xref: /linux/drivers/block/aoe/aoecmd.c (revision da51bbcdbace8f43adf6066934c3926b656376e5)
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/moduleparam.h>
14 #include <linux/workqueue.h>
15 #include <linux/kthread.h>
16 #include <net/net_namespace.h>
17 #include <asm/unaligned.h>
18 #include <linux/uio.h>
19 #include "aoe.h"
20 
21 #define MAXIOC (8192)	/* default meant to avoid most soft lockups */
22 
23 static void ktcomplete(struct frame *, struct sk_buff *);
24 static int count_targets(struct aoedev *d, int *untainted);
25 
26 static struct buf *nextbuf(struct aoedev *);
27 
28 static int aoe_deadsecs = 60 * 3;
29 module_param(aoe_deadsecs, int, 0644);
30 MODULE_PARM_DESC(aoe_deadsecs, "After aoe_deadsecs seconds, give up and fail dev.");
31 
32 static int aoe_maxout = 64;
33 module_param(aoe_maxout, int, 0644);
34 MODULE_PARM_DESC(aoe_maxout,
35 	"Only aoe_maxout outstanding packets for every MAC on eX.Y.");
36 
37 /* The number of online cpus during module initialization gives us a
38  * convenient heuristic cap on the parallelism used for ktio threads
39  * doing I/O completion.  It is not important that the cap equal the
40  * actual number of running CPUs at any given time, but because of CPU
41  * hotplug, we take care to use ncpus instead of using
42  * num_online_cpus() after module initialization.
43  */
44 static int ncpus;
45 
46 /* mutex lock used for synchronization while thread spawning */
47 static DEFINE_MUTEX(ktio_spawn_lock);
48 
49 static wait_queue_head_t *ktiowq;
50 static struct ktstate *kts;
51 
52 /* io completion queue */
53 struct iocq_ktio {
54 	struct list_head head;
55 	spinlock_t lock;
56 };
57 static struct iocq_ktio *iocq;
58 
59 static struct page *empty_page;
60 
61 static struct sk_buff *
62 new_skb(ulong len)
63 {
64 	struct sk_buff *skb;
65 
66 	skb = alloc_skb(len + MAX_HEADER, GFP_ATOMIC);
67 	if (skb) {
68 		skb_reserve(skb, MAX_HEADER);
69 		skb_reset_mac_header(skb);
70 		skb_reset_network_header(skb);
71 		skb->protocol = __constant_htons(ETH_P_AOE);
72 		skb_checksum_none_assert(skb);
73 	}
74 	return skb;
75 }
76 
77 static struct frame *
78 getframe_deferred(struct aoedev *d, u32 tag)
79 {
80 	struct list_head *head, *pos, *nx;
81 	struct frame *f;
82 
83 	head = &d->rexmitq;
84 	list_for_each_safe(pos, nx, head) {
85 		f = list_entry(pos, struct frame, head);
86 		if (f->tag == tag) {
87 			list_del(pos);
88 			return f;
89 		}
90 	}
91 	return NULL;
92 }
93 
94 static struct frame *
95 getframe(struct aoedev *d, u32 tag)
96 {
97 	struct frame *f;
98 	struct list_head *head, *pos, *nx;
99 	u32 n;
100 
101 	n = tag % NFACTIVE;
102 	head = &d->factive[n];
103 	list_for_each_safe(pos, nx, head) {
104 		f = list_entry(pos, struct frame, head);
105 		if (f->tag == tag) {
106 			list_del(pos);
107 			return f;
108 		}
109 	}
110 	return NULL;
111 }
112 
113 /*
114  * Leave the top bit clear so we have tagspace for userland.
115  * The bottom 16 bits are the xmit tick for rexmit/rttavg processing.
116  * This driver reserves tag -1 to mean "unused frame."
117  */
118 static int
119 newtag(struct aoedev *d)
120 {
121 	register ulong n;
122 
123 	n = jiffies & 0xffff;
124 	return n | (++d->lasttag & 0x7fff) << 16;
125 }
126 
127 static u32
128 aoehdr_atainit(struct aoedev *d, struct aoetgt *t, struct aoe_hdr *h)
129 {
130 	u32 host_tag = newtag(d);
131 
132 	memcpy(h->src, t->ifp->nd->dev_addr, sizeof h->src);
133 	memcpy(h->dst, t->addr, sizeof h->dst);
134 	h->type = __constant_cpu_to_be16(ETH_P_AOE);
135 	h->verfl = AOE_HVER;
136 	h->major = cpu_to_be16(d->aoemajor);
137 	h->minor = d->aoeminor;
138 	h->cmd = AOECMD_ATA;
139 	h->tag = cpu_to_be32(host_tag);
140 
141 	return host_tag;
142 }
143 
144 static inline void
145 put_lba(struct aoe_atahdr *ah, sector_t lba)
146 {
147 	ah->lba0 = lba;
148 	ah->lba1 = lba >>= 8;
149 	ah->lba2 = lba >>= 8;
150 	ah->lba3 = lba >>= 8;
151 	ah->lba4 = lba >>= 8;
152 	ah->lba5 = lba >>= 8;
153 }
154 
155 static struct aoeif *
156 ifrotate(struct aoetgt *t)
157 {
158 	struct aoeif *ifp;
159 
160 	ifp = t->ifp;
161 	ifp++;
162 	if (ifp >= &t->ifs[NAOEIFS] || ifp->nd == NULL)
163 		ifp = t->ifs;
164 	if (ifp->nd == NULL)
165 		return NULL;
166 	return t->ifp = ifp;
167 }
168 
169 static void
170 skb_pool_put(struct aoedev *d, struct sk_buff *skb)
171 {
172 	__skb_queue_tail(&d->skbpool, skb);
173 }
174 
175 static struct sk_buff *
176 skb_pool_get(struct aoedev *d)
177 {
178 	struct sk_buff *skb = skb_peek(&d->skbpool);
179 
180 	if (skb && atomic_read(&skb_shinfo(skb)->dataref) == 1) {
181 		__skb_unlink(skb, &d->skbpool);
182 		return skb;
183 	}
184 	if (skb_queue_len(&d->skbpool) < NSKBPOOLMAX &&
185 	    (skb = new_skb(ETH_ZLEN)))
186 		return skb;
187 
188 	return NULL;
189 }
190 
191 void
192 aoe_freetframe(struct frame *f)
193 {
194 	struct aoetgt *t;
195 
196 	t = f->t;
197 	f->buf = NULL;
198 	memset(&f->iter, 0, sizeof(f->iter));
199 	f->r_skb = NULL;
200 	f->flags = 0;
201 	list_add(&f->head, &t->ffree);
202 }
203 
204 static struct frame *
205 newtframe(struct aoedev *d, struct aoetgt *t)
206 {
207 	struct frame *f;
208 	struct sk_buff *skb;
209 	struct list_head *pos;
210 
211 	if (list_empty(&t->ffree)) {
212 		if (t->falloc >= NSKBPOOLMAX*2)
213 			return NULL;
214 		f = kcalloc(1, sizeof(*f), GFP_ATOMIC);
215 		if (f == NULL)
216 			return NULL;
217 		t->falloc++;
218 		f->t = t;
219 	} else {
220 		pos = t->ffree.next;
221 		list_del(pos);
222 		f = list_entry(pos, struct frame, head);
223 	}
224 
225 	skb = f->skb;
226 	if (skb == NULL) {
227 		f->skb = skb = new_skb(ETH_ZLEN);
228 		if (!skb) {
229 bail:			aoe_freetframe(f);
230 			return NULL;
231 		}
232 	}
233 
234 	if (atomic_read(&skb_shinfo(skb)->dataref) != 1) {
235 		skb = skb_pool_get(d);
236 		if (skb == NULL)
237 			goto bail;
238 		skb_pool_put(d, f->skb);
239 		f->skb = skb;
240 	}
241 
242 	skb->truesize -= skb->data_len;
243 	skb_shinfo(skb)->nr_frags = skb->data_len = 0;
244 	skb_trim(skb, 0);
245 	return f;
246 }
247 
248 static struct frame *
249 newframe(struct aoedev *d)
250 {
251 	struct frame *f;
252 	struct aoetgt *t, **tt;
253 	int totout = 0;
254 	int use_tainted;
255 	int has_untainted;
256 
257 	if (!d->targets || !d->targets[0]) {
258 		printk(KERN_ERR "aoe: NULL TARGETS!\n");
259 		return NULL;
260 	}
261 	tt = d->tgt;	/* last used target */
262 	for (use_tainted = 0, has_untainted = 0;;) {
263 		tt++;
264 		if (tt >= &d->targets[d->ntargets] || !*tt)
265 			tt = d->targets;
266 		t = *tt;
267 		if (!t->taint) {
268 			has_untainted = 1;
269 			totout += t->nout;
270 		}
271 		if (t->nout < t->maxout
272 		&& (use_tainted || !t->taint)
273 		&& t->ifp->nd) {
274 			f = newtframe(d, t);
275 			if (f) {
276 				ifrotate(t);
277 				d->tgt = tt;
278 				return f;
279 			}
280 		}
281 		if (tt == d->tgt) {	/* we've looped and found nada */
282 			if (!use_tainted && !has_untainted)
283 				use_tainted = 1;
284 			else
285 				break;
286 		}
287 	}
288 	if (totout == 0) {
289 		d->kicked++;
290 		d->flags |= DEVFL_KICKME;
291 	}
292 	return NULL;
293 }
294 
295 static void
296 skb_fillup(struct sk_buff *skb, struct bio *bio, struct bvec_iter iter)
297 {
298 	int frag = 0;
299 	struct bio_vec bv;
300 
301 	__bio_for_each_segment(bv, bio, iter, iter)
302 		skb_fill_page_desc(skb, frag++, bv.bv_page,
303 				   bv.bv_offset, bv.bv_len);
304 }
305 
306 static void
307 fhash(struct frame *f)
308 {
309 	struct aoedev *d = f->t->d;
310 	u32 n;
311 
312 	n = f->tag % NFACTIVE;
313 	list_add_tail(&f->head, &d->factive[n]);
314 }
315 
316 static void
317 ata_rw_frameinit(struct frame *f)
318 {
319 	struct aoetgt *t;
320 	struct aoe_hdr *h;
321 	struct aoe_atahdr *ah;
322 	struct sk_buff *skb;
323 	char writebit, extbit;
324 
325 	skb = f->skb;
326 	h = (struct aoe_hdr *) skb_mac_header(skb);
327 	ah = (struct aoe_atahdr *) (h + 1);
328 	skb_put(skb, sizeof(*h) + sizeof(*ah));
329 	memset(h, 0, skb->len);
330 
331 	writebit = 0x10;
332 	extbit = 0x4;
333 
334 	t = f->t;
335 	f->tag = aoehdr_atainit(t->d, t, h);
336 	fhash(f);
337 	t->nout++;
338 	f->waited = 0;
339 	f->waited_total = 0;
340 
341 	/* set up ata header */
342 	ah->scnt = f->iter.bi_size >> 9;
343 	put_lba(ah, f->iter.bi_sector);
344 	if (t->d->flags & DEVFL_EXT) {
345 		ah->aflags |= AOEAFL_EXT;
346 	} else {
347 		extbit = 0;
348 		ah->lba3 &= 0x0f;
349 		ah->lba3 |= 0xe0;	/* LBA bit + obsolete 0xa0 */
350 	}
351 	if (f->buf && bio_data_dir(f->buf->bio) == WRITE) {
352 		skb_fillup(skb, f->buf->bio, f->iter);
353 		ah->aflags |= AOEAFL_WRITE;
354 		skb->len += f->iter.bi_size;
355 		skb->data_len = f->iter.bi_size;
356 		skb->truesize += f->iter.bi_size;
357 		t->wpkts++;
358 	} else {
359 		t->rpkts++;
360 		writebit = 0;
361 	}
362 
363 	ah->cmdstat = ATA_CMD_PIO_READ | writebit | extbit;
364 	skb->dev = t->ifp->nd;
365 }
366 
367 static int
368 aoecmd_ata_rw(struct aoedev *d)
369 {
370 	struct frame *f;
371 	struct buf *buf;
372 	struct sk_buff *skb;
373 	struct sk_buff_head queue;
374 
375 	buf = nextbuf(d);
376 	if (buf == NULL)
377 		return 0;
378 	f = newframe(d);
379 	if (f == NULL)
380 		return 0;
381 
382 	/* initialize the headers & frame */
383 	f->buf = buf;
384 	f->iter = buf->iter;
385 	f->iter.bi_size = min_t(unsigned long,
386 				d->maxbcnt ?: DEFAULTBCNT,
387 				f->iter.bi_size);
388 	bio_advance_iter(buf->bio, &buf->iter, f->iter.bi_size);
389 
390 	if (!buf->iter.bi_size)
391 		d->ip.buf = NULL;
392 
393 	/* mark all tracking fields and load out */
394 	buf->nframesout += 1;
395 
396 	ata_rw_frameinit(f);
397 
398 	skb = skb_clone(f->skb, GFP_ATOMIC);
399 	if (skb) {
400 		f->sent = ktime_get();
401 		__skb_queue_head_init(&queue);
402 		__skb_queue_tail(&queue, skb);
403 		aoenet_xmit(&queue);
404 	}
405 	return 1;
406 }
407 
408 /* some callers cannot sleep, and they can call this function,
409  * transmitting the packets later, when interrupts are on
410  */
411 static void
412 aoecmd_cfg_pkts(ushort aoemajor, unsigned char aoeminor, struct sk_buff_head *queue)
413 {
414 	struct aoe_hdr *h;
415 	struct aoe_cfghdr *ch;
416 	struct sk_buff *skb;
417 	struct net_device *ifp;
418 
419 	rcu_read_lock();
420 	for_each_netdev_rcu(&init_net, ifp) {
421 		dev_hold(ifp);
422 		if (!is_aoe_netif(ifp)) {
423 			dev_put(ifp);
424 			continue;
425 		}
426 
427 		skb = new_skb(sizeof *h + sizeof *ch);
428 		if (skb == NULL) {
429 			printk(KERN_INFO "aoe: skb alloc failure\n");
430 			dev_put(ifp);
431 			continue;
432 		}
433 		skb_put(skb, sizeof *h + sizeof *ch);
434 		skb->dev = ifp;
435 		__skb_queue_tail(queue, skb);
436 		h = (struct aoe_hdr *) skb_mac_header(skb);
437 		memset(h, 0, sizeof *h + sizeof *ch);
438 
439 		memset(h->dst, 0xff, sizeof h->dst);
440 		memcpy(h->src, ifp->dev_addr, sizeof h->src);
441 		h->type = __constant_cpu_to_be16(ETH_P_AOE);
442 		h->verfl = AOE_HVER;
443 		h->major = cpu_to_be16(aoemajor);
444 		h->minor = aoeminor;
445 		h->cmd = AOECMD_CFG;
446 	}
447 	rcu_read_unlock();
448 }
449 
450 static void
451 resend(struct aoedev *d, struct frame *f)
452 {
453 	struct sk_buff *skb;
454 	struct sk_buff_head queue;
455 	struct aoe_hdr *h;
456 	struct aoetgt *t;
457 	char buf[128];
458 	u32 n;
459 
460 	t = f->t;
461 	n = newtag(d);
462 	skb = f->skb;
463 	if (ifrotate(t) == NULL) {
464 		/* probably can't happen, but set it up to fail anyway */
465 		pr_info("aoe: resend: no interfaces to rotate to.\n");
466 		ktcomplete(f, NULL);
467 		return;
468 	}
469 	h = (struct aoe_hdr *) skb_mac_header(skb);
470 
471 	if (!(f->flags & FFL_PROBE)) {
472 		snprintf(buf, sizeof(buf),
473 			"%15s e%ld.%d oldtag=%08x@%08lx newtag=%08x s=%pm d=%pm nout=%d\n",
474 			"retransmit", d->aoemajor, d->aoeminor,
475 			f->tag, jiffies, n,
476 			h->src, h->dst, t->nout);
477 		aoechr_error(buf);
478 	}
479 
480 	f->tag = n;
481 	fhash(f);
482 	h->tag = cpu_to_be32(n);
483 	memcpy(h->dst, t->addr, sizeof h->dst);
484 	memcpy(h->src, t->ifp->nd->dev_addr, sizeof h->src);
485 
486 	skb->dev = t->ifp->nd;
487 	skb = skb_clone(skb, GFP_ATOMIC);
488 	if (skb == NULL)
489 		return;
490 	f->sent = ktime_get();
491 	__skb_queue_head_init(&queue);
492 	__skb_queue_tail(&queue, skb);
493 	aoenet_xmit(&queue);
494 }
495 
496 static int
497 tsince_hr(struct frame *f)
498 {
499 	u64 delta = ktime_to_ns(ktime_sub(ktime_get(), f->sent));
500 
501 	/* delta is normally under 4.2 seconds, avoid 64-bit division */
502 	if (likely(delta <= UINT_MAX))
503 		return (u32)delta / NSEC_PER_USEC;
504 
505 	/* avoid overflow after 71 minutes */
506 	if (delta > ((u64)INT_MAX * NSEC_PER_USEC))
507 		return INT_MAX;
508 
509 	return div_u64(delta, NSEC_PER_USEC);
510 }
511 
512 static int
513 tsince(u32 tag)
514 {
515 	int n;
516 
517 	n = jiffies & 0xffff;
518 	n -= tag & 0xffff;
519 	if (n < 0)
520 		n += 1<<16;
521 	return jiffies_to_usecs(n + 1);
522 }
523 
524 static struct aoeif *
525 getif(struct aoetgt *t, struct net_device *nd)
526 {
527 	struct aoeif *p, *e;
528 
529 	p = t->ifs;
530 	e = p + NAOEIFS;
531 	for (; p < e; p++)
532 		if (p->nd == nd)
533 			return p;
534 	return NULL;
535 }
536 
537 static void
538 ejectif(struct aoetgt *t, struct aoeif *ifp)
539 {
540 	struct aoeif *e;
541 	struct net_device *nd;
542 	ulong n;
543 
544 	nd = ifp->nd;
545 	e = t->ifs + NAOEIFS - 1;
546 	n = (e - ifp) * sizeof *ifp;
547 	memmove(ifp, ifp+1, n);
548 	e->nd = NULL;
549 	dev_put(nd);
550 }
551 
552 static struct frame *
553 reassign_frame(struct frame *f)
554 {
555 	struct frame *nf;
556 	struct sk_buff *skb;
557 
558 	nf = newframe(f->t->d);
559 	if (!nf)
560 		return NULL;
561 	if (nf->t == f->t) {
562 		aoe_freetframe(nf);
563 		return NULL;
564 	}
565 
566 	skb = nf->skb;
567 	nf->skb = f->skb;
568 	nf->buf = f->buf;
569 	nf->iter = f->iter;
570 	nf->waited = 0;
571 	nf->waited_total = f->waited_total;
572 	nf->sent = f->sent;
573 	f->skb = skb;
574 
575 	return nf;
576 }
577 
578 static void
579 probe(struct aoetgt *t)
580 {
581 	struct aoedev *d;
582 	struct frame *f;
583 	struct sk_buff *skb;
584 	struct sk_buff_head queue;
585 	size_t n, m;
586 	int frag;
587 
588 	d = t->d;
589 	f = newtframe(d, t);
590 	if (!f) {
591 		pr_err("%s %pm for e%ld.%d: %s\n",
592 			"aoe: cannot probe remote address",
593 			t->addr,
594 			(long) d->aoemajor, d->aoeminor,
595 			"no frame available");
596 		return;
597 	}
598 	f->flags |= FFL_PROBE;
599 	ifrotate(t);
600 	f->iter.bi_size = t->d->maxbcnt ? t->d->maxbcnt : DEFAULTBCNT;
601 	ata_rw_frameinit(f);
602 	skb = f->skb;
603 	for (frag = 0, n = f->iter.bi_size; n > 0; ++frag, n -= m) {
604 		if (n < PAGE_SIZE)
605 			m = n;
606 		else
607 			m = PAGE_SIZE;
608 		skb_fill_page_desc(skb, frag, empty_page, 0, m);
609 	}
610 	skb->len += f->iter.bi_size;
611 	skb->data_len = f->iter.bi_size;
612 	skb->truesize += f->iter.bi_size;
613 
614 	skb = skb_clone(f->skb, GFP_ATOMIC);
615 	if (skb) {
616 		f->sent = ktime_get();
617 		__skb_queue_head_init(&queue);
618 		__skb_queue_tail(&queue, skb);
619 		aoenet_xmit(&queue);
620 	}
621 }
622 
623 static long
624 rto(struct aoedev *d)
625 {
626 	long t;
627 
628 	t = 2 * d->rttavg >> RTTSCALE;
629 	t += 8 * d->rttdev >> RTTDSCALE;
630 	if (t == 0)
631 		t = 1;
632 
633 	return t;
634 }
635 
636 static void
637 rexmit_deferred(struct aoedev *d)
638 {
639 	struct aoetgt *t;
640 	struct frame *f;
641 	struct frame *nf;
642 	struct list_head *pos, *nx, *head;
643 	int since;
644 	int untainted;
645 
646 	count_targets(d, &untainted);
647 
648 	head = &d->rexmitq;
649 	list_for_each_safe(pos, nx, head) {
650 		f = list_entry(pos, struct frame, head);
651 		t = f->t;
652 		if (t->taint) {
653 			if (!(f->flags & FFL_PROBE)) {
654 				nf = reassign_frame(f);
655 				if (nf) {
656 					if (t->nout_probes == 0
657 					&& untainted > 0) {
658 						probe(t);
659 						t->nout_probes++;
660 					}
661 					list_replace(&f->head, &nf->head);
662 					pos = &nf->head;
663 					aoe_freetframe(f);
664 					f = nf;
665 					t = f->t;
666 				}
667 			} else if (untainted < 1) {
668 				/* don't probe w/o other untainted aoetgts */
669 				goto stop_probe;
670 			} else if (tsince_hr(f) < t->taint * rto(d)) {
671 				/* reprobe slowly when taint is high */
672 				continue;
673 			}
674 		} else if (f->flags & FFL_PROBE) {
675 stop_probe:		/* don't probe untainted aoetgts */
676 			list_del(pos);
677 			aoe_freetframe(f);
678 			/* leaving d->kicked, because this is routine */
679 			f->t->d->flags |= DEVFL_KICKME;
680 			continue;
681 		}
682 		if (t->nout >= t->maxout)
683 			continue;
684 		list_del(pos);
685 		t->nout++;
686 		if (f->flags & FFL_PROBE)
687 			t->nout_probes++;
688 		since = tsince_hr(f);
689 		f->waited += since;
690 		f->waited_total += since;
691 		resend(d, f);
692 	}
693 }
694 
695 /* An aoetgt accumulates demerits quickly, and successful
696  * probing redeems the aoetgt slowly.
697  */
698 static void
699 scorn(struct aoetgt *t)
700 {
701 	int n;
702 
703 	n = t->taint++;
704 	t->taint += t->taint * 2;
705 	if (n > t->taint)
706 		t->taint = n;
707 	if (t->taint > MAX_TAINT)
708 		t->taint = MAX_TAINT;
709 }
710 
711 static int
712 count_targets(struct aoedev *d, int *untainted)
713 {
714 	int i, good;
715 
716 	for (i = good = 0; i < d->ntargets && d->targets[i]; ++i)
717 		if (d->targets[i]->taint == 0)
718 			good++;
719 
720 	if (untainted)
721 		*untainted = good;
722 	return i;
723 }
724 
725 static void
726 rexmit_timer(struct timer_list *timer)
727 {
728 	struct aoedev *d;
729 	struct aoetgt *t;
730 	struct aoeif *ifp;
731 	struct frame *f;
732 	struct list_head *head, *pos, *nx;
733 	LIST_HEAD(flist);
734 	register long timeout;
735 	ulong flags, n;
736 	int i;
737 	int utgts;	/* number of aoetgt descriptors (not slots) */
738 	int since;
739 
740 	d = from_timer(d, timer, timer);
741 
742 	spin_lock_irqsave(&d->lock, flags);
743 
744 	/* timeout based on observed timings and variations */
745 	timeout = rto(d);
746 
747 	utgts = count_targets(d, NULL);
748 
749 	if (d->flags & DEVFL_TKILL) {
750 		spin_unlock_irqrestore(&d->lock, flags);
751 		return;
752 	}
753 
754 	/* collect all frames to rexmit into flist */
755 	for (i = 0; i < NFACTIVE; i++) {
756 		head = &d->factive[i];
757 		list_for_each_safe(pos, nx, head) {
758 			f = list_entry(pos, struct frame, head);
759 			if (tsince_hr(f) < timeout)
760 				break;	/* end of expired frames */
761 			/* move to flist for later processing */
762 			list_move_tail(pos, &flist);
763 		}
764 	}
765 
766 	/* process expired frames */
767 	while (!list_empty(&flist)) {
768 		pos = flist.next;
769 		f = list_entry(pos, struct frame, head);
770 		since = tsince_hr(f);
771 		n = f->waited_total + since;
772 		n /= USEC_PER_SEC;
773 		if (aoe_deadsecs
774 		&& n > aoe_deadsecs
775 		&& !(f->flags & FFL_PROBE)) {
776 			/* Waited too long.  Device failure.
777 			 * Hang all frames on first hash bucket for downdev
778 			 * to clean up.
779 			 */
780 			list_splice(&flist, &d->factive[0]);
781 			aoedev_downdev(d);
782 			goto out;
783 		}
784 
785 		t = f->t;
786 		n = f->waited + since;
787 		n /= USEC_PER_SEC;
788 		if (aoe_deadsecs && utgts > 0
789 		&& (n > aoe_deadsecs / utgts || n > HARD_SCORN_SECS))
790 			scorn(t); /* avoid this target */
791 
792 		if (t->maxout != 1) {
793 			t->ssthresh = t->maxout / 2;
794 			t->maxout = 1;
795 		}
796 
797 		if (f->flags & FFL_PROBE) {
798 			t->nout_probes--;
799 		} else {
800 			ifp = getif(t, f->skb->dev);
801 			if (ifp && ++ifp->lost > (t->nframes << 1)
802 			&& (ifp != t->ifs || t->ifs[1].nd)) {
803 				ejectif(t, ifp);
804 				ifp = NULL;
805 			}
806 		}
807 		list_move_tail(pos, &d->rexmitq);
808 		t->nout--;
809 	}
810 	rexmit_deferred(d);
811 
812 out:
813 	if ((d->flags & DEVFL_KICKME) && d->blkq) {
814 		d->flags &= ~DEVFL_KICKME;
815 		blk_mq_run_hw_queues(d->blkq, true);
816 	}
817 
818 	d->timer.expires = jiffies + TIMERTICK;
819 	add_timer(&d->timer);
820 
821 	spin_unlock_irqrestore(&d->lock, flags);
822 }
823 
824 static void
825 bufinit(struct buf *buf, struct request *rq, struct bio *bio)
826 {
827 	memset(buf, 0, sizeof(*buf));
828 	buf->rq = rq;
829 	buf->bio = bio;
830 	buf->iter = bio->bi_iter;
831 }
832 
833 static struct buf *
834 nextbuf(struct aoedev *d)
835 {
836 	struct request *rq;
837 	struct request_queue *q;
838 	struct aoe_req *req;
839 	struct buf *buf;
840 	struct bio *bio;
841 
842 	q = d->blkq;
843 	if (q == NULL)
844 		return NULL;	/* initializing */
845 	if (d->ip.buf)
846 		return d->ip.buf;
847 	rq = d->ip.rq;
848 	if (rq == NULL) {
849 		rq = list_first_entry_or_null(&d->rq_list, struct request,
850 						queuelist);
851 		if (rq == NULL)
852 			return NULL;
853 		list_del_init(&rq->queuelist);
854 		blk_mq_start_request(rq);
855 		d->ip.rq = rq;
856 		d->ip.nxbio = rq->bio;
857 
858 		req = blk_mq_rq_to_pdu(rq);
859 		req->nr_bios = 0;
860 		__rq_for_each_bio(bio, rq)
861 			req->nr_bios++;
862 	}
863 	buf = mempool_alloc(d->bufpool, GFP_ATOMIC);
864 	if (buf == NULL) {
865 		pr_err("aoe: nextbuf: unable to mempool_alloc!\n");
866 		return NULL;
867 	}
868 	bio = d->ip.nxbio;
869 	bufinit(buf, rq, bio);
870 	bio = bio->bi_next;
871 	d->ip.nxbio = bio;
872 	if (bio == NULL)
873 		d->ip.rq = NULL;
874 	return d->ip.buf = buf;
875 }
876 
877 /* enters with d->lock held */
878 void
879 aoecmd_work(struct aoedev *d)
880 {
881 	rexmit_deferred(d);
882 	while (aoecmd_ata_rw(d))
883 		;
884 }
885 
886 /* this function performs work that has been deferred until sleeping is OK
887  */
888 void
889 aoecmd_sleepwork(struct work_struct *work)
890 {
891 	struct aoedev *d = container_of(work, struct aoedev, work);
892 
893 	if (d->flags & DEVFL_GDALLOC)
894 		aoeblk_gdalloc(d);
895 
896 	if (d->flags & DEVFL_NEWSIZE) {
897 		set_capacity_and_notify(d->gd, d->ssize);
898 
899 		spin_lock_irq(&d->lock);
900 		d->flags |= DEVFL_UP;
901 		d->flags &= ~DEVFL_NEWSIZE;
902 		spin_unlock_irq(&d->lock);
903 	}
904 }
905 
906 static void
907 ata_ident_fixstring(u16 *id, int ns)
908 {
909 	u16 s;
910 
911 	while (ns-- > 0) {
912 		s = *id;
913 		*id++ = s >> 8 | s << 8;
914 	}
915 }
916 
917 static void
918 ataid_complete(struct aoedev *d, struct aoetgt *t, unsigned char *id)
919 {
920 	u64 ssize;
921 	u16 n;
922 
923 	/* word 83: command set supported */
924 	n = get_unaligned_le16(&id[83 << 1]);
925 
926 	/* word 86: command set/feature enabled */
927 	n |= get_unaligned_le16(&id[86 << 1]);
928 
929 	if (n & (1<<10)) {	/* bit 10: LBA 48 */
930 		d->flags |= DEVFL_EXT;
931 
932 		/* word 100: number lba48 sectors */
933 		ssize = get_unaligned_le64(&id[100 << 1]);
934 
935 		/* set as in ide-disk.c:init_idedisk_capacity */
936 		d->geo.cylinders = ssize;
937 		d->geo.cylinders /= (255 * 63);
938 		d->geo.heads = 255;
939 		d->geo.sectors = 63;
940 	} else {
941 		d->flags &= ~DEVFL_EXT;
942 
943 		/* number lba28 sectors */
944 		ssize = get_unaligned_le32(&id[60 << 1]);
945 
946 		/* NOTE: obsolete in ATA 6 */
947 		d->geo.cylinders = get_unaligned_le16(&id[54 << 1]);
948 		d->geo.heads = get_unaligned_le16(&id[55 << 1]);
949 		d->geo.sectors = get_unaligned_le16(&id[56 << 1]);
950 	}
951 
952 	ata_ident_fixstring((u16 *) &id[10<<1], 10);	/* serial */
953 	ata_ident_fixstring((u16 *) &id[23<<1], 4);	/* firmware */
954 	ata_ident_fixstring((u16 *) &id[27<<1], 20);	/* model */
955 	memcpy(d->ident, id, sizeof(d->ident));
956 
957 	if (d->ssize != ssize)
958 		printk(KERN_INFO
959 			"aoe: %pm e%ld.%d v%04x has %llu sectors\n",
960 			t->addr,
961 			d->aoemajor, d->aoeminor,
962 			d->fw_ver, (long long)ssize);
963 	d->ssize = ssize;
964 	d->geo.start = 0;
965 	if (d->flags & (DEVFL_GDALLOC|DEVFL_NEWSIZE))
966 		return;
967 	if (d->gd != NULL)
968 		d->flags |= DEVFL_NEWSIZE;
969 	else
970 		d->flags |= DEVFL_GDALLOC;
971 	queue_work(aoe_wq, &d->work);
972 }
973 
974 static void
975 calc_rttavg(struct aoedev *d, struct aoetgt *t, int rtt)
976 {
977 	register long n;
978 
979 	n = rtt;
980 
981 	/* cf. Congestion Avoidance and Control, Jacobson & Karels, 1988 */
982 	n -= d->rttavg >> RTTSCALE;
983 	d->rttavg += n;
984 	if (n < 0)
985 		n = -n;
986 	n -= d->rttdev >> RTTDSCALE;
987 	d->rttdev += n;
988 
989 	if (!t || t->maxout >= t->nframes)
990 		return;
991 	if (t->maxout < t->ssthresh)
992 		t->maxout += 1;
993 	else if (t->nout == t->maxout && t->next_cwnd-- == 0) {
994 		t->maxout += 1;
995 		t->next_cwnd = t->maxout;
996 	}
997 }
998 
999 static struct aoetgt *
1000 gettgt(struct aoedev *d, char *addr)
1001 {
1002 	struct aoetgt **t, **e;
1003 
1004 	t = d->targets;
1005 	e = t + d->ntargets;
1006 	for (; t < e && *t; t++)
1007 		if (memcmp((*t)->addr, addr, sizeof((*t)->addr)) == 0)
1008 			return *t;
1009 	return NULL;
1010 }
1011 
1012 static void
1013 bvcpy(struct sk_buff *skb, struct bio *bio, struct bvec_iter iter, long cnt)
1014 {
1015 	int soff = 0;
1016 	struct bio_vec bv;
1017 
1018 	iter.bi_size = cnt;
1019 
1020 	__bio_for_each_segment(bv, bio, iter, iter) {
1021 		char *p = bvec_kmap_local(&bv);
1022 		skb_copy_bits(skb, soff, p, bv.bv_len);
1023 		kunmap_local(p);
1024 		soff += bv.bv_len;
1025 	}
1026 }
1027 
1028 void
1029 aoe_end_request(struct aoedev *d, struct request *rq, int fastfail)
1030 {
1031 	struct bio *bio;
1032 	int bok;
1033 	struct request_queue *q;
1034 	blk_status_t err = BLK_STS_OK;
1035 
1036 	q = d->blkq;
1037 	if (rq == d->ip.rq)
1038 		d->ip.rq = NULL;
1039 	do {
1040 		bio = rq->bio;
1041 		bok = !fastfail && !bio->bi_status;
1042 		if (!bok)
1043 			err = BLK_STS_IOERR;
1044 	} while (blk_update_request(rq, bok ? BLK_STS_OK : BLK_STS_IOERR, bio->bi_iter.bi_size));
1045 
1046 	__blk_mq_end_request(rq, err);
1047 
1048 	/* cf. https://lore.kernel.org/lkml/20061031071040.GS14055@kernel.dk/ */
1049 	if (!fastfail)
1050 		blk_mq_run_hw_queues(q, true);
1051 }
1052 
1053 static void
1054 aoe_end_buf(struct aoedev *d, struct buf *buf)
1055 {
1056 	struct request *rq = buf->rq;
1057 	struct aoe_req *req = blk_mq_rq_to_pdu(rq);
1058 
1059 	if (buf == d->ip.buf)
1060 		d->ip.buf = NULL;
1061 	mempool_free(buf, d->bufpool);
1062 	if (--req->nr_bios == 0)
1063 		aoe_end_request(d, rq, 0);
1064 }
1065 
1066 static void
1067 ktiocomplete(struct frame *f)
1068 {
1069 	struct aoe_hdr *hin, *hout;
1070 	struct aoe_atahdr *ahin, *ahout;
1071 	struct buf *buf;
1072 	struct sk_buff *skb;
1073 	struct aoetgt *t;
1074 	struct aoeif *ifp;
1075 	struct aoedev *d;
1076 	long n;
1077 	int untainted;
1078 
1079 	if (f == NULL)
1080 		return;
1081 
1082 	t = f->t;
1083 	d = t->d;
1084 	skb = f->r_skb;
1085 	buf = f->buf;
1086 	if (f->flags & FFL_PROBE)
1087 		goto out;
1088 	if (!skb)		/* just fail the buf. */
1089 		goto noskb;
1090 
1091 	hout = (struct aoe_hdr *) skb_mac_header(f->skb);
1092 	ahout = (struct aoe_atahdr *) (hout+1);
1093 
1094 	hin = (struct aoe_hdr *) skb->data;
1095 	skb_pull(skb, sizeof(*hin));
1096 	ahin = (struct aoe_atahdr *) skb->data;
1097 	skb_pull(skb, sizeof(*ahin));
1098 	if (ahin->cmdstat & 0xa9) {	/* these bits cleared on success */
1099 		pr_err("aoe: ata error cmd=%2.2Xh stat=%2.2Xh from e%ld.%d\n",
1100 			ahout->cmdstat, ahin->cmdstat,
1101 			d->aoemajor, d->aoeminor);
1102 noskb:		if (buf)
1103 			buf->bio->bi_status = BLK_STS_IOERR;
1104 		goto out;
1105 	}
1106 
1107 	n = ahout->scnt << 9;
1108 	switch (ahout->cmdstat) {
1109 	case ATA_CMD_PIO_READ:
1110 	case ATA_CMD_PIO_READ_EXT:
1111 		if (skb->len < n) {
1112 			pr_err("%s e%ld.%d.  skb->len=%d need=%ld\n",
1113 				"aoe: runt data size in read from",
1114 				(long) d->aoemajor, d->aoeminor,
1115 			       skb->len, n);
1116 			buf->bio->bi_status = BLK_STS_IOERR;
1117 			break;
1118 		}
1119 		if (n > f->iter.bi_size) {
1120 			pr_err_ratelimited("%s e%ld.%d.  bytes=%ld need=%u\n",
1121 				"aoe: too-large data size in read from",
1122 				(long) d->aoemajor, d->aoeminor,
1123 				n, f->iter.bi_size);
1124 			buf->bio->bi_status = BLK_STS_IOERR;
1125 			break;
1126 		}
1127 		bvcpy(skb, f->buf->bio, f->iter, n);
1128 		fallthrough;
1129 	case ATA_CMD_PIO_WRITE:
1130 	case ATA_CMD_PIO_WRITE_EXT:
1131 		spin_lock_irq(&d->lock);
1132 		ifp = getif(t, skb->dev);
1133 		if (ifp)
1134 			ifp->lost = 0;
1135 		spin_unlock_irq(&d->lock);
1136 		break;
1137 	case ATA_CMD_ID_ATA:
1138 		if (skb->len < 512) {
1139 			pr_info("%s e%ld.%d.  skb->len=%d need=512\n",
1140 				"aoe: runt data size in ataid from",
1141 				(long) d->aoemajor, d->aoeminor,
1142 				skb->len);
1143 			break;
1144 		}
1145 		if (skb_linearize(skb))
1146 			break;
1147 		spin_lock_irq(&d->lock);
1148 		ataid_complete(d, t, skb->data);
1149 		spin_unlock_irq(&d->lock);
1150 		break;
1151 	default:
1152 		pr_info("aoe: unrecognized ata command %2.2Xh for %d.%d\n",
1153 			ahout->cmdstat,
1154 			be16_to_cpu(get_unaligned(&hin->major)),
1155 			hin->minor);
1156 	}
1157 out:
1158 	spin_lock_irq(&d->lock);
1159 	if (t->taint > 0
1160 	&& --t->taint > 0
1161 	&& t->nout_probes == 0) {
1162 		count_targets(d, &untainted);
1163 		if (untainted > 0) {
1164 			probe(t);
1165 			t->nout_probes++;
1166 		}
1167 	}
1168 
1169 	aoe_freetframe(f);
1170 
1171 	if (buf && --buf->nframesout == 0 && buf->iter.bi_size == 0)
1172 		aoe_end_buf(d, buf);
1173 
1174 	spin_unlock_irq(&d->lock);
1175 	aoedev_put(d);
1176 	dev_kfree_skb(skb);
1177 }
1178 
1179 /* Enters with iocq.lock held.
1180  * Returns true iff responses needing processing remain.
1181  */
1182 static int
1183 ktio(int id)
1184 {
1185 	struct frame *f;
1186 	struct list_head *pos;
1187 	int i;
1188 	int actual_id;
1189 
1190 	for (i = 0; ; ++i) {
1191 		if (i == MAXIOC)
1192 			return 1;
1193 		if (list_empty(&iocq[id].head))
1194 			return 0;
1195 		pos = iocq[id].head.next;
1196 		list_del(pos);
1197 		f = list_entry(pos, struct frame, head);
1198 		spin_unlock_irq(&iocq[id].lock);
1199 		ktiocomplete(f);
1200 
1201 		/* Figure out if extra threads are required. */
1202 		actual_id = f->t->d->aoeminor % ncpus;
1203 
1204 		if (!kts[actual_id].active) {
1205 			BUG_ON(id != 0);
1206 			mutex_lock(&ktio_spawn_lock);
1207 			if (!kts[actual_id].active
1208 				&& aoe_ktstart(&kts[actual_id]) == 0)
1209 				kts[actual_id].active = 1;
1210 			mutex_unlock(&ktio_spawn_lock);
1211 		}
1212 		spin_lock_irq(&iocq[id].lock);
1213 	}
1214 }
1215 
1216 static int
1217 kthread(void *vp)
1218 {
1219 	struct ktstate *k;
1220 	DECLARE_WAITQUEUE(wait, current);
1221 	int more;
1222 
1223 	k = vp;
1224 	current->flags |= PF_NOFREEZE;
1225 	set_user_nice(current, -10);
1226 	complete(&k->rendez);	/* tell spawner we're running */
1227 	do {
1228 		spin_lock_irq(k->lock);
1229 		more = k->fn(k->id);
1230 		if (!more) {
1231 			add_wait_queue(k->waitq, &wait);
1232 			__set_current_state(TASK_INTERRUPTIBLE);
1233 		}
1234 		spin_unlock_irq(k->lock);
1235 		if (!more) {
1236 			schedule();
1237 			remove_wait_queue(k->waitq, &wait);
1238 		} else
1239 			cond_resched();
1240 	} while (!kthread_should_stop());
1241 	complete(&k->rendez);	/* tell spawner we're stopping */
1242 	return 0;
1243 }
1244 
1245 void
1246 aoe_ktstop(struct ktstate *k)
1247 {
1248 	kthread_stop(k->task);
1249 	wait_for_completion(&k->rendez);
1250 }
1251 
1252 int
1253 aoe_ktstart(struct ktstate *k)
1254 {
1255 	struct task_struct *task;
1256 
1257 	init_completion(&k->rendez);
1258 	task = kthread_run(kthread, k, "%s", k->name);
1259 	if (task == NULL || IS_ERR(task))
1260 		return -ENOMEM;
1261 	k->task = task;
1262 	wait_for_completion(&k->rendez); /* allow kthread to start */
1263 	init_completion(&k->rendez);	/* for waiting for exit later */
1264 	return 0;
1265 }
1266 
1267 /* pass it off to kthreads for processing */
1268 static void
1269 ktcomplete(struct frame *f, struct sk_buff *skb)
1270 {
1271 	int id;
1272 	ulong flags;
1273 
1274 	f->r_skb = skb;
1275 	id = f->t->d->aoeminor % ncpus;
1276 	spin_lock_irqsave(&iocq[id].lock, flags);
1277 	if (!kts[id].active) {
1278 		spin_unlock_irqrestore(&iocq[id].lock, flags);
1279 		/* The thread with id has not been spawned yet,
1280 		 * so delegate the work to the main thread and
1281 		 * try spawning a new thread.
1282 		 */
1283 		id = 0;
1284 		spin_lock_irqsave(&iocq[id].lock, flags);
1285 	}
1286 	list_add_tail(&f->head, &iocq[id].head);
1287 	spin_unlock_irqrestore(&iocq[id].lock, flags);
1288 	wake_up(&ktiowq[id]);
1289 }
1290 
1291 struct sk_buff *
1292 aoecmd_ata_rsp(struct sk_buff *skb)
1293 {
1294 	struct aoedev *d;
1295 	struct aoe_hdr *h;
1296 	struct frame *f;
1297 	u32 n;
1298 	ulong flags;
1299 	char ebuf[128];
1300 	u16 aoemajor;
1301 
1302 	h = (struct aoe_hdr *) skb->data;
1303 	aoemajor = be16_to_cpu(get_unaligned(&h->major));
1304 	d = aoedev_by_aoeaddr(aoemajor, h->minor, 0);
1305 	if (d == NULL) {
1306 		snprintf(ebuf, sizeof ebuf, "aoecmd_ata_rsp: ata response "
1307 			"for unknown device %d.%d\n",
1308 			aoemajor, h->minor);
1309 		aoechr_error(ebuf);
1310 		return skb;
1311 	}
1312 
1313 	spin_lock_irqsave(&d->lock, flags);
1314 
1315 	n = be32_to_cpu(get_unaligned(&h->tag));
1316 	f = getframe(d, n);
1317 	if (f) {
1318 		calc_rttavg(d, f->t, tsince_hr(f));
1319 		f->t->nout--;
1320 		if (f->flags & FFL_PROBE)
1321 			f->t->nout_probes--;
1322 	} else {
1323 		f = getframe_deferred(d, n);
1324 		if (f) {
1325 			calc_rttavg(d, NULL, tsince_hr(f));
1326 		} else {
1327 			calc_rttavg(d, NULL, tsince(n));
1328 			spin_unlock_irqrestore(&d->lock, flags);
1329 			aoedev_put(d);
1330 			snprintf(ebuf, sizeof(ebuf),
1331 				 "%15s e%d.%d    tag=%08x@%08lx s=%pm d=%pm\n",
1332 				 "unexpected rsp",
1333 				 get_unaligned_be16(&h->major),
1334 				 h->minor,
1335 				 get_unaligned_be32(&h->tag),
1336 				 jiffies,
1337 				 h->src,
1338 				 h->dst);
1339 			aoechr_error(ebuf);
1340 			return skb;
1341 		}
1342 	}
1343 	aoecmd_work(d);
1344 
1345 	spin_unlock_irqrestore(&d->lock, flags);
1346 
1347 	ktcomplete(f, skb);
1348 
1349 	/*
1350 	 * Note here that we do not perform an aoedev_put, as we are
1351 	 * leaving this reference for the ktio to release.
1352 	 */
1353 	return NULL;
1354 }
1355 
1356 void
1357 aoecmd_cfg(ushort aoemajor, unsigned char aoeminor)
1358 {
1359 	struct sk_buff_head queue;
1360 
1361 	__skb_queue_head_init(&queue);
1362 	aoecmd_cfg_pkts(aoemajor, aoeminor, &queue);
1363 	aoenet_xmit(&queue);
1364 }
1365 
1366 struct sk_buff *
1367 aoecmd_ata_id(struct aoedev *d)
1368 {
1369 	struct aoe_hdr *h;
1370 	struct aoe_atahdr *ah;
1371 	struct frame *f;
1372 	struct sk_buff *skb;
1373 	struct aoetgt *t;
1374 
1375 	f = newframe(d);
1376 	if (f == NULL)
1377 		return NULL;
1378 
1379 	t = *d->tgt;
1380 
1381 	/* initialize the headers & frame */
1382 	skb = f->skb;
1383 	h = (struct aoe_hdr *) skb_mac_header(skb);
1384 	ah = (struct aoe_atahdr *) (h+1);
1385 	skb_put(skb, sizeof *h + sizeof *ah);
1386 	memset(h, 0, skb->len);
1387 	f->tag = aoehdr_atainit(d, t, h);
1388 	fhash(f);
1389 	t->nout++;
1390 	f->waited = 0;
1391 	f->waited_total = 0;
1392 
1393 	/* set up ata header */
1394 	ah->scnt = 1;
1395 	ah->cmdstat = ATA_CMD_ID_ATA;
1396 	ah->lba3 = 0xa0;
1397 
1398 	skb->dev = t->ifp->nd;
1399 
1400 	d->rttavg = RTTAVG_INIT;
1401 	d->rttdev = RTTDEV_INIT;
1402 	d->timer.function = rexmit_timer;
1403 
1404 	skb = skb_clone(skb, GFP_ATOMIC);
1405 	if (skb)
1406 		f->sent = ktime_get();
1407 
1408 	return skb;
1409 }
1410 
1411 static struct aoetgt **
1412 grow_targets(struct aoedev *d)
1413 {
1414 	ulong oldn, newn;
1415 	struct aoetgt **tt;
1416 
1417 	oldn = d->ntargets;
1418 	newn = oldn * 2;
1419 	tt = kcalloc(newn, sizeof(*d->targets), GFP_ATOMIC);
1420 	if (!tt)
1421 		return NULL;
1422 	memmove(tt, d->targets, sizeof(*d->targets) * oldn);
1423 	d->tgt = tt + (d->tgt - d->targets);
1424 	kfree(d->targets);
1425 	d->targets = tt;
1426 	d->ntargets = newn;
1427 
1428 	return &d->targets[oldn];
1429 }
1430 
1431 static struct aoetgt *
1432 addtgt(struct aoedev *d, char *addr, ulong nframes)
1433 {
1434 	struct aoetgt *t, **tt, **te;
1435 
1436 	tt = d->targets;
1437 	te = tt + d->ntargets;
1438 	for (; tt < te && *tt; tt++)
1439 		;
1440 
1441 	if (tt == te) {
1442 		tt = grow_targets(d);
1443 		if (!tt)
1444 			goto nomem;
1445 	}
1446 	t = kzalloc(sizeof(*t), GFP_ATOMIC);
1447 	if (!t)
1448 		goto nomem;
1449 	t->nframes = nframes;
1450 	t->d = d;
1451 	memcpy(t->addr, addr, sizeof t->addr);
1452 	t->ifp = t->ifs;
1453 	aoecmd_wreset(t);
1454 	t->maxout = t->nframes / 2;
1455 	INIT_LIST_HEAD(&t->ffree);
1456 	return *tt = t;
1457 
1458  nomem:
1459 	pr_info("aoe: cannot allocate memory to add target\n");
1460 	return NULL;
1461 }
1462 
1463 static void
1464 setdbcnt(struct aoedev *d)
1465 {
1466 	struct aoetgt **t, **e;
1467 	int bcnt = 0;
1468 
1469 	t = d->targets;
1470 	e = t + d->ntargets;
1471 	for (; t < e && *t; t++)
1472 		if (bcnt == 0 || bcnt > (*t)->minbcnt)
1473 			bcnt = (*t)->minbcnt;
1474 	if (bcnt != d->maxbcnt) {
1475 		d->maxbcnt = bcnt;
1476 		pr_info("aoe: e%ld.%d: setting %d byte data frames\n",
1477 			d->aoemajor, d->aoeminor, bcnt);
1478 	}
1479 }
1480 
1481 static void
1482 setifbcnt(struct aoetgt *t, struct net_device *nd, int bcnt)
1483 {
1484 	struct aoedev *d;
1485 	struct aoeif *p, *e;
1486 	int minbcnt;
1487 
1488 	d = t->d;
1489 	minbcnt = bcnt;
1490 	p = t->ifs;
1491 	e = p + NAOEIFS;
1492 	for (; p < e; p++) {
1493 		if (p->nd == NULL)
1494 			break;		/* end of the valid interfaces */
1495 		if (p->nd == nd) {
1496 			p->bcnt = bcnt;	/* we're updating */
1497 			nd = NULL;
1498 		} else if (minbcnt > p->bcnt)
1499 			minbcnt = p->bcnt; /* find the min interface */
1500 	}
1501 	if (nd) {
1502 		if (p == e) {
1503 			pr_err("aoe: device setifbcnt failure; too many interfaces.\n");
1504 			return;
1505 		}
1506 		dev_hold(nd);
1507 		p->nd = nd;
1508 		p->bcnt = bcnt;
1509 	}
1510 	t->minbcnt = minbcnt;
1511 	setdbcnt(d);
1512 }
1513 
1514 void
1515 aoecmd_cfg_rsp(struct sk_buff *skb)
1516 {
1517 	struct aoedev *d;
1518 	struct aoe_hdr *h;
1519 	struct aoe_cfghdr *ch;
1520 	struct aoetgt *t;
1521 	ulong flags, aoemajor;
1522 	struct sk_buff *sl;
1523 	struct sk_buff_head queue;
1524 	u16 n;
1525 
1526 	sl = NULL;
1527 	h = (struct aoe_hdr *) skb_mac_header(skb);
1528 	ch = (struct aoe_cfghdr *) (h+1);
1529 
1530 	/*
1531 	 * Enough people have their dip switches set backwards to
1532 	 * warrant a loud message for this special case.
1533 	 */
1534 	aoemajor = get_unaligned_be16(&h->major);
1535 	if (aoemajor == 0xfff) {
1536 		printk(KERN_ERR "aoe: Warning: shelf address is all ones.  "
1537 			"Check shelf dip switches.\n");
1538 		return;
1539 	}
1540 	if (aoemajor == 0xffff) {
1541 		pr_info("aoe: e%ld.%d: broadcast shelf number invalid\n",
1542 			aoemajor, (int) h->minor);
1543 		return;
1544 	}
1545 	if (h->minor == 0xff) {
1546 		pr_info("aoe: e%ld.%d: broadcast slot number invalid\n",
1547 			aoemajor, (int) h->minor);
1548 		return;
1549 	}
1550 
1551 	n = be16_to_cpu(ch->bufcnt);
1552 	if (n > aoe_maxout)	/* keep it reasonable */
1553 		n = aoe_maxout;
1554 
1555 	d = aoedev_by_aoeaddr(aoemajor, h->minor, 1);
1556 	if (d == NULL) {
1557 		pr_info("aoe: device allocation failure\n");
1558 		return;
1559 	}
1560 
1561 	spin_lock_irqsave(&d->lock, flags);
1562 
1563 	t = gettgt(d, h->src);
1564 	if (t) {
1565 		t->nframes = n;
1566 		if (n < t->maxout)
1567 			aoecmd_wreset(t);
1568 	} else {
1569 		t = addtgt(d, h->src, n);
1570 		if (!t)
1571 			goto bail;
1572 	}
1573 	n = skb->dev->mtu;
1574 	n -= sizeof(struct aoe_hdr) + sizeof(struct aoe_atahdr);
1575 	n /= 512;
1576 	if (n > ch->scnt)
1577 		n = ch->scnt;
1578 	n = n ? n * 512 : DEFAULTBCNT;
1579 	setifbcnt(t, skb->dev, n);
1580 
1581 	/* don't change users' perspective */
1582 	if (d->nopen == 0) {
1583 		d->fw_ver = be16_to_cpu(ch->fwver);
1584 		sl = aoecmd_ata_id(d);
1585 	}
1586 bail:
1587 	spin_unlock_irqrestore(&d->lock, flags);
1588 	aoedev_put(d);
1589 	if (sl) {
1590 		__skb_queue_head_init(&queue);
1591 		__skb_queue_tail(&queue, sl);
1592 		aoenet_xmit(&queue);
1593 	}
1594 }
1595 
1596 void
1597 aoecmd_wreset(struct aoetgt *t)
1598 {
1599 	t->maxout = 1;
1600 	t->ssthresh = t->nframes / 2;
1601 	t->next_cwnd = t->nframes;
1602 }
1603 
1604 void
1605 aoecmd_cleanslate(struct aoedev *d)
1606 {
1607 	struct aoetgt **t, **te;
1608 
1609 	d->rttavg = RTTAVG_INIT;
1610 	d->rttdev = RTTDEV_INIT;
1611 	d->maxbcnt = 0;
1612 
1613 	t = d->targets;
1614 	te = t + d->ntargets;
1615 	for (; t < te && *t; t++)
1616 		aoecmd_wreset(*t);
1617 }
1618 
1619 void
1620 aoe_failbuf(struct aoedev *d, struct buf *buf)
1621 {
1622 	if (buf == NULL)
1623 		return;
1624 	buf->iter.bi_size = 0;
1625 	buf->bio->bi_status = BLK_STS_IOERR;
1626 	if (buf->nframesout == 0)
1627 		aoe_end_buf(d, buf);
1628 }
1629 
1630 void
1631 aoe_flush_iocq(void)
1632 {
1633 	int i;
1634 
1635 	for (i = 0; i < ncpus; i++) {
1636 		if (kts[i].active)
1637 			aoe_flush_iocq_by_index(i);
1638 	}
1639 }
1640 
1641 void
1642 aoe_flush_iocq_by_index(int id)
1643 {
1644 	struct frame *f;
1645 	struct aoedev *d;
1646 	LIST_HEAD(flist);
1647 	struct list_head *pos;
1648 	struct sk_buff *skb;
1649 	ulong flags;
1650 
1651 	spin_lock_irqsave(&iocq[id].lock, flags);
1652 	list_splice_init(&iocq[id].head, &flist);
1653 	spin_unlock_irqrestore(&iocq[id].lock, flags);
1654 	while (!list_empty(&flist)) {
1655 		pos = flist.next;
1656 		list_del(pos);
1657 		f = list_entry(pos, struct frame, head);
1658 		d = f->t->d;
1659 		skb = f->r_skb;
1660 		spin_lock_irqsave(&d->lock, flags);
1661 		if (f->buf) {
1662 			f->buf->nframesout--;
1663 			aoe_failbuf(d, f->buf);
1664 		}
1665 		aoe_freetframe(f);
1666 		spin_unlock_irqrestore(&d->lock, flags);
1667 		dev_kfree_skb(skb);
1668 		aoedev_put(d);
1669 	}
1670 }
1671 
1672 int __init
1673 aoecmd_init(void)
1674 {
1675 	void *p;
1676 	int i;
1677 	int ret;
1678 
1679 	/* get_zeroed_page returns page with ref count 1 */
1680 	p = (void *) get_zeroed_page(GFP_KERNEL);
1681 	if (!p)
1682 		return -ENOMEM;
1683 	empty_page = virt_to_page(p);
1684 
1685 	ncpus = num_online_cpus();
1686 
1687 	iocq = kcalloc(ncpus, sizeof(struct iocq_ktio), GFP_KERNEL);
1688 	if (!iocq)
1689 		return -ENOMEM;
1690 
1691 	kts = kcalloc(ncpus, sizeof(struct ktstate), GFP_KERNEL);
1692 	if (!kts) {
1693 		ret = -ENOMEM;
1694 		goto kts_fail;
1695 	}
1696 
1697 	ktiowq = kcalloc(ncpus, sizeof(wait_queue_head_t), GFP_KERNEL);
1698 	if (!ktiowq) {
1699 		ret = -ENOMEM;
1700 		goto ktiowq_fail;
1701 	}
1702 
1703 	for (i = 0; i < ncpus; i++) {
1704 		INIT_LIST_HEAD(&iocq[i].head);
1705 		spin_lock_init(&iocq[i].lock);
1706 		init_waitqueue_head(&ktiowq[i]);
1707 		snprintf(kts[i].name, sizeof(kts[i].name), "aoe_ktio%d", i);
1708 		kts[i].fn = ktio;
1709 		kts[i].waitq = &ktiowq[i];
1710 		kts[i].lock = &iocq[i].lock;
1711 		kts[i].id = i;
1712 		kts[i].active = 0;
1713 	}
1714 	kts[0].active = 1;
1715 	if (aoe_ktstart(&kts[0])) {
1716 		ret = -ENOMEM;
1717 		goto ktstart_fail;
1718 	}
1719 	return 0;
1720 
1721 ktstart_fail:
1722 	kfree(ktiowq);
1723 ktiowq_fail:
1724 	kfree(kts);
1725 kts_fail:
1726 	kfree(iocq);
1727 
1728 	return ret;
1729 }
1730 
1731 void
1732 aoecmd_exit(void)
1733 {
1734 	int i;
1735 
1736 	for (i = 0; i < ncpus; i++)
1737 		if (kts[i].active)
1738 			aoe_ktstop(&kts[i]);
1739 
1740 	aoe_flush_iocq();
1741 
1742 	/* Free up the iocq and thread speicific configuration
1743 	* allocated during startup.
1744 	*/
1745 	kfree(iocq);
1746 	kfree(kts);
1747 	kfree(ktiowq);
1748 
1749 	free_page((unsigned long) page_address(empty_page));
1750 	empty_page = NULL;
1751 }
1752