xref: /openbmc/linux/drivers/nvdimm/pmem.c (revision 0c874100)
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/set_memory.h>
24 #include <linux/module.h>
25 #include <linux/moduleparam.h>
26 #include <linux/badblocks.h>
27 #include <linux/memremap.h>
28 #include <linux/vmalloc.h>
29 #include <linux/blk-mq.h>
30 #include <linux/pfn_t.h>
31 #include <linux/slab.h>
32 #include <linux/uio.h>
33 #include <linux/dax.h>
34 #include <linux/nd.h>
35 #include <linux/backing-dev.h>
36 #include "pmem.h"
37 #include "pfn.h"
38 #include "nd.h"
39 #include "nd-core.h"
40 
41 static struct device *to_dev(struct pmem_device *pmem)
42 {
43 	/*
44 	 * nvdimm bus services need a 'dev' parameter, and we record the device
45 	 * at init in bb.dev.
46 	 */
47 	return pmem->bb.dev;
48 }
49 
50 static struct nd_region *to_region(struct pmem_device *pmem)
51 {
52 	return to_nd_region(to_dev(pmem)->parent);
53 }
54 
55 static void hwpoison_clear(struct pmem_device *pmem,
56 		phys_addr_t phys, unsigned int len)
57 {
58 	unsigned long pfn_start, pfn_end, pfn;
59 
60 	/* only pmem in the linear map supports HWPoison */
61 	if (is_vmalloc_addr(pmem->virt_addr))
62 		return;
63 
64 	pfn_start = PHYS_PFN(phys);
65 	pfn_end = pfn_start + PHYS_PFN(len);
66 	for (pfn = pfn_start; pfn < pfn_end; pfn++) {
67 		struct page *page = pfn_to_page(pfn);
68 
69 		/*
70 		 * Note, no need to hold a get_dev_pagemap() reference
71 		 * here since we're in the driver I/O path and
72 		 * outstanding I/O requests pin the dev_pagemap.
73 		 */
74 		if (test_and_clear_pmem_poison(page))
75 			clear_mce_nospec(pfn);
76 	}
77 }
78 
79 static blk_status_t pmem_clear_poison(struct pmem_device *pmem,
80 		phys_addr_t offset, unsigned int len)
81 {
82 	struct device *dev = to_dev(pmem);
83 	sector_t sector;
84 	long cleared;
85 	blk_status_t rc = BLK_STS_OK;
86 
87 	sector = (offset - pmem->data_offset) / 512;
88 
89 	cleared = nvdimm_clear_poison(dev, pmem->phys_addr + offset, len);
90 	if (cleared < len)
91 		rc = BLK_STS_IOERR;
92 	if (cleared > 0 && cleared / 512) {
93 		hwpoison_clear(pmem, pmem->phys_addr + offset, cleared);
94 		cleared /= 512;
95 		dev_dbg(dev, "%#llx clear %ld sector%s\n",
96 				(unsigned long long) sector, cleared,
97 				cleared > 1 ? "s" : "");
98 		badblocks_clear(&pmem->bb, sector, cleared);
99 		if (pmem->bb_state)
100 			sysfs_notify_dirent(pmem->bb_state);
101 	}
102 
103 	arch_invalidate_pmem(pmem->virt_addr + offset, len);
104 
105 	return rc;
106 }
107 
108 static void write_pmem(void *pmem_addr, struct page *page,
109 		unsigned int off, unsigned int len)
110 {
111 	unsigned int chunk;
112 	void *mem;
113 
114 	while (len) {
115 		mem = kmap_atomic(page);
116 		chunk = min_t(unsigned int, len, PAGE_SIZE);
117 		memcpy_flushcache(pmem_addr, mem + off, chunk);
118 		kunmap_atomic(mem);
119 		len -= chunk;
120 		off = 0;
121 		page++;
122 		pmem_addr += PAGE_SIZE;
123 	}
124 }
125 
126 static blk_status_t read_pmem(struct page *page, unsigned int off,
127 		void *pmem_addr, unsigned int len)
128 {
129 	unsigned int chunk;
130 	unsigned long rem;
131 	void *mem;
132 
133 	while (len) {
134 		mem = kmap_atomic(page);
135 		chunk = min_t(unsigned int, len, PAGE_SIZE);
136 		rem = memcpy_mcsafe(mem + off, pmem_addr, chunk);
137 		kunmap_atomic(mem);
138 		if (rem)
139 			return BLK_STS_IOERR;
140 		len -= chunk;
141 		off = 0;
142 		page++;
143 		pmem_addr += PAGE_SIZE;
144 	}
145 	return BLK_STS_OK;
146 }
147 
148 static blk_status_t pmem_do_bvec(struct pmem_device *pmem, struct page *page,
149 			unsigned int len, unsigned int off, unsigned int op,
150 			sector_t sector)
151 {
152 	blk_status_t rc = BLK_STS_OK;
153 	bool bad_pmem = false;
154 	phys_addr_t pmem_off = sector * 512 + pmem->data_offset;
155 	void *pmem_addr = pmem->virt_addr + pmem_off;
156 
157 	if (unlikely(is_bad_pmem(&pmem->bb, sector, len)))
158 		bad_pmem = true;
159 
160 	if (!op_is_write(op)) {
161 		if (unlikely(bad_pmem))
162 			rc = BLK_STS_IOERR;
163 		else {
164 			rc = read_pmem(page, off, pmem_addr, len);
165 			flush_dcache_page(page);
166 		}
167 	} else {
168 		/*
169 		 * Note that we write the data both before and after
170 		 * clearing poison.  The write before clear poison
171 		 * handles situations where the latest written data is
172 		 * preserved and the clear poison operation simply marks
173 		 * the address range as valid without changing the data.
174 		 * In this case application software can assume that an
175 		 * interrupted write will either return the new good
176 		 * data or an error.
177 		 *
178 		 * However, if pmem_clear_poison() leaves the data in an
179 		 * indeterminate state we need to perform the write
180 		 * after clear poison.
181 		 */
182 		flush_dcache_page(page);
183 		write_pmem(pmem_addr, page, off, len);
184 		if (unlikely(bad_pmem)) {
185 			rc = pmem_clear_poison(pmem, pmem_off, len);
186 			write_pmem(pmem_addr, page, off, len);
187 		}
188 	}
189 
190 	return rc;
191 }
192 
193 static blk_qc_t pmem_make_request(struct request_queue *q, struct bio *bio)
194 {
195 	blk_status_t rc = 0;
196 	bool do_acct;
197 	unsigned long start;
198 	struct bio_vec bvec;
199 	struct bvec_iter iter;
200 	struct pmem_device *pmem = q->queuedata;
201 	struct nd_region *nd_region = to_region(pmem);
202 
203 	if (bio->bi_opf & REQ_PREFLUSH)
204 		nvdimm_flush(nd_region);
205 
206 	do_acct = nd_iostat_start(bio, &start);
207 	bio_for_each_segment(bvec, bio, iter) {
208 		rc = pmem_do_bvec(pmem, bvec.bv_page, bvec.bv_len,
209 				bvec.bv_offset, bio_op(bio), iter.bi_sector);
210 		if (rc) {
211 			bio->bi_status = rc;
212 			break;
213 		}
214 	}
215 	if (do_acct)
216 		nd_iostat_end(bio, start);
217 
218 	if (bio->bi_opf & REQ_FUA)
219 		nvdimm_flush(nd_region);
220 
221 	bio_endio(bio);
222 	return BLK_QC_T_NONE;
223 }
224 
225 static int pmem_rw_page(struct block_device *bdev, sector_t sector,
226 		       struct page *page, unsigned int op)
227 {
228 	struct pmem_device *pmem = bdev->bd_queue->queuedata;
229 	blk_status_t rc;
230 
231 	rc = pmem_do_bvec(pmem, page, hpage_nr_pages(page) * PAGE_SIZE,
232 			  0, op, sector);
233 
234 	/*
235 	 * The ->rw_page interface is subtle and tricky.  The core
236 	 * retries on any error, so we can only invoke page_endio() in
237 	 * the successful completion case.  Otherwise, we'll see crashes
238 	 * caused by double completion.
239 	 */
240 	if (rc == 0)
241 		page_endio(page, op_is_write(op), 0);
242 
243 	return blk_status_to_errno(rc);
244 }
245 
246 /* see "strong" declaration in tools/testing/nvdimm/pmem-dax.c */
247 __weak long __pmem_direct_access(struct pmem_device *pmem, pgoff_t pgoff,
248 		long nr_pages, void **kaddr, pfn_t *pfn)
249 {
250 	resource_size_t offset = PFN_PHYS(pgoff) + pmem->data_offset;
251 
252 	if (unlikely(is_bad_pmem(&pmem->bb, PFN_PHYS(pgoff) / 512,
253 					PFN_PHYS(nr_pages))))
254 		return -EIO;
255 
256 	if (kaddr)
257 		*kaddr = pmem->virt_addr + offset;
258 	if (pfn)
259 		*pfn = phys_to_pfn_t(pmem->phys_addr + offset, pmem->pfn_flags);
260 
261 	/*
262 	 * If badblocks are present, limit known good range to the
263 	 * requested range.
264 	 */
265 	if (unlikely(pmem->bb.count))
266 		return nr_pages;
267 	return PHYS_PFN(pmem->size - pmem->pfn_pad - offset);
268 }
269 
270 static const struct block_device_operations pmem_fops = {
271 	.owner =		THIS_MODULE,
272 	.rw_page =		pmem_rw_page,
273 	.revalidate_disk =	nvdimm_revalidate_disk,
274 };
275 
276 static long pmem_dax_direct_access(struct dax_device *dax_dev,
277 		pgoff_t pgoff, long nr_pages, void **kaddr, pfn_t *pfn)
278 {
279 	struct pmem_device *pmem = dax_get_private(dax_dev);
280 
281 	return __pmem_direct_access(pmem, pgoff, nr_pages, kaddr, pfn);
282 }
283 
284 static size_t pmem_copy_from_iter(struct dax_device *dax_dev, pgoff_t pgoff,
285 		void *addr, size_t bytes, struct iov_iter *i)
286 {
287 	return copy_from_iter_flushcache(addr, bytes, i);
288 }
289 
290 static size_t pmem_copy_to_iter(struct dax_device *dax_dev, pgoff_t pgoff,
291 		void *addr, size_t bytes, struct iov_iter *i)
292 {
293 	return copy_to_iter_mcsafe(addr, bytes, i);
294 }
295 
296 static const struct dax_operations pmem_dax_ops = {
297 	.direct_access = pmem_dax_direct_access,
298 	.copy_from_iter = pmem_copy_from_iter,
299 	.copy_to_iter = pmem_copy_to_iter,
300 };
301 
302 static const struct attribute_group *pmem_attribute_groups[] = {
303 	&dax_attribute_group,
304 	NULL,
305 };
306 
307 static void pmem_release_queue(void *q)
308 {
309 	blk_cleanup_queue(q);
310 }
311 
312 static void pmem_freeze_queue(void *q)
313 {
314 	blk_freeze_queue_start(q);
315 }
316 
317 static void pmem_release_disk(void *__pmem)
318 {
319 	struct pmem_device *pmem = __pmem;
320 
321 	kill_dax(pmem->dax_dev);
322 	put_dax(pmem->dax_dev);
323 	del_gendisk(pmem->disk);
324 	put_disk(pmem->disk);
325 }
326 
327 static void pmem_release_pgmap_ops(void *__pgmap)
328 {
329 	dev_pagemap_put_ops();
330 }
331 
332 static void fsdax_pagefree(struct page *page, void *data)
333 {
334 	wake_up_var(&page->_refcount);
335 }
336 
337 static int setup_pagemap_fsdax(struct device *dev, struct dev_pagemap *pgmap)
338 {
339 	dev_pagemap_get_ops();
340 	if (devm_add_action_or_reset(dev, pmem_release_pgmap_ops, pgmap))
341 		return -ENOMEM;
342 	pgmap->type = MEMORY_DEVICE_FS_DAX;
343 	pgmap->page_free = fsdax_pagefree;
344 
345 	return 0;
346 }
347 
348 static int pmem_attach_disk(struct device *dev,
349 		struct nd_namespace_common *ndns)
350 {
351 	struct nd_namespace_io *nsio = to_nd_namespace_io(&ndns->dev);
352 	struct nd_region *nd_region = to_nd_region(dev->parent);
353 	int nid = dev_to_node(dev), fua;
354 	struct resource *res = &nsio->res;
355 	struct resource bb_res;
356 	struct nd_pfn *nd_pfn = NULL;
357 	struct dax_device *dax_dev;
358 	struct nd_pfn_sb *pfn_sb;
359 	struct pmem_device *pmem;
360 	struct request_queue *q;
361 	struct device *gendev;
362 	struct gendisk *disk;
363 	void *addr;
364 	int rc;
365 
366 	pmem = devm_kzalloc(dev, sizeof(*pmem), GFP_KERNEL);
367 	if (!pmem)
368 		return -ENOMEM;
369 
370 	/* while nsio_rw_bytes is active, parse a pfn info block if present */
371 	if (is_nd_pfn(dev)) {
372 		nd_pfn = to_nd_pfn(dev);
373 		rc = nvdimm_setup_pfn(nd_pfn, &pmem->pgmap);
374 		if (rc)
375 			return rc;
376 	}
377 
378 	/* we're attaching a block device, disable raw namespace access */
379 	devm_nsio_disable(dev, nsio);
380 
381 	dev_set_drvdata(dev, pmem);
382 	pmem->phys_addr = res->start;
383 	pmem->size = resource_size(res);
384 	fua = nvdimm_has_flush(nd_region);
385 	if (!IS_ENABLED(CONFIG_ARCH_HAS_UACCESS_FLUSHCACHE) || fua < 0) {
386 		dev_warn(dev, "unable to guarantee persistence of writes\n");
387 		fua = 0;
388 	}
389 
390 	if (!devm_request_mem_region(dev, res->start, resource_size(res),
391 				dev_name(&ndns->dev))) {
392 		dev_warn(dev, "could not reserve region %pR\n", res);
393 		return -EBUSY;
394 	}
395 
396 	q = blk_alloc_queue_node(GFP_KERNEL, dev_to_node(dev), NULL);
397 	if (!q)
398 		return -ENOMEM;
399 
400 	if (devm_add_action_or_reset(dev, pmem_release_queue, q))
401 		return -ENOMEM;
402 
403 	pmem->pfn_flags = PFN_DEV;
404 	pmem->pgmap.ref = &q->q_usage_counter;
405 	if (is_nd_pfn(dev)) {
406 		if (setup_pagemap_fsdax(dev, &pmem->pgmap))
407 			return -ENOMEM;
408 		addr = devm_memremap_pages(dev, &pmem->pgmap);
409 		pfn_sb = nd_pfn->pfn_sb;
410 		pmem->data_offset = le64_to_cpu(pfn_sb->dataoff);
411 		pmem->pfn_pad = resource_size(res) -
412 			resource_size(&pmem->pgmap.res);
413 		pmem->pfn_flags |= PFN_MAP;
414 		memcpy(&bb_res, &pmem->pgmap.res, sizeof(bb_res));
415 		bb_res.start += pmem->data_offset;
416 	} else if (pmem_should_map_pages(dev)) {
417 		memcpy(&pmem->pgmap.res, &nsio->res, sizeof(pmem->pgmap.res));
418 		pmem->pgmap.altmap_valid = false;
419 		if (setup_pagemap_fsdax(dev, &pmem->pgmap))
420 			return -ENOMEM;
421 		addr = devm_memremap_pages(dev, &pmem->pgmap);
422 		pmem->pfn_flags |= PFN_MAP;
423 		memcpy(&bb_res, &pmem->pgmap.res, sizeof(bb_res));
424 	} else {
425 		addr = devm_memremap(dev, pmem->phys_addr,
426 				pmem->size, ARCH_MEMREMAP_PMEM);
427 		memcpy(&bb_res, &nsio->res, sizeof(bb_res));
428 	}
429 
430 	/*
431 	 * At release time the queue must be frozen before
432 	 * devm_memremap_pages is unwound
433 	 */
434 	if (devm_add_action_or_reset(dev, pmem_freeze_queue, q))
435 		return -ENOMEM;
436 
437 	if (IS_ERR(addr))
438 		return PTR_ERR(addr);
439 	pmem->virt_addr = addr;
440 
441 	blk_queue_write_cache(q, true, fua);
442 	blk_queue_make_request(q, pmem_make_request);
443 	blk_queue_physical_block_size(q, PAGE_SIZE);
444 	blk_queue_logical_block_size(q, pmem_sector_size(ndns));
445 	blk_queue_max_hw_sectors(q, UINT_MAX);
446 	blk_queue_flag_set(QUEUE_FLAG_NONROT, q);
447 	if (pmem->pfn_flags & PFN_MAP)
448 		blk_queue_flag_set(QUEUE_FLAG_DAX, q);
449 	q->queuedata = pmem;
450 
451 	disk = alloc_disk_node(0, nid);
452 	if (!disk)
453 		return -ENOMEM;
454 	pmem->disk = disk;
455 
456 	disk->fops		= &pmem_fops;
457 	disk->queue		= q;
458 	disk->flags		= GENHD_FL_EXT_DEVT;
459 	disk->queue->backing_dev_info->capabilities |= BDI_CAP_SYNCHRONOUS_IO;
460 	nvdimm_namespace_disk_name(ndns, disk->disk_name);
461 	set_capacity(disk, (pmem->size - pmem->pfn_pad - pmem->data_offset)
462 			/ 512);
463 	if (devm_init_badblocks(dev, &pmem->bb))
464 		return -ENOMEM;
465 	nvdimm_badblocks_populate(nd_region, &pmem->bb, &bb_res);
466 	disk->bb = &pmem->bb;
467 
468 	dax_dev = alloc_dax(pmem, disk->disk_name, &pmem_dax_ops);
469 	if (!dax_dev) {
470 		put_disk(disk);
471 		return -ENOMEM;
472 	}
473 	dax_write_cache(dax_dev, nvdimm_has_cache(nd_region));
474 	pmem->dax_dev = dax_dev;
475 
476 	gendev = disk_to_dev(disk);
477 	gendev->groups = pmem_attribute_groups;
478 
479 	device_add_disk(dev, disk, NULL);
480 	if (devm_add_action_or_reset(dev, pmem_release_disk, pmem))
481 		return -ENOMEM;
482 
483 	revalidate_disk(disk);
484 
485 	pmem->bb_state = sysfs_get_dirent(disk_to_dev(disk)->kobj.sd,
486 					  "badblocks");
487 	if (!pmem->bb_state)
488 		dev_warn(dev, "'badblocks' notification disabled\n");
489 
490 	return 0;
491 }
492 
493 static int nd_pmem_probe(struct device *dev)
494 {
495 	struct nd_namespace_common *ndns;
496 
497 	ndns = nvdimm_namespace_common_probe(dev);
498 	if (IS_ERR(ndns))
499 		return PTR_ERR(ndns);
500 
501 	if (devm_nsio_enable(dev, to_nd_namespace_io(&ndns->dev)))
502 		return -ENXIO;
503 
504 	if (is_nd_btt(dev))
505 		return nvdimm_namespace_attach_btt(ndns);
506 
507 	if (is_nd_pfn(dev))
508 		return pmem_attach_disk(dev, ndns);
509 
510 	/* if we find a valid info-block we'll come back as that personality */
511 	if (nd_btt_probe(dev, ndns) == 0 || nd_pfn_probe(dev, ndns) == 0
512 			|| nd_dax_probe(dev, ndns) == 0)
513 		return -ENXIO;
514 
515 	/* ...otherwise we're just a raw pmem device */
516 	return pmem_attach_disk(dev, ndns);
517 }
518 
519 static int nd_pmem_remove(struct device *dev)
520 {
521 	struct pmem_device *pmem = dev_get_drvdata(dev);
522 
523 	if (is_nd_btt(dev))
524 		nvdimm_namespace_detach_btt(to_nd_btt(dev));
525 	else {
526 		/*
527 		 * Note, this assumes device_lock() context to not race
528 		 * nd_pmem_notify()
529 		 */
530 		sysfs_put(pmem->bb_state);
531 		pmem->bb_state = NULL;
532 	}
533 	nvdimm_flush(to_nd_region(dev->parent));
534 
535 	return 0;
536 }
537 
538 static void nd_pmem_shutdown(struct device *dev)
539 {
540 	nvdimm_flush(to_nd_region(dev->parent));
541 }
542 
543 static void nd_pmem_notify(struct device *dev, enum nvdimm_event event)
544 {
545 	struct nd_region *nd_region;
546 	resource_size_t offset = 0, end_trunc = 0;
547 	struct nd_namespace_common *ndns;
548 	struct nd_namespace_io *nsio;
549 	struct resource res;
550 	struct badblocks *bb;
551 	struct kernfs_node *bb_state;
552 
553 	if (event != NVDIMM_REVALIDATE_POISON)
554 		return;
555 
556 	if (is_nd_btt(dev)) {
557 		struct nd_btt *nd_btt = to_nd_btt(dev);
558 
559 		ndns = nd_btt->ndns;
560 		nd_region = to_nd_region(ndns->dev.parent);
561 		nsio = to_nd_namespace_io(&ndns->dev);
562 		bb = &nsio->bb;
563 		bb_state = NULL;
564 	} else {
565 		struct pmem_device *pmem = dev_get_drvdata(dev);
566 
567 		nd_region = to_region(pmem);
568 		bb = &pmem->bb;
569 		bb_state = pmem->bb_state;
570 
571 		if (is_nd_pfn(dev)) {
572 			struct nd_pfn *nd_pfn = to_nd_pfn(dev);
573 			struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
574 
575 			ndns = nd_pfn->ndns;
576 			offset = pmem->data_offset +
577 					__le32_to_cpu(pfn_sb->start_pad);
578 			end_trunc = __le32_to_cpu(pfn_sb->end_trunc);
579 		} else {
580 			ndns = to_ndns(dev);
581 		}
582 
583 		nsio = to_nd_namespace_io(&ndns->dev);
584 	}
585 
586 	res.start = nsio->res.start + offset;
587 	res.end = nsio->res.end - end_trunc;
588 	nvdimm_badblocks_populate(nd_region, bb, &res);
589 	if (bb_state)
590 		sysfs_notify_dirent(bb_state);
591 }
592 
593 MODULE_ALIAS("pmem");
594 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_IO);
595 MODULE_ALIAS_ND_DEVICE(ND_DEVICE_NAMESPACE_PMEM);
596 static struct nd_device_driver nd_pmem_driver = {
597 	.probe = nd_pmem_probe,
598 	.remove = nd_pmem_remove,
599 	.notify = nd_pmem_notify,
600 	.shutdown = nd_pmem_shutdown,
601 	.drv = {
602 		.name = "nd_pmem",
603 	},
604 	.type = ND_DRIVER_NAMESPACE_IO | ND_DRIVER_NAMESPACE_PMEM,
605 };
606 
607 module_nd_driver(nd_pmem_driver);
608 
609 MODULE_AUTHOR("Ross Zwisler <ross.zwisler@linux.intel.com>");
610 MODULE_LICENSE("GPL v2");
611