drivers/block/aoe/aoecmd.c

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