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