2 * Copyright(c) 2016 - 2017 Intel Corporation. All rights reserved.
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.
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.
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>
22 #include "dax-private.h"
25 static struct class *dax_class
;
28 * Rely on the fact that drvdata is set before the attributes are
29 * registered, and that the attributes are unregistered before drvdata
30 * is cleared to assume that drvdata is always valid.
32 static ssize_t
id_show(struct device
*dev
,
33 struct device_attribute
*attr
, char *buf
)
35 struct dax_region
*dax_region
= dev_get_drvdata(dev
);
37 return sprintf(buf
, "%d\n", dax_region
->id
);
39 static DEVICE_ATTR_RO(id
);
41 static ssize_t
region_size_show(struct device
*dev
,
42 struct device_attribute
*attr
, char *buf
)
44 struct dax_region
*dax_region
= dev_get_drvdata(dev
);
46 return sprintf(buf
, "%llu\n", (unsigned long long)
47 resource_size(&dax_region
->res
));
49 static struct device_attribute dev_attr_region_size
= __ATTR(size
, 0444,
50 region_size_show
, NULL
);
52 static ssize_t
align_show(struct device
*dev
,
53 struct device_attribute
*attr
, char *buf
)
55 struct dax_region
*dax_region
= dev_get_drvdata(dev
);
57 return sprintf(buf
, "%u\n", dax_region
->align
);
59 static DEVICE_ATTR_RO(align
);
61 static struct attribute
*dax_region_attributes
[] = {
62 &dev_attr_region_size
.attr
,
68 static const struct attribute_group dax_region_attribute_group
= {
70 .attrs
= dax_region_attributes
,
73 static const struct attribute_group
*dax_region_attribute_groups
[] = {
74 &dax_region_attribute_group
,
78 static void dax_region_free(struct kref
*kref
)
80 struct dax_region
*dax_region
;
82 dax_region
= container_of(kref
, struct dax_region
, kref
);
86 void dax_region_put(struct dax_region
*dax_region
)
88 kref_put(&dax_region
->kref
, dax_region_free
);
90 EXPORT_SYMBOL_GPL(dax_region_put
);
92 static void dax_region_unregister(void *region
)
94 struct dax_region
*dax_region
= region
;
96 sysfs_remove_groups(&dax_region
->dev
->kobj
,
97 dax_region_attribute_groups
);
98 dax_region_put(dax_region
);
101 struct dax_region
*alloc_dax_region(struct device
*parent
, int region_id
,
102 struct resource
*res
, unsigned int align
, void *addr
,
103 unsigned long pfn_flags
)
105 struct dax_region
*dax_region
;
108 * The DAX core assumes that it can store its private data in
109 * parent->driver_data. This WARN is a reminder / safeguard for
110 * developers of device-dax drivers.
112 if (dev_get_drvdata(parent
)) {
113 dev_WARN(parent
, "dax core failed to setup private data\n");
117 if (!IS_ALIGNED(res
->start
, align
)
118 || !IS_ALIGNED(resource_size(res
), align
))
121 dax_region
= kzalloc(sizeof(*dax_region
), GFP_KERNEL
);
125 dev_set_drvdata(parent
, dax_region
);
126 memcpy(&dax_region
->res
, res
, sizeof(*res
));
127 dax_region
->pfn_flags
= pfn_flags
;
128 kref_init(&dax_region
->kref
);
129 dax_region
->id
= region_id
;
130 ida_init(&dax_region
->ida
);
131 dax_region
->align
= align
;
132 dax_region
->dev
= parent
;
133 dax_region
->base
= addr
;
134 if (sysfs_create_groups(&parent
->kobj
, dax_region_attribute_groups
)) {
139 kref_get(&dax_region
->kref
);
140 if (devm_add_action_or_reset(parent
, dax_region_unregister
, dax_region
))
144 EXPORT_SYMBOL_GPL(alloc_dax_region
);
146 static struct dev_dax
*to_dev_dax(struct device
*dev
)
148 return container_of(dev
, struct dev_dax
, dev
);
151 static ssize_t
size_show(struct device
*dev
,
152 struct device_attribute
*attr
, char *buf
)
154 struct dev_dax
*dev_dax
= to_dev_dax(dev
);
155 unsigned long long size
= 0;
158 for (i
= 0; i
< dev_dax
->num_resources
; i
++)
159 size
+= resource_size(&dev_dax
->res
[i
]);
161 return sprintf(buf
, "%llu\n", size
);
163 static DEVICE_ATTR_RO(size
);
165 static struct attribute
*dev_dax_attributes
[] = {
170 static const struct attribute_group dev_dax_attribute_group
= {
171 .attrs
= dev_dax_attributes
,
174 static const struct attribute_group
*dax_attribute_groups
[] = {
175 &dev_dax_attribute_group
,
179 static int check_vma(struct dev_dax
*dev_dax
, struct vm_area_struct
*vma
,
182 struct dax_region
*dax_region
= dev_dax
->region
;
183 struct device
*dev
= &dev_dax
->dev
;
186 if (!dax_alive(dev_dax
->dax_dev
))
189 /* prevent private mappings from being established */
190 if ((vma
->vm_flags
& VM_MAYSHARE
) != VM_MAYSHARE
) {
191 dev_info_ratelimited(dev
,
192 "%s: %s: fail, attempted private mapping\n",
193 current
->comm
, func
);
197 mask
= dax_region
->align
- 1;
198 if (vma
->vm_start
& mask
|| vma
->vm_end
& mask
) {
199 dev_info_ratelimited(dev
,
200 "%s: %s: fail, unaligned vma (%#lx - %#lx, %#lx)\n",
201 current
->comm
, func
, vma
->vm_start
, vma
->vm_end
,
206 if ((dax_region
->pfn_flags
& (PFN_DEV
|PFN_MAP
)) == PFN_DEV
207 && (vma
->vm_flags
& VM_DONTCOPY
) == 0) {
208 dev_info_ratelimited(dev
,
209 "%s: %s: fail, dax range requires MADV_DONTFORK\n",
210 current
->comm
, func
);
214 if (!vma_is_dax(vma
)) {
215 dev_info_ratelimited(dev
,
216 "%s: %s: fail, vma is not DAX capable\n",
217 current
->comm
, func
);
224 /* see "strong" declaration in tools/testing/nvdimm/dax-dev.c */
225 __weak phys_addr_t
dax_pgoff_to_phys(struct dev_dax
*dev_dax
, pgoff_t pgoff
,
228 struct resource
*res
;
229 /* gcc-4.6.3-nolibc for i386 complains that this is uninitialized */
230 phys_addr_t
uninitialized_var(phys
);
233 for (i
= 0; i
< dev_dax
->num_resources
; i
++) {
234 res
= &dev_dax
->res
[i
];
235 phys
= pgoff
* PAGE_SIZE
+ res
->start
;
236 if (phys
>= res
->start
&& phys
<= res
->end
)
238 pgoff
-= PHYS_PFN(resource_size(res
));
241 if (i
< dev_dax
->num_resources
) {
242 res
= &dev_dax
->res
[i
];
243 if (phys
+ size
- 1 <= res
->end
)
250 static int __dev_dax_pte_fault(struct dev_dax
*dev_dax
, struct vm_fault
*vmf
)
252 struct device
*dev
= &dev_dax
->dev
;
253 struct dax_region
*dax_region
;
254 int rc
= VM_FAULT_SIGBUS
;
257 unsigned int fault_size
= PAGE_SIZE
;
259 if (check_vma(dev_dax
, vmf
->vma
, __func__
))
260 return VM_FAULT_SIGBUS
;
262 dax_region
= dev_dax
->region
;
263 if (dax_region
->align
> PAGE_SIZE
) {
264 dev_dbg(dev
, "%s: alignment (%#x) > fault size (%#x)\n",
265 __func__
, dax_region
->align
, fault_size
);
266 return VM_FAULT_SIGBUS
;
269 if (fault_size
!= dax_region
->align
)
270 return VM_FAULT_SIGBUS
;
272 phys
= dax_pgoff_to_phys(dev_dax
, vmf
->pgoff
, PAGE_SIZE
);
274 dev_dbg(dev
, "%s: pgoff_to_phys(%#lx) failed\n", __func__
,
276 return VM_FAULT_SIGBUS
;
279 pfn
= phys_to_pfn_t(phys
, dax_region
->pfn_flags
);
281 rc
= vm_insert_mixed(vmf
->vma
, vmf
->address
, pfn
);
285 if (rc
< 0 && rc
!= -EBUSY
)
286 return VM_FAULT_SIGBUS
;
288 return VM_FAULT_NOPAGE
;
291 static int __dev_dax_pmd_fault(struct dev_dax
*dev_dax
, struct vm_fault
*vmf
)
293 unsigned long pmd_addr
= vmf
->address
& PMD_MASK
;
294 struct device
*dev
= &dev_dax
->dev
;
295 struct dax_region
*dax_region
;
299 unsigned int fault_size
= PMD_SIZE
;
301 if (check_vma(dev_dax
, vmf
->vma
, __func__
))
302 return VM_FAULT_SIGBUS
;
304 dax_region
= dev_dax
->region
;
305 if (dax_region
->align
> PMD_SIZE
) {
306 dev_dbg(dev
, "%s: alignment (%#x) > fault size (%#x)\n",
307 __func__
, dax_region
->align
, fault_size
);
308 return VM_FAULT_SIGBUS
;
311 /* dax pmd mappings require pfn_t_devmap() */
312 if ((dax_region
->pfn_flags
& (PFN_DEV
|PFN_MAP
)) != (PFN_DEV
|PFN_MAP
)) {
313 dev_dbg(dev
, "%s: region lacks devmap flags\n", __func__
);
314 return VM_FAULT_SIGBUS
;
317 if (fault_size
< dax_region
->align
)
318 return VM_FAULT_SIGBUS
;
319 else if (fault_size
> dax_region
->align
)
320 return VM_FAULT_FALLBACK
;
322 /* if we are outside of the VMA */
323 if (pmd_addr
< vmf
->vma
->vm_start
||
324 (pmd_addr
+ PMD_SIZE
) > vmf
->vma
->vm_end
)
325 return VM_FAULT_SIGBUS
;
327 pgoff
= linear_page_index(vmf
->vma
, pmd_addr
);
328 phys
= dax_pgoff_to_phys(dev_dax
, pgoff
, PMD_SIZE
);
330 dev_dbg(dev
, "%s: pgoff_to_phys(%#lx) failed\n", __func__
,
332 return VM_FAULT_SIGBUS
;
335 pfn
= phys_to_pfn_t(phys
, dax_region
->pfn_flags
);
337 return vmf_insert_pfn_pmd(vmf
->vma
, vmf
->address
, vmf
->pmd
, pfn
,
338 vmf
->flags
& FAULT_FLAG_WRITE
);
341 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
342 static int __dev_dax_pud_fault(struct dev_dax
*dev_dax
, struct vm_fault
*vmf
)
344 unsigned long pud_addr
= vmf
->address
& PUD_MASK
;
345 struct device
*dev
= &dev_dax
->dev
;
346 struct dax_region
*dax_region
;
350 unsigned int fault_size
= PUD_SIZE
;
353 if (check_vma(dev_dax
, vmf
->vma
, __func__
))
354 return VM_FAULT_SIGBUS
;
356 dax_region
= dev_dax
->region
;
357 if (dax_region
->align
> PUD_SIZE
) {
358 dev_dbg(dev
, "%s: alignment (%#x) > fault size (%#x)\n",
359 __func__
, dax_region
->align
, fault_size
);
360 return VM_FAULT_SIGBUS
;
363 /* dax pud mappings require pfn_t_devmap() */
364 if ((dax_region
->pfn_flags
& (PFN_DEV
|PFN_MAP
)) != (PFN_DEV
|PFN_MAP
)) {
365 dev_dbg(dev
, "%s: region lacks devmap flags\n", __func__
);
366 return VM_FAULT_SIGBUS
;
369 if (fault_size
< dax_region
->align
)
370 return VM_FAULT_SIGBUS
;
371 else if (fault_size
> dax_region
->align
)
372 return VM_FAULT_FALLBACK
;
374 /* if we are outside of the VMA */
375 if (pud_addr
< vmf
->vma
->vm_start
||
376 (pud_addr
+ PUD_SIZE
) > vmf
->vma
->vm_end
)
377 return VM_FAULT_SIGBUS
;
379 pgoff
= linear_page_index(vmf
->vma
, pud_addr
);
380 phys
= dax_pgoff_to_phys(dev_dax
, pgoff
, PUD_SIZE
);
382 dev_dbg(dev
, "%s: pgoff_to_phys(%#lx) failed\n", __func__
,
384 return VM_FAULT_SIGBUS
;
387 pfn
= phys_to_pfn_t(phys
, dax_region
->pfn_flags
);
389 return vmf_insert_pfn_pud(vmf
->vma
, vmf
->address
, vmf
->pud
, pfn
,
390 vmf
->flags
& FAULT_FLAG_WRITE
);
393 static int __dev_dax_pud_fault(struct dev_dax
*dev_dax
, struct vm_fault
*vmf
)
395 return VM_FAULT_FALLBACK
;
397 #endif /* !CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
399 static int dev_dax_huge_fault(struct vm_fault
*vmf
,
400 enum page_entry_size pe_size
)
403 struct file
*filp
= vmf
->vma
->vm_file
;
404 struct dev_dax
*dev_dax
= filp
->private_data
;
406 dev_dbg(&dev_dax
->dev
, "%s: %s: %s (%#lx - %#lx) size = %d\n", __func__
,
407 current
->comm
, (vmf
->flags
& FAULT_FLAG_WRITE
)
409 vmf
->vma
->vm_start
, vmf
->vma
->vm_end
, pe_size
);
411 id
= dax_read_lock();
414 rc
= __dev_dax_pte_fault(dev_dax
, vmf
);
417 rc
= __dev_dax_pmd_fault(dev_dax
, vmf
);
420 rc
= __dev_dax_pud_fault(dev_dax
, vmf
);
423 rc
= VM_FAULT_SIGBUS
;
430 static int dev_dax_fault(struct vm_fault
*vmf
)
432 return dev_dax_huge_fault(vmf
, PE_SIZE_PTE
);
435 static int dev_dax_split(struct vm_area_struct
*vma
, unsigned long addr
)
437 struct file
*filp
= vma
->vm_file
;
438 struct dev_dax
*dev_dax
= filp
->private_data
;
439 struct dax_region
*dax_region
= dev_dax
->region
;
441 if (!IS_ALIGNED(addr
, dax_region
->align
))
446 static const struct vm_operations_struct dax_vm_ops
= {
447 .fault
= dev_dax_fault
,
448 .huge_fault
= dev_dax_huge_fault
,
449 .split
= dev_dax_split
,
452 static int dax_mmap(struct file
*filp
, struct vm_area_struct
*vma
)
454 struct dev_dax
*dev_dax
= filp
->private_data
;
457 dev_dbg(&dev_dax
->dev
, "%s\n", __func__
);
460 * We lock to check dax_dev liveness and will re-check at
463 id
= dax_read_lock();
464 rc
= check_vma(dev_dax
, vma
, __func__
);
469 vma
->vm_ops
= &dax_vm_ops
;
470 vma
->vm_flags
|= VM_MIXEDMAP
| VM_HUGEPAGE
;
474 /* return an unmapped area aligned to the dax region specified alignment */
475 static unsigned long dax_get_unmapped_area(struct file
*filp
,
476 unsigned long addr
, unsigned long len
, unsigned long pgoff
,
479 unsigned long off
, off_end
, off_align
, len_align
, addr_align
, align
;
480 struct dev_dax
*dev_dax
= filp
? filp
->private_data
: NULL
;
481 struct dax_region
*dax_region
;
483 if (!dev_dax
|| addr
)
486 dax_region
= dev_dax
->region
;
487 align
= dax_region
->align
;
488 off
= pgoff
<< PAGE_SHIFT
;
490 off_align
= round_up(off
, align
);
492 if ((off_end
<= off_align
) || ((off_end
- off_align
) < align
))
495 len_align
= len
+ align
;
496 if ((off
+ len_align
) < off
)
499 addr_align
= current
->mm
->get_unmapped_area(filp
, addr
, len_align
,
501 if (!IS_ERR_VALUE(addr_align
)) {
502 addr_align
+= (off
- addr_align
) & (align
- 1);
506 return current
->mm
->get_unmapped_area(filp
, addr
, len
, pgoff
, flags
);
509 static int dax_open(struct inode
*inode
, struct file
*filp
)
511 struct dax_device
*dax_dev
= inode_dax(inode
);
512 struct inode
*__dax_inode
= dax_inode(dax_dev
);
513 struct dev_dax
*dev_dax
= dax_get_private(dax_dev
);
515 dev_dbg(&dev_dax
->dev
, "%s\n", __func__
);
516 inode
->i_mapping
= __dax_inode
->i_mapping
;
517 inode
->i_mapping
->host
= __dax_inode
;
518 filp
->f_mapping
= inode
->i_mapping
;
519 filp
->f_wb_err
= filemap_sample_wb_err(filp
->f_mapping
);
520 filp
->private_data
= dev_dax
;
521 inode
->i_flags
= S_DAX
;
526 static int dax_release(struct inode
*inode
, struct file
*filp
)
528 struct dev_dax
*dev_dax
= filp
->private_data
;
530 dev_dbg(&dev_dax
->dev
, "%s\n", __func__
);
534 static const struct file_operations dax_fops
= {
535 .llseek
= noop_llseek
,
536 .owner
= THIS_MODULE
,
538 .release
= dax_release
,
539 .get_unmapped_area
= dax_get_unmapped_area
,
543 static void dev_dax_release(struct device
*dev
)
545 struct dev_dax
*dev_dax
= to_dev_dax(dev
);
546 struct dax_region
*dax_region
= dev_dax
->region
;
547 struct dax_device
*dax_dev
= dev_dax
->dax_dev
;
549 if (dev_dax
->id
>= 0)
550 ida_simple_remove(&dax_region
->ida
, dev_dax
->id
);
551 dax_region_put(dax_region
);
556 static void kill_dev_dax(struct dev_dax
*dev_dax
)
558 struct dax_device
*dax_dev
= dev_dax
->dax_dev
;
559 struct inode
*inode
= dax_inode(dax_dev
);
562 unmap_mapping_range(inode
->i_mapping
, 0, 0, 1);
565 static void unregister_dev_dax(void *dev
)
567 struct dev_dax
*dev_dax
= to_dev_dax(dev
);
568 struct dax_device
*dax_dev
= dev_dax
->dax_dev
;
569 struct inode
*inode
= dax_inode(dax_dev
);
570 struct cdev
*cdev
= inode
->i_cdev
;
572 dev_dbg(dev
, "%s\n", __func__
);
574 kill_dev_dax(dev_dax
);
575 cdev_device_del(cdev
, dev
);
579 struct dev_dax
*devm_create_dev_dax(struct dax_region
*dax_region
,
580 int id
, struct resource
*res
, int count
)
582 struct device
*parent
= dax_region
->dev
;
583 struct dax_device
*dax_dev
;
584 struct dev_dax
*dev_dax
;
591 return ERR_PTR(-EINVAL
);
593 dev_dax
= kzalloc(sizeof(*dev_dax
) + sizeof(*res
) * count
, GFP_KERNEL
);
595 return ERR_PTR(-ENOMEM
);
597 for (i
= 0; i
< count
; i
++) {
598 if (!IS_ALIGNED(res
[i
].start
, dax_region
->align
)
599 || !IS_ALIGNED(resource_size(&res
[i
]),
600 dax_region
->align
)) {
604 dev_dax
->res
[i
].start
= res
[i
].start
;
605 dev_dax
->res
[i
].end
= res
[i
].end
;
612 id
= ida_simple_get(&dax_region
->ida
, 0, 0, GFP_KERNEL
);
619 /* region provider owns @id lifetime */
624 * No 'host' or dax_operations since there is no access to this
625 * device outside of mmap of the resulting character device.
627 dax_dev
= alloc_dax(dev_dax
, NULL
, NULL
);
633 /* from here on we're committed to teardown via dax_dev_release() */
635 device_initialize(dev
);
637 inode
= dax_inode(dax_dev
);
638 cdev
= inode
->i_cdev
;
639 cdev_init(cdev
, &dax_fops
);
640 cdev
->owner
= parent
->driver
->owner
;
642 dev_dax
->num_resources
= count
;
643 dev_dax
->dax_dev
= dax_dev
;
644 dev_dax
->region
= dax_region
;
645 kref_get(&dax_region
->kref
);
647 dev
->devt
= inode
->i_rdev
;
648 dev
->class = dax_class
;
649 dev
->parent
= parent
;
650 dev
->groups
= dax_attribute_groups
;
651 dev
->release
= dev_dax_release
;
652 dev_set_name(dev
, "dax%d.%d", dax_region
->id
, id
);
654 rc
= cdev_device_add(cdev
, dev
);
656 kill_dev_dax(dev_dax
);
661 rc
= devm_add_action_or_reset(dax_region
->dev
, unregister_dev_dax
, dev
);
668 if (dev_dax
->id
>= 0)
669 ida_simple_remove(&dax_region
->ida
, dev_dax
->id
);
675 EXPORT_SYMBOL_GPL(devm_create_dev_dax
);
677 static int __init
dax_init(void)
679 dax_class
= class_create(THIS_MODULE
, "dax");
680 return PTR_ERR_OR_ZERO(dax_class
);
683 static void __exit
dax_exit(void)
685 class_destroy(dax_class
);
688 MODULE_AUTHOR("Intel Corporation");
689 MODULE_LICENSE("GPL v2");
690 subsys_initcall(dax_init
);
691 module_exit(dax_exit
);