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 struct request *rq, *rqnext;
202 int i;
203 unsigned long flags;
204
205 spin_lock_irqsave(&d->lock, flags);
206 d->flags &= ~(DEVFL_UP | DEVFL_DEAD);
207 spin_unlock_irqrestore(&d->lock, flags);
208
209 /* clean out active and to-be-retransmitted buffers */
210 for (i = 0; i < NFACTIVE; i++) {
211 head = &d->factive[i];
212 list_for_each_safe(pos, nx, head)
213 downdev_frame(pos);
214 }
215 head = &d->rexmitq;
216 list_for_each_safe(pos, nx, head)
217 downdev_frame(pos);
218
219 /* reset window dressings */
220 tt = d->targets;
221 te = tt + d->ntargets;
222 for (; tt < te && (t = *tt); tt++) {
223 aoecmd_wreset(t);
224 t->nout = 0;
225 }
226
227 /* clean out the in-process request (if any) */
228 aoe_failip(d);
229
230 /* clean out any queued block requests */
231 list_for_each_entry_safe(rq, rqnext, &d->rq_list, queuelist) {
232 list_del_init(&rq->queuelist);
233 blk_mq_start_request(rq);
234 blk_mq_end_request(rq, BLK_STS_IOERR);
235 }
236
237 /* fast fail all pending I/O */
238 if (d->blkq) {
239 /* UP is cleared, freeze+quiesce to insure all are errored */
240 blk_mq_freeze_queue(d->blkq);
241 blk_mq_quiesce_queue(d->blkq);
242 blk_mq_unquiesce_queue(d->blkq);
243 blk_mq_unfreeze_queue(d->blkq);
244 }
245
246 if (d->gd)
247 set_capacity(d->gd, 0);
248 }
249
250 /* return whether the user asked for this particular
251 * device to be flushed
252 */
253 static int
user_req(char * s,size_t slen,struct aoedev * d)254 user_req(char *s, size_t slen, struct aoedev *d)
255 {
256 const char *p;
257 size_t lim;
258
259 if (!d->gd)
260 return 0;
261 p = kbasename(d->gd->disk_name);
262 lim = sizeof(d->gd->disk_name);
263 lim -= p - d->gd->disk_name;
264 if (slen < lim)
265 lim = slen;
266
267 return !strncmp(s, p, lim);
268 }
269
270 static void
freedev(struct aoedev * d)271 freedev(struct aoedev *d)
272 {
273 struct aoetgt **t, **e;
274 int freeing = 0;
275 unsigned long flags;
276
277 spin_lock_irqsave(&d->lock, flags);
278 if (d->flags & DEVFL_TKILL
279 && !(d->flags & DEVFL_FREEING)) {
280 d->flags |= DEVFL_FREEING;
281 freeing = 1;
282 }
283 spin_unlock_irqrestore(&d->lock, flags);
284 if (!freeing)
285 return;
286
287 del_timer_sync(&d->timer);
288 if (d->gd) {
289 aoedisk_rm_debugfs(d);
290 del_gendisk(d->gd);
291 put_disk(d->gd);
292 blk_mq_free_tag_set(&d->tag_set);
293 }
294 t = d->targets;
295 e = t + d->ntargets;
296 for (; t < e && *t; t++)
297 freetgt(d, *t);
298
299 mempool_destroy(d->bufpool);
300 skbpoolfree(d);
301 minor_free(d->sysminor);
302
303 spin_lock_irqsave(&d->lock, flags);
304 d->flags |= DEVFL_FREED;
305 spin_unlock_irqrestore(&d->lock, flags);
306 }
307
308 enum flush_parms {
309 NOT_EXITING = 0,
310 EXITING = 1,
311 };
312
313 static int
flush(const char __user * str,size_t cnt,int exiting)314 flush(const char __user *str, size_t cnt, int exiting)
315 {
316 ulong flags;
317 struct aoedev *d, **dd;
318 char buf[16];
319 int all = 0;
320 int specified = 0; /* flush a specific device */
321 unsigned int skipflags;
322
323 skipflags = DEVFL_GDALLOC | DEVFL_NEWSIZE | DEVFL_TKILL;
324
325 if (!exiting && cnt >= 3) {
326 if (cnt > sizeof buf)
327 cnt = sizeof buf;
328 if (copy_from_user(buf, str, cnt))
329 return -EFAULT;
330 all = !strncmp(buf, "all", 3);
331 if (!all)
332 specified = 1;
333 }
334
335 flush_workqueue(aoe_wq);
336 /* pass one: do aoedev_downdev, which might sleep */
337 restart1:
338 spin_lock_irqsave(&devlist_lock, flags);
339 for (d = devlist; d; d = d->next) {
340 spin_lock(&d->lock);
341 if (d->flags & DEVFL_TKILL)
342 goto cont;
343
344 if (exiting) {
345 /* unconditionally take each device down */
346 } else if (specified) {
347 if (!user_req(buf, cnt, d))
348 goto cont;
349 } else if ((!all && (d->flags & DEVFL_UP))
350 || d->flags & skipflags
351 || d->nopen
352 || d->ref)
353 goto cont;
354
355 spin_unlock(&d->lock);
356 spin_unlock_irqrestore(&devlist_lock, flags);
357 aoedev_downdev(d);
358 d->flags |= DEVFL_TKILL;
359 goto restart1;
360 cont:
361 spin_unlock(&d->lock);
362 }
363 spin_unlock_irqrestore(&devlist_lock, flags);
364
365 /* pass two: call freedev, which might sleep,
366 * for aoedevs marked with DEVFL_TKILL
367 */
368 restart2:
369 spin_lock_irqsave(&devlist_lock, flags);
370 for (d = devlist; d; d = d->next) {
371 spin_lock(&d->lock);
372 if (d->flags & DEVFL_TKILL
373 && !(d->flags & DEVFL_FREEING)) {
374 spin_unlock(&d->lock);
375 spin_unlock_irqrestore(&devlist_lock, flags);
376 freedev(d);
377 goto restart2;
378 }
379 spin_unlock(&d->lock);
380 }
381
382 /* pass three: remove aoedevs marked with DEVFL_FREED */
383 for (dd = &devlist, d = *dd; d; d = *dd) {
384 struct aoedev *doomed = NULL;
385
386 spin_lock(&d->lock);
387 if (d->flags & DEVFL_FREED) {
388 *dd = d->next;
389 doomed = d;
390 } else {
391 dd = &d->next;
392 }
393 spin_unlock(&d->lock);
394 if (doomed)
395 kfree(doomed->targets);
396 kfree(doomed);
397 }
398 spin_unlock_irqrestore(&devlist_lock, flags);
399
400 return 0;
401 }
402
403 int
aoedev_flush(const char __user * str,size_t cnt)404 aoedev_flush(const char __user *str, size_t cnt)
405 {
406 return flush(str, cnt, NOT_EXITING);
407 }
408
409 /* This has been confirmed to occur once with Tms=3*1000 due to the
410 * driver changing link and not processing its transmit ring. The
411 * problem is hard enough to solve by returning an error that I'm
412 * still punting on "solving" this.
413 */
414 static void
skbfree(struct sk_buff * skb)415 skbfree(struct sk_buff *skb)
416 {
417 enum { Sms = 250, Tms = 30 * 1000};
418 int i = Tms / Sms;
419
420 if (skb == NULL)
421 return;
422 while (atomic_read(&skb_shinfo(skb)->dataref) != 1 && i-- > 0)
423 msleep(Sms);
424 if (i < 0) {
425 printk(KERN_ERR
426 "aoe: %s holds ref: %s\n",
427 skb->dev ? skb->dev->name : "netif",
428 "cannot free skb -- memory leaked.");
429 return;
430 }
431 skb->truesize -= skb->data_len;
432 skb_shinfo(skb)->nr_frags = skb->data_len = 0;
433 skb_trim(skb, 0);
434 dev_kfree_skb(skb);
435 }
436
437 static void
skbpoolfree(struct aoedev * d)438 skbpoolfree(struct aoedev *d)
439 {
440 struct sk_buff *skb, *tmp;
441
442 skb_queue_walk_safe(&d->skbpool, skb, tmp)
443 skbfree(skb);
444
445 __skb_queue_head_init(&d->skbpool);
446 }
447
448 /* find it or allocate it */
449 struct aoedev *
aoedev_by_aoeaddr(ulong maj,int min,int do_alloc)450 aoedev_by_aoeaddr(ulong maj, int min, int do_alloc)
451 {
452 struct aoedev *d;
453 int i;
454 ulong flags;
455 ulong sysminor = 0;
456
457 spin_lock_irqsave(&devlist_lock, flags);
458
459 for (d=devlist; d; d=d->next)
460 if (d->aoemajor == maj && d->aoeminor == min) {
461 spin_lock(&d->lock);
462 if (d->flags & DEVFL_TKILL) {
463 spin_unlock(&d->lock);
464 d = NULL;
465 goto out;
466 }
467 d->ref++;
468 spin_unlock(&d->lock);
469 break;
470 }
471 if (d || !do_alloc || minor_get(&sysminor, maj, min) < 0)
472 goto out;
473 d = kcalloc(1, sizeof *d, GFP_ATOMIC);
474 if (!d)
475 goto out;
476 d->targets = kcalloc(NTARGETS, sizeof(*d->targets), GFP_ATOMIC);
477 if (!d->targets) {
478 kfree(d);
479 d = NULL;
480 goto out;
481 }
482 d->ntargets = NTARGETS;
483 INIT_WORK(&d->work, aoecmd_sleepwork);
484 spin_lock_init(&d->lock);
485 INIT_LIST_HEAD(&d->rq_list);
486 skb_queue_head_init(&d->skbpool);
487 timer_setup(&d->timer, dummy_timer, 0);
488 d->timer.expires = jiffies + HZ;
489 add_timer(&d->timer);
490 d->bufpool = NULL; /* defer to aoeblk_gdalloc */
491 d->tgt = d->targets;
492 d->ref = 1;
493 for (i = 0; i < NFACTIVE; i++)
494 INIT_LIST_HEAD(&d->factive[i]);
495 INIT_LIST_HEAD(&d->rexmitq);
496 d->sysminor = sysminor;
497 d->aoemajor = maj;
498 d->aoeminor = min;
499 d->rttavg = RTTAVG_INIT;
500 d->rttdev = RTTDEV_INIT;
501 d->next = devlist;
502 devlist = d;
503 out:
504 spin_unlock_irqrestore(&devlist_lock, flags);
505 return d;
506 }
507
508 static void
freetgt(struct aoedev * d,struct aoetgt * t)509 freetgt(struct aoedev *d, struct aoetgt *t)
510 {
511 struct frame *f;
512 struct list_head *pos, *nx, *head;
513 struct aoeif *ifp;
514
515 for (ifp = t->ifs; ifp < &t->ifs[NAOEIFS]; ++ifp) {
516 if (!ifp->nd)
517 break;
518 dev_put(ifp->nd);
519 }
520
521 head = &t->ffree;
522 list_for_each_safe(pos, nx, head) {
523 list_del(pos);
524 f = list_entry(pos, struct frame, head);
525 skbfree(f->skb);
526 kfree(f);
527 }
528 kfree(t);
529 }
530
531 void
aoedev_exit(void)532 aoedev_exit(void)
533 {
534 flush_workqueue(aoe_wq);
535 flush(NULL, 0, EXITING);
536 }
537
538 int __init
aoedev_init(void)539 aoedev_init(void)
540 {
541 return 0;
542 }
543