xref: /openbmc/linux/drivers/block/aoe/aoedev.c (revision e0b2bbab)
1 /* Copyright (c) 2012 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/blkdev.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 "aoe.h"
16 
17 static void dummy_timer(ulong);
18 static void aoedev_freedev(struct aoedev *);
19 static void freetgt(struct aoedev *d, struct aoetgt *t);
20 static void skbpoolfree(struct aoedev *d);
21 
22 static int aoe_dyndevs = 1;
23 module_param(aoe_dyndevs, int, 0644);
24 MODULE_PARM_DESC(aoe_dyndevs, "Use dynamic minor numbers for devices.");
25 
26 static struct aoedev *devlist;
27 static DEFINE_SPINLOCK(devlist_lock);
28 
29 /* Because some systems will have one, many, or no
30  *   - partitions,
31  *   - slots per shelf,
32  *   - or shelves,
33  * we need some flexibility in the way the minor numbers
34  * are allocated.  So they are dynamic.
35  */
36 #define N_DEVS ((1U<<MINORBITS)/AOE_PARTITIONS)
37 
38 static DEFINE_SPINLOCK(used_minors_lock);
39 static DECLARE_BITMAP(used_minors, N_DEVS);
40 
41 static int
42 minor_get_dyn(ulong *sysminor)
43 {
44 	ulong flags;
45 	ulong n;
46 	int error = 0;
47 
48 	spin_lock_irqsave(&used_minors_lock, flags);
49 	n = find_first_zero_bit(used_minors, N_DEVS);
50 	if (n < N_DEVS)
51 		set_bit(n, used_minors);
52 	else
53 		error = -1;
54 	spin_unlock_irqrestore(&used_minors_lock, flags);
55 
56 	*sysminor = n * AOE_PARTITIONS;
57 	return error;
58 }
59 
60 static int
61 minor_get_static(ulong *sysminor, ulong aoemaj, int aoemin)
62 {
63 	ulong flags;
64 	ulong n;
65 	int error = 0;
66 	enum {
67 		/* for backwards compatibility when !aoe_dyndevs,
68 		 * a static number of supported slots per shelf */
69 		NPERSHELF = 16,
70 	};
71 
72 	if (aoemin >= NPERSHELF) {
73 		pr_err("aoe: %s %d slots per shelf\n",
74 			"static minor device numbers support only",
75 			NPERSHELF);
76 		error = -1;
77 		goto out;
78 	}
79 
80 	n = aoemaj * NPERSHELF + aoemin;
81 	if (n >= N_DEVS) {
82 		pr_err("aoe: %s with e%ld.%d\n",
83 			"cannot use static minor device numbers",
84 			aoemaj, aoemin);
85 		error = -1;
86 		goto out;
87 	}
88 
89 	spin_lock_irqsave(&used_minors_lock, flags);
90 	if (test_bit(n, used_minors)) {
91 		pr_err("aoe: %s %lu\n",
92 			"existing device already has static minor number",
93 			n);
94 		error = -1;
95 	} else
96 		set_bit(n, used_minors);
97 	spin_unlock_irqrestore(&used_minors_lock, flags);
98 	*sysminor = n;
99 out:
100 	return error;
101 }
102 
103 static int
104 minor_get(ulong *sysminor, ulong aoemaj, int aoemin)
105 {
106 	if (aoe_dyndevs)
107 		return minor_get_dyn(sysminor);
108 	else
109 		return minor_get_static(sysminor, aoemaj, aoemin);
110 }
111 
112 static void
113 minor_free(ulong minor)
114 {
115 	ulong flags;
116 
117 	minor /= AOE_PARTITIONS;
118 	BUG_ON(minor >= N_DEVS);
119 
120 	spin_lock_irqsave(&used_minors_lock, flags);
121 	BUG_ON(!test_bit(minor, used_minors));
122 	clear_bit(minor, used_minors);
123 	spin_unlock_irqrestore(&used_minors_lock, flags);
124 }
125 
126 /*
127  * Users who grab a pointer to the device with aoedev_by_aoeaddr
128  * automatically get a reference count and must be responsible
129  * for performing a aoedev_put.  With the addition of async
130  * kthread processing I'm no longer confident that we can
131  * guarantee consistency in the face of device flushes.
132  *
133  * For the time being, we only bother to add extra references for
134  * frames sitting on the iocq.  When the kthreads finish processing
135  * these frames, they will aoedev_put the device.
136  */
137 
138 void
139 aoedev_put(struct aoedev *d)
140 {
141 	ulong flags;
142 
143 	spin_lock_irqsave(&devlist_lock, flags);
144 	d->ref--;
145 	spin_unlock_irqrestore(&devlist_lock, flags);
146 }
147 
148 static void
149 dummy_timer(ulong vp)
150 {
151 	struct aoedev *d;
152 
153 	d = (struct aoedev *)vp;
154 	if (d->flags & DEVFL_TKILL)
155 		return;
156 	d->timer.expires = jiffies + HZ;
157 	add_timer(&d->timer);
158 }
159 
160 static void
161 aoe_failip(struct aoedev *d)
162 {
163 	struct request *rq;
164 	struct bio *bio;
165 	unsigned long n;
166 
167 	aoe_failbuf(d, d->ip.buf);
168 
169 	rq = d->ip.rq;
170 	if (rq == NULL)
171 		return;
172 	while ((bio = d->ip.nxbio)) {
173 		clear_bit(BIO_UPTODATE, &bio->bi_flags);
174 		d->ip.nxbio = bio->bi_next;
175 		n = (unsigned long) rq->special;
176 		rq->special = (void *) --n;
177 	}
178 	if ((unsigned long) rq->special == 0)
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 	struct request *rq;
202 	int i;
203 
204 	d->flags &= ~DEVFL_UP;
205 
206 	/* clean out active and to-be-retransmitted buffers */
207 	for (i = 0; i < NFACTIVE; i++) {
208 		head = &d->factive[i];
209 		list_for_each_safe(pos, nx, head)
210 			downdev_frame(pos);
211 	}
212 	head = &d->rexmitq;
213 	list_for_each_safe(pos, nx, head)
214 		downdev_frame(pos);
215 
216 	/* reset window dressings */
217 	tt = d->targets;
218 	te = tt + NTARGETS;
219 	for (; tt < te && (t = *tt); tt++) {
220 		aoecmd_wreset(t);
221 		t->nout = 0;
222 	}
223 
224 	/* clean out the in-process request (if any) */
225 	aoe_failip(d);
226 	d->htgt = NULL;
227 
228 	/* fast fail all pending I/O */
229 	if (d->blkq) {
230 		while ((rq = blk_peek_request(d->blkq))) {
231 			blk_start_request(rq);
232 			aoe_end_request(d, rq, 1);
233 		}
234 	}
235 
236 	if (d->gd)
237 		set_capacity(d->gd, 0);
238 }
239 
240 static void
241 aoedev_freedev(struct aoedev *d)
242 {
243 	struct aoetgt **t, **e;
244 
245 	cancel_work_sync(&d->work);
246 	if (d->gd) {
247 		aoedisk_rm_sysfs(d);
248 		del_gendisk(d->gd);
249 		put_disk(d->gd);
250 		blk_cleanup_queue(d->blkq);
251 	}
252 	t = d->targets;
253 	e = t + NTARGETS;
254 	for (; t < e && *t; t++)
255 		freetgt(d, *t);
256 	if (d->bufpool)
257 		mempool_destroy(d->bufpool);
258 	skbpoolfree(d);
259 	minor_free(d->sysminor);
260 	kfree(d);
261 }
262 
263 /* return whether the user asked for this particular
264  * device to be flushed
265  */
266 static int
267 user_req(char *s, size_t slen, struct aoedev *d)
268 {
269 	char *p;
270 	size_t lim;
271 
272 	if (!d->gd)
273 		return 0;
274 	p = strrchr(d->gd->disk_name, '/');
275 	if (!p)
276 		p = d->gd->disk_name;
277 	else
278 		p += 1;
279 	lim = sizeof(d->gd->disk_name);
280 	lim -= p - d->gd->disk_name;
281 	if (slen < lim)
282 		lim = slen;
283 
284 	return !strncmp(s, p, lim);
285 }
286 
287 int
288 aoedev_flush(const char __user *str, size_t cnt)
289 {
290 	ulong flags;
291 	struct aoedev *d, **dd;
292 	struct aoedev *rmd = NULL;
293 	char buf[16];
294 	int all = 0;
295 	int specified = 0;	/* flush a specific device */
296 
297 	if (cnt >= 3) {
298 		if (cnt > sizeof buf)
299 			cnt = sizeof buf;
300 		if (copy_from_user(buf, str, cnt))
301 			return -EFAULT;
302 		all = !strncmp(buf, "all", 3);
303 		if (!all)
304 			specified = 1;
305 	}
306 
307 	spin_lock_irqsave(&devlist_lock, flags);
308 	dd = &devlist;
309 	while ((d = *dd)) {
310 		spin_lock(&d->lock);
311 		if (specified) {
312 			if (!user_req(buf, cnt, d))
313 				goto skip;
314 		} else if ((!all && (d->flags & DEVFL_UP))
315 		|| (d->flags & (DEVFL_GDALLOC|DEVFL_NEWSIZE))
316 		|| d->nopen
317 		|| d->ref)
318 			goto skip;
319 
320 		*dd = d->next;
321 		aoedev_downdev(d);
322 		d->flags |= DEVFL_TKILL;
323 		spin_unlock(&d->lock);
324 		d->next = rmd;
325 		rmd = d;
326 		continue;
327 skip:
328 		spin_unlock(&d->lock);
329 		dd = &d->next;
330 	}
331 	spin_unlock_irqrestore(&devlist_lock, flags);
332 	while ((d = rmd)) {
333 		rmd = d->next;
334 		del_timer_sync(&d->timer);
335 		aoedev_freedev(d);	/* must be able to sleep */
336 	}
337 	return 0;
338 }
339 
340 /* This has been confirmed to occur once with Tms=3*1000 due to the
341  * driver changing link and not processing its transmit ring.  The
342  * problem is hard enough to solve by returning an error that I'm
343  * still punting on "solving" this.
344  */
345 static void
346 skbfree(struct sk_buff *skb)
347 {
348 	enum { Sms = 250, Tms = 30 * 1000};
349 	int i = Tms / Sms;
350 
351 	if (skb == NULL)
352 		return;
353 	while (atomic_read(&skb_shinfo(skb)->dataref) != 1 && i-- > 0)
354 		msleep(Sms);
355 	if (i < 0) {
356 		printk(KERN_ERR
357 			"aoe: %s holds ref: %s\n",
358 			skb->dev ? skb->dev->name : "netif",
359 			"cannot free skb -- memory leaked.");
360 		return;
361 	}
362 	skb->truesize -= skb->data_len;
363 	skb_shinfo(skb)->nr_frags = skb->data_len = 0;
364 	skb_trim(skb, 0);
365 	dev_kfree_skb(skb);
366 }
367 
368 static void
369 skbpoolfree(struct aoedev *d)
370 {
371 	struct sk_buff *skb, *tmp;
372 
373 	skb_queue_walk_safe(&d->skbpool, skb, tmp)
374 		skbfree(skb);
375 
376 	__skb_queue_head_init(&d->skbpool);
377 }
378 
379 /* find it or allocate it */
380 struct aoedev *
381 aoedev_by_aoeaddr(ulong maj, int min, int do_alloc)
382 {
383 	struct aoedev *d;
384 	int i;
385 	ulong flags;
386 	ulong sysminor;
387 
388 	spin_lock_irqsave(&devlist_lock, flags);
389 
390 	for (d=devlist; d; d=d->next)
391 		if (d->aoemajor == maj && d->aoeminor == min) {
392 			d->ref++;
393 			break;
394 		}
395 	if (d || !do_alloc || minor_get(&sysminor, maj, min) < 0)
396 		goto out;
397 	d = kcalloc(1, sizeof *d, GFP_ATOMIC);
398 	if (!d)
399 		goto out;
400 	INIT_WORK(&d->work, aoecmd_sleepwork);
401 	spin_lock_init(&d->lock);
402 	skb_queue_head_init(&d->skbpool);
403 	init_timer(&d->timer);
404 	d->timer.data = (ulong) d;
405 	d->timer.function = dummy_timer;
406 	d->timer.expires = jiffies + HZ;
407 	add_timer(&d->timer);
408 	d->bufpool = NULL;	/* defer to aoeblk_gdalloc */
409 	d->tgt = d->targets;
410 	d->ref = 1;
411 	for (i = 0; i < NFACTIVE; i++)
412 		INIT_LIST_HEAD(&d->factive[i]);
413 	INIT_LIST_HEAD(&d->rexmitq);
414 	d->sysminor = sysminor;
415 	d->aoemajor = maj;
416 	d->aoeminor = min;
417 	d->rttavg = RTTAVG_INIT;
418 	d->rttdev = RTTDEV_INIT;
419 	d->next = devlist;
420 	devlist = d;
421  out:
422 	spin_unlock_irqrestore(&devlist_lock, flags);
423 	return d;
424 }
425 
426 static void
427 freetgt(struct aoedev *d, struct aoetgt *t)
428 {
429 	struct frame *f;
430 	struct list_head *pos, *nx, *head;
431 	struct aoeif *ifp;
432 
433 	for (ifp = t->ifs; ifp < &t->ifs[NAOEIFS]; ++ifp) {
434 		if (!ifp->nd)
435 			break;
436 		dev_put(ifp->nd);
437 	}
438 
439 	head = &t->ffree;
440 	list_for_each_safe(pos, nx, head) {
441 		list_del(pos);
442 		f = list_entry(pos, struct frame, head);
443 		skbfree(f->skb);
444 		kfree(f);
445 	}
446 	kfree(t);
447 }
448 
449 void
450 aoedev_exit(void)
451 {
452 	struct aoedev *d;
453 	ulong flags;
454 
455 	aoe_flush_iocq();
456 	while ((d = devlist)) {
457 		devlist = d->next;
458 
459 		spin_lock_irqsave(&d->lock, flags);
460 		aoedev_downdev(d);
461 		d->flags |= DEVFL_TKILL;
462 		spin_unlock_irqrestore(&d->lock, flags);
463 
464 		del_timer_sync(&d->timer);
465 		aoedev_freedev(d);
466 	}
467 }
468 
469 int __init
470 aoedev_init(void)
471 {
472 	return 0;
473 }
474