drivers/dax/device.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Copyright(c) 2016-2018 Intel Corporation. All rights reserved. */
   3 #include <linux/memremap.h>
   4 #include <linux/pagemap.h>
   5 #include <linux/module.h>
   6 #include <linux/device.h>
   7 #include <linux/pfn_t.h>
   8 #include <linux/cdev.h>
   9 #include <linux/slab.h>
  10 #include <linux/dax.h>
  11 #include <linux/fs.h>
  12 #include <linux/mm.h>
  13 #include <linux/mman.h>
  14 #include "dax-private.h"
  15 #include "bus.h"
  16
  17 static int check_vma(struct dev_dax *dev_dax, struct vm_area_struct *vma,
  18                 const char *func)
  19 {
  20         struct device *dev = &dev_dax->dev;
  21         unsigned long mask;
  22
  23         if (!dax_alive(dev_dax->dax_dev))
  24                 return -ENXIO;
  25
  26         /* prevent private mappings from being established */
  27         if ((vma->vm_flags & VM_MAYSHARE) != VM_MAYSHARE) {
  28                 dev_info_ratelimited(dev,
  29                                 "%s: %s: fail, attempted private mapping\n",
  30                                 current->comm, func);
  31                 return -EINVAL;
  32         }
  33
  34         mask = dev_dax->align - 1;
  35         if (vma->vm_start & mask || vma->vm_end & mask) {
  36                 dev_info_ratelimited(dev,
  37                                 "%s: %s: fail, unaligned vma (%#lx - %#lx, %#lx)\n",
  38                                 current->comm, func, vma->vm_start, vma->vm_end,
  39                                 mask);
  40                 return -EINVAL;
  41         }
  42
  43         if (!vma_is_dax(vma)) {
  44                 dev_info_ratelimited(dev,
  45                                 "%s: %s: fail, vma is not DAX capable\n",
  46                                 current->comm, func);
  47                 return -EINVAL;
  48         }
  49
  50         return 0;
  51 }
  52
  53 /* see "strong" declaration in tools/testing/nvdimm/dax-dev.c */
  54 __weak phys_addr_t dax_pgoff_to_phys(struct dev_dax *dev_dax, pgoff_t pgoff,
  55                 unsigned long size)
  56 {
  57         int i;
  58
  59         for (i = 0; i < dev_dax->nr_range; i++) {
  60                 struct dev_dax_range *dax_range = &dev_dax->ranges[i];
  61                 struct range *range = &dax_range->range;
  62                 unsigned long long pgoff_end;
  63                 phys_addr_t phys;
  64
  65                 pgoff_end = dax_range->pgoff + PHYS_PFN(range_len(range)) - 1;
  66                 if (pgoff < dax_range->pgoff || pgoff > pgoff_end)
  67                         continue;
  68                 phys = PFN_PHYS(pgoff - dax_range->pgoff) + range->start;
  69                 if (phys + size - 1 <= range->end)
  70                         return phys;
  71                 break;
  72         }
  73         return -1;
  74 }
  75
  76 static vm_fault_t __dev_dax_pte_fault(struct dev_dax *dev_dax,
  77                                 struct vm_fault *vmf, pfn_t *pfn)
  78 {
  79         struct device *dev = &dev_dax->dev;
  80         phys_addr_t phys;
  81         unsigned int fault_size = PAGE_SIZE;
  82
  83         if (check_vma(dev_dax, vmf->vma, __func__))
  84                 return VM_FAULT_SIGBUS;
  85
  86         if (dev_dax->align > PAGE_SIZE) {
  87                 dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
  88                         dev_dax->align, fault_size);
  89                 return VM_FAULT_SIGBUS;
  90         }
  91
  92         if (fault_size != dev_dax->align)
  93                 return VM_FAULT_SIGBUS;
  94
  95         phys = dax_pgoff_to_phys(dev_dax, vmf->pgoff, PAGE_SIZE);
  96         if (phys == -1) {
  97                 dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", vmf->pgoff);
  98                 return VM_FAULT_SIGBUS;
  99         }
 100
 101         *pfn = phys_to_pfn_t(phys, PFN_DEV|PFN_MAP);
 102
 103         return vmf_insert_mixed(vmf->vma, vmf->address, *pfn);
 104 }
 105
 106 static vm_fault_t __dev_dax_pmd_fault(struct dev_dax *dev_dax,
 107                                 struct vm_fault *vmf, pfn_t *pfn)
 108 {
 109         unsigned long pmd_addr = vmf->address & PMD_MASK;
 110         struct device *dev = &dev_dax->dev;
 111         phys_addr_t phys;
 112         pgoff_t pgoff;
 113         unsigned int fault_size = PMD_SIZE;
 114
 115         if (check_vma(dev_dax, vmf->vma, __func__))
 116                 return VM_FAULT_SIGBUS;
 117
 118         if (dev_dax->align > PMD_SIZE) {
 119                 dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
 120                         dev_dax->align, fault_size);
 121                 return VM_FAULT_SIGBUS;
 122         }
 123
 124         if (fault_size < dev_dax->align)
 125                 return VM_FAULT_SIGBUS;
 126         else if (fault_size > dev_dax->align)
 127                 return VM_FAULT_FALLBACK;
 128
 129         /* if we are outside of the VMA */
 130         if (pmd_addr < vmf->vma->vm_start ||
 131                         (pmd_addr + PMD_SIZE) > vmf->vma->vm_end)
 132                 return VM_FAULT_SIGBUS;
 133
 134         pgoff = linear_page_index(vmf->vma, pmd_addr);
 135         phys = dax_pgoff_to_phys(dev_dax, pgoff, PMD_SIZE);
 136         if (phys == -1) {
 137                 dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", pgoff);
 138                 return VM_FAULT_SIGBUS;
 139         }
 140
 141         *pfn = phys_to_pfn_t(phys, PFN_DEV|PFN_MAP);
 142
 143         return vmf_insert_pfn_pmd(vmf, *pfn, vmf->flags & FAULT_FLAG_WRITE);
 144 }
 145
 146 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 147 static vm_fault_t __dev_dax_pud_fault(struct dev_dax *dev_dax,
 148                                 struct vm_fault *vmf, pfn_t *pfn)
 149 {
 150         unsigned long pud_addr = vmf->address & PUD_MASK;
 151         struct device *dev = &dev_dax->dev;
 152         phys_addr_t phys;
 153         pgoff_t pgoff;
 154         unsigned int fault_size = PUD_SIZE;
 155
 156
 157         if (check_vma(dev_dax, vmf->vma, __func__))
 158                 return VM_FAULT_SIGBUS;
 159
 160         if (dev_dax->align > PUD_SIZE) {
 161                 dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
 162                         dev_dax->align, fault_size);
 163                 return VM_FAULT_SIGBUS;
 164         }
 165
 166         if (fault_size < dev_dax->align)
 167                 return VM_FAULT_SIGBUS;
 168         else if (fault_size > dev_dax->align)
 169                 return VM_FAULT_FALLBACK;
 170
 171         /* if we are outside of the VMA */
 172         if (pud_addr < vmf->vma->vm_start ||
 173                         (pud_addr + PUD_SIZE) > vmf->vma->vm_end)
 174                 return VM_FAULT_SIGBUS;
 175
 176         pgoff = linear_page_index(vmf->vma, pud_addr);
 177         phys = dax_pgoff_to_phys(dev_dax, pgoff, PUD_SIZE);
 178         if (phys == -1) {
 179                 dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", pgoff);
 180                 return VM_FAULT_SIGBUS;
 181         }
 182
 183         *pfn = phys_to_pfn_t(phys, PFN_DEV|PFN_MAP);
 184
 185         return vmf_insert_pfn_pud(vmf, *pfn, vmf->flags & FAULT_FLAG_WRITE);
 186 }
 187 #else
 188 static vm_fault_t __dev_dax_pud_fault(struct dev_dax *dev_dax,
 189                                 struct vm_fault *vmf, pfn_t *pfn)
 190 {
 191         return VM_FAULT_FALLBACK;
 192 }
 193 #endif /* !CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 194
 195 static vm_fault_t dev_dax_huge_fault(struct vm_fault *vmf,
 196                 enum page_entry_size pe_size)
 197 {
 198         struct file *filp = vmf->vma->vm_file;
 199         unsigned long fault_size;
 200         vm_fault_t rc = VM_FAULT_SIGBUS;
 201         int id;
 202         pfn_t pfn;
 203         struct dev_dax *dev_dax = filp->private_data;
 204
 205         dev_dbg(&dev_dax->dev, "%s: %s (%#lx - %#lx) size = %d\n", current->comm,
 206                         (vmf->flags & FAULT_FLAG_WRITE) ? "write" : "read",
 207                         vmf->vma->vm_start, vmf->vma->vm_end, pe_size);
 208
 209         id = dax_read_lock();
 210         switch (pe_size) {
 211         case PE_SIZE_PTE:
 212                 fault_size = PAGE_SIZE;
 213                 rc = __dev_dax_pte_fault(dev_dax, vmf, &pfn);
 214                 break;
 215         case PE_SIZE_PMD:
 216                 fault_size = PMD_SIZE;
 217                 rc = __dev_dax_pmd_fault(dev_dax, vmf, &pfn);
 218                 break;
 219         case PE_SIZE_PUD:
 220                 fault_size = PUD_SIZE;
 221                 rc = __dev_dax_pud_fault(dev_dax, vmf, &pfn);
 222                 break;
 223         default:
 224                 rc = VM_FAULT_SIGBUS;
 225         }
 226
 227         if (rc == VM_FAULT_NOPAGE) {
 228                 unsigned long i;
 229                 pgoff_t pgoff;
 230
 231                 /*
 232                  * In the device-dax case the only possibility for a
 233                  * VM_FAULT_NOPAGE result is when device-dax capacity is
 234                  * mapped. No need to consider the zero page, or racing
 235                  * conflicting mappings.
 236                  */
 237                 pgoff = linear_page_index(vmf->vma, vmf->address
 238                                 & ~(fault_size - 1));
 239                 for (i = 0; i < fault_size / PAGE_SIZE; i++) {
 240                         struct page *page;
 241
 242                         page = pfn_to_page(pfn_t_to_pfn(pfn) + i);
 243                         if (page->mapping)
 244                                 continue;
 245                         page->mapping = filp->f_mapping;
 246                         page->index = pgoff + i;
 247                 }
 248         }
 249         dax_read_unlock(id);
 250
 251         return rc;
 252 }
 253
 254 static vm_fault_t dev_dax_fault(struct vm_fault *vmf)
 255 {
 256         return dev_dax_huge_fault(vmf, PE_SIZE_PTE);
 257 }
 258
 259 static int dev_dax_may_split(struct vm_area_struct *vma, unsigned long addr)
 260 {
 261         struct file *filp = vma->vm_file;
 262         struct dev_dax *dev_dax = filp->private_data;
 263
 264         if (!IS_ALIGNED(addr, dev_dax->align))
 265                 return -EINVAL;
 266         return 0;
 267 }
 268
 269 static unsigned long dev_dax_pagesize(struct vm_area_struct *vma)
 270 {
 271         struct file *filp = vma->vm_file;
 272         struct dev_dax *dev_dax = filp->private_data;
 273
 274         return dev_dax->align;
 275 }
 276
 277 static const struct vm_operations_struct dax_vm_ops = {
 278         .fault = dev_dax_fault,
 279         .huge_fault = dev_dax_huge_fault,
 280         .may_split = dev_dax_may_split,
 281         .pagesize = dev_dax_pagesize,
 282 };
 283
 284 static int dax_mmap(struct file *filp, struct vm_area_struct *vma)
 285 {
 286         struct dev_dax *dev_dax = filp->private_data;
 287         int rc, id;
 288
 289         dev_dbg(&dev_dax->dev, "trace\n");
 290
 291         /*
 292          * We lock to check dax_dev liveness and will re-check at
 293          * fault time.
 294          */
 295         id = dax_read_lock();
 296         rc = check_vma(dev_dax, vma, __func__);
 297         dax_read_unlock(id);
 298         if (rc)
 299                 return rc;
 300
 301         vma->vm_ops = &dax_vm_ops;
 302         vma->vm_flags |= VM_HUGEPAGE;
 303         return 0;
 304 }
 305
 306 /* return an unmapped area aligned to the dax region specified alignment */
 307 static unsigned long dax_get_unmapped_area(struct file *filp,
 308                 unsigned long addr, unsigned long len, unsigned long pgoff,
 309                 unsigned long flags)
 310 {
 311         unsigned long off, off_end, off_align, len_align, addr_align, align;
 312         struct dev_dax *dev_dax = filp ? filp->private_data : NULL;
 313
 314         if (!dev_dax || addr)
 315                 goto out;
 316
 317         align = dev_dax->align;
 318         off = pgoff << PAGE_SHIFT;
 319         off_end = off + len;
 320         off_align = round_up(off, align);
 321
 322         if ((off_end <= off_align) || ((off_end - off_align) < align))
 323                 goto out;
 324
 325         len_align = len + align;
 326         if ((off + len_align) < off)
 327                 goto out;
 328
 329         addr_align = current->mm->get_unmapped_area(filp, addr, len_align,
 330                         pgoff, flags);
 331         if (!IS_ERR_VALUE(addr_align)) {
 332                 addr_align += (off - addr_align) & (align - 1);
 333                 return addr_align;
 334         }
 335  out:
 336         return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
 337 }
 338
 339 static const struct address_space_operations dev_dax_aops = {
 340         .set_page_dirty         = __set_page_dirty_no_writeback,
 341         .invalidatepage         = noop_invalidatepage,
 342 };
 343
 344 static int dax_open(struct inode *inode, struct file *filp)
 345 {
 346         struct dax_device *dax_dev = inode_dax(inode);
 347         struct inode *__dax_inode = dax_inode(dax_dev);
 348         struct dev_dax *dev_dax = dax_get_private(dax_dev);
 349
 350         dev_dbg(&dev_dax->dev, "trace\n");
 351         inode->i_mapping = __dax_inode->i_mapping;
 352         inode->i_mapping->host = __dax_inode;
 353         inode->i_mapping->a_ops = &dev_dax_aops;
 354         filp->f_mapping = inode->i_mapping;
 355         filp->f_wb_err = filemap_sample_wb_err(filp->f_mapping);
 356         filp->f_sb_err = file_sample_sb_err(filp);
 357         filp->private_data = dev_dax;
 358         inode->i_flags = S_DAX;
 359
 360         return 0;
 361 }
 362
 363 static int dax_release(struct inode *inode, struct file *filp)
 364 {
 365         struct dev_dax *dev_dax = filp->private_data;
 366
 367         dev_dbg(&dev_dax->dev, "trace\n");
 368         return 0;
 369 }
 370
 371 static const struct file_operations dax_fops = {
 372         .llseek = noop_llseek,
 373         .owner = THIS_MODULE,
 374         .open = dax_open,
 375         .release = dax_release,
 376         .get_unmapped_area = dax_get_unmapped_area,
 377         .mmap = dax_mmap,
 378         .mmap_supported_flags = MAP_SYNC,
 379 };
 380
 381 static void dev_dax_cdev_del(void *cdev)
 382 {
 383         cdev_del(cdev);
 384 }
 385
 386 static void dev_dax_kill(void *dev_dax)
 387 {
 388         kill_dev_dax(dev_dax);
 389 }
 390
 391 int dev_dax_probe(struct dev_dax *dev_dax)
 392 {
 393         struct dax_device *dax_dev = dev_dax->dax_dev;
 394         struct device *dev = &dev_dax->dev;
 395         struct dev_pagemap *pgmap;
 396         struct inode *inode;
 397         struct cdev *cdev;
 398         void *addr;
 399         int rc, i;
 400
 401         pgmap = dev_dax->pgmap;
 402         if (dev_WARN_ONCE(dev, pgmap && dev_dax->nr_range > 1,
 403                         "static pgmap / multi-range device conflict\n"))
 404                 return -EINVAL;
 405
 406         if (!pgmap) {
 407                 pgmap = devm_kzalloc(dev, sizeof(*pgmap) + sizeof(struct range)
 408                                 * (dev_dax->nr_range - 1), GFP_KERNEL);
 409                 if (!pgmap)
 410                         return -ENOMEM;
 411                 pgmap->nr_range = dev_dax->nr_range;
 412         }
 413
 414         for (i = 0; i < dev_dax->nr_range; i++) {
 415                 struct range *range = &dev_dax->ranges[i].range;
 416
 417                 if (!devm_request_mem_region(dev, range->start,
 418                                         range_len(range), dev_name(dev))) {
 419                         dev_warn(dev, "mapping%d: %#llx-%#llx could not reserve range\n",
 420                                         i, range->start, range->end);
 421                         return -EBUSY;
 422                 }
 423                 /* don't update the range for static pgmap */
 424                 if (!dev_dax->pgmap)
 425                         pgmap->ranges[i] = *range;
 426         }
 427
 428         pgmap->type = MEMORY_DEVICE_GENERIC;
 429         addr = devm_memremap_pages(dev, pgmap);
 430         if (IS_ERR(addr))
 431                 return PTR_ERR(addr);
 432
 433         inode = dax_inode(dax_dev);
 434         cdev = inode->i_cdev;
 435         cdev_init(cdev, &dax_fops);
 436         if (dev->class) {
 437                 /* for the CONFIG_DEV_DAX_PMEM_COMPAT case */
 438                 cdev->owner = dev->parent->driver->owner;
 439         } else
 440                 cdev->owner = dev->driver->owner;
 441         cdev_set_parent(cdev, &dev->kobj);
 442         rc = cdev_add(cdev, dev->devt, 1);
 443         if (rc)
 444                 return rc;
 445
 446         rc = devm_add_action_or_reset(dev, dev_dax_cdev_del, cdev);
 447         if (rc)
 448                 return rc;
 449
 450         run_dax(dax_dev);
 451         return devm_add_action_or_reset(dev, dev_dax_kill, dev_dax);
 452 }
 453 EXPORT_SYMBOL_GPL(dev_dax_probe);
 454
 455 static struct dax_device_driver device_dax_driver = {
 456         .probe = dev_dax_probe,
 457         /* all probe actions are unwound by devm, so .remove isn't necessary */
 458         .match_always = 1,
 459 };
 460
 461 static int __init dax_init(void)
 462 {
 463         return dax_driver_register(&device_dax_driver);
 464 }
 465
 466 static void __exit dax_exit(void)
 467 {
 468         dax_driver_unregister(&device_dax_driver);
 469 }
 470
 471 MODULE_AUTHOR("Intel Corporation");
 472 MODULE_LICENSE("GPL v2");
 473 module_init(dax_init);
 474 module_exit(dax_exit);
 475 MODULE_ALIAS_DAX_DEVICE(0);