Loading...
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
17static 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
29struct 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}
36EXPORT_SYMBOL(to_nd_pfn);
37
38static 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
53static 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}
85static DEVICE_ATTR_RW(mode);
86
87static 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
95static 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 */
112static 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
120static 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}
138static DEVICE_ATTR_RW(align);
139
140static 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
150static 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}
164static DEVICE_ATTR_RW(uuid);
165
166static 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
179static 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}
195static DEVICE_ATTR_RW(namespace);
196
197static 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}
221static DEVICE_ATTR_ADMIN_RO(resource);
222
223static 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}
249static DEVICE_ATTR_RO(size);
250
251static 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}
259static DEVICE_ATTR_RO(supported_alignments);
260
261static 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
272static struct attribute_group nd_pfn_attribute_group = {
273 .attrs = nd_pfn_attributes,
274};
275
276const 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
283static 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
289bool is_nd_pfn(struct device *dev)
290{
291 return dev ? dev->type == &nd_pfn_device_type : false;
292}
293EXPORT_SYMBOL(is_nd_pfn);
294
295struct 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
316static 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
339struct 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 */
360static 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
422static 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 */
446int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig)
447{
448 u64 checksum, offset;
449 struct resource *res;
450 enum nd_pfn_mode mode;
451 struct nd_namespace_io *nsio;
452 unsigned long align, start_pad;
453 struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
454 struct nd_namespace_common *ndns = nd_pfn->ndns;
455 const u8 *parent_uuid = nd_dev_to_uuid(&ndns->dev);
456
457 if (!pfn_sb || !ndns)
458 return -ENODEV;
459
460 if (!is_memory(nd_pfn->dev.parent))
461 return -ENODEV;
462
463 if (nvdimm_read_bytes(ndns, SZ_4K, pfn_sb, sizeof(*pfn_sb), 0))
464 return -ENXIO;
465
466 if (memcmp(pfn_sb->signature, sig, PFN_SIG_LEN) != 0)
467 return -ENODEV;
468
469 checksum = le64_to_cpu(pfn_sb->checksum);
470 pfn_sb->checksum = 0;
471 if (checksum != nd_sb_checksum((struct nd_gen_sb *) pfn_sb))
472 return -ENODEV;
473 pfn_sb->checksum = cpu_to_le64(checksum);
474
475 if (memcmp(pfn_sb->parent_uuid, parent_uuid, 16) != 0)
476 return -ENODEV;
477
478 if (__le16_to_cpu(pfn_sb->version_minor) < 1) {
479 pfn_sb->start_pad = 0;
480 pfn_sb->end_trunc = 0;
481 }
482
483 if (__le16_to_cpu(pfn_sb->version_minor) < 2)
484 pfn_sb->align = 0;
485
486 if (__le16_to_cpu(pfn_sb->version_minor) < 4) {
487 pfn_sb->page_struct_size = cpu_to_le16(64);
488 pfn_sb->page_size = cpu_to_le32(PAGE_SIZE);
489 }
490
491 switch (le32_to_cpu(pfn_sb->mode)) {
492 case PFN_MODE_RAM:
493 case PFN_MODE_PMEM:
494 break;
495 default:
496 return -ENXIO;
497 }
498
499 align = le32_to_cpu(pfn_sb->align);
500 offset = le64_to_cpu(pfn_sb->dataoff);
501 start_pad = le32_to_cpu(pfn_sb->start_pad);
502 if (align == 0)
503 align = 1UL << ilog2(offset);
504 mode = le32_to_cpu(pfn_sb->mode);
505
506 if ((le32_to_cpu(pfn_sb->page_size) > PAGE_SIZE) &&
507 (mode == PFN_MODE_PMEM)) {
508 dev_err(&nd_pfn->dev,
509 "init failed, page size mismatch %d\n",
510 le32_to_cpu(pfn_sb->page_size));
511 return -EOPNOTSUPP;
512 }
513
514 if ((le16_to_cpu(pfn_sb->page_struct_size) < sizeof(struct page)) &&
515 (mode == PFN_MODE_PMEM)) {
516 dev_err(&nd_pfn->dev,
517 "init failed, struct page size mismatch %d\n",
518 le16_to_cpu(pfn_sb->page_struct_size));
519 return -EOPNOTSUPP;
520 }
521
522 /*
523 * Check whether the we support the alignment. For Dax if the
524 * superblock alignment is not matching, we won't initialize
525 * the device.
526 */
527 if (!nd_supported_alignment(align) &&
528 !memcmp(pfn_sb->signature, DAX_SIG, PFN_SIG_LEN)) {
529 dev_err(&nd_pfn->dev, "init failed, alignment mismatch: "
530 "%ld:%ld\n", nd_pfn->align, align);
531 return -EOPNOTSUPP;
532 }
533
534 if (!nd_pfn->uuid) {
535 /*
536 * When probing a namepace via nd_pfn_probe() the uuid
537 * is NULL (see: nd_pfn_devinit()) we init settings from
538 * pfn_sb
539 */
540 nd_pfn->uuid = kmemdup(pfn_sb->uuid, 16, GFP_KERNEL);
541 if (!nd_pfn->uuid)
542 return -ENOMEM;
543 nd_pfn->align = align;
544 nd_pfn->mode = mode;
545 } else {
546 /*
547 * When probing a pfn / dax instance we validate the
548 * live settings against the pfn_sb
549 */
550 if (memcmp(nd_pfn->uuid, pfn_sb->uuid, 16) != 0)
551 return -ENODEV;
552
553 /*
554 * If the uuid validates, but other settings mismatch
555 * return EINVAL because userspace has managed to change
556 * the configuration without specifying new
557 * identification.
558 */
559 if (nd_pfn->align != align || nd_pfn->mode != mode) {
560 dev_err(&nd_pfn->dev,
561 "init failed, settings mismatch\n");
562 dev_dbg(&nd_pfn->dev, "align: %lx:%lx mode: %d:%d\n",
563 nd_pfn->align, align, nd_pfn->mode,
564 mode);
565 return -EOPNOTSUPP;
566 }
567 }
568
569 if (align > nvdimm_namespace_capacity(ndns)) {
570 dev_err(&nd_pfn->dev, "alignment: %lx exceeds capacity %llx\n",
571 align, nvdimm_namespace_capacity(ndns));
572 return -EOPNOTSUPP;
573 }
574
575 /*
576 * These warnings are verbose because they can only trigger in
577 * the case where the physical address alignment of the
578 * namespace has changed since the pfn superblock was
579 * established.
580 */
581 nsio = to_nd_namespace_io(&ndns->dev);
582 res = &nsio->res;
583 if (offset >= resource_size(res)) {
584 dev_err(&nd_pfn->dev, "pfn array size exceeds capacity of %s\n",
585 dev_name(&ndns->dev));
586 return -EOPNOTSUPP;
587 }
588
589 if ((align && !IS_ALIGNED(res->start + offset + start_pad, align))
590 || !IS_ALIGNED(offset, PAGE_SIZE)) {
591 dev_err(&nd_pfn->dev,
592 "bad offset: %#llx dax disabled align: %#lx\n",
593 offset, align);
594 return -EOPNOTSUPP;
595 }
596
597 if (!IS_ALIGNED(res->start + le32_to_cpu(pfn_sb->start_pad),
598 memremap_compat_align())) {
599 dev_err(&nd_pfn->dev, "resource start misaligned\n");
600 return -EOPNOTSUPP;
601 }
602
603 if (!IS_ALIGNED(res->end + 1 - le32_to_cpu(pfn_sb->end_trunc),
604 memremap_compat_align())) {
605 dev_err(&nd_pfn->dev, "resource end misaligned\n");
606 return -EOPNOTSUPP;
607 }
608
609 return 0;
610}
611EXPORT_SYMBOL(nd_pfn_validate);
612
613int nd_pfn_probe(struct device *dev, struct nd_namespace_common *ndns)
614{
615 int rc;
616 struct nd_pfn *nd_pfn;
617 struct device *pfn_dev;
618 struct nd_pfn_sb *pfn_sb;
619 struct nd_region *nd_region = to_nd_region(ndns->dev.parent);
620
621 if (ndns->force_raw)
622 return -ENODEV;
623
624 switch (ndns->claim_class) {
625 case NVDIMM_CCLASS_NONE:
626 case NVDIMM_CCLASS_PFN:
627 break;
628 default:
629 return -ENODEV;
630 }
631
632 nvdimm_bus_lock(&ndns->dev);
633 nd_pfn = nd_pfn_alloc(nd_region);
634 pfn_dev = nd_pfn_devinit(nd_pfn, ndns);
635 nvdimm_bus_unlock(&ndns->dev);
636 if (!pfn_dev)
637 return -ENOMEM;
638 pfn_sb = devm_kmalloc(dev, sizeof(*pfn_sb), GFP_KERNEL);
639 nd_pfn = to_nd_pfn(pfn_dev);
640 nd_pfn->pfn_sb = pfn_sb;
641 rc = nd_pfn_validate(nd_pfn, PFN_SIG);
642 dev_dbg(dev, "pfn: %s\n", rc == 0 ? dev_name(pfn_dev) : "<none>");
643 if (rc < 0) {
644 nd_detach_ndns(pfn_dev, &nd_pfn->ndns);
645 put_device(pfn_dev);
646 } else
647 __nd_device_register(pfn_dev);
648
649 return rc;
650}
651EXPORT_SYMBOL(nd_pfn_probe);
652
653/*
654 * We hotplug memory at sub-section granularity, pad the reserved area
655 * from the previous section base to the namespace base address.
656 */
657static unsigned long init_altmap_base(resource_size_t base)
658{
659 unsigned long base_pfn = PHYS_PFN(base);
660
661 return SUBSECTION_ALIGN_DOWN(base_pfn);
662}
663
664static unsigned long init_altmap_reserve(resource_size_t base)
665{
666 unsigned long reserve = nd_info_block_reserve() >> PAGE_SHIFT;
667 unsigned long base_pfn = PHYS_PFN(base);
668
669 reserve += base_pfn - SUBSECTION_ALIGN_DOWN(base_pfn);
670 return reserve;
671}
672
673static int __nvdimm_setup_pfn(struct nd_pfn *nd_pfn, struct dev_pagemap *pgmap)
674{
675 struct resource *res = &pgmap->res;
676 struct vmem_altmap *altmap = &pgmap->altmap;
677 struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
678 u64 offset = le64_to_cpu(pfn_sb->dataoff);
679 u32 start_pad = __le32_to_cpu(pfn_sb->start_pad);
680 u32 end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
681 u32 reserve = nd_info_block_reserve();
682 struct nd_namespace_common *ndns = nd_pfn->ndns;
683 struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
684 resource_size_t base = nsio->res.start + start_pad;
685 resource_size_t end = nsio->res.end - end_trunc;
686 struct vmem_altmap __altmap = {
687 .base_pfn = init_altmap_base(base),
688 .reserve = init_altmap_reserve(base),
689 .end_pfn = PHYS_PFN(end),
690 };
691
692 memcpy(res, &nsio->res, sizeof(*res));
693 res->start += start_pad;
694 res->end -= end_trunc;
695
696 if (nd_pfn->mode == PFN_MODE_RAM) {
697 if (offset < reserve)
698 return -EINVAL;
699 nd_pfn->npfns = le64_to_cpu(pfn_sb->npfns);
700 } else if (nd_pfn->mode == PFN_MODE_PMEM) {
701 nd_pfn->npfns = PHYS_PFN((resource_size(res) - offset));
702 if (le64_to_cpu(nd_pfn->pfn_sb->npfns) > nd_pfn->npfns)
703 dev_info(&nd_pfn->dev,
704 "number of pfns truncated from %lld to %ld\n",
705 le64_to_cpu(nd_pfn->pfn_sb->npfns),
706 nd_pfn->npfns);
707 memcpy(altmap, &__altmap, sizeof(*altmap));
708 altmap->free = PHYS_PFN(offset - reserve);
709 altmap->alloc = 0;
710 pgmap->flags |= PGMAP_ALTMAP_VALID;
711 } else
712 return -ENXIO;
713
714 return 0;
715}
716
717static int nd_pfn_init(struct nd_pfn *nd_pfn)
718{
719 struct nd_namespace_common *ndns = nd_pfn->ndns;
720 struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
721 resource_size_t start, size;
722 struct nd_region *nd_region;
723 unsigned long npfns, align;
724 u32 end_trunc;
725 struct nd_pfn_sb *pfn_sb;
726 phys_addr_t offset;
727 const char *sig;
728 u64 checksum;
729 int rc;
730
731 pfn_sb = devm_kmalloc(&nd_pfn->dev, sizeof(*pfn_sb), GFP_KERNEL);
732 if (!pfn_sb)
733 return -ENOMEM;
734
735 nd_pfn->pfn_sb = pfn_sb;
736 if (is_nd_dax(&nd_pfn->dev))
737 sig = DAX_SIG;
738 else
739 sig = PFN_SIG;
740
741 rc = nd_pfn_validate(nd_pfn, sig);
742 if (rc == 0)
743 return nd_pfn_clear_memmap_errors(nd_pfn);
744 if (rc != -ENODEV)
745 return rc;
746
747 /* no info block, do init */;
748 memset(pfn_sb, 0, sizeof(*pfn_sb));
749
750 nd_region = to_nd_region(nd_pfn->dev.parent);
751 if (nd_region->ro) {
752 dev_info(&nd_pfn->dev,
753 "%s is read-only, unable to init metadata\n",
754 dev_name(&nd_region->dev));
755 return -ENXIO;
756 }
757
758 /*
759 * Note, we use 64 here for the standard size of struct page,
760 * debugging options may cause it to be larger in which case the
761 * implementation will limit the pfns advertised through
762 * ->direct_access() to those that are included in the memmap.
763 */
764 start = nsio->res.start;
765 size = resource_size(&nsio->res);
766 npfns = PHYS_PFN(size - SZ_8K);
767 align = max(nd_pfn->align, memremap_compat_align());
768
769 /*
770 * When @start is misaligned fail namespace creation. See
771 * the 'struct nd_pfn_sb' commentary on why ->start_pad is not
772 * an option.
773 */
774 if (!IS_ALIGNED(start, memremap_compat_align())) {
775 dev_err(&nd_pfn->dev, "%s: start %pa misaligned to %#lx\n",
776 dev_name(&ndns->dev), &start,
777 memremap_compat_align());
778 return -EINVAL;
779 }
780 end_trunc = start + size - ALIGN_DOWN(start + size, align);
781 if (nd_pfn->mode == PFN_MODE_PMEM) {
782 /*
783 * The altmap should be padded out to the block size used
784 * when populating the vmemmap. This *should* be equal to
785 * PMD_SIZE for most architectures.
786 *
787 * Also make sure size of struct page is less than 64. We
788 * want to make sure we use large enough size here so that
789 * we don't have a dynamic reserve space depending on
790 * struct page size. But we also want to make sure we notice
791 * when we end up adding new elements to struct page.
792 */
793 BUILD_BUG_ON(sizeof(struct page) > MAX_STRUCT_PAGE_SIZE);
794 offset = ALIGN(start + SZ_8K + MAX_STRUCT_PAGE_SIZE * npfns, align)
795 - start;
796 } else if (nd_pfn->mode == PFN_MODE_RAM)
797 offset = ALIGN(start + SZ_8K, align) - start;
798 else
799 return -ENXIO;
800
801 if (offset >= size) {
802 dev_err(&nd_pfn->dev, "%s unable to satisfy requested alignment\n",
803 dev_name(&ndns->dev));
804 return -ENXIO;
805 }
806
807 npfns = PHYS_PFN(size - offset - end_trunc);
808 pfn_sb->mode = cpu_to_le32(nd_pfn->mode);
809 pfn_sb->dataoff = cpu_to_le64(offset);
810 pfn_sb->npfns = cpu_to_le64(npfns);
811 memcpy(pfn_sb->signature, sig, PFN_SIG_LEN);
812 memcpy(pfn_sb->uuid, nd_pfn->uuid, 16);
813 memcpy(pfn_sb->parent_uuid, nd_dev_to_uuid(&ndns->dev), 16);
814 pfn_sb->version_major = cpu_to_le16(1);
815 pfn_sb->version_minor = cpu_to_le16(4);
816 pfn_sb->end_trunc = cpu_to_le32(end_trunc);
817 pfn_sb->align = cpu_to_le32(nd_pfn->align);
818 pfn_sb->page_struct_size = cpu_to_le16(MAX_STRUCT_PAGE_SIZE);
819 pfn_sb->page_size = cpu_to_le32(PAGE_SIZE);
820 checksum = nd_sb_checksum((struct nd_gen_sb *) pfn_sb);
821 pfn_sb->checksum = cpu_to_le64(checksum);
822
823 rc = nd_pfn_clear_memmap_errors(nd_pfn);
824 if (rc)
825 return rc;
826
827 return nvdimm_write_bytes(ndns, SZ_4K, pfn_sb, sizeof(*pfn_sb), 0);
828}
829
830/*
831 * Determine the effective resource range and vmem_altmap from an nd_pfn
832 * instance.
833 */
834int nvdimm_setup_pfn(struct nd_pfn *nd_pfn, struct dev_pagemap *pgmap)
835{
836 int rc;
837
838 if (!nd_pfn->uuid || !nd_pfn->ndns)
839 return -ENODEV;
840
841 rc = nd_pfn_init(nd_pfn);
842 if (rc)
843 return rc;
844
845 /* we need a valid pfn_sb before we can init a dev_pagemap */
846 return __nvdimm_setup_pfn(nd_pfn, pgmap);
847}
848EXPORT_SYMBOL_GPL(nvdimm_setup_pfn);
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
16static const bool page_struct_override = IS_ENABLED(CONFIG_NVDIMM_KMSAN);
17
18static 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_free(&nd_region->pfn_ida, nd_pfn->id);
26 kfree(nd_pfn->uuid);
27 kfree(nd_pfn);
28}
29
30struct 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}
37EXPORT_SYMBOL(to_nd_pfn);
38
39static 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
54static 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}
86static DEVICE_ATTR_RW(mode);
87
88static 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
96static 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 */
113static 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
121static 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}
139static DEVICE_ATTR_RW(align);
140
141static 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
151static 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}
165static DEVICE_ATTR_RW(uuid);
166
167static 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
180static 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}
196static DEVICE_ATTR_RW(namespace);
197
198static 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}
222static DEVICE_ATTR_ADMIN_RO(resource);
223
224static 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}
250static DEVICE_ATTR_RO(size);
251
252static 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}
260static DEVICE_ATTR_RO(supported_alignments);
261
262static 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
273static struct attribute_group nd_pfn_attribute_group = {
274 .attrs = nd_pfn_attributes,
275};
276
277const 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
284static 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
290bool is_nd_pfn(struct device *dev)
291{
292 return dev ? dev->type == &nd_pfn_device_type : false;
293}
294EXPORT_SYMBOL(is_nd_pfn);
295
296static struct lock_class_key nvdimm_pfn_key;
297
298struct 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
320static 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_alloc(&nd_region->pfn_ida, 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
343struct 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 */
364static 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
426static 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 */
450int 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 namespace 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}
620EXPORT_SYMBOL(nd_pfn_validate);
621
622int 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}
660EXPORT_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 */
666static 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
673static 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
682static 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
727static 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 */
855int 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}
869EXPORT_SYMBOL_GPL(nvdimm_setup_pfn);