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
minor_get_dyn(ulong * sysminor)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
minor_get_static(ulong * sysminor,ulong aoemaj,int aoemin)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
minor_get(ulong * sysminor,ulong aoemaj,int aoemin)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
minor_free(ulong minor)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
aoedev_put(struct aoedev * d)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
dummy_timer(struct timer_list * t)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
aoe_failip(struct aoedev * d)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
downdev_frame(struct list_head * pos)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
aoedev_downdev(struct aoedev * d)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
user_req(char * s,size_t slen,struct aoedev * d)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
freedev(struct aoedev * d)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 blk_cleanup_disk(d->gd);
281 blk_mq_free_tag_set(&d->tag_set);
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
flush(const char __user * str,size_t cnt,int exiting)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: do aoedev_downdev, which might sleep */
326 restart1:
327 spin_lock_irqsave(&devlist_lock, flags);
328 for (d = devlist; d; d = d->next) {
329 spin_lock(&d->lock);
330 if (d->flags & DEVFL_TKILL)
331 goto cont;
332
333 if (exiting) {
334 /* unconditionally take each device down */
335 } else if (specified) {
336 if (!user_req(buf, cnt, d))
337 goto cont;
338 } else if ((!all && (d->flags & DEVFL_UP))
339 || d->flags & skipflags
340 || d->nopen
341 || d->ref)
342 goto cont;
343
344 spin_unlock(&d->lock);
345 spin_unlock_irqrestore(&devlist_lock, flags);
346 aoedev_downdev(d);
347 d->flags |= DEVFL_TKILL;
348 goto restart1;
349 cont:
350 spin_unlock(&d->lock);
351 }
352 spin_unlock_irqrestore(&devlist_lock, flags);
353
354 /* pass two: call freedev, which might sleep,
355 * for aoedevs marked with DEVFL_TKILL
356 */
357 restart2:
358 spin_lock_irqsave(&devlist_lock, flags);
359 for (d = devlist; d; d = d->next) {
360 spin_lock(&d->lock);
361 if (d->flags & DEVFL_TKILL
362 && !(d->flags & DEVFL_FREEING)) {
363 spin_unlock(&d->lock);
364 spin_unlock_irqrestore(&devlist_lock, flags);
365 freedev(d);
366 goto restart2;
367 }
368 spin_unlock(&d->lock);
369 }
370
371 /* pass three: remove aoedevs marked with DEVFL_FREED */
372 for (dd = &devlist, d = *dd; d; d = *dd) {
373 struct aoedev *doomed = NULL;
374
375 spin_lock(&d->lock);
376 if (d->flags & DEVFL_FREED) {
377 *dd = d->next;
378 doomed = d;
379 } else {
380 dd = &d->next;
381 }
382 spin_unlock(&d->lock);
383 if (doomed)
384 kfree(doomed->targets);
385 kfree(doomed);
386 }
387 spin_unlock_irqrestore(&devlist_lock, flags);
388
389 return 0;
390 }
391
392 int
aoedev_flush(const char __user * str,size_t cnt)393 aoedev_flush(const char __user *str, size_t cnt)
394 {
395 return flush(str, cnt, NOT_EXITING);
396 }
397
398 /* This has been confirmed to occur once with Tms=3*1000 due to the
399 * driver changing link and not processing its transmit ring. The
400 * problem is hard enough to solve by returning an error that I'm
401 * still punting on "solving" this.
402 */
403 static void
skbfree(struct sk_buff * skb)404 skbfree(struct sk_buff *skb)
405 {
406 enum { Sms = 250, Tms = 30 * 1000};
407 int i = Tms / Sms;
408
409 if (skb == NULL)
410 return;
411 while (atomic_read(&skb_shinfo(skb)->dataref) != 1 && i-- > 0)
412 msleep(Sms);
413 if (i < 0) {
414 printk(KERN_ERR
415 "aoe: %s holds ref: %s\n",
416 skb->dev ? skb->dev->name : "netif",
417 "cannot free skb -- memory leaked.");
418 return;
419 }
420 skb->truesize -= skb->data_len;
421 skb_shinfo(skb)->nr_frags = skb->data_len = 0;
422 skb_trim(skb, 0);
423 dev_kfree_skb(skb);
424 }
425
426 static void
skbpoolfree(struct aoedev * d)427 skbpoolfree(struct aoedev *d)
428 {
429 struct sk_buff *skb, *tmp;
430
431 skb_queue_walk_safe(&d->skbpool, skb, tmp)
432 skbfree(skb);
433
434 __skb_queue_head_init(&d->skbpool);
435 }
436
437 /* find it or allocate it */
438 struct aoedev *
aoedev_by_aoeaddr(ulong maj,int min,int do_alloc)439 aoedev_by_aoeaddr(ulong maj, int min, int do_alloc)
440 {
441 struct aoedev *d;
442 int i;
443 ulong flags;
444 ulong sysminor = 0;
445
446 spin_lock_irqsave(&devlist_lock, flags);
447
448 for (d=devlist; d; d=d->next)
449 if (d->aoemajor == maj && d->aoeminor == min) {
450 spin_lock(&d->lock);
451 if (d->flags & DEVFL_TKILL) {
452 spin_unlock(&d->lock);
453 d = NULL;
454 goto out;
455 }
456 d->ref++;
457 spin_unlock(&d->lock);
458 break;
459 }
460 if (d || !do_alloc || minor_get(&sysminor, maj, min) < 0)
461 goto out;
462 d = kcalloc(1, sizeof *d, GFP_ATOMIC);
463 if (!d)
464 goto out;
465 d->targets = kcalloc(NTARGETS, sizeof(*d->targets), GFP_ATOMIC);
466 if (!d->targets) {
467 kfree(d);
468 d = NULL;
469 goto out;
470 }
471 d->ntargets = NTARGETS;
472 INIT_WORK(&d->work, aoecmd_sleepwork);
473 spin_lock_init(&d->lock);
474 INIT_LIST_HEAD(&d->rq_list);
475 skb_queue_head_init(&d->skbpool);
476 timer_setup(&d->timer, dummy_timer, 0);
477 d->timer.expires = jiffies + HZ;
478 add_timer(&d->timer);
479 d->bufpool = NULL; /* defer to aoeblk_gdalloc */
480 d->tgt = d->targets;
481 d->ref = 1;
482 for (i = 0; i < NFACTIVE; i++)
483 INIT_LIST_HEAD(&d->factive[i]);
484 INIT_LIST_HEAD(&d->rexmitq);
485 d->sysminor = sysminor;
486 d->aoemajor = maj;
487 d->aoeminor = min;
488 d->rttavg = RTTAVG_INIT;
489 d->rttdev = RTTDEV_INIT;
490 d->next = devlist;
491 devlist = d;
492 out:
493 spin_unlock_irqrestore(&devlist_lock, flags);
494 return d;
495 }
496
497 static void
freetgt(struct aoedev * d,struct aoetgt * t)498 freetgt(struct aoedev *d, struct aoetgt *t)
499 {
500 struct frame *f;
501 struct list_head *pos, *nx, *head;
502 struct aoeif *ifp;
503
504 for (ifp = t->ifs; ifp < &t->ifs[NAOEIFS]; ++ifp) {
505 if (!ifp->nd)
506 break;
507 dev_put(ifp->nd);
508 }
509
510 head = &t->ffree;
511 list_for_each_safe(pos, nx, head) {
512 list_del(pos);
513 f = list_entry(pos, struct frame, head);
514 skbfree(f->skb);
515 kfree(f);
516 }
517 kfree(t);
518 }
519
520 void
aoedev_exit(void)521 aoedev_exit(void)
522 {
523 flush_scheduled_work();
524 flush(NULL, 0, EXITING);
525 }
526
527 int __init
aoedev_init(void)528 aoedev_init(void)
529 {
530 return 0;
531 }
532