Fix rmmod/read/write races in /proc entries
authorAlexey Dobriyan <adobriyan@sw.ru>
Mon, 16 Jul 2007 06:39:00 +0000 (23:39 -0700)
committerLinus Torvalds <torvalds@woody.linux-foundation.org>
Mon, 16 Jul 2007 16:05:39 +0000 (09:05 -0700)
Fix following races:
===========================================
1. Write via ->write_proc sleeps in copy_from_user(). Module disappears
   meanwhile. Or, more generically, system call done on /proc file, method
   supplied by module is called, module dissapeares meanwhile.

   pde = create_proc_entry()
   if (!pde)
return -ENOMEM;
   pde->write_proc = ...
open
write
copy_from_user
   pde = create_proc_entry();
   if (!pde) {
remove_proc_entry();
return -ENOMEM;
/* module unloaded */
   }
*boom*
==========================================
2. bogo-revoke aka proc_kill_inodes()

  remove_proc_entry vfs_read
  proc_kill_inodes [check ->f_op validness]
[check ->f_op->read validness]
[verify_area, security permissions checks]
->f_op = NULL;
if (file->f_op->read)
/* ->f_op dereference, boom */

NOTE, NOTE, NOTE: file_operations are proxied for regular files only. Let's
see how this scheme behaves, then extend if needed for directories.
Directories creators in /proc only set ->owner for them, so proxying for
directories may be unneeded.

NOTE, NOTE, NOTE: methods being proxied are ->llseek, ->read, ->write,
->poll, ->unlocked_ioctl, ->ioctl, ->compat_ioctl, ->open, ->release.
If your in-tree module uses something else, yell on me. Full audit pending.

[akpm@linux-foundation.org: build fix]
Signed-off-by: Alexey Dobriyan <adobriyan@sw.ru>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/proc/generic.c
fs/proc/inode.c
include/linux/proc_fs.h

index 8a40e15f5ecb90333f7eca84ede39c40e7adb945..4f8e53568b22d24ba1334fbc95363b8d5b2fd765 100644 (file)
@@ -20,6 +20,7 @@
 #include <linux/namei.h>
 #include <linux/bitops.h>
 #include <linux/spinlock.h>
+#include <linux/completion.h>
 #include <asm/uaccess.h>
 
 #include "internal.h"
@@ -613,6 +614,9 @@ static struct proc_dir_entry *proc_create(struct proc_dir_entry **parent,
        ent->namelen = len;
        ent->mode = mode;
        ent->nlink = nlink;
+       ent->pde_users = 0;
+       spin_lock_init(&ent->pde_unload_lock);
+       ent->pde_unload_completion = NULL;
  out:
        return ent;
 }
@@ -734,9 +738,35 @@ void remove_proc_entry(const char *name, struct proc_dir_entry *parent)
                de = *p;
                *p = de->next;
                de->next = NULL;
+
+               spin_lock(&de->pde_unload_lock);
+               /*
+                * Stop accepting new callers into module. If you're
+                * dynamically allocating ->proc_fops, save a pointer somewhere.
+                */
+               de->proc_fops = NULL;
+               /* Wait until all existing callers into module are done. */
+               if (de->pde_users > 0) {
+                       DECLARE_COMPLETION_ONSTACK(c);
+
+                       if (!de->pde_unload_completion)
+                               de->pde_unload_completion = &c;
+
+                       spin_unlock(&de->pde_unload_lock);
+                       spin_unlock(&proc_subdir_lock);
+
+                       wait_for_completion(de->pde_unload_completion);
+
+                       spin_lock(&proc_subdir_lock);
+                       goto continue_removing;
+               }
+               spin_unlock(&de->pde_unload_lock);
+
+continue_removing:
                if (S_ISDIR(de->mode))
                        parent->nlink--;
-               proc_kill_inodes(de);
+               if (!S_ISREG(de->mode))
+                       proc_kill_inodes(de);
                de->nlink = 0;
                WARN_ON(de->subdir);
                if (!atomic_read(&de->count))
index d5ce65c68d7b2b538a87c0a58de0863941db6ff3..dd28e86ab422fe23e06969cebf14c3677a6c8902 100644 (file)
@@ -10,6 +10,7 @@
 #include <linux/mm.h>
 #include <linux/string.h>
 #include <linux/stat.h>
+#include <linux/completion.h>
 #include <linux/file.h>
 #include <linux/limits.h>
 #include <linux/init.h>
@@ -140,6 +141,251 @@ static const struct super_operations proc_sops = {
        .remount_fs     = proc_remount,
 };
 
+static void pde_users_dec(struct proc_dir_entry *pde)
+{
+       spin_lock(&pde->pde_unload_lock);
+       pde->pde_users--;
+       if (pde->pde_unload_completion && pde->pde_users == 0)
+               complete(pde->pde_unload_completion);
+       spin_unlock(&pde->pde_unload_lock);
+}
+
+static loff_t proc_reg_llseek(struct file *file, loff_t offset, int whence)
+{
+       struct proc_dir_entry *pde = PDE(file->f_path.dentry->d_inode);
+       loff_t rv = -EINVAL;
+       loff_t (*llseek)(struct file *, loff_t, int);
+
+       spin_lock(&pde->pde_unload_lock);
+       /*
+        * remove_proc_entry() is going to delete PDE (as part of module
+        * cleanup sequence). No new callers into module allowed.
+        */
+       if (!pde->proc_fops) {
+               spin_unlock(&pde->pde_unload_lock);
+               return rv;
+       }
+       /*
+        * Bump refcount so that remove_proc_entry will wail for ->llseek to
+        * complete.
+        */
+       pde->pde_users++;
+       /*
+        * Save function pointer under lock, to protect against ->proc_fops
+        * NULL'ifying right after ->pde_unload_lock is dropped.
+        */
+       llseek = pde->proc_fops->llseek;
+       spin_unlock(&pde->pde_unload_lock);
+
+       if (!llseek)
+               llseek = default_llseek;
+       rv = llseek(file, offset, whence);
+
+       pde_users_dec(pde);
+       return rv;
+}
+
+static ssize_t proc_reg_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
+{
+       struct proc_dir_entry *pde = PDE(file->f_path.dentry->d_inode);
+       ssize_t rv = -EIO;
+       ssize_t (*read)(struct file *, char __user *, size_t, loff_t *);
+
+       spin_lock(&pde->pde_unload_lock);
+       if (!pde->proc_fops) {
+               spin_unlock(&pde->pde_unload_lock);
+               return rv;
+       }
+       pde->pde_users++;
+       read = pde->proc_fops->read;
+       spin_unlock(&pde->pde_unload_lock);
+
+       if (read)
+               rv = read(file, buf, count, ppos);
+
+       pde_users_dec(pde);
+       return rv;
+}
+
+static ssize_t proc_reg_write(struct file *file, const char __user *buf, size_t count, loff_t *ppos)
+{
+       struct proc_dir_entry *pde = PDE(file->f_path.dentry->d_inode);
+       ssize_t rv = -EIO;
+       ssize_t (*write)(struct file *, const char __user *, size_t, loff_t *);
+
+       spin_lock(&pde->pde_unload_lock);
+       if (!pde->proc_fops) {
+               spin_unlock(&pde->pde_unload_lock);
+               return rv;
+       }
+       pde->pde_users++;
+       write = pde->proc_fops->write;
+       spin_unlock(&pde->pde_unload_lock);
+
+       if (write)
+               rv = write(file, buf, count, ppos);
+
+       pde_users_dec(pde);
+       return rv;
+}
+
+static unsigned int proc_reg_poll(struct file *file, struct poll_table_struct *pts)
+{
+       struct proc_dir_entry *pde = PDE(file->f_path.dentry->d_inode);
+       unsigned int rv = 0;
+       unsigned int (*poll)(struct file *, struct poll_table_struct *);
+
+       spin_lock(&pde->pde_unload_lock);
+       if (!pde->proc_fops) {
+               spin_unlock(&pde->pde_unload_lock);
+               return rv;
+       }
+       pde->pde_users++;
+       poll = pde->proc_fops->poll;
+       spin_unlock(&pde->pde_unload_lock);
+
+       if (poll)
+               rv = poll(file, pts);
+
+       pde_users_dec(pde);
+       return rv;
+}
+
+static long proc_reg_unlocked_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
+{
+       struct proc_dir_entry *pde = PDE(file->f_path.dentry->d_inode);
+       long rv = -ENOTTY;
+       long (*unlocked_ioctl)(struct file *, unsigned int, unsigned long);
+       int (*ioctl)(struct inode *, struct file *, unsigned int, unsigned long);
+
+       spin_lock(&pde->pde_unload_lock);
+       if (!pde->proc_fops) {
+               spin_unlock(&pde->pde_unload_lock);
+               return rv;
+       }
+       pde->pde_users++;
+       unlocked_ioctl = pde->proc_fops->unlocked_ioctl;
+       ioctl = pde->proc_fops->ioctl;
+       spin_unlock(&pde->pde_unload_lock);
+
+       if (unlocked_ioctl) {
+               rv = unlocked_ioctl(file, cmd, arg);
+               if (rv == -ENOIOCTLCMD)
+                       rv = -EINVAL;
+       } else if (ioctl) {
+               lock_kernel();
+               rv = ioctl(file->f_path.dentry->d_inode, file, cmd, arg);
+               unlock_kernel();
+       }
+
+       pde_users_dec(pde);
+       return rv;
+}
+
+#ifdef CONFIG_COMPAT
+static long proc_reg_compat_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
+{
+       struct proc_dir_entry *pde = PDE(file->f_path.dentry->d_inode);
+       long rv = -ENOTTY;
+       long (*compat_ioctl)(struct file *, unsigned int, unsigned long);
+
+       spin_lock(&pde->pde_unload_lock);
+       if (!pde->proc_fops) {
+               spin_unlock(&pde->pde_unload_lock);
+               return rv;
+       }
+       pde->pde_users++;
+       compat_ioctl = pde->proc_fops->compat_ioctl;
+       spin_unlock(&pde->pde_unload_lock);
+
+       if (compat_ioctl)
+               rv = compat_ioctl(file, cmd, arg);
+
+       pde_users_dec(pde);
+       return rv;
+}
+#endif
+
+static int proc_reg_mmap(struct file *file, struct vm_area_struct *vma)
+{
+       struct proc_dir_entry *pde = PDE(file->f_path.dentry->d_inode);
+       int rv = -EIO;
+       int (*mmap)(struct file *, struct vm_area_struct *);
+
+       spin_lock(&pde->pde_unload_lock);
+       if (!pde->proc_fops) {
+               spin_unlock(&pde->pde_unload_lock);
+               return rv;
+       }
+       pde->pde_users++;
+       mmap = pde->proc_fops->mmap;
+       spin_unlock(&pde->pde_unload_lock);
+
+       if (mmap)
+               rv = mmap(file, vma);
+
+       pde_users_dec(pde);
+       return rv;
+}
+
+static int proc_reg_open(struct inode *inode, struct file *file)
+{
+       struct proc_dir_entry *pde = PDE(inode);
+       int rv = 0;
+       int (*open)(struct inode *, struct file *);
+
+       spin_lock(&pde->pde_unload_lock);
+       if (!pde->proc_fops) {
+               spin_unlock(&pde->pde_unload_lock);
+               return rv;
+       }
+       pde->pde_users++;
+       open = pde->proc_fops->open;
+       spin_unlock(&pde->pde_unload_lock);
+
+       if (open)
+               rv = open(inode, file);
+
+       pde_users_dec(pde);
+       return rv;
+}
+
+static int proc_reg_release(struct inode *inode, struct file *file)
+{
+       struct proc_dir_entry *pde = PDE(inode);
+       int rv = 0;
+       int (*release)(struct inode *, struct file *);
+
+       spin_lock(&pde->pde_unload_lock);
+       if (!pde->proc_fops) {
+               spin_unlock(&pde->pde_unload_lock);
+               return rv;
+       }
+       pde->pde_users++;
+       release = pde->proc_fops->release;
+       spin_unlock(&pde->pde_unload_lock);
+
+       if (release)
+               rv = release(inode, file);
+
+       pde_users_dec(pde);
+       return rv;
+}
+
+static const struct file_operations proc_reg_file_ops = {
+       .llseek         = proc_reg_llseek,
+       .read           = proc_reg_read,
+       .write          = proc_reg_write,
+       .poll           = proc_reg_poll,
+       .unlocked_ioctl = proc_reg_unlocked_ioctl,
+#ifdef CONFIG_COMPAT
+       .compat_ioctl   = proc_reg_compat_ioctl,
+#endif
+       .mmap           = proc_reg_mmap,
+       .open           = proc_reg_open,
+       .release        = proc_reg_release,
+};
+
 struct inode *proc_get_inode(struct super_block *sb, unsigned int ino,
                                struct proc_dir_entry *de)
 {
@@ -166,8 +412,12 @@ struct inode *proc_get_inode(struct super_block *sb, unsigned int ino,
                        inode->i_nlink = de->nlink;
                if (de->proc_iops)
                        inode->i_op = de->proc_iops;
-               if (de->proc_fops)
-                       inode->i_fop = de->proc_fops;
+               if (de->proc_fops) {
+                       if (S_ISREG(inode->i_mode))
+                               inode->i_fop = &proc_reg_file_ops;
+                       else
+                               inode->i_fop = de->proc_fops;
+               }
        }
 
        return inode;
index 3469f96bc8b2d0e11543d57246f4c4ae75f64115..28e3664fdf1b860155157b3e1317b83a0203864f 100644 (file)
@@ -7,6 +7,8 @@
 #include <linux/magic.h>
 #include <asm/atomic.h>
 
+struct completion;
+
 /*
  * The proc filesystem constants/structures
  */
@@ -56,6 +58,14 @@ struct proc_dir_entry {
        gid_t gid;
        loff_t size;
        const struct inode_operations *proc_iops;
+       /*
+        * NULL ->proc_fops means "PDE is going away RSN" or
+        * "PDE is just created". In either case, e.g. ->read_proc won't be
+        * called because it's too late or too early, respectively.
+        *
+        * If you're allocating ->proc_fops dynamically, save a pointer
+        * somewhere.
+        */
        const struct file_operations *proc_fops;
        get_info_t *get_info;
        struct module *owner;
@@ -66,6 +76,9 @@ struct proc_dir_entry {
        atomic_t count;         /* use count */
        int deleted;            /* delete flag */
        void *set;
+       int pde_users;  /* number of callers into module in progress */
+       spinlock_t pde_unload_lock; /* proc_fops checks and pde_users bumps */
+       struct completion *pde_unload_completion;
 };
 
 struct kcore_list {