In order to reduce the impact of the VPT parsing happening on the GIC,
we can split the vcpu reseidency in two phases:
- programming GICR_VPENDBASER: this still happens in vcpu_load()
- checking for the VPT parsing to be complete: this can happen
on vcpu entry (in kvm_vgic_flush_hwstate())
This allows the GIC and the CPU to work in parallel, rewmoving some
of the entry overhead.
Suggested-by: Marc Zyngier <maz@kernel.org>
Signed-off-by: Shenming Lu <lushenming@huawei.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20201128141857.983-3-lushenming@huawei.com
[maz: The GICv3 spec has evolved quite a bit since the draft the Linux
driver was written against, and some register definitions are simply gone]
As per the GICv3 specification, GIC{D,R}_SEIR are not assigned and the
locations (0x0068) are actually Reserved. GICR_MOV{LPI,ALL}R are two IMP
DEF registers and might be defined by some specific micro-architecture.
As they're not used anywhere in the kernel, just drop all of them.
Signed-off-by: Zenghui Yu <yuzenghui@huawei.com>
[maz: added context explaination]
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200630134126.880-1-yuzenghui@huawei.com
When a vPE is made resident, the GIC starts parsing the virtual pending
table to deliver pending interrupts. This takes place asynchronously,
and can at times take a long while. Long enough that the vcpu enters
the guest and hits WFI before any interrupt has been signaled yet.
The vcpu then exits, blocks, and now gets a doorbell. Rince, repeat.
In order to avoid the above, a (optional on GICv4, mandatory on v4.1)
feature allows the GIC to feedback to the hypervisor whether it is
done parsing the VPT by clearing the GICR_VPENDBASER.Dirty bit.
The hypervisor can then wait until the GIC is ready before actually
running the vPE.
Plug the detection code as well as polling on vPE schedule. While
at it, tidy-up the kernel message that displays the GICv4 optional
features.
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
In order to hide some of the differences between v4.0 and v4.1, move
the doorbell management out of the KVM code, and into the GICv4-specific
layer. This allows the calling code to ask for the doorbell when blocking,
and otherwise to leave the doorbell permanently disabled.
This matches the v4.1 code perfectly, and only results in a minor
refactoring of the v4.0 code.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20200304203330.4967-14-maz@kernel.org
Just like for vLPIs, there is some configuration information that cannot
be directly communicated through the normal irqchip API, and we have to
use our good old friend set_vcpu_affinity as a side-band communication
mechanism.
This is used to configure group and priority for a given vSGI.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Link: https://lore.kernel.org/r/20200304203330.4967-13-maz@kernel.org
To implement the get/set_irqchip_state callbacks (limited to the
PENDING state), we have to use a particular set of hacks:
- Reading the pending state is done by using a pair of new redistributor
registers (GICR_VSGIR, GICR_VSGIPENDR), which allow the 16 interrupts
state to be retrieved.
- Setting the pending state is done by generating it as we'd otherwise do
for a guest (writing to GITS_SGIR).
- Clearing the pending state is done by emitting a VSGI command with the
"clear" bit set.
This requires some interesting locking though:
- When talking to the redistributor, we must make sure that the VPE
affinity doesn't change, hence taking the VPE lock.
- At the same time, we must ensure that nobody accesses the same
redistributor's GICR_VSGIR registers for a different VPE, which
would corrupt the reading of the pending bits. We thus take the
per-RD spinlock. Much fun.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20200304203330.4967-12-maz@kernel.org
The GICv4.1 ITS has yet another new command (VSGI) which allows
a VPE-targeted SGI to be configured (or have its pending state
cleared). Add support for this command and plumb it into the
activate irqdomain callback so that it is ready to be used.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20200304203330.4967-10-maz@kernel.org
Since GICv4.1 has the capability to inject 16 SGIs into each VPE,
and that I'm keen not to invent too many specific interfaces to
manipulate these interrupts, let's pretend that each of these SGIs
is an actual Linux interrupt.
For that matter, let's introduce a minimal irqchip and irqdomain
setup that will get fleshed up in the following patches.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Link: https://lore.kernel.org/r/20200304203330.4967-9-maz@kernel.org
There is no special reason to set virtual LPI pending table as
non-shareable. If we choose to hard code the shareability without
probing, Inner-Shareable is likely to be a better choice, as the
VPEs can move around and benefit from having the redistributors
snooping each other's cache, if that's something they can do.
Furthermore, Hisilicon hip08 ends up with unspecified errors when
mixing shareability attributes. So let's move to IS attributes for
the VPT. This has also been tested on D05 and didn't show any
regression.
Signed-off-by: Heyi Guo <guoheyi@huawei.com>
[maz: rewrote commit message]
Signed-off-by: Marc Zyngier <maz@kernel.org>
Tested-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20191130073849.38378-1-guoheyi@huawei.com
The GICv4.1 spec says that it is CONTRAINED UNPREDICTABLE to write to
any of the GICR_INV{LPI,ALL}R registers if GICR_SYNCR.Busy == 1.
To deal with it, we must ensure that only a single invalidation can
happen at a time for a given redistributor. Add a per-RD lock to that
effect and take it around the invalidation/syncr-read to deal with this.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Link: https://lore.kernel.org/r/20200304203330.4967-6-maz@kernel.org
Before GICv4.1, all operations would be serialized with the affinity
changes by virtue of using the same ITS command queue. With v4.1, things
change, as invalidations (and a number of other operations) are issued
using the redistributor MMIO frame.
We must thus make sure that these redistributor accesses cannot race
against aginst the affinity change, or we may end-up talking to the
wrong redistributor.
To ensure this, we expand the irq_to_cpuid() helper to take a spinlock
when the LPI is mapped to a vLPI (a new per-VPE lock) on each operation
that requires mutual exclusion.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20200304203330.4967-4-maz@kernel.org
To allow the direct injection of SGIs into a guest, the GICv4.1
architecture has to sacrifice the Active state so that SGIs look
a lot like LPIs (they are injected by the same mechanism).
In order not to break existing software, the architecture gives
offers guests OSs the choice: SGIs with or without an active
state. It is the hypervisors duty to honor the guest's choice.
For this, the architecture offers a discovery bit indicating whether
the GIC supports GICv4.1 SGIs (GICD_TYPER2.nASSGIcap), and another
bit indicating whether the guest wants Active-less SGIs or not
(controlled by GICD_CTLR.nASSGIreq).
A hypervisor not supporting GICv4.1 SGIs would leave nASSGIcap
clear, and a guest not knowing about GICv4.1 SGIs (or definitely
wanting an Active state) would leave nASSGIreq clear (both being
thankfully backward compatible with older revisions of the GIC).
Since Linux is perfectly happy without an active state on SGIs,
inform the hypervisor that we'll use that if offered.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20200304203330.4967-2-maz@kernel.org
Currently, we will not set vpe_l1_page for the current RD if we can
inherit the vPE configuration table from another RD (or ITS), which
results in an inconsistency between RDs within the same CommonLPIAff
group.
Let's rename it to vpe_l1_base to indicate the base address of the
vPE configuration table of this RD, and set it properly for *all*
v4.1 redistributors.
Signed-off-by: Zenghui Yu <yuzenghui@huawei.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200206075711.1275-3-yuzenghui@huawei.com
