On Thu, 25 Aug 2022 16:23:48 +0100,
Pierre Gondois <pierre.gondois@xxxxxxx> wrote:
>
> In KvmTool, on a ThunderX2, when running a preemp_rt kernel based on
How is this specific to kvmtool? Or to running as a guest?
> v5.19-rc3-rt4, the following happens:
> [ 4.070739] [ BUG: Invalid wait context ]
> [ 4.070740] 5.19.0-rc3-rt4-00001-g1a6597c0bdcf #153 Not tainted
> [ 4.070742] -----------------------------
> [ 4.070743] swapper/0/1 is trying to lock:
> [ 4.070744] ffff0000ab7405d8 ((&c->lock)){+.+.}-{3:3}, at: ___slab_alloc (mm/slub.c:2954)
> [ 4.070757] other info that might help us debug this:
> [ 4.070758] context-{5:5}
> [ 4.070759] 5 locks held by swapper/0/1:
> [ 4.070760] #0: ffff0000811491e0 (&dev->mutex){....}-{4:4}, at: __device_driver_lock (drivers/base/dd.c:1055)
> [ 4.070769] #1: ffff0000846c5670 (&desc->request_mutex){+.+.}-{4:4}, at: __setup_irq (kernel/irq/internals.h:147)
> [ 4.070778] #2: ffff0000846c54c8 (&irq_desc_lock_class){....}-{2:2}, at: __setup_irq (kernel/irq/manage.c:1612)
> [ 4.070784] #3: ffff80000b23ea78 (mask_lock){....}-{2:2}, at: irq_setup_affinity (./include/linux/irq.h:381)
> [ 4.070791] #4: ffff80000b23ea38 (tmp_mask_lock){....}-{2:2}, at: irq_do_set_affinity (./include/linux/irq.h:381)
> [ 4.070797] stack backtrace:
> [ 4.070801] CPU: 2 PID: 1 Comm: swapper/0 Not tainted 5.19.0-rc3-rt4-00001-g1a6597c0bdcf #153
> [ 4.070805] Call trace:
> [ 4.070806] dump_backtrace (arch/arm64/kernel/stacktrace.c:200)
> [ 4.070811] show_stack (arch/arm64/kernel/stacktrace.c:207)
> [ 4.070813] dump_stack_lvl (lib/dump_stack.c:107)
> [ 4.070818] dump_stack (lib/dump_stack.c:114)
> [ 4.070820] __lock_acquire (kernel/locking/lockdep.c:4707)
> [ 4.070823] lock_acquire (kernel/locking/lockdep.c:466)
> [ 4.070825] rt_spin_lock (./arch/arm64/include/asm/current.h:19 (discriminator 4))
> [ 4.070830] ___slab_alloc (mm/slub.c:2954)
> [ 4.070832] __slab_alloc.isra.0 (mm/slub.c:3116)
> [ 4.070835] __kmalloc_node (mm/slub.c:3207)
> [ 4.070837] alloc_cpumask_var_node (lib/cpumask.c:115)
> [ 4.070843] alloc_cpumask_var (lib/cpumask.c:147)
> [ 4.070846] its_select_cpu (drivers/irqchip/irq-gic-v3-its.c:1580)
> [ 4.070850] its_set_affinity (drivers/irqchip/irq-gic-v3-its.c:1659)
> [ 4.070853] msi_domain_set_affinity (kernel/irq/msi.c:501)
> [ 4.070857] irq_do_set_affinity (kernel/irq/manage.c:276)
> [ 4.070860] irq_setup_affinity (kernel/irq/manage.c:633)
> [ 4.070863] irq_startup (kernel/irq/chip.c:280)
> [ 4.070865] __setup_irq (kernel/irq/manage.c:1777)
> [ 4.070869] request_threaded_irq (kernel/irq/manage.c:2206)
> [ 4.070872] vp_find_vqs_msix (./include/linux/interrupt.h:168)
> [ 4.070876] vp_find_vqs (drivers/virtio/virtio_pci_common.c:400)
> [ 4.070878] vp_modern_find_vqs (drivers/virtio/virtio_pci_modern.c:259)
> [ 4.070880] init_vq (./include/linux/virtio_config.h:213)
> [ 4.070885] virtblk_probe (drivers/block/virtio_blk.c:936)
> [ 4.070887] virtio_dev_probe (drivers/virtio/virtio.c:303)
> [ 4.070892] really_probe (drivers/base/dd.c:555)
> [ 4.070895] __driver_probe_device (drivers/base/dd.c:764)
> [ 4.070897] driver_probe_device (drivers/base/dd.c:794)
> [ 4.070899] __driver_attach (drivers/base/dd.c:1164)
> [ 4.070901] bus_for_each_dev (drivers/base/bus.c:301)
> [ 4.070904] driver_attach (drivers/base/dd.c:1181)
> [ 4.070906] bus_add_driver (drivers/base/bus.c:618)
> [ 4.070908] driver_register (drivers/base/driver.c:240)
> [ 4.070910] register_virtio_driver (drivers/virtio/virtio.c:356 (discriminator 4))
> [ 4.070913] virtio_blk_init (drivers/block/virtio_blk.c:1218)
> [ 4.070918] do_one_initcall (init/main.c:1295)
> [ 4.070921] kernel_init_freeable (init/main.c:1367)
> [ 4.070924] kernel_init (init/main.c:1503)
> [ 4.070927] ret_from_fork (arch/arm64/kernel/entry.S:868)
Consider trimming the trace to the useful part.
>
> commit cba4235e6031e ("genirq: Remove mask argument from
> setup_affinity()")
> and
> commit 11ea68f553e24 ("genirq, sched/isolation: Isolate from handling
> managed interrupts")
> overcome this issue by defining a static struct cpumask and protecting
> it by a raw spinlock. The code in these commits is executed with IRQs
> disabled.
> its_select_cpu() can be executed with IRQs enabled or disabled. Thus
> disabling IRQs is necesserary to avoid deadlocking.
Only if this code can be preempted by an interrupt and subsequently
called from interrupt context. Can you describe this code path?
>
> This patch:
Please refer to the documentation and the section about avoiding
things like "this patch" in a commit message.
> - makes tmpmask a 'static struct cpumask'. This prevents storing it on
> the stack and having to dynamically allocate it
> - protects tmpmask with a raw spinlock
> - disables IRQs around the spinlock for the case its_select_cpu() is
> called with IRQs enabled
>
> Signed-off-by: Pierre Gondois <pierre.gondois@xxxxxxx>
> ---
> drivers/irqchip/irq-gic-v3-its.c | 34 +++++++++++++++++---------------
> 1 file changed, 18 insertions(+), 16 deletions(-)
>
> diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c
> index 5ff09de6c48f..3cf89e59b036 100644
> --- a/drivers/irqchip/irq-gic-v3-its.c
> +++ b/drivers/irqchip/irq-gic-v3-its.c
> @@ -1574,14 +1574,15 @@ static int its_select_cpu(struct irq_data *d,
> const struct cpumask *aff_mask)
> {
> struct its_device *its_dev = irq_data_get_irq_chip_data(d);
> - cpumask_var_t tmpmask;
> + static DEFINE_RAW_SPINLOCK(tmpmask_lock);
> + static struct cpumask tmpmask;
> + unsigned long flags;
> int cpu, node;
> -
> - if (!alloc_cpumask_var(&tmpmask, GFP_ATOMIC))
> - return -ENOMEM;
> -
> node = its_dev->its->numa_node;
>
> + local_irq_save(flags);
> + raw_spin_lock(&tmpmask_lock);
Why is this done in two steps? What is wrong with raw_spin_lock_irqsave()?
More importantly, a number of call paths reaching its_set_affinity()
already execute under a raw spinlock, with interrupts disabled. Is it
worth not hitting this lock at all times?
> +
> if (!irqd_affinity_is_managed(d)) {
> /* First try the NUMA node */
> if (node != NUMA_NO_NODE) {
> @@ -1589,8 +1590,8 @@ static int its_select_cpu(struct irq_data *d,
> * Try the intersection of the affinity mask and the
> * node mask (and the online mask, just to be safe).
> */
> - cpumask_and(tmpmask, cpumask_of_node(node), aff_mask);
> - cpumask_and(tmpmask, tmpmask, cpu_online_mask);
> + cpumask_and(&tmpmask, cpumask_of_node(node), aff_mask);
> + cpumask_and(&tmpmask, &tmpmask, cpu_online_mask);
Consider reducing the churn by writing something like:
static struct cpumask __tmpmask;
struct cpumask *tmpmask = &__tmpmask;
which is strictly equivalent, and makes the patch much smaller.
Thanks,
M.
--
Without deviation from the norm, progress is not possible.
