xref: /openbmc/linux/drivers/dax/device.c (revision 5b4cb650)
1 /*
2  * Copyright(c) 2016 - 2017 Intel Corporation. All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or modify
5  * it under the terms of version 2 of the GNU General Public License as
6  * published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful, but
9  * WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  */
13 #include <linux/pagemap.h>
14 #include <linux/module.h>
15 #include <linux/device.h>
16 #include <linux/pfn_t.h>
17 #include <linux/cdev.h>
18 #include <linux/slab.h>
19 #include <linux/dax.h>
20 #include <linux/fs.h>
21 #include <linux/mm.h>
22 #include <linux/mman.h>
23 #include "dax-private.h"
24 #include "dax.h"
25 
26 static struct class *dax_class;
27 
28 /*
29  * Rely on the fact that drvdata is set before the attributes are
30  * registered, and that the attributes are unregistered before drvdata
31  * is cleared to assume that drvdata is always valid.
32  */
33 static ssize_t id_show(struct device *dev,
34 		struct device_attribute *attr, char *buf)
35 {
36 	struct dax_region *dax_region = dev_get_drvdata(dev);
37 
38 	return sprintf(buf, "%d\n", dax_region->id);
39 }
40 static DEVICE_ATTR_RO(id);
41 
42 static ssize_t region_size_show(struct device *dev,
43 		struct device_attribute *attr, char *buf)
44 {
45 	struct dax_region *dax_region = dev_get_drvdata(dev);
46 
47 	return sprintf(buf, "%llu\n", (unsigned long long)
48 			resource_size(&dax_region->res));
49 }
50 static struct device_attribute dev_attr_region_size = __ATTR(size, 0444,
51 		region_size_show, NULL);
52 
53 static ssize_t align_show(struct device *dev,
54 		struct device_attribute *attr, char *buf)
55 {
56 	struct dax_region *dax_region = dev_get_drvdata(dev);
57 
58 	return sprintf(buf, "%u\n", dax_region->align);
59 }
60 static DEVICE_ATTR_RO(align);
61 
62 static struct attribute *dax_region_attributes[] = {
63 	&dev_attr_region_size.attr,
64 	&dev_attr_align.attr,
65 	&dev_attr_id.attr,
66 	NULL,
67 };
68 
69 static const struct attribute_group dax_region_attribute_group = {
70 	.name = "dax_region",
71 	.attrs = dax_region_attributes,
72 };
73 
74 static const struct attribute_group *dax_region_attribute_groups[] = {
75 	&dax_region_attribute_group,
76 	NULL,
77 };
78 
79 static void dax_region_free(struct kref *kref)
80 {
81 	struct dax_region *dax_region;
82 
83 	dax_region = container_of(kref, struct dax_region, kref);
84 	kfree(dax_region);
85 }
86 
87 void dax_region_put(struct dax_region *dax_region)
88 {
89 	kref_put(&dax_region->kref, dax_region_free);
90 }
91 EXPORT_SYMBOL_GPL(dax_region_put);
92 
93 static void dax_region_unregister(void *region)
94 {
95 	struct dax_region *dax_region = region;
96 
97 	sysfs_remove_groups(&dax_region->dev->kobj,
98 			dax_region_attribute_groups);
99 	dax_region_put(dax_region);
100 }
101 
102 struct dax_region *alloc_dax_region(struct device *parent, int region_id,
103 		struct resource *res, unsigned int align, void *addr,
104 		unsigned long pfn_flags)
105 {
106 	struct dax_region *dax_region;
107 
108 	/*
109 	 * The DAX core assumes that it can store its private data in
110 	 * parent->driver_data. This WARN is a reminder / safeguard for
111 	 * developers of device-dax drivers.
112 	 */
113 	if (dev_get_drvdata(parent)) {
114 		dev_WARN(parent, "dax core failed to setup private data\n");
115 		return NULL;
116 	}
117 
118 	if (!IS_ALIGNED(res->start, align)
119 			|| !IS_ALIGNED(resource_size(res), align))
120 		return NULL;
121 
122 	dax_region = kzalloc(sizeof(*dax_region), GFP_KERNEL);
123 	if (!dax_region)
124 		return NULL;
125 
126 	dev_set_drvdata(parent, dax_region);
127 	memcpy(&dax_region->res, res, sizeof(*res));
128 	dax_region->pfn_flags = pfn_flags;
129 	kref_init(&dax_region->kref);
130 	dax_region->id = region_id;
131 	ida_init(&dax_region->ida);
132 	dax_region->align = align;
133 	dax_region->dev = parent;
134 	dax_region->base = addr;
135 	if (sysfs_create_groups(&parent->kobj, dax_region_attribute_groups)) {
136 		kfree(dax_region);
137 		return NULL;
138 	}
139 
140 	kref_get(&dax_region->kref);
141 	if (devm_add_action_or_reset(parent, dax_region_unregister, dax_region))
142 		return NULL;
143 	return dax_region;
144 }
145 EXPORT_SYMBOL_GPL(alloc_dax_region);
146 
147 static struct dev_dax *to_dev_dax(struct device *dev)
148 {
149 	return container_of(dev, struct dev_dax, dev);
150 }
151 
152 static ssize_t size_show(struct device *dev,
153 		struct device_attribute *attr, char *buf)
154 {
155 	struct dev_dax *dev_dax = to_dev_dax(dev);
156 	unsigned long long size = 0;
157 	int i;
158 
159 	for (i = 0; i < dev_dax->num_resources; i++)
160 		size += resource_size(&dev_dax->res[i]);
161 
162 	return sprintf(buf, "%llu\n", size);
163 }
164 static DEVICE_ATTR_RO(size);
165 
166 static struct attribute *dev_dax_attributes[] = {
167 	&dev_attr_size.attr,
168 	NULL,
169 };
170 
171 static const struct attribute_group dev_dax_attribute_group = {
172 	.attrs = dev_dax_attributes,
173 };
174 
175 static const struct attribute_group *dax_attribute_groups[] = {
176 	&dev_dax_attribute_group,
177 	NULL,
178 };
179 
180 static int check_vma(struct dev_dax *dev_dax, struct vm_area_struct *vma,
181 		const char *func)
182 {
183 	struct dax_region *dax_region = dev_dax->region;
184 	struct device *dev = &dev_dax->dev;
185 	unsigned long mask;
186 
187 	if (!dax_alive(dev_dax->dax_dev))
188 		return -ENXIO;
189 
190 	/* prevent private mappings from being established */
191 	if ((vma->vm_flags & VM_MAYSHARE) != VM_MAYSHARE) {
192 		dev_info_ratelimited(dev,
193 				"%s: %s: fail, attempted private mapping\n",
194 				current->comm, func);
195 		return -EINVAL;
196 	}
197 
198 	mask = dax_region->align - 1;
199 	if (vma->vm_start & mask || vma->vm_end & mask) {
200 		dev_info_ratelimited(dev,
201 				"%s: %s: fail, unaligned vma (%#lx - %#lx, %#lx)\n",
202 				current->comm, func, vma->vm_start, vma->vm_end,
203 				mask);
204 		return -EINVAL;
205 	}
206 
207 	if ((dax_region->pfn_flags & (PFN_DEV|PFN_MAP)) == PFN_DEV
208 			&& (vma->vm_flags & VM_DONTCOPY) == 0) {
209 		dev_info_ratelimited(dev,
210 				"%s: %s: fail, dax range requires MADV_DONTFORK\n",
211 				current->comm, func);
212 		return -EINVAL;
213 	}
214 
215 	if (!vma_is_dax(vma)) {
216 		dev_info_ratelimited(dev,
217 				"%s: %s: fail, vma is not DAX capable\n",
218 				current->comm, func);
219 		return -EINVAL;
220 	}
221 
222 	return 0;
223 }
224 
225 /* see "strong" declaration in tools/testing/nvdimm/dax-dev.c */
226 __weak phys_addr_t dax_pgoff_to_phys(struct dev_dax *dev_dax, pgoff_t pgoff,
227 		unsigned long size)
228 {
229 	struct resource *res;
230 	/* gcc-4.6.3-nolibc for i386 complains that this is uninitialized */
231 	phys_addr_t uninitialized_var(phys);
232 	int i;
233 
234 	for (i = 0; i < dev_dax->num_resources; i++) {
235 		res = &dev_dax->res[i];
236 		phys = pgoff * PAGE_SIZE + res->start;
237 		if (phys >= res->start && phys <= res->end)
238 			break;
239 		pgoff -= PHYS_PFN(resource_size(res));
240 	}
241 
242 	if (i < dev_dax->num_resources) {
243 		res = &dev_dax->res[i];
244 		if (phys + size - 1 <= res->end)
245 			return phys;
246 	}
247 
248 	return -1;
249 }
250 
251 static vm_fault_t __dev_dax_pte_fault(struct dev_dax *dev_dax,
252 				struct vm_fault *vmf, pfn_t *pfn)
253 {
254 	struct device *dev = &dev_dax->dev;
255 	struct dax_region *dax_region;
256 	phys_addr_t phys;
257 	unsigned int fault_size = PAGE_SIZE;
258 
259 	if (check_vma(dev_dax, vmf->vma, __func__))
260 		return VM_FAULT_SIGBUS;
261 
262 	dax_region = dev_dax->region;
263 	if (dax_region->align > PAGE_SIZE) {
264 		dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
265 			dax_region->align, fault_size);
266 		return VM_FAULT_SIGBUS;
267 	}
268 
269 	if (fault_size != dax_region->align)
270 		return VM_FAULT_SIGBUS;
271 
272 	phys = dax_pgoff_to_phys(dev_dax, vmf->pgoff, PAGE_SIZE);
273 	if (phys == -1) {
274 		dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", vmf->pgoff);
275 		return VM_FAULT_SIGBUS;
276 	}
277 
278 	*pfn = phys_to_pfn_t(phys, dax_region->pfn_flags);
279 
280 	return vmf_insert_mixed(vmf->vma, vmf->address, *pfn);
281 }
282 
283 static vm_fault_t __dev_dax_pmd_fault(struct dev_dax *dev_dax,
284 				struct vm_fault *vmf, pfn_t *pfn)
285 {
286 	unsigned long pmd_addr = vmf->address & PMD_MASK;
287 	struct device *dev = &dev_dax->dev;
288 	struct dax_region *dax_region;
289 	phys_addr_t phys;
290 	pgoff_t pgoff;
291 	unsigned int fault_size = PMD_SIZE;
292 
293 	if (check_vma(dev_dax, vmf->vma, __func__))
294 		return VM_FAULT_SIGBUS;
295 
296 	dax_region = dev_dax->region;
297 	if (dax_region->align > PMD_SIZE) {
298 		dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
299 			dax_region->align, fault_size);
300 		return VM_FAULT_SIGBUS;
301 	}
302 
303 	/* dax pmd mappings require pfn_t_devmap() */
304 	if ((dax_region->pfn_flags & (PFN_DEV|PFN_MAP)) != (PFN_DEV|PFN_MAP)) {
305 		dev_dbg(dev, "region lacks devmap flags\n");
306 		return VM_FAULT_SIGBUS;
307 	}
308 
309 	if (fault_size < dax_region->align)
310 		return VM_FAULT_SIGBUS;
311 	else if (fault_size > dax_region->align)
312 		return VM_FAULT_FALLBACK;
313 
314 	/* if we are outside of the VMA */
315 	if (pmd_addr < vmf->vma->vm_start ||
316 			(pmd_addr + PMD_SIZE) > vmf->vma->vm_end)
317 		return VM_FAULT_SIGBUS;
318 
319 	pgoff = linear_page_index(vmf->vma, pmd_addr);
320 	phys = dax_pgoff_to_phys(dev_dax, pgoff, PMD_SIZE);
321 	if (phys == -1) {
322 		dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", pgoff);
323 		return VM_FAULT_SIGBUS;
324 	}
325 
326 	*pfn = phys_to_pfn_t(phys, dax_region->pfn_flags);
327 
328 	return vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd, *pfn,
329 			vmf->flags & FAULT_FLAG_WRITE);
330 }
331 
332 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
333 static vm_fault_t __dev_dax_pud_fault(struct dev_dax *dev_dax,
334 				struct vm_fault *vmf, pfn_t *pfn)
335 {
336 	unsigned long pud_addr = vmf->address & PUD_MASK;
337 	struct device *dev = &dev_dax->dev;
338 	struct dax_region *dax_region;
339 	phys_addr_t phys;
340 	pgoff_t pgoff;
341 	unsigned int fault_size = PUD_SIZE;
342 
343 
344 	if (check_vma(dev_dax, vmf->vma, __func__))
345 		return VM_FAULT_SIGBUS;
346 
347 	dax_region = dev_dax->region;
348 	if (dax_region->align > PUD_SIZE) {
349 		dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
350 			dax_region->align, fault_size);
351 		return VM_FAULT_SIGBUS;
352 	}
353 
354 	/* dax pud mappings require pfn_t_devmap() */
355 	if ((dax_region->pfn_flags & (PFN_DEV|PFN_MAP)) != (PFN_DEV|PFN_MAP)) {
356 		dev_dbg(dev, "region lacks devmap flags\n");
357 		return VM_FAULT_SIGBUS;
358 	}
359 
360 	if (fault_size < dax_region->align)
361 		return VM_FAULT_SIGBUS;
362 	else if (fault_size > dax_region->align)
363 		return VM_FAULT_FALLBACK;
364 
365 	/* if we are outside of the VMA */
366 	if (pud_addr < vmf->vma->vm_start ||
367 			(pud_addr + PUD_SIZE) > vmf->vma->vm_end)
368 		return VM_FAULT_SIGBUS;
369 
370 	pgoff = linear_page_index(vmf->vma, pud_addr);
371 	phys = dax_pgoff_to_phys(dev_dax, pgoff, PUD_SIZE);
372 	if (phys == -1) {
373 		dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", pgoff);
374 		return VM_FAULT_SIGBUS;
375 	}
376 
377 	*pfn = phys_to_pfn_t(phys, dax_region->pfn_flags);
378 
379 	return vmf_insert_pfn_pud(vmf->vma, vmf->address, vmf->pud, *pfn,
380 			vmf->flags & FAULT_FLAG_WRITE);
381 }
382 #else
383 static vm_fault_t __dev_dax_pud_fault(struct dev_dax *dev_dax,
384 				struct vm_fault *vmf, pfn_t *pfn)
385 {
386 	return VM_FAULT_FALLBACK;
387 }
388 #endif /* !CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
389 
390 static vm_fault_t dev_dax_huge_fault(struct vm_fault *vmf,
391 		enum page_entry_size pe_size)
392 {
393 	struct file *filp = vmf->vma->vm_file;
394 	unsigned long fault_size;
395 	vm_fault_t rc = VM_FAULT_SIGBUS;
396 	int id;
397 	pfn_t pfn;
398 	struct dev_dax *dev_dax = filp->private_data;
399 
400 	dev_dbg(&dev_dax->dev, "%s: %s (%#lx - %#lx) size = %d\n", current->comm,
401 			(vmf->flags & FAULT_FLAG_WRITE) ? "write" : "read",
402 			vmf->vma->vm_start, vmf->vma->vm_end, pe_size);
403 
404 	id = dax_read_lock();
405 	switch (pe_size) {
406 	case PE_SIZE_PTE:
407 		fault_size = PAGE_SIZE;
408 		rc = __dev_dax_pte_fault(dev_dax, vmf, &pfn);
409 		break;
410 	case PE_SIZE_PMD:
411 		fault_size = PMD_SIZE;
412 		rc = __dev_dax_pmd_fault(dev_dax, vmf, &pfn);
413 		break;
414 	case PE_SIZE_PUD:
415 		fault_size = PUD_SIZE;
416 		rc = __dev_dax_pud_fault(dev_dax, vmf, &pfn);
417 		break;
418 	default:
419 		rc = VM_FAULT_SIGBUS;
420 	}
421 
422 	if (rc == VM_FAULT_NOPAGE) {
423 		unsigned long i;
424 		pgoff_t pgoff;
425 
426 		/*
427 		 * In the device-dax case the only possibility for a
428 		 * VM_FAULT_NOPAGE result is when device-dax capacity is
429 		 * mapped. No need to consider the zero page, or racing
430 		 * conflicting mappings.
431 		 */
432 		pgoff = linear_page_index(vmf->vma, vmf->address
433 				& ~(fault_size - 1));
434 		for (i = 0; i < fault_size / PAGE_SIZE; i++) {
435 			struct page *page;
436 
437 			page = pfn_to_page(pfn_t_to_pfn(pfn) + i);
438 			if (page->mapping)
439 				continue;
440 			page->mapping = filp->f_mapping;
441 			page->index = pgoff + i;
442 		}
443 	}
444 	dax_read_unlock(id);
445 
446 	return rc;
447 }
448 
449 static vm_fault_t dev_dax_fault(struct vm_fault *vmf)
450 {
451 	return dev_dax_huge_fault(vmf, PE_SIZE_PTE);
452 }
453 
454 static int dev_dax_split(struct vm_area_struct *vma, unsigned long addr)
455 {
456 	struct file *filp = vma->vm_file;
457 	struct dev_dax *dev_dax = filp->private_data;
458 	struct dax_region *dax_region = dev_dax->region;
459 
460 	if (!IS_ALIGNED(addr, dax_region->align))
461 		return -EINVAL;
462 	return 0;
463 }
464 
465 static unsigned long dev_dax_pagesize(struct vm_area_struct *vma)
466 {
467 	struct file *filp = vma->vm_file;
468 	struct dev_dax *dev_dax = filp->private_data;
469 	struct dax_region *dax_region = dev_dax->region;
470 
471 	return dax_region->align;
472 }
473 
474 static const struct vm_operations_struct dax_vm_ops = {
475 	.fault = dev_dax_fault,
476 	.huge_fault = dev_dax_huge_fault,
477 	.split = dev_dax_split,
478 	.pagesize = dev_dax_pagesize,
479 };
480 
481 static int dax_mmap(struct file *filp, struct vm_area_struct *vma)
482 {
483 	struct dev_dax *dev_dax = filp->private_data;
484 	int rc, id;
485 
486 	dev_dbg(&dev_dax->dev, "trace\n");
487 
488 	/*
489 	 * We lock to check dax_dev liveness and will re-check at
490 	 * fault time.
491 	 */
492 	id = dax_read_lock();
493 	rc = check_vma(dev_dax, vma, __func__);
494 	dax_read_unlock(id);
495 	if (rc)
496 		return rc;
497 
498 	vma->vm_ops = &dax_vm_ops;
499 	vma->vm_flags |= VM_HUGEPAGE;
500 	return 0;
501 }
502 
503 /* return an unmapped area aligned to the dax region specified alignment */
504 static unsigned long dax_get_unmapped_area(struct file *filp,
505 		unsigned long addr, unsigned long len, unsigned long pgoff,
506 		unsigned long flags)
507 {
508 	unsigned long off, off_end, off_align, len_align, addr_align, align;
509 	struct dev_dax *dev_dax = filp ? filp->private_data : NULL;
510 	struct dax_region *dax_region;
511 
512 	if (!dev_dax || addr)
513 		goto out;
514 
515 	dax_region = dev_dax->region;
516 	align = dax_region->align;
517 	off = pgoff << PAGE_SHIFT;
518 	off_end = off + len;
519 	off_align = round_up(off, align);
520 
521 	if ((off_end <= off_align) || ((off_end - off_align) < align))
522 		goto out;
523 
524 	len_align = len + align;
525 	if ((off + len_align) < off)
526 		goto out;
527 
528 	addr_align = current->mm->get_unmapped_area(filp, addr, len_align,
529 			pgoff, flags);
530 	if (!IS_ERR_VALUE(addr_align)) {
531 		addr_align += (off - addr_align) & (align - 1);
532 		return addr_align;
533 	}
534  out:
535 	return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
536 }
537 
538 static const struct address_space_operations dev_dax_aops = {
539 	.set_page_dirty		= noop_set_page_dirty,
540 	.invalidatepage		= noop_invalidatepage,
541 };
542 
543 static int dax_open(struct inode *inode, struct file *filp)
544 {
545 	struct dax_device *dax_dev = inode_dax(inode);
546 	struct inode *__dax_inode = dax_inode(dax_dev);
547 	struct dev_dax *dev_dax = dax_get_private(dax_dev);
548 
549 	dev_dbg(&dev_dax->dev, "trace\n");
550 	inode->i_mapping = __dax_inode->i_mapping;
551 	inode->i_mapping->host = __dax_inode;
552 	inode->i_mapping->a_ops = &dev_dax_aops;
553 	filp->f_mapping = inode->i_mapping;
554 	filp->f_wb_err = filemap_sample_wb_err(filp->f_mapping);
555 	filp->private_data = dev_dax;
556 	inode->i_flags = S_DAX;
557 
558 	return 0;
559 }
560 
561 static int dax_release(struct inode *inode, struct file *filp)
562 {
563 	struct dev_dax *dev_dax = filp->private_data;
564 
565 	dev_dbg(&dev_dax->dev, "trace\n");
566 	return 0;
567 }
568 
569 static const struct file_operations dax_fops = {
570 	.llseek = noop_llseek,
571 	.owner = THIS_MODULE,
572 	.open = dax_open,
573 	.release = dax_release,
574 	.get_unmapped_area = dax_get_unmapped_area,
575 	.mmap = dax_mmap,
576 	.mmap_supported_flags = MAP_SYNC,
577 };
578 
579 static void dev_dax_release(struct device *dev)
580 {
581 	struct dev_dax *dev_dax = to_dev_dax(dev);
582 	struct dax_region *dax_region = dev_dax->region;
583 	struct dax_device *dax_dev = dev_dax->dax_dev;
584 
585 	if (dev_dax->id >= 0)
586 		ida_simple_remove(&dax_region->ida, dev_dax->id);
587 	dax_region_put(dax_region);
588 	put_dax(dax_dev);
589 	kfree(dev_dax);
590 }
591 
592 static void kill_dev_dax(struct dev_dax *dev_dax)
593 {
594 	struct dax_device *dax_dev = dev_dax->dax_dev;
595 	struct inode *inode = dax_inode(dax_dev);
596 
597 	kill_dax(dax_dev);
598 	unmap_mapping_range(inode->i_mapping, 0, 0, 1);
599 }
600 
601 static void unregister_dev_dax(void *dev)
602 {
603 	struct dev_dax *dev_dax = to_dev_dax(dev);
604 	struct dax_device *dax_dev = dev_dax->dax_dev;
605 	struct inode *inode = dax_inode(dax_dev);
606 	struct cdev *cdev = inode->i_cdev;
607 
608 	dev_dbg(dev, "trace\n");
609 
610 	kill_dev_dax(dev_dax);
611 	cdev_device_del(cdev, dev);
612 	put_device(dev);
613 }
614 
615 struct dev_dax *devm_create_dev_dax(struct dax_region *dax_region,
616 		int id, struct resource *res, int count)
617 {
618 	struct device *parent = dax_region->dev;
619 	struct dax_device *dax_dev;
620 	struct dev_dax *dev_dax;
621 	struct inode *inode;
622 	struct device *dev;
623 	struct cdev *cdev;
624 	int rc, i;
625 
626 	if (!count)
627 		return ERR_PTR(-EINVAL);
628 
629 	dev_dax = kzalloc(struct_size(dev_dax, res, count), GFP_KERNEL);
630 	if (!dev_dax)
631 		return ERR_PTR(-ENOMEM);
632 
633 	for (i = 0; i < count; i++) {
634 		if (!IS_ALIGNED(res[i].start, dax_region->align)
635 				|| !IS_ALIGNED(resource_size(&res[i]),
636 					dax_region->align)) {
637 			rc = -EINVAL;
638 			break;
639 		}
640 		dev_dax->res[i].start = res[i].start;
641 		dev_dax->res[i].end = res[i].end;
642 	}
643 
644 	if (i < count)
645 		goto err_id;
646 
647 	if (id < 0) {
648 		id = ida_simple_get(&dax_region->ida, 0, 0, GFP_KERNEL);
649 		dev_dax->id = id;
650 		if (id < 0) {
651 			rc = id;
652 			goto err_id;
653 		}
654 	} else {
655 		/* region provider owns @id lifetime */
656 		dev_dax->id = -1;
657 	}
658 
659 	/*
660 	 * No 'host' or dax_operations since there is no access to this
661 	 * device outside of mmap of the resulting character device.
662 	 */
663 	dax_dev = alloc_dax(dev_dax, NULL, NULL);
664 	if (!dax_dev) {
665 		rc = -ENOMEM;
666 		goto err_dax;
667 	}
668 
669 	/* from here on we're committed to teardown via dax_dev_release() */
670 	dev = &dev_dax->dev;
671 	device_initialize(dev);
672 
673 	inode = dax_inode(dax_dev);
674 	cdev = inode->i_cdev;
675 	cdev_init(cdev, &dax_fops);
676 	cdev->owner = parent->driver->owner;
677 
678 	dev_dax->num_resources = count;
679 	dev_dax->dax_dev = dax_dev;
680 	dev_dax->region = dax_region;
681 	kref_get(&dax_region->kref);
682 
683 	dev->devt = inode->i_rdev;
684 	dev->class = dax_class;
685 	dev->parent = parent;
686 	dev->groups = dax_attribute_groups;
687 	dev->release = dev_dax_release;
688 	dev_set_name(dev, "dax%d.%d", dax_region->id, id);
689 
690 	rc = cdev_device_add(cdev, dev);
691 	if (rc) {
692 		kill_dev_dax(dev_dax);
693 		put_device(dev);
694 		return ERR_PTR(rc);
695 	}
696 
697 	rc = devm_add_action_or_reset(dax_region->dev, unregister_dev_dax, dev);
698 	if (rc)
699 		return ERR_PTR(rc);
700 
701 	return dev_dax;
702 
703  err_dax:
704 	if (dev_dax->id >= 0)
705 		ida_simple_remove(&dax_region->ida, dev_dax->id);
706  err_id:
707 	kfree(dev_dax);
708 
709 	return ERR_PTR(rc);
710 }
711 EXPORT_SYMBOL_GPL(devm_create_dev_dax);
712 
713 static int __init dax_init(void)
714 {
715 	dax_class = class_create(THIS_MODULE, "dax");
716 	return PTR_ERR_OR_ZERO(dax_class);
717 }
718 
719 static void __exit dax_exit(void)
720 {
721 	class_destroy(dax_class);
722 }
723 
724 MODULE_AUTHOR("Intel Corporation");
725 MODULE_LICENSE("GPL v2");
726 subsys_initcall(dax_init);
727 module_exit(dax_exit);
728