Re: [PATCH v3 3/3] xen/sched: fix cpu hotplug

[Andrew Cooper <Andrew.Cooper3@xxxxxxxxxx>]

On 16/08/2022 11:13, Juergen Gross wrote:
> Cpu cpu unplugging is calling schedule_cpu_rm() via stop_machine_run()

Cpu cpu.

> with interrupts disabled, thus any memory allocation or freeing must
> be avoided.
>
> Since commit 5047cd1d5dea ("xen/common: Use enhanced
> ASSERT_ALLOC_CONTEXT in xmalloc()") this restriction is being enforced
> via an assertion, which will now fail.
>
> Before that commit cpu unplugging in normal configurations was working
> just by chance as only the cpu performing schedule_cpu_rm() was doing
> active work. With core scheduling enabled, however, failures could
> result from memory allocations not being properly propagated to other
> cpus' TLBs.

This isn't accurate, is it?  The problem with initiating a TLB flush
with IRQs disabled is that you can deadlock against a remote CPU which
is waiting for you to enable IRQs first to take a TLB flush IPI.

How does a memory allocation out of the xenheap result in a TLB flush? 
Even with split heaps, you're only potentially allocating into a new
slot which was unused...

> diff --git a/xen/common/sched/core.c b/xen/common/sched/core.c
> index 228470ac41..ffb2d6202b 100644
> --- a/xen/common/sched/core.c
> +++ b/xen/common/sched/core.c
> @@ -3260,6 +3260,17 @@ static struct cpu_rm_data 
> *schedule_cpu_rm_alloc(unsigned int cpu)
>      if ( !data )
>          goto out;
>  
> +    if ( aff_alloc )
> +    {
> +        if ( !update_node_aff_alloc(&data->affinity) )

I spent ages trying to figure out what this was doing, before realising
the problem is the function name.

alloc (as with free) is the critical piece of information and needs to
come first.  The fact we typically pass the result to
update_node_aff(inity) isn't relevant, and becomes actively wrong here
when we're nowhere near.

Patch 1 needs to name these helpers:

bool alloc_affinity_masks(struct affinity_masks *affinity);
void free_affinity_masks(struct affinity_masks *affinity);

and then patches 2 and 3 become far easier to follow.

Similarly in patch 2, the new helpers need to be
{alloc,free}_cpu_rm_data() to make sense.  These have nothing to do with
scheduling.

Also, you shouldn't introduce the helpers static in patch 2 and then
turn them non-static in patch 3.  That just adds unnecessary churn to
the complicated patch.

> +        {
> +            XFREE(data);
> +            goto out;
> +        }
> +    }
> +    else
> +        memset(&data->affinity, 0, sizeof(data->affinity));

I honestly don't think it is worth optimising xzalloc() -> xmalloc() 
for the cognitive complexity of having this logic here.

> diff --git a/xen/common/sched/cpupool.c b/xen/common/sched/cpupool.c
> index 58e082eb4c..2506861e4f 100644
> --- a/xen/common/sched/cpupool.c
> +++ b/xen/common/sched/cpupool.c
> @@ -411,22 +411,28 @@ int cpupool_move_domain(struct domain *d, struct 
> cpupool *c)
>  }
>  
>  /* Update affinities of all domains in a cpupool. */
> -static void cpupool_update_node_affinity(const struct cpupool *c)
> +static void cpupool_update_node_affinity(const struct cpupool *c,
> +                                         struct affinity_masks *masks)
>  {
> -    struct affinity_masks masks;
> +    struct affinity_masks local_masks;
>      struct domain *d;
>  
> -    if ( !update_node_aff_alloc(&masks) )
> -        return;
> +    if ( !masks )
> +    {
> +        if ( !update_node_aff_alloc(&local_masks) )
> +            return;
> +        masks = &local_masks;
> +    }
>  
>      rcu_read_lock(&domlist_read_lock);
>  
>      for_each_domain_in_cpupool(d, c)
> -        domain_update_node_aff(d, &masks);
> +        domain_update_node_aff(d, masks);
>  
>      rcu_read_unlock(&domlist_read_lock);
>  
> -    update_node_aff_free(&masks);
> +    if ( masks == &local_masks )
> +        update_node_aff_free(masks);
>  }
>  
>  /*

Why do we need this at all?  domain_update_node_aff() already knows what
to do when passed NULL, so this seems like an awfully complicated no-op.

> @@ -1008,10 +1016,21 @@ static int cf_check cpu_callback(
>  {
>      unsigned int cpu = (unsigned long)hcpu;
>      int rc = 0;
> +    static struct cpu_rm_data *mem;
>  
>      switch ( action )
>      {
>      case CPU_DOWN_FAILED:
> +        if ( system_state <= SYS_STATE_active )
> +        {
> +            if ( mem )
> +            {

So, this does compile (and indeed I've tested the result), but I can't
see how it should.

mem is guaranteed to be uninitialised at this point, and ...

> +                schedule_cpu_rm_free(mem, cpu);
> +                mem = NULL;
> +            }
> +            rc = cpupool_cpu_add(cpu);
> +        }
> +        break;
>      case CPU_ONLINE:
>          if ( system_state <= SYS_STATE_active )
>              rc = cpupool_cpu_add(cpu);
> @@ -1019,12 +1038,31 @@ static int cf_check cpu_callback(
>      case CPU_DOWN_PREPARE:
>          /* Suspend/Resume don't change assignments of cpus to cpupools. */
>          if ( system_state <= SYS_STATE_active )
> +        {
>              rc = cpupool_cpu_remove_prologue(cpu);
> +            if ( !rc )
> +            {
> +                ASSERT(!mem);

... here, and each subsequent assertion too.

Given that I tested the patch and it does fix the IRQ assertion, I can
only imagine that it works by deterministically finding stack rubble
which happens to be 0.

~Andrew