Re: [PATCH v4] irqchip: gic-v3-its: irq_pipeline: Fix OOB context in-band lock warning for ITS lock

[Philippe Gerum <rpm@xenomai.org>]

Florian Bezdeka <florian.bezdeka@siemens.com> writes:

> On Tue, 2026-08-04 at 17:30 +0800, linz wrote:
>> At 2026-08-04 16:59:00, "Florian Bezdeka" <florian.bezdeka@siemens.com> wrote:
>> > On Tue, 2026-08-04 at 16:53 +0800, linz wrote:
>> > > The warning shown below could be observed on Dovetail 6.6 with arm64
>> > > architecture when CONFIG_DEBUG_IRQ_PIPELINE is enabled.
>> > > 
>> > > The reason is the usage of a raw_spinlock_t in hard IRQ masking code
>> > > path. A migration to hard_spinlock_t fixes this issue.
>> > > 
>> > > [    1.510903] IRQ pipeline: some code running in oob context 'Xenomai'
>> > >                              called an in-band only routine
>> > > [    1.510912] CPU: 0 PID: 0 Comm: swapper/0 Tainted: G S                 6.6.63-dovetail2 #8
>> > > [    1.510918] Hardware name: Pe2204 DEMO DDR4 (DT)
>> > > [    1.510920] IRQ stage: Xenomai
>> > > [    1.510923] Call trace:
>> > > [    1.510926]  dump_backtrace+0x90/0xe4
>> > > [    1.510940]  show_stack+0x14/0x1c
>> > > [    1.510946]  dump_stack_lvl+0x84/0xc8
>> > > [    1.510953]  dump_stack+0x14/0x1c
>> > > [    1.510957]  check_inband_stage+0xb0/0xc8
>> > > [    1.510965]  inband_irq_save+0xc/0x28
>> > > [    1.510971]  _raw_spin_lock_irqsave+0x14/0x90
>> > > [    1.510977]  its_send_single_command+0x24/0x154
>> > > [    1.510984]  lpi_update_config+0x9c/0x144
>> > > [    1.510990]  its_mask_irq+0x2c/0x64
>> > > [    1.510997]  irq_chip_mask_parent+0x18/0x20
>> > > [    1.511005]  its_mask_msi_irq+0x1c/0x28
>> > > [    1.511012]  handle_fasteoi_irq+0x1cc/0x2b4
>> > > [    1.511016]  generic_pipeline_irq_desc+0x6c/0xa4
>> > > [    1.511021]  generic_handle_domain_irq+0x18/0x20
>> > > [    1.511028]  gic_handle_irq+0x4c/0x120
>> > > [    1.511032]  handle_irq_pipelined+0x40/0x64
>> > > [    1.511038]  call_on_irq_stack+0x24/0x30
>> > > [    1.511044]  do_interrupt_handler+0x138/0x158
>> > > [    1.511050]  el1_interrupt+0x40/0x110
>> > > [    1.511055]  el1h_64_irq_handler+0x14/0x1c
>> > > [    1.511061]  el1h_64_irq+0x64/0x68
>> > > [    1.511064]  default_idle_call+0x30/0x78
>> > > [    1.511071]  do_idle+0x128/0x150
>> > > [    1.511077]  cpu_startup_entry+0x34/0x38
>> > > [    1.511081]  kernel_init+0x0/0x1d4
>> > > [    1.511088]  arch_post_acpi_subsys_init+0x0/0x8
>> > > [    1.511096]  start_kernel+0x504/0x5cc
>> > > [    1.511103]  __primary_switched+0xbc/0xc4
>> > > 
>> > > Beyond that, note that IRQ chip handlers such as irq_mask / irq_unmask run in
>> > > OOB context. For instance, any code path calling lpi_update_config() must
>> > > be oob‑capable, so gic_data_rdist_cpu(cpu)->rd_lock and its_vpe->vpe_lock are
>> > > modified in this patch.
>> > > 
>> > > Convert:
>> > > - its_node->lock
>> > > - rdists.rd_lock
>> > > - its_vpe->vpe_lock
>> > > from raw_spinlock_t to hard_spinlock_t.
>> > > 
>> > > Signed-off-by: linz <powertree@163.com>
>> > > ---
>> > >  drivers/irqchip/irq-gic-v3-its.c   | 2 +-
>> > >  include/linux/irqchip/arm-gic-v3.h | 2 +-
>> > >  include/linux/irqchip/arm-gic-v4.h | 2 +-
>> > >  3 files changed, 3 insertions(+), 3 deletions(-)
>> > > 
>> > > diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c
>> > > index 0e57735eac..3f2b02b8f4 100644
>> > > --- a/drivers/irqchip/irq-gic-v3-its.c
>> > > +++ b/drivers/irqchip/irq-gic-v3-its.c
>> > > @@ -94,7 +94,7 @@ struct its_device;
>> > >   * list.
>> > >   */
>> > >  struct its_node {
>> > > -    raw_spinlock_t        lock;
>> > > +    hard_spinlock_t        lock;
>> > >      struct mutex        dev_alloc_lock;
>> > >      struct list_head    entry;
>> > >      void __iomem        *base;
>> > > diff --git a/include/linux/irqchip/arm-gic-v3.h b/include/linux/irqchip/arm-gic-v3.h
>> > > index 7286913654..9dcbf4b331 100644
>> > > --- a/include/linux/irqchip/arm-gic-v3.h
>> > > +++ b/include/linux/irqchip/arm-gic-v3.h
>> > > @@ -613,7 +613,7 @@
>> > >  
>> > >  struct rdists {
>> > >      struct {
>> > > -        raw_spinlock_t    rd_lock;
>> > > +        hard_spinlock_t    rd_lock;
>> > >          void __iomem    *rd_base;
>> > >          struct page    *pend_page;
>> > >          phys_addr_t    phys_base;
>> > > diff --git a/include/linux/irqchip/arm-gic-v4.h b/include/linux/irqchip/arm-gic-v4.h
>> > > index bf9e064028..d7afd9d2cd 100644
>> > > --- a/include/linux/irqchip/arm-gic-v4.h
>> > > +++ b/include/linux/irqchip/arm-gic-v4.h
>> > > @@ -68,7 +68,7 @@ struct its_vpe {
>> > >       * Ensures mutual exclusion between affinity setting of the
>> > >       * vPE and vLPI operations using vpe->col_idx.
>> > >       */
>> > > -    raw_spinlock_t        vpe_lock;
>> > > +    hard_spinlock_t        vpe_lock;
>> > 
>> > Can you please share the call stack where this lock is involved in an
>> > OOB code path?
>> > 
>> > The other two locks are fine. I'm unsure about this one here.
>> 
>> > 
>> The vpe_lock can be taken in OOB context via the following call chain when
>> handling vLPI interrupt masking:
>> 
>> its_mask_irq()
>>     lpi_update_config()
>>         direct_lpi_inv()
>
> This one here is protected by is_v4_1(), which seems gic-v4 specific.
> The location (arm-gic-v4.h) is also pointing in that direction. Most
> likely the reason why you never saw any warning for that lock. You are
> on v3, right?
>
> The change itself is correct but v4 specific and merged into a patch
> titled v3. That's confusing.
>
> Proposal: Skip that lock for now and revisit it once the v4 irq chip is
> made OOB aware. (Nice additional contribution possibility ;-))
>
> Philippe, any comments from your side?
>

gic-v3 and v4 implementations share some type definitions via
arm-gic-v4.h. Since we need to fix up anything down the
lpi_update_config() path, it looks like the impact of enabling v3 for
oob context is going to spread to v4.

Looking at the its_vm definition in arm-gic-v4.h, there is a helpful
comment stating that the following lock order applies: vmapp_lock ->
vpe_lock ->vmovp_lock.  For this reason, if we are going to convert
vpe_lock to a hard lock, then vmovp_lock must be converted in the same
move too due to nesting constraints. Which means in turn that any code
path going through its_send_vmovp() needs to be inspected.

-- 
Philippe.