1 /* Copyright (c) 2013 Coraid, Inc. See COPYING for GPL terms. */ 2 /* 3 * aoedev.c 4 * AoE device utility functions; maintains device list. 5 */ 6 7 #include <linux/hdreg.h> 8 #include <linux/blk-mq.h> 9 #include <linux/netdevice.h> 10 #include <linux/delay.h> 11 #include <linux/slab.h> 12 #include <linux/bitmap.h> 13 #include <linux/kdev_t.h> 14 #include <linux/moduleparam.h> 15 #include <linux/string.h> 16 #include "aoe.h" 17 18 static void freetgt(struct aoedev *d, struct aoetgt *t); 19 static void skbpoolfree(struct aoedev *d); 20 21 static int aoe_dyndevs = 1; 22 module_param(aoe_dyndevs, int, 0644); 23 MODULE_PARM_DESC(aoe_dyndevs, "Use dynamic minor numbers for devices."); 24 25 static struct aoedev *devlist; 26 static DEFINE_SPINLOCK(devlist_lock); 27 28 /* Because some systems will have one, many, or no 29 * - partitions, 30 * - slots per shelf, 31 * - or shelves, 32 * we need some flexibility in the way the minor numbers 33 * are allocated. So they are dynamic. 34 */ 35 #define N_DEVS ((1U<<MINORBITS)/AOE_PARTITIONS) 36 37 static DEFINE_SPINLOCK(used_minors_lock); 38 static DECLARE_BITMAP(used_minors, N_DEVS); 39 40 static int 41 minor_get_dyn(ulong *sysminor) 42 { 43 ulong flags; 44 ulong n; 45 int error = 0; 46 47 spin_lock_irqsave(&used_minors_lock, flags); 48 n = find_first_zero_bit(used_minors, N_DEVS); 49 if (n < N_DEVS) 50 set_bit(n, used_minors); 51 else 52 error = -1; 53 spin_unlock_irqrestore(&used_minors_lock, flags); 54 55 *sysminor = n * AOE_PARTITIONS; 56 return error; 57 } 58 59 static int 60 minor_get_static(ulong *sysminor, ulong aoemaj, int aoemin) 61 { 62 ulong flags; 63 ulong n; 64 int error = 0; 65 enum { 66 /* for backwards compatibility when !aoe_dyndevs, 67 * a static number of supported slots per shelf */ 68 NPERSHELF = 16, 69 }; 70 71 if (aoemin >= NPERSHELF) { 72 pr_err("aoe: %s %d slots per shelf\n", 73 "static minor device numbers support only", 74 NPERSHELF); 75 error = -1; 76 goto out; 77 } 78 79 n = aoemaj * NPERSHELF + aoemin; 80 if (n >= N_DEVS) { 81 pr_err("aoe: %s with e%ld.%d\n", 82 "cannot use static minor device numbers", 83 aoemaj, aoemin); 84 error = -1; 85 goto out; 86 } 87 88 spin_lock_irqsave(&used_minors_lock, flags); 89 if (test_bit(n, used_minors)) { 90 pr_err("aoe: %s %lu\n", 91 "existing device already has static minor number", 92 n); 93 error = -1; 94 } else 95 set_bit(n, used_minors); 96 spin_unlock_irqrestore(&used_minors_lock, flags); 97 *sysminor = n * AOE_PARTITIONS; 98 out: 99 return error; 100 } 101 102 static int 103 minor_get(ulong *sysminor, ulong aoemaj, int aoemin) 104 { 105 if (aoe_dyndevs) 106 return minor_get_dyn(sysminor); 107 else 108 return minor_get_static(sysminor, aoemaj, aoemin); 109 } 110 111 static void 112 minor_free(ulong minor) 113 { 114 ulong flags; 115 116 minor /= AOE_PARTITIONS; 117 BUG_ON(minor >= N_DEVS); 118 119 spin_lock_irqsave(&used_minors_lock, flags); 120 BUG_ON(!test_bit(minor, used_minors)); 121 clear_bit(minor, used_minors); 122 spin_unlock_irqrestore(&used_minors_lock, flags); 123 } 124 125 /* 126 * Users who grab a pointer to the device with aoedev_by_aoeaddr 127 * automatically get a reference count and must be responsible 128 * for performing a aoedev_put. With the addition of async 129 * kthread processing I'm no longer confident that we can 130 * guarantee consistency in the face of device flushes. 131 * 132 * For the time being, we only bother to add extra references for 133 * frames sitting on the iocq. When the kthreads finish processing 134 * these frames, they will aoedev_put the device. 135 */ 136 137 void 138 aoedev_put(struct aoedev *d) 139 { 140 ulong flags; 141 142 spin_lock_irqsave(&devlist_lock, flags); 143 d->ref--; 144 spin_unlock_irqrestore(&devlist_lock, flags); 145 } 146 147 static void 148 dummy_timer(struct timer_list *t) 149 { 150 struct aoedev *d; 151 152 d = from_timer(d, t, timer); 153 if (d->flags & DEVFL_TKILL) 154 return; 155 d->timer.expires = jiffies + HZ; 156 add_timer(&d->timer); 157 } 158 159 static void 160 aoe_failip(struct aoedev *d) 161 { 162 struct request *rq; 163 struct bio *bio; 164 unsigned long n; 165 166 aoe_failbuf(d, d->ip.buf); 167 168 rq = d->ip.rq; 169 if (rq == NULL) 170 return; 171 while ((bio = d->ip.nxbio)) { 172 bio->bi_status = BLK_STS_IOERR; 173 d->ip.nxbio = bio->bi_next; 174 n = (unsigned long) rq->special; 175 rq->special = (void *) --n; 176 } 177 if ((unsigned long) rq->special == 0) 178 aoe_end_request(d, rq, 0); 179 } 180 181 static void 182 downdev_frame(struct list_head *pos) 183 { 184 struct frame *f; 185 186 f = list_entry(pos, struct frame, head); 187 list_del(pos); 188 if (f->buf) { 189 f->buf->nframesout--; 190 aoe_failbuf(f->t->d, f->buf); 191 } 192 aoe_freetframe(f); 193 } 194 195 void 196 aoedev_downdev(struct aoedev *d) 197 { 198 struct aoetgt *t, **tt, **te; 199 struct list_head *head, *pos, *nx; 200 int i; 201 202 d->flags &= ~DEVFL_UP; 203 204 /* clean out active and to-be-retransmitted buffers */ 205 for (i = 0; i < NFACTIVE; i++) { 206 head = &d->factive[i]; 207 list_for_each_safe(pos, nx, head) 208 downdev_frame(pos); 209 } 210 head = &d->rexmitq; 211 list_for_each_safe(pos, nx, head) 212 downdev_frame(pos); 213 214 /* reset window dressings */ 215 tt = d->targets; 216 te = tt + d->ntargets; 217 for (; tt < te && (t = *tt); tt++) { 218 aoecmd_wreset(t); 219 t->nout = 0; 220 } 221 222 /* clean out the in-process request (if any) */ 223 aoe_failip(d); 224 225 /* fast fail all pending I/O */ 226 if (d->blkq) { 227 /* UP is cleared, freeze+quiesce to insure all are errored */ 228 blk_mq_freeze_queue(d->blkq); 229 blk_mq_quiesce_queue(d->blkq); 230 blk_mq_unquiesce_queue(d->blkq); 231 blk_mq_unfreeze_queue(d->blkq); 232 } 233 234 if (d->gd) 235 set_capacity(d->gd, 0); 236 } 237 238 /* return whether the user asked for this particular 239 * device to be flushed 240 */ 241 static int 242 user_req(char *s, size_t slen, struct aoedev *d) 243 { 244 const char *p; 245 size_t lim; 246 247 if (!d->gd) 248 return 0; 249 p = kbasename(d->gd->disk_name); 250 lim = sizeof(d->gd->disk_name); 251 lim -= p - d->gd->disk_name; 252 if (slen < lim) 253 lim = slen; 254 255 return !strncmp(s, p, lim); 256 } 257 258 static void 259 freedev(struct aoedev *d) 260 { 261 struct aoetgt **t, **e; 262 int freeing = 0; 263 unsigned long flags; 264 265 spin_lock_irqsave(&d->lock, flags); 266 if (d->flags & DEVFL_TKILL 267 && !(d->flags & DEVFL_FREEING)) { 268 d->flags |= DEVFL_FREEING; 269 freeing = 1; 270 } 271 spin_unlock_irqrestore(&d->lock, flags); 272 if (!freeing) 273 return; 274 275 del_timer_sync(&d->timer); 276 if (d->gd) { 277 aoedisk_rm_debugfs(d); 278 del_gendisk(d->gd); 279 put_disk(d->gd); 280 blk_mq_free_tag_set(&d->tag_set); 281 blk_cleanup_queue(d->blkq); 282 } 283 t = d->targets; 284 e = t + d->ntargets; 285 for (; t < e && *t; t++) 286 freetgt(d, *t); 287 288 mempool_destroy(d->bufpool); 289 skbpoolfree(d); 290 minor_free(d->sysminor); 291 292 spin_lock_irqsave(&d->lock, flags); 293 d->flags |= DEVFL_FREED; 294 spin_unlock_irqrestore(&d->lock, flags); 295 } 296 297 enum flush_parms { 298 NOT_EXITING = 0, 299 EXITING = 1, 300 }; 301 302 static int 303 flush(const char __user *str, size_t cnt, int exiting) 304 { 305 ulong flags; 306 struct aoedev *d, **dd; 307 char buf[16]; 308 int all = 0; 309 int specified = 0; /* flush a specific device */ 310 unsigned int skipflags; 311 312 skipflags = DEVFL_GDALLOC | DEVFL_NEWSIZE | DEVFL_TKILL; 313 314 if (!exiting && cnt >= 3) { 315 if (cnt > sizeof buf) 316 cnt = sizeof buf; 317 if (copy_from_user(buf, str, cnt)) 318 return -EFAULT; 319 all = !strncmp(buf, "all", 3); 320 if (!all) 321 specified = 1; 322 } 323 324 flush_scheduled_work(); 325 /* pass one: without sleeping, do aoedev_downdev */ 326 spin_lock_irqsave(&devlist_lock, flags); 327 for (d = devlist; d; d = d->next) { 328 spin_lock(&d->lock); 329 if (exiting) { 330 /* unconditionally take each device down */ 331 } else if (specified) { 332 if (!user_req(buf, cnt, d)) 333 goto cont; 334 } else if ((!all && (d->flags & DEVFL_UP)) 335 || d->flags & skipflags 336 || d->nopen 337 || d->ref) 338 goto cont; 339 340 aoedev_downdev(d); 341 d->flags |= DEVFL_TKILL; 342 cont: 343 spin_unlock(&d->lock); 344 } 345 spin_unlock_irqrestore(&devlist_lock, flags); 346 347 /* pass two: call freedev, which might sleep, 348 * for aoedevs marked with DEVFL_TKILL 349 */ 350 restart: 351 spin_lock_irqsave(&devlist_lock, flags); 352 for (d = devlist; d; d = d->next) { 353 spin_lock(&d->lock); 354 if (d->flags & DEVFL_TKILL 355 && !(d->flags & DEVFL_FREEING)) { 356 spin_unlock(&d->lock); 357 spin_unlock_irqrestore(&devlist_lock, flags); 358 freedev(d); 359 goto restart; 360 } 361 spin_unlock(&d->lock); 362 } 363 364 /* pass three: remove aoedevs marked with DEVFL_FREED */ 365 for (dd = &devlist, d = *dd; d; d = *dd) { 366 struct aoedev *doomed = NULL; 367 368 spin_lock(&d->lock); 369 if (d->flags & DEVFL_FREED) { 370 *dd = d->next; 371 doomed = d; 372 } else { 373 dd = &d->next; 374 } 375 spin_unlock(&d->lock); 376 if (doomed) 377 kfree(doomed->targets); 378 kfree(doomed); 379 } 380 spin_unlock_irqrestore(&devlist_lock, flags); 381 382 return 0; 383 } 384 385 int 386 aoedev_flush(const char __user *str, size_t cnt) 387 { 388 return flush(str, cnt, NOT_EXITING); 389 } 390 391 /* This has been confirmed to occur once with Tms=3*1000 due to the 392 * driver changing link and not processing its transmit ring. The 393 * problem is hard enough to solve by returning an error that I'm 394 * still punting on "solving" this. 395 */ 396 static void 397 skbfree(struct sk_buff *skb) 398 { 399 enum { Sms = 250, Tms = 30 * 1000}; 400 int i = Tms / Sms; 401 402 if (skb == NULL) 403 return; 404 while (atomic_read(&skb_shinfo(skb)->dataref) != 1 && i-- > 0) 405 msleep(Sms); 406 if (i < 0) { 407 printk(KERN_ERR 408 "aoe: %s holds ref: %s\n", 409 skb->dev ? skb->dev->name : "netif", 410 "cannot free skb -- memory leaked."); 411 return; 412 } 413 skb->truesize -= skb->data_len; 414 skb_shinfo(skb)->nr_frags = skb->data_len = 0; 415 skb_trim(skb, 0); 416 dev_kfree_skb(skb); 417 } 418 419 static void 420 skbpoolfree(struct aoedev *d) 421 { 422 struct sk_buff *skb, *tmp; 423 424 skb_queue_walk_safe(&d->skbpool, skb, tmp) 425 skbfree(skb); 426 427 __skb_queue_head_init(&d->skbpool); 428 } 429 430 /* find it or allocate it */ 431 struct aoedev * 432 aoedev_by_aoeaddr(ulong maj, int min, int do_alloc) 433 { 434 struct aoedev *d; 435 int i; 436 ulong flags; 437 ulong sysminor = 0; 438 439 spin_lock_irqsave(&devlist_lock, flags); 440 441 for (d=devlist; d; d=d->next) 442 if (d->aoemajor == maj && d->aoeminor == min) { 443 spin_lock(&d->lock); 444 if (d->flags & DEVFL_TKILL) { 445 spin_unlock(&d->lock); 446 d = NULL; 447 goto out; 448 } 449 d->ref++; 450 spin_unlock(&d->lock); 451 break; 452 } 453 if (d || !do_alloc || minor_get(&sysminor, maj, min) < 0) 454 goto out; 455 d = kcalloc(1, sizeof *d, GFP_ATOMIC); 456 if (!d) 457 goto out; 458 d->targets = kcalloc(NTARGETS, sizeof(*d->targets), GFP_ATOMIC); 459 if (!d->targets) { 460 kfree(d); 461 d = NULL; 462 goto out; 463 } 464 d->ntargets = NTARGETS; 465 INIT_WORK(&d->work, aoecmd_sleepwork); 466 spin_lock_init(&d->lock); 467 INIT_LIST_HEAD(&d->rq_list); 468 skb_queue_head_init(&d->skbpool); 469 timer_setup(&d->timer, dummy_timer, 0); 470 d->timer.expires = jiffies + HZ; 471 add_timer(&d->timer); 472 d->bufpool = NULL; /* defer to aoeblk_gdalloc */ 473 d->tgt = d->targets; 474 d->ref = 1; 475 for (i = 0; i < NFACTIVE; i++) 476 INIT_LIST_HEAD(&d->factive[i]); 477 INIT_LIST_HEAD(&d->rexmitq); 478 d->sysminor = sysminor; 479 d->aoemajor = maj; 480 d->aoeminor = min; 481 d->rttavg = RTTAVG_INIT; 482 d->rttdev = RTTDEV_INIT; 483 d->next = devlist; 484 devlist = d; 485 out: 486 spin_unlock_irqrestore(&devlist_lock, flags); 487 return d; 488 } 489 490 static void 491 freetgt(struct aoedev *d, struct aoetgt *t) 492 { 493 struct frame *f; 494 struct list_head *pos, *nx, *head; 495 struct aoeif *ifp; 496 497 for (ifp = t->ifs; ifp < &t->ifs[NAOEIFS]; ++ifp) { 498 if (!ifp->nd) 499 break; 500 dev_put(ifp->nd); 501 } 502 503 head = &t->ffree; 504 list_for_each_safe(pos, nx, head) { 505 list_del(pos); 506 f = list_entry(pos, struct frame, head); 507 skbfree(f->skb); 508 kfree(f); 509 } 510 kfree(t); 511 } 512 513 void 514 aoedev_exit(void) 515 { 516 flush_scheduled_work(); 517 flush(NULL, 0, EXITING); 518 } 519 520 int __init 521 aoedev_init(void) 522 { 523 return 0; 524 } 525