linux/drivers/irqchip/irq-gic-v4.c
Marc Zyngier 8e01d9a396 KVM: arm64: vgic-v4: Move the GICv4 residency flow to be driven by vcpu_load/put
When the VHE code was reworked, a lot of the vgic stuff was moved around,
but the GICv4 residency code did stay untouched, meaning that we come
in and out of residency on each flush/sync, which is obviously suboptimal.

To address this, let's move things around a bit:

- Residency entry (flush) moves to vcpu_load
- Residency exit (sync) moves to vcpu_put
- On blocking (entry to WFI), we "put"
- On unblocking (exit from WFI), we "load"

Because these can nest (load/block/put/load/unblock/put, for example),
we now have per-VPE tracking of the residency state.

Additionally, vgic_v4_put gains a "need doorbell" parameter, which only
gets set to true when blocking because of a WFI. This allows a finer
control of the doorbell, which now also gets disabled as soon as
it gets signaled.

Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20191027144234.8395-2-maz@kernel.org
2019-10-28 16:20:58 +00:00

231 lines
6.5 KiB
C

// SPDX-License-Identifier: GPL-2.0-only
/*
* Copyright (C) 2016,2017 ARM Limited, All Rights Reserved.
* Author: Marc Zyngier <marc.zyngier@arm.com>
*/
#include <linux/interrupt.h>
#include <linux/irq.h>
#include <linux/irqdomain.h>
#include <linux/msi.h>
#include <linux/sched.h>
#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;
int its_alloc_vcpu_irqs(struct its_vm *vm)
{
int vpe_base_irq, i;
vm->fwnode = irq_domain_alloc_named_id_fwnode("GICv4-vpe",
task_pid_nr(current));
if (!vm->fwnode)
goto err;
vm->domain = irq_domain_create_hierarchy(gic_domain, 0, vm->nr_vpes,
vm->fwnode, vpe_domain_ops,
vm);
if (!vm->domain)
goto err;
for (i = 0; i < vm->nr_vpes; i++) {
vm->vpes[i]->its_vm = vm;
vm->vpes[i]->idai = true;
}
vpe_base_irq = __irq_domain_alloc_irqs(vm->domain, -1, vm->nr_vpes,
NUMA_NO_NODE, vm,
false, NULL);
if (vpe_base_irq <= 0)
goto err;
for (i = 0; i < vm->nr_vpes; i++)
vm->vpes[i]->irq = vpe_base_irq + i;
return 0;
err:
if (vm->domain)
irq_domain_remove(vm->domain);
if (vm->fwnode)
irq_domain_free_fwnode(vm->fwnode);
return -ENOMEM;
}
void its_free_vcpu_irqs(struct its_vm *vm)
{
irq_domain_free_irqs(vm->vpes[0]->irq, vm->nr_vpes);
irq_domain_remove(vm->domain);
irq_domain_free_fwnode(vm->fwnode);
}
static int its_send_vpe_cmd(struct its_vpe *vpe, struct its_cmd_info *info)
{
return irq_set_vcpu_affinity(vpe->irq, info);
}
int its_schedule_vpe(struct its_vpe *vpe, bool on)
{
struct its_cmd_info info;
int ret;
WARN_ON(preemptible());
info.cmd_type = on ? SCHEDULE_VPE : DESCHEDULE_VPE;
ret = its_send_vpe_cmd(vpe, &info);
if (!ret)
vpe->resident = on;
return ret;
}
int its_invall_vpe(struct its_vpe *vpe)
{
struct its_cmd_info info = {
.cmd_type = INVALL_VPE,
};
return its_send_vpe_cmd(vpe, &info);
}
int its_map_vlpi(int irq, struct its_vlpi_map *map)
{
struct its_cmd_info info = {
.cmd_type = MAP_VLPI,
{
.map = map,
},
};
int ret;
/*
* The host will never see that interrupt firing again, so it
* is vital that we don't do any lazy masking.
*/
irq_set_status_flags(irq, IRQ_DISABLE_UNLAZY);
ret = irq_set_vcpu_affinity(irq, &info);
if (ret)
irq_clear_status_flags(irq, IRQ_DISABLE_UNLAZY);
return ret;
}
int its_get_vlpi(int irq, struct its_vlpi_map *map)
{
struct its_cmd_info info = {
.cmd_type = GET_VLPI,
{
.map = map,
},
};
return irq_set_vcpu_affinity(irq, &info);
}
int its_unmap_vlpi(int irq)
{
irq_clear_status_flags(irq, IRQ_DISABLE_UNLAZY);
return irq_set_vcpu_affinity(irq, NULL);
}
int its_prop_update_vlpi(int irq, u8 config, bool inv)
{
struct its_cmd_info info = {
.cmd_type = inv ? PROP_UPDATE_AND_INV_VLPI : PROP_UPDATE_VLPI,
{
.config = config,
},
};
return irq_set_vcpu_affinity(irq, &info);
}
int its_init_v4(struct irq_domain *domain, const struct irq_domain_ops *ops)
{
if (domain) {
pr_info("ITS: Enabling GICv4 support\n");
gic_domain = domain;
vpe_domain_ops = ops;
return 0;
}
pr_err("ITS: No GICv4 VPE domain allocated\n");
return -ENODEV;
}