xref: /linux/drivers/s390/block/dcssblk.c (revision fcc8487d477a3452a1d0ccbdd4c5e0e1e3cb8bed)
1 /*
2  * dcssblk.c -- the S/390 block driver for dcss memory
3  *
4  * Authors: Carsten Otte, Stefan Weinhuber, Gerald Schaefer
5  */
6 
7 #define KMSG_COMPONENT "dcssblk"
8 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
9 
10 #include <linux/module.h>
11 #include <linux/moduleparam.h>
12 #include <linux/ctype.h>
13 #include <linux/errno.h>
14 #include <linux/init.h>
15 #include <linux/slab.h>
16 #include <linux/blkdev.h>
17 #include <linux/completion.h>
18 #include <linux/interrupt.h>
19 #include <linux/platform_device.h>
20 #include <linux/pfn_t.h>
21 #include <linux/dax.h>
22 #include <asm/extmem.h>
23 #include <asm/io.h>
24 
25 #define DCSSBLK_NAME "dcssblk"
26 #define DCSSBLK_MINORS_PER_DISK 1
27 #define DCSSBLK_PARM_LEN 400
28 #define DCSS_BUS_ID_SIZE 20
29 
30 static int dcssblk_open(struct block_device *bdev, fmode_t mode);
31 static void dcssblk_release(struct gendisk *disk, fmode_t mode);
32 static blk_qc_t dcssblk_make_request(struct request_queue *q,
33 						struct bio *bio);
34 static long dcssblk_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
35 		long nr_pages, void **kaddr, pfn_t *pfn);
36 
37 static char dcssblk_segments[DCSSBLK_PARM_LEN] = "\0";
38 
39 static int dcssblk_major;
40 static const struct block_device_operations dcssblk_devops = {
41 	.owner   	= THIS_MODULE,
42 	.open    	= dcssblk_open,
43 	.release 	= dcssblk_release,
44 };
45 
46 static const struct dax_operations dcssblk_dax_ops = {
47 	.direct_access = dcssblk_dax_direct_access,
48 };
49 
50 struct dcssblk_dev_info {
51 	struct list_head lh;
52 	struct device dev;
53 	char segment_name[DCSS_BUS_ID_SIZE];
54 	atomic_t use_count;
55 	struct gendisk *gd;
56 	unsigned long start;
57 	unsigned long end;
58 	int segment_type;
59 	unsigned char save_pending;
60 	unsigned char is_shared;
61 	struct request_queue *dcssblk_queue;
62 	int num_of_segments;
63 	struct list_head seg_list;
64 	struct dax_device *dax_dev;
65 };
66 
67 struct segment_info {
68 	struct list_head lh;
69 	char segment_name[DCSS_BUS_ID_SIZE];
70 	unsigned long start;
71 	unsigned long end;
72 	int segment_type;
73 };
74 
75 static ssize_t dcssblk_add_store(struct device * dev, struct device_attribute *attr, const char * buf,
76 				  size_t count);
77 static ssize_t dcssblk_remove_store(struct device * dev, struct device_attribute *attr, const char * buf,
78 				  size_t count);
79 
80 static DEVICE_ATTR(add, S_IWUSR, NULL, dcssblk_add_store);
81 static DEVICE_ATTR(remove, S_IWUSR, NULL, dcssblk_remove_store);
82 
83 static struct device *dcssblk_root_dev;
84 
85 static LIST_HEAD(dcssblk_devices);
86 static struct rw_semaphore dcssblk_devices_sem;
87 
88 /*
89  * release function for segment device.
90  */
91 static void
92 dcssblk_release_segment(struct device *dev)
93 {
94 	struct dcssblk_dev_info *dev_info;
95 	struct segment_info *entry, *temp;
96 
97 	dev_info = container_of(dev, struct dcssblk_dev_info, dev);
98 	list_for_each_entry_safe(entry, temp, &dev_info->seg_list, lh) {
99 		list_del(&entry->lh);
100 		kfree(entry);
101 	}
102 	kfree(dev_info);
103 	module_put(THIS_MODULE);
104 }
105 
106 /*
107  * get a minor number. needs to be called with
108  * down_write(&dcssblk_devices_sem) and the
109  * device needs to be enqueued before the semaphore is
110  * freed.
111  */
112 static int
113 dcssblk_assign_free_minor(struct dcssblk_dev_info *dev_info)
114 {
115 	int minor, found;
116 	struct dcssblk_dev_info *entry;
117 
118 	if (dev_info == NULL)
119 		return -EINVAL;
120 	for (minor = 0; minor < (1<<MINORBITS); minor++) {
121 		found = 0;
122 		// test if minor available
123 		list_for_each_entry(entry, &dcssblk_devices, lh)
124 			if (minor == entry->gd->first_minor)
125 				found++;
126 		if (!found) break; // got unused minor
127 	}
128 	if (found)
129 		return -EBUSY;
130 	dev_info->gd->first_minor = minor;
131 	return 0;
132 }
133 
134 /*
135  * get the struct dcssblk_dev_info from dcssblk_devices
136  * for the given name.
137  * down_read(&dcssblk_devices_sem) must be held.
138  */
139 static struct dcssblk_dev_info *
140 dcssblk_get_device_by_name(char *name)
141 {
142 	struct dcssblk_dev_info *entry;
143 
144 	list_for_each_entry(entry, &dcssblk_devices, lh) {
145 		if (!strcmp(name, entry->segment_name)) {
146 			return entry;
147 		}
148 	}
149 	return NULL;
150 }
151 
152 /*
153  * get the struct segment_info from seg_list
154  * for the given name.
155  * down_read(&dcssblk_devices_sem) must be held.
156  */
157 static struct segment_info *
158 dcssblk_get_segment_by_name(char *name)
159 {
160 	struct dcssblk_dev_info *dev_info;
161 	struct segment_info *entry;
162 
163 	list_for_each_entry(dev_info, &dcssblk_devices, lh) {
164 		list_for_each_entry(entry, &dev_info->seg_list, lh) {
165 			if (!strcmp(name, entry->segment_name))
166 				return entry;
167 		}
168 	}
169 	return NULL;
170 }
171 
172 /*
173  * get the highest address of the multi-segment block.
174  */
175 static unsigned long
176 dcssblk_find_highest_addr(struct dcssblk_dev_info *dev_info)
177 {
178 	unsigned long highest_addr;
179 	struct segment_info *entry;
180 
181 	highest_addr = 0;
182 	list_for_each_entry(entry, &dev_info->seg_list, lh) {
183 		if (highest_addr < entry->end)
184 			highest_addr = entry->end;
185 	}
186 	return highest_addr;
187 }
188 
189 /*
190  * get the lowest address of the multi-segment block.
191  */
192 static unsigned long
193 dcssblk_find_lowest_addr(struct dcssblk_dev_info *dev_info)
194 {
195 	int set_first;
196 	unsigned long lowest_addr;
197 	struct segment_info *entry;
198 
199 	set_first = 0;
200 	lowest_addr = 0;
201 	list_for_each_entry(entry, &dev_info->seg_list, lh) {
202 		if (set_first == 0) {
203 			lowest_addr = entry->start;
204 			set_first = 1;
205 		} else {
206 			if (lowest_addr > entry->start)
207 				lowest_addr = entry->start;
208 		}
209 	}
210 	return lowest_addr;
211 }
212 
213 /*
214  * Check continuity of segments.
215  */
216 static int
217 dcssblk_is_continuous(struct dcssblk_dev_info *dev_info)
218 {
219 	int i, j, rc;
220 	struct segment_info *sort_list, *entry, temp;
221 
222 	if (dev_info->num_of_segments <= 1)
223 		return 0;
224 
225 	sort_list = kzalloc(
226 			sizeof(struct segment_info) * dev_info->num_of_segments,
227 			GFP_KERNEL);
228 	if (sort_list == NULL)
229 		return -ENOMEM;
230 	i = 0;
231 	list_for_each_entry(entry, &dev_info->seg_list, lh) {
232 		memcpy(&sort_list[i], entry, sizeof(struct segment_info));
233 		i++;
234 	}
235 
236 	/* sort segments */
237 	for (i = 0; i < dev_info->num_of_segments; i++)
238 		for (j = 0; j < dev_info->num_of_segments; j++)
239 			if (sort_list[j].start > sort_list[i].start) {
240 				memcpy(&temp, &sort_list[i],
241 					sizeof(struct segment_info));
242 				memcpy(&sort_list[i], &sort_list[j],
243 					sizeof(struct segment_info));
244 				memcpy(&sort_list[j], &temp,
245 					sizeof(struct segment_info));
246 			}
247 
248 	/* check continuity */
249 	for (i = 0; i < dev_info->num_of_segments - 1; i++) {
250 		if ((sort_list[i].end + 1) != sort_list[i+1].start) {
251 			pr_err("Adjacent DCSSs %s and %s are not "
252 			       "contiguous\n", sort_list[i].segment_name,
253 			       sort_list[i+1].segment_name);
254 			rc = -EINVAL;
255 			goto out;
256 		}
257 		/* EN and EW are allowed in a block device */
258 		if (sort_list[i].segment_type != sort_list[i+1].segment_type) {
259 			if (!(sort_list[i].segment_type & SEGMENT_EXCLUSIVE) ||
260 				(sort_list[i].segment_type == SEG_TYPE_ER) ||
261 				!(sort_list[i+1].segment_type &
262 				SEGMENT_EXCLUSIVE) ||
263 				(sort_list[i+1].segment_type == SEG_TYPE_ER)) {
264 				pr_err("DCSS %s and DCSS %s have "
265 				       "incompatible types\n",
266 				       sort_list[i].segment_name,
267 				       sort_list[i+1].segment_name);
268 				rc = -EINVAL;
269 				goto out;
270 			}
271 		}
272 	}
273 	rc = 0;
274 out:
275 	kfree(sort_list);
276 	return rc;
277 }
278 
279 /*
280  * Load a segment
281  */
282 static int
283 dcssblk_load_segment(char *name, struct segment_info **seg_info)
284 {
285 	int rc;
286 
287 	/* already loaded? */
288 	down_read(&dcssblk_devices_sem);
289 	*seg_info = dcssblk_get_segment_by_name(name);
290 	up_read(&dcssblk_devices_sem);
291 	if (*seg_info != NULL)
292 		return -EEXIST;
293 
294 	/* get a struct segment_info */
295 	*seg_info = kzalloc(sizeof(struct segment_info), GFP_KERNEL);
296 	if (*seg_info == NULL)
297 		return -ENOMEM;
298 
299 	strcpy((*seg_info)->segment_name, name);
300 
301 	/* load the segment */
302 	rc = segment_load(name, SEGMENT_SHARED,
303 			&(*seg_info)->start, &(*seg_info)->end);
304 	if (rc < 0) {
305 		segment_warning(rc, (*seg_info)->segment_name);
306 		kfree(*seg_info);
307 	} else {
308 		INIT_LIST_HEAD(&(*seg_info)->lh);
309 		(*seg_info)->segment_type = rc;
310 	}
311 	return rc;
312 }
313 
314 /*
315  * device attribute for switching shared/nonshared (exclusive)
316  * operation (show + store)
317  */
318 static ssize_t
319 dcssblk_shared_show(struct device *dev, struct device_attribute *attr, char *buf)
320 {
321 	struct dcssblk_dev_info *dev_info;
322 
323 	dev_info = container_of(dev, struct dcssblk_dev_info, dev);
324 	return sprintf(buf, dev_info->is_shared ? "1\n" : "0\n");
325 }
326 
327 static ssize_t
328 dcssblk_shared_store(struct device *dev, struct device_attribute *attr, const char *inbuf, size_t count)
329 {
330 	struct dcssblk_dev_info *dev_info;
331 	struct segment_info *entry, *temp;
332 	int rc;
333 
334 	if ((count > 1) && (inbuf[1] != '\n') && (inbuf[1] != '\0'))
335 		return -EINVAL;
336 	down_write(&dcssblk_devices_sem);
337 	dev_info = container_of(dev, struct dcssblk_dev_info, dev);
338 	if (atomic_read(&dev_info->use_count)) {
339 		rc = -EBUSY;
340 		goto out;
341 	}
342 	if (inbuf[0] == '1') {
343 		/* reload segments in shared mode */
344 		list_for_each_entry(entry, &dev_info->seg_list, lh) {
345 			rc = segment_modify_shared(entry->segment_name,
346 						SEGMENT_SHARED);
347 			if (rc < 0) {
348 				BUG_ON(rc == -EINVAL);
349 				if (rc != -EAGAIN)
350 					goto removeseg;
351 			}
352 		}
353 		dev_info->is_shared = 1;
354 		switch (dev_info->segment_type) {
355 		case SEG_TYPE_SR:
356 		case SEG_TYPE_ER:
357 		case SEG_TYPE_SC:
358 			set_disk_ro(dev_info->gd, 1);
359 		}
360 	} else if (inbuf[0] == '0') {
361 		/* reload segments in exclusive mode */
362 		if (dev_info->segment_type == SEG_TYPE_SC) {
363 			pr_err("DCSS %s is of type SC and cannot be "
364 			       "loaded as exclusive-writable\n",
365 			       dev_info->segment_name);
366 			rc = -EINVAL;
367 			goto out;
368 		}
369 		list_for_each_entry(entry, &dev_info->seg_list, lh) {
370 			rc = segment_modify_shared(entry->segment_name,
371 						   SEGMENT_EXCLUSIVE);
372 			if (rc < 0) {
373 				BUG_ON(rc == -EINVAL);
374 				if (rc != -EAGAIN)
375 					goto removeseg;
376 			}
377 		}
378 		dev_info->is_shared = 0;
379 		set_disk_ro(dev_info->gd, 0);
380 	} else {
381 		rc = -EINVAL;
382 		goto out;
383 	}
384 	rc = count;
385 	goto out;
386 
387 removeseg:
388 	pr_err("DCSS device %s is removed after a failed access mode "
389 	       "change\n", dev_info->segment_name);
390 	temp = entry;
391 	list_for_each_entry(entry, &dev_info->seg_list, lh) {
392 		if (entry != temp)
393 			segment_unload(entry->segment_name);
394 	}
395 	list_del(&dev_info->lh);
396 
397 	kill_dax(dev_info->dax_dev);
398 	put_dax(dev_info->dax_dev);
399 	del_gendisk(dev_info->gd);
400 	blk_cleanup_queue(dev_info->dcssblk_queue);
401 	dev_info->gd->queue = NULL;
402 	put_disk(dev_info->gd);
403 	up_write(&dcssblk_devices_sem);
404 
405 	if (device_remove_file_self(dev, attr)) {
406 		device_unregister(dev);
407 		put_device(dev);
408 	}
409 	return rc;
410 out:
411 	up_write(&dcssblk_devices_sem);
412 	return rc;
413 }
414 static DEVICE_ATTR(shared, S_IWUSR | S_IRUSR, dcssblk_shared_show,
415 		   dcssblk_shared_store);
416 
417 /*
418  * device attribute for save operation on current copy
419  * of the segment. If the segment is busy, saving will
420  * become pending until it gets released, which can be
421  * undone by storing a non-true value to this entry.
422  * (show + store)
423  */
424 static ssize_t
425 dcssblk_save_show(struct device *dev, struct device_attribute *attr, char *buf)
426 {
427 	struct dcssblk_dev_info *dev_info;
428 
429 	dev_info = container_of(dev, struct dcssblk_dev_info, dev);
430 	return sprintf(buf, dev_info->save_pending ? "1\n" : "0\n");
431 }
432 
433 static ssize_t
434 dcssblk_save_store(struct device *dev, struct device_attribute *attr, const char *inbuf, size_t count)
435 {
436 	struct dcssblk_dev_info *dev_info;
437 	struct segment_info *entry;
438 
439 	if ((count > 1) && (inbuf[1] != '\n') && (inbuf[1] != '\0'))
440 		return -EINVAL;
441 	dev_info = container_of(dev, struct dcssblk_dev_info, dev);
442 
443 	down_write(&dcssblk_devices_sem);
444 	if (inbuf[0] == '1') {
445 		if (atomic_read(&dev_info->use_count) == 0) {
446 			// device is idle => we save immediately
447 			pr_info("All DCSSs that map to device %s are "
448 				"saved\n", dev_info->segment_name);
449 			list_for_each_entry(entry, &dev_info->seg_list, lh) {
450 				if (entry->segment_type == SEG_TYPE_EN ||
451 				    entry->segment_type == SEG_TYPE_SN)
452 					pr_warn("DCSS %s is of type SN or EN"
453 						" and cannot be saved\n",
454 						entry->segment_name);
455 				else
456 					segment_save(entry->segment_name);
457 			}
458 		}  else {
459 			// device is busy => we save it when it becomes
460 			// idle in dcssblk_release
461 			pr_info("Device %s is in use, its DCSSs will be "
462 				"saved when it becomes idle\n",
463 				dev_info->segment_name);
464 			dev_info->save_pending = 1;
465 		}
466 	} else if (inbuf[0] == '0') {
467 		if (dev_info->save_pending) {
468 			// device is busy & the user wants to undo his save
469 			// request
470 			dev_info->save_pending = 0;
471 			pr_info("A pending save request for device %s "
472 				"has been canceled\n",
473 				dev_info->segment_name);
474 		}
475 	} else {
476 		up_write(&dcssblk_devices_sem);
477 		return -EINVAL;
478 	}
479 	up_write(&dcssblk_devices_sem);
480 	return count;
481 }
482 static DEVICE_ATTR(save, S_IWUSR | S_IRUSR, dcssblk_save_show,
483 		   dcssblk_save_store);
484 
485 /*
486  * device attribute for showing all segments in a device
487  */
488 static ssize_t
489 dcssblk_seglist_show(struct device *dev, struct device_attribute *attr,
490 		char *buf)
491 {
492 	int i;
493 
494 	struct dcssblk_dev_info *dev_info;
495 	struct segment_info *entry;
496 
497 	down_read(&dcssblk_devices_sem);
498 	dev_info = container_of(dev, struct dcssblk_dev_info, dev);
499 	i = 0;
500 	buf[0] = '\0';
501 	list_for_each_entry(entry, &dev_info->seg_list, lh) {
502 		strcpy(&buf[i], entry->segment_name);
503 		i += strlen(entry->segment_name);
504 		buf[i] = '\n';
505 		i++;
506 	}
507 	up_read(&dcssblk_devices_sem);
508 	return i;
509 }
510 static DEVICE_ATTR(seglist, S_IRUSR, dcssblk_seglist_show, NULL);
511 
512 static struct attribute *dcssblk_dev_attrs[] = {
513 	&dev_attr_shared.attr,
514 	&dev_attr_save.attr,
515 	&dev_attr_seglist.attr,
516 	NULL,
517 };
518 static struct attribute_group dcssblk_dev_attr_group = {
519 	.attrs = dcssblk_dev_attrs,
520 };
521 static const struct attribute_group *dcssblk_dev_attr_groups[] = {
522 	&dcssblk_dev_attr_group,
523 	NULL,
524 };
525 
526 /*
527  * device attribute for adding devices
528  */
529 static ssize_t
530 dcssblk_add_store(struct device *dev, struct device_attribute *attr, const char *buf, size_t count)
531 {
532 	int rc, i, j, num_of_segments;
533 	struct dcssblk_dev_info *dev_info;
534 	struct segment_info *seg_info, *temp;
535 	char *local_buf;
536 	unsigned long seg_byte_size;
537 
538 	dev_info = NULL;
539 	seg_info = NULL;
540 	if (dev != dcssblk_root_dev) {
541 		rc = -EINVAL;
542 		goto out_nobuf;
543 	}
544 	if ((count < 1) || (buf[0] == '\0') || (buf[0] == '\n')) {
545 		rc = -ENAMETOOLONG;
546 		goto out_nobuf;
547 	}
548 
549 	local_buf = kmalloc(count + 1, GFP_KERNEL);
550 	if (local_buf == NULL) {
551 		rc = -ENOMEM;
552 		goto out_nobuf;
553 	}
554 
555 	/*
556 	 * parse input
557 	 */
558 	num_of_segments = 0;
559 	for (i = 0; (i < count && (buf[i] != '\0') && (buf[i] != '\n')); i++) {
560 		for (j = i; j < count &&
561 			(buf[j] != ':') &&
562 			(buf[j] != '\0') &&
563 			(buf[j] != '\n'); j++) {
564 			local_buf[j-i] = toupper(buf[j]);
565 		}
566 		local_buf[j-i] = '\0';
567 		if (((j - i) == 0) || ((j - i) > 8)) {
568 			rc = -ENAMETOOLONG;
569 			goto seg_list_del;
570 		}
571 
572 		rc = dcssblk_load_segment(local_buf, &seg_info);
573 		if (rc < 0)
574 			goto seg_list_del;
575 		/*
576 		 * get a struct dcssblk_dev_info
577 		 */
578 		if (num_of_segments == 0) {
579 			dev_info = kzalloc(sizeof(struct dcssblk_dev_info),
580 					GFP_KERNEL);
581 			if (dev_info == NULL) {
582 				rc = -ENOMEM;
583 				goto out;
584 			}
585 			strcpy(dev_info->segment_name, local_buf);
586 			dev_info->segment_type = seg_info->segment_type;
587 			INIT_LIST_HEAD(&dev_info->seg_list);
588 		}
589 		list_add_tail(&seg_info->lh, &dev_info->seg_list);
590 		num_of_segments++;
591 		i = j;
592 
593 		if ((buf[j] == '\0') || (buf[j] == '\n'))
594 			break;
595 	}
596 
597 	/* no trailing colon at the end of the input */
598 	if ((i > 0) && (buf[i-1] == ':')) {
599 		rc = -ENAMETOOLONG;
600 		goto seg_list_del;
601 	}
602 	strlcpy(local_buf, buf, i + 1);
603 	dev_info->num_of_segments = num_of_segments;
604 	rc = dcssblk_is_continuous(dev_info);
605 	if (rc < 0)
606 		goto seg_list_del;
607 
608 	dev_info->start = dcssblk_find_lowest_addr(dev_info);
609 	dev_info->end = dcssblk_find_highest_addr(dev_info);
610 
611 	dev_set_name(&dev_info->dev, "%s", dev_info->segment_name);
612 	dev_info->dev.release = dcssblk_release_segment;
613 	dev_info->dev.groups = dcssblk_dev_attr_groups;
614 	INIT_LIST_HEAD(&dev_info->lh);
615 	dev_info->gd = alloc_disk(DCSSBLK_MINORS_PER_DISK);
616 	if (dev_info->gd == NULL) {
617 		rc = -ENOMEM;
618 		goto seg_list_del;
619 	}
620 	dev_info->gd->major = dcssblk_major;
621 	dev_info->gd->fops = &dcssblk_devops;
622 	dev_info->dcssblk_queue = blk_alloc_queue(GFP_KERNEL);
623 	dev_info->gd->queue = dev_info->dcssblk_queue;
624 	dev_info->gd->private_data = dev_info;
625 	blk_queue_make_request(dev_info->dcssblk_queue, dcssblk_make_request);
626 	blk_queue_logical_block_size(dev_info->dcssblk_queue, 4096);
627 	queue_flag_set_unlocked(QUEUE_FLAG_DAX, dev_info->dcssblk_queue);
628 
629 	seg_byte_size = (dev_info->end - dev_info->start + 1);
630 	set_capacity(dev_info->gd, seg_byte_size >> 9); // size in sectors
631 	pr_info("Loaded %s with total size %lu bytes and capacity %lu "
632 		"sectors\n", local_buf, seg_byte_size, seg_byte_size >> 9);
633 
634 	dev_info->save_pending = 0;
635 	dev_info->is_shared = 1;
636 	dev_info->dev.parent = dcssblk_root_dev;
637 
638 	/*
639 	 *get minor, add to list
640 	 */
641 	down_write(&dcssblk_devices_sem);
642 	if (dcssblk_get_segment_by_name(local_buf)) {
643 		rc = -EEXIST;
644 		goto release_gd;
645 	}
646 	rc = dcssblk_assign_free_minor(dev_info);
647 	if (rc)
648 		goto release_gd;
649 	sprintf(dev_info->gd->disk_name, "dcssblk%d",
650 		dev_info->gd->first_minor);
651 	list_add_tail(&dev_info->lh, &dcssblk_devices);
652 
653 	if (!try_module_get(THIS_MODULE)) {
654 		rc = -ENODEV;
655 		goto dev_list_del;
656 	}
657 	/*
658 	 * register the device
659 	 */
660 	rc = device_register(&dev_info->dev);
661 	if (rc)
662 		goto put_dev;
663 
664 	dev_info->dax_dev = alloc_dax(dev_info, dev_info->gd->disk_name,
665 			&dcssblk_dax_ops);
666 	if (!dev_info->dax_dev) {
667 		rc = -ENOMEM;
668 		goto put_dev;
669 	}
670 
671 	get_device(&dev_info->dev);
672 	device_add_disk(&dev_info->dev, dev_info->gd);
673 
674 	switch (dev_info->segment_type) {
675 		case SEG_TYPE_SR:
676 		case SEG_TYPE_ER:
677 		case SEG_TYPE_SC:
678 			set_disk_ro(dev_info->gd,1);
679 			break;
680 		default:
681 			set_disk_ro(dev_info->gd,0);
682 			break;
683 	}
684 	up_write(&dcssblk_devices_sem);
685 	rc = count;
686 	goto out;
687 
688 put_dev:
689 	list_del(&dev_info->lh);
690 	blk_cleanup_queue(dev_info->dcssblk_queue);
691 	dev_info->gd->queue = NULL;
692 	put_disk(dev_info->gd);
693 	list_for_each_entry(seg_info, &dev_info->seg_list, lh) {
694 		segment_unload(seg_info->segment_name);
695 	}
696 	put_device(&dev_info->dev);
697 	up_write(&dcssblk_devices_sem);
698 	goto out;
699 dev_list_del:
700 	list_del(&dev_info->lh);
701 release_gd:
702 	blk_cleanup_queue(dev_info->dcssblk_queue);
703 	dev_info->gd->queue = NULL;
704 	put_disk(dev_info->gd);
705 	up_write(&dcssblk_devices_sem);
706 seg_list_del:
707 	if (dev_info == NULL)
708 		goto out;
709 	list_for_each_entry_safe(seg_info, temp, &dev_info->seg_list, lh) {
710 		list_del(&seg_info->lh);
711 		segment_unload(seg_info->segment_name);
712 		kfree(seg_info);
713 	}
714 	kfree(dev_info);
715 out:
716 	kfree(local_buf);
717 out_nobuf:
718 	return rc;
719 }
720 
721 /*
722  * device attribute for removing devices
723  */
724 static ssize_t
725 dcssblk_remove_store(struct device *dev, struct device_attribute *attr, const char *buf, size_t count)
726 {
727 	struct dcssblk_dev_info *dev_info;
728 	struct segment_info *entry;
729 	int rc, i;
730 	char *local_buf;
731 
732 	if (dev != dcssblk_root_dev) {
733 		return -EINVAL;
734 	}
735 	local_buf = kmalloc(count + 1, GFP_KERNEL);
736 	if (local_buf == NULL) {
737 		return -ENOMEM;
738 	}
739 	/*
740 	 * parse input
741 	 */
742 	for (i = 0; (i < count && (*(buf+i)!='\0') && (*(buf+i)!='\n')); i++) {
743 		local_buf[i] = toupper(buf[i]);
744 	}
745 	local_buf[i] = '\0';
746 	if ((i == 0) || (i > 8)) {
747 		rc = -ENAMETOOLONG;
748 		goto out_buf;
749 	}
750 
751 	down_write(&dcssblk_devices_sem);
752 	dev_info = dcssblk_get_device_by_name(local_buf);
753 	if (dev_info == NULL) {
754 		up_write(&dcssblk_devices_sem);
755 		pr_warn("Device %s cannot be removed because it is not a known device\n",
756 			local_buf);
757 		rc = -ENODEV;
758 		goto out_buf;
759 	}
760 	if (atomic_read(&dev_info->use_count) != 0) {
761 		up_write(&dcssblk_devices_sem);
762 		pr_warn("Device %s cannot be removed while it is in use\n",
763 			local_buf);
764 		rc = -EBUSY;
765 		goto out_buf;
766 	}
767 
768 	list_del(&dev_info->lh);
769 	kill_dax(dev_info->dax_dev);
770 	put_dax(dev_info->dax_dev);
771 	del_gendisk(dev_info->gd);
772 	blk_cleanup_queue(dev_info->dcssblk_queue);
773 	dev_info->gd->queue = NULL;
774 	put_disk(dev_info->gd);
775 
776 	/* unload all related segments */
777 	list_for_each_entry(entry, &dev_info->seg_list, lh)
778 		segment_unload(entry->segment_name);
779 
780 	up_write(&dcssblk_devices_sem);
781 
782 	device_unregister(&dev_info->dev);
783 	put_device(&dev_info->dev);
784 
785 	rc = count;
786 out_buf:
787 	kfree(local_buf);
788 	return rc;
789 }
790 
791 static int
792 dcssblk_open(struct block_device *bdev, fmode_t mode)
793 {
794 	struct dcssblk_dev_info *dev_info;
795 	int rc;
796 
797 	dev_info = bdev->bd_disk->private_data;
798 	if (NULL == dev_info) {
799 		rc = -ENODEV;
800 		goto out;
801 	}
802 	atomic_inc(&dev_info->use_count);
803 	bdev->bd_block_size = 4096;
804 	rc = 0;
805 out:
806 	return rc;
807 }
808 
809 static void
810 dcssblk_release(struct gendisk *disk, fmode_t mode)
811 {
812 	struct dcssblk_dev_info *dev_info = disk->private_data;
813 	struct segment_info *entry;
814 
815 	if (!dev_info) {
816 		WARN_ON(1);
817 		return;
818 	}
819 	down_write(&dcssblk_devices_sem);
820 	if (atomic_dec_and_test(&dev_info->use_count)
821 	    && (dev_info->save_pending)) {
822 		pr_info("Device %s has become idle and is being saved "
823 			"now\n", dev_info->segment_name);
824 		list_for_each_entry(entry, &dev_info->seg_list, lh) {
825 			if (entry->segment_type == SEG_TYPE_EN ||
826 			    entry->segment_type == SEG_TYPE_SN)
827 				pr_warn("DCSS %s is of type SN or EN and cannot"
828 					" be saved\n", entry->segment_name);
829 			else
830 				segment_save(entry->segment_name);
831 		}
832 		dev_info->save_pending = 0;
833 	}
834 	up_write(&dcssblk_devices_sem);
835 }
836 
837 static blk_qc_t
838 dcssblk_make_request(struct request_queue *q, struct bio *bio)
839 {
840 	struct dcssblk_dev_info *dev_info;
841 	struct bio_vec bvec;
842 	struct bvec_iter iter;
843 	unsigned long index;
844 	unsigned long page_addr;
845 	unsigned long source_addr;
846 	unsigned long bytes_done;
847 
848 	blk_queue_split(q, &bio, q->bio_split);
849 
850 	bytes_done = 0;
851 	dev_info = bio->bi_bdev->bd_disk->private_data;
852 	if (dev_info == NULL)
853 		goto fail;
854 	if ((bio->bi_iter.bi_sector & 7) != 0 ||
855 	    (bio->bi_iter.bi_size & 4095) != 0)
856 		/* Request is not page-aligned. */
857 		goto fail;
858 	if (bio_end_sector(bio) > get_capacity(bio->bi_bdev->bd_disk)) {
859 		/* Request beyond end of DCSS segment. */
860 		goto fail;
861 	}
862 	/* verify data transfer direction */
863 	if (dev_info->is_shared) {
864 		switch (dev_info->segment_type) {
865 		case SEG_TYPE_SR:
866 		case SEG_TYPE_ER:
867 		case SEG_TYPE_SC:
868 			/* cannot write to these segments */
869 			if (bio_data_dir(bio) == WRITE) {
870 				pr_warn("Writing to %s failed because it is a read-only device\n",
871 					dev_name(&dev_info->dev));
872 				goto fail;
873 			}
874 		}
875 	}
876 
877 	index = (bio->bi_iter.bi_sector >> 3);
878 	bio_for_each_segment(bvec, bio, iter) {
879 		page_addr = (unsigned long)
880 			page_address(bvec.bv_page) + bvec.bv_offset;
881 		source_addr = dev_info->start + (index<<12) + bytes_done;
882 		if (unlikely((page_addr & 4095) != 0) || (bvec.bv_len & 4095) != 0)
883 			// More paranoia.
884 			goto fail;
885 		if (bio_data_dir(bio) == READ) {
886 			memcpy((void*)page_addr, (void*)source_addr,
887 				bvec.bv_len);
888 		} else {
889 			memcpy((void*)source_addr, (void*)page_addr,
890 				bvec.bv_len);
891 		}
892 		bytes_done += bvec.bv_len;
893 	}
894 	bio_endio(bio);
895 	return BLK_QC_T_NONE;
896 fail:
897 	bio_io_error(bio);
898 	return BLK_QC_T_NONE;
899 }
900 
901 static long
902 __dcssblk_direct_access(struct dcssblk_dev_info *dev_info, pgoff_t pgoff,
903 		long nr_pages, void **kaddr, pfn_t *pfn)
904 {
905 	resource_size_t offset = pgoff * PAGE_SIZE;
906 	unsigned long dev_sz;
907 
908 	dev_sz = dev_info->end - dev_info->start + 1;
909 	*kaddr = (void *) dev_info->start + offset;
910 	*pfn = __pfn_to_pfn_t(PFN_DOWN(dev_info->start + offset), PFN_DEV);
911 
912 	return (dev_sz - offset) / PAGE_SIZE;
913 }
914 
915 static long
916 dcssblk_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
917 		long nr_pages, void **kaddr, pfn_t *pfn)
918 {
919 	struct dcssblk_dev_info *dev_info = dax_get_private(dax_dev);
920 
921 	return __dcssblk_direct_access(dev_info, pgoff, nr_pages, kaddr, pfn);
922 }
923 
924 static void
925 dcssblk_check_params(void)
926 {
927 	int rc, i, j, k;
928 	char buf[DCSSBLK_PARM_LEN + 1];
929 	struct dcssblk_dev_info *dev_info;
930 
931 	for (i = 0; (i < DCSSBLK_PARM_LEN) && (dcssblk_segments[i] != '\0');
932 	     i++) {
933 		for (j = i; (j < DCSSBLK_PARM_LEN) &&
934 			    (dcssblk_segments[j] != ',')  &&
935 			    (dcssblk_segments[j] != '\0') &&
936 			    (dcssblk_segments[j] != '('); j++)
937 		{
938 			buf[j-i] = dcssblk_segments[j];
939 		}
940 		buf[j-i] = '\0';
941 		rc = dcssblk_add_store(dcssblk_root_dev, NULL, buf, j-i);
942 		if ((rc >= 0) && (dcssblk_segments[j] == '(')) {
943 			for (k = 0; (buf[k] != ':') && (buf[k] != '\0'); k++)
944 				buf[k] = toupper(buf[k]);
945 			buf[k] = '\0';
946 			if (!strncmp(&dcssblk_segments[j], "(local)", 7)) {
947 				down_read(&dcssblk_devices_sem);
948 				dev_info = dcssblk_get_device_by_name(buf);
949 				up_read(&dcssblk_devices_sem);
950 				if (dev_info)
951 					dcssblk_shared_store(&dev_info->dev,
952 							     NULL, "0\n", 2);
953 			}
954 		}
955 		while ((dcssblk_segments[j] != ',') &&
956 		       (dcssblk_segments[j] != '\0'))
957 		{
958 			j++;
959 		}
960 		if (dcssblk_segments[j] == '\0')
961 			break;
962 		i = j;
963 	}
964 }
965 
966 /*
967  * Suspend / Resume
968  */
969 static int dcssblk_freeze(struct device *dev)
970 {
971 	struct dcssblk_dev_info *dev_info;
972 	int rc = 0;
973 
974 	list_for_each_entry(dev_info, &dcssblk_devices, lh) {
975 		switch (dev_info->segment_type) {
976 			case SEG_TYPE_SR:
977 			case SEG_TYPE_ER:
978 			case SEG_TYPE_SC:
979 				if (!dev_info->is_shared)
980 					rc = -EINVAL;
981 				break;
982 			default:
983 				rc = -EINVAL;
984 				break;
985 		}
986 		if (rc)
987 			break;
988 	}
989 	if (rc)
990 		pr_err("Suspending the system failed because DCSS device %s "
991 		       "is writable\n",
992 		       dev_info->segment_name);
993 	return rc;
994 }
995 
996 static int dcssblk_restore(struct device *dev)
997 {
998 	struct dcssblk_dev_info *dev_info;
999 	struct segment_info *entry;
1000 	unsigned long start, end;
1001 	int rc = 0;
1002 
1003 	list_for_each_entry(dev_info, &dcssblk_devices, lh) {
1004 		list_for_each_entry(entry, &dev_info->seg_list, lh) {
1005 			segment_unload(entry->segment_name);
1006 			rc = segment_load(entry->segment_name, SEGMENT_SHARED,
1007 					  &start, &end);
1008 			if (rc < 0) {
1009 // TODO in_use check ?
1010 				segment_warning(rc, entry->segment_name);
1011 				goto out_panic;
1012 			}
1013 			if (start != entry->start || end != entry->end) {
1014 				pr_err("The address range of DCSS %s changed "
1015 				       "while the system was suspended\n",
1016 				       entry->segment_name);
1017 				goto out_panic;
1018 			}
1019 		}
1020 	}
1021 	return 0;
1022 out_panic:
1023 	panic("fatal dcssblk resume error\n");
1024 }
1025 
1026 static int dcssblk_thaw(struct device *dev)
1027 {
1028 	return 0;
1029 }
1030 
1031 static const struct dev_pm_ops dcssblk_pm_ops = {
1032 	.freeze		= dcssblk_freeze,
1033 	.thaw		= dcssblk_thaw,
1034 	.restore	= dcssblk_restore,
1035 };
1036 
1037 static struct platform_driver dcssblk_pdrv = {
1038 	.driver = {
1039 		.name	= "dcssblk",
1040 		.pm	= &dcssblk_pm_ops,
1041 	},
1042 };
1043 
1044 static struct platform_device *dcssblk_pdev;
1045 
1046 
1047 /*
1048  * The init/exit functions.
1049  */
1050 static void __exit
1051 dcssblk_exit(void)
1052 {
1053 	platform_device_unregister(dcssblk_pdev);
1054 	platform_driver_unregister(&dcssblk_pdrv);
1055 	root_device_unregister(dcssblk_root_dev);
1056 	unregister_blkdev(dcssblk_major, DCSSBLK_NAME);
1057 }
1058 
1059 static int __init
1060 dcssblk_init(void)
1061 {
1062 	int rc;
1063 
1064 	rc = platform_driver_register(&dcssblk_pdrv);
1065 	if (rc)
1066 		return rc;
1067 
1068 	dcssblk_pdev = platform_device_register_simple("dcssblk", -1, NULL,
1069 							0);
1070 	if (IS_ERR(dcssblk_pdev)) {
1071 		rc = PTR_ERR(dcssblk_pdev);
1072 		goto out_pdrv;
1073 	}
1074 
1075 	dcssblk_root_dev = root_device_register("dcssblk");
1076 	if (IS_ERR(dcssblk_root_dev)) {
1077 		rc = PTR_ERR(dcssblk_root_dev);
1078 		goto out_pdev;
1079 	}
1080 	rc = device_create_file(dcssblk_root_dev, &dev_attr_add);
1081 	if (rc)
1082 		goto out_root;
1083 	rc = device_create_file(dcssblk_root_dev, &dev_attr_remove);
1084 	if (rc)
1085 		goto out_root;
1086 	rc = register_blkdev(0, DCSSBLK_NAME);
1087 	if (rc < 0)
1088 		goto out_root;
1089 	dcssblk_major = rc;
1090 	init_rwsem(&dcssblk_devices_sem);
1091 
1092 	dcssblk_check_params();
1093 	return 0;
1094 
1095 out_root:
1096 	root_device_unregister(dcssblk_root_dev);
1097 out_pdev:
1098 	platform_device_unregister(dcssblk_pdev);
1099 out_pdrv:
1100 	platform_driver_unregister(&dcssblk_pdrv);
1101 	return rc;
1102 }
1103 
1104 module_init(dcssblk_init);
1105 module_exit(dcssblk_exit);
1106 
1107 module_param_string(segments, dcssblk_segments, DCSSBLK_PARM_LEN, 0444);
1108 MODULE_PARM_DESC(segments, "Name of DCSS segment(s) to be loaded, "
1109 		 "comma-separated list, names in each set separated "
1110 		 "by commas are separated by colons, each set contains "
1111 		 "names of contiguous segments and each name max. 8 chars.\n"
1112 		 "Adding \"(local)\" to the end of each set equals echoing 0 "
1113 		 "to /sys/devices/dcssblk/<device name>/shared after loading "
1114 		 "the contiguous segments - \n"
1115 		 "e.g. segments=\"mydcss1,mydcss2:mydcss3,mydcss4(local)\"");
1116 
1117 MODULE_LICENSE("GPL");
1118