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 aoe_req *req; 164 struct bio *bio; 165 166 aoe_failbuf(d, d->ip.buf); 167 rq = d->ip.rq; 168 if (rq == NULL) 169 return; 170 171 req = blk_mq_rq_to_pdu(rq); 172 while ((bio = d->ip.nxbio)) { 173 bio->bi_status = BLK_STS_IOERR; 174 d->ip.nxbio = bio->bi_next; 175 req->nr_bios--; 176 } 177 178 if (!req->nr_bios) 179 aoe_end_request(d, rq, 0); 180 } 181 182 static void 183 downdev_frame(struct list_head *pos) 184 { 185 struct frame *f; 186 187 f = list_entry(pos, struct frame, head); 188 list_del(pos); 189 if (f->buf) { 190 f->buf->nframesout--; 191 aoe_failbuf(f->t->d, f->buf); 192 } 193 aoe_freetframe(f); 194 } 195 196 void 197 aoedev_downdev(struct aoedev *d) 198 { 199 struct aoetgt *t, **tt, **te; 200 struct list_head *head, *pos, *nx; 201 int i; 202 203 d->flags &= ~DEVFL_UP; 204 205 /* clean out active and to-be-retransmitted buffers */ 206 for (i = 0; i < NFACTIVE; i++) { 207 head = &d->factive[i]; 208 list_for_each_safe(pos, nx, head) 209 downdev_frame(pos); 210 } 211 head = &d->rexmitq; 212 list_for_each_safe(pos, nx, head) 213 downdev_frame(pos); 214 215 /* reset window dressings */ 216 tt = d->targets; 217 te = tt + d->ntargets; 218 for (; tt < te && (t = *tt); tt++) { 219 aoecmd_wreset(t); 220 t->nout = 0; 221 } 222 223 /* clean out the in-process request (if any) */ 224 aoe_failip(d); 225 226 /* fast fail all pending I/O */ 227 if (d->blkq) { 228 /* UP is cleared, freeze+quiesce to insure all are errored */ 229 blk_mq_freeze_queue(d->blkq); 230 blk_mq_quiesce_queue(d->blkq); 231 blk_mq_unquiesce_queue(d->blkq); 232 blk_mq_unfreeze_queue(d->blkq); 233 } 234 235 if (d->gd) 236 set_capacity(d->gd, 0); 237 } 238 239 /* return whether the user asked for this particular 240 * device to be flushed 241 */ 242 static int 243 user_req(char *s, size_t slen, struct aoedev *d) 244 { 245 const char *p; 246 size_t lim; 247 248 if (!d->gd) 249 return 0; 250 p = kbasename(d->gd->disk_name); 251 lim = sizeof(d->gd->disk_name); 252 lim -= p - d->gd->disk_name; 253 if (slen < lim) 254 lim = slen; 255 256 return !strncmp(s, p, lim); 257 } 258 259 static void 260 freedev(struct aoedev *d) 261 { 262 struct aoetgt **t, **e; 263 int freeing = 0; 264 unsigned long flags; 265 266 spin_lock_irqsave(&d->lock, flags); 267 if (d->flags & DEVFL_TKILL 268 && !(d->flags & DEVFL_FREEING)) { 269 d->flags |= DEVFL_FREEING; 270 freeing = 1; 271 } 272 spin_unlock_irqrestore(&d->lock, flags); 273 if (!freeing) 274 return; 275 276 del_timer_sync(&d->timer); 277 if (d->gd) { 278 aoedisk_rm_debugfs(d); 279 del_gendisk(d->gd); 280 put_disk(d->gd); 281 blk_mq_free_tag_set(&d->tag_set); 282 blk_cleanup_queue(d->blkq); 283 } 284 t = d->targets; 285 e = t + d->ntargets; 286 for (; t < e && *t; t++) 287 freetgt(d, *t); 288 289 mempool_destroy(d->bufpool); 290 skbpoolfree(d); 291 minor_free(d->sysminor); 292 293 spin_lock_irqsave(&d->lock, flags); 294 d->flags |= DEVFL_FREED; 295 spin_unlock_irqrestore(&d->lock, flags); 296 } 297 298 enum flush_parms { 299 NOT_EXITING = 0, 300 EXITING = 1, 301 }; 302 303 static int 304 flush(const char __user *str, size_t cnt, int exiting) 305 { 306 ulong flags; 307 struct aoedev *d, **dd; 308 char buf[16]; 309 int all = 0; 310 int specified = 0; /* flush a specific device */ 311 unsigned int skipflags; 312 313 skipflags = DEVFL_GDALLOC | DEVFL_NEWSIZE | DEVFL_TKILL; 314 315 if (!exiting && cnt >= 3) { 316 if (cnt > sizeof buf) 317 cnt = sizeof buf; 318 if (copy_from_user(buf, str, cnt)) 319 return -EFAULT; 320 all = !strncmp(buf, "all", 3); 321 if (!all) 322 specified = 1; 323 } 324 325 flush_scheduled_work(); 326 /* pass one: without sleeping, do aoedev_downdev */ 327 spin_lock_irqsave(&devlist_lock, flags); 328 for (d = devlist; d; d = d->next) { 329 spin_lock(&d->lock); 330 if (exiting) { 331 /* unconditionally take each device down */ 332 } else if (specified) { 333 if (!user_req(buf, cnt, d)) 334 goto cont; 335 } else if ((!all && (d->flags & DEVFL_UP)) 336 || d->flags & skipflags 337 || d->nopen 338 || d->ref) 339 goto cont; 340 341 aoedev_downdev(d); 342 d->flags |= DEVFL_TKILL; 343 cont: 344 spin_unlock(&d->lock); 345 } 346 spin_unlock_irqrestore(&devlist_lock, flags); 347 348 /* pass two: call freedev, which might sleep, 349 * for aoedevs marked with DEVFL_TKILL 350 */ 351 restart: 352 spin_lock_irqsave(&devlist_lock, flags); 353 for (d = devlist; d; d = d->next) { 354 spin_lock(&d->lock); 355 if (d->flags & DEVFL_TKILL 356 && !(d->flags & DEVFL_FREEING)) { 357 spin_unlock(&d->lock); 358 spin_unlock_irqrestore(&devlist_lock, flags); 359 freedev(d); 360 goto restart; 361 } 362 spin_unlock(&d->lock); 363 } 364 365 /* pass three: remove aoedevs marked with DEVFL_FREED */ 366 for (dd = &devlist, d = *dd; d; d = *dd) { 367 struct aoedev *doomed = NULL; 368 369 spin_lock(&d->lock); 370 if (d->flags & DEVFL_FREED) { 371 *dd = d->next; 372 doomed = d; 373 } else { 374 dd = &d->next; 375 } 376 spin_unlock(&d->lock); 377 if (doomed) 378 kfree(doomed->targets); 379 kfree(doomed); 380 } 381 spin_unlock_irqrestore(&devlist_lock, flags); 382 383 return 0; 384 } 385 386 int 387 aoedev_flush(const char __user *str, size_t cnt) 388 { 389 return flush(str, cnt, NOT_EXITING); 390 } 391 392 /* This has been confirmed to occur once with Tms=3*1000 due to the 393 * driver changing link and not processing its transmit ring. The 394 * problem is hard enough to solve by returning an error that I'm 395 * still punting on "solving" this. 396 */ 397 static void 398 skbfree(struct sk_buff *skb) 399 { 400 enum { Sms = 250, Tms = 30 * 1000}; 401 int i = Tms / Sms; 402 403 if (skb == NULL) 404 return; 405 while (atomic_read(&skb_shinfo(skb)->dataref) != 1 && i-- > 0) 406 msleep(Sms); 407 if (i < 0) { 408 printk(KERN_ERR 409 "aoe: %s holds ref: %s\n", 410 skb->dev ? skb->dev->name : "netif", 411 "cannot free skb -- memory leaked."); 412 return; 413 } 414 skb->truesize -= skb->data_len; 415 skb_shinfo(skb)->nr_frags = skb->data_len = 0; 416 skb_trim(skb, 0); 417 dev_kfree_skb(skb); 418 } 419 420 static void 421 skbpoolfree(struct aoedev *d) 422 { 423 struct sk_buff *skb, *tmp; 424 425 skb_queue_walk_safe(&d->skbpool, skb, tmp) 426 skbfree(skb); 427 428 __skb_queue_head_init(&d->skbpool); 429 } 430 431 /* find it or allocate it */ 432 struct aoedev * 433 aoedev_by_aoeaddr(ulong maj, int min, int do_alloc) 434 { 435 struct aoedev *d; 436 int i; 437 ulong flags; 438 ulong sysminor = 0; 439 440 spin_lock_irqsave(&devlist_lock, flags); 441 442 for (d=devlist; d; d=d->next) 443 if (d->aoemajor == maj && d->aoeminor == min) { 444 spin_lock(&d->lock); 445 if (d->flags & DEVFL_TKILL) { 446 spin_unlock(&d->lock); 447 d = NULL; 448 goto out; 449 } 450 d->ref++; 451 spin_unlock(&d->lock); 452 break; 453 } 454 if (d || !do_alloc || minor_get(&sysminor, maj, min) < 0) 455 goto out; 456 d = kcalloc(1, sizeof *d, GFP_ATOMIC); 457 if (!d) 458 goto out; 459 d->targets = kcalloc(NTARGETS, sizeof(*d->targets), GFP_ATOMIC); 460 if (!d->targets) { 461 kfree(d); 462 d = NULL; 463 goto out; 464 } 465 d->ntargets = NTARGETS; 466 INIT_WORK(&d->work, aoecmd_sleepwork); 467 spin_lock_init(&d->lock); 468 INIT_LIST_HEAD(&d->rq_list); 469 skb_queue_head_init(&d->skbpool); 470 timer_setup(&d->timer, dummy_timer, 0); 471 d->timer.expires = jiffies + HZ; 472 add_timer(&d->timer); 473 d->bufpool = NULL; /* defer to aoeblk_gdalloc */ 474 d->tgt = d->targets; 475 d->ref = 1; 476 for (i = 0; i < NFACTIVE; i++) 477 INIT_LIST_HEAD(&d->factive[i]); 478 INIT_LIST_HEAD(&d->rexmitq); 479 d->sysminor = sysminor; 480 d->aoemajor = maj; 481 d->aoeminor = min; 482 d->rttavg = RTTAVG_INIT; 483 d->rttdev = RTTDEV_INIT; 484 d->next = devlist; 485 devlist = d; 486 out: 487 spin_unlock_irqrestore(&devlist_lock, flags); 488 return d; 489 } 490 491 static void 492 freetgt(struct aoedev *d, struct aoetgt *t) 493 { 494 struct frame *f; 495 struct list_head *pos, *nx, *head; 496 struct aoeif *ifp; 497 498 for (ifp = t->ifs; ifp < &t->ifs[NAOEIFS]; ++ifp) { 499 if (!ifp->nd) 500 break; 501 dev_put(ifp->nd); 502 } 503 504 head = &t->ffree; 505 list_for_each_safe(pos, nx, head) { 506 list_del(pos); 507 f = list_entry(pos, struct frame, head); 508 skbfree(f->skb); 509 kfree(f); 510 } 511 kfree(t); 512 } 513 514 void 515 aoedev_exit(void) 516 { 517 flush_scheduled_work(); 518 flush(NULL, 0, EXITING); 519 } 520 521 int __init 522 aoedev_init(void) 523 { 524 return 0; 525 } 526