mm/secretmem.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright IBM Corporation, 2021
   4  *
   5  * Author: Mike Rapoport <rppt@linux.ibm.com>
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/fs.h>
  10 #include <linux/swap.h>
  11 #include <linux/mount.h>
  12 #include <linux/memfd.h>
  13 #include <linux/bitops.h>
  14 #include <linux/printk.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/syscalls.h>
  17 #include <linux/pseudo_fs.h>
  18 #include <linux/secretmem.h>
  19 #include <linux/set_memory.h>
  20 #include <linux/sched/signal.h>
  21
  22 #include <uapi/linux/magic.h>
  23
  24 #include <asm/tlbflush.h>
  25
  26 #include "internal.h"
  27
  28 #undef pr_fmt
  29 #define pr_fmt(fmt) "secretmem: " fmt
  30
  31 /*
  32  * Define mode and flag masks to allow validation of the system call
  33  * parameters.
  34  */
  35 #define SECRETMEM_MODE_MASK     (0x0)
  36 #define SECRETMEM_FLAGS_MASK    SECRETMEM_MODE_MASK
  37
  38 static bool secretmem_enable __ro_after_init;
  39 module_param_named(enable, secretmem_enable, bool, 0400);
  40 MODULE_PARM_DESC(secretmem_enable,
  41                  "Enable secretmem and memfd_secret(2) system call");
  42
  43 static atomic_t secretmem_users;
  44
  45 bool secretmem_active(void)
  46 {
  47         return !!atomic_read(&secretmem_users);
  48 }
  49
  50 static vm_fault_t secretmem_fault(struct vm_fault *vmf)
  51 {
  52         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
  53         struct inode *inode = file_inode(vmf->vma->vm_file);
  54         pgoff_t offset = vmf->pgoff;
  55         gfp_t gfp = vmf->gfp_mask;
  56         unsigned long addr;
  57         struct page *page;
  58         int err;
  59
  60         if (((loff_t)vmf->pgoff << PAGE_SHIFT) >= i_size_read(inode))
  61                 return vmf_error(-EINVAL);
  62
  63 retry:
  64         page = find_lock_page(mapping, offset);
  65         if (!page) {
  66                 page = alloc_page(gfp | __GFP_ZERO);
  67                 if (!page)
  68                         return VM_FAULT_OOM;
  69
  70                 err = set_direct_map_invalid_noflush(page);
  71                 if (err) {
  72                         put_page(page);
  73                         return vmf_error(err);
  74                 }
  75
  76                 __SetPageUptodate(page);
  77                 err = add_to_page_cache_lru(page, mapping, offset, gfp);
  78                 if (unlikely(err)) {
  79                         put_page(page);
  80                         /*
  81                          * If a split of large page was required, it
  82                          * already happened when we marked the page invalid
  83                          * which guarantees that this call won't fail
  84                          */
  85                         set_direct_map_default_noflush(page);
  86                         if (err == -EEXIST)
  87                                 goto retry;
  88
  89                         return vmf_error(err);
  90                 }
  91
  92                 addr = (unsigned long)page_address(page);
  93                 flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
  94         }
  95
  96         vmf->page = page;
  97         return VM_FAULT_LOCKED;
  98 }
  99
 100 static const struct vm_operations_struct secretmem_vm_ops = {
 101         .fault = secretmem_fault,
 102 };
 103
 104 static int secretmem_release(struct inode *inode, struct file *file)
 105 {
 106         atomic_dec(&secretmem_users);
 107         return 0;
 108 }
 109
 110 static int secretmem_mmap(struct file *file, struct vm_area_struct *vma)
 111 {
 112         unsigned long len = vma->vm_end - vma->vm_start;
 113
 114         if ((vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) == 0)
 115                 return -EINVAL;
 116
 117         if (mlock_future_check(vma->vm_mm, vma->vm_flags | VM_LOCKED, len))
 118                 return -EAGAIN;
 119
 120         vma->vm_flags |= VM_LOCKED | VM_DONTDUMP;
 121         vma->vm_ops = &secretmem_vm_ops;
 122
 123         return 0;
 124 }
 125
 126 bool vma_is_secretmem(struct vm_area_struct *vma)
 127 {
 128         return vma->vm_ops == &secretmem_vm_ops;
 129 }
 130
 131 static const struct file_operations secretmem_fops = {
 132         .release        = secretmem_release,
 133         .mmap           = secretmem_mmap,
 134 };
 135
 136 static int secretmem_migratepage(struct address_space *mapping,
 137                                  struct page *newpage, struct page *page,
 138                                  enum migrate_mode mode)
 139 {
 140         return -EBUSY;
 141 }
 142
 143 static void secretmem_free_folio(struct folio *folio)
 144 {
 145         set_direct_map_default_noflush(&folio->page);
 146         folio_zero_segment(folio, 0, folio_size(folio));
 147 }
 148
 149 const struct address_space_operations secretmem_aops = {
 150         .dirty_folio    = noop_dirty_folio,
 151         .free_folio     = secretmem_free_folio,
 152         .migratepage    = secretmem_migratepage,
 153 };
 154
 155 static int secretmem_setattr(struct user_namespace *mnt_userns,
 156                              struct dentry *dentry, struct iattr *iattr)
 157 {
 158         struct inode *inode = d_inode(dentry);
 159         unsigned int ia_valid = iattr->ia_valid;
 160
 161         if ((ia_valid & ATTR_SIZE) && inode->i_size)
 162                 return -EINVAL;
 163
 164         return simple_setattr(mnt_userns, dentry, iattr);
 165 }
 166
 167 static const struct inode_operations secretmem_iops = {
 168         .setattr = secretmem_setattr,
 169 };
 170
 171 static struct vfsmount *secretmem_mnt;
 172
 173 static struct file *secretmem_file_create(unsigned long flags)
 174 {
 175         struct file *file = ERR_PTR(-ENOMEM);
 176         struct inode *inode;
 177
 178         inode = alloc_anon_inode(secretmem_mnt->mnt_sb);
 179         if (IS_ERR(inode))
 180                 return ERR_CAST(inode);
 181
 182         file = alloc_file_pseudo(inode, secretmem_mnt, "secretmem",
 183                                  O_RDWR, &secretmem_fops);
 184         if (IS_ERR(file))
 185                 goto err_free_inode;
 186
 187         mapping_set_gfp_mask(inode->i_mapping, GFP_HIGHUSER);
 188         mapping_set_unevictable(inode->i_mapping);
 189
 190         inode->i_op = &secretmem_iops;
 191         inode->i_mapping->a_ops = &secretmem_aops;
 192
 193         /* pretend we are a normal file with zero size */
 194         inode->i_mode |= S_IFREG;
 195         inode->i_size = 0;
 196
 197         return file;
 198
 199 err_free_inode:
 200         iput(inode);
 201         return file;
 202 }
 203
 204 SYSCALL_DEFINE1(memfd_secret, unsigned int, flags)
 205 {
 206         struct file *file;
 207         int fd, err;
 208
 209         /* make sure local flags do not confict with global fcntl.h */
 210         BUILD_BUG_ON(SECRETMEM_FLAGS_MASK & O_CLOEXEC);
 211
 212         if (!secretmem_enable)
 213                 return -ENOSYS;
 214
 215         if (flags & ~(SECRETMEM_FLAGS_MASK | O_CLOEXEC))
 216                 return -EINVAL;
 217         if (atomic_read(&secretmem_users) < 0)
 218                 return -ENFILE;
 219
 220         fd = get_unused_fd_flags(flags & O_CLOEXEC);
 221         if (fd < 0)
 222                 return fd;
 223
 224         file = secretmem_file_create(flags);
 225         if (IS_ERR(file)) {
 226                 err = PTR_ERR(file);
 227                 goto err_put_fd;
 228         }
 229
 230         file->f_flags |= O_LARGEFILE;
 231
 232         atomic_inc(&secretmem_users);
 233         fd_install(fd, file);
 234         return fd;
 235
 236 err_put_fd:
 237         put_unused_fd(fd);
 238         return err;
 239 }
 240
 241 static int secretmem_init_fs_context(struct fs_context *fc)
 242 {
 243         return init_pseudo(fc, SECRETMEM_MAGIC) ? 0 : -ENOMEM;
 244 }
 245
 246 static struct file_system_type secretmem_fs = {
 247         .name           = "secretmem",
 248         .init_fs_context = secretmem_init_fs_context,
 249         .kill_sb        = kill_anon_super,
 250 };
 251
 252 static int secretmem_init(void)
 253 {
 254         int ret = 0;
 255
 256         if (!secretmem_enable)
 257                 return ret;
 258
 259         secretmem_mnt = kern_mount(&secretmem_fs);
 260         if (IS_ERR(secretmem_mnt))
 261                 ret = PTR_ERR(secretmem_mnt);
 262
 263         /* prevent secretmem mappings from ever getting PROT_EXEC */
 264         secretmem_mnt->mnt_flags |= MNT_NOEXEC;
 265
 266         return ret;
 267 }
 268 fs_initcall(secretmem_init);