[PATCH v4 1/2] ARM: mm: fault: Enable interrupts before invoking __do_user_fault()

[Russell King (Oracle)]

On Fri, Jan 16, 2026 at 08:01:20PM +0000, Russell King (Oracle) wrote:
> On Fri, Jan 16, 2026 at 07:12:01PM +0100, Sebastian Andrzej Siewior wrote:
> > On 2026-01-16 17:33:48 [+0000], Russell King (Oracle) wrote:
> > > On Fri, Jan 16, 2026 at 06:00:40PM +0100, Sebastian Andrzej Siewior wrote:
> > > > From: "Yadi.hu" <yadi.hu at windriver.com>
> > > > 
> > > > A page fault from userland for a kernel address originates from from
> > > > do_sect_fault() (!LPAE) or do_page_fault() and ends in __do_user_fault()
> > > > by sending a signal.
> > > > 
> > > > Sending a signal requires to acquire sighand_struct::siglock which is a
> > > > spinlock_t. On PREEMPT_RT spinlock_t becomes a sleeping spin lock which
> > > > requires interrupts to be enabled. Since the calling context is user
> > > > land, interrupts must have been enabled so it is fine to enable them in
> > > > this case.
> > > > 
> > > > Enable interrupts in do_kernel_address_page_fault() unconditional in the
> > > > user_mode case().
> > > > Enable interrupts in do_sect_fault() if they were previously enabled.
> > > 
> > > Do you need any of this? __do_user_fault() now calls
> > > local_irq_enable() as almost the first thing it does.
> > 
> > Different path(s):
> > 
> > LPAE
> > | BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:48
> > | in_atomic(): 0, irqs_disabled(): 128, non_block: 0, pid: 639, name: arm-segfault
> > | preempt_count: 0, expected: 0
> > | RCU nest depth: 0, expected: 0
> > | CPU: 0 UID: 0 PID: 639 Comm: arm-segfault3 Tainted: G        W           6.19.0-rc5-dirty #7 PREEMPT_RT
> > | Tainted: [W]=WARN
> > | Hardware name: QEMU QEMU Virtual Machine, BIOS unknown 02/02/2022
> > | Call trace:
> > |  unwind_backtrace from show_stack+0x18/0x1c
> > |  show_stack from dump_stack_lvl+0x34/0x44
> > |  dump_stack_lvl from __might_resched+0x180/0x1c0
> > |  __might_resched from rt_spin_lock+0x3c/0x1f0
> > |  rt_spin_lock from force_sig_info_to_task+0x24/0x184
> > |  force_sig_info_to_task from force_sig_fault+0x50/0x74
> > |  force_sig_fault from do_kernel_address_page_fault+0xa8/0xb4
> > |  do_kernel_address_page_fault from do_DataAbort+0x38/0xac
> > |  do_DataAbort from __dabt_usr+0x50/0x60
> > 
> > !LPAE
> > | BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:48
> > | in_atomic(): 0, irqs_disabled(): 128, non_block: 0, pid: 622, name: arm-segfault
> > | preempt_count: 0, expected: 0
> > | RCU nest depth: 0, expected: 0
> > | CPU: 0 UID: 0 PID: 622 Comm: arm-segfault Tainted: G        W           6.19.0-rc5-dirty #8 PREEMPT_RT
> > | Tainted: [W]=WARN
> > | Hardware name: QEMU QEMU Virtual Machine, BIOS unknown 02/02/2022
> > | Call trace:
> > |  unwind_backtrace from show_stack+0x18/0x1c
> > |  show_stack from dump_stack_lvl+0x34/0x44
> > |  dump_stack_lvl from __might_resched+0x180/0x1c0
> > |  __might_resched from rt_spin_lock+0x3c/0x1f0
> > |  rt_spin_lock from force_sig_info_to_task+0x24/0x184
> > |  force_sig_info_to_task from force_sig_fault+0x50/0x74
> > |  force_sig_fault from do_sect_fault+0x30/0x80
> > |  do_sect_fault from do_DataAbort+0x44/0xb8
> > |  do_DataAbort from __dabt_usr+0x50/0x60
> 
> Right, because I haven't pushed all the patches out. What was pushed out
> was the basic set of fixes during the last merge window:
> 
> fd2dee1c6e22 ARM: fix branch predictor hardening (fixes)
> 7733bc7d299d ARM: fix hash_name() fault
> 40b466db1dff ARM: allow __do_kernel_fault() to report execution of memory faults
> dea20281ac88 ARM: group is_permission_fault() with is_translation_fault()
> edb924a7211c ARM: 9464/1: fix input-only operand modification in load_unaligned_zeropad()
> 
> which were sent out on the 8th December, but I still have these:
> 
> 1055c9d604f9 ARM: provide individual is_translation_fault() and is_permission_fault()
> 47ce12a37f82 ARM: move FSR fault status definitions before fsr_fs()
> eefbf2e97ad5 ARM: use BIT() and GENMASK() for fault status register fields
> 5f1e55bed37f ARM: move is_permission_fault() and is_translation_fault() to fault.h
> ad1c1212ae15 ARM: move vmalloc() lazy-page table population
> 92442d814d99 ARM: ensure interrupts are enabled in __do_user_fault()
> 
> and, because they weren't fixes, it would've been inappropriate to
> post these. However, I was on vacation from the 11th December through
> to the 2nd January, had the madness of the huge email mountain, and
> been working on stmmac and internal work issues since, I've not had any
> time to look at anything else.
> 
> However, these changes will negate the need for your patch 1.

Oh, and, as I've said on netdev recently, I have lots of patches. I
don't have time to push all patches all the time. Those that I think
are the highest priority get the attention - so if something goes
quiet, the patches don't get progressed (because something else has
taken over.)

I haven't published these yet to my external git tree because they're
currently on top of my raw private "development" branch containing...

$ git lg origin..rmk | wc -l
473

many of those behind those I mentioned above. At some point I need to
move them to their own separate branch, or to the "misc" branch after
sending them to the mailing list... but the latter means spending time
writing a covering message summarising the changes, and the key thing
is "time".

I've wasted almost all of today running a bisect for a stmmac issue that
takes out my _entire_ network at home. It's wasted because what I
thought was a good commit turns out, on re-testing, to have been bad,
and right now I have no idea what a good commit is, and whether that
even exists before the platform support was merged.

-- 
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTP is here! 80Mbps down 10Mbps up. Decent connectivity at last!