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 void dax_set_mapping(struct vm_fault *vmf, pfn_t pfn,
  77                               unsigned long fault_size)
  78 {
  79         unsigned long i, nr_pages = fault_size / PAGE_SIZE;
  80         struct file *filp = vmf->vma->vm_file;
  81         struct dev_dax *dev_dax = filp->private_data;
  82         pgoff_t pgoff;
  83
  84         /* mapping is only set on the head */
  85         if (dev_dax->pgmap->vmemmap_shift)
  86                 nr_pages = 1;
  87
  88         pgoff = linear_page_index(vmf->vma,
  89                         ALIGN_DOWN(vmf->address, fault_size));
  90
  91         for (i = 0; i < nr_pages; i++) {
  92                 struct folio *folio = pfn_folio(pfn_t_to_pfn(pfn) + i);
  93
  94                 if (folio->mapping)
  95                         continue;
  96
  97                 folio->mapping = filp->f_mapping;
  98                 folio->index = pgoff + i;
  99         }
 100 }
 101
 102 static vm_fault_t __dev_dax_pte_fault(struct dev_dax *dev_dax,
 103                                 struct vm_fault *vmf)
 104 {
 105         struct device *dev = &dev_dax->dev;
 106         phys_addr_t phys;
 107         pfn_t pfn;
 108         unsigned int fault_size = PAGE_SIZE;
 109
 110         if (check_vma(dev_dax, vmf->vma, __func__))
 111                 return VM_FAULT_SIGBUS;
 112
 113         if (dev_dax->align > PAGE_SIZE) {
 114                 dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
 115                         dev_dax->align, fault_size);
 116                 return VM_FAULT_SIGBUS;
 117         }
 118
 119         if (fault_size != dev_dax->align)
 120                 return VM_FAULT_SIGBUS;
 121
 122         phys = dax_pgoff_to_phys(dev_dax, vmf->pgoff, PAGE_SIZE);
 123         if (phys == -1) {
 124                 dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", vmf->pgoff);
 125                 return VM_FAULT_SIGBUS;
 126         }
 127
 128         pfn = phys_to_pfn_t(phys, 0);
 129
 130         dax_set_mapping(vmf, pfn, fault_size);
 131
 132         return vmf_insert_page_mkwrite(vmf, pfn_t_to_page(pfn),
 133                                         vmf->flags & FAULT_FLAG_WRITE);
 134 }
 135
 136 static vm_fault_t __dev_dax_pmd_fault(struct dev_dax *dev_dax,
 137                                 struct vm_fault *vmf)
 138 {
 139         unsigned long pmd_addr = vmf->address & PMD_MASK;
 140         struct device *dev = &dev_dax->dev;
 141         phys_addr_t phys;
 142         pgoff_t pgoff;
 143         pfn_t pfn;
 144         unsigned int fault_size = PMD_SIZE;
 145
 146         if (check_vma(dev_dax, vmf->vma, __func__))
 147                 return VM_FAULT_SIGBUS;
 148
 149         if (dev_dax->align > PMD_SIZE) {
 150                 dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
 151                         dev_dax->align, fault_size);
 152                 return VM_FAULT_SIGBUS;
 153         }
 154
 155         if (fault_size < dev_dax->align)
 156                 return VM_FAULT_SIGBUS;
 157         else if (fault_size > dev_dax->align)
 158                 return VM_FAULT_FALLBACK;
 159
 160         /* if we are outside of the VMA */
 161         if (pmd_addr < vmf->vma->vm_start ||
 162                         (pmd_addr + PMD_SIZE) > vmf->vma->vm_end)
 163                 return VM_FAULT_SIGBUS;
 164
 165         pgoff = linear_page_index(vmf->vma, pmd_addr);
 166         phys = dax_pgoff_to_phys(dev_dax, pgoff, PMD_SIZE);
 167         if (phys == -1) {
 168                 dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", pgoff);
 169                 return VM_FAULT_SIGBUS;
 170         }
 171
 172         pfn = phys_to_pfn_t(phys, 0);
 173
 174         dax_set_mapping(vmf, pfn, fault_size);
 175
 176         return vmf_insert_folio_pmd(vmf, page_folio(pfn_t_to_page(pfn)),
 177                                 vmf->flags & FAULT_FLAG_WRITE);
 178 }
 179
 180 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 181 static vm_fault_t __dev_dax_pud_fault(struct dev_dax *dev_dax,
 182                                 struct vm_fault *vmf)
 183 {
 184         unsigned long pud_addr = vmf->address & PUD_MASK;
 185         struct device *dev = &dev_dax->dev;
 186         phys_addr_t phys;
 187         pgoff_t pgoff;
 188         pfn_t pfn;
 189         unsigned int fault_size = PUD_SIZE;
 190
 191
 192         if (check_vma(dev_dax, vmf->vma, __func__))
 193                 return VM_FAULT_SIGBUS;
 194
 195         if (dev_dax->align > PUD_SIZE) {
 196                 dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
 197                         dev_dax->align, fault_size);
 198                 return VM_FAULT_SIGBUS;
 199         }
 200
 201         if (fault_size < dev_dax->align)
 202                 return VM_FAULT_SIGBUS;
 203         else if (fault_size > dev_dax->align)
 204                 return VM_FAULT_FALLBACK;
 205
 206         /* if we are outside of the VMA */
 207         if (pud_addr < vmf->vma->vm_start ||
 208                         (pud_addr + PUD_SIZE) > vmf->vma->vm_end)
 209                 return VM_FAULT_SIGBUS;
 210
 211         pgoff = linear_page_index(vmf->vma, pud_addr);
 212         phys = dax_pgoff_to_phys(dev_dax, pgoff, PUD_SIZE);
 213         if (phys == -1) {
 214                 dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", pgoff);
 215                 return VM_FAULT_SIGBUS;
 216         }
 217
 218         pfn = phys_to_pfn_t(phys, 0);
 219
 220         dax_set_mapping(vmf, pfn, fault_size);
 221
 222         return vmf_insert_folio_pud(vmf, page_folio(pfn_t_to_page(pfn)),
 223                                 vmf->flags & FAULT_FLAG_WRITE);
 224 }
 225 #else
 226 static vm_fault_t __dev_dax_pud_fault(struct dev_dax *dev_dax,
 227                                 struct vm_fault *vmf)
 228 {
 229         return VM_FAULT_FALLBACK;
 230 }
 231 #endif /* !CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 232
 233 static vm_fault_t dev_dax_huge_fault(struct vm_fault *vmf, unsigned int order)
 234 {
 235         struct file *filp = vmf->vma->vm_file;
 236         vm_fault_t rc = VM_FAULT_SIGBUS;
 237         int id;
 238         struct dev_dax *dev_dax = filp->private_data;
 239
 240         dev_dbg(&dev_dax->dev, "%s: op=%s addr=%#lx order=%d\n", current->comm,
 241                 (vmf->flags & FAULT_FLAG_WRITE) ? "write" : "read",
 242                 vmf->address & ~((1UL << (order + PAGE_SHIFT)) - 1), order);
 243
 244         id = dax_read_lock();
 245         if (order == 0)
 246                 rc = __dev_dax_pte_fault(dev_dax, vmf);
 247         else if (order == PMD_ORDER)
 248                 rc = __dev_dax_pmd_fault(dev_dax, vmf);
 249         else if (order == PUD_ORDER)
 250                 rc = __dev_dax_pud_fault(dev_dax, vmf);
 251         else
 252                 rc = VM_FAULT_SIGBUS;
 253
 254         dax_read_unlock(id);
 255
 256         return rc;
 257 }
 258
 259 static vm_fault_t dev_dax_fault(struct vm_fault *vmf)
 260 {
 261         return dev_dax_huge_fault(vmf, 0);
 262 }
 263
 264 static int dev_dax_may_split(struct vm_area_struct *vma, unsigned long addr)
 265 {
 266         struct file *filp = vma->vm_file;
 267         struct dev_dax *dev_dax = filp->private_data;
 268
 269         if (!IS_ALIGNED(addr, dev_dax->align))
 270                 return -EINVAL;
 271         return 0;
 272 }
 273
 274 static unsigned long dev_dax_pagesize(struct vm_area_struct *vma)
 275 {
 276         struct file *filp = vma->vm_file;
 277         struct dev_dax *dev_dax = filp->private_data;
 278
 279         return dev_dax->align;
 280 }
 281
 282 static const struct vm_operations_struct dax_vm_ops = {
 283         .fault = dev_dax_fault,
 284         .huge_fault = dev_dax_huge_fault,
 285         .may_split = dev_dax_may_split,
 286         .pagesize = dev_dax_pagesize,
 287 };
 288
 289 static int dax_mmap(struct file *filp, struct vm_area_struct *vma)
 290 {
 291         struct dev_dax *dev_dax = filp->private_data;
 292         int rc, id;
 293
 294         dev_dbg(&dev_dax->dev, "trace\n");
 295
 296         /*
 297          * We lock to check dax_dev liveness and will re-check at
 298          * fault time.
 299          */
 300         id = dax_read_lock();
 301         rc = check_vma(dev_dax, vma, __func__);
 302         dax_read_unlock(id);
 303         if (rc)
 304                 return rc;
 305
 306         vma->vm_ops = &dax_vm_ops;
 307         vm_flags_set(vma, VM_HUGEPAGE);
 308         return 0;
 309 }
 310
 311 /* return an unmapped area aligned to the dax region specified alignment */
 312 static unsigned long dax_get_unmapped_area(struct file *filp,
 313                 unsigned long addr, unsigned long len, unsigned long pgoff,
 314                 unsigned long flags)
 315 {
 316         unsigned long off, off_end, off_align, len_align, addr_align, align;
 317         struct dev_dax *dev_dax = filp ? filp->private_data : NULL;
 318
 319         if (!dev_dax || addr)
 320                 goto out;
 321
 322         align = dev_dax->align;
 323         off = pgoff << PAGE_SHIFT;
 324         off_end = off + len;
 325         off_align = round_up(off, align);
 326
 327         if ((off_end <= off_align) || ((off_end - off_align) < align))
 328                 goto out;
 329
 330         len_align = len + align;
 331         if ((off + len_align) < off)
 332                 goto out;
 333
 334         addr_align = mm_get_unmapped_area(current->mm, filp, addr, len_align,
 335                                           pgoff, flags);
 336         if (!IS_ERR_VALUE(addr_align)) {
 337                 addr_align += (off - addr_align) & (align - 1);
 338                 return addr_align;
 339         }
 340  out:
 341         return mm_get_unmapped_area(current->mm, filp, addr, len, pgoff, flags);
 342 }
 343
 344 static const struct address_space_operations dev_dax_aops = {
 345         .dirty_folio    = noop_dirty_folio,
 346 };
 347
 348 static int dax_open(struct inode *inode, struct file *filp)
 349 {
 350         struct dax_device *dax_dev = inode_dax(inode);
 351         struct inode *__dax_inode = dax_inode(dax_dev);
 352         struct dev_dax *dev_dax = dax_get_private(dax_dev);
 353
 354         dev_dbg(&dev_dax->dev, "trace\n");
 355         inode->i_mapping = __dax_inode->i_mapping;
 356         inode->i_mapping->host = __dax_inode;
 357         inode->i_mapping->a_ops = &dev_dax_aops;
 358         filp->f_mapping = inode->i_mapping;
 359         filp->f_wb_err = filemap_sample_wb_err(filp->f_mapping);
 360         filp->f_sb_err = file_sample_sb_err(filp);
 361         filp->private_data = dev_dax;
 362         inode->i_flags = S_DAX;
 363
 364         return 0;
 365 }
 366
 367 static int dax_release(struct inode *inode, struct file *filp)
 368 {
 369         struct dev_dax *dev_dax = filp->private_data;
 370
 371         dev_dbg(&dev_dax->dev, "trace\n");
 372         return 0;
 373 }
 374
 375 static const struct file_operations dax_fops = {
 376         .llseek = noop_llseek,
 377         .owner = THIS_MODULE,
 378         .open = dax_open,
 379         .release = dax_release,
 380         .get_unmapped_area = dax_get_unmapped_area,
 381         .mmap = dax_mmap,
 382         .fop_flags = FOP_MMAP_SYNC,
 383 };
 384
 385 static void dev_dax_cdev_del(void *cdev)
 386 {
 387         cdev_del(cdev);
 388 }
 389
 390 static void dev_dax_kill(void *dev_dax)
 391 {
 392         kill_dev_dax(dev_dax);
 393 }
 394
 395 static int dev_dax_probe(struct dev_dax *dev_dax)
 396 {
 397         struct dax_device *dax_dev = dev_dax->dax_dev;
 398         struct device *dev = &dev_dax->dev;
 399         struct dev_pagemap *pgmap;
 400         struct inode *inode;
 401         struct cdev *cdev;
 402         void *addr;
 403         int rc, i;
 404
 405         if (static_dev_dax(dev_dax))  {
 406                 if (dev_dax->nr_range > 1) {
 407                         dev_warn(dev,
 408                                 "static pgmap / multi-range device conflict\n");
 409                         return -EINVAL;
 410                 }
 411
 412                 pgmap = dev_dax->pgmap;
 413         } else {
 414                 if (dev_dax->pgmap) {
 415                         dev_warn(dev,
 416                                  "dynamic-dax with pre-populated page map\n");
 417                         return -EINVAL;
 418                 }
 419
 420                 pgmap = devm_kzalloc(dev,
 421                        struct_size(pgmap, ranges, dev_dax->nr_range - 1),
 422                        GFP_KERNEL);
 423                 if (!pgmap)
 424                         return -ENOMEM;
 425
 426                 pgmap->nr_range = dev_dax->nr_range;
 427                 dev_dax->pgmap = pgmap;
 428
 429                 for (i = 0; i < dev_dax->nr_range; i++) {
 430                         struct range *range = &dev_dax->ranges[i].range;
 431                         pgmap->ranges[i] = *range;
 432                 }
 433         }
 434
 435         for (i = 0; i < dev_dax->nr_range; i++) {
 436                 struct range *range = &dev_dax->ranges[i].range;
 437
 438                 if (!devm_request_mem_region(dev, range->start,
 439                                         range_len(range), dev_name(dev))) {
 440                         dev_warn(dev, "mapping%d: %#llx-%#llx could not reserve range\n",
 441                                         i, range->start, range->end);
 442                         return -EBUSY;
 443                 }
 444         }
 445
 446         pgmap->type = MEMORY_DEVICE_GENERIC;
 447         if (dev_dax->align > PAGE_SIZE)
 448                 pgmap->vmemmap_shift =
 449                         order_base_2(dev_dax->align >> PAGE_SHIFT);
 450         addr = devm_memremap_pages(dev, pgmap);
 451         if (IS_ERR(addr))
 452                 return PTR_ERR(addr);
 453
 454         inode = dax_inode(dax_dev);
 455         cdev = inode->i_cdev;
 456         cdev_init(cdev, &dax_fops);
 457         cdev->owner = dev->driver->owner;
 458         cdev_set_parent(cdev, &dev->kobj);
 459         rc = cdev_add(cdev, dev->devt, 1);
 460         if (rc)
 461                 return rc;
 462
 463         rc = devm_add_action_or_reset(dev, dev_dax_cdev_del, cdev);
 464         if (rc)
 465                 return rc;
 466
 467         run_dax(dax_dev);
 468         return devm_add_action_or_reset(dev, dev_dax_kill, dev_dax);
 469 }
 470
 471 static struct dax_device_driver device_dax_driver = {
 472         .probe = dev_dax_probe,
 473         .type = DAXDRV_DEVICE_TYPE,
 474 };
 475
 476 static int __init dax_init(void)
 477 {
 478         return dax_driver_register(&device_dax_driver);
 479 }
 480
 481 static void __exit dax_exit(void)
 482 {
 483         dax_driver_unregister(&device_dax_driver);
 484 }
 485
 486 MODULE_AUTHOR("Intel Corporation");
 487 MODULE_DESCRIPTION("Device DAX: direct access device driver");
 488 MODULE_LICENSE("GPL v2");
 489 module_init(dax_init);
 490 module_exit(dax_exit);
 491 MODULE_ALIAS_DAX_DEVICE(0);