Creating PDEs with refcount 0 and "deleted" flag has problems (see below).
Switch to usual scheme:
* PDE is created with refcount 1
* every de_get does +1
* every de_put() and remove_proc_entry() do -1
* once refcount reaches 0, PDE is freed.
This elegantly fixes at least two following races (both observed) without
introducing new locks, without abusing old locks, without spreading
lock_kernel():
1) PDE leak
remove_proc_entry			de_put
-----------------			------
			[refcnt = 1]
if (atomic_read(&de->count) == 0)
					if (atomic_dec_and_test(&de->count))
						if (de->deleted)
							/* also not taken! */
							free_proc_entry(de);
else
	de->deleted = 1;
		[refcount=0, deleted=1]
2) use after free
remove_proc_entry			de_put
-----------------			------
			[refcnt = 1]
					if (atomic_dec_and_test(&de->count))
if (atomic_read(&de->count) == 0)
	free_proc_entry(de);
						/* boom! */
						if (de->deleted)
							free_proc_entry(de);
BUG: unable to handle kernel paging request at virtual address 
6b6b6b6b
printing eip: 
c10acdda *pdpt = 
00000000338f8001 *pde = 
0000000000000000
Oops: 0000 [#1] PREEMPT SMP
Modules linked in: af_packet ipv6 cpufreq_ondemand loop serio_raw psmouse k8temp hwmon sr_mod cdrom
Pid: 23161, comm: cat Not tainted (2.6.24-rc2-
8c0863403f109a43d7000b4646da4818220d501f #4)
EIP: 0060:[<
c10acdda>] EFLAGS: 
00210097 CPU: 1
EIP is at strnlen+0x6/0x18
EAX: 
6b6b6b6b EBX: 
6b6b6b6b ECX: 
6b6b6b6b EDX: 
fffffffe
ESI: 
c128fa3b EDI: 
f380bf34 EBP: 
ffffffff ESP: 
f380be44
 DS: 007b ES: 007b FS: 00d8 GS: 0033 SS: 0068
Process cat (pid: 23161, ti=
f380b000 task=
f38f2570 task.ti=
f380b000)
Stack: 
c10ac4f0 00000278 c12ce000 f43cd2a8 00000163 00000000 7da86067 00000400
       c128fa20 00896b18 f38325a8 c128fe20 ffffffff 00000000 c11f291e 00000400
       f75be300 c128fa20 f769c9a0 c10ac779 f380bf34 f7bfee70 c1018e6b f380bf34
Call Trace:
 [<
c10ac4f0>] vsnprintf+0x2ad/0x49b
 [<
c10ac779>] vscnprintf+0x14/0x1f
 [<
c1018e6b>] vprintk+0xc5/0x2f9
 [<
c10379f1>] handle_fasteoi_irq+0x0/0xab
 [<
c1004f44>] do_IRQ+0x9f/0xb7
 [<
c117db3b>] preempt_schedule_irq+0x3f/0x5b
 [<
c100264e>] need_resched+0x1f/0x21
 [<
c10190ba>] printk+0x1b/0x1f
 [<
c107c8ad>] de_put+0x3d/0x50
 [<
c107c8f8>] proc_delete_inode+0x38/0x41
 [<
c107c8c0>] proc_delete_inode+0x0/0x41
 [<
c1066298>] generic_delete_inode+0x5e/0xc6
 [<
c1065aa9>] iput+0x60/0x62
 [<
c1063c8e>] d_kill+0x2d/0x46
 [<
c1063fa9>] dput+0xdc/0xe4
 [<
c10571a1>] __fput+0xb0/0xcd
 [<
c1054e49>] filp_close+0x48/0x4f
 [<
c1055ee9>] sys_close+0x67/0xa5
 [<
c10026b6>] sysenter_past_esp+0x5f/0x85
=======================
Code: c9 74 0c f2 ae 74 05 bf 01 00 00 00 4f 89 fa 5f 89 d0 c3 85 c9 57 89 c7 89 d0 74 05 f2 ae 75 01 4f 89 f8 5f c3 89 c1 89 c8 eb 06 <80> 38 00 74 07 40 4a 83 fa ff 75 f4 29 c8 c3 90 90 90 57 83 c9
EIP: [<
c10acdda>] strnlen+0x6/0x18 SS:ESP 0068:
f380be44
Also, remove broken usage of ->deleted from reiserfs: if sget() succeeds,
module is already pinned and remove_proc_entry() can't happen => nobody
can mark PDE deleted.
Dummy proc root in netns code is not marked with refcount 1. AFAICS, we
never get it, it's just for proper /proc/net removal. I double checked
CLONE_NETNS continues to work.
Patch survives many hours of modprobe/rmmod/cat loops without new bugs
which can be attributed to refcounting.
Signed-off-by: Alexey Dobriyan <adobriyan@sw.ru>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
        ent->namelen = len;
        ent->mode = mode;
        ent->nlink = nlink;
+       atomic_set(&ent->count, 1);
        ent->pde_users = 0;
        spin_lock_init(&ent->pde_unload_lock);
        ent->pde_unload_completion = NULL;
 
 /*
  * Remove a /proc entry and free it if it's not currently in use.
- * If it is in use, we set the 'deleted' flag.
  */
 void remove_proc_entry(const char *name, struct proc_dir_entry *parent)
 {
                        parent->nlink--;
                de->nlink = 0;
                WARN_ON(de->subdir);
-               if (!atomic_read(&de->count))
+               if (atomic_dec_and_test(&de->count))
                        free_proc_entry(de);
-               else {
-                       de->deleted = 1;
-                       printk("remove_proc_entry: %s/%s busy, count=%d\n",
-                               parent->name, de->name, atomic_read(&de->count));
-               }
                break;
        }
        spin_unlock(&proc_subdir_lock);
 
                        return;
                }
 
-               if (atomic_dec_and_test(&de->count)) {
-                       if (de->deleted) {
-                               printk("de_put: deferred delete of %s\n",
-                                       de->name);
-                               free_proc_entry(de);
-                       }
-               }               
+               if (atomic_dec_and_test(&de->count))
+                       free_proc_entry(de);
                unlock_kernel();
        }
 }
 
        .name           = "/proc",
        .mode           = S_IFDIR | S_IRUGO | S_IXUGO, 
        .nlink          = 2, 
+       .count          = ATOMIC_INIT(1),
        .proc_iops      = &proc_root_inode_operations, 
        .proc_fops      = &proc_root_operations,
        .parent         = &proc_root,
 
                return NULL;
 
        up_write(&s->s_umount);
-
-       if (de->deleted) {
-               deactivate_super(s);
-               return NULL;
-       }
-
        return s;
 }
 
 
        read_proc_t *read_proc;
        write_proc_t *write_proc;
        atomic_t count;         /* use count */
-       int deleted;            /* delete flag */
        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;