Skip to contentSkip to conversation

'strace -f' regression, bisected to tracehook

PATCH New Eduardo Habkost 2 participants 3 messages Aug 7, 2008 – Aug 8, 2008

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

Hi,

I have just hit a problem with strace when following forks, using recent trees. I have bisected the problem to commit 09a05394 (tracehook:
clone).

'strace -f' is not being able to trace child processes just after fork, and traces them only after the child has run for some time. I am getting the following output, when tracing a test program whose child exits just after returning from fork:

clone(Process 399 attached (waiting for parent)
* resume: ptrace(PTRACE_SYSCALL, ...): No such process
child_stack=0, flags=CLONE_CHILD_CLEARTID|CLONE_CHILD_SETTID|SIGCHLD, child_tidptr=0x7f8df681a780) = 399
[pid 398] --- SIGCHLD (Child exited) @ 0 (0) ---
[pid 398] rt_sigprocmask(SIG_BLOCK, [CHLD], [], 8) = 0
[...]

What I expect to get (and was getting on 2.6.26 and before the bisected commit) is:

clone(Process 391 attached (waiting for parent)
* Process 391 resumed (parent 390 ready)
child_stack=0, flags=CLONE_CHILD_CLEARTID|CLONE_CHILD_SETTID|SIGCHLD, child_tidptr=0x7fa84cf3c780) = 391
* [pid 391] exit_group(1) = ?
* Process 391 detached
--- SIGCHLD (Child exited) @ 0 (0) ---
rt_sigprocmask(SIG_BLOCK, [CHLD], [], 8) = 0
[...]

strace uses a trick to set the CLONE_PTRACE flag on clone() syscalls made by the traced process. I don't know if the trick used by strace is broken, or the handling of CLONE_PTRACE itself is broken.

The bisected commit is this:

commit 09a05394fe2448a4139b014936330af23fa7ec83

Author: Roland McGrath <[email protected]>
Date:   Fri Jul 25 19:45:47 2008 -0700

tracehook: clone

This moves all the ptrace initialization and tracing logic for task creation into tracehook.h and ptrace.h inlines. It reorganizes the code slightly, but should not change any behavior.

There are four tracehook entry points, at each important stage of task creation. This keeps the interface from the core fork.c code fairly clean, while supporting the complex setup required for ptrace or something like it.

include/linux/ptrace.h +22 -0
index c74abfc..dae6d85 100644
@@ -154,6 +154,28 @@ static inline int ptrace_event(int mask, int event, unsigned long message)
 	return 1;
 }
 
+/**
+ * ptrace_init_task - initialize ptrace state for a new child
+ * @child:		new child task
+ * @ptrace:		true if child should be ptrace'd by parent's tracer
+ *
+ * This is called immediately after adding @child to its parent's children
+ * list.  @ptrace is false in the normal case, and true to ptrace @child.
+ *
+ * Called with current's siglock and write_lock_irq(&tasklist_lock) held.
+ */
+static inline void ptrace_init_task(struct task_struct *child, bool ptrace)
+{
+	INIT_LIST_HEAD(&child->ptrace_entry);
+	INIT_LIST_HEAD(&child->ptraced);
+	child->parent = child->real_parent;
+	child->ptrace = 0;
+	if (unlikely(ptrace)) {
+		child->ptrace = current->ptrace;
+		__ptrace_link(child, current->parent);
+	}
+}
+
 #ifndef force_successful_syscall_return
 /*
  * System call handlers that, upon successful completion, need to return a
include/linux/tracehook.h +100 -0

104 lines not on this page · raw message

kernel/fork.c +28 -41

139 lines not on this page · raw message

Signature · 1 line
Eduardo
#1 lore raw

CLONE_PTRACE Oops (was Re: 'strace -f' regression, bisected to tracehook)

Quoted 32 lines · Eduardo Habkost

Hi,

I have just hit a problem with strace when following forks, using recent trees. I have bisected the problem to commit 09a05394 (tracehook:
clone).

'strace -f' is not being able to trace child processes just after fork, and traces them only after the child has run for some time. I am getting the following output, when tracing a test program whose child exits just after returning from fork:

clone(Process 399 attached (waiting for parent)
* resume: ptrace(PTRACE_SYSCALL, ...): No such process
child_stack=0, flags=CLONE_CHILD_CLEARTID|CLONE_CHILD_SETTID|SIGCHLD, child_tidptr=0x7f8df681a780) = 399
[pid 398] --- SIGCHLD (Child exited) @ 0 (0) ---
[pid 398] rt_sigprocmask(SIG_BLOCK, [CHLD], [], 8) = 0
[...]

What I expect to get (and was getting on 2.6.26 and before the bisected commit) is:

clone(Process 391 attached (waiting for parent)
* Process 391 resumed (parent 390 ready)
child_stack=0, flags=CLONE_CHILD_CLEARTID|CLONE_CHILD_SETTID|SIGCHLD, child_tidptr=0x7fa84cf3c780) = 391
* [pid 391] exit_group(1) = ?
* Process 391 detached
--- SIGCHLD (Child exited) @ 0 (0) ---
rt_sigprocmask(SIG_BLOCK, [CHLD], [], 8) = 0
[...]

Quoted 3 lines

strace uses a trick to set the CLONE_PTRACE flag on clone() syscalls made by the traced process. I don't know if the trick used by strace is broken, or the handling of CLONE_PTRACE itself is broken.

While trying to investigate this, I have hit a BUG_ON() that can be triggered by user-space code.

Steps to reproduce:

Compile the C program below. It will call clone() with the CLONE_PTRACE flag set. Run it from bash (_not_ under strace).

====================
#include <sched.h>
#include <stdlib.h>

int e(void *p)

{
	exit(1);
}

char stack[4096*2];

int main()

{
	int r = clone(e, stack+4096, CLONE_PTRACE, 0);
	if (r < 0) {
		perror("clone");
		return 1;
	}
	return 0;
}

====================

When running the program, bash hangs on a wait4() loop. Probably because it is getting notified of the termination of the CLONE_PTRACE child but doesn't know anything about it.

Send SIGTERM to bash. It won't have any effect.

Send SIGKILL to bash. It will trigger the BUG_ON(!child->ptrace) at __ptrace_unlink():

------------[ cut here ]------------
kernel BUG at kernel/ptrace.c:69!
invalid opcode: 0000 [1] SMP
CPU 0
Modules linked in:
Pid: 1784, comm: bash Not tainted 2.6.26-kvm #47
RIP: 0010:[<ffffffff802393d0>]  [<ffffffff802393d0>] __ptrace_unlink+0xa/0x5b
RSP: 0018:ffff88007e9dbc88  EFLAGS: 00010046
RAX: ffff88007f900328 RBX: ffff88007e9dbcc8 RCX: ffffffff8066a5a0
RDX: ffff88007ea982d8 RSI: ffff88007e9dbc58 RDI: ffff88007f900080
RBP: ffff88007e9dbc88 R08: ffffffff80681880 R09: ffffffff806817e0
R10: ffff88007e9dbc58 R11: 0000000000000282 R12: ffff88007ea98040
R13: ffff88007f900080 R14: ffff88007f9ad440 R15: 00000000ffffffff
FS:  0000000000000000(0000) GS:ffffffff806a5a80(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b
CR2: 0000000000483400 CR3: 0000000000201000 CR4: 00000000000006a0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
Process bash (pid: 1784, threadinfo ffff88007e9da000, task ffff88007ea98040)
Stack:  ffff88007e9dbd08 ffffffff80234cad 0000000100000087 ffff88007f9ad438
 ffff88007ea982d8 ffff88007f900328 ffff88007f9ad450 ffff88007ea98030
 ffff88007e9dbcc8 ffff88007e9dbcc8 ffff88007e9dbd18 ffff88007f9ad440
Call Trace:
 [<ffffffff80234cad>] do_exit+0x34c/0x7f3
 [<ffffffff802351d1>] do_group_exit+0x7d/0xaa
 [<ffffffff8023de2c>] get_signal_to_deliver+0x31a/0x342
 [<ffffffff8020b4d5>] ? sysret_signal+0x3d/0x67
 [<ffffffff8020a68e>] do_notify_resume+0x7b/0x89f
 [<ffffffff80209b47>] ? __switch_to+0x1b6/0x3b2
 [<ffffffff80225e69>] ? set_next_entity+0x62/0xb2
 [<ffffffff804ef652>] ? thread_return+0x3d/0xc5
 [<ffffffff8020b4d5>] ? sysret_signal+0x3d/0x67
 [<ffffffff8020b877>] ptregscall_common+0x67/0xb0


Code: 48 89 df e8 2b 2e 00 00 48 8b bb 68 05 00 00 48 81 c7 08 08 00 00 e8 2e 0f fe ff 90 41 5b 5b c9 c3 83 7f 18 00 55 48 89 e5 75 04 <0f> 0b eb fe 48 8b 87 60 02 00 00 48 8b 97 a8 02 00 00 48 8d 8f
RIP  [<ffffffff802393d0>] __ptrace_unlink+0xa/0x5b
 RSP <ffff88007e9dbc88>
---[ end trace 9740fb23e0450ea6 ]---

The problem was reproduced on commit 09a05394, and not reproduced on the commit immediately before it.

Quoted 25 lines

The bisected commit is this:

commit 09a05394fe2448a4139b014936330af23fa7ec83

Author: Roland McGrath <[email protected]>
Date:   Fri Jul 25 19:45:47 2008 -0700

tracehook: clone

This moves all the ptrace initialization and tracing logic for task creation into tracehook.h and ptrace.h inlines. It reorganizes the code slightly, but should not change any behavior.

There are four tracehook entry points, at each important stage of task creation. This keeps the interface from the core fork.c code fairly clean, while supporting the complex setup required for ptrace or something like it.

Signature · 1 line
Eduardo
#2 lore raw

PATCH 1/1 tracehook: fix CLONE_PTRACE

In reply to Eduardo Habkost

The following changes since commit 685d87f7ccc649ab92b55e18e507a65d0e694eb9:
Linus Torvalds (1):
Revert "pcm_native.c: remove unused label"

are available in the git repository at:

git://git.kernel.org/pub/scm/linux/kernel/git/frob/linux-2.6-utrace.git tracehook

Roland McGrath (1):
tracehook: fix CLONE_PTRACE

 include/linux/ptrace.h    |    2 +-
 include/linux/tracehook.h |    2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

Thanks,
Roland

---
[PATCH] tracehook: fix CLONE_PTRACE

In the change in commit 09a05394fe2448a4139b014936330af23fa7ec83, I overlooked two nits in the logic and this broke using CLONE_PTRACE when PTRACE_O_TRACE* are not being used.

A parent that is itself traced at all but not using PTRACE_O_TRACE*, using CLONE_PTRACE would have its new child fail to be traced.

A parent that is not itself traced at all that uses CLONE_PTRACE (which should be a no-op in this case) would confuse the bookkeeping and lead to a crash at exit time.

This restores the missing checks and fixes both failure modes.

---
 include/linux/ptrace.h    |    2 +-
 include/linux/tracehook.h |    2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)
include/linux/ptrace.h +1 -1
index fd31756..ea7416c 100644
@@ -172,7 +172,7 @@ static inline void ptrace_init_task(struct task_struct *child, bool ptrace)
 	child->ptrace = 0;
 	if (unlikely(ptrace)) {
 		child->ptrace = current->ptrace;
-		__ptrace_link(child, current->parent);
+		ptrace_link(child, current->parent);
 	}
 }
 
include/linux/tracehook.h +1 -1
index ab3ef7a..b48d819 100644
@@ -280,7 +280,7 @@ static inline void tracehook_report_clone(int trace, struct pt_regs *regs,
 					  unsigned long clone_flags,
 					  pid_t pid, struct task_struct *child)
 {
-	if (unlikely(trace)) {
+	if (unlikely(trace) || unlikely(clone_flags & CLONE_PTRACE)) {
 		/*
 		 * The child starts up with an immediate SIGSTOP.
 		 */