xref: /openbmc/linux/drivers/block/aoe/aoecmd.c (revision 6f4eaea2)
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 
894 	if (d->flags & DEVFL_GDALLOC)
895 		aoeblk_gdalloc(d);
896 
897 	if (d->flags & DEVFL_NEWSIZE) {
898 		set_capacity_and_notify(d->gd, d->ssize);
899 
900 		spin_lock_irq(&d->lock);
901 		d->flags |= DEVFL_UP;
902 		d->flags &= ~DEVFL_NEWSIZE;
903 		spin_unlock_irq(&d->lock);
904 	}
905 }
906 
907 static void
908 ata_ident_fixstring(u16 *id, int ns)
909 {
910 	u16 s;
911 
912 	while (ns-- > 0) {
913 		s = *id;
914 		*id++ = s >> 8 | s << 8;
915 	}
916 }
917 
918 static void
919 ataid_complete(struct aoedev *d, struct aoetgt *t, unsigned char *id)
920 {
921 	u64 ssize;
922 	u16 n;
923 
924 	/* word 83: command set supported */
925 	n = get_unaligned_le16(&id[83 << 1]);
926 
927 	/* word 86: command set/feature enabled */
928 	n |= get_unaligned_le16(&id[86 << 1]);
929 
930 	if (n & (1<<10)) {	/* bit 10: LBA 48 */
931 		d->flags |= DEVFL_EXT;
932 
933 		/* word 100: number lba48 sectors */
934 		ssize = get_unaligned_le64(&id[100 << 1]);
935 
936 		/* set as in ide-disk.c:init_idedisk_capacity */
937 		d->geo.cylinders = ssize;
938 		d->geo.cylinders /= (255 * 63);
939 		d->geo.heads = 255;
940 		d->geo.sectors = 63;
941 	} else {
942 		d->flags &= ~DEVFL_EXT;
943 
944 		/* number lba28 sectors */
945 		ssize = get_unaligned_le32(&id[60 << 1]);
946 
947 		/* NOTE: obsolete in ATA 6 */
948 		d->geo.cylinders = get_unaligned_le16(&id[54 << 1]);
949 		d->geo.heads = get_unaligned_le16(&id[55 << 1]);
950 		d->geo.sectors = get_unaligned_le16(&id[56 << 1]);
951 	}
952 
953 	ata_ident_fixstring((u16 *) &id[10<<1], 10);	/* serial */
954 	ata_ident_fixstring((u16 *) &id[23<<1], 4);	/* firmware */
955 	ata_ident_fixstring((u16 *) &id[27<<1], 20);	/* model */
956 	memcpy(d->ident, id, sizeof(d->ident));
957 
958 	if (d->ssize != ssize)
959 		printk(KERN_INFO
960 			"aoe: %pm e%ld.%d v%04x has %llu sectors\n",
961 			t->addr,
962 			d->aoemajor, d->aoeminor,
963 			d->fw_ver, (long long)ssize);
964 	d->ssize = ssize;
965 	d->geo.start = 0;
966 	if (d->flags & (DEVFL_GDALLOC|DEVFL_NEWSIZE))
967 		return;
968 	if (d->gd != NULL)
969 		d->flags |= DEVFL_NEWSIZE;
970 	else
971 		d->flags |= DEVFL_GDALLOC;
972 	schedule_work(&d->work);
973 }
974 
975 static void
976 calc_rttavg(struct aoedev *d, struct aoetgt *t, int rtt)
977 {
978 	register long n;
979 
980 	n = rtt;
981 
982 	/* cf. Congestion Avoidance and Control, Jacobson & Karels, 1988 */
983 	n -= d->rttavg >> RTTSCALE;
984 	d->rttavg += n;
985 	if (n < 0)
986 		n = -n;
987 	n -= d->rttdev >> RTTDSCALE;
988 	d->rttdev += n;
989 
990 	if (!t || t->maxout >= t->nframes)
991 		return;
992 	if (t->maxout < t->ssthresh)
993 		t->maxout += 1;
994 	else if (t->nout == t->maxout && t->next_cwnd-- == 0) {
995 		t->maxout += 1;
996 		t->next_cwnd = t->maxout;
997 	}
998 }
999 
1000 static struct aoetgt *
1001 gettgt(struct aoedev *d, char *addr)
1002 {
1003 	struct aoetgt **t, **e;
1004 
1005 	t = d->targets;
1006 	e = t + d->ntargets;
1007 	for (; t < e && *t; t++)
1008 		if (memcmp((*t)->addr, addr, sizeof((*t)->addr)) == 0)
1009 			return *t;
1010 	return NULL;
1011 }
1012 
1013 static void
1014 bvcpy(struct sk_buff *skb, struct bio *bio, struct bvec_iter iter, long cnt)
1015 {
1016 	int soff = 0;
1017 	struct bio_vec bv;
1018 
1019 	iter.bi_size = cnt;
1020 
1021 	__bio_for_each_segment(bv, bio, iter, iter) {
1022 		char *p = kmap_atomic(bv.bv_page) + bv.bv_offset;
1023 		skb_copy_bits(skb, soff, p, bv.bv_len);
1024 		kunmap_atomic(p);
1025 		soff += bv.bv_len;
1026 	}
1027 }
1028 
1029 void
1030 aoe_end_request(struct aoedev *d, struct request *rq, int fastfail)
1031 {
1032 	struct bio *bio;
1033 	int bok;
1034 	struct request_queue *q;
1035 	blk_status_t err = BLK_STS_OK;
1036 
1037 	q = d->blkq;
1038 	if (rq == d->ip.rq)
1039 		d->ip.rq = NULL;
1040 	do {
1041 		bio = rq->bio;
1042 		bok = !fastfail && !bio->bi_status;
1043 		if (!bok)
1044 			err = BLK_STS_IOERR;
1045 	} while (blk_update_request(rq, bok ? BLK_STS_OK : BLK_STS_IOERR, bio->bi_iter.bi_size));
1046 
1047 	__blk_mq_end_request(rq, err);
1048 
1049 	/* cf. https://lore.kernel.org/lkml/20061031071040.GS14055@kernel.dk/ */
1050 	if (!fastfail)
1051 		blk_mq_run_hw_queues(q, true);
1052 }
1053 
1054 static void
1055 aoe_end_buf(struct aoedev *d, struct buf *buf)
1056 {
1057 	struct request *rq = buf->rq;
1058 	struct aoe_req *req = blk_mq_rq_to_pdu(rq);
1059 
1060 	if (buf == d->ip.buf)
1061 		d->ip.buf = NULL;
1062 	mempool_free(buf, d->bufpool);
1063 	if (--req->nr_bios == 0)
1064 		aoe_end_request(d, rq, 0);
1065 }
1066 
1067 static void
1068 ktiocomplete(struct frame *f)
1069 {
1070 	struct aoe_hdr *hin, *hout;
1071 	struct aoe_atahdr *ahin, *ahout;
1072 	struct buf *buf;
1073 	struct sk_buff *skb;
1074 	struct aoetgt *t;
1075 	struct aoeif *ifp;
1076 	struct aoedev *d;
1077 	long n;
1078 	int untainted;
1079 
1080 	if (f == NULL)
1081 		return;
1082 
1083 	t = f->t;
1084 	d = t->d;
1085 	skb = f->r_skb;
1086 	buf = f->buf;
1087 	if (f->flags & FFL_PROBE)
1088 		goto out;
1089 	if (!skb)		/* just fail the buf. */
1090 		goto noskb;
1091 
1092 	hout = (struct aoe_hdr *) skb_mac_header(f->skb);
1093 	ahout = (struct aoe_atahdr *) (hout+1);
1094 
1095 	hin = (struct aoe_hdr *) skb->data;
1096 	skb_pull(skb, sizeof(*hin));
1097 	ahin = (struct aoe_atahdr *) skb->data;
1098 	skb_pull(skb, sizeof(*ahin));
1099 	if (ahin->cmdstat & 0xa9) {	/* these bits cleared on success */
1100 		pr_err("aoe: ata error cmd=%2.2Xh stat=%2.2Xh from e%ld.%d\n",
1101 			ahout->cmdstat, ahin->cmdstat,
1102 			d->aoemajor, d->aoeminor);
1103 noskb:		if (buf)
1104 			buf->bio->bi_status = BLK_STS_IOERR;
1105 		goto out;
1106 	}
1107 
1108 	n = ahout->scnt << 9;
1109 	switch (ahout->cmdstat) {
1110 	case ATA_CMD_PIO_READ:
1111 	case ATA_CMD_PIO_READ_EXT:
1112 		if (skb->len < n) {
1113 			pr_err("%s e%ld.%d.  skb->len=%d need=%ld\n",
1114 				"aoe: runt data size in read from",
1115 				(long) d->aoemajor, d->aoeminor,
1116 			       skb->len, n);
1117 			buf->bio->bi_status = BLK_STS_IOERR;
1118 			break;
1119 		}
1120 		if (n > f->iter.bi_size) {
1121 			pr_err_ratelimited("%s e%ld.%d.  bytes=%ld need=%u\n",
1122 				"aoe: too-large data size in read from",
1123 				(long) d->aoemajor, d->aoeminor,
1124 				n, f->iter.bi_size);
1125 			buf->bio->bi_status = BLK_STS_IOERR;
1126 			break;
1127 		}
1128 		bvcpy(skb, f->buf->bio, f->iter, n);
1129 		fallthrough;
1130 	case ATA_CMD_PIO_WRITE:
1131 	case ATA_CMD_PIO_WRITE_EXT:
1132 		spin_lock_irq(&d->lock);
1133 		ifp = getif(t, skb->dev);
1134 		if (ifp)
1135 			ifp->lost = 0;
1136 		spin_unlock_irq(&d->lock);
1137 		break;
1138 	case ATA_CMD_ID_ATA:
1139 		if (skb->len < 512) {
1140 			pr_info("%s e%ld.%d.  skb->len=%d need=512\n",
1141 				"aoe: runt data size in ataid from",
1142 				(long) d->aoemajor, d->aoeminor,
1143 				skb->len);
1144 			break;
1145 		}
1146 		if (skb_linearize(skb))
1147 			break;
1148 		spin_lock_irq(&d->lock);
1149 		ataid_complete(d, t, skb->data);
1150 		spin_unlock_irq(&d->lock);
1151 		break;
1152 	default:
1153 		pr_info("aoe: unrecognized ata command %2.2Xh for %d.%d\n",
1154 			ahout->cmdstat,
1155 			be16_to_cpu(get_unaligned(&hin->major)),
1156 			hin->minor);
1157 	}
1158 out:
1159 	spin_lock_irq(&d->lock);
1160 	if (t->taint > 0
1161 	&& --t->taint > 0
1162 	&& t->nout_probes == 0) {
1163 		count_targets(d, &untainted);
1164 		if (untainted > 0) {
1165 			probe(t);
1166 			t->nout_probes++;
1167 		}
1168 	}
1169 
1170 	aoe_freetframe(f);
1171 
1172 	if (buf && --buf->nframesout == 0 && buf->iter.bi_size == 0)
1173 		aoe_end_buf(d, buf);
1174 
1175 	spin_unlock_irq(&d->lock);
1176 	aoedev_put(d);
1177 	dev_kfree_skb(skb);
1178 }
1179 
1180 /* Enters with iocq.lock held.
1181  * Returns true iff responses needing processing remain.
1182  */
1183 static int
1184 ktio(int id)
1185 {
1186 	struct frame *f;
1187 	struct list_head *pos;
1188 	int i;
1189 	int actual_id;
1190 
1191 	for (i = 0; ; ++i) {
1192 		if (i == MAXIOC)
1193 			return 1;
1194 		if (list_empty(&iocq[id].head))
1195 			return 0;
1196 		pos = iocq[id].head.next;
1197 		list_del(pos);
1198 		f = list_entry(pos, struct frame, head);
1199 		spin_unlock_irq(&iocq[id].lock);
1200 		ktiocomplete(f);
1201 
1202 		/* Figure out if extra threads are required. */
1203 		actual_id = f->t->d->aoeminor % ncpus;
1204 
1205 		if (!kts[actual_id].active) {
1206 			BUG_ON(id != 0);
1207 			mutex_lock(&ktio_spawn_lock);
1208 			if (!kts[actual_id].active
1209 				&& aoe_ktstart(&kts[actual_id]) == 0)
1210 				kts[actual_id].active = 1;
1211 			mutex_unlock(&ktio_spawn_lock);
1212 		}
1213 		spin_lock_irq(&iocq[id].lock);
1214 	}
1215 }
1216 
1217 static int
1218 kthread(void *vp)
1219 {
1220 	struct ktstate *k;
1221 	DECLARE_WAITQUEUE(wait, current);
1222 	int more;
1223 
1224 	k = vp;
1225 	current->flags |= PF_NOFREEZE;
1226 	set_user_nice(current, -10);
1227 	complete(&k->rendez);	/* tell spawner we're running */
1228 	do {
1229 		spin_lock_irq(k->lock);
1230 		more = k->fn(k->id);
1231 		if (!more) {
1232 			add_wait_queue(k->waitq, &wait);
1233 			__set_current_state(TASK_INTERRUPTIBLE);
1234 		}
1235 		spin_unlock_irq(k->lock);
1236 		if (!more) {
1237 			schedule();
1238 			remove_wait_queue(k->waitq, &wait);
1239 		} else
1240 			cond_resched();
1241 	} while (!kthread_should_stop());
1242 	complete(&k->rendez);	/* tell spawner we're stopping */
1243 	return 0;
1244 }
1245 
1246 void
1247 aoe_ktstop(struct ktstate *k)
1248 {
1249 	kthread_stop(k->task);
1250 	wait_for_completion(&k->rendez);
1251 }
1252 
1253 int
1254 aoe_ktstart(struct ktstate *k)
1255 {
1256 	struct task_struct *task;
1257 
1258 	init_completion(&k->rendez);
1259 	task = kthread_run(kthread, k, "%s", k->name);
1260 	if (task == NULL || IS_ERR(task))
1261 		return -ENOMEM;
1262 	k->task = task;
1263 	wait_for_completion(&k->rendez); /* allow kthread to start */
1264 	init_completion(&k->rendez);	/* for waiting for exit later */
1265 	return 0;
1266 }
1267 
1268 /* pass it off to kthreads for processing */
1269 static void
1270 ktcomplete(struct frame *f, struct sk_buff *skb)
1271 {
1272 	int id;
1273 	ulong flags;
1274 
1275 	f->r_skb = skb;
1276 	id = f->t->d->aoeminor % ncpus;
1277 	spin_lock_irqsave(&iocq[id].lock, flags);
1278 	if (!kts[id].active) {
1279 		spin_unlock_irqrestore(&iocq[id].lock, flags);
1280 		/* The thread with id has not been spawned yet,
1281 		 * so delegate the work to the main thread and
1282 		 * try spawning a new thread.
1283 		 */
1284 		id = 0;
1285 		spin_lock_irqsave(&iocq[id].lock, flags);
1286 	}
1287 	list_add_tail(&f->head, &iocq[id].head);
1288 	spin_unlock_irqrestore(&iocq[id].lock, flags);
1289 	wake_up(&ktiowq[id]);
1290 }
1291 
1292 struct sk_buff *
1293 aoecmd_ata_rsp(struct sk_buff *skb)
1294 {
1295 	struct aoedev *d;
1296 	struct aoe_hdr *h;
1297 	struct frame *f;
1298 	u32 n;
1299 	ulong flags;
1300 	char ebuf[128];
1301 	u16 aoemajor;
1302 
1303 	h = (struct aoe_hdr *) skb->data;
1304 	aoemajor = be16_to_cpu(get_unaligned(&h->major));
1305 	d = aoedev_by_aoeaddr(aoemajor, h->minor, 0);
1306 	if (d == NULL) {
1307 		snprintf(ebuf, sizeof ebuf, "aoecmd_ata_rsp: ata response "
1308 			"for unknown device %d.%d\n",
1309 			aoemajor, h->minor);
1310 		aoechr_error(ebuf);
1311 		return skb;
1312 	}
1313 
1314 	spin_lock_irqsave(&d->lock, flags);
1315 
1316 	n = be32_to_cpu(get_unaligned(&h->tag));
1317 	f = getframe(d, n);
1318 	if (f) {
1319 		calc_rttavg(d, f->t, tsince_hr(f));
1320 		f->t->nout--;
1321 		if (f->flags & FFL_PROBE)
1322 			f->t->nout_probes--;
1323 	} else {
1324 		f = getframe_deferred(d, n);
1325 		if (f) {
1326 			calc_rttavg(d, NULL, tsince_hr(f));
1327 		} else {
1328 			calc_rttavg(d, NULL, tsince(n));
1329 			spin_unlock_irqrestore(&d->lock, flags);
1330 			aoedev_put(d);
1331 			snprintf(ebuf, sizeof(ebuf),
1332 				 "%15s e%d.%d    tag=%08x@%08lx s=%pm d=%pm\n",
1333 				 "unexpected rsp",
1334 				 get_unaligned_be16(&h->major),
1335 				 h->minor,
1336 				 get_unaligned_be32(&h->tag),
1337 				 jiffies,
1338 				 h->src,
1339 				 h->dst);
1340 			aoechr_error(ebuf);
1341 			return skb;
1342 		}
1343 	}
1344 	aoecmd_work(d);
1345 
1346 	spin_unlock_irqrestore(&d->lock, flags);
1347 
1348 	ktcomplete(f, skb);
1349 
1350 	/*
1351 	 * Note here that we do not perform an aoedev_put, as we are
1352 	 * leaving this reference for the ktio to release.
1353 	 */
1354 	return NULL;
1355 }
1356 
1357 void
1358 aoecmd_cfg(ushort aoemajor, unsigned char aoeminor)
1359 {
1360 	struct sk_buff_head queue;
1361 
1362 	__skb_queue_head_init(&queue);
1363 	aoecmd_cfg_pkts(aoemajor, aoeminor, &queue);
1364 	aoenet_xmit(&queue);
1365 }
1366 
1367 struct sk_buff *
1368 aoecmd_ata_id(struct aoedev *d)
1369 {
1370 	struct aoe_hdr *h;
1371 	struct aoe_atahdr *ah;
1372 	struct frame *f;
1373 	struct sk_buff *skb;
1374 	struct aoetgt *t;
1375 
1376 	f = newframe(d);
1377 	if (f == NULL)
1378 		return NULL;
1379 
1380 	t = *d->tgt;
1381 
1382 	/* initialize the headers & frame */
1383 	skb = f->skb;
1384 	h = (struct aoe_hdr *) skb_mac_header(skb);
1385 	ah = (struct aoe_atahdr *) (h+1);
1386 	skb_put(skb, sizeof *h + sizeof *ah);
1387 	memset(h, 0, skb->len);
1388 	f->tag = aoehdr_atainit(d, t, h);
1389 	fhash(f);
1390 	t->nout++;
1391 	f->waited = 0;
1392 	f->waited_total = 0;
1393 
1394 	/* set up ata header */
1395 	ah->scnt = 1;
1396 	ah->cmdstat = ATA_CMD_ID_ATA;
1397 	ah->lba3 = 0xa0;
1398 
1399 	skb->dev = t->ifp->nd;
1400 
1401 	d->rttavg = RTTAVG_INIT;
1402 	d->rttdev = RTTDEV_INIT;
1403 	d->timer.function = rexmit_timer;
1404 
1405 	skb = skb_clone(skb, GFP_ATOMIC);
1406 	if (skb)
1407 		f->sent = ktime_get();
1408 
1409 	return skb;
1410 }
1411 
1412 static struct aoetgt **
1413 grow_targets(struct aoedev *d)
1414 {
1415 	ulong oldn, newn;
1416 	struct aoetgt **tt;
1417 
1418 	oldn = d->ntargets;
1419 	newn = oldn * 2;
1420 	tt = kcalloc(newn, sizeof(*d->targets), GFP_ATOMIC);
1421 	if (!tt)
1422 		return NULL;
1423 	memmove(tt, d->targets, sizeof(*d->targets) * oldn);
1424 	d->tgt = tt + (d->tgt - d->targets);
1425 	kfree(d->targets);
1426 	d->targets = tt;
1427 	d->ntargets = newn;
1428 
1429 	return &d->targets[oldn];
1430 }
1431 
1432 static struct aoetgt *
1433 addtgt(struct aoedev *d, char *addr, ulong nframes)
1434 {
1435 	struct aoetgt *t, **tt, **te;
1436 
1437 	tt = d->targets;
1438 	te = tt + d->ntargets;
1439 	for (; tt < te && *tt; tt++)
1440 		;
1441 
1442 	if (tt == te) {
1443 		tt = grow_targets(d);
1444 		if (!tt)
1445 			goto nomem;
1446 	}
1447 	t = kzalloc(sizeof(*t), GFP_ATOMIC);
1448 	if (!t)
1449 		goto nomem;
1450 	t->nframes = nframes;
1451 	t->d = d;
1452 	memcpy(t->addr, addr, sizeof t->addr);
1453 	t->ifp = t->ifs;
1454 	aoecmd_wreset(t);
1455 	t->maxout = t->nframes / 2;
1456 	INIT_LIST_HEAD(&t->ffree);
1457 	return *tt = t;
1458 
1459  nomem:
1460 	pr_info("aoe: cannot allocate memory to add target\n");
1461 	return NULL;
1462 }
1463 
1464 static void
1465 setdbcnt(struct aoedev *d)
1466 {
1467 	struct aoetgt **t, **e;
1468 	int bcnt = 0;
1469 
1470 	t = d->targets;
1471 	e = t + d->ntargets;
1472 	for (; t < e && *t; t++)
1473 		if (bcnt == 0 || bcnt > (*t)->minbcnt)
1474 			bcnt = (*t)->minbcnt;
1475 	if (bcnt != d->maxbcnt) {
1476 		d->maxbcnt = bcnt;
1477 		pr_info("aoe: e%ld.%d: setting %d byte data frames\n",
1478 			d->aoemajor, d->aoeminor, bcnt);
1479 	}
1480 }
1481 
1482 static void
1483 setifbcnt(struct aoetgt *t, struct net_device *nd, int bcnt)
1484 {
1485 	struct aoedev *d;
1486 	struct aoeif *p, *e;
1487 	int minbcnt;
1488 
1489 	d = t->d;
1490 	minbcnt = bcnt;
1491 	p = t->ifs;
1492 	e = p + NAOEIFS;
1493 	for (; p < e; p++) {
1494 		if (p->nd == NULL)
1495 			break;		/* end of the valid interfaces */
1496 		if (p->nd == nd) {
1497 			p->bcnt = bcnt;	/* we're updating */
1498 			nd = NULL;
1499 		} else if (minbcnt > p->bcnt)
1500 			minbcnt = p->bcnt; /* find the min interface */
1501 	}
1502 	if (nd) {
1503 		if (p == e) {
1504 			pr_err("aoe: device setifbcnt failure; too many interfaces.\n");
1505 			return;
1506 		}
1507 		dev_hold(nd);
1508 		p->nd = nd;
1509 		p->bcnt = bcnt;
1510 	}
1511 	t->minbcnt = minbcnt;
1512 	setdbcnt(d);
1513 }
1514 
1515 void
1516 aoecmd_cfg_rsp(struct sk_buff *skb)
1517 {
1518 	struct aoedev *d;
1519 	struct aoe_hdr *h;
1520 	struct aoe_cfghdr *ch;
1521 	struct aoetgt *t;
1522 	ulong flags, aoemajor;
1523 	struct sk_buff *sl;
1524 	struct sk_buff_head queue;
1525 	u16 n;
1526 
1527 	sl = NULL;
1528 	h = (struct aoe_hdr *) skb_mac_header(skb);
1529 	ch = (struct aoe_cfghdr *) (h+1);
1530 
1531 	/*
1532 	 * Enough people have their dip switches set backwards to
1533 	 * warrant a loud message for this special case.
1534 	 */
1535 	aoemajor = get_unaligned_be16(&h->major);
1536 	if (aoemajor == 0xfff) {
1537 		printk(KERN_ERR "aoe: Warning: shelf address is all ones.  "
1538 			"Check shelf dip switches.\n");
1539 		return;
1540 	}
1541 	if (aoemajor == 0xffff) {
1542 		pr_info("aoe: e%ld.%d: broadcast shelf number invalid\n",
1543 			aoemajor, (int) h->minor);
1544 		return;
1545 	}
1546 	if (h->minor == 0xff) {
1547 		pr_info("aoe: e%ld.%d: broadcast slot number invalid\n",
1548 			aoemajor, (int) h->minor);
1549 		return;
1550 	}
1551 
1552 	n = be16_to_cpu(ch->bufcnt);
1553 	if (n > aoe_maxout)	/* keep it reasonable */
1554 		n = aoe_maxout;
1555 
1556 	d = aoedev_by_aoeaddr(aoemajor, h->minor, 1);
1557 	if (d == NULL) {
1558 		pr_info("aoe: device allocation failure\n");
1559 		return;
1560 	}
1561 
1562 	spin_lock_irqsave(&d->lock, flags);
1563 
1564 	t = gettgt(d, h->src);
1565 	if (t) {
1566 		t->nframes = n;
1567 		if (n < t->maxout)
1568 			aoecmd_wreset(t);
1569 	} else {
1570 		t = addtgt(d, h->src, n);
1571 		if (!t)
1572 			goto bail;
1573 	}
1574 	n = skb->dev->mtu;
1575 	n -= sizeof(struct aoe_hdr) + sizeof(struct aoe_atahdr);
1576 	n /= 512;
1577 	if (n > ch->scnt)
1578 		n = ch->scnt;
1579 	n = n ? n * 512 : DEFAULTBCNT;
1580 	setifbcnt(t, skb->dev, n);
1581 
1582 	/* don't change users' perspective */
1583 	if (d->nopen == 0) {
1584 		d->fw_ver = be16_to_cpu(ch->fwver);
1585 		sl = aoecmd_ata_id(d);
1586 	}
1587 bail:
1588 	spin_unlock_irqrestore(&d->lock, flags);
1589 	aoedev_put(d);
1590 	if (sl) {
1591 		__skb_queue_head_init(&queue);
1592 		__skb_queue_tail(&queue, sl);
1593 		aoenet_xmit(&queue);
1594 	}
1595 }
1596 
1597 void
1598 aoecmd_wreset(struct aoetgt *t)
1599 {
1600 	t->maxout = 1;
1601 	t->ssthresh = t->nframes / 2;
1602 	t->next_cwnd = t->nframes;
1603 }
1604 
1605 void
1606 aoecmd_cleanslate(struct aoedev *d)
1607 {
1608 	struct aoetgt **t, **te;
1609 
1610 	d->rttavg = RTTAVG_INIT;
1611 	d->rttdev = RTTDEV_INIT;
1612 	d->maxbcnt = 0;
1613 
1614 	t = d->targets;
1615 	te = t + d->ntargets;
1616 	for (; t < te && *t; t++)
1617 		aoecmd_wreset(*t);
1618 }
1619 
1620 void
1621 aoe_failbuf(struct aoedev *d, struct buf *buf)
1622 {
1623 	if (buf == NULL)
1624 		return;
1625 	buf->iter.bi_size = 0;
1626 	buf->bio->bi_status = BLK_STS_IOERR;
1627 	if (buf->nframesout == 0)
1628 		aoe_end_buf(d, buf);
1629 }
1630 
1631 void
1632 aoe_flush_iocq(void)
1633 {
1634 	int i;
1635 
1636 	for (i = 0; i < ncpus; i++) {
1637 		if (kts[i].active)
1638 			aoe_flush_iocq_by_index(i);
1639 	}
1640 }
1641 
1642 void
1643 aoe_flush_iocq_by_index(int id)
1644 {
1645 	struct frame *f;
1646 	struct aoedev *d;
1647 	LIST_HEAD(flist);
1648 	struct list_head *pos;
1649 	struct sk_buff *skb;
1650 	ulong flags;
1651 
1652 	spin_lock_irqsave(&iocq[id].lock, flags);
1653 	list_splice_init(&iocq[id].head, &flist);
1654 	spin_unlock_irqrestore(&iocq[id].lock, flags);
1655 	while (!list_empty(&flist)) {
1656 		pos = flist.next;
1657 		list_del(pos);
1658 		f = list_entry(pos, struct frame, head);
1659 		d = f->t->d;
1660 		skb = f->r_skb;
1661 		spin_lock_irqsave(&d->lock, flags);
1662 		if (f->buf) {
1663 			f->buf->nframesout--;
1664 			aoe_failbuf(d, f->buf);
1665 		}
1666 		aoe_freetframe(f);
1667 		spin_unlock_irqrestore(&d->lock, flags);
1668 		dev_kfree_skb(skb);
1669 		aoedev_put(d);
1670 	}
1671 }
1672 
1673 int __init
1674 aoecmd_init(void)
1675 {
1676 	void *p;
1677 	int i;
1678 	int ret;
1679 
1680 	/* get_zeroed_page returns page with ref count 1 */
1681 	p = (void *) get_zeroed_page(GFP_KERNEL);
1682 	if (!p)
1683 		return -ENOMEM;
1684 	empty_page = virt_to_page(p);
1685 
1686 	ncpus = num_online_cpus();
1687 
1688 	iocq = kcalloc(ncpus, sizeof(struct iocq_ktio), GFP_KERNEL);
1689 	if (!iocq)
1690 		return -ENOMEM;
1691 
1692 	kts = kcalloc(ncpus, sizeof(struct ktstate), GFP_KERNEL);
1693 	if (!kts) {
1694 		ret = -ENOMEM;
1695 		goto kts_fail;
1696 	}
1697 
1698 	ktiowq = kcalloc(ncpus, sizeof(wait_queue_head_t), GFP_KERNEL);
1699 	if (!ktiowq) {
1700 		ret = -ENOMEM;
1701 		goto ktiowq_fail;
1702 	}
1703 
1704 	mutex_init(&ktio_spawn_lock);
1705 
1706 	for (i = 0; i < ncpus; i++) {
1707 		INIT_LIST_HEAD(&iocq[i].head);
1708 		spin_lock_init(&iocq[i].lock);
1709 		init_waitqueue_head(&ktiowq[i]);
1710 		snprintf(kts[i].name, sizeof(kts[i].name), "aoe_ktio%d", i);
1711 		kts[i].fn = ktio;
1712 		kts[i].waitq = &ktiowq[i];
1713 		kts[i].lock = &iocq[i].lock;
1714 		kts[i].id = i;
1715 		kts[i].active = 0;
1716 	}
1717 	kts[0].active = 1;
1718 	if (aoe_ktstart(&kts[0])) {
1719 		ret = -ENOMEM;
1720 		goto ktstart_fail;
1721 	}
1722 	return 0;
1723 
1724 ktstart_fail:
1725 	kfree(ktiowq);
1726 ktiowq_fail:
1727 	kfree(kts);
1728 kts_fail:
1729 	kfree(iocq);
1730 
1731 	return ret;
1732 }
1733 
1734 void
1735 aoecmd_exit(void)
1736 {
1737 	int i;
1738 
1739 	for (i = 0; i < ncpus; i++)
1740 		if (kts[i].active)
1741 			aoe_ktstop(&kts[i]);
1742 
1743 	aoe_flush_iocq();
1744 
1745 	/* Free up the iocq and thread speicific configuration
1746 	* allocated during startup.
1747 	*/
1748 	kfree(iocq);
1749 	kfree(kts);
1750 	kfree(ktiowq);
1751 
1752 	free_page((unsigned long) page_address(empty_page));
1753 	empty_page = NULL;
1754 }
1755