xref: /openbmc/linux/drivers/s390/block/dcssblk.c (revision 0c874100)
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/platform_device.h>
21 #include <linux/pfn_t.h>
22 #include <linux/uio.h>
23 #include <linux/dax.h>
24 #include <asm/extmem.h>
25 #include <asm/io.h>
26 
27 #define DCSSBLK_NAME "dcssblk"
28 #define DCSSBLK_MINORS_PER_DISK 1
29 #define DCSSBLK_PARM_LEN 400
30 #define DCSS_BUS_ID_SIZE 20
31 
32 static int dcssblk_open(struct block_device *bdev, fmode_t mode);
33 static void dcssblk_release(struct gendisk *disk, fmode_t mode);
34 static blk_qc_t dcssblk_make_request(struct request_queue *q,
35 						struct bio *bio);
36 static long dcssblk_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
37 		long nr_pages, void **kaddr, pfn_t *pfn);
38 
39 static char dcssblk_segments[DCSSBLK_PARM_LEN] = "\0";
40 
41 static int dcssblk_major;
42 static const struct block_device_operations dcssblk_devops = {
43 	.owner   	= THIS_MODULE,
44 	.open    	= dcssblk_open,
45 	.release 	= dcssblk_release,
46 };
47 
48 static size_t dcssblk_dax_copy_from_iter(struct dax_device *dax_dev,
49 		pgoff_t pgoff, void *addr, size_t bytes, struct iov_iter *i)
50 {
51 	return copy_from_iter(addr, bytes, i);
52 }
53 
54 static size_t dcssblk_dax_copy_to_iter(struct dax_device *dax_dev,
55 		pgoff_t pgoff, void *addr, size_t bytes, struct iov_iter *i)
56 {
57 	return copy_to_iter(addr, bytes, i);
58 }
59 
60 static const struct dax_operations dcssblk_dax_ops = {
61 	.direct_access = dcssblk_dax_direct_access,
62 	.copy_from_iter = dcssblk_dax_copy_from_iter,
63 	.copy_to_iter = dcssblk_dax_copy_to_iter,
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 	struct request_queue *dcssblk_queue;
78 	int num_of_segments;
79 	struct list_head seg_list;
80 	struct dax_device *dax_dev;
81 };
82 
83 struct segment_info {
84 	struct list_head lh;
85 	char segment_name[DCSS_BUS_ID_SIZE];
86 	unsigned long start;
87 	unsigned long end;
88 	int segment_type;
89 };
90 
91 static ssize_t dcssblk_add_store(struct device * dev, struct device_attribute *attr, const char * buf,
92 				  size_t count);
93 static ssize_t dcssblk_remove_store(struct device * dev, struct device_attribute *attr, const char * buf,
94 				  size_t count);
95 
96 static DEVICE_ATTR(add, S_IWUSR, NULL, dcssblk_add_store);
97 static DEVICE_ATTR(remove, S_IWUSR, NULL, dcssblk_remove_store);
98 
99 static struct device *dcssblk_root_dev;
100 
101 static LIST_HEAD(dcssblk_devices);
102 static struct rw_semaphore dcssblk_devices_sem;
103 
104 /*
105  * release function for segment device.
106  */
107 static void
108 dcssblk_release_segment(struct device *dev)
109 {
110 	struct dcssblk_dev_info *dev_info;
111 	struct segment_info *entry, *temp;
112 
113 	dev_info = container_of(dev, struct dcssblk_dev_info, dev);
114 	list_for_each_entry_safe(entry, temp, &dev_info->seg_list, lh) {
115 		list_del(&entry->lh);
116 		kfree(entry);
117 	}
118 	kfree(dev_info);
119 	module_put(THIS_MODULE);
120 }
121 
122 /*
123  * get a minor number. needs to be called with
124  * down_write(&dcssblk_devices_sem) and the
125  * device needs to be enqueued before the semaphore is
126  * freed.
127  */
128 static int
129 dcssblk_assign_free_minor(struct dcssblk_dev_info *dev_info)
130 {
131 	int minor, found;
132 	struct dcssblk_dev_info *entry;
133 
134 	if (dev_info == NULL)
135 		return -EINVAL;
136 	for (minor = 0; minor < (1<<MINORBITS); minor++) {
137 		found = 0;
138 		// test if minor available
139 		list_for_each_entry(entry, &dcssblk_devices, lh)
140 			if (minor == entry->gd->first_minor)
141 				found++;
142 		if (!found) break; // got unused minor
143 	}
144 	if (found)
145 		return -EBUSY;
146 	dev_info->gd->first_minor = minor;
147 	return 0;
148 }
149 
150 /*
151  * get the struct dcssblk_dev_info from dcssblk_devices
152  * for the given name.
153  * down_read(&dcssblk_devices_sem) must be held.
154  */
155 static struct dcssblk_dev_info *
156 dcssblk_get_device_by_name(char *name)
157 {
158 	struct dcssblk_dev_info *entry;
159 
160 	list_for_each_entry(entry, &dcssblk_devices, lh) {
161 		if (!strcmp(name, entry->segment_name)) {
162 			return entry;
163 		}
164 	}
165 	return NULL;
166 }
167 
168 /*
169  * get the struct segment_info from seg_list
170  * for the given name.
171  * down_read(&dcssblk_devices_sem) must be held.
172  */
173 static struct segment_info *
174 dcssblk_get_segment_by_name(char *name)
175 {
176 	struct dcssblk_dev_info *dev_info;
177 	struct segment_info *entry;
178 
179 	list_for_each_entry(dev_info, &dcssblk_devices, lh) {
180 		list_for_each_entry(entry, &dev_info->seg_list, lh) {
181 			if (!strcmp(name, entry->segment_name))
182 				return entry;
183 		}
184 	}
185 	return NULL;
186 }
187 
188 /*
189  * get the highest address of the multi-segment block.
190  */
191 static unsigned long
192 dcssblk_find_highest_addr(struct dcssblk_dev_info *dev_info)
193 {
194 	unsigned long highest_addr;
195 	struct segment_info *entry;
196 
197 	highest_addr = 0;
198 	list_for_each_entry(entry, &dev_info->seg_list, lh) {
199 		if (highest_addr < entry->end)
200 			highest_addr = entry->end;
201 	}
202 	return highest_addr;
203 }
204 
205 /*
206  * get the lowest address of the multi-segment block.
207  */
208 static unsigned long
209 dcssblk_find_lowest_addr(struct dcssblk_dev_info *dev_info)
210 {
211 	int set_first;
212 	unsigned long lowest_addr;
213 	struct segment_info *entry;
214 
215 	set_first = 0;
216 	lowest_addr = 0;
217 	list_for_each_entry(entry, &dev_info->seg_list, lh) {
218 		if (set_first == 0) {
219 			lowest_addr = entry->start;
220 			set_first = 1;
221 		} else {
222 			if (lowest_addr > entry->start)
223 				lowest_addr = entry->start;
224 		}
225 	}
226 	return lowest_addr;
227 }
228 
229 /*
230  * Check continuity of segments.
231  */
232 static int
233 dcssblk_is_continuous(struct dcssblk_dev_info *dev_info)
234 {
235 	int i, j, rc;
236 	struct segment_info *sort_list, *entry, temp;
237 
238 	if (dev_info->num_of_segments <= 1)
239 		return 0;
240 
241 	sort_list = kcalloc(dev_info->num_of_segments,
242 			    sizeof(struct segment_info),
243 			    GFP_KERNEL);
244 	if (sort_list == NULL)
245 		return -ENOMEM;
246 	i = 0;
247 	list_for_each_entry(entry, &dev_info->seg_list, lh) {
248 		memcpy(&sort_list[i], entry, sizeof(struct segment_info));
249 		i++;
250 	}
251 
252 	/* sort segments */
253 	for (i = 0; i < dev_info->num_of_segments; i++)
254 		for (j = 0; j < dev_info->num_of_segments; j++)
255 			if (sort_list[j].start > sort_list[i].start) {
256 				memcpy(&temp, &sort_list[i],
257 					sizeof(struct segment_info));
258 				memcpy(&sort_list[i], &sort_list[j],
259 					sizeof(struct segment_info));
260 				memcpy(&sort_list[j], &temp,
261 					sizeof(struct segment_info));
262 			}
263 
264 	/* check continuity */
265 	for (i = 0; i < dev_info->num_of_segments - 1; i++) {
266 		if ((sort_list[i].end + 1) != sort_list[i+1].start) {
267 			pr_err("Adjacent DCSSs %s and %s are not "
268 			       "contiguous\n", sort_list[i].segment_name,
269 			       sort_list[i+1].segment_name);
270 			rc = -EINVAL;
271 			goto out;
272 		}
273 		/* EN and EW are allowed in a block device */
274 		if (sort_list[i].segment_type != sort_list[i+1].segment_type) {
275 			if (!(sort_list[i].segment_type & SEGMENT_EXCLUSIVE) ||
276 				(sort_list[i].segment_type == SEG_TYPE_ER) ||
277 				!(sort_list[i+1].segment_type &
278 				SEGMENT_EXCLUSIVE) ||
279 				(sort_list[i+1].segment_type == SEG_TYPE_ER)) {
280 				pr_err("DCSS %s and DCSS %s have "
281 				       "incompatible types\n",
282 				       sort_list[i].segment_name,
283 				       sort_list[i+1].segment_name);
284 				rc = -EINVAL;
285 				goto out;
286 			}
287 		}
288 	}
289 	rc = 0;
290 out:
291 	kfree(sort_list);
292 	return rc;
293 }
294 
295 /*
296  * Load a segment
297  */
298 static int
299 dcssblk_load_segment(char *name, struct segment_info **seg_info)
300 {
301 	int rc;
302 
303 	/* already loaded? */
304 	down_read(&dcssblk_devices_sem);
305 	*seg_info = dcssblk_get_segment_by_name(name);
306 	up_read(&dcssblk_devices_sem);
307 	if (*seg_info != NULL)
308 		return -EEXIST;
309 
310 	/* get a struct segment_info */
311 	*seg_info = kzalloc(sizeof(struct segment_info), GFP_KERNEL);
312 	if (*seg_info == NULL)
313 		return -ENOMEM;
314 
315 	strcpy((*seg_info)->segment_name, name);
316 
317 	/* load the segment */
318 	rc = segment_load(name, SEGMENT_SHARED,
319 			&(*seg_info)->start, &(*seg_info)->end);
320 	if (rc < 0) {
321 		segment_warning(rc, (*seg_info)->segment_name);
322 		kfree(*seg_info);
323 	} else {
324 		INIT_LIST_HEAD(&(*seg_info)->lh);
325 		(*seg_info)->segment_type = rc;
326 	}
327 	return rc;
328 }
329 
330 /*
331  * device attribute for switching shared/nonshared (exclusive)
332  * operation (show + store)
333  */
334 static ssize_t
335 dcssblk_shared_show(struct device *dev, struct device_attribute *attr, char *buf)
336 {
337 	struct dcssblk_dev_info *dev_info;
338 
339 	dev_info = container_of(dev, struct dcssblk_dev_info, dev);
340 	return sprintf(buf, dev_info->is_shared ? "1\n" : "0\n");
341 }
342 
343 static ssize_t
344 dcssblk_shared_store(struct device *dev, struct device_attribute *attr, const char *inbuf, size_t count)
345 {
346 	struct dcssblk_dev_info *dev_info;
347 	struct segment_info *entry, *temp;
348 	int rc;
349 
350 	if ((count > 1) && (inbuf[1] != '\n') && (inbuf[1] != '\0'))
351 		return -EINVAL;
352 	down_write(&dcssblk_devices_sem);
353 	dev_info = container_of(dev, struct dcssblk_dev_info, dev);
354 	if (atomic_read(&dev_info->use_count)) {
355 		rc = -EBUSY;
356 		goto out;
357 	}
358 	if (inbuf[0] == '1') {
359 		/* reload segments in shared mode */
360 		list_for_each_entry(entry, &dev_info->seg_list, lh) {
361 			rc = segment_modify_shared(entry->segment_name,
362 						SEGMENT_SHARED);
363 			if (rc < 0) {
364 				BUG_ON(rc == -EINVAL);
365 				if (rc != -EAGAIN)
366 					goto removeseg;
367 			}
368 		}
369 		dev_info->is_shared = 1;
370 		switch (dev_info->segment_type) {
371 		case SEG_TYPE_SR:
372 		case SEG_TYPE_ER:
373 		case SEG_TYPE_SC:
374 			set_disk_ro(dev_info->gd, 1);
375 		}
376 	} else if (inbuf[0] == '0') {
377 		/* reload segments in exclusive mode */
378 		if (dev_info->segment_type == SEG_TYPE_SC) {
379 			pr_err("DCSS %s is of type SC and cannot be "
380 			       "loaded as exclusive-writable\n",
381 			       dev_info->segment_name);
382 			rc = -EINVAL;
383 			goto out;
384 		}
385 		list_for_each_entry(entry, &dev_info->seg_list, lh) {
386 			rc = segment_modify_shared(entry->segment_name,
387 						   SEGMENT_EXCLUSIVE);
388 			if (rc < 0) {
389 				BUG_ON(rc == -EINVAL);
390 				if (rc != -EAGAIN)
391 					goto removeseg;
392 			}
393 		}
394 		dev_info->is_shared = 0;
395 		set_disk_ro(dev_info->gd, 0);
396 	} else {
397 		rc = -EINVAL;
398 		goto out;
399 	}
400 	rc = count;
401 	goto out;
402 
403 removeseg:
404 	pr_err("DCSS device %s is removed after a failed access mode "
405 	       "change\n", dev_info->segment_name);
406 	temp = entry;
407 	list_for_each_entry(entry, &dev_info->seg_list, lh) {
408 		if (entry != temp)
409 			segment_unload(entry->segment_name);
410 	}
411 	list_del(&dev_info->lh);
412 
413 	kill_dax(dev_info->dax_dev);
414 	put_dax(dev_info->dax_dev);
415 	del_gendisk(dev_info->gd);
416 	blk_cleanup_queue(dev_info->dcssblk_queue);
417 	dev_info->gd->queue = NULL;
418 	put_disk(dev_info->gd);
419 	up_write(&dcssblk_devices_sem);
420 
421 	if (device_remove_file_self(dev, attr)) {
422 		device_unregister(dev);
423 		put_device(dev);
424 	}
425 	return rc;
426 out:
427 	up_write(&dcssblk_devices_sem);
428 	return rc;
429 }
430 static DEVICE_ATTR(shared, S_IWUSR | S_IRUSR, dcssblk_shared_show,
431 		   dcssblk_shared_store);
432 
433 /*
434  * device attribute for save operation on current copy
435  * of the segment. If the segment is busy, saving will
436  * become pending until it gets released, which can be
437  * undone by storing a non-true value to this entry.
438  * (show + store)
439  */
440 static ssize_t
441 dcssblk_save_show(struct device *dev, struct device_attribute *attr, char *buf)
442 {
443 	struct dcssblk_dev_info *dev_info;
444 
445 	dev_info = container_of(dev, struct dcssblk_dev_info, dev);
446 	return sprintf(buf, dev_info->save_pending ? "1\n" : "0\n");
447 }
448 
449 static ssize_t
450 dcssblk_save_store(struct device *dev, struct device_attribute *attr, const char *inbuf, size_t count)
451 {
452 	struct dcssblk_dev_info *dev_info;
453 	struct segment_info *entry;
454 
455 	if ((count > 1) && (inbuf[1] != '\n') && (inbuf[1] != '\0'))
456 		return -EINVAL;
457 	dev_info = container_of(dev, struct dcssblk_dev_info, dev);
458 
459 	down_write(&dcssblk_devices_sem);
460 	if (inbuf[0] == '1') {
461 		if (atomic_read(&dev_info->use_count) == 0) {
462 			// device is idle => we save immediately
463 			pr_info("All DCSSs that map to device %s are "
464 				"saved\n", dev_info->segment_name);
465 			list_for_each_entry(entry, &dev_info->seg_list, lh) {
466 				if (entry->segment_type == SEG_TYPE_EN ||
467 				    entry->segment_type == SEG_TYPE_SN)
468 					pr_warn("DCSS %s is of type SN or EN"
469 						" and cannot be saved\n",
470 						entry->segment_name);
471 				else
472 					segment_save(entry->segment_name);
473 			}
474 		}  else {
475 			// device is busy => we save it when it becomes
476 			// idle in dcssblk_release
477 			pr_info("Device %s is in use, its DCSSs will be "
478 				"saved when it becomes idle\n",
479 				dev_info->segment_name);
480 			dev_info->save_pending = 1;
481 		}
482 	} else if (inbuf[0] == '0') {
483 		if (dev_info->save_pending) {
484 			// device is busy & the user wants to undo his save
485 			// request
486 			dev_info->save_pending = 0;
487 			pr_info("A pending save request for device %s "
488 				"has been canceled\n",
489 				dev_info->segment_name);
490 		}
491 	} else {
492 		up_write(&dcssblk_devices_sem);
493 		return -EINVAL;
494 	}
495 	up_write(&dcssblk_devices_sem);
496 	return count;
497 }
498 static DEVICE_ATTR(save, S_IWUSR | S_IRUSR, dcssblk_save_show,
499 		   dcssblk_save_store);
500 
501 /*
502  * device attribute for showing all segments in a device
503  */
504 static ssize_t
505 dcssblk_seglist_show(struct device *dev, struct device_attribute *attr,
506 		char *buf)
507 {
508 	int i;
509 
510 	struct dcssblk_dev_info *dev_info;
511 	struct segment_info *entry;
512 
513 	down_read(&dcssblk_devices_sem);
514 	dev_info = container_of(dev, struct dcssblk_dev_info, dev);
515 	i = 0;
516 	buf[0] = '\0';
517 	list_for_each_entry(entry, &dev_info->seg_list, lh) {
518 		strcpy(&buf[i], entry->segment_name);
519 		i += strlen(entry->segment_name);
520 		buf[i] = '\n';
521 		i++;
522 	}
523 	up_read(&dcssblk_devices_sem);
524 	return i;
525 }
526 static DEVICE_ATTR(seglist, S_IRUSR, dcssblk_seglist_show, NULL);
527 
528 static struct attribute *dcssblk_dev_attrs[] = {
529 	&dev_attr_shared.attr,
530 	&dev_attr_save.attr,
531 	&dev_attr_seglist.attr,
532 	NULL,
533 };
534 static struct attribute_group dcssblk_dev_attr_group = {
535 	.attrs = dcssblk_dev_attrs,
536 };
537 static const struct attribute_group *dcssblk_dev_attr_groups[] = {
538 	&dcssblk_dev_attr_group,
539 	NULL,
540 };
541 
542 /*
543  * device attribute for adding devices
544  */
545 static ssize_t
546 dcssblk_add_store(struct device *dev, struct device_attribute *attr, const char *buf, size_t count)
547 {
548 	int rc, i, j, num_of_segments;
549 	struct dcssblk_dev_info *dev_info;
550 	struct segment_info *seg_info, *temp;
551 	char *local_buf;
552 	unsigned long seg_byte_size;
553 
554 	dev_info = NULL;
555 	seg_info = NULL;
556 	if (dev != dcssblk_root_dev) {
557 		rc = -EINVAL;
558 		goto out_nobuf;
559 	}
560 	if ((count < 1) || (buf[0] == '\0') || (buf[0] == '\n')) {
561 		rc = -ENAMETOOLONG;
562 		goto out_nobuf;
563 	}
564 
565 	local_buf = kmalloc(count + 1, GFP_KERNEL);
566 	if (local_buf == NULL) {
567 		rc = -ENOMEM;
568 		goto out_nobuf;
569 	}
570 
571 	/*
572 	 * parse input
573 	 */
574 	num_of_segments = 0;
575 	for (i = 0; (i < count && (buf[i] != '\0') && (buf[i] != '\n')); i++) {
576 		for (j = i; j < count &&
577 			(buf[j] != ':') &&
578 			(buf[j] != '\0') &&
579 			(buf[j] != '\n'); j++) {
580 			local_buf[j-i] = toupper(buf[j]);
581 		}
582 		local_buf[j-i] = '\0';
583 		if (((j - i) == 0) || ((j - i) > 8)) {
584 			rc = -ENAMETOOLONG;
585 			goto seg_list_del;
586 		}
587 
588 		rc = dcssblk_load_segment(local_buf, &seg_info);
589 		if (rc < 0)
590 			goto seg_list_del;
591 		/*
592 		 * get a struct dcssblk_dev_info
593 		 */
594 		if (num_of_segments == 0) {
595 			dev_info = kzalloc(sizeof(struct dcssblk_dev_info),
596 					GFP_KERNEL);
597 			if (dev_info == NULL) {
598 				rc = -ENOMEM;
599 				goto out;
600 			}
601 			strcpy(dev_info->segment_name, local_buf);
602 			dev_info->segment_type = seg_info->segment_type;
603 			INIT_LIST_HEAD(&dev_info->seg_list);
604 		}
605 		list_add_tail(&seg_info->lh, &dev_info->seg_list);
606 		num_of_segments++;
607 		i = j;
608 
609 		if ((buf[j] == '\0') || (buf[j] == '\n'))
610 			break;
611 	}
612 
613 	/* no trailing colon at the end of the input */
614 	if ((i > 0) && (buf[i-1] == ':')) {
615 		rc = -ENAMETOOLONG;
616 		goto seg_list_del;
617 	}
618 	strlcpy(local_buf, buf, i + 1);
619 	dev_info->num_of_segments = num_of_segments;
620 	rc = dcssblk_is_continuous(dev_info);
621 	if (rc < 0)
622 		goto seg_list_del;
623 
624 	dev_info->start = dcssblk_find_lowest_addr(dev_info);
625 	dev_info->end = dcssblk_find_highest_addr(dev_info);
626 
627 	dev_set_name(&dev_info->dev, "%s", dev_info->segment_name);
628 	dev_info->dev.release = dcssblk_release_segment;
629 	dev_info->dev.groups = dcssblk_dev_attr_groups;
630 	INIT_LIST_HEAD(&dev_info->lh);
631 	dev_info->gd = alloc_disk(DCSSBLK_MINORS_PER_DISK);
632 	if (dev_info->gd == NULL) {
633 		rc = -ENOMEM;
634 		goto seg_list_del;
635 	}
636 	dev_info->gd->major = dcssblk_major;
637 	dev_info->gd->fops = &dcssblk_devops;
638 	dev_info->dcssblk_queue = blk_alloc_queue(GFP_KERNEL);
639 	dev_info->gd->queue = dev_info->dcssblk_queue;
640 	dev_info->gd->private_data = dev_info;
641 	blk_queue_make_request(dev_info->dcssblk_queue, dcssblk_make_request);
642 	blk_queue_logical_block_size(dev_info->dcssblk_queue, 4096);
643 	blk_queue_flag_set(QUEUE_FLAG_DAX, dev_info->dcssblk_queue);
644 
645 	seg_byte_size = (dev_info->end - dev_info->start + 1);
646 	set_capacity(dev_info->gd, seg_byte_size >> 9); // size in sectors
647 	pr_info("Loaded %s with total size %lu bytes and capacity %lu "
648 		"sectors\n", local_buf, seg_byte_size, seg_byte_size >> 9);
649 
650 	dev_info->save_pending = 0;
651 	dev_info->is_shared = 1;
652 	dev_info->dev.parent = dcssblk_root_dev;
653 
654 	/*
655 	 *get minor, add to list
656 	 */
657 	down_write(&dcssblk_devices_sem);
658 	if (dcssblk_get_segment_by_name(local_buf)) {
659 		rc = -EEXIST;
660 		goto release_gd;
661 	}
662 	rc = dcssblk_assign_free_minor(dev_info);
663 	if (rc)
664 		goto release_gd;
665 	sprintf(dev_info->gd->disk_name, "dcssblk%d",
666 		dev_info->gd->first_minor);
667 	list_add_tail(&dev_info->lh, &dcssblk_devices);
668 
669 	if (!try_module_get(THIS_MODULE)) {
670 		rc = -ENODEV;
671 		goto dev_list_del;
672 	}
673 	/*
674 	 * register the device
675 	 */
676 	rc = device_register(&dev_info->dev);
677 	if (rc)
678 		goto put_dev;
679 
680 	dev_info->dax_dev = alloc_dax(dev_info, dev_info->gd->disk_name,
681 			&dcssblk_dax_ops);
682 	if (!dev_info->dax_dev) {
683 		rc = -ENOMEM;
684 		goto put_dev;
685 	}
686 
687 	get_device(&dev_info->dev);
688 	device_add_disk(&dev_info->dev, dev_info->gd, NULL);
689 
690 	switch (dev_info->segment_type) {
691 		case SEG_TYPE_SR:
692 		case SEG_TYPE_ER:
693 		case SEG_TYPE_SC:
694 			set_disk_ro(dev_info->gd,1);
695 			break;
696 		default:
697 			set_disk_ro(dev_info->gd,0);
698 			break;
699 	}
700 	up_write(&dcssblk_devices_sem);
701 	rc = count;
702 	goto out;
703 
704 put_dev:
705 	list_del(&dev_info->lh);
706 	blk_cleanup_queue(dev_info->dcssblk_queue);
707 	dev_info->gd->queue = NULL;
708 	put_disk(dev_info->gd);
709 	list_for_each_entry(seg_info, &dev_info->seg_list, lh) {
710 		segment_unload(seg_info->segment_name);
711 	}
712 	put_device(&dev_info->dev);
713 	up_write(&dcssblk_devices_sem);
714 	goto out;
715 dev_list_del:
716 	list_del(&dev_info->lh);
717 release_gd:
718 	blk_cleanup_queue(dev_info->dcssblk_queue);
719 	dev_info->gd->queue = NULL;
720 	put_disk(dev_info->gd);
721 	up_write(&dcssblk_devices_sem);
722 seg_list_del:
723 	if (dev_info == NULL)
724 		goto out;
725 	list_for_each_entry_safe(seg_info, temp, &dev_info->seg_list, lh) {
726 		list_del(&seg_info->lh);
727 		segment_unload(seg_info->segment_name);
728 		kfree(seg_info);
729 	}
730 	kfree(dev_info);
731 out:
732 	kfree(local_buf);
733 out_nobuf:
734 	return rc;
735 }
736 
737 /*
738  * device attribute for removing devices
739  */
740 static ssize_t
741 dcssblk_remove_store(struct device *dev, struct device_attribute *attr, const char *buf, size_t count)
742 {
743 	struct dcssblk_dev_info *dev_info;
744 	struct segment_info *entry;
745 	int rc, i;
746 	char *local_buf;
747 
748 	if (dev != dcssblk_root_dev) {
749 		return -EINVAL;
750 	}
751 	local_buf = kmalloc(count + 1, GFP_KERNEL);
752 	if (local_buf == NULL) {
753 		return -ENOMEM;
754 	}
755 	/*
756 	 * parse input
757 	 */
758 	for (i = 0; (i < count && (*(buf+i)!='\0') && (*(buf+i)!='\n')); i++) {
759 		local_buf[i] = toupper(buf[i]);
760 	}
761 	local_buf[i] = '\0';
762 	if ((i == 0) || (i > 8)) {
763 		rc = -ENAMETOOLONG;
764 		goto out_buf;
765 	}
766 
767 	down_write(&dcssblk_devices_sem);
768 	dev_info = dcssblk_get_device_by_name(local_buf);
769 	if (dev_info == NULL) {
770 		up_write(&dcssblk_devices_sem);
771 		pr_warn("Device %s cannot be removed because it is not a known device\n",
772 			local_buf);
773 		rc = -ENODEV;
774 		goto out_buf;
775 	}
776 	if (atomic_read(&dev_info->use_count) != 0) {
777 		up_write(&dcssblk_devices_sem);
778 		pr_warn("Device %s cannot be removed while it is in use\n",
779 			local_buf);
780 		rc = -EBUSY;
781 		goto out_buf;
782 	}
783 
784 	list_del(&dev_info->lh);
785 	kill_dax(dev_info->dax_dev);
786 	put_dax(dev_info->dax_dev);
787 	del_gendisk(dev_info->gd);
788 	blk_cleanup_queue(dev_info->dcssblk_queue);
789 	dev_info->gd->queue = NULL;
790 	put_disk(dev_info->gd);
791 
792 	/* unload all related segments */
793 	list_for_each_entry(entry, &dev_info->seg_list, lh)
794 		segment_unload(entry->segment_name);
795 
796 	up_write(&dcssblk_devices_sem);
797 
798 	device_unregister(&dev_info->dev);
799 	put_device(&dev_info->dev);
800 
801 	rc = count;
802 out_buf:
803 	kfree(local_buf);
804 	return rc;
805 }
806 
807 static int
808 dcssblk_open(struct block_device *bdev, fmode_t mode)
809 {
810 	struct dcssblk_dev_info *dev_info;
811 	int rc;
812 
813 	dev_info = bdev->bd_disk->private_data;
814 	if (NULL == dev_info) {
815 		rc = -ENODEV;
816 		goto out;
817 	}
818 	atomic_inc(&dev_info->use_count);
819 	bdev->bd_block_size = 4096;
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 blk_qc_t
854 dcssblk_make_request(struct request_queue *q, 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(q, &bio);
865 
866 	bytes_done = 0;
867 	dev_info = bio->bi_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 	if (bio_end_sector(bio) > get_capacity(bio->bi_disk)) {
875 		/* Request beyond end of DCSS segment. */
876 		goto fail;
877 	}
878 	/* verify data transfer direction */
879 	if (dev_info->is_shared) {
880 		switch (dev_info->segment_type) {
881 		case SEG_TYPE_SR:
882 		case SEG_TYPE_ER:
883 		case SEG_TYPE_SC:
884 			/* cannot write to these segments */
885 			if (bio_data_dir(bio) == WRITE) {
886 				pr_warn("Writing to %s failed because it is a read-only device\n",
887 					dev_name(&dev_info->dev));
888 				goto fail;
889 			}
890 		}
891 	}
892 
893 	index = (bio->bi_iter.bi_sector >> 3);
894 	bio_for_each_segment(bvec, bio, iter) {
895 		page_addr = (unsigned long)
896 			page_address(bvec.bv_page) + bvec.bv_offset;
897 		source_addr = dev_info->start + (index<<12) + bytes_done;
898 		if (unlikely((page_addr & 4095) != 0) || (bvec.bv_len & 4095) != 0)
899 			// More paranoia.
900 			goto fail;
901 		if (bio_data_dir(bio) == READ) {
902 			memcpy((void*)page_addr, (void*)source_addr,
903 				bvec.bv_len);
904 		} else {
905 			memcpy((void*)source_addr, (void*)page_addr,
906 				bvec.bv_len);
907 		}
908 		bytes_done += bvec.bv_len;
909 	}
910 	bio_endio(bio);
911 	return BLK_QC_T_NONE;
912 fail:
913 	bio_io_error(bio);
914 	return BLK_QC_T_NONE;
915 }
916 
917 static long
918 __dcssblk_direct_access(struct dcssblk_dev_info *dev_info, pgoff_t pgoff,
919 		long nr_pages, void **kaddr, pfn_t *pfn)
920 {
921 	resource_size_t offset = pgoff * PAGE_SIZE;
922 	unsigned long dev_sz;
923 
924 	dev_sz = dev_info->end - dev_info->start + 1;
925 	if (kaddr)
926 		*kaddr = (void *) dev_info->start + offset;
927 	if (pfn)
928 		*pfn = __pfn_to_pfn_t(PFN_DOWN(dev_info->start + offset),
929 				PFN_DEV|PFN_SPECIAL);
930 
931 	return (dev_sz - offset) / PAGE_SIZE;
932 }
933 
934 static long
935 dcssblk_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff,
936 		long nr_pages, void **kaddr, pfn_t *pfn)
937 {
938 	struct dcssblk_dev_info *dev_info = dax_get_private(dax_dev);
939 
940 	return __dcssblk_direct_access(dev_info, pgoff, nr_pages, kaddr, pfn);
941 }
942 
943 static void
944 dcssblk_check_params(void)
945 {
946 	int rc, i, j, k;
947 	char buf[DCSSBLK_PARM_LEN + 1];
948 	struct dcssblk_dev_info *dev_info;
949 
950 	for (i = 0; (i < DCSSBLK_PARM_LEN) && (dcssblk_segments[i] != '\0');
951 	     i++) {
952 		for (j = i; (j < DCSSBLK_PARM_LEN) &&
953 			    (dcssblk_segments[j] != ',')  &&
954 			    (dcssblk_segments[j] != '\0') &&
955 			    (dcssblk_segments[j] != '('); j++)
956 		{
957 			buf[j-i] = dcssblk_segments[j];
958 		}
959 		buf[j-i] = '\0';
960 		rc = dcssblk_add_store(dcssblk_root_dev, NULL, buf, j-i);
961 		if ((rc >= 0) && (dcssblk_segments[j] == '(')) {
962 			for (k = 0; (buf[k] != ':') && (buf[k] != '\0'); k++)
963 				buf[k] = toupper(buf[k]);
964 			buf[k] = '\0';
965 			if (!strncmp(&dcssblk_segments[j], "(local)", 7)) {
966 				down_read(&dcssblk_devices_sem);
967 				dev_info = dcssblk_get_device_by_name(buf);
968 				up_read(&dcssblk_devices_sem);
969 				if (dev_info)
970 					dcssblk_shared_store(&dev_info->dev,
971 							     NULL, "0\n", 2);
972 			}
973 		}
974 		while ((dcssblk_segments[j] != ',') &&
975 		       (dcssblk_segments[j] != '\0'))
976 		{
977 			j++;
978 		}
979 		if (dcssblk_segments[j] == '\0')
980 			break;
981 		i = j;
982 	}
983 }
984 
985 /*
986  * Suspend / Resume
987  */
988 static int dcssblk_freeze(struct device *dev)
989 {
990 	struct dcssblk_dev_info *dev_info;
991 	int rc = 0;
992 
993 	list_for_each_entry(dev_info, &dcssblk_devices, lh) {
994 		switch (dev_info->segment_type) {
995 			case SEG_TYPE_SR:
996 			case SEG_TYPE_ER:
997 			case SEG_TYPE_SC:
998 				if (!dev_info->is_shared)
999 					rc = -EINVAL;
1000 				break;
1001 			default:
1002 				rc = -EINVAL;
1003 				break;
1004 		}
1005 		if (rc)
1006 			break;
1007 	}
1008 	if (rc)
1009 		pr_err("Suspending the system failed because DCSS device %s "
1010 		       "is writable\n",
1011 		       dev_info->segment_name);
1012 	return rc;
1013 }
1014 
1015 static int dcssblk_restore(struct device *dev)
1016 {
1017 	struct dcssblk_dev_info *dev_info;
1018 	struct segment_info *entry;
1019 	unsigned long start, end;
1020 	int rc = 0;
1021 
1022 	list_for_each_entry(dev_info, &dcssblk_devices, lh) {
1023 		list_for_each_entry(entry, &dev_info->seg_list, lh) {
1024 			segment_unload(entry->segment_name);
1025 			rc = segment_load(entry->segment_name, SEGMENT_SHARED,
1026 					  &start, &end);
1027 			if (rc < 0) {
1028 // TODO in_use check ?
1029 				segment_warning(rc, entry->segment_name);
1030 				goto out_panic;
1031 			}
1032 			if (start != entry->start || end != entry->end) {
1033 				pr_err("The address range of DCSS %s changed "
1034 				       "while the system was suspended\n",
1035 				       entry->segment_name);
1036 				goto out_panic;
1037 			}
1038 		}
1039 	}
1040 	return 0;
1041 out_panic:
1042 	panic("fatal dcssblk resume error\n");
1043 }
1044 
1045 static int dcssblk_thaw(struct device *dev)
1046 {
1047 	return 0;
1048 }
1049 
1050 static const struct dev_pm_ops dcssblk_pm_ops = {
1051 	.freeze		= dcssblk_freeze,
1052 	.thaw		= dcssblk_thaw,
1053 	.restore	= dcssblk_restore,
1054 };
1055 
1056 static struct platform_driver dcssblk_pdrv = {
1057 	.driver = {
1058 		.name	= "dcssblk",
1059 		.pm	= &dcssblk_pm_ops,
1060 	},
1061 };
1062 
1063 static struct platform_device *dcssblk_pdev;
1064 
1065 
1066 /*
1067  * The init/exit functions.
1068  */
1069 static void __exit
1070 dcssblk_exit(void)
1071 {
1072 	platform_device_unregister(dcssblk_pdev);
1073 	platform_driver_unregister(&dcssblk_pdrv);
1074 	root_device_unregister(dcssblk_root_dev);
1075 	unregister_blkdev(dcssblk_major, DCSSBLK_NAME);
1076 }
1077 
1078 static int __init
1079 dcssblk_init(void)
1080 {
1081 	int rc;
1082 
1083 	rc = platform_driver_register(&dcssblk_pdrv);
1084 	if (rc)
1085 		return rc;
1086 
1087 	dcssblk_pdev = platform_device_register_simple("dcssblk", -1, NULL,
1088 							0);
1089 	if (IS_ERR(dcssblk_pdev)) {
1090 		rc = PTR_ERR(dcssblk_pdev);
1091 		goto out_pdrv;
1092 	}
1093 
1094 	dcssblk_root_dev = root_device_register("dcssblk");
1095 	if (IS_ERR(dcssblk_root_dev)) {
1096 		rc = PTR_ERR(dcssblk_root_dev);
1097 		goto out_pdev;
1098 	}
1099 	rc = device_create_file(dcssblk_root_dev, &dev_attr_add);
1100 	if (rc)
1101 		goto out_root;
1102 	rc = device_create_file(dcssblk_root_dev, &dev_attr_remove);
1103 	if (rc)
1104 		goto out_root;
1105 	rc = register_blkdev(0, DCSSBLK_NAME);
1106 	if (rc < 0)
1107 		goto out_root;
1108 	dcssblk_major = rc;
1109 	init_rwsem(&dcssblk_devices_sem);
1110 
1111 	dcssblk_check_params();
1112 	return 0;
1113 
1114 out_root:
1115 	root_device_unregister(dcssblk_root_dev);
1116 out_pdev:
1117 	platform_device_unregister(dcssblk_pdev);
1118 out_pdrv:
1119 	platform_driver_unregister(&dcssblk_pdrv);
1120 	return rc;
1121 }
1122 
1123 module_init(dcssblk_init);
1124 module_exit(dcssblk_exit);
1125 
1126 module_param_string(segments, dcssblk_segments, DCSSBLK_PARM_LEN, 0444);
1127 MODULE_PARM_DESC(segments, "Name of DCSS segment(s) to be loaded, "
1128 		 "comma-separated list, names in each set separated "
1129 		 "by commas are separated by colons, each set contains "
1130 		 "names of contiguous segments and each name max. 8 chars.\n"
1131 		 "Adding \"(local)\" to the end of each set equals echoing 0 "
1132 		 "to /sys/devices/dcssblk/<device name>/shared after loading "
1133 		 "the contiguous segments - \n"
1134 		 "e.g. segments=\"mydcss1,mydcss2:mydcss3,mydcss4(local)\"");
1135 
1136 MODULE_LICENSE("GPL");
1137