Re: [PATCH 00/10] Cover letter: fix a race in release_task when flushing the dentry

From: Wen Yang
Date: Sun Jan 03 2021 - 23:16:40 EST




在 2020/12/31 下午5:22, Greg Kroah-Hartman 写道:
On Thu, Dec 17, 2020 at 10:26:23AM +0800, Wen Yang wrote:


在 2020/12/4 上午2:31, Wen Yang 写道:
The dentries such as /proc/<pid>/ns/ have the DCACHE_OP_DELETE flag, they
should be deleted when the process exits.

Suppose the following race appears:

release_task                 dput
-> proc_flush_task
                            -> dentry->d_op->d_delete(dentry)
-> __exit_signal
                             -> dentry->d_lockref.count--  and return.

In the proc_flush_task(), if another process is using this dentry, it will
not be deleted. At the same time, in dput(), d_op->d_delete() can be executed
before __exit_signal(pid has not been hashed), d_delete returns false, so
this dentry still cannot be deleted.

This dentry will always be cached (although its count is 0 and the
DCACHE_OP_DELETE flag is set), its parent denry will also be cached too, and
these dentries can only be deleted when drop_caches is manually triggered.

This will result in wasted memory. What's more troublesome is that these
dentries reference pid, according to the commit f333c700c610 ("pidns: Add a
limit on the number of pid namespaces"), if the pid cannot be released, it
may result in the inability to create a new pid_ns.

This problem occurred in our cluster environment (Linux 4.9 LTS).
We could reproduce it by manually constructing a test program + adding some
debugging switches in the kernel:
* A test program to open the directory (/proc/<pid>/ns) [1]
* Adding some debugging switches to the kernel, adding a delay between
proc_flush_task and __exit_signal in release_task() [2]

The test process is as follows:

A, terminal #1

Turn on the debug switch:
echo 1> /proc/sys/vm/dentry_debug_trace

Execute the following unshare command:
sudo unshare --pid --fork --mount-proc bash


B, terminal #2

Find the pid of the unshare process:

# pstree -p | grep unshare
| `-sshd(716)---bash(718)--sudo(816)---unshare(817)---bash(818)


Find the corresponding dentry:
# dmesg | grep pid=818
[70.424722] XXX proc_pid_instantiate:3119 pid=818 tid=818 entry=818/ffff8802c7b670e8


C, terminal #3

Execute the opendir program, it will always open the /proc/818/ns/ directory:

# ./a.out /proc/818/ns/
pid: 876
.
..
net
uts
ipc
pid
user
mnt
cgroup

D, go back to terminal #2

Turn on the debugging switches to construct the race:
# echo 818> /proc/sys/vm/dentry_debug_pid
# echo 1> /proc/sys/vm/dentry_debug_delay

Kill the unshare process (pid 818). Since the debugging switches have been
turned on, it will get stuck in release_task():
# kill -9 818

Then kill the process that opened the /proc/818/ns/ directory:
# kill -9 876

Then turn off these debugging switches to allow the 818 process to exit:
# echo 0> /proc/sys/vm/dentry_debug_delay
# echo 0> /proc/sys/vm/dentry_debug_pid

Checking the dmesg, we will find that the dentry(/proc/818/ns) ’s count is 0,
and the flag is 2800cc (#define DCACHE_OP_DELETE 0x00000008), but it is still
cached:
# dmesg | grep ffff8802a3999548

[565.559156] XXX dput:853 dentry=ns/ffff8802bea7b528, flag=2800cc, cnt=0, inode=ffff8802b38c2010, pdentry=818/ffff8802c7b670e8, pflag=20008c, pcnt=1, pinode=ffff8802c7812010, keywords: be cached


It could also be verified via the crash tool:

crash> dentry.d_flags,d_iname,d_inode,d_lockref -x ffff8802bea7b528
d_flags = 0x2800cc
d_iname = "ns\000kkkkkkkkkkkkkkkkkkkkkkkkkkkk"
d_inode = 0xffff8802b38c2010
d_lockref = {
{
lock_count = 0x0,
{
lock = {
{
rlock = {
raw_lock = {
{
val = {
counter = 0x0
},
{
locked = 0x0,
pending = 0x0
},
{
locked_pending = 0x0,
tail = 0x0
}
}
}
}
}
},
count = 0x0
}
}
}
crash> kmem ffff8802bea7b528
CACHE OBJSIZE ALLOCATED TOTAL SLABS SSIZE NAME
ffff8802dd5f5900 192 23663 26130 871 16k dentry
SLAB MEMORY NODE TOTAL ALLOCATED FREE
ffffea000afa9e00 ffff8802bea78000 0 30 25 5
FREE / [ALLOCATED]
[ffff8802bea7b520]

PAGE PHYSICAL MAPPING INDEX CNT FLAGS
ffffea000afa9ec0 2bea7b000 dead000000000400 0 0 2fffff80000000
crash>

This series of patches is to fix this issue.

Regards,
Wen

Alexey Dobriyan (1):
proc: use %u for pid printing and slightly less stack

Andreas Gruenbacher (1):
proc: Pass file mode to proc_pid_make_inode

Christian Brauner (1):
clone: add CLONE_PIDFD

Eric W. Biederman (6):
proc: Better ownership of files for non-dumpable tasks in user
namespaces
proc: Rename in proc_inode rename sysctl_inodes sibling_inodes
proc: Generalize proc_sys_prune_dcache into proc_prune_siblings_dcache
proc: Clear the pieces of proc_inode that proc_evict_inode cares about
proc: Use d_invalidate in proc_prune_siblings_dcache
proc: Use a list of inodes to flush from proc

Joel Fernandes (Google) (1):
pidfd: add polling support

fs/proc/base.c | 242 ++++++++++++++++++++-------------------------
fs/proc/fd.c | 20 +---
fs/proc/inode.c | 67 ++++++++++++-
fs/proc/internal.h | 22 ++---
fs/proc/namespaces.c | 3 +-
fs/proc/proc_sysctl.c | 45 ++-------
fs/proc/self.c | 6 +-
fs/proc/thread_self.c | 5 +-
include/linux/pid.h | 5 +
include/linux/proc_fs.h | 4 +-
include/uapi/linux/sched.h | 1 +
kernel/exit.c | 5 +-
kernel/fork.c | 145 ++++++++++++++++++++++++++-
kernel/pid.c | 3 +
kernel/signal.c | 11 +++
security/selinux/hooks.c | 1 +
16 files changed, 357 insertions(+), 228 deletions(-)

[1] A test program to open the directory (/proc/<pid>/ns)
#include <stdio.h>
#include <sys/types.h>
#include <dirent.h>
#include <errno.h>

int main(int argc, char *argv[])
{
DIR *dip;
struct dirent *dit;

if (argc < 2) {
printf("Usage :%s <directory>\n", argv[0]);
return -1;
}

if ((dip = opendir(argv[1])) == NULL) {
perror("opendir");
return -1;
}

printf("pid: %d\n", getpid());
while((dit = readdir (dip)) != NULL) {
printf("%s\n", dit->d_name);
}

while (1)
sleep (1);

return 0;
}

[2] Adding some debugging switches to the kernel, also adding a delay between
proc_flush_task and __exit_signal in release_task():

diff --git a/fs/dcache.c b/fs/dcache.c
index 05bad55..fafad37 100644
--- a/fs/dcache.c
+++ b/fs/dcache.c
@@ -84,6 +84,9 @@
int sysctl_vfs_cache_pressure __read_mostly = 100;
EXPORT_SYMBOL_GPL(sysctl_vfs_cache_pressure);

+int sysctl_dentry_debug_trace __read_mostly = 0;
+EXPORT_SYMBOL_GPL(sysctl_dentry_debug_trace);
+
__cacheline_aligned_in_smp DEFINE_SEQLOCK(rename_lock);

EXPORT_SYMBOL(rename_lock);
@@ -758,6 +761,26 @@ static inline bool fast_dput(struct dentry *dentry)
return 0;
}

+#define DENTRY_DEBUG_TRACE(dentry, keywords) \
+do { \
+ if (sysctl_dentry_debug_trace) \
+ printk("XXX %s:%d " \
+ "dentry=%s/%p, flag=%x, cnt=%d, inode=%p, " \
+ "pdentry=%s/%p, pflag=%x, pcnt=%d, pinode=%p, " \
+ "keywords: %s\n", \
+ __func__, __LINE__, \
+ dentry->d_name.name, \
+ dentry, \
+ dentry->d_flags, \
+ dentry->d_lockref.count, \
+ dentry->d_inode, \
+ dentry->d_parent->d_name.name, \
+ dentry->d_parent, \
+ dentry->d_parent->d_flags, \
+ dentry->d_parent->d_lockref.count, \
+ dentry->d_parent->d_inode, \
+ keywords); \
+} while (0)

/*
* This is dput
@@ -804,6 +827,8 @@ void dput(struct dentry *dentry)

WARN_ON(d_in_lookup(dentry));

+ DENTRY_DEBUG_TRACE(dentry, "be checked");
+
/* Unreachable? Get rid of it */
if (unlikely(d_unhashed(dentry)))
goto kill_it;
@@ -812,8 +837,10 @@ void dput(struct dentry *dentry)
goto kill_it;

if (unlikely(dentry->d_flags & DCACHE_OP_DELETE)) {
- if (dentry->d_op->d_delete(dentry))
+ if (dentry->d_op->d_delete(dentry)) {
+ DENTRY_DEBUG_TRACE(dentry, "be killed");
goto kill_it;
+ }
}

if (!(dentry->d_flags & DCACHE_REFERENCED))
@@ -822,6 +849,9 @@ void dput(struct dentry *dentry)

dentry->d_lockref.count--;
spin_unlock(&dentry->d_lock);
+
+ DENTRY_DEBUG_TRACE(dentry, "be cached");
+
return;

kill_it:
diff --git a/fs/proc/base.c b/fs/proc/base.c
index b9e4183..419a409 100644
--- a/fs/proc/base.c
+++ b/fs/proc/base.c
@@ -3090,6 +3090,8 @@ void proc_flush_task(struct task_struct *task)
}
}

+extern int sysctl_dentry_debug_trace;
+
static int proc_pid_instantiate(struct inode *dir,
struct dentry * dentry,
struct task_struct *task, const void *ptr)
@@ -3111,6 +3113,12 @@ static int proc_pid_instantiate(struct inode *dir,
d_set_d_op(dentry, &pid_dentry_operations);

d_add(dentry, inode);
+
+ if (sysctl_dentry_debug_trace)
+ printk("XXX %s:%d pid=%d tid=%d entry=%s/%p\n",
+ __func__, __LINE__, task->pid, task->tgid,
+ dentry->d_name.name, dentry);
+
/* Close the race of the process dying before we return the dentry */
if (pid_revalidate(dentry, 0))
return 0;
diff --git a/kernel/exit.c b/kernel/exit.c
index 27f4168..2b3e1b6 100644
--- a/kernel/exit.c
+++ b/kernel/exit.c
@@ -55,6 +55,8 @@
#include <linux/shm.h>
#include <linux/kcov.h>

+#include <linux/delay.h>
+
#include <asm/uaccess.h>
#include <asm/unistd.h>
#include <asm/pgtable.h>
@@ -164,6 +166,8 @@ static void delayed_put_task_struct(struct rcu_head *rhp)
put_task_struct(tsk);
}

+int sysctl_dentry_debug_delay __read_mostly = 0;
+int sysctl_dentry_debug_pid __read_mostly = 0;

void release_task(struct task_struct *p)
{
@@ -178,6 +182,11 @@ void release_task(struct task_struct *p)

proc_flush_task(p);

+ if (sysctl_dentry_debug_delay && p->pid == sysctl_dentry_debug_pid) {
+ while (sysctl_dentry_debug_delay)
+ mdelay(1);
+ }
+
write_lock_irq(&tasklist_lock);
ptrace_release_task(p);
__exit_signal(p);
diff --git a/kernel/sysctl.c b/kernel/sysctl.c
index 513e6da..27f1395 100644
--- a/kernel/sysctl.c
+++ b/kernel/sysctl.c
@@ -282,6 +282,10 @@ static int sysrq_sysctl_handler(struct ctl_table *table, int write,
static int max_extfrag_threshold = 1000;
#endif

+extern int sysctl_dentry_debug_trace;
+extern int sysctl_dentry_debug_delay;
+extern int sysctl_dentry_debug_pid;
+
static struct ctl_table kern_table[] = {
{
.procname = "sched_child_runs_first",
@@ -1498,6 +1502,30 @@ static int sysrq_sysctl_handler(struct ctl_table *table, int write,
.proc_handler = proc_dointvec,
.extra1 = &zero,
},
+ {
+ .procname = "dentry_debug_trace",
+ .data = &sysctl_dentry_debug_trace,
+ .maxlen = sizeof(sysctl_dentry_debug_trace),
+ .mode = 0644,
+ .proc_handler = proc_dointvec,
+ .extra1 = &zero,
+ },
+ {
+ .procname = "dentry_debug_delay",
+ .data = &sysctl_dentry_debug_delay,
+ .maxlen = sizeof(sysctl_dentry_debug_delay),
+ .mode = 0644,
+ .proc_handler = proc_dointvec,
+ .extra1 = &zero,
+ },
+ {
+ .procname = "dentry_debug_pid",
+ .data = &sysctl_dentry_debug_pid,
+ .maxlen = sizeof(sysctl_dentry_debug_pid),
+ .mode = 0644,
+ .proc_handler = proc_dointvec,
+ .extra1 = &zero,
+ },
#ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
{
.procname = "legacy_va_layout",


Signed-off-by: Wen Yang <wenyang@xxxxxxxxxxxxxxxxx>
Cc: Pavel Emelyanov <xemul@xxxxxxxxxx>
Cc: Oleg Nesterov <oleg@xxxxxxxxxx>
Cc: Sukadev Bhattiprolu <sukadev@xxxxxxxxxx>
Cc: Paul Menage <menage@xxxxxxxxxx>
Cc: "Eric W. Biederman" <ebiederm@xxxxxxxxxxxx>
Cc: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
Cc: <stable@xxxxxxxxxxxxxxx>


Hi Greg,

Could you kindly give some suggestions?

I'm sorry, but I don't really understand what this patch series is for.

Why is there a patch in the 00/10 email? What stable kernel(s) should
this be backported to? And why can't you just use a newer kernel (like
4.19) if you are hitting this issue with much older ones?


This bug was introduced by 60347f6716aa ("pid namespaces: prepare proc_flust_task() to flush entries from multiple proc trees"), exposed by f333c700c610 ("pidns: Add a limit on the number of pid namespaces"), and then fixed by 7bc3e6e55acf (“proc: Use a list of inodes to flush from proc”)

4.19 LTS did not solve it either.


The one that fixes the bug is this patch (10/10):
https://lkml.org/lkml/2020/12/3/1046
The previous 9 patches (from 01/10 to 09/10) are its pre-dependencies.

The 00/10 is some of our test programs, including a user mode program (open the /proc/<pid>/ns directory), and some hacks added to the kernel (just add log printing and some delays for easy construction this race).

Thanks.

--
Best wishes,
Wen