xref: /openbmc/linux/drivers/s390/block/dcssblk.c (revision 94d964e5)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * dcssblk.c -- the S/390 block driver for dcss memory
4  *
5  * Authors: Carsten Otte, Stefan Weinhuber, Gerald Schaefer
6  */
7 
8 #define KMSG_COMPONENT "dcssblk"
9 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
10 
11 #include <linux/module.h>
12 #include <linux/moduleparam.h>
13 #include <linux/ctype.h>
14 #include <linux/errno.h>
15 #include <linux/init.h>
16 #include <linux/slab.h>
17 #include <linux/blkdev.h>
18 #include <linux/completion.h>
19 #include <linux/interrupt.h>
20 #include <linux/pfn_t.h>
21 #include <linux/uio.h>
22 #include <linux/dax.h>
23 #include <asm/extmem.h>
24 #include <asm/io.h>
25 
26 #define DCSSBLK_NAME "dcssblk"
27 #define DCSSBLK_MINORS_PER_DISK 1
28 #define DCSSBLK_PARM_LEN 400
29 #define DCSS_BUS_ID_SIZE 20
30 
31 static int dcssblk_open(struct block_device *bdev, fmode_t mode);
32 static void dcssblk_release(struct gendisk *disk, fmode_t mode);
33 static void dcssblk_submit_bio(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 	.submit_bio	= dcssblk_submit_bio,
43 	.open    	= dcssblk_open,
44 	.release 	= dcssblk_release,
45 };
46 
47 static int dcssblk_dax_zero_page_range(struct dax_device *dax_dev,
48 				       pgoff_t pgoff, size_t nr_pages)
49 {
50 	long rc;
51 	void *kaddr;
52 
53 	rc = dax_direct_access(dax_dev, pgoff, nr_pages, &kaddr, NULL);
54 	if (rc < 0)
55 		return rc;
56 	memset(kaddr, 0, nr_pages << PAGE_SHIFT);
57 	dax_flush(dax_dev, kaddr, nr_pages << PAGE_SHIFT);
58 	return 0;
59 }
60 
61 static const struct dax_operations dcssblk_dax_ops = {
62 	.direct_access = dcssblk_dax_direct_access,
63 	.zero_page_range = dcssblk_dax_zero_page_range,
64 };
65 
66 struct dcssblk_dev_info {
67 	struct list_head lh;
68 	struct device dev;
69 	char segment_name[DCSS_BUS_ID_SIZE];
70 	atomic_t use_count;
71 	struct gendisk *gd;
72 	unsigned long start;
73 	unsigned long end;
74 	int segment_type;
75 	unsigned char save_pending;
76 	unsigned char is_shared;
77 	int num_of_segments;
78 	struct list_head seg_list;
79 	struct dax_device *dax_dev;
80 };
81 
82 struct segment_info {
83 	struct list_head lh;
84 	char segment_name[DCSS_BUS_ID_SIZE];
85 	unsigned long start;
86 	unsigned long end;
87 	int segment_type;
88 };
89 
90 static ssize_t dcssblk_add_store(struct device * dev, struct device_attribute *attr, const char * buf,
91 				  size_t count);
92 static ssize_t dcssblk_remove_store(struct device * dev, struct device_attribute *attr, const char * buf,
93 				  size_t count);
94 
95 static DEVICE_ATTR(add, S_IWUSR, NULL, dcssblk_add_store);
96 static DEVICE_ATTR(remove, S_IWUSR, NULL, dcssblk_remove_store);
97 
98 static struct device *dcssblk_root_dev;
99 
100 static LIST_HEAD(dcssblk_devices);
101 static struct rw_semaphore dcssblk_devices_sem;
102 
103 /*
104  * release function for segment device.
105  */
106 static void
107 dcssblk_release_segment(struct device *dev)
108 {
109 	struct dcssblk_dev_info *dev_info;
110 	struct segment_info *entry, *temp;
111 
112 	dev_info = container_of(dev, struct dcssblk_dev_info, dev);
113 	list_for_each_entry_safe(entry, temp, &dev_info->seg_list, lh) {
114 		list_del(&entry->lh);
115 		kfree(entry);
116 	}
117 	kfree(dev_info);
118 	module_put(THIS_MODULE);
119 }
120 
121 /*
122  * get a minor number. needs to be called with
123  * down_write(&dcssblk_devices_sem) and the
124  * device needs to be enqueued before the semaphore is
125  * freed.
126  */
127 static int
128 dcssblk_assign_free_minor(struct dcssblk_dev_info *dev_info)
129 {
130 	int minor, found;
131 	struct dcssblk_dev_info *entry;
132 
133 	if (dev_info == NULL)
134 		return -EINVAL;
135 	for (minor = 0; minor < (1<<MINORBITS); minor++) {
136 		found = 0;
137 		// test if minor available
138 		list_for_each_entry(entry, &dcssblk_devices, lh)
139 			if (minor == entry->gd->first_minor)
140 				found++;
141 		if (!found) break; // got unused minor
142 	}
143 	if (found)
144 		return -EBUSY;
145 	dev_info->gd->first_minor = minor;
146 	return 0;
147 }
148 
149 /*
150  * get the struct dcssblk_dev_info from dcssblk_devices
151  * for the given name.
152  * down_read(&dcssblk_devices_sem) must be held.
153  */
154 static struct dcssblk_dev_info *
155 dcssblk_get_device_by_name(char *name)
156 {
157 	struct dcssblk_dev_info *entry;
158 
159 	list_for_each_entry(entry, &dcssblk_devices, lh) {
160 		if (!strcmp(name, entry->segment_name)) {
161 			return entry;
162 		}
163 	}
164 	return NULL;
165 }
166 
167 /*
168  * get the struct segment_info from seg_list
169  * for the given name.
170  * down_read(&dcssblk_devices_sem) must be held.
171  */
172 static struct segment_info *
173 dcssblk_get_segment_by_name(char *name)
174 {
175 	struct dcssblk_dev_info *dev_info;
176 	struct segment_info *entry;
177 
178 	list_for_each_entry(dev_info, &dcssblk_devices, lh) {
179 		list_for_each_entry(entry, &dev_info->seg_list, lh) {
180 			if (!strcmp(name, entry->segment_name))
181 				return entry;
182 		}
183 	}
184 	return NULL;
185 }
186 
187 /*
188  * get the highest address of the multi-segment block.
189  */
190 static unsigned long
191 dcssblk_find_highest_addr(struct dcssblk_dev_info *dev_info)
192 {
193 	unsigned long highest_addr;
194 	struct segment_info *entry;
195 
196 	highest_addr = 0;
197 	list_for_each_entry(entry, &dev_info->seg_list, lh) {
198 		if (highest_addr < entry->end)
199 			highest_addr = entry->end;
200 	}
201 	return highest_addr;
202 }
203 
204 /*
205  * get the lowest address of the multi-segment block.
206  */
207 static unsigned long
208 dcssblk_find_lowest_addr(struct dcssblk_dev_info *dev_info)
209 {
210 	int set_first;
211 	unsigned long lowest_addr;
212 	struct segment_info *entry;
213 
214 	set_first = 0;
215 	lowest_addr = 0;
216 	list_for_each_entry(entry, &dev_info->seg_list, lh) {
217 		if (set_first == 0) {
218 			lowest_addr = entry->start;
219 			set_first = 1;
220 		} else {
221 			if (lowest_addr > entry->start)
222 				lowest_addr = entry->start;
223 		}
224 	}
225 	return lowest_addr;
226 }
227 
228 /*
229  * Check continuity of segments.
230  */
231 static int
232 dcssblk_is_continuous(struct dcssblk_dev_info *dev_info)
233 {
234 	int i, j, rc;
235 	struct segment_info *sort_list, *entry, temp;
236 
237 	if (dev_info->num_of_segments <= 1)
238 		return 0;
239 
240 	sort_list = kcalloc(dev_info->num_of_segments,
241 			    sizeof(struct segment_info),
242 			    GFP_KERNEL);
243 	if (sort_list == NULL)
244 		return -ENOMEM;
245 	i = 0;
246 	list_for_each_entry(entry, &dev_info->seg_list, lh) {
247 		memcpy(&sort_list[i], entry, sizeof(struct segment_info));
248 		i++;
249 	}
250 
251 	/* sort segments */
252 	for (i = 0; i < dev_info->num_of_segments; i++)
253 		for (j = 0; j < dev_info->num_of_segments; j++)
254 			if (sort_list[j].start > sort_list[i].start) {
255 				memcpy(&temp, &sort_list[i],
256 					sizeof(struct segment_info));
257 				memcpy(&sort_list[i], &sort_list[j],
258 					sizeof(struct segment_info));
259 				memcpy(&sort_list[j], &temp,
260 					sizeof(struct segment_info));
261 			}
262 
263 	/* check continuity */
264 	for (i = 0; i < dev_info->num_of_segments - 1; i++) {
265 		if ((sort_list[i].end + 1) != sort_list[i+1].start) {
266 			pr_err("Adjacent DCSSs %s and %s are not "
267 			       "contiguous\n", sort_list[i].segment_name,
268 			       sort_list[i+1].segment_name);
269 			rc = -EINVAL;
270 			goto out;
271 		}
272 		/* EN and EW are allowed in a block device */
273 		if (sort_list[i].segment_type != sort_list[i+1].segment_type) {
274 			if (!(sort_list[i].segment_type & SEGMENT_EXCLUSIVE) ||
275 				(sort_list[i].segment_type == SEG_TYPE_ER) ||
276 				!(sort_list[i+1].segment_type &
277 				SEGMENT_EXCLUSIVE) ||
278 				(sort_list[i+1].segment_type == SEG_TYPE_ER)) {
279 				pr_err("DCSS %s and DCSS %s have "
280 				       "incompatible types\n",
281 				       sort_list[i].segment_name,
282 				       sort_list[i+1].segment_name);
283 				rc = -EINVAL;
284 				goto out;
285 			}
286 		}
287 	}
288 	rc = 0;
289 out:
290 	kfree(sort_list);
291 	return rc;
292 }
293 
294 /*
295  * Load a segment
296  */
297 static int
298 dcssblk_load_segment(char *name, struct segment_info **seg_info)
299 {
300 	int rc;
301 
302 	/* already loaded? */
303 	down_read(&dcssblk_devices_sem);
304 	*seg_info = dcssblk_get_segment_by_name(name);
305 	up_read(&dcssblk_devices_sem);
306 	if (*seg_info != NULL)
307 		return -EEXIST;
308 
309 	/* get a struct segment_info */
310 	*seg_info = kzalloc(sizeof(struct segment_info), GFP_KERNEL);
311 	if (*seg_info == NULL)
312 		return -ENOMEM;
313 
314 	strcpy((*seg_info)->segment_name, name);
315 
316 	/* load the segment */
317 	rc = segment_load(name, SEGMENT_SHARED,
318 			&(*seg_info)->start, &(*seg_info)->end);
319 	if (rc < 0) {
320 		segment_warning(rc, (*seg_info)->segment_name);
321 		kfree(*seg_info);
322 	} else {
323 		INIT_LIST_HEAD(&(*seg_info)->lh);
324 		(*seg_info)->segment_type = rc;
325 	}
326 	return rc;
327 }
328 
329 /*
330  * device attribute for switching shared/nonshared (exclusive)
331  * operation (show + store)
332  */
333 static ssize_t
334 dcssblk_shared_show(struct device *dev, struct device_attribute *attr, char *buf)
335 {
336 	struct dcssblk_dev_info *dev_info;
337 
338 	dev_info = container_of(dev, struct dcssblk_dev_info, dev);
339 	return sprintf(buf, dev_info->is_shared ? "1\n" : "0\n");
340 }
341 
342 static ssize_t
343 dcssblk_shared_store(struct device *dev, struct device_attribute *attr, const char *inbuf, size_t count)
344 {
345 	struct dcssblk_dev_info *dev_info;
346 	struct segment_info *entry, *temp;
347 	int rc;
348 
349 	if ((count > 1) && (inbuf[1] != '\n') && (inbuf[1] != '\0'))
350 		return -EINVAL;
351 	down_write(&dcssblk_devices_sem);
352 	dev_info = container_of(dev, struct dcssblk_dev_info, dev);
353 	if (atomic_read(&dev_info->use_count)) {
354 		rc = -EBUSY;
355 		goto out;
356 	}
357 	if (inbuf[0] == '1') {
358 		/* reload segments in shared mode */
359 		list_for_each_entry(entry, &dev_info->seg_list, lh) {
360 			rc = segment_modify_shared(entry->segment_name,
361 						SEGMENT_SHARED);
362 			if (rc < 0) {
363 				BUG_ON(rc == -EINVAL);
364 				if (rc != -EAGAIN)
365 					goto removeseg;
366 			}
367 		}
368 		dev_info->is_shared = 1;
369 		switch (dev_info->segment_type) {
370 		case SEG_TYPE_SR:
371 		case SEG_TYPE_ER:
372 		case SEG_TYPE_SC:
373 			set_disk_ro(dev_info->gd, 1);
374 		}
375 	} else if (inbuf[0] == '0') {
376 		/* reload segments in exclusive mode */
377 		if (dev_info->segment_type == SEG_TYPE_SC) {
378 			pr_err("DCSS %s is of type SC and cannot be "
379 			       "loaded as exclusive-writable\n",
380 			       dev_info->segment_name);
381 			rc = -EINVAL;
382 			goto out;
383 		}
384 		list_for_each_entry(entry, &dev_info->seg_list, lh) {
385 			rc = segment_modify_shared(entry->segment_name,
386 						   SEGMENT_EXCLUSIVE);
387 			if (rc < 0) {
388 				BUG_ON(rc == -EINVAL);
389 				if (rc != -EAGAIN)
390 					goto removeseg;
391 			}
392 		}
393 		dev_info->is_shared = 0;
394 		set_disk_ro(dev_info->gd, 0);
395 	} else {
396 		rc = -EINVAL;
397 		goto out;
398 	}
399 	rc = count;
400 	goto out;
401 
402 removeseg:
403 	pr_err("DCSS device %s is removed after a failed access mode "
404 	       "change\n", dev_info->segment_name);
405 	temp = entry;
406 	list_for_each_entry(entry, &dev_info->seg_list, lh) {
407 		if (entry != temp)
408 			segment_unload(entry->segment_name);
409 	}
410 	list_del(&dev_info->lh);
411 
412 	kill_dax(dev_info->dax_dev);
413 	put_dax(dev_info->dax_dev);
414 	del_gendisk(dev_info->gd);
415 	blk_cleanup_disk(dev_info->gd);
416 	up_write(&dcssblk_devices_sem);
417 
418 	if (device_remove_file_self(dev, attr)) {
419 		device_unregister(dev);
420 		put_device(dev);
421 	}
422 	return rc;
423 out:
424 	up_write(&dcssblk_devices_sem);
425 	return rc;
426 }
427 static DEVICE_ATTR(shared, S_IWUSR | S_IRUSR, dcssblk_shared_show,
428 		   dcssblk_shared_store);
429 
430 /*
431  * device attribute for save operation on current copy
432  * of the segment. If the segment is busy, saving will
433  * become pending until it gets released, which can be
434  * undone by storing a non-true value to this entry.
435  * (show + store)
436  */
437 static ssize_t
438 dcssblk_save_show(struct device *dev, struct device_attribute *attr, char *buf)
439 {
440 	struct dcssblk_dev_info *dev_info;
441 
442 	dev_info = container_of(dev, struct dcssblk_dev_info, dev);
443 	return sprintf(buf, dev_info->save_pending ? "1\n" : "0\n");
444 }
445 
446 static ssize_t
447 dcssblk_save_store(struct device *dev, struct device_attribute *attr, const char *inbuf, size_t count)
448 {
449 	struct dcssblk_dev_info *dev_info;
450 	struct segment_info *entry;
451 
452 	if ((count > 1) && (inbuf[1] != '\n') && (inbuf[1] != '\0'))
453 		return -EINVAL;
454 	dev_info = container_of(dev, struct dcssblk_dev_info, dev);
455 
456 	down_write(&dcssblk_devices_sem);
457 	if (inbuf[0] == '1') {
458 		if (atomic_read(&dev_info->use_count) == 0) {
459 			// device is idle => we save immediately
460 			pr_info("All DCSSs that map to device %s are "
461 				"saved\n", dev_info->segment_name);
462 			list_for_each_entry(entry, &dev_info->seg_list, lh) {
463 				if (entry->segment_type == SEG_TYPE_EN ||
464 				    entry->segment_type == SEG_TYPE_SN)
465 					pr_warn("DCSS %s is of type SN or EN"
466 						" and cannot be saved\n",
467 						entry->segment_name);
468 				else
469 					segment_save(entry->segment_name);
470 			}
471 		}  else {
472 			// device is busy => we save it when it becomes
473 			// idle in dcssblk_release
474 			pr_info("Device %s is in use, its DCSSs will be "
475 				"saved when it becomes idle\n",
476 				dev_info->segment_name);
477 			dev_info->save_pending = 1;
478 		}
479 	} else if (inbuf[0] == '0') {
480 		if (dev_info->save_pending) {
481 			// device is busy & the user wants to undo his save
482 			// request
483 			dev_info->save_pending = 0;
484 			pr_info("A pending save request for device %s "
485 				"has been canceled\n",
486 				dev_info->segment_name);
487 		}
488 	} else {
489 		up_write(&dcssblk_devices_sem);
490 		return -EINVAL;
491 	}
492 	up_write(&dcssblk_devices_sem);
493 	return count;
494 }
495 static DEVICE_ATTR(save, S_IWUSR | S_IRUSR, dcssblk_save_show,
496 		   dcssblk_save_store);
497 
498 /*
499  * device attribute for showing all segments in a device
500  */
501 static ssize_t
502 dcssblk_seglist_show(struct device *dev, struct device_attribute *attr,
503 		char *buf)
504 {
505 	int i;
506 
507 	struct dcssblk_dev_info *dev_info;
508 	struct segment_info *entry;
509 
510 	down_read(&dcssblk_devices_sem);
511 	dev_info = container_of(dev, struct dcssblk_dev_info, dev);
512 	i = 0;
513 	buf[0] = '\0';
514 	list_for_each_entry(entry, &dev_info->seg_list, lh) {
515 		strcpy(&buf[i], entry->segment_name);
516 		i += strlen(entry->segment_name);
517 		buf[i] = '\n';
518 		i++;
519 	}
520 	up_read(&dcssblk_devices_sem);
521 	return i;
522 }
523 static DEVICE_ATTR(seglist, S_IRUSR, dcssblk_seglist_show, NULL);
524 
525 static struct attribute *dcssblk_dev_attrs[] = {
526 	&dev_attr_shared.attr,
527 	&dev_attr_save.attr,
528 	&dev_attr_seglist.attr,
529 	NULL,
530 };
531 static struct attribute_group dcssblk_dev_attr_group = {
532 	.attrs = dcssblk_dev_attrs,
533 };
534 static const struct attribute_group *dcssblk_dev_attr_groups[] = {
535 	&dcssblk_dev_attr_group,
536 	NULL,
537 };
538 
539 /*
540  * device attribute for adding devices
541  */
542 static ssize_t
543 dcssblk_add_store(struct device *dev, struct device_attribute *attr, const char *buf, size_t count)
544 {
545 	int rc, i, j, num_of_segments;
546 	struct dcssblk_dev_info *dev_info;
547 	struct segment_info *seg_info, *temp;
548 	char *local_buf;
549 	unsigned long seg_byte_size;
550 
551 	dev_info = NULL;
552 	seg_info = NULL;
553 	if (dev != dcssblk_root_dev) {
554 		rc = -EINVAL;
555 		goto out_nobuf;
556 	}
557 	if ((count < 1) || (buf[0] == '\0') || (buf[0] == '\n')) {
558 		rc = -ENAMETOOLONG;
559 		goto out_nobuf;
560 	}
561 
562 	local_buf = kmalloc(count + 1, GFP_KERNEL);
563 	if (local_buf == NULL) {
564 		rc = -ENOMEM;
565 		goto out_nobuf;
566 	}
567 
568 	/*
569 	 * parse input
570 	 */
571 	num_of_segments = 0;
572 	for (i = 0; (i < count && (buf[i] != '\0') && (buf[i] != '\n')); i++) {
573 		for (j = i; j < count &&
574 			(buf[j] != ':') &&
575 			(buf[j] != '\0') &&
576 			(buf[j] != '\n'); j++) {
577 			local_buf[j-i] = toupper(buf[j]);
578 		}
579 		local_buf[j-i] = '\0';
580 		if (((j - i) == 0) || ((j - i) > 8)) {
581 			rc = -ENAMETOOLONG;
582 			goto seg_list_del;
583 		}
584 
585 		rc = dcssblk_load_segment(local_buf, &seg_info);
586 		if (rc < 0)
587 			goto seg_list_del;
588 		/*
589 		 * get a struct dcssblk_dev_info
590 		 */
591 		if (num_of_segments == 0) {
592 			dev_info = kzalloc(sizeof(struct dcssblk_dev_info),
593 					GFP_KERNEL);
594 			if (dev_info == NULL) {
595 				rc = -ENOMEM;
596 				goto out;
597 			}
598 			strcpy(dev_info->segment_name, local_buf);
599 			dev_info->segment_type = seg_info->segment_type;
600 			INIT_LIST_HEAD(&dev_info->seg_list);
601 		}
602 		list_add_tail(&seg_info->lh, &dev_info->seg_list);
603 		num_of_segments++;
604 		i = j;
605 
606 		if ((buf[j] == '\0') || (buf[j] == '\n'))
607 			break;
608 	}
609 
610 	/* no trailing colon at the end of the input */
611 	if ((i > 0) && (buf[i-1] == ':')) {
612 		rc = -ENAMETOOLONG;
613 		goto seg_list_del;
614 	}
615 	strlcpy(local_buf, buf, i + 1);
616 	dev_info->num_of_segments = num_of_segments;
617 	rc = dcssblk_is_continuous(dev_info);
618 	if (rc < 0)
619 		goto seg_list_del;
620 
621 	dev_info->start = dcssblk_find_lowest_addr(dev_info);
622 	dev_info->end = dcssblk_find_highest_addr(dev_info);
623 
624 	dev_set_name(&dev_info->dev, "%s", dev_info->segment_name);
625 	dev_info->dev.release = dcssblk_release_segment;
626 	dev_info->dev.groups = dcssblk_dev_attr_groups;
627 	INIT_LIST_HEAD(&dev_info->lh);
628 	dev_info->gd = blk_alloc_disk(NUMA_NO_NODE);
629 	if (dev_info->gd == NULL) {
630 		rc = -ENOMEM;
631 		goto seg_list_del;
632 	}
633 	dev_info->gd->major = dcssblk_major;
634 	dev_info->gd->minors = DCSSBLK_MINORS_PER_DISK;
635 	dev_info->gd->fops = &dcssblk_devops;
636 	dev_info->gd->private_data = dev_info;
637 	blk_queue_logical_block_size(dev_info->gd->queue, 4096);
638 	blk_queue_flag_set(QUEUE_FLAG_DAX, dev_info->gd->queue);
639 
640 	seg_byte_size = (dev_info->end - dev_info->start + 1);
641 	set_capacity(dev_info->gd, seg_byte_size >> 9); // size in sectors
642 	pr_info("Loaded %s with total size %lu bytes and capacity %lu "
643 		"sectors\n", local_buf, seg_byte_size, seg_byte_size >> 9);
644 
645 	dev_info->save_pending = 0;
646 	dev_info->is_shared = 1;
647 	dev_info->dev.parent = dcssblk_root_dev;
648 
649 	/*
650 	 *get minor, add to list
651 	 */
652 	down_write(&dcssblk_devices_sem);
653 	if (dcssblk_get_segment_by_name(local_buf)) {
654 		rc = -EEXIST;
655 		goto release_gd;
656 	}
657 	rc = dcssblk_assign_free_minor(dev_info);
658 	if (rc)
659 		goto release_gd;
660 	sprintf(dev_info->gd->disk_name, "dcssblk%d",
661 		dev_info->gd->first_minor);
662 	list_add_tail(&dev_info->lh, &dcssblk_devices);
663 
664 	if (!try_module_get(THIS_MODULE)) {
665 		rc = -ENODEV;
666 		goto dev_list_del;
667 	}
668 	/*
669 	 * register the device
670 	 */
671 	rc = device_register(&dev_info->dev);
672 	if (rc)
673 		goto put_dev;
674 
675 	dev_info->dax_dev = alloc_dax(dev_info, &dcssblk_dax_ops);
676 	if (IS_ERR(dev_info->dax_dev)) {
677 		rc = PTR_ERR(dev_info->dax_dev);
678 		dev_info->dax_dev = NULL;
679 		goto put_dev;
680 	}
681 	set_dax_synchronous(dev_info->dax_dev);
682 	rc = dax_add_host(dev_info->dax_dev, dev_info->gd);
683 	if (rc)
684 		goto out_dax;
685 
686 	get_device(&dev_info->dev);
687 	rc = device_add_disk(&dev_info->dev, dev_info->gd, NULL);
688 	if (rc)
689 		goto out_dax_host;
690 
691 	switch (dev_info->segment_type) {
692 		case SEG_TYPE_SR:
693 		case SEG_TYPE_ER:
694 		case SEG_TYPE_SC:
695 			set_disk_ro(dev_info->gd,1);
696 			break;
697 		default:
698 			set_disk_ro(dev_info->gd,0);
699 			break;
700 	}
701 	up_write(&dcssblk_devices_sem);
702 	rc = count;
703 	goto out;
704 
705 out_dax_host:
706 	dax_remove_host(dev_info->gd);
707 out_dax:
708 	put_device(&dev_info->dev);
709 	kill_dax(dev_info->dax_dev);
710 	put_dax(dev_info->dax_dev);
711 put_dev:
712 	list_del(&dev_info->lh);
713 	blk_cleanup_disk(dev_info->gd);
714 	list_for_each_entry(seg_info, &dev_info->seg_list, lh) {
715 		segment_unload(seg_info->segment_name);
716 	}
717 	put_device(&dev_info->dev);
718 	up_write(&dcssblk_devices_sem);
719 	goto out;
720 dev_list_del:
721 	list_del(&dev_info->lh);
722 release_gd:
723 	blk_cleanup_disk(dev_info->gd);
724 	up_write(&dcssblk_devices_sem);
725 seg_list_del:
726 	if (dev_info == NULL)
727 		goto out;
728 	list_for_each_entry_safe(seg_info, temp, &dev_info->seg_list, lh) {
729 		list_del(&seg_info->lh);
730 		segment_unload(seg_info->segment_name);
731 		kfree(seg_info);
732 	}
733 	kfree(dev_info);
734 out:
735 	kfree(local_buf);
736 out_nobuf:
737 	return rc;
738 }
739 
740 /*
741  * device attribute for removing devices
742  */
743 static ssize_t
744 dcssblk_remove_store(struct device *dev, struct device_attribute *attr, const char *buf, size_t count)
745 {
746 	struct dcssblk_dev_info *dev_info;
747 	struct segment_info *entry;
748 	int rc, i;
749 	char *local_buf;
750 
751 	if (dev != dcssblk_root_dev) {
752 		return -EINVAL;
753 	}
754 	local_buf = kmalloc(count + 1, GFP_KERNEL);
755 	if (local_buf == NULL) {
756 		return -ENOMEM;
757 	}
758 	/*
759 	 * parse input
760 	 */
761 	for (i = 0; (i < count && (*(buf+i)!='\0') && (*(buf+i)!='\n')); i++) {
762 		local_buf[i] = toupper(buf[i]);
763 	}
764 	local_buf[i] = '\0';
765 	if ((i == 0) || (i > 8)) {
766 		rc = -ENAMETOOLONG;
767 		goto out_buf;
768 	}
769 
770 	down_write(&dcssblk_devices_sem);
771 	dev_info = dcssblk_get_device_by_name(local_buf);
772 	if (dev_info == NULL) {
773 		up_write(&dcssblk_devices_sem);
774 		pr_warn("Device %s cannot be removed because it is not a known device\n",
775 			local_buf);
776 		rc = -ENODEV;
777 		goto out_buf;
778 	}
779 	if (atomic_read(&dev_info->use_count) != 0) {
780 		up_write(&dcssblk_devices_sem);
781 		pr_warn("Device %s cannot be removed while it is in use\n",
782 			local_buf);
783 		rc = -EBUSY;
784 		goto out_buf;
785 	}
786 
787 	list_del(&dev_info->lh);
788 	kill_dax(dev_info->dax_dev);
789 	put_dax(dev_info->dax_dev);
790 	del_gendisk(dev_info->gd);
791 	blk_cleanup_disk(dev_info->gd);
792 
793 	/* unload all related segments */
794 	list_for_each_entry(entry, &dev_info->seg_list, lh)
795 		segment_unload(entry->segment_name);
796 
797 	up_write(&dcssblk_devices_sem);
798 
799 	device_unregister(&dev_info->dev);
800 	put_device(&dev_info->dev);
801 
802 	rc = count;
803 out_buf:
804 	kfree(local_buf);
805 	return rc;
806 }
807 
808 static int
809 dcssblk_open(struct block_device *bdev, fmode_t mode)
810 {
811 	struct dcssblk_dev_info *dev_info;
812 	int rc;
813 
814 	dev_info = bdev->bd_disk->private_data;
815 	if (NULL == dev_info) {
816 		rc = -ENODEV;
817 		goto out;
818 	}
819 	atomic_inc(&dev_info->use_count);
820 	rc = 0;
821 out:
822 	return rc;
823 }
824 
825 static void
826 dcssblk_release(struct gendisk *disk, fmode_t mode)
827 {
828 	struct dcssblk_dev_info *dev_info = disk->private_data;
829 	struct segment_info *entry;
830 
831 	if (!dev_info) {
832 		WARN_ON(1);
833 		return;
834 	}
835 	down_write(&dcssblk_devices_sem);
836 	if (atomic_dec_and_test(&dev_info->use_count)
837 	    && (dev_info->save_pending)) {
838 		pr_info("Device %s has become idle and is being saved "
839 			"now\n", dev_info->segment_name);
840 		list_for_each_entry(entry, &dev_info->seg_list, lh) {
841 			if (entry->segment_type == SEG_TYPE_EN ||
842 			    entry->segment_type == SEG_TYPE_SN)
843 				pr_warn("DCSS %s is of type SN or EN and cannot"
844 					" be saved\n", entry->segment_name);
845 			else
846 				segment_save(entry->segment_name);
847 		}
848 		dev_info->save_pending = 0;
849 	}
850 	up_write(&dcssblk_devices_sem);
851 }
852 
853 static void
854 dcssblk_submit_bio(struct bio *bio)
855 {
856 	struct dcssblk_dev_info *dev_info;
857 	struct bio_vec bvec;
858 	struct bvec_iter iter;
859 	unsigned long index;
860 	unsigned long page_addr;
861 	unsigned long source_addr;
862 	unsigned long bytes_done;
863 
864 	blk_queue_split(&bio);
865 
866 	bytes_done = 0;
867 	dev_info = bio->bi_bdev->bd_disk->private_data;
868 	if (dev_info == NULL)
869 		goto fail;
870 	if ((bio->bi_iter.bi_sector & 7) != 0 ||
871 	    (bio->bi_iter.bi_size & 4095) != 0)
872 		/* Request is not page-aligned. */
873 		goto fail;
874 	/* verify data transfer direction */
875 	if (dev_info->is_shared) {
876 		switch (dev_info->segment_type) {
877 		case SEG_TYPE_SR:
878 		case SEG_TYPE_ER:
879 		case SEG_TYPE_SC:
880 			/* cannot write to these segments */
881 			if (bio_data_dir(bio) == WRITE) {
882 				pr_warn("Writing to %s failed because it is a read-only device\n",
883 					dev_name(&dev_info->dev));
884 				goto fail;
885 			}
886 		}
887 	}
888 
889 	index = (bio->bi_iter.bi_sector >> 3);
890 	bio_for_each_segment(bvec, bio, iter) {
891 		page_addr = (unsigned long)bvec_virt(&bvec);
892 		source_addr = dev_info->start + (index<<12) + bytes_done;
893 		if (unlikely((page_addr & 4095) != 0) || (bvec.bv_len & 4095) != 0)
894 			// More paranoia.
895 			goto fail;
896 		if (bio_data_dir(bio) == READ) {
897 			memcpy((void*)page_addr, (void*)source_addr,
898 				bvec.bv_len);
899 		} else {
900 			memcpy((void*)source_addr, (void*)page_addr,
901 				bvec.bv_len);
902 		}
903 		bytes_done += bvec.bv_len;
904 	}
905 	bio_endio(bio);
906 	return;
907 fail:
908 	bio_io_error(bio);
909 }
910 
911 static long
912 __dcssblk_direct_access(struct dcssblk_dev_info *dev_info, pgoff_t pgoff,
913 		long nr_pages, void **kaddr, pfn_t *pfn)
914 {
915 	resource_size_t offset = pgoff * PAGE_SIZE;
916 	unsigned long dev_sz;
917 
918 	dev_sz = dev_info->end - dev_info->start + 1;
919 	if (kaddr)
920 		*kaddr = (void *) dev_info->start + offset;
921 	if (pfn)
922 		*pfn = __pfn_to_pfn_t(PFN_DOWN(dev_info->start + offset),
923 				PFN_DEV|PFN_SPECIAL);
924 
925 	return (dev_sz - offset) / PAGE_SIZE;
926 }
927 
928 static long
929 dcssblk_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
930 		long nr_pages, void **kaddr, pfn_t *pfn)
931 {
932 	struct dcssblk_dev_info *dev_info = dax_get_private(dax_dev);
933 
934 	return __dcssblk_direct_access(dev_info, pgoff, nr_pages, kaddr, pfn);
935 }
936 
937 static void
938 dcssblk_check_params(void)
939 {
940 	int rc, i, j, k;
941 	char buf[DCSSBLK_PARM_LEN + 1];
942 	struct dcssblk_dev_info *dev_info;
943 
944 	for (i = 0; (i < DCSSBLK_PARM_LEN) && (dcssblk_segments[i] != '\0');
945 	     i++) {
946 		for (j = i; (j < DCSSBLK_PARM_LEN) &&
947 			    (dcssblk_segments[j] != ',')  &&
948 			    (dcssblk_segments[j] != '\0') &&
949 			    (dcssblk_segments[j] != '('); j++)
950 		{
951 			buf[j-i] = dcssblk_segments[j];
952 		}
953 		buf[j-i] = '\0';
954 		rc = dcssblk_add_store(dcssblk_root_dev, NULL, buf, j-i);
955 		if ((rc >= 0) && (dcssblk_segments[j] == '(')) {
956 			for (k = 0; (buf[k] != ':') && (buf[k] != '\0'); k++)
957 				buf[k] = toupper(buf[k]);
958 			buf[k] = '\0';
959 			if (!strncmp(&dcssblk_segments[j], "(local)", 7)) {
960 				down_read(&dcssblk_devices_sem);
961 				dev_info = dcssblk_get_device_by_name(buf);
962 				up_read(&dcssblk_devices_sem);
963 				if (dev_info)
964 					dcssblk_shared_store(&dev_info->dev,
965 							     NULL, "0\n", 2);
966 			}
967 		}
968 		while ((dcssblk_segments[j] != ',') &&
969 		       (dcssblk_segments[j] != '\0'))
970 		{
971 			j++;
972 		}
973 		if (dcssblk_segments[j] == '\0')
974 			break;
975 		i = j;
976 	}
977 }
978 
979 /*
980  * The init/exit functions.
981  */
982 static void __exit
983 dcssblk_exit(void)
984 {
985 	root_device_unregister(dcssblk_root_dev);
986 	unregister_blkdev(dcssblk_major, DCSSBLK_NAME);
987 }
988 
989 static int __init
990 dcssblk_init(void)
991 {
992 	int rc;
993 
994 	dcssblk_root_dev = root_device_register("dcssblk");
995 	if (IS_ERR(dcssblk_root_dev))
996 		return PTR_ERR(dcssblk_root_dev);
997 	rc = device_create_file(dcssblk_root_dev, &dev_attr_add);
998 	if (rc)
999 		goto out_root;
1000 	rc = device_create_file(dcssblk_root_dev, &dev_attr_remove);
1001 	if (rc)
1002 		goto out_root;
1003 	rc = register_blkdev(0, DCSSBLK_NAME);
1004 	if (rc < 0)
1005 		goto out_root;
1006 	dcssblk_major = rc;
1007 	init_rwsem(&dcssblk_devices_sem);
1008 
1009 	dcssblk_check_params();
1010 	return 0;
1011 
1012 out_root:
1013 	root_device_unregister(dcssblk_root_dev);
1014 
1015 	return rc;
1016 }
1017 
1018 module_init(dcssblk_init);
1019 module_exit(dcssblk_exit);
1020 
1021 module_param_string(segments, dcssblk_segments, DCSSBLK_PARM_LEN, 0444);
1022 MODULE_PARM_DESC(segments, "Name of DCSS segment(s) to be loaded, "
1023 		 "comma-separated list, names in each set separated "
1024 		 "by commas are separated by colons, each set contains "
1025 		 "names of contiguous segments and each name max. 8 chars.\n"
1026 		 "Adding \"(local)\" to the end of each set equals echoing 0 "
1027 		 "to /sys/devices/dcssblk/<device name>/shared after loading "
1028 		 "the contiguous segments - \n"
1029 		 "e.g. segments=\"mydcss1,mydcss2:mydcss3,mydcss4(local)\"");
1030 
1031 MODULE_LICENSE("GPL");
1032