drivers/dax/device.c

0001 // SPDX-License-Identifier: GPL-2.0
0002 /* Copyright(c) 2016-2018 Intel Corporation. All rights reserved. */
0003 #include <linux/memremap.h>
0004 #include <linux/pagemap.h>
0005 #include <linux/module.h>
0006 #include <linux/device.h>
0007 #include <linux/pfn_t.h>
0008 #include <linux/cdev.h>
0009 #include <linux/slab.h>
0010 #include <linux/dax.h>
0011 #include <linux/fs.h>
0012 #include <linux/mm.h>
0013 #include <linux/mman.h>
0014 #include "dax-private.h"
0015 #include "bus.h"
0016
0017 static int check_vma(struct dev_dax *dev_dax, struct vm_area_struct *vma,
0018         const char *func)
0019 {
0020     struct device *dev = &dev_dax->dev;
0021     unsigned long mask;
0022
0023     if (!dax_alive(dev_dax->dax_dev))
0024         return -ENXIO;
0025
0026     /* prevent private mappings from being established */
0027     if ((vma->vm_flags & VM_MAYSHARE) != VM_MAYSHARE) {
0028         dev_info_ratelimited(dev,
0029                 "%s: %s: fail, attempted private mapping\n",
0030                 current->comm, func);
0031         return -EINVAL;
0032     }
0033
0034     mask = dev_dax->align - 1;
0035     if (vma->vm_start & mask || vma->vm_end & mask) {
0036         dev_info_ratelimited(dev,
0037                 "%s: %s: fail, unaligned vma (%#lx - %#lx, %#lx)\n",
0038                 current->comm, func, vma->vm_start, vma->vm_end,
0039                 mask);
0040         return -EINVAL;
0041     }
0042
0043     if (!vma_is_dax(vma)) {
0044         dev_info_ratelimited(dev,
0045                 "%s: %s: fail, vma is not DAX capable\n",
0046                 current->comm, func);
0047         return -EINVAL;
0048     }
0049
0050     return 0;
0051 }
0052
0053 /* see "strong" declaration in tools/testing/nvdimm/dax-dev.c */
0054 __weak phys_addr_t dax_pgoff_to_phys(struct dev_dax *dev_dax, pgoff_t pgoff,
0055         unsigned long size)
0056 {
0057     int i;
0058
0059     for (i = 0; i < dev_dax->nr_range; i++) {
0060         struct dev_dax_range *dax_range = &dev_dax->ranges[i];
0061         struct range *range = &dax_range->range;
0062         unsigned long long pgoff_end;
0063         phys_addr_t phys;
0064
0065         pgoff_end = dax_range->pgoff + PHYS_PFN(range_len(range)) - 1;
0066         if (pgoff < dax_range->pgoff || pgoff > pgoff_end)
0067             continue;
0068         phys = PFN_PHYS(pgoff - dax_range->pgoff) + range->start;
0069         if (phys + size - 1 <= range->end)
0070             return phys;
0071         break;
0072     }
0073     return -1;
0074 }
0075
0076 static void dax_set_mapping(struct vm_fault *vmf, pfn_t pfn,
0077                   unsigned long fault_size)
0078 {
0079     unsigned long i, nr_pages = fault_size / PAGE_SIZE;
0080     struct file *filp = vmf->vma->vm_file;
0081     struct dev_dax *dev_dax = filp->private_data;
0082     pgoff_t pgoff;
0083
0084     /* mapping is only set on the head */
0085     if (dev_dax->pgmap->vmemmap_shift)
0086         nr_pages = 1;
0087
0088     pgoff = linear_page_index(vmf->vma,
0089             ALIGN(vmf->address, fault_size));
0090
0091     for (i = 0; i < nr_pages; i++) {
0092         struct page *page = pfn_to_page(pfn_t_to_pfn(pfn) + i);
0093
0094         page = compound_head(page);
0095         if (page->mapping)
0096             continue;
0097
0098         page->mapping = filp->f_mapping;
0099         page->index = pgoff + i;
0100     }
0101 }
0102
0103 static vm_fault_t __dev_dax_pte_fault(struct dev_dax *dev_dax,
0104                 struct vm_fault *vmf)
0105 {
0106     struct device *dev = &dev_dax->dev;
0107     phys_addr_t phys;
0108     pfn_t pfn;
0109     unsigned int fault_size = PAGE_SIZE;
0110
0111     if (check_vma(dev_dax, vmf->vma, __func__))
0112         return VM_FAULT_SIGBUS;
0113
0114     if (dev_dax->align > PAGE_SIZE) {
0115         dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
0116             dev_dax->align, fault_size);
0117         return VM_FAULT_SIGBUS;
0118     }
0119
0120     if (fault_size != dev_dax->align)
0121         return VM_FAULT_SIGBUS;
0122
0123     phys = dax_pgoff_to_phys(dev_dax, vmf->pgoff, PAGE_SIZE);
0124     if (phys == -1) {
0125         dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", vmf->pgoff);
0126         return VM_FAULT_SIGBUS;
0127     }
0128
0129     pfn = phys_to_pfn_t(phys, PFN_DEV|PFN_MAP);
0130
0131     dax_set_mapping(vmf, pfn, fault_size);
0132
0133     return vmf_insert_mixed(vmf->vma, vmf->address, pfn);
0134 }
0135
0136 static vm_fault_t __dev_dax_pmd_fault(struct dev_dax *dev_dax,
0137                 struct vm_fault *vmf)
0138 {
0139     unsigned long pmd_addr = vmf->address & PMD_MASK;
0140     struct device *dev = &dev_dax->dev;
0141     phys_addr_t phys;
0142     pgoff_t pgoff;
0143     pfn_t pfn;
0144     unsigned int fault_size = PMD_SIZE;
0145
0146     if (check_vma(dev_dax, vmf->vma, __func__))
0147         return VM_FAULT_SIGBUS;
0148
0149     if (dev_dax->align > PMD_SIZE) {
0150         dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
0151             dev_dax->align, fault_size);
0152         return VM_FAULT_SIGBUS;
0153     }
0154
0155     if (fault_size < dev_dax->align)
0156         return VM_FAULT_SIGBUS;
0157     else if (fault_size > dev_dax->align)
0158         return VM_FAULT_FALLBACK;
0159
0160     /* if we are outside of the VMA */
0161     if (pmd_addr < vmf->vma->vm_start ||
0162             (pmd_addr + PMD_SIZE) > vmf->vma->vm_end)
0163         return VM_FAULT_SIGBUS;
0164
0165     pgoff = linear_page_index(vmf->vma, pmd_addr);
0166     phys = dax_pgoff_to_phys(dev_dax, pgoff, PMD_SIZE);
0167     if (phys == -1) {
0168         dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", pgoff);
0169         return VM_FAULT_SIGBUS;
0170     }
0171
0172     pfn = phys_to_pfn_t(phys, PFN_DEV|PFN_MAP);
0173
0174     dax_set_mapping(vmf, pfn, fault_size);
0175
0176     return vmf_insert_pfn_pmd(vmf, pfn, vmf->flags & FAULT_FLAG_WRITE);
0177 }
0178
0179 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
0180 static vm_fault_t __dev_dax_pud_fault(struct dev_dax *dev_dax,
0181                 struct vm_fault *vmf)
0182 {
0183     unsigned long pud_addr = vmf->address & PUD_MASK;
0184     struct device *dev = &dev_dax->dev;
0185     phys_addr_t phys;
0186     pgoff_t pgoff;
0187     pfn_t pfn;
0188     unsigned int fault_size = PUD_SIZE;
0189
0190
0191     if (check_vma(dev_dax, vmf->vma, __func__))
0192         return VM_FAULT_SIGBUS;
0193
0194     if (dev_dax->align > PUD_SIZE) {
0195         dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
0196             dev_dax->align, fault_size);
0197         return VM_FAULT_SIGBUS;
0198     }
0199
0200     if (fault_size < dev_dax->align)
0201         return VM_FAULT_SIGBUS;
0202     else if (fault_size > dev_dax->align)
0203         return VM_FAULT_FALLBACK;
0204
0205     /* if we are outside of the VMA */
0206     if (pud_addr < vmf->vma->vm_start ||
0207             (pud_addr + PUD_SIZE) > vmf->vma->vm_end)
0208         return VM_FAULT_SIGBUS;
0209
0210     pgoff = linear_page_index(vmf->vma, pud_addr);
0211     phys = dax_pgoff_to_phys(dev_dax, pgoff, PUD_SIZE);
0212     if (phys == -1) {
0213         dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", pgoff);
0214         return VM_FAULT_SIGBUS;
0215     }
0216
0217     pfn = phys_to_pfn_t(phys, PFN_DEV|PFN_MAP);
0218
0219     dax_set_mapping(vmf, pfn, fault_size);
0220
0221     return vmf_insert_pfn_pud(vmf, pfn, vmf->flags & FAULT_FLAG_WRITE);
0222 }
0223 #else
0224 static vm_fault_t __dev_dax_pud_fault(struct dev_dax *dev_dax,
0225                 struct vm_fault *vmf)
0226 {
0227     return VM_FAULT_FALLBACK;
0228 }
0229 #endif /* !CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
0230
0231 static vm_fault_t dev_dax_huge_fault(struct vm_fault *vmf,
0232         enum page_entry_size pe_size)
0233 {
0234     struct file *filp = vmf->vma->vm_file;
0235     vm_fault_t rc = VM_FAULT_SIGBUS;
0236     int id;
0237     struct dev_dax *dev_dax = filp->private_data;
0238
0239     dev_dbg(&dev_dax->dev, "%s: %s (%#lx - %#lx) size = %d\n", current->comm,
0240             (vmf->flags & FAULT_FLAG_WRITE) ? "write" : "read",
0241             vmf->vma->vm_start, vmf->vma->vm_end, pe_size);
0242
0243     id = dax_read_lock();
0244     switch (pe_size) {
0245     case PE_SIZE_PTE:
0246         rc = __dev_dax_pte_fault(dev_dax, vmf);
0247         break;
0248     case PE_SIZE_PMD:
0249         rc = __dev_dax_pmd_fault(dev_dax, vmf);
0250         break;
0251     case PE_SIZE_PUD:
0252         rc = __dev_dax_pud_fault(dev_dax, vmf);
0253         break;
0254     default:
0255         rc = VM_FAULT_SIGBUS;
0256     }
0257
0258     dax_read_unlock(id);
0259
0260     return rc;
0261 }
0262
0263 static vm_fault_t dev_dax_fault(struct vm_fault *vmf)
0264 {
0265     return dev_dax_huge_fault(vmf, PE_SIZE_PTE);
0266 }
0267
0268 static int dev_dax_may_split(struct vm_area_struct *vma, unsigned long addr)
0269 {
0270     struct file *filp = vma->vm_file;
0271     struct dev_dax *dev_dax = filp->private_data;
0272
0273     if (!IS_ALIGNED(addr, dev_dax->align))
0274         return -EINVAL;
0275     return 0;
0276 }
0277
0278 static unsigned long dev_dax_pagesize(struct vm_area_struct *vma)
0279 {
0280     struct file *filp = vma->vm_file;
0281     struct dev_dax *dev_dax = filp->private_data;
0282
0283     return dev_dax->align;
0284 }
0285
0286 static const struct vm_operations_struct dax_vm_ops = {
0287     .fault = dev_dax_fault,
0288     .huge_fault = dev_dax_huge_fault,
0289     .may_split = dev_dax_may_split,
0290     .pagesize = dev_dax_pagesize,
0291 };
0292
0293 static int dax_mmap(struct file *filp, struct vm_area_struct *vma)
0294 {
0295     struct dev_dax *dev_dax = filp->private_data;
0296     int rc, id;
0297
0298     dev_dbg(&dev_dax->dev, "trace\n");
0299
0300     /*
0301      * We lock to check dax_dev liveness and will re-check at
0302      * fault time.
0303      */
0304     id = dax_read_lock();
0305     rc = check_vma(dev_dax, vma, __func__);
0306     dax_read_unlock(id);
0307     if (rc)
0308         return rc;
0309
0310     vma->vm_ops = &dax_vm_ops;
0311     vma->vm_flags |= VM_HUGEPAGE;
0312     return 0;
0313 }
0314
0315 /* return an unmapped area aligned to the dax region specified alignment */
0316 static unsigned long dax_get_unmapped_area(struct file *filp,
0317         unsigned long addr, unsigned long len, unsigned long pgoff,
0318         unsigned long flags)
0319 {
0320     unsigned long off, off_end, off_align, len_align, addr_align, align;
0321     struct dev_dax *dev_dax = filp ? filp->private_data : NULL;
0322
0323     if (!dev_dax || addr)
0324         goto out;
0325
0326     align = dev_dax->align;
0327     off = pgoff << PAGE_SHIFT;
0328     off_end = off + len;
0329     off_align = round_up(off, align);
0330
0331     if ((off_end <= off_align) || ((off_end - off_align) < align))
0332         goto out;
0333
0334     len_align = len + align;
0335     if ((off + len_align) < off)
0336         goto out;
0337
0338     addr_align = current->mm->get_unmapped_area(filp, addr, len_align,
0339             pgoff, flags);
0340     if (!IS_ERR_VALUE(addr_align)) {
0341         addr_align += (off - addr_align) & (align - 1);
0342         return addr_align;
0343     }
0344  out:
0345     return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
0346 }
0347
0348 static const struct address_space_operations dev_dax_aops = {
0349     .dirty_folio    = noop_dirty_folio,
0350 };
0351
0352 static int dax_open(struct inode *inode, struct file *filp)
0353 {
0354     struct dax_device *dax_dev = inode_dax(inode);
0355     struct inode *__dax_inode = dax_inode(dax_dev);
0356     struct dev_dax *dev_dax = dax_get_private(dax_dev);
0357
0358     dev_dbg(&dev_dax->dev, "trace\n");
0359     inode->i_mapping = __dax_inode->i_mapping;
0360     inode->i_mapping->host = __dax_inode;
0361     inode->i_mapping->a_ops = &dev_dax_aops;
0362     filp->f_mapping = inode->i_mapping;
0363     filp->f_wb_err = filemap_sample_wb_err(filp->f_mapping);
0364     filp->f_sb_err = file_sample_sb_err(filp);
0365     filp->private_data = dev_dax;
0366     inode->i_flags = S_DAX;
0367
0368     return 0;
0369 }
0370
0371 static int dax_release(struct inode *inode, struct file *filp)
0372 {
0373     struct dev_dax *dev_dax = filp->private_data;
0374
0375     dev_dbg(&dev_dax->dev, "trace\n");
0376     return 0;
0377 }
0378
0379 static const struct file_operations dax_fops = {
0380     .llseek = noop_llseek,
0381     .owner = THIS_MODULE,
0382     .open = dax_open,
0383     .release = dax_release,
0384     .get_unmapped_area = dax_get_unmapped_area,
0385     .mmap = dax_mmap,
0386     .mmap_supported_flags = MAP_SYNC,
0387 };
0388
0389 static void dev_dax_cdev_del(void *cdev)
0390 {
0391     cdev_del(cdev);
0392 }
0393
0394 static void dev_dax_kill(void *dev_dax)
0395 {
0396     kill_dev_dax(dev_dax);
0397 }
0398
0399 int dev_dax_probe(struct dev_dax *dev_dax)
0400 {
0401     struct dax_device *dax_dev = dev_dax->dax_dev;
0402     struct device *dev = &dev_dax->dev;
0403     struct dev_pagemap *pgmap;
0404     struct inode *inode;
0405     struct cdev *cdev;
0406     void *addr;
0407     int rc, i;
0408
0409     if (static_dev_dax(dev_dax))  {
0410         if (dev_dax->nr_range > 1) {
0411             dev_warn(dev,
0412                 "static pgmap / multi-range device conflict\n");
0413             return -EINVAL;
0414         }
0415
0416         pgmap = dev_dax->pgmap;
0417     } else {
0418         if (dev_dax->pgmap) {
0419             dev_warn(dev,
0420                  "dynamic-dax with pre-populated page map\n");
0421             return -EINVAL;
0422         }
0423
0424         pgmap = devm_kzalloc(dev,
0425                        struct_size(pgmap, ranges, dev_dax->nr_range - 1),
0426                        GFP_KERNEL);
0427         if (!pgmap)
0428             return -ENOMEM;
0429
0430         pgmap->nr_range = dev_dax->nr_range;
0431         dev_dax->pgmap = pgmap;
0432
0433         for (i = 0; i < dev_dax->nr_range; i++) {
0434             struct range *range = &dev_dax->ranges[i].range;
0435             pgmap->ranges[i] = *range;
0436         }
0437     }
0438
0439     for (i = 0; i < dev_dax->nr_range; i++) {
0440         struct range *range = &dev_dax->ranges[i].range;
0441
0442         if (!devm_request_mem_region(dev, range->start,
0443                     range_len(range), dev_name(dev))) {
0444             dev_warn(dev, "mapping%d: %#llx-%#llx could not reserve range\n",
0445                     i, range->start, range->end);
0446             return -EBUSY;
0447         }
0448     }
0449
0450     pgmap->type = MEMORY_DEVICE_GENERIC;
0451     if (dev_dax->align > PAGE_SIZE)
0452         pgmap->vmemmap_shift =
0453             order_base_2(dev_dax->align >> PAGE_SHIFT);
0454     addr = devm_memremap_pages(dev, pgmap);
0455     if (IS_ERR(addr))
0456         return PTR_ERR(addr);
0457
0458     inode = dax_inode(dax_dev);
0459     cdev = inode->i_cdev;
0460     cdev_init(cdev, &dax_fops);
0461     cdev->owner = dev->driver->owner;
0462     cdev_set_parent(cdev, &dev->kobj);
0463     rc = cdev_add(cdev, dev->devt, 1);
0464     if (rc)
0465         return rc;
0466
0467     rc = devm_add_action_or_reset(dev, dev_dax_cdev_del, cdev);
0468     if (rc)
0469         return rc;
0470
0471     run_dax(dax_dev);
0472     return devm_add_action_or_reset(dev, dev_dax_kill, dev_dax);
0473 }
0474 EXPORT_SYMBOL_GPL(dev_dax_probe);
0475
0476 static struct dax_device_driver device_dax_driver = {
0477     .probe = dev_dax_probe,
0478     /* all probe actions are unwound by devm, so .remove isn't necessary */
0479     .match_always = 1,
0480 };
0481
0482 static int __init dax_init(void)
0483 {
0484     return dax_driver_register(&device_dax_driver);
0485 }
0486
0487 static void __exit dax_exit(void)
0488 {
0489     dax_driver_unregister(&device_dax_driver);
0490 }
0491
0492 MODULE_AUTHOR("Intel Corporation");
0493 MODULE_LICENSE("GPL v2");
0494 module_init(dax_init);
0495 module_exit(dax_exit);
0496 MODULE_ALIAS_DAX_DEVICE(0);