Skip to contentSkip to conversation

s390/mm: Fix in_atomic() handling in do_secure_storage_access()

PATCH In review Heiko Carstens 4 participants 10 messages Jun 3, 2025 – Jun 5, 2025

Subsystems (3)
lore
Share & tools
lore

Heiko Carstens

author maintainer
#0 lore raw

PATCH 1/1 s390/mm: Fix in_atomic() handling in do_secure_storage_access()

Kernel user spaces accesses to not exported pages in atomic context incorrectly try to resolve the page fault.
With debug options enabled call traces like this can be seen:

BUG: sleeping function called from invalid context at kernel/locking/rwsem.c:1523

in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 419074, name: qemu-system-s39
preempt_count: 1, expected: 0
RCU nest depth: 0, expected: 0
INFO: lockdep is turned off.
Preemption disabled at:

[<00000383ea47cfa2>] copy_page_from_iter_atomic+0xa2/0x8a0
CPU: 12 UID: 0 PID: 419074 Comm: qemu-system-s39
Tainted: G        W           6.16.0-20250531.rc0.git0.69b3a602feac.63.fc42.s390x+debug #1 PREEMPT
Tainted: [W]=WARN
Hardware name: IBM 3931 A01 703 (LPAR)
Call Trace:
 [<00000383e990d282>] dump_stack_lvl+0xa2/0xe8
 [<00000383e99bf152>] __might_resched+0x292/0x2d0
 [<00000383eaa7c374>] down_read+0x34/0x2d0
 [<00000383e99432f8>] do_secure_storage_access+0x108/0x360
 [<00000383eaa724b0>] __do_pgm_check+0x130/0x220
 [<00000383eaa842e4>] pgm_check_handler+0x114/0x160
 [<00000383ea47d028>] copy_page_from_iter_atomic+0x128/0x8a0
([<00000383ea47d016>] copy_page_from_iter_atomic+0x116/0x8a0)
 [<00000383e9c45eae>] generic_perform_write+0x16e/0x310
 [<00000383e9eb87f4>] ext4_buffered_write_iter+0x84/0x160
 [<00000383e9da0de4>] vfs_write+0x1c4/0x460
 [<00000383e9da123c>] ksys_write+0x7c/0x100
 [<00000383eaa7284e>] __do_syscall+0x15e/0x280
 [<00000383eaa8417e>] system_call+0x6e/0x90

INFO: lockdep is turned off.

It is not allowed to take the mmap_lock while in atomic context. Therefore handle such a secure storage access fault as if the accessed page is not
mapped: the uaccess function will return -EFAULT, and the caller has to deal with this. Usually this means that the access is retried in process context, which allows to resolve the page fault (or in this case export the page).

---
 arch/s390/mm/fault.c | 2 ++
 1 file changed, 2 insertions(+)
arch/s390/mm/fault.c +2 -0
index 3829521450dd..e1ad05bfd28a 100644
@@ -441,6 +441,8 @@ void do_secure_storage_access(struct pt_regs *regs)
 		if (rc)
 			BUG();
 	} else {
+		if (faulthandler_disabled())
+			return handle_fault_error_nolock(regs, 0);
 		mm = current->mm;
 		mmap_read_lock(mm);
 		vma = find_vma(mm, addr);
Signature · 1 line
2.45.2
Reviewed-by on v1 1/1
#1 lore raw
Quoted 32 lines · On Tue, 3 Jun 2025 15:49:36 +0200 Heiko Carstens <[email protected]> wrote:

Kernel user spaces accesses to not exported pages in atomic context incorrectly try to resolve the page fault.
With debug options enabled call traces like this can be seen:

BUG: sleeping function called from invalid context at kernel/locking/rwsem.c:1523

in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 419074, name: qemu-system-s39
preempt_count: 1, expected: 0
RCU nest depth: 0, expected: 0
INFO: lockdep is turned off.
Preemption disabled at:

[<00000383ea47cfa2>] copy_page_from_iter_atomic+0xa2/0x8a0
CPU: 12 UID: 0 PID: 419074 Comm: qemu-system-s39
Tainted: G        W           6.16.0-20250531.rc0.git0.69b3a602feac.63.fc42.s390x+debug #1 PREEMPT
Tainted: [W]=WARN
Hardware name: IBM 3931 A01 703 (LPAR)
Call Trace:
 [<00000383e990d282>] dump_stack_lvl+0xa2/0xe8
 [<00000383e99bf152>] __might_resched+0x292/0x2d0
 [<00000383eaa7c374>] down_read+0x34/0x2d0
 [<00000383e99432f8>] do_secure_storage_access+0x108/0x360
 [<00000383eaa724b0>] __do_pgm_check+0x130/0x220
 [<00000383eaa842e4>] pgm_check_handler+0x114/0x160
 [<00000383ea47d028>] copy_page_from_iter_atomic+0x128/0x8a0
([<00000383ea47d016>] copy_page_from_iter_atomic+0x116/0x8a0)
 [<00000383e9c45eae>] generic_perform_write+0x16e/0x310
 [<00000383e9eb87f4>] ext4_buffered_write_iter+0x84/0x160
 [<00000383e9da0de4>] vfs_write+0x1c4/0x460
 [<00000383e9da123c>] ksys_write+0x7c/0x100
 [<00000383eaa7284e>] __do_syscall+0x15e/0x280
 [<00000383eaa8417e>] system_call+0x6e/0x90

INFO: lockdep is turned off.

Quoted 8 lines

It is not allowed to take the mmap_lock while in atomic context. Therefore handle such a secure storage access fault as if the accessed page is not
mapped: the uaccess function will return -EFAULT, and the caller has to deal with this. Usually this means that the access is retried in process context, which allows to resolve the page fault (or in this case export the page).

Quoted 17 lines
---
 arch/s390/mm/fault.c | 2 ++
 1 file changed, 2 insertions(+)
arch/s390/mm/fault.c +2 -0
index 3829521450dd..e1ad05bfd28a 100644
@@ -441,6 +441,8 @@ void do_secure_storage_access(struct pt_regs *regs)
 		if (rc)
 			BUG();
 	} else {
+		if (faulthandler_disabled())
+			return handle_fault_error_nolock(regs, 0);
 		mm = current->mm;
 		mmap_read_lock(mm);
 		vma = find_vma(mm, addr);
#2 lore raw
In reply to Heiko Carstens

arch/s390/mm/fault.c void do_secure_storage_access(struct pt_regs *regs) · line 445

Quoted 10 lines · Heiko Carstens
arch/s390/mm/fault.c +2 -0
index 3829521450dd..e1ad05bfd28a 100644
@@ -441,6 +441,8 @@ void do_secure_storage_access(struct pt_regs *regs)
 		if (rc)
 			BUG();
 	} else {
+		if (faulthandler_disabled())
+			return handle_fault_error_nolock(regs, 0);

This could trigger WARN_ON_ONCE() in handle_fault_error_nolock():

		if (WARN_ON_ONCE(!si_code))
			si_code = SEGV_MAPERR;

Would this warning be justified in this case (aka user_mode(regs) == true)?

Quoted 3 lines
arch/s390/mm/fault.c +0 -0
 		mm = current->mm;
 		mmap_read_lock(mm);
 		vma = find_vma(mm, addr);
#3 lore raw
Quoted 15 lines · On Wed, 4 Jun 2025 18:27:49 +0200 Alexander Gordeev <[email protected]>…

arch/s390/mm/fault.c void do_secure_storage_access(struct pt_regs *regs) · line 446

Quoted 11 lines · Heiko Carstens
arch/s390/mm/fault.c +2 -0
index 3829521450dd..e1ad05bfd28a 100644
@@ -441,6 +441,8 @@ void do_secure_storage_access(struct pt_regs *regs)
 		if (rc)
 			BUG();
 	} else {
+		if (faulthandler_disabled())
+			return handle_fault_error_nolock(regs, 0);
 

This could trigger WARN_ON_ONCE() in handle_fault_error_nolock():

Quoted 5 lines
		if (WARN_ON_ONCE(!si_code))
			si_code = SEGV_MAPERR;

Would this warning be justified in this case (aka user_mode(regs) == true)?

I think so, because if we are in usermode, we should never trigger faulthandler_disabled()

Quoted 4 lines
Quoted 4 lines
arch/s390/mm/fault.c +0 -0
 		mm = current->mm;
 		mmap_read_lock(mm);

vma = find_vma(mm, addr);

#4 lore raw
Quoted 10 lines · Claudio Imbrenda
Quoted 10 lines

arch/s390/mm/fault.c void do_secure_storage_access(struct pt_regs *regs) · line 446

Quoted 7 lines
arch/s390/mm/fault.c +2 -0
@@ -441,6 +441,8 @@ void do_secure_storage_access(struct pt_regs *regs)
 		if (rc)
 			BUG();
 	} else {
+		if (faulthandler_disabled())
+			return handle_fault_error_nolock(regs, 0);
 

This could trigger WARN_ON_ONCE() in handle_fault_error_nolock():

Quoted 8 lines
Quoted 5 lines
		if (WARN_ON_ONCE(!si_code))
			si_code = SEGV_MAPERR;

Would this warning be justified in this case (aka user_mode(regs) == true)?

I think so, because if we are in usermode, we should never trigger faulthandler_disabled()

I think I do not get you. We are in a system call and also in_atomic(), so faulthandler_disabled() is true and handle_fault_error_nolock(regs, 0) is called (above).

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
arch/s390/mm/fault.c +0 -0
 		mm = current->mm;
 		mmap_read_lock(mm);

vma = find_vma(mm, addr);

#5 lore raw
Quoted 17 lines · On Wed, 4 Jun 2025 19:29:31 +0200 Alexander Gordeev <[email protected]>…
Quoted 16 lines · Claudio Imbrenda
Quoted 15 lines

arch/s390/mm/fault.c void do_secure_storage_access(struct pt_regs *regs) · line 446

Quoted 7 lines
arch/s390/mm/fault.c +2 -0
@@ -441,6 +441,8 @@ void do_secure_storage_access(struct pt_regs *regs)
 		if (rc)
 			BUG();
 	} else {
+		if (faulthandler_disabled())
+			return handle_fault_error_nolock(regs, 0);
 

This could trigger WARN_ON_ONCE() in handle_fault_error_nolock():

		if (WARN_ON_ONCE(!si_code))
			si_code = SEGV_MAPERR;

Would this warning be justified in this case (aka user_mode(regs) == true)?

Quoted 6 lines
Quoted 2 lines

I think so, because if we are in usermode, we should never trigger faulthandler_disabled()

I think I do not get you. We are in a system call and also in_atomic(), so faulthandler_disabled() is true and handle_fault_error_nolock(regs, 0) is called (above).

what is the psw in regs?
is it not the one that was being used when the exception was triggered?

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
arch/s390/mm/fault.c +0 -0
 		mm = current->mm;
 		mmap_read_lock(mm);

vma = find_vma(mm, addr);

maintainer Acked-by
#6 lore raw
Quoted 11 lines · Claudio Imbrenda
Quoted 11 lines
Quoted 10 lines
Quoted 7 lines

This could trigger WARN_ON_ONCE() in handle_fault_error_nolock():

		if (WARN_ON_ONCE(!si_code))
			si_code = SEGV_MAPERR;

Would this warning be justified in this case (aka user_mode(regs) == true)?

I think so, because if we are in usermode, we should never trigger faulthandler_disabled()

Quoted 6 lines
Quoted 3 lines

I think I do not get you. We are in a system call and also in_atomic(), so faulthandler_disabled() is true and handle_fault_error_nolock(regs, 0) is called (above).

what is the psw in regs?
is it not the one that was being used when the exception was triggered?

Hmm, right. I assume is_kernel_fault() returns false not because user_mode(regs) is true, but because we access the secondary AS.

Still, to me it feels wrong to trigger that warning due to a user process activity. But anyway:

#7 lore raw

Am 05.06.25 um 11:04 schrieb Alexander Gordeev:

Quoted 19 lines · Claudio Imbrenda
Quoted 11 lines
Quoted 11 lines
Quoted 10 lines
Quoted 7 lines

This could trigger WARN_ON_ONCE() in handle_fault_error_nolock():

		if (WARN_ON_ONCE(!si_code))
			si_code = SEGV_MAPERR;

Would this warning be justified in this case (aka user_mode(regs) == true)?

I think so, because if we are in usermode, we should never trigger faulthandler_disabled()

Quoted 6 lines
Quoted 3 lines

I think I do not get you. We are in a system call and also in_atomic(), so faulthandler_disabled() is true and handle_fault_error_nolock(regs, 0) is called (above).

what is the psw in regs?
is it not the one that was being used when the exception was triggered?

Quoted 7 lines

Hmm, right. I assume is_kernel_fault() returns false not because user_mode(regs) is true, but because we access the secondary AS.

Still, to me it feels wrong to trigger that warning due to a user process activity. But anyway:

Can we trigger a WARN from userspace?

author maintainer
#8 lore raw
Quoted 23 lines · Christian Borntraeger

Am 05.06.25 um 11:04 schrieb Alexander Gordeev:

Quoted 22 lines · Claudio Imbrenda
Quoted 11 lines
Quoted 11 lines
Quoted 10 lines
Quoted 7 lines

This could trigger WARN_ON_ONCE() in handle_fault_error_nolock():

		if (WARN_ON_ONCE(!si_code))
			si_code = SEGV_MAPERR;

Would this warning be justified in this case (aka user_mode(regs) == true)?

I think so, because if we are in usermode, we should never trigger faulthandler_disabled()

Quoted 6 lines
Quoted 3 lines

I think I do not get you. We are in a system call and also in_atomic(), so faulthandler_disabled() is true and handle_fault_error_nolock(regs, 0) is called (above).

what is the psw in regs?
is it not the one that was being used when the exception was triggered?

Hmm, right. I assume is_kernel_fault() returns false not because user_mode(regs) is true, but because we access the secondary AS.

Quoted 6 lines
Quoted 4 lines

Still, to me it feels wrong to trigger that warning due to a user process activity. But anyway:

Can we trigger a WARN from userspace?

No. If the warning triggers, then this indicates a bug in the kernel (exit to user with faulthandler_disabled() == true). I managed to screw up the kernel exactly with such a bug. See commit 588a9836a4ef ("s390/stacktrace: Use break instead of return statement"), which lead to random unexplainable user space crashes.

Note that we have the identical check/code in do_exception().

reviewer Acked-by
#9 lore raw

Am 05.06.25 um 12:07 schrieb Heiko Carstens:

Quoted 29 lines · Christian Borntraeger
Quoted 28 lines

Am 05.06.25 um 11:04 schrieb Alexander Gordeev:

Quoted 26 lines · Claudio Imbrenda
Quoted 11 lines
Quoted 11 lines
Quoted 10 lines
Quoted 7 lines

This could trigger WARN_ON_ONCE() in handle_fault_error_nolock():

		if (WARN_ON_ONCE(!si_code))
			si_code = SEGV_MAPERR;

Would this warning be justified in this case (aka user_mode(regs) == true)?

I think so, because if we are in usermode, we should never trigger faulthandler_disabled()

Quoted 6 lines
Quoted 3 lines

I think I do not get you. We are in a system call and also in_atomic(), so faulthandler_disabled() is true and handle_fault_error_nolock(regs, 0) is called (above).

what is the psw in regs?
is it not the one that was being used when the exception was triggered?

Hmm, right. I assume is_kernel_fault() returns false not because user_mode(regs) is true, but because we access the secondary AS.

Still, to me it feels wrong to trigger that warning due to a user process activity. But anyway:

Quoted 7 lines
Quoted 1 line

Can we trigger a WARN from userspace?

No. If the warning triggers, then this indicates a bug in the kernel (exit to user with faulthandler_disabled() == true). I managed to screw up the kernel exactly with such a bug. See commit 588a9836a4ef ("s390/stacktrace: Use break instead of return statement"), which lead to random unexplainable user space crashes.

Ok, then this makes a lot of sense to WARN.