fault-inject: support systematic fault injection

Add /proc/self/task/<current-tid>/fail-nth file that allows failing
0-th, 1-st, 2-nd and so on calls systematically.
Excerpt from the added documentation:

 "Write to this file of integer N makes N-th call in the current task
  fail (N is 0-based). Read from this file returns a single char 'Y' or
  'N' that says if the fault setup with a previous write to this file
  was injected or not, and disables the fault if it wasn't yet injected.
  Note that this file enables all types of faults (slab, futex, etc).
  This setting takes precedence over all other generic settings like
  probability, interval, times, etc. But per-capability settings (e.g.
  fail_futex/ignore-private) take precedence over it. This feature is
  intended for systematic testing of faults in a single system call. See
  an example below"

Why add a new setting:
1. Existing settings are global rather than per-task.
   So parallel testing is not possible.
2. attr->interval is close but it depends on attr->count
   which is non reset to 0, so interval does not work as expected.
3. Trying to model this with existing settings requires manipulations
   of all of probability, interval, times, space, task-filter and
   unexposed count and per-task make-it-fail files.
4. Existing settings are per-failure-type, and the set of failure
   types is potentially expanding.
5. make-it-fail can't be changed by unprivileged user and aggressive
   stress testing better be done from an unprivileged user.
   Similarly, this would require opening the debugfs files to the
   unprivileged user, as he would need to reopen at least times file
   (not possible to pre-open before dropping privs).

The proposed interface solves all of the above (see the example).

We want to integrate this into syzkaller fuzzer.  A prototype has found
10 bugs in kernel in first day of usage:

  https://groups.google.com/forum/#!searchin/syzkaller/%22FAULT_INJECTION%22%7Csort:relevance

I've made the current interface work with all types of our sandboxes.
For setuid the secret sauce was prctl(PR_SET_DUMPABLE, 1, 0, 0, 0) to
make /proc entries non-root owned.  So I am fine with the current
version of the code.

[akpm@linux-foundation.org: fix build]
Link: http://lkml.kernel.org/r/20170328130128.101773-1-dvyukov@google.com
Signed-off-by: Dmitry Vyukov <dvyukov@google.com>
Cc: Akinobu Mita <akinobu.mita@gmail.com>
Cc: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

5 files changed, 142 insertions, 0 deletions

diff --git a/Documentation/fault-injection/fault-injection.txt b/Documentation/fault-injection/fault-injection.txt
index 415484f..192d8cb 100644
--- a/Documentation/fault-injection/fault-injection.txt
+++ b/Documentation/fault-injection/fault-injection.txt
@@ -134,6 +134,22 @@ use the boot option:
 	fail_futex=
 	mmc_core.fail_request=<interval>,<probability>,<space>,<times>
 
+o proc entries
+
+- /proc/self/task/<current-tid>/fail-nth:
+
+	Write to this file of integer N makes N-th call in the current task fail
+	(N is 0-based). Read from this file returns a single char 'Y' or 'N'
+	that says if the fault setup with a previous write to this file was
+	injected or not, and disables the fault if it wasn't yet injected.
+	Note that this file enables all types of faults (slab, futex, etc).
+	This setting takes precedence over all other generic debugfs settings
+	like probability, interval, times, etc. But per-capability settings
+	(e.g. fail_futex/ignore-private) take precedence over it.
+
+	This feature is intended for systematic testing of faults in a single
+	system call. See an example below.
+
 How to add new fault injection capability
 -----------------------------------------
 
@@ -278,3 +294,65 @@ allocation failure.
 	# env FAILCMD_TYPE=fail_page_alloc \
 		./tools/testing/fault-injection/failcmd.sh --times=100 \
                 -- make -C tools/testing/selftests/ run_tests
+
+Systematic faults using fail-nth
+---------------------------------
+
+The following code systematically faults 0-th, 1-st, 2-nd and so on
+capabilities in the socketpair() system call.
+
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <sys/socket.h>
+#include <sys/syscall.h>
+#include <fcntl.h>
+#include <unistd.h>
+#include <string.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <errno.h>
+
+int main()
+{
+	int i, err, res, fail_nth, fds[2];
+	char buf[128];
+
+	system("echo N > /sys/kernel/debug/failslab/ignore-gfp-wait");
+	sprintf(buf, "/proc/self/task/%ld/fail-nth", syscall(SYS_gettid));
+	fail_nth = open(buf, O_RDWR);
+	for (i = 0;; i++) {
+		sprintf(buf, "%d", i);
+		write(fail_nth, buf, strlen(buf));
+		res = socketpair(AF_LOCAL, SOCK_STREAM, 0, fds);
+		err = errno;
+		read(fail_nth, buf, 1);
+		if (res == 0) {
+			close(fds[0]);
+			close(fds[1]);
+		}
+		printf("%d-th fault %c: res=%d/%d\n", i, buf[0], res, err);
+		if (buf[0] != 'Y')
+			break;
+	}
+	return 0;
+}
+
+An example output:
+
+0-th fault Y: res=-1/23
+1-th fault Y: res=-1/23
+2-th fault Y: res=-1/23
+3-th fault Y: res=-1/12
+4-th fault Y: res=-1/12
+5-th fault Y: res=-1/23
+6-th fault Y: res=-1/23
+7-th fault Y: res=-1/23
+8-th fault Y: res=-1/12
+9-th fault Y: res=-1/12
+10-th fault Y: res=-1/12
+11-th fault Y: res=-1/12
+12-th fault Y: res=-1/12
+13-th fault Y: res=-1/12
+14-th fault Y: res=-1/12
+15-th fault Y: res=-1/12
+16-th fault N: res=0/12
diff --git a/fs/proc/base.c b/fs/proc/base.c
index f1e1927..88b773f 100644
--- a/fs/proc/base.c
+++ b/fs/proc/base.c
@@ -1355,6 +1355,53 @@ static const struct file_operations proc_fault_inject_operations = {
 	.write		= proc_fault_inject_write,
 	.llseek		= generic_file_llseek,
 };
+
+static ssize_t proc_fail_nth_write(struct file *file, const char __user *buf,
+				   size_t count, loff_t *ppos)
+{
+	struct task_struct *task;
+	int err, n;
+
+	task = get_proc_task(file_inode(file));
+	if (!task)
+		return -ESRCH;
+	put_task_struct(task);
+	if (task != current)
+		return -EPERM;
+	err = kstrtoint_from_user(buf, count, 10, &n);
+	if (err)
+		return err;
+	if (n < 0 || n == INT_MAX)
+		return -EINVAL;
+	current->fail_nth = n + 1;
+	return count;
+}
+
+static ssize_t proc_fail_nth_read(struct file *file, char __user *buf,
+				  size_t count, loff_t *ppos)
+{
+	struct task_struct *task;
+	int err;
+
+	task = get_proc_task(file_inode(file));
+	if (!task)
+		return -ESRCH;
+	put_task_struct(task);
+	if (task != current)
+		return -EPERM;
+	if (count < 1)
+		return -EINVAL;
+	err = put_user((char)(current->fail_nth ? 'N' : 'Y'), buf);
+	if (err)
+		return err;
+	current->fail_nth = 0;
+	return 1;
+}
+
+static const struct file_operations proc_fail_nth_operations = {
+	.read		= proc_fail_nth_read,
+	.write		= proc_fail_nth_write,
+};
 #endif
 
 
@@ -3311,6 +3358,11 @@ static const struct pid_entry tid_base_stuff[] = {
 #endif
 #ifdef CONFIG_FAULT_INJECTION
 	REG("make-it-fail", S_IRUGO|S_IWUSR, proc_fault_inject_operations),
+	/*
+	 * Operations on the file check that the task is current,
+	 * so we create it with 0666 to support testing under unprivileged user.
+	 */
+	REG("fail-nth", 0666, proc_fail_nth_operations),
 #endif
 #ifdef CONFIG_TASK_IO_ACCOUNTING
 	ONE("io",	S_IRUSR, proc_tid_io_accounting),
diff --git a/include/linux/sched.h b/include/linux/sched.h
index 20814b7..3822d74 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -974,6 +974,7 @@ struct task_struct {
 
 #ifdef CONFIG_FAULT_INJECTION
 	int				make_it_fail;
+	int fail_nth;
 #endif
 	/*
 	 * When (nr_dirtied >= nr_dirtied_pause), it's time to call
diff --git a/kernel/fork.c b/kernel/fork.c
index d2b9d7c..ade237a 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -573,6 +573,10 @@ static struct task_struct *dup_task_struct(struct task_struct *orig, int node)
 
 	kcov_task_init(tsk);
 
+#ifdef CONFIG_FAULT_INJECTION
+	tsk->fail_nth = 0;
+#endif
+
 	return tsk;
 
 free_stack:
diff --git a/lib/fault-inject.c b/lib/fault-inject.c
index 4ff1571..09ac73c1 100644
--- a/lib/fault-inject.c
+++ b/lib/fault-inject.c
@@ -107,6 +107,12 @@ static inline bool fail_stacktrace(struct fault_attr *attr)
 
 bool should_fail(struct fault_attr *attr, ssize_t size)
 {
+	if (in_task() && current->fail_nth) {
+		if (--current->fail_nth == 0)
+			goto fail;
+		return false;
+	}
+
 	/* No need to check any other properties if the probability is 0 */
 	if (attr->probability == 0)
 		return false;
@@ -134,6 +140,7 @@ bool should_fail(struct fault_attr *attr, ssize_t size)
 	if (!fail_stacktrace(attr))
 		return false;
 
+fail:
 	fail_dump(attr);
 
 	if (atomic_read(&attr->times) != -1)