xref: /openbmc/linux/drivers/nvdimm/pfn_devs.c (revision 2985bed6)
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/genhd.h>
9 #include <linux/sizes.h>
10 #include <linux/slab.h>
11 #include <linux/fs.h>
12 #include <linux/mm.h>
13 #include "nd-core.h"
14 #include "pfn.h"
15 #include "nd.h"
16 
17 static void nd_pfn_release(struct device *dev)
18 {
19 	struct nd_region *nd_region = to_nd_region(dev->parent);
20 	struct nd_pfn *nd_pfn = to_nd_pfn(dev);
21 
22 	dev_dbg(dev, "trace\n");
23 	nd_detach_ndns(&nd_pfn->dev, &nd_pfn->ndns);
24 	ida_simple_remove(&nd_region->pfn_ida, nd_pfn->id);
25 	kfree(nd_pfn->uuid);
26 	kfree(nd_pfn);
27 }
28 
29 struct nd_pfn *to_nd_pfn(struct device *dev)
30 {
31 	struct nd_pfn *nd_pfn = container_of(dev, struct nd_pfn, dev);
32 
33 	WARN_ON(!is_nd_pfn(dev));
34 	return nd_pfn;
35 }
36 EXPORT_SYMBOL(to_nd_pfn);
37 
38 static ssize_t mode_show(struct device *dev,
39 		struct device_attribute *attr, char *buf)
40 {
41 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
42 
43 	switch (nd_pfn->mode) {
44 	case PFN_MODE_RAM:
45 		return sprintf(buf, "ram\n");
46 	case PFN_MODE_PMEM:
47 		return sprintf(buf, "pmem\n");
48 	default:
49 		return sprintf(buf, "none\n");
50 	}
51 }
52 
53 static ssize_t mode_store(struct device *dev,
54 		struct device_attribute *attr, const char *buf, size_t len)
55 {
56 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
57 	ssize_t rc = 0;
58 
59 	nd_device_lock(dev);
60 	nvdimm_bus_lock(dev);
61 	if (dev->driver)
62 		rc = -EBUSY;
63 	else {
64 		size_t n = len - 1;
65 
66 		if (strncmp(buf, "pmem\n", n) == 0
67 				|| strncmp(buf, "pmem", n) == 0) {
68 			nd_pfn->mode = PFN_MODE_PMEM;
69 		} else if (strncmp(buf, "ram\n", n) == 0
70 				|| strncmp(buf, "ram", n) == 0)
71 			nd_pfn->mode = PFN_MODE_RAM;
72 		else if (strncmp(buf, "none\n", n) == 0
73 				|| strncmp(buf, "none", n) == 0)
74 			nd_pfn->mode = PFN_MODE_NONE;
75 		else
76 			rc = -EINVAL;
77 	}
78 	dev_dbg(dev, "result: %zd wrote: %s%s", rc, buf,
79 			buf[len - 1] == '\n' ? "" : "\n");
80 	nvdimm_bus_unlock(dev);
81 	nd_device_unlock(dev);
82 
83 	return rc ? rc : len;
84 }
85 static DEVICE_ATTR_RW(mode);
86 
87 static ssize_t align_show(struct device *dev,
88 		struct device_attribute *attr, char *buf)
89 {
90 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
91 
92 	return sprintf(buf, "%ld\n", nd_pfn->align);
93 }
94 
95 static unsigned long *nd_pfn_supported_alignments(unsigned long *alignments)
96 {
97 
98 	alignments[0] = PAGE_SIZE;
99 
100 	if (has_transparent_hugepage()) {
101 		alignments[1] = HPAGE_PMD_SIZE;
102 		if (IS_ENABLED(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD))
103 			alignments[2] = HPAGE_PUD_SIZE;
104 	}
105 
106 	return alignments;
107 }
108 
109 /*
110  * Use pmd mapping if supported as default alignment
111  */
112 static unsigned long nd_pfn_default_alignment(void)
113 {
114 
115 	if (has_transparent_hugepage())
116 		return HPAGE_PMD_SIZE;
117 	return PAGE_SIZE;
118 }
119 
120 static ssize_t align_store(struct device *dev,
121 		struct device_attribute *attr, const char *buf, size_t len)
122 {
123 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
124 	unsigned long aligns[MAX_NVDIMM_ALIGN] = { [0] = 0, };
125 	ssize_t rc;
126 
127 	nd_device_lock(dev);
128 	nvdimm_bus_lock(dev);
129 	rc = nd_size_select_store(dev, buf, &nd_pfn->align,
130 			nd_pfn_supported_alignments(aligns));
131 	dev_dbg(dev, "result: %zd wrote: %s%s", rc, buf,
132 			buf[len - 1] == '\n' ? "" : "\n");
133 	nvdimm_bus_unlock(dev);
134 	nd_device_unlock(dev);
135 
136 	return rc ? rc : len;
137 }
138 static DEVICE_ATTR_RW(align);
139 
140 static ssize_t uuid_show(struct device *dev,
141 		struct device_attribute *attr, char *buf)
142 {
143 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
144 
145 	if (nd_pfn->uuid)
146 		return sprintf(buf, "%pUb\n", nd_pfn->uuid);
147 	return sprintf(buf, "\n");
148 }
149 
150 static ssize_t uuid_store(struct device *dev,
151 		struct device_attribute *attr, const char *buf, size_t len)
152 {
153 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
154 	ssize_t rc;
155 
156 	nd_device_lock(dev);
157 	rc = nd_uuid_store(dev, &nd_pfn->uuid, buf, len);
158 	dev_dbg(dev, "result: %zd wrote: %s%s", rc, buf,
159 			buf[len - 1] == '\n' ? "" : "\n");
160 	nd_device_unlock(dev);
161 
162 	return rc ? rc : len;
163 }
164 static DEVICE_ATTR_RW(uuid);
165 
166 static ssize_t namespace_show(struct device *dev,
167 		struct device_attribute *attr, char *buf)
168 {
169 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
170 	ssize_t rc;
171 
172 	nvdimm_bus_lock(dev);
173 	rc = sprintf(buf, "%s\n", nd_pfn->ndns
174 			? dev_name(&nd_pfn->ndns->dev) : "");
175 	nvdimm_bus_unlock(dev);
176 	return rc;
177 }
178 
179 static ssize_t namespace_store(struct device *dev,
180 		struct device_attribute *attr, const char *buf, size_t len)
181 {
182 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
183 	ssize_t rc;
184 
185 	nd_device_lock(dev);
186 	nvdimm_bus_lock(dev);
187 	rc = nd_namespace_store(dev, &nd_pfn->ndns, buf, len);
188 	dev_dbg(dev, "result: %zd wrote: %s%s", rc, buf,
189 			buf[len - 1] == '\n' ? "" : "\n");
190 	nvdimm_bus_unlock(dev);
191 	nd_device_unlock(dev);
192 
193 	return rc;
194 }
195 static DEVICE_ATTR_RW(namespace);
196 
197 static ssize_t resource_show(struct device *dev,
198 		struct device_attribute *attr, char *buf)
199 {
200 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
201 	ssize_t rc;
202 
203 	nd_device_lock(dev);
204 	if (dev->driver) {
205 		struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
206 		u64 offset = __le64_to_cpu(pfn_sb->dataoff);
207 		struct nd_namespace_common *ndns = nd_pfn->ndns;
208 		u32 start_pad = __le32_to_cpu(pfn_sb->start_pad);
209 		struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
210 
211 		rc = sprintf(buf, "%#llx\n", (unsigned long long) nsio->res.start
212 				+ start_pad + offset);
213 	} else {
214 		/* no address to convey if the pfn instance is disabled */
215 		rc = -ENXIO;
216 	}
217 	nd_device_unlock(dev);
218 
219 	return rc;
220 }
221 static DEVICE_ATTR(resource, 0400, resource_show, NULL);
222 
223 static ssize_t size_show(struct device *dev,
224 		struct device_attribute *attr, char *buf)
225 {
226 	struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
227 	ssize_t rc;
228 
229 	nd_device_lock(dev);
230 	if (dev->driver) {
231 		struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
232 		u64 offset = __le64_to_cpu(pfn_sb->dataoff);
233 		struct nd_namespace_common *ndns = nd_pfn->ndns;
234 		u32 start_pad = __le32_to_cpu(pfn_sb->start_pad);
235 		u32 end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
236 		struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
237 
238 		rc = sprintf(buf, "%llu\n", (unsigned long long)
239 				resource_size(&nsio->res) - start_pad
240 				- end_trunc - offset);
241 	} else {
242 		/* no size to convey if the pfn instance is disabled */
243 		rc = -ENXIO;
244 	}
245 	nd_device_unlock(dev);
246 
247 	return rc;
248 }
249 static DEVICE_ATTR_RO(size);
250 
251 static ssize_t supported_alignments_show(struct device *dev,
252 		struct device_attribute *attr, char *buf)
253 {
254 	unsigned long aligns[MAX_NVDIMM_ALIGN] = { [0] = 0, };
255 
256 	return nd_size_select_show(0,
257 			nd_pfn_supported_alignments(aligns), buf);
258 }
259 static DEVICE_ATTR_RO(supported_alignments);
260 
261 static struct attribute *nd_pfn_attributes[] = {
262 	&dev_attr_mode.attr,
263 	&dev_attr_namespace.attr,
264 	&dev_attr_uuid.attr,
265 	&dev_attr_align.attr,
266 	&dev_attr_resource.attr,
267 	&dev_attr_size.attr,
268 	&dev_attr_supported_alignments.attr,
269 	NULL,
270 };
271 
272 static struct attribute_group nd_pfn_attribute_group = {
273 	.attrs = nd_pfn_attributes,
274 };
275 
276 const struct attribute_group *nd_pfn_attribute_groups[] = {
277 	&nd_pfn_attribute_group,
278 	&nd_device_attribute_group,
279 	&nd_numa_attribute_group,
280 	NULL,
281 };
282 
283 static const struct device_type nd_pfn_device_type = {
284 	.name = "nd_pfn",
285 	.release = nd_pfn_release,
286 	.groups = nd_pfn_attribute_groups,
287 };
288 
289 bool is_nd_pfn(struct device *dev)
290 {
291 	return dev ? dev->type == &nd_pfn_device_type : false;
292 }
293 EXPORT_SYMBOL(is_nd_pfn);
294 
295 struct device *nd_pfn_devinit(struct nd_pfn *nd_pfn,
296 		struct nd_namespace_common *ndns)
297 {
298 	struct device *dev;
299 
300 	if (!nd_pfn)
301 		return NULL;
302 
303 	nd_pfn->mode = PFN_MODE_NONE;
304 	nd_pfn->align = nd_pfn_default_alignment();
305 	dev = &nd_pfn->dev;
306 	device_initialize(&nd_pfn->dev);
307 	if (ndns && !__nd_attach_ndns(&nd_pfn->dev, ndns, &nd_pfn->ndns)) {
308 		dev_dbg(&ndns->dev, "failed, already claimed by %s\n",
309 				dev_name(ndns->claim));
310 		put_device(dev);
311 		return NULL;
312 	}
313 	return dev;
314 }
315 
316 static struct nd_pfn *nd_pfn_alloc(struct nd_region *nd_region)
317 {
318 	struct nd_pfn *nd_pfn;
319 	struct device *dev;
320 
321 	nd_pfn = kzalloc(sizeof(*nd_pfn), GFP_KERNEL);
322 	if (!nd_pfn)
323 		return NULL;
324 
325 	nd_pfn->id = ida_simple_get(&nd_region->pfn_ida, 0, 0, GFP_KERNEL);
326 	if (nd_pfn->id < 0) {
327 		kfree(nd_pfn);
328 		return NULL;
329 	}
330 
331 	dev = &nd_pfn->dev;
332 	dev_set_name(dev, "pfn%d.%d", nd_region->id, nd_pfn->id);
333 	dev->type = &nd_pfn_device_type;
334 	dev->parent = &nd_region->dev;
335 
336 	return nd_pfn;
337 }
338 
339 struct device *nd_pfn_create(struct nd_region *nd_region)
340 {
341 	struct nd_pfn *nd_pfn;
342 	struct device *dev;
343 
344 	if (!is_memory(&nd_region->dev))
345 		return NULL;
346 
347 	nd_pfn = nd_pfn_alloc(nd_region);
348 	dev = nd_pfn_devinit(nd_pfn, NULL);
349 
350 	__nd_device_register(dev);
351 	return dev;
352 }
353 
354 /*
355  * nd_pfn_clear_memmap_errors() clears any errors in the volatile memmap
356  * space associated with the namespace. If the memmap is set to DRAM, then
357  * this is a no-op. Since the memmap area is freshly initialized during
358  * probe, we have an opportunity to clear any badblocks in this area.
359  */
360 static int nd_pfn_clear_memmap_errors(struct nd_pfn *nd_pfn)
361 {
362 	struct nd_region *nd_region = to_nd_region(nd_pfn->dev.parent);
363 	struct nd_namespace_common *ndns = nd_pfn->ndns;
364 	void *zero_page = page_address(ZERO_PAGE(0));
365 	struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
366 	int num_bad, meta_num, rc, bb_present;
367 	sector_t first_bad, meta_start;
368 	struct nd_namespace_io *nsio;
369 
370 	if (nd_pfn->mode != PFN_MODE_PMEM)
371 		return 0;
372 
373 	nsio = to_nd_namespace_io(&ndns->dev);
374 	meta_start = (SZ_4K + sizeof(*pfn_sb)) >> 9;
375 	meta_num = (le64_to_cpu(pfn_sb->dataoff) >> 9) - meta_start;
376 
377 	/*
378 	 * re-enable the namespace with correct size so that we can access
379 	 * the device memmap area.
380 	 */
381 	devm_namespace_disable(&nd_pfn->dev, ndns);
382 	rc = devm_namespace_enable(&nd_pfn->dev, ndns, le64_to_cpu(pfn_sb->dataoff));
383 	if (rc)
384 		return rc;
385 
386 	do {
387 		unsigned long zero_len;
388 		u64 nsoff;
389 
390 		bb_present = badblocks_check(&nd_region->bb, meta_start,
391 				meta_num, &first_bad, &num_bad);
392 		if (bb_present) {
393 			dev_dbg(&nd_pfn->dev, "meta: %x badblocks at %llx\n",
394 					num_bad, first_bad);
395 			nsoff = ALIGN_DOWN((nd_region->ndr_start
396 					+ (first_bad << 9)) - nsio->res.start,
397 					PAGE_SIZE);
398 			zero_len = ALIGN(num_bad << 9, PAGE_SIZE);
399 			while (zero_len) {
400 				unsigned long chunk = min(zero_len, PAGE_SIZE);
401 
402 				rc = nvdimm_write_bytes(ndns, nsoff, zero_page,
403 							chunk, 0);
404 				if (rc)
405 					break;
406 
407 				zero_len -= chunk;
408 				nsoff += chunk;
409 			}
410 			if (rc) {
411 				dev_err(&nd_pfn->dev,
412 					"error clearing %x badblocks at %llx\n",
413 					num_bad, first_bad);
414 				return rc;
415 			}
416 		}
417 	} while (bb_present);
418 
419 	return 0;
420 }
421 
422 static bool nd_supported_alignment(unsigned long align)
423 {
424 	int i;
425 	unsigned long supported[MAX_NVDIMM_ALIGN] = { [0] = 0, };
426 
427 	if (align == 0)
428 		return false;
429 
430 	nd_pfn_supported_alignments(supported);
431 	for (i = 0; supported[i]; i++)
432 		if (align == supported[i])
433 			return true;
434 	return false;
435 }
436 
437 /**
438  * nd_pfn_validate - read and validate info-block
439  * @nd_pfn: fsdax namespace runtime state / properties
440  * @sig: 'devdax' or 'fsdax' signature
441  *
442  * Upon return the info-block buffer contents (->pfn_sb) are
443  * indeterminate when validation fails, and a coherent info-block
444  * otherwise.
445  */
446 int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig)
447 {
448 	u64 checksum, offset;
449 	enum nd_pfn_mode mode;
450 	struct nd_namespace_io *nsio;
451 	unsigned long align, start_pad;
452 	struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
453 	struct nd_namespace_common *ndns = nd_pfn->ndns;
454 	const u8 *parent_uuid = nd_dev_to_uuid(&ndns->dev);
455 
456 	if (!pfn_sb || !ndns)
457 		return -ENODEV;
458 
459 	if (!is_memory(nd_pfn->dev.parent))
460 		return -ENODEV;
461 
462 	if (nvdimm_read_bytes(ndns, SZ_4K, pfn_sb, sizeof(*pfn_sb), 0))
463 		return -ENXIO;
464 
465 	if (memcmp(pfn_sb->signature, sig, PFN_SIG_LEN) != 0)
466 		return -ENODEV;
467 
468 	checksum = le64_to_cpu(pfn_sb->checksum);
469 	pfn_sb->checksum = 0;
470 	if (checksum != nd_sb_checksum((struct nd_gen_sb *) pfn_sb))
471 		return -ENODEV;
472 	pfn_sb->checksum = cpu_to_le64(checksum);
473 
474 	if (memcmp(pfn_sb->parent_uuid, parent_uuid, 16) != 0)
475 		return -ENODEV;
476 
477 	if (__le16_to_cpu(pfn_sb->version_minor) < 1) {
478 		pfn_sb->start_pad = 0;
479 		pfn_sb->end_trunc = 0;
480 	}
481 
482 	if (__le16_to_cpu(pfn_sb->version_minor) < 2)
483 		pfn_sb->align = 0;
484 
485 	if (__le16_to_cpu(pfn_sb->version_minor) < 4) {
486 		pfn_sb->page_struct_size = cpu_to_le16(64);
487 		pfn_sb->page_size = cpu_to_le32(PAGE_SIZE);
488 	}
489 
490 	switch (le32_to_cpu(pfn_sb->mode)) {
491 	case PFN_MODE_RAM:
492 	case PFN_MODE_PMEM:
493 		break;
494 	default:
495 		return -ENXIO;
496 	}
497 
498 	align = le32_to_cpu(pfn_sb->align);
499 	offset = le64_to_cpu(pfn_sb->dataoff);
500 	start_pad = le32_to_cpu(pfn_sb->start_pad);
501 	if (align == 0)
502 		align = 1UL << ilog2(offset);
503 	mode = le32_to_cpu(pfn_sb->mode);
504 
505 	if ((le32_to_cpu(pfn_sb->page_size) > PAGE_SIZE) &&
506 			(mode == PFN_MODE_PMEM)) {
507 		dev_err(&nd_pfn->dev,
508 				"init failed, page size mismatch %d\n",
509 				le32_to_cpu(pfn_sb->page_size));
510 		return -EOPNOTSUPP;
511 	}
512 
513 	if ((le16_to_cpu(pfn_sb->page_struct_size) < sizeof(struct page)) &&
514 			(mode == PFN_MODE_PMEM)) {
515 		dev_err(&nd_pfn->dev,
516 				"init failed, struct page size mismatch %d\n",
517 				le16_to_cpu(pfn_sb->page_struct_size));
518 		return -EOPNOTSUPP;
519 	}
520 
521 	/*
522 	 * Check whether the we support the alignment. For Dax if the
523 	 * superblock alignment is not matching, we won't initialize
524 	 * the device.
525 	 */
526 	if (!nd_supported_alignment(align) &&
527 			!memcmp(pfn_sb->signature, DAX_SIG, PFN_SIG_LEN)) {
528 		dev_err(&nd_pfn->dev, "init failed, alignment mismatch: "
529 				"%ld:%ld\n", nd_pfn->align, align);
530 		return -EOPNOTSUPP;
531 	}
532 
533 	if (!nd_pfn->uuid) {
534 		/*
535 		 * When probing a namepace via nd_pfn_probe() the uuid
536 		 * is NULL (see: nd_pfn_devinit()) we init settings from
537 		 * pfn_sb
538 		 */
539 		nd_pfn->uuid = kmemdup(pfn_sb->uuid, 16, GFP_KERNEL);
540 		if (!nd_pfn->uuid)
541 			return -ENOMEM;
542 		nd_pfn->align = align;
543 		nd_pfn->mode = mode;
544 	} else {
545 		/*
546 		 * When probing a pfn / dax instance we validate the
547 		 * live settings against the pfn_sb
548 		 */
549 		if (memcmp(nd_pfn->uuid, pfn_sb->uuid, 16) != 0)
550 			return -ENODEV;
551 
552 		/*
553 		 * If the uuid validates, but other settings mismatch
554 		 * return EINVAL because userspace has managed to change
555 		 * the configuration without specifying new
556 		 * identification.
557 		 */
558 		if (nd_pfn->align != align || nd_pfn->mode != mode) {
559 			dev_err(&nd_pfn->dev,
560 					"init failed, settings mismatch\n");
561 			dev_dbg(&nd_pfn->dev, "align: %lx:%lx mode: %d:%d\n",
562 					nd_pfn->align, align, nd_pfn->mode,
563 					mode);
564 			return -EINVAL;
565 		}
566 	}
567 
568 	if (align > nvdimm_namespace_capacity(ndns)) {
569 		dev_err(&nd_pfn->dev, "alignment: %lx exceeds capacity %llx\n",
570 				align, nvdimm_namespace_capacity(ndns));
571 		return -EINVAL;
572 	}
573 
574 	/*
575 	 * These warnings are verbose because they can only trigger in
576 	 * the case where the physical address alignment of the
577 	 * namespace has changed since the pfn superblock was
578 	 * established.
579 	 */
580 	nsio = to_nd_namespace_io(&ndns->dev);
581 	if (offset >= resource_size(&nsio->res)) {
582 		dev_err(&nd_pfn->dev, "pfn array size exceeds capacity of %s\n",
583 				dev_name(&ndns->dev));
584 		return -EBUSY;
585 	}
586 
587 	if ((align && !IS_ALIGNED(nsio->res.start + offset + start_pad, align))
588 			|| !IS_ALIGNED(offset, PAGE_SIZE)) {
589 		dev_err(&nd_pfn->dev,
590 				"bad offset: %#llx dax disabled align: %#lx\n",
591 				offset, align);
592 		return -ENXIO;
593 	}
594 
595 	return 0;
596 }
597 EXPORT_SYMBOL(nd_pfn_validate);
598 
599 int nd_pfn_probe(struct device *dev, struct nd_namespace_common *ndns)
600 {
601 	int rc;
602 	struct nd_pfn *nd_pfn;
603 	struct device *pfn_dev;
604 	struct nd_pfn_sb *pfn_sb;
605 	struct nd_region *nd_region = to_nd_region(ndns->dev.parent);
606 
607 	if (ndns->force_raw)
608 		return -ENODEV;
609 
610 	switch (ndns->claim_class) {
611 	case NVDIMM_CCLASS_NONE:
612 	case NVDIMM_CCLASS_PFN:
613 		break;
614 	default:
615 		return -ENODEV;
616 	}
617 
618 	nvdimm_bus_lock(&ndns->dev);
619 	nd_pfn = nd_pfn_alloc(nd_region);
620 	pfn_dev = nd_pfn_devinit(nd_pfn, ndns);
621 	nvdimm_bus_unlock(&ndns->dev);
622 	if (!pfn_dev)
623 		return -ENOMEM;
624 	pfn_sb = devm_kmalloc(dev, sizeof(*pfn_sb), GFP_KERNEL);
625 	nd_pfn = to_nd_pfn(pfn_dev);
626 	nd_pfn->pfn_sb = pfn_sb;
627 	rc = nd_pfn_validate(nd_pfn, PFN_SIG);
628 	dev_dbg(dev, "pfn: %s\n", rc == 0 ? dev_name(pfn_dev) : "<none>");
629 	if (rc < 0) {
630 		nd_detach_ndns(pfn_dev, &nd_pfn->ndns);
631 		put_device(pfn_dev);
632 	} else
633 		__nd_device_register(pfn_dev);
634 
635 	return rc;
636 }
637 EXPORT_SYMBOL(nd_pfn_probe);
638 
639 /*
640  * We hotplug memory at sub-section granularity, pad the reserved area
641  * from the previous section base to the namespace base address.
642  */
643 static unsigned long init_altmap_base(resource_size_t base)
644 {
645 	unsigned long base_pfn = PHYS_PFN(base);
646 
647 	return SUBSECTION_ALIGN_DOWN(base_pfn);
648 }
649 
650 static unsigned long init_altmap_reserve(resource_size_t base)
651 {
652 	unsigned long reserve = nd_info_block_reserve() >> PAGE_SHIFT;
653 	unsigned long base_pfn = PHYS_PFN(base);
654 
655 	reserve += base_pfn - SUBSECTION_ALIGN_DOWN(base_pfn);
656 	return reserve;
657 }
658 
659 static int __nvdimm_setup_pfn(struct nd_pfn *nd_pfn, struct dev_pagemap *pgmap)
660 {
661 	struct resource *res = &pgmap->res;
662 	struct vmem_altmap *altmap = &pgmap->altmap;
663 	struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
664 	u64 offset = le64_to_cpu(pfn_sb->dataoff);
665 	u32 start_pad = __le32_to_cpu(pfn_sb->start_pad);
666 	u32 end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
667 	u32 reserve = nd_info_block_reserve();
668 	struct nd_namespace_common *ndns = nd_pfn->ndns;
669 	struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
670 	resource_size_t base = nsio->res.start + start_pad;
671 	resource_size_t end = nsio->res.end - end_trunc;
672 	struct vmem_altmap __altmap = {
673 		.base_pfn = init_altmap_base(base),
674 		.reserve = init_altmap_reserve(base),
675 		.end_pfn = PHYS_PFN(end),
676 	};
677 
678 	memcpy(res, &nsio->res, sizeof(*res));
679 	res->start += start_pad;
680 	res->end -= end_trunc;
681 
682 	if (nd_pfn->mode == PFN_MODE_RAM) {
683 		if (offset < reserve)
684 			return -EINVAL;
685 		nd_pfn->npfns = le64_to_cpu(pfn_sb->npfns);
686 	} else if (nd_pfn->mode == PFN_MODE_PMEM) {
687 		nd_pfn->npfns = PHYS_PFN((resource_size(res) - offset));
688 		if (le64_to_cpu(nd_pfn->pfn_sb->npfns) > nd_pfn->npfns)
689 			dev_info(&nd_pfn->dev,
690 					"number of pfns truncated from %lld to %ld\n",
691 					le64_to_cpu(nd_pfn->pfn_sb->npfns),
692 					nd_pfn->npfns);
693 		memcpy(altmap, &__altmap, sizeof(*altmap));
694 		altmap->free = PHYS_PFN(offset - reserve);
695 		altmap->alloc = 0;
696 		pgmap->flags |= PGMAP_ALTMAP_VALID;
697 	} else
698 		return -ENXIO;
699 
700 	return 0;
701 }
702 
703 static int nd_pfn_init(struct nd_pfn *nd_pfn)
704 {
705 	struct nd_namespace_common *ndns = nd_pfn->ndns;
706 	struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
707 	resource_size_t start, size;
708 	struct nd_region *nd_region;
709 	unsigned long npfns, align;
710 	u32 end_trunc;
711 	struct nd_pfn_sb *pfn_sb;
712 	phys_addr_t offset;
713 	const char *sig;
714 	u64 checksum;
715 	int rc;
716 
717 	pfn_sb = devm_kmalloc(&nd_pfn->dev, sizeof(*pfn_sb), GFP_KERNEL);
718 	if (!pfn_sb)
719 		return -ENOMEM;
720 
721 	nd_pfn->pfn_sb = pfn_sb;
722 	if (is_nd_dax(&nd_pfn->dev))
723 		sig = DAX_SIG;
724 	else
725 		sig = PFN_SIG;
726 
727 	rc = nd_pfn_validate(nd_pfn, sig);
728 	if (rc == 0)
729 		return nd_pfn_clear_memmap_errors(nd_pfn);
730 	if (rc != -ENODEV)
731 		return rc;
732 
733 	/* no info block, do init */;
734 	memset(pfn_sb, 0, sizeof(*pfn_sb));
735 
736 	nd_region = to_nd_region(nd_pfn->dev.parent);
737 	if (nd_region->ro) {
738 		dev_info(&nd_pfn->dev,
739 				"%s is read-only, unable to init metadata\n",
740 				dev_name(&nd_region->dev));
741 		return -ENXIO;
742 	}
743 
744 	/*
745 	 * Note, we use 64 here for the standard size of struct page,
746 	 * debugging options may cause it to be larger in which case the
747 	 * implementation will limit the pfns advertised through
748 	 * ->direct_access() to those that are included in the memmap.
749 	 */
750 	start = nsio->res.start;
751 	size = resource_size(&nsio->res);
752 	npfns = PHYS_PFN(size - SZ_8K);
753 	align = max(nd_pfn->align, (1UL << SUBSECTION_SHIFT));
754 	end_trunc = start + size - ALIGN_DOWN(start + size, align);
755 	if (nd_pfn->mode == PFN_MODE_PMEM) {
756 		/*
757 		 * The altmap should be padded out to the block size used
758 		 * when populating the vmemmap. This *should* be equal to
759 		 * PMD_SIZE for most architectures.
760 		 *
761 		 * Also make sure size of struct page is less than 64. We
762 		 * want to make sure we use large enough size here so that
763 		 * we don't have a dynamic reserve space depending on
764 		 * struct page size. But we also want to make sure we notice
765 		 * when we end up adding new elements to struct page.
766 		 */
767 		BUILD_BUG_ON(sizeof(struct page) > MAX_STRUCT_PAGE_SIZE);
768 		offset = ALIGN(start + SZ_8K + MAX_STRUCT_PAGE_SIZE * npfns, align)
769 			- start;
770 	} else if (nd_pfn->mode == PFN_MODE_RAM)
771 		offset = ALIGN(start + SZ_8K, align) - start;
772 	else
773 		return -ENXIO;
774 
775 	if (offset >= size) {
776 		dev_err(&nd_pfn->dev, "%s unable to satisfy requested alignment\n",
777 				dev_name(&ndns->dev));
778 		return -ENXIO;
779 	}
780 
781 	npfns = PHYS_PFN(size - offset - end_trunc);
782 	pfn_sb->mode = cpu_to_le32(nd_pfn->mode);
783 	pfn_sb->dataoff = cpu_to_le64(offset);
784 	pfn_sb->npfns = cpu_to_le64(npfns);
785 	memcpy(pfn_sb->signature, sig, PFN_SIG_LEN);
786 	memcpy(pfn_sb->uuid, nd_pfn->uuid, 16);
787 	memcpy(pfn_sb->parent_uuid, nd_dev_to_uuid(&ndns->dev), 16);
788 	pfn_sb->version_major = cpu_to_le16(1);
789 	pfn_sb->version_minor = cpu_to_le16(4);
790 	pfn_sb->end_trunc = cpu_to_le32(end_trunc);
791 	pfn_sb->align = cpu_to_le32(nd_pfn->align);
792 	pfn_sb->page_struct_size = cpu_to_le16(MAX_STRUCT_PAGE_SIZE);
793 	pfn_sb->page_size = cpu_to_le32(PAGE_SIZE);
794 	checksum = nd_sb_checksum((struct nd_gen_sb *) pfn_sb);
795 	pfn_sb->checksum = cpu_to_le64(checksum);
796 
797 	rc = nd_pfn_clear_memmap_errors(nd_pfn);
798 	if (rc)
799 		return rc;
800 
801 	return nvdimm_write_bytes(ndns, SZ_4K, pfn_sb, sizeof(*pfn_sb), 0);
802 }
803 
804 /*
805  * Determine the effective resource range and vmem_altmap from an nd_pfn
806  * instance.
807  */
808 int nvdimm_setup_pfn(struct nd_pfn *nd_pfn, struct dev_pagemap *pgmap)
809 {
810 	int rc;
811 
812 	if (!nd_pfn->uuid || !nd_pfn->ndns)
813 		return -ENODEV;
814 
815 	rc = nd_pfn_init(nd_pfn);
816 	if (rc)
817 		return rc;
818 
819 	/* we need a valid pfn_sb before we can init a dev_pagemap */
820 	return __nvdimm_setup_pfn(nd_pfn, pgmap);
821 }
822 EXPORT_SYMBOL_GPL(nvdimm_setup_pfn);
823