diff mbox series

proc: align mnt_id in /proc/pid/fdinfo and /proc/pid/mountinfo

Message ID 20191121070613.4286-1-hu1.chen@intel.com (mailing list archive)
State New, archived
Headers show
Series proc: align mnt_id in /proc/pid/fdinfo and /proc/pid/mountinfo | expand

Commit Message

Chen, Hu1 Nov. 21, 2019, 7:06 a.m. UTC 
For Android application process, we found that the mnt_id read from
/proc/pid/fdinfo doesn't exist in /proc/pid/mountinfo. Thus CRIU fails
to dump such process and it complains

"(00.019206) Error (criu/files-reg.c:1299): Can't lookup mount=42 for
fd=-3 path=/data/dalvik-cache/x86_64/system@framework@boot.art"

This is due to how Android application is launched. In Android, there is
a special process called Zygote which handles the forking of each new
application process:
0. Zygote opens and maps some files, for example
   "/data/dalvik-cache/x86_64/system@framework@boot.art" in its current
   mount namespace, say "old mnt ns".
1. Zygote waits for the request to fork a new application.
2. Zygote gets a request, it forks and run the new process in a new
   mount namespace, say "new mnt ns".

The file opened in step 0 ties to the mount point in "old mnt ns". The
mnt_id of that mount is listed in /proc/pid/fdinfo. However,
/proc/pid/mountinfo points to current ns, i.e., "new mnt ns".

Althgouh this issue is exposed in Android, we believe it's generic.
Prcoess may open files and enter new mnt ns.

To address it, this patch searches the mirror mount in current ns with
MAJOR and MINOR and shows the mirror's mnt_id.

Signed-off-by: Chen, Hu <hu1.chen@intel.com>

Comments

Miklos Szeredi Nov. 21, 2019, 10:03 a.m. UTC | #1 
On Thu, Nov 21, 2019 at 8:28 AM Chen, Hu <hu1.chen@intel.com> wrote:
>
> For Android application process, we found that the mnt_id read from
> /proc/pid/fdinfo doesn't exist in /proc/pid/mountinfo. Thus CRIU fails
> to dump such process and it complains
>
> "(00.019206) Error (criu/files-reg.c:1299): Can't lookup mount=42 for
> fd=-3 path=/data/dalvik-cache/x86_64/system@framework@boot.art"
>
> This is due to how Android application is launched. In Android, there is
> a special process called Zygote which handles the forking of each new
> application process:
> 0. Zygote opens and maps some files, for example
>    "/data/dalvik-cache/x86_64/system@framework@boot.art" in its current
>    mount namespace, say "old mnt ns".
> 1. Zygote waits for the request to fork a new application.
> 2. Zygote gets a request, it forks and run the new process in a new
>    mount namespace, say "new mnt ns".
>
> The file opened in step 0 ties to the mount point in "old mnt ns". The
> mnt_id of that mount is listed in /proc/pid/fdinfo. However,
> /proc/pid/mountinfo points to current ns, i.e., "new mnt ns".
>
> Althgouh this issue is exposed in Android, we believe it's generic.
> Prcoess may open files and enter new mnt ns.
>
> To address it, this patch searches the mirror mount in current ns with
> MAJOR and MINOR and shows the mirror's mnt_id.

This is a hack.   I suggest instead to add a new line to fdinfo with
the MAJOR:MINOR number of the device.

Thanks,
Miklos
diff mbox series

Patch

diff --git a/fs/mount.h b/fs/mount.h
index 711a4093e475..6bbfc2b3b8ba 100644
--- a/fs/mount.h
+++ b/fs/mount.h
@@ -153,3 +153,5 @@  static inline bool is_anon_ns(struct mnt_namespace *ns)
 {
 	return ns->seq == 0;
 }
+
+extern struct mount *lookup_mirror_mnt(const struct mount *mnt);
diff --git a/fs/namespace.c b/fs/namespace.c
index 2adfe7b166a3..4ea9b4464cd0 100644
--- a/fs/namespace.c
+++ b/fs/namespace.c
@@ -683,6 +683,36 @@  bool __is_local_mountpoint(struct dentry *dentry)
 	return is_covered;
 }
 
+/*
+ * lookup_mirror_mnt - Return @mnt's mirror mount in the current/local mount
+ * namespace. If mirror isn't found, just return NULL.
+ */
+struct mount *lookup_mirror_mnt(const struct mount *mnt)
+{
+	struct mnt_namespace *ns = current->nsproxy->mnt_ns;
+	struct mount *mnt_local;
+	bool is_matched = false;
+
+	/* mnt belongs to current namesapce */
+	if (mnt->mnt_ns == ns)
+		return mnt;
+
+	down_read(&namespace_sem);
+	list_for_each_entry(mnt_local, &ns->list, mnt_list) {
+		struct super_block *sb = mnt->mnt.mnt_sb;
+		struct super_block *sb_local = mnt_local->mnt.mnt_sb;
+
+		if (MAJOR(sb->s_dev) == MAJOR(sb_local->s_dev) &&
+		    MINOR(sb->s_dev) == MINOR(sb_local->s_dev)) {
+			is_matched = true;
+			break;
+		}
+	}
+	up_read(&namespace_sem);
+
+	return is_matched ? mnt_local : NULL;
+}
+
 static struct mountpoint *lookup_mountpoint(struct dentry *dentry)
 {
 	struct hlist_head *chain = mp_hash(dentry);
diff --git a/fs/proc/fd.c b/fs/proc/fd.c
index 81882a13212d..cbf2571b0620 100644
--- a/fs/proc/fd.c
+++ b/fs/proc/fd.c
@@ -23,6 +23,7 @@  static int seq_show(struct seq_file *m, void *v)
 	int f_flags = 0, ret = -ENOENT;
 	struct file *file = NULL;
 	struct task_struct *task;
+	struct mount *mount = NULL;
 
 	task = get_proc_task(m->private);
 	if (!task)
@@ -53,9 +54,16 @@  static int seq_show(struct seq_file *m, void *v)
 	if (ret)
 		return ret;
 
+	/* After unshare -m, real_mount(file->f_path.mnt) is not meaningful in
+	 * current mount namesapce. We want to know the mnt_id in current mount
+	 * namespace
+	 */
+	mount = lookup_mirror_mnt(real_mount(file->f_path.mnt));
+	if (!mount)
+		mount = real_mount(file->f_path.mnt);
+
 	seq_printf(m, "pos:\t%lli\nflags:\t0%o\nmnt_id:\t%i\n",
-		   (long long)file->f_pos, f_flags,
-		   real_mount(file->f_path.mnt)->mnt_id);
+		   (long long)file->f_pos, f_flags, mount->mnt_id);
 
 	show_fd_locks(m, file, files);
 	if (seq_has_overflowed(m))