From mboxrd@z Thu Jan 1 00:00:00 1970 From: "Paul E. McKenney" Subject: Re: 2.6.34-rc5-git7 (plus all patches) -- another suspicious rcu_dereference_check() usage. Date: Thu, 29 Apr 2010 17:48:29 -0700 Message-ID: <20100430004829.GA18979@linux.vnet.ibm.com> References: <20100428175426.GK2540@linux.vnet.ibm.com> <1272483491.2201.9.camel@edumazet-laptop> <20100428200904.GS2540@linux.vnet.ibm.com> <1272485923.2201.22.camel@edumazet-laptop> <20100428204403.GT2540@linux.vnet.ibm.com> Reply-To: paulmck@linux.vnet.ibm.com Mime-Version: 1.0 Content-Type: text/plain; charset=iso-8859-1 Content-Transfer-Encoding: QUOTED-PRINTABLE Cc: Miles Lane , Vivek Goyal , Eric Paris , Lai Jiangshan , Ingo Molnar , Peter Zijlstra , LKML , nauman@google.com, netdev@vger.kernel.org, Jens Axboe , Gui Jianfeng , Li Zefan , Johannes Berg , shemminger@vyatta.com To: Eric Dumazet Return-path: Content-Disposition: inline In-Reply-To: <20100428204403.GT2540@linux.vnet.ibm.com> Sender: linux-kernel-owner@vger.kernel.org List-Id: netdev.vger.kernel.org On Wed, Apr 28, 2010 at 01:44:03PM -0700, Paul E. McKenney wrote: > On Wed, Apr 28, 2010 at 10:18:43PM +0200, Eric Dumazet wrote: > > Le mercredi 28 avril 2010 =E0 13:09 -0700, Paul E. McKenney a =E9cr= it : > > > On Wed, Apr 28, 2010 at 09:38:11PM +0200, Eric Dumazet wrote: > > > > Le mercredi 28 avril 2010 =E0 10:54 -0700, Paul E. McKenney a =E9= crit : > > > > > On Mon, Apr 26, 2010 at 08:51:06PM -0400, Miles Lane wrote: > > > > > > This one occurred during the wakeup from suspend to RAM. > > > > > >=20 > > > > > > [ 984.724697] [ INFO: suspicious rcu_dereference_check() u= sage. ] > > > > > > [ 984.724700] --------------------------------------------= ------- > > > > > > [ 984.724703] include/linux/fdtable.h:88 invoked > > > > > > rcu_dereference_check() without protection! > > > > > > [ 984.724706] > > > > > > [ 984.724707] other info that might help us debug this: > > > > > > [ 984.724708] > > > > > > [ 984.724711] > > > > > > [ 984.724711] rcu_scheduler_active =3D 1, debug_locks =3D = 1 > > > > > > [ 984.724714] no locks held by dbus-daemon/4680. > > > > > > [ 984.724717] > > > > > > [ 984.724717] stack backtrace: > > > > > > [ 984.724721] Pid: 4680, comm: dbus-daemon Not tainted 2.6= =2E34-rc5-git7 #33 > > > > > > [ 984.724724] Call Trace: > > > > > > [ 984.724734] [] lockdep_rcu_dereferenc= e+0x9d/0xa6 > > > > > > [ 984.724740] [] fcheck_files+0xb1/0xc9 > > > > > > [ 984.724745] [] fget_light+0x35/0xab > > > > > > [ 984.724751] [] ? sock_poll_wait+0x13/= 0x18 > > > > > > [ 984.724755] [] ? unix_poll+0x19/0x95 > > > > > > [ 984.724762] [] do_sys_poll+0x1ff/0x3e= 5 > > > > > > [ 984.724766] [] ? __pollwait+0x0/0xc7 > > > > > > [ 984.724771] [] ? pollwake+0x0/0x4f > > > > > > [ 984.724776] [] ? pollwake+0x0/0x4f > > > > > > [ 984.724780] [] ? pollwake+0x0/0x4f > > > > > > [ 984.724784] [] ? pollwake+0x0/0x4f > > > > > > [ 984.724788] [] ? pollwake+0x0/0x4f > > > > > > [ 984.724793] [] ? pollwake+0x0/0x4f > > > > > > [ 984.724797] [] ? pollwake+0x0/0x4f > > > > > > [ 984.724802] [] ? pollwake+0x0/0x4f > > > > > > [ 984.724806] [] ? pollwake+0x0/0x4f > > > > > > [ 984.724812] [] sys_poll+0x50/0xbb > > > > > > [ 984.724818] [] system_call_fastpath+0= x16/0x1b > > > > >=20 > > > > > Hmmm... I am not convinced that this is a false positive. C= ouldn't > > > > > there be a multi-threaded process where one thread is invokin= g poll() > > > > > on a UNIX socket just as another thread is calling close() on= it? > > > > >=20 > > > > > The current fcheck_files() logic requires that the caller eit= her (1) be in > > > > > an RCU read-side critical section, (2) hold ->files_lock, or = (3) passing > > > > > in a files_struct with ->count equal to 1 (initialization or = cleanup). > > > > >=20 > > > > > So I don't feel comfortable just slapping an RCU read-side cr= itical > > > > > section around this one, at least not unless someone who unde= rstands > > > > > the locking says that doing so is OK. > > > > >=20 > > > > > =09 > > > >=20 > > > > Its a single threaded program. > > > >=20 > > > > So fget_light() calls fcheck_files(files, fd); without rcu lock= , > > > > but some /proc/pid/fd/... user temporarly raised files->count j= ust > > > > before we perform the condition check. > > >=20 > > > So I should add a single-threaded check. My first thought was to= use > > > current_is_single_threaded(), but the bit about scanning the full= list > > > of processes does give me pause. However, thread_group_empty() l= ooks > > > like a much lighter-weight alternative. > > >=20 > > > I believe that it is possible for a pair of single-threaded proce= sses > > > to share a file descriptor, but that should not be a problem, as = both > > > of them would need to close it for it to go away. > > >=20 > > > But what happens if someone does a clone() with CLONE_FILES, as s= ome > > > of the AIO stuff seems to do? Won't that allow one of the result= ing > > > processes to close the file for both of them, even though both ar= e > > > otherwise single-threaded? And the ->count seems to be the only > > > distinction between these two cases. > > >=20 > > > And AIO does CLONE_VM as well as CLONE_FILES, but that seems to m= ean that > > > the check must scan the processes with current_is_single_threaded= (). > > > Besides which, a user could invoke clone() with only CLONE_FILES > > > specified, right? > > >=20 > > > Or am I just confused here? Not confused, just stupid. You would think that after almost 20 years doing parallel programming, I would understand that concurrent debuggin= g assertions usually cannot be exact. In this case, I really don't want false positives, so I must accept some false negatives. -That- I shoul= d be able to do without adding runtime overhead when !PROVE_RCU. I will put together a patch, test it, and send it out. Thanx, Paul > > If a program is mono threaded, and doing a fget_light() syscall, it > > cannot possibly do a clone() in // ;) >=20 > The sequence of events that I am worried about is as follows: >=20 > 1. Single-threaded process does clone(CLONE_FILES). The > result is a pair of single-threaded processes that share > file descriptors. >=20 > 2. One of these processes does files_fdtable(i) at the same > time as the other process closes file descriptor i. >=20 > So, clone and -then- do fget_light(). >=20 > > If we want to be picky, we could add a user provided condition, aka= "we > > are sure we are allowed to do this because we are the owner of the = files > > struct". >=20 > But yes, if I understand your trick below, the race conditions from > the above sequence of events would simply force the processes off > of the fget_light() path, which should be OK. >=20 > Thanx, Paul >=20 > > diff --git a/drivers/char/tty_io.c b/drivers/char/tty_io.c > > index 6da962c..027f5e1 100644 > > --- a/drivers/char/tty_io.c > > +++ b/drivers/char/tty_io.c > > @@ -2694,7 +2694,7 @@ void __do_SAK(struct tty_struct *tty) > > spin_lock(&p->files->file_lock); > > fdt =3D files_fdtable(p->files); > > for (i =3D 0; i < fdt->max_fds; i++) { > > - filp =3D fcheck_files(p->files, i); > > + filp =3D fcheck_files(p->files, i, false); > > if (!filp) > > continue; > > if (filp->f_op->read =3D=3D tty_read && > > diff --git a/fs/fcntl.c b/fs/fcntl.c > > index 452d02f..dabf4d8 100644 > > --- a/fs/fcntl.c > > +++ b/fs/fcntl.c > > @@ -119,7 +119,7 @@ SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsi= gned int, newfd) > > int retval =3D oldfd; > >=20 > > rcu_read_lock(); > > - if (!fcheck_files(files, oldfd)) > > + if (!fcheck_files(files, oldfd, false)) > > retval =3D -EBADF; > > rcu_read_unlock(); > > return retval; > > diff --git a/fs/file_table.c b/fs/file_table.c > > index 32d12b7..2865f72 100644 > > --- a/fs/file_table.c > > +++ b/fs/file_table.c > > @@ -274,7 +274,7 @@ struct file *fget(unsigned int fd) > > struct files_struct *files =3D current->files; > >=20 > > rcu_read_lock(); > > - file =3D fcheck_files(files, fd); > > + file =3D fcheck_files(files, fd, false); > > if (file) { > > if (!atomic_long_inc_not_zero(&file->f_count)) { > > /* File object ref couldn't be taken */ > > @@ -303,10 +303,10 @@ struct file *fget_light(unsigned int fd, int = *fput_needed) > >=20 > > *fput_needed =3D 0; > > if (likely((atomic_read(&files->count) =3D=3D 1))) { > > - file =3D fcheck_files(files, fd); > > + file =3D fcheck_files(files, fd, true); > > } else { > > rcu_read_lock(); > > - file =3D fcheck_files(files, fd); > > + file =3D fcheck_files(files, fd, false); > > if (file) { > > if (atomic_long_inc_not_zero(&file->f_count)) > > *fput_needed =3D 1; > > diff --git a/fs/proc/base.c b/fs/proc/base.c > > index 8418fcc..0e89448 100644 > > --- a/fs/proc/base.c > > +++ b/fs/proc/base.c > > @@ -1716,7 +1716,7 @@ static int proc_fd_info(struct inode *inode, = struct path *path, char *info) > > * hold ->file_lock. > > */ > > spin_lock(&files->file_lock); > > - file =3D fcheck_files(files, fd); > > + file =3D fcheck_files(files, fd, false); > > if (file) { > > if (path) { > > *path =3D file->f_path; > > @@ -1755,7 +1755,7 @@ static int tid_fd_revalidate(struct dentry *d= entry, struct nameidata *nd) > > files =3D get_files_struct(task); > > if (files) { > > rcu_read_lock(); > > - if (fcheck_files(files, fd)) { > > + if (fcheck_files(files, fd, false)) { > > rcu_read_unlock(); > > put_files_struct(files); > > if (task_dumpable(task)) { > > @@ -1813,7 +1813,7 @@ static struct dentry *proc_fd_instantiate(str= uct inode *dir, > > * hold ->file_lock. > > */ > > spin_lock(&files->file_lock); > > - file =3D fcheck_files(files, fd); > > + file =3D fcheck_files(files, fd, false); > > if (!file) > > goto out_unlock; > > if (file->f_mode & FMODE_READ) > > @@ -1899,7 +1899,7 @@ static int proc_readfd_common(struct file * f= ilp, void * dirent, > > char name[PROC_NUMBUF]; > > int len; > >=20 > > - if (!fcheck_files(files, fd)) > > + if (!fcheck_files(files, fd, false)) > > continue; > > rcu_read_unlock(); > >=20 > > diff --git a/include/linux/fdtable.h b/include/linux/fdtable.h > > index 013dc52..76423ad 100644 > > --- a/include/linux/fdtable.h > > +++ b/include/linux/fdtable.h > > @@ -57,11 +57,12 @@ struct files_struct { > > struct file * fd_array[NR_OPEN_DEFAULT]; > > }; > >=20 > > -#define rcu_dereference_check_fdtable(files, fdtfd) \ > > +#define rcu_dereference_check_fdtable(files, fdtfd, cond) \ > > (rcu_dereference_check((fdtfd), \ > > rcu_read_lock_held() || \ > > lockdep_is_held(&(files)->file_lock) || \ > > - atomic_read(&(files)->count) =3D=3D 1)) > > + atomic_read(&(files)->count) =3D=3D 1 || \ > > + cond)) > >=20 > > #define files_fdtable(files) \ > > (rcu_dereference_check_fdtable((files), (files)->fdt)) > > @@ -79,13 +80,13 @@ static inline void free_fdtable(struct fdtable = *fdt) > > call_rcu(&fdt->rcu, free_fdtable_rcu); > > } > >=20 > > -static inline struct file * fcheck_files(struct files_struct *file= s, unsigned int fd) > > +static inline struct file * fcheck_files(struct files_struct *file= s, unsigned int fd, bool cond) > > { > > struct file * file =3D NULL; > > struct fdtable *fdt =3D files_fdtable(files); > >=20 > > if (fd < fdt->max_fds) > > - file =3D rcu_dereference_check_fdtable(files, fdt->fd[fd]); > > + file =3D rcu_dereference_check_fdtable(files, fdt->fd[fd], cond)= ; > > return file; > > } > >=20 > >=20 > >=20