xref: /openbmc/linux/drivers/nvdimm/pfn_devs.c (revision 6db6b729)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright(c) 2013-2016 Intel Corporation. All rights reserved.
4  */
5 #include <linux/memremap.h>
6 #include <linux/blkdev.h>
7 #include <linux/device.h>
8 #include <linux/sizes.h>
9 #include <linux/slab.h>
10 #include <linux/fs.h>
11 #include <linux/mm.h>
12 #include "nd-core.h"
13 #include "pfn.h"
14 #include "nd.h"
15 
16 static const bool page_struct_override = IS_ENABLED(CONFIG_NVDIMM_KMSAN);
17 
18 static void nd_pfn_release(struct device *dev)
19 {
20 	struct nd_region *nd_region = to_nd_region(dev->parent);
21 	struct nd_pfn *nd_pfn = to_nd_pfn(dev);
22 
23 	dev_dbg(dev, "trace\n");
24 	nd_detach_ndns(&nd_pfn->dev, &nd_pfn->ndns);
25 	ida_simple_remove(&nd_region->pfn_ida, nd_pfn->id);
26 	kfree(nd_pfn->uuid);
27 	kfree(nd_pfn);
28 }
29 
30 struct nd_pfn *to_nd_pfn(struct device *dev)
31 {
32 	struct nd_pfn *nd_pfn = container_of(dev, struct nd_pfn, dev);
33 
34 	WARN_ON(!is_nd_pfn(dev));
35 	return nd_pfn;
36 }
37 EXPORT_SYMBOL(to_nd_pfn);
38 
39 static ssize_t mode_show(struct device *dev,
40 		struct device_attribute *attr, char *buf)
41 {
42 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
43 
44 	switch (nd_pfn->mode) {
45 	case PFN_MODE_RAM:
46 		return sprintf(buf, "ram\n");
47 	case PFN_MODE_PMEM:
48 		return sprintf(buf, "pmem\n");
49 	default:
50 		return sprintf(buf, "none\n");
51 	}
52 }
53 
54 static ssize_t mode_store(struct device *dev,
55 		struct device_attribute *attr, const char *buf, size_t len)
56 {
57 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
58 	ssize_t rc = 0;
59 
60 	device_lock(dev);
61 	nvdimm_bus_lock(dev);
62 	if (dev->driver)
63 		rc = -EBUSY;
64 	else {
65 		size_t n = len - 1;
66 
67 		if (strncmp(buf, "pmem\n", n) == 0
68 				|| strncmp(buf, "pmem", n) == 0) {
69 			nd_pfn->mode = PFN_MODE_PMEM;
70 		} else if (strncmp(buf, "ram\n", n) == 0
71 				|| strncmp(buf, "ram", n) == 0)
72 			nd_pfn->mode = PFN_MODE_RAM;
73 		else if (strncmp(buf, "none\n", n) == 0
74 				|| strncmp(buf, "none", n) == 0)
75 			nd_pfn->mode = PFN_MODE_NONE;
76 		else
77 			rc = -EINVAL;
78 	}
79 	dev_dbg(dev, "result: %zd wrote: %s%s", rc, buf,
80 			buf[len - 1] == '\n' ? "" : "\n");
81 	nvdimm_bus_unlock(dev);
82 	device_unlock(dev);
83 
84 	return rc ? rc : len;
85 }
86 static DEVICE_ATTR_RW(mode);
87 
88 static ssize_t align_show(struct device *dev,
89 		struct device_attribute *attr, char *buf)
90 {
91 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
92 
93 	return sprintf(buf, "%ld\n", nd_pfn->align);
94 }
95 
96 static unsigned long *nd_pfn_supported_alignments(unsigned long *alignments)
97 {
98 
99 	alignments[0] = PAGE_SIZE;
100 
101 	if (has_transparent_hugepage()) {
102 		alignments[1] = HPAGE_PMD_SIZE;
103 		if (has_transparent_pud_hugepage())
104 			alignments[2] = HPAGE_PUD_SIZE;
105 	}
106 
107 	return alignments;
108 }
109 
110 /*
111  * Use pmd mapping if supported as default alignment
112  */
113 static unsigned long nd_pfn_default_alignment(void)
114 {
115 
116 	if (has_transparent_hugepage())
117 		return HPAGE_PMD_SIZE;
118 	return PAGE_SIZE;
119 }
120 
121 static ssize_t align_store(struct device *dev,
122 		struct device_attribute *attr, const char *buf, size_t len)
123 {
124 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
125 	unsigned long aligns[MAX_NVDIMM_ALIGN] = { [0] = 0, };
126 	ssize_t rc;
127 
128 	device_lock(dev);
129 	nvdimm_bus_lock(dev);
130 	rc = nd_size_select_store(dev, buf, &nd_pfn->align,
131 			nd_pfn_supported_alignments(aligns));
132 	dev_dbg(dev, "result: %zd wrote: %s%s", rc, buf,
133 			buf[len - 1] == '\n' ? "" : "\n");
134 	nvdimm_bus_unlock(dev);
135 	device_unlock(dev);
136 
137 	return rc ? rc : len;
138 }
139 static DEVICE_ATTR_RW(align);
140 
141 static ssize_t uuid_show(struct device *dev,
142 		struct device_attribute *attr, char *buf)
143 {
144 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
145 
146 	if (nd_pfn->uuid)
147 		return sprintf(buf, "%pUb\n", nd_pfn->uuid);
148 	return sprintf(buf, "\n");
149 }
150 
151 static ssize_t uuid_store(struct device *dev,
152 		struct device_attribute *attr, const char *buf, size_t len)
153 {
154 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
155 	ssize_t rc;
156 
157 	device_lock(dev);
158 	rc = nd_uuid_store(dev, &nd_pfn->uuid, buf, len);
159 	dev_dbg(dev, "result: %zd wrote: %s%s", rc, buf,
160 			buf[len - 1] == '\n' ? "" : "\n");
161 	device_unlock(dev);
162 
163 	return rc ? rc : len;
164 }
165 static DEVICE_ATTR_RW(uuid);
166 
167 static ssize_t namespace_show(struct device *dev,
168 		struct device_attribute *attr, char *buf)
169 {
170 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
171 	ssize_t rc;
172 
173 	nvdimm_bus_lock(dev);
174 	rc = sprintf(buf, "%s\n", nd_pfn->ndns
175 			? dev_name(&nd_pfn->ndns->dev) : "");
176 	nvdimm_bus_unlock(dev);
177 	return rc;
178 }
179 
180 static ssize_t namespace_store(struct device *dev,
181 		struct device_attribute *attr, const char *buf, size_t len)
182 {
183 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
184 	ssize_t rc;
185 
186 	device_lock(dev);
187 	nvdimm_bus_lock(dev);
188 	rc = nd_namespace_store(dev, &nd_pfn->ndns, buf, len);
189 	dev_dbg(dev, "result: %zd wrote: %s%s", rc, buf,
190 			buf[len - 1] == '\n' ? "" : "\n");
191 	nvdimm_bus_unlock(dev);
192 	device_unlock(dev);
193 
194 	return rc;
195 }
196 static DEVICE_ATTR_RW(namespace);
197 
198 static ssize_t resource_show(struct device *dev,
199 		struct device_attribute *attr, char *buf)
200 {
201 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
202 	ssize_t rc;
203 
204 	device_lock(dev);
205 	if (dev->driver) {
206 		struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
207 		u64 offset = __le64_to_cpu(pfn_sb->dataoff);
208 		struct nd_namespace_common *ndns = nd_pfn->ndns;
209 		u32 start_pad = __le32_to_cpu(pfn_sb->start_pad);
210 		struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
211 
212 		rc = sprintf(buf, "%#llx\n", (unsigned long long) nsio->res.start
213 				+ start_pad + offset);
214 	} else {
215 		/* no address to convey if the pfn instance is disabled */
216 		rc = -ENXIO;
217 	}
218 	device_unlock(dev);
219 
220 	return rc;
221 }
222 static DEVICE_ATTR_ADMIN_RO(resource);
223 
224 static ssize_t size_show(struct device *dev,
225 		struct device_attribute *attr, char *buf)
226 {
227 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
228 	ssize_t rc;
229 
230 	device_lock(dev);
231 	if (dev->driver) {
232 		struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
233 		u64 offset = __le64_to_cpu(pfn_sb->dataoff);
234 		struct nd_namespace_common *ndns = nd_pfn->ndns;
235 		u32 start_pad = __le32_to_cpu(pfn_sb->start_pad);
236 		u32 end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
237 		struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
238 
239 		rc = sprintf(buf, "%llu\n", (unsigned long long)
240 				resource_size(&nsio->res) - start_pad
241 				- end_trunc - offset);
242 	} else {
243 		/* no size to convey if the pfn instance is disabled */
244 		rc = -ENXIO;
245 	}
246 	device_unlock(dev);
247 
248 	return rc;
249 }
250 static DEVICE_ATTR_RO(size);
251 
252 static ssize_t supported_alignments_show(struct device *dev,
253 		struct device_attribute *attr, char *buf)
254 {
255 	unsigned long aligns[MAX_NVDIMM_ALIGN] = { [0] = 0, };
256 
257 	return nd_size_select_show(0,
258 			nd_pfn_supported_alignments(aligns), buf);
259 }
260 static DEVICE_ATTR_RO(supported_alignments);
261 
262 static struct attribute *nd_pfn_attributes[] = {
263 	&dev_attr_mode.attr,
264 	&dev_attr_namespace.attr,
265 	&dev_attr_uuid.attr,
266 	&dev_attr_align.attr,
267 	&dev_attr_resource.attr,
268 	&dev_attr_size.attr,
269 	&dev_attr_supported_alignments.attr,
270 	NULL,
271 };
272 
273 static struct attribute_group nd_pfn_attribute_group = {
274 	.attrs = nd_pfn_attributes,
275 };
276 
277 const struct attribute_group *nd_pfn_attribute_groups[] = {
278 	&nd_pfn_attribute_group,
279 	&nd_device_attribute_group,
280 	&nd_numa_attribute_group,
281 	NULL,
282 };
283 
284 static const struct device_type nd_pfn_device_type = {
285 	.name = "nd_pfn",
286 	.release = nd_pfn_release,
287 	.groups = nd_pfn_attribute_groups,
288 };
289 
290 bool is_nd_pfn(struct device *dev)
291 {
292 	return dev ? dev->type == &nd_pfn_device_type : false;
293 }
294 EXPORT_SYMBOL(is_nd_pfn);
295 
296 static struct lock_class_key nvdimm_pfn_key;
297 
298 struct device *nd_pfn_devinit(struct nd_pfn *nd_pfn,
299 		struct nd_namespace_common *ndns)
300 {
301 	struct device *dev;
302 
303 	if (!nd_pfn)
304 		return NULL;
305 
306 	nd_pfn->mode = PFN_MODE_NONE;
307 	nd_pfn->align = nd_pfn_default_alignment();
308 	dev = &nd_pfn->dev;
309 	device_initialize(&nd_pfn->dev);
310 	lockdep_set_class(&nd_pfn->dev.mutex, &nvdimm_pfn_key);
311 	if (ndns && !__nd_attach_ndns(&nd_pfn->dev, ndns, &nd_pfn->ndns)) {
312 		dev_dbg(&ndns->dev, "failed, already claimed by %s\n",
313 				dev_name(ndns->claim));
314 		put_device(dev);
315 		return NULL;
316 	}
317 	return dev;
318 }
319 
320 static struct nd_pfn *nd_pfn_alloc(struct nd_region *nd_region)
321 {
322 	struct nd_pfn *nd_pfn;
323 	struct device *dev;
324 
325 	nd_pfn = kzalloc(sizeof(*nd_pfn), GFP_KERNEL);
326 	if (!nd_pfn)
327 		return NULL;
328 
329 	nd_pfn->id = ida_simple_get(&nd_region->pfn_ida, 0, 0, GFP_KERNEL);
330 	if (nd_pfn->id < 0) {
331 		kfree(nd_pfn);
332 		return NULL;
333 	}
334 
335 	dev = &nd_pfn->dev;
336 	dev_set_name(dev, "pfn%d.%d", nd_region->id, nd_pfn->id);
337 	dev->type = &nd_pfn_device_type;
338 	dev->parent = &nd_region->dev;
339 
340 	return nd_pfn;
341 }
342 
343 struct device *nd_pfn_create(struct nd_region *nd_region)
344 {
345 	struct nd_pfn *nd_pfn;
346 	struct device *dev;
347 
348 	if (!is_memory(&nd_region->dev))
349 		return NULL;
350 
351 	nd_pfn = nd_pfn_alloc(nd_region);
352 	dev = nd_pfn_devinit(nd_pfn, NULL);
353 
354 	nd_device_register(dev);
355 	return dev;
356 }
357 
358 /*
359  * nd_pfn_clear_memmap_errors() clears any errors in the volatile memmap
360  * space associated with the namespace. If the memmap is set to DRAM, then
361  * this is a no-op. Since the memmap area is freshly initialized during
362  * probe, we have an opportunity to clear any badblocks in this area.
363  */
364 static int nd_pfn_clear_memmap_errors(struct nd_pfn *nd_pfn)
365 {
366 	struct nd_region *nd_region = to_nd_region(nd_pfn->dev.parent);
367 	struct nd_namespace_common *ndns = nd_pfn->ndns;
368 	void *zero_page = page_address(ZERO_PAGE(0));
369 	struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
370 	int num_bad, meta_num, rc, bb_present;
371 	sector_t first_bad, meta_start;
372 	struct nd_namespace_io *nsio;
373 
374 	if (nd_pfn->mode != PFN_MODE_PMEM)
375 		return 0;
376 
377 	nsio = to_nd_namespace_io(&ndns->dev);
378 	meta_start = (SZ_4K + sizeof(*pfn_sb)) >> 9;
379 	meta_num = (le64_to_cpu(pfn_sb->dataoff) >> 9) - meta_start;
380 
381 	/*
382 	 * re-enable the namespace with correct size so that we can access
383 	 * the device memmap area.
384 	 */
385 	devm_namespace_disable(&nd_pfn->dev, ndns);
386 	rc = devm_namespace_enable(&nd_pfn->dev, ndns, le64_to_cpu(pfn_sb->dataoff));
387 	if (rc)
388 		return rc;
389 
390 	do {
391 		unsigned long zero_len;
392 		u64 nsoff;
393 
394 		bb_present = badblocks_check(&nd_region->bb, meta_start,
395 				meta_num, &first_bad, &num_bad);
396 		if (bb_present) {
397 			dev_dbg(&nd_pfn->dev, "meta: %x badblocks at %llx\n",
398 					num_bad, first_bad);
399 			nsoff = ALIGN_DOWN((nd_region->ndr_start
400 					+ (first_bad << 9)) - nsio->res.start,
401 					PAGE_SIZE);
402 			zero_len = ALIGN(num_bad << 9, PAGE_SIZE);
403 			while (zero_len) {
404 				unsigned long chunk = min(zero_len, PAGE_SIZE);
405 
406 				rc = nvdimm_write_bytes(ndns, nsoff, zero_page,
407 							chunk, 0);
408 				if (rc)
409 					break;
410 
411 				zero_len -= chunk;
412 				nsoff += chunk;
413 			}
414 			if (rc) {
415 				dev_err(&nd_pfn->dev,
416 					"error clearing %x badblocks at %llx\n",
417 					num_bad, first_bad);
418 				return rc;
419 			}
420 		}
421 	} while (bb_present);
422 
423 	return 0;
424 }
425 
426 static bool nd_supported_alignment(unsigned long align)
427 {
428 	int i;
429 	unsigned long supported[MAX_NVDIMM_ALIGN] = { [0] = 0, };
430 
431 	if (align == 0)
432 		return false;
433 
434 	nd_pfn_supported_alignments(supported);
435 	for (i = 0; supported[i]; i++)
436 		if (align == supported[i])
437 			return true;
438 	return false;
439 }
440 
441 /**
442  * nd_pfn_validate - read and validate info-block
443  * @nd_pfn: fsdax namespace runtime state / properties
444  * @sig: 'devdax' or 'fsdax' signature
445  *
446  * Upon return the info-block buffer contents (->pfn_sb) are
447  * indeterminate when validation fails, and a coherent info-block
448  * otherwise.
449  */
450 int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig)
451 {
452 	u64 checksum, offset;
453 	struct resource *res;
454 	enum nd_pfn_mode mode;
455 	resource_size_t res_size;
456 	struct nd_namespace_io *nsio;
457 	unsigned long align, start_pad, end_trunc;
458 	struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
459 	struct nd_namespace_common *ndns = nd_pfn->ndns;
460 	const uuid_t *parent_uuid = nd_dev_to_uuid(&ndns->dev);
461 
462 	if (!pfn_sb || !ndns)
463 		return -ENODEV;
464 
465 	if (!is_memory(nd_pfn->dev.parent))
466 		return -ENODEV;
467 
468 	if (nvdimm_read_bytes(ndns, SZ_4K, pfn_sb, sizeof(*pfn_sb), 0))
469 		return -ENXIO;
470 
471 	if (memcmp(pfn_sb->signature, sig, PFN_SIG_LEN) != 0)
472 		return -ENODEV;
473 
474 	checksum = le64_to_cpu(pfn_sb->checksum);
475 	pfn_sb->checksum = 0;
476 	if (checksum != nd_sb_checksum((struct nd_gen_sb *) pfn_sb))
477 		return -ENODEV;
478 	pfn_sb->checksum = cpu_to_le64(checksum);
479 
480 	if (memcmp(pfn_sb->parent_uuid, parent_uuid, 16) != 0)
481 		return -ENODEV;
482 
483 	if (__le16_to_cpu(pfn_sb->version_minor) < 1) {
484 		pfn_sb->start_pad = 0;
485 		pfn_sb->end_trunc = 0;
486 	}
487 
488 	if (__le16_to_cpu(pfn_sb->version_minor) < 2)
489 		pfn_sb->align = 0;
490 
491 	if (__le16_to_cpu(pfn_sb->version_minor) < 4) {
492 		pfn_sb->page_struct_size = cpu_to_le16(64);
493 		pfn_sb->page_size = cpu_to_le32(PAGE_SIZE);
494 	}
495 
496 	switch (le32_to_cpu(pfn_sb->mode)) {
497 	case PFN_MODE_RAM:
498 	case PFN_MODE_PMEM:
499 		break;
500 	default:
501 		return -ENXIO;
502 	}
503 
504 	align = le32_to_cpu(pfn_sb->align);
505 	offset = le64_to_cpu(pfn_sb->dataoff);
506 	start_pad = le32_to_cpu(pfn_sb->start_pad);
507 	end_trunc = le32_to_cpu(pfn_sb->end_trunc);
508 	if (align == 0)
509 		align = 1UL << ilog2(offset);
510 	mode = le32_to_cpu(pfn_sb->mode);
511 
512 	if ((le32_to_cpu(pfn_sb->page_size) > PAGE_SIZE) &&
513 			(mode == PFN_MODE_PMEM)) {
514 		dev_err(&nd_pfn->dev,
515 				"init failed, page size mismatch %d\n",
516 				le32_to_cpu(pfn_sb->page_size));
517 		return -EOPNOTSUPP;
518 	}
519 
520 	if ((le16_to_cpu(pfn_sb->page_struct_size) < sizeof(struct page)) &&
521 			(mode == PFN_MODE_PMEM)) {
522 		dev_err(&nd_pfn->dev,
523 				"init failed, struct page size mismatch %d\n",
524 				le16_to_cpu(pfn_sb->page_struct_size));
525 		return -EOPNOTSUPP;
526 	}
527 
528 	/*
529 	 * Check whether the we support the alignment. For Dax if the
530 	 * superblock alignment is not matching, we won't initialize
531 	 * the device.
532 	 */
533 	if (!nd_supported_alignment(align) &&
534 			!memcmp(pfn_sb->signature, DAX_SIG, PFN_SIG_LEN)) {
535 		dev_err(&nd_pfn->dev, "init failed, alignment mismatch: "
536 				"%ld:%ld\n", nd_pfn->align, align);
537 		return -EOPNOTSUPP;
538 	}
539 
540 	if (!nd_pfn->uuid) {
541 		/*
542 		 * When probing a namepace via nd_pfn_probe() the uuid
543 		 * is NULL (see: nd_pfn_devinit()) we init settings from
544 		 * pfn_sb
545 		 */
546 		nd_pfn->uuid = kmemdup(pfn_sb->uuid, 16, GFP_KERNEL);
547 		if (!nd_pfn->uuid)
548 			return -ENOMEM;
549 		nd_pfn->align = align;
550 		nd_pfn->mode = mode;
551 	} else {
552 		/*
553 		 * When probing a pfn / dax instance we validate the
554 		 * live settings against the pfn_sb
555 		 */
556 		if (memcmp(nd_pfn->uuid, pfn_sb->uuid, 16) != 0)
557 			return -ENODEV;
558 
559 		/*
560 		 * If the uuid validates, but other settings mismatch
561 		 * return EINVAL because userspace has managed to change
562 		 * the configuration without specifying new
563 		 * identification.
564 		 */
565 		if (nd_pfn->align != align || nd_pfn->mode != mode) {
566 			dev_err(&nd_pfn->dev,
567 					"init failed, settings mismatch\n");
568 			dev_dbg(&nd_pfn->dev, "align: %lx:%lx mode: %d:%d\n",
569 					nd_pfn->align, align, nd_pfn->mode,
570 					mode);
571 			return -EOPNOTSUPP;
572 		}
573 	}
574 
575 	if (align > nvdimm_namespace_capacity(ndns)) {
576 		dev_err(&nd_pfn->dev, "alignment: %lx exceeds capacity %llx\n",
577 				align, nvdimm_namespace_capacity(ndns));
578 		return -EOPNOTSUPP;
579 	}
580 
581 	/*
582 	 * These warnings are verbose because they can only trigger in
583 	 * the case where the physical address alignment of the
584 	 * namespace has changed since the pfn superblock was
585 	 * established.
586 	 */
587 	nsio = to_nd_namespace_io(&ndns->dev);
588 	res = &nsio->res;
589 	res_size = resource_size(res);
590 	if (offset >= res_size) {
591 		dev_err(&nd_pfn->dev, "pfn array size exceeds capacity of %s\n",
592 				dev_name(&ndns->dev));
593 		return -EOPNOTSUPP;
594 	}
595 
596 	if ((align && !IS_ALIGNED(res->start + offset + start_pad, align))
597 			|| !IS_ALIGNED(offset, PAGE_SIZE)) {
598 		dev_err(&nd_pfn->dev,
599 				"bad offset: %#llx dax disabled align: %#lx\n",
600 				offset, align);
601 		return -EOPNOTSUPP;
602 	}
603 
604 	if (!IS_ALIGNED(res->start + start_pad, memremap_compat_align())) {
605 		dev_err(&nd_pfn->dev, "resource start misaligned\n");
606 		return -EOPNOTSUPP;
607 	}
608 
609 	if (!IS_ALIGNED(res->end + 1 - end_trunc, memremap_compat_align())) {
610 		dev_err(&nd_pfn->dev, "resource end misaligned\n");
611 		return -EOPNOTSUPP;
612 	}
613 
614 	if (offset >= (res_size - start_pad - end_trunc)) {
615 		dev_err(&nd_pfn->dev, "bad offset with small namespace\n");
616 		return -EOPNOTSUPP;
617 	}
618 	return 0;
619 }
620 EXPORT_SYMBOL(nd_pfn_validate);
621 
622 int nd_pfn_probe(struct device *dev, struct nd_namespace_common *ndns)
623 {
624 	int rc;
625 	struct nd_pfn *nd_pfn;
626 	struct device *pfn_dev;
627 	struct nd_pfn_sb *pfn_sb;
628 	struct nd_region *nd_region = to_nd_region(ndns->dev.parent);
629 
630 	if (ndns->force_raw)
631 		return -ENODEV;
632 
633 	switch (ndns->claim_class) {
634 	case NVDIMM_CCLASS_NONE:
635 	case NVDIMM_CCLASS_PFN:
636 		break;
637 	default:
638 		return -ENODEV;
639 	}
640 
641 	nvdimm_bus_lock(&ndns->dev);
642 	nd_pfn = nd_pfn_alloc(nd_region);
643 	pfn_dev = nd_pfn_devinit(nd_pfn, ndns);
644 	nvdimm_bus_unlock(&ndns->dev);
645 	if (!pfn_dev)
646 		return -ENOMEM;
647 	pfn_sb = devm_kmalloc(dev, sizeof(*pfn_sb), GFP_KERNEL);
648 	nd_pfn = to_nd_pfn(pfn_dev);
649 	nd_pfn->pfn_sb = pfn_sb;
650 	rc = nd_pfn_validate(nd_pfn, PFN_SIG);
651 	dev_dbg(dev, "pfn: %s\n", rc == 0 ? dev_name(pfn_dev) : "<none>");
652 	if (rc < 0) {
653 		nd_detach_ndns(pfn_dev, &nd_pfn->ndns);
654 		put_device(pfn_dev);
655 	} else
656 		nd_device_register(pfn_dev);
657 
658 	return rc;
659 }
660 EXPORT_SYMBOL(nd_pfn_probe);
661 
662 /*
663  * We hotplug memory at sub-section granularity, pad the reserved area
664  * from the previous section base to the namespace base address.
665  */
666 static unsigned long init_altmap_base(resource_size_t base)
667 {
668 	unsigned long base_pfn = PHYS_PFN(base);
669 
670 	return SUBSECTION_ALIGN_DOWN(base_pfn);
671 }
672 
673 static unsigned long init_altmap_reserve(resource_size_t base)
674 {
675 	unsigned long reserve = nd_info_block_reserve() >> PAGE_SHIFT;
676 	unsigned long base_pfn = PHYS_PFN(base);
677 
678 	reserve += base_pfn - SUBSECTION_ALIGN_DOWN(base_pfn);
679 	return reserve;
680 }
681 
682 static int __nvdimm_setup_pfn(struct nd_pfn *nd_pfn, struct dev_pagemap *pgmap)
683 {
684 	struct range *range = &pgmap->range;
685 	struct vmem_altmap *altmap = &pgmap->altmap;
686 	struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
687 	u64 offset = le64_to_cpu(pfn_sb->dataoff);
688 	u32 start_pad = __le32_to_cpu(pfn_sb->start_pad);
689 	u32 end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
690 	u32 reserve = nd_info_block_reserve();
691 	struct nd_namespace_common *ndns = nd_pfn->ndns;
692 	struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
693 	resource_size_t base = nsio->res.start + start_pad;
694 	resource_size_t end = nsio->res.end - end_trunc;
695 	struct vmem_altmap __altmap = {
696 		.base_pfn = init_altmap_base(base),
697 		.reserve = init_altmap_reserve(base),
698 		.end_pfn = PHYS_PFN(end),
699 	};
700 
701 	*range = (struct range) {
702 		.start = nsio->res.start + start_pad,
703 		.end = nsio->res.end - end_trunc,
704 	};
705 	pgmap->nr_range = 1;
706 	if (nd_pfn->mode == PFN_MODE_RAM) {
707 		if (offset < reserve)
708 			return -EINVAL;
709 		nd_pfn->npfns = le64_to_cpu(pfn_sb->npfns);
710 	} else if (nd_pfn->mode == PFN_MODE_PMEM) {
711 		nd_pfn->npfns = PHYS_PFN((range_len(range) - offset));
712 		if (le64_to_cpu(nd_pfn->pfn_sb->npfns) > nd_pfn->npfns)
713 			dev_info(&nd_pfn->dev,
714 					"number of pfns truncated from %lld to %ld\n",
715 					le64_to_cpu(nd_pfn->pfn_sb->npfns),
716 					nd_pfn->npfns);
717 		memcpy(altmap, &__altmap, sizeof(*altmap));
718 		altmap->free = PHYS_PFN(offset - reserve);
719 		altmap->alloc = 0;
720 		pgmap->flags |= PGMAP_ALTMAP_VALID;
721 	} else
722 		return -ENXIO;
723 
724 	return 0;
725 }
726 
727 static int nd_pfn_init(struct nd_pfn *nd_pfn)
728 {
729 	struct nd_namespace_common *ndns = nd_pfn->ndns;
730 	struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
731 	resource_size_t start, size;
732 	struct nd_region *nd_region;
733 	unsigned long npfns, align;
734 	u32 end_trunc;
735 	struct nd_pfn_sb *pfn_sb;
736 	phys_addr_t offset;
737 	const char *sig;
738 	u64 checksum;
739 	int rc;
740 
741 	pfn_sb = devm_kmalloc(&nd_pfn->dev, sizeof(*pfn_sb), GFP_KERNEL);
742 	if (!pfn_sb)
743 		return -ENOMEM;
744 
745 	nd_pfn->pfn_sb = pfn_sb;
746 	if (is_nd_dax(&nd_pfn->dev))
747 		sig = DAX_SIG;
748 	else
749 		sig = PFN_SIG;
750 
751 	rc = nd_pfn_validate(nd_pfn, sig);
752 	if (rc == 0)
753 		return nd_pfn_clear_memmap_errors(nd_pfn);
754 	if (rc != -ENODEV)
755 		return rc;
756 
757 	/* no info block, do init */;
758 	memset(pfn_sb, 0, sizeof(*pfn_sb));
759 
760 	nd_region = to_nd_region(nd_pfn->dev.parent);
761 	if (nd_region->ro) {
762 		dev_info(&nd_pfn->dev,
763 				"%s is read-only, unable to init metadata\n",
764 				dev_name(&nd_region->dev));
765 		return -ENXIO;
766 	}
767 
768 	start = nsio->res.start;
769 	size = resource_size(&nsio->res);
770 	npfns = PHYS_PFN(size - SZ_8K);
771 	align = max(nd_pfn->align, memremap_compat_align());
772 
773 	/*
774 	 * When @start is misaligned fail namespace creation. See
775 	 * the 'struct nd_pfn_sb' commentary on why ->start_pad is not
776 	 * an option.
777 	 */
778 	if (!IS_ALIGNED(start, memremap_compat_align())) {
779 		dev_err(&nd_pfn->dev, "%s: start %pa misaligned to %#lx\n",
780 				dev_name(&ndns->dev), &start,
781 				memremap_compat_align());
782 		return -EINVAL;
783 	}
784 	end_trunc = start + size - ALIGN_DOWN(start + size, align);
785 	if (nd_pfn->mode == PFN_MODE_PMEM) {
786 		unsigned long page_map_size = MAX_STRUCT_PAGE_SIZE * npfns;
787 
788 		/*
789 		 * The altmap should be padded out to the block size used
790 		 * when populating the vmemmap. This *should* be equal to
791 		 * PMD_SIZE for most architectures.
792 		 *
793 		 * Also make sure size of struct page is less than
794 		 * MAX_STRUCT_PAGE_SIZE. The goal here is compatibility in the
795 		 * face of production kernel configurations that reduce the
796 		 * 'struct page' size below MAX_STRUCT_PAGE_SIZE. For debug
797 		 * kernel configurations that increase the 'struct page' size
798 		 * above MAX_STRUCT_PAGE_SIZE, the page_struct_override allows
799 		 * for continuing with the capacity that will be wasted when
800 		 * reverting to a production kernel configuration. Otherwise,
801 		 * those configurations are blocked by default.
802 		 */
803 		if (sizeof(struct page) > MAX_STRUCT_PAGE_SIZE) {
804 			if (page_struct_override)
805 				page_map_size = sizeof(struct page) * npfns;
806 			else {
807 				dev_err(&nd_pfn->dev,
808 					"Memory debug options prevent using pmem for the page map\n");
809 				return -EINVAL;
810 			}
811 		}
812 		offset = ALIGN(start + SZ_8K + page_map_size, align) - start;
813 	} else if (nd_pfn->mode == PFN_MODE_RAM)
814 		offset = ALIGN(start + SZ_8K, align) - start;
815 	else
816 		return -ENXIO;
817 
818 	if (offset >= (size - end_trunc)) {
819 		/* This results in zero size devices */
820 		dev_err(&nd_pfn->dev, "%s unable to satisfy requested alignment\n",
821 				dev_name(&ndns->dev));
822 		return -ENXIO;
823 	}
824 
825 	npfns = PHYS_PFN(size - offset - end_trunc);
826 	pfn_sb->mode = cpu_to_le32(nd_pfn->mode);
827 	pfn_sb->dataoff = cpu_to_le64(offset);
828 	pfn_sb->npfns = cpu_to_le64(npfns);
829 	memcpy(pfn_sb->signature, sig, PFN_SIG_LEN);
830 	memcpy(pfn_sb->uuid, nd_pfn->uuid, 16);
831 	memcpy(pfn_sb->parent_uuid, nd_dev_to_uuid(&ndns->dev), 16);
832 	pfn_sb->version_major = cpu_to_le16(1);
833 	pfn_sb->version_minor = cpu_to_le16(4);
834 	pfn_sb->end_trunc = cpu_to_le32(end_trunc);
835 	pfn_sb->align = cpu_to_le32(nd_pfn->align);
836 	if (sizeof(struct page) > MAX_STRUCT_PAGE_SIZE && page_struct_override)
837 		pfn_sb->page_struct_size = cpu_to_le16(sizeof(struct page));
838 	else
839 		pfn_sb->page_struct_size = cpu_to_le16(MAX_STRUCT_PAGE_SIZE);
840 	pfn_sb->page_size = cpu_to_le32(PAGE_SIZE);
841 	checksum = nd_sb_checksum((struct nd_gen_sb *) pfn_sb);
842 	pfn_sb->checksum = cpu_to_le64(checksum);
843 
844 	rc = nd_pfn_clear_memmap_errors(nd_pfn);
845 	if (rc)
846 		return rc;
847 
848 	return nvdimm_write_bytes(ndns, SZ_4K, pfn_sb, sizeof(*pfn_sb), 0);
849 }
850 
851 /*
852  * Determine the effective resource range and vmem_altmap from an nd_pfn
853  * instance.
854  */
855 int nvdimm_setup_pfn(struct nd_pfn *nd_pfn, struct dev_pagemap *pgmap)
856 {
857 	int rc;
858 
859 	if (!nd_pfn->uuid || !nd_pfn->ndns)
860 		return -ENODEV;
861 
862 	rc = nd_pfn_init(nd_pfn);
863 	if (rc)
864 		return rc;
865 
866 	/* we need a valid pfn_sb before we can init a dev_pagemap */
867 	return __nvdimm_setup_pfn(nd_pfn, pgmap);
868 }
869 EXPORT_SYMBOL_GPL(nvdimm_setup_pfn);
870