2 * Copyright(c) 2016 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/mount.h>
17 #include <linux/pfn_t.h>
18 #include <linux/hash.h>
19 #include <linux/cdev.h>
20 #include <linux/slab.h>
21 #include <linux/dax.h>
26 static dev_t dax_devt;
27 static struct class *dax_class;
28 static DEFINE_IDA(dax_minor_ida);
29 static int nr_dax = CONFIG_NR_DEV_DAX;
30 module_param(nr_dax, int, S_IRUGO);
31 static struct vfsmount *dax_mnt;
32 static struct kmem_cache *dax_cache __read_mostly;
33 static struct super_block *dax_superblock __read_mostly;
34 MODULE_PARM_DESC(nr_dax, "max number of device-dax instances");
37 * struct dax_region - mapping infrastructure for dax devices
38 * @id: kernel-wide unique region for a memory range
39 * @base: linear address corresponding to @res
40 * @kref: to pin while other agents have a need to do lookups
41 * @dev: parent device backing this region
42 * @align: allocation and mapping alignment for child dax devices
43 * @res: physical address range of the region
44 * @pfn_flags: identify whether the pfns are paged back or not
54 unsigned long pfn_flags;
58 * struct dax_dev - subdivision of a dax region
59 * @region - parent region
60 * @dev - device backing the character device
61 * @cdev - core chardev data
62 * @alive - !alive + rcu grace period == no new mappings can be established
63 * @id - child id in the region
64 * @num_resources - number of physical address extents in this device
65 * @res - array of physical address ranges
68 struct dax_region *region;
75 struct resource res[0];
78 static ssize_t id_show(struct device *dev,
79 struct device_attribute *attr, char *buf)
81 struct dax_region *dax_region;
85 dax_region = dev_get_drvdata(dev);
87 rc = sprintf(buf, "%d\n", dax_region->id);
92 static DEVICE_ATTR_RO(id);
94 static ssize_t region_size_show(struct device *dev,
95 struct device_attribute *attr, char *buf)
97 struct dax_region *dax_region;
101 dax_region = dev_get_drvdata(dev);
103 rc = sprintf(buf, "%llu\n", (unsigned long long)
104 resource_size(&dax_region->res));
109 static struct device_attribute dev_attr_region_size = __ATTR(size, 0444,
110 region_size_show, NULL);
112 static ssize_t align_show(struct device *dev,
113 struct device_attribute *attr, char *buf)
115 struct dax_region *dax_region;
119 dax_region = dev_get_drvdata(dev);
121 rc = sprintf(buf, "%u\n", dax_region->align);
126 static DEVICE_ATTR_RO(align);
128 static struct attribute *dax_region_attributes[] = {
129 &dev_attr_region_size.attr,
130 &dev_attr_align.attr,
135 static const struct attribute_group dax_region_attribute_group = {
136 .name = "dax_region",
137 .attrs = dax_region_attributes,
140 static const struct attribute_group *dax_region_attribute_groups[] = {
141 &dax_region_attribute_group,
145 static struct inode *dax_alloc_inode(struct super_block *sb)
147 return kmem_cache_alloc(dax_cache, GFP_KERNEL);
150 static void dax_i_callback(struct rcu_head *head)
152 struct inode *inode = container_of(head, struct inode, i_rcu);
154 kmem_cache_free(dax_cache, inode);
157 static void dax_destroy_inode(struct inode *inode)
159 call_rcu(&inode->i_rcu, dax_i_callback);
162 static const struct super_operations dax_sops = {
163 .statfs = simple_statfs,
164 .alloc_inode = dax_alloc_inode,
165 .destroy_inode = dax_destroy_inode,
166 .drop_inode = generic_delete_inode,
169 static struct dentry *dax_mount(struct file_system_type *fs_type,
170 int flags, const char *dev_name, void *data)
172 return mount_pseudo(fs_type, "dax:", &dax_sops, NULL, DAXFS_MAGIC);
175 static struct file_system_type dax_type = {
178 .kill_sb = kill_anon_super,
181 static int dax_test(struct inode *inode, void *data)
183 return inode->i_cdev == data;
186 static int dax_set(struct inode *inode, void *data)
188 inode->i_cdev = data;
192 static struct inode *dax_inode_get(struct cdev *cdev, dev_t devt)
196 inode = iget5_locked(dax_superblock, hash_32(devt + DAXFS_MAGIC, 31),
197 dax_test, dax_set, cdev);
202 if (inode->i_state & I_NEW) {
203 inode->i_mode = S_IFCHR;
204 inode->i_flags = S_DAX;
205 inode->i_rdev = devt;
206 mapping_set_gfp_mask(&inode->i_data, GFP_USER);
207 unlock_new_inode(inode);
212 static void init_once(void *inode)
214 inode_init_once(inode);
217 static int dax_inode_init(void)
221 dax_cache = kmem_cache_create("dax_cache", sizeof(struct inode), 0,
222 (SLAB_HWCACHE_ALIGN|SLAB_RECLAIM_ACCOUNT|
223 SLAB_MEM_SPREAD|SLAB_ACCOUNT),
228 rc = register_filesystem(&dax_type);
230 goto err_register_fs;
232 dax_mnt = kern_mount(&dax_type);
233 if (IS_ERR(dax_mnt)) {
234 rc = PTR_ERR(dax_mnt);
237 dax_superblock = dax_mnt->mnt_sb;
242 unregister_filesystem(&dax_type);
244 kmem_cache_destroy(dax_cache);
249 static void dax_inode_exit(void)
251 kern_unmount(dax_mnt);
252 unregister_filesystem(&dax_type);
253 kmem_cache_destroy(dax_cache);
256 static void dax_region_free(struct kref *kref)
258 struct dax_region *dax_region;
260 dax_region = container_of(kref, struct dax_region, kref);
264 void dax_region_put(struct dax_region *dax_region)
266 kref_put(&dax_region->kref, dax_region_free);
268 EXPORT_SYMBOL_GPL(dax_region_put);
270 static void dax_region_unregister(void *region)
272 struct dax_region *dax_region = region;
274 sysfs_remove_groups(&dax_region->dev->kobj,
275 dax_region_attribute_groups);
276 dax_region_put(dax_region);
279 struct dax_region *alloc_dax_region(struct device *parent, int region_id,
280 struct resource *res, unsigned int align, void *addr,
281 unsigned long pfn_flags)
283 struct dax_region *dax_region;
286 * The DAX core assumes that it can store its private data in
287 * parent->driver_data. This WARN is a reminder / safeguard for
288 * developers of device-dax drivers.
290 if (dev_get_drvdata(parent)) {
291 dev_WARN(parent, "dax core failed to setup private data\n");
295 if (!IS_ALIGNED(res->start, align)
296 || !IS_ALIGNED(resource_size(res), align))
299 dax_region = kzalloc(sizeof(*dax_region), GFP_KERNEL);
303 dev_set_drvdata(parent, dax_region);
304 memcpy(&dax_region->res, res, sizeof(*res));
305 dax_region->pfn_flags = pfn_flags;
306 kref_init(&dax_region->kref);
307 dax_region->id = region_id;
308 ida_init(&dax_region->ida);
309 dax_region->align = align;
310 dax_region->dev = parent;
311 dax_region->base = addr;
312 if (sysfs_create_groups(&parent->kobj, dax_region_attribute_groups)) {
317 kref_get(&dax_region->kref);
318 if (devm_add_action_or_reset(parent, dax_region_unregister, dax_region))
322 EXPORT_SYMBOL_GPL(alloc_dax_region);
324 static struct dax_dev *to_dax_dev(struct device *dev)
326 return container_of(dev, struct dax_dev, dev);
329 static ssize_t size_show(struct device *dev,
330 struct device_attribute *attr, char *buf)
332 struct dax_dev *dax_dev = to_dax_dev(dev);
333 unsigned long long size = 0;
336 for (i = 0; i < dax_dev->num_resources; i++)
337 size += resource_size(&dax_dev->res[i]);
339 return sprintf(buf, "%llu\n", size);
341 static DEVICE_ATTR_RO(size);
343 static struct attribute *dax_device_attributes[] = {
348 static const struct attribute_group dax_device_attribute_group = {
349 .attrs = dax_device_attributes,
352 static const struct attribute_group *dax_attribute_groups[] = {
353 &dax_device_attribute_group,
357 static int check_vma(struct dax_dev *dax_dev, struct vm_area_struct *vma,
360 struct dax_region *dax_region = dax_dev->region;
361 struct device *dev = &dax_dev->dev;
367 /* prevent private mappings from being established */
368 if ((vma->vm_flags & VM_MAYSHARE) != VM_MAYSHARE) {
369 dev_info(dev, "%s: %s: fail, attempted private mapping\n",
370 current->comm, func);
374 mask = dax_region->align - 1;
375 if (vma->vm_start & mask || vma->vm_end & mask) {
376 dev_info(dev, "%s: %s: fail, unaligned vma (%#lx - %#lx, %#lx)\n",
377 current->comm, func, vma->vm_start, vma->vm_end,
382 if ((dax_region->pfn_flags & (PFN_DEV|PFN_MAP)) == PFN_DEV
383 && (vma->vm_flags & VM_DONTCOPY) == 0) {
384 dev_info(dev, "%s: %s: fail, dax range requires MADV_DONTFORK\n",
385 current->comm, func);
389 if (!vma_is_dax(vma)) {
390 dev_info(dev, "%s: %s: fail, vma is not DAX capable\n",
391 current->comm, func);
398 static phys_addr_t pgoff_to_phys(struct dax_dev *dax_dev, pgoff_t pgoff,
401 struct resource *res;
405 for (i = 0; i < dax_dev->num_resources; i++) {
406 res = &dax_dev->res[i];
407 phys = pgoff * PAGE_SIZE + res->start;
408 if (phys >= res->start && phys <= res->end)
410 pgoff -= PHYS_PFN(resource_size(res));
413 if (i < dax_dev->num_resources) {
414 res = &dax_dev->res[i];
415 if (phys + size - 1 <= res->end)
422 static int __dax_dev_pte_fault(struct dax_dev *dax_dev, struct vm_fault *vmf)
424 struct device *dev = &dax_dev->dev;
425 struct dax_region *dax_region;
426 int rc = VM_FAULT_SIGBUS;
430 if (check_vma(dax_dev, vmf->vma, __func__))
431 return VM_FAULT_SIGBUS;
433 dax_region = dax_dev->region;
434 if (dax_region->align > PAGE_SIZE) {
435 dev_dbg(dev, "%s: alignment > fault size\n", __func__);
436 return VM_FAULT_SIGBUS;
439 phys = pgoff_to_phys(dax_dev, vmf->pgoff, PAGE_SIZE);
441 dev_dbg(dev, "%s: phys_to_pgoff(%#lx) failed\n", __func__,
443 return VM_FAULT_SIGBUS;
446 pfn = phys_to_pfn_t(phys, dax_region->pfn_flags);
448 rc = vm_insert_mixed(vmf->vma, vmf->address, pfn);
452 if (rc < 0 && rc != -EBUSY)
453 return VM_FAULT_SIGBUS;
455 return VM_FAULT_NOPAGE;
458 static int __dax_dev_pmd_fault(struct dax_dev *dax_dev, struct vm_fault *vmf)
460 unsigned long pmd_addr = vmf->address & PMD_MASK;
461 struct device *dev = &dax_dev->dev;
462 struct dax_region *dax_region;
467 if (check_vma(dax_dev, vmf->vma, __func__))
468 return VM_FAULT_SIGBUS;
470 dax_region = dax_dev->region;
471 if (dax_region->align > PMD_SIZE) {
472 dev_dbg(dev, "%s: alignment > fault size\n", __func__);
473 return VM_FAULT_SIGBUS;
476 /* dax pmd mappings require pfn_t_devmap() */
477 if ((dax_region->pfn_flags & (PFN_DEV|PFN_MAP)) != (PFN_DEV|PFN_MAP)) {
478 dev_dbg(dev, "%s: alignment > fault size\n", __func__);
479 return VM_FAULT_SIGBUS;
482 pgoff = linear_page_index(vmf->vma, pmd_addr);
483 phys = pgoff_to_phys(dax_dev, pgoff, PMD_SIZE);
485 dev_dbg(dev, "%s: phys_to_pgoff(%#lx) failed\n", __func__,
487 return VM_FAULT_SIGBUS;
490 pfn = phys_to_pfn_t(phys, dax_region->pfn_flags);
492 return vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd, pfn,
493 vmf->flags & FAULT_FLAG_WRITE);
496 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
497 static int __dax_dev_pud_fault(struct dax_dev *dax_dev, struct vm_fault *vmf)
499 unsigned long pud_addr = vmf->address & PUD_MASK;
500 struct device *dev = &dax_dev->dev;
501 struct dax_region *dax_region;
506 if (check_vma(dax_dev, vmf->vma, __func__))
507 return VM_FAULT_SIGBUS;
509 dax_region = dax_dev->region;
510 if (dax_region->align > PUD_SIZE) {
511 dev_dbg(dev, "%s: alignment > fault size\n", __func__);
512 return VM_FAULT_SIGBUS;
515 /* dax pud mappings require pfn_t_devmap() */
516 if ((dax_region->pfn_flags & (PFN_DEV|PFN_MAP)) != (PFN_DEV|PFN_MAP)) {
517 dev_dbg(dev, "%s: alignment > fault size\n", __func__);
518 return VM_FAULT_SIGBUS;
521 pgoff = linear_page_index(vmf->vma, pud_addr);
522 phys = pgoff_to_phys(dax_dev, pgoff, PUD_SIZE);
524 dev_dbg(dev, "%s: phys_to_pgoff(%#lx) failed\n", __func__,
526 return VM_FAULT_SIGBUS;
529 pfn = phys_to_pfn_t(phys, dax_region->pfn_flags);
531 return vmf_insert_pfn_pud(vmf->vma, vmf->address, vmf->pud, pfn,
532 vmf->flags & FAULT_FLAG_WRITE);
535 static int __dax_dev_pud_fault(struct dax_dev *dax_dev, struct vm_fault *vmf)
537 return VM_FAULT_FALLBACK;
539 #endif /* !CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
541 static int dax_dev_fault(struct vm_fault *vmf)
544 struct file *filp = vmf->vma->vm_file;
545 struct dax_dev *dax_dev = filp->private_data;
547 dev_dbg(&dax_dev->dev, "%s: %s: %s (%#lx - %#lx)\n", __func__,
548 current->comm, (vmf->flags & FAULT_FLAG_WRITE)
550 vmf->vma->vm_start, vmf->vma->vm_end);
553 switch (vmf->flags & FAULT_FLAG_SIZE_MASK) {
554 case FAULT_FLAG_SIZE_PTE:
555 rc = __dax_dev_pte_fault(dax_dev, vmf);
557 case FAULT_FLAG_SIZE_PMD:
558 rc = __dax_dev_pmd_fault(dax_dev, vmf);
560 case FAULT_FLAG_SIZE_PUD:
561 rc = __dax_dev_pud_fault(dax_dev, vmf);
564 return VM_FAULT_FALLBACK;
571 static const struct vm_operations_struct dax_dev_vm_ops = {
572 .fault = dax_dev_fault,
573 .huge_fault = dax_dev_fault,
576 static int dax_mmap(struct file *filp, struct vm_area_struct *vma)
578 struct dax_dev *dax_dev = filp->private_data;
581 dev_dbg(&dax_dev->dev, "%s\n", __func__);
583 rc = check_vma(dax_dev, vma, __func__);
587 vma->vm_ops = &dax_dev_vm_ops;
588 vma->vm_flags |= VM_MIXEDMAP | VM_HUGEPAGE;
592 /* return an unmapped area aligned to the dax region specified alignment */
593 static unsigned long dax_get_unmapped_area(struct file *filp,
594 unsigned long addr, unsigned long len, unsigned long pgoff,
597 unsigned long off, off_end, off_align, len_align, addr_align, align;
598 struct dax_dev *dax_dev = filp ? filp->private_data : NULL;
599 struct dax_region *dax_region;
601 if (!dax_dev || addr)
604 dax_region = dax_dev->region;
605 align = dax_region->align;
606 off = pgoff << PAGE_SHIFT;
608 off_align = round_up(off, align);
610 if ((off_end <= off_align) || ((off_end - off_align) < align))
613 len_align = len + align;
614 if ((off + len_align) < off)
617 addr_align = current->mm->get_unmapped_area(filp, addr, len_align,
619 if (!IS_ERR_VALUE(addr_align)) {
620 addr_align += (off - addr_align) & (align - 1);
624 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
627 static int dax_open(struct inode *inode, struct file *filp)
629 struct dax_dev *dax_dev;
631 dax_dev = container_of(inode->i_cdev, struct dax_dev, cdev);
632 dev_dbg(&dax_dev->dev, "%s\n", __func__);
633 inode->i_mapping = dax_dev->inode->i_mapping;
634 inode->i_mapping->host = dax_dev->inode;
635 filp->f_mapping = inode->i_mapping;
636 filp->private_data = dax_dev;
637 inode->i_flags = S_DAX;
642 static int dax_release(struct inode *inode, struct file *filp)
644 struct dax_dev *dax_dev = filp->private_data;
646 dev_dbg(&dax_dev->dev, "%s\n", __func__);
650 static const struct file_operations dax_fops = {
651 .llseek = noop_llseek,
652 .owner = THIS_MODULE,
654 .release = dax_release,
655 .get_unmapped_area = dax_get_unmapped_area,
659 static void dax_dev_release(struct device *dev)
661 struct dax_dev *dax_dev = to_dax_dev(dev);
662 struct dax_region *dax_region = dax_dev->region;
664 ida_simple_remove(&dax_region->ida, dax_dev->id);
665 ida_simple_remove(&dax_minor_ida, MINOR(dev->devt));
666 dax_region_put(dax_region);
667 iput(dax_dev->inode);
671 static void unregister_dax_dev(void *dev)
673 struct dax_dev *dax_dev = to_dax_dev(dev);
674 struct cdev *cdev = &dax_dev->cdev;
676 dev_dbg(dev, "%s\n", __func__);
679 * Note, rcu is not protecting the liveness of dax_dev, rcu is
680 * ensuring that any fault handlers that might have seen
681 * dax_dev->alive == true, have completed. Any fault handlers
682 * that start after synchronize_rcu() has started will abort
683 * upon seeing dax_dev->alive == false.
685 dax_dev->alive = false;
687 unmap_mapping_range(dax_dev->inode->i_mapping, 0, 0, 1);
689 device_unregister(dev);
692 struct dax_dev *devm_create_dax_dev(struct dax_region *dax_region,
693 struct resource *res, int count)
695 struct device *parent = dax_region->dev;
696 struct dax_dev *dax_dev;
697 int rc = 0, minor, i;
702 dax_dev = kzalloc(sizeof(*dax_dev) + sizeof(*res) * count, GFP_KERNEL);
704 return ERR_PTR(-ENOMEM);
706 for (i = 0; i < count; i++) {
707 if (!IS_ALIGNED(res[i].start, dax_region->align)
708 || !IS_ALIGNED(resource_size(&res[i]),
709 dax_region->align)) {
713 dax_dev->res[i].start = res[i].start;
714 dax_dev->res[i].end = res[i].end;
720 dax_dev->id = ida_simple_get(&dax_region->ida, 0, 0, GFP_KERNEL);
721 if (dax_dev->id < 0) {
726 minor = ida_simple_get(&dax_minor_ida, 0, 0, GFP_KERNEL);
732 dev_t = MKDEV(MAJOR(dax_devt), minor);
734 dax_dev->inode = dax_inode_get(&dax_dev->cdev, dev_t);
735 if (!dax_dev->inode) {
740 /* device_initialize() so cdev can reference kobj parent */
741 device_initialize(dev);
743 cdev = &dax_dev->cdev;
744 cdev_init(cdev, &dax_fops);
745 cdev->owner = parent->driver->owner;
746 cdev->kobj.parent = &dev->kobj;
747 rc = cdev_add(&dax_dev->cdev, dev_t, 1);
751 /* from here on we're committed to teardown via dax_dev_release() */
752 dax_dev->num_resources = count;
753 dax_dev->alive = true;
754 dax_dev->region = dax_region;
755 kref_get(&dax_region->kref);
758 dev->class = dax_class;
759 dev->parent = parent;
760 dev->groups = dax_attribute_groups;
761 dev->release = dax_dev_release;
762 dev_set_name(dev, "dax%d.%d", dax_region->id, dax_dev->id);
763 rc = device_add(dev);
769 rc = devm_add_action_or_reset(dax_region->dev, unregister_dax_dev, dev);
776 iput(dax_dev->inode);
778 ida_simple_remove(&dax_minor_ida, minor);
780 ida_simple_remove(&dax_region->ida, dax_dev->id);
786 EXPORT_SYMBOL_GPL(devm_create_dax_dev);
788 static int __init dax_init(void)
792 rc = dax_inode_init();
796 nr_dax = max(nr_dax, 256);
797 rc = alloc_chrdev_region(&dax_devt, 0, nr_dax, "dax");
801 dax_class = class_create(THIS_MODULE, "dax");
802 if (IS_ERR(dax_class)) {
803 rc = PTR_ERR(dax_class);
810 unregister_chrdev_region(dax_devt, nr_dax);
816 static void __exit dax_exit(void)
818 class_destroy(dax_class);
819 unregister_chrdev_region(dax_devt, nr_dax);
820 ida_destroy(&dax_minor_ida);
824 MODULE_AUTHOR("Intel Corporation");
825 MODULE_LICENSE("GPL v2");
826 subsys_initcall(dax_init);
827 module_exit(dax_exit);