Skip to contentSkip to conversation

s390: fix double free of GS and RI CBs on fork() failure

PATCH Applied Brian Foster 4 participants 5 messages Aug 16, 2022 – Aug 24, 2022 s390

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 s390: fix double free of GS and RI CBs on fork() failure

The pointers for guarded storage and runtime instrumentation control blocks are stored in the thread_struct of the associated task. These pointers are initially copied on fork() via arch_dup_task_struct() and then cleared via copy_thread() before fork() returns. If fork() happens to fail after the initial task dup and before copy_thread(), the newly allocated task and associated thread_struct memory are freed via free_task() -> arch_release_task_struct(). This results in a double free of the guarded storage and runtime info structs because the fields in the failed task still refer to memory associated with the source task.

This problem can manifest as a BUG_ON() in set_freepointer() (with CONFIG_SLAB_FREELIST_HARDENED enabled) or KASAN splat (if enabled) when running trinity syscall fuzz tests on s390x. To avoid this problem, clear the associated pointer fields in arch_dup_task_struct() immediately after the new task is copied. Note that the RI flag is still cleared in copy_thread() because it resides in thread stack memory and that is where stack info is copied.

---

Hi all,

Note that I'm not subscribed to the list so please CC on reply. Further, I'm not terribly familiar with these associated features and so have not run any kind of functional testing here. My testing was purely around producing/preventing the double free issue. Any thoughts, reviews or further testing is appreciated. Thanks.

Brian

 arch/s390/kernel/process.c | 22 ++++++++++++++++------
 1 file changed, 16 insertions(+), 6 deletions(-)
arch/s390/kernel/process.c +16 -6
index 89949b9f3cf8..d5119e039d85 100644
@@ -91,6 +91,18 @@ int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
 
 	memcpy(dst, src, arch_task_struct_size);
 	dst->thread.fpu.regs = dst->thread.fpu.fprs;
+
+	/*
+	 * Don't transfer over the runtime instrumentation or the guarded
+	 * storage control block pointers. These fields are cleared here instead
+	 * of in copy_thread() to avoid premature freeing of associated memory
+	 * on fork() failure. Wait to clear the RI flag because ->stack still
+	 * refers to the source thread.
+	 */
+	dst->thread.ri_cb = NULL;
+	dst->thread.gs_cb = NULL;
+	dst->thread.gs_bc_cb = NULL;
+
 	return 0;
 }
 
@@ -150,13 +162,11 @@ int copy_thread(struct task_struct *p, const struct kernel_clone_args *args)
 	frame->childregs.flags = 0;
 	if (new_stackp)
 		frame->childregs.gprs[15] = new_stackp;
-
-	/* Don't copy runtime instrumentation info */
-	p->thread.ri_cb = NULL;
+	/*
+	 * Clear the runtime instrumentation flag after the above childregs
+	 * copy. The CB pointer was already cleared in arch_dup_task_struct().
+	 */
 	frame->childregs.psw.mask &= ~PSW_MASK_RI;
-	/* Don't copy guarded storage control block */
-	p->thread.gs_cb = NULL;
-	p->thread.gs_bc_cb = NULL;
 
 	/* Set a new TLS ?  */
 	if (clone_flags & CLONE_SETTLS) {
Signature · 1 line
2.37.1
Reviewed-by on v1 1/1
#1 lore raw
Quoted 67 lines · On Tue, 16 Aug 2022 11:54:07 -0400 Brian Foster <[email protected]> wrote:

The pointers for guarded storage and runtime instrumentation control blocks are stored in the thread_struct of the associated task. These pointers are initially copied on fork() via arch_dup_task_struct() and then cleared via copy_thread() before fork() returns. If fork() happens to fail after the initial task dup and before copy_thread(), the newly allocated task and associated thread_struct memory are freed via free_task() -> arch_release_task_struct(). This results in a double free of the guarded storage and runtime info structs because the fields in the failed task still refer to memory associated with the source task.

This problem can manifest as a BUG_ON() in set_freepointer() (with CONFIG_SLAB_FREELIST_HARDENED enabled) or KASAN splat (if enabled) when running trinity syscall fuzz tests on s390x. To avoid this problem, clear the associated pointer fields in arch_dup_task_struct() immediately after the new task is copied. Note that the RI flag is still cleared in copy_thread() because it resides in thread stack memory and that is where stack info is copied.

---

Hi all,

Note that I'm not subscribed to the list so please CC on reply. Further, I'm not terribly familiar with these associated features and so have not run any kind of functional testing here. My testing was purely around producing/preventing the double free issue. Any thoughts, reviews or further testing is appreciated. Thanks.

Brian

 arch/s390/kernel/process.c | 22 ++++++++++++++++------
 1 file changed, 16 insertions(+), 6 deletions(-)
arch/s390/kernel/process.c +13 -3
index 89949b9f3cf8..d5119e039d85 100644
@@ -91,6 +91,18 @@ int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
 
 	memcpy(dst, src, arch_task_struct_size);
 	dst->thread.fpu.regs = dst->thread.fpu.fprs;
+
+	/*
+	 * Don't transfer over the runtime instrumentation or the guarded
+	 * storage control block pointers. These fields are cleared here instead
+	 * of in copy_thread() to avoid premature freeing of associated memory
+	 * on fork() failure. Wait to clear the RI flag because ->stack still
+	 * refers to the source thread.
+	 */
+	dst->thread.ri_cb = NULL;
+	dst->thread.gs_cb = NULL;
+	dst->thread.gs_bc_cb = NULL;
+
 	return 0;
 }
 
@@ -150,13 +162,11 @@ int copy_thread(struct task_struct *p, const struct kernel_clone_args *args)
 	frame->childregs.flags = 0;
 	if (new_stackp)
 		frame->childregs.gprs[15] = new_stackp;
-
-	/* Don't copy runtime instrumentation info */
-	p->thread.ri_cb = NULL;
+	/*

arch/s390/kernel/process.c int copy_thread(struct task_struct *p, const struct kernel_clone_args *args) · line 172

Quoted 10 lines
arch/s390/kernel/process.c +3 -3
+	 * Clear the runtime instrumentation flag after the above childregs
+	 * copy. The CB pointer was already cleared in arch_dup_task_struct().
+	 */
 	frame->childregs.psw.mask &= ~PSW_MASK_RI;
-	/* Don't copy guarded storage control block */
-	p->thread.gs_cb = NULL;
-	p->thread.gs_bc_cb = NULL;
 
 	/* Set a new TLS ?  */
 	if (clone_flags & CLONE_SETTLS) {

Thanks Brian, nice catch! Looks good to me. For completeness, we should add stable / Fixes tags, like this:

Fixes 8d9047f8b967c ("s390/runtime instrumentation: simplify task exit handling")
Fixes 7b83c6297d2fc ("s390/guarded storage: simplify task exit handling")
Cc <[email protected]> # 4.15

Not 100% sure about the Fixes tags, Heiko should also have a look when he returns next week.

#2 lore raw
Quoted 84 lines · Gerald Schaefer
Quoted 67 lines · On Tue, 16 Aug 2022 11:54:07 -0400 Brian Foster <[email protected]> wrote:

The pointers for guarded storage and runtime instrumentation control blocks are stored in the thread_struct of the associated task. These pointers are initially copied on fork() via arch_dup_task_struct() and then cleared via copy_thread() before fork() returns. If fork() happens to fail after the initial task dup and before copy_thread(), the newly allocated task and associated thread_struct memory are freed via free_task() -> arch_release_task_struct(). This results in a double free of the guarded storage and runtime info structs because the fields in the failed task still refer to memory associated with the source task.

This problem can manifest as a BUG_ON() in set_freepointer() (with CONFIG_SLAB_FREELIST_HARDENED enabled) or KASAN splat (if enabled) when running trinity syscall fuzz tests on s390x. To avoid this problem, clear the associated pointer fields in arch_dup_task_struct() immediately after the new task is copied. Note that the RI flag is still cleared in copy_thread() because it resides in thread stack memory and that is where stack info is copied.

---

Hi all,

Note that I'm not subscribed to the list so please CC on reply. Further, I'm not terribly familiar with these associated features and so have not run any kind of functional testing here. My testing was purely around producing/preventing the double free issue. Any thoughts, reviews or further testing is appreciated. Thanks.

Brian

 arch/s390/kernel/process.c | 22 ++++++++++++++++------
 1 file changed, 16 insertions(+), 6 deletions(-)
arch/s390/kernel/process.c +13 -3
index 89949b9f3cf8..d5119e039d85 100644
@@ -91,6 +91,18 @@ int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
 
 	memcpy(dst, src, arch_task_struct_size);
 	dst->thread.fpu.regs = dst->thread.fpu.fprs;
+
+	/*
+	 * Don't transfer over the runtime instrumentation or the guarded
+	 * storage control block pointers. These fields are cleared here instead
+	 * of in copy_thread() to avoid premature freeing of associated memory
+	 * on fork() failure. Wait to clear the RI flag because ->stack still
+	 * refers to the source thread.
+	 */
+	dst->thread.ri_cb = NULL;
+	dst->thread.gs_cb = NULL;
+	dst->thread.gs_bc_cb = NULL;
+
 	return 0;
 }
 
@@ -150,13 +162,11 @@ int copy_thread(struct task_struct *p, const struct kernel_clone_args *args)
 	frame->childregs.flags = 0;
 	if (new_stackp)
 		frame->childregs.gprs[15] = new_stackp;
-
-	/* Don't copy runtime instrumentation info */
-	p->thread.ri_cb = NULL;
+	/*

arch/s390/kernel/process.c int copy_thread(struct task_struct *p, const struct kernel_clone_args *args) · line 172

Quoted 10 lines
arch/s390/kernel/process.c +3 -3
+	 * Clear the runtime instrumentation flag after the above childregs
+	 * copy. The CB pointer was already cleared in arch_dup_task_struct().
+	 */
 	frame->childregs.psw.mask &= ~PSW_MASK_RI;
-	/* Don't copy guarded storage control block */
-	p->thread.gs_cb = NULL;
-	p->thread.gs_bc_cb = NULL;
 
 	/* Set a new TLS ?  */
 	if (clone_flags & CLONE_SETTLS) {

Thanks Brian, nice catch! Looks good to me. For completeness, we should add stable / Fixes tags, like this:

Quoted 9 lines
Fixes 8d9047f8b967c ("s390/runtime instrumentation: simplify task exit handling")
Fixes 7b83c6297d2fc ("s390/guarded storage: simplify task exit handling")
Cc <[email protected]> # 4.15

Not 100% sure about the Fixes tags, Heiko should also have a look when he returns next week.

Hi Gerald,

Sounds good, thanks for the review. I'll wait a bit so Heiko has a chance to take a look and then follow up with the tag updates and any other necessary changes..

Brian

maintainer Reviewed-by on v1 1/1
#3 lore raw
In reply to Brian Foster
Quoted 20 lines · Brian Foster

The pointers for guarded storage and runtime instrumentation control blocks are stored in the thread_struct of the associated task. These pointers are initially copied on fork() via arch_dup_task_struct() and then cleared via copy_thread() before fork() returns. If fork() happens to fail after the initial task dup and before copy_thread(), the newly allocated task and associated thread_struct memory are freed via free_task() -> arch_release_task_struct(). This results in a double free of the guarded storage and runtime info structs because the fields in the failed task still refer to memory associated with the source task.

This problem can manifest as a BUG_ON() in set_freepointer() (with CONFIG_SLAB_FREELIST_HARDENED enabled) or KASAN splat (if enabled) when running trinity syscall fuzz tests on s390x. To avoid this problem, clear the associated pointer fields in arch_dup_task_struct() immediately after the new task is copied. Note that the RI flag is still cleared in copy_thread() because it resides in thread stack memory and that is where stack info is copied.

Quoted 4 lines
---
 arch/s390/kernel/process.c | 22 ++++++++++++++++------
 1 file changed, 16 insertions(+), 6 deletions(-)

Thanks for debugging and fixing this!

Vasily, could you pick this one up, and also add the Fixes: and Cc:
stable tags as proposed by Gerald?
Thanks!

maintainer Applied
#4 lore raw
Quoted 26 lines · Heiko Carstens
Quoted 20 lines · Brian Foster

The pointers for guarded storage and runtime instrumentation control blocks are stored in the thread_struct of the associated task. These pointers are initially copied on fork() via arch_dup_task_struct() and then cleared via copy_thread() before fork() returns. If fork() happens to fail after the initial task dup and before copy_thread(), the newly allocated task and associated thread_struct memory are freed via free_task() -> arch_release_task_struct(). This results in a double free of the guarded storage and runtime info structs because the fields in the failed task still refer to memory associated with the source task.

This problem can manifest as a BUG_ON() in set_freepointer() (with CONFIG_SLAB_FREELIST_HARDENED enabled) or KASAN splat (if enabled) when running trinity syscall fuzz tests on s390x. To avoid this problem, clear the associated pointer fields in arch_dup_task_struct() immediately after the new task is copied. Note that the RI flag is still cleared in copy_thread() because it resides in thread stack memory and that is where stack info is copied.

Quoted 4 lines
---
 arch/s390/kernel/process.c | 22 ++++++++++++++++------
 1 file changed, 16 insertions(+), 6 deletions(-)
Quoted 7 lines

Thanks for debugging and fixing this!

Vasily, could you pick this one up, and also add the Fixes: and Cc:
stable tags as proposed by Gerald?
Thanks!

Applied, Thank you!