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_write_dax(void *addr, struct page *page, 32 unsigned int off, unsigned int len) 33 { 34 while (len) { 35 void *mem = kmap_local_page(page); 36 unsigned int chunk = min_t(unsigned int, len, PAGE_SIZE - off); 37 38 memcpy_flushcache(addr, mem + off, chunk); 39 kunmap_local(mem); 40 len -= chunk; 41 off = 0; 42 page++; 43 addr += chunk; 44 } 45 } 46 47 static long __fsdev_dax_direct_access(struct dax_device *dax_dev, pgoff_t pgoff, 48 long nr_pages, enum dax_access_mode mode, void **kaddr, 49 unsigned long *pfn) 50 { 51 struct dev_dax *dev_dax = dax_get_private(dax_dev); 52 size_t size = nr_pages << PAGE_SHIFT; 53 size_t offset = pgoff << PAGE_SHIFT; 54 void *virt_addr = dev_dax->virt_addr + offset; 55 phys_addr_t phys; 56 unsigned long local_pfn; 57 58 phys = dax_pgoff_to_phys(dev_dax, pgoff, size); 59 if (phys == -1) { 60 dev_dbg(&dev_dax->dev, 61 "pgoff (%#lx) out of range\n", pgoff); 62 return -EFAULT; 63 } 64 65 if (kaddr) 66 *kaddr = virt_addr; 67 68 local_pfn = PHYS_PFN(phys); 69 if (pfn) 70 *pfn = local_pfn; 71 72 /* 73 * Use cached_size which was computed at probe time. The size cannot 74 * change while the driver is bound (resize returns -EBUSY). 75 */ 76 return PHYS_PFN(min(size, dev_dax->cached_size - offset)); 77 } 78 79 static int fsdev_dax_zero_page_range(struct dax_device *dax_dev, 80 pgoff_t pgoff, size_t nr_pages) 81 { 82 void *kaddr; 83 84 WARN_ONCE(nr_pages > 1, "%s: nr_pages > 1\n", __func__); 85 __fsdev_dax_direct_access(dax_dev, pgoff, 1, DAX_ACCESS, &kaddr, NULL); 86 fsdev_write_dax(kaddr, ZERO_PAGE(0), 0, PAGE_SIZE); 87 return 0; 88 } 89 90 static long fsdev_dax_direct_access(struct dax_device *dax_dev, 91 pgoff_t pgoff, long nr_pages, enum dax_access_mode mode, 92 void **kaddr, unsigned long *pfn) 93 { 94 return __fsdev_dax_direct_access(dax_dev, pgoff, nr_pages, mode, 95 kaddr, pfn); 96 } 97 98 static size_t fsdev_dax_recovery_write(struct dax_device *dax_dev, pgoff_t pgoff, 99 void *addr, size_t bytes, struct iov_iter *i) 100 { 101 return _copy_from_iter_flushcache(addr, bytes, i); 102 } 103 104 static const struct dax_operations dev_dax_ops = { 105 .direct_access = fsdev_dax_direct_access, 106 .zero_page_range = fsdev_dax_zero_page_range, 107 .recovery_write = fsdev_dax_recovery_write, 108 }; 109 110 static void fsdev_cdev_del(void *cdev) 111 { 112 cdev_del(cdev); 113 } 114 115 static void fsdev_kill(void *dev_dax) 116 { 117 kill_dev_dax(dev_dax); 118 } 119 120 /* 121 * Page map operations for FS-DAX mode 122 * Similar to fsdax_pagemap_ops in drivers/nvdimm/pmem.c 123 * 124 * Note: folio_free callback is not needed for MEMORY_DEVICE_FS_DAX. 125 * The core mm code in free_zone_device_folio() handles the wake_up_var() 126 * directly for this memory type. 127 */ 128 static int fsdev_pagemap_memory_failure(struct dev_pagemap *pgmap, 129 unsigned long pfn, unsigned long nr_pages, int mf_flags) 130 { 131 struct dev_dax *dev_dax = pgmap->owner; 132 u64 offset = PFN_PHYS(pfn) - dev_dax->ranges[0].range.start; 133 u64 len = nr_pages << PAGE_SHIFT; 134 135 return dax_holder_notify_failure(dev_dax->dax_dev, offset, 136 len, mf_flags); 137 } 138 139 static const struct dev_pagemap_ops fsdev_pagemap_ops = { 140 .memory_failure = fsdev_pagemap_memory_failure, 141 }; 142 143 /* 144 * Clear any stale folio state from pages in the given range. 145 * This is necessary because device_dax pre-initializes compound folios 146 * based on vmemmap_shift, and that state may persist after driver unbind. 147 * Since fsdev_dax uses MEMORY_DEVICE_FS_DAX without vmemmap_shift, fs-dax 148 * expects to find clean order-0 folios that it can build into compound 149 * folios on demand. 150 * 151 * At probe time, no filesystem should be mounted yet, so all mappings 152 * are stale and must be cleared along with compound state. 153 */ 154 static void fsdev_clear_folio_state(struct dev_dax *dev_dax) 155 { 156 for (int i = 0; i < dev_dax->nr_range; i++) { 157 struct range *range = &dev_dax->ranges[i].range; 158 unsigned long pfn = PHYS_PFN(range->start); 159 unsigned long end_pfn = PHYS_PFN(range->end) + 1; 160 161 while (pfn < end_pfn) { 162 struct folio *folio = pfn_folio(pfn); 163 int order = dax_folio_reset_order(folio); 164 165 pfn += 1UL << order; 166 } 167 } 168 } 169 170 static void fsdev_clear_folio_state_action(void *data) 171 { 172 fsdev_clear_folio_state(data); 173 } 174 175 static int fsdev_open(struct inode *inode, struct file *filp) 176 { 177 struct dax_device *dax_dev = inode_dax(inode); 178 struct dev_dax *dev_dax = dax_get_private(dax_dev); 179 180 filp->private_data = dev_dax; 181 182 return 0; 183 } 184 185 static int fsdev_release(struct inode *inode, struct file *filp) 186 { 187 return 0; 188 } 189 190 static const struct file_operations fsdev_fops = { 191 .llseek = noop_llseek, 192 .owner = THIS_MODULE, 193 .open = fsdev_open, 194 .release = fsdev_release, 195 }; 196 197 static int fsdev_dax_probe(struct dev_dax *dev_dax) 198 { 199 struct dax_device *dax_dev = dev_dax->dax_dev; 200 struct device *dev = &dev_dax->dev; 201 struct dev_pagemap *pgmap; 202 struct inode *inode; 203 u64 data_offset = 0; 204 struct cdev *cdev; 205 void *addr; 206 int rc, i; 207 208 if (static_dev_dax(dev_dax)) { 209 if (dev_dax->nr_range > 1) { 210 dev_warn(dev, "static pgmap / multi-range device conflict\n"); 211 return -EINVAL; 212 } 213 214 pgmap = dev_dax->pgmap; 215 } else { 216 size_t pgmap_size; 217 218 if (dev_dax->pgmap) { 219 dev_warn(dev, "dynamic-dax with pre-populated page map\n"); 220 return -EINVAL; 221 } 222 223 pgmap_size = struct_size(pgmap, ranges, dev_dax->nr_range - 1); 224 pgmap = devm_kzalloc(dev, pgmap_size, GFP_KERNEL); 225 if (!pgmap) 226 return -ENOMEM; 227 228 pgmap->nr_range = dev_dax->nr_range; 229 dev_dax->pgmap = pgmap; 230 231 for (i = 0; i < dev_dax->nr_range; i++) { 232 struct range *range = &dev_dax->ranges[i].range; 233 234 pgmap->ranges[i] = *range; 235 } 236 } 237 238 for (i = 0; i < dev_dax->nr_range; i++) { 239 struct range *range = &dev_dax->ranges[i].range; 240 241 if (!devm_request_mem_region(dev, range->start, 242 range_len(range), dev_name(dev))) { 243 dev_warn(dev, "mapping%d: %#llx-%#llx could not reserve range\n", 244 i, range->start, range->end); 245 return -EBUSY; 246 } 247 } 248 249 /* Cache size now; it cannot change while driver is bound */ 250 dev_dax->cached_size = 0; 251 for (i = 0; i < dev_dax->nr_range; i++) 252 dev_dax->cached_size += range_len(&dev_dax->ranges[i].range); 253 254 /* 255 * Use MEMORY_DEVICE_FS_DAX without setting vmemmap_shift, leaving 256 * folios at order-0. Unlike device.c (MEMORY_DEVICE_GENERIC), this 257 * lets fs-dax dynamically build compound folios as needed, similar 258 * to pmem behavior. 259 */ 260 pgmap->type = MEMORY_DEVICE_FS_DAX; 261 pgmap->ops = &fsdev_pagemap_ops; 262 pgmap->owner = dev_dax; 263 264 addr = devm_memremap_pages(dev, pgmap); 265 if (IS_ERR(addr)) 266 return PTR_ERR(addr); 267 268 /* 269 * Clear any stale compound folio state left over from a previous 270 * driver (e.g., device_dax with vmemmap_shift). Also register this 271 * as a devm action so folio state is cleared on unbind, ensuring 272 * clean pages for subsequent drivers (e.g., kmem for system-ram). 273 */ 274 fsdev_clear_folio_state(dev_dax); 275 rc = devm_add_action_or_reset(dev, fsdev_clear_folio_state_action, 276 dev_dax); 277 if (rc) 278 return rc; 279 280 /* Detect whether the data is at a non-zero offset into the memory */ 281 if (pgmap->range.start != dev_dax->ranges[0].range.start) { 282 u64 phys = dev_dax->ranges[0].range.start; 283 u64 pgmap_phys = dev_dax->pgmap[0].range.start; 284 285 if (!WARN_ON(pgmap_phys > phys)) 286 data_offset = phys - pgmap_phys; 287 288 pr_debug("%s: offset detected phys=%llx pgmap_phys=%llx offset=%llx\n", 289 __func__, phys, pgmap_phys, data_offset); 290 } 291 dev_dax->virt_addr = addr + data_offset; 292 293 inode = dax_inode(dax_dev); 294 cdev = inode->i_cdev; 295 cdev_init(cdev, &fsdev_fops); 296 cdev->owner = dev->driver->owner; 297 cdev_set_parent(cdev, &dev->kobj); 298 rc = cdev_add(cdev, dev->devt, 1); 299 if (rc) 300 return rc; 301 302 rc = devm_add_action_or_reset(dev, fsdev_cdev_del, cdev); 303 if (rc) 304 return rc; 305 306 run_dax(dax_dev); 307 return devm_add_action_or_reset(dev, fsdev_kill, dev_dax); 308 } 309 310 static struct dax_device_driver fsdev_dax_driver = { 311 .probe = fsdev_dax_probe, 312 .type = DAXDRV_FSDEV_TYPE, 313 }; 314 315 static int __init dax_init(void) 316 { 317 return dax_driver_register(&fsdev_dax_driver); 318 } 319 320 static void __exit dax_exit(void) 321 { 322 dax_driver_unregister(&fsdev_dax_driver); 323 } 324 325 MODULE_AUTHOR("John Groves"); 326 MODULE_DESCRIPTION("FS-DAX Device: fs-dax compatible devdax driver"); 327 MODULE_LICENSE("GPL"); 328 module_init(dax_init); 329 module_exit(dax_exit); 330 MODULE_ALIAS_DAX_DEVICE(0); 331