[PATCH v3 22/25] irq_domain/x86: Convert x86 (embedded) to use common irq_domain

[Grant Likely]

On Sat, Jan 28, 2012 at 5:35 PM, Grant Likely <grant.likely at secretlab.ca> wrote:
> On Sat, Jan 28, 2012 at 05:44:05PM +0100, Sebastian Andrzej Siewior wrote:
>> * Grant Likely | 2012-01-27 14:36:16 [-0700]:
>>
>> >This patch removes the x86-specific definition of irq_domain and replaces
>> >it with the common implementation.
>>
>> I pulled your devicetree/next tree. After this patch I get:
>>
>> |Hierarchical RCU implementation.
>> |NR_IRQS:2304 nr_irqs:256 16
>> |------------[ cut here ]------------
>> |WARNING: at /home/bigeasy/work/shiva/git/linux-2.6-tip/kernel/irq/irqdomain.c:114 irq_domain_add_legacy+0x75/0x150()
>> |Modules linked in:
>> |Pid: 0, comm: swapper/0 Not tainted 3.3.0-rc1+ #65
>> |Call Trace:
>> | [<c15044e0>] ? printk+0x18/0x1a
>> | [<c102cdbd>] warn_slowpath_common+0x6d/0xa0
>> | [<c1095575>] ? irq_domain_add_legacy+0x75/0x150
>> | [<c1095575>] ? irq_domain_add_legacy+0x75/0x150
>> | [<c102ce0d>] warn_slowpath_null+0x1d/0x20
>> | [<c1095575>] irq_domain_add_legacy+0x75/0x150
>> | [<c1714824>] x86_add_irq_domains+0x96/0xd6
>> | [<c1708df2>] init_IRQ+0x8/0x33
>> | [<c170557f>] start_kernel+0x191/0x2e1
>> | [<c170517f>] ? loglevel+0x2b/0x2b
>> | [<c1705081>] i386_start_kernel+0x81/0x86
>> |---[ end trace 4eaa2a86a8e2da22 ]---
>> |------------[ cut here ]------------
>> |kernel BUG at /home/bigeasy/work/shiva/git/linux-2.6-tip/arch/x86/kernel/devicetree.c:367!
>>
>> The warning is comming from this piece in irq_domain_add_legacy()
>> |for (i = 0; i < size; i++) {
>> |                 int irq = first_irq + i;
>> |                 struct irq_data *irq_data = irq_get_irq_data(irq);
>> |
>> |                 if (WARN_ON(!irq_data || irq_data->domain)) {
>>
>> irq_data is NULL here.
>>
>> |                         mutex_unlock(&irq_domain_mutex);
>> |                         of_node_put(domain->of_node);
>> |                         kfree(domain);
>> |                         return NULL;
>> |                 }
>> |         }
>> |
>>
>> This is not always the case. arch_early_irq_init() in [0] sets up the
>> first 16 entries. The reminaing few (there is a toal of 24 irqs for
>> first ioapic and a second ioapic) are not initialized. This happens
>> later via ->xlate, ioapic_xlate() => io_apic_setup_irq_pin() =>
>> alloc_irq_and_cfg_at() calls irq_set_chip_data() on demand.
>>
>> [0] arch/x86/kernel/apic/io_apic.c
>
> That's not going to work then.  The irqs must not be set up in ->xlate.
> They need to be set up either before creating the irq_domain, or in the
> .map hook.  Inside the map hook is preferred, but having some
> configured and some not at initialization time does make it difficult

Actually, that's still not right.  irq_set_chip_data() (sets
desc->irq_data.chip_data) has nothing to do with irq_get_irq_data()
(which returns desc->irq_data).  It's the   allocation of the irq_desc
that is at issue here.  Where does the allocation occur?  (I haven't
dug in to find it yet)

g.