2 * Persistent Memory Driver
4 * Copyright (c) 2014-2015, Intel Corporation.
5 * Copyright (c) 2015, Christoph Hellwig <hch@lst.de>.
6 * Copyright (c) 2015, Boaz Harrosh <boaz@plexistor.com>.
8 * This program is free software; you can redistribute it and/or modify it
9 * under the terms and conditions of the GNU General Public License,
10 * version 2, as published by the Free Software Foundation.
12 * This program is distributed in the hope it will be useful, but WITHOUT
13 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
18 #include <asm/cacheflush.h>
19 #include <linux/blkdev.h>
20 #include <linux/hdreg.h>
21 #include <linux/init.h>
22 #include <linux/platform_device.h>
23 #include <linux/module.h>
24 #include <linux/moduleparam.h>
25 #include <linux/badblocks.h>
26 #include <linux/memremap.h>
27 #include <linux/vmalloc.h>
28 #include <linux/pfn_t.h>
29 #include <linux/slab.h>
30 #include <linux/pmem.h>
36 struct request_queue *pmem_queue;
37 struct gendisk *pmem_disk;
38 struct nd_namespace_common *ndns;
40 /* One contiguous memory region per device */
41 phys_addr_t phys_addr;
42 /* when non-zero this device is hosting a 'pfn' instance */
43 phys_addr_t data_offset;
45 void __pmem *virt_addr;
46 /* immutable base size of the namespace */
48 /* trim size when namespace capacity has been section aligned */
53 static bool is_bad_pmem(struct badblocks *bb, sector_t sector, unsigned int len)
59 return !!badblocks_check(bb, sector, len / 512, &first_bad,
66 static void pmem_clear_poison(struct pmem_device *pmem, phys_addr_t offset,
69 struct device *dev = disk_to_dev(pmem->pmem_disk);
73 sector = (offset - pmem->data_offset) / 512;
74 cleared = nvdimm_clear_poison(dev, pmem->phys_addr + offset, len);
76 if (cleared > 0 && cleared / 512) {
77 dev_dbg(dev, "%s: %llx clear %ld sector%s\n",
78 __func__, (unsigned long long) sector,
79 cleared / 512, cleared / 512 > 1 ? "s" : "");
80 badblocks_clear(&pmem->bb, sector, cleared / 512);
82 invalidate_pmem(pmem->virt_addr + offset, len);
85 static int pmem_do_bvec(struct pmem_device *pmem, struct page *page,
86 unsigned int len, unsigned int off, int rw,
90 bool bad_pmem = false;
91 void *mem = kmap_atomic(page);
92 phys_addr_t pmem_off = sector * 512 + pmem->data_offset;
93 void __pmem *pmem_addr = pmem->virt_addr + pmem_off;
95 if (unlikely(is_bad_pmem(&pmem->bb, sector, len)))
99 if (unlikely(bad_pmem))
102 rc = memcpy_from_pmem(mem + off, pmem_addr, len);
103 flush_dcache_page(page);
107 * Note that we write the data both before and after
108 * clearing poison. The write before clear poison
109 * handles situations where the latest written data is
110 * preserved and the clear poison operation simply marks
111 * the address range as valid without changing the data.
112 * In this case application software can assume that an
113 * interrupted write will either return the new good
116 * However, if pmem_clear_poison() leaves the data in an
117 * indeterminate state we need to perform the write
118 * after clear poison.
120 flush_dcache_page(page);
121 memcpy_to_pmem(pmem_addr, mem + off, len);
122 if (unlikely(bad_pmem)) {
123 pmem_clear_poison(pmem, pmem_off, len);
124 memcpy_to_pmem(pmem_addr, mem + off, len);
132 static blk_qc_t pmem_make_request(struct request_queue *q, struct bio *bio)
138 struct bvec_iter iter;
139 struct block_device *bdev = bio->bi_bdev;
140 struct pmem_device *pmem = bdev->bd_disk->private_data;
142 do_acct = nd_iostat_start(bio, &start);
143 bio_for_each_segment(bvec, bio, iter) {
144 rc = pmem_do_bvec(pmem, bvec.bv_page, bvec.bv_len,
145 bvec.bv_offset, bio_data_dir(bio),
153 nd_iostat_end(bio, start);
155 if (bio_data_dir(bio))
159 return BLK_QC_T_NONE;
162 static int pmem_rw_page(struct block_device *bdev, sector_t sector,
163 struct page *page, int rw)
165 struct pmem_device *pmem = bdev->bd_disk->private_data;
168 rc = pmem_do_bvec(pmem, page, PAGE_SIZE, 0, rw, sector);
173 * The ->rw_page interface is subtle and tricky. The core
174 * retries on any error, so we can only invoke page_endio() in
175 * the successful completion case. Otherwise, we'll see crashes
176 * caused by double completion.
179 page_endio(page, rw & WRITE, 0);
184 static long pmem_direct_access(struct block_device *bdev, sector_t sector,
185 void __pmem **kaddr, pfn_t *pfn)
187 struct pmem_device *pmem = bdev->bd_disk->private_data;
188 resource_size_t offset = sector * 512 + pmem->data_offset;
190 *kaddr = pmem->virt_addr + offset;
191 *pfn = phys_to_pfn_t(pmem->phys_addr + offset, pmem->pfn_flags);
193 return pmem->size - pmem->pfn_pad - offset;
196 static const struct block_device_operations pmem_fops = {
197 .owner = THIS_MODULE,
198 .rw_page = pmem_rw_page,
199 .direct_access = pmem_direct_access,
200 .revalidate_disk = nvdimm_revalidate_disk,
203 static struct pmem_device *pmem_alloc(struct device *dev,
204 struct resource *res, int id)
206 struct pmem_device *pmem;
207 struct request_queue *q;
209 pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL);
211 return ERR_PTR(-ENOMEM);
213 pmem->phys_addr = res->start;
214 pmem->size = resource_size(res);
215 if (!arch_has_wmb_pmem())
216 dev_warn(dev, "unable to guarantee persistence of writes\n");
218 if (!devm_request_mem_region(dev, pmem->phys_addr, pmem->size,
220 dev_warn(dev, "could not reserve region [0x%pa:0x%zx]\n",
221 &pmem->phys_addr, pmem->size);
222 return ERR_PTR(-EBUSY);
225 q = blk_alloc_queue_node(GFP_KERNEL, dev_to_node(dev));
227 return ERR_PTR(-ENOMEM);
229 pmem->pfn_flags = PFN_DEV;
230 if (pmem_should_map_pages(dev)) {
231 pmem->virt_addr = (void __pmem *) devm_memremap_pages(dev, res,
232 &q->q_usage_counter, NULL);
233 pmem->pfn_flags |= PFN_MAP;
235 pmem->virt_addr = (void __pmem *) devm_memremap(dev,
236 pmem->phys_addr, pmem->size,
239 if (IS_ERR(pmem->virt_addr)) {
240 blk_cleanup_queue(q);
241 return (void __force *) pmem->virt_addr;
244 pmem->pmem_queue = q;
248 static void pmem_detach_disk(struct pmem_device *pmem)
250 if (!pmem->pmem_disk)
253 del_gendisk(pmem->pmem_disk);
254 put_disk(pmem->pmem_disk);
255 blk_cleanup_queue(pmem->pmem_queue);
258 static int pmem_attach_disk(struct device *dev,
259 struct nd_namespace_common *ndns, struct pmem_device *pmem)
261 struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
262 int nid = dev_to_node(dev);
263 struct resource bb_res;
264 struct gendisk *disk;
266 blk_queue_make_request(pmem->pmem_queue, pmem_make_request);
267 blk_queue_physical_block_size(pmem->pmem_queue, PAGE_SIZE);
268 blk_queue_max_hw_sectors(pmem->pmem_queue, UINT_MAX);
269 blk_queue_bounce_limit(pmem->pmem_queue, BLK_BOUNCE_ANY);
270 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, pmem->pmem_queue);
272 disk = alloc_disk_node(0, nid);
274 blk_cleanup_queue(pmem->pmem_queue);
278 disk->fops = &pmem_fops;
279 disk->private_data = pmem;
280 disk->queue = pmem->pmem_queue;
281 disk->flags = GENHD_FL_EXT_DEVT;
282 nvdimm_namespace_disk_name(ndns, disk->disk_name);
283 disk->driverfs_dev = dev;
284 set_capacity(disk, (pmem->size - pmem->pfn_pad - pmem->data_offset)
286 pmem->pmem_disk = disk;
287 devm_exit_badblocks(dev, &pmem->bb);
288 if (devm_init_badblocks(dev, &pmem->bb))
290 bb_res.start = nsio->res.start + pmem->data_offset;
291 bb_res.end = nsio->res.end;
292 if (is_nd_pfn(dev)) {
293 struct nd_pfn *nd_pfn = to_nd_pfn(dev);
294 struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
296 bb_res.start += __le32_to_cpu(pfn_sb->start_pad);
297 bb_res.end -= __le32_to_cpu(pfn_sb->end_trunc);
299 nvdimm_badblocks_populate(to_nd_region(dev->parent), &pmem->bb,
301 disk->bb = &pmem->bb;
303 revalidate_disk(disk);
308 static int pmem_rw_bytes(struct nd_namespace_common *ndns,
309 resource_size_t offset, void *buf, size_t size, int rw)
311 struct pmem_device *pmem = dev_get_drvdata(ndns->claim);
313 if (unlikely(offset + size > pmem->size)) {
314 dev_WARN_ONCE(&ndns->dev, 1, "request out of range\n");
319 unsigned int sz_align = ALIGN(size + (offset & (512 - 1)), 512);
321 if (unlikely(is_bad_pmem(&pmem->bb, offset / 512, sz_align)))
323 return memcpy_from_pmem(buf, pmem->virt_addr + offset, size);
325 memcpy_to_pmem(pmem->virt_addr + offset, buf, size);
332 static int nd_pfn_init(struct nd_pfn *nd_pfn)
334 struct nd_pfn_sb *pfn_sb = kzalloc(sizeof(*pfn_sb), GFP_KERNEL);
335 struct pmem_device *pmem = dev_get_drvdata(&nd_pfn->dev);
336 struct nd_namespace_common *ndns = nd_pfn->ndns;
337 u32 start_pad = 0, end_trunc = 0;
338 resource_size_t start, size;
339 struct nd_namespace_io *nsio;
340 struct nd_region *nd_region;
349 nd_pfn->pfn_sb = pfn_sb;
350 rc = nd_pfn_validate(nd_pfn);
352 /* no info block, do init */;
356 nd_region = to_nd_region(nd_pfn->dev.parent);
358 dev_info(&nd_pfn->dev,
359 "%s is read-only, unable to init metadata\n",
360 dev_name(&nd_region->dev));
364 memset(pfn_sb, 0, sizeof(*pfn_sb));
367 * Check if pmem collides with 'System RAM' when section aligned and
368 * trim it accordingly
370 nsio = to_nd_namespace_io(&ndns->dev);
371 start = PHYS_SECTION_ALIGN_DOWN(nsio->res.start);
372 size = resource_size(&nsio->res);
373 if (region_intersects(start, size, IORESOURCE_SYSTEM_RAM,
374 IORES_DESC_NONE) == REGION_MIXED) {
376 start = nsio->res.start;
377 start_pad = PHYS_SECTION_ALIGN_UP(start) - start;
380 start = nsio->res.start;
381 size = PHYS_SECTION_ALIGN_UP(start + size) - start;
382 if (region_intersects(start, size, IORESOURCE_SYSTEM_RAM,
383 IORES_DESC_NONE) == REGION_MIXED) {
384 size = resource_size(&nsio->res);
385 end_trunc = start + size - PHYS_SECTION_ALIGN_DOWN(start + size);
388 if (start_pad + end_trunc)
389 dev_info(&nd_pfn->dev, "%s section collision, truncate %d bytes\n",
390 dev_name(&ndns->dev), start_pad + end_trunc);
393 * Note, we use 64 here for the standard size of struct page,
394 * debugging options may cause it to be larger in which case the
395 * implementation will limit the pfns advertised through
396 * ->direct_access() to those that are included in the memmap.
399 npfns = (pmem->size - start_pad - end_trunc - SZ_8K) / SZ_4K;
400 if (nd_pfn->mode == PFN_MODE_PMEM) {
401 unsigned long memmap_size;
404 * vmemmap_populate_hugepages() allocates the memmap array in
407 memmap_size = ALIGN(64 * npfns, HPAGE_SIZE);
408 offset = ALIGN(start + SZ_8K + memmap_size, nd_pfn->align)
410 } else if (nd_pfn->mode == PFN_MODE_RAM)
411 offset = ALIGN(start + SZ_8K, nd_pfn->align) - start;
415 if (offset + start_pad + end_trunc >= pmem->size) {
416 dev_err(&nd_pfn->dev, "%s unable to satisfy requested alignment\n",
417 dev_name(&ndns->dev));
421 npfns = (pmem->size - offset - start_pad - end_trunc) / SZ_4K;
422 pfn_sb->mode = cpu_to_le32(nd_pfn->mode);
423 pfn_sb->dataoff = cpu_to_le64(offset);
424 pfn_sb->npfns = cpu_to_le64(npfns);
425 memcpy(pfn_sb->signature, PFN_SIG, PFN_SIG_LEN);
426 memcpy(pfn_sb->uuid, nd_pfn->uuid, 16);
427 memcpy(pfn_sb->parent_uuid, nd_dev_to_uuid(&ndns->dev), 16);
428 pfn_sb->version_major = cpu_to_le16(1);
429 pfn_sb->version_minor = cpu_to_le16(1);
430 pfn_sb->start_pad = cpu_to_le32(start_pad);
431 pfn_sb->end_trunc = cpu_to_le32(end_trunc);
432 checksum = nd_sb_checksum((struct nd_gen_sb *) pfn_sb);
433 pfn_sb->checksum = cpu_to_le64(checksum);
435 rc = nvdimm_write_bytes(ndns, SZ_4K, pfn_sb, sizeof(*pfn_sb));
441 nd_pfn->pfn_sb = NULL;
446 static int nvdimm_namespace_detach_pfn(struct nd_namespace_common *ndns)
448 struct nd_pfn *nd_pfn = to_nd_pfn(ndns->claim);
449 struct pmem_device *pmem;
452 pmem = dev_get_drvdata(&nd_pfn->dev);
453 pmem_detach_disk(pmem);
455 /* release nd_pfn resources */
456 kfree(nd_pfn->pfn_sb);
457 nd_pfn->pfn_sb = NULL;
463 * We hotplug memory at section granularity, pad the reserved area from
464 * the previous section base to the namespace base address.
466 static unsigned long init_altmap_base(resource_size_t base)
468 unsigned long base_pfn = PHYS_PFN(base);
470 return PFN_SECTION_ALIGN_DOWN(base_pfn);
473 static unsigned long init_altmap_reserve(resource_size_t base)
475 unsigned long reserve = PHYS_PFN(SZ_8K);
476 unsigned long base_pfn = PHYS_PFN(base);
478 reserve += base_pfn - PFN_SECTION_ALIGN_DOWN(base_pfn);
482 static int __nvdimm_namespace_attach_pfn(struct nd_pfn *nd_pfn)
486 struct request_queue *q;
487 struct pmem_device *pmem;
488 struct vmem_altmap *altmap;
489 struct device *dev = &nd_pfn->dev;
490 struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
491 struct nd_namespace_common *ndns = nd_pfn->ndns;
492 u32 start_pad = __le32_to_cpu(pfn_sb->start_pad);
493 u32 end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
494 struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
495 resource_size_t base = nsio->res.start + start_pad;
496 struct vmem_altmap __altmap = {
497 .base_pfn = init_altmap_base(base),
498 .reserve = init_altmap_reserve(base),
501 pmem = dev_get_drvdata(dev);
502 pmem->data_offset = le64_to_cpu(pfn_sb->dataoff);
503 pmem->pfn_pad = start_pad + end_trunc;
504 nd_pfn->mode = le32_to_cpu(nd_pfn->pfn_sb->mode);
505 if (nd_pfn->mode == PFN_MODE_RAM) {
506 if (pmem->data_offset < SZ_8K)
508 nd_pfn->npfns = le64_to_cpu(pfn_sb->npfns);
510 } else if (nd_pfn->mode == PFN_MODE_PMEM) {
511 nd_pfn->npfns = (pmem->size - pmem->pfn_pad - pmem->data_offset)
513 if (le64_to_cpu(nd_pfn->pfn_sb->npfns) > nd_pfn->npfns)
514 dev_info(&nd_pfn->dev,
515 "number of pfns truncated from %lld to %ld\n",
516 le64_to_cpu(nd_pfn->pfn_sb->npfns),
519 altmap->free = PHYS_PFN(pmem->data_offset - SZ_8K);
526 /* establish pfn range for lookup, and switch to direct map */
527 q = pmem->pmem_queue;
528 memcpy(&res, &nsio->res, sizeof(res));
529 res.start += start_pad;
530 res.end -= end_trunc;
531 devm_memunmap(dev, (void __force *) pmem->virt_addr);
532 pmem->virt_addr = (void __pmem *) devm_memremap_pages(dev, &res,
533 &q->q_usage_counter, altmap);
534 pmem->pfn_flags |= PFN_MAP;
535 if (IS_ERR(pmem->virt_addr)) {
536 rc = PTR_ERR(pmem->virt_addr);
540 /* attach pmem disk in "pfn-mode" */
541 rc = pmem_attach_disk(dev, ndns, pmem);
547 nvdimm_namespace_detach_pfn(ndns);
552 static int nvdimm_namespace_attach_pfn(struct nd_namespace_common *ndns)
554 struct nd_pfn *nd_pfn = to_nd_pfn(ndns->claim);
557 if (!nd_pfn->uuid || !nd_pfn->ndns)
560 rc = nd_pfn_init(nd_pfn);
563 /* we need a valid pfn_sb before we can init a vmem_altmap */
564 return __nvdimm_namespace_attach_pfn(nd_pfn);
567 static int nd_pmem_probe(struct device *dev)
569 struct nd_region *nd_region = to_nd_region(dev->parent);
570 struct nd_namespace_common *ndns;
571 struct nd_namespace_io *nsio;
572 struct pmem_device *pmem;
574 ndns = nvdimm_namespace_common_probe(dev);
576 return PTR_ERR(ndns);
578 nsio = to_nd_namespace_io(&ndns->dev);
579 pmem = pmem_alloc(dev, &nsio->res, nd_region->id);
581 return PTR_ERR(pmem);
584 dev_set_drvdata(dev, pmem);
585 ndns->rw_bytes = pmem_rw_bytes;
586 if (devm_init_badblocks(dev, &pmem->bb))
588 nvdimm_badblocks_populate(nd_region, &pmem->bb, &nsio->res);
590 if (is_nd_btt(dev)) {
591 /* btt allocates its own request_queue */
592 blk_cleanup_queue(pmem->pmem_queue);
593 pmem->pmem_queue = NULL;
594 return nvdimm_namespace_attach_btt(ndns);
598 return nvdimm_namespace_attach_pfn(ndns);
600 if (nd_btt_probe(ndns, pmem) == 0 || nd_pfn_probe(ndns, pmem) == 0) {
602 * We'll come back as either btt-pmem, or pfn-pmem, so
603 * drop the queue allocation for now.
605 blk_cleanup_queue(pmem->pmem_queue);
609 return pmem_attach_disk(dev, ndns, pmem);
612 static int nd_pmem_remove(struct device *dev)
614 struct pmem_device *pmem = dev_get_drvdata(dev);
617 nvdimm_namespace_detach_btt(pmem->ndns);
618 else if (is_nd_pfn(dev))
619 nvdimm_namespace_detach_pfn(pmem->ndns);
621 pmem_detach_disk(pmem);
626 static void nd_pmem_notify(struct device *dev, enum nvdimm_event event)
628 struct pmem_device *pmem = dev_get_drvdata(dev);
629 struct nd_namespace_common *ndns = pmem->ndns;
630 struct nd_region *nd_region = to_nd_region(dev->parent);
631 struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
632 struct resource res = {
633 .start = nsio->res.start + pmem->data_offset,
634 .end = nsio->res.end,
637 if (event != NVDIMM_REVALIDATE_POISON)
640 if (is_nd_pfn(dev)) {
641 struct nd_pfn *nd_pfn = to_nd_pfn(dev);
642 struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
644 res.start += __le32_to_cpu(pfn_sb->start_pad);
645 res.end -= __le32_to_cpu(pfn_sb->end_trunc);
648 nvdimm_badblocks_populate(nd_region, &pmem->bb, &res);
651 MODULE_ALIAS("pmem");
652 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO);
653 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM);
654 static struct nd_device_driver nd_pmem_driver = {
655 .probe = nd_pmem_probe,
656 .remove = nd_pmem_remove,
657 .notify = nd_pmem_notify,
661 .type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM,
664 static int __init pmem_init(void)
666 return nd_driver_register(&nd_pmem_driver);
668 module_init(pmem_init);
670 static void pmem_exit(void)
672 driver_unregister(&nd_pmem_driver.drv);
674 module_exit(pmem_exit);
676 MODULE_AUTHOR("Ross Zwisler <ross.zwisler@linux.intel.com>");
677 MODULE_LICENSE("GPL v2");