xref: /openbmc/linux/drivers/nvdimm/pfn_devs.c (revision 6de298ff)
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 (IS_ENABLED(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD))
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 	struct nd_namespace_io *nsio;
456 	unsigned long align, start_pad;
457 	struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
458 	struct nd_namespace_common *ndns = nd_pfn->ndns;
459 	const uuid_t *parent_uuid = nd_dev_to_uuid(&ndns->dev);
460 
461 	if (!pfn_sb || !ndns)
462 		return -ENODEV;
463 
464 	if (!is_memory(nd_pfn->dev.parent))
465 		return -ENODEV;
466 
467 	if (nvdimm_read_bytes(ndns, SZ_4K, pfn_sb, sizeof(*pfn_sb), 0))
468 		return -ENXIO;
469 
470 	if (memcmp(pfn_sb->signature, sig, PFN_SIG_LEN) != 0)
471 		return -ENODEV;
472 
473 	checksum = le64_to_cpu(pfn_sb->checksum);
474 	pfn_sb->checksum = 0;
475 	if (checksum != nd_sb_checksum((struct nd_gen_sb *) pfn_sb))
476 		return -ENODEV;
477 	pfn_sb->checksum = cpu_to_le64(checksum);
478 
479 	if (memcmp(pfn_sb->parent_uuid, parent_uuid, 16) != 0)
480 		return -ENODEV;
481 
482 	if (__le16_to_cpu(pfn_sb->version_minor) < 1) {
483 		pfn_sb->start_pad = 0;
484 		pfn_sb->end_trunc = 0;
485 	}
486 
487 	if (__le16_to_cpu(pfn_sb->version_minor) < 2)
488 		pfn_sb->align = 0;
489 
490 	if (__le16_to_cpu(pfn_sb->version_minor) < 4) {
491 		pfn_sb->page_struct_size = cpu_to_le16(64);
492 		pfn_sb->page_size = cpu_to_le32(PAGE_SIZE);
493 	}
494 
495 	switch (le32_to_cpu(pfn_sb->mode)) {
496 	case PFN_MODE_RAM:
497 	case PFN_MODE_PMEM:
498 		break;
499 	default:
500 		return -ENXIO;
501 	}
502 
503 	align = le32_to_cpu(pfn_sb->align);
504 	offset = le64_to_cpu(pfn_sb->dataoff);
505 	start_pad = le32_to_cpu(pfn_sb->start_pad);
506 	if (align == 0)
507 		align = 1UL << ilog2(offset);
508 	mode = le32_to_cpu(pfn_sb->mode);
509 
510 	if ((le32_to_cpu(pfn_sb->page_size) > PAGE_SIZE) &&
511 			(mode == PFN_MODE_PMEM)) {
512 		dev_err(&nd_pfn->dev,
513 				"init failed, page size mismatch %d\n",
514 				le32_to_cpu(pfn_sb->page_size));
515 		return -EOPNOTSUPP;
516 	}
517 
518 	if ((le16_to_cpu(pfn_sb->page_struct_size) < sizeof(struct page)) &&
519 			(mode == PFN_MODE_PMEM)) {
520 		dev_err(&nd_pfn->dev,
521 				"init failed, struct page size mismatch %d\n",
522 				le16_to_cpu(pfn_sb->page_struct_size));
523 		return -EOPNOTSUPP;
524 	}
525 
526 	/*
527 	 * Check whether the we support the alignment. For Dax if the
528 	 * superblock alignment is not matching, we won't initialize
529 	 * the device.
530 	 */
531 	if (!nd_supported_alignment(align) &&
532 			!memcmp(pfn_sb->signature, DAX_SIG, PFN_SIG_LEN)) {
533 		dev_err(&nd_pfn->dev, "init failed, alignment mismatch: "
534 				"%ld:%ld\n", nd_pfn->align, align);
535 		return -EOPNOTSUPP;
536 	}
537 
538 	if (!nd_pfn->uuid) {
539 		/*
540 		 * When probing a namepace via nd_pfn_probe() the uuid
541 		 * is NULL (see: nd_pfn_devinit()) we init settings from
542 		 * pfn_sb
543 		 */
544 		nd_pfn->uuid = kmemdup(pfn_sb->uuid, 16, GFP_KERNEL);
545 		if (!nd_pfn->uuid)
546 			return -ENOMEM;
547 		nd_pfn->align = align;
548 		nd_pfn->mode = mode;
549 	} else {
550 		/*
551 		 * When probing a pfn / dax instance we validate the
552 		 * live settings against the pfn_sb
553 		 */
554 		if (memcmp(nd_pfn->uuid, pfn_sb->uuid, 16) != 0)
555 			return -ENODEV;
556 
557 		/*
558 		 * If the uuid validates, but other settings mismatch
559 		 * return EINVAL because userspace has managed to change
560 		 * the configuration without specifying new
561 		 * identification.
562 		 */
563 		if (nd_pfn->align != align || nd_pfn->mode != mode) {
564 			dev_err(&nd_pfn->dev,
565 					"init failed, settings mismatch\n");
566 			dev_dbg(&nd_pfn->dev, "align: %lx:%lx mode: %d:%d\n",
567 					nd_pfn->align, align, nd_pfn->mode,
568 					mode);
569 			return -EOPNOTSUPP;
570 		}
571 	}
572 
573 	if (align > nvdimm_namespace_capacity(ndns)) {
574 		dev_err(&nd_pfn->dev, "alignment: %lx exceeds capacity %llx\n",
575 				align, nvdimm_namespace_capacity(ndns));
576 		return -EOPNOTSUPP;
577 	}
578 
579 	/*
580 	 * These warnings are verbose because they can only trigger in
581 	 * the case where the physical address alignment of the
582 	 * namespace has changed since the pfn superblock was
583 	 * established.
584 	 */
585 	nsio = to_nd_namespace_io(&ndns->dev);
586 	res = &nsio->res;
587 	if (offset >= resource_size(res)) {
588 		dev_err(&nd_pfn->dev, "pfn array size exceeds capacity of %s\n",
589 				dev_name(&ndns->dev));
590 		return -EOPNOTSUPP;
591 	}
592 
593 	if ((align && !IS_ALIGNED(res->start + offset + start_pad, align))
594 			|| !IS_ALIGNED(offset, PAGE_SIZE)) {
595 		dev_err(&nd_pfn->dev,
596 				"bad offset: %#llx dax disabled align: %#lx\n",
597 				offset, align);
598 		return -EOPNOTSUPP;
599 	}
600 
601 	if (!IS_ALIGNED(res->start + le32_to_cpu(pfn_sb->start_pad),
602 				memremap_compat_align())) {
603 		dev_err(&nd_pfn->dev, "resource start misaligned\n");
604 		return -EOPNOTSUPP;
605 	}
606 
607 	if (!IS_ALIGNED(res->end + 1 - le32_to_cpu(pfn_sb->end_trunc),
608 				memremap_compat_align())) {
609 		dev_err(&nd_pfn->dev, "resource end misaligned\n");
610 		return -EOPNOTSUPP;
611 	}
612 
613 	return 0;
614 }
615 EXPORT_SYMBOL(nd_pfn_validate);
616 
617 int nd_pfn_probe(struct device *dev, struct nd_namespace_common *ndns)
618 {
619 	int rc;
620 	struct nd_pfn *nd_pfn;
621 	struct device *pfn_dev;
622 	struct nd_pfn_sb *pfn_sb;
623 	struct nd_region *nd_region = to_nd_region(ndns->dev.parent);
624 
625 	if (ndns->force_raw)
626 		return -ENODEV;
627 
628 	switch (ndns->claim_class) {
629 	case NVDIMM_CCLASS_NONE:
630 	case NVDIMM_CCLASS_PFN:
631 		break;
632 	default:
633 		return -ENODEV;
634 	}
635 
636 	nvdimm_bus_lock(&ndns->dev);
637 	nd_pfn = nd_pfn_alloc(nd_region);
638 	pfn_dev = nd_pfn_devinit(nd_pfn, ndns);
639 	nvdimm_bus_unlock(&ndns->dev);
640 	if (!pfn_dev)
641 		return -ENOMEM;
642 	pfn_sb = devm_kmalloc(dev, sizeof(*pfn_sb), GFP_KERNEL);
643 	nd_pfn = to_nd_pfn(pfn_dev);
644 	nd_pfn->pfn_sb = pfn_sb;
645 	rc = nd_pfn_validate(nd_pfn, PFN_SIG);
646 	dev_dbg(dev, "pfn: %s\n", rc == 0 ? dev_name(pfn_dev) : "<none>");
647 	if (rc < 0) {
648 		nd_detach_ndns(pfn_dev, &nd_pfn->ndns);
649 		put_device(pfn_dev);
650 	} else
651 		nd_device_register(pfn_dev);
652 
653 	return rc;
654 }
655 EXPORT_SYMBOL(nd_pfn_probe);
656 
657 /*
658  * We hotplug memory at sub-section granularity, pad the reserved area
659  * from the previous section base to the namespace base address.
660  */
661 static unsigned long init_altmap_base(resource_size_t base)
662 {
663 	unsigned long base_pfn = PHYS_PFN(base);
664 
665 	return SUBSECTION_ALIGN_DOWN(base_pfn);
666 }
667 
668 static unsigned long init_altmap_reserve(resource_size_t base)
669 {
670 	unsigned long reserve = nd_info_block_reserve() >> PAGE_SHIFT;
671 	unsigned long base_pfn = PHYS_PFN(base);
672 
673 	reserve += base_pfn - SUBSECTION_ALIGN_DOWN(base_pfn);
674 	return reserve;
675 }
676 
677 static int __nvdimm_setup_pfn(struct nd_pfn *nd_pfn, struct dev_pagemap *pgmap)
678 {
679 	struct range *range = &pgmap->range;
680 	struct vmem_altmap *altmap = &pgmap->altmap;
681 	struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
682 	u64 offset = le64_to_cpu(pfn_sb->dataoff);
683 	u32 start_pad = __le32_to_cpu(pfn_sb->start_pad);
684 	u32 end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
685 	u32 reserve = nd_info_block_reserve();
686 	struct nd_namespace_common *ndns = nd_pfn->ndns;
687 	struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
688 	resource_size_t base = nsio->res.start + start_pad;
689 	resource_size_t end = nsio->res.end - end_trunc;
690 	struct vmem_altmap __altmap = {
691 		.base_pfn = init_altmap_base(base),
692 		.reserve = init_altmap_reserve(base),
693 		.end_pfn = PHYS_PFN(end),
694 	};
695 
696 	*range = (struct range) {
697 		.start = nsio->res.start + start_pad,
698 		.end = nsio->res.end - end_trunc,
699 	};
700 	pgmap->nr_range = 1;
701 	if (nd_pfn->mode == PFN_MODE_RAM) {
702 		if (offset < reserve)
703 			return -EINVAL;
704 		nd_pfn->npfns = le64_to_cpu(pfn_sb->npfns);
705 	} else if (nd_pfn->mode == PFN_MODE_PMEM) {
706 		nd_pfn->npfns = PHYS_PFN((range_len(range) - offset));
707 		if (le64_to_cpu(nd_pfn->pfn_sb->npfns) > nd_pfn->npfns)
708 			dev_info(&nd_pfn->dev,
709 					"number of pfns truncated from %lld to %ld\n",
710 					le64_to_cpu(nd_pfn->pfn_sb->npfns),
711 					nd_pfn->npfns);
712 		memcpy(altmap, &__altmap, sizeof(*altmap));
713 		altmap->free = PHYS_PFN(offset - reserve);
714 		altmap->alloc = 0;
715 		pgmap->flags |= PGMAP_ALTMAP_VALID;
716 	} else
717 		return -ENXIO;
718 
719 	return 0;
720 }
721 
722 static int nd_pfn_init(struct nd_pfn *nd_pfn)
723 {
724 	struct nd_namespace_common *ndns = nd_pfn->ndns;
725 	struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
726 	resource_size_t start, size;
727 	struct nd_region *nd_region;
728 	unsigned long npfns, align;
729 	u32 end_trunc;
730 	struct nd_pfn_sb *pfn_sb;
731 	phys_addr_t offset;
732 	const char *sig;
733 	u64 checksum;
734 	int rc;
735 
736 	pfn_sb = devm_kmalloc(&nd_pfn->dev, sizeof(*pfn_sb), GFP_KERNEL);
737 	if (!pfn_sb)
738 		return -ENOMEM;
739 
740 	nd_pfn->pfn_sb = pfn_sb;
741 	if (is_nd_dax(&nd_pfn->dev))
742 		sig = DAX_SIG;
743 	else
744 		sig = PFN_SIG;
745 
746 	rc = nd_pfn_validate(nd_pfn, sig);
747 	if (rc == 0)
748 		return nd_pfn_clear_memmap_errors(nd_pfn);
749 	if (rc != -ENODEV)
750 		return rc;
751 
752 	/* no info block, do init */;
753 	memset(pfn_sb, 0, sizeof(*pfn_sb));
754 
755 	nd_region = to_nd_region(nd_pfn->dev.parent);
756 	if (nd_region->ro) {
757 		dev_info(&nd_pfn->dev,
758 				"%s is read-only, unable to init metadata\n",
759 				dev_name(&nd_region->dev));
760 		return -ENXIO;
761 	}
762 
763 	start = nsio->res.start;
764 	size = resource_size(&nsio->res);
765 	npfns = PHYS_PFN(size - SZ_8K);
766 	align = max(nd_pfn->align, memremap_compat_align());
767 
768 	/*
769 	 * When @start is misaligned fail namespace creation. See
770 	 * the 'struct nd_pfn_sb' commentary on why ->start_pad is not
771 	 * an option.
772 	 */
773 	if (!IS_ALIGNED(start, memremap_compat_align())) {
774 		dev_err(&nd_pfn->dev, "%s: start %pa misaligned to %#lx\n",
775 				dev_name(&ndns->dev), &start,
776 				memremap_compat_align());
777 		return -EINVAL;
778 	}
779 	end_trunc = start + size - ALIGN_DOWN(start + size, align);
780 	if (nd_pfn->mode == PFN_MODE_PMEM) {
781 		unsigned long page_map_size = MAX_STRUCT_PAGE_SIZE * npfns;
782 
783 		/*
784 		 * The altmap should be padded out to the block size used
785 		 * when populating the vmemmap. This *should* be equal to
786 		 * PMD_SIZE for most architectures.
787 		 *
788 		 * Also make sure size of struct page is less than
789 		 * MAX_STRUCT_PAGE_SIZE. The goal here is compatibility in the
790 		 * face of production kernel configurations that reduce the
791 		 * 'struct page' size below MAX_STRUCT_PAGE_SIZE. For debug
792 		 * kernel configurations that increase the 'struct page' size
793 		 * above MAX_STRUCT_PAGE_SIZE, the page_struct_override allows
794 		 * for continuing with the capacity that will be wasted when
795 		 * reverting to a production kernel configuration. Otherwise,
796 		 * those configurations are blocked by default.
797 		 */
798 		if (sizeof(struct page) > MAX_STRUCT_PAGE_SIZE) {
799 			if (page_struct_override)
800 				page_map_size = sizeof(struct page) * npfns;
801 			else {
802 				dev_err(&nd_pfn->dev,
803 					"Memory debug options prevent using pmem for the page map\n");
804 				return -EINVAL;
805 			}
806 		}
807 		offset = ALIGN(start + SZ_8K + page_map_size, align) - start;
808 	} else if (nd_pfn->mode == PFN_MODE_RAM)
809 		offset = ALIGN(start + SZ_8K, align) - start;
810 	else
811 		return -ENXIO;
812 
813 	if (offset >= size) {
814 		dev_err(&nd_pfn->dev, "%s unable to satisfy requested alignment\n",
815 				dev_name(&ndns->dev));
816 		return -ENXIO;
817 	}
818 
819 	npfns = PHYS_PFN(size - offset - end_trunc);
820 	pfn_sb->mode = cpu_to_le32(nd_pfn->mode);
821 	pfn_sb->dataoff = cpu_to_le64(offset);
822 	pfn_sb->npfns = cpu_to_le64(npfns);
823 	memcpy(pfn_sb->signature, sig, PFN_SIG_LEN);
824 	memcpy(pfn_sb->uuid, nd_pfn->uuid, 16);
825 	memcpy(pfn_sb->parent_uuid, nd_dev_to_uuid(&ndns->dev), 16);
826 	pfn_sb->version_major = cpu_to_le16(1);
827 	pfn_sb->version_minor = cpu_to_le16(4);
828 	pfn_sb->end_trunc = cpu_to_le32(end_trunc);
829 	pfn_sb->align = cpu_to_le32(nd_pfn->align);
830 	if (sizeof(struct page) > MAX_STRUCT_PAGE_SIZE && page_struct_override)
831 		pfn_sb->page_struct_size = cpu_to_le16(sizeof(struct page));
832 	else
833 		pfn_sb->page_struct_size = cpu_to_le16(MAX_STRUCT_PAGE_SIZE);
834 	pfn_sb->page_size = cpu_to_le32(PAGE_SIZE);
835 	checksum = nd_sb_checksum((struct nd_gen_sb *) pfn_sb);
836 	pfn_sb->checksum = cpu_to_le64(checksum);
837 
838 	rc = nd_pfn_clear_memmap_errors(nd_pfn);
839 	if (rc)
840 		return rc;
841 
842 	return nvdimm_write_bytes(ndns, SZ_4K, pfn_sb, sizeof(*pfn_sb), 0);
843 }
844 
845 /*
846  * Determine the effective resource range and vmem_altmap from an nd_pfn
847  * instance.
848  */
849 int nvdimm_setup_pfn(struct nd_pfn *nd_pfn, struct dev_pagemap *pgmap)
850 {
851 	int rc;
852 
853 	if (!nd_pfn->uuid || !nd_pfn->ndns)
854 		return -ENODEV;
855 
856 	rc = nd_pfn_init(nd_pfn);
857 	if (rc)
858 		return rc;
859 
860 	/* we need a valid pfn_sb before we can init a dev_pagemap */
861 	return __nvdimm_setup_pfn(nd_pfn, pgmap);
862 }
863 EXPORT_SYMBOL_GPL(nvdimm_setup_pfn);
864