Do a braindump of the way things are supposed to work.
Reviewed-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
Reviewed-by: Eric Auger <eric.auger@xxxxxxxxxx>
Signed-off-by: Marc Zyngier <marc.zyngier@xxxxxxx>
---
drivers/irqchip/irq-gic-v4.c | 71 ++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 71 insertions(+)
diff --git a/drivers/irqchip/irq-gic-v4.c b/drivers/irqchip/irq-gic-v4.c
index 6fe2e09172ed..d10d73477d01 100644
--- a/drivers/irqchip/irq-gic-v4.c
+++ b/drivers/irqchip/irq-gic-v4.c
@@ -22,6 +22,77 @@
#include <linux/irqchip/arm-gic-v4.h>
+/*
+ * WARNING: The blurb below assumes that you understand the
+ * intricacies of GICv3, GICv4, and how a guest's view of a GICv3 gets
+ * translated into GICv4 commands. So it effectively targets at most
+ * two individuals. You know who you are.
+ *
+ * The core GICv4 code is designed to *avoid* exposing too much of the
+ * core GIC code (that would in turn leak into the hypervisor code),
+ * and instead provide a hypervisor agnostic interface to the HW (of
+ * course, the astute reader will quickly realize that hypervisor
+ * agnostic actually means KVM-specific - what were you thinking?).
+ *
+ * In order to achieve a modicum of isolation, we try to hide most of
+ * the GICv4 "stuff" behind normal irqchip operations:
+ *
+ * - Any guest-visible VLPI is backed by a Linux interrupt (and a
+ * physical LPI which gets unmapped when the guest maps the
+ * VLPI). This allows the same DevID/EventID pair to be either
+ * mapped to the LPI (host) or the VLPI (guest). Note that this is
+ * exclusive, and you cannot have both.
+ *
+ * - Enabling/disabling a VLPI is done by issuing mask/unmask calls.
+ *
+ * - Guest INT/CLEAR commands are implemented through
+ * irq_set_irqchip_state().
+ *
+ * - The *bizarre* stuff (mapping/unmapping an interrupt to a VLPI, or
+ * issuing an INV after changing a priority) gets shoved into the
+ * irq_set_vcpu_affinity() method. While this is quite horrible
+ * (let's face it, this is the irqchip version of an ioctl), it
+ * confines the crap to a single location. And map/unmap really is
+ * about setting the affinity of a VLPI to a vcpu, so only INV is
+ * majorly out of place. So there.
+ *
+ * A number of commands are simply not provided by this interface, as
+ * they do not make direct sense. For example, MAPD is purely local to
+ * the virtual ITS (because it references a virtual device, and the
+ * physical ITS is still very much in charge of the physical
+ * device). Same goes for things like MAPC (the physical ITS deals
+ * with the actual vPE affinity, and not the braindead concept of
+ * collection). SYNC is not provided either, as each and every command
+ * is followed by a VSYNC. This could be relaxed in the future, should
+ * this be seen as a bottleneck (yes, this means *never*).
+ *
+ * But handling VLPIs is only one side of the job of the GICv4
+ * code. The other (darker) side is to take care of the doorbell
+ * interrupts which are delivered when a VLPI targeting a non-running
+ * vcpu is being made pending.
+ *
+ * The choice made here is that each vcpu (VPE in old northern GICv4
+ * dialect) gets a single doorbell LPI, no matter how many interrupts
+ * are targeting it. This has a nice property, which is that the
+ * interrupt becomes a handle for the VPE, and that the hypervisor
+ * code can manipulate it through the normal interrupt API:
+ *
+ * - VMs (or rather the VM abstraction that matters to the GIC)
+ * contain an irq domain where each interrupt maps to a VPE. In
+ * turn, this domain sits on top of the normal LPI allocator, and a
+ * specially crafted irq_chip implementation.
+ *
+ * - mask/unmask do what is expected on the doorbell interrupt.
+ *
+ * - irq_set_affinity is used to move a VPE from one redistributor to
+ * another.
+ *
+ * - irq_set_vcpu_affinity once again gets hijacked for the purpose of
+ * creating a new sub-API, namely scheduling/descheduling a VPE
+ * (which involves programming GICR_V{PROP,PEND}BASER) and
+ * performing INVALL operations.
+ */
+
static struct irq_domain *gic_domain;
static const struct irq_domain_ops *vpe_domain_ops;
--
2.11.0
