1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright(c) 2013-2016 Intel Corporation. All rights reserved.
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>
17 static void nd_pfn_release(struct device *dev)
19 struct nd_region *nd_region = to_nd_region(dev->parent);
20 struct nd_pfn *nd_pfn = to_nd_pfn(dev);
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);
29 static struct device_type nd_pfn_device_type = {
31 .release = nd_pfn_release,
34 bool is_nd_pfn(struct device *dev)
36 return dev ? dev->type == &nd_pfn_device_type : false;
38 EXPORT_SYMBOL(is_nd_pfn);
40 struct nd_pfn *to_nd_pfn(struct device *dev)
42 struct nd_pfn *nd_pfn = container_of(dev, struct nd_pfn, dev);
44 WARN_ON(!is_nd_pfn(dev));
47 EXPORT_SYMBOL(to_nd_pfn);
49 static ssize_t mode_show(struct device *dev,
50 struct device_attribute *attr, char *buf)
52 struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
54 switch (nd_pfn->mode) {
56 return sprintf(buf, "ram\n");
58 return sprintf(buf, "pmem\n");
60 return sprintf(buf, "none\n");
64 static ssize_t mode_store(struct device *dev,
65 struct device_attribute *attr, const char *buf, size_t len)
67 struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
77 if (strncmp(buf, "pmem\n", n) == 0
78 || strncmp(buf, "pmem", n) == 0) {
79 nd_pfn->mode = PFN_MODE_PMEM;
80 } else if (strncmp(buf, "ram\n", n) == 0
81 || strncmp(buf, "ram", n) == 0)
82 nd_pfn->mode = PFN_MODE_RAM;
83 else if (strncmp(buf, "none\n", n) == 0
84 || strncmp(buf, "none", n) == 0)
85 nd_pfn->mode = PFN_MODE_NONE;
89 dev_dbg(dev, "result: %zd wrote: %s%s", rc, buf,
90 buf[len - 1] == '\n' ? "" : "\n");
91 nvdimm_bus_unlock(dev);
92 nd_device_unlock(dev);
96 static DEVICE_ATTR_RW(mode);
98 static ssize_t align_show(struct device *dev,
99 struct device_attribute *attr, char *buf)
101 struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
103 return sprintf(buf, "%ld\n", nd_pfn->align);
106 static const unsigned long *nd_pfn_supported_alignments(void)
109 * This needs to be a non-static variable because the *_SIZE
110 * macros aren't always constants.
112 const unsigned long supported_alignments[] = {
114 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
116 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
122 static unsigned long data[ARRAY_SIZE(supported_alignments)];
124 memcpy(data, supported_alignments, sizeof(data));
129 static ssize_t align_store(struct device *dev,
130 struct device_attribute *attr, const char *buf, size_t len)
132 struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
136 nvdimm_bus_lock(dev);
137 rc = nd_size_select_store(dev, buf, &nd_pfn->align,
138 nd_pfn_supported_alignments());
139 dev_dbg(dev, "result: %zd wrote: %s%s", rc, buf,
140 buf[len - 1] == '\n' ? "" : "\n");
141 nvdimm_bus_unlock(dev);
142 nd_device_unlock(dev);
144 return rc ? rc : len;
146 static DEVICE_ATTR_RW(align);
148 static ssize_t uuid_show(struct device *dev,
149 struct device_attribute *attr, char *buf)
151 struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
154 return sprintf(buf, "%pUb\n", nd_pfn->uuid);
155 return sprintf(buf, "\n");
158 static ssize_t uuid_store(struct device *dev,
159 struct device_attribute *attr, const char *buf, size_t len)
161 struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
165 rc = nd_uuid_store(dev, &nd_pfn->uuid, buf, len);
166 dev_dbg(dev, "result: %zd wrote: %s%s", rc, buf,
167 buf[len - 1] == '\n' ? "" : "\n");
168 nd_device_unlock(dev);
170 return rc ? rc : len;
172 static DEVICE_ATTR_RW(uuid);
174 static ssize_t namespace_show(struct device *dev,
175 struct device_attribute *attr, char *buf)
177 struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
180 nvdimm_bus_lock(dev);
181 rc = sprintf(buf, "%s\n", nd_pfn->ndns
182 ? dev_name(&nd_pfn->ndns->dev) : "");
183 nvdimm_bus_unlock(dev);
187 static ssize_t namespace_store(struct device *dev,
188 struct device_attribute *attr, const char *buf, size_t len)
190 struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
194 nvdimm_bus_lock(dev);
195 rc = nd_namespace_store(dev, &nd_pfn->ndns, buf, len);
196 dev_dbg(dev, "result: %zd wrote: %s%s", rc, buf,
197 buf[len - 1] == '\n' ? "" : "\n");
198 nvdimm_bus_unlock(dev);
199 nd_device_unlock(dev);
203 static DEVICE_ATTR_RW(namespace);
205 static ssize_t resource_show(struct device *dev,
206 struct device_attribute *attr, char *buf)
208 struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
213 struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
214 u64 offset = __le64_to_cpu(pfn_sb->dataoff);
215 struct nd_namespace_common *ndns = nd_pfn->ndns;
216 u32 start_pad = __le32_to_cpu(pfn_sb->start_pad);
217 struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
219 rc = sprintf(buf, "%#llx\n", (unsigned long long) nsio->res.start
220 + start_pad + offset);
222 /* no address to convey if the pfn instance is disabled */
225 nd_device_unlock(dev);
229 static DEVICE_ATTR_RO(resource);
231 static ssize_t size_show(struct device *dev,
232 struct device_attribute *attr, char *buf)
234 struct nd_pfn *nd_pfn = to_nd_pfn_safe(dev);
239 struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
240 u64 offset = __le64_to_cpu(pfn_sb->dataoff);
241 struct nd_namespace_common *ndns = nd_pfn->ndns;
242 u32 start_pad = __le32_to_cpu(pfn_sb->start_pad);
243 u32 end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
244 struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
246 rc = sprintf(buf, "%llu\n", (unsigned long long)
247 resource_size(&nsio->res) - start_pad
248 - end_trunc - offset);
250 /* no size to convey if the pfn instance is disabled */
253 nd_device_unlock(dev);
257 static DEVICE_ATTR_RO(size);
259 static ssize_t supported_alignments_show(struct device *dev,
260 struct device_attribute *attr, char *buf)
262 return nd_size_select_show(0, nd_pfn_supported_alignments(), buf);
264 static DEVICE_ATTR_RO(supported_alignments);
266 static struct attribute *nd_pfn_attributes[] = {
268 &dev_attr_namespace.attr,
270 &dev_attr_align.attr,
271 &dev_attr_resource.attr,
273 &dev_attr_supported_alignments.attr,
277 static umode_t pfn_visible(struct kobject *kobj, struct attribute *a, int n)
279 if (a == &dev_attr_resource.attr)
284 struct attribute_group nd_pfn_attribute_group = {
285 .attrs = nd_pfn_attributes,
286 .is_visible = pfn_visible,
289 static const struct attribute_group *nd_pfn_attribute_groups[] = {
290 &nd_pfn_attribute_group,
291 &nd_device_attribute_group,
292 &nd_numa_attribute_group,
296 struct device *nd_pfn_devinit(struct nd_pfn *nd_pfn,
297 struct nd_namespace_common *ndns)
304 nd_pfn->mode = PFN_MODE_NONE;
305 nd_pfn->align = PFN_DEFAULT_ALIGNMENT;
307 device_initialize(&nd_pfn->dev);
308 if (ndns && !__nd_attach_ndns(&nd_pfn->dev, ndns, &nd_pfn->ndns)) {
309 dev_dbg(&ndns->dev, "failed, already claimed by %s\n",
310 dev_name(ndns->claim));
317 static struct nd_pfn *nd_pfn_alloc(struct nd_region *nd_region)
319 struct nd_pfn *nd_pfn;
322 nd_pfn = kzalloc(sizeof(*nd_pfn), GFP_KERNEL);
326 nd_pfn->id = ida_simple_get(&nd_region->pfn_ida, 0, 0, GFP_KERNEL);
327 if (nd_pfn->id < 0) {
333 dev_set_name(dev, "pfn%d.%d", nd_region->id, nd_pfn->id);
334 dev->groups = nd_pfn_attribute_groups;
335 dev->type = &nd_pfn_device_type;
336 dev->parent = &nd_region->dev;
341 struct device *nd_pfn_create(struct nd_region *nd_region)
343 struct nd_pfn *nd_pfn;
346 if (!is_memory(&nd_region->dev))
349 nd_pfn = nd_pfn_alloc(nd_region);
350 dev = nd_pfn_devinit(nd_pfn, NULL);
352 __nd_device_register(dev);
357 * nd_pfn_clear_memmap_errors() clears any errors in the volatile memmap
358 * space associated with the namespace. If the memmap is set to DRAM, then
359 * this is a no-op. Since the memmap area is freshly initialized during
360 * probe, we have an opportunity to clear any badblocks in this area.
362 static int nd_pfn_clear_memmap_errors(struct nd_pfn *nd_pfn)
364 struct nd_region *nd_region = to_nd_region(nd_pfn->dev.parent);
365 struct nd_namespace_common *ndns = nd_pfn->ndns;
366 void *zero_page = page_address(ZERO_PAGE(0));
367 struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
368 int num_bad, meta_num, rc, bb_present;
369 sector_t first_bad, meta_start;
370 struct nd_namespace_io *nsio;
372 if (nd_pfn->mode != PFN_MODE_PMEM)
375 nsio = to_nd_namespace_io(&ndns->dev);
376 meta_start = (SZ_4K + sizeof(*pfn_sb)) >> 9;
377 meta_num = (le64_to_cpu(pfn_sb->dataoff) >> 9) - meta_start;
380 unsigned long zero_len;
383 bb_present = badblocks_check(&nd_region->bb, meta_start,
384 meta_num, &first_bad, &num_bad);
386 dev_dbg(&nd_pfn->dev, "meta: %x badblocks at %llx\n",
388 nsoff = ALIGN_DOWN((nd_region->ndr_start
389 + (first_bad << 9)) - nsio->res.start,
391 zero_len = ALIGN(num_bad << 9, PAGE_SIZE);
393 unsigned long chunk = min(zero_len, PAGE_SIZE);
395 rc = nvdimm_write_bytes(ndns, nsoff, zero_page,
404 dev_err(&nd_pfn->dev,
405 "error clearing %x badblocks at %llx\n",
410 } while (bb_present);
416 * nd_pfn_validate - read and validate info-block
417 * @nd_pfn: fsdax namespace runtime state / properties
418 * @sig: 'devdax' or 'fsdax' signature
420 * Upon return the info-block buffer contents (->pfn_sb) are
421 * indeterminate when validation fails, and a coherent info-block
424 int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig)
426 u64 checksum, offset;
427 enum nd_pfn_mode mode;
428 struct nd_namespace_io *nsio;
429 unsigned long align, start_pad;
430 struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
431 struct nd_namespace_common *ndns = nd_pfn->ndns;
432 const u8 *parent_uuid = nd_dev_to_uuid(&ndns->dev);
434 if (!pfn_sb || !ndns)
437 if (!is_memory(nd_pfn->dev.parent))
440 if (nvdimm_read_bytes(ndns, SZ_4K, pfn_sb, sizeof(*pfn_sb), 0))
443 if (memcmp(pfn_sb->signature, sig, PFN_SIG_LEN) != 0)
446 checksum = le64_to_cpu(pfn_sb->checksum);
447 pfn_sb->checksum = 0;
448 if (checksum != nd_sb_checksum((struct nd_gen_sb *) pfn_sb))
450 pfn_sb->checksum = cpu_to_le64(checksum);
452 if (memcmp(pfn_sb->parent_uuid, parent_uuid, 16) != 0)
455 if (__le16_to_cpu(pfn_sb->version_minor) < 1) {
456 pfn_sb->start_pad = 0;
457 pfn_sb->end_trunc = 0;
460 if (__le16_to_cpu(pfn_sb->version_minor) < 2)
463 switch (le32_to_cpu(pfn_sb->mode)) {
471 align = le32_to_cpu(pfn_sb->align);
472 offset = le64_to_cpu(pfn_sb->dataoff);
473 start_pad = le32_to_cpu(pfn_sb->start_pad);
475 align = 1UL << ilog2(offset);
476 mode = le32_to_cpu(pfn_sb->mode);
480 * When probing a namepace via nd_pfn_probe() the uuid
481 * is NULL (see: nd_pfn_devinit()) we init settings from
484 nd_pfn->uuid = kmemdup(pfn_sb->uuid, 16, GFP_KERNEL);
487 nd_pfn->align = align;
491 * When probing a pfn / dax instance we validate the
492 * live settings against the pfn_sb
494 if (memcmp(nd_pfn->uuid, pfn_sb->uuid, 16) != 0)
498 * If the uuid validates, but other settings mismatch
499 * return EINVAL because userspace has managed to change
500 * the configuration without specifying new
503 if (nd_pfn->align != align || nd_pfn->mode != mode) {
504 dev_err(&nd_pfn->dev,
505 "init failed, settings mismatch\n");
506 dev_dbg(&nd_pfn->dev, "align: %lx:%lx mode: %d:%d\n",
507 nd_pfn->align, align, nd_pfn->mode,
513 if (align > nvdimm_namespace_capacity(ndns)) {
514 dev_err(&nd_pfn->dev, "alignment: %lx exceeds capacity %llx\n",
515 align, nvdimm_namespace_capacity(ndns));
520 * These warnings are verbose because they can only trigger in
521 * the case where the physical address alignment of the
522 * namespace has changed since the pfn superblock was
525 nsio = to_nd_namespace_io(&ndns->dev);
526 if (offset >= resource_size(&nsio->res)) {
527 dev_err(&nd_pfn->dev, "pfn array size exceeds capacity of %s\n",
528 dev_name(&ndns->dev));
532 if ((align && !IS_ALIGNED(nsio->res.start + offset + start_pad, align))
533 || !IS_ALIGNED(offset, PAGE_SIZE)) {
534 dev_err(&nd_pfn->dev,
535 "bad offset: %#llx dax disabled align: %#lx\n",
540 return nd_pfn_clear_memmap_errors(nd_pfn);
542 EXPORT_SYMBOL(nd_pfn_validate);
544 int nd_pfn_probe(struct device *dev, struct nd_namespace_common *ndns)
547 struct nd_pfn *nd_pfn;
548 struct device *pfn_dev;
549 struct nd_pfn_sb *pfn_sb;
550 struct nd_region *nd_region = to_nd_region(ndns->dev.parent);
555 switch (ndns->claim_class) {
556 case NVDIMM_CCLASS_NONE:
557 case NVDIMM_CCLASS_PFN:
563 nvdimm_bus_lock(&ndns->dev);
564 nd_pfn = nd_pfn_alloc(nd_region);
565 pfn_dev = nd_pfn_devinit(nd_pfn, ndns);
566 nvdimm_bus_unlock(&ndns->dev);
569 pfn_sb = devm_kmalloc(dev, sizeof(*pfn_sb), GFP_KERNEL);
570 nd_pfn = to_nd_pfn(pfn_dev);
571 nd_pfn->pfn_sb = pfn_sb;
572 rc = nd_pfn_validate(nd_pfn, PFN_SIG);
573 dev_dbg(dev, "pfn: %s\n", rc == 0 ? dev_name(pfn_dev) : "<none>");
575 nd_detach_ndns(pfn_dev, &nd_pfn->ndns);
578 __nd_device_register(pfn_dev);
582 EXPORT_SYMBOL(nd_pfn_probe);
584 static u32 info_block_reserve(void)
586 return ALIGN(SZ_8K, PAGE_SIZE);
590 * We hotplug memory at sub-section granularity, pad the reserved area
591 * from the previous section base to the namespace base address.
593 static unsigned long init_altmap_base(resource_size_t base)
595 unsigned long base_pfn = PHYS_PFN(base);
597 return SUBSECTION_ALIGN_DOWN(base_pfn);
600 static unsigned long init_altmap_reserve(resource_size_t base)
602 unsigned long reserve = info_block_reserve() >> PAGE_SHIFT;
603 unsigned long base_pfn = PHYS_PFN(base);
605 reserve += base_pfn - SUBSECTION_ALIGN_DOWN(base_pfn);
609 static int __nvdimm_setup_pfn(struct nd_pfn *nd_pfn, struct dev_pagemap *pgmap)
611 struct resource *res = &pgmap->res;
612 struct vmem_altmap *altmap = &pgmap->altmap;
613 struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
614 u64 offset = le64_to_cpu(pfn_sb->dataoff);
615 u32 start_pad = __le32_to_cpu(pfn_sb->start_pad);
616 u32 end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
617 u32 reserve = info_block_reserve();
618 struct nd_namespace_common *ndns = nd_pfn->ndns;
619 struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
620 resource_size_t base = nsio->res.start + start_pad;
621 struct vmem_altmap __altmap = {
622 .base_pfn = init_altmap_base(base),
623 .reserve = init_altmap_reserve(base),
626 memcpy(res, &nsio->res, sizeof(*res));
627 res->start += start_pad;
628 res->end -= end_trunc;
630 if (nd_pfn->mode == PFN_MODE_RAM) {
631 if (offset < reserve)
633 nd_pfn->npfns = le64_to_cpu(pfn_sb->npfns);
634 } else if (nd_pfn->mode == PFN_MODE_PMEM) {
635 nd_pfn->npfns = PHYS_PFN((resource_size(res) - offset));
636 if (le64_to_cpu(nd_pfn->pfn_sb->npfns) > nd_pfn->npfns)
637 dev_info(&nd_pfn->dev,
638 "number of pfns truncated from %lld to %ld\n",
639 le64_to_cpu(nd_pfn->pfn_sb->npfns),
641 memcpy(altmap, &__altmap, sizeof(*altmap));
642 altmap->free = PHYS_PFN(offset - reserve);
644 pgmap->flags |= PGMAP_ALTMAP_VALID;
651 static int nd_pfn_init(struct nd_pfn *nd_pfn)
653 struct nd_namespace_common *ndns = nd_pfn->ndns;
654 struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
655 resource_size_t start, size;
656 struct nd_region *nd_region;
657 unsigned long npfns, align;
659 struct nd_pfn_sb *pfn_sb;
665 pfn_sb = devm_kmalloc(&nd_pfn->dev, sizeof(*pfn_sb), GFP_KERNEL);
669 nd_pfn->pfn_sb = pfn_sb;
670 if (is_nd_dax(&nd_pfn->dev))
675 rc = nd_pfn_validate(nd_pfn, sig);
679 /* no info block, do init */;
680 memset(pfn_sb, 0, sizeof(*pfn_sb));
682 nd_region = to_nd_region(nd_pfn->dev.parent);
684 dev_info(&nd_pfn->dev,
685 "%s is read-only, unable to init metadata\n",
686 dev_name(&nd_region->dev));
691 * Note, we use 64 here for the standard size of struct page,
692 * debugging options may cause it to be larger in which case the
693 * implementation will limit the pfns advertised through
694 * ->direct_access() to those that are included in the memmap.
696 start = nsio->res.start;
697 size = resource_size(&nsio->res);
698 npfns = PHYS_PFN(size - SZ_8K);
699 align = max(nd_pfn->align, (1UL << SUBSECTION_SHIFT));
700 end_trunc = start + size - ALIGN_DOWN(start + size, align);
701 if (nd_pfn->mode == PFN_MODE_PMEM) {
703 * The altmap should be padded out to the block size used
704 * when populating the vmemmap. This *should* be equal to
705 * PMD_SIZE for most architectures.
707 offset = ALIGN(start + SZ_8K + 64 * npfns, align) - start;
708 } else if (nd_pfn->mode == PFN_MODE_RAM)
709 offset = ALIGN(start + SZ_8K, align) - start;
713 if (offset >= size) {
714 dev_err(&nd_pfn->dev, "%s unable to satisfy requested alignment\n",
715 dev_name(&ndns->dev));
719 npfns = PHYS_PFN(size - offset - end_trunc);
720 pfn_sb->mode = cpu_to_le32(nd_pfn->mode);
721 pfn_sb->dataoff = cpu_to_le64(offset);
722 pfn_sb->npfns = cpu_to_le64(npfns);
723 memcpy(pfn_sb->signature, sig, PFN_SIG_LEN);
724 memcpy(pfn_sb->uuid, nd_pfn->uuid, 16);
725 memcpy(pfn_sb->parent_uuid, nd_dev_to_uuid(&ndns->dev), 16);
726 pfn_sb->version_major = cpu_to_le16(1);
727 pfn_sb->version_minor = cpu_to_le16(3);
728 pfn_sb->end_trunc = cpu_to_le32(end_trunc);
729 pfn_sb->align = cpu_to_le32(nd_pfn->align);
730 checksum = nd_sb_checksum((struct nd_gen_sb *) pfn_sb);
731 pfn_sb->checksum = cpu_to_le64(checksum);
733 return nvdimm_write_bytes(ndns, SZ_4K, pfn_sb, sizeof(*pfn_sb), 0);
737 * Determine the effective resource range and vmem_altmap from an nd_pfn
740 int nvdimm_setup_pfn(struct nd_pfn *nd_pfn, struct dev_pagemap *pgmap)
744 if (!nd_pfn->uuid || !nd_pfn->ndns)
747 rc = nd_pfn_init(nd_pfn);
751 /* we need a valid pfn_sb before we can init a dev_pagemap */
752 return __nvdimm_setup_pfn(nd_pfn, pgmap);
754 EXPORT_SYMBOL_GPL(nvdimm_setup_pfn);