xref: /linux/drivers/dax/fsdev.c (revision 759455848df0b9ac3acabdbedcdc4a55af67935f)
1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright(c) 2026 Micron Technology, Inc. */
3 #include <linux/memremap.h>
4 #include <linux/pagemap.h>
5 #include <linux/module.h>
6 #include <linux/device.h>
7 #include <linux/cdev.h>
8 #include <linux/slab.h>
9 #include <linux/dax.h>
10 #include <linux/uio.h>
11 #include <linux/fs.h>
12 #include <linux/mm.h>
13 #include "dax-private.h"
14 #include "bus.h"
15 
16 /*
17  * FS-DAX compatible devdax driver
18  *
19  * Unlike drivers/dax/device.c which pre-initializes compound folios based
20  * on device alignment (via vmemmap_shift), this driver leaves folios
21  * uninitialized similar to pmem. This allows fs-dax filesystems like famfs
22  * to work without needing special handling for pre-initialized folios.
23  *
24  * Key differences from device.c:
25  * - pgmap type is MEMORY_DEVICE_FS_DAX (not MEMORY_DEVICE_GENERIC)
26  * - vmemmap_shift is NOT set (folios remain order-0)
27  * - fs-dax can dynamically create compound folios as needed
28  * - No mmap support - all access is through fs-dax/iomap
29  */
30 
31 static void fsdev_cdev_del(void *cdev)
32 {
33 	cdev_del(cdev);
34 }
35 
36 static void fsdev_kill(void *dev_dax)
37 {
38 	kill_dev_dax(dev_dax);
39 }
40 
41 /*
42  * Page map operations for FS-DAX mode
43  * Similar to fsdax_pagemap_ops in drivers/nvdimm/pmem.c
44  *
45  * Note: folio_free callback is not needed for MEMORY_DEVICE_FS_DAX.
46  * The core mm code in free_zone_device_folio() handles the wake_up_var()
47  * directly for this memory type.
48  */
49 static int fsdev_pagemap_memory_failure(struct dev_pagemap *pgmap,
50 		unsigned long pfn, unsigned long nr_pages, int mf_flags)
51 {
52 	struct dev_dax *dev_dax = pgmap->owner;
53 	u64 offset = PFN_PHYS(pfn) - dev_dax->ranges[0].range.start;
54 	u64 len = nr_pages << PAGE_SHIFT;
55 
56 	return dax_holder_notify_failure(dev_dax->dax_dev, offset,
57 					 len, mf_flags);
58 }
59 
60 static const struct dev_pagemap_ops fsdev_pagemap_ops = {
61 	.memory_failure		= fsdev_pagemap_memory_failure,
62 };
63 
64 /*
65  * Clear any stale folio state from pages in the given range.
66  * This is necessary because device_dax pre-initializes compound folios
67  * based on vmemmap_shift, and that state may persist after driver unbind.
68  * Since fsdev_dax uses MEMORY_DEVICE_FS_DAX without vmemmap_shift, fs-dax
69  * expects to find clean order-0 folios that it can build into compound
70  * folios on demand.
71  *
72  * At probe time, no filesystem should be mounted yet, so all mappings
73  * are stale and must be cleared along with compound state.
74  */
75 static void fsdev_clear_folio_state(struct dev_dax *dev_dax)
76 {
77 	for (int i = 0; i < dev_dax->nr_range; i++) {
78 		struct range *range = &dev_dax->ranges[i].range;
79 		unsigned long pfn = PHYS_PFN(range->start);
80 		unsigned long end_pfn = PHYS_PFN(range->end) + 1;
81 
82 		while (pfn < end_pfn) {
83 			struct folio *folio = pfn_folio(pfn);
84 			int order = dax_folio_reset_order(folio);
85 
86 			pfn += 1UL << order;
87 		}
88 	}
89 }
90 
91 static void fsdev_clear_folio_state_action(void *data)
92 {
93 	fsdev_clear_folio_state(data);
94 }
95 
96 static int fsdev_open(struct inode *inode, struct file *filp)
97 {
98 	struct dax_device *dax_dev = inode_dax(inode);
99 	struct dev_dax *dev_dax = dax_get_private(dax_dev);
100 
101 	filp->private_data = dev_dax;
102 
103 	return 0;
104 }
105 
106 static int fsdev_release(struct inode *inode, struct file *filp)
107 {
108 	return 0;
109 }
110 
111 static const struct file_operations fsdev_fops = {
112 	.llseek = noop_llseek,
113 	.owner = THIS_MODULE,
114 	.open = fsdev_open,
115 	.release = fsdev_release,
116 };
117 
118 static int fsdev_dax_probe(struct dev_dax *dev_dax)
119 {
120 	struct dax_device *dax_dev = dev_dax->dax_dev;
121 	struct device *dev = &dev_dax->dev;
122 	struct dev_pagemap *pgmap;
123 	struct inode *inode;
124 	u64 data_offset = 0;
125 	struct cdev *cdev;
126 	void *addr;
127 	int rc, i;
128 
129 	if (static_dev_dax(dev_dax)) {
130 		if (dev_dax->nr_range > 1) {
131 			dev_warn(dev, "static pgmap / multi-range device conflict\n");
132 			return -EINVAL;
133 		}
134 
135 		pgmap = dev_dax->pgmap;
136 	} else {
137 		size_t pgmap_size;
138 
139 		if (dev_dax->pgmap) {
140 			dev_warn(dev, "dynamic-dax with pre-populated page map\n");
141 			return -EINVAL;
142 		}
143 
144 		pgmap_size = struct_size(pgmap, ranges, dev_dax->nr_range - 1);
145 		pgmap = devm_kzalloc(dev, pgmap_size, GFP_KERNEL);
146 		if (!pgmap)
147 			return -ENOMEM;
148 
149 		pgmap->nr_range = dev_dax->nr_range;
150 		dev_dax->pgmap = pgmap;
151 
152 		for (i = 0; i < dev_dax->nr_range; i++) {
153 			struct range *range = &dev_dax->ranges[i].range;
154 
155 			pgmap->ranges[i] = *range;
156 		}
157 	}
158 
159 	for (i = 0; i < dev_dax->nr_range; i++) {
160 		struct range *range = &dev_dax->ranges[i].range;
161 
162 		if (!devm_request_mem_region(dev, range->start,
163 					range_len(range), dev_name(dev))) {
164 			dev_warn(dev, "mapping%d: %#llx-%#llx could not reserve range\n",
165 				 i, range->start, range->end);
166 			return -EBUSY;
167 		}
168 	}
169 
170 	/*
171 	 * Use MEMORY_DEVICE_FS_DAX without setting vmemmap_shift, leaving
172 	 * folios at order-0. Unlike device.c (MEMORY_DEVICE_GENERIC), this
173 	 * lets fs-dax dynamically build compound folios as needed, similar
174 	 * to pmem behavior.
175 	 */
176 	pgmap->type = MEMORY_DEVICE_FS_DAX;
177 	pgmap->ops = &fsdev_pagemap_ops;
178 	pgmap->owner = dev_dax;
179 
180 	addr = devm_memremap_pages(dev, pgmap);
181 	if (IS_ERR(addr))
182 		return PTR_ERR(addr);
183 
184 	/*
185 	 * Clear any stale compound folio state left over from a previous
186 	 * driver (e.g., device_dax with vmemmap_shift). Also register this
187 	 * as a devm action so folio state is cleared on unbind, ensuring
188 	 * clean pages for subsequent drivers (e.g., kmem for system-ram).
189 	 */
190 	fsdev_clear_folio_state(dev_dax);
191 	rc = devm_add_action_or_reset(dev, fsdev_clear_folio_state_action,
192 				      dev_dax);
193 	if (rc)
194 		return rc;
195 
196 	/* Detect whether the data is at a non-zero offset into the memory */
197 	if (pgmap->range.start != dev_dax->ranges[0].range.start) {
198 		u64 phys = dev_dax->ranges[0].range.start;
199 		u64 pgmap_phys = dev_dax->pgmap[0].range.start;
200 
201 		if (!WARN_ON(pgmap_phys > phys))
202 			data_offset = phys - pgmap_phys;
203 
204 		pr_debug("%s: offset detected phys=%llx pgmap_phys=%llx offset=%llx\n",
205 		       __func__, phys, pgmap_phys, data_offset);
206 	}
207 	dev_dax->virt_addr = addr + data_offset;
208 
209 	inode = dax_inode(dax_dev);
210 	cdev = inode->i_cdev;
211 	cdev_init(cdev, &fsdev_fops);
212 	cdev->owner = dev->driver->owner;
213 	cdev_set_parent(cdev, &dev->kobj);
214 	rc = cdev_add(cdev, dev->devt, 1);
215 	if (rc)
216 		return rc;
217 
218 	rc = devm_add_action_or_reset(dev, fsdev_cdev_del, cdev);
219 	if (rc)
220 		return rc;
221 
222 	run_dax(dax_dev);
223 	return devm_add_action_or_reset(dev, fsdev_kill, dev_dax);
224 }
225 
226 static struct dax_device_driver fsdev_dax_driver = {
227 	.probe = fsdev_dax_probe,
228 	.type = DAXDRV_FSDEV_TYPE,
229 };
230 
231 static int __init dax_init(void)
232 {
233 	return dax_driver_register(&fsdev_dax_driver);
234 }
235 
236 static void __exit dax_exit(void)
237 {
238 	dax_driver_unregister(&fsdev_dax_driver);
239 }
240 
241 MODULE_AUTHOR("John Groves");
242 MODULE_DESCRIPTION("FS-DAX Device: fs-dax compatible devdax driver");
243 MODULE_LICENSE("GPL");
244 module_init(dax_init);
245 module_exit(dax_exit);
246 MODULE_ALIAS_DAX_DEVICE(0);
247