	if (d->gd != NULL) {
1070 		set_capacity(d->gd, ssize);
1071 		d->flags |= DEVFL_NEWSIZE;
1072 	} else
1073 		d->flags |= DEVFL_GDALLOC;
1074 	schedule_work(&d->work);
1075 }
1076 
1077 static void
1078 calc_rttavg(struct aoedev *d, struct aoetgt *t, int rtt)
1079 {
1080 	register long n;
1081 
1082 	n = rtt;
1083 
1084 	/* cf. Congestion Avoidance and Control, Jacobson & Karels, 1988 */
1085 	n -= d->rttavg >> RTTSCALE;
1086 	d->rttavg += n;
1087 	if (n < 0)
1088 		n = -n;
1089 	n -= d->rttdev >> RTTDSCALE;
1090 	d->rttdev += n;
1091 
1092 	if (!t || t->maxout >= t->nframes)
1093 		return;
1094 	if (t->maxout < t->ssthresh)
1095 		t->maxout += 1;
1096 	else if (t->nout == t->maxout && t->next_cwnd-- == 0) {
1097 		t->maxout += 1;
1098 		t->next_cwnd = t->maxout;
1099 	}
1100 }
1101 
1102 static struct aoetgt *
1103 gettgt(struct aoedev *d, char *addr)
1104 {
1105 	struct aoetgt **t, **e;
1106 
1107 	t = d->targets;
1108 	e = t + d->ntargets;
1109 	for (; t < e && *t; t++)
1110 		if (memcmp((*t)->addr, addr, sizeof((*t)->addr)) == 0)
1111 			return *t;
1112 	return NULL;
1113 }
1114 
1115 static void
1116 bvcpy(struct bio_vec *bv, ulong off, struct sk_buff *skb, long cnt)
1117 {
1118 	ulong fcnt;
1119 	char *p;
1120 	int soff = 0;
1121 loop:
1122 	fcnt = bv->bv_len - (off - bv->bv_offset);
1123 	if (fcnt > cnt)
1124 		fcnt = cnt;
1125 	p = page_address(bv->bv_page) + off;
1126 	skb_copy_bits(skb, soff, p, fcnt);
1127 	soff += fcnt;
1128 	cnt -= fcnt;
1129 	if (cnt <= 0)
1130 		return;
1131 	bv++;
1132 	off = bv->bv_offset;
1133 	goto loop;
1134 }
1135 
1136 void
1137 aoe_end_request(struct aoedev *d, struct request *rq, int fastfail)
1138 {
1139 	struct bio *bio;
1140 	int bok;
1141 	struct request_queue *q;
1142 
1143 	q = d->blkq;
1144 	if (rq == d->ip.rq)
1145 		d->ip.rq = NULL;
1146 	do {
1147 		bio = rq->bio;
1148 		bok = !fastfail && test_bit(BIO_UPTODATE, &bio->bi_flags);
1149 	} while (__blk_end_request(rq, bok ? 0 : -EIO, bio->bi_size));
1150 
1151 	/* cf. http://lkml.org/lkml/2006/10/31/28 */
1152 	if (!fastfail)
1153 		__blk_run_queue(q);
1154 }
1155 
1156 static void
1157 aoe_end_buf(struct aoedev *d, struct buf *buf)
1158 {
1159 	struct request *rq;
1160 	unsigned long n;
1161 
1162 	if (buf == d->ip.buf)
1163 		d->ip.buf = NULL;
1164 	rq = buf->rq;
1165 	bio_pagedec(buf->bio);
1166 	mempool_free(buf, d->bufpool);
1167 	n = (unsigned long) rq->special;
1168 	rq->special = (void *) --n;
1169 	if (n == 0)
1170 		aoe_end_request(d, rq, 0);
1171 }
1172 
1173 static void
1174 ktiocomplete(struct frame *f)
1175 {
1176 	struct aoe_hdr *hin, *hout;
1177 	struct aoe_atahdr *ahin, *ahout;
1178 	struct buf *buf;
1179 	struct sk_buff *skb;
1180 	struct aoetgt *t;
1181 	struct aoeif *ifp;
1182 	struct aoedev *d;
1183 	long n;
1184 	int untainted;
1185 
1186 	if (f == NULL)
1187 		return;
1188 
1189 	t = f->t;
1190 	d = t->d;
1191 	skb = f->r_skb;
1192 	buf = f->buf;
1193 	if (f->flags & FFL_PROBE)
1194 		goto out;
1195 	if (!skb)		/* just fail the buf. */
1196 		goto noskb;
1197 
1198 	hout = (struct aoe_hdr *) skb_mac_header(f->skb);
1199 	ahout = (struct aoe_atahdr *) (hout+1);
1200 
1201 	hin = (struct aoe_hdr *) skb->data;
1202 	skb_pull(skb, sizeof(*hin));
1203 	ahin = (struct aoe_atahdr *) skb->data;
1204 	skb_pull(skb, sizeof(*ahin));
1205 	if (ahin->cmdstat & 0xa9) {	/* these bits cleared on success */
1206 		pr_err("aoe: ata error cmd=%2.2Xh stat=%2.2Xh from e%ld.%d\n",
1207 			ahout->cmdstat, ahin->cmdstat,
1208 			d->aoemajor, d->aoeminor);
1209 noskb:		if (buf)
1210 			clear_bit(BIO_UPTODATE, &buf->bio->bi_flags);
1211 		goto out;
1212 	}
1213 
1214 	n = ahout->scnt << 9;
1215 	switch (ahout->cmdstat) {
1216 	case ATA_CMD_PIO_READ:
1217 	case ATA_CMD_PIO_READ_EXT:
1218 		if (skb->len < n) {
1219 			pr_err("%s e%ld.%d.  skb->len=%d need=%ld\n",
1220 				"aoe: runt data size in read from",
1221 				(long) d->aoemajor, d->aoeminor,
1222 			       skb->len, n);
1223 			clear_bit(BIO_UPTODATE, &buf->bio->bi_flags);
1224 			break;
1225 		}
1226 		bvcpy(f->bv, f->bv_off, skb, n);
1227 	case ATA_CMD_PIO_WRITE:
1228 	case ATA_CMD_PIO_WRITE_EXT:
1229 		spin_lock_irq(&d->lock);
1230 		ifp = getif(t, skb->dev);
1231 		if (ifp)
1232 			ifp->lost = 0;
1233 		spin_unlock_irq(&d->lock);
1234 		break;
1235 	case ATA_CMD_ID_ATA:
1236 		if (skb->len < 512) {
1237 			pr_info("%s e%ld.%d.  skb->len=%d need=512\n",
1238 				"aoe: runt data size in ataid from",
1239 				(long) d->aoemajor, d->aoeminor,
1240 				skb->len);
1241 			break;
1242 		}
1243 		if (skb_linearize(skb))
1244 			break;
1245 		spin_lock_irq(&d->lock);
1246 		ataid_complete(d, t, skb->data);
1247 		spin_unlock_irq(&d->lock);
1248 		break;
1249 	default:
1250 		pr_info("aoe: unrecognized ata command %2.2Xh for %d.%d\n",
1251 			ahout->cmdstat,
1252 			be16_to_cpu(get_unaligned(&hin->major)),
1253 			hin->minor);
1254 	}
1255 out:
1256 	spin_lock_irq(&d->lock);
1257 	if (t->taint > 0
1258 	&& --t->taint > 0
1259 	&& t->nout_probes == 0) {
1260 		count_targets(d, &untainted);
1261 		if (untainted > 0) {
1262 			probe(t);
1263 			t->nout_probes++;
1264 		}
1265 	}
1266 
1267 	aoe_freetframe(f);
1268 
1269 	if (buf && --buf->nframesout == 0 && buf->resid == 0)
1270 		aoe_end_buf(d, buf);
1271 
1272 	spin_unlock_irq(&d->lock);
1273 	aoedev_put(d);
1274 	dev_kfree_skb(skb);
1275 }
1276 
1277 /* Enters with iocq.lock held.
1278  * Returns true iff responses needing processing remain.
1279  */
1280 static int
1281 ktio(void)
1282 {
1283 	struct frame *f;
1284 	struct list_head *pos;
1285 	int i;
1286 
1287 	for (i = 0; ; ++i) {
1288 		if (i == MAXIOC)
1289 			return 1;
1290 		if (list_empty(&iocq.head))
1291 			return 0;
1292 		pos = iocq.head.next;
1293 		list_del(pos);
1294 		spin_unlock_irq(&iocq.lock);
1295 		f = list_entry(pos, struct frame, head);
1296 		ktiocomplete(f);
1297 		spin_lock_irq(&iocq.lock);
1298 	}
1299 }
1300 
1301 static int
1302 kthread(void *vp)
1303 {
1304 	struct ktstate *k;
1305 	DECLARE_WAITQUEUE(wait, current);
1306 	int more;
1307 
1308 	k = vp;
1309 	current->flags |= PF_NOFREEZE;
1310 	set_user_nice(current, -10);
1311 	complete(&k->rendez);	/* tell spawner we're running */
1312 	do {
1313 		spin_lock_irq(k->lock);
1314 		more = k->fn();
1315 		if (!more) {
1316 			add_wait_queue(k->waitq, &wait);
1317 			__set_current_state(TASK_INTERRUPTIBLE);
1318 		}
1319 		spin_unlock_irq(k->lock);
1320 		if (!more) {
1321 			schedule();
1322 			remove_wait_queue(k->waitq, &wait);
1323 		} else
1324 			cond_resched();
1325 	} while (!kthread_should_stop());
1326 	complete(&k->rendez);	/* tell spawner we're stopping */
1327 	return 0;
1328 }
1329 
1330 void
1331 aoe_ktstop(struct ktstate *k)
1332 {
1333 	kthread_stop(k->task);
1334 	wait_for_completion(&k->rendez);
1335 }
1336 
1337 int
1338 aoe_ktstart(struct ktstate *k)
1339 {
1340 	struct task_struct *task;
1341 
1342 	init_completion(&k->rendez);
1343 	task = kthread_run(kthread, k, k->name);
1344 	if (task == NULL || IS_ERR(task))
1345 		return -ENOMEM;
1346 	k->task = task;
1347 	wait_for_completion(&k->rendez); /* allow kthread to start */
1348 	init_completion(&k->rendez);	/* for waiting for exit later */
1349 	return 0;
1350 }
1351 
1352 /* pass it off to kthreads for processing */
1353 static void
1354 ktcomplete(struct frame *f, struct sk_buff *skb)
1355 {
1356 	ulong flags;
1357 
1358 	f->r_skb = skb;
1359 	spin_lock_irqsave(&iocq.lock, flags);
1360 	list_add_tail(&f->head, &iocq.head);
1361 	spin_unlock_irqrestore(&iocq.lock, flags);
1362 	wake_up(&ktiowq);
1363 }
1364 
1365 struct sk_buff *
1366 aoecmd_ata_rsp(struct sk_buff *skb)
1367 {
1368 	struct aoedev *d;
1369 	struct aoe_hdr *h;
1370 	struct frame *f;
1371 	u32 n;
1372 	ulong flags;
1373 	char ebuf[128];
1374 	u16 aoemajor;
1375 
1376 	h = (struct aoe_hdr *) skb->data;
1377 	aoemajor = be16_to_cpu(get_unaligned(&h->major));
1378 	d = aoedev_by_aoeaddr(aoemajor, h->minor, 0);
1379 	if (d == NULL) {
1380 		snprintf(ebuf, sizeof ebuf, "aoecmd_ata_rsp: ata response "
1381 			"for unknown device %d.%d\n",
1382 			aoemajor, h->minor);
1383 		aoechr_error(ebuf);
1384 		return skb;
1385 	}
1386 
1387 	spin_lock_irqsave(&d->lock, flags);
1388 
1389 	n = be32_to_cpu(get_unaligned(&h->tag));
1390 	f = getframe(d, n);
1391 	if (f) {
1392 		calc_rttavg(d, f->t, tsince_hr(f));
1393 		f->t->nout--;
1394 		if (f->flags & FFL_PROBE)
1395 			f->t->nout_probes--;
1396 	} else {
1397 		f = getframe_deferred(d, n);
1398 		if (f) {
1399 			calc_rttavg(d, NULL, tsince_hr(f));
1400 		} else {
1401 			calc_rttavg(d, NULL, tsince(n));
1402 			spin_unlock_irqrestore(&d->lock, flags);
1403 			aoedev_put(d);
1404 			snprintf(ebuf, sizeof(ebuf),
1405 				 "%15s e%d.%d    tag=%08x@%08lx s=%pm d=%pm\n",
1406 				 "unexpected rsp",
1407 				 get_unaligned_be16(&h->major),
1408 				 h->minor,
1409 				 get_unaligned_be32(&h->tag),
1410 				 jiffies,
1411 				 h->src,
1412 				 h->dst);
1413 			aoechr_error(ebuf);
1414 			return skb;
1415 		}
1416 	}
1417 	aoecmd_work(d);
1418 
1419 	spin_unlock_irqrestore(&d->lock, flags);
1420 
1421 	ktcomplete(f, skb);
1422 
1423 	/*
1424 	 * Note here that we do not perform an aoedev_put, as we are
1425 	 * leaving this reference for the ktio to release.
1426 	 */
1427 	return NULL;
1428 }
1429 
1430 void
1431 aoecmd_cfg(ushort aoemajor, unsigned char aoeminor)
1432 {
1433 	struct sk_buff_head queue;
1434 
1435 	__skb_queue_head_init(&queue);
1436 	aoecmd_cfg_pkts(aoemajor, aoeminor, &queue);
1437 	aoenet_xmit(&queue);
1438 }
1439 
1440 struct sk_buff *
1441 aoecmd_ata_id(struct aoedev *d)
1442 {
1443 	struct aoe_hdr *h;
1444 	struct aoe_atahdr *ah;
1445 	struct frame *f;
1446 	struct sk_buff *skb;
1447 	struct aoetgt *t;
1448 
1449 	f = newframe(d);
1450 	if (f == NULL)
1451 		return NULL;
1452 
1453 	t = *d->tgt;
1454 
1455 	/* initialize the headers & frame */
1456 	skb = f->skb;
1457 	h = (struct aoe_hdr *) skb_mac_header(skb);
1458 	ah = (struct aoe_atahdr *) (h+1);
1459 	skb_put(skb, sizeof *h + sizeof *ah);
1460 	memset(h, 0, skb->len);
1461 	f->tag = aoehdr_atainit(d, t, h);
1462 	fhash(f);
1463 	t->nout++;
1464 	f->waited = 0;
1465 	f->waited_total = 0;
1466 
1467 	/* set up ata header */
1468 	ah->scnt = 1;
1469 	ah->cmdstat = ATA_CMD_ID_ATA;
1470 	ah->lba3 = 0xa0;
1471 
1472 	skb->dev = t->ifp->nd;
1473 
1474 	d->rttavg = RTTAVG_INIT;
1475 	d->rttdev = RTTDEV_INIT;
1476 	d->timer.function = rexmit_timer;
1477 
1478 	skb = skb_clone(skb, GFP_ATOMIC);
1479 	if (skb) {
1480 		do_gettimeofday(&f->sent);
1481 		f->sent_jiffs = (u32) jiffies;
1482 	}
1483 
1484 	return skb;
1485 }
1486 
1487 static struct aoetgt **
1488 grow_targets(struct aoedev *d)
1489 {
1490 	ulong oldn, newn;
1491 	struct aoetgt **tt;
1492 
1493 	oldn = d->ntargets;
1494 	newn = oldn * 2;
1495 	tt = kcalloc(newn, sizeof(*d->targets), GFP_ATOMIC);
1496 	if (!tt)
1497 		return NULL;
1498 	memmove(tt, d->targets, sizeof(*d->targets) * oldn);
1499 	d->tgt = tt + (d->tgt - d->targets);
1500 	kfree(d->targets);
1501 	d->targets = tt;
1502 	d->ntargets = newn;
1503 
1504 	return &d->targets[oldn];
1505 }
1506 
1507 static struct aoetgt *
1508 addtgt(struct aoedev *d, char *addr, ulong nframes)
1509 {
1510 	struct aoetgt *t, **tt, **te;
1511 
1512 	tt = d->targets;
1513 	te = tt + d->ntargets;
1514 	for (; tt < te && *tt; tt++)
1515 		;
1516 
1517 	if (tt == te) {
1518 		tt = grow_targets(d);
1519 		if (!tt)
1520 			goto nomem;
1521 	}
1522 	t = kzalloc(sizeof(*t), GFP_ATOMIC);
1523 	if (!t)
1524 		goto nomem;
1525 	t->nframes = nframes;
1526 	t->d = d;
1527 	memcpy(t->addr, addr, sizeof t->addr);
1528 	t->ifp = t->ifs;
1529 	aoecmd_wreset(t);
1530 	t->maxout = t->nframes / 2;
1531 	INIT_LIST_HEAD(&t->ffree);
1532 	return *tt = t;
1533 
1534  nomem:
1535 	pr_info("aoe: cannot allocate memory to add target\n");
1536 	return NULL;
1537 }
1538 
1539 static void
1540 setdbcnt(struct aoedev *d)
1541 {
1542 	struct aoetgt **t, **e;
1543 	int bcnt = 0;
1544 
1545 	t = d->targets;
1546 	e = t + d->ntargets;
1547 	for (; t < e && *t; t++)
1548 		if (bcnt == 0 || bcnt > (*t)->minbcnt)
1549 			bcnt = (*t)->minbcnt;
1550 	if (bcnt != d->maxbcnt) {
1551 		d->maxbcnt = bcnt;
1552 		pr_info("aoe: e%ld.%d: setting %d byte data frames\n",
1553 			d->aoemajor, d->aoeminor, bcnt);
1554 	}
1555 }
1556 
1557 static void
1558 setifbcnt(struct aoetgt *t, struct net_device *nd, int bcnt)
1559 {
1560 	struct aoedev *d;
1561 	struct aoeif *p, *e;
1562 	int minbcnt;
1563 
1564 	d = t->d;
1565 	minbcnt = bcnt;
1566 	p = t->ifs;
1567 	e = p + NAOEIFS;
1568 	for (; p < e; p++) {
1569 		if (p->nd == NULL)
1570 			break;		/* end of the valid interfaces */
1571 		if (p->nd == nd) {
1572 			p->bcnt = bcnt;	/* we're updating */
1573 			nd = NULL;
1574 		} else if (minbcnt > p->bcnt)
1575 			minbcnt = p->bcnt; /* find the min interface */
1576 	}
1577 	if (nd) {
1578 		if (p == e) {
1579 			pr_err("aoe: device setifbcnt failure; too many interfaces.\n");
1580 			return;
1581 		}
1582 		dev_hold(nd);
1583 		p->nd = nd;
1584 		p->bcnt = bcnt;
1585 	}
1586 	t->minbcnt = minbcnt;
1587 	setdbcnt(d);
1588 }
1589 
1590 void
1591 aoecmd_cfg_rsp(struct sk_buff *skb)
1592 {
1593 	struct aoedev *d;
1594 	struct aoe_hdr *h;
1595 	struct aoe_cfghdr *ch;
1596 	struct aoetgt *t;
1597 	ulong flags, aoemajor;
1598 	struct sk_buff *sl;
1599 	struct sk_buff_head queue;
1600 	u16 n;
1601 
1602 	sl = NULL;
1603 	h = (struct aoe_hdr *) skb_mac_header(skb);
1604 	ch = (struct aoe_cfghdr *) (h+1);
1605 
1606 	/*
1607 	 * Enough people have their dip switches set backwards to
1608 	 * warrant a loud message for this special case.
1609 	 */
1610 	aoemajor = get_unaligned_be16(&h->major);
1611 	if (aoemajor == 0xfff) {
1612 		printk(KERN_ERR "aoe: Warning: shelf address is all ones.  "
1613 			"Check shelf dip switches.\n");
1614 		return;
1615 	}
1616 	if (aoemajor == 0xffff) {
1617 		pr_info("aoe: e%ld.%d: broadcast shelf number invalid\n",
1618 			aoemajor, (int) h->minor);
1619 		return;
1620 	}
1621 	if (h->minor == 0xff) {
1622 		pr_info("aoe: e%ld.%d: broadcast slot number invalid\n",
1623 			aoemajor, (int) h->minor);
1624 		return;
1625 	}
1626 
1627 	n = be16_to_cpu(ch->bufcnt);
1628 	if (n > aoe_maxout)	/* keep it reasonable */
1629 		n = aoe_maxout;
1630 
1631 	d = aoedev_by_aoeaddr(aoemajor, h->minor, 1);
1632 	if (d == NULL) {
1633 		pr_info("aoe: device allocation failure\n");
1634 		return;
1635 	}
1636 
1637 	spin_lock_irqsave(&d->lock, flags);
1638 
1639 	t = gettgt(d, h->src);
1640 	if (t) {
1641 		t->nframes = n;
1642 		if (n < t->maxout)
1643 			aoecmd_wreset(t);
1644 	} else {
1645 		t = addtgt(d, h->src, n);
1646 		if (!t)
1647 			goto bail;
1648 	}
1649 	n = skb->dev->mtu;
1650 	n -= sizeof(struct aoe_hdr) + sizeof(struct aoe_atahdr);
1651 	n /= 512;
1652 	if (n > ch->scnt)
1653 		n = ch->scnt;
1654 	n = n ? n * 512 : DEFAULTBCNT;
1655 	setifbcnt(t, skb->dev, n);
1656 
1657 	/* don't change users' perspective */
1658 	if (d->nopen == 0) {
1659 		d->fw_ver = be16_to_cpu(ch->fwver);
1660 		sl = aoecmd_ata_id(d);
1661 	}
1662 bail:
1663 	spin_unlock_irqrestore(&d->lock, flags);
1664 	aoedev_put(d);
1665 	if (sl) {
1666 		__skb_queue_head_init(&queue);
1667 		__skb_queue_tail(&queue, sl);
1668 		aoenet_xmit(&queue);
1669 	}
1670 }
1671 
1672 void
1673 aoecmd_wreset(struct aoetgt *t)
1674 {
1675 	t->maxout = 1;
1676 	t->ssthresh = t->nframes / 2;
1677 	t->next_cwnd = t->nframes;
1678 }
1679 
1680 void
1681 aoecmd_cleanslate(struct aoedev *d)
1682 {
1683 	struct aoetgt **t, **te;
1684 
1685 	d->rttavg = RTTAVG_INIT;
1686 	d->rttdev = RTTDEV_INIT;
1687 	d->maxbcnt = 0;
1688 
1689 	t = d->targets;
1690 	te = t + d->ntargets;
1691 	for (; t < te && *t; t++)
1692 		aoecmd_wreset(*t);
1693 }
1694 
1695 void
1696 aoe_failbuf(struct aoedev *d, struct buf *buf)
1697 {
1698 	if (buf == NULL)
1699 		return;
1700 	buf->resid = 0;
1701 	clear_bit(BIO_UPTODATE, &buf->bio->bi_flags);
1702 	if (buf->nframesout == 0)
1703 		aoe_end_buf(d, buf);
1704 }
1705 
1706 void
1707 aoe_flush_iocq(void)
1708 {
1709 	struct frame *f;
1710 	struct aoedev *d;
1711 	LIST_HEAD(flist);
1712 	struct list_head *pos;
1713 	struct sk_buff *skb;
1714 	ulong flags;
1715 
1716 	spin_lock_irqsave(&iocq.lock, flags);
1717 	list_splice_init(&iocq.head, &flist);
1718 	spin_unlock_irqrestore(&iocq.lock, flags);
1719 	while (!list_empty(&flist)) {
1720 		pos = flist.next;
1721 		list_del(pos);
1722 		f = list_entry(pos, struct frame, head);
1723 		d = f->t->d;
1724 		skb = f->r_skb;
1725 		spin_lock_irqsave(&d->lock, flags);
1726 		if (f->buf) {
1727 			f->buf->nframesout--;
1728 			aoe_failbuf(d, f->buf);
1729 		}
1730 		aoe_freetframe(f);
1731 		spin_unlock_irqrestore(&d->lock, flags);
1732 		dev_kfree_skb(skb);
1733 		aoedev_put(d);
1734 	}
1735 }
1736 
1737 int __init
1738 aoecmd_init(void)
1739 {
1740 	void *p;
1741 
1742 	/* get_zeroed_page returns page with ref count 1 */
1743 	p = (void *) get_zeroed_page(GFP_KERNEL | __GFP_REPEAT);
1744 	if (!p)
1745 		return -ENOMEM;
1746 	empty_page = virt_to_page(p);
1747 
1748 	INIT_LIST_HEAD(&iocq.head);
1749 	spin_lock_init(&iocq.lock);
1750 	init_waitqueue_head(&ktiowq);
1751 	kts.name = "aoe_ktio";
1752 	kts.fn = ktio;
1753 	kts.waitq = &ktiowq;
1754 	kts.lock = &iocq.lock;
1755 	return aoe_ktstart(&kts);
1756 }
1757 
1758 void
1759 aoecmd_exit(void)
1760 {
1761 	aoe_ktstop(&kts);
1762 	aoe_flush_iocq();
1763 
1764 	free_page((unsigned long) page_address(empty_page));
1765 	empty_page = NULL;
1766 }
1767