8:32 a.m.
Hi,
I can confirm that hooking in generic_show_interrupts() and
generic_get_irq_affinity() results in meaningful output for ARM as
well, using PATCH v3.
I've got another suggestion however. How about mimicking the kernel
code a little closer?
fs/proc/interrupts.c
will keep going for one more iteration, i.e. while <=nr_irqs and
kernel/irq/proc.c:show_interrupts()
will call arch_show_interrupts() when called with ==nr_irqs.
In Crash we could do this in cmd_irq
for (i = 0; i <= nr_irqs; i++)
machdep->show_interrupts(i, cpus);
and this at the start of generic_show_interrupts():
if (irq >= machdep->nr_irqs) {
return;
}
making it possible to have machdep->show_interrupts point to a wrapper
function like:
static void arm_show_interrupts(int irq, ulong *cpus)
{
if (irq == machdep->nr_irqs) {
// Show FIQ
// Show IPI
}
else {
generic_show_interrupts(irq, cpus);
}
}
Just an idea.
Regards,
Per
On Wed, Jan 18, 2012 at 9:11 PM, Dave Anderson <anderson(a)redhat.com> wrote:
----- Original Message -----
> Hello, Dave
>
> Re-posting two patches as attachments. Very sorry for incorrectly
> changing the behaviour of the original "irq" command in the last
> patches.
>
> Thanks,
> Zhang Yanfei
I should first note that I would have preferred that the patch were
written to address the other actively-supported architectures, given
that the irq_data and irq_desc structures are not x86/x86_64-specific.
However, I understand that you may not have access to the other
architectures. That being the case, I invite the s390x, ia64, arm,
ppc32 and ppc64 maintainers to look at this patch, and if you are
interested, please apply it for those architectures as well.
Anyway, I made a number of changes to the patch as written.
(1) I changed the "irq" option letters to something more meaningful:
(a) For IRQ affinity, I changed it from "irq -s" to "irq -a".
(b) For IRQ stats, I changed it from "irq -S" to "irq -s".
(2) In generic_show_interrupts(), I changed the output so that it
does not append additional unnecessary spaces to the end of
the output line, i.e., from this:
fprintf(fp, " %-20s ", name_buf);
fprintf(fp, "\n");
to simply this:
fprintf(fp, " %s\n", name_buf);
(3) In cmd_irq(), I changed the two new command_not_supported() calls
to be option_not_supported(c).
(4) The "irq -s" option fails on earlier kernels, for example, with
the RHEL4 kernel:
crash> irq -s
IRQ NAME AFFINITY
irq: invalid structure member offset: irq_desc_t_affinity
FILE: kernel.c LINE: 5543 FUNCTION: generic_get_irq_affinity()
[/home/anderson/crash-6.0.2/crash] error trace: 459c05 => 4bf432 => 4c1d31
=> 4fb70e
4fb70e: OFFSET_verify+222
4c1d31: generic_get_irq_affinity+1153
4bf432: cmd_irq+530
459c05: exec_command+821
irq: invalid structure member offset: irq_desc_t_affinity
The "irq -s" option would also fail on recent kernels if they are
not configured with CONFIG_SMP, because irq_data.affinity
member would not exist:
struct irq_data {
unsigned int irq;
unsigned long hwirq;
unsigned int node;
unsigned int state_use_accessors;
struct irq_chip *chip;
struct irq_domain *domain;
void *handler_data;
void *chip_data;
struct msi_desc *msi_desc;
#ifdef CONFIG_SMP
cpumask_var_t affinity;
#endif
};
To handle those two possibilities, I added this to cmd_irq()
for the "-a" option:
if (VALID_STRUCT(irq_data)) {
if (INVALID_MEMBER(irq_data_affinity))
option_not_supported(c);
} else if (INVALID_MEMBER(irq_desc_t_affinity))
option_not_supported(c);
(5) In x86.c, it makes no sense to set machdep->get_irq_affinity
and machdep->show_interrupts in x86_init_hyper() given that the
"irq" command is not supported for the Xen hypervisor analysis.
(See the xen_hyper_command_table[] in xen_hyper_global_data.c)
(6) In defs.h, I moved:
(a) get_irq_affinity and show_interrupts to the end of the
machdep_table structure
(b) irq_desc_t_irq_data, irq_desc_t_kstat_irqs, irq_desc_t_affinity,
irq_data_chip, irq_data_affinity and kernel_stat_irqs to the
end of the offset_table structure.
(c) kernel_stat to the end of the size_table structure.
This would allow an extension module built with an earlier version
of the crash utility to still work without recompiling due to
structure member offset changes.
(7) I updated dump_offset_table() to show the new members of the
offset_table and the size_table for the "help -o" command.
(8) And lastly, I reworked the "help irq" page.
I will continue testing it on various kernels and architectures,
and if all goes well, I will queue it for crash-6.0.3.
Dave
--
Crash-utility mailing list
Crash-utility(a)redhat.com
https://www.redhat.com/mailman/listinfo/crash-utility