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
 38static bool secretmem_enable __ro_after_init = 1;
 39module_param_named(enable, secretmem_enable, bool, 0400);
 40MODULE_PARM_DESC(secretmem_enable,
 41		 "Enable secretmem and memfd_secret(2) system call");
 42
 43static atomic_t secretmem_users;
 44
 45bool secretmem_active(void)
 46{
 47	return !!atomic_read(&secretmem_users);
 48}
 49
 50static 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 folio *folio;
 58	vm_fault_t ret;
 59	int err;
 60
 61	if (((loff_t)vmf->pgoff << PAGE_SHIFT) >= i_size_read(inode))
 62		return vmf_error(-EINVAL);
 63
 64	filemap_invalidate_lock_shared(mapping);
 65
 66retry:
 67	folio = filemap_lock_folio(mapping, offset);
 68	if (IS_ERR(folio)) {
 69		folio = folio_alloc(gfp | __GFP_ZERO, 0);
 70		if (!folio) {
 71			ret = VM_FAULT_OOM;
 72			goto out;
 73		}
 74
 75		err = set_direct_map_invalid_noflush(folio_page(folio, 0));
 76		if (err) {
 77			folio_put(folio);
 78			ret = vmf_error(err);
 79			goto out;
 80		}
 81
 82		__folio_mark_uptodate(folio);
 83		err = filemap_add_folio(mapping, folio, offset, gfp);
 84		if (unlikely(err)) {
 85			/*
 86			 * If a split of large page was required, it
 87			 * already happened when we marked the page invalid
 88			 * which guarantees that this call won't fail
 89			 */
 90			set_direct_map_default_noflush(folio_page(folio, 0));
 91			folio_put(folio);
 92			if (err == -EEXIST)
 93				goto retry;
 94
 95			ret = vmf_error(err);
 96			goto out;
 97		}
 98
 99		addr = (unsigned long)folio_address(folio);
100		flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
101	}
102
103	vmf->page = folio_file_page(folio, vmf->pgoff);
104	ret = VM_FAULT_LOCKED;
105
106out:
107	filemap_invalidate_unlock_shared(mapping);
108	return ret;
109}
110
111static const struct vm_operations_struct secretmem_vm_ops = {
112	.fault = secretmem_fault,
113};
114
115static int secretmem_release(struct inode *inode, struct file *file)
116{
117	atomic_dec(&secretmem_users);
118	return 0;
119}
120
121static int secretmem_mmap_prepare(struct vm_area_desc *desc)
122{
123	const unsigned long len = vma_desc_size(desc);
124
125	if ((desc->vm_flags & (VM_SHARED | VM_MAYSHARE)) == 0)
126		return -EINVAL;
127
128	if (!mlock_future_ok(desc->mm, desc->vm_flags | VM_LOCKED, len))
129		return -EAGAIN;
130
131	desc->vm_flags |= VM_LOCKED | VM_DONTDUMP;
132	desc->vm_ops = &secretmem_vm_ops;
133
134	return 0;
135}
136
137bool vma_is_secretmem(struct vm_area_struct *vma)
138{
139	return vma->vm_ops == &secretmem_vm_ops;
140}
141
142static const struct file_operations secretmem_fops = {
143	.release	= secretmem_release,
144	.mmap_prepare	= secretmem_mmap_prepare,
145};
146
147static int secretmem_migrate_folio(struct address_space *mapping,
148		struct folio *dst, struct folio *src, enum migrate_mode mode)
149{
150	return -EBUSY;
151}
152
153static void secretmem_free_folio(struct folio *folio)
154{
155	set_direct_map_default_noflush(folio_page(folio, 0));
156	folio_zero_segment(folio, 0, folio_size(folio));
157}
158
159const struct address_space_operations secretmem_aops = {
160	.dirty_folio	= noop_dirty_folio,
161	.free_folio	= secretmem_free_folio,
162	.migrate_folio	= secretmem_migrate_folio,
163};
164
165static int secretmem_setattr(struct mnt_idmap *idmap,
166			     struct dentry *dentry, struct iattr *iattr)
167{
168	struct inode *inode = d_inode(dentry);
169	struct address_space *mapping = inode->i_mapping;
170	unsigned int ia_valid = iattr->ia_valid;
171	int ret;
172
173	filemap_invalidate_lock(mapping);
174
175	if ((ia_valid & ATTR_SIZE) && inode->i_size)
176		ret = -EINVAL;
177	else
178		ret = simple_setattr(idmap, dentry, iattr);
179
180	filemap_invalidate_unlock(mapping);
181
182	return ret;
183}
184
185static const struct inode_operations secretmem_iops = {
186	.setattr = secretmem_setattr,
187};
188
189static struct vfsmount *secretmem_mnt;
190
191static struct file *secretmem_file_create(unsigned long flags)
192{
193	struct file *file;
194	struct inode *inode;
195	const char *anon_name = "[secretmem]";
196
197	inode = anon_inode_make_secure_inode(secretmem_mnt->mnt_sb, anon_name, NULL);
198	if (IS_ERR(inode))
199		return ERR_CAST(inode);
200
201	file = alloc_file_pseudo(inode, secretmem_mnt, "secretmem",
202				 O_RDWR | O_LARGEFILE, &secretmem_fops);
203	if (IS_ERR(file))
204		goto err_free_inode;
205
206	mapping_set_gfp_mask(inode->i_mapping, GFP_HIGHUSER);
207	mapping_set_unevictable(inode->i_mapping);
208
209	inode->i_op = &secretmem_iops;
210	inode->i_mapping->a_ops = &secretmem_aops;
211
212	/* pretend we are a normal file with zero size */
213	inode->i_mode |= S_IFREG;
214	inode->i_size = 0;
215
216	atomic_inc(&secretmem_users);
217
218	return file;
219
220err_free_inode:
221	iput(inode);
222	return file;
223}
224
225SYSCALL_DEFINE1(memfd_secret, unsigned int, flags)
226{
227	/* make sure local flags do not conflict with global fcntl.h */
228	BUILD_BUG_ON(SECRETMEM_FLAGS_MASK & O_CLOEXEC);
229
230	if (!secretmem_enable || !can_set_direct_map())
231		return -ENOSYS;
232
233	if (flags & ~(SECRETMEM_FLAGS_MASK | O_CLOEXEC))
234		return -EINVAL;
235	if (atomic_read(&secretmem_users) < 0)
236		return -ENFILE;
237
238	return FD_ADD(flags & O_CLOEXEC, secretmem_file_create(flags));
239}
240
241static int secretmem_init_fs_context(struct fs_context *fc)
242{
243	struct pseudo_fs_context *ctx;
244
245	ctx = init_pseudo(fc, SECRETMEM_MAGIC);
246	if (!ctx)
247		return -ENOMEM;
248
249	fc->s_iflags |= SB_I_NOEXEC;
250	fc->s_iflags |= SB_I_NODEV;
251	return 0;
252}
253
254static struct file_system_type secretmem_fs = {
255	.name		= "secretmem",
256	.init_fs_context = secretmem_init_fs_context,
257	.kill_sb	= kill_anon_super,
258};
259
260static int __init secretmem_init(void)
261{
262	if (!secretmem_enable || !can_set_direct_map())
263		return 0;
264
265	secretmem_mnt = kern_mount(&secretmem_fs);
266	if (IS_ERR(secretmem_mnt))
267		return PTR_ERR(secretmem_mnt);
268
269	return 0;
270}
271fs_initcall(secretmem_init);