xref: /linux/drivers/nvdimm/pmem.c (revision 856e7c4b619af622d56b3b454f7bec32a170ac99)
1 /*
2  * Persistent Memory Driver
3  *
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>.
7  *
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.
11  *
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
15  * more details.
16  */
17 
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/blk-mq.h>
29 #include <linux/pfn_t.h>
30 #include <linux/slab.h>
31 #include <linux/uio.h>
32 #include <linux/dax.h>
33 #include <linux/nd.h>
34 #include <linux/backing-dev.h>
35 #include "pmem.h"
36 #include "pfn.h"
37 #include "nd.h"
38 #include "nd-core.h"
39 
40 static struct device *to_dev(struct pmem_device *pmem)
41 {
42 	/*
43 	 * nvdimm bus services need a 'dev' parameter, and we record the device
44 	 * at init in bb.dev.
45 	 */
46 	return pmem->bb.dev;
47 }
48 
49 static struct nd_region *to_region(struct pmem_device *pmem)
50 {
51 	return to_nd_region(to_dev(pmem)->parent);
52 }
53 
54 static blk_status_t pmem_clear_poison(struct pmem_device *pmem,
55 		phys_addr_t offset, unsigned int len)
56 {
57 	struct device *dev = to_dev(pmem);
58 	sector_t sector;
59 	long cleared;
60 	blk_status_t rc = BLK_STS_OK;
61 
62 	sector = (offset - pmem->data_offset) / 512;
63 
64 	cleared = nvdimm_clear_poison(dev, pmem->phys_addr + offset, len);
65 	if (cleared < len)
66 		rc = BLK_STS_IOERR;
67 	if (cleared > 0 && cleared / 512) {
68 		cleared /= 512;
69 		dev_dbg(dev, "%#llx clear %ld sector%s\n",
70 				(unsigned long long) sector, cleared,
71 				cleared > 1 ? "s" : "");
72 		badblocks_clear(&pmem->bb, sector, cleared);
73 		if (pmem->bb_state)
74 			sysfs_notify_dirent(pmem->bb_state);
75 	}
76 
77 	arch_invalidate_pmem(pmem->virt_addr + offset, len);
78 
79 	return rc;
80 }
81 
82 static void write_pmem(void *pmem_addr, struct page *page,
83 		unsigned int off, unsigned int len)
84 {
85 	unsigned int chunk;
86 	void *mem;
87 
88 	while (len) {
89 		mem = kmap_atomic(page);
90 		chunk = min_t(unsigned int, len, PAGE_SIZE);
91 		memcpy_flushcache(pmem_addr, mem + off, chunk);
92 		kunmap_atomic(mem);
93 		len -= chunk;
94 		off = 0;
95 		page++;
96 		pmem_addr += PAGE_SIZE;
97 	}
98 }
99 
100 static blk_status_t read_pmem(struct page *page, unsigned int off,
101 		void *pmem_addr, unsigned int len)
102 {
103 	unsigned int chunk;
104 	unsigned long rem;
105 	void *mem;
106 
107 	while (len) {
108 		mem = kmap_atomic(page);
109 		chunk = min_t(unsigned int, len, PAGE_SIZE);
110 		rem = memcpy_mcsafe(mem + off, pmem_addr, chunk);
111 		kunmap_atomic(mem);
112 		if (rem)
113 			return BLK_STS_IOERR;
114 		len -= chunk;
115 		off = 0;
116 		page++;
117 		pmem_addr += PAGE_SIZE;
118 	}
119 	return BLK_STS_OK;
120 }
121 
122 static blk_status_t pmem_do_bvec(struct pmem_device *pmem, struct page *page,
123 			unsigned int len, unsigned int off, bool is_write,
124 			sector_t sector)
125 {
126 	blk_status_t rc = BLK_STS_OK;
127 	bool bad_pmem = false;
128 	phys_addr_t pmem_off = sector * 512 + pmem->data_offset;
129 	void *pmem_addr = pmem->virt_addr + pmem_off;
130 
131 	if (unlikely(is_bad_pmem(&pmem->bb, sector, len)))
132 		bad_pmem = true;
133 
134 	if (!is_write) {
135 		if (unlikely(bad_pmem))
136 			rc = BLK_STS_IOERR;
137 		else {
138 			rc = read_pmem(page, off, pmem_addr, len);
139 			flush_dcache_page(page);
140 		}
141 	} else {
142 		/*
143 		 * Note that we write the data both before and after
144 		 * clearing poison.  The write before clear poison
145 		 * handles situations where the latest written data is
146 		 * preserved and the clear poison operation simply marks
147 		 * the address range as valid without changing the data.
148 		 * In this case application software can assume that an
149 		 * interrupted write will either return the new good
150 		 * data or an error.
151 		 *
152 		 * However, if pmem_clear_poison() leaves the data in an
153 		 * indeterminate state we need to perform the write
154 		 * after clear poison.
155 		 */
156 		flush_dcache_page(page);
157 		write_pmem(pmem_addr, page, off, len);
158 		if (unlikely(bad_pmem)) {
159 			rc = pmem_clear_poison(pmem, pmem_off, len);
160 			write_pmem(pmem_addr, page, off, len);
161 		}
162 	}
163 
164 	return rc;
165 }
166 
167 static blk_qc_t pmem_make_request(struct request_queue *q, struct bio *bio)
168 {
169 	blk_status_t rc = 0;
170 	bool do_acct;
171 	unsigned long start;
172 	struct bio_vec bvec;
173 	struct bvec_iter iter;
174 	struct pmem_device *pmem = q->queuedata;
175 	struct nd_region *nd_region = to_region(pmem);
176 
177 	if (bio->bi_opf & REQ_PREFLUSH)
178 		nvdimm_flush(nd_region);
179 
180 	do_acct = nd_iostat_start(bio, &start);
181 	bio_for_each_segment(bvec, bio, iter) {
182 		rc = pmem_do_bvec(pmem, bvec.bv_page, bvec.bv_len,
183 				bvec.bv_offset, op_is_write(bio_op(bio)),
184 				iter.bi_sector);
185 		if (rc) {
186 			bio->bi_status = rc;
187 			break;
188 		}
189 	}
190 	if (do_acct)
191 		nd_iostat_end(bio, start);
192 
193 	if (bio->bi_opf & REQ_FUA)
194 		nvdimm_flush(nd_region);
195 
196 	bio_endio(bio);
197 	return BLK_QC_T_NONE;
198 }
199 
200 static int pmem_rw_page(struct block_device *bdev, sector_t sector,
201 		       struct page *page, bool is_write)
202 {
203 	struct pmem_device *pmem = bdev->bd_queue->queuedata;
204 	blk_status_t rc;
205 
206 	rc = pmem_do_bvec(pmem, page, hpage_nr_pages(page) * PAGE_SIZE,
207 			  0, is_write, sector);
208 
209 	/*
210 	 * The ->rw_page interface is subtle and tricky.  The core
211 	 * retries on any error, so we can only invoke page_endio() in
212 	 * the successful completion case.  Otherwise, we'll see crashes
213 	 * caused by double completion.
214 	 */
215 	if (rc == 0)
216 		page_endio(page, is_write, 0);
217 
218 	return blk_status_to_errno(rc);
219 }
220 
221 /* see "strong" declaration in tools/testing/nvdimm/pmem-dax.c */
222 __weak long __pmem_direct_access(struct pmem_device *pmem, pgoff_t pgoff,
223 		long nr_pages, void **kaddr, pfn_t *pfn)
224 {
225 	resource_size_t offset = PFN_PHYS(pgoff) + pmem->data_offset;
226 
227 	if (unlikely(is_bad_pmem(&pmem->bb, PFN_PHYS(pgoff) / 512,
228 					PFN_PHYS(nr_pages))))
229 		return -EIO;
230 	*kaddr = pmem->virt_addr + offset;
231 	*pfn = phys_to_pfn_t(pmem->phys_addr + offset, pmem->pfn_flags);
232 
233 	/*
234 	 * If badblocks are present, limit known good range to the
235 	 * requested range.
236 	 */
237 	if (unlikely(pmem->bb.count))
238 		return nr_pages;
239 	return PHYS_PFN(pmem->size - pmem->pfn_pad - offset);
240 }
241 
242 static const struct block_device_operations pmem_fops = {
243 	.owner =		THIS_MODULE,
244 	.rw_page =		pmem_rw_page,
245 	.revalidate_disk =	nvdimm_revalidate_disk,
246 };
247 
248 static long pmem_dax_direct_access(struct dax_device *dax_dev,
249 		pgoff_t pgoff, long nr_pages, void **kaddr, pfn_t *pfn)
250 {
251 	struct pmem_device *pmem = dax_get_private(dax_dev);
252 
253 	return __pmem_direct_access(pmem, pgoff, nr_pages, kaddr, pfn);
254 }
255 
256 static size_t pmem_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff,
257 		void *addr, size_t bytes, struct iov_iter *i)
258 {
259 	return copy_from_iter_flushcache(addr, bytes, i);
260 }
261 
262 static size_t pmem_copy_to_iter(struct dax_device *dax_dev, pgoff_t pgoff,
263 		void *addr, size_t bytes, struct iov_iter *i)
264 {
265 	return copy_to_iter_mcsafe(addr, bytes, i);
266 }
267 
268 static const struct dax_operations pmem_dax_ops = {
269 	.direct_access = pmem_dax_direct_access,
270 	.copy_from_iter = pmem_copy_from_iter,
271 	.copy_to_iter = pmem_copy_to_iter,
272 };
273 
274 static const struct attribute_group *pmem_attribute_groups[] = {
275 	&dax_attribute_group,
276 	NULL,
277 };
278 
279 static void pmem_release_queue(void *q)
280 {
281 	blk_cleanup_queue(q);
282 }
283 
284 static void pmem_freeze_queue(void *q)
285 {
286 	blk_freeze_queue_start(q);
287 }
288 
289 static void pmem_release_disk(void *__pmem)
290 {
291 	struct pmem_device *pmem = __pmem;
292 
293 	kill_dax(pmem->dax_dev);
294 	put_dax(pmem->dax_dev);
295 	del_gendisk(pmem->disk);
296 	put_disk(pmem->disk);
297 }
298 
299 static void pmem_release_pgmap_ops(void *__pgmap)
300 {
301 	dev_pagemap_put_ops();
302 }
303 
304 static void fsdax_pagefree(struct page *page, void *data)
305 {
306 	wake_up_var(&page->_refcount);
307 }
308 
309 static int setup_pagemap_fsdax(struct device *dev, struct dev_pagemap *pgmap)
310 {
311 	dev_pagemap_get_ops();
312 	if (devm_add_action_or_reset(dev, pmem_release_pgmap_ops, pgmap))
313 		return -ENOMEM;
314 	pgmap->type = MEMORY_DEVICE_FS_DAX;
315 	pgmap->page_free = fsdax_pagefree;
316 
317 	return 0;
318 }
319 
320 static int pmem_attach_disk(struct device *dev,
321 		struct nd_namespace_common *ndns)
322 {
323 	struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
324 	struct nd_region *nd_region = to_nd_region(dev->parent);
325 	int nid = dev_to_node(dev), fua;
326 	struct resource *res = &nsio->res;
327 	struct resource bb_res;
328 	struct nd_pfn *nd_pfn = NULL;
329 	struct dax_device *dax_dev;
330 	struct nd_pfn_sb *pfn_sb;
331 	struct pmem_device *pmem;
332 	struct request_queue *q;
333 	struct device *gendev;
334 	struct gendisk *disk;
335 	void *addr;
336 	int rc;
337 
338 	pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL);
339 	if (!pmem)
340 		return -ENOMEM;
341 
342 	/* while nsio_rw_bytes is active, parse a pfn info block if present */
343 	if (is_nd_pfn(dev)) {
344 		nd_pfn = to_nd_pfn(dev);
345 		rc = nvdimm_setup_pfn(nd_pfn, &pmem->pgmap);
346 		if (rc)
347 			return rc;
348 	}
349 
350 	/* we're attaching a block device, disable raw namespace access */
351 	devm_nsio_disable(dev, nsio);
352 
353 	dev_set_drvdata(dev, pmem);
354 	pmem->phys_addr = res->start;
355 	pmem->size = resource_size(res);
356 	fua = nvdimm_has_flush(nd_region);
357 	if (!IS_ENABLED(CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE) || fua < 0) {
358 		dev_warn(dev, "unable to guarantee persistence of writes\n");
359 		fua = 0;
360 	}
361 
362 	if (!devm_request_mem_region(dev, res->start, resource_size(res),
363 				dev_name(&ndns->dev))) {
364 		dev_warn(dev, "could not reserve region %pR\n", res);
365 		return -EBUSY;
366 	}
367 
368 	q = blk_alloc_queue_node(GFP_KERNEL, dev_to_node(dev), NULL);
369 	if (!q)
370 		return -ENOMEM;
371 
372 	if (devm_add_action_or_reset(dev, pmem_release_queue, q))
373 		return -ENOMEM;
374 
375 	pmem->pfn_flags = PFN_DEV;
376 	pmem->pgmap.ref = &q->q_usage_counter;
377 	if (is_nd_pfn(dev)) {
378 		if (setup_pagemap_fsdax(dev, &pmem->pgmap))
379 			return -ENOMEM;
380 		addr = devm_memremap_pages(dev, &pmem->pgmap);
381 		pfn_sb = nd_pfn->pfn_sb;
382 		pmem->data_offset = le64_to_cpu(pfn_sb->dataoff);
383 		pmem->pfn_pad = resource_size(res) -
384 			resource_size(&pmem->pgmap.res);
385 		pmem->pfn_flags |= PFN_MAP;
386 		memcpy(&bb_res, &pmem->pgmap.res, sizeof(bb_res));
387 		bb_res.start += pmem->data_offset;
388 	} else if (pmem_should_map_pages(dev)) {
389 		memcpy(&pmem->pgmap.res, &nsio->res, sizeof(pmem->pgmap.res));
390 		pmem->pgmap.altmap_valid = false;
391 		if (setup_pagemap_fsdax(dev, &pmem->pgmap))
392 			return -ENOMEM;
393 		addr = devm_memremap_pages(dev, &pmem->pgmap);
394 		pmem->pfn_flags |= PFN_MAP;
395 		memcpy(&bb_res, &pmem->pgmap.res, sizeof(bb_res));
396 	} else
397 		addr = devm_memremap(dev, pmem->phys_addr,
398 				pmem->size, ARCH_MEMREMAP_PMEM);
399 
400 	/*
401 	 * At release time the queue must be frozen before
402 	 * devm_memremap_pages is unwound
403 	 */
404 	if (devm_add_action_or_reset(dev, pmem_freeze_queue, q))
405 		return -ENOMEM;
406 
407 	if (IS_ERR(addr))
408 		return PTR_ERR(addr);
409 	pmem->virt_addr = addr;
410 
411 	blk_queue_write_cache(q, true, fua);
412 	blk_queue_make_request(q, pmem_make_request);
413 	blk_queue_physical_block_size(q, PAGE_SIZE);
414 	blk_queue_logical_block_size(q, pmem_sector_size(ndns));
415 	blk_queue_max_hw_sectors(q, UINT_MAX);
416 	blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
417 	blk_queue_flag_set(QUEUE_FLAG_DAX, q);
418 	q->queuedata = pmem;
419 
420 	disk = alloc_disk_node(0, nid);
421 	if (!disk)
422 		return -ENOMEM;
423 	pmem->disk = disk;
424 
425 	disk->fops		= &pmem_fops;
426 	disk->queue		= q;
427 	disk->flags		= GENHD_FL_EXT_DEVT;
428 	disk->queue->backing_dev_info->capabilities |= BDI_CAP_SYNCHRONOUS_IO;
429 	nvdimm_namespace_disk_name(ndns, disk->disk_name);
430 	set_capacity(disk, (pmem->size - pmem->pfn_pad - pmem->data_offset)
431 			/ 512);
432 	if (devm_init_badblocks(dev, &pmem->bb))
433 		return -ENOMEM;
434 	nvdimm_badblocks_populate(nd_region, &pmem->bb, &bb_res);
435 	disk->bb = &pmem->bb;
436 
437 	dax_dev = alloc_dax(pmem, disk->disk_name, &pmem_dax_ops);
438 	if (!dax_dev) {
439 		put_disk(disk);
440 		return -ENOMEM;
441 	}
442 	dax_write_cache(dax_dev, nvdimm_has_cache(nd_region));
443 	pmem->dax_dev = dax_dev;
444 
445 	gendev = disk_to_dev(disk);
446 	gendev->groups = pmem_attribute_groups;
447 
448 	device_add_disk(dev, disk);
449 	if (devm_add_action_or_reset(dev, pmem_release_disk, pmem))
450 		return -ENOMEM;
451 
452 	revalidate_disk(disk);
453 
454 	pmem->bb_state = sysfs_get_dirent(disk_to_dev(disk)->kobj.sd,
455 					  "badblocks");
456 	if (!pmem->bb_state)
457 		dev_warn(dev, "'badblocks' notification disabled\n");
458 
459 	return 0;
460 }
461 
462 static int nd_pmem_probe(struct device *dev)
463 {
464 	struct nd_namespace_common *ndns;
465 
466 	ndns = nvdimm_namespace_common_probe(dev);
467 	if (IS_ERR(ndns))
468 		return PTR_ERR(ndns);
469 
470 	if (devm_nsio_enable(dev, to_nd_namespace_io(&ndns->dev)))
471 		return -ENXIO;
472 
473 	if (is_nd_btt(dev))
474 		return nvdimm_namespace_attach_btt(ndns);
475 
476 	if (is_nd_pfn(dev))
477 		return pmem_attach_disk(dev, ndns);
478 
479 	/* if we find a valid info-block we'll come back as that personality */
480 	if (nd_btt_probe(dev, ndns) == 0 || nd_pfn_probe(dev, ndns) == 0
481 			|| nd_dax_probe(dev, ndns) == 0)
482 		return -ENXIO;
483 
484 	/* ...otherwise we're just a raw pmem device */
485 	return pmem_attach_disk(dev, ndns);
486 }
487 
488 static int nd_pmem_remove(struct device *dev)
489 {
490 	struct pmem_device *pmem = dev_get_drvdata(dev);
491 
492 	if (is_nd_btt(dev))
493 		nvdimm_namespace_detach_btt(to_nd_btt(dev));
494 	else {
495 		/*
496 		 * Note, this assumes device_lock() context to not race
497 		 * nd_pmem_notify()
498 		 */
499 		sysfs_put(pmem->bb_state);
500 		pmem->bb_state = NULL;
501 	}
502 	nvdimm_flush(to_nd_region(dev->parent));
503 
504 	return 0;
505 }
506 
507 static void nd_pmem_shutdown(struct device *dev)
508 {
509 	nvdimm_flush(to_nd_region(dev->parent));
510 }
511 
512 static void nd_pmem_notify(struct device *dev, enum nvdimm_event event)
513 {
514 	struct nd_region *nd_region;
515 	resource_size_t offset = 0, end_trunc = 0;
516 	struct nd_namespace_common *ndns;
517 	struct nd_namespace_io *nsio;
518 	struct resource res;
519 	struct badblocks *bb;
520 	struct kernfs_node *bb_state;
521 
522 	if (event != NVDIMM_REVALIDATE_POISON)
523 		return;
524 
525 	if (is_nd_btt(dev)) {
526 		struct nd_btt *nd_btt = to_nd_btt(dev);
527 
528 		ndns = nd_btt->ndns;
529 		nd_region = to_nd_region(ndns->dev.parent);
530 		nsio = to_nd_namespace_io(&ndns->dev);
531 		bb = &nsio->bb;
532 		bb_state = NULL;
533 	} else {
534 		struct pmem_device *pmem = dev_get_drvdata(dev);
535 
536 		nd_region = to_region(pmem);
537 		bb = &pmem->bb;
538 		bb_state = pmem->bb_state;
539 
540 		if (is_nd_pfn(dev)) {
541 			struct nd_pfn *nd_pfn = to_nd_pfn(dev);
542 			struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
543 
544 			ndns = nd_pfn->ndns;
545 			offset = pmem->data_offset +
546 					__le32_to_cpu(pfn_sb->start_pad);
547 			end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
548 		} else {
549 			ndns = to_ndns(dev);
550 		}
551 
552 		nsio = to_nd_namespace_io(&ndns->dev);
553 	}
554 
555 	res.start = nsio->res.start + offset;
556 	res.end = nsio->res.end - end_trunc;
557 	nvdimm_badblocks_populate(nd_region, bb, &res);
558 	if (bb_state)
559 		sysfs_notify_dirent(bb_state);
560 }
561 
562 MODULE_ALIAS("pmem");
563 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO);
564 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM);
565 static struct nd_device_driver nd_pmem_driver = {
566 	.probe = nd_pmem_probe,
567 	.remove = nd_pmem_remove,
568 	.notify = nd_pmem_notify,
569 	.shutdown = nd_pmem_shutdown,
570 	.drv = {
571 		.name = "nd_pmem",
572 	},
573 	.type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM,
574 };
575 
576 module_nd_driver(nd_pmem_driver);
577 
578 MODULE_AUTHOR("Ross Zwisler <ross.zwisler@linux.intel.com>");
579 MODULE_LICENSE("GPL v2");
580