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/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 <linux/string.h>
16#include "aoe.h"
17
18static void dummy_timer(ulong);
19static void freetgt(struct aoedev *d, struct aoetgt *t);
20static void skbpoolfree(struct aoedev *d);
21
22static int aoe_dyndevs = 1;
23module_param(aoe_dyndevs, int, 0644);
24MODULE_PARM_DESC(aoe_dyndevs, "Use dynamic minor numbers for devices.");
25
26static struct aoedev *devlist;
27static 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
38static DEFINE_SPINLOCK(used_minors_lock);
39static DECLARE_BITMAP(used_minors, N_DEVS);
40
41static int
42minor_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
60static int
61minor_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 * AOE_PARTITIONS;
99out:
100	return error;
101}
102
103static int
104minor_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
112static void
113minor_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
138void
139aoedev_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
148static void
149dummy_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
160static void
161aoe_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		bio->bi_error = -EIO;
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
182static void
183downdev_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
196void
197aoedev_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 + d->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
227	/* fast fail all pending I/O */
228	if (d->blkq) {
229		while ((rq = blk_peek_request(d->blkq))) {
230			blk_start_request(rq);
231			aoe_end_request(d, rq, 1);
232		}
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 */
242static int
243user_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
259static void
260freedev(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		aoedisk_rm_sysfs(d);
280		del_gendisk(d->gd);
281		put_disk(d->gd);
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	if (d->bufpool)
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
298enum flush_parms {
299	NOT_EXITING = 0,
300	EXITING = 1,
301};
302
303static int
304flush(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;
343cont:
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	 */
351restart:
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
386int
387aoedev_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 */
397static void
398skbfree(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
420static void
421skbpoolfree(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 */
432struct aoedev *
433aoedev_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	skb_queue_head_init(&d->skbpool);
469	init_timer(&d->timer);
470	d->timer.data = (ulong) d;
471	d->timer.function = dummy_timer;
472	d->timer.expires = jiffies + HZ;
473	add_timer(&d->timer);
474	d->bufpool = NULL;	/* defer to aoeblk_gdalloc */
475	d->tgt = d->targets;
476	d->ref = 1;
477	for (i = 0; i < NFACTIVE; i++)
478		INIT_LIST_HEAD(&d->factive[i]);
479	INIT_LIST_HEAD(&d->rexmitq);
480	d->sysminor = sysminor;
481	d->aoemajor = maj;
482	d->aoeminor = min;
483	d->rttavg = RTTAVG_INIT;
484	d->rttdev = RTTDEV_INIT;
485	d->next = devlist;
486	devlist = d;
487 out:
488	spin_unlock_irqrestore(&devlist_lock, flags);
489	return d;
490}
491
492static void
493freetgt(struct aoedev *d, struct aoetgt *t)
494{
495	struct frame *f;
496	struct list_head *pos, *nx, *head;
497	struct aoeif *ifp;
498
499	for (ifp = t->ifs; ifp < &t->ifs[NAOEIFS]; ++ifp) {
500		if (!ifp->nd)
501			break;
502		dev_put(ifp->nd);
503	}
504
505	head = &t->ffree;
506	list_for_each_safe(pos, nx, head) {
507		list_del(pos);
508		f = list_entry(pos, struct frame, head);
509		skbfree(f->skb);
510		kfree(f);
511	}
512	kfree(t);
513}
514
515void
516aoedev_exit(void)
517{
518	flush_scheduled_work();
519	flush(NULL, 0, EXITING);
520}
521
522int __init
523aoedev_init(void)
524{
525	return 0;
526}
527