Le 24/03/2018 à 09:14, Benjamin Herrenschmidt a écrit :
On Fri, 2018-03-23 at 17:17 +0100, christophe lombard wrote:
Le 23/03/2018 à 03:14, Benjamin Herrenschmidt a écrit :
On Thu, 2018-03-22 at 17:37 +0100, Christophe Lombard wrote:
The cxl driver cannot disable the interrupt at the device level and has
to use disable_irq[_nosync] instead.
To avoid the implementation of the lazy optimisation (the interrupt is
marked disabled, but the hardware is left unmasked), we can disable it,
for a particular irq line, by calling
'irq_set_status_flags(irq, IRQ_DISABLE_UNLAZY)'.
Why do you need that ? What's wrong with the lazy approach ? It makes
disable_irq/enable_irq faster...
You shouldn't need that unless your device is generating a *LOT* of
irqs while disabled.
An issue on POWERVM (CAPI) has been introduced with the following patch
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=bf22ff45bed664aefb5c4e43029057a199b7070c
The PSL or AFU interrupts are never received by the cxl driver because
the interrupts are never unmasked.
Without this patch (genirq: Avoid unnecessary low level irq function
calls), the callback desc->irq_data.chip->irq_unmask(&desc->irq_data);
(= ics_rtas_unmask_irq()) is called by default through irq_enable().
I don't see why this would change with the patch...
The cxl driver disables the interrupts before attaching the process
element and enables the interrupts after that.
How ? Using disable_irq or something else ?
In the current code, irq_enable() unmasks the irq only if the irq state
is IRQD_IRQ_MASKED but it does not.
Sorry I don't really parse your sentence. You mean there's a disconnect
between the "HW" (or pHyp) state of the interrupt and the
IRQD_IRQ_MASKED flag ? That shouldn't happen... if that's the case the
bug is elsewhere, what is causing the disconnect in the first place ?
Call irq_set_status_flags(irq, IRQ_DISABLE_UNLAZY) allows forcing
irq_disable() to update the irq state to IRQD_IRQ_MASKED and by default
irq_enable() will unmask the irq through ics_rtas_unmask_irq().
Hrm. I don't quite understand. You shouldn't need that, I suspect you
are papering over a different bug but I'm not 100% certain as I don't
completely understand what's happening.
Cheers,
Ben.
Hi Ben,
To answer to your questions, here is the timeline in the cxl driver
1. call disable_irq()
2. call plpar_hcall9() to attach a process element
During this phase, phyp (as described in CAPI PAPR document)
disables the virtual interrupts provided in the process-element-
struct.
Then the partition must use ibm,set-xive to enable the virtual
interrupt source after H_ATTACH_CA_PROCESS completes
successfully.
3. call enable_irq()
Before the following patch (genirq: Avoid unnecessary low level
irq function calls) be pushed in mainline, the unmask api
(= ics_rtas_unmask_irq()) was called by default and everything
was ok.
With this patch, the unmask api is now called only if
IRQD_IRQ_MASKED is set and in our case, it's never done and the
irqs remain masked.
So, there is a disconnect between the 'HW' (phyp) state of the
interrupt and the IRQD_IRQ_MASKED flag.
In the timeline of the cxl driver, the IRQD_IRQ_MASKED flag is never
set.
This flag was not used in enable_irq(), before that the patch "genirq:
Avoid ..." was pushed. So ics_rtas_unmask_irq() was called and
everything was ok.
Thanks
Signed-off-by: Christophe Lombard <clomb...@linux.vnet.ibm.com>
---
drivers/misc/cxl/guest.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/drivers/misc/cxl/guest.c b/drivers/misc/cxl/guest.c
index f58b4b6c..dc476e1 100644
--- a/drivers/misc/cxl/guest.c
+++ b/drivers/misc/cxl/guest.c
@@ -389,6 +389,7 @@ static void disable_afu_irqs(struct cxl_context *ctx)
hwirq = ctx->irqs.offset[r];
for (i = 0; i < ctx->irqs.range[r]; hwirq++, i++) {
virq = irq_find_mapping(NULL, hwirq);
+ irq_set_status_flags(virq, IRQ_DISABLE_UNLAZY);
disable_irq(virq);
}
}